linux/mm/page_alloc.c

5822 lines
160 KiB
C
Raw Normal View History

/*
* linux/mm/page_alloc.c
*
* Manages the free list, the system allocates free pages here.
* Note that kmalloc() lives in slab.c
*
* Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
* Swap reorganised 29.12.95, Stephen Tweedie
* Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
* Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
* Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
* Zone balancing, Kanoj Sarcar, SGI, Jan 2000
* Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
* (lots of bits borrowed from Ingo Molnar & Andrew Morton)
*/
#include <linux/stddef.h>
#include <linux/mm.h>
#include <linux/swap.h>
#include <linux/interrupt.h>
#include <linux/pagemap.h>
#include <linux/jiffies.h>
#include <linux/bootmem.h>
#include <linux/memblock.h>
#include <linux/compiler.h>
#include <linux/kernel.h>
#include <linux/kmemcheck.h>
#include <linux/module.h>
#include <linux/suspend.h>
#include <linux/pagevec.h>
#include <linux/blkdev.h>
#include <linux/slab.h>
#include <linux/ratelimit.h>
#include <linux/oom.h>
#include <linux/notifier.h>
#include <linux/topology.h>
#include <linux/sysctl.h>
#include <linux/cpu.h>
#include <linux/cpuset.h>
#include <linux/memory_hotplug.h>
#include <linux/nodemask.h>
#include <linux/vmalloc.h>
#include <linux/vmstat.h>
#include <linux/mempolicy.h>
#include <linux/stop_machine.h>
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#include <linux/sort.h>
#include <linux/pfn.h>
#include <linux/backing-dev.h>
#include <linux/fault-inject.h>
#include <linux/page-isolation.h>
#include <linux/page_cgroup.h>
infrastructure to debug (dynamic) objects We can see an ever repeating problem pattern with objects of any kind in the kernel: 1) freeing of active objects 2) reinitialization of active objects Both problems can be hard to debug because the crash happens at a point where we have no chance to decode the root cause anymore. One problem spot are kernel timers, where the detection of the problem often happens in interrupt context and usually causes the machine to panic. While working on a timer related bug report I had to hack specialized code into the timer subsystem to get a reasonable hint for the root cause. This debug hack was fine for temporary use, but far from a mergeable solution due to the intrusiveness into the timer code. The code further lacked the ability to detect and report the root cause instantly and keep the system operational. Keeping the system operational is important to get hold of the debug information without special debugging aids like serial consoles and special knowledge of the bug reporter. The problems described above are not restricted to timers, but timers tend to expose it usually in a full system crash. Other objects are less explosive, but the symptoms caused by such mistakes can be even harder to debug. Instead of creating specialized debugging code for the timer subsystem a generic infrastructure is created which allows developers to verify their code and provides an easy to enable debug facility for users in case of trouble. The debugobjects core code keeps track of operations on static and dynamic objects by inserting them into a hashed list and sanity checking them on object operations and provides additional checks whenever kernel memory is freed. The tracked object operations are: - initializing an object - adding an object to a subsystem list - deleting an object from a subsystem list Each operation is sanity checked before the operation is executed and the subsystem specific code can provide a fixup function which allows to prevent the damage of the operation. When the sanity check triggers a warning message and a stack trace is printed. The list of operations can be extended if the need arises. For now it's limited to the requirements of the first user (timers). The core code enqueues the objects into hash buckets. The hash index is generated from the address of the object to simplify the lookup for the check on kfree/vfree. Each bucket has it's own spinlock to avoid contention on a global lock. The debug code can be compiled in without being active. The runtime overhead is minimal and could be optimized by asm alternatives. A kernel command line option enables the debugging code. Thanks to Ingo Molnar for review, suggestions and cleanup patches. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu> Cc: Greg KH <greg@kroah.com> Cc: Randy Dunlap <randy.dunlap@oracle.com> Cc: Kay Sievers <kay.sievers@vrfy.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 07:55:01 +00:00
#include <linux/debugobjects.h>
#include <linux/kmemleak.h>
#include <linux/memory.h>
#include <linux/compaction.h>
tracing, page-allocator: add trace event for page traffic related to the buddy lists The page allocation trace event reports that a page was successfully allocated but it does not specify where it came from. When analysing performance, it can be important to distinguish between pages coming from the per-cpu allocator and pages coming from the buddy lists as the latter requires the zone lock to the taken and more data structures to be examined. This patch adds a trace event for __rmqueue reporting when a page is being allocated from the buddy lists. It distinguishes between being called to refill the per-cpu lists or whether it is a high-order allocation. Similarly, this patch adds an event to catch when the PCP lists are being drained a little and pages are going back to the buddy lists. This is trickier to draw conclusions from but high activity on those events could explain why there were a large number of cache misses on a page-allocator-intensive workload. The coalescing and splitting of buddies involves a lot of writing of page metadata and cache line bounces not to mention the acquisition of an interrupt-safe lock necessary to enter this path. [akpm@linux-foundation.org: fix build] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Ingo Molnar <mingo@elte.hu> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Li Ming Chun <macli@brc.ubc.ca> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:02:44 +00:00
#include <trace/events/kmem.h>
#include <linux/ftrace_event.h>
#include <linux/memcontrol.h>
#include <linux/prefetch.h>
#include <asm/tlbflush.h>
#include <asm/div64.h>
#include "internal.h"
numa: add generic percpu var numa_node_id() implementation Rework the generic version of the numa_node_id() function to use the new generic percpu variable infrastructure. Guard the new implementation with a new config option: CONFIG_USE_PERCPU_NUMA_NODE_ID. Archs which support this new implemention will default this option to 'y' when NUMA is configured. This config option could be removed if/when all archs switch over to the generic percpu implementation of numa_node_id(). Arch support involves: 1) converting any existing per cpu variable implementations to use this implementation. x86_64 is an instance of such an arch. 2) archs that don't use a per cpu variable for numa_node_id() will need to initialize the new per cpu variable "numa_node" as cpus are brought on-line. ia64 is an example. 3) Defining USE_PERCPU_NUMA_NODE_ID in arch dependent Kconfig--e.g., when NUMA is configured. This is required because I have retained the old implementation by default to allow archs to be modified incrementally, as desired. Subsequent patches will convert x86_64 and ia64 to use this implemenation. Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Tejun Heo <tj@kernel.org> Cc: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Cc: Nick Piggin <npiggin@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Eric Whitney <eric.whitney@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-26 21:44:56 +00:00
#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
DEFINE_PER_CPU(int, numa_node);
EXPORT_PER_CPU_SYMBOL(numa_node);
#endif
numa: introduce numa_mem_id()- effective local memory node id Introduce numa_mem_id(), based on generic percpu variable infrastructure to track "nearest node with memory" for archs that support memoryless nodes. Define API in <linux/topology.h> when CONFIG_HAVE_MEMORYLESS_NODES defined, else stubs. Architectures will define HAVE_MEMORYLESS_NODES if/when they support them. Archs can override definitions of: numa_mem_id() - returns node number of "local memory" node set_numa_mem() - initialize [this cpus'] per cpu variable 'numa_mem' cpu_to_mem() - return numa_mem for specified cpu; may be used as lvalue Generic initialization of 'numa_mem' occurs in __build_all_zonelists(). This will initialize the boot cpu at boot time, and all cpus on change of numa_zonelist_order, or when node or memory hot-plug requires zonelist rebuild. Archs that support memoryless nodes will need to initialize 'numa_mem' for secondary cpus as they're brought on-line. [akpm@linux-foundation.org: fix build] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Christoph Lameter <cl@linux-foundation.org> Cc: Tejun Heo <tj@kernel.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Nick Piggin <npiggin@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Eric Whitney <eric.whitney@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-26 21:45:00 +00:00
#ifdef CONFIG_HAVE_MEMORYLESS_NODES
/*
* N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
* It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
* Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
* defined in <linux/topology.h>.
*/
DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
EXPORT_PER_CPU_SYMBOL(_numa_mem_);
#endif
/*
Memoryless nodes: Generic management of nodemasks for various purposes Why do we need to support memoryless nodes? KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote: > For fujitsu, problem is called "empty" node. > > When ACPI's SRAT table includes "possible nodes", ia64 bootstrap(acpi_numa_init) > creates nodes, which includes no memory, no cpu. > > I tried to remove empty-node in past, but that was denied. > It was because we can hot-add cpu to the empty node. > (node-hotplug triggered by cpu is not implemented now. and it will be ugly.) > > > For HP, (Lee can comment on this later), they have memory-less-node. > As far as I hear, HP's machine can have following configration. > > (example) > Node0: CPU0 memory AAA MB > Node1: CPU1 memory AAA MB > Node2: CPU2 memory AAA MB > Node3: CPU3 memory AAA MB > Node4: Memory XXX GB > > AAA is very small value (below 16MB) and will be omitted by ia64 bootstrap. > After boot, only Node 4 has valid memory (but have no cpu.) > > Maybe this is memory-interleave by firmware config. Christoph Lameter <clameter@sgi.com> wrote: > Future SGI platforms (actually also current one can have but nothing like > that is deployed to my knowledge) have nodes with only cpus. Current SGI > platforms have nodes with just I/O that we so far cannot manage in the > core. So the arch code maps them to the nearest memory node. Lee Schermerhorn <Lee.Schermerhorn@hp.com> wrote: > For the HP platforms, we can configure each cell with from 0% to 100% > "cell local memory". When we configure with <100% CLM, the "missing > percentages" are interleaved by hardware on a cache-line granularity to > improve bandwidth at the expense of latency for numa-challenged > applications [and OSes, but not our problem ;-)]. When we boot Linux on > such a config, all of the real nodes have no memory--it all resides in a > single interleaved pseudo-node. > > When we boot Linux on a 100% CLM configuration [== NUMA], we still have > the interleaved pseudo-node. It contains a few hundred MB stolen from > the real nodes to contain the DMA zone. [Interleaved memory resides at > phys addr 0]. The memoryless-nodes patches, along with the zoneorder > patches, support this config as well. > > Also, when we boot a NUMA config with the "mem=" command line, > specifying less memory than actually exists, Linux takes the excluded > memory "off the top" rather than distributing it across the nodes. This > can result in memoryless nodes, as well. > This patch: Preparation for memoryless node patches. Provide a generic way to keep nodemasks describing various characteristics of NUMA nodes. Remove the node_online_map and the node_possible map and realize the same functionality using two nodes stats: N_POSSIBLE and N_ONLINE. [Lee.Schermerhorn@hp.com: Initialize N_*_MEMORY and N_CPU masks for non-NUMA config] Signed-off-by: Christoph Lameter <clameter@sgi.com> Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Bob Picco <bob.picco@hp.com> Cc: Nishanth Aravamudan <nacc@us.ibm.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@skynet.ie> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: "Serge E. Hallyn" <serge@hallyn.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:27 +00:00
* Array of node states.
*/
Memoryless nodes: Generic management of nodemasks for various purposes Why do we need to support memoryless nodes? KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote: > For fujitsu, problem is called "empty" node. > > When ACPI's SRAT table includes "possible nodes", ia64 bootstrap(acpi_numa_init) > creates nodes, which includes no memory, no cpu. > > I tried to remove empty-node in past, but that was denied. > It was because we can hot-add cpu to the empty node. > (node-hotplug triggered by cpu is not implemented now. and it will be ugly.) > > > For HP, (Lee can comment on this later), they have memory-less-node. > As far as I hear, HP's machine can have following configration. > > (example) > Node0: CPU0 memory AAA MB > Node1: CPU1 memory AAA MB > Node2: CPU2 memory AAA MB > Node3: CPU3 memory AAA MB > Node4: Memory XXX GB > > AAA is very small value (below 16MB) and will be omitted by ia64 bootstrap. > After boot, only Node 4 has valid memory (but have no cpu.) > > Maybe this is memory-interleave by firmware config. Christoph Lameter <clameter@sgi.com> wrote: > Future SGI platforms (actually also current one can have but nothing like > that is deployed to my knowledge) have nodes with only cpus. Current SGI > platforms have nodes with just I/O that we so far cannot manage in the > core. So the arch code maps them to the nearest memory node. Lee Schermerhorn <Lee.Schermerhorn@hp.com> wrote: > For the HP platforms, we can configure each cell with from 0% to 100% > "cell local memory". When we configure with <100% CLM, the "missing > percentages" are interleaved by hardware on a cache-line granularity to > improve bandwidth at the expense of latency for numa-challenged > applications [and OSes, but not our problem ;-)]. When we boot Linux on > such a config, all of the real nodes have no memory--it all resides in a > single interleaved pseudo-node. > > When we boot Linux on a 100% CLM configuration [== NUMA], we still have > the interleaved pseudo-node. It contains a few hundred MB stolen from > the real nodes to contain the DMA zone. [Interleaved memory resides at > phys addr 0]. The memoryless-nodes patches, along with the zoneorder > patches, support this config as well. > > Also, when we boot a NUMA config with the "mem=" command line, > specifying less memory than actually exists, Linux takes the excluded > memory "off the top" rather than distributing it across the nodes. This > can result in memoryless nodes, as well. > This patch: Preparation for memoryless node patches. Provide a generic way to keep nodemasks describing various characteristics of NUMA nodes. Remove the node_online_map and the node_possible map and realize the same functionality using two nodes stats: N_POSSIBLE and N_ONLINE. [Lee.Schermerhorn@hp.com: Initialize N_*_MEMORY and N_CPU masks for non-NUMA config] Signed-off-by: Christoph Lameter <clameter@sgi.com> Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Bob Picco <bob.picco@hp.com> Cc: Nishanth Aravamudan <nacc@us.ibm.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@skynet.ie> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: "Serge E. Hallyn" <serge@hallyn.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:27 +00:00
nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
[N_POSSIBLE] = NODE_MASK_ALL,
[N_ONLINE] = { { [0] = 1UL } },
#ifndef CONFIG_NUMA
[N_NORMAL_MEMORY] = { { [0] = 1UL } },
#ifdef CONFIG_HIGHMEM
[N_HIGH_MEMORY] = { { [0] = 1UL } },
#endif
[N_CPU] = { { [0] = 1UL } },
#endif /* NUMA */
};
EXPORT_SYMBOL(node_states);
unsigned long totalram_pages __read_mostly;
[PATCH] overcommit: add calculate_totalreserve_pages() These patches are an enhancement of OVERCOMMIT_GUESS algorithm in __vm_enough_memory(). - why the kernel needed patching When the kernel can't allocate anonymous pages in practice, currnet OVERCOMMIT_GUESS could return success. This implementation might be the cause of oom kill in memory pressure situation. If the Linux runs with page reservation features like /proc/sys/vm/lowmem_reserve_ratio and without swap region, I think the oom kill occurs easily. - the overall design approach in the patch When the OVERCOMMET_GUESS algorithm calculates number of free pages, the reserved free pages are regarded as non-free pages. This change helps to avoid the pitfall that the number of free pages become less than the number which the kernel tries to keep free. - testing results I tested the patches using my test kernel module. If the patches aren't applied to the kernel, __vm_enough_memory() returns success in the situation but autual page allocation is failed. On the other hand, if the patches are applied to the kernel, memory allocation failure is avoided since __vm_enough_memory() returns failure in the situation. I checked that on i386 SMP 16GB memory machine. I haven't tested on nommu environment currently. This patch adds totalreserve_pages for __vm_enough_memory(). Calculate_totalreserve_pages() checks maximum lowmem_reserve pages and pages_high in each zone. Finally, the function stores the sum of each zone to totalreserve_pages. The totalreserve_pages is calculated when the VM is initilized. And the variable is updated when /proc/sys/vm/lowmem_reserve_raito or /proc/sys/vm/min_free_kbytes are changed. Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-11 05:52:59 +00:00
unsigned long totalreserve_pages __read_mostly;
int percpu_pagelist_fraction;
gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
#ifdef CONFIG_PM_SLEEP
/*
* The following functions are used by the suspend/hibernate code to temporarily
* change gfp_allowed_mask in order to avoid using I/O during memory allocations
* while devices are suspended. To avoid races with the suspend/hibernate code,
* they should always be called with pm_mutex held (gfp_allowed_mask also should
* only be modified with pm_mutex held, unless the suspend/hibernate code is
* guaranteed not to run in parallel with that modification).
*/
static gfp_t saved_gfp_mask;
void pm_restore_gfp_mask(void)
{
WARN_ON(!mutex_is_locked(&pm_mutex));
if (saved_gfp_mask) {
gfp_allowed_mask = saved_gfp_mask;
saved_gfp_mask = 0;
}
}
void pm_restrict_gfp_mask(void)
{
WARN_ON(!mutex_is_locked(&pm_mutex));
WARN_ON(saved_gfp_mask);
saved_gfp_mask = gfp_allowed_mask;
gfp_allowed_mask &= ~GFP_IOFS;
}
#endif /* CONFIG_PM_SLEEP */
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
int pageblock_order __read_mostly;
#endif
static void __free_pages_ok(struct page *page, unsigned int order);
/*
* results with 256, 32 in the lowmem_reserve sysctl:
* 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
* 1G machine -> (16M dma, 784M normal, 224M high)
* NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
* HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
* HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
[PATCH] x86_64: Add 4GB DMA32 zone Add a new 4GB GFP_DMA32 zone between the GFP_DMA and GFP_NORMAL zones. As a bit of historical background: when the x86-64 port was originally designed we had some discussion if we should use a 16MB DMA zone like i386 or a 4GB DMA zone like IA64 or both. Both was ruled out at this point because it was in early 2.4 when VM is still quite shakey and had bad troubles even dealing with one DMA zone. We settled on the 16MB DMA zone mainly because we worried about older soundcards and the floppy. But this has always caused problems since then because device drivers had trouble getting enough DMA able memory. These days the VM works much better and the wide use of NUMA has proven it can deal with many zones successfully. So this patch adds both zones. This helps drivers who need a lot of memory below 4GB because their hardware is not accessing more (graphic drivers - proprietary and free ones, video frame buffer drivers, sound drivers etc.). Previously they could only use IOMMU+16MB GFP_DMA, which was not enough memory. Another common problem is that hardware who has full memory addressing for >4GB misses it for some control structures in memory (like transmit rings or other metadata). They tended to allocate memory in the 16MB GFP_DMA or the IOMMU/swiotlb then using pci_alloc_consistent, but that can tie up a lot of precious 16MB GFPDMA/IOMMU/swiotlb memory (even on AMD systems the IOMMU tends to be quite small) especially if you have many devices. With the new zone pci_alloc_consistent can just put this stuff into memory below 4GB which works better. One argument was still if the zone should be 4GB or 2GB. The main motivation for 2GB would be an unnamed not so unpopular hardware raid controller (mostly found in older machines from a particular four letter company) who has a strange 2GB restriction in firmware. But that one works ok with swiotlb/IOMMU anyways, so it doesn't really need GFP_DMA32. I chose 4GB to be compatible with IA64 and because it seems to be the most common restriction. The new zone is so far added only for x86-64. For other architectures who don't set up this new zone nothing changes. Architectures can set a compatibility define in Kconfig CONFIG_DMA_IS_DMA32 that will define GFP_DMA32 as GFP_DMA. Otherwise it's a nop because on 32bit architectures it's normally not needed because GFP_NORMAL (=0) is DMA able enough. One problem is still that GFP_DMA means different things on different architectures. e.g. some drivers used to have #ifdef ia64 use GFP_DMA (trusting it to be 4GB) #elif __x86_64__ (use other hacks like the swiotlb because 16MB is not enough) ... . This was quite ugly and is now obsolete. These should be now converted to use GFP_DMA32 unconditionally. I haven't done this yet. Or best only use pci_alloc_consistent/dma_alloc_coherent which will use GFP_DMA32 transparently. Signed-off-by: Andi Kleen <ak@suse.de> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-11-05 16:25:53 +00:00
*
* TBD: should special case ZONE_DMA32 machines here - in those we normally
* don't need any ZONE_NORMAL reservation
*/
int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
#ifdef CONFIG_ZONE_DMA
256,
#endif
#ifdef CONFIG_ZONE_DMA32
256,
#endif
#ifdef CONFIG_HIGHMEM
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
32,
#endif
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
32,
};
EXPORT_SYMBOL(totalram_pages);
static char * const zone_names[MAX_NR_ZONES] = {
#ifdef CONFIG_ZONE_DMA
"DMA",
#endif
#ifdef CONFIG_ZONE_DMA32
"DMA32",
#endif
"Normal",
#ifdef CONFIG_HIGHMEM
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
"HighMem",
#endif
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
"Movable",
};
int min_free_kbytes = 1024;
static unsigned long __meminitdata nr_kernel_pages;
static unsigned long __meminitdata nr_all_pages;
static unsigned long __meminitdata dma_reserve;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
/*
* MAX_ACTIVE_REGIONS determines the maximum number of distinct
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
* ranges of memory (RAM) that may be registered with add_active_range().
* Ranges passed to add_active_range() will be merged if possible
* so the number of times add_active_range() can be called is
* related to the number of nodes and the number of holes
*/
#ifdef CONFIG_MAX_ACTIVE_REGIONS
/* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
#define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
#else
#if MAX_NUMNODES >= 32
/* If there can be many nodes, allow up to 50 holes per node */
#define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
#else
/* By default, allow up to 256 distinct regions */
#define MAX_ACTIVE_REGIONS 256
#endif
#endif
static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
static int __meminitdata nr_nodemap_entries;
static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
static unsigned long __initdata required_kernelcore;
static unsigned long __initdata required_movablecore;
static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
int movable_zone;
EXPORT_SYMBOL(movable_zone);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
#if MAX_NUMNODES > 1
int nr_node_ids __read_mostly = MAX_NUMNODES;
int nr_online_nodes __read_mostly = 1;
EXPORT_SYMBOL(nr_node_ids);
EXPORT_SYMBOL(nr_online_nodes);
#endif
int page_group_by_mobility_disabled __read_mostly;
static void set_pageblock_migratetype(struct page *page, int migratetype)
{
if (unlikely(page_group_by_mobility_disabled))
migratetype = MIGRATE_UNMOVABLE;
set_pageblock_flags_group(page, (unsigned long)migratetype,
PB_migrate, PB_migrate_end);
}
mm, PM/Freezer: Disable OOM killer when tasks are frozen Currently, the following scenario appears to be possible in theory: * Tasks are frozen for hibernation or suspend. * Free pages are almost exhausted. * Certain piece of code in the suspend code path attempts to allocate some memory using GFP_KERNEL and allocation order less than or equal to PAGE_ALLOC_COSTLY_ORDER. * __alloc_pages_internal() cannot find a free page so it invokes the OOM killer. * The OOM killer attempts to kill a task, but the task is frozen, so it doesn't die immediately. * __alloc_pages_internal() jumps to 'restart', unsuccessfully tries to find a free page and invokes the OOM killer. * No progress can be made. Although it is now hard to trigger during hibernation due to the memory shrinking carried out by the hibernation code, it is theoretically possible to trigger during suspend after the memory shrinking has been removed from that code path. Moreover, since memory allocations are going to be used for the hibernation memory shrinking, it will be even more likely to happen during hibernation. To prevent it from happening, introduce the oom_killer_disabled switch that will cause __alloc_pages_internal() to fail in the situations in which the OOM killer would have been called and make the freezer set this switch after tasks have been successfully frozen. [akpm@linux-foundation.org: be nicer to the namespace] Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl> Cc: Fengguang Wu <fengguang.wu@gmail.com> Cc: David Rientjes <rientjes@google.com> Acked-by: Pavel Machek <pavel@ucw.cz> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:32:41 +00:00
bool oom_killer_disabled __read_mostly;
#ifdef CONFIG_DEBUG_VM
static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
{
int ret = 0;
unsigned seq;
unsigned long pfn = page_to_pfn(page);
do {
seq = zone_span_seqbegin(zone);
if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
ret = 1;
else if (pfn < zone->zone_start_pfn)
ret = 1;
} while (zone_span_seqretry(zone, seq));
return ret;
}
static int page_is_consistent(struct zone *zone, struct page *page)
{
if (!pfn_valid_within(page_to_pfn(page)))
return 0;
if (zone != page_zone(page))
return 0;
return 1;
}
/*
* Temporary debugging check for pages not lying within a given zone.
*/
static int bad_range(struct zone *zone, struct page *page)
{
if (page_outside_zone_boundaries(zone, page))
return 1;
if (!page_is_consistent(zone, page))
return 1;
return 0;
}
#else
static inline int bad_range(struct zone *zone, struct page *page)
{
return 0;
}
#endif
static void bad_page(struct page *page)
{
static unsigned long resume;
static unsigned long nr_shown;
static unsigned long nr_unshown;
/* Don't complain about poisoned pages */
if (PageHWPoison(page)) {
reset_page_mapcount(page); /* remove PageBuddy */
return;
}
/*
* Allow a burst of 60 reports, then keep quiet for that minute;
* or allow a steady drip of one report per second.
*/
if (nr_shown == 60) {
if (time_before(jiffies, resume)) {
nr_unshown++;
goto out;
}
if (nr_unshown) {
printk(KERN_ALERT
"BUG: Bad page state: %lu messages suppressed\n",
nr_unshown);
nr_unshown = 0;
}
nr_shown = 0;
}
if (nr_shown++ == 0)
resume = jiffies + 60 * HZ;
printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
badpage: replace page_remove_rmap Eeek and BUG Now that bad pages are kept out of circulation, there is no need for the infamous page_remove_rmap() BUG() - once that page is freed, its negative mapcount will issue a "Bad page state" message and the page won't be freed. Removing the BUG() allows more info, on subsequent pages, to be gathered. We do have more info about the page at this point than bad_page() can know - notably, what the pmd is, which might pinpoint something like low 64kB corruption - but page_remove_rmap() isn't given the address to find that. In practice, there is only one call to page_remove_rmap() which has ever reported anything, that from zap_pte_range() (usually on exit, sometimes on munmap). It has all the info, so remove page_remove_rmap()'s "Eeek" message and leave it all to zap_pte_range(). mm/memory.c already has a hardly used print_bad_pte() function, showing some of the appropriate info: extend it to show what we want for the rmap case: pte info, page info (when there is a page) and vma info to compare. zap_pte_range() already knows the pmd, but print_bad_pte() is easier to use if it works that out for itself. Some of this info is also shown in bad_page()'s "Bad page state" message. Keep them separate, but adjust them to match each other as far as possible. Say "Bad page map" in print_bad_pte(), and add a TAINT_BAD_PAGE there too. print_bad_pte() show current->comm unconditionally (though it should get repeated in the usually irrelevant stack trace): sorry, I misled Nick Piggin to make it conditional on vm_mm == current->mm, but current->mm is already NULL in the exit case. Usually current->comm is good, though exceptionally it may not be that of the mm (when "swapoff" for example). Signed-off-by: Hugh Dickins <hugh@veritas.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:40:08 +00:00
current->comm, page_to_pfn(page));
dump_page(page);
badpage: replace page_remove_rmap Eeek and BUG Now that bad pages are kept out of circulation, there is no need for the infamous page_remove_rmap() BUG() - once that page is freed, its negative mapcount will issue a "Bad page state" message and the page won't be freed. Removing the BUG() allows more info, on subsequent pages, to be gathered. We do have more info about the page at this point than bad_page() can know - notably, what the pmd is, which might pinpoint something like low 64kB corruption - but page_remove_rmap() isn't given the address to find that. In practice, there is only one call to page_remove_rmap() which has ever reported anything, that from zap_pte_range() (usually on exit, sometimes on munmap). It has all the info, so remove page_remove_rmap()'s "Eeek" message and leave it all to zap_pte_range(). mm/memory.c already has a hardly used print_bad_pte() function, showing some of the appropriate info: extend it to show what we want for the rmap case: pte info, page info (when there is a page) and vma info to compare. zap_pte_range() already knows the pmd, but print_bad_pte() is easier to use if it works that out for itself. Some of this info is also shown in bad_page()'s "Bad page state" message. Keep them separate, but adjust them to match each other as far as possible. Say "Bad page map" in print_bad_pte(), and add a TAINT_BAD_PAGE there too. print_bad_pte() show current->comm unconditionally (though it should get repeated in the usually irrelevant stack trace): sorry, I misled Nick Piggin to make it conditional on vm_mm == current->mm, but current->mm is already NULL in the exit case. Usually current->comm is good, though exceptionally it may not be that of the mm (when "swapoff" for example). Signed-off-by: Hugh Dickins <hugh@veritas.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:40:08 +00:00
print_modules();
dump_stack();
out:
/* Leave bad fields for debug, except PageBuddy could make trouble */
reset_page_mapcount(page); /* remove PageBuddy */
add_taint(TAINT_BAD_PAGE);
}
/*
* Higher-order pages are called "compound pages". They are structured thusly:
*
* The first PAGE_SIZE page is called the "head page".
*
* The remaining PAGE_SIZE pages are called "tail pages".
*
* All pages have PG_compound set. All pages have their ->private pointing at
* the head page (even the head page has this).
*
[PATCH] compound page: use page[1].lru If a compound page has its own put_page_testzero destructor (the only current example is free_huge_page), that is noted in page[1].mapping of the compound page. But that's rather a poor place to keep it: functions which call set_page_dirty_lock after get_user_pages (e.g. Infiniband's __ib_umem_release) ought to be checking first, otherwise set_page_dirty is liable to crash on what's not the address of a struct address_space. And now I'm about to make that worse: it turns out that every compound page needs a destructor, so we can no longer rely on hugetlb pages going their own special way, to avoid further problems of page->mapping reuse. For example, not many people know that: on 50% of i386 -Os builds, the first tail page of a compound page purports to be PageAnon (when its destructor has an odd address), which surprises page_add_file_rmap. Keep the compound page destructor in page[1].lru.next instead. And to free up the common pairing of mapping and index, also move compound page order from index to lru.prev. Slab reuses page->lru too: but if we ever need slab to use compound pages, it can easily stack its use above this. (akpm: decoded version of the above: the tail pages of a compound page now have ->mapping==NULL, so there's no need for the set_page_dirty[_lock]() caller to check that they're not compund pages before doing the dirty). Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-14 21:52:58 +00:00
* The first tail page's ->lru.next holds the address of the compound page's
* put_page() function. Its ->lru.prev holds the order of allocation.
* This usage means that zero-order pages may not be compound.
*/
static void free_compound_page(struct page *page)
{
2007-05-06 21:49:39 +00:00
__free_pages_ok(page, compound_order(page));
}
void prep_compound_page(struct page *page, unsigned long order)
{
int i;
int nr_pages = 1 << order;
set_compound_page_dtor(page, free_compound_page);
set_compound_order(page, order);
__SetPageHead(page);
for (i = 1; i < nr_pages; i++) {
struct page *p = page + i;
__SetPageTail(p);
p->first_page = page;
}
}
/* update __split_huge_page_refcount if you change this function */
static int destroy_compound_page(struct page *page, unsigned long order)
{
int i;
int nr_pages = 1 << order;
int bad = 0;
if (unlikely(compound_order(page) != order) ||
unlikely(!PageHead(page))) {
bad_page(page);
bad++;
}
__ClearPageHead(page);
for (i = 1; i < nr_pages; i++) {
struct page *p = page + i;
if (unlikely(!PageTail(p) || (p->first_page != page))) {
bad_page(page);
bad++;
}
2007-05-06 21:49:39 +00:00
__ClearPageTail(p);
}
return bad;
}
static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
{
int i;
/*
* clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
* and __GFP_HIGHMEM from hard or soft interrupt context.
*/
VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
for (i = 0; i < (1 << order); i++)
clear_highpage(page + i);
}
static inline void set_page_order(struct page *page, int order)
{
[PATCH] mm: split page table lock Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with a many-threaded application which concurrently initializes different parts of a large anonymous area. This patch corrects that, by using a separate spinlock per page table page, to guard the page table entries in that page, instead of using the mm's single page_table_lock. (But even then, page_table_lock is still used to guard page table allocation, and anon_vma allocation.) In this implementation, the spinlock is tucked inside the struct page of the page table page: with a BUILD_BUG_ON in case it overflows - which it would in the case of 32-bit PA-RISC with spinlock debugging enabled. Splitting the lock is not quite for free: another cacheline access. Ideally, I suppose we would use split ptlock only for multi-threaded processes on multi-cpu machines; but deciding that dynamically would have its own costs. So for now enable it by config, at some number of cpus - since the Kconfig language doesn't support inequalities, let preprocessor compare that with NR_CPUS. But I don't think it's worth being user-configurable: for good testing of both split and unsplit configs, split now at 4 cpus, and perhaps change that to 8 later. There is a benefit even for singly threaded processes: kswapd can be attacking one part of the mm while another part is busy faulting. Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 01:16:40 +00:00
set_page_private(page, order);
__SetPageBuddy(page);
}
static inline void rmv_page_order(struct page *page)
{
__ClearPageBuddy(page);
[PATCH] mm: split page table lock Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with a many-threaded application which concurrently initializes different parts of a large anonymous area. This patch corrects that, by using a separate spinlock per page table page, to guard the page table entries in that page, instead of using the mm's single page_table_lock. (But even then, page_table_lock is still used to guard page table allocation, and anon_vma allocation.) In this implementation, the spinlock is tucked inside the struct page of the page table page: with a BUILD_BUG_ON in case it overflows - which it would in the case of 32-bit PA-RISC with spinlock debugging enabled. Splitting the lock is not quite for free: another cacheline access. Ideally, I suppose we would use split ptlock only for multi-threaded processes on multi-cpu machines; but deciding that dynamically would have its own costs. So for now enable it by config, at some number of cpus - since the Kconfig language doesn't support inequalities, let preprocessor compare that with NR_CPUS. But I don't think it's worth being user-configurable: for good testing of both split and unsplit configs, split now at 4 cpus, and perhaps change that to 8 later. There is a benefit even for singly threaded processes: kswapd can be attacking one part of the mm while another part is busy faulting. Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 01:16:40 +00:00
set_page_private(page, 0);
}
/*
* Locate the struct page for both the matching buddy in our
* pair (buddy1) and the combined O(n+1) page they form (page).
*
* 1) Any buddy B1 will have an order O twin B2 which satisfies
* the following equation:
* B2 = B1 ^ (1 << O)
* For example, if the starting buddy (buddy2) is #8 its order
* 1 buddy is #10:
* B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
*
* 2) Any buddy B will have an order O+1 parent P which
* satisfies the following equation:
* P = B & ~(1 << O)
*
* Assumption: *_mem_map is contiguous at least up to MAX_ORDER
*/
static inline unsigned long
__find_buddy_index(unsigned long page_idx, unsigned int order)
{
return page_idx ^ (1 << order);
}
/*
* This function checks whether a page is free && is the buddy
* we can do coalesce a page and its buddy if
* (a) the buddy is not in a hole &&
* (b) the buddy is in the buddy system &&
* (c) a page and its buddy have the same order &&
* (d) a page and its buddy are in the same zone.
*
* For recording whether a page is in the buddy system, we set ->_mapcount -2.
* Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
*
* For recording page's order, we use page_private(page).
*/
static inline int page_is_buddy(struct page *page, struct page *buddy,
int order)
{
if (!pfn_valid_within(page_to_pfn(buddy)))
return 0;
if (page_zone_id(page) != page_zone_id(buddy))
return 0;
if (PageBuddy(buddy) && page_order(buddy) == order) {
VM_BUG_ON(page_count(buddy) != 0);
return 1;
}
return 0;
}
/*
* Freeing function for a buddy system allocator.
*
* The concept of a buddy system is to maintain direct-mapped table
* (containing bit values) for memory blocks of various "orders".
* The bottom level table contains the map for the smallest allocatable
* units of memory (here, pages), and each level above it describes
* pairs of units from the levels below, hence, "buddies".
* At a high level, all that happens here is marking the table entry
* at the bottom level available, and propagating the changes upward
* as necessary, plus some accounting needed to play nicely with other
* parts of the VM system.
* At each level, we keep a list of pages, which are heads of continuous
* free pages of length of (1 << order) and marked with _mapcount -2. Page's
[PATCH] mm: split page table lock Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with a many-threaded application which concurrently initializes different parts of a large anonymous area. This patch corrects that, by using a separate spinlock per page table page, to guard the page table entries in that page, instead of using the mm's single page_table_lock. (But even then, page_table_lock is still used to guard page table allocation, and anon_vma allocation.) In this implementation, the spinlock is tucked inside the struct page of the page table page: with a BUILD_BUG_ON in case it overflows - which it would in the case of 32-bit PA-RISC with spinlock debugging enabled. Splitting the lock is not quite for free: another cacheline access. Ideally, I suppose we would use split ptlock only for multi-threaded processes on multi-cpu machines; but deciding that dynamically would have its own costs. So for now enable it by config, at some number of cpus - since the Kconfig language doesn't support inequalities, let preprocessor compare that with NR_CPUS. But I don't think it's worth being user-configurable: for good testing of both split and unsplit configs, split now at 4 cpus, and perhaps change that to 8 later. There is a benefit even for singly threaded processes: kswapd can be attacking one part of the mm while another part is busy faulting. Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 01:16:40 +00:00
* order is recorded in page_private(page) field.
* So when we are allocating or freeing one, we can derive the state of the
* other. That is, if we allocate a small block, and both were
* free, the remainder of the region must be split into blocks.
* If a block is freed, and its buddy is also free, then this
* triggers coalescing into a block of larger size.
*
* -- wli
*/
static inline void __free_one_page(struct page *page,
struct zone *zone, unsigned int order,
int migratetype)
{
unsigned long page_idx;
page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists In order to reduce fragmentation, this patch classifies freed pages in two groups according to their probability of being part of a high order merge. Pages belonging to a compound whose next-highest buddy is free are more likely to be part of a high order merge in the near future, so they will be added at the tail of the freelist. The remaining pages are put at the front of the freelist. In this way, the pages that are more likely to cause a big merge are kept free longer. Consequently there is a tendency to aggregate the long-living allocations on a subset of the compounds, reducing the fragmentation. This heuristic was tested on three machines, x86, x86-64 and ppc64 with 3GB of RAM in each machine. The tests were kernbench, netperf, sysbench and STREAM for performance and a high-order stress test for huge page allocations. KernBench X86 Elapsed mean 374.77 ( 0.00%) 375.10 (-0.09%) User mean 649.53 ( 0.00%) 650.44 (-0.14%) System mean 54.75 ( 0.00%) 54.18 ( 1.05%) CPU mean 187.75 ( 0.00%) 187.25 ( 0.27%) KernBench X86-64 Elapsed mean 94.45 ( 0.00%) 94.01 ( 0.47%) User mean 323.27 ( 0.00%) 322.66 ( 0.19%) System mean 36.71 ( 0.00%) 36.50 ( 0.57%) CPU mean 380.75 ( 0.00%) 381.75 (-0.26%) KernBench PPC64 Elapsed mean 173.45 ( 0.00%) 173.74 (-0.17%) User mean 587.99 ( 0.00%) 587.95 ( 0.01%) System mean 60.60 ( 0.00%) 60.57 ( 0.05%) CPU mean 373.50 ( 0.00%) 372.75 ( 0.20%) Nothing notable for kernbench. NetPerf UDP X86 64 42.68 ( 0.00%) 42.77 ( 0.21%) 128 85.62 ( 0.00%) 85.32 (-0.35%) 256 170.01 ( 0.00%) 168.76 (-0.74%) 1024 655.68 ( 0.00%) 652.33 (-0.51%) 2048 1262.39 ( 0.00%) 1248.61 (-1.10%) 3312 1958.41 ( 0.00%) 1944.61 (-0.71%) 4096 2345.63 ( 0.00%) 2318.83 (-1.16%) 8192 4132.90 ( 0.00%) 4089.50 (-1.06%) 16384 6770.88 ( 0.00%) 6642.05 (-1.94%)* NetPerf UDP X86-64 64 148.82 ( 0.00%) 154.92 ( 3.94%) 128 298.96 ( 0.00%) 312.95 ( 4.47%) 256 583.67 ( 0.00%) 626.39 ( 6.82%) 1024 2293.18 ( 0.00%) 2371.10 ( 3.29%) 2048 4274.16 ( 0.00%) 4396.83 ( 2.79%) 3312 6356.94 ( 0.00%) 6571.35 ( 3.26%) 4096 7422.68 ( 0.00%) 7635.42 ( 2.79%)* 8192 12114.81 ( 0.00%)* 12346.88 ( 1.88%) 16384 17022.28 ( 0.00%)* 17033.19 ( 0.06%)* 1.64% 2.73% NetPerf UDP PPC64 64 49.98 ( 0.00%) 50.25 ( 0.54%) 128 98.66 ( 0.00%) 100.95 ( 2.27%) 256 197.33 ( 0.00%) 191.03 (-3.30%) 1024 761.98 ( 0.00%) 785.07 ( 2.94%) 2048 1493.50 ( 0.00%) 1510.85 ( 1.15%) 3312 2303.95 ( 0.00%) 2271.72 (-1.42%) 4096 2774.56 ( 0.00%) 2773.06 (-0.05%) 8192 4918.31 ( 0.00%) 4793.59 (-2.60%) 16384 7497.98 ( 0.00%) 7749.52 ( 3.25%) The tests are run to have confidence limits within 1%. Results marked with a * were not confident although in this case, it's only outside by small amounts. Even with some results that were not confident, the netperf UDP results were generally positive. NetPerf TCP X86 64 652.25 ( 0.00%)* 648.12 (-0.64%)* 23.80% 22.82% 128 1229.98 ( 0.00%)* 1220.56 (-0.77%)* 21.03% 18.90% 256 2105.88 ( 0.00%) 1872.03 (-12.49%)* 1.00% 16.46% 1024 3476.46 ( 0.00%)* 3548.28 ( 2.02%)* 13.37% 11.39% 2048 4023.44 ( 0.00%)* 4231.45 ( 4.92%)* 9.76% 12.48% 3312 4348.88 ( 0.00%)* 4396.96 ( 1.09%)* 6.49% 8.75% 4096 4726.56 ( 0.00%)* 4877.71 ( 3.10%)* 9.85% 8.50% 8192 4732.28 ( 0.00%)* 5777.77 (18.10%)* 9.13% 13.04% 16384 5543.05 ( 0.00%)* 5906.24 ( 6.15%)* 7.73% 8.68% NETPERF TCP X86-64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 1895.87 ( 0.00%)* 1775.07 (-6.81%)* 5.79% 4.78% 128 3571.03 ( 0.00%)* 3342.20 (-6.85%)* 3.68% 6.06% 256 5097.21 ( 0.00%)* 4859.43 (-4.89%)* 3.02% 2.10% 1024 8919.10 ( 0.00%)* 8892.49 (-0.30%)* 5.89% 6.55% 2048 10255.46 ( 0.00%)* 10449.39 ( 1.86%)* 7.08% 7.44% 3312 10839.90 ( 0.00%)* 10740.15 (-0.93%)* 6.87% 7.33% 4096 10814.84 ( 0.00%)* 10766.97 (-0.44%)* 6.86% 8.18% 8192 11606.89 ( 0.00%)* 11189.28 (-3.73%)* 7.49% 5.55% 16384 12554.88 ( 0.00%)* 12361.22 (-1.57%)* 7.36% 6.49% NETPERF TCP PPC64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 594.17 ( 0.00%) 596.04 ( 0.31%)* 1.00% 2.29% 128 1064.87 ( 0.00%)* 1074.77 ( 0.92%)* 1.30% 1.40% 256 1852.46 ( 0.00%)* 1856.95 ( 0.24%) 1.25% 1.00% 1024 3839.46 ( 0.00%)* 3813.05 (-0.69%) 1.02% 1.00% 2048 4885.04 ( 0.00%)* 4881.97 (-0.06%)* 1.15% 1.04% 3312 5506.90 ( 0.00%) 5459.72 (-0.86%) 4096 6449.19 ( 0.00%) 6345.46 (-1.63%) 8192 7501.17 ( 0.00%) 7508.79 ( 0.10%) 16384 9618.65 ( 0.00%) 9490.10 (-1.35%) There was a distinct lack of confidence in the X86* figures so I included what the devation was where the results were not confident. Many of the results, whether gains or losses were within the standard deviation so no solid conclusion can be reached on performance impact. Looking at the figures, only the X86-64 ones look suspicious with a few losses that were outside the noise. However, the results were so unstable that without knowing why they vary so much, a solid conclusion cannot be reached. SYSBENCH X86 sysbench-vanilla pgalloc-delay 1 7722.85 ( 0.00%) 7756.79 ( 0.44%) 2 14901.11 ( 0.00%) 13683.44 (-8.90%) 3 15171.71 ( 0.00%) 14888.25 (-1.90%) 4 14966.98 ( 0.00%) 15029.67 ( 0.42%) 5 14370.47 ( 0.00%) 14865.00 ( 3.33%) 6 14870.33 ( 0.00%) 14845.57 (-0.17%) 7 14429.45 ( 0.00%) 14520.85 ( 0.63%) 8 14354.35 ( 0.00%) 14362.31 ( 0.06%) SYSBENCH X86-64 1 17448.70 ( 0.00%) 17484.41 ( 0.20%) 2 34276.39 ( 0.00%) 34251.00 (-0.07%) 3 50805.25 ( 0.00%) 50854.80 ( 0.10%) 4 66667.10 ( 0.00%) 66174.69 (-0.74%) 5 66003.91 ( 0.00%) 65685.25 (-0.49%) 6 64981.90 ( 0.00%) 65125.60 ( 0.22%) 7 64933.16 ( 0.00%) 64379.23 (-0.86%) 8 63353.30 ( 0.00%) 63281.22 (-0.11%) 9 63511.84 ( 0.00%) 63570.37 ( 0.09%) 10 62708.27 ( 0.00%) 63166.25 ( 0.73%) 11 62092.81 ( 0.00%) 61787.75 (-0.49%) 12 61330.11 ( 0.00%) 61036.34 (-0.48%) 13 61438.37 ( 0.00%) 61994.47 ( 0.90%) 14 62304.48 ( 0.00%) 62064.90 (-0.39%) 15 63296.48 ( 0.00%) 62875.16 (-0.67%) 16 63951.76 ( 0.00%) 63769.09 (-0.29%) SYSBENCH PPC64 -sysbench-pgalloc-delay-sysbench sysbench-vanilla pgalloc-delay 1 7645.08 ( 0.00%) 7467.43 (-2.38%) 2 14856.67 ( 0.00%) 14558.73 (-2.05%) 3 21952.31 ( 0.00%) 21683.64 (-1.24%) 4 27946.09 ( 0.00%) 28623.29 ( 2.37%) 5 28045.11 ( 0.00%) 28143.69 ( 0.35%) 6 27477.10 ( 0.00%) 27337.45 (-0.51%) 7 26489.17 ( 0.00%) 26590.06 ( 0.38%) 8 26642.91 ( 0.00%) 25274.33 (-5.41%) 9 25137.27 ( 0.00%) 24810.06 (-1.32%) 10 24451.99 ( 0.00%) 24275.85 (-0.73%) 11 23262.20 ( 0.00%) 23674.88 ( 1.74%) 12 24234.81 ( 0.00%) 23640.89 (-2.51%) 13 24577.75 ( 0.00%) 24433.50 (-0.59%) 14 25640.19 ( 0.00%) 25116.52 (-2.08%) 15 26188.84 ( 0.00%) 26181.36 (-0.03%) 16 26782.37 ( 0.00%) 26255.99 (-2.00%) Again, there is little to conclude here. While there are a few losses, the results vary by +/- 8% in some cases. They are the results of most concern as there are some large losses but it's also within the variance typically seen between kernel releases. The STREAM results varied so little and are so verbose that I didn't include them here. The final test stressed how many huge pages can be allocated. The absolute number of huge pages allocated are the same with or without the page. However, the "unusability free space index" which is a measure of external fragmentation was slightly lower (lower is better) throughout the lifetime of the system. I also measured the latency of how long it took to successfully allocate a huge page. The latency was slightly lower and on X86 and PPC64, more huge pages were allocated almost immediately from the free lists. The improvement is slight but there. [mel@csn.ul.ie: Tested, reworked for less branches] [czoccolo@gmail.com: fix oops by checking pfn_valid_within()] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Corrado Zoccolo <czoccolo@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:31:54 +00:00
unsigned long combined_idx;
unsigned long uninitialized_var(buddy_idx);
page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists In order to reduce fragmentation, this patch classifies freed pages in two groups according to their probability of being part of a high order merge. Pages belonging to a compound whose next-highest buddy is free are more likely to be part of a high order merge in the near future, so they will be added at the tail of the freelist. The remaining pages are put at the front of the freelist. In this way, the pages that are more likely to cause a big merge are kept free longer. Consequently there is a tendency to aggregate the long-living allocations on a subset of the compounds, reducing the fragmentation. This heuristic was tested on three machines, x86, x86-64 and ppc64 with 3GB of RAM in each machine. The tests were kernbench, netperf, sysbench and STREAM for performance and a high-order stress test for huge page allocations. KernBench X86 Elapsed mean 374.77 ( 0.00%) 375.10 (-0.09%) User mean 649.53 ( 0.00%) 650.44 (-0.14%) System mean 54.75 ( 0.00%) 54.18 ( 1.05%) CPU mean 187.75 ( 0.00%) 187.25 ( 0.27%) KernBench X86-64 Elapsed mean 94.45 ( 0.00%) 94.01 ( 0.47%) User mean 323.27 ( 0.00%) 322.66 ( 0.19%) System mean 36.71 ( 0.00%) 36.50 ( 0.57%) CPU mean 380.75 ( 0.00%) 381.75 (-0.26%) KernBench PPC64 Elapsed mean 173.45 ( 0.00%) 173.74 (-0.17%) User mean 587.99 ( 0.00%) 587.95 ( 0.01%) System mean 60.60 ( 0.00%) 60.57 ( 0.05%) CPU mean 373.50 ( 0.00%) 372.75 ( 0.20%) Nothing notable for kernbench. NetPerf UDP X86 64 42.68 ( 0.00%) 42.77 ( 0.21%) 128 85.62 ( 0.00%) 85.32 (-0.35%) 256 170.01 ( 0.00%) 168.76 (-0.74%) 1024 655.68 ( 0.00%) 652.33 (-0.51%) 2048 1262.39 ( 0.00%) 1248.61 (-1.10%) 3312 1958.41 ( 0.00%) 1944.61 (-0.71%) 4096 2345.63 ( 0.00%) 2318.83 (-1.16%) 8192 4132.90 ( 0.00%) 4089.50 (-1.06%) 16384 6770.88 ( 0.00%) 6642.05 (-1.94%)* NetPerf UDP X86-64 64 148.82 ( 0.00%) 154.92 ( 3.94%) 128 298.96 ( 0.00%) 312.95 ( 4.47%) 256 583.67 ( 0.00%) 626.39 ( 6.82%) 1024 2293.18 ( 0.00%) 2371.10 ( 3.29%) 2048 4274.16 ( 0.00%) 4396.83 ( 2.79%) 3312 6356.94 ( 0.00%) 6571.35 ( 3.26%) 4096 7422.68 ( 0.00%) 7635.42 ( 2.79%)* 8192 12114.81 ( 0.00%)* 12346.88 ( 1.88%) 16384 17022.28 ( 0.00%)* 17033.19 ( 0.06%)* 1.64% 2.73% NetPerf UDP PPC64 64 49.98 ( 0.00%) 50.25 ( 0.54%) 128 98.66 ( 0.00%) 100.95 ( 2.27%) 256 197.33 ( 0.00%) 191.03 (-3.30%) 1024 761.98 ( 0.00%) 785.07 ( 2.94%) 2048 1493.50 ( 0.00%) 1510.85 ( 1.15%) 3312 2303.95 ( 0.00%) 2271.72 (-1.42%) 4096 2774.56 ( 0.00%) 2773.06 (-0.05%) 8192 4918.31 ( 0.00%) 4793.59 (-2.60%) 16384 7497.98 ( 0.00%) 7749.52 ( 3.25%) The tests are run to have confidence limits within 1%. Results marked with a * were not confident although in this case, it's only outside by small amounts. Even with some results that were not confident, the netperf UDP results were generally positive. NetPerf TCP X86 64 652.25 ( 0.00%)* 648.12 (-0.64%)* 23.80% 22.82% 128 1229.98 ( 0.00%)* 1220.56 (-0.77%)* 21.03% 18.90% 256 2105.88 ( 0.00%) 1872.03 (-12.49%)* 1.00% 16.46% 1024 3476.46 ( 0.00%)* 3548.28 ( 2.02%)* 13.37% 11.39% 2048 4023.44 ( 0.00%)* 4231.45 ( 4.92%)* 9.76% 12.48% 3312 4348.88 ( 0.00%)* 4396.96 ( 1.09%)* 6.49% 8.75% 4096 4726.56 ( 0.00%)* 4877.71 ( 3.10%)* 9.85% 8.50% 8192 4732.28 ( 0.00%)* 5777.77 (18.10%)* 9.13% 13.04% 16384 5543.05 ( 0.00%)* 5906.24 ( 6.15%)* 7.73% 8.68% NETPERF TCP X86-64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 1895.87 ( 0.00%)* 1775.07 (-6.81%)* 5.79% 4.78% 128 3571.03 ( 0.00%)* 3342.20 (-6.85%)* 3.68% 6.06% 256 5097.21 ( 0.00%)* 4859.43 (-4.89%)* 3.02% 2.10% 1024 8919.10 ( 0.00%)* 8892.49 (-0.30%)* 5.89% 6.55% 2048 10255.46 ( 0.00%)* 10449.39 ( 1.86%)* 7.08% 7.44% 3312 10839.90 ( 0.00%)* 10740.15 (-0.93%)* 6.87% 7.33% 4096 10814.84 ( 0.00%)* 10766.97 (-0.44%)* 6.86% 8.18% 8192 11606.89 ( 0.00%)* 11189.28 (-3.73%)* 7.49% 5.55% 16384 12554.88 ( 0.00%)* 12361.22 (-1.57%)* 7.36% 6.49% NETPERF TCP PPC64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 594.17 ( 0.00%) 596.04 ( 0.31%)* 1.00% 2.29% 128 1064.87 ( 0.00%)* 1074.77 ( 0.92%)* 1.30% 1.40% 256 1852.46 ( 0.00%)* 1856.95 ( 0.24%) 1.25% 1.00% 1024 3839.46 ( 0.00%)* 3813.05 (-0.69%) 1.02% 1.00% 2048 4885.04 ( 0.00%)* 4881.97 (-0.06%)* 1.15% 1.04% 3312 5506.90 ( 0.00%) 5459.72 (-0.86%) 4096 6449.19 ( 0.00%) 6345.46 (-1.63%) 8192 7501.17 ( 0.00%) 7508.79 ( 0.10%) 16384 9618.65 ( 0.00%) 9490.10 (-1.35%) There was a distinct lack of confidence in the X86* figures so I included what the devation was where the results were not confident. Many of the results, whether gains or losses were within the standard deviation so no solid conclusion can be reached on performance impact. Looking at the figures, only the X86-64 ones look suspicious with a few losses that were outside the noise. However, the results were so unstable that without knowing why they vary so much, a solid conclusion cannot be reached. SYSBENCH X86 sysbench-vanilla pgalloc-delay 1 7722.85 ( 0.00%) 7756.79 ( 0.44%) 2 14901.11 ( 0.00%) 13683.44 (-8.90%) 3 15171.71 ( 0.00%) 14888.25 (-1.90%) 4 14966.98 ( 0.00%) 15029.67 ( 0.42%) 5 14370.47 ( 0.00%) 14865.00 ( 3.33%) 6 14870.33 ( 0.00%) 14845.57 (-0.17%) 7 14429.45 ( 0.00%) 14520.85 ( 0.63%) 8 14354.35 ( 0.00%) 14362.31 ( 0.06%) SYSBENCH X86-64 1 17448.70 ( 0.00%) 17484.41 ( 0.20%) 2 34276.39 ( 0.00%) 34251.00 (-0.07%) 3 50805.25 ( 0.00%) 50854.80 ( 0.10%) 4 66667.10 ( 0.00%) 66174.69 (-0.74%) 5 66003.91 ( 0.00%) 65685.25 (-0.49%) 6 64981.90 ( 0.00%) 65125.60 ( 0.22%) 7 64933.16 ( 0.00%) 64379.23 (-0.86%) 8 63353.30 ( 0.00%) 63281.22 (-0.11%) 9 63511.84 ( 0.00%) 63570.37 ( 0.09%) 10 62708.27 ( 0.00%) 63166.25 ( 0.73%) 11 62092.81 ( 0.00%) 61787.75 (-0.49%) 12 61330.11 ( 0.00%) 61036.34 (-0.48%) 13 61438.37 ( 0.00%) 61994.47 ( 0.90%) 14 62304.48 ( 0.00%) 62064.90 (-0.39%) 15 63296.48 ( 0.00%) 62875.16 (-0.67%) 16 63951.76 ( 0.00%) 63769.09 (-0.29%) SYSBENCH PPC64 -sysbench-pgalloc-delay-sysbench sysbench-vanilla pgalloc-delay 1 7645.08 ( 0.00%) 7467.43 (-2.38%) 2 14856.67 ( 0.00%) 14558.73 (-2.05%) 3 21952.31 ( 0.00%) 21683.64 (-1.24%) 4 27946.09 ( 0.00%) 28623.29 ( 2.37%) 5 28045.11 ( 0.00%) 28143.69 ( 0.35%) 6 27477.10 ( 0.00%) 27337.45 (-0.51%) 7 26489.17 ( 0.00%) 26590.06 ( 0.38%) 8 26642.91 ( 0.00%) 25274.33 (-5.41%) 9 25137.27 ( 0.00%) 24810.06 (-1.32%) 10 24451.99 ( 0.00%) 24275.85 (-0.73%) 11 23262.20 ( 0.00%) 23674.88 ( 1.74%) 12 24234.81 ( 0.00%) 23640.89 (-2.51%) 13 24577.75 ( 0.00%) 24433.50 (-0.59%) 14 25640.19 ( 0.00%) 25116.52 (-2.08%) 15 26188.84 ( 0.00%) 26181.36 (-0.03%) 16 26782.37 ( 0.00%) 26255.99 (-2.00%) Again, there is little to conclude here. While there are a few losses, the results vary by +/- 8% in some cases. They are the results of most concern as there are some large losses but it's also within the variance typically seen between kernel releases. The STREAM results varied so little and are so verbose that I didn't include them here. The final test stressed how many huge pages can be allocated. The absolute number of huge pages allocated are the same with or without the page. However, the "unusability free space index" which is a measure of external fragmentation was slightly lower (lower is better) throughout the lifetime of the system. I also measured the latency of how long it took to successfully allocate a huge page. The latency was slightly lower and on X86 and PPC64, more huge pages were allocated almost immediately from the free lists. The improvement is slight but there. [mel@csn.ul.ie: Tested, reworked for less branches] [czoccolo@gmail.com: fix oops by checking pfn_valid_within()] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Corrado Zoccolo <czoccolo@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:31:54 +00:00
struct page *buddy;
if (unlikely(PageCompound(page)))
if (unlikely(destroy_compound_page(page, order)))
return;
VM_BUG_ON(migratetype == -1);
page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
VM_BUG_ON(page_idx & ((1 << order) - 1));
VM_BUG_ON(bad_range(zone, page));
while (order < MAX_ORDER-1) {
buddy_idx = __find_buddy_index(page_idx, order);
buddy = page + (buddy_idx - page_idx);
if (!page_is_buddy(page, buddy, order))
break;
/* Our buddy is free, merge with it and move up one order. */
list_del(&buddy->lru);
zone->free_area[order].nr_free--;
rmv_page_order(buddy);
combined_idx = buddy_idx & page_idx;
page = page + (combined_idx - page_idx);
page_idx = combined_idx;
order++;
}
set_page_order(page, order);
page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists In order to reduce fragmentation, this patch classifies freed pages in two groups according to their probability of being part of a high order merge. Pages belonging to a compound whose next-highest buddy is free are more likely to be part of a high order merge in the near future, so they will be added at the tail of the freelist. The remaining pages are put at the front of the freelist. In this way, the pages that are more likely to cause a big merge are kept free longer. Consequently there is a tendency to aggregate the long-living allocations on a subset of the compounds, reducing the fragmentation. This heuristic was tested on three machines, x86, x86-64 and ppc64 with 3GB of RAM in each machine. The tests were kernbench, netperf, sysbench and STREAM for performance and a high-order stress test for huge page allocations. KernBench X86 Elapsed mean 374.77 ( 0.00%) 375.10 (-0.09%) User mean 649.53 ( 0.00%) 650.44 (-0.14%) System mean 54.75 ( 0.00%) 54.18 ( 1.05%) CPU mean 187.75 ( 0.00%) 187.25 ( 0.27%) KernBench X86-64 Elapsed mean 94.45 ( 0.00%) 94.01 ( 0.47%) User mean 323.27 ( 0.00%) 322.66 ( 0.19%) System mean 36.71 ( 0.00%) 36.50 ( 0.57%) CPU mean 380.75 ( 0.00%) 381.75 (-0.26%) KernBench PPC64 Elapsed mean 173.45 ( 0.00%) 173.74 (-0.17%) User mean 587.99 ( 0.00%) 587.95 ( 0.01%) System mean 60.60 ( 0.00%) 60.57 ( 0.05%) CPU mean 373.50 ( 0.00%) 372.75 ( 0.20%) Nothing notable for kernbench. NetPerf UDP X86 64 42.68 ( 0.00%) 42.77 ( 0.21%) 128 85.62 ( 0.00%) 85.32 (-0.35%) 256 170.01 ( 0.00%) 168.76 (-0.74%) 1024 655.68 ( 0.00%) 652.33 (-0.51%) 2048 1262.39 ( 0.00%) 1248.61 (-1.10%) 3312 1958.41 ( 0.00%) 1944.61 (-0.71%) 4096 2345.63 ( 0.00%) 2318.83 (-1.16%) 8192 4132.90 ( 0.00%) 4089.50 (-1.06%) 16384 6770.88 ( 0.00%) 6642.05 (-1.94%)* NetPerf UDP X86-64 64 148.82 ( 0.00%) 154.92 ( 3.94%) 128 298.96 ( 0.00%) 312.95 ( 4.47%) 256 583.67 ( 0.00%) 626.39 ( 6.82%) 1024 2293.18 ( 0.00%) 2371.10 ( 3.29%) 2048 4274.16 ( 0.00%) 4396.83 ( 2.79%) 3312 6356.94 ( 0.00%) 6571.35 ( 3.26%) 4096 7422.68 ( 0.00%) 7635.42 ( 2.79%)* 8192 12114.81 ( 0.00%)* 12346.88 ( 1.88%) 16384 17022.28 ( 0.00%)* 17033.19 ( 0.06%)* 1.64% 2.73% NetPerf UDP PPC64 64 49.98 ( 0.00%) 50.25 ( 0.54%) 128 98.66 ( 0.00%) 100.95 ( 2.27%) 256 197.33 ( 0.00%) 191.03 (-3.30%) 1024 761.98 ( 0.00%) 785.07 ( 2.94%) 2048 1493.50 ( 0.00%) 1510.85 ( 1.15%) 3312 2303.95 ( 0.00%) 2271.72 (-1.42%) 4096 2774.56 ( 0.00%) 2773.06 (-0.05%) 8192 4918.31 ( 0.00%) 4793.59 (-2.60%) 16384 7497.98 ( 0.00%) 7749.52 ( 3.25%) The tests are run to have confidence limits within 1%. Results marked with a * were not confident although in this case, it's only outside by small amounts. Even with some results that were not confident, the netperf UDP results were generally positive. NetPerf TCP X86 64 652.25 ( 0.00%)* 648.12 (-0.64%)* 23.80% 22.82% 128 1229.98 ( 0.00%)* 1220.56 (-0.77%)* 21.03% 18.90% 256 2105.88 ( 0.00%) 1872.03 (-12.49%)* 1.00% 16.46% 1024 3476.46 ( 0.00%)* 3548.28 ( 2.02%)* 13.37% 11.39% 2048 4023.44 ( 0.00%)* 4231.45 ( 4.92%)* 9.76% 12.48% 3312 4348.88 ( 0.00%)* 4396.96 ( 1.09%)* 6.49% 8.75% 4096 4726.56 ( 0.00%)* 4877.71 ( 3.10%)* 9.85% 8.50% 8192 4732.28 ( 0.00%)* 5777.77 (18.10%)* 9.13% 13.04% 16384 5543.05 ( 0.00%)* 5906.24 ( 6.15%)* 7.73% 8.68% NETPERF TCP X86-64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 1895.87 ( 0.00%)* 1775.07 (-6.81%)* 5.79% 4.78% 128 3571.03 ( 0.00%)* 3342.20 (-6.85%)* 3.68% 6.06% 256 5097.21 ( 0.00%)* 4859.43 (-4.89%)* 3.02% 2.10% 1024 8919.10 ( 0.00%)* 8892.49 (-0.30%)* 5.89% 6.55% 2048 10255.46 ( 0.00%)* 10449.39 ( 1.86%)* 7.08% 7.44% 3312 10839.90 ( 0.00%)* 10740.15 (-0.93%)* 6.87% 7.33% 4096 10814.84 ( 0.00%)* 10766.97 (-0.44%)* 6.86% 8.18% 8192 11606.89 ( 0.00%)* 11189.28 (-3.73%)* 7.49% 5.55% 16384 12554.88 ( 0.00%)* 12361.22 (-1.57%)* 7.36% 6.49% NETPERF TCP PPC64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 594.17 ( 0.00%) 596.04 ( 0.31%)* 1.00% 2.29% 128 1064.87 ( 0.00%)* 1074.77 ( 0.92%)* 1.30% 1.40% 256 1852.46 ( 0.00%)* 1856.95 ( 0.24%) 1.25% 1.00% 1024 3839.46 ( 0.00%)* 3813.05 (-0.69%) 1.02% 1.00% 2048 4885.04 ( 0.00%)* 4881.97 (-0.06%)* 1.15% 1.04% 3312 5506.90 ( 0.00%) 5459.72 (-0.86%) 4096 6449.19 ( 0.00%) 6345.46 (-1.63%) 8192 7501.17 ( 0.00%) 7508.79 ( 0.10%) 16384 9618.65 ( 0.00%) 9490.10 (-1.35%) There was a distinct lack of confidence in the X86* figures so I included what the devation was where the results were not confident. Many of the results, whether gains or losses were within the standard deviation so no solid conclusion can be reached on performance impact. Looking at the figures, only the X86-64 ones look suspicious with a few losses that were outside the noise. However, the results were so unstable that without knowing why they vary so much, a solid conclusion cannot be reached. SYSBENCH X86 sysbench-vanilla pgalloc-delay 1 7722.85 ( 0.00%) 7756.79 ( 0.44%) 2 14901.11 ( 0.00%) 13683.44 (-8.90%) 3 15171.71 ( 0.00%) 14888.25 (-1.90%) 4 14966.98 ( 0.00%) 15029.67 ( 0.42%) 5 14370.47 ( 0.00%) 14865.00 ( 3.33%) 6 14870.33 ( 0.00%) 14845.57 (-0.17%) 7 14429.45 ( 0.00%) 14520.85 ( 0.63%) 8 14354.35 ( 0.00%) 14362.31 ( 0.06%) SYSBENCH X86-64 1 17448.70 ( 0.00%) 17484.41 ( 0.20%) 2 34276.39 ( 0.00%) 34251.00 (-0.07%) 3 50805.25 ( 0.00%) 50854.80 ( 0.10%) 4 66667.10 ( 0.00%) 66174.69 (-0.74%) 5 66003.91 ( 0.00%) 65685.25 (-0.49%) 6 64981.90 ( 0.00%) 65125.60 ( 0.22%) 7 64933.16 ( 0.00%) 64379.23 (-0.86%) 8 63353.30 ( 0.00%) 63281.22 (-0.11%) 9 63511.84 ( 0.00%) 63570.37 ( 0.09%) 10 62708.27 ( 0.00%) 63166.25 ( 0.73%) 11 62092.81 ( 0.00%) 61787.75 (-0.49%) 12 61330.11 ( 0.00%) 61036.34 (-0.48%) 13 61438.37 ( 0.00%) 61994.47 ( 0.90%) 14 62304.48 ( 0.00%) 62064.90 (-0.39%) 15 63296.48 ( 0.00%) 62875.16 (-0.67%) 16 63951.76 ( 0.00%) 63769.09 (-0.29%) SYSBENCH PPC64 -sysbench-pgalloc-delay-sysbench sysbench-vanilla pgalloc-delay 1 7645.08 ( 0.00%) 7467.43 (-2.38%) 2 14856.67 ( 0.00%) 14558.73 (-2.05%) 3 21952.31 ( 0.00%) 21683.64 (-1.24%) 4 27946.09 ( 0.00%) 28623.29 ( 2.37%) 5 28045.11 ( 0.00%) 28143.69 ( 0.35%) 6 27477.10 ( 0.00%) 27337.45 (-0.51%) 7 26489.17 ( 0.00%) 26590.06 ( 0.38%) 8 26642.91 ( 0.00%) 25274.33 (-5.41%) 9 25137.27 ( 0.00%) 24810.06 (-1.32%) 10 24451.99 ( 0.00%) 24275.85 (-0.73%) 11 23262.20 ( 0.00%) 23674.88 ( 1.74%) 12 24234.81 ( 0.00%) 23640.89 (-2.51%) 13 24577.75 ( 0.00%) 24433.50 (-0.59%) 14 25640.19 ( 0.00%) 25116.52 (-2.08%) 15 26188.84 ( 0.00%) 26181.36 (-0.03%) 16 26782.37 ( 0.00%) 26255.99 (-2.00%) Again, there is little to conclude here. While there are a few losses, the results vary by +/- 8% in some cases. They are the results of most concern as there are some large losses but it's also within the variance typically seen between kernel releases. The STREAM results varied so little and are so verbose that I didn't include them here. The final test stressed how many huge pages can be allocated. The absolute number of huge pages allocated are the same with or without the page. However, the "unusability free space index" which is a measure of external fragmentation was slightly lower (lower is better) throughout the lifetime of the system. I also measured the latency of how long it took to successfully allocate a huge page. The latency was slightly lower and on X86 and PPC64, more huge pages were allocated almost immediately from the free lists. The improvement is slight but there. [mel@csn.ul.ie: Tested, reworked for less branches] [czoccolo@gmail.com: fix oops by checking pfn_valid_within()] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Corrado Zoccolo <czoccolo@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:31:54 +00:00
/*
* If this is not the largest possible page, check if the buddy
* of the next-highest order is free. If it is, it's possible
* that pages are being freed that will coalesce soon. In case,
* that is happening, add the free page to the tail of the list
* so it's less likely to be used soon and more likely to be merged
* as a higher order page
*/
mm, page-allocator: do not check the state of a non-existant buddy during free There is a bug in commit 6dda9d55 ("page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists") that means a buddy at order MAX_ORDER is checked for merging. A page of this order never exists so at times, an effectively random piece of memory is being checked. Alan Curry has reported that this is causing memory corruption in userspace data on a PPC32 platform (http://lkml.org/lkml/2010/10/9/32). It is not clear why this is happening. It could be a cache coherency problem where pages mapped in both user and kernel space are getting different cache lines due to the bad read from kernel space (http://lkml.org/lkml/2010/10/13/179). It could also be that there are some special registers being io-remapped at the end of the memmap array and that a read has special meaning on them. Compiler bugs have been ruled out because the assembly before and after the patch looks relatively harmless. This patch fixes the problem by ensuring we are not reading a possibly invalid location of memory. It's not clear why the read causes corruption but one way or the other it is a buggy read. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Corrado Zoccolo <czoccolo@gmail.com> Reported-by: Alan Curry <pacman@kosh.dhis.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Rik van Riel <riel@redhat.com> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-26 21:21:11 +00:00
if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists In order to reduce fragmentation, this patch classifies freed pages in two groups according to their probability of being part of a high order merge. Pages belonging to a compound whose next-highest buddy is free are more likely to be part of a high order merge in the near future, so they will be added at the tail of the freelist. The remaining pages are put at the front of the freelist. In this way, the pages that are more likely to cause a big merge are kept free longer. Consequently there is a tendency to aggregate the long-living allocations on a subset of the compounds, reducing the fragmentation. This heuristic was tested on three machines, x86, x86-64 and ppc64 with 3GB of RAM in each machine. The tests were kernbench, netperf, sysbench and STREAM for performance and a high-order stress test for huge page allocations. KernBench X86 Elapsed mean 374.77 ( 0.00%) 375.10 (-0.09%) User mean 649.53 ( 0.00%) 650.44 (-0.14%) System mean 54.75 ( 0.00%) 54.18 ( 1.05%) CPU mean 187.75 ( 0.00%) 187.25 ( 0.27%) KernBench X86-64 Elapsed mean 94.45 ( 0.00%) 94.01 ( 0.47%) User mean 323.27 ( 0.00%) 322.66 ( 0.19%) System mean 36.71 ( 0.00%) 36.50 ( 0.57%) CPU mean 380.75 ( 0.00%) 381.75 (-0.26%) KernBench PPC64 Elapsed mean 173.45 ( 0.00%) 173.74 (-0.17%) User mean 587.99 ( 0.00%) 587.95 ( 0.01%) System mean 60.60 ( 0.00%) 60.57 ( 0.05%) CPU mean 373.50 ( 0.00%) 372.75 ( 0.20%) Nothing notable for kernbench. NetPerf UDP X86 64 42.68 ( 0.00%) 42.77 ( 0.21%) 128 85.62 ( 0.00%) 85.32 (-0.35%) 256 170.01 ( 0.00%) 168.76 (-0.74%) 1024 655.68 ( 0.00%) 652.33 (-0.51%) 2048 1262.39 ( 0.00%) 1248.61 (-1.10%) 3312 1958.41 ( 0.00%) 1944.61 (-0.71%) 4096 2345.63 ( 0.00%) 2318.83 (-1.16%) 8192 4132.90 ( 0.00%) 4089.50 (-1.06%) 16384 6770.88 ( 0.00%) 6642.05 (-1.94%)* NetPerf UDP X86-64 64 148.82 ( 0.00%) 154.92 ( 3.94%) 128 298.96 ( 0.00%) 312.95 ( 4.47%) 256 583.67 ( 0.00%) 626.39 ( 6.82%) 1024 2293.18 ( 0.00%) 2371.10 ( 3.29%) 2048 4274.16 ( 0.00%) 4396.83 ( 2.79%) 3312 6356.94 ( 0.00%) 6571.35 ( 3.26%) 4096 7422.68 ( 0.00%) 7635.42 ( 2.79%)* 8192 12114.81 ( 0.00%)* 12346.88 ( 1.88%) 16384 17022.28 ( 0.00%)* 17033.19 ( 0.06%)* 1.64% 2.73% NetPerf UDP PPC64 64 49.98 ( 0.00%) 50.25 ( 0.54%) 128 98.66 ( 0.00%) 100.95 ( 2.27%) 256 197.33 ( 0.00%) 191.03 (-3.30%) 1024 761.98 ( 0.00%) 785.07 ( 2.94%) 2048 1493.50 ( 0.00%) 1510.85 ( 1.15%) 3312 2303.95 ( 0.00%) 2271.72 (-1.42%) 4096 2774.56 ( 0.00%) 2773.06 (-0.05%) 8192 4918.31 ( 0.00%) 4793.59 (-2.60%) 16384 7497.98 ( 0.00%) 7749.52 ( 3.25%) The tests are run to have confidence limits within 1%. Results marked with a * were not confident although in this case, it's only outside by small amounts. Even with some results that were not confident, the netperf UDP results were generally positive. NetPerf TCP X86 64 652.25 ( 0.00%)* 648.12 (-0.64%)* 23.80% 22.82% 128 1229.98 ( 0.00%)* 1220.56 (-0.77%)* 21.03% 18.90% 256 2105.88 ( 0.00%) 1872.03 (-12.49%)* 1.00% 16.46% 1024 3476.46 ( 0.00%)* 3548.28 ( 2.02%)* 13.37% 11.39% 2048 4023.44 ( 0.00%)* 4231.45 ( 4.92%)* 9.76% 12.48% 3312 4348.88 ( 0.00%)* 4396.96 ( 1.09%)* 6.49% 8.75% 4096 4726.56 ( 0.00%)* 4877.71 ( 3.10%)* 9.85% 8.50% 8192 4732.28 ( 0.00%)* 5777.77 (18.10%)* 9.13% 13.04% 16384 5543.05 ( 0.00%)* 5906.24 ( 6.15%)* 7.73% 8.68% NETPERF TCP X86-64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 1895.87 ( 0.00%)* 1775.07 (-6.81%)* 5.79% 4.78% 128 3571.03 ( 0.00%)* 3342.20 (-6.85%)* 3.68% 6.06% 256 5097.21 ( 0.00%)* 4859.43 (-4.89%)* 3.02% 2.10% 1024 8919.10 ( 0.00%)* 8892.49 (-0.30%)* 5.89% 6.55% 2048 10255.46 ( 0.00%)* 10449.39 ( 1.86%)* 7.08% 7.44% 3312 10839.90 ( 0.00%)* 10740.15 (-0.93%)* 6.87% 7.33% 4096 10814.84 ( 0.00%)* 10766.97 (-0.44%)* 6.86% 8.18% 8192 11606.89 ( 0.00%)* 11189.28 (-3.73%)* 7.49% 5.55% 16384 12554.88 ( 0.00%)* 12361.22 (-1.57%)* 7.36% 6.49% NETPERF TCP PPC64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 594.17 ( 0.00%) 596.04 ( 0.31%)* 1.00% 2.29% 128 1064.87 ( 0.00%)* 1074.77 ( 0.92%)* 1.30% 1.40% 256 1852.46 ( 0.00%)* 1856.95 ( 0.24%) 1.25% 1.00% 1024 3839.46 ( 0.00%)* 3813.05 (-0.69%) 1.02% 1.00% 2048 4885.04 ( 0.00%)* 4881.97 (-0.06%)* 1.15% 1.04% 3312 5506.90 ( 0.00%) 5459.72 (-0.86%) 4096 6449.19 ( 0.00%) 6345.46 (-1.63%) 8192 7501.17 ( 0.00%) 7508.79 ( 0.10%) 16384 9618.65 ( 0.00%) 9490.10 (-1.35%) There was a distinct lack of confidence in the X86* figures so I included what the devation was where the results were not confident. Many of the results, whether gains or losses were within the standard deviation so no solid conclusion can be reached on performance impact. Looking at the figures, only the X86-64 ones look suspicious with a few losses that were outside the noise. However, the results were so unstable that without knowing why they vary so much, a solid conclusion cannot be reached. SYSBENCH X86 sysbench-vanilla pgalloc-delay 1 7722.85 ( 0.00%) 7756.79 ( 0.44%) 2 14901.11 ( 0.00%) 13683.44 (-8.90%) 3 15171.71 ( 0.00%) 14888.25 (-1.90%) 4 14966.98 ( 0.00%) 15029.67 ( 0.42%) 5 14370.47 ( 0.00%) 14865.00 ( 3.33%) 6 14870.33 ( 0.00%) 14845.57 (-0.17%) 7 14429.45 ( 0.00%) 14520.85 ( 0.63%) 8 14354.35 ( 0.00%) 14362.31 ( 0.06%) SYSBENCH X86-64 1 17448.70 ( 0.00%) 17484.41 ( 0.20%) 2 34276.39 ( 0.00%) 34251.00 (-0.07%) 3 50805.25 ( 0.00%) 50854.80 ( 0.10%) 4 66667.10 ( 0.00%) 66174.69 (-0.74%) 5 66003.91 ( 0.00%) 65685.25 (-0.49%) 6 64981.90 ( 0.00%) 65125.60 ( 0.22%) 7 64933.16 ( 0.00%) 64379.23 (-0.86%) 8 63353.30 ( 0.00%) 63281.22 (-0.11%) 9 63511.84 ( 0.00%) 63570.37 ( 0.09%) 10 62708.27 ( 0.00%) 63166.25 ( 0.73%) 11 62092.81 ( 0.00%) 61787.75 (-0.49%) 12 61330.11 ( 0.00%) 61036.34 (-0.48%) 13 61438.37 ( 0.00%) 61994.47 ( 0.90%) 14 62304.48 ( 0.00%) 62064.90 (-0.39%) 15 63296.48 ( 0.00%) 62875.16 (-0.67%) 16 63951.76 ( 0.00%) 63769.09 (-0.29%) SYSBENCH PPC64 -sysbench-pgalloc-delay-sysbench sysbench-vanilla pgalloc-delay 1 7645.08 ( 0.00%) 7467.43 (-2.38%) 2 14856.67 ( 0.00%) 14558.73 (-2.05%) 3 21952.31 ( 0.00%) 21683.64 (-1.24%) 4 27946.09 ( 0.00%) 28623.29 ( 2.37%) 5 28045.11 ( 0.00%) 28143.69 ( 0.35%) 6 27477.10 ( 0.00%) 27337.45 (-0.51%) 7 26489.17 ( 0.00%) 26590.06 ( 0.38%) 8 26642.91 ( 0.00%) 25274.33 (-5.41%) 9 25137.27 ( 0.00%) 24810.06 (-1.32%) 10 24451.99 ( 0.00%) 24275.85 (-0.73%) 11 23262.20 ( 0.00%) 23674.88 ( 1.74%) 12 24234.81 ( 0.00%) 23640.89 (-2.51%) 13 24577.75 ( 0.00%) 24433.50 (-0.59%) 14 25640.19 ( 0.00%) 25116.52 (-2.08%) 15 26188.84 ( 0.00%) 26181.36 (-0.03%) 16 26782.37 ( 0.00%) 26255.99 (-2.00%) Again, there is little to conclude here. While there are a few losses, the results vary by +/- 8% in some cases. They are the results of most concern as there are some large losses but it's also within the variance typically seen between kernel releases. The STREAM results varied so little and are so verbose that I didn't include them here. The final test stressed how many huge pages can be allocated. The absolute number of huge pages allocated are the same with or without the page. However, the "unusability free space index" which is a measure of external fragmentation was slightly lower (lower is better) throughout the lifetime of the system. I also measured the latency of how long it took to successfully allocate a huge page. The latency was slightly lower and on X86 and PPC64, more huge pages were allocated almost immediately from the free lists. The improvement is slight but there. [mel@csn.ul.ie: Tested, reworked for less branches] [czoccolo@gmail.com: fix oops by checking pfn_valid_within()] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Corrado Zoccolo <czoccolo@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:31:54 +00:00
struct page *higher_page, *higher_buddy;
combined_idx = buddy_idx & page_idx;
higher_page = page + (combined_idx - page_idx);
buddy_idx = __find_buddy_index(combined_idx, order + 1);
higher_buddy = page + (buddy_idx - combined_idx);
page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists In order to reduce fragmentation, this patch classifies freed pages in two groups according to their probability of being part of a high order merge. Pages belonging to a compound whose next-highest buddy is free are more likely to be part of a high order merge in the near future, so they will be added at the tail of the freelist. The remaining pages are put at the front of the freelist. In this way, the pages that are more likely to cause a big merge are kept free longer. Consequently there is a tendency to aggregate the long-living allocations on a subset of the compounds, reducing the fragmentation. This heuristic was tested on three machines, x86, x86-64 and ppc64 with 3GB of RAM in each machine. The tests were kernbench, netperf, sysbench and STREAM for performance and a high-order stress test for huge page allocations. KernBench X86 Elapsed mean 374.77 ( 0.00%) 375.10 (-0.09%) User mean 649.53 ( 0.00%) 650.44 (-0.14%) System mean 54.75 ( 0.00%) 54.18 ( 1.05%) CPU mean 187.75 ( 0.00%) 187.25 ( 0.27%) KernBench X86-64 Elapsed mean 94.45 ( 0.00%) 94.01 ( 0.47%) User mean 323.27 ( 0.00%) 322.66 ( 0.19%) System mean 36.71 ( 0.00%) 36.50 ( 0.57%) CPU mean 380.75 ( 0.00%) 381.75 (-0.26%) KernBench PPC64 Elapsed mean 173.45 ( 0.00%) 173.74 (-0.17%) User mean 587.99 ( 0.00%) 587.95 ( 0.01%) System mean 60.60 ( 0.00%) 60.57 ( 0.05%) CPU mean 373.50 ( 0.00%) 372.75 ( 0.20%) Nothing notable for kernbench. NetPerf UDP X86 64 42.68 ( 0.00%) 42.77 ( 0.21%) 128 85.62 ( 0.00%) 85.32 (-0.35%) 256 170.01 ( 0.00%) 168.76 (-0.74%) 1024 655.68 ( 0.00%) 652.33 (-0.51%) 2048 1262.39 ( 0.00%) 1248.61 (-1.10%) 3312 1958.41 ( 0.00%) 1944.61 (-0.71%) 4096 2345.63 ( 0.00%) 2318.83 (-1.16%) 8192 4132.90 ( 0.00%) 4089.50 (-1.06%) 16384 6770.88 ( 0.00%) 6642.05 (-1.94%)* NetPerf UDP X86-64 64 148.82 ( 0.00%) 154.92 ( 3.94%) 128 298.96 ( 0.00%) 312.95 ( 4.47%) 256 583.67 ( 0.00%) 626.39 ( 6.82%) 1024 2293.18 ( 0.00%) 2371.10 ( 3.29%) 2048 4274.16 ( 0.00%) 4396.83 ( 2.79%) 3312 6356.94 ( 0.00%) 6571.35 ( 3.26%) 4096 7422.68 ( 0.00%) 7635.42 ( 2.79%)* 8192 12114.81 ( 0.00%)* 12346.88 ( 1.88%) 16384 17022.28 ( 0.00%)* 17033.19 ( 0.06%)* 1.64% 2.73% NetPerf UDP PPC64 64 49.98 ( 0.00%) 50.25 ( 0.54%) 128 98.66 ( 0.00%) 100.95 ( 2.27%) 256 197.33 ( 0.00%) 191.03 (-3.30%) 1024 761.98 ( 0.00%) 785.07 ( 2.94%) 2048 1493.50 ( 0.00%) 1510.85 ( 1.15%) 3312 2303.95 ( 0.00%) 2271.72 (-1.42%) 4096 2774.56 ( 0.00%) 2773.06 (-0.05%) 8192 4918.31 ( 0.00%) 4793.59 (-2.60%) 16384 7497.98 ( 0.00%) 7749.52 ( 3.25%) The tests are run to have confidence limits within 1%. Results marked with a * were not confident although in this case, it's only outside by small amounts. Even with some results that were not confident, the netperf UDP results were generally positive. NetPerf TCP X86 64 652.25 ( 0.00%)* 648.12 (-0.64%)* 23.80% 22.82% 128 1229.98 ( 0.00%)* 1220.56 (-0.77%)* 21.03% 18.90% 256 2105.88 ( 0.00%) 1872.03 (-12.49%)* 1.00% 16.46% 1024 3476.46 ( 0.00%)* 3548.28 ( 2.02%)* 13.37% 11.39% 2048 4023.44 ( 0.00%)* 4231.45 ( 4.92%)* 9.76% 12.48% 3312 4348.88 ( 0.00%)* 4396.96 ( 1.09%)* 6.49% 8.75% 4096 4726.56 ( 0.00%)* 4877.71 ( 3.10%)* 9.85% 8.50% 8192 4732.28 ( 0.00%)* 5777.77 (18.10%)* 9.13% 13.04% 16384 5543.05 ( 0.00%)* 5906.24 ( 6.15%)* 7.73% 8.68% NETPERF TCP X86-64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 1895.87 ( 0.00%)* 1775.07 (-6.81%)* 5.79% 4.78% 128 3571.03 ( 0.00%)* 3342.20 (-6.85%)* 3.68% 6.06% 256 5097.21 ( 0.00%)* 4859.43 (-4.89%)* 3.02% 2.10% 1024 8919.10 ( 0.00%)* 8892.49 (-0.30%)* 5.89% 6.55% 2048 10255.46 ( 0.00%)* 10449.39 ( 1.86%)* 7.08% 7.44% 3312 10839.90 ( 0.00%)* 10740.15 (-0.93%)* 6.87% 7.33% 4096 10814.84 ( 0.00%)* 10766.97 (-0.44%)* 6.86% 8.18% 8192 11606.89 ( 0.00%)* 11189.28 (-3.73%)* 7.49% 5.55% 16384 12554.88 ( 0.00%)* 12361.22 (-1.57%)* 7.36% 6.49% NETPERF TCP PPC64 netperf-tcp-vanilla-netperf netperf-tcp tcp-vanilla pgalloc-delay 64 594.17 ( 0.00%) 596.04 ( 0.31%)* 1.00% 2.29% 128 1064.87 ( 0.00%)* 1074.77 ( 0.92%)* 1.30% 1.40% 256 1852.46 ( 0.00%)* 1856.95 ( 0.24%) 1.25% 1.00% 1024 3839.46 ( 0.00%)* 3813.05 (-0.69%) 1.02% 1.00% 2048 4885.04 ( 0.00%)* 4881.97 (-0.06%)* 1.15% 1.04% 3312 5506.90 ( 0.00%) 5459.72 (-0.86%) 4096 6449.19 ( 0.00%) 6345.46 (-1.63%) 8192 7501.17 ( 0.00%) 7508.79 ( 0.10%) 16384 9618.65 ( 0.00%) 9490.10 (-1.35%) There was a distinct lack of confidence in the X86* figures so I included what the devation was where the results were not confident. Many of the results, whether gains or losses were within the standard deviation so no solid conclusion can be reached on performance impact. Looking at the figures, only the X86-64 ones look suspicious with a few losses that were outside the noise. However, the results were so unstable that without knowing why they vary so much, a solid conclusion cannot be reached. SYSBENCH X86 sysbench-vanilla pgalloc-delay 1 7722.85 ( 0.00%) 7756.79 ( 0.44%) 2 14901.11 ( 0.00%) 13683.44 (-8.90%) 3 15171.71 ( 0.00%) 14888.25 (-1.90%) 4 14966.98 ( 0.00%) 15029.67 ( 0.42%) 5 14370.47 ( 0.00%) 14865.00 ( 3.33%) 6 14870.33 ( 0.00%) 14845.57 (-0.17%) 7 14429.45 ( 0.00%) 14520.85 ( 0.63%) 8 14354.35 ( 0.00%) 14362.31 ( 0.06%) SYSBENCH X86-64 1 17448.70 ( 0.00%) 17484.41 ( 0.20%) 2 34276.39 ( 0.00%) 34251.00 (-0.07%) 3 50805.25 ( 0.00%) 50854.80 ( 0.10%) 4 66667.10 ( 0.00%) 66174.69 (-0.74%) 5 66003.91 ( 0.00%) 65685.25 (-0.49%) 6 64981.90 ( 0.00%) 65125.60 ( 0.22%) 7 64933.16 ( 0.00%) 64379.23 (-0.86%) 8 63353.30 ( 0.00%) 63281.22 (-0.11%) 9 63511.84 ( 0.00%) 63570.37 ( 0.09%) 10 62708.27 ( 0.00%) 63166.25 ( 0.73%) 11 62092.81 ( 0.00%) 61787.75 (-0.49%) 12 61330.11 ( 0.00%) 61036.34 (-0.48%) 13 61438.37 ( 0.00%) 61994.47 ( 0.90%) 14 62304.48 ( 0.00%) 62064.90 (-0.39%) 15 63296.48 ( 0.00%) 62875.16 (-0.67%) 16 63951.76 ( 0.00%) 63769.09 (-0.29%) SYSBENCH PPC64 -sysbench-pgalloc-delay-sysbench sysbench-vanilla pgalloc-delay 1 7645.08 ( 0.00%) 7467.43 (-2.38%) 2 14856.67 ( 0.00%) 14558.73 (-2.05%) 3 21952.31 ( 0.00%) 21683.64 (-1.24%) 4 27946.09 ( 0.00%) 28623.29 ( 2.37%) 5 28045.11 ( 0.00%) 28143.69 ( 0.35%) 6 27477.10 ( 0.00%) 27337.45 (-0.51%) 7 26489.17 ( 0.00%) 26590.06 ( 0.38%) 8 26642.91 ( 0.00%) 25274.33 (-5.41%) 9 25137.27 ( 0.00%) 24810.06 (-1.32%) 10 24451.99 ( 0.00%) 24275.85 (-0.73%) 11 23262.20 ( 0.00%) 23674.88 ( 1.74%) 12 24234.81 ( 0.00%) 23640.89 (-2.51%) 13 24577.75 ( 0.00%) 24433.50 (-0.59%) 14 25640.19 ( 0.00%) 25116.52 (-2.08%) 15 26188.84 ( 0.00%) 26181.36 (-0.03%) 16 26782.37 ( 0.00%) 26255.99 (-2.00%) Again, there is little to conclude here. While there are a few losses, the results vary by +/- 8% in some cases. They are the results of most concern as there are some large losses but it's also within the variance typically seen between kernel releases. The STREAM results varied so little and are so verbose that I didn't include them here. The final test stressed how many huge pages can be allocated. The absolute number of huge pages allocated are the same with or without the page. However, the "unusability free space index" which is a measure of external fragmentation was slightly lower (lower is better) throughout the lifetime of the system. I also measured the latency of how long it took to successfully allocate a huge page. The latency was slightly lower and on X86 and PPC64, more huge pages were allocated almost immediately from the free lists. The improvement is slight but there. [mel@csn.ul.ie: Tested, reworked for less branches] [czoccolo@gmail.com: fix oops by checking pfn_valid_within()] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Corrado Zoccolo <czoccolo@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:31:54 +00:00
if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
list_add_tail(&page->lru,
&zone->free_area[order].free_list[migratetype]);
goto out;
}
}
list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
out:
zone->free_area[order].nr_free++;
}
/*
* free_page_mlock() -- clean up attempts to free and mlocked() page.
* Page should not be on lru, so no need to fix that up.
* free_pages_check() will verify...
*/
static inline void free_page_mlock(struct page *page)
{
__dec_zone_page_state(page, NR_MLOCK);
__count_vm_event(UNEVICTABLE_MLOCKFREED);
}
static inline int free_pages_check(struct page *page)
{
if (unlikely(page_mapcount(page) |
(page->mapping != NULL) |
(atomic_read(&page->_count) != 0) |
(page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
(mem_cgroup_bad_page_check(page)))) {
bad_page(page);
return 1;
}
if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
return 0;
}
/*
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
* Frees a number of pages from the PCP lists
* Assumes all pages on list are in same zone, and of same order.
* count is the number of pages to free.
*
* If the zone was previously in an "all pages pinned" state then look to
* see if this freeing clears that state.
*
* And clear the zone's pages_scanned counter, to hold off the "all pages are
* pinned" detection logic.
*/
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
static void free_pcppages_bulk(struct zone *zone, int count,
struct per_cpu_pages *pcp)
{
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
int migratetype = 0;
int batch_free = 0;
mm: page allocator: update free page counters after pages are placed on the free list When allocating a page, the system uses NR_FREE_PAGES counters to determine if watermarks would remain intact after the allocation was made. This check is made without interrupts disabled or the zone lock held and so is race-prone by nature. Unfortunately, when pages are being freed in batch, the counters are updated before the pages are added on the list. During this window, the counters are misleading as the pages do not exist yet. When under significant pressure on systems with large numbers of CPUs, it's possible for processes to make progress even though they should have been stalled. This is particularly problematic if a number of the processes are using GFP_ATOMIC as the min watermark can be accidentally breached and in extreme cases, the system can livelock. This patch updates the counters after the pages have been added to the list. This makes the allocator more cautious with respect to preserving the watermarks and mitigates livelock possibilities. [akpm@linux-foundation.org: avoid modifying incoming args] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Christoph Lameter <cl@linux.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-09-09 23:38:16 +00:00
int to_free = count;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
spin_lock(&zone->lock);
zone->all_unreclaimable = 0;
zone->pages_scanned = 0;
mm: page allocator: update free page counters after pages are placed on the free list When allocating a page, the system uses NR_FREE_PAGES counters to determine if watermarks would remain intact after the allocation was made. This check is made without interrupts disabled or the zone lock held and so is race-prone by nature. Unfortunately, when pages are being freed in batch, the counters are updated before the pages are added on the list. During this window, the counters are misleading as the pages do not exist yet. When under significant pressure on systems with large numbers of CPUs, it's possible for processes to make progress even though they should have been stalled. This is particularly problematic if a number of the processes are using GFP_ATOMIC as the min watermark can be accidentally breached and in extreme cases, the system can livelock. This patch updates the counters after the pages have been added to the list. This makes the allocator more cautious with respect to preserving the watermarks and mitigates livelock possibilities. [akpm@linux-foundation.org: avoid modifying incoming args] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Christoph Lameter <cl@linux.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-09-09 23:38:16 +00:00
while (to_free) {
struct page *page;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
struct list_head *list;
/*
* Remove pages from lists in a round-robin fashion. A
* batch_free count is maintained that is incremented when an
* empty list is encountered. This is so more pages are freed
* off fuller lists instead of spinning excessively around empty
* lists
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
*/
do {
batch_free++;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
if (++migratetype == MIGRATE_PCPTYPES)
migratetype = 0;
list = &pcp->lists[migratetype];
} while (list_empty(list));
/* This is the only non-empty list. Free them all. */
if (batch_free == MIGRATE_PCPTYPES)
batch_free = to_free;
do {
page = list_entry(list->prev, struct page, lru);
/* must delete as __free_one_page list manipulates */
list_del(&page->lru);
mm: fix migratetype bug which slowed swapping After memory pressure has forced it to dip into the reserves, 2.6.32's 5f8dcc21211a3d4e3a7a5ca366b469fb88117f61 "page-allocator: split per-cpu list into one-list-per-migrate-type" has been returning MIGRATE_RESERVE pages to the MIGRATE_MOVABLE free_list: in some sense depleting reserves. Fix that in the most straightforward way (which, considering the overheads of alternative approaches, is Mel's preference): the right migratetype is already in page_private(page), but free_pcppages_bulk() wasn't using it. How did this bug show up? As a 20% slowdown in my tmpfs loop kbuild swapping tests, on PowerMac G5 with SLUB allocator. Bisecting to that commit was easy, but explaining the magnitude of the slowdown not easy. The same effect appears, but much less markedly, with SLAB, and even less markedly on other machines (the PowerMac divides into fewer zones than x86, I think that may be a factor). We guess that lumpy reclaim of short-lived high-order pages is implicated in some way, and probably this bug has been tickling a poor decision somewhere in page reclaim. But instrumentation hasn't told me much, I've run out of time and imagination to determine exactly what's going on, and shouldn't hold up the fix any longer: it's valid, and might even fix other misbehaviours. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Acked-by: Mel Gorman <mel@csn.ul.ie> Cc: stable@kernel.org Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-01-29 17:46:34 +00:00
/* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
__free_one_page(page, zone, 0, page_private(page));
trace_mm_page_pcpu_drain(page, 0, page_private(page));
mm: page allocator: update free page counters after pages are placed on the free list When allocating a page, the system uses NR_FREE_PAGES counters to determine if watermarks would remain intact after the allocation was made. This check is made without interrupts disabled or the zone lock held and so is race-prone by nature. Unfortunately, when pages are being freed in batch, the counters are updated before the pages are added on the list. During this window, the counters are misleading as the pages do not exist yet. When under significant pressure on systems with large numbers of CPUs, it's possible for processes to make progress even though they should have been stalled. This is particularly problematic if a number of the processes are using GFP_ATOMIC as the min watermark can be accidentally breached and in extreme cases, the system can livelock. This patch updates the counters after the pages have been added to the list. This makes the allocator more cautious with respect to preserving the watermarks and mitigates livelock possibilities. [akpm@linux-foundation.org: avoid modifying incoming args] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Christoph Lameter <cl@linux.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-09-09 23:38:16 +00:00
} while (--to_free && --batch_free && !list_empty(list));
}
mm: page allocator: update free page counters after pages are placed on the free list When allocating a page, the system uses NR_FREE_PAGES counters to determine if watermarks would remain intact after the allocation was made. This check is made without interrupts disabled or the zone lock held and so is race-prone by nature. Unfortunately, when pages are being freed in batch, the counters are updated before the pages are added on the list. During this window, the counters are misleading as the pages do not exist yet. When under significant pressure on systems with large numbers of CPUs, it's possible for processes to make progress even though they should have been stalled. This is particularly problematic if a number of the processes are using GFP_ATOMIC as the min watermark can be accidentally breached and in extreme cases, the system can livelock. This patch updates the counters after the pages have been added to the list. This makes the allocator more cautious with respect to preserving the watermarks and mitigates livelock possibilities. [akpm@linux-foundation.org: avoid modifying incoming args] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Christoph Lameter <cl@linux.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-09-09 23:38:16 +00:00
__mod_zone_page_state(zone, NR_FREE_PAGES, count);
spin_unlock(&zone->lock);
}
static void free_one_page(struct zone *zone, struct page *page, int order,
int migratetype)
{
spin_lock(&zone->lock);
zone->all_unreclaimable = 0;
zone->pages_scanned = 0;
__free_one_page(page, zone, order, migratetype);
mm: page allocator: update free page counters after pages are placed on the free list When allocating a page, the system uses NR_FREE_PAGES counters to determine if watermarks would remain intact after the allocation was made. This check is made without interrupts disabled or the zone lock held and so is race-prone by nature. Unfortunately, when pages are being freed in batch, the counters are updated before the pages are added on the list. During this window, the counters are misleading as the pages do not exist yet. When under significant pressure on systems with large numbers of CPUs, it's possible for processes to make progress even though they should have been stalled. This is particularly problematic if a number of the processes are using GFP_ATOMIC as the min watermark can be accidentally breached and in extreme cases, the system can livelock. This patch updates the counters after the pages have been added to the list. This makes the allocator more cautious with respect to preserving the watermarks and mitigates livelock possibilities. [akpm@linux-foundation.org: avoid modifying incoming args] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Christoph Lameter <cl@linux.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-09-09 23:38:16 +00:00
__mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
spin_unlock(&zone->lock);
}
static bool free_pages_prepare(struct page *page, unsigned int order)
{
int i;
int bad = 0;
trace_mm_page_free_direct(page, order);
kmemcheck_free_shadow(page, order);
if (PageAnon(page))
page->mapping = NULL;
for (i = 0; i < (1 << order); i++)
bad += free_pages_check(page + i);
if (bad)
return false;
infrastructure to debug (dynamic) objects We can see an ever repeating problem pattern with objects of any kind in the kernel: 1) freeing of active objects 2) reinitialization of active objects Both problems can be hard to debug because the crash happens at a point where we have no chance to decode the root cause anymore. One problem spot are kernel timers, where the detection of the problem often happens in interrupt context and usually causes the machine to panic. While working on a timer related bug report I had to hack specialized code into the timer subsystem to get a reasonable hint for the root cause. This debug hack was fine for temporary use, but far from a mergeable solution due to the intrusiveness into the timer code. The code further lacked the ability to detect and report the root cause instantly and keep the system operational. Keeping the system operational is important to get hold of the debug information without special debugging aids like serial consoles and special knowledge of the bug reporter. The problems described above are not restricted to timers, but timers tend to expose it usually in a full system crash. Other objects are less explosive, but the symptoms caused by such mistakes can be even harder to debug. Instead of creating specialized debugging code for the timer subsystem a generic infrastructure is created which allows developers to verify their code and provides an easy to enable debug facility for users in case of trouble. The debugobjects core code keeps track of operations on static and dynamic objects by inserting them into a hashed list and sanity checking them on object operations and provides additional checks whenever kernel memory is freed. The tracked object operations are: - initializing an object - adding an object to a subsystem list - deleting an object from a subsystem list Each operation is sanity checked before the operation is executed and the subsystem specific code can provide a fixup function which allows to prevent the damage of the operation. When the sanity check triggers a warning message and a stack trace is printed. The list of operations can be extended if the need arises. For now it's limited to the requirements of the first user (timers). The core code enqueues the objects into hash buckets. The hash index is generated from the address of the object to simplify the lookup for the check on kfree/vfree. Each bucket has it's own spinlock to avoid contention on a global lock. The debug code can be compiled in without being active. The runtime overhead is minimal and could be optimized by asm alternatives. A kernel command line option enables the debugging code. Thanks to Ingo Molnar for review, suggestions and cleanup patches. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu> Cc: Greg KH <greg@kroah.com> Cc: Randy Dunlap <randy.dunlap@oracle.com> Cc: Kay Sievers <kay.sievers@vrfy.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 07:55:01 +00:00
if (!PageHighMem(page)) {
debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
infrastructure to debug (dynamic) objects We can see an ever repeating problem pattern with objects of any kind in the kernel: 1) freeing of active objects 2) reinitialization of active objects Both problems can be hard to debug because the crash happens at a point where we have no chance to decode the root cause anymore. One problem spot are kernel timers, where the detection of the problem often happens in interrupt context and usually causes the machine to panic. While working on a timer related bug report I had to hack specialized code into the timer subsystem to get a reasonable hint for the root cause. This debug hack was fine for temporary use, but far from a mergeable solution due to the intrusiveness into the timer code. The code further lacked the ability to detect and report the root cause instantly and keep the system operational. Keeping the system operational is important to get hold of the debug information without special debugging aids like serial consoles and special knowledge of the bug reporter. The problems described above are not restricted to timers, but timers tend to expose it usually in a full system crash. Other objects are less explosive, but the symptoms caused by such mistakes can be even harder to debug. Instead of creating specialized debugging code for the timer subsystem a generic infrastructure is created which allows developers to verify their code and provides an easy to enable debug facility for users in case of trouble. The debugobjects core code keeps track of operations on static and dynamic objects by inserting them into a hashed list and sanity checking them on object operations and provides additional checks whenever kernel memory is freed. The tracked object operations are: - initializing an object - adding an object to a subsystem list - deleting an object from a subsystem list Each operation is sanity checked before the operation is executed and the subsystem specific code can provide a fixup function which allows to prevent the damage of the operation. When the sanity check triggers a warning message and a stack trace is printed. The list of operations can be extended if the need arises. For now it's limited to the requirements of the first user (timers). The core code enqueues the objects into hash buckets. The hash index is generated from the address of the object to simplify the lookup for the check on kfree/vfree. Each bucket has it's own spinlock to avoid contention on a global lock. The debug code can be compiled in without being active. The runtime overhead is minimal and could be optimized by asm alternatives. A kernel command line option enables the debugging code. Thanks to Ingo Molnar for review, suggestions and cleanup patches. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu> Cc: Greg KH <greg@kroah.com> Cc: Randy Dunlap <randy.dunlap@oracle.com> Cc: Kay Sievers <kay.sievers@vrfy.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 07:55:01 +00:00
debug_check_no_obj_freed(page_address(page),
PAGE_SIZE << order);
}
arch_free_page(page, order);
kernel_map_pages(page, 1 << order, 0);
return true;
}
static void __free_pages_ok(struct page *page, unsigned int order)
{
unsigned long flags;
int wasMlocked = __TestClearPageMlocked(page);
if (!free_pages_prepare(page, order))
return;
local_irq_save(flags);
if (unlikely(wasMlocked))
free_page_mlock(page);
[PATCH] Light weight event counters The remaining counters in page_state after the zoned VM counter patches have been applied are all just for show in /proc/vmstat. They have no essential function for the VM. We use a simple increment of per cpu variables. In order to avoid the most severe races we disable preempt. Preempt does not prevent the race between an increment and an interrupt handler incrementing the same statistics counter. However, that race is exceedingly rare, we may only loose one increment or so and there is no requirement (at least not in kernel) that the vm event counters have to be accurate. In the non preempt case this results in a simple increment for each counter. For many architectures this will be reduced by the compiler to a single instruction. This single instruction is atomic for i386 and x86_64. And therefore even the rare race condition in an interrupt is avoided for both architectures in most cases. The patchset also adds an off switch for embedded systems that allows a building of linux kernels without these counters. The implementation of these counters is through inline code that hopefully results in only a single instruction increment instruction being emitted (i386, x86_64) or in the increment being hidden though instruction concurrency (EPIC architectures such as ia64 can get that done). Benefits: - VM event counter operations usually reduce to a single inline instruction on i386 and x86_64. - No interrupt disable, only preempt disable for the preempt case. Preempt disable can also be avoided by moving the counter into a spinlock. - Handling is similar to zoned VM counters. - Simple and easily extendable. - Can be omitted to reduce memory use for embedded use. References: RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=113512330605497&w=2 RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=114988082814934&w=2 local_t http://marc.theaimsgroup.com/?l=linux-kernel&m=114991748606690&w=2 V2 http://marc.theaimsgroup.com/?t=115014808400007&r=1&w=2 V3 http://marc.theaimsgroup.com/?l=linux-kernel&m=115024767022346&w=2 V4 http://marc.theaimsgroup.com/?l=linux-kernel&m=115047968808926&w=2 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-30 08:55:45 +00:00
__count_vm_events(PGFREE, 1 << order);
free_one_page(page_zone(page), page, order,
get_pageblock_migratetype(page));
local_irq_restore(flags);
}
/*
* permit the bootmem allocator to evade page validation on high-order frees
*/
void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
{
if (order == 0) {
__ClearPageReserved(page);
set_page_count(page, 0);
set_page_refcounted(page);
__free_page(page);
} else {
int loop;
prefetchw(page);
for (loop = 0; loop < BITS_PER_LONG; loop++) {
struct page *p = &page[loop];
if (loop + 1 < BITS_PER_LONG)
prefetchw(p + 1);
__ClearPageReserved(p);
set_page_count(p, 0);
}
set_page_refcounted(page);
__free_pages(page, order);
}
}
/*
* The order of subdivision here is critical for the IO subsystem.
* Please do not alter this order without good reasons and regression
* testing. Specifically, as large blocks of memory are subdivided,
* the order in which smaller blocks are delivered depends on the order
* they're subdivided in this function. This is the primary factor
* influencing the order in which pages are delivered to the IO
* subsystem according to empirical testing, and this is also justified
* by considering the behavior of a buddy system containing a single
* large block of memory acted on by a series of small allocations.
* This behavior is a critical factor in sglist merging's success.
*
* -- wli
*/
static inline void expand(struct zone *zone, struct page *page,
int low, int high, struct free_area *area,
int migratetype)
{
unsigned long size = 1 << high;
while (high > low) {
area--;
high--;
size >>= 1;
VM_BUG_ON(bad_range(zone, &page[size]));
list_add(&page[size].lru, &area->free_list[migratetype]);
area->nr_free++;
set_page_order(&page[size], high);
}
}
/*
* This page is about to be returned from the page allocator
*/
static inline int check_new_page(struct page *page)
{
if (unlikely(page_mapcount(page) |
(page->mapping != NULL) |
(atomic_read(&page->_count) != 0) |
(page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
(mem_cgroup_bad_page_check(page)))) {
bad_page(page);
return 1;
}
return 0;
}
static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
{
int i;
for (i = 0; i < (1 << order); i++) {
struct page *p = page + i;
if (unlikely(check_new_page(p)))
return 1;
}
[PATCH] mm: split page table lock Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with a many-threaded application which concurrently initializes different parts of a large anonymous area. This patch corrects that, by using a separate spinlock per page table page, to guard the page table entries in that page, instead of using the mm's single page_table_lock. (But even then, page_table_lock is still used to guard page table allocation, and anon_vma allocation.) In this implementation, the spinlock is tucked inside the struct page of the page table page: with a BUILD_BUG_ON in case it overflows - which it would in the case of 32-bit PA-RISC with spinlock debugging enabled. Splitting the lock is not quite for free: another cacheline access. Ideally, I suppose we would use split ptlock only for multi-threaded processes on multi-cpu machines; but deciding that dynamically would have its own costs. So for now enable it by config, at some number of cpus - since the Kconfig language doesn't support inequalities, let preprocessor compare that with NR_CPUS. But I don't think it's worth being user-configurable: for good testing of both split and unsplit configs, split now at 4 cpus, and perhaps change that to 8 later. There is a benefit even for singly threaded processes: kswapd can be attacking one part of the mm while another part is busy faulting. Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 01:16:40 +00:00
set_page_private(page, 0);
set_page_refcounted(page);
arch_alloc_page(page, order);
kernel_map_pages(page, 1 << order, 1);
if (gfp_flags & __GFP_ZERO)
prep_zero_page(page, order, gfp_flags);
if (order && (gfp_flags & __GFP_COMP))
prep_compound_page(page, order);
return 0;
}
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
/*
* Go through the free lists for the given migratetype and remove
* the smallest available page from the freelists
*/
static inline
struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
int migratetype)
{
unsigned int current_order;
struct free_area * area;
struct page *page;
/* Find a page of the appropriate size in the preferred list */
for (current_order = order; current_order < MAX_ORDER; ++current_order) {
area = &(zone->free_area[current_order]);
if (list_empty(&area->free_list[migratetype]))
continue;
page = list_entry(area->free_list[migratetype].next,
struct page, lru);
list_del(&page->lru);
rmv_page_order(page);
area->nr_free--;
expand(zone, page, order, current_order, area, migratetype);
return page;
}
return NULL;
}
/*
* This array describes the order lists are fallen back to when
* the free lists for the desirable migrate type are depleted
*/
static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
don't group high order atomic allocations Grouping high-order atomic allocations together was intended to allow bursty users of atomic allocations to work such as e1000 in situations where their preallocated buffers were depleted. This did not work in at least one case with a wireless network adapter needing order-1 allocations frequently. To resolve that, the free pages used for min_free_kbytes were moved to separate contiguous blocks with the patch bias-the-location-of-pages-freed-for-min_free_kbytes-in-the-same-max_order_nr_pages-blocks. It is felt that keeping the free pages in the same contiguous blocks should be sufficient for bursty short-lived high-order atomic allocations to succeed, maybe even with the e1000. Even if there is a failure, increasing the value of min_free_kbytes will free pages as contiguous bloks in contrast to the standard buddy allocator which makes no attempt to keep the minimum number of free pages contiguous. This patch backs out grouping high order atomic allocations together to determine if it is really needed or not. If a new report comes in about high-order atomic allocations failing, the feature can be reintroduced to determine if it fixes the problem or not. As a side-effect, this patch reduces by 1 the number of bits required to track the mobility type of pages within a MAX_ORDER_NR_PAGES block. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:59 +00:00
[MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
[MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
[MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
[MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
};
/*
* Move the free pages in a range to the free lists of the requested type.
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
* Note that start_page and end_pages are not aligned on a pageblock
* boundary. If alignment is required, use move_freepages_block()
*/
static int move_freepages(struct zone *zone,
struct page *start_page, struct page *end_page,
int migratetype)
{
struct page *page;
unsigned long order;
int pages_moved = 0;
#ifndef CONFIG_HOLES_IN_ZONE
/*
* page_zone is not safe to call in this context when
* CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
* anyway as we check zone boundaries in move_freepages_block().
* Remove at a later date when no bug reports exist related to
* grouping pages by mobility
*/
BUG_ON(page_zone(start_page) != page_zone(end_page));
#endif
for (page = start_page; page <= end_page;) {
/* Make sure we are not inadvertently changing nodes */
VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
if (!pfn_valid_within(page_to_pfn(page))) {
page++;
continue;
}
if (!PageBuddy(page)) {
page++;
continue;
}
order = page_order(page);
list_move(&page->lru,
&zone->free_area[order].free_list[migratetype]);
page += 1 << order;
pages_moved += 1 << order;
}
return pages_moved;
}
static int move_freepages_block(struct zone *zone, struct page *page,
int migratetype)
{
unsigned long start_pfn, end_pfn;
struct page *start_page, *end_page;
start_pfn = page_to_pfn(page);
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
start_pfn = start_pfn & ~(pageblock_nr_pages-1);
start_page = pfn_to_page(start_pfn);
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
end_page = start_page + pageblock_nr_pages - 1;
end_pfn = start_pfn + pageblock_nr_pages - 1;
/* Do not cross zone boundaries */
if (start_pfn < zone->zone_start_pfn)
start_page = page;
if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
return 0;
return move_freepages(zone, start_page, end_page, migratetype);
}
static void change_pageblock_range(struct page *pageblock_page,
int start_order, int migratetype)
{
int nr_pageblocks = 1 << (start_order - pageblock_order);
while (nr_pageblocks--) {
set_pageblock_migratetype(pageblock_page, migratetype);
pageblock_page += pageblock_nr_pages;
}
}
/* Remove an element from the buddy allocator from the fallback list */
static inline struct page *
__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
{
struct free_area * area;
int current_order;
struct page *page;
int migratetype, i;
/* Find the largest possible block of pages in the other list */
for (current_order = MAX_ORDER-1; current_order >= order;
--current_order) {
for (i = 0; i < MIGRATE_TYPES - 1; i++) {
migratetype = fallbacks[start_migratetype][i];
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
/* MIGRATE_RESERVE handled later if necessary */
if (migratetype == MIGRATE_RESERVE)
continue;
area = &(zone->free_area[current_order]);
if (list_empty(&area->free_list[migratetype]))
continue;
page = list_entry(area->free_list[migratetype].next,
struct page, lru);
area->nr_free--;
/*
* If breaking a large block of pages, move all free
* pages to the preferred allocation list. If falling
* back for a reclaimable kernel allocation, be more
* aggressive about taking ownership of free pages
*/
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
if (unlikely(current_order >= (pageblock_order >> 1)) ||
start_migratetype == MIGRATE_RECLAIMABLE ||
page_group_by_mobility_disabled) {
unsigned long pages;
pages = move_freepages_block(zone, page,
start_migratetype);
/* Claim the whole block if over half of it is free */
if (pages >= (1 << (pageblock_order-1)) ||
page_group_by_mobility_disabled)
set_pageblock_migratetype(page,
start_migratetype);
migratetype = start_migratetype;
}
/* Remove the page from the freelists */
list_del(&page->lru);
rmv_page_order(page);
/* Take ownership for orders >= pageblock_order */
if (current_order >= pageblock_order)
change_pageblock_range(page, current_order,
start_migratetype);
expand(zone, page, order, current_order, area, migratetype);
trace_mm_page_alloc_extfrag(page, order, current_order,
start_migratetype, migratetype);
return page;
}
}
return NULL;
}
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
/*
* Do the hard work of removing an element from the buddy allocator.
* Call me with the zone->lock already held.
*/
static struct page *__rmqueue(struct zone *zone, unsigned int order,
int migratetype)
{
struct page *page;
retry_reserve:
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
page = __rmqueue_smallest(zone, order, migratetype);
if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
page = __rmqueue_fallback(zone, order, migratetype);
/*
* Use MIGRATE_RESERVE rather than fail an allocation. goto
* is used because __rmqueue_smallest is an inline function
* and we want just one call site
*/
if (!page) {
migratetype = MIGRATE_RESERVE;
goto retry_reserve;
}
}
tracing, page-allocator: add trace event for page traffic related to the buddy lists The page allocation trace event reports that a page was successfully allocated but it does not specify where it came from. When analysing performance, it can be important to distinguish between pages coming from the per-cpu allocator and pages coming from the buddy lists as the latter requires the zone lock to the taken and more data structures to be examined. This patch adds a trace event for __rmqueue reporting when a page is being allocated from the buddy lists. It distinguishes between being called to refill the per-cpu lists or whether it is a high-order allocation. Similarly, this patch adds an event to catch when the PCP lists are being drained a little and pages are going back to the buddy lists. This is trickier to draw conclusions from but high activity on those events could explain why there were a large number of cache misses on a page-allocator-intensive workload. The coalescing and splitting of buddies involves a lot of writing of page metadata and cache line bounces not to mention the acquisition of an interrupt-safe lock necessary to enter this path. [akpm@linux-foundation.org: fix build] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Rik van Riel <riel@redhat.com> Reviewed-by: Ingo Molnar <mingo@elte.hu> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Li Ming Chun <macli@brc.ubc.ca> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:02:44 +00:00
trace_mm_page_alloc_zone_locked(page, order, migratetype);
return page;
}
/*
* Obtain a specified number of elements from the buddy allocator, all under
* a single hold of the lock, for efficiency. Add them to the supplied list.
* Returns the number of new pages which were placed at *list.
*/
static int rmqueue_bulk(struct zone *zone, unsigned int order,
unsigned long count, struct list_head *list,
int migratetype, int cold)
{
int i;
spin_lock(&zone->lock);
for (i = 0; i < count; ++i) {
struct page *page = __rmqueue(zone, order, migratetype);
if (unlikely(page == NULL))
break;
/*
* Split buddy pages returned by expand() are received here
* in physical page order. The page is added to the callers and
* list and the list head then moves forward. From the callers
* perspective, the linked list is ordered by page number in
* some conditions. This is useful for IO devices that can
* merge IO requests if the physical pages are ordered
* properly.
*/
if (likely(cold == 0))
list_add(&page->lru, list);
else
list_add_tail(&page->lru, list);
set_page_private(page, migratetype);
list = &page->lru;
}
__mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
spin_unlock(&zone->lock);
return i;
}
#ifdef CONFIG_NUMA
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
/*
Move remote node draining out of slab allocators Currently the slab allocators contain callbacks into the page allocator to perform the draining of pagesets on remote nodes. This requires SLUB to have a whole subsystem in order to be compatible with SLAB. Moving node draining out of the slab allocators avoids a section of code in SLUB. Move the node draining so that is is done when the vm statistics are updated. At that point we are already touching all the cachelines with the pagesets of a processor. Add a expire counter there. If we have to update per zone or global vm statistics then assume that the pageset will require subsequent draining. The expire counter will be decremented on each vm stats update pass until it reaches zero. Then we will drain one batch from the pageset. The draining will cause vm counter updates which will then cause another expiration until the pcp is empty. So we will drain a batch every 3 seconds. Note that remote node draining is a somewhat esoteric feature that is required on large NUMA systems because otherwise significant portions of system memory can become trapped in pcp queues. The number of pcp is determined by the number of processors and nodes in a system. A system with 4 processors and 2 nodes has 8 pcps which is okay. But a system with 1024 processors and 512 nodes has 512k pcps with a high potential for large amount of memory being caught in them. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:35:14 +00:00
* Called from the vmstat counter updater to drain pagesets of this
* currently executing processor on remote nodes after they have
* expired.
*
* Note that this function must be called with the thread pinned to
* a single processor.
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
*/
Move remote node draining out of slab allocators Currently the slab allocators contain callbacks into the page allocator to perform the draining of pagesets on remote nodes. This requires SLUB to have a whole subsystem in order to be compatible with SLAB. Moving node draining out of the slab allocators avoids a section of code in SLUB. Move the node draining so that is is done when the vm statistics are updated. At that point we are already touching all the cachelines with the pagesets of a processor. Add a expire counter there. If we have to update per zone or global vm statistics then assume that the pageset will require subsequent draining. The expire counter will be decremented on each vm stats update pass until it reaches zero. Then we will drain one batch from the pageset. The draining will cause vm counter updates which will then cause another expiration until the pcp is empty. So we will drain a batch every 3 seconds. Note that remote node draining is a somewhat esoteric feature that is required on large NUMA systems because otherwise significant portions of system memory can become trapped in pcp queues. The number of pcp is determined by the number of processors and nodes in a system. A system with 4 processors and 2 nodes has 8 pcps which is okay. But a system with 1024 processors and 512 nodes has 512k pcps with a high potential for large amount of memory being caught in them. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:35:14 +00:00
void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
{
unsigned long flags;
Move remote node draining out of slab allocators Currently the slab allocators contain callbacks into the page allocator to perform the draining of pagesets on remote nodes. This requires SLUB to have a whole subsystem in order to be compatible with SLAB. Moving node draining out of the slab allocators avoids a section of code in SLUB. Move the node draining so that is is done when the vm statistics are updated. At that point we are already touching all the cachelines with the pagesets of a processor. Add a expire counter there. If we have to update per zone or global vm statistics then assume that the pageset will require subsequent draining. The expire counter will be decremented on each vm stats update pass until it reaches zero. Then we will drain one batch from the pageset. The draining will cause vm counter updates which will then cause another expiration until the pcp is empty. So we will drain a batch every 3 seconds. Note that remote node draining is a somewhat esoteric feature that is required on large NUMA systems because otherwise significant portions of system memory can become trapped in pcp queues. The number of pcp is determined by the number of processors and nodes in a system. A system with 4 processors and 2 nodes has 8 pcps which is okay. But a system with 1024 processors and 512 nodes has 512k pcps with a high potential for large amount of memory being caught in them. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:35:14 +00:00
int to_drain;
Move remote node draining out of slab allocators Currently the slab allocators contain callbacks into the page allocator to perform the draining of pagesets on remote nodes. This requires SLUB to have a whole subsystem in order to be compatible with SLAB. Moving node draining out of the slab allocators avoids a section of code in SLUB. Move the node draining so that is is done when the vm statistics are updated. At that point we are already touching all the cachelines with the pagesets of a processor. Add a expire counter there. If we have to update per zone or global vm statistics then assume that the pageset will require subsequent draining. The expire counter will be decremented on each vm stats update pass until it reaches zero. Then we will drain one batch from the pageset. The draining will cause vm counter updates which will then cause another expiration until the pcp is empty. So we will drain a batch every 3 seconds. Note that remote node draining is a somewhat esoteric feature that is required on large NUMA systems because otherwise significant portions of system memory can become trapped in pcp queues. The number of pcp is determined by the number of processors and nodes in a system. A system with 4 processors and 2 nodes has 8 pcps which is okay. But a system with 1024 processors and 512 nodes has 512k pcps with a high potential for large amount of memory being caught in them. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:35:14 +00:00
local_irq_save(flags);
if (pcp->count >= pcp->batch)
to_drain = pcp->batch;
else
to_drain = pcp->count;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
free_pcppages_bulk(zone, to_drain, pcp);
Move remote node draining out of slab allocators Currently the slab allocators contain callbacks into the page allocator to perform the draining of pagesets on remote nodes. This requires SLUB to have a whole subsystem in order to be compatible with SLAB. Moving node draining out of the slab allocators avoids a section of code in SLUB. Move the node draining so that is is done when the vm statistics are updated. At that point we are already touching all the cachelines with the pagesets of a processor. Add a expire counter there. If we have to update per zone or global vm statistics then assume that the pageset will require subsequent draining. The expire counter will be decremented on each vm stats update pass until it reaches zero. Then we will drain one batch from the pageset. The draining will cause vm counter updates which will then cause another expiration until the pcp is empty. So we will drain a batch every 3 seconds. Note that remote node draining is a somewhat esoteric feature that is required on large NUMA systems because otherwise significant portions of system memory can become trapped in pcp queues. The number of pcp is determined by the number of processors and nodes in a system. A system with 4 processors and 2 nodes has 8 pcps which is okay. But a system with 1024 processors and 512 nodes has 512k pcps with a high potential for large amount of memory being caught in them. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:35:14 +00:00
pcp->count -= to_drain;
local_irq_restore(flags);
}
#endif
/*
* Drain pages of the indicated processor.
*
* The processor must either be the current processor and the
* thread pinned to the current processor or a processor that
* is not online.
*/
static void drain_pages(unsigned int cpu)
{
unsigned long flags;
struct zone *zone;
for_each_populated_zone(zone) {
struct per_cpu_pageset *pset;
struct per_cpu_pages *pcp;
local_irq_save(flags);
pset = per_cpu_ptr(zone->pageset, cpu);
pcp = &pset->pcp;
mm: clear pages_scanned only if draining a pcp adds pages to the buddy allocator Commit 0e093d99763e ("writeback: do not sleep on the congestion queue if there are no congested BDIs or if significant congestion is not being encountered in the current zone") uncovered a livelock in the page allocator that resulted in tasks infinitely looping trying to find memory and kswapd running at 100% cpu. The issue occurs because drain_all_pages() is called immediately following direct reclaim when no memory is freed and try_to_free_pages() returns non-zero because all zones in the zonelist do not have their all_unreclaimable flag set. When draining the per-cpu pagesets back to the buddy allocator for each zone, the zone->pages_scanned counter is cleared to avoid erroneously setting zone->all_unreclaimable later. The problem is that no pages may actually be drained and, thus, the unreclaimable logic never fails direct reclaim so the oom killer may be invoked. This apparently only manifested after wait_iff_congested() was introduced and the zone was full of anonymous memory that would not congest the backing store. The page allocator would infinitely loop if there were no other tasks waiting to be scheduled and clear zone->pages_scanned because of drain_all_pages() as the result of this change before kswapd could scan enough pages to trigger the reclaim logic. Additionally, with every loop of the page allocator and in the reclaim path, kswapd would be kicked and would end up running at 100% cpu. In this scenario, current and kswapd are all running continuously with kswapd incrementing zone->pages_scanned and current clearing it. The problem is even more pronounced when current swaps some of its memory to swap cache and the reclaimable logic then considers all active anonymous memory in the all_unreclaimable logic, which requires a much higher zone->pages_scanned value for try_to_free_pages() to return zero that is never attainable in this scenario. Before wait_iff_congested(), the page allocator would incur an unconditional timeout and allow kswapd to elevate zone->pages_scanned to a level that the oom killer would be called the next time it loops. The fix is to only attempt to drain pcp pages if there is actually a quantity to be drained. The unconditional clearing of zone->pages_scanned in free_pcppages_bulk() need not be changed since other callers already ensure that draining will occur. This patch ensures that free_pcppages_bulk() will actually free memory before calling into it from drain_all_pages() so zone->pages_scanned is only cleared if appropriate. Signed-off-by: David Rientjes <rientjes@google.com> Cc: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-25 23:07:23 +00:00
if (pcp->count) {
free_pcppages_bulk(zone, pcp->count, pcp);
pcp->count = 0;
}
local_irq_restore(flags);
}
}
/*
* Spill all of this CPU's per-cpu pages back into the buddy allocator.
*/
void drain_local_pages(void *arg)
{
drain_pages(smp_processor_id());
}
/*
* Spill all the per-cpu pages from all CPUs back into the buddy allocator
*/
void drain_all_pages(void)
{
on_each_cpu(drain_local_pages, NULL, 1);
}
#ifdef CONFIG_HIBERNATION
void mark_free_pages(struct zone *zone)
{
unsigned long pfn, max_zone_pfn;
unsigned long flags;
int order, t;
struct list_head *curr;
if (!zone->spanned_pages)
return;
spin_lock_irqsave(&zone->lock, flags);
max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
if (pfn_valid(pfn)) {
struct page *page = pfn_to_page(pfn);
if (!swsusp_page_is_forbidden(page))
swsusp_unset_page_free(page);
}
for_each_migratetype_order(order, t) {
list_for_each(curr, &zone->free_area[order].free_list[t]) {
unsigned long i;
pfn = page_to_pfn(list_entry(curr, struct page, lru));
for (i = 0; i < (1UL << order); i++)
swsusp_set_page_free(pfn_to_page(pfn + i));
}
}
spin_unlock_irqrestore(&zone->lock, flags);
}
#endif /* CONFIG_PM */
/*
* Free a 0-order page
* cold == 1 ? free a cold page : free a hot page
*/
void free_hot_cold_page(struct page *page, int cold)
{
struct zone *zone = page_zone(page);
struct per_cpu_pages *pcp;
unsigned long flags;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
int migratetype;
int wasMlocked = __TestClearPageMlocked(page);
if (!free_pages_prepare(page, 0))
return;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
migratetype = get_pageblock_migratetype(page);
set_page_private(page, migratetype);
local_irq_save(flags);
if (unlikely(wasMlocked))
free_page_mlock(page);
[PATCH] Light weight event counters The remaining counters in page_state after the zoned VM counter patches have been applied are all just for show in /proc/vmstat. They have no essential function for the VM. We use a simple increment of per cpu variables. In order to avoid the most severe races we disable preempt. Preempt does not prevent the race between an increment and an interrupt handler incrementing the same statistics counter. However, that race is exceedingly rare, we may only loose one increment or so and there is no requirement (at least not in kernel) that the vm event counters have to be accurate. In the non preempt case this results in a simple increment for each counter. For many architectures this will be reduced by the compiler to a single instruction. This single instruction is atomic for i386 and x86_64. And therefore even the rare race condition in an interrupt is avoided for both architectures in most cases. The patchset also adds an off switch for embedded systems that allows a building of linux kernels without these counters. The implementation of these counters is through inline code that hopefully results in only a single instruction increment instruction being emitted (i386, x86_64) or in the increment being hidden though instruction concurrency (EPIC architectures such as ia64 can get that done). Benefits: - VM event counter operations usually reduce to a single inline instruction on i386 and x86_64. - No interrupt disable, only preempt disable for the preempt case. Preempt disable can also be avoided by moving the counter into a spinlock. - Handling is similar to zoned VM counters. - Simple and easily extendable. - Can be omitted to reduce memory use for embedded use. References: RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=113512330605497&w=2 RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=114988082814934&w=2 local_t http://marc.theaimsgroup.com/?l=linux-kernel&m=114991748606690&w=2 V2 http://marc.theaimsgroup.com/?t=115014808400007&r=1&w=2 V3 http://marc.theaimsgroup.com/?l=linux-kernel&m=115024767022346&w=2 V4 http://marc.theaimsgroup.com/?l=linux-kernel&m=115047968808926&w=2 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-30 08:55:45 +00:00
__count_vm_event(PGFREE);
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
/*
* We only track unmovable, reclaimable and movable on pcp lists.
* Free ISOLATE pages back to the allocator because they are being
* offlined but treat RESERVE as movable pages so we can get those
* areas back if necessary. Otherwise, we may have to free
* excessively into the page allocator
*/
if (migratetype >= MIGRATE_PCPTYPES) {
if (unlikely(migratetype == MIGRATE_ISOLATE)) {
free_one_page(zone, page, 0, migratetype);
goto out;
}
migratetype = MIGRATE_MOVABLE;
}
pcp = &this_cpu_ptr(zone->pageset)->pcp;
if (cold)
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
list_add_tail(&page->lru, &pcp->lists[migratetype]);
else
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
list_add(&page->lru, &pcp->lists[migratetype]);
pcp->count++;
if (pcp->count >= pcp->high) {
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
free_pcppages_bulk(zone, pcp->batch, pcp);
pcp->count -= pcp->batch;
}
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
out:
local_irq_restore(flags);
}
/*
* split_page takes a non-compound higher-order page, and splits it into
* n (1<<order) sub-pages: page[0..n]
* Each sub-page must be freed individually.
*
* Note: this is probably too low level an operation for use in drivers.
* Please consult with lkml before using this in your driver.
*/
void split_page(struct page *page, unsigned int order)
{
int i;
VM_BUG_ON(PageCompound(page));
VM_BUG_ON(!page_count(page));
#ifdef CONFIG_KMEMCHECK
/*
* Split shadow pages too, because free(page[0]) would
* otherwise free the whole shadow.
*/
if (kmemcheck_page_is_tracked(page))
split_page(virt_to_page(page[0].shadow), order);
#endif
for (i = 1; i < (1 << order); i++)
set_page_refcounted(page + i);
}
/*
* Similar to split_page except the page is already free. As this is only
* being used for migration, the migratetype of the block also changes.
* As this is called with interrupts disabled, the caller is responsible
* for calling arch_alloc_page() and kernel_map_page() after interrupts
* are enabled.
*
* Note: this is probably too low level an operation for use in drivers.
* Please consult with lkml before using this in your driver.
*/
int split_free_page(struct page *page)
{
unsigned int order;
unsigned long watermark;
struct zone *zone;
BUG_ON(!PageBuddy(page));
zone = page_zone(page);
order = page_order(page);
/* Obey watermarks as if the page was being allocated */
watermark = low_wmark_pages(zone) + (1 << order);
if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
return 0;
/* Remove page from free list */
list_del(&page->lru);
zone->free_area[order].nr_free--;
rmv_page_order(page);
__mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
/* Split into individual pages */
set_page_refcounted(page);
split_page(page, order);
if (order >= pageblock_order - 1) {
struct page *endpage = page + (1 << order) - 1;
for (; page < endpage; page += pageblock_nr_pages)
set_pageblock_migratetype(page, MIGRATE_MOVABLE);
}
return 1 << order;
}
/*
* Really, prep_compound_page() should be called from __rmqueue_bulk(). But
* we cheat by calling it from here, in the order > 0 path. Saves a branch
* or two.
*/
static inline
struct page *buffered_rmqueue(struct zone *preferred_zone,
struct zone *zone, int order, gfp_t gfp_flags,
int migratetype)
{
unsigned long flags;
struct page *page;
int cold = !!(gfp_flags & __GFP_COLD);
again:
if (likely(order == 0)) {
struct per_cpu_pages *pcp;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
struct list_head *list;
local_irq_save(flags);
pcp = &this_cpu_ptr(zone->pageset)->pcp;
list = &pcp->lists[migratetype];
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
if (list_empty(list)) {
pcp->count += rmqueue_bulk(zone, 0,
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
pcp->batch, list,
migratetype, cold);
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
if (unlikely(list_empty(list)))
goto failed;
}
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
if (cold)
page = list_entry(list->prev, struct page, lru);
else
page = list_entry(list->next, struct page, lru);
list_del(&page->lru);
pcp->count--;
} else {
if (unlikely(gfp_flags & __GFP_NOFAIL)) {
/*
* __GFP_NOFAIL is not to be used in new code.
*
* All __GFP_NOFAIL callers should be fixed so that they
* properly detect and handle allocation failures.
*
* We most definitely don't want callers attempting to
* allocate greater than order-1 page units with
* __GFP_NOFAIL.
*/
WARN_ON_ONCE(order > 1);
}
spin_lock_irqsave(&zone->lock, flags);
page = __rmqueue(zone, order, migratetype);
spin_unlock(&zone->lock);
if (!page)
goto failed;
__mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
}
[PATCH] Light weight event counters The remaining counters in page_state after the zoned VM counter patches have been applied are all just for show in /proc/vmstat. They have no essential function for the VM. We use a simple increment of per cpu variables. In order to avoid the most severe races we disable preempt. Preempt does not prevent the race between an increment and an interrupt handler incrementing the same statistics counter. However, that race is exceedingly rare, we may only loose one increment or so and there is no requirement (at least not in kernel) that the vm event counters have to be accurate. In the non preempt case this results in a simple increment for each counter. For many architectures this will be reduced by the compiler to a single instruction. This single instruction is atomic for i386 and x86_64. And therefore even the rare race condition in an interrupt is avoided for both architectures in most cases. The patchset also adds an off switch for embedded systems that allows a building of linux kernels without these counters. The implementation of these counters is through inline code that hopefully results in only a single instruction increment instruction being emitted (i386, x86_64) or in the increment being hidden though instruction concurrency (EPIC architectures such as ia64 can get that done). Benefits: - VM event counter operations usually reduce to a single inline instruction on i386 and x86_64. - No interrupt disable, only preempt disable for the preempt case. Preempt disable can also be avoided by moving the counter into a spinlock. - Handling is similar to zoned VM counters. - Simple and easily extendable. - Can be omitted to reduce memory use for embedded use. References: RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=113512330605497&w=2 RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=114988082814934&w=2 local_t http://marc.theaimsgroup.com/?l=linux-kernel&m=114991748606690&w=2 V2 http://marc.theaimsgroup.com/?t=115014808400007&r=1&w=2 V3 http://marc.theaimsgroup.com/?l=linux-kernel&m=115024767022346&w=2 V4 http://marc.theaimsgroup.com/?l=linux-kernel&m=115047968808926&w=2 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-30 08:55:45 +00:00
__count_zone_vm_events(PGALLOC, zone, 1 << order);
zone_statistics(preferred_zone, zone, gfp_flags);
local_irq_restore(flags);
VM_BUG_ON(bad_range(zone, page));
if (prep_new_page(page, order, gfp_flags))
goto again;
return page;
failed:
local_irq_restore(flags);
return NULL;
}
/* The ALLOC_WMARK bits are used as an index to zone->watermark */
#define ALLOC_WMARK_MIN WMARK_MIN
#define ALLOC_WMARK_LOW WMARK_LOW
#define ALLOC_WMARK_HIGH WMARK_HIGH
#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
/* Mask to get the watermark bits */
#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
#define ALLOC_HARDER 0x10 /* try to alloc harder */
#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
#define ALLOC_CPUSET 0x40 /* check for correct cpuset */
#ifdef CONFIG_FAIL_PAGE_ALLOC
static struct {
struct fault_attr attr;
u32 ignore_gfp_highmem;
u32 ignore_gfp_wait;
u32 min_order;
} fail_page_alloc = {
.attr = FAULT_ATTR_INITIALIZER,
.ignore_gfp_wait = 1,
.ignore_gfp_highmem = 1,
.min_order = 1,
};
static int __init setup_fail_page_alloc(char *str)
{
return setup_fault_attr(&fail_page_alloc.attr, str);
}
__setup("fail_page_alloc=", setup_fail_page_alloc);
static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
{
if (order < fail_page_alloc.min_order)
return 0;
if (gfp_mask & __GFP_NOFAIL)
return 0;
if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
return 0;
if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
return 0;
return should_fail(&fail_page_alloc.attr, 1 << order);
}
#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
static int __init fail_page_alloc_debugfs(void)
{
mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
struct dentry *dir;
dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
&fail_page_alloc.attr);
if (IS_ERR(dir))
return PTR_ERR(dir);
if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
&fail_page_alloc.ignore_gfp_wait))
goto fail;
if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
&fail_page_alloc.ignore_gfp_highmem))
goto fail;
if (!debugfs_create_u32("min-order", mode, dir,
&fail_page_alloc.min_order))
goto fail;
return 0;
fail:
debugfs_remove_recursive(dir);
return -ENOMEM;
}
late_initcall(fail_page_alloc_debugfs);
#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
#else /* CONFIG_FAIL_PAGE_ALLOC */
static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
{
return 0;
}
#endif /* CONFIG_FAIL_PAGE_ALLOC */
/*
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
* Return true if free pages are above 'mark'. This takes into account the order
* of the allocation.
*/
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
int classzone_idx, int alloc_flags, long free_pages)
{
/* free_pages my go negative - that's OK */
long min = mark;
int o;
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
free_pages -= (1 << order) + 1;
if (alloc_flags & ALLOC_HIGH)
min -= min / 2;
if (alloc_flags & ALLOC_HARDER)
min -= min / 4;
if (free_pages <= min + z->lowmem_reserve[classzone_idx])
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
return false;
for (o = 0; o < order; o++) {
/* At the next order, this order's pages become unavailable */
free_pages -= z->free_area[o].nr_free << o;
/* Require fewer higher order pages to be free */
min >>= 1;
if (free_pages <= min)
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
return false;
}
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
return true;
}
bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
int classzone_idx, int alloc_flags)
{
return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
zone_page_state(z, NR_FREE_PAGES));
}
bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
int classzone_idx, int alloc_flags)
{
long free_pages = zone_page_state(z, NR_FREE_PAGES);
if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
free_pages);
}
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
#ifdef CONFIG_NUMA
/*
* zlc_setup - Setup for "zonelist cache". Uses cached zone data to
* skip over zones that are not allowed by the cpuset, or that have
* been recently (in last second) found to be nearly full. See further
* comments in mmzone.h. Reduces cache footprint of zonelist scans
* that have to skip over a lot of full or unallowed zones.
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
*
* If the zonelist cache is present in the passed in zonelist, then
* returns a pointer to the allowed node mask (either the current
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
* tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
*
* If the zonelist cache is not available for this zonelist, does
* nothing and returns NULL.
*
* If the fullzones BITMAP in the zonelist cache is stale (more than
* a second since last zap'd) then we zap it out (clear its bits.)
*
* We hold off even calling zlc_setup, until after we've checked the
* first zone in the zonelist, on the theory that most allocations will
* be satisfied from that first zone, so best to examine that zone as
* quickly as we can.
*/
static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
{
struct zonelist_cache *zlc; /* cached zonelist speedup info */
nodemask_t *allowednodes; /* zonelist_cache approximation */
zlc = zonelist->zlcache_ptr;
if (!zlc)
return NULL;
if (time_after(jiffies, zlc->last_full_zap + HZ)) {
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
zlc->last_full_zap = jiffies;
}
allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
&cpuset_current_mems_allowed :
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
&node_states[N_HIGH_MEMORY];
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
return allowednodes;
}
/*
* Given 'z' scanning a zonelist, run a couple of quick checks to see
* if it is worth looking at further for free memory:
* 1) Check that the zone isn't thought to be full (doesn't have its
* bit set in the zonelist_cache fullzones BITMAP).
* 2) Check that the zones node (obtained from the zonelist_cache
* z_to_n[] mapping) is allowed in the passed in allowednodes mask.
* Return true (non-zero) if zone is worth looking at further, or
* else return false (zero) if it is not.
*
* This check -ignores- the distinction between various watermarks,
* such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
* found to be full for any variation of these watermarks, it will
* be considered full for up to one second by all requests, unless
* we are so low on memory on all allowed nodes that we are forced
* into the second scan of the zonelist.
*
* In the second scan we ignore this zonelist cache and exactly
* apply the watermarks to all zones, even it is slower to do so.
* We are low on memory in the second scan, and should leave no stone
* unturned looking for a free page.
*/
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
nodemask_t *allowednodes)
{
struct zonelist_cache *zlc; /* cached zonelist speedup info */
int i; /* index of *z in zonelist zones */
int n; /* node that zone *z is on */
zlc = zonelist->zlcache_ptr;
if (!zlc)
return 1;
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
i = z - zonelist->_zonerefs;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
n = zlc->z_to_n[i];
/* This zone is worth trying if it is allowed but not full */
return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
}
/*
* Given 'z' scanning a zonelist, set the corresponding bit in
* zlc->fullzones, so that subsequent attempts to allocate a page
* from that zone don't waste time re-examining it.
*/
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
{
struct zonelist_cache *zlc; /* cached zonelist speedup info */
int i; /* index of *z in zonelist zones */
zlc = zonelist->zlcache_ptr;
if (!zlc)
return;
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
i = z - zonelist->_zonerefs;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
set_bit(i, zlc->fullzones);
}
/*
* clear all zones full, called after direct reclaim makes progress so that
* a zone that was recently full is not skipped over for up to a second
*/
static void zlc_clear_zones_full(struct zonelist *zonelist)
{
struct zonelist_cache *zlc; /* cached zonelist speedup info */
zlc = zonelist->zlcache_ptr;
if (!zlc)
return;
bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
}
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
#else /* CONFIG_NUMA */
static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
{
return NULL;
}
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
nodemask_t *allowednodes)
{
return 1;
}
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
{
}
static void zlc_clear_zones_full(struct zonelist *zonelist)
{
}
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
#endif /* CONFIG_NUMA */
/*
* get_page_from_freelist goes through the zonelist trying to allocate
* a page.
*/
static struct page *
get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
struct zone *preferred_zone, int migratetype)
[PATCH] VM: early zone reclaim This is the core of the (much simplified) early reclaim. The goal of this patch is to reclaim some easily-freed pages from a zone before falling back onto another zone. One of the major uses of this is NUMA machines. With the default allocator behavior the allocator would look for memory in another zone, which might be off-node, before trying to reclaim from the current zone. This adds a zone tuneable to enable early zone reclaim. It is selected on a per-zone basis and is turned on/off via syscall. Adding some extra throttling on the reclaim was also required (patch 4/4). Without the machine would grind to a crawl when doing a "make -j" kernel build. Even with this patch the System Time is higher on average, but it seems tolerable. Here are some numbers for kernbench runs on a 2-node, 4cpu, 8Gig RAM Altix in the "make -j" run: wall user sys %cpu ctx sw. sleeps ---- ---- --- ---- ------ ------ No patch 1009 1384 847 258 298170 504402 w/patch, no reclaim 880 1376 667 288 254064 396745 w/patch & reclaim 1079 1385 926 252 291625 548873 These numbers are the average of 2 runs of 3 "make -j" runs done right after system boot. Run-to-run variability for "make -j" is huge, so these numbers aren't terribly useful except to seee that with reclaim the benchmark still finishes in a reasonable amount of time. I also looked at the NUMA hit/miss stats for the "make -j" runs and the reclaim doesn't make any difference when the machine is thrashing away. Doing a "make -j8" on a single node that is filled with page cache pages takes 700 seconds with reclaim turned on and 735 seconds without reclaim (due to remote memory accesses). The simple zone_reclaim syscall program is at http://www.bork.org/~mort/sgi/zone_reclaim.c Signed-off-by: Martin Hicks <mort@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:41 +00:00
{
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
struct zoneref *z;
struct page *page = NULL;
int classzone_idx;
struct zone *zone;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
int zlc_active = 0; /* set if using zonelist_cache */
int did_zlc_setup = 0; /* just call zlc_setup() one time */
classzone_idx = zone_idx(preferred_zone);
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
zonelist_scan:
/*
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
* Scan zonelist, looking for a zone with enough free.
* See also cpuset_zone_allowed() comment in kernel/cpuset.c.
*/
for_each_zone_zonelist_nodemask(zone, z, zonelist,
high_zoneidx, nodemask) {
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
if (NUMA_BUILD && zlc_active &&
!zlc_zone_worth_trying(zonelist, z, allowednodes))
continue;
if ((alloc_flags & ALLOC_CPUSET) &&
[PATCH] cpuset: rework cpuset_zone_allowed api Elaborate the API for calling cpuset_zone_allowed(), so that users have to explicitly choose between the two variants: cpuset_zone_allowed_hardwall() cpuset_zone_allowed_softwall() Until now, whether or not you got the hardwall flavor depended solely on whether or not you or'd in the __GFP_HARDWALL gfp flag to the gfp_mask argument. If you didn't specify __GFP_HARDWALL, you implicitly got the softwall version. Unfortunately, this meant that users would end up with the softwall version without thinking about it. Since only the softwall version might sleep, this led to bugs with possible sleeping in interrupt context on more than one occassion. The hardwall version requires that the current tasks mems_allowed allows the node of the specified zone (or that you're in interrupt or that __GFP_THISNODE is set or that you're on a one cpuset system.) The softwall version, depending on the gfp_mask, might allow a node if it was allowed in the nearest enclusing cpuset marked mem_exclusive (which requires taking the cpuset lock 'callback_mutex' to evaluate.) This patch removes the cpuset_zone_allowed() call, and forces the caller to explicitly choose between the hardwall and the softwall case. If the caller wants the gfp_mask to determine this choice, they should (1) be sure they can sleep or that __GFP_HARDWALL is set, and (2) invoke the cpuset_zone_allowed_softwall() routine. This adds another 100 or 200 bytes to the kernel text space, due to the few lines of nearly duplicate code at the top of both cpuset_zone_allowed_* routines. It should save a few instructions executed for the calls that turned into calls of cpuset_zone_allowed_hardwall, thanks to not having to set (before the call) then check (within the call) the __GFP_HARDWALL flag. For the most critical call, from get_page_from_freelist(), the same instructions are executed as before -- the old cpuset_zone_allowed() routine it used to call is the same code as the cpuset_zone_allowed_softwall() routine that it calls now. Not a perfect win, but seems worth it, to reduce this chance of hitting a sleeping with irq off complaint again. Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-13 08:34:25 +00:00
!cpuset_zone_allowed_softwall(zone, gfp_mask))
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim There have been a small number of complaints about significant stalls while copying large amounts of data on NUMA machines reported on a distribution bugzilla. In these cases, zone_reclaim was enabled by default due to large NUMA distances. In general, the complaints have not been about the workload itself unless it was a file server (in which case the recommendation was disable zone_reclaim). The stalls are mostly due to significant amounts of time spent scanning the preferred zone for pages to free. After a failure, it might fallback to another node (as zonelists are often node-ordered rather than zone-ordered) but stall quickly again when the next allocation attempt occurs. In bad cases, each page allocated results in a full scan of the preferred zone. Patch 1 checks the preferred zone for recent allocation failure which is particularly important if zone_reclaim has failed recently. This avoids rescanning the zone in the near future and instead falling back to another node. This may hurt node locality in some cases but a failure to zone_reclaim is more expensive than a remote access. Patch 2 clears the zlc information after direct reclaim. Otherwise, zone_reclaim can mark zones full, direct reclaim can reclaim enough pages but the zone is still not considered for allocation. This was tested on a 24-thread 2-node x86_64 machine. The tests were focused on large amounts of IO. All tests were bound to the CPUs on node-0 to avoid disturbances due to processes being scheduled on different nodes. The kernels tested are 3.0-rc6-vanilla Vanilla 3.0-rc6 zlcfirst Patch 1 applied zlcreconsider Patches 1+2 applied FS-Mark ./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288 fsmark-3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirs zlcreconsider Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%) Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%) Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%) Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%) Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%) Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%) Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%) Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 501.49 493.91 499.93 Total Elapsed Time (seconds) 2451.57 2257.48 2215.92 MMTests Statistics: vmstat Page Ins 46268 63840 66008 Page Outs 90821596 90671128 88043732 Swap Ins 0 0 0 Swap Outs 0 0 0 Direct pages scanned 13091697 8966863 8971790 Kswapd pages scanned 0 1830011 1831116 Kswapd pages reclaimed 0 1829068 1829930 Direct pages reclaimed 13037777 8956828 8648314 Kswapd efficiency 100% 99% 99% Kswapd velocity 0.000 810.643 826.346 Direct efficiency 99% 99% 96% Direct velocity 5340.128 3972.068 4048.788 Percentage direct scans 100% 83% 83% Page writes by reclaim 0 3 0 Slabs scanned 796672 720640 720256 Direct inode steals 7422667 7160012 7088638 Kswapd inode steals 0 1736840 2021238 Test completes far faster with a large increase in the number of files created per second. Standard deviation is high as a small number of iterations were much higher than the mean. The number of pages scanned by zone_reclaim is reduced and kswapd is used for more work. LARGE DD 3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirst zlcreconsider download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%) dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%) delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 125.03 118.98 122.01 Total Elapsed Time (seconds) 624.56 375.02 398.06 MMTests Statistics: vmstat Page Ins 3594216 439368 407032 Page Outs 23380832 23380488 23377444 Swap Ins 0 0 0 Swap Outs 0 436 287 Direct pages scanned 17482342 69315973 82864918 Kswapd pages scanned 0 519123 575425 Kswapd pages reclaimed 0 466501 522487 Direct pages reclaimed 5858054 2732949 2712547 Kswapd efficiency 100% 89% 90% Kswapd velocity 0.000 1384.254 1445.574 Direct efficiency 33% 3% 3% Direct velocity 27991.453 184832.737 208171.929 Percentage direct scans 100% 99% 99% Page writes by reclaim 0 5082 13917 Slabs scanned 17280 29952 35328 Direct inode steals 115257 1431122 332201 Kswapd inode steals 0 0 979532 This test downloads a large tarfile and copies it with dd a number of times - similar to the most recent bug report I've dealt with. Time to completion is reduced. The number of pages scanned directly is still disturbingly high with a low efficiency but this is likely due to the number of dirty pages encountered. The figures could probably be improved with more work around how kswapd is used and how dirty pages are handled but that is separate work and this result is significant on its own. Streaming Mapped Writer MMTests Statistics: duration User/Sys Time Running Test (seconds) 124.47 111.67 112.64 Total Elapsed Time (seconds) 2138.14 1816.30 1867.56 MMTests Statistics: vmstat Page Ins 90760 89124 89516 Page Outs 121028340 120199524 120736696 Swap Ins 0 86 55 Swap Outs 0 0 0 Direct pages scanned 114989363 96461439 96330619 Kswapd pages scanned 56430948 56965763 57075875 Kswapd pages reclaimed 27743219 27752044 27766606 Direct pages reclaimed 49777 46884 36655 Kswapd efficiency 49% 48% 48% Kswapd velocity 26392.541 31363.631 30561.736 Direct efficiency 0% 0% 0% Direct velocity 53780.091 53108.759 51581.004 Percentage direct scans 67% 62% 62% Page writes by reclaim 385 122 1513 Slabs scanned 43008 39040 42112 Direct inode steals 0 10 8 Kswapd inode steals 733 534 477 This test just creates a large file mapping and writes to it linearly. Time to completion is again reduced. The gains are mostly down to two things. In many cases, there is less scanning as zone_reclaim simply gives up faster due to recent failures. The second reason is that memory is used more efficiently. Instead of scanning the preferred zone every time, the allocator falls back to another zone and uses it instead improving overall memory utilisation. This patch: initialise ZLC for first zone eligible for zone_reclaim. The zonelist cache (ZLC) is used among other things to record if zone_reclaim() failed for a particular zone recently. The intention is to avoid a high cost scanning extremely long zonelists or scanning within the zone uselessly. Currently the zonelist cache is setup only after the first zone has been considered and zone_reclaim() has been called. The objective was to avoid a costly setup but zone_reclaim is itself quite expensive. If it is failing regularly such as the first eligible zone having mostly mapped pages, the cost in scanning and allocation stalls is far higher than the ZLC initialisation step. This patch initialises ZLC before the first eligible zone calls zone_reclaim(). Once initialised, it is checked whether the zone failed zone_reclaim recently. If it has, the zone is skipped. As the first zone is now being checked, additional care has to be taken about zones marked full. A zone can be marked "full" because it should not have enough unmapped pages for zone_reclaim but this is excessive as direct reclaim or kswapd may succeed where zone_reclaim fails. Only mark zones "full" after zone_reclaim fails if it failed to reclaim enough pages after scanning. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-26 00:12:29 +00:00
continue;
BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
unsigned long mark;
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
int ret;
mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
if (zone_watermark_ok(zone, order, mark,
classzone_idx, alloc_flags))
goto try_this_zone;
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim There have been a small number of complaints about significant stalls while copying large amounts of data on NUMA machines reported on a distribution bugzilla. In these cases, zone_reclaim was enabled by default due to large NUMA distances. In general, the complaints have not been about the workload itself unless it was a file server (in which case the recommendation was disable zone_reclaim). The stalls are mostly due to significant amounts of time spent scanning the preferred zone for pages to free. After a failure, it might fallback to another node (as zonelists are often node-ordered rather than zone-ordered) but stall quickly again when the next allocation attempt occurs. In bad cases, each page allocated results in a full scan of the preferred zone. Patch 1 checks the preferred zone for recent allocation failure which is particularly important if zone_reclaim has failed recently. This avoids rescanning the zone in the near future and instead falling back to another node. This may hurt node locality in some cases but a failure to zone_reclaim is more expensive than a remote access. Patch 2 clears the zlc information after direct reclaim. Otherwise, zone_reclaim can mark zones full, direct reclaim can reclaim enough pages but the zone is still not considered for allocation. This was tested on a 24-thread 2-node x86_64 machine. The tests were focused on large amounts of IO. All tests were bound to the CPUs on node-0 to avoid disturbances due to processes being scheduled on different nodes. The kernels tested are 3.0-rc6-vanilla Vanilla 3.0-rc6 zlcfirst Patch 1 applied zlcreconsider Patches 1+2 applied FS-Mark ./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288 fsmark-3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirs zlcreconsider Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%) Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%) Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%) Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%) Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%) Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%) Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%) Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 501.49 493.91 499.93 Total Elapsed Time (seconds) 2451.57 2257.48 2215.92 MMTests Statistics: vmstat Page Ins 46268 63840 66008 Page Outs 90821596 90671128 88043732 Swap Ins 0 0 0 Swap Outs 0 0 0 Direct pages scanned 13091697 8966863 8971790 Kswapd pages scanned 0 1830011 1831116 Kswapd pages reclaimed 0 1829068 1829930 Direct pages reclaimed 13037777 8956828 8648314 Kswapd efficiency 100% 99% 99% Kswapd velocity 0.000 810.643 826.346 Direct efficiency 99% 99% 96% Direct velocity 5340.128 3972.068 4048.788 Percentage direct scans 100% 83% 83% Page writes by reclaim 0 3 0 Slabs scanned 796672 720640 720256 Direct inode steals 7422667 7160012 7088638 Kswapd inode steals 0 1736840 2021238 Test completes far faster with a large increase in the number of files created per second. Standard deviation is high as a small number of iterations were much higher than the mean. The number of pages scanned by zone_reclaim is reduced and kswapd is used for more work. LARGE DD 3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirst zlcreconsider download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%) dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%) delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 125.03 118.98 122.01 Total Elapsed Time (seconds) 624.56 375.02 398.06 MMTests Statistics: vmstat Page Ins 3594216 439368 407032 Page Outs 23380832 23380488 23377444 Swap Ins 0 0 0 Swap Outs 0 436 287 Direct pages scanned 17482342 69315973 82864918 Kswapd pages scanned 0 519123 575425 Kswapd pages reclaimed 0 466501 522487 Direct pages reclaimed 5858054 2732949 2712547 Kswapd efficiency 100% 89% 90% Kswapd velocity 0.000 1384.254 1445.574 Direct efficiency 33% 3% 3% Direct velocity 27991.453 184832.737 208171.929 Percentage direct scans 100% 99% 99% Page writes by reclaim 0 5082 13917 Slabs scanned 17280 29952 35328 Direct inode steals 115257 1431122 332201 Kswapd inode steals 0 0 979532 This test downloads a large tarfile and copies it with dd a number of times - similar to the most recent bug report I've dealt with. Time to completion is reduced. The number of pages scanned directly is still disturbingly high with a low efficiency but this is likely due to the number of dirty pages encountered. The figures could probably be improved with more work around how kswapd is used and how dirty pages are handled but that is separate work and this result is significant on its own. Streaming Mapped Writer MMTests Statistics: duration User/Sys Time Running Test (seconds) 124.47 111.67 112.64 Total Elapsed Time (seconds) 2138.14 1816.30 1867.56 MMTests Statistics: vmstat Page Ins 90760 89124 89516 Page Outs 121028340 120199524 120736696 Swap Ins 0 86 55 Swap Outs 0 0 0 Direct pages scanned 114989363 96461439 96330619 Kswapd pages scanned 56430948 56965763 57075875 Kswapd pages reclaimed 27743219 27752044 27766606 Direct pages reclaimed 49777 46884 36655 Kswapd efficiency 49% 48% 48% Kswapd velocity 26392.541 31363.631 30561.736 Direct efficiency 0% 0% 0% Direct velocity 53780.091 53108.759 51581.004 Percentage direct scans 67% 62% 62% Page writes by reclaim 385 122 1513 Slabs scanned 43008 39040 42112 Direct inode steals 0 10 8 Kswapd inode steals 733 534 477 This test just creates a large file mapping and writes to it linearly. Time to completion is again reduced. The gains are mostly down to two things. In many cases, there is less scanning as zone_reclaim simply gives up faster due to recent failures. The second reason is that memory is used more efficiently. Instead of scanning the preferred zone every time, the allocator falls back to another zone and uses it instead improving overall memory utilisation. This patch: initialise ZLC for first zone eligible for zone_reclaim. The zonelist cache (ZLC) is used among other things to record if zone_reclaim() failed for a particular zone recently. The intention is to avoid a high cost scanning extremely long zonelists or scanning within the zone uselessly. Currently the zonelist cache is setup only after the first zone has been considered and zone_reclaim() has been called. The objective was to avoid a costly setup but zone_reclaim is itself quite expensive. If it is failing regularly such as the first eligible zone having mostly mapped pages, the cost in scanning and allocation stalls is far higher than the ZLC initialisation step. This patch initialises ZLC before the first eligible zone calls zone_reclaim(). Once initialised, it is checked whether the zone failed zone_reclaim recently. If it has, the zone is skipped. As the first zone is now being checked, additional care has to be taken about zones marked full. A zone can be marked "full" because it should not have enough unmapped pages for zone_reclaim but this is excessive as direct reclaim or kswapd may succeed where zone_reclaim fails. Only mark zones "full" after zone_reclaim fails if it failed to reclaim enough pages after scanning. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-26 00:12:29 +00:00
if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
/*
* we do zlc_setup if there are multiple nodes
* and before considering the first zone allowed
* by the cpuset.
*/
allowednodes = zlc_setup(zonelist, alloc_flags);
zlc_active = 1;
did_zlc_setup = 1;
}
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
if (zone_reclaim_mode == 0)
goto this_zone_full;
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim There have been a small number of complaints about significant stalls while copying large amounts of data on NUMA machines reported on a distribution bugzilla. In these cases, zone_reclaim was enabled by default due to large NUMA distances. In general, the complaints have not been about the workload itself unless it was a file server (in which case the recommendation was disable zone_reclaim). The stalls are mostly due to significant amounts of time spent scanning the preferred zone for pages to free. After a failure, it might fallback to another node (as zonelists are often node-ordered rather than zone-ordered) but stall quickly again when the next allocation attempt occurs. In bad cases, each page allocated results in a full scan of the preferred zone. Patch 1 checks the preferred zone for recent allocation failure which is particularly important if zone_reclaim has failed recently. This avoids rescanning the zone in the near future and instead falling back to another node. This may hurt node locality in some cases but a failure to zone_reclaim is more expensive than a remote access. Patch 2 clears the zlc information after direct reclaim. Otherwise, zone_reclaim can mark zones full, direct reclaim can reclaim enough pages but the zone is still not considered for allocation. This was tested on a 24-thread 2-node x86_64 machine. The tests were focused on large amounts of IO. All tests were bound to the CPUs on node-0 to avoid disturbances due to processes being scheduled on different nodes. The kernels tested are 3.0-rc6-vanilla Vanilla 3.0-rc6 zlcfirst Patch 1 applied zlcreconsider Patches 1+2 applied FS-Mark ./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288 fsmark-3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirs zlcreconsider Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%) Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%) Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%) Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%) Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%) Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%) Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%) Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 501.49 493.91 499.93 Total Elapsed Time (seconds) 2451.57 2257.48 2215.92 MMTests Statistics: vmstat Page Ins 46268 63840 66008 Page Outs 90821596 90671128 88043732 Swap Ins 0 0 0 Swap Outs 0 0 0 Direct pages scanned 13091697 8966863 8971790 Kswapd pages scanned 0 1830011 1831116 Kswapd pages reclaimed 0 1829068 1829930 Direct pages reclaimed 13037777 8956828 8648314 Kswapd efficiency 100% 99% 99% Kswapd velocity 0.000 810.643 826.346 Direct efficiency 99% 99% 96% Direct velocity 5340.128 3972.068 4048.788 Percentage direct scans 100% 83% 83% Page writes by reclaim 0 3 0 Slabs scanned 796672 720640 720256 Direct inode steals 7422667 7160012 7088638 Kswapd inode steals 0 1736840 2021238 Test completes far faster with a large increase in the number of files created per second. Standard deviation is high as a small number of iterations were much higher than the mean. The number of pages scanned by zone_reclaim is reduced and kswapd is used for more work. LARGE DD 3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirst zlcreconsider download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%) dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%) delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 125.03 118.98 122.01 Total Elapsed Time (seconds) 624.56 375.02 398.06 MMTests Statistics: vmstat Page Ins 3594216 439368 407032 Page Outs 23380832 23380488 23377444 Swap Ins 0 0 0 Swap Outs 0 436 287 Direct pages scanned 17482342 69315973 82864918 Kswapd pages scanned 0 519123 575425 Kswapd pages reclaimed 0 466501 522487 Direct pages reclaimed 5858054 2732949 2712547 Kswapd efficiency 100% 89% 90% Kswapd velocity 0.000 1384.254 1445.574 Direct efficiency 33% 3% 3% Direct velocity 27991.453 184832.737 208171.929 Percentage direct scans 100% 99% 99% Page writes by reclaim 0 5082 13917 Slabs scanned 17280 29952 35328 Direct inode steals 115257 1431122 332201 Kswapd inode steals 0 0 979532 This test downloads a large tarfile and copies it with dd a number of times - similar to the most recent bug report I've dealt with. Time to completion is reduced. The number of pages scanned directly is still disturbingly high with a low efficiency but this is likely due to the number of dirty pages encountered. The figures could probably be improved with more work around how kswapd is used and how dirty pages are handled but that is separate work and this result is significant on its own. Streaming Mapped Writer MMTests Statistics: duration User/Sys Time Running Test (seconds) 124.47 111.67 112.64 Total Elapsed Time (seconds) 2138.14 1816.30 1867.56 MMTests Statistics: vmstat Page Ins 90760 89124 89516 Page Outs 121028340 120199524 120736696 Swap Ins 0 86 55 Swap Outs 0 0 0 Direct pages scanned 114989363 96461439 96330619 Kswapd pages scanned 56430948 56965763 57075875 Kswapd pages reclaimed 27743219 27752044 27766606 Direct pages reclaimed 49777 46884 36655 Kswapd efficiency 49% 48% 48% Kswapd velocity 26392.541 31363.631 30561.736 Direct efficiency 0% 0% 0% Direct velocity 53780.091 53108.759 51581.004 Percentage direct scans 67% 62% 62% Page writes by reclaim 385 122 1513 Slabs scanned 43008 39040 42112 Direct inode steals 0 10 8 Kswapd inode steals 733 534 477 This test just creates a large file mapping and writes to it linearly. Time to completion is again reduced. The gains are mostly down to two things. In many cases, there is less scanning as zone_reclaim simply gives up faster due to recent failures. The second reason is that memory is used more efficiently. Instead of scanning the preferred zone every time, the allocator falls back to another zone and uses it instead improving overall memory utilisation. This patch: initialise ZLC for first zone eligible for zone_reclaim. The zonelist cache (ZLC) is used among other things to record if zone_reclaim() failed for a particular zone recently. The intention is to avoid a high cost scanning extremely long zonelists or scanning within the zone uselessly. Currently the zonelist cache is setup only after the first zone has been considered and zone_reclaim() has been called. The objective was to avoid a costly setup but zone_reclaim is itself quite expensive. If it is failing regularly such as the first eligible zone having mostly mapped pages, the cost in scanning and allocation stalls is far higher than the ZLC initialisation step. This patch initialises ZLC before the first eligible zone calls zone_reclaim(). Once initialised, it is checked whether the zone failed zone_reclaim recently. If it has, the zone is skipped. As the first zone is now being checked, additional care has to be taken about zones marked full. A zone can be marked "full" because it should not have enough unmapped pages for zone_reclaim but this is excessive as direct reclaim or kswapd may succeed where zone_reclaim fails. Only mark zones "full" after zone_reclaim fails if it failed to reclaim enough pages after scanning. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-26 00:12:29 +00:00
/*
* As we may have just activated ZLC, check if the first
* eligible zone has failed zone_reclaim recently.
*/
if (NUMA_BUILD && zlc_active &&
!zlc_zone_worth_trying(zonelist, z, allowednodes))
continue;
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
ret = zone_reclaim(zone, gfp_mask, order);
switch (ret) {
case ZONE_RECLAIM_NOSCAN:
/* did not scan */
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim There have been a small number of complaints about significant stalls while copying large amounts of data on NUMA machines reported on a distribution bugzilla. In these cases, zone_reclaim was enabled by default due to large NUMA distances. In general, the complaints have not been about the workload itself unless it was a file server (in which case the recommendation was disable zone_reclaim). The stalls are mostly due to significant amounts of time spent scanning the preferred zone for pages to free. After a failure, it might fallback to another node (as zonelists are often node-ordered rather than zone-ordered) but stall quickly again when the next allocation attempt occurs. In bad cases, each page allocated results in a full scan of the preferred zone. Patch 1 checks the preferred zone for recent allocation failure which is particularly important if zone_reclaim has failed recently. This avoids rescanning the zone in the near future and instead falling back to another node. This may hurt node locality in some cases but a failure to zone_reclaim is more expensive than a remote access. Patch 2 clears the zlc information after direct reclaim. Otherwise, zone_reclaim can mark zones full, direct reclaim can reclaim enough pages but the zone is still not considered for allocation. This was tested on a 24-thread 2-node x86_64 machine. The tests were focused on large amounts of IO. All tests were bound to the CPUs on node-0 to avoid disturbances due to processes being scheduled on different nodes. The kernels tested are 3.0-rc6-vanilla Vanilla 3.0-rc6 zlcfirst Patch 1 applied zlcreconsider Patches 1+2 applied FS-Mark ./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288 fsmark-3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirs zlcreconsider Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%) Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%) Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%) Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%) Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%) Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%) Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%) Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 501.49 493.91 499.93 Total Elapsed Time (seconds) 2451.57 2257.48 2215.92 MMTests Statistics: vmstat Page Ins 46268 63840 66008 Page Outs 90821596 90671128 88043732 Swap Ins 0 0 0 Swap Outs 0 0 0 Direct pages scanned 13091697 8966863 8971790 Kswapd pages scanned 0 1830011 1831116 Kswapd pages reclaimed 0 1829068 1829930 Direct pages reclaimed 13037777 8956828 8648314 Kswapd efficiency 100% 99% 99% Kswapd velocity 0.000 810.643 826.346 Direct efficiency 99% 99% 96% Direct velocity 5340.128 3972.068 4048.788 Percentage direct scans 100% 83% 83% Page writes by reclaim 0 3 0 Slabs scanned 796672 720640 720256 Direct inode steals 7422667 7160012 7088638 Kswapd inode steals 0 1736840 2021238 Test completes far faster with a large increase in the number of files created per second. Standard deviation is high as a small number of iterations were much higher than the mean. The number of pages scanned by zone_reclaim is reduced and kswapd is used for more work. LARGE DD 3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirst zlcreconsider download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%) dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%) delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 125.03 118.98 122.01 Total Elapsed Time (seconds) 624.56 375.02 398.06 MMTests Statistics: vmstat Page Ins 3594216 439368 407032 Page Outs 23380832 23380488 23377444 Swap Ins 0 0 0 Swap Outs 0 436 287 Direct pages scanned 17482342 69315973 82864918 Kswapd pages scanned 0 519123 575425 Kswapd pages reclaimed 0 466501 522487 Direct pages reclaimed 5858054 2732949 2712547 Kswapd efficiency 100% 89% 90% Kswapd velocity 0.000 1384.254 1445.574 Direct efficiency 33% 3% 3% Direct velocity 27991.453 184832.737 208171.929 Percentage direct scans 100% 99% 99% Page writes by reclaim 0 5082 13917 Slabs scanned 17280 29952 35328 Direct inode steals 115257 1431122 332201 Kswapd inode steals 0 0 979532 This test downloads a large tarfile and copies it with dd a number of times - similar to the most recent bug report I've dealt with. Time to completion is reduced. The number of pages scanned directly is still disturbingly high with a low efficiency but this is likely due to the number of dirty pages encountered. The figures could probably be improved with more work around how kswapd is used and how dirty pages are handled but that is separate work and this result is significant on its own. Streaming Mapped Writer MMTests Statistics: duration User/Sys Time Running Test (seconds) 124.47 111.67 112.64 Total Elapsed Time (seconds) 2138.14 1816.30 1867.56 MMTests Statistics: vmstat Page Ins 90760 89124 89516 Page Outs 121028340 120199524 120736696 Swap Ins 0 86 55 Swap Outs 0 0 0 Direct pages scanned 114989363 96461439 96330619 Kswapd pages scanned 56430948 56965763 57075875 Kswapd pages reclaimed 27743219 27752044 27766606 Direct pages reclaimed 49777 46884 36655 Kswapd efficiency 49% 48% 48% Kswapd velocity 26392.541 31363.631 30561.736 Direct efficiency 0% 0% 0% Direct velocity 53780.091 53108.759 51581.004 Percentage direct scans 67% 62% 62% Page writes by reclaim 385 122 1513 Slabs scanned 43008 39040 42112 Direct inode steals 0 10 8 Kswapd inode steals 733 534 477 This test just creates a large file mapping and writes to it linearly. Time to completion is again reduced. The gains are mostly down to two things. In many cases, there is less scanning as zone_reclaim simply gives up faster due to recent failures. The second reason is that memory is used more efficiently. Instead of scanning the preferred zone every time, the allocator falls back to another zone and uses it instead improving overall memory utilisation. This patch: initialise ZLC for first zone eligible for zone_reclaim. The zonelist cache (ZLC) is used among other things to record if zone_reclaim() failed for a particular zone recently. The intention is to avoid a high cost scanning extremely long zonelists or scanning within the zone uselessly. Currently the zonelist cache is setup only after the first zone has been considered and zone_reclaim() has been called. The objective was to avoid a costly setup but zone_reclaim is itself quite expensive. If it is failing regularly such as the first eligible zone having mostly mapped pages, the cost in scanning and allocation stalls is far higher than the ZLC initialisation step. This patch initialises ZLC before the first eligible zone calls zone_reclaim(). Once initialised, it is checked whether the zone failed zone_reclaim recently. If it has, the zone is skipped. As the first zone is now being checked, additional care has to be taken about zones marked full. A zone can be marked "full" because it should not have enough unmapped pages for zone_reclaim but this is excessive as direct reclaim or kswapd may succeed where zone_reclaim fails. Only mark zones "full" after zone_reclaim fails if it failed to reclaim enough pages after scanning. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-26 00:12:29 +00:00
continue;
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
case ZONE_RECLAIM_FULL:
/* scanned but unreclaimable */
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim There have been a small number of complaints about significant stalls while copying large amounts of data on NUMA machines reported on a distribution bugzilla. In these cases, zone_reclaim was enabled by default due to large NUMA distances. In general, the complaints have not been about the workload itself unless it was a file server (in which case the recommendation was disable zone_reclaim). The stalls are mostly due to significant amounts of time spent scanning the preferred zone for pages to free. After a failure, it might fallback to another node (as zonelists are often node-ordered rather than zone-ordered) but stall quickly again when the next allocation attempt occurs. In bad cases, each page allocated results in a full scan of the preferred zone. Patch 1 checks the preferred zone for recent allocation failure which is particularly important if zone_reclaim has failed recently. This avoids rescanning the zone in the near future and instead falling back to another node. This may hurt node locality in some cases but a failure to zone_reclaim is more expensive than a remote access. Patch 2 clears the zlc information after direct reclaim. Otherwise, zone_reclaim can mark zones full, direct reclaim can reclaim enough pages but the zone is still not considered for allocation. This was tested on a 24-thread 2-node x86_64 machine. The tests were focused on large amounts of IO. All tests were bound to the CPUs on node-0 to avoid disturbances due to processes being scheduled on different nodes. The kernels tested are 3.0-rc6-vanilla Vanilla 3.0-rc6 zlcfirst Patch 1 applied zlcreconsider Patches 1+2 applied FS-Mark ./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288 fsmark-3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirs zlcreconsider Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%) Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%) Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%) Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%) Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%) Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%) Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%) Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 501.49 493.91 499.93 Total Elapsed Time (seconds) 2451.57 2257.48 2215.92 MMTests Statistics: vmstat Page Ins 46268 63840 66008 Page Outs 90821596 90671128 88043732 Swap Ins 0 0 0 Swap Outs 0 0 0 Direct pages scanned 13091697 8966863 8971790 Kswapd pages scanned 0 1830011 1831116 Kswapd pages reclaimed 0 1829068 1829930 Direct pages reclaimed 13037777 8956828 8648314 Kswapd efficiency 100% 99% 99% Kswapd velocity 0.000 810.643 826.346 Direct efficiency 99% 99% 96% Direct velocity 5340.128 3972.068 4048.788 Percentage direct scans 100% 83% 83% Page writes by reclaim 0 3 0 Slabs scanned 796672 720640 720256 Direct inode steals 7422667 7160012 7088638 Kswapd inode steals 0 1736840 2021238 Test completes far faster with a large increase in the number of files created per second. Standard deviation is high as a small number of iterations were much higher than the mean. The number of pages scanned by zone_reclaim is reduced and kswapd is used for more work. LARGE DD 3.0-rc6 3.0-rc6 3.0-rc6 vanilla zlcfirst zlcreconsider download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%) dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%) delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 125.03 118.98 122.01 Total Elapsed Time (seconds) 624.56 375.02 398.06 MMTests Statistics: vmstat Page Ins 3594216 439368 407032 Page Outs 23380832 23380488 23377444 Swap Ins 0 0 0 Swap Outs 0 436 287 Direct pages scanned 17482342 69315973 82864918 Kswapd pages scanned 0 519123 575425 Kswapd pages reclaimed 0 466501 522487 Direct pages reclaimed 5858054 2732949 2712547 Kswapd efficiency 100% 89% 90% Kswapd velocity 0.000 1384.254 1445.574 Direct efficiency 33% 3% 3% Direct velocity 27991.453 184832.737 208171.929 Percentage direct scans 100% 99% 99% Page writes by reclaim 0 5082 13917 Slabs scanned 17280 29952 35328 Direct inode steals 115257 1431122 332201 Kswapd inode steals 0 0 979532 This test downloads a large tarfile and copies it with dd a number of times - similar to the most recent bug report I've dealt with. Time to completion is reduced. The number of pages scanned directly is still disturbingly high with a low efficiency but this is likely due to the number of dirty pages encountered. The figures could probably be improved with more work around how kswapd is used and how dirty pages are handled but that is separate work and this result is significant on its own. Streaming Mapped Writer MMTests Statistics: duration User/Sys Time Running Test (seconds) 124.47 111.67 112.64 Total Elapsed Time (seconds) 2138.14 1816.30 1867.56 MMTests Statistics: vmstat Page Ins 90760 89124 89516 Page Outs 121028340 120199524 120736696 Swap Ins 0 86 55 Swap Outs 0 0 0 Direct pages scanned 114989363 96461439 96330619 Kswapd pages scanned 56430948 56965763 57075875 Kswapd pages reclaimed 27743219 27752044 27766606 Direct pages reclaimed 49777 46884 36655 Kswapd efficiency 49% 48% 48% Kswapd velocity 26392.541 31363.631 30561.736 Direct efficiency 0% 0% 0% Direct velocity 53780.091 53108.759 51581.004 Percentage direct scans 67% 62% 62% Page writes by reclaim 385 122 1513 Slabs scanned 43008 39040 42112 Direct inode steals 0 10 8 Kswapd inode steals 733 534 477 This test just creates a large file mapping and writes to it linearly. Time to completion is again reduced. The gains are mostly down to two things. In many cases, there is less scanning as zone_reclaim simply gives up faster due to recent failures. The second reason is that memory is used more efficiently. Instead of scanning the preferred zone every time, the allocator falls back to another zone and uses it instead improving overall memory utilisation. This patch: initialise ZLC for first zone eligible for zone_reclaim. The zonelist cache (ZLC) is used among other things to record if zone_reclaim() failed for a particular zone recently. The intention is to avoid a high cost scanning extremely long zonelists or scanning within the zone uselessly. Currently the zonelist cache is setup only after the first zone has been considered and zone_reclaim() has been called. The objective was to avoid a costly setup but zone_reclaim is itself quite expensive. If it is failing regularly such as the first eligible zone having mostly mapped pages, the cost in scanning and allocation stalls is far higher than the ZLC initialisation step. This patch initialises ZLC before the first eligible zone calls zone_reclaim(). Once initialised, it is checked whether the zone failed zone_reclaim recently. If it has, the zone is skipped. As the first zone is now being checked, additional care has to be taken about zones marked full. A zone can be marked "full" because it should not have enough unmapped pages for zone_reclaim but this is excessive as direct reclaim or kswapd may succeed where zone_reclaim fails. Only mark zones "full" after zone_reclaim fails if it failed to reclaim enough pages after scanning. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-26 00:12:29 +00:00
continue;
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
default:
/* did we reclaim enough */
if (!zone_watermark_ok(zone, order, mark,
classzone_idx, alloc_flags))
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
goto this_zone_full;
}
}
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full On NUMA machines, the administrator can configure zone_reclaim_mode that is a more targetted form of direct reclaim. On machines with large NUMA distances for example, a zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be reclaimed if the zone watermarks are not being met. The problem is that zone_reclaim() failing at all means the zone gets marked full. This can cause situations where a zone is usable, but is being skipped because it has been considered full. Take a situation where a large tmpfs mount is occuping a large percentage of memory overall. The pages do not get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full and the zonelist cache considers them not worth trying in the future. This patch makes zone_reclaim() return more fine-grained information about what occured when zone_reclaim() failued. The zone only gets marked full if it really is unreclaimable. If it's a case that the scan did not occur or if enough pages were not reclaimed with the limited reclaim_mode, then the zone is simply skipped. There is a side-effect to this patch. Currently, if zone_reclaim() successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go ahead. With this patch applied, zone watermarks are rechecked after zone_reclaim() does some work. This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26 ("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the zonelist_cache was introduced. It was not intended that zone_reclaim() aggressively consider the zone to be full when it failed as full direct reclaim can still be an option. Due to the age of the bug, it should be considered a -stable candidate. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Rik van Riel <riel@redhat.com> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:22 +00:00
try_this_zone:
page = buffered_rmqueue(preferred_zone, zone, order,
gfp_mask, migratetype);
if (page)
break;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
this_zone_full:
if (NUMA_BUILD)
zlc_mark_zone_full(zonelist, z);
}
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
/* Disable zlc cache for second zonelist scan */
zlc_active = 0;
goto zonelist_scan;
}
return page;
[PATCH] VM: early zone reclaim This is the core of the (much simplified) early reclaim. The goal of this patch is to reclaim some easily-freed pages from a zone before falling back onto another zone. One of the major uses of this is NUMA machines. With the default allocator behavior the allocator would look for memory in another zone, which might be off-node, before trying to reclaim from the current zone. This adds a zone tuneable to enable early zone reclaim. It is selected on a per-zone basis and is turned on/off via syscall. Adding some extra throttling on the reclaim was also required (patch 4/4). Without the machine would grind to a crawl when doing a "make -j" kernel build. Even with this patch the System Time is higher on average, but it seems tolerable. Here are some numbers for kernbench runs on a 2-node, 4cpu, 8Gig RAM Altix in the "make -j" run: wall user sys %cpu ctx sw. sleeps ---- ---- --- ---- ------ ------ No patch 1009 1384 847 258 298170 504402 w/patch, no reclaim 880 1376 667 288 254064 396745 w/patch & reclaim 1079 1385 926 252 291625 548873 These numbers are the average of 2 runs of 3 "make -j" runs done right after system boot. Run-to-run variability for "make -j" is huge, so these numbers aren't terribly useful except to seee that with reclaim the benchmark still finishes in a reasonable amount of time. I also looked at the NUMA hit/miss stats for the "make -j" runs and the reclaim doesn't make any difference when the machine is thrashing away. Doing a "make -j8" on a single node that is filled with page cache pages takes 700 seconds with reclaim turned on and 735 seconds without reclaim (due to remote memory accesses). The simple zone_reclaim syscall program is at http://www.bork.org/~mort/sgi/zone_reclaim.c Signed-off-by: Martin Hicks <mort@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:41 +00:00
}
/*
* Large machines with many possible nodes should not always dump per-node
* meminfo in irq context.
*/
static inline bool should_suppress_show_mem(void)
{
bool ret = false;
#if NODES_SHIFT > 8
ret = in_interrupt();
#endif
return ret;
}
static DEFINE_RATELIMIT_STATE(nopage_rs,
DEFAULT_RATELIMIT_INTERVAL,
DEFAULT_RATELIMIT_BURST);
void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
{
unsigned int filter = SHOW_MEM_FILTER_NODES;
if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
return;
/*
* This documents exceptions given to allocations in certain
* contexts that are allowed to allocate outside current's set
* of allowed nodes.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
if (test_thread_flag(TIF_MEMDIE) ||
(current->flags & (PF_MEMALLOC | PF_EXITING)))
filter &= ~SHOW_MEM_FILTER_NODES;
if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
filter &= ~SHOW_MEM_FILTER_NODES;
if (fmt) {
struct va_format vaf;
va_list args;
va_start(args, fmt);
vaf.fmt = fmt;
vaf.va = &args;
pr_warn("%pV", &vaf);
va_end(args);
}
pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
current->comm, order, gfp_mask);
dump_stack();
if (!should_suppress_show_mem())
show_mem(filter);
}
static inline int
should_alloc_retry(gfp_t gfp_mask, unsigned int order,
unsigned long pages_reclaimed)
{
/* Do not loop if specifically requested */
if (gfp_mask & __GFP_NORETRY)
return 0;
/*
* In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
* means __GFP_NOFAIL, but that may not be true in other
* implementations.
*/
if (order <= PAGE_ALLOC_COSTLY_ORDER)
return 1;
/*
* For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
* specified, then we retry until we no longer reclaim any pages
* (above), or we've reclaimed an order of pages at least as
* large as the allocation's order. In both cases, if the
* allocation still fails, we stop retrying.
*/
if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
return 1;
lockdep: annotate reclaim context (__GFP_NOFS) Here is another version, with the incremental patch rolled up, and added reclaim context annotation to kswapd, and allocation tracing to slab allocators (which may only ever reach the page allocator in rare cases, so it is good to put annotations here too). Haven't tested this version as such, but it should be getting closer to merge worthy ;) -- After noticing some code in mm/filemap.c accidentally perform a __GFP_FS allocation when it should not have been, I thought it might be a good idea to try to catch this kind of thing with lockdep. I coded up a little idea that seems to work. Unfortunately the system has to actually be in __GFP_FS page reclaim, then take the lock, before it will mark it. But at least that might still be some orders of magnitude more common (and more debuggable) than an actual deadlock condition, so we have some improvement I hope (the concept is no less complete than discovery of a lock's interrupt contexts). I guess we could even do the same thing with __GFP_IO (normal reclaim), and even GFP_NOIO locks too... but filesystems will have the most locks and fiddly code paths, so let's start there and see how it goes. It *seems* to work. I did a quick test. ================================= [ INFO: inconsistent lock state ] 2.6.28-rc6-00007-ged31348-dirty #26 --------------------------------- inconsistent {in-reclaim-W} -> {ov-reclaim-W} usage. modprobe/8526 [HC0[0]:SC0[0]:HE1:SE1] takes: (testlock){--..}, at: [<ffffffffa0020055>] brd_init+0x55/0x216 [brd] {in-reclaim-W} state was registered at: [<ffffffff80267bdb>] __lock_acquire+0x75b/0x1a60 [<ffffffff80268f71>] lock_acquire+0x91/0xc0 [<ffffffff8070f0e1>] mutex_lock_nested+0xb1/0x310 [<ffffffffa002002b>] brd_init+0x2b/0x216 [brd] [<ffffffff8020903b>] _stext+0x3b/0x170 [<ffffffff80272ebf>] sys_init_module+0xaf/0x1e0 [<ffffffff8020c3fb>] system_call_fastpath+0x16/0x1b [<ffffffffffffffff>] 0xffffffffffffffff irq event stamp: 3929 hardirqs last enabled at (3929): [<ffffffff8070f2b5>] mutex_lock_nested+0x285/0x310 hardirqs last disabled at (3928): [<ffffffff8070f089>] mutex_lock_nested+0x59/0x310 softirqs last enabled at (3732): [<ffffffff8061f623>] sk_filter+0x83/0xe0 softirqs last disabled at (3730): [<ffffffff8061f5b6>] sk_filter+0x16/0xe0 other info that might help us debug this: 1 lock held by modprobe/8526: #0: (testlock){--..}, at: [<ffffffffa0020055>] brd_init+0x55/0x216 [brd] stack backtrace: Pid: 8526, comm: modprobe Not tainted 2.6.28-rc6-00007-ged31348-dirty #26 Call Trace: [<ffffffff80265483>] print_usage_bug+0x193/0x1d0 [<ffffffff80266530>] mark_lock+0xaf0/0xca0 [<ffffffff80266735>] mark_held_locks+0x55/0xc0 [<ffffffffa0020000>] ? brd_init+0x0/0x216 [brd] [<ffffffff802667ca>] trace_reclaim_fs+0x2a/0x60 [<ffffffff80285005>] __alloc_pages_internal+0x475/0x580 [<ffffffff8070f29e>] ? mutex_lock_nested+0x26e/0x310 [<ffffffffa0020000>] ? brd_init+0x0/0x216 [brd] [<ffffffffa002006a>] brd_init+0x6a/0x216 [brd] [<ffffffffa0020000>] ? brd_init+0x0/0x216 [brd] [<ffffffff8020903b>] _stext+0x3b/0x170 [<ffffffff8070f8b9>] ? mutex_unlock+0x9/0x10 [<ffffffff8070f83d>] ? __mutex_unlock_slowpath+0x10d/0x180 [<ffffffff802669ec>] ? trace_hardirqs_on_caller+0x12c/0x190 [<ffffffff80272ebf>] sys_init_module+0xaf/0x1e0 [<ffffffff8020c3fb>] system_call_fastpath+0x16/0x1b Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-21 07:12:39 +00:00
/*
* Don't let big-order allocations loop unless the caller
* explicitly requests that.
*/
if (gfp_mask & __GFP_NOFAIL)
return 1;
return 0;
}
static inline struct page *
__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, enum zone_type high_zoneidx,
nodemask_t *nodemask, struct zone *preferred_zone,
int migratetype)
{
struct page *page;
/* Acquire the OOM killer lock for the zones in zonelist */
if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
schedule_timeout_uninterruptible(1);
return NULL;
}
/*
* Go through the zonelist yet one more time, keep very high watermark
* here, this is only to catch a parallel oom killing, we must fail if
* we're still under heavy pressure.
*/
page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
order, zonelist, high_zoneidx,
ALLOC_WMARK_HIGH|ALLOC_CPUSET,
preferred_zone, migratetype);
if (page)
goto out;
if (!(gfp_mask & __GFP_NOFAIL)) {
/* The OOM killer will not help higher order allocs */
if (order > PAGE_ALLOC_COSTLY_ORDER)
goto out;
2010-08-10 00:18:54 +00:00
/* The OOM killer does not needlessly kill tasks for lowmem */
if (high_zoneidx < ZONE_NORMAL)
goto out;
/*
* GFP_THISNODE contains __GFP_NORETRY and we never hit this.
* Sanity check for bare calls of __GFP_THISNODE, not real OOM.
* The caller should handle page allocation failure by itself if
* it specifies __GFP_THISNODE.
* Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
*/
if (gfp_mask & __GFP_THISNODE)
goto out;
}
/* Exhausted what can be done so it's blamo time */
out_of_memory(zonelist, gfp_mask, order, nodemask);
out:
clear_zonelist_oom(zonelist, gfp_mask);
return page;
}
#ifdef CONFIG_COMPACTION
/* Try memory compaction for high-order allocations before reclaim */
static struct page *
__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, enum zone_type high_zoneidx,
nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
int migratetype, unsigned long *did_some_progress,
bool sync_migration)
{
struct page *page;
if (!order || compaction_deferred(preferred_zone))
return NULL;
current->flags |= PF_MEMALLOC;
*did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
nodemask, sync_migration);
current->flags &= ~PF_MEMALLOC;
if (*did_some_progress != COMPACT_SKIPPED) {
/* Page migration frees to the PCP lists but we want merging */
drain_pages(get_cpu());
put_cpu();
page = get_page_from_freelist(gfp_mask, nodemask,
order, zonelist, high_zoneidx,
alloc_flags, preferred_zone,
migratetype);
if (page) {
preferred_zone->compact_considered = 0;
preferred_zone->compact_defer_shift = 0;
count_vm_event(COMPACTSUCCESS);
return page;
}
/*
* It's bad if compaction run occurs and fails.
* The most likely reason is that pages exist,
* but not enough to satisfy watermarks.
*/
count_vm_event(COMPACTFAIL);
defer_compaction(preferred_zone);
cond_resched();
}
return NULL;
}
#else
static inline struct page *
__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, enum zone_type high_zoneidx,
nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
int migratetype, unsigned long *did_some_progress,
bool sync_migration)
{
return NULL;
}
#endif /* CONFIG_COMPACTION */
/* The really slow allocator path where we enter direct reclaim */
static inline struct page *
__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, enum zone_type high_zoneidx,
nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
int migratetype, unsigned long *did_some_progress)
{
struct page *page = NULL;
struct reclaim_state reclaim_state;
bool drained = false;
cond_resched();
/* We now go into synchronous reclaim */
cpuset_memory_pressure_bump();
current->flags |= PF_MEMALLOC;
lockdep_set_current_reclaim_state(gfp_mask);
reclaim_state.reclaimed_slab = 0;
current->reclaim_state = &reclaim_state;
*did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
current->reclaim_state = NULL;
lockdep_clear_current_reclaim_state();
current->flags &= ~PF_MEMALLOC;
cond_resched();
if (unlikely(!(*did_some_progress)))
return NULL;
/* After successful reclaim, reconsider all zones for allocation */
if (NUMA_BUILD)
zlc_clear_zones_full(zonelist);
retry:
page = get_page_from_freelist(gfp_mask, nodemask, order,
zonelist, high_zoneidx,
alloc_flags, preferred_zone,
migratetype);
/*
* If an allocation failed after direct reclaim, it could be because
* pages are pinned on the per-cpu lists. Drain them and try again
*/
if (!page && !drained) {
drain_all_pages();
drained = true;
goto retry;
}
return page;
}
/*
* This is called in the allocator slow-path if the allocation request is of
* sufficient urgency to ignore watermarks and take other desperate measures
*/
static inline struct page *
__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, enum zone_type high_zoneidx,
nodemask_t *nodemask, struct zone *preferred_zone,
int migratetype)
{
struct page *page;
do {
page = get_page_from_freelist(gfp_mask, nodemask, order,
zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
preferred_zone, migratetype);
if (!page && gfp_mask & __GFP_NOFAIL)
wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
} while (!page && (gfp_mask & __GFP_NOFAIL));
return page;
}
static inline
void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
mm: kswapd: stop high-order balancing when any suitable zone is balanced Simon Kirby reported the following problem We're seeing cases on a number of servers where cache never fully grows to use all available memory. Sometimes we see servers with 4 GB of memory that never seem to have less than 1.5 GB free, even with a constantly-active VM. In some cases, these servers also swap out while this happens, even though they are constantly reading the working set into memory. We have been seeing this happening for a long time; I don't think it's anything recent, and it still happens on 2.6.36. After some debugging work by Simon, Dave Hansen and others, the prevaling theory became that kswapd is reclaiming order-3 pages requested by SLUB too aggressive about it. There are two apparent problems here. On the target machine, there is a small Normal zone in comparison to DMA32. As kswapd tries to balance all zones, it would continually try reclaiming for Normal even though DMA32 was balanced enough for callers. The second problem is that sleeping_prematurely() does not use the same logic as balance_pgdat() when deciding whether to sleep or not. This keeps kswapd artifically awake. A number of tests were run and the figures from previous postings will look very different for a few reasons. One, the old figures were forcing my network card to use GFP_ATOMIC in attempt to replicate Simon's problem. Second, I previous specified slub_min_order=3 again in an attempt to reproduce Simon's problem. In this posting, I'm depending on Simon to say whether his problem is fixed or not and these figures are to show the impact to the ordinary cases. Finally, the "vmscan" figures are taken from /proc/vmstat instead of the tracepoints. There is less information but recording is less disruptive. The first test of relevance was postmark with a process running in the background reading a large amount of anonymous memory in blocks. The objective was to vaguely simulate what was happening on Simon's machine and it's memory intensive enough to have kswapd awake. POSTMARK traceonly kanyzone Transactions per second: 156.00 ( 0.00%) 153.00 (-1.96%) Data megabytes read per second: 21.51 ( 0.00%) 21.52 ( 0.05%) Data megabytes written per second: 29.28 ( 0.00%) 29.11 (-0.58%) Files created alone per second: 250.00 ( 0.00%) 416.00 (39.90%) Files create/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) Files deleted alone per second: 520.00 ( 0.00%) 420.00 (-23.81%) Files delete/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 16.58 17.4 Total Elapsed Time (seconds) 218.48 222.47 VMstat Reclaim Statistics: vmscan Direct reclaims 0 4 Direct reclaim pages scanned 0 203 Direct reclaim pages reclaimed 0 184 Kswapd pages scanned 326631 322018 Kswapd pages reclaimed 312632 309784 Kswapd low wmark quickly 1 4 Kswapd high wmark quickly 122 475 Kswapd skip congestion_wait 1 0 Pages activated 700040 705317 Pages deactivated 212113 203922 Pages written 9875 6363 Total pages scanned 326631 322221 Total pages reclaimed 312632 309968 %age total pages scanned/reclaimed 95.71% 96.20% %age total pages scanned/written 3.02% 1.97% proc vmstat: Faults Major Faults 300 254 Minor Faults 645183 660284 Page ins 493588 486704 Page outs 4960088 4986704 Swap ins 1230 661 Swap outs 9869 6355 Performance is mildly affected because kswapd is no longer doing as much work and the background memory consumer process is getting in the way. Note that kswapd scanned and reclaimed fewer pages as it's less aggressive and overall fewer pages were scanned and reclaimed. Swap in/out is particularly reduced again reflecting kswapd throwing out fewer pages. The slight performance impact is unfortunate here but it looks like a direct result of kswapd being less aggressive. As the bug report is about too many pages being freed by kswapd, it may have to be accepted for now. The second test is a streaming IO benchmark that was previously used by Johannes to show regressions in page reclaim. MICRO traceonly kanyzone User/Sys Time Running Test (seconds) 29.29 28.87 Total Elapsed Time (seconds) 492.18 488.79 VMstat Reclaim Statistics: vmscan Direct reclaims 2128 1460 Direct reclaim pages scanned 2284822 1496067 Direct reclaim pages reclaimed 148919 110937 Kswapd pages scanned 15450014 16202876 Kswapd pages reclaimed 8503697 8537897 Kswapd low wmark quickly 3100 3397 Kswapd high wmark quickly 1860 7243 Kswapd skip congestion_wait 708 801 Pages activated 9635 9573 Pages deactivated 1432 1271 Pages written 223 1130 Total pages scanned 17734836 17698943 Total pages reclaimed 8652616 8648834 %age total pages scanned/reclaimed 48.79% 48.87% %age total pages scanned/written 0.00% 0.01% proc vmstat: Faults Major Faults 165 221 Minor Faults 9655785 9656506 Page ins 3880 7228 Page outs 37692940 37480076 Swap ins 0 69 Swap outs 19 15 Again fewer pages are scanned and reclaimed as expected and this time the test completed faster. Note that kswapd is hitting its watermarks faster (low and high wmark quickly) which I expect is due to kswapd reclaiming fewer pages. I also ran fs-mark, iozone and sysbench but there is nothing interesting to report in the figures. Performance is not significantly changed and the reclaim statistics look reasonable. Tgis patch: When the allocator enters its slow path, kswapd is woken up to balance the node. It continues working until all zones within the node are balanced. For order-0 allocations, this makes perfect sense but for higher orders it can have unintended side-effects. If the zone sizes are imbalanced, kswapd may reclaim heavily within a smaller zone discarding an excessive number of pages. The user-visible behaviour is that kswapd is awake and reclaiming even though plenty of pages are free from a suitable zone. This patch alters the "balance" logic for high-order reclaim allowing kswapd to stop if any suitable zone becomes balanced to reduce the number of pages it reclaims from other zones. kswapd still tries to ensure that order-0 watermarks for all zones are met before sleeping. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Eric B Munson <emunson@mgebm.net> Cc: Simon Kirby <sim@hostway.ca> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Shaohua Li <shaohua.li@intel.com> Cc: Dave Hansen <dave@linux.vnet.ibm.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:46:20 +00:00
enum zone_type high_zoneidx,
enum zone_type classzone_idx)
{
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
struct zoneref *z;
struct zone *zone;
for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
mm: kswapd: stop high-order balancing when any suitable zone is balanced Simon Kirby reported the following problem We're seeing cases on a number of servers where cache never fully grows to use all available memory. Sometimes we see servers with 4 GB of memory that never seem to have less than 1.5 GB free, even with a constantly-active VM. In some cases, these servers also swap out while this happens, even though they are constantly reading the working set into memory. We have been seeing this happening for a long time; I don't think it's anything recent, and it still happens on 2.6.36. After some debugging work by Simon, Dave Hansen and others, the prevaling theory became that kswapd is reclaiming order-3 pages requested by SLUB too aggressive about it. There are two apparent problems here. On the target machine, there is a small Normal zone in comparison to DMA32. As kswapd tries to balance all zones, it would continually try reclaiming for Normal even though DMA32 was balanced enough for callers. The second problem is that sleeping_prematurely() does not use the same logic as balance_pgdat() when deciding whether to sleep or not. This keeps kswapd artifically awake. A number of tests were run and the figures from previous postings will look very different for a few reasons. One, the old figures were forcing my network card to use GFP_ATOMIC in attempt to replicate Simon's problem. Second, I previous specified slub_min_order=3 again in an attempt to reproduce Simon's problem. In this posting, I'm depending on Simon to say whether his problem is fixed or not and these figures are to show the impact to the ordinary cases. Finally, the "vmscan" figures are taken from /proc/vmstat instead of the tracepoints. There is less information but recording is less disruptive. The first test of relevance was postmark with a process running in the background reading a large amount of anonymous memory in blocks. The objective was to vaguely simulate what was happening on Simon's machine and it's memory intensive enough to have kswapd awake. POSTMARK traceonly kanyzone Transactions per second: 156.00 ( 0.00%) 153.00 (-1.96%) Data megabytes read per second: 21.51 ( 0.00%) 21.52 ( 0.05%) Data megabytes written per second: 29.28 ( 0.00%) 29.11 (-0.58%) Files created alone per second: 250.00 ( 0.00%) 416.00 (39.90%) Files create/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) Files deleted alone per second: 520.00 ( 0.00%) 420.00 (-23.81%) Files delete/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 16.58 17.4 Total Elapsed Time (seconds) 218.48 222.47 VMstat Reclaim Statistics: vmscan Direct reclaims 0 4 Direct reclaim pages scanned 0 203 Direct reclaim pages reclaimed 0 184 Kswapd pages scanned 326631 322018 Kswapd pages reclaimed 312632 309784 Kswapd low wmark quickly 1 4 Kswapd high wmark quickly 122 475 Kswapd skip congestion_wait 1 0 Pages activated 700040 705317 Pages deactivated 212113 203922 Pages written 9875 6363 Total pages scanned 326631 322221 Total pages reclaimed 312632 309968 %age total pages scanned/reclaimed 95.71% 96.20% %age total pages scanned/written 3.02% 1.97% proc vmstat: Faults Major Faults 300 254 Minor Faults 645183 660284 Page ins 493588 486704 Page outs 4960088 4986704 Swap ins 1230 661 Swap outs 9869 6355 Performance is mildly affected because kswapd is no longer doing as much work and the background memory consumer process is getting in the way. Note that kswapd scanned and reclaimed fewer pages as it's less aggressive and overall fewer pages were scanned and reclaimed. Swap in/out is particularly reduced again reflecting kswapd throwing out fewer pages. The slight performance impact is unfortunate here but it looks like a direct result of kswapd being less aggressive. As the bug report is about too many pages being freed by kswapd, it may have to be accepted for now. The second test is a streaming IO benchmark that was previously used by Johannes to show regressions in page reclaim. MICRO traceonly kanyzone User/Sys Time Running Test (seconds) 29.29 28.87 Total Elapsed Time (seconds) 492.18 488.79 VMstat Reclaim Statistics: vmscan Direct reclaims 2128 1460 Direct reclaim pages scanned 2284822 1496067 Direct reclaim pages reclaimed 148919 110937 Kswapd pages scanned 15450014 16202876 Kswapd pages reclaimed 8503697 8537897 Kswapd low wmark quickly 3100 3397 Kswapd high wmark quickly 1860 7243 Kswapd skip congestion_wait 708 801 Pages activated 9635 9573 Pages deactivated 1432 1271 Pages written 223 1130 Total pages scanned 17734836 17698943 Total pages reclaimed 8652616 8648834 %age total pages scanned/reclaimed 48.79% 48.87% %age total pages scanned/written 0.00% 0.01% proc vmstat: Faults Major Faults 165 221 Minor Faults 9655785 9656506 Page ins 3880 7228 Page outs 37692940 37480076 Swap ins 0 69 Swap outs 19 15 Again fewer pages are scanned and reclaimed as expected and this time the test completed faster. Note that kswapd is hitting its watermarks faster (low and high wmark quickly) which I expect is due to kswapd reclaiming fewer pages. I also ran fs-mark, iozone and sysbench but there is nothing interesting to report in the figures. Performance is not significantly changed and the reclaim statistics look reasonable. Tgis patch: When the allocator enters its slow path, kswapd is woken up to balance the node. It continues working until all zones within the node are balanced. For order-0 allocations, this makes perfect sense but for higher orders it can have unintended side-effects. If the zone sizes are imbalanced, kswapd may reclaim heavily within a smaller zone discarding an excessive number of pages. The user-visible behaviour is that kswapd is awake and reclaiming even though plenty of pages are free from a suitable zone. This patch alters the "balance" logic for high-order reclaim allowing kswapd to stop if any suitable zone becomes balanced to reduce the number of pages it reclaims from other zones. kswapd still tries to ensure that order-0 watermarks for all zones are met before sleeping. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Eric B Munson <emunson@mgebm.net> Cc: Simon Kirby <sim@hostway.ca> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Shaohua Li <shaohua.li@intel.com> Cc: Dave Hansen <dave@linux.vnet.ibm.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:46:20 +00:00
wakeup_kswapd(zone, order, classzone_idx);
}
lockdep: annotate reclaim context (__GFP_NOFS) Here is another version, with the incremental patch rolled up, and added reclaim context annotation to kswapd, and allocation tracing to slab allocators (which may only ever reach the page allocator in rare cases, so it is good to put annotations here too). Haven't tested this version as such, but it should be getting closer to merge worthy ;) -- After noticing some code in mm/filemap.c accidentally perform a __GFP_FS allocation when it should not have been, I thought it might be a good idea to try to catch this kind of thing with lockdep. I coded up a little idea that seems to work. Unfortunately the system has to actually be in __GFP_FS page reclaim, then take the lock, before it will mark it. But at least that might still be some orders of magnitude more common (and more debuggable) than an actual deadlock condition, so we have some improvement I hope (the concept is no less complete than discovery of a lock's interrupt contexts). I guess we could even do the same thing with __GFP_IO (normal reclaim), and even GFP_NOIO locks too... but filesystems will have the most locks and fiddly code paths, so let's start there and see how it goes. It *seems* to work. I did a quick test. ================================= [ INFO: inconsistent lock state ] 2.6.28-rc6-00007-ged31348-dirty #26 --------------------------------- inconsistent {in-reclaim-W} -> {ov-reclaim-W} usage. modprobe/8526 [HC0[0]:SC0[0]:HE1:SE1] takes: (testlock){--..}, at: [<ffffffffa0020055>] brd_init+0x55/0x216 [brd] {in-reclaim-W} state was registered at: [<ffffffff80267bdb>] __lock_acquire+0x75b/0x1a60 [<ffffffff80268f71>] lock_acquire+0x91/0xc0 [<ffffffff8070f0e1>] mutex_lock_nested+0xb1/0x310 [<ffffffffa002002b>] brd_init+0x2b/0x216 [brd] [<ffffffff8020903b>] _stext+0x3b/0x170 [<ffffffff80272ebf>] sys_init_module+0xaf/0x1e0 [<ffffffff8020c3fb>] system_call_fastpath+0x16/0x1b [<ffffffffffffffff>] 0xffffffffffffffff irq event stamp: 3929 hardirqs last enabled at (3929): [<ffffffff8070f2b5>] mutex_lock_nested+0x285/0x310 hardirqs last disabled at (3928): [<ffffffff8070f089>] mutex_lock_nested+0x59/0x310 softirqs last enabled at (3732): [<ffffffff8061f623>] sk_filter+0x83/0xe0 softirqs last disabled at (3730): [<ffffffff8061f5b6>] sk_filter+0x16/0xe0 other info that might help us debug this: 1 lock held by modprobe/8526: #0: (testlock){--..}, at: [<ffffffffa0020055>] brd_init+0x55/0x216 [brd] stack backtrace: Pid: 8526, comm: modprobe Not tainted 2.6.28-rc6-00007-ged31348-dirty #26 Call Trace: [<ffffffff80265483>] print_usage_bug+0x193/0x1d0 [<ffffffff80266530>] mark_lock+0xaf0/0xca0 [<ffffffff80266735>] mark_held_locks+0x55/0xc0 [<ffffffffa0020000>] ? brd_init+0x0/0x216 [brd] [<ffffffff802667ca>] trace_reclaim_fs+0x2a/0x60 [<ffffffff80285005>] __alloc_pages_internal+0x475/0x580 [<ffffffff8070f29e>] ? mutex_lock_nested+0x26e/0x310 [<ffffffffa0020000>] ? brd_init+0x0/0x216 [brd] [<ffffffffa002006a>] brd_init+0x6a/0x216 [brd] [<ffffffffa0020000>] ? brd_init+0x0/0x216 [brd] [<ffffffff8020903b>] _stext+0x3b/0x170 [<ffffffff8070f8b9>] ? mutex_unlock+0x9/0x10 [<ffffffff8070f83d>] ? __mutex_unlock_slowpath+0x10d/0x180 [<ffffffff802669ec>] ? trace_hardirqs_on_caller+0x12c/0x190 [<ffffffff80272ebf>] sys_init_module+0xaf/0x1e0 [<ffffffff8020c3fb>] system_call_fastpath+0x16/0x1b Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-21 07:12:39 +00:00
static inline int
gfp_to_alloc_flags(gfp_t gfp_mask)
{
int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
const gfp_t wait = gfp_mask & __GFP_WAIT;
/* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
/*
* The caller may dip into page reserves a bit more if the caller
* cannot run direct reclaim, or if the caller has realtime scheduling
* policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
* set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
*/
alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
if (!wait) {
/*
* Not worth trying to allocate harder for
* __GFP_NOMEMALLOC even if it can't schedule.
*/
if (!(gfp_mask & __GFP_NOMEMALLOC))
alloc_flags |= ALLOC_HARDER;
/*
* Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
* See also cpuset_zone_allowed() comment in kernel/cpuset.c.
*/
alloc_flags &= ~ALLOC_CPUSET;
} else if (unlikely(rt_task(current)) && !in_interrupt())
alloc_flags |= ALLOC_HARDER;
if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
if (!in_interrupt() &&
((current->flags & PF_MEMALLOC) ||
unlikely(test_thread_flag(TIF_MEMDIE))))
alloc_flags |= ALLOC_NO_WATERMARKS;
}
return alloc_flags;
}
static inline struct page *
__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, enum zone_type high_zoneidx,
nodemask_t *nodemask, struct zone *preferred_zone,
int migratetype)
{
const gfp_t wait = gfp_mask & __GFP_WAIT;
struct page *page = NULL;
int alloc_flags;
unsigned long pages_reclaimed = 0;
unsigned long did_some_progress;
bool sync_migration = false;
/*
* In the slowpath, we sanity check order to avoid ever trying to
* reclaim >= MAX_ORDER areas which will never succeed. Callers may
* be using allocators in order of preference for an area that is
* too large.
*/
if (order >= MAX_ORDER) {
WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
return NULL;
}
/*
* GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
* __GFP_NOWARN set) should not cause reclaim since the subsystem
* (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
* using a larger set of nodes after it has established that the
* allowed per node queues are empty and that nodes are
* over allocated.
*/
if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
goto nopage;
restart:
if (!(gfp_mask & __GFP_NO_KSWAPD))
wake_all_kswapd(order, zonelist, high_zoneidx,
mm: kswapd: stop high-order balancing when any suitable zone is balanced Simon Kirby reported the following problem We're seeing cases on a number of servers where cache never fully grows to use all available memory. Sometimes we see servers with 4 GB of memory that never seem to have less than 1.5 GB free, even with a constantly-active VM. In some cases, these servers also swap out while this happens, even though they are constantly reading the working set into memory. We have been seeing this happening for a long time; I don't think it's anything recent, and it still happens on 2.6.36. After some debugging work by Simon, Dave Hansen and others, the prevaling theory became that kswapd is reclaiming order-3 pages requested by SLUB too aggressive about it. There are two apparent problems here. On the target machine, there is a small Normal zone in comparison to DMA32. As kswapd tries to balance all zones, it would continually try reclaiming for Normal even though DMA32 was balanced enough for callers. The second problem is that sleeping_prematurely() does not use the same logic as balance_pgdat() when deciding whether to sleep or not. This keeps kswapd artifically awake. A number of tests were run and the figures from previous postings will look very different for a few reasons. One, the old figures were forcing my network card to use GFP_ATOMIC in attempt to replicate Simon's problem. Second, I previous specified slub_min_order=3 again in an attempt to reproduce Simon's problem. In this posting, I'm depending on Simon to say whether his problem is fixed or not and these figures are to show the impact to the ordinary cases. Finally, the "vmscan" figures are taken from /proc/vmstat instead of the tracepoints. There is less information but recording is less disruptive. The first test of relevance was postmark with a process running in the background reading a large amount of anonymous memory in blocks. The objective was to vaguely simulate what was happening on Simon's machine and it's memory intensive enough to have kswapd awake. POSTMARK traceonly kanyzone Transactions per second: 156.00 ( 0.00%) 153.00 (-1.96%) Data megabytes read per second: 21.51 ( 0.00%) 21.52 ( 0.05%) Data megabytes written per second: 29.28 ( 0.00%) 29.11 (-0.58%) Files created alone per second: 250.00 ( 0.00%) 416.00 (39.90%) Files create/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) Files deleted alone per second: 520.00 ( 0.00%) 420.00 (-23.81%) Files delete/transact per second: 79.00 ( 0.00%) 76.00 (-3.95%) MMTests Statistics: duration User/Sys Time Running Test (seconds) 16.58 17.4 Total Elapsed Time (seconds) 218.48 222.47 VMstat Reclaim Statistics: vmscan Direct reclaims 0 4 Direct reclaim pages scanned 0 203 Direct reclaim pages reclaimed 0 184 Kswapd pages scanned 326631 322018 Kswapd pages reclaimed 312632 309784 Kswapd low wmark quickly 1 4 Kswapd high wmark quickly 122 475 Kswapd skip congestion_wait 1 0 Pages activated 700040 705317 Pages deactivated 212113 203922 Pages written 9875 6363 Total pages scanned 326631 322221 Total pages reclaimed 312632 309968 %age total pages scanned/reclaimed 95.71% 96.20% %age total pages scanned/written 3.02% 1.97% proc vmstat: Faults Major Faults 300 254 Minor Faults 645183 660284 Page ins 493588 486704 Page outs 4960088 4986704 Swap ins 1230 661 Swap outs 9869 6355 Performance is mildly affected because kswapd is no longer doing as much work and the background memory consumer process is getting in the way. Note that kswapd scanned and reclaimed fewer pages as it's less aggressive and overall fewer pages were scanned and reclaimed. Swap in/out is particularly reduced again reflecting kswapd throwing out fewer pages. The slight performance impact is unfortunate here but it looks like a direct result of kswapd being less aggressive. As the bug report is about too many pages being freed by kswapd, it may have to be accepted for now. The second test is a streaming IO benchmark that was previously used by Johannes to show regressions in page reclaim. MICRO traceonly kanyzone User/Sys Time Running Test (seconds) 29.29 28.87 Total Elapsed Time (seconds) 492.18 488.79 VMstat Reclaim Statistics: vmscan Direct reclaims 2128 1460 Direct reclaim pages scanned 2284822 1496067 Direct reclaim pages reclaimed 148919 110937 Kswapd pages scanned 15450014 16202876 Kswapd pages reclaimed 8503697 8537897 Kswapd low wmark quickly 3100 3397 Kswapd high wmark quickly 1860 7243 Kswapd skip congestion_wait 708 801 Pages activated 9635 9573 Pages deactivated 1432 1271 Pages written 223 1130 Total pages scanned 17734836 17698943 Total pages reclaimed 8652616 8648834 %age total pages scanned/reclaimed 48.79% 48.87% %age total pages scanned/written 0.00% 0.01% proc vmstat: Faults Major Faults 165 221 Minor Faults 9655785 9656506 Page ins 3880 7228 Page outs 37692940 37480076 Swap ins 0 69 Swap outs 19 15 Again fewer pages are scanned and reclaimed as expected and this time the test completed faster. Note that kswapd is hitting its watermarks faster (low and high wmark quickly) which I expect is due to kswapd reclaiming fewer pages. I also ran fs-mark, iozone and sysbench but there is nothing interesting to report in the figures. Performance is not significantly changed and the reclaim statistics look reasonable. Tgis patch: When the allocator enters its slow path, kswapd is woken up to balance the node. It continues working until all zones within the node are balanced. For order-0 allocations, this makes perfect sense but for higher orders it can have unintended side-effects. If the zone sizes are imbalanced, kswapd may reclaim heavily within a smaller zone discarding an excessive number of pages. The user-visible behaviour is that kswapd is awake and reclaiming even though plenty of pages are free from a suitable zone. This patch alters the "balance" logic for high-order reclaim allowing kswapd to stop if any suitable zone becomes balanced to reduce the number of pages it reclaims from other zones. kswapd still tries to ensure that order-0 watermarks for all zones are met before sleeping. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Reviewed-by: Eric B Munson <emunson@mgebm.net> Cc: Simon Kirby <sim@hostway.ca> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Shaohua Li <shaohua.li@intel.com> Cc: Dave Hansen <dave@linux.vnet.ibm.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:46:20 +00:00
zone_idx(preferred_zone));
[PATCH] cpusets: formalize intermediate GFP_KERNEL containment This patch makes use of the previously underutilized cpuset flag 'mem_exclusive' to provide what amounts to another layer of memory placement resolution. With this patch, there are now the following four layers of memory placement available: 1) The whole system (interrupt and GFP_ATOMIC allocations can use this), 2) The nearest enclosing mem_exclusive cpuset (GFP_KERNEL allocations can use), 3) The current tasks cpuset (GFP_USER allocations constrained to here), and 4) Specific node placement, using mbind and set_mempolicy. These nest - each layer is a subset (same or within) of the previous. Layer (2) above is new, with this patch. The call used to check whether a zone (its node, actually) is in a cpuset (in its mems_allowed, actually) is extended to take a gfp_mask argument, and its logic is extended, in the case that __GFP_HARDWALL is not set in the flag bits, to look up the cpuset hierarchy for the nearest enclosing mem_exclusive cpuset, to determine if placement is allowed. The definition of GFP_USER, which used to be identical to GFP_KERNEL, is changed to also set the __GFP_HARDWALL bit, in the previous cpuset_gfp_hardwall_flag patch. GFP_ATOMIC and GFP_KERNEL allocations will stay within the current tasks cpuset, so long as any node therein is not too tight on memory, but will escape to the larger layer, if need be. The intended use is to allow something like a batch manager to handle several jobs, each job in its own cpuset, but using common kernel memory for caches and such. Swapper and oom_kill activity is also constrained to Layer (2). A task in or below one mem_exclusive cpuset should not cause swapping on nodes in another non-overlapping mem_exclusive cpuset, nor provoke oom_killing of a task in another such cpuset. Heavy use of kernel memory for i/o caching and such by one job should not impact the memory available to jobs in other non-overlapping mem_exclusive cpusets. This patch enables providing hardwall, inescapable cpusets for memory allocations of each job, while sharing kernel memory allocations between several jobs, in an enclosing mem_exclusive cpuset. Like Dinakar's patch earlier to enable administering sched domains using the cpu_exclusive flag, this patch also provides a useful meaning to a cpuset flag that had previously done nothing much useful other than restrict what cpuset configurations were allowed. Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 22:18:12 +00:00
/*
* OK, we're below the kswapd watermark and have kicked background
* reclaim. Now things get more complex, so set up alloc_flags according
* to how we want to proceed.
[PATCH] cpusets: formalize intermediate GFP_KERNEL containment This patch makes use of the previously underutilized cpuset flag 'mem_exclusive' to provide what amounts to another layer of memory placement resolution. With this patch, there are now the following four layers of memory placement available: 1) The whole system (interrupt and GFP_ATOMIC allocations can use this), 2) The nearest enclosing mem_exclusive cpuset (GFP_KERNEL allocations can use), 3) The current tasks cpuset (GFP_USER allocations constrained to here), and 4) Specific node placement, using mbind and set_mempolicy. These nest - each layer is a subset (same or within) of the previous. Layer (2) above is new, with this patch. The call used to check whether a zone (its node, actually) is in a cpuset (in its mems_allowed, actually) is extended to take a gfp_mask argument, and its logic is extended, in the case that __GFP_HARDWALL is not set in the flag bits, to look up the cpuset hierarchy for the nearest enclosing mem_exclusive cpuset, to determine if placement is allowed. The definition of GFP_USER, which used to be identical to GFP_KERNEL, is changed to also set the __GFP_HARDWALL bit, in the previous cpuset_gfp_hardwall_flag patch. GFP_ATOMIC and GFP_KERNEL allocations will stay within the current tasks cpuset, so long as any node therein is not too tight on memory, but will escape to the larger layer, if need be. The intended use is to allow something like a batch manager to handle several jobs, each job in its own cpuset, but using common kernel memory for caches and such. Swapper and oom_kill activity is also constrained to Layer (2). A task in or below one mem_exclusive cpuset should not cause swapping on nodes in another non-overlapping mem_exclusive cpuset, nor provoke oom_killing of a task in another such cpuset. Heavy use of kernel memory for i/o caching and such by one job should not impact the memory available to jobs in other non-overlapping mem_exclusive cpusets. This patch enables providing hardwall, inescapable cpusets for memory allocations of each job, while sharing kernel memory allocations between several jobs, in an enclosing mem_exclusive cpuset. Like Dinakar's patch earlier to enable administering sched domains using the cpu_exclusive flag, this patch also provides a useful meaning to a cpuset flag that had previously done nothing much useful other than restrict what cpuset configurations were allowed. Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-06 22:18:12 +00:00
*/
alloc_flags = gfp_to_alloc_flags(gfp_mask);
mm: fix deferred congestion timeout if preferred zone is not allowed Before 0e093d99763e ("writeback: do not sleep on the congestion queue if there are no congested BDIs or if significant congestion is not being encountered in the current zone"), preferred_zone was only used for NUMA statistics, to determine the zoneidx from which to allocate from given the type requested, and whether to utilize memory compaction. wait_iff_congested(), though, uses preferred_zone to determine if the congestion wait should be deferred because its dirty pages are backed by a congested bdi. This incorrectly defers the timeout and busy loops in the page allocator with various cond_resched() calls if preferred_zone is not allowed in the current context, usually consuming 100% of a cpu. This patch ensures preferred_zone is an allowed zone in the fastpath depending on whether current is constrained by its cpuset or nodes in its mempolicy (when the nodemask passed is non-NULL). This is correct since the fastpath allocation always passes ALLOC_CPUSET when trying to allocate memory. In the slowpath, this patch resets preferred_zone to the first zone of the allowed type when the allocation is not constrained by current's cpuset, i.e. it does not pass ALLOC_CPUSET. This patch also ensures preferred_zone is from the set of allowed nodes when called from within direct reclaim since allocations are always constrained by cpusets in this context (it is blockable). Both of these uses of cpuset_current_mems_allowed are protected by get_mems_allowed(). Signed-off-by: David Rientjes <rientjes@google.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-25 23:07:20 +00:00
/*
* Find the true preferred zone if the allocation is unconstrained by
* cpusets.
*/
if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
first_zones_zonelist(zonelist, high_zoneidx, NULL,
&preferred_zone);
mm/page_alloc.c: prevent unending loop in __alloc_pages_slowpath() I believe I found a problem in __alloc_pages_slowpath, which allows a process to get stuck endlessly looping, even when lots of memory is available. Running an I/O and memory intensive stress-test I see a 0-order page allocation with __GFP_IO and __GFP_WAIT, running on a system with very little free memory. Right about the same time that the stress-test gets killed by the OOM-killer, the utility trying to allocate memory gets stuck in __alloc_pages_slowpath even though most of the systems memory was freed by the oom-kill of the stress-test. The utility ends up looping from the rebalance label down through the wait_iff_congested continiously. Because order=0, __alloc_pages_direct_compact skips the call to get_page_from_freelist. Because all of the reclaimable memory on the system has already been reclaimed, __alloc_pages_direct_reclaim skips the call to get_page_from_freelist. Since there is no __GFP_FS flag, the block with __alloc_pages_may_oom is skipped. The loop hits the wait_iff_congested, then jumps back to rebalance without ever trying to get_page_from_freelist. This loop repeats infinitely. The test case is pretty pathological. Running a mix of I/O stress-tests that do a lot of fork() and consume all of the system memory, I can pretty reliably hit this on 600 nodes, in about 12 hours. 32GB/node. Signed-off-by: Andrew Barry <abarry@cray.com> Signed-off-by: Minchan Kim <minchan.kim@gmail.com> Reviewed-by: Rik van Riel<riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-25 00:12:52 +00:00
rebalance:
/* This is the last chance, in general, before the goto nopage. */
page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
preferred_zone, migratetype);
if (page)
goto got_pg;
/* Allocate without watermarks if the context allows */
if (alloc_flags & ALLOC_NO_WATERMARKS) {
page = __alloc_pages_high_priority(gfp_mask, order,
zonelist, high_zoneidx, nodemask,
preferred_zone, migratetype);
if (page)
goto got_pg;
}
/* Atomic allocations - we can't balance anything */
if (!wait)
goto nopage;
/* Avoid recursion of direct reclaim */
if (current->flags & PF_MEMALLOC)
goto nopage;
/* Avoid allocations with no watermarks from looping endlessly */
if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
goto nopage;
/*
* Try direct compaction. The first pass is asynchronous. Subsequent
* attempts after direct reclaim are synchronous
*/
page = __alloc_pages_direct_compact(gfp_mask, order,
zonelist, high_zoneidx,
nodemask,
alloc_flags, preferred_zone,
migratetype, &did_some_progress,
sync_migration);
if (page)
goto got_pg;
sync_migration = true;
/* Try direct reclaim and then allocating */
page = __alloc_pages_direct_reclaim(gfp_mask, order,
zonelist, high_zoneidx,
nodemask,
alloc_flags, preferred_zone,
migratetype, &did_some_progress);
if (page)
goto got_pg;
/*
* If we failed to make any progress reclaiming, then we are
* running out of options and have to consider going OOM
*/
if (!did_some_progress) {
if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
mm, PM/Freezer: Disable OOM killer when tasks are frozen Currently, the following scenario appears to be possible in theory: * Tasks are frozen for hibernation or suspend. * Free pages are almost exhausted. * Certain piece of code in the suspend code path attempts to allocate some memory using GFP_KERNEL and allocation order less than or equal to PAGE_ALLOC_COSTLY_ORDER. * __alloc_pages_internal() cannot find a free page so it invokes the OOM killer. * The OOM killer attempts to kill a task, but the task is frozen, so it doesn't die immediately. * __alloc_pages_internal() jumps to 'restart', unsuccessfully tries to find a free page and invokes the OOM killer. * No progress can be made. Although it is now hard to trigger during hibernation due to the memory shrinking carried out by the hibernation code, it is theoretically possible to trigger during suspend after the memory shrinking has been removed from that code path. Moreover, since memory allocations are going to be used for the hibernation memory shrinking, it will be even more likely to happen during hibernation. To prevent it from happening, introduce the oom_killer_disabled switch that will cause __alloc_pages_internal() to fail in the situations in which the OOM killer would have been called and make the freezer set this switch after tasks have been successfully frozen. [akpm@linux-foundation.org: be nicer to the namespace] Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl> Cc: Fengguang Wu <fengguang.wu@gmail.com> Cc: David Rientjes <rientjes@google.com> Acked-by: Pavel Machek <pavel@ucw.cz> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:32:41 +00:00
if (oom_killer_disabled)
goto nopage;
page = __alloc_pages_may_oom(gfp_mask, order,
zonelist, high_zoneidx,
nodemask, preferred_zone,
migratetype);
if (page)
goto got_pg;
2010-08-10 00:18:54 +00:00
if (!(gfp_mask & __GFP_NOFAIL)) {
/*
* The oom killer is not called for high-order
* allocations that may fail, so if no progress
* is being made, there are no other options and
* retrying is unlikely to help.
*/
if (order > PAGE_ALLOC_COSTLY_ORDER)
goto nopage;
/*
* The oom killer is not called for lowmem
* allocations to prevent needlessly killing
* innocent tasks.
*/
if (high_zoneidx < ZONE_NORMAL)
goto nopage;
}
goto restart;
}
}
/* Check if we should retry the allocation */
page allocator: smarter retry of costly-order allocations Because of page order checks in __alloc_pages(), hugepage (and similarly large order) allocations will not retry unless explicitly marked __GFP_REPEAT. However, the current retry logic is nearly an infinite loop (or until reclaim does no progress whatsoever). For these costly allocations, that seems like overkill and could potentially never terminate. Mel observed that allowing current __GFP_REPEAT semantics for hugepage allocations essentially killed the system. I believe this is because we may continue to reclaim small orders of pages all over, but never have enough to satisfy the hugepage allocation request. This is clearly only a problem for large order allocations, of which hugepages are the most obvious (to me). Modify try_to_free_pages() to indicate how many pages were reclaimed. Use that information in __alloc_pages() to eventually fail a large __GFP_REPEAT allocation when we've reclaimed an order of pages equal to or greater than the allocation's order. This relies on lumpy reclaim functioning as advertised. Due to fragmentation, lumpy reclaim may not be able to free up the order needed in one invocation, so multiple iterations may be requred. In other words, the more fragmented memory is, the more retry attempts __GFP_REPEAT will make (particularly for higher order allocations). This changes the semantics of __GFP_REPEAT subtly, but *only* for allocations > PAGE_ALLOC_COSTLY_ORDER. With this patch, for those size allocations, we will try up to some point (at least 1<<order reclaimed pages), rather than forever (which is the case for allocations <= PAGE_ALLOC_COSTLY_ORDER). This change improves the /proc/sys/vm/nr_hugepages interface with a follow-on patch that makes pool allocations use __GFP_REPEAT. Rather than administrators repeatedly echo'ing a particular value into the sysctl, and forcing reclaim into action manually, this change allows for the sysctl to attempt a reasonable effort itself. Similarly, dynamic pool growth should be more successful under load, as lumpy reclaim can try to free up pages, rather than failing right away. Choosing to reclaim only up to the order of the requested allocation strikes a balance between not failing hugepage allocations and returning to the caller when it's unlikely to every succeed. Because of lumpy reclaim, if we have freed the order requested, hopefully it has been in big chunks and those chunks will allow our allocation to succeed. If that isn't the case after freeing up the current order, I don't think it is likely to succeed in the future, although it is possible given a particular fragmentation pattern. Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Tested-by: Mel Gorman <mel@csn.ul.ie> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-29 07:58:25 +00:00
pages_reclaimed += did_some_progress;
if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
/* Wait for some write requests to complete then retry */
wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
goto rebalance;
} else {
/*
* High-order allocations do not necessarily loop after
* direct reclaim and reclaim/compaction depends on compaction
* being called after reclaim so call directly if necessary
*/
page = __alloc_pages_direct_compact(gfp_mask, order,
zonelist, high_zoneidx,
nodemask,
alloc_flags, preferred_zone,
migratetype, &did_some_progress,
sync_migration);
if (page)
goto got_pg;
}
nopage:
warn_alloc_failed(gfp_mask, order, NULL);
return page;
got_pg:
if (kmemcheck_enabled)
kmemcheck_pagealloc_alloc(page, order, gfp_mask);
return page;
}
/*
* This is the 'heart' of the zoned buddy allocator.
*/
struct page *
__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
struct zonelist *zonelist, nodemask_t *nodemask)
{
enum zone_type high_zoneidx = gfp_zone(gfp_mask);
struct zone *preferred_zone;
struct page *page;
int migratetype = allocflags_to_migratetype(gfp_mask);
gfp_mask &= gfp_allowed_mask;
lockdep_trace_alloc(gfp_mask);
might_sleep_if(gfp_mask & __GFP_WAIT);
if (should_fail_alloc_page(gfp_mask, order))
return NULL;
/*
* Check the zones suitable for the gfp_mask contain at least one
* valid zone. It's possible to have an empty zonelist as a result
* of GFP_THISNODE and a memoryless node
*/
if (unlikely(!zonelist->_zonerefs->zone))
return NULL;
cpuset,mm: fix no node to alloc memory when changing cpuset's mems Before applying this patch, cpuset updates task->mems_allowed and mempolicy by setting all new bits in the nodemask first, and clearing all old unallowed bits later. But in the way, the allocator may find that there is no node to alloc memory. The reason is that cpuset rebinds the task's mempolicy, it cleans the nodes which the allocater can alloc pages on, for example: (mpol: mempolicy) task1 task1's mpol task2 alloc page 1 alloc on node0? NO 1 1 change mems from 1 to 0 1 rebind task1's mpol 0-1 set new bits 0 clear disallowed bits alloc on node1? NO 0 ... can't alloc page goto oom This patch fixes this problem by expanding the nodes range first(set newly allowed bits) and shrink it lazily(clear newly disallowed bits). So we use a variable to tell the write-side task that read-side task is reading nodemask, and the write-side task clears newly disallowed nodes after read-side task ends the current memory allocation. [akpm@linux-foundation.org: fix spello] Signed-off-by: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Paul Menage <menage@google.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Andi Kleen <andi@firstfloor.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:08 +00:00
get_mems_allowed();
/* The preferred zone is used for statistics later */
mm: fix deferred congestion timeout if preferred zone is not allowed Before 0e093d99763e ("writeback: do not sleep on the congestion queue if there are no congested BDIs or if significant congestion is not being encountered in the current zone"), preferred_zone was only used for NUMA statistics, to determine the zoneidx from which to allocate from given the type requested, and whether to utilize memory compaction. wait_iff_congested(), though, uses preferred_zone to determine if the congestion wait should be deferred because its dirty pages are backed by a congested bdi. This incorrectly defers the timeout and busy loops in the page allocator with various cond_resched() calls if preferred_zone is not allowed in the current context, usually consuming 100% of a cpu. This patch ensures preferred_zone is an allowed zone in the fastpath depending on whether current is constrained by its cpuset or nodes in its mempolicy (when the nodemask passed is non-NULL). This is correct since the fastpath allocation always passes ALLOC_CPUSET when trying to allocate memory. In the slowpath, this patch resets preferred_zone to the first zone of the allowed type when the allocation is not constrained by current's cpuset, i.e. it does not pass ALLOC_CPUSET. This patch also ensures preferred_zone is from the set of allowed nodes when called from within direct reclaim since allocations are always constrained by cpusets in this context (it is blockable). Both of these uses of cpuset_current_mems_allowed are protected by get_mems_allowed(). Signed-off-by: David Rientjes <rientjes@google.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-25 23:07:20 +00:00
first_zones_zonelist(zonelist, high_zoneidx,
nodemask ? : &cpuset_current_mems_allowed,
&preferred_zone);
cpuset,mm: fix no node to alloc memory when changing cpuset's mems Before applying this patch, cpuset updates task->mems_allowed and mempolicy by setting all new bits in the nodemask first, and clearing all old unallowed bits later. But in the way, the allocator may find that there is no node to alloc memory. The reason is that cpuset rebinds the task's mempolicy, it cleans the nodes which the allocater can alloc pages on, for example: (mpol: mempolicy) task1 task1's mpol task2 alloc page 1 alloc on node0? NO 1 1 change mems from 1 to 0 1 rebind task1's mpol 0-1 set new bits 0 clear disallowed bits alloc on node1? NO 0 ... can't alloc page goto oom This patch fixes this problem by expanding the nodes range first(set newly allowed bits) and shrink it lazily(clear newly disallowed bits). So we use a variable to tell the write-side task that read-side task is reading nodemask, and the write-side task clears newly disallowed nodes after read-side task ends the current memory allocation. [akpm@linux-foundation.org: fix spello] Signed-off-by: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Paul Menage <menage@google.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Andi Kleen <andi@firstfloor.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:08 +00:00
if (!preferred_zone) {
put_mems_allowed();
return NULL;
cpuset,mm: fix no node to alloc memory when changing cpuset's mems Before applying this patch, cpuset updates task->mems_allowed and mempolicy by setting all new bits in the nodemask first, and clearing all old unallowed bits later. But in the way, the allocator may find that there is no node to alloc memory. The reason is that cpuset rebinds the task's mempolicy, it cleans the nodes which the allocater can alloc pages on, for example: (mpol: mempolicy) task1 task1's mpol task2 alloc page 1 alloc on node0? NO 1 1 change mems from 1 to 0 1 rebind task1's mpol 0-1 set new bits 0 clear disallowed bits alloc on node1? NO 0 ... can't alloc page goto oom This patch fixes this problem by expanding the nodes range first(set newly allowed bits) and shrink it lazily(clear newly disallowed bits). So we use a variable to tell the write-side task that read-side task is reading nodemask, and the write-side task clears newly disallowed nodes after read-side task ends the current memory allocation. [akpm@linux-foundation.org: fix spello] Signed-off-by: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Paul Menage <menage@google.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Andi Kleen <andi@firstfloor.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:08 +00:00
}
/* First allocation attempt */
page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
preferred_zone, migratetype);
if (unlikely(!page))
page = __alloc_pages_slowpath(gfp_mask, order,
zonelist, high_zoneidx, nodemask,
preferred_zone, migratetype);
cpuset,mm: fix no node to alloc memory when changing cpuset's mems Before applying this patch, cpuset updates task->mems_allowed and mempolicy by setting all new bits in the nodemask first, and clearing all old unallowed bits later. But in the way, the allocator may find that there is no node to alloc memory. The reason is that cpuset rebinds the task's mempolicy, it cleans the nodes which the allocater can alloc pages on, for example: (mpol: mempolicy) task1 task1's mpol task2 alloc page 1 alloc on node0? NO 1 1 change mems from 1 to 0 1 rebind task1's mpol 0-1 set new bits 0 clear disallowed bits alloc on node1? NO 0 ... can't alloc page goto oom This patch fixes this problem by expanding the nodes range first(set newly allowed bits) and shrink it lazily(clear newly disallowed bits). So we use a variable to tell the write-side task that read-side task is reading nodemask, and the write-side task clears newly disallowed nodes after read-side task ends the current memory allocation. [akpm@linux-foundation.org: fix spello] Signed-off-by: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Paul Menage <menage@google.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Andi Kleen <andi@firstfloor.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:08 +00:00
put_mems_allowed();
trace_mm_page_alloc(page, order, gfp_mask, migratetype);
return page;
}
page allocator: replace __alloc_pages_internal() with __alloc_pages_nodemask() The start of a large patch series to clean up and optimise the page allocator. The performance improvements are in a wide range depending on the exact machine but the results I've seen so fair are approximately; kernbench: 0 to 0.12% (elapsed time) 0.49% to 3.20% (sys time) aim9: -4% to 30% (for page_test and brk_test) tbench: -1% to 4% hackbench: -2.5% to 3.45% (mostly within the noise though) netperf-udp -1.34% to 4.06% (varies between machines a bit) netperf-tcp -0.44% to 5.22% (varies between machines a bit) I haven't sysbench figures at hand, but previously they were within the -0.5% to 2% range. On netperf, the client and server were bound to opposite number CPUs to maximise the problems with cache line bouncing of the struct pages so I expect different people to report different results for netperf depending on their exact machine and how they ran the test (different machines, same cpus client/server, shared cache but two threads client/server, different socket client/server etc). I also measured the vmlinux sizes for a single x86-based config with CONFIG_DEBUG_INFO enabled but not CONFIG_DEBUG_VM. The core of the .config is based on the Debian Lenny kernel config so I expect it to be reasonably typical. This patch: __alloc_pages_internal is the core page allocator function but essentially it is an alias of __alloc_pages_nodemask. Naming a publicly available and exported function "internal" is also a big ugly. This patch renames __alloc_pages_internal() to __alloc_pages_nodemask() and deletes the old nodemask function. Warning - This patch renames an exported symbol. No kernel driver is affected by external drivers calling __alloc_pages_internal() should change the call to __alloc_pages_nodemask() without any alteration of parameters. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Dave Hansen <dave@linux.vnet.ibm.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:31:52 +00:00
EXPORT_SYMBOL(__alloc_pages_nodemask);
/*
* Common helper functions.
*/
unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
{
struct page *page;
/*
* __get_free_pages() returns a 32-bit address, which cannot represent
* a highmem page
*/
VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
page = alloc_pages(gfp_mask, order);
if (!page)
return 0;
return (unsigned long) page_address(page);
}
EXPORT_SYMBOL(__get_free_pages);
unsigned long get_zeroed_page(gfp_t gfp_mask)
{
return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
}
EXPORT_SYMBOL(get_zeroed_page);
void __pagevec_free(struct pagevec *pvec)
{
int i = pagevec_count(pvec);
while (--i >= 0) {
trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
free_hot_cold_page(pvec->pages[i], pvec->cold);
}
}
void __free_pages(struct page *page, unsigned int order)
{
2005-10-30 01:16:12 +00:00
if (put_page_testzero(page)) {
if (order == 0)
free_hot_cold_page(page, 0);
else
__free_pages_ok(page, order);
}
}
EXPORT_SYMBOL(__free_pages);
void free_pages(unsigned long addr, unsigned int order)
{
if (addr != 0) {
VM_BUG_ON(!virt_addr_valid((void *)addr));
__free_pages(virt_to_page((void *)addr), order);
}
}
EXPORT_SYMBOL(free_pages);
static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
{
if (addr) {
unsigned long alloc_end = addr + (PAGE_SIZE << order);
unsigned long used = addr + PAGE_ALIGN(size);
split_page(virt_to_page((void *)addr), order);
while (used < alloc_end) {
free_page(used);
used += PAGE_SIZE;
}
}
return (void *)addr;
}
/**
* alloc_pages_exact - allocate an exact number physically-contiguous pages.
* @size: the number of bytes to allocate
* @gfp_mask: GFP flags for the allocation
*
* This function is similar to alloc_pages(), except that it allocates the
* minimum number of pages to satisfy the request. alloc_pages() can only
* allocate memory in power-of-two pages.
*
* This function is also limited by MAX_ORDER.
*
* Memory allocated by this function must be released by free_pages_exact().
*/
void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
{
unsigned int order = get_order(size);
unsigned long addr;
addr = __get_free_pages(gfp_mask, order);
return make_alloc_exact(addr, order, size);
}
EXPORT_SYMBOL(alloc_pages_exact);
/**
* alloc_pages_exact_nid - allocate an exact number of physically-contiguous
* pages on a node.
* @nid: the preferred node ID where memory should be allocated
* @size: the number of bytes to allocate
* @gfp_mask: GFP flags for the allocation
*
* Like alloc_pages_exact(), but try to allocate on node nid first before falling
* back.
* Note this is not alloc_pages_exact_node() which allocates on a specific node,
* but is not exact.
*/
void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
{
unsigned order = get_order(size);
struct page *p = alloc_pages_node(nid, gfp_mask, order);
if (!p)
return NULL;
return make_alloc_exact((unsigned long)page_address(p), order, size);
}
EXPORT_SYMBOL(alloc_pages_exact_nid);
/**
* free_pages_exact - release memory allocated via alloc_pages_exact()
* @virt: the value returned by alloc_pages_exact.
* @size: size of allocation, same value as passed to alloc_pages_exact().
*
* Release the memory allocated by a previous call to alloc_pages_exact.
*/
void free_pages_exact(void *virt, size_t size)
{
unsigned long addr = (unsigned long)virt;
unsigned long end = addr + PAGE_ALIGN(size);
while (addr < end) {
free_page(addr);
addr += PAGE_SIZE;
}
}
EXPORT_SYMBOL(free_pages_exact);
static unsigned int nr_free_zone_pages(int offset)
{
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
struct zoneref *z;
struct zone *zone;
/* Just pick one node, since fallback list is circular */
unsigned int sum = 0;
struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
for_each_zone_zonelist(zone, z, zonelist, offset) {
unsigned long size = zone->present_pages;
unsigned long high = high_wmark_pages(zone);
if (size > high)
sum += size - high;
}
return sum;
}
/*
* Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
*/
unsigned int nr_free_buffer_pages(void)
{
return nr_free_zone_pages(gfp_zone(GFP_USER));
}
EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
/*
* Amount of free RAM allocatable within all zones
*/
unsigned int nr_free_pagecache_pages(void)
{
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
}
static inline void show_node(struct zone *zone)
{
if (NUMA_BUILD)
printk("Node %d ", zone_to_nid(zone));
}
void si_meminfo(struct sysinfo *val)
{
val->totalram = totalram_pages;
val->sharedram = 0;
val->freeram = global_page_state(NR_FREE_PAGES);
val->bufferram = nr_blockdev_pages();
val->totalhigh = totalhigh_pages;
val->freehigh = nr_free_highpages();
val->mem_unit = PAGE_SIZE;
}
EXPORT_SYMBOL(si_meminfo);
#ifdef CONFIG_NUMA
void si_meminfo_node(struct sysinfo *val, int nid)
{
pg_data_t *pgdat = NODE_DATA(nid);
val->totalram = pgdat->node_present_pages;
val->freeram = node_page_state(nid, NR_FREE_PAGES);
#ifdef CONFIG_HIGHMEM
val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
NR_FREE_PAGES);
#else
val->totalhigh = 0;
val->freehigh = 0;
#endif
val->mem_unit = PAGE_SIZE;
}
#endif
/*
* Determine whether the node should be displayed or not, depending on whether
* SHOW_MEM_FILTER_NODES was passed to show_free_areas().
*/
bool skip_free_areas_node(unsigned int flags, int nid)
{
bool ret = false;
if (!(flags & SHOW_MEM_FILTER_NODES))
goto out;
get_mems_allowed();
ret = !node_isset(nid, cpuset_current_mems_allowed);
put_mems_allowed();
out:
return ret;
}
#define K(x) ((x) << (PAGE_SHIFT-10))
/*
* Show free area list (used inside shift_scroll-lock stuff)
* We also calculate the percentage fragmentation. We do this by counting the
* memory on each free list with the exception of the first item on the list.
* Suppresses nodes that are not allowed by current's cpuset if
* SHOW_MEM_FILTER_NODES is passed.
*/
void show_free_areas(unsigned int filter)
{
[PATCH] Condense output of show_free_areas() On larger systems, the amount of output dumped on the console when you do SysRq-M is beyond insane. This patch is trying to reduce it somewhat as even with the smaller NUMA systems that have hit the desktop this seems to be a fair thing to do. The philosophy I have taken is as follows: 1) If a zone is empty, don't tell, we don't need yet another line telling us so. The information is available since one can look up the fact how many zones were initialized in the first place. 2) Put as much information on a line is possible, if it can be done in one line, rahter than two, then do it in one. I tried to format the temperature stuff for easy reading. Change show_free_areas() to not print lines for empty zones. If no zone output is printed, the zone is empty. This reduces the number of lines dumped to the console in sysrq on a large system by several thousand lines. Change the zone temperature printouts to use one line per CPU instead of two lines (one hot, one cold). On a 1024 CPU, 1024 node system, this reduces the console output by over a million lines of output. While this is a bigger problem on large NUMA systems, it is also applicable to smaller desktop sized and mid range NUMA systems. Old format: Mem-info: Node 0 DMA per-cpu: cpu 0 hot: high 42, batch 7 used:24 cpu 0 cold: high 14, batch 3 used:1 cpu 1 hot: high 42, batch 7 used:34 cpu 1 cold: high 14, batch 3 used:0 cpu 2 hot: high 42, batch 7 used:0 cpu 2 cold: high 14, batch 3 used:0 cpu 3 hot: high 42, batch 7 used:0 cpu 3 cold: high 14, batch 3 used:0 cpu 4 hot: high 42, batch 7 used:0 cpu 4 cold: high 14, batch 3 used:0 cpu 5 hot: high 42, batch 7 used:0 cpu 5 cold: high 14, batch 3 used:0 cpu 6 hot: high 42, batch 7 used:0 cpu 6 cold: high 14, batch 3 used:0 cpu 7 hot: high 42, batch 7 used:0 cpu 7 cold: high 14, batch 3 used:0 Node 0 DMA32 per-cpu: empty Node 0 Normal per-cpu: empty Node 0 HighMem per-cpu: empty Node 1 DMA per-cpu: [snip] Free pages: 5410688kB (0kB HighMem) Active:9536 inactive:4261 dirty:6 writeback:0 unstable:0 free:338168 slab:1931 mapped:1900 pagetables:208 Node 0 DMA free:1676304kB min:3264kB low:4080kB high:4896kB active:128048kB inactive:61568kB present:1970880kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 DMA32 free:0kB min:0kB low:0kB high:0kB active:0kB inactive:0kB present:0kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 Normal free:0kB min:0kB low:0kB high:0kB active:0kB inactive:0kB present:0kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 HighMem free:0kB min:512kB low:512kB high:512kB active:0kB inactive:0kB present:0kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 1 DMA free:1951728kB min:3280kB low:4096kB high:4912kB active:5632kB inactive:1504kB present:1982464kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 .... New format: Mem-info: Node 0 DMA per-cpu: CPU 0: Hot: hi: 42, btch: 7 usd: 41 Cold: hi: 14, btch: 3 usd: 2 CPU 1: Hot: hi: 42, btch: 7 usd: 40 Cold: hi: 14, btch: 3 usd: 1 CPU 2: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 3: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 4: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 5: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 6: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 7: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 Node 1 DMA per-cpu: [snip] Free pages: 5411088kB (0kB HighMem) Active:9558 inactive:4233 dirty:6 writeback:0 unstable:0 free:338193 slab:1942 mapped:1918 pagetables:208 Node 0 DMA free:1677648kB min:3264kB low:4080kB high:4896kB active:129296kB inactive:58864kB present:1970880kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 1 DMA free:1948448kB min:3280kB low:4096kB high:4912kB active:6864kB inactive:3536kB present:1982464kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Signed-off-by: Jes Sorensen <jes@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:50:05 +00:00
int cpu;
struct zone *zone;
for_each_populated_zone(zone) {
if (skip_free_areas_node(filter, zone_to_nid(zone)))
continue;
[PATCH] Condense output of show_free_areas() On larger systems, the amount of output dumped on the console when you do SysRq-M is beyond insane. This patch is trying to reduce it somewhat as even with the smaller NUMA systems that have hit the desktop this seems to be a fair thing to do. The philosophy I have taken is as follows: 1) If a zone is empty, don't tell, we don't need yet another line telling us so. The information is available since one can look up the fact how many zones were initialized in the first place. 2) Put as much information on a line is possible, if it can be done in one line, rahter than two, then do it in one. I tried to format the temperature stuff for easy reading. Change show_free_areas() to not print lines for empty zones. If no zone output is printed, the zone is empty. This reduces the number of lines dumped to the console in sysrq on a large system by several thousand lines. Change the zone temperature printouts to use one line per CPU instead of two lines (one hot, one cold). On a 1024 CPU, 1024 node system, this reduces the console output by over a million lines of output. While this is a bigger problem on large NUMA systems, it is also applicable to smaller desktop sized and mid range NUMA systems. Old format: Mem-info: Node 0 DMA per-cpu: cpu 0 hot: high 42, batch 7 used:24 cpu 0 cold: high 14, batch 3 used:1 cpu 1 hot: high 42, batch 7 used:34 cpu 1 cold: high 14, batch 3 used:0 cpu 2 hot: high 42, batch 7 used:0 cpu 2 cold: high 14, batch 3 used:0 cpu 3 hot: high 42, batch 7 used:0 cpu 3 cold: high 14, batch 3 used:0 cpu 4 hot: high 42, batch 7 used:0 cpu 4 cold: high 14, batch 3 used:0 cpu 5 hot: high 42, batch 7 used:0 cpu 5 cold: high 14, batch 3 used:0 cpu 6 hot: high 42, batch 7 used:0 cpu 6 cold: high 14, batch 3 used:0 cpu 7 hot: high 42, batch 7 used:0 cpu 7 cold: high 14, batch 3 used:0 Node 0 DMA32 per-cpu: empty Node 0 Normal per-cpu: empty Node 0 HighMem per-cpu: empty Node 1 DMA per-cpu: [snip] Free pages: 5410688kB (0kB HighMem) Active:9536 inactive:4261 dirty:6 writeback:0 unstable:0 free:338168 slab:1931 mapped:1900 pagetables:208 Node 0 DMA free:1676304kB min:3264kB low:4080kB high:4896kB active:128048kB inactive:61568kB present:1970880kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 DMA32 free:0kB min:0kB low:0kB high:0kB active:0kB inactive:0kB present:0kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 Normal free:0kB min:0kB low:0kB high:0kB active:0kB inactive:0kB present:0kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 HighMem free:0kB min:512kB low:512kB high:512kB active:0kB inactive:0kB present:0kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 1 DMA free:1951728kB min:3280kB low:4096kB high:4912kB active:5632kB inactive:1504kB present:1982464kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 .... New format: Mem-info: Node 0 DMA per-cpu: CPU 0: Hot: hi: 42, btch: 7 usd: 41 Cold: hi: 14, btch: 3 usd: 2 CPU 1: Hot: hi: 42, btch: 7 usd: 40 Cold: hi: 14, btch: 3 usd: 1 CPU 2: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 3: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 4: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 5: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 6: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 CPU 7: Hot: hi: 42, btch: 7 usd: 0 Cold: hi: 14, btch: 3 usd: 0 Node 1 DMA per-cpu: [snip] Free pages: 5411088kB (0kB HighMem) Active:9558 inactive:4233 dirty:6 writeback:0 unstable:0 free:338193 slab:1942 mapped:1918 pagetables:208 Node 0 DMA free:1677648kB min:3264kB low:4080kB high:4896kB active:129296kB inactive:58864kB present:1970880kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 1 DMA free:1948448kB min:3280kB low:4096kB high:4912kB active:6864kB inactive:3536kB present:1982464kB pages_scanned:0 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Signed-off-by: Jes Sorensen <jes@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:50:05 +00:00
show_node(zone);
printk("%s per-cpu:\n", zone->name);
for_each_online_cpu(cpu) {
struct per_cpu_pageset *pageset;
pageset = per_cpu_ptr(zone->pageset, cpu);
printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
cpu, pageset->pcp.high,
pageset->pcp.batch, pageset->pcp.count);
}
}
printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
" active_file:%lu inactive_file:%lu isolated_file:%lu\n"
" unevictable:%lu"
" dirty:%lu writeback:%lu unstable:%lu\n"
" free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
" mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:32 +00:00
global_page_state(NR_ACTIVE_ANON),
global_page_state(NR_INACTIVE_ANON),
global_page_state(NR_ISOLATED_ANON),
global_page_state(NR_ACTIVE_FILE),
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:32 +00:00
global_page_state(NR_INACTIVE_FILE),
global_page_state(NR_ISOLATED_FILE),
global_page_state(NR_UNEVICTABLE),
global_page_state(NR_FILE_DIRTY),
global_page_state(NR_WRITEBACK),
global_page_state(NR_UNSTABLE_NFS),
global_page_state(NR_FREE_PAGES),
global_page_state(NR_SLAB_RECLAIMABLE),
global_page_state(NR_SLAB_UNRECLAIMABLE),
global_page_state(NR_FILE_MAPPED),
global_page_state(NR_SHMEM),
global_page_state(NR_PAGETABLE),
global_page_state(NR_BOUNCE));
for_each_populated_zone(zone) {
int i;
if (skip_free_areas_node(filter, zone_to_nid(zone)))
continue;
show_node(zone);
printk("%s"
" free:%lukB"
" min:%lukB"
" low:%lukB"
" high:%lukB"
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:32 +00:00
" active_anon:%lukB"
" inactive_anon:%lukB"
" active_file:%lukB"
" inactive_file:%lukB"
" unevictable:%lukB"
" isolated(anon):%lukB"
" isolated(file):%lukB"
" present:%lukB"
" mlocked:%lukB"
" dirty:%lukB"
" writeback:%lukB"
" mapped:%lukB"
" shmem:%lukB"
" slab_reclaimable:%lukB"
" slab_unreclaimable:%lukB"
" kernel_stack:%lukB"
" pagetables:%lukB"
" unstable:%lukB"
" bounce:%lukB"
" writeback_tmp:%lukB"
" pages_scanned:%lu"
" all_unreclaimable? %s"
"\n",
zone->name,
mm: page allocator: adjust the per-cpu counter threshold when memory is low Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To avoid synchronization overhead, these counters are maintained on a per-cpu basis and drained both periodically and when a threshold is above a threshold. On large CPU systems, the difference between the estimate and real value of NR_FREE_PAGES can be very high. The system can get into a case where pages are allocated far below the min watermark potentially causing livelock issues. The commit solved the problem by taking a better reading of NR_FREE_PAGES when memory was low. Unfortately, as reported by Shaohua Li this accurate reading can consume a large amount of CPU time on systems with many sockets due to cache line bouncing. This patch takes a different approach. For large machines where counter drift might be unsafe and while kswapd is awake, the per-cpu thresholds for the target pgdat are reduced to limit the level of drift to what should be a safe level. This incurs a performance penalty in heavy memory pressure by a factor that depends on the workload and the machine but the machine should function correctly without accidentally exhausting all memory on a node. There is an additional cost when kswapd wakes and sleeps but the event is not expected to be frequent - in Shaohua's test case, there was one recorded sleep and wake event at least. To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is introduced that takes a more accurate reading of NR_FREE_PAGES when called from wakeup_kswapd, when deciding whether it is really safe to go back to sleep in sleeping_prematurely() and when deciding if a zone is really balanced or not in balance_pgdat(). We are still using an expensive function but limiting how often it is called. When the test case is reproduced, the time spent in the watermark functions is reduced. The following report is on the percentage of time spent cumulatively spent in the functions zone_nr_free_pages(), zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(), zone_page_state_snapshot(), zone_page_state(). vanilla 11.6615% disable-threshold 0.2584% David said: : We had to pull aa454840 "mm: page allocator: calculate a better estimate : of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36 : internally because tests showed that it would cause the machine to stall : as the result of heavy kswapd activity. I merged it back with this fix as : it is pending in the -mm tree and it solves the issue we were seeing, so I : definitely think this should be pushed to -stable (and I would seriously : consider it for 2.6.37 inclusion even at this late date). Signed-off-by: Mel Gorman <mel@csn.ul.ie> Reported-by: Shaohua Li <shaohua.li@intel.com> Reviewed-by: Christoph Lameter <cl@linux.com> Tested-by: Nicolas Bareil <nico@chdir.org> Cc: David Rientjes <rientjes@google.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
K(zone_page_state(zone, NR_FREE_PAGES)),
K(min_wmark_pages(zone)),
K(low_wmark_pages(zone)),
K(high_wmark_pages(zone)),
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:32 +00:00
K(zone_page_state(zone, NR_ACTIVE_ANON)),
K(zone_page_state(zone, NR_INACTIVE_ANON)),
K(zone_page_state(zone, NR_ACTIVE_FILE)),
K(zone_page_state(zone, NR_INACTIVE_FILE)),
K(zone_page_state(zone, NR_UNEVICTABLE)),
K(zone_page_state(zone, NR_ISOLATED_ANON)),
K(zone_page_state(zone, NR_ISOLATED_FILE)),
K(zone->present_pages),
K(zone_page_state(zone, NR_MLOCK)),
K(zone_page_state(zone, NR_FILE_DIRTY)),
K(zone_page_state(zone, NR_WRITEBACK)),
K(zone_page_state(zone, NR_FILE_MAPPED)),
K(zone_page_state(zone, NR_SHMEM)),
K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
zone_page_state(zone, NR_KERNEL_STACK) *
THREAD_SIZE / 1024,
K(zone_page_state(zone, NR_PAGETABLE)),
K(zone_page_state(zone, NR_UNSTABLE_NFS)),
K(zone_page_state(zone, NR_BOUNCE)),
K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
zone->pages_scanned,
(zone->all_unreclaimable ? "yes" : "no")
);
printk("lowmem_reserve[]:");
for (i = 0; i < MAX_NR_ZONES; i++)
printk(" %lu", zone->lowmem_reserve[i]);
printk("\n");
}
for_each_populated_zone(zone) {
unsigned long nr[MAX_ORDER], flags, order, total = 0;
if (skip_free_areas_node(filter, zone_to_nid(zone)))
continue;
show_node(zone);
printk("%s: ", zone->name);
spin_lock_irqsave(&zone->lock, flags);
for (order = 0; order < MAX_ORDER; order++) {
nr[order] = zone->free_area[order].nr_free;
total += nr[order] << order;
}
spin_unlock_irqrestore(&zone->lock, flags);
for (order = 0; order < MAX_ORDER; order++)
printk("%lu*%lukB ", nr[order], K(1UL) << order);
printk("= %lukB\n", K(total));
}
printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
show_swap_cache_info();
}
static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
{
zoneref->zone = zone;
zoneref->zone_idx = zone_idx(zone);
}
/*
* Builds allocation fallback zone lists.
*
* Add all populated zones of a node to the zonelist.
*/
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
int nr_zones, enum zone_type zone_type)
{
struct zone *zone;
BUG_ON(zone_type >= MAX_NR_ZONES);
zone_type++;
do {
zone_type--;
zone = pgdat->node_zones + zone_type;
if (populated_zone(zone)) {
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zoneref_set_zone(zone,
&zonelist->_zonerefs[nr_zones++]);
check_highest_zone(zone_type);
}
} while (zone_type);
return nr_zones;
}
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
/*
* zonelist_order:
* 0 = automatic detection of better ordering.
* 1 = order by ([node] distance, -zonetype)
* 2 = order by (-zonetype, [node] distance)
*
* If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
* the same zonelist. So only NUMA can configure this param.
*/
#define ZONELIST_ORDER_DEFAULT 0
#define ZONELIST_ORDER_NODE 1
#define ZONELIST_ORDER_ZONE 2
/* zonelist order in the kernel.
* set_zonelist_order() will set this to NODE or ZONE.
*/
static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
#ifdef CONFIG_NUMA
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
/* The value user specified ....changed by config */
static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
/* string for sysctl */
#define NUMA_ZONELIST_ORDER_LEN 16
char numa_zonelist_order[16] = "default";
/*
* interface for configure zonelist ordering.
* command line option "numa_zonelist_order"
* = "[dD]efault - default, automatic configuration.
* = "[nN]ode - order by node locality, then by zone within node
* = "[zZ]one - order by zone, then by locality within zone
*/
static int __parse_numa_zonelist_order(char *s)
{
if (*s == 'd' || *s == 'D') {
user_zonelist_order = ZONELIST_ORDER_DEFAULT;
} else if (*s == 'n' || *s == 'N') {
user_zonelist_order = ZONELIST_ORDER_NODE;
} else if (*s == 'z' || *s == 'Z') {
user_zonelist_order = ZONELIST_ORDER_ZONE;
} else {
printk(KERN_WARNING
"Ignoring invalid numa_zonelist_order value: "
"%s\n", s);
return -EINVAL;
}
return 0;
}
static __init int setup_numa_zonelist_order(char *s)
{
int ret;
if (!s)
return 0;
ret = __parse_numa_zonelist_order(s);
if (ret == 0)
strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
return ret;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
early_param("numa_zonelist_order", setup_numa_zonelist_order);
/*
* sysctl handler for numa_zonelist_order
*/
int numa_zonelist_order_handler(ctl_table *table, int write,
void __user *buffer, size_t *length,
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
loff_t *ppos)
{
char saved_string[NUMA_ZONELIST_ORDER_LEN];
int ret;
static DEFINE_MUTEX(zl_order_mutex);
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
mutex_lock(&zl_order_mutex);
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
if (write)
strcpy(saved_string, (char*)table->data);
ret = proc_dostring(table, write, buffer, length, ppos);
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
if (ret)
goto out;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
if (write) {
int oldval = user_zonelist_order;
if (__parse_numa_zonelist_order((char*)table->data)) {
/*
* bogus value. restore saved string
*/
strncpy((char*)table->data, saved_string,
NUMA_ZONELIST_ORDER_LEN);
user_zonelist_order = oldval;
} else if (oldval != user_zonelist_order) {
mutex_lock(&zonelists_mutex);
mem-hotplug: avoid multiple zones sharing same boot strapping boot_pageset For each new populated zone of hotadded node, need to update its pagesets with dynamically allocated per_cpu_pageset struct for all possible CPUs: 1) Detach zone->pageset from the shared boot_pageset at end of __build_all_zonelists(). 2) Use mutex to protect zone->pageset when it's still shared in onlined_pages() Otherwises, multiple zones of different nodes would share same boot strapping boot_pageset for same CPU, which will finally cause below kernel panic: ------------[ cut here ]------------ kernel BUG at mm/page_alloc.c:1239! invalid opcode: 0000 [#1] SMP ... Call Trace: [<ffffffff811300c1>] __alloc_pages_nodemask+0x131/0x7b0 [<ffffffff81162e67>] alloc_pages_current+0x87/0xd0 [<ffffffff81128407>] __page_cache_alloc+0x67/0x70 [<ffffffff811325f0>] __do_page_cache_readahead+0x120/0x260 [<ffffffff81132751>] ra_submit+0x21/0x30 [<ffffffff811329c6>] ondemand_readahead+0x166/0x2c0 [<ffffffff81132ba0>] page_cache_async_readahead+0x80/0xa0 [<ffffffff8112a0e4>] generic_file_aio_read+0x364/0x670 [<ffffffff81266cfa>] nfs_file_read+0xca/0x130 [<ffffffff8117b20a>] do_sync_read+0xfa/0x140 [<ffffffff8117bf75>] vfs_read+0xb5/0x1a0 [<ffffffff8117c151>] sys_read+0x51/0x80 [<ffffffff8103c032>] system_call_fastpath+0x16/0x1b RIP [<ffffffff8112ff13>] get_page_from_freelist+0x883/0x900 RSP <ffff88000d1e78a8> ---[ end trace 4bda28328b9990db ] [akpm@linux-foundation.org: merge fix] Signed-off-by: Haicheng Li <haicheng.li@linux.intel.com> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Andi Kleen <andi.kleen@intel.com> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Tejun Heo <tj@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:51 +00:00
build_all_zonelists(NULL);
mutex_unlock(&zonelists_mutex);
}
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
out:
mutex_unlock(&zl_order_mutex);
return ret;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
#define MAX_NODE_LOAD (nr_online_nodes)
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
static int node_load[MAX_NUMNODES];
/**
[PATCH] DocBook: changes and extensions to the kernel documentation I have recompiled Linux kernel 2.6.11.5 documentation for me and our university students again. The documentation could be extended for more sources which are equipped by structured comments for recent 2.6 kernels. I have tried to proceed with that task. I have done that more times from 2.6.0 time and it gets boring to do same changes again and again. Linux kernel compiles after changes for i386 and ARM targets. I have added references to some more files into kernel-api book, I have added some section names as well. So please, check that changes do not break something and that categories are not too much skewed. I have changed kernel-doc to accept "fastcall" and "asmlinkage" words reserved by kernel convention. Most of the other changes are modifications in the comments to make kernel-doc happy, accept some parameters description and do not bail out on errors. Changed <pid> to @pid in the description, moved some #ifdef before comments to correct function to comments bindings, etc. You can see result of the modified documentation build at http://cmp.felk.cvut.cz/~pisa/linux/lkdb-2.6.11.tar.gz Some more sources are ready to be included into kernel-doc generated documentation. Sources has been added into kernel-api for now. Some more section names added and probably some more chaos introduced as result of quick cleanup work. Signed-off-by: Pavel Pisa <pisa@cmp.felk.cvut.cz> Signed-off-by: Martin Waitz <tali@admingilde.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-05-01 15:59:25 +00:00
* find_next_best_node - find the next node that should appear in a given node's fallback list
* @node: node whose fallback list we're appending
* @used_node_mask: nodemask_t of already used nodes
*
* We use a number of factors to determine which is the next node that should
* appear on a given node's fallback list. The node should not have appeared
* already in @node's fallback list, and it should be the next closest node
* according to the distance array (which contains arbitrary distance values
* from each node to each node in the system), and should also prefer nodes
* with no CPUs, since presumably they'll have very little allocation pressure
* on them otherwise.
* It returns -1 if no node is found.
*/
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
static int find_next_best_node(int node, nodemask_t *used_node_mask)
{
int n, val;
int min_val = INT_MAX;
int best_node = -1;
const struct cpumask *tmp = cpumask_of_node(0);
/* Use the local node if we haven't already */
if (!node_isset(node, *used_node_mask)) {
node_set(node, *used_node_mask);
return node;
}
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
for_each_node_state(n, N_HIGH_MEMORY) {
/* Don't want a node to appear more than once */
if (node_isset(n, *used_node_mask))
continue;
/* Use the distance array to find the distance */
val = node_distance(node, n);
/* Penalize nodes under us ("prefer the next node") */
val += (n < node);
/* Give preference to headless and unused nodes */
tmp = cpumask_of_node(n);
if (!cpumask_empty(tmp))
val += PENALTY_FOR_NODE_WITH_CPUS;
/* Slight preference for less loaded node */
val *= (MAX_NODE_LOAD*MAX_NUMNODES);
val += node_load[n];
if (val < min_val) {
min_val = val;
best_node = n;
}
}
if (best_node >= 0)
node_set(best_node, *used_node_mask);
return best_node;
}
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
/*
* Build zonelists ordered by node and zones within node.
* This results in maximum locality--normal zone overflows into local
* DMA zone, if any--but risks exhausting DMA zone.
*/
static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
{
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
int j;
struct zonelist *zonelist;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
zonelist = &pgdat->node_zonelists[0];
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
;
j = build_zonelists_node(NODE_DATA(node), zonelist, j,
MAX_NR_ZONES - 1);
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zonelist->_zonerefs[j].zone = NULL;
zonelist->_zonerefs[j].zone_idx = 0;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
/*
* Build gfp_thisnode zonelists
*/
static void build_thisnode_zonelists(pg_data_t *pgdat)
{
int j;
struct zonelist *zonelist;
zonelist = &pgdat->node_zonelists[1];
j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zonelist->_zonerefs[j].zone = NULL;
zonelist->_zonerefs[j].zone_idx = 0;
}
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
/*
* Build zonelists ordered by zone and nodes within zones.
* This results in conserving DMA zone[s] until all Normal memory is
* exhausted, but results in overflowing to remote node while memory
* may still exist in local DMA zone.
*/
static int node_order[MAX_NUMNODES];
static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
{
int pos, j, node;
int zone_type; /* needs to be signed */
struct zone *z;
struct zonelist *zonelist;
zonelist = &pgdat->node_zonelists[0];
pos = 0;
for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
for (j = 0; j < nr_nodes; j++) {
node = node_order[j];
z = &NODE_DATA(node)->node_zones[zone_type];
if (populated_zone(z)) {
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zoneref_set_zone(z,
&zonelist->_zonerefs[pos++]);
check_highest_zone(zone_type);
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
}
}
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zonelist->_zonerefs[pos].zone = NULL;
zonelist->_zonerefs[pos].zone_idx = 0;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
static int default_zonelist_order(void)
{
int nid, zone_type;
unsigned long low_kmem_size,total_size;
struct zone *z;
int average_size;
/*
* ZONE_DMA and ZONE_DMA32 can be very small area in the system.
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
* If they are really small and used heavily, the system can fall
* into OOM very easily.
* This function detect ZONE_DMA/DMA32 size and configures zone order.
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
*/
/* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
low_kmem_size = 0;
total_size = 0;
for_each_online_node(nid) {
for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
z = &NODE_DATA(nid)->node_zones[zone_type];
if (populated_zone(z)) {
if (zone_type < ZONE_NORMAL)
low_kmem_size += z->present_pages;
total_size += z->present_pages;
} else if (zone_type == ZONE_NORMAL) {
/*
* If any node has only lowmem, then node order
* is preferred to allow kernel allocations
* locally; otherwise, they can easily infringe
* on other nodes when there is an abundance of
* lowmem available to allocate from.
*/
return ZONELIST_ORDER_NODE;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
}
}
}
if (!low_kmem_size || /* there are no DMA area. */
low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
return ZONELIST_ORDER_NODE;
/*
* look into each node's config.
* If there is a node whose DMA/DMA32 memory is very big area on
* local memory, NODE_ORDER may be suitable.
*/
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
average_size = total_size /
(nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
for_each_online_node(nid) {
low_kmem_size = 0;
total_size = 0;
for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
z = &NODE_DATA(nid)->node_zones[zone_type];
if (populated_zone(z)) {
if (zone_type < ZONE_NORMAL)
low_kmem_size += z->present_pages;
total_size += z->present_pages;
}
}
if (low_kmem_size &&
total_size > average_size && /* ignore small node */
low_kmem_size > total_size * 70/100)
return ZONELIST_ORDER_NODE;
}
return ZONELIST_ORDER_ZONE;
}
static void set_zonelist_order(void)
{
if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
current_zonelist_order = default_zonelist_order();
else
current_zonelist_order = user_zonelist_order;
}
static void build_zonelists(pg_data_t *pgdat)
{
int j, node, load;
enum zone_type i;
nodemask_t used_mask;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
int local_node, prev_node;
struct zonelist *zonelist;
int order = current_zonelist_order;
/* initialize zonelists */
for (i = 0; i < MAX_ZONELISTS; i++) {
zonelist = pgdat->node_zonelists + i;
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zonelist->_zonerefs[0].zone = NULL;
zonelist->_zonerefs[0].zone_idx = 0;
}
/* NUMA-aware ordering of nodes */
local_node = pgdat->node_id;
load = nr_online_nodes;
prev_node = local_node;
nodes_clear(used_mask);
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
memset(node_order, 0, sizeof(node_order));
j = 0;
while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
int distance = node_distance(local_node, node);
/*
* If another node is sufficiently far away then it is better
* to reclaim pages in a zone before going off node.
*/
if (distance > RECLAIM_DISTANCE)
zone_reclaim_mode = 1;
/*
* We don't want to pressure a particular node.
* So adding penalty to the first node in same
* distance group to make it round-robin.
*/
if (distance != node_distance(local_node, prev_node))
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
node_load[node] = load;
prev_node = node;
load--;
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
if (order == ZONELIST_ORDER_NODE)
build_zonelists_in_node_order(pgdat, node);
else
node_order[j++] = node; /* remember order */
}
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
if (order == ZONELIST_ORDER_ZONE) {
/* calculate node order -- i.e., DMA last! */
build_zonelists_in_zone_order(pgdat, j);
}
build_thisnode_zonelists(pgdat);
}
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
/* Construct the zonelist performance cache - see further mmzone.h */
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
static void build_zonelist_cache(pg_data_t *pgdat)
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
{
struct zonelist *zonelist;
struct zonelist_cache *zlc;
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
struct zoneref *z;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
zonelist = &pgdat->node_zonelists[0];
zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
for (z = zonelist->_zonerefs; z->zone; z++)
zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
}
numa: introduce numa_mem_id()- effective local memory node id Introduce numa_mem_id(), based on generic percpu variable infrastructure to track "nearest node with memory" for archs that support memoryless nodes. Define API in <linux/topology.h> when CONFIG_HAVE_MEMORYLESS_NODES defined, else stubs. Architectures will define HAVE_MEMORYLESS_NODES if/when they support them. Archs can override definitions of: numa_mem_id() - returns node number of "local memory" node set_numa_mem() - initialize [this cpus'] per cpu variable 'numa_mem' cpu_to_mem() - return numa_mem for specified cpu; may be used as lvalue Generic initialization of 'numa_mem' occurs in __build_all_zonelists(). This will initialize the boot cpu at boot time, and all cpus on change of numa_zonelist_order, or when node or memory hot-plug requires zonelist rebuild. Archs that support memoryless nodes will need to initialize 'numa_mem' for secondary cpus as they're brought on-line. [akpm@linux-foundation.org: fix build] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Christoph Lameter <cl@linux-foundation.org> Cc: Tejun Heo <tj@kernel.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Nick Piggin <npiggin@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Eric Whitney <eric.whitney@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-26 21:45:00 +00:00
#ifdef CONFIG_HAVE_MEMORYLESS_NODES
/*
* Return node id of node used for "local" allocations.
* I.e., first node id of first zone in arg node's generic zonelist.
* Used for initializing percpu 'numa_mem', which is used primarily
* for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
*/
int local_memory_node(int node)
{
struct zone *zone;
(void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
gfp_zone(GFP_KERNEL),
NULL,
&zone);
return zone->node;
}
#endif
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
#else /* CONFIG_NUMA */
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
static void set_zonelist_order(void)
{
current_zonelist_order = ZONELIST_ORDER_ZONE;
}
static void build_zonelists(pg_data_t *pgdat)
{
int node, local_node;
enum zone_type j;
struct zonelist *zonelist;
local_node = pgdat->node_id;
zonelist = &pgdat->node_zonelists[0];
j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
/*
* Now we build the zonelist so that it contains the zones
* of all the other nodes.
* We don't want to pressure a particular node, so when
* building the zones for node N, we make sure that the
* zones coming right after the local ones are those from
* node N+1 (modulo N)
*/
for (node = local_node + 1; node < MAX_NUMNODES; node++) {
if (!node_online(node))
continue;
j = build_zonelists_node(NODE_DATA(node), zonelist, j,
MAX_NR_ZONES - 1);
}
for (node = 0; node < local_node; node++) {
if (!node_online(node))
continue;
j = build_zonelists_node(NODE_DATA(node), zonelist, j,
MAX_NR_ZONES - 1);
}
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:12:17 +00:00
zonelist->_zonerefs[j].zone = NULL;
zonelist->_zonerefs[j].zone_idx = 0;
}
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
static void build_zonelist_cache(pg_data_t *pgdat)
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
{
pgdat->node_zonelists[0].zlcache_ptr = NULL;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
}
#endif /* CONFIG_NUMA */
/*
* Boot pageset table. One per cpu which is going to be used for all
* zones and all nodes. The parameters will be set in such a way
* that an item put on a list will immediately be handed over to
* the buddy list. This is safe since pageset manipulation is done
* with interrupts disabled.
*
* The boot_pagesets must be kept even after bootup is complete for
* unused processors and/or zones. They do play a role for bootstrapping
* hotplugged processors.
*
* zoneinfo_show() and maybe other functions do
* not check if the processor is online before following the pageset pointer.
* Other parts of the kernel may not check if the zone is available.
*/
static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
mem-hotplug: avoid multiple zones sharing same boot strapping boot_pageset For each new populated zone of hotadded node, need to update its pagesets with dynamically allocated per_cpu_pageset struct for all possible CPUs: 1) Detach zone->pageset from the shared boot_pageset at end of __build_all_zonelists(). 2) Use mutex to protect zone->pageset when it's still shared in onlined_pages() Otherwises, multiple zones of different nodes would share same boot strapping boot_pageset for same CPU, which will finally cause below kernel panic: ------------[ cut here ]------------ kernel BUG at mm/page_alloc.c:1239! invalid opcode: 0000 [#1] SMP ... Call Trace: [<ffffffff811300c1>] __alloc_pages_nodemask+0x131/0x7b0 [<ffffffff81162e67>] alloc_pages_current+0x87/0xd0 [<ffffffff81128407>] __page_cache_alloc+0x67/0x70 [<ffffffff811325f0>] __do_page_cache_readahead+0x120/0x260 [<ffffffff81132751>] ra_submit+0x21/0x30 [<ffffffff811329c6>] ondemand_readahead+0x166/0x2c0 [<ffffffff81132ba0>] page_cache_async_readahead+0x80/0xa0 [<ffffffff8112a0e4>] generic_file_aio_read+0x364/0x670 [<ffffffff81266cfa>] nfs_file_read+0xca/0x130 [<ffffffff8117b20a>] do_sync_read+0xfa/0x140 [<ffffffff8117bf75>] vfs_read+0xb5/0x1a0 [<ffffffff8117c151>] sys_read+0x51/0x80 [<ffffffff8103c032>] system_call_fastpath+0x16/0x1b RIP [<ffffffff8112ff13>] get_page_from_freelist+0x883/0x900 RSP <ffff88000d1e78a8> ---[ end trace 4bda28328b9990db ] [akpm@linux-foundation.org: merge fix] Signed-off-by: Haicheng Li <haicheng.li@linux.intel.com> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Andi Kleen <andi.kleen@intel.com> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Tejun Heo <tj@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:51 +00:00
static void setup_zone_pageset(struct zone *zone);
/*
* Global mutex to protect against size modification of zonelists
* as well as to serialize pageset setup for the new populated zone.
*/
DEFINE_MUTEX(zonelists_mutex);
/* return values int ....just for stop_machine() */
mem-hotplug: avoid multiple zones sharing same boot strapping boot_pageset For each new populated zone of hotadded node, need to update its pagesets with dynamically allocated per_cpu_pageset struct for all possible CPUs: 1) Detach zone->pageset from the shared boot_pageset at end of __build_all_zonelists(). 2) Use mutex to protect zone->pageset when it's still shared in onlined_pages() Otherwises, multiple zones of different nodes would share same boot strapping boot_pageset for same CPU, which will finally cause below kernel panic: ------------[ cut here ]------------ kernel BUG at mm/page_alloc.c:1239! invalid opcode: 0000 [#1] SMP ... Call Trace: [<ffffffff811300c1>] __alloc_pages_nodemask+0x131/0x7b0 [<ffffffff81162e67>] alloc_pages_current+0x87/0xd0 [<ffffffff81128407>] __page_cache_alloc+0x67/0x70 [<ffffffff811325f0>] __do_page_cache_readahead+0x120/0x260 [<ffffffff81132751>] ra_submit+0x21/0x30 [<ffffffff811329c6>] ondemand_readahead+0x166/0x2c0 [<ffffffff81132ba0>] page_cache_async_readahead+0x80/0xa0 [<ffffffff8112a0e4>] generic_file_aio_read+0x364/0x670 [<ffffffff81266cfa>] nfs_file_read+0xca/0x130 [<ffffffff8117b20a>] do_sync_read+0xfa/0x140 [<ffffffff8117bf75>] vfs_read+0xb5/0x1a0 [<ffffffff8117c151>] sys_read+0x51/0x80 [<ffffffff8103c032>] system_call_fastpath+0x16/0x1b RIP [<ffffffff8112ff13>] get_page_from_freelist+0x883/0x900 RSP <ffff88000d1e78a8> ---[ end trace 4bda28328b9990db ] [akpm@linux-foundation.org: merge fix] Signed-off-by: Haicheng Li <haicheng.li@linux.intel.com> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com> Reviewed-by: Andi Kleen <andi.kleen@intel.com> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Tejun Heo <tj@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:51 +00:00
static __init_refok int __build_all_zonelists(void *data)
{
int nid;
int cpu;
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
#ifdef CONFIG_NUMA
memset(node_load, 0, sizeof(node_load));
#endif
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
for_each_online_node(nid) {
pg_data_t *pgdat = NODE_DATA(nid);
build_zonelists(pgdat);
build_zonelist_cache(pgdat);
[PATCH] memory page_alloc zonelist caching speedup Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them. Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.) Recent changes: This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full. Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.) See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and (2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below. I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist(). There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone: 1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though. The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof. 2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging. The following approach has no new dependency on node distances or zone sorting. See comment in the patch below for a description of what it actually does. Technical details of note (or controversy): - See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking. - Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND". - Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay. - This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines. - If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted. - I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim. - I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node. =============== Performance - some benchmark results and analysis: This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can. Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below. Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel. number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent GBs N ------------ -------------- ---------------- systime mem threads sys user real sys user real sys user real better 12 1 153 24 177 151 24 176 -2 0 -1 1% 12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5% Notes: 1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread. 2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds. 3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere. 4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM. 5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not. 6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload. 7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch. Conclusions: * This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations. * For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent. * For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.) Signed-off-by: Paul Jackson <pj@sgi.com> Cc: Rohit Seth <rohitseth@google.com> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: David Rientjes <rientjes@cs.washington.edu> Cc: Paul Menage <menage@google.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:31:48 +00:00
}
/*
* Initialize the boot_pagesets that are going to be used
* for bootstrapping processors. The real pagesets for
* each zone will be allocated later when the per cpu
* allocator is available.
*
* boot_pagesets are used also for bootstrapping offline
* cpus if the system is already booted because the pagesets
* are needed to initialize allocators on a specific cpu too.
* F.e. the percpu allocator needs the page allocator which
* needs the percpu allocator in order to allocate its pagesets
* (a chicken-egg dilemma).
*/
numa: introduce numa_mem_id()- effective local memory node id Introduce numa_mem_id(), based on generic percpu variable infrastructure to track "nearest node with memory" for archs that support memoryless nodes. Define API in <linux/topology.h> when CONFIG_HAVE_MEMORYLESS_NODES defined, else stubs. Architectures will define HAVE_MEMORYLESS_NODES if/when they support them. Archs can override definitions of: numa_mem_id() - returns node number of "local memory" node set_numa_mem() - initialize [this cpus'] per cpu variable 'numa_mem' cpu_to_mem() - return numa_mem for specified cpu; may be used as lvalue Generic initialization of 'numa_mem' occurs in __build_all_zonelists(). This will initialize the boot cpu at boot time, and all cpus on change of numa_zonelist_order, or when node or memory hot-plug requires zonelist rebuild. Archs that support memoryless nodes will need to initialize 'numa_mem' for secondary cpus as they're brought on-line. [akpm@linux-foundation.org: fix build] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Christoph Lameter <cl@linux-foundation.org> Cc: Tejun Heo <tj@kernel.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Nick Piggin <npiggin@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Eric Whitney <eric.whitney@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-26 21:45:00 +00:00
for_each_possible_cpu(cpu) {
setup_pageset(&per_cpu(boot_pageset, cpu), 0);
numa: introduce numa_mem_id()- effective local memory node id Introduce numa_mem_id(), based on generic percpu variable infrastructure to track "nearest node with memory" for archs that support memoryless nodes. Define API in <linux/topology.h> when CONFIG_HAVE_MEMORYLESS_NODES defined, else stubs. Architectures will define HAVE_MEMORYLESS_NODES if/when they support them. Archs can override definitions of: numa_mem_id() - returns node number of "local memory" node set_numa_mem() - initialize [this cpus'] per cpu variable 'numa_mem' cpu_to_mem() - return numa_mem for specified cpu; may be used as lvalue Generic initialization of 'numa_mem' occurs in __build_all_zonelists(). This will initialize the boot cpu at boot time, and all cpus on change of numa_zonelist_order, or when node or memory hot-plug requires zonelist rebuild. Archs that support memoryless nodes will need to initialize 'numa_mem' for secondary cpus as they're brought on-line. [akpm@linux-foundation.org: fix build] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Christoph Lameter <cl@linux-foundation.org> Cc: Tejun Heo <tj@kernel.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Nick Piggin <npiggin@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Eric Whitney <eric.whitney@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-26 21:45:00 +00:00
#ifdef CONFIG_HAVE_MEMORYLESS_NODES
/*
* We now know the "local memory node" for each node--
* i.e., the node of the first zone in the generic zonelist.
* Set up numa_mem percpu variable for on-line cpus. During
* boot, only the boot cpu should be on-line; we'll init the
* secondary cpus' numa_mem as they come on-line. During
* node/memory hotplug, we'll fixup all on-line cpus.
*/
if (cpu_online(cpu))
set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
#endif
}
return 0;
}
/*
* Called with zonelists_mutex held always
* unless system_state == SYSTEM_BOOTING.
*/
void __ref build_all_zonelists(void *data)
{
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
set_zonelist_order();
if (system_state == SYSTEM_BOOTING) {
__build_all_zonelists(NULL);
mminit_verify_zonelist();
cpuset_init_current_mems_allowed();
} else {
/* we have to stop all cpus to guarantee there is no user
of zonelist */
mm/page_alloc.c: fix build_all_zonelist() where percpu_alloc() is wrongly called under stop_machine_run() During memory hotplug, build_allzonelists() may be called under stop_machine_run(). In this function, setup_zone_pageset() is called. But it's bug because it will do page allocation under stop_machine_run(). Here is a report from Alok Kataria. BUG: sleeping function called from invalid context at kernel/mutex.c:94 in_atomic(): 0, irqs_disabled(): 1, pid: 4, name: migration/0 Pid: 4, comm: migration/0 Not tainted 2.6.35.6-45.fc14.x86_64 #1 Call Trace: [<ffffffff8103d12b>] __might_sleep+0xeb/0xf0 [<ffffffff81468245>] mutex_lock+0x24/0x50 [<ffffffff8110eaa6>] pcpu_alloc+0x6d/0x7ee [<ffffffff81048888>] ? load_balance+0xbe/0x60e [<ffffffff8103a1b3>] ? rt_se_boosted+0x21/0x2f [<ffffffff8103e1cf>] ? dequeue_rt_stack+0x18b/0x1ed [<ffffffff8110f237>] __alloc_percpu+0x10/0x12 [<ffffffff81465e22>] setup_zone_pageset+0x38/0xbe [<ffffffff810d6d81>] ? build_zonelists_node.clone.58+0x79/0x8c [<ffffffff81452539>] __build_all_zonelists+0x419/0x46c [<ffffffff8108ef01>] ? cpu_stopper_thread+0xb2/0x198 [<ffffffff8108f075>] stop_machine_cpu_stop+0x8e/0xc5 [<ffffffff8108efe7>] ? stop_machine_cpu_stop+0x0/0xc5 [<ffffffff8108ef57>] cpu_stopper_thread+0x108/0x198 [<ffffffff81467a37>] ? schedule+0x5b2/0x5cc [<ffffffff8108ee4f>] ? cpu_stopper_thread+0x0/0x198 [<ffffffff81065f29>] kthread+0x7f/0x87 [<ffffffff8100aae4>] kernel_thread_helper+0x4/0x10 [<ffffffff81065eaa>] ? kthread+0x0/0x87 [<ffffffff8100aae0>] ? kernel_thread_helper+0x0/0x10 Built 5 zonelists in Node order, mobility grouping on. Total pages: 289456 Policy zone: Normal This patch tries to fix the issue by moving setup_zone_pageset() out from stop_machine_run(). It's obviously not necessary to be called under stop_machine_run(). [akpm@linux-foundation.org: remove unneeded local] Reported-by: Alok Kataria <akataria@vmware.com> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Tejun Heo <tj@kernel.org> Cc: Petr Vandrovec <petr@vmware.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Reviewed-by: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-11-24 20:57:09 +00:00
#ifdef CONFIG_MEMORY_HOTPLUG
if (data)
setup_zone_pageset((struct zone *)data);
#endif
stop_machine(__build_all_zonelists, NULL, NULL);
/* cpuset refresh routine should be here */
}
vm_total_pages = nr_free_pagecache_pages();
/*
* Disable grouping by mobility if the number of pages in the
* system is too low to allow the mechanism to work. It would be
* more accurate, but expensive to check per-zone. This check is
* made on memory-hotadd so a system can start with mobility
* disabled and enable it later
*/
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
page_group_by_mobility_disabled = 1;
else
page_group_by_mobility_disabled = 0;
printk("Built %i zonelists in %s order, mobility grouping %s. "
"Total pages: %ld\n",
nr_online_nodes,
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
zonelist_order_name[current_zonelist_order],
page_group_by_mobility_disabled ? "off" : "on",
change zonelist order: zonelist order selection logic Make zonelist creation policy selectable from sysctl/boot option v6. This patch makes NUMA's zonelist (of pgdat) order selectable. Available order are Default(automatic)/ Node-based / Zone-based. [Default Order] The kernel selects Node-based or Zone-based order automatically. [Node-based Order] This policy treats the locality of memory as the most important parameter. Zonelist order is created by each zone's locality. This means lower zones (ex. ZONE_DMA) can be used before higher zone (ex. ZONE_NORMAL) exhausion. IOW. ZONE_DMA will be in the middle of zonelist. current 2.6.21 kernel uses this. Pros. * A user can expect local memory as much as possible. Cons. * lower zone will be exhansted before higher zone. This may cause OOM_KILL. Maybe suitable if ZONE_DMA is relatively big and you never see OOM_KILL because of ZONE_DMA exhaution and you need the best locality. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(0)'s DMA -> node(1)'s NORMAL. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. [Zone-based order] This policy treats the zone type as the most important parameter. Zonelist order is created by zone-type order. This means lower zone never be used bofere higher zone exhaustion. IOW. ZONE_DMA will be always at the tail of zonelist. Pros. * OOM_KILL(bacause of lower zone) occurs only if the whole zones are exhausted. Cons. * memory locality may not be best. (example) assume 2 node NUMA. node(0) has ZONE_DMA/ZONE_NORMAL, node(1) has ZONE_NORMAL. *node(0)'s memory allocation order: node(0)'s NORMAL -> node(1)'s NORMAL -> node(0)'s DMA. *node(1)'s memory allocation order: node(1)'s NORMAL -> node(0)'s NORMAL -> node(0)'s DMA. bootoption "numa_zonelist_order=" and proc/sysctl is supporetd. command: %echo N > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Node-based order. command: %echo Z > /proc/sys/vm/numa_zonelist_order Will rebuild zonelist in Zone-based order. Thanks to Lee Schermerhorn, he gives me much help and codes. [Lee.Schermerhorn@hp.com: add check_highest_zone to build_zonelists_in_zone_order] [akpm@linux-foundation.org: build fix] Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "jesse.barnes@intel.com" <jesse.barnes@intel.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-16 06:38:01 +00:00
vm_total_pages);
#ifdef CONFIG_NUMA
printk("Policy zone: %s\n", zone_names[policy_zone]);
#endif
}
/*
* Helper functions to size the waitqueue hash table.
* Essentially these want to choose hash table sizes sufficiently
* large so that collisions trying to wait on pages are rare.
* But in fact, the number of active page waitqueues on typical
* systems is ridiculously low, less than 200. So this is even
* conservative, even though it seems large.
*
* The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
* waitqueues, i.e. the size of the waitq table given the number of pages.
*/
#define PAGES_PER_WAITQUEUE 256
#ifndef CONFIG_MEMORY_HOTPLUG
static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
{
unsigned long size = 1;
pages /= PAGES_PER_WAITQUEUE;
while (size < pages)
size <<= 1;
/*
* Once we have dozens or even hundreds of threads sleeping
* on IO we've got bigger problems than wait queue collision.
* Limit the size of the wait table to a reasonable size.
*/
size = min(size, 4096UL);
return max(size, 4UL);
}
#else
/*
* A zone's size might be changed by hot-add, so it is not possible to determine
* a suitable size for its wait_table. So we use the maximum size now.
*
* The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
*
* i386 (preemption config) : 4096 x 16 = 64Kbyte.
* ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
* ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
*
* The maximum entries are prepared when a zone's memory is (512K + 256) pages
* or more by the traditional way. (See above). It equals:
*
* i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
* ia64(16K page size) : = ( 8G + 4M)byte.
* powerpc (64K page size) : = (32G +16M)byte.
*/
static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
{
return 4096UL;
}
#endif
/*
* This is an integer logarithm so that shifts can be used later
* to extract the more random high bits from the multiplicative
* hash function before the remainder is taken.
*/
static inline unsigned long wait_table_bits(unsigned long size)
{
return ffz(~size);
}
#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
/*
* Check if a pageblock contains reserved pages
*/
static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
{
unsigned long pfn;
for (pfn = start_pfn; pfn < end_pfn; pfn++) {
if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
return 1;
}
return 0;
}
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
/*
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
* Mark a number of pageblocks as MIGRATE_RESERVE. The number
* of blocks reserved is based on min_wmark_pages(zone). The memory within
* the reserve will tend to store contiguous free pages. Setting min_free_kbytes
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
* higher will lead to a bigger reserve which will get freed as contiguous
* blocks as reclaim kicks in
*/
static void setup_zone_migrate_reserve(struct zone *zone)
{
unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
struct page *page;
page-allocator: limit the number of MIGRATE_RESERVE pageblocks per zone After anti-fragmentation was merged, a bug was reported whereby devices that depended on high-order atomic allocations were failing. The solution was to preserve a property in the buddy allocator which tended to keep the minimum number of free pages in the zone at the lower physical addresses and contiguous. To preserve this property, MIGRATE_RESERVE was introduced and a number of pageblocks at the start of a zone would be marked "reserve", the number of which depended on min_free_kbytes. Anti-fragmentation works by avoiding the mixing of page migratetypes within the same pageblock. One way of helping this is to increase min_free_kbytes because it becomes less like that it will be necessary to place pages of of MIGRATE_RESERVE is unbounded, the free memory is kept there in large contiguous blocks instead of helping anti-fragmentation as much as it should. With the page-allocator tracepoint patches applied, it was found during anti-fragmentation tests that the number of fragmentation-related events were far higher than expected even with min_free_kbytes at higher values. This patch limits the number of MIGRATE_RESERVE blocks that exist per zone to two. For example, with a sufficient min_free_kbytes, 4MB of memory will be kept aside on an x86-64 and remain more or less free and contiguous for the systems uptime. This should be sufficient for devices depending on high-order atomic allocations while helping fragmentation control when min_free_kbytes is tuned appropriately. As side-effect of this patch is that the reserve variable is converted to int as unsigned long was the wrong type to use when ensuring that only the required number of reserve blocks are created. With the patches applied, fragmentation-related events as measured by the page allocator tracepoints were significantly reduced when running some fragmentation stress-tests on systems with min_free_kbytes tuned to a value appropriate for hugepage allocations at runtime. On x86, the events recorded were reduced by 99.8%, on x86-64 by 99.72% and on ppc64 by 99.83%. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:02 +00:00
unsigned long block_migratetype;
int reserve;
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
/* Get the start pfn, end pfn and the number of blocks to reserve */
start_pfn = zone->zone_start_pfn;
end_pfn = start_pfn + zone->spanned_pages;
reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
pageblock_order;
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
page-allocator: limit the number of MIGRATE_RESERVE pageblocks per zone After anti-fragmentation was merged, a bug was reported whereby devices that depended on high-order atomic allocations were failing. The solution was to preserve a property in the buddy allocator which tended to keep the minimum number of free pages in the zone at the lower physical addresses and contiguous. To preserve this property, MIGRATE_RESERVE was introduced and a number of pageblocks at the start of a zone would be marked "reserve", the number of which depended on min_free_kbytes. Anti-fragmentation works by avoiding the mixing of page migratetypes within the same pageblock. One way of helping this is to increase min_free_kbytes because it becomes less like that it will be necessary to place pages of of MIGRATE_RESERVE is unbounded, the free memory is kept there in large contiguous blocks instead of helping anti-fragmentation as much as it should. With the page-allocator tracepoint patches applied, it was found during anti-fragmentation tests that the number of fragmentation-related events were far higher than expected even with min_free_kbytes at higher values. This patch limits the number of MIGRATE_RESERVE blocks that exist per zone to two. For example, with a sufficient min_free_kbytes, 4MB of memory will be kept aside on an x86-64 and remain more or less free and contiguous for the systems uptime. This should be sufficient for devices depending on high-order atomic allocations while helping fragmentation control when min_free_kbytes is tuned appropriately. As side-effect of this patch is that the reserve variable is converted to int as unsigned long was the wrong type to use when ensuring that only the required number of reserve blocks are created. With the patches applied, fragmentation-related events as measured by the page allocator tracepoints were significantly reduced when running some fragmentation stress-tests on systems with min_free_kbytes tuned to a value appropriate for hugepage allocations at runtime. On x86, the events recorded were reduced by 99.8%, on x86-64 by 99.72% and on ppc64 by 99.83%. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:02 +00:00
/*
* Reserve blocks are generally in place to help high-order atomic
* allocations that are short-lived. A min_free_kbytes value that
* would result in more than 2 reserve blocks for atomic allocations
* is assumed to be in place to help anti-fragmentation for the
* future allocation of hugepages at runtime.
*/
reserve = min(2, reserve);
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
if (!pfn_valid(pfn))
continue;
page = pfn_to_page(pfn);
/* Watch out for overlapping nodes */
if (page_to_nid(page) != zone_to_nid(zone))
continue;
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
/* Blocks with reserved pages will never free, skip them. */
block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
if (pageblock_is_reserved(pfn, block_end_pfn))
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
continue;
block_migratetype = get_pageblock_migratetype(page);
/* If this block is reserved, account for it */
if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
reserve--;
continue;
}
/* Suitable for reserving if this block is movable */
if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
set_pageblock_migratetype(page, MIGRATE_RESERVE);
move_freepages_block(zone, page, MIGRATE_RESERVE);
reserve--;
continue;
}
/*
* If the reserve is met and this is a previous reserved block,
* take it back
*/
if (block_migratetype == MIGRATE_RESERVE) {
set_pageblock_migratetype(page, MIGRATE_MOVABLE);
move_freepages_block(zone, page, MIGRATE_MOVABLE);
}
}
}
/*
* Initially all pages are reserved - free ones are freed
* up by free_all_bootmem() once the early boot process is
* done. Non-atomic initialization, single-pass.
*/
void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
unsigned long start_pfn, enum memmap_context context)
{
struct page *page;
[PATCH] sparsemem hotplug base Make sparse's initalization be accessible at runtime. This allows sparse mappings to be created after boot in a hotplug situation. This patch is separated from the previous one just to give an indication how much of the sparse infrastructure is *just* for hotplug memory. The section_mem_map doesn't really store a pointer. It stores something that is convenient to do some math against to get a pointer. It isn't valid to just do *section_mem_map, so I don't think it should be stored as a pointer. There are a couple of things I'd like to store about a section. First of all, the fact that it is !NULL does not mean that it is present. There could be such a combination where section_mem_map *is* NULL, but the math gets you properly to a real mem_map. So, I don't think that check is safe. Since we're storing 32-bit-aligned structures, we have a few bits in the bottom of the pointer to play with. Use one bit to encode whether there's really a mem_map there, and the other one to tell whether there's a valid section there. We need to distinguish between the two because sometimes there's a gap between when a section is discovered to be present and when we can get the mem_map for it. Signed-off-by: Dave Hansen <haveblue@us.ibm.com> Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Jack Steiner <steiner@sgi.com> Signed-off-by: Bob Picco <bob.picco@hp.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:00 +00:00
unsigned long end_pfn = start_pfn + size;
unsigned long pfn;
struct zone *z;
if (highest_memmap_pfn < end_pfn - 1)
highest_memmap_pfn = end_pfn - 1;
z = &NODE_DATA(nid)->node_zones[zone];
for (pfn = start_pfn; pfn < end_pfn; pfn++) {
/*
* There can be holes in boot-time mem_map[]s
* handed to this function. They do not
* exist on hotplugged memory.
*/
if (context == MEMMAP_EARLY) {
if (!early_pfn_valid(pfn))
continue;
if (!early_pfn_in_nid(pfn, nid))
continue;
}
[PATCH] sparsemem memory model Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of mem_map[] is needed by discontiguous memory machines (like in the old CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually become a complete replacement. A significant advantage over DISCONTIGMEM is that it's completely separated from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA and DISCONTIG are often confused. Another advantage is that sparse doesn't require each NUMA node's ranges to be contiguous. It can handle overlapping ranges between nodes with no problems, where DISCONTIGMEM currently throws away that memory. Sparsemem uses an array to provide different pfn_to_page() translations for each SECTION_SIZE area of physical memory. This is what allows the mem_map[] to be chopped up. In order to do quick pfn_to_page() operations, the section number of the page is encoded in page->flags. Part of the sparsemem infrastructure enables sharing of these bits more dynamically (at compile-time) between the page_zone() and sparsemem operations. However, on 32-bit architectures, the number of bits is quite limited, and may require growing the size of the page->flags type in certain conditions. Several things might force this to occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of memory), an increase in the physical address space, or an increase in the number of used page->flags. One thing to note is that, once sparsemem is present, the NUMA node information no longer needs to be stored in the page->flags. It might provide speed increases on certain platforms and will be stored there if there is room. But, if out of room, an alternate (theoretically slower) mechanism is used. This patch introduces CONFIG_FLATMEM. It is used in almost all cases where there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM often have to compile out the same areas of code. Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Dave Hansen <haveblue@us.ibm.com> Signed-off-by: Martin Bligh <mbligh@aracnet.com> Signed-off-by: Adrian Bunk <bunk@stusta.de> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Bob Picco <bob.picco@hp.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:07:54 +00:00
page = pfn_to_page(pfn);
set_page_links(page, zone, nid, pfn);
mminit_verify_page_links(page, zone, nid, pfn);
init_page_count(page);
reset_page_mapcount(page);
SetPageReserved(page);
/*
* Mark the block movable so that blocks are reserved for
* movable at startup. This will force kernel allocations
* to reserve their blocks rather than leaking throughout
* the address space during boot when many long-lived
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
* kernel allocations are made. Later some blocks near
* the start are marked MIGRATE_RESERVE by
* setup_zone_migrate_reserve()
*
* bitmap is created for zone's valid pfn range. but memmap
* can be created for invalid pages (for alignment)
* check here not to call set_pageblock_migratetype() against
* pfn out of zone.
*/
if ((z->zone_start_pfn <= pfn)
&& (pfn < z->zone_start_pfn + z->spanned_pages)
&& !(pfn & (pageblock_nr_pages - 1)))
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
set_pageblock_migratetype(page, MIGRATE_MOVABLE);
INIT_LIST_HEAD(&page->lru);
#ifdef WANT_PAGE_VIRTUAL
/* The shift won't overflow because ZONE_NORMAL is below 4G. */
if (!is_highmem_idx(zone))
set_page_address(page, __va(pfn << PAGE_SHIFT));
#endif
}
}
static void __meminit zone_init_free_lists(struct zone *zone)
{
int order, t;
for_each_migratetype_order(order, t) {
INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
zone->free_area[order].nr_free = 0;
}
}
#ifndef __HAVE_ARCH_MEMMAP_INIT
#define memmap_init(size, nid, zone, start_pfn) \
memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
#endif
static int zone_batchsize(struct zone *zone)
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
{
#ifdef CONFIG_MMU
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
int batch;
/*
* The per-cpu-pages pools are set to around 1000th of the
* size of the zone. But no more than 1/2 of a meg.
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
*
* OK, so we don't know how big the cache is. So guess.
*/
batch = zone->present_pages / 1024;
if (batch * PAGE_SIZE > 512 * 1024)
batch = (512 * 1024) / PAGE_SIZE;
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
batch /= 4; /* We effectively *= 4 below */
if (batch < 1)
batch = 1;
/*
* Clamp the batch to a 2^n - 1 value. Having a power
* of 2 value was found to be more likely to have
* suboptimal cache aliasing properties in some cases.
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
*
* For example if 2 tasks are alternately allocating
* batches of pages, one task can end up with a lot
* of pages of one half of the possible page colors
* and the other with pages of the other colors.
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
*/
batch = rounddown_pow_of_two(batch + batch/2) - 1;
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
return batch;
#else
/* The deferral and batching of frees should be suppressed under NOMMU
* conditions.
*
* The problem is that NOMMU needs to be able to allocate large chunks
* of contiguous memory as there's no hardware page translation to
* assemble apparent contiguous memory from discontiguous pages.
*
* Queueing large contiguous runs of pages for batching, however,
* causes the pages to actually be freed in smaller chunks. As there
* can be a significant delay between the individual batches being
* recycled, this leads to the once large chunks of space being
* fragmented and becoming unavailable for high-order allocations.
*/
return 0;
#endif
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
}
static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
{
struct per_cpu_pages *pcp;
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
int migratetype;
memset(p, 0, sizeof(*p));
pcp = &p->pcp;
pcp->count = 0;
pcp->high = 6 * batch;
pcp->batch = max(1UL, 1 * batch);
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
INIT_LIST_HEAD(&pcp->lists[migratetype]);
}
/*
* setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
* to the value high for the pageset p.
*/
static void setup_pagelist_highmark(struct per_cpu_pageset *p,
unsigned long high)
{
struct per_cpu_pages *pcp;
pcp = &p->pcp;
pcp->high = high;
pcp->batch = max(1UL, high/4);
if ((high/4) > (PAGE_SHIFT * 8))
pcp->batch = PAGE_SHIFT * 8;
}
static void setup_zone_pageset(struct zone *zone)
{
int cpu;
zone->pageset = alloc_percpu(struct per_cpu_pageset);
for_each_possible_cpu(cpu) {
struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
setup_pageset(pcp, zone_batchsize(zone));
if (percpu_pagelist_fraction)
setup_pagelist_highmark(pcp,
(zone->present_pages /
percpu_pagelist_fraction));
}
}
/*
* Allocate per cpu pagesets and initialize them.
* Before this call only boot pagesets were available.
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
*/
void __init setup_per_cpu_pageset(void)
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
{
struct zone *zone;
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
for_each_populated_zone(zone)
setup_zone_pageset(zone);
[PATCH] node local per-cpu-pages This patch modifies the way pagesets in struct zone are managed. Each zone has a per-cpu array of pagesets. So any particular CPU has some memory in each zone structure which belongs to itself. Even if that CPU is not local to that zone. So the patch relocates the pagesets for each cpu to the node that is nearest to the cpu instead of allocating the pagesets in the (possibly remote) target zone. This means that the operations to manage pages on remote zone can be done with information available locally. We play a macro trick so that non-NUMA pmachines avoid the additional pointer chase on the page allocator fastpath. AIM7 benchmark on a 32 CPU SGI Altix w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.68 100 484.6769 12.01 1.97 Fri Mar 25 11:01:42 2005 100 27140.46 89 271.4046 21.44 148.71 Fri Mar 25 11:02:04 2005 200 30792.02 82 153.9601 37.80 296.72 Fri Mar 25 11:02:42 2005 300 32209.27 81 107.3642 54.21 451.34 Fri Mar 25 11:03:37 2005 400 34962.83 78 87.4071 66.59 588.97 Fri Mar 25 11:04:44 2005 500 31676.92 75 63.3538 91.87 742.71 Fri Mar 25 11:06:16 2005 600 36032.69 73 60.0545 96.91 885.44 Fri Mar 25 11:07:54 2005 700 35540.43 77 50.7720 114.63 1024.28 Fri Mar 25 11:09:49 2005 800 33906.70 74 42.3834 137.32 1181.65 Fri Mar 25 11:12:06 2005 900 34120.67 73 37.9119 153.51 1325.26 Fri Mar 25 11:14:41 2005 1000 34802.37 74 34.8024 167.23 1465.26 Fri Mar 25 11:17:28 2005 with slab API changes and pageset patch: Tasks jobs/min jti jobs/min/task real cpu 1 485.00 100 485.0000 12.00 1.96 Fri Mar 25 11:46:18 2005 100 28000.96 89 280.0096 20.79 150.45 Fri Mar 25 11:46:39 2005 200 32285.80 79 161.4290 36.05 293.37 Fri Mar 25 11:47:16 2005 300 40424.15 84 134.7472 43.19 438.42 Fri Mar 25 11:47:59 2005 400 39155.01 79 97.8875 59.46 590.05 Fri Mar 25 11:48:59 2005 500 37881.25 82 75.7625 76.82 730.19 Fri Mar 25 11:50:16 2005 600 39083.14 78 65.1386 89.35 872.79 Fri Mar 25 11:51:46 2005 700 38627.83 77 55.1826 105.47 1022.46 Fri Mar 25 11:53:32 2005 800 39631.94 78 49.5399 117.48 1169.94 Fri Mar 25 11:55:30 2005 900 36903.70 79 41.0041 141.94 1310.78 Fri Mar 25 11:57:53 2005 1000 36201.23 77 36.2012 160.77 1458.31 Fri Mar 25 12:00:34 2005 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-22 00:14:47 +00:00
}
static noinline __init_refok
int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
{
int i;
struct pglist_data *pgdat = zone->zone_pgdat;
size_t alloc_size;
/*
* The per-page waitqueue mechanism uses hashed waitqueues
* per zone.
*/
zone->wait_table_hash_nr_entries =
wait_table_hash_nr_entries(zone_size_pages);
zone->wait_table_bits =
wait_table_bits(zone->wait_table_hash_nr_entries);
alloc_size = zone->wait_table_hash_nr_entries
* sizeof(wait_queue_head_t);
if (!slab_is_available()) {
zone->wait_table = (wait_queue_head_t *)
mm: use alloc_bootmem_node_nopanic() on really needed path Stefan found nobootmem does not work on his system that has only 8M of RAM. This causes an early panic: BIOS-provided physical RAM map: BIOS-88: 0000000000000000 - 000000000009f000 (usable) BIOS-88: 0000000000100000 - 0000000000840000 (usable) bootconsole [earlyser0] enabled Notice: NX (Execute Disable) protection missing in CPU or disabled in BIOS! DMI not present or invalid. last_pfn = 0x840 max_arch_pfn = 0x100000 init_memory_mapping: 0000000000000000-0000000000840000 8MB LOWMEM available. mapped low ram: 0 - 00840000 low ram: 0 - 00840000 Zone PFN ranges: DMA 0x00000001 -> 0x00001000 Normal empty Movable zone start PFN for each node early_node_map[2] active PFN ranges 0: 0x00000001 -> 0x0000009f 0: 0x00000100 -> 0x00000840 BUG: Int 6: CR2 (null) EDI c034663c ESI (null) EBP c0329f38 ESP c0329ef4 EBX c0346380 EDX 00000006 ECX ffffffff EAX fffffff4 err (null) EIP c0353191 CS c0320060 flg 00010082 Stack: (null) c030c533 000007cd (null) c030c533 00000001 (null) (null) 00000003 0000083f 00000018 00000002 00000002 c0329f6c c03534d6 (null) (null) 00000100 00000840 (null) c0329f64 00000001 00001000 (null) Pid: 0, comm: swapper Not tainted 2.6.36 #5 Call Trace: [<c02e3707>] ? 0xc02e3707 [<c035e6e5>] 0xc035e6e5 [<c0353191>] ? 0xc0353191 [<c03534d6>] 0xc03534d6 [<c034f1cd>] 0xc034f1cd [<c034a824>] 0xc034a824 [<c03513cb>] ? 0xc03513cb [<c0349432>] 0xc0349432 [<c0349066>] 0xc0349066 It turns out that we should ignore the low limit of 16M. Use alloc_bootmem_node_nopanic() in this case. [akpm@linux-foundation.org: less mess] Signed-off-by: Yinghai LU <yinghai@kernel.org> Reported-by: Stefan Hellermann <stefan@the2masters.de> Tested-by: Stefan Hellermann <stefan@the2masters.de> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@linux.intel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: <stable@kernel.org> [2.6.34+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-11 22:13:32 +00:00
alloc_bootmem_node_nopanic(pgdat, alloc_size);
} else {
/*
* This case means that a zone whose size was 0 gets new memory
* via memory hot-add.
* But it may be the case that a new node was hot-added. In
* this case vmalloc() will not be able to use this new node's
* memory - this wait_table must be initialized to use this new
* node itself as well.
* To use this new node's memory, further consideration will be
* necessary.
*/
zone->wait_table = vmalloc(alloc_size);
}
if (!zone->wait_table)
return -ENOMEM;
for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
init_waitqueue_head(zone->wait_table + i);
return 0;
}
static int __zone_pcp_update(void *data)
{
struct zone *zone = data;
int cpu;
unsigned long batch = zone_batchsize(zone), flags;
for_each_possible_cpu(cpu) {
struct per_cpu_pageset *pset;
struct per_cpu_pages *pcp;
pset = per_cpu_ptr(zone->pageset, cpu);
pcp = &pset->pcp;
local_irq_save(flags);
page-allocator: split per-cpu list into one-list-per-migrate-type The following two patches remove searching in the page allocator fast-path by maintaining multiple free-lists in the per-cpu structure. At the time the search was introduced, increasing the per-cpu structures would waste a lot of memory as per-cpu structures were statically allocated at compile-time. This is no longer the case. The patches are as follows. They are based on mmotm-2009-08-27. Patch 1 adds multiple lists to struct per_cpu_pages, one per migratetype that can be stored on the PCP lists. Patch 2 notes that the pcpu drain path check empty lists multiple times. The patch reduces the number of checks by maintaining a count of free lists encountered. Lists containing pages will then free multiple pages in batch The patches were tested with kernbench, netperf udp/tcp, hackbench and sysbench. The netperf tests were not bound to any CPU in particular and were run such that the results should be 99% confidence that the reported results are within 1% of the estimated mean. sysbench was run with a postgres background and read-only tests. Similar to netperf, it was run multiple times so that it's 99% confidence results are within 1%. The patches were tested on x86, x86-64 and ppc64 as x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.34% to 2.28% gain netperf-tcp - 0.45% to 1.22% gain hackbench - Small variances, very close to noise sysbench - Very small gains x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine) kernbench - No significant difference, variance well within noise netperf-udp - 1.83% to 10.42% gains netperf-tcp - No conclusive until buffer >= PAGE_SIZE 4096 +15.83% 8192 + 0.34% (not significant) 16384 + 1% hackbench - Small gains, very close to noise sysbench - 0.79% to 1.6% gain ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation) kernbench - No significant difference, variance well within noise netperf-udp - 2-3% gain for almost all buffer sizes tested netperf-tcp - losses on small buffers, gains on larger buffers possibly indicates some bad caching effect. hackbench - No significant difference sysbench - 2-4% gain This patch: Currently the per-cpu page allocator searches the PCP list for pages of the correct migrate-type to reduce the possibility of pages being inappropriate placed from a fragmentation perspective. This search is potentially expensive in a fast-path and undesirable. Splitting the per-cpu list into multiple lists increases the size of a per-cpu structure and this was potentially a major problem at the time the search was introduced. These problem has been mitigated as now only the necessary number of structures is allocated for the running system. This patch replaces a list search in the per-cpu allocator with one list per migrate type. The potential snag with this approach is when bulk freeing pages. We round-robin free pages based on migrate type which has little bearing on the cache hotness of the page and potentially checks empty lists repeatedly in the event the majority of PCP pages are of one type. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Nick Piggin <npiggin@suse.de> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-22 00:03:19 +00:00
free_pcppages_bulk(zone, pcp->count, pcp);
setup_pageset(pset, batch);
local_irq_restore(flags);
}
return 0;
}
void zone_pcp_update(struct zone *zone)
{
stop_machine(__zone_pcp_update, zone, NULL);
}
static __meminit void zone_pcp_init(struct zone *zone)
{
/*
* per cpu subsystem is not up at this point. The following code
* relies on the ability of the linker to provide the
* offset of a (static) per cpu variable into the per cpu area.
*/
zone->pageset = &boot_pageset;
if (zone->present_pages)
printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
zone->name, zone->present_pages,
zone_batchsize(zone));
}
__meminit int init_currently_empty_zone(struct zone *zone,
unsigned long zone_start_pfn,
unsigned long size,
enum memmap_context context)
{
struct pglist_data *pgdat = zone->zone_pgdat;
int ret;
ret = zone_wait_table_init(zone, size);
if (ret)
return ret;
pgdat->nr_zones = zone_idx(zone) + 1;
zone->zone_start_pfn = zone_start_pfn;
mminit_dprintk(MMINIT_TRACE, "memmap_init",
"Initialising map node %d zone %lu pfns %lu -> %lu\n",
pgdat->node_id,
(unsigned long)zone_idx(zone),
zone_start_pfn, (zone_start_pfn + size));
zone_init_free_lists(zone);
return 0;
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
/*
* Basic iterator support. Return the first range of PFNs for a node
* Note: nid == MAX_NUMNODES returns first region regardless of node
*/
static int __meminit first_active_region_index_in_nid(int nid)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
int i;
for (i = 0; i < nr_nodemap_entries; i++)
if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
return i;
return -1;
}
/*
* Basic iterator support. Return the next active range of PFNs for a node
* Note: nid == MAX_NUMNODES returns next region regardless of node
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
static int __meminit next_active_region_index_in_nid(int index, int nid)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
for (index = index + 1; index < nr_nodemap_entries; index++)
if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
return index;
return -1;
}
#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
/*
* Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
* Architectures may implement their own version but if add_active_range()
* was used and there are no special requirements, this is a convenient
* alternative
*/
mm: clean up for early_pfn_to_nid() What's happening is that the assertion in mm/page_alloc.c:move_freepages() is triggering: BUG_ON(page_zone(start_page) != page_zone(end_page)); Once I knew this is what was happening, I added some annotations: if (unlikely(page_zone(start_page) != page_zone(end_page))) { printk(KERN_ERR "move_freepages: Bogus zones: " "start_page[%p] end_page[%p] zone[%p]\n", start_page, end_page, zone); printk(KERN_ERR "move_freepages: " "start_zone[%p] end_zone[%p]\n", page_zone(start_page), page_zone(end_page)); printk(KERN_ERR "move_freepages: " "start_pfn[0x%lx] end_pfn[0x%lx]\n", page_to_pfn(start_page), page_to_pfn(end_page)); printk(KERN_ERR "move_freepages: " "start_nid[%d] end_nid[%d]\n", page_to_nid(start_page), page_to_nid(end_page)); ... And here's what I got: move_freepages: Bogus zones: start_page[2207d0000] end_page[2207dffc0] zone[fffff8103effcb00] move_freepages: start_zone[fffff8103effcb00] end_zone[fffff8003fffeb00] move_freepages: start_pfn[0x81f600] end_pfn[0x81f7ff] move_freepages: start_nid[1] end_nid[0] My memory layout on this box is: [ 0.000000] Zone PFN ranges: [ 0.000000] Normal 0x00000000 -> 0x0081ff5d [ 0.000000] Movable zone start PFN for each node [ 0.000000] early_node_map[8] active PFN ranges [ 0.000000] 0: 0x00000000 -> 0x00020000 [ 0.000000] 1: 0x00800000 -> 0x0081f7ff [ 0.000000] 1: 0x0081f800 -> 0x0081fe50 [ 0.000000] 1: 0x0081fed1 -> 0x0081fed8 [ 0.000000] 1: 0x0081feda -> 0x0081fedb [ 0.000000] 1: 0x0081fedd -> 0x0081fee5 [ 0.000000] 1: 0x0081fee7 -> 0x0081ff51 [ 0.000000] 1: 0x0081ff59 -> 0x0081ff5d So it's a block move in that 0x81f600-->0x81f7ff region which triggers the problem. This patch: Declaration of early_pfn_to_nid() is scattered over per-arch include files, and it seems it's complicated to know when the declaration is used. I think it makes fix-for-memmap-init not easy. This patch moves all declaration to include/linux/mm.h After this, if !CONFIG_NODES_POPULATES_NODE_MAP && !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use static definition in include/linux/mm.h else if !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use generic definition in mm/page_alloc.c else -> per-arch back end function will be called. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reported-by: David Miller <davem@davemlloft.net> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: <stable@kernel.org> [2.6.25.x, 2.6.26.x, 2.6.27.x, 2.6.28.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-18 22:48:32 +00:00
int __meminit __early_pfn_to_nid(unsigned long pfn)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
int i;
for (i = 0; i < nr_nodemap_entries; i++) {
unsigned long start_pfn = early_node_map[i].start_pfn;
unsigned long end_pfn = early_node_map[i].end_pfn;
if (start_pfn <= pfn && pfn < end_pfn)
return early_node_map[i].nid;
}
/* This is a memory hole */
return -1;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
}
#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
mm: clean up for early_pfn_to_nid() What's happening is that the assertion in mm/page_alloc.c:move_freepages() is triggering: BUG_ON(page_zone(start_page) != page_zone(end_page)); Once I knew this is what was happening, I added some annotations: if (unlikely(page_zone(start_page) != page_zone(end_page))) { printk(KERN_ERR "move_freepages: Bogus zones: " "start_page[%p] end_page[%p] zone[%p]\n", start_page, end_page, zone); printk(KERN_ERR "move_freepages: " "start_zone[%p] end_zone[%p]\n", page_zone(start_page), page_zone(end_page)); printk(KERN_ERR "move_freepages: " "start_pfn[0x%lx] end_pfn[0x%lx]\n", page_to_pfn(start_page), page_to_pfn(end_page)); printk(KERN_ERR "move_freepages: " "start_nid[%d] end_nid[%d]\n", page_to_nid(start_page), page_to_nid(end_page)); ... And here's what I got: move_freepages: Bogus zones: start_page[2207d0000] end_page[2207dffc0] zone[fffff8103effcb00] move_freepages: start_zone[fffff8103effcb00] end_zone[fffff8003fffeb00] move_freepages: start_pfn[0x81f600] end_pfn[0x81f7ff] move_freepages: start_nid[1] end_nid[0] My memory layout on this box is: [ 0.000000] Zone PFN ranges: [ 0.000000] Normal 0x00000000 -> 0x0081ff5d [ 0.000000] Movable zone start PFN for each node [ 0.000000] early_node_map[8] active PFN ranges [ 0.000000] 0: 0x00000000 -> 0x00020000 [ 0.000000] 1: 0x00800000 -> 0x0081f7ff [ 0.000000] 1: 0x0081f800 -> 0x0081fe50 [ 0.000000] 1: 0x0081fed1 -> 0x0081fed8 [ 0.000000] 1: 0x0081feda -> 0x0081fedb [ 0.000000] 1: 0x0081fedd -> 0x0081fee5 [ 0.000000] 1: 0x0081fee7 -> 0x0081ff51 [ 0.000000] 1: 0x0081ff59 -> 0x0081ff5d So it's a block move in that 0x81f600-->0x81f7ff region which triggers the problem. This patch: Declaration of early_pfn_to_nid() is scattered over per-arch include files, and it seems it's complicated to know when the declaration is used. I think it makes fix-for-memmap-init not easy. This patch moves all declaration to include/linux/mm.h After this, if !CONFIG_NODES_POPULATES_NODE_MAP && !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use static definition in include/linux/mm.h else if !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use generic definition in mm/page_alloc.c else -> per-arch back end function will be called. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reported-by: David Miller <davem@davemlloft.net> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: <stable@kernel.org> [2.6.25.x, 2.6.26.x, 2.6.27.x, 2.6.28.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-18 22:48:32 +00:00
int __meminit early_pfn_to_nid(unsigned long pfn)
{
int nid;
nid = __early_pfn_to_nid(pfn);
if (nid >= 0)
return nid;
/* just returns 0 */
return 0;
mm: clean up for early_pfn_to_nid() What's happening is that the assertion in mm/page_alloc.c:move_freepages() is triggering: BUG_ON(page_zone(start_page) != page_zone(end_page)); Once I knew this is what was happening, I added some annotations: if (unlikely(page_zone(start_page) != page_zone(end_page))) { printk(KERN_ERR "move_freepages: Bogus zones: " "start_page[%p] end_page[%p] zone[%p]\n", start_page, end_page, zone); printk(KERN_ERR "move_freepages: " "start_zone[%p] end_zone[%p]\n", page_zone(start_page), page_zone(end_page)); printk(KERN_ERR "move_freepages: " "start_pfn[0x%lx] end_pfn[0x%lx]\n", page_to_pfn(start_page), page_to_pfn(end_page)); printk(KERN_ERR "move_freepages: " "start_nid[%d] end_nid[%d]\n", page_to_nid(start_page), page_to_nid(end_page)); ... And here's what I got: move_freepages: Bogus zones: start_page[2207d0000] end_page[2207dffc0] zone[fffff8103effcb00] move_freepages: start_zone[fffff8103effcb00] end_zone[fffff8003fffeb00] move_freepages: start_pfn[0x81f600] end_pfn[0x81f7ff] move_freepages: start_nid[1] end_nid[0] My memory layout on this box is: [ 0.000000] Zone PFN ranges: [ 0.000000] Normal 0x00000000 -> 0x0081ff5d [ 0.000000] Movable zone start PFN for each node [ 0.000000] early_node_map[8] active PFN ranges [ 0.000000] 0: 0x00000000 -> 0x00020000 [ 0.000000] 1: 0x00800000 -> 0x0081f7ff [ 0.000000] 1: 0x0081f800 -> 0x0081fe50 [ 0.000000] 1: 0x0081fed1 -> 0x0081fed8 [ 0.000000] 1: 0x0081feda -> 0x0081fedb [ 0.000000] 1: 0x0081fedd -> 0x0081fee5 [ 0.000000] 1: 0x0081fee7 -> 0x0081ff51 [ 0.000000] 1: 0x0081ff59 -> 0x0081ff5d So it's a block move in that 0x81f600-->0x81f7ff region which triggers the problem. This patch: Declaration of early_pfn_to_nid() is scattered over per-arch include files, and it seems it's complicated to know when the declaration is used. I think it makes fix-for-memmap-init not easy. This patch moves all declaration to include/linux/mm.h After this, if !CONFIG_NODES_POPULATES_NODE_MAP && !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use static definition in include/linux/mm.h else if !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use generic definition in mm/page_alloc.c else -> per-arch back end function will be called. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reported-by: David Miller <davem@davemlloft.net> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: <stable@kernel.org> [2.6.25.x, 2.6.26.x, 2.6.27.x, 2.6.28.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-18 22:48:32 +00:00
}
#ifdef CONFIG_NODES_SPAN_OTHER_NODES
bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
{
int nid;
nid = __early_pfn_to_nid(pfn);
if (nid >= 0 && nid != node)
return false;
return true;
}
#endif
mm: clean up for early_pfn_to_nid() What's happening is that the assertion in mm/page_alloc.c:move_freepages() is triggering: BUG_ON(page_zone(start_page) != page_zone(end_page)); Once I knew this is what was happening, I added some annotations: if (unlikely(page_zone(start_page) != page_zone(end_page))) { printk(KERN_ERR "move_freepages: Bogus zones: " "start_page[%p] end_page[%p] zone[%p]\n", start_page, end_page, zone); printk(KERN_ERR "move_freepages: " "start_zone[%p] end_zone[%p]\n", page_zone(start_page), page_zone(end_page)); printk(KERN_ERR "move_freepages: " "start_pfn[0x%lx] end_pfn[0x%lx]\n", page_to_pfn(start_page), page_to_pfn(end_page)); printk(KERN_ERR "move_freepages: " "start_nid[%d] end_nid[%d]\n", page_to_nid(start_page), page_to_nid(end_page)); ... And here's what I got: move_freepages: Bogus zones: start_page[2207d0000] end_page[2207dffc0] zone[fffff8103effcb00] move_freepages: start_zone[fffff8103effcb00] end_zone[fffff8003fffeb00] move_freepages: start_pfn[0x81f600] end_pfn[0x81f7ff] move_freepages: start_nid[1] end_nid[0] My memory layout on this box is: [ 0.000000] Zone PFN ranges: [ 0.000000] Normal 0x00000000 -> 0x0081ff5d [ 0.000000] Movable zone start PFN for each node [ 0.000000] early_node_map[8] active PFN ranges [ 0.000000] 0: 0x00000000 -> 0x00020000 [ 0.000000] 1: 0x00800000 -> 0x0081f7ff [ 0.000000] 1: 0x0081f800 -> 0x0081fe50 [ 0.000000] 1: 0x0081fed1 -> 0x0081fed8 [ 0.000000] 1: 0x0081feda -> 0x0081fedb [ 0.000000] 1: 0x0081fedd -> 0x0081fee5 [ 0.000000] 1: 0x0081fee7 -> 0x0081ff51 [ 0.000000] 1: 0x0081ff59 -> 0x0081ff5d So it's a block move in that 0x81f600-->0x81f7ff region which triggers the problem. This patch: Declaration of early_pfn_to_nid() is scattered over per-arch include files, and it seems it's complicated to know when the declaration is used. I think it makes fix-for-memmap-init not easy. This patch moves all declaration to include/linux/mm.h After this, if !CONFIG_NODES_POPULATES_NODE_MAP && !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use static definition in include/linux/mm.h else if !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID -> Use generic definition in mm/page_alloc.c else -> per-arch back end function will be called. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reported-by: David Miller <davem@davemlloft.net> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: <stable@kernel.org> [2.6.25.x, 2.6.26.x, 2.6.27.x, 2.6.28.x] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-18 22:48:32 +00:00
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Basic iterator support to walk early_node_map[] */
#define for_each_active_range_index_in_nid(i, nid) \
for (i = first_active_region_index_in_nid(nid); i != -1; \
i = next_active_region_index_in_nid(i, nid))
/**
* free_bootmem_with_active_regions - Call free_bootmem_node for each active range
* @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
* @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*
* If an architecture guarantees that all ranges registered with
* add_active_ranges() contain no holes and may be freed, this
* this function may be used instead of calling free_bootmem() manually.
*/
void __init free_bootmem_with_active_regions(int nid,
unsigned long max_low_pfn)
{
int i;
for_each_active_range_index_in_nid(i, nid) {
unsigned long size_pages = 0;
unsigned long end_pfn = early_node_map[i].end_pfn;
if (early_node_map[i].start_pfn >= max_low_pfn)
continue;
if (end_pfn > max_low_pfn)
end_pfn = max_low_pfn;
size_pages = end_pfn - early_node_map[i].start_pfn;
free_bootmem_node(NODE_DATA(early_node_map[i].nid),
PFN_PHYS(early_node_map[i].start_pfn),
size_pages << PAGE_SHIFT);
}
}
#ifdef CONFIG_HAVE_MEMBLOCK
/*
* Basic iterator support. Return the last range of PFNs for a node
* Note: nid == MAX_NUMNODES returns last region regardless of node
*/
static int __meminit last_active_region_index_in_nid(int nid)
{
int i;
for (i = nr_nodemap_entries - 1; i >= 0; i--)
if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
return i;
return -1;
}
/*
* Basic iterator support. Return the previous active range of PFNs for a node
* Note: nid == MAX_NUMNODES returns next region regardless of node
*/
static int __meminit previous_active_region_index_in_nid(int index, int nid)
{
for (index = index - 1; index >= 0; index--)
if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
return index;
return -1;
}
#define for_each_active_range_index_in_nid_reverse(i, nid) \
for (i = last_active_region_index_in_nid(nid); i != -1; \
i = previous_active_region_index_in_nid(i, nid))
u64 __init find_memory_core_early(int nid, u64 size, u64 align,
u64 goal, u64 limit)
{
int i;
/* Need to go over early_node_map to find out good range for node */
for_each_active_range_index_in_nid_reverse(i, nid) {
u64 addr;
u64 ei_start, ei_last;
u64 final_start, final_end;
ei_last = early_node_map[i].end_pfn;
ei_last <<= PAGE_SHIFT;
ei_start = early_node_map[i].start_pfn;
ei_start <<= PAGE_SHIFT;
final_start = max(ei_start, goal);
final_end = min(ei_last, limit);
if (final_start >= final_end)
continue;
addr = memblock_find_in_range(final_start, final_end, size, align);
if (addr == MEMBLOCK_ERROR)
continue;
return addr;
}
return MEMBLOCK_ERROR;
}
#endif
int __init add_from_early_node_map(struct range *range, int az,
int nr_range, int nid)
{
int i;
u64 start, end;
/* need to go over early_node_map to find out good range for node */
for_each_active_range_index_in_nid(i, nid) {
start = early_node_map[i].start_pfn;
end = early_node_map[i].end_pfn;
nr_range = add_range(range, az, nr_range, start, end);
}
return nr_range;
}
void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
{
int i;
int ret;
for_each_active_range_index_in_nid(i, nid) {
ret = work_fn(early_node_map[i].start_pfn,
early_node_map[i].end_pfn, data);
if (ret)
break;
}
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/**
* sparse_memory_present_with_active_regions - Call memory_present for each active range
* @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*
* If an architecture guarantees that all ranges registered with
* add_active_ranges() contain no holes and may be freed, this
* function may be used instead of calling memory_present() manually.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
void __init sparse_memory_present_with_active_regions(int nid)
{
int i;
for_each_active_range_index_in_nid(i, nid)
memory_present(early_node_map[i].nid,
early_node_map[i].start_pfn,
early_node_map[i].end_pfn);
}
/**
* get_pfn_range_for_nid - Return the start and end page frames for a node
* @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
* @start_pfn: Passed by reference. On return, it will have the node start_pfn.
* @end_pfn: Passed by reference. On return, it will have the node end_pfn.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*
* It returns the start and end page frame of a node based on information
* provided by an arch calling add_active_range(). If called for a node
* with no available memory, a warning is printed and the start and end
* PFNs will be 0.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
void __meminit get_pfn_range_for_nid(unsigned int nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long *start_pfn, unsigned long *end_pfn)
{
int i;
*start_pfn = -1UL;
*end_pfn = 0;
for_each_active_range_index_in_nid(i, nid) {
*start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
*end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
}
if (*start_pfn == -1UL)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*start_pfn = 0;
}
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/*
* This finds a zone that can be used for ZONE_MOVABLE pages. The
* assumption is made that zones within a node are ordered in monotonic
* increasing memory addresses so that the "highest" populated zone is used
*/
static void __init find_usable_zone_for_movable(void)
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
{
int zone_index;
for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
if (zone_index == ZONE_MOVABLE)
continue;
if (arch_zone_highest_possible_pfn[zone_index] >
arch_zone_lowest_possible_pfn[zone_index])
break;
}
VM_BUG_ON(zone_index == -1);
movable_zone = zone_index;
}
/*
* The zone ranges provided by the architecture do not include ZONE_MOVABLE
* because it is sized independent of architecture. Unlike the other zones,
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
* the starting point for ZONE_MOVABLE is not fixed. It may be different
* in each node depending on the size of each node and how evenly kernelcore
* is distributed. This helper function adjusts the zone ranges
* provided by the architecture for a given node by using the end of the
* highest usable zone for ZONE_MOVABLE. This preserves the assumption that
* zones within a node are in order of monotonic increases memory addresses
*/
static void __meminit adjust_zone_range_for_zone_movable(int nid,
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
unsigned long zone_type,
unsigned long node_start_pfn,
unsigned long node_end_pfn,
unsigned long *zone_start_pfn,
unsigned long *zone_end_pfn)
{
/* Only adjust if ZONE_MOVABLE is on this node */
if (zone_movable_pfn[nid]) {
/* Size ZONE_MOVABLE */
if (zone_type == ZONE_MOVABLE) {
*zone_start_pfn = zone_movable_pfn[nid];
*zone_end_pfn = min(node_end_pfn,
arch_zone_highest_possible_pfn[movable_zone]);
/* Adjust for ZONE_MOVABLE starting within this range */
} else if (*zone_start_pfn < zone_movable_pfn[nid] &&
*zone_end_pfn > zone_movable_pfn[nid]) {
*zone_end_pfn = zone_movable_pfn[nid];
/* Check if this whole range is within ZONE_MOVABLE */
} else if (*zone_start_pfn >= zone_movable_pfn[nid])
*zone_start_pfn = *zone_end_pfn;
}
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/*
* Return the number of pages a zone spans in a node, including holes
* present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
*/
static unsigned long __meminit zone_spanned_pages_in_node(int nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long zone_type,
unsigned long *ignored)
{
unsigned long node_start_pfn, node_end_pfn;
unsigned long zone_start_pfn, zone_end_pfn;
/* Get the start and end of the node and zone */
get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
adjust_zone_range_for_zone_movable(nid, zone_type,
node_start_pfn, node_end_pfn,
&zone_start_pfn, &zone_end_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Check that this node has pages within the zone's required range */
if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
return 0;
/* Move the zone boundaries inside the node if necessary */
zone_end_pfn = min(zone_end_pfn, node_end_pfn);
zone_start_pfn = max(zone_start_pfn, node_start_pfn);
/* Return the spanned pages */
return zone_end_pfn - zone_start_pfn;
}
/*
* Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
* then all holes in the requested range will be accounted for.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
x86: Fix checking of SRAT when node 0 ram is not from 0 Found one system that boot from socket1 instead of socket0, SRAT get rejected... [ 0.000000] SRAT: Node 1 PXM 0 0-a0000 [ 0.000000] SRAT: Node 1 PXM 0 100000-80000000 [ 0.000000] SRAT: Node 1 PXM 0 100000000-2080000000 [ 0.000000] SRAT: Node 0 PXM 1 2080000000-4080000000 [ 0.000000] SRAT: Node 2 PXM 2 4080000000-6080000000 [ 0.000000] SRAT: Node 3 PXM 3 6080000000-8080000000 [ 0.000000] SRAT: Node 4 PXM 4 8080000000-a080000000 [ 0.000000] SRAT: Node 5 PXM 5 a080000000-c080000000 [ 0.000000] SRAT: Node 6 PXM 6 c080000000-e080000000 [ 0.000000] SRAT: Node 7 PXM 7 e080000000-10080000000 ... [ 0.000000] NUMA: Allocated memnodemap from 500000 - 701040 [ 0.000000] NUMA: Using 20 for the hash shift. [ 0.000000] Adding active range (0, 0x2080000, 0x4080000) 0 entries of 3200 used [ 0.000000] Adding active range (1, 0x0, 0x96) 1 entries of 3200 used [ 0.000000] Adding active range (1, 0x100, 0x7f750) 2 entries of 3200 used [ 0.000000] Adding active range (1, 0x100000, 0x2080000) 3 entries of 3200 used [ 0.000000] Adding active range (2, 0x4080000, 0x6080000) 4 entries of 3200 used [ 0.000000] Adding active range (3, 0x6080000, 0x8080000) 5 entries of 3200 used [ 0.000000] Adding active range (4, 0x8080000, 0xa080000) 6 entries of 3200 used [ 0.000000] Adding active range (5, 0xa080000, 0xc080000) 7 entries of 3200 used [ 0.000000] Adding active range (6, 0xc080000, 0xe080000) 8 entries of 3200 used [ 0.000000] Adding active range (7, 0xe080000, 0x10080000) 9 entries of 3200 used [ 0.000000] SRAT: PXMs only cover 917504MB of your 1048566MB e820 RAM. Not used. [ 0.000000] SRAT: SRAT not used. the early_node_map is not sorted because node0 with non zero start come first. so try to sort it right away after all regions are registered. also fixs refression by 8716273c (x86: Export srat physical topology) -v2: make it more solid to handle cross node case like node0 [0,4g), [8,12g) and node1 [4g, 8g), [12g, 16g) -v3: update comments. Reported-and-tested-by: Jens Axboe <jens.axboe@oracle.com> Signed-off-by: Yinghai Lu <yinghai@kernel.org> LKML-Reference: <4B2579D2.3010201@kernel.org> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2009-12-16 01:59:02 +00:00
unsigned long __meminit __absent_pages_in_range(int nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long range_start_pfn,
unsigned long range_end_pfn)
{
int i = 0;
unsigned long prev_end_pfn = 0, hole_pages = 0;
unsigned long start_pfn;
/* Find the end_pfn of the first active range of pfns in the node */
i = first_active_region_index_in_nid(nid);
if (i == -1)
return 0;
prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
/* Account for ranges before physical memory on this node */
if (early_node_map[i].start_pfn > range_start_pfn)
hole_pages = prev_end_pfn - range_start_pfn;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Find all holes for the zone within the node */
for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
/* No need to continue if prev_end_pfn is outside the zone */
if (prev_end_pfn >= range_end_pfn)
break;
/* Make sure the end of the zone is not within the hole */
start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
prev_end_pfn = max(prev_end_pfn, range_start_pfn);
/* Update the hole size cound and move on */
if (start_pfn > range_start_pfn) {
BUG_ON(prev_end_pfn > start_pfn);
hole_pages += start_pfn - prev_end_pfn;
}
prev_end_pfn = early_node_map[i].end_pfn;
}
/* Account for ranges past physical memory on this node */
if (range_end_pfn > prev_end_pfn)
hole_pages += range_end_pfn -
max(range_start_pfn, prev_end_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
return hole_pages;
}
/**
* absent_pages_in_range - Return number of page frames in holes within a range
* @start_pfn: The start PFN to start searching for holes
* @end_pfn: The end PFN to stop searching for holes
*
* It returns the number of pages frames in memory holes within a range.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
unsigned long __init absent_pages_in_range(unsigned long start_pfn,
unsigned long end_pfn)
{
return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
}
/* Return the number of page frames in holes in a zone on a node */
static unsigned long __meminit zone_absent_pages_in_node(int nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long zone_type,
unsigned long *ignored)
{
unsigned long node_start_pfn, node_end_pfn;
unsigned long zone_start_pfn, zone_end_pfn;
get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
node_start_pfn);
zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
node_end_pfn);
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
adjust_zone_range_for_zone_movable(nid, zone_type,
node_start_pfn, node_end_pfn,
&zone_start_pfn, &zone_end_pfn);
return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#else
static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long zone_type,
unsigned long *zones_size)
{
return zones_size[zone_type];
}
static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long zone_type,
unsigned long *zholes_size)
{
if (!zholes_size)
return 0;
return zholes_size[zone_type];
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#endif
static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
unsigned long *zones_size, unsigned long *zholes_size)
{
unsigned long realtotalpages, totalpages = 0;
enum zone_type i;
for (i = 0; i < MAX_NR_ZONES; i++)
totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
zones_size);
pgdat->node_spanned_pages = totalpages;
realtotalpages = totalpages;
for (i = 0; i < MAX_NR_ZONES; i++)
realtotalpages -=
zone_absent_pages_in_node(pgdat->node_id, i,
zholes_size);
pgdat->node_present_pages = realtotalpages;
printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
realtotalpages);
}
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
#ifndef CONFIG_SPARSEMEM
/*
* Calculate the size of the zone->blockflags rounded to an unsigned long
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
* Start by making sure zonesize is a multiple of pageblock_order by rounding
* up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
* round what is now in bits to nearest long in bits, then return it in
* bytes.
*/
static unsigned long __init usemap_size(unsigned long zonesize)
{
unsigned long usemapsize;
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
usemapsize = roundup(zonesize, pageblock_nr_pages);
usemapsize = usemapsize >> pageblock_order;
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
usemapsize *= NR_PAGEBLOCK_BITS;
usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
return usemapsize / 8;
}
static void __init setup_usemap(struct pglist_data *pgdat,
struct zone *zone, unsigned long zonesize)
{
unsigned long usemapsize = usemap_size(zonesize);
zone->pageblock_flags = NULL;
if (usemapsize)
mm: use alloc_bootmem_node_nopanic() on really needed path Stefan found nobootmem does not work on his system that has only 8M of RAM. This causes an early panic: BIOS-provided physical RAM map: BIOS-88: 0000000000000000 - 000000000009f000 (usable) BIOS-88: 0000000000100000 - 0000000000840000 (usable) bootconsole [earlyser0] enabled Notice: NX (Execute Disable) protection missing in CPU or disabled in BIOS! DMI not present or invalid. last_pfn = 0x840 max_arch_pfn = 0x100000 init_memory_mapping: 0000000000000000-0000000000840000 8MB LOWMEM available. mapped low ram: 0 - 00840000 low ram: 0 - 00840000 Zone PFN ranges: DMA 0x00000001 -> 0x00001000 Normal empty Movable zone start PFN for each node early_node_map[2] active PFN ranges 0: 0x00000001 -> 0x0000009f 0: 0x00000100 -> 0x00000840 BUG: Int 6: CR2 (null) EDI c034663c ESI (null) EBP c0329f38 ESP c0329ef4 EBX c0346380 EDX 00000006 ECX ffffffff EAX fffffff4 err (null) EIP c0353191 CS c0320060 flg 00010082 Stack: (null) c030c533 000007cd (null) c030c533 00000001 (null) (null) 00000003 0000083f 00000018 00000002 00000002 c0329f6c c03534d6 (null) (null) 00000100 00000840 (null) c0329f64 00000001 00001000 (null) Pid: 0, comm: swapper Not tainted 2.6.36 #5 Call Trace: [<c02e3707>] ? 0xc02e3707 [<c035e6e5>] 0xc035e6e5 [<c0353191>] ? 0xc0353191 [<c03534d6>] 0xc03534d6 [<c034f1cd>] 0xc034f1cd [<c034a824>] 0xc034a824 [<c03513cb>] ? 0xc03513cb [<c0349432>] 0xc0349432 [<c0349066>] 0xc0349066 It turns out that we should ignore the low limit of 16M. Use alloc_bootmem_node_nopanic() in this case. [akpm@linux-foundation.org: less mess] Signed-off-by: Yinghai LU <yinghai@kernel.org> Reported-by: Stefan Hellermann <stefan@the2masters.de> Tested-by: Stefan Hellermann <stefan@the2masters.de> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@linux.intel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: <stable@kernel.org> [2.6.34+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-11 22:13:32 +00:00
zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
usemapsize);
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
}
#else
static inline void setup_usemap(struct pglist_data *pgdat,
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
struct zone *zone, unsigned long zonesize) {}
#endif /* CONFIG_SPARSEMEM */
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
/* Return a sensible default order for the pageblock size. */
static inline int pageblock_default_order(void)
{
if (HPAGE_SHIFT > PAGE_SHIFT)
return HUGETLB_PAGE_ORDER;
return MAX_ORDER-1;
}
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
static inline void __init set_pageblock_order(unsigned int order)
{
/* Check that pageblock_nr_pages has not already been setup */
if (pageblock_order)
return;
/*
* Assume the largest contiguous order of interest is a huge page.
* This value may be variable depending on boot parameters on IA64
*/
pageblock_order = order;
}
#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
/*
* When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
* and pageblock_default_order() are unused as pageblock_order is set
* at compile-time. See include/linux/pageblock-flags.h for the values of
* pageblock_order based on the kernel config
*/
static inline int pageblock_default_order(unsigned int order)
{
return MAX_ORDER-1;
}
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
#define set_pageblock_order(x) do {} while (0)
#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
/*
* Set up the zone data structures:
* - mark all pages reserved
* - mark all memory queues empty
* - clear the memory bitmaps
*/
Solve section mismatch for free_area_init_core. WARNING: vmlinux.o(.meminit.text+0x649): Section mismatch in reference from the function free_area_init_core() to the function .init.text:setup_usemap() The function __meminit free_area_init_core() references a function __init setup_usemap(). If free_area_init_core is only used by setup_usemap then annotate free_area_init_core with a matching annotation. The warning is covers this stack of functions in mm/page_alloc.c: alloc_bootmem_node must be marked __init. alloc_bootmem_node is used by setup_usemap, if !SPARSEMEM. (usemap_size is only used by setup_usemap, if !SPARSEMEM.) setup_usemap is only used by free_area_init_core. free_area_init_core is only used by free_area_init_node. free_area_init_node is used by: arch/alpha/mm/numa.c: __init paging_init() arch/arm/mm/init.c: __init bootmem_init_node() arch/avr32/mm/init.c: __init paging_init() arch/cris/arch-v10/mm/init.c: __init paging_init() arch/cris/arch-v32/mm/init.c: __init paging_init() arch/m32r/mm/discontig.c: __init zone_sizes_init() arch/m32r/mm/init.c: __init zone_sizes_init() arch/m68k/mm/motorola.c: __init paging_init() arch/m68k/mm/sun3mmu.c: __init paging_init() arch/mips/sgi-ip27/ip27-memory.c: __init paging_init() arch/parisc/mm/init.c: __init paging_init() arch/sparc/mm/srmmu.c: __init srmmu_paging_init() arch/sparc/mm/sun4c.c: __init sun4c_paging_init() arch/sparc64/mm/init.c: __init paging_init() mm/page_alloc.c: __init free_area_init_nodes() mm/page_alloc.c: __init free_area_init() and mm/memory_hotplug.c: hotadd_new_pgdat() hotadd_new_pgdat can not be an __init function, but: It is compiled for MEMORY_HOTPLUG configurations only MEMORY_HOTPLUG depends on SPARSEMEM || X86_64_ACPI_NUMA X86_64_ACPI_NUMA depends on X86_64 ARCH_FLATMEM_ENABLE depends on X86_32 ARCH_DISCONTIGMEM_ENABLE depends on X86_32 So X86_64_ACPI_NUMA implies SPARSEMEM, right? So we can mark the stack of functions __init for !SPARSEMEM, but we must mark them __meminit for SPARSEMEM configurations. This is ok, because then the calls to alloc_bootmem_node are also avoided. Compile-tested on: silly minimal config defconfig x86_32 defconfig x86_64 defconfig x86_64 -HIBERNATION +MEMORY_HOTPLUG Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm> Reviewed-by: Sam Ravnborg <sam@ravnborg.org> Acked-by: Geert Uytterhoeven <geert@linux-m68k.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-23 23:24:06 +00:00
static void __paginginit free_area_init_core(struct pglist_data *pgdat,
unsigned long *zones_size, unsigned long *zholes_size)
{
enum zone_type j;
int nid = pgdat->node_id;
unsigned long zone_start_pfn = pgdat->node_start_pfn;
int ret;
pgdat_resize_init(pgdat);
pgdat->nr_zones = 0;
init_waitqueue_head(&pgdat->kswapd_wait);
pgdat->kswapd_max_order = 0;
pgdat_page_cgroup_init(pgdat);
for (j = 0; j < MAX_NR_ZONES; j++) {
struct zone *zone = pgdat->node_zones + j;
unsigned long size, realsize, memmap_pages;
enum lru_list l;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
size = zone_spanned_pages_in_node(nid, j, zones_size);
realsize = size - zone_absent_pages_in_node(nid, j,
zholes_size);
/*
* Adjust realsize so that it accounts for how much memory
* is used by this zone for memmap. This affects the watermark
* and per-cpu initialisations
*/
memmap_pages =
PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
if (realsize >= memmap_pages) {
realsize -= memmap_pages;
if (memmap_pages)
printk(KERN_DEBUG
" %s zone: %lu pages used for memmap\n",
zone_names[j], memmap_pages);
} else
printk(KERN_WARNING
" %s zone: %lu pages exceeds realsize %lu\n",
zone_names[j], memmap_pages, realsize);
[PATCH] optional ZONE_DMA: deal with cases of ZONE_DMA meaning the first zone This patchset follows up on the earlier work in Andrew's tree to reduce the number of zones. The patches allow to go to a minimum of 2 zones. This one allows also to make ZONE_DMA optional and therefore the number of zones can be reduced to one. ZONE_DMA is usually used for ISA DMA devices. There are a number of reasons why we would not want to have ZONE_DMA 1. Some arches do not need ZONE_DMA at all. 2. With the advent of IOMMUs DMA zones are no longer needed. The necessity of DMA zones may drastically be reduced in the future. This patchset allows a compilation of a kernel without that overhead. 3. Devices that require ISA DMA get rare these days. All my systems do not have any need for ISA DMA. 4. The presence of an additional zone unecessarily complicates VM operations because it must be scanned and balancing logic must operate on its. 5. With only ZONE_NORMAL one can reach the situation where we have only one zone. This will allow the unrolling of many loops in the VM and allows the optimization of varous code paths in the VM. 6. Having only a single zone in a NUMA system results in a 1-1 correspondence between nodes and zones. Various additional optimizations to critical VM paths become possible. Many systems today can operate just fine with a single zone. If you look at what is in ZONE_DMA then one usually sees that nothing uses it. The DMA slabs are empty (Some arches use ZONE_DMA instead of ZONE_NORMAL, then ZONE_NORMAL will be empty instead). On all of my systems (i386, x86_64, ia64) ZONE_DMA is completely empty. Why constantly look at an empty zone in /proc/zoneinfo and empty slab in /proc/slabinfo? Non i386 also frequently have no need for ZONE_DMA and zones stay empty. The patchset was tested on i386 (UP / SMP), x86_64 (UP, NUMA) and ia64 (NUMA). The RFC posted earlier (see http://marc.theaimsgroup.com/?l=linux-kernel&m=115231723513008&w=2) had lots of #ifdefs in them. An effort has been made to minize the number of #ifdefs and make this as compact as possible. The job was made much easier by the ongoing efforts of others to extract common arch specific functionality. I have been running this for awhile now on my desktop and finally Linux is using all my available RAM instead of leaving the 16MB in ZONE_DMA untouched: christoph@pentium940:~$ cat /proc/zoneinfo Node 0, zone Normal pages free 4435 min 1448 low 1810 high 2172 active 241786 inactive 210170 scanned 0 (a: 0 i: 0) spanned 524224 present 524224 nr_anon_pages 61680 nr_mapped 14271 nr_file_pages 390264 nr_slab_reclaimable 27564 nr_slab_unreclaimable 1793 nr_page_table_pages 449 nr_dirty 39 nr_writeback 0 nr_unstable 0 nr_bounce 0 cpu: 0 pcp: 0 count: 156 high: 186 batch: 31 cpu: 0 pcp: 1 count: 9 high: 62 batch: 15 vm stats threshold: 20 cpu: 1 pcp: 0 count: 177 high: 186 batch: 31 cpu: 1 pcp: 1 count: 12 high: 62 batch: 15 vm stats threshold: 20 all_unreclaimable: 0 prev_priority: 12 temp_priority: 12 start_pfn: 0 This patch: In two places in the VM we use ZONE_DMA to refer to the first zone. If ZONE_DMA is optional then other zones may be first. So simply replace ZONE_DMA with zone 0. This also fixes ZONETABLE_PGSHIFT. If we have only a single zone then ZONES_PGSHIFT may become 0 because there is no need anymore to encode the zone number related to a pgdat. However, we still need a zonetable to index all the zones for each node if this is a NUMA system. Therefore define ZONETABLE_SHIFT unconditionally as the offset of the ZONE field in page flags. [apw@shadowen.org: fix mismerge] Acked-by: Christoph Hellwig <hch@infradead.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: Matthew Wilcox <willy@debian.org> Cc: James Bottomley <James.Bottomley@steeleye.com> Cc: Paul Mundt <lethal@linux-sh.org> Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-10 09:43:07 +00:00
/* Account for reserved pages */
if (j == 0 && realsize > dma_reserve) {
realsize -= dma_reserve;
printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
[PATCH] optional ZONE_DMA: deal with cases of ZONE_DMA meaning the first zone This patchset follows up on the earlier work in Andrew's tree to reduce the number of zones. The patches allow to go to a minimum of 2 zones. This one allows also to make ZONE_DMA optional and therefore the number of zones can be reduced to one. ZONE_DMA is usually used for ISA DMA devices. There are a number of reasons why we would not want to have ZONE_DMA 1. Some arches do not need ZONE_DMA at all. 2. With the advent of IOMMUs DMA zones are no longer needed. The necessity of DMA zones may drastically be reduced in the future. This patchset allows a compilation of a kernel without that overhead. 3. Devices that require ISA DMA get rare these days. All my systems do not have any need for ISA DMA. 4. The presence of an additional zone unecessarily complicates VM operations because it must be scanned and balancing logic must operate on its. 5. With only ZONE_NORMAL one can reach the situation where we have only one zone. This will allow the unrolling of many loops in the VM and allows the optimization of varous code paths in the VM. 6. Having only a single zone in a NUMA system results in a 1-1 correspondence between nodes and zones. Various additional optimizations to critical VM paths become possible. Many systems today can operate just fine with a single zone. If you look at what is in ZONE_DMA then one usually sees that nothing uses it. The DMA slabs are empty (Some arches use ZONE_DMA instead of ZONE_NORMAL, then ZONE_NORMAL will be empty instead). On all of my systems (i386, x86_64, ia64) ZONE_DMA is completely empty. Why constantly look at an empty zone in /proc/zoneinfo and empty slab in /proc/slabinfo? Non i386 also frequently have no need for ZONE_DMA and zones stay empty. The patchset was tested on i386 (UP / SMP), x86_64 (UP, NUMA) and ia64 (NUMA). The RFC posted earlier (see http://marc.theaimsgroup.com/?l=linux-kernel&m=115231723513008&w=2) had lots of #ifdefs in them. An effort has been made to minize the number of #ifdefs and make this as compact as possible. The job was made much easier by the ongoing efforts of others to extract common arch specific functionality. I have been running this for awhile now on my desktop and finally Linux is using all my available RAM instead of leaving the 16MB in ZONE_DMA untouched: christoph@pentium940:~$ cat /proc/zoneinfo Node 0, zone Normal pages free 4435 min 1448 low 1810 high 2172 active 241786 inactive 210170 scanned 0 (a: 0 i: 0) spanned 524224 present 524224 nr_anon_pages 61680 nr_mapped 14271 nr_file_pages 390264 nr_slab_reclaimable 27564 nr_slab_unreclaimable 1793 nr_page_table_pages 449 nr_dirty 39 nr_writeback 0 nr_unstable 0 nr_bounce 0 cpu: 0 pcp: 0 count: 156 high: 186 batch: 31 cpu: 0 pcp: 1 count: 9 high: 62 batch: 15 vm stats threshold: 20 cpu: 1 pcp: 0 count: 177 high: 186 batch: 31 cpu: 1 pcp: 1 count: 12 high: 62 batch: 15 vm stats threshold: 20 all_unreclaimable: 0 prev_priority: 12 temp_priority: 12 start_pfn: 0 This patch: In two places in the VM we use ZONE_DMA to refer to the first zone. If ZONE_DMA is optional then other zones may be first. So simply replace ZONE_DMA with zone 0. This also fixes ZONETABLE_PGSHIFT. If we have only a single zone then ZONES_PGSHIFT may become 0 because there is no need anymore to encode the zone number related to a pgdat. However, we still need a zonetable to index all the zones for each node if this is a NUMA system. Therefore define ZONETABLE_SHIFT unconditionally as the offset of the ZONE field in page flags. [apw@shadowen.org: fix mismerge] Acked-by: Christoph Hellwig <hch@infradead.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Andi Kleen <ak@suse.de> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: Matthew Wilcox <willy@debian.org> Cc: James Bottomley <James.Bottomley@steeleye.com> Cc: Paul Mundt <lethal@linux-sh.org> Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-10 09:43:07 +00:00
zone_names[0], dma_reserve);
}
if (!is_highmem_idx(j))
nr_kernel_pages += realsize;
nr_all_pages += realsize;
zone->spanned_pages = size;
zone->present_pages = realsize;
#ifdef CONFIG_NUMA
zone->node = nid;
zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
/ 100;
[PATCH] zone_reclaim: dynamic slab reclaim Currently one can enable slab reclaim by setting an explicit option in /proc/sys/vm/zone_reclaim_mode. Slab reclaim is then used as a final option if the freeing of unmapped file backed pages is not enough to free enough pages to allow a local allocation. However, that means that the slab can grow excessively and that most memory of a node may be used by slabs. We have had a case where a machine with 46GB of memory was using 40-42GB for slab. Zone reclaim was effective in dealing with pagecache pages. However, slab reclaim was only done during global reclaim (which is a bit rare on NUMA systems). This patch implements slab reclaim during zone reclaim. Zone reclaim occurs if there is a danger of an off node allocation. At that point we 1. Shrink the per node page cache if the number of pagecache pages is more than min_unmapped_ratio percent of pages in a zone. 2. Shrink the slab cache if the number of the nodes reclaimable slab pages (patch depends on earlier one that implements that counter) are more than min_slab_ratio (a new /proc/sys/vm tunable). The shrinking of the slab cache is a bit problematic since it is not node specific. So we simply calculate what point in the slab we want to reach (current per node slab use minus the number of pages that neeed to be allocated) and then repeately run the global reclaim until that is unsuccessful or we have reached the limit. I hope we will have zone based slab reclaim at some point which will make that easier. The default for the min_slab_ratio is 5% Also remove the slab option from /proc/sys/vm/zone_reclaim_mode. [akpm@osdl.org: cleanups] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-26 06:31:52 +00:00
zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
#endif
zone->name = zone_names[j];
spin_lock_init(&zone->lock);
spin_lock_init(&zone->lru_lock);
zone_seqlock_init(zone);
zone->zone_pgdat = pgdat;
zone_pcp_init(zone);
memcg: fix get_scan_count() for small targets During memory reclaim we determine the number of pages to be scanned per zone as (anon + file) >> priority. Assume scan = (anon + file) >> priority. If scan < SWAP_CLUSTER_MAX, the scan will be skipped for this time and priority gets higher. This has some problems. 1. This increases priority as 1 without any scan. To do scan in this priority, amount of pages should be larger than 512M. If pages>>priority < SWAP_CLUSTER_MAX, it's recorded and scan will be batched, later. (But we lose 1 priority.) If memory size is below 16M, pages >> priority is 0 and no scan in DEF_PRIORITY forever. 2. If zone->all_unreclaimabe==true, it's scanned only when priority==0. So, x86's ZONE_DMA will never be recoverred until the user of pages frees memory by itself. 3. With memcg, the limit of memory can be small. When using small memcg, it gets priority < DEF_PRIORITY-2 very easily and need to call wait_iff_congested(). For doing scan before priorty=9, 64MB of memory should be used. Then, this patch tries to scan SWAP_CLUSTER_MAX of pages in force...when 1. the target is enough small. 2. it's kswapd or memcg reclaim. Then we can avoid rapid priority drop and may be able to recover all_unreclaimable in a small zones. And this patch removes nr_saved_scan. This will allow scanning in this priority even when pages >> priority is very small. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Ying Han <yinghan@google.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-26 23:25:34 +00:00
for_each_lru(l)
INIT_LIST_HEAD(&zone->lru[l].list);
zone->reclaim_stat.recent_rotated[0] = 0;
zone->reclaim_stat.recent_rotated[1] = 0;
zone->reclaim_stat.recent_scanned[0] = 0;
zone->reclaim_stat.recent_scanned[1] = 0;
[PATCH] zoned vm counters: basic ZVC (zoned vm counter) implementation Per zone counter infrastructure The counters that we currently have for the VM are split per processor. The processor however has not much to do with the zone these pages belong to. We cannot tell f.e. how many ZONE_DMA pages are dirty. So we are blind to potentially inbalances in the usage of memory in various zones. F.e. in a NUMA system we cannot tell how many pages are dirty on a particular node. If we knew then we could put measures into the VM to balance the use of memory between different zones and different nodes in a NUMA system. For example it would be possible to limit the dirty pages per node so that fast local memory is kept available even if a process is dirtying huge amounts of pages. Another example is zone reclaim. We do not know how many unmapped pages exist per zone. So we just have to try to reclaim. If it is not working then we pause and try again later. It would be better if we knew when it makes sense to reclaim unmapped pages from a zone. This patchset allows the determination of the number of unmapped pages per zone. We can remove the zone reclaim interval with the counters introduced here. Futhermore the ability to have various usage statistics available will allow the development of new NUMA balancing algorithms that may be able to improve the decision making in the scheduler of when to move a process to another node and hopefully will also enable automatic page migration through a user space program that can analyse the memory load distribution and then rebalance memory use in order to increase performance. The counter framework here implements differential counters for each processor in struct zone. The differential counters are consolidated when a threshold is exceeded (like done in the current implementation for nr_pageache), when slab reaping occurs or when a consolidation function is called. Consolidation uses atomic operations and accumulates counters per zone in the zone structure and also globally in the vm_stat array. VM functions can access the counts by simply indexing a global or zone specific array. The arrangement of counters in an array also simplifies processing when output has to be generated for /proc/*. Counters can be updated by calling inc/dec_zone_page_state or _inc/dec_zone_page_state analogous to *_page_state. The second group of functions can be called if it is known that interrupts are disabled. Special optimized increment and decrement functions are provided. These can avoid certain checks and use increment or decrement instructions that an architecture may provide. We also add a new CONFIG_DMA_IS_NORMAL that signifies that an architecture can do DMA to all memory and therefore ZONE_NORMAL will not be populated. This is only currently set for IA64 SGI SN2 and currently only affects node_page_state(). In the best case node_page_state can be reduced to retrieving a single counter for the one zone on the node. [akpm@osdl.org: cleanups] [akpm@osdl.org: export vm_stat[] for filesystems] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-30 08:55:33 +00:00
zap_zone_vm_stats(zone);
zone->flags = 0;
if (!size)
continue;
set_pageblock_order(pageblock_default_order());
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
setup_usemap(pgdat, zone, size);
ret = init_currently_empty_zone(zone, zone_start_pfn,
size, MEMMAP_EARLY);
BUG_ON(ret);
memory_hotplug: always initialize pageblock bitmap Trying to online a new memory section that was added via memory hotplug sometimes results in crashes when the new pages are added via __free_page. Reason for that is that the pageblock bitmap isn't initialized and hence contains random stuff. That means that get_pageblock_migratetype() returns also random stuff and therefore list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); in __free_one_page() tries to do a list_add to something that isn't even necessarily a list. This happens since 86051ca5eaf5e560113ec7673462804c54284456 ("mm: fix usemap initialization") which makes sure that the pageblock bitmap gets only initialized for pages present in a zone. Unfortunately for hot-added memory the zones "grow" after the memmap and the pageblock memmap have been initialized. Which means that the new pages have an unitialized bitmap. To solve this the calls to grow_zone_span() and grow_pgdat_span() are moved to __add_zone() just before the initialization happens. The patch also moves the two functions since __add_zone() is the only caller and I didn't want to add a forward declaration. Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-05-14 23:05:52 +00:00
memmap_init(size, nid, j, zone_start_pfn);
zone_start_pfn += size;
}
}
static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
{
/* Skip empty nodes */
if (!pgdat->node_spanned_pages)
return;
[PATCH] sparsemem memory model Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of mem_map[] is needed by discontiguous memory machines (like in the old CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually become a complete replacement. A significant advantage over DISCONTIGMEM is that it's completely separated from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA and DISCONTIG are often confused. Another advantage is that sparse doesn't require each NUMA node's ranges to be contiguous. It can handle overlapping ranges between nodes with no problems, where DISCONTIGMEM currently throws away that memory. Sparsemem uses an array to provide different pfn_to_page() translations for each SECTION_SIZE area of physical memory. This is what allows the mem_map[] to be chopped up. In order to do quick pfn_to_page() operations, the section number of the page is encoded in page->flags. Part of the sparsemem infrastructure enables sharing of these bits more dynamically (at compile-time) between the page_zone() and sparsemem operations. However, on 32-bit architectures, the number of bits is quite limited, and may require growing the size of the page->flags type in certain conditions. Several things might force this to occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of memory), an increase in the physical address space, or an increase in the number of used page->flags. One thing to note is that, once sparsemem is present, the NUMA node information no longer needs to be stored in the page->flags. It might provide speed increases on certain platforms and will be stored there if there is room. But, if out of room, an alternate (theoretically slower) mechanism is used. This patch introduces CONFIG_FLATMEM. It is used in almost all cases where there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM often have to compile out the same areas of code. Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Dave Hansen <haveblue@us.ibm.com> Signed-off-by: Martin Bligh <mbligh@aracnet.com> Signed-off-by: Adrian Bunk <bunk@stusta.de> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Bob Picco <bob.picco@hp.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:07:54 +00:00
#ifdef CONFIG_FLAT_NODE_MEM_MAP
/* ia64 gets its own node_mem_map, before this, without bootmem */
if (!pgdat->node_mem_map) {
unsigned long size, start, end;
[PATCH] sparsemem memory model Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of mem_map[] is needed by discontiguous memory machines (like in the old CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually become a complete replacement. A significant advantage over DISCONTIGMEM is that it's completely separated from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA and DISCONTIG are often confused. Another advantage is that sparse doesn't require each NUMA node's ranges to be contiguous. It can handle overlapping ranges between nodes with no problems, where DISCONTIGMEM currently throws away that memory. Sparsemem uses an array to provide different pfn_to_page() translations for each SECTION_SIZE area of physical memory. This is what allows the mem_map[] to be chopped up. In order to do quick pfn_to_page() operations, the section number of the page is encoded in page->flags. Part of the sparsemem infrastructure enables sharing of these bits more dynamically (at compile-time) between the page_zone() and sparsemem operations. However, on 32-bit architectures, the number of bits is quite limited, and may require growing the size of the page->flags type in certain conditions. Several things might force this to occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of memory), an increase in the physical address space, or an increase in the number of used page->flags. One thing to note is that, once sparsemem is present, the NUMA node information no longer needs to be stored in the page->flags. It might provide speed increases on certain platforms and will be stored there if there is room. But, if out of room, an alternate (theoretically slower) mechanism is used. This patch introduces CONFIG_FLATMEM. It is used in almost all cases where there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM often have to compile out the same areas of code. Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Dave Hansen <haveblue@us.ibm.com> Signed-off-by: Martin Bligh <mbligh@aracnet.com> Signed-off-by: Adrian Bunk <bunk@stusta.de> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Bob Picco <bob.picco@hp.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:07:54 +00:00
struct page *map;
/*
* The zone's endpoints aren't required to be MAX_ORDER
* aligned but the node_mem_map endpoints must be in order
* for the buddy allocator to function correctly.
*/
start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
end = ALIGN(end, MAX_ORDER_NR_PAGES);
size = (end - start) * sizeof(struct page);
map = alloc_remap(pgdat->node_id, size);
if (!map)
mm: use alloc_bootmem_node_nopanic() on really needed path Stefan found nobootmem does not work on his system that has only 8M of RAM. This causes an early panic: BIOS-provided physical RAM map: BIOS-88: 0000000000000000 - 000000000009f000 (usable) BIOS-88: 0000000000100000 - 0000000000840000 (usable) bootconsole [earlyser0] enabled Notice: NX (Execute Disable) protection missing in CPU or disabled in BIOS! DMI not present or invalid. last_pfn = 0x840 max_arch_pfn = 0x100000 init_memory_mapping: 0000000000000000-0000000000840000 8MB LOWMEM available. mapped low ram: 0 - 00840000 low ram: 0 - 00840000 Zone PFN ranges: DMA 0x00000001 -> 0x00001000 Normal empty Movable zone start PFN for each node early_node_map[2] active PFN ranges 0: 0x00000001 -> 0x0000009f 0: 0x00000100 -> 0x00000840 BUG: Int 6: CR2 (null) EDI c034663c ESI (null) EBP c0329f38 ESP c0329ef4 EBX c0346380 EDX 00000006 ECX ffffffff EAX fffffff4 err (null) EIP c0353191 CS c0320060 flg 00010082 Stack: (null) c030c533 000007cd (null) c030c533 00000001 (null) (null) 00000003 0000083f 00000018 00000002 00000002 c0329f6c c03534d6 (null) (null) 00000100 00000840 (null) c0329f64 00000001 00001000 (null) Pid: 0, comm: swapper Not tainted 2.6.36 #5 Call Trace: [<c02e3707>] ? 0xc02e3707 [<c035e6e5>] 0xc035e6e5 [<c0353191>] ? 0xc0353191 [<c03534d6>] 0xc03534d6 [<c034f1cd>] 0xc034f1cd [<c034a824>] 0xc034a824 [<c03513cb>] ? 0xc03513cb [<c0349432>] 0xc0349432 [<c0349066>] 0xc0349066 It turns out that we should ignore the low limit of 16M. Use alloc_bootmem_node_nopanic() in this case. [akpm@linux-foundation.org: less mess] Signed-off-by: Yinghai LU <yinghai@kernel.org> Reported-by: Stefan Hellermann <stefan@the2masters.de> Tested-by: Stefan Hellermann <stefan@the2masters.de> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@linux.intel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: <stable@kernel.org> [2.6.34+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-11 22:13:32 +00:00
map = alloc_bootmem_node_nopanic(pgdat, size);
pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
}
#ifndef CONFIG_NEED_MULTIPLE_NODES
/*
* With no DISCONTIG, the global mem_map is just set as node 0's
*/
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
if (pgdat == NODE_DATA(0)) {
mem_map = NODE_DATA(0)->node_mem_map;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
}
#endif
[PATCH] sparsemem memory model Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of mem_map[] is needed by discontiguous memory machines (like in the old CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually become a complete replacement. A significant advantage over DISCONTIGMEM is that it's completely separated from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA and DISCONTIG are often confused. Another advantage is that sparse doesn't require each NUMA node's ranges to be contiguous. It can handle overlapping ranges between nodes with no problems, where DISCONTIGMEM currently throws away that memory. Sparsemem uses an array to provide different pfn_to_page() translations for each SECTION_SIZE area of physical memory. This is what allows the mem_map[] to be chopped up. In order to do quick pfn_to_page() operations, the section number of the page is encoded in page->flags. Part of the sparsemem infrastructure enables sharing of these bits more dynamically (at compile-time) between the page_zone() and sparsemem operations. However, on 32-bit architectures, the number of bits is quite limited, and may require growing the size of the page->flags type in certain conditions. Several things might force this to occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of memory), an increase in the physical address space, or an increase in the number of used page->flags. One thing to note is that, once sparsemem is present, the NUMA node information no longer needs to be stored in the page->flags. It might provide speed increases on certain platforms and will be stored there if there is room. But, if out of room, an alternate (theoretically slower) mechanism is used. This patch introduces CONFIG_FLATMEM. It is used in almost all cases where there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM often have to compile out the same areas of code. Signed-off-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Dave Hansen <haveblue@us.ibm.com> Signed-off-by: Martin Bligh <mbligh@aracnet.com> Signed-off-by: Adrian Bunk <bunk@stusta.de> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Bob Picco <bob.picco@hp.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:07:54 +00:00
#endif /* CONFIG_FLAT_NODE_MEM_MAP */
}
void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
unsigned long node_start_pfn, unsigned long *zholes_size)
{
pg_data_t *pgdat = NODE_DATA(nid);
pgdat->node_id = nid;
pgdat->node_start_pfn = node_start_pfn;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
calculate_node_totalpages(pgdat, zones_size, zholes_size);
alloc_node_mem_map(pgdat);
#ifdef CONFIG_FLAT_NODE_MEM_MAP
printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
nid, (unsigned long)pgdat,
(unsigned long)pgdat->node_mem_map);
#endif
free_area_init_core(pgdat, zones_size, zholes_size);
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
#if MAX_NUMNODES > 1
/*
* Figure out the number of possible node ids.
*/
static void __init setup_nr_node_ids(void)
{
unsigned int node;
unsigned int highest = 0;
for_each_node_mask(node, node_possible_map)
highest = node;
nr_node_ids = highest + 1;
}
#else
static inline void setup_nr_node_ids(void)
{
}
#endif
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/**
* add_active_range - Register a range of PFNs backed by physical memory
* @nid: The node ID the range resides on
* @start_pfn: The start PFN of the available physical memory
* @end_pfn: The end PFN of the available physical memory
*
* These ranges are stored in an early_node_map[] and later used by
* free_area_init_nodes() to calculate zone sizes and holes. If the
* range spans a memory hole, it is up to the architecture to ensure
* the memory is not freed by the bootmem allocator. If possible
* the range being registered will be merged with existing ranges.
*/
void __init add_active_range(unsigned int nid, unsigned long start_pfn,
unsigned long end_pfn)
{
int i;
mminit_dprintk(MMINIT_TRACE, "memory_register",
"Entering add_active_range(%d, %#lx, %#lx) "
"%d entries of %d used\n",
nid, start_pfn, end_pfn,
nr_nodemap_entries, MAX_ACTIVE_REGIONS);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Merge with existing active regions if possible */
for (i = 0; i < nr_nodemap_entries; i++) {
if (early_node_map[i].nid != nid)
continue;
/* Skip if an existing region covers this new one */
if (start_pfn >= early_node_map[i].start_pfn &&
end_pfn <= early_node_map[i].end_pfn)
return;
/* Merge forward if suitable */
if (start_pfn <= early_node_map[i].end_pfn &&
end_pfn > early_node_map[i].end_pfn) {
early_node_map[i].end_pfn = end_pfn;
return;
}
/* Merge backward if suitable */
if (start_pfn < early_node_map[i].start_pfn &&
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
end_pfn >= early_node_map[i].start_pfn) {
early_node_map[i].start_pfn = start_pfn;
return;
}
}
/* Check that early_node_map is large enough */
if (i >= MAX_ACTIVE_REGIONS) {
printk(KERN_CRIT "More than %d memory regions, truncating\n",
MAX_ACTIVE_REGIONS);
return;
}
early_node_map[i].nid = nid;
early_node_map[i].start_pfn = start_pfn;
early_node_map[i].end_pfn = end_pfn;
nr_nodemap_entries = i + 1;
}
/**
* remove_active_range - Shrink an existing registered range of PFNs
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
* @nid: The node id the range is on that should be shrunk
* @start_pfn: The new PFN of the range
* @end_pfn: The new PFN of the range
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*
* i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
* The map is kept near the end physical page range that has already been
* registered. This function allows an arch to shrink an existing registered
* range.
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
unsigned long end_pfn)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
int i, j;
int removed = 0;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
nid, start_pfn, end_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Find the old active region end and shrink */
for_each_active_range_index_in_nid(i, nid) {
if (early_node_map[i].start_pfn >= start_pfn &&
early_node_map[i].end_pfn <= end_pfn) {
/* clear it */
early_node_map[i].start_pfn = 0;
early_node_map[i].end_pfn = 0;
removed = 1;
continue;
}
if (early_node_map[i].start_pfn < start_pfn &&
early_node_map[i].end_pfn > start_pfn) {
unsigned long temp_end_pfn = early_node_map[i].end_pfn;
early_node_map[i].end_pfn = start_pfn;
if (temp_end_pfn > end_pfn)
add_active_range(nid, end_pfn, temp_end_pfn);
continue;
}
if (early_node_map[i].start_pfn >= start_pfn &&
early_node_map[i].end_pfn > end_pfn &&
early_node_map[i].start_pfn < end_pfn) {
early_node_map[i].start_pfn = end_pfn;
continue;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
}
}
if (!removed)
return;
/* remove the blank ones */
for (i = nr_nodemap_entries - 1; i > 0; i--) {
if (early_node_map[i].nid != nid)
continue;
if (early_node_map[i].end_pfn)
continue;
/* we found it, get rid of it */
for (j = i; j < nr_nodemap_entries - 1; j++)
memcpy(&early_node_map[j], &early_node_map[j+1],
sizeof(early_node_map[j]));
j = nr_nodemap_entries - 1;
memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
nr_nodemap_entries--;
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
}
/**
* remove_all_active_ranges - Remove all currently registered regions
*
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
* During discovery, it may be found that a table like SRAT is invalid
* and an alternative discovery method must be used. This function removes
* all currently registered regions.
*/
void __init remove_all_active_ranges(void)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
memset(early_node_map, 0, sizeof(early_node_map));
nr_nodemap_entries = 0;
}
/* Compare two active node_active_regions */
static int __init cmp_node_active_region(const void *a, const void *b)
{
struct node_active_region *arange = (struct node_active_region *)a;
struct node_active_region *brange = (struct node_active_region *)b;
/* Done this way to avoid overflows */
if (arange->start_pfn > brange->start_pfn)
return 1;
if (arange->start_pfn < brange->start_pfn)
return -1;
return 0;
}
/* sort the node_map by start_pfn */
x86: Fix checking of SRAT when node 0 ram is not from 0 Found one system that boot from socket1 instead of socket0, SRAT get rejected... [ 0.000000] SRAT: Node 1 PXM 0 0-a0000 [ 0.000000] SRAT: Node 1 PXM 0 100000-80000000 [ 0.000000] SRAT: Node 1 PXM 0 100000000-2080000000 [ 0.000000] SRAT: Node 0 PXM 1 2080000000-4080000000 [ 0.000000] SRAT: Node 2 PXM 2 4080000000-6080000000 [ 0.000000] SRAT: Node 3 PXM 3 6080000000-8080000000 [ 0.000000] SRAT: Node 4 PXM 4 8080000000-a080000000 [ 0.000000] SRAT: Node 5 PXM 5 a080000000-c080000000 [ 0.000000] SRAT: Node 6 PXM 6 c080000000-e080000000 [ 0.000000] SRAT: Node 7 PXM 7 e080000000-10080000000 ... [ 0.000000] NUMA: Allocated memnodemap from 500000 - 701040 [ 0.000000] NUMA: Using 20 for the hash shift. [ 0.000000] Adding active range (0, 0x2080000, 0x4080000) 0 entries of 3200 used [ 0.000000] Adding active range (1, 0x0, 0x96) 1 entries of 3200 used [ 0.000000] Adding active range (1, 0x100, 0x7f750) 2 entries of 3200 used [ 0.000000] Adding active range (1, 0x100000, 0x2080000) 3 entries of 3200 used [ 0.000000] Adding active range (2, 0x4080000, 0x6080000) 4 entries of 3200 used [ 0.000000] Adding active range (3, 0x6080000, 0x8080000) 5 entries of 3200 used [ 0.000000] Adding active range (4, 0x8080000, 0xa080000) 6 entries of 3200 used [ 0.000000] Adding active range (5, 0xa080000, 0xc080000) 7 entries of 3200 used [ 0.000000] Adding active range (6, 0xc080000, 0xe080000) 8 entries of 3200 used [ 0.000000] Adding active range (7, 0xe080000, 0x10080000) 9 entries of 3200 used [ 0.000000] SRAT: PXMs only cover 917504MB of your 1048566MB e820 RAM. Not used. [ 0.000000] SRAT: SRAT not used. the early_node_map is not sorted because node0 with non zero start come first. so try to sort it right away after all regions are registered. also fixs refression by 8716273c (x86: Export srat physical topology) -v2: make it more solid to handle cross node case like node0 [0,4g), [8,12g) and node1 [4g, 8g), [12g, 16g) -v3: update comments. Reported-and-tested-by: Jens Axboe <jens.axboe@oracle.com> Signed-off-by: Yinghai Lu <yinghai@kernel.org> LKML-Reference: <4B2579D2.3010201@kernel.org> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2009-12-16 01:59:02 +00:00
void __init sort_node_map(void)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
sort(early_node_map, (size_t)nr_nodemap_entries,
sizeof(struct node_active_region),
cmp_node_active_region, NULL);
}
x86, numa: Implement pfn -> nid mapping granularity check SPARSEMEM w/o VMEMMAP and DISCONTIGMEM, both used only on 32bit, use sections array to map pfn to nid which is limited in granularity. If NUMA nodes are laid out such that the mapping cannot be accurate, boot will fail triggering BUG_ON() in mminit_verify_page_links(). On 32bit, it's 512MiB w/ PAE and SPARSEMEM. This seems to have been granular enough until commit 2706a0bf7b (x86, NUMA: Enable CONFIG_AMD_NUMA on 32bit too). Apparently, there is a machine which aligns NUMA nodes to 128MiB and has only AMD NUMA but not SRAT. This led to the following BUG_ON(). On node 0 totalpages: 2096615 DMA zone: 32 pages used for memmap DMA zone: 0 pages reserved DMA zone: 3927 pages, LIFO batch:0 Normal zone: 1740 pages used for memmap Normal zone: 220978 pages, LIFO batch:31 HighMem zone: 16405 pages used for memmap HighMem zone: 1853533 pages, LIFO batch:31 BUG: Int 6: CR2 (null) EDI (null) ESI 00000002 EBP 00000002 ESP c1543ecc EBX f2400000 EDX 00000006 ECX (null) EAX 00000001 err (null) EIP c16209aa CS 00000060 flg 00010002 Stack: f2400000 00220000 f7200800 c1620613 00220000 01000000 04400000 00238000 (null) f7200000 00000002 f7200b58 f7200800 c1620929 000375fe (null) f7200b80 c16395f0 00200a02 f7200a80 (null) 000375fe 00000002 (null) Pid: 0, comm: swapper Not tainted 2.6.39-rc5-00181-g2706a0b #17 Call Trace: [<c136b1e5>] ? early_fault+0x2e/0x2e [<c16209aa>] ? mminit_verify_page_links+0x12/0x42 [<c1620613>] ? memmap_init_zone+0xaf/0x10c [<c1620929>] ? free_area_init_node+0x2b9/0x2e3 [<c1607e99>] ? free_area_init_nodes+0x3f2/0x451 [<c1601d80>] ? paging_init+0x112/0x118 [<c15f578d>] ? setup_arch+0x791/0x82f [<c15f43d9>] ? start_kernel+0x6a/0x257 This patch implements node_map_pfn_alignment() which determines maximum internode alignment and update numa_register_memblks() to reject NUMA configuration if alignment exceeds the pfn -> nid mapping granularity of the memory model as determined by PAGES_PER_SECTION. This makes the problematic machine boot w/ flatmem by rejecting the NUMA config and provides protection against crazy NUMA configurations. Signed-off-by: Tejun Heo <tj@kernel.org> Link: http://lkml.kernel.org/r/20110712074534.GB2872@htj.dyndns.org LKML-Reference: <20110628174613.GP478@escobedo.osrc.amd.com> Reported-and-Tested-by: Hans Rosenfeld <hans.rosenfeld@amd.com> Cc: Conny Seidel <conny.seidel@amd.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-07-12 07:45:34 +00:00
/**
* node_map_pfn_alignment - determine the maximum internode alignment
*
* This function should be called after node map is populated and sorted.
* It calculates the maximum power of two alignment which can distinguish
* all the nodes.
*
* For example, if all nodes are 1GiB and aligned to 1GiB, the return value
* would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
* nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
* shifted, 1GiB is enough and this function will indicate so.
*
* This is used to test whether pfn -> nid mapping of the chosen memory
* model has fine enough granularity to avoid incorrect mapping for the
* populated node map.
*
* Returns the determined alignment in pfn's. 0 if there is no alignment
* requirement (single node).
*/
unsigned long __init node_map_pfn_alignment(void)
{
unsigned long accl_mask = 0, last_end = 0;
int last_nid = -1;
int i;
for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
int nid = early_node_map[i].nid;
unsigned long start = early_node_map[i].start_pfn;
unsigned long end = early_node_map[i].end_pfn;
unsigned long mask;
if (!start || last_nid < 0 || last_nid == nid) {
last_nid = nid;
last_end = end;
continue;
}
/*
* Start with a mask granular enough to pin-point to the
* start pfn and tick off bits one-by-one until it becomes
* too coarse to separate the current node from the last.
*/
mask = ~((1 << __ffs(start)) - 1);
while (mask && last_end <= (start & (mask << 1)))
mask <<= 1;
/* accumulate all internode masks */
accl_mask |= mask;
}
/* convert mask to number of pages */
return ~accl_mask + 1;
}
/* Find the lowest pfn for a node */
static unsigned long __init find_min_pfn_for_node(int nid)
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
{
int i;
unsigned long min_pfn = ULONG_MAX;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Assuming a sorted map, the first range found has the starting pfn */
for_each_active_range_index_in_nid(i, nid)
min_pfn = min(min_pfn, early_node_map[i].start_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
if (min_pfn == ULONG_MAX) {
printk(KERN_WARNING
"Could not find start_pfn for node %d\n", nid);
return 0;
}
return min_pfn;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
}
/**
* find_min_pfn_with_active_regions - Find the minimum PFN registered
*
* It returns the minimum PFN based on information provided via
* add_active_range().
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*/
unsigned long __init find_min_pfn_with_active_regions(void)
{
return find_min_pfn_for_node(MAX_NUMNODES);
}
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
/*
* early_calculate_totalpages()
* Sum pages in active regions for movable zone.
* Populate N_HIGH_MEMORY for calculating usable_nodes.
*/
static unsigned long __init early_calculate_totalpages(void)
{
int i;
unsigned long totalpages = 0;
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
for (i = 0; i < nr_nodemap_entries; i++) {
unsigned long pages = early_node_map[i].end_pfn -
early_node_map[i].start_pfn;
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
totalpages += pages;
if (pages)
node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
}
return totalpages;
}
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/*
* Find the PFN the Movable zone begins in each node. Kernel memory
* is spread evenly between nodes as long as the nodes have enough
* memory. When they don't, some nodes will have more kernelcore than
* others
*/
static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
{
int i, nid;
unsigned long usable_startpfn;
unsigned long kernelcore_node, kernelcore_remaining;
/* save the state before borrow the nodemask */
nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
unsigned long totalpages = early_calculate_totalpages();
int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/*
* If movablecore was specified, calculate what size of
* kernelcore that corresponds so that memory usable for
* any allocation type is evenly spread. If both kernelcore
* and movablecore are specified, then the value of kernelcore
* will be used for required_kernelcore if it's greater than
* what movablecore would have allowed.
*/
if (required_movablecore) {
unsigned long corepages;
/*
* Round-up so that ZONE_MOVABLE is at least as large as what
* was requested by the user
*/
required_movablecore =
roundup(required_movablecore, MAX_ORDER_NR_PAGES);
corepages = totalpages - required_movablecore;
required_kernelcore = max(required_kernelcore, corepages);
}
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/* If kernelcore was not specified, there is no ZONE_MOVABLE */
if (!required_kernelcore)
goto out;
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
find_usable_zone_for_movable();
usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
restart:
/* Spread kernelcore memory as evenly as possible throughout nodes */
kernelcore_node = required_kernelcore / usable_nodes;
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
for_each_node_state(nid, N_HIGH_MEMORY) {
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/*
* Recalculate kernelcore_node if the division per node
* now exceeds what is necessary to satisfy the requested
* amount of memory for the kernel
*/
if (required_kernelcore < kernelcore_node)
kernelcore_node = required_kernelcore / usable_nodes;
/*
* As the map is walked, we track how much memory is usable
* by the kernel using kernelcore_remaining. When it is
* 0, the rest of the node is usable by ZONE_MOVABLE
*/
kernelcore_remaining = kernelcore_node;
/* Go through each range of PFNs within this node */
for_each_active_range_index_in_nid(i, nid) {
unsigned long start_pfn, end_pfn;
unsigned long size_pages;
start_pfn = max(early_node_map[i].start_pfn,
zone_movable_pfn[nid]);
end_pfn = early_node_map[i].end_pfn;
if (start_pfn >= end_pfn)
continue;
/* Account for what is only usable for kernelcore */
if (start_pfn < usable_startpfn) {
unsigned long kernel_pages;
kernel_pages = min(end_pfn, usable_startpfn)
- start_pfn;
kernelcore_remaining -= min(kernel_pages,
kernelcore_remaining);
required_kernelcore -= min(kernel_pages,
required_kernelcore);
/* Continue if range is now fully accounted */
if (end_pfn <= usable_startpfn) {
/*
* Push zone_movable_pfn to the end so
* that if we have to rebalance
* kernelcore across nodes, we will
* not double account here
*/
zone_movable_pfn[nid] = end_pfn;
continue;
}
start_pfn = usable_startpfn;
}
/*
* The usable PFN range for ZONE_MOVABLE is from
* start_pfn->end_pfn. Calculate size_pages as the
* number of pages used as kernelcore
*/
size_pages = end_pfn - start_pfn;
if (size_pages > kernelcore_remaining)
size_pages = kernelcore_remaining;
zone_movable_pfn[nid] = start_pfn + size_pages;
/*
* Some kernelcore has been met, update counts and
* break if the kernelcore for this node has been
* satisified
*/
required_kernelcore -= min(required_kernelcore,
size_pages);
kernelcore_remaining -= size_pages;
if (!kernelcore_remaining)
break;
}
}
/*
* If there is still required_kernelcore, we do another pass with one
* less node in the count. This will push zone_movable_pfn[nid] further
* along on the nodes that still have memory until kernelcore is
* satisified
*/
usable_nodes--;
if (usable_nodes && required_kernelcore > usable_nodes)
goto restart;
/* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
for (nid = 0; nid < MAX_NUMNODES; nid++)
zone_movable_pfn[nid] =
roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
out:
/* restore the node_state */
node_states[N_HIGH_MEMORY] = saved_node_state;
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
}
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
/* Any regular memory on that node ? */
static void check_for_regular_memory(pg_data_t *pgdat)
{
#ifdef CONFIG_HIGHMEM
enum zone_type zone_type;
for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
struct zone *zone = &pgdat->node_zones[zone_type];
if (zone->present_pages)
node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
}
#endif
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/**
* free_area_init_nodes - Initialise all pg_data_t and zone data
* @max_zone_pfn: an array of max PFNs for each zone
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
*
* This will call free_area_init_node() for each active node in the system.
* Using the page ranges provided by add_active_range(), the size of each
* zone in each node and their holes is calculated. If the maximum PFN
* between two adjacent zones match, it is assumed that the zone is empty.
* For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
* that arch_max_dma32_pfn has no pages. It is also assumed that a zone
* starts where the previous one ended. For example, ZONE_DMA32 starts
* at arch_max_dma_pfn.
*/
void __init free_area_init_nodes(unsigned long *max_zone_pfn)
{
unsigned long nid;
int i;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Sort early_node_map as initialisation assumes it is sorted */
sort_node_map();
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Record where the zone boundaries are */
memset(arch_zone_lowest_possible_pfn, 0,
sizeof(arch_zone_lowest_possible_pfn));
memset(arch_zone_highest_possible_pfn, 0,
sizeof(arch_zone_highest_possible_pfn));
arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
for (i = 1; i < MAX_NR_ZONES; i++) {
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
if (i == ZONE_MOVABLE)
continue;
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
arch_zone_lowest_possible_pfn[i] =
arch_zone_highest_possible_pfn[i-1];
arch_zone_highest_possible_pfn[i] =
max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
}
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
/* Find the PFNs that ZONE_MOVABLE begins at in each node */
memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
find_zone_movable_pfns_for_nodes(zone_movable_pfn);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Print out the zone ranges */
printk("Zone PFN ranges:\n");
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
for (i = 0; i < MAX_NR_ZONES; i++) {
if (i == ZONE_MOVABLE)
continue;
printk(" %-8s ", zone_names[i]);
if (arch_zone_lowest_possible_pfn[i] ==
arch_zone_highest_possible_pfn[i])
printk("empty\n");
else
printk("%0#10lx -> %0#10lx\n",
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
arch_zone_lowest_possible_pfn[i],
arch_zone_highest_possible_pfn[i]);
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
}
/* Print out the PFNs ZONE_MOVABLE begins at in each node */
printk("Movable zone start PFN for each node\n");
for (i = 0; i < MAX_NUMNODES; i++) {
if (zone_movable_pfn[i])
printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
}
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
/* Print out the early_node_map[] */
printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
for (i = 0; i < nr_nodemap_entries; i++)
printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
early_node_map[i].start_pfn,
early_node_map[i].end_pfn);
/* Initialise every node */
mminit_verify_pageflags_layout();
setup_nr_node_ids();
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
for_each_online_node(nid) {
pg_data_t *pgdat = NODE_DATA(nid);
free_area_init_node(nid, NULL,
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
find_min_pfn_for_node(nid), NULL);
memoryless nodes: fixup uses of node_online_map in generic code Here's a cut at fixing up uses of the online node map in generic code. mm/shmem.c:shmem_parse_mpol() Ensure nodelist is subset of nodes with memory. Use node_states[N_HIGH_MEMORY] as default for missing nodelist for interleave policy. mm/shmem.c:shmem_fill_super() initialize policy_nodes to node_states[N_HIGH_MEMORY] mm/page-writeback.c:highmem_dirtyable_memory() sum over nodes with memory mm/page_alloc.c:zlc_setup() allowednodes - use nodes with memory. mm/page_alloc.c:default_zonelist_order() average over nodes with memory. mm/page_alloc.c:find_next_best_node() skip nodes w/o memory. N_HIGH_MEMORY state mask may not be initialized at this time, unless we want to depend on early_calculate_totalpages() [see below]. Will ZONE_MOVABLE ever be configurable? mm/page_alloc.c:find_zone_movable_pfns_for_nodes() spread kernelcore over nodes with memory. This required calling early_calculate_totalpages() unconditionally, and populating N_HIGH_MEMORY node state therein from nodes in the early_node_map[]. If we can depend on this, we can eliminate the population of N_HIGH_MEMORY mask from __build_all_zonelists() and use the N_HIGH_MEMORY mask in find_next_best_node(). mm/mempolicy.c:mpol_check_policy() Ensure nodes specified for policy are subset of nodes with memory. [akpm@linux-foundation.org: fix warnings] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Acked-by: Christoph Lameter <clameter@sgi.com> Cc: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:39 +00:00
/* Any memory on that node */
if (pgdat->node_present_pages)
node_set_state(nid, N_HIGH_MEMORY);
check_for_regular_memory(pgdat);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
}
}
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
static int __init cmdline_parse_core(char *p, unsigned long *core)
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
{
unsigned long long coremem;
if (!p)
return -EINVAL;
coremem = memparse(p, &p);
*core = coremem >> PAGE_SHIFT;
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
/* Paranoid check that UL is enough for the coremem value */
Create the ZONE_MOVABLE zone The following 8 patches against 2.6.20-mm2 create a zone called ZONE_MOVABLE that is only usable by allocations that specify both __GFP_HIGHMEM and __GFP_MOVABLE. This has the effect of keeping all non-movable pages within a single memory partition while allowing movable allocations to be satisfied from either partition. The patches may be applied with the list-based anti-fragmentation patches that groups pages together based on mobility. The size of the zone is determined by a kernelcore= parameter specified at boot-time. This specifies how much memory is usable by non-movable allocations and the remainder is used for ZONE_MOVABLE. Any range of pages within ZONE_MOVABLE can be released by migrating the pages or by reclaiming. When selecting a zone to take pages from for ZONE_MOVABLE, there are two things to consider. First, only memory from the highest populated zone is used for ZONE_MOVABLE. On the x86, this is probably going to be ZONE_HIGHMEM but it would be ZONE_DMA on ppc64 or possibly ZONE_DMA32 on x86_64. Second, the amount of memory usable by the kernel will be spread evenly throughout NUMA nodes where possible. If the nodes are not of equal size, the amount of memory usable by the kernel on some nodes may be greater than others. By default, the zone is not as useful for hugetlb allocations because they are pinned and non-migratable (currently at least). A sysctl is provided that allows huge pages to be allocated from that zone. This means that the huge page pool can be resized to the size of ZONE_MOVABLE during the lifetime of the system assuming that pages are not mlocked. Despite huge pages being non-movable, we do not introduce additional external fragmentation of note as huge pages are always the largest contiguous block we care about. Credit goes to Andy Whitcroft for catching a large variety of problems during review of the patches. This patch creates an additional zone, ZONE_MOVABLE. This zone is only usable by allocations which specify both __GFP_HIGHMEM and __GFP_MOVABLE. Hot-added memory continues to be placed in their existing destination as there is no mechanism to redirect them to a specific zone. [y-goto@jp.fujitsu.com: Fix section mismatch of memory hotplug related code] [akpm@linux-foundation.org: various fixes] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com> Cc: William Lee Irwin III <wli@holomorphy.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-17 11:03:12 +00:00
WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
return 0;
}
/*
* kernelcore=size sets the amount of memory for use for allocations that
* cannot be reclaimed or migrated.
*/
static int __init cmdline_parse_kernelcore(char *p)
{
return cmdline_parse_core(p, &required_kernelcore);
}
/*
* movablecore=size sets the amount of memory for use for allocations that
* can be reclaimed or migrated.
*/
static int __init cmdline_parse_movablecore(char *p)
{
return cmdline_parse_core(p, &required_movablecore);
}
early_param("kernelcore", cmdline_parse_kernelcore);
early_param("movablecore", cmdline_parse_movablecore);
[PATCH] Introduce mechanism for registering active regions of memory At a basic level, architectures define structures to record where active ranges of page frames are located. Once located, the code to calculate zone sizes and holes in each architecture is very similar. Some of this zone and hole sizing code is difficult to read for no good reason. This set of patches eliminates the similar-looking architecture-specific code. The patches introduce a mechanism where architectures register where the active ranges of page frames are with add_active_range(). When all areas have been discovered, free_area_init_nodes() is called to initialise the pgdat and zones. The zone sizes and holes are then calculated in an architecture independent manner. Patch 1 introduces the mechanism for registering and initialising PFN ranges Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable. It adjusts the watermarks slightly Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig, gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based machine. These were on patches against 2.6.17-rc1 and release 3 of these patches but there have been no ia64-changes since release 3. There are differences in the zone sizes for x86_64 as the arch-specific code for x86_64 accounts the kernel image and the starting mem_maps as memory holes but the architecture-independent code accounts the memory as present. The big benefit of this set of patches is a sizable reduction of architecture-specific code, some of which is very hairy. There should be a greater reduction when other architectures use the same mechanisms for zone and hole sizing but I lack the hardware to test on. Additional credit; Dave Hansen for the initial suggestion and comments on early patches Andy Whitcroft for reviewing early versions and catching numerous errors Tony Luck for testing and debugging on IA64 Bob Picco for fixing bugs related to pfn registration, reviewing a number of patch revisions, providing a number of suggestions on future direction and testing heavily Jack Steiner and Robin Holt for testing on IA64 and clarifying issues related to memory holes Yasunori for testing on IA64 Andi Kleen for reviewing and feeding back about x86_64 Christian Kujau for providing valuable information related to ACPI problems on x86_64 and testing potential fixes This patch: Define the structure to represent an active range of page frames within a node in an architecture independent manner. Architectures are expected to register active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call free_area_init_nodes() passing the PFNs of the end of each zone. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Bob Picco <bob.picco@hp.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Andy Whitcroft <apw@shadowen.org> Cc: Andi Kleen <ak@muc.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Keith Mannthey" <kmannth@gmail.com> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Yasunori Goto <y-goto@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 08:49:43 +00:00
#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
/**
* set_dma_reserve - set the specified number of pages reserved in the first zone
* @new_dma_reserve: The number of pages to mark reserved
*
* The per-cpu batchsize and zone watermarks are determined by present_pages.
* In the DMA zone, a significant percentage may be consumed by kernel image
* and other unfreeable allocations which can skew the watermarks badly. This
* function may optionally be used to account for unfreeable pages in the
* first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
* smaller per-cpu batchsize.
*/
void __init set_dma_reserve(unsigned long new_dma_reserve)
{
dma_reserve = new_dma_reserve;
}
void __init free_area_init(unsigned long *zones_size)
{
free_area_init_node(0, zones_size,
__pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
}
static int page_alloc_cpu_notify(struct notifier_block *self,
unsigned long action, void *hcpu)
{
int cpu = (unsigned long)hcpu;
if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
drain_pages(cpu);
/*
* Spill the event counters of the dead processor
* into the current processors event counters.
* This artificially elevates the count of the current
* processor.
*/
[PATCH] Light weight event counters The remaining counters in page_state after the zoned VM counter patches have been applied are all just for show in /proc/vmstat. They have no essential function for the VM. We use a simple increment of per cpu variables. In order to avoid the most severe races we disable preempt. Preempt does not prevent the race between an increment and an interrupt handler incrementing the same statistics counter. However, that race is exceedingly rare, we may only loose one increment or so and there is no requirement (at least not in kernel) that the vm event counters have to be accurate. In the non preempt case this results in a simple increment for each counter. For many architectures this will be reduced by the compiler to a single instruction. This single instruction is atomic for i386 and x86_64. And therefore even the rare race condition in an interrupt is avoided for both architectures in most cases. The patchset also adds an off switch for embedded systems that allows a building of linux kernels without these counters. The implementation of these counters is through inline code that hopefully results in only a single instruction increment instruction being emitted (i386, x86_64) or in the increment being hidden though instruction concurrency (EPIC architectures such as ia64 can get that done). Benefits: - VM event counter operations usually reduce to a single inline instruction on i386 and x86_64. - No interrupt disable, only preempt disable for the preempt case. Preempt disable can also be avoided by moving the counter into a spinlock. - Handling is similar to zoned VM counters. - Simple and easily extendable. - Can be omitted to reduce memory use for embedded use. References: RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=113512330605497&w=2 RFC http://marc.theaimsgroup.com/?l=linux-kernel&m=114988082814934&w=2 local_t http://marc.theaimsgroup.com/?l=linux-kernel&m=114991748606690&w=2 V2 http://marc.theaimsgroup.com/?t=115014808400007&r=1&w=2 V3 http://marc.theaimsgroup.com/?l=linux-kernel&m=115024767022346&w=2 V4 http://marc.theaimsgroup.com/?l=linux-kernel&m=115047968808926&w=2 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-30 08:55:45 +00:00
vm_events_fold_cpu(cpu);
/*
* Zero the differential counters of the dead processor
* so that the vm statistics are consistent.
*
* This is only okay since the processor is dead and cannot
* race with what we are doing.
*/
[PATCH] zoned vm counters: basic ZVC (zoned vm counter) implementation Per zone counter infrastructure The counters that we currently have for the VM are split per processor. The processor however has not much to do with the zone these pages belong to. We cannot tell f.e. how many ZONE_DMA pages are dirty. So we are blind to potentially inbalances in the usage of memory in various zones. F.e. in a NUMA system we cannot tell how many pages are dirty on a particular node. If we knew then we could put measures into the VM to balance the use of memory between different zones and different nodes in a NUMA system. For example it would be possible to limit the dirty pages per node so that fast local memory is kept available even if a process is dirtying huge amounts of pages. Another example is zone reclaim. We do not know how many unmapped pages exist per zone. So we just have to try to reclaim. If it is not working then we pause and try again later. It would be better if we knew when it makes sense to reclaim unmapped pages from a zone. This patchset allows the determination of the number of unmapped pages per zone. We can remove the zone reclaim interval with the counters introduced here. Futhermore the ability to have various usage statistics available will allow the development of new NUMA balancing algorithms that may be able to improve the decision making in the scheduler of when to move a process to another node and hopefully will also enable automatic page migration through a user space program that can analyse the memory load distribution and then rebalance memory use in order to increase performance. The counter framework here implements differential counters for each processor in struct zone. The differential counters are consolidated when a threshold is exceeded (like done in the current implementation for nr_pageache), when slab reaping occurs or when a consolidation function is called. Consolidation uses atomic operations and accumulates counters per zone in the zone structure and also globally in the vm_stat array. VM functions can access the counts by simply indexing a global or zone specific array. The arrangement of counters in an array also simplifies processing when output has to be generated for /proc/*. Counters can be updated by calling inc/dec_zone_page_state or _inc/dec_zone_page_state analogous to *_page_state. The second group of functions can be called if it is known that interrupts are disabled. Special optimized increment and decrement functions are provided. These can avoid certain checks and use increment or decrement instructions that an architecture may provide. We also add a new CONFIG_DMA_IS_NORMAL that signifies that an architecture can do DMA to all memory and therefore ZONE_NORMAL will not be populated. This is only currently set for IA64 SGI SN2 and currently only affects node_page_state(). In the best case node_page_state can be reduced to retrieving a single counter for the one zone on the node. [akpm@osdl.org: cleanups] [akpm@osdl.org: export vm_stat[] for filesystems] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-30 08:55:33 +00:00
refresh_cpu_vm_stats(cpu);
}
return NOTIFY_OK;
}
void __init page_alloc_init(void)
{
hotcpu_notifier(page_alloc_cpu_notify, 0);
}
[PATCH] overcommit: add calculate_totalreserve_pages() These patches are an enhancement of OVERCOMMIT_GUESS algorithm in __vm_enough_memory(). - why the kernel needed patching When the kernel can't allocate anonymous pages in practice, currnet OVERCOMMIT_GUESS could return success. This implementation might be the cause of oom kill in memory pressure situation. If the Linux runs with page reservation features like /proc/sys/vm/lowmem_reserve_ratio and without swap region, I think the oom kill occurs easily. - the overall design approach in the patch When the OVERCOMMET_GUESS algorithm calculates number of free pages, the reserved free pages are regarded as non-free pages. This change helps to avoid the pitfall that the number of free pages become less than the number which the kernel tries to keep free. - testing results I tested the patches using my test kernel module. If the patches aren't applied to the kernel, __vm_enough_memory() returns success in the situation but autual page allocation is failed. On the other hand, if the patches are applied to the kernel, memory allocation failure is avoided since __vm_enough_memory() returns failure in the situation. I checked that on i386 SMP 16GB memory machine. I haven't tested on nommu environment currently. This patch adds totalreserve_pages for __vm_enough_memory(). Calculate_totalreserve_pages() checks maximum lowmem_reserve pages and pages_high in each zone. Finally, the function stores the sum of each zone to totalreserve_pages. The totalreserve_pages is calculated when the VM is initilized. And the variable is updated when /proc/sys/vm/lowmem_reserve_raito or /proc/sys/vm/min_free_kbytes are changed. Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-11 05:52:59 +00:00
/*
* calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
* or min_free_kbytes changes.
*/
static void calculate_totalreserve_pages(void)
{
struct pglist_data *pgdat;
unsigned long reserve_pages = 0;
enum zone_type i, j;
[PATCH] overcommit: add calculate_totalreserve_pages() These patches are an enhancement of OVERCOMMIT_GUESS algorithm in __vm_enough_memory(). - why the kernel needed patching When the kernel can't allocate anonymous pages in practice, currnet OVERCOMMIT_GUESS could return success. This implementation might be the cause of oom kill in memory pressure situation. If the Linux runs with page reservation features like /proc/sys/vm/lowmem_reserve_ratio and without swap region, I think the oom kill occurs easily. - the overall design approach in the patch When the OVERCOMMET_GUESS algorithm calculates number of free pages, the reserved free pages are regarded as non-free pages. This change helps to avoid the pitfall that the number of free pages become less than the number which the kernel tries to keep free. - testing results I tested the patches using my test kernel module. If the patches aren't applied to the kernel, __vm_enough_memory() returns success in the situation but autual page allocation is failed. On the other hand, if the patches are applied to the kernel, memory allocation failure is avoided since __vm_enough_memory() returns failure in the situation. I checked that on i386 SMP 16GB memory machine. I haven't tested on nommu environment currently. This patch adds totalreserve_pages for __vm_enough_memory(). Calculate_totalreserve_pages() checks maximum lowmem_reserve pages and pages_high in each zone. Finally, the function stores the sum of each zone to totalreserve_pages. The totalreserve_pages is calculated when the VM is initilized. And the variable is updated when /proc/sys/vm/lowmem_reserve_raito or /proc/sys/vm/min_free_kbytes are changed. Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-11 05:52:59 +00:00
for_each_online_pgdat(pgdat) {
for (i = 0; i < MAX_NR_ZONES; i++) {
struct zone *zone = pgdat->node_zones + i;
unsigned long max = 0;
/* Find valid and maximum lowmem_reserve in the zone */
for (j = i; j < MAX_NR_ZONES; j++) {
if (zone->lowmem_reserve[j] > max)
max = zone->lowmem_reserve[j];
}
/* we treat the high watermark as reserved pages. */
max += high_wmark_pages(zone);
[PATCH] overcommit: add calculate_totalreserve_pages() These patches are an enhancement of OVERCOMMIT_GUESS algorithm in __vm_enough_memory(). - why the kernel needed patching When the kernel can't allocate anonymous pages in practice, currnet OVERCOMMIT_GUESS could return success. This implementation might be the cause of oom kill in memory pressure situation. If the Linux runs with page reservation features like /proc/sys/vm/lowmem_reserve_ratio and without swap region, I think the oom kill occurs easily. - the overall design approach in the patch When the OVERCOMMET_GUESS algorithm calculates number of free pages, the reserved free pages are regarded as non-free pages. This change helps to avoid the pitfall that the number of free pages become less than the number which the kernel tries to keep free. - testing results I tested the patches using my test kernel module. If the patches aren't applied to the kernel, __vm_enough_memory() returns success in the situation but autual page allocation is failed. On the other hand, if the patches are applied to the kernel, memory allocation failure is avoided since __vm_enough_memory() returns failure in the situation. I checked that on i386 SMP 16GB memory machine. I haven't tested on nommu environment currently. This patch adds totalreserve_pages for __vm_enough_memory(). Calculate_totalreserve_pages() checks maximum lowmem_reserve pages and pages_high in each zone. Finally, the function stores the sum of each zone to totalreserve_pages. The totalreserve_pages is calculated when the VM is initilized. And the variable is updated when /proc/sys/vm/lowmem_reserve_raito or /proc/sys/vm/min_free_kbytes are changed. Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-11 05:52:59 +00:00
if (max > zone->present_pages)
max = zone->present_pages;
reserve_pages += max;
}
}
totalreserve_pages = reserve_pages;
}
/*
* setup_per_zone_lowmem_reserve - called whenever
* sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
* has a correct pages reserved value, so an adequate number of
* pages are left in the zone after a successful __alloc_pages().
*/
static void setup_per_zone_lowmem_reserve(void)
{
struct pglist_data *pgdat;
enum zone_type j, idx;
for_each_online_pgdat(pgdat) {
for (j = 0; j < MAX_NR_ZONES; j++) {
struct zone *zone = pgdat->node_zones + j;
unsigned long present_pages = zone->present_pages;
zone->lowmem_reserve[j] = 0;
idx = j;
while (idx) {
struct zone *lower_zone;
idx--;
if (sysctl_lowmem_reserve_ratio[idx] < 1)
sysctl_lowmem_reserve_ratio[idx] = 1;
lower_zone = pgdat->node_zones + idx;
lower_zone->lowmem_reserve[j] = present_pages /
sysctl_lowmem_reserve_ratio[idx];
present_pages += lower_zone->present_pages;
}
}
}
[PATCH] overcommit: add calculate_totalreserve_pages() These patches are an enhancement of OVERCOMMIT_GUESS algorithm in __vm_enough_memory(). - why the kernel needed patching When the kernel can't allocate anonymous pages in practice, currnet OVERCOMMIT_GUESS could return success. This implementation might be the cause of oom kill in memory pressure situation. If the Linux runs with page reservation features like /proc/sys/vm/lowmem_reserve_ratio and without swap region, I think the oom kill occurs easily. - the overall design approach in the patch When the OVERCOMMET_GUESS algorithm calculates number of free pages, the reserved free pages are regarded as non-free pages. This change helps to avoid the pitfall that the number of free pages become less than the number which the kernel tries to keep free. - testing results I tested the patches using my test kernel module. If the patches aren't applied to the kernel, __vm_enough_memory() returns success in the situation but autual page allocation is failed. On the other hand, if the patches are applied to the kernel, memory allocation failure is avoided since __vm_enough_memory() returns failure in the situation. I checked that on i386 SMP 16GB memory machine. I haven't tested on nommu environment currently. This patch adds totalreserve_pages for __vm_enough_memory(). Calculate_totalreserve_pages() checks maximum lowmem_reserve pages and pages_high in each zone. Finally, the function stores the sum of each zone to totalreserve_pages. The totalreserve_pages is calculated when the VM is initilized. And the variable is updated when /proc/sys/vm/lowmem_reserve_raito or /proc/sys/vm/min_free_kbytes are changed. Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-11 05:52:59 +00:00
/* update totalreserve_pages */
calculate_totalreserve_pages();
}
/**
* setup_per_zone_wmarks - called when min_free_kbytes changes
* or when memory is hot-{added|removed}
*
* Ensures that the watermark[min,low,high] values for each zone are set
* correctly with respect to min_free_kbytes.
*/
void setup_per_zone_wmarks(void)
{
unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
unsigned long lowmem_pages = 0;
struct zone *zone;
unsigned long flags;
/* Calculate total number of !ZONE_HIGHMEM pages */
for_each_zone(zone) {
if (!is_highmem(zone))
lowmem_pages += zone->present_pages;
}
for_each_zone(zone) {
u64 tmp;
spin_lock_irqsave(&zone->lock, flags);
tmp = (u64)pages_min * zone->present_pages;
do_div(tmp, lowmem_pages);
if (is_highmem(zone)) {
/*
* __GFP_HIGH and PF_MEMALLOC allocations usually don't
* need highmem pages, so cap pages_min to a small
* value here.
*
* The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
* deltas controls asynch page reclaim, and so should
* not be capped for highmem.
*/
int min_pages;
min_pages = zone->present_pages / 1024;
if (min_pages < SWAP_CLUSTER_MAX)
min_pages = SWAP_CLUSTER_MAX;
if (min_pages > 128)
min_pages = 128;
zone->watermark[WMARK_MIN] = min_pages;
} else {
/*
* If it's a lowmem zone, reserve a number of pages
* proportionate to the zone's size.
*/
zone->watermark[WMARK_MIN] = tmp;
}
zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks The standard buddy allocator always favours the smallest block of pages. The effect of this is that the pages free to satisfy min_free_kbytes tends to be preserved since boot time at the same location of memory ffor a very long time and as a contiguous block. When an administrator sets the reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks that remain free. This allows the occasional high atomic allocation to succeed up until the point the blocks are split. In practice, it is difficult to split these blocks but when they do split, the benefit of having min_free_kbytes for contiguous blocks disappears. Additionally, increasing min_free_kbytes once the system has been running for some time has no guarantee of creating contiguous blocks. On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large blocks when there are no free pages of the appropriate type available. A side-effect of this is that all blocks in memory tends to be used up and the contiguous free blocks from boot time are not preserved like in the vanilla allocator. This can cause a problem if a new caller is unwilling to reclaim or does not reclaim for long enough. A failure scenario was found for a wireless network device allocating order-1 atomic allocations but the allocations were not intense or frequent enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC. This was reproduced on a desktop by booting with mem=256mb, forcing the driver to allocate at order-1, running a bittorrent client (downloading a debian ISO) and building a kernel with -j2. This patch addresses the problem on the desktop machine booted with mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES blocks, the number of which depends on the value of min_free_kbytes. These blocks are only fallen back to when there is no other free pages. Then the smallest possible page is used just like the normal buddy allocator instead of the largest possible page to preserve contiguous pages The pages in free lists in the reserve blocks are never taken for another migrate type. The results is that even if min_free_kbytes is set to a low value, contiguous blocks will be preserved in the MIGRATE_RESERVE blocks. This works better than the vanilla allocator because if min_free_kbytes is increased, a new reserve block will be chosen based on the location of reclaimable pages and the block will free up as contiguous pages. In the vanilla allocator, no effort is made to target a block of pages to free as contiguous pages and min_free_kbytes pages are scattered randomly. This effect has been observed on the test machine. min_free_kbytes was set initially low but it was kept as a contiguous free block within MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a period of time, the free blocks were within the reserve and coalescing. How long it takes to free up depends on how quickly LRU is rotating. Amusingly, this means that more activity will free the blocks faster. This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more effective than grouping contiguous free pages together. It all depends on whether the number of active atomic high allocations exceeds min_free_kbytes or not. If the number of active allocations exceeds min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes should be set higher. Once there are no more reports of allocation failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and see if the reports stay missing. Credit to Mariusz Kozlowski for discovering the problem, describing the failure scenario and testing patches and scenarios. [akpm@linux-foundation.org: cleanups] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:58 +00:00
setup_zone_migrate_reserve(zone);
spin_unlock_irqrestore(&zone->lock, flags);
}
[PATCH] overcommit: add calculate_totalreserve_pages() These patches are an enhancement of OVERCOMMIT_GUESS algorithm in __vm_enough_memory(). - why the kernel needed patching When the kernel can't allocate anonymous pages in practice, currnet OVERCOMMIT_GUESS could return success. This implementation might be the cause of oom kill in memory pressure situation. If the Linux runs with page reservation features like /proc/sys/vm/lowmem_reserve_ratio and without swap region, I think the oom kill occurs easily. - the overall design approach in the patch When the OVERCOMMET_GUESS algorithm calculates number of free pages, the reserved free pages are regarded as non-free pages. This change helps to avoid the pitfall that the number of free pages become less than the number which the kernel tries to keep free. - testing results I tested the patches using my test kernel module. If the patches aren't applied to the kernel, __vm_enough_memory() returns success in the situation but autual page allocation is failed. On the other hand, if the patches are applied to the kernel, memory allocation failure is avoided since __vm_enough_memory() returns failure in the situation. I checked that on i386 SMP 16GB memory machine. I haven't tested on nommu environment currently. This patch adds totalreserve_pages for __vm_enough_memory(). Calculate_totalreserve_pages() checks maximum lowmem_reserve pages and pages_high in each zone. Finally, the function stores the sum of each zone to totalreserve_pages. The totalreserve_pages is calculated when the VM is initilized. And the variable is updated when /proc/sys/vm/lowmem_reserve_raito or /proc/sys/vm/min_free_kbytes are changed. Signed-off-by: Hideo Aoki <haoki@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-11 05:52:59 +00:00
/* update totalreserve_pages */
calculate_totalreserve_pages();
}
/*
* The inactive anon list should be small enough that the VM never has to
* do too much work, but large enough that each inactive page has a chance
* to be referenced again before it is swapped out.
*
* The inactive_anon ratio is the target ratio of ACTIVE_ANON to
* INACTIVE_ANON pages on this zone's LRU, maintained by the
* pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
* the anonymous pages are kept on the inactive list.
*
* total target max
* memory ratio inactive anon
* -------------------------------------
* 10MB 1 5MB
* 100MB 1 50MB
* 1GB 3 250MB
* 10GB 10 0.9GB
* 100GB 31 3GB
* 1TB 101 10GB
* 10TB 320 32GB
*/
static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
{
unsigned int gb, ratio;
/* Zone size in gigabytes */
gb = zone->present_pages >> (30 - PAGE_SHIFT);
if (gb)
ratio = int_sqrt(10 * gb);
else
ratio = 1;
zone->inactive_ratio = ratio;
}
static void __meminit setup_per_zone_inactive_ratio(void)
{
struct zone *zone;
for_each_zone(zone)
calculate_zone_inactive_ratio(zone);
}
/*
* Initialise min_free_kbytes.
*
* For small machines we want it small (128k min). For large machines
* we want it large (64MB max). But it is not linear, because network
* bandwidth does not increase linearly with machine size. We use
*
* min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
* min_free_kbytes = sqrt(lowmem_kbytes * 16)
*
* which yields
*
* 16MB: 512k
* 32MB: 724k
* 64MB: 1024k
* 128MB: 1448k
* 256MB: 2048k
* 512MB: 2896k
* 1024MB: 4096k
* 2048MB: 5792k
* 4096MB: 8192k
* 8192MB: 11584k
* 16384MB: 16384k
*/
int __meminit init_per_zone_wmark_min(void)
{
unsigned long lowmem_kbytes;
lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
if (min_free_kbytes < 128)
min_free_kbytes = 128;
if (min_free_kbytes > 65536)
min_free_kbytes = 65536;
setup_per_zone_wmarks();
refresh_zone_stat_thresholds();
setup_per_zone_lowmem_reserve();
setup_per_zone_inactive_ratio();
return 0;
}
module_init(init_per_zone_wmark_min)
/*
* min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
* that we can call two helper functions whenever min_free_kbytes
* changes.
*/
int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
void __user *buffer, size_t *length, loff_t *ppos)
{
proc_dointvec(table, write, buffer, length, ppos);
if (write)
setup_per_zone_wmarks();
return 0;
}
#ifdef CONFIG_NUMA
int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
void __user *buffer, size_t *length, loff_t *ppos)
{
struct zone *zone;
int rc;
rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
if (rc)
return rc;
for_each_zone(zone)
zone->min_unmapped_pages = (zone->present_pages *
sysctl_min_unmapped_ratio) / 100;
return 0;
}
[PATCH] zone_reclaim: dynamic slab reclaim Currently one can enable slab reclaim by setting an explicit option in /proc/sys/vm/zone_reclaim_mode. Slab reclaim is then used as a final option if the freeing of unmapped file backed pages is not enough to free enough pages to allow a local allocation. However, that means that the slab can grow excessively and that most memory of a node may be used by slabs. We have had a case where a machine with 46GB of memory was using 40-42GB for slab. Zone reclaim was effective in dealing with pagecache pages. However, slab reclaim was only done during global reclaim (which is a bit rare on NUMA systems). This patch implements slab reclaim during zone reclaim. Zone reclaim occurs if there is a danger of an off node allocation. At that point we 1. Shrink the per node page cache if the number of pagecache pages is more than min_unmapped_ratio percent of pages in a zone. 2. Shrink the slab cache if the number of the nodes reclaimable slab pages (patch depends on earlier one that implements that counter) are more than min_slab_ratio (a new /proc/sys/vm tunable). The shrinking of the slab cache is a bit problematic since it is not node specific. So we simply calculate what point in the slab we want to reach (current per node slab use minus the number of pages that neeed to be allocated) and then repeately run the global reclaim until that is unsuccessful or we have reached the limit. I hope we will have zone based slab reclaim at some point which will make that easier. The default for the min_slab_ratio is 5% Also remove the slab option from /proc/sys/vm/zone_reclaim_mode. [akpm@osdl.org: cleanups] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-26 06:31:52 +00:00
int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
void __user *buffer, size_t *length, loff_t *ppos)
[PATCH] zone_reclaim: dynamic slab reclaim Currently one can enable slab reclaim by setting an explicit option in /proc/sys/vm/zone_reclaim_mode. Slab reclaim is then used as a final option if the freeing of unmapped file backed pages is not enough to free enough pages to allow a local allocation. However, that means that the slab can grow excessively and that most memory of a node may be used by slabs. We have had a case where a machine with 46GB of memory was using 40-42GB for slab. Zone reclaim was effective in dealing with pagecache pages. However, slab reclaim was only done during global reclaim (which is a bit rare on NUMA systems). This patch implements slab reclaim during zone reclaim. Zone reclaim occurs if there is a danger of an off node allocation. At that point we 1. Shrink the per node page cache if the number of pagecache pages is more than min_unmapped_ratio percent of pages in a zone. 2. Shrink the slab cache if the number of the nodes reclaimable slab pages (patch depends on earlier one that implements that counter) are more than min_slab_ratio (a new /proc/sys/vm tunable). The shrinking of the slab cache is a bit problematic since it is not node specific. So we simply calculate what point in the slab we want to reach (current per node slab use minus the number of pages that neeed to be allocated) and then repeately run the global reclaim until that is unsuccessful or we have reached the limit. I hope we will have zone based slab reclaim at some point which will make that easier. The default for the min_slab_ratio is 5% Also remove the slab option from /proc/sys/vm/zone_reclaim_mode. [akpm@osdl.org: cleanups] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-26 06:31:52 +00:00
{
struct zone *zone;
int rc;
rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
[PATCH] zone_reclaim: dynamic slab reclaim Currently one can enable slab reclaim by setting an explicit option in /proc/sys/vm/zone_reclaim_mode. Slab reclaim is then used as a final option if the freeing of unmapped file backed pages is not enough to free enough pages to allow a local allocation. However, that means that the slab can grow excessively and that most memory of a node may be used by slabs. We have had a case where a machine with 46GB of memory was using 40-42GB for slab. Zone reclaim was effective in dealing with pagecache pages. However, slab reclaim was only done during global reclaim (which is a bit rare on NUMA systems). This patch implements slab reclaim during zone reclaim. Zone reclaim occurs if there is a danger of an off node allocation. At that point we 1. Shrink the per node page cache if the number of pagecache pages is more than min_unmapped_ratio percent of pages in a zone. 2. Shrink the slab cache if the number of the nodes reclaimable slab pages (patch depends on earlier one that implements that counter) are more than min_slab_ratio (a new /proc/sys/vm tunable). The shrinking of the slab cache is a bit problematic since it is not node specific. So we simply calculate what point in the slab we want to reach (current per node slab use minus the number of pages that neeed to be allocated) and then repeately run the global reclaim until that is unsuccessful or we have reached the limit. I hope we will have zone based slab reclaim at some point which will make that easier. The default for the min_slab_ratio is 5% Also remove the slab option from /proc/sys/vm/zone_reclaim_mode. [akpm@osdl.org: cleanups] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-26 06:31:52 +00:00
if (rc)
return rc;
for_each_zone(zone)
zone->min_slab_pages = (zone->present_pages *
sysctl_min_slab_ratio) / 100;
return 0;
}
#endif
/*
* lowmem_reserve_ratio_sysctl_handler - just a wrapper around
* proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
* whenever sysctl_lowmem_reserve_ratio changes.
*
* The reserve ratio obviously has absolutely no relation with the
* minimum watermarks. The lowmem reserve ratio can only make sense
* if in function of the boot time zone sizes.
*/
int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
void __user *buffer, size_t *length, loff_t *ppos)
{
proc_dointvec_minmax(table, write, buffer, length, ppos);
setup_per_zone_lowmem_reserve();
return 0;
}
/*
* percpu_pagelist_fraction - changes the pcp->high for each zone on each
* cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
* can have before it gets flushed back to buddy allocator.
*/
int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
void __user *buffer, size_t *length, loff_t *ppos)
{
struct zone *zone;
unsigned int cpu;
int ret;
ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
if (!write || (ret == -EINVAL))
return ret;
for_each_populated_zone(zone) {
for_each_possible_cpu(cpu) {
unsigned long high;
high = zone->present_pages / percpu_pagelist_fraction;
setup_pagelist_highmark(
per_cpu_ptr(zone->pageset, cpu), high);
}
}
return 0;
}
int hashdist = HASHDIST_DEFAULT;
#ifdef CONFIG_NUMA
static int __init set_hashdist(char *str)
{
if (!str)
return 0;
hashdist = simple_strtoul(str, &str, 0);
return 1;
}
__setup("hashdist=", set_hashdist);
#endif
/*
* allocate a large system hash table from bootmem
* - it is assumed that the hash table must contain an exact power-of-2
* quantity of entries
* - limit is the number of hash buckets, not the total allocation size
*/
void *__init alloc_large_system_hash(const char *tablename,
unsigned long bucketsize,
unsigned long numentries,
int scale,
int flags,
unsigned int *_hash_shift,
unsigned int *_hash_mask,
unsigned long limit)
{
unsigned long long max = limit;
unsigned long log2qty, size;
void *table = NULL;
/* allow the kernel cmdline to have a say */
if (!numentries) {
/* round applicable memory size up to nearest megabyte */
numentries = nr_kernel_pages;
numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
numentries >>= 20 - PAGE_SHIFT;
numentries <<= 20 - PAGE_SHIFT;
/* limit to 1 bucket per 2^scale bytes of low memory */
if (scale > PAGE_SHIFT)
numentries >>= (scale - PAGE_SHIFT);
else
numentries <<= (PAGE_SHIFT - scale);
/* Make sure we've got at least a 0-order allocation.. */
if (unlikely(flags & HASH_SMALL)) {
/* Makes no sense without HASH_EARLY */
WARN_ON(!(flags & HASH_EARLY));
if (!(numentries >> *_hash_shift)) {
numentries = 1UL << *_hash_shift;
BUG_ON(!numentries);
}
} else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
numentries = PAGE_SIZE / bucketsize;
}
numentries = roundup_pow_of_two(numentries);
/* limit allocation size to 1/16 total memory by default */
if (max == 0) {
max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
do_div(max, bucketsize);
}
if (numentries > max)
numentries = max;
log2qty = ilog2(numentries);
do {
size = bucketsize << log2qty;
if (flags & HASH_EARLY)
table = alloc_bootmem_nopanic(size);
else if (hashdist)
table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
else {
/*
* If bucketsize is not a power-of-two, we may free
* some pages at the end of hash table which
* alloc_pages_exact() automatically does
*/
if (get_order(size) < MAX_ORDER) {
table = alloc_pages_exact(size, GFP_ATOMIC);
kmemleak_alloc(table, size, 1, GFP_ATOMIC);
}
}
} while (!table && size > PAGE_SIZE && --log2qty);
if (!table)
panic("Failed to allocate %s hash table\n", tablename);
printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
tablename,
(1UL << log2qty),
ilog2(size) - PAGE_SHIFT,
size);
if (_hash_shift)
*_hash_shift = log2qty;
if (_hash_mask)
*_hash_mask = (1 << log2qty) - 1;
return table;
}
[PATCH] unify pfn_to_page: generic functions There are 3 memory models, FLATMEM, DISCONTIGMEM, SPARSEMEM. Each arch has its own page_to_pfn(), pfn_to_page() for each models. But most of them can use the same arithmetic. This patch adds asm-generic/memory_model.h, which includes generic page_to_pfn(), pfn_to_page() definitions for each memory model. When CONFIG_OUT_OF_LINE_PFN_TO_PAGE=y, out-of-line functions are used instead of macro. This is enabled by some archs and reduces text size. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Hugh Dickins <hugh@veritas.com> Cc: Andi Kleen <ak@muc.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Richard Henderson <rth@twiddle.net> Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru> Cc: Russell King <rmk@arm.linux.org.uk> Cc: Ian Molton <spyro@f2s.com> Cc: Mikael Starvik <starvik@axis.com> Cc: David Howells <dhowells@redhat.com> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Hirokazu Takata <takata.hirokazu@renesas.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Paul Mundt <lethal@linux-sh.org> Cc: Kazumoto Kojima <kkojima@rr.iij4u.or.jp> Cc: Richard Curnow <rc@rc0.org.uk> Cc: William Lee Irwin III <wli@holomorphy.com> Cc: "David S. Miller" <davem@davemloft.net> Cc: Jeff Dike <jdike@addtoit.com> Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Cc: Miles Bader <uclinux-v850@lsi.nec.co.jp> Cc: Chris Zankel <chris@zankel.net> Cc: "Luck, Tony" <tony.luck@intel.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-27 09:15:25 +00:00
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
/* Return a pointer to the bitmap storing bits affecting a block of pages */
static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
unsigned long pfn)
{
#ifdef CONFIG_SPARSEMEM
return __pfn_to_section(pfn)->pageblock_flags;
#else
return zone->pageblock_flags;
#endif /* CONFIG_SPARSEMEM */
}
static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
{
#ifdef CONFIG_SPARSEMEM
pfn &= (PAGES_PER_SECTION-1);
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
#else
pfn = pfn - zone->zone_start_pfn;
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
#endif /* CONFIG_SPARSEMEM */
}
/**
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
* get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
* @page: The page within the block of interest
* @start_bitidx: The first bit of interest to retrieve
* @end_bitidx: The last bit of interest
* returns pageblock_bits flags
*/
unsigned long get_pageblock_flags_group(struct page *page,
int start_bitidx, int end_bitidx)
{
struct zone *zone;
unsigned long *bitmap;
unsigned long pfn, bitidx;
unsigned long flags = 0;
unsigned long value = 1;
zone = page_zone(page);
pfn = page_to_pfn(page);
bitmap = get_pageblock_bitmap(zone, pfn);
bitidx = pfn_to_bitidx(zone, pfn);
for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
if (test_bit(bitidx + start_bitidx, bitmap))
flags |= value;
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
return flags;
}
/**
Do not depend on MAX_ORDER when grouping pages by mobility Currently mobility grouping works at the MAX_ORDER_NR_PAGES level. This makes sense for the majority of users where this is also the huge page size. However, on platforms like ia64 where the huge page size is runtime configurable it is desirable to group at a lower order. On x86_64 and occasionally on x86, the hugepage size may not always be MAX_ORDER_NR_PAGES. This patch groups pages together based on the value of HUGETLB_PAGE_ORDER. It uses a compile-time constant if possible and a variable where the huge page size is runtime configurable. It is assumed that grouping should be done at the lowest sensible order and that the user would not want to override this. If this is not true, page_block order could be forced to a variable initialised via a boot-time kernel parameter. One potential issue with this patch is that IA64 now parses hugepagesz with early_param() instead of __setup(). __setup() is called after the memory allocator has been initialised and the pageblock bitmaps already setup. In tests on one IA64 there did not seem to be any problem with using early_param() and in fact may be more correct as it guarantees the parameter is handled before the parsing of hugepages=. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Andy Whitcroft <apw@shadowen.org> Acked-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:26:01 +00:00
* set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
* @page: The page within the block of interest
* @start_bitidx: The first bit of interest
* @end_bitidx: The last bit of interest
* @flags: The flags to set
*/
void set_pageblock_flags_group(struct page *page, unsigned long flags,
int start_bitidx, int end_bitidx)
{
struct zone *zone;
unsigned long *bitmap;
unsigned long pfn, bitidx;
unsigned long value = 1;
zone = page_zone(page);
pfn = page_to_pfn(page);
bitmap = get_pageblock_bitmap(zone, pfn);
bitidx = pfn_to_bitidx(zone, pfn);
VM_BUG_ON(pfn < zone->zone_start_pfn);
VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
Add a bitmap that is used to track flags affecting a block of pages Here is the latest revision of the anti-fragmentation patches. Of particular note in this version is special treatment of high-order atomic allocations. Care is taken to group them together and avoid grouping pages of other types near them. Artifical tests imply that it works. I'm trying to get the hardware together that would allow setting up of a "real" test. If anyone already has a setup and test that can trigger the atomic-allocation problem, I'd appreciate a test of these patches and a report. The second major change is that these patches will apply cleanly with patches that implement anti-fragmentation through zones. kernbench shows effectively no performance difference varying between -0.2% and +2% on a variety of test machines. Success rates for huge page allocation are dramatically increased. For example, on a ppc64 machine, the vanilla kernel was only able to allocate 1% of memory as a hugepage and this was due to a single hugepage reserved as min_free_kbytes. With these patches applied, 17% was allocatable as superpages. With reclaim-related fixes from Andy Whitcroft, it was 40% and further reclaim-related improvements should increase this further. Changelog Since V28 o Group high-order atomic allocations together o It is no longer required to set min_free_kbytes to 10% of memory. A value of 16384 in most cases will be sufficient o Now applied with zone-based anti-fragmentation o Fix incorrect VM_BUG_ON within buffered_rmqueue() o Reorder the stack so later patches do not back out work from earlier patches o Fix bug were journal pages were being treated as movable o Bias placement of non-movable pages to lower PFNs o More agressive clustering of reclaimable pages in reactions to workloads like updatedb that flood the size of inode caches Changelog Since V27 o Renamed anti-fragmentation to Page Clustering. Anti-fragmentation was giving the mistaken impression that it was the 100% solution for high order allocations. Instead, it greatly increases the chances high-order allocations will succeed and lays the foundation for defragmentation and memory hot-remove to work properly o Redefine page groupings based on ability to migrate or reclaim instead of basing on reclaimability alone o Get rid of spurious inits o Per-cpu lists are no longer split up per-type. Instead the per-cpu list is searched for a page of the appropriate type o Added more explanation commentary o Fix up bug in pageblock code where bitmap was used before being initalised Changelog Since V26 o Fix double init of lists in setup_pageset Changelog Since V25 o Fix loop order of for_each_rclmtype_order so that order of loop matches args o gfpflags_to_rclmtype uses gfp_t instead of unsigned long o Rename get_pageblock_type() to get_page_rclmtype() o Fix alignment problem in move_freepages() o Add mechanism for assigning flags to blocks of pages instead of page->flags o On fallback, do not examine the preferred list of free pages a second time The purpose of these patches is to reduce external fragmentation by grouping pages of related types together. When pages are migrated (or reclaimed under memory pressure), large contiguous pages will be freed. This patch works by categorising allocations by their ability to migrate; Movable - The pages may be moved with the page migration mechanism. These are generally userspace pages. Reclaimable - These are allocations for some kernel caches that are reclaimable or allocations that are known to be very short-lived. Unmovable - These are pages that are allocated by the kernel that are not trivially reclaimed. For example, the memory allocated for a loaded module would be in this category. By default, allocations are considered to be of this type HighAtomic - These are high-order allocations belonging to callers that cannot sleep or perform any IO. In practice, this is restricted to jumbo frame allocation for network receive. It is assumed that the allocations are short-lived Instead of having one MAX_ORDER-sized array of free lists in struct free_area, there is one for each type of reclaimability. Once a 2^MAX_ORDER block of pages is split for a type of allocation, it is added to the free-lists for that type, in effect reserving it. Hence, over time, pages of the different types can be clustered together. When the preferred freelists are expired, the largest possible block is taken from an alternative list. Buddies that are split from that large block are placed on the preferred allocation-type freelists to mitigate fragmentation. This implementation gives best-effort for low fragmentation in all zones. Ideally, min_free_kbytes needs to be set to a value equal to 4 * (1 << (MAX_ORDER-1)) pages in most cases. This would be 16384 on x86 and x86_64 for example. Our tests show that about 60-70% of physical memory can be allocated on a desktop after a few days uptime. In benchmarks and stress tests, we are finding that 80% of memory is available as contiguous blocks at the end of the test. To compare, a standard kernel was getting < 1% of memory as large pages on a desktop and about 8-12% of memory as large pages at the end of stress tests. Following this email are 12 patches that implement thie page grouping feature. The first patch introduces a mechanism for storing flags related to a whole block of pages. Then allocations are split between movable and all other allocations. Following that are patches to deal with per-cpu pages and make the mechanism configurable. The next patch moves free pages between lists when partially allocated blocks are used for pages of another migrate type. The second last patch groups reclaimable kernel allocations such as inode caches together. The final patch related to groupings keeps high-order atomic allocations. The last two patches are more concerned with control of fragmentation. The second last patch biases placement of non-movable allocations towards the start of memory. This is with a view of supporting memory hot-remove of DIMMs with higher PFNs in the future. The biasing could be enforced a lot heavier but it would cost. The last patch agressively clusters reclaimable pages like inode caches together. The fragmentation reduction strategy needs to track if pages within a block can be moved or reclaimed so that pages are freed to the appropriate list. This patch adds a bitmap for flags affecting a whole a MAX_ORDER block of pages. In non-SPARSEMEM configurations, the bitmap is stored in the struct zone and allocated during initialisation. SPARSEMEM statically allocates the bitmap in a struct mem_section so that bitmaps do not have to be resized during memory hotadd. This wastes a small amount of memory per unused section (usually sizeof(unsigned long)) but the complexity of dynamically allocating the memory is quite high. Additional credit to Andy Whitcroft who reviewed up an earlier implementation of the mechanism an suggested how to make it a *lot* cleaner. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 08:25:47 +00:00
for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
if (flags & value)
__set_bit(bitidx + start_bitidx, bitmap);
else
__clear_bit(bitidx + start_bitidx, bitmap);
}
/*
* This is designed as sub function...plz see page_isolation.c also.
* set/clear page block's type to be ISOLATE.
* page allocater never alloc memory from ISOLATE block.
*/
static int
__count_immobile_pages(struct zone *zone, struct page *page, int count)
{
unsigned long pfn, iter, found;
/*
* For avoiding noise data, lru_add_drain_all() should be called
* If ZONE_MOVABLE, the zone never contains immobile pages
*/
if (zone_idx(zone) == ZONE_MOVABLE)
return true;
if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
return true;
pfn = page_to_pfn(page);
for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
unsigned long check = pfn + iter;
if (!pfn_valid_within(check))
continue;
page = pfn_to_page(check);
if (!page_count(page)) {
if (PageBuddy(page))
iter += (1 << page_order(page)) - 1;
continue;
}
if (!PageLRU(page))
found++;
/*
* If there are RECLAIMABLE pages, we need to check it.
* But now, memory offline itself doesn't call shrink_slab()
* and it still to be fixed.
*/
/*
* If the page is not RAM, page_count()should be 0.
* we don't need more check. This is an _used_ not-movable page.
*
* The problematic thing here is PG_reserved pages. PG_reserved
* is set to both of a memory hole page and a _used_ kernel
* page at boot.
*/
if (found > count)
return false;
}
return true;
}
bool is_pageblock_removable_nolock(struct page *page)
{
struct zone *zone = page_zone(page);
return __count_immobile_pages(zone, page, 0);
}
int set_migratetype_isolate(struct page *page)
{
struct zone *zone;
unsigned long flags, pfn;
struct memory_isolate_notify arg;
int notifier_ret;
int ret = -EBUSY;
zone = page_zone(page);
spin_lock_irqsave(&zone->lock, flags);
pfn = page_to_pfn(page);
arg.start_pfn = pfn;
arg.nr_pages = pageblock_nr_pages;
arg.pages_found = 0;
/*
* It may be possible to isolate a pageblock even if the
* migratetype is not MIGRATE_MOVABLE. The memory isolation
* notifier chain is used by balloon drivers to return the
* number of pages in a range that are held by the balloon
* driver to shrink memory. If all the pages are accounted for
* by balloons, are free, or on the LRU, isolation can continue.
* Later, for example, when memory hotplug notifier runs, these
* pages reported as "can be isolated" should be isolated(freed)
* by the balloon driver through the memory notifier chain.
*/
notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
notifier_ret = notifier_to_errno(notifier_ret);
if (notifier_ret)
goto out;
/*
* FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
* We just check MOVABLE pages.
*/
if (__count_immobile_pages(zone, page, arg.pages_found))
ret = 0;
/*
* immobile means "not-on-lru" paes. If immobile is larger than
* removable-by-driver pages reported by notifier, we'll fail.
*/
out:
if (!ret) {
set_pageblock_migratetype(page, MIGRATE_ISOLATE);
move_freepages_block(zone, page, MIGRATE_ISOLATE);
}
spin_unlock_irqrestore(&zone->lock, flags);
if (!ret)
drain_all_pages();
return ret;
}
void unset_migratetype_isolate(struct page *page)
{
struct zone *zone;
unsigned long flags;
zone = page_zone(page);
spin_lock_irqsave(&zone->lock, flags);
if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
goto out;
set_pageblock_migratetype(page, MIGRATE_MOVABLE);
move_freepages_block(zone, page, MIGRATE_MOVABLE);
out:
spin_unlock_irqrestore(&zone->lock, flags);
}
#ifdef CONFIG_MEMORY_HOTREMOVE
/*
* All pages in the range must be isolated before calling this.
*/
void
__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
{
struct page *page;
struct zone *zone;
int order, i;
unsigned long pfn;
unsigned long flags;
/* find the first valid pfn */
for (pfn = start_pfn; pfn < end_pfn; pfn++)
if (pfn_valid(pfn))
break;
if (pfn == end_pfn)
return;
zone = page_zone(pfn_to_page(pfn));
spin_lock_irqsave(&zone->lock, flags);
pfn = start_pfn;
while (pfn < end_pfn) {
if (!pfn_valid(pfn)) {
pfn++;
continue;
}
page = pfn_to_page(pfn);
BUG_ON(page_count(page));
BUG_ON(!PageBuddy(page));
order = page_order(page);
#ifdef CONFIG_DEBUG_VM
printk(KERN_INFO "remove from free list %lx %d %lx\n",
pfn, 1 << order, end_pfn);
#endif
list_del(&page->lru);
rmv_page_order(page);
zone->free_area[order].nr_free--;
__mod_zone_page_state(zone, NR_FREE_PAGES,
- (1UL << order));
for (i = 0; i < (1 << order); i++)
SetPageReserved((page+i));
pfn += (1 << order);
}
spin_unlock_irqrestore(&zone->lock, flags);
}
#endif
#ifdef CONFIG_MEMORY_FAILURE
bool is_free_buddy_page(struct page *page)
{
struct zone *zone = page_zone(page);
unsigned long pfn = page_to_pfn(page);
unsigned long flags;
int order;
spin_lock_irqsave(&zone->lock, flags);
for (order = 0; order < MAX_ORDER; order++) {
struct page *page_head = page - (pfn & ((1 << order) - 1));
if (PageBuddy(page_head) && page_order(page_head) >= order)
break;
}
spin_unlock_irqrestore(&zone->lock, flags);
return order < MAX_ORDER;
}
#endif
static struct trace_print_flags pageflag_names[] = {
{1UL << PG_locked, "locked" },
{1UL << PG_error, "error" },
{1UL << PG_referenced, "referenced" },
{1UL << PG_uptodate, "uptodate" },
{1UL << PG_dirty, "dirty" },
{1UL << PG_lru, "lru" },
{1UL << PG_active, "active" },
{1UL << PG_slab, "slab" },
{1UL << PG_owner_priv_1, "owner_priv_1" },
{1UL << PG_arch_1, "arch_1" },
{1UL << PG_reserved, "reserved" },
{1UL << PG_private, "private" },
{1UL << PG_private_2, "private_2" },
{1UL << PG_writeback, "writeback" },
#ifdef CONFIG_PAGEFLAGS_EXTENDED
{1UL << PG_head, "head" },
{1UL << PG_tail, "tail" },
#else
{1UL << PG_compound, "compound" },
#endif
{1UL << PG_swapcache, "swapcache" },
{1UL << PG_mappedtodisk, "mappedtodisk" },
{1UL << PG_reclaim, "reclaim" },
{1UL << PG_swapbacked, "swapbacked" },
{1UL << PG_unevictable, "unevictable" },
#ifdef CONFIG_MMU
{1UL << PG_mlocked, "mlocked" },
#endif
#ifdef CONFIG_ARCH_USES_PG_UNCACHED
{1UL << PG_uncached, "uncached" },
#endif
#ifdef CONFIG_MEMORY_FAILURE
{1UL << PG_hwpoison, "hwpoison" },
#endif
{-1UL, NULL },
};
static void dump_page_flags(unsigned long flags)
{
const char *delim = "";
unsigned long mask;
int i;
printk(KERN_ALERT "page flags: %#lx(", flags);
/* remove zone id */
flags &= (1UL << NR_PAGEFLAGS) - 1;
for (i = 0; pageflag_names[i].name && flags; i++) {
mask = pageflag_names[i].mask;
if ((flags & mask) != mask)
continue;
flags &= ~mask;
printk("%s%s", delim, pageflag_names[i].name);
delim = "|";
}
/* check for left over flags */
if (flags)
printk("%s%#lx", delim, flags);
printk(")\n");
}
void dump_page(struct page *page)
{
printk(KERN_ALERT
"page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
page, atomic_read(&page->_count), page_mapcount(page),
page->mapping, page->index);
dump_page_flags(page->flags);
mem_cgroup_print_bad_page(page);
}