Rename _bss to __bss_start as on other architectures. That makes it
possible to use the <linux/sections.h> instead of own declarations. Also
add __bss_stop because that symbol exists on other architectures.
Signed-off-by: Bernhard Walle <bwalle@suse.de>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Clean up the process for presenting the "physical id" field in
/proc/cpuinfo.
- remove global smp_num_cpucores, as it is mostly useless
- remove check_for_logical_procs(), since we do the same
functionality in identify_siblings()
- reflow logic in identify_siblings(). If an older CPU
does not implement PAL_LOGICAL_TO_PHYSICAL, we may still
be able to get useful information from SAL_PHYSICAL_ID_INFO
- in identify_siblings(), threads/cores are a property of
the CPU, not the platform
- remove useless printk's about multi-core / thread
capability in identify_siblings(), as that information
is readily available in /proc/cpuinfo, and printing for
the BSP only adds little value
- smp_num_siblings is now meaningful if any CPU in the
system supports threads, not just the BSP
- expose "physical id" field, even on CPUs that are not
multi-core / multi-threaded (as long as we have a valid
value). Now we know what sockets Madisons live in too.
Signed-off-by: Alex Chiang <achiang@hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Add the BSS to the resource tree just as kernel text and kernel data are in
the resource tree. The main reason behind this is to avoid crashkernel
reservation in that area.
While it's not strictly necessary to have the BSS in the resource tree (the
actual collision detection is done in the reserve_bootmem() function before),
the usage of the BSS resource should be presented to the user in /proc/iomem
just as Kernel data and Kernel code.
Note: The patch currently is only implemented for x86 and ia64 (because
efi_initialize_iomem_resources() has the same signature on i386 and ia64).
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Bernhard Walle <bwalle@suse.de>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Vivek Goyal <vgoyal@in.ibm.com>
Cc: <linux-arch@vger.kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch adapts IA64 to use the generic parse_crashkernel() function instead
of its own parsing for the crashkernel command line.
Because the total amount of System RAM must be known when calling this
function, efi_memmap_init() is modified to return its accumulated total_memory
variable.
Also, the crashkernel handling is moved in an own function in
arch/ia64/kernel/setup.c to make the code more readable.
[kamalesh@linux.vnet.ibm.com: build fix]
Signed-off-by: Bernhard Walle <bwalle@suse.de>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Vivek Goyal <vgoyal@in.ibm.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
d5a7430ddc missed a spot where we
use cpu_sibling_map and cpu_core_map. These don't exist on a
uni-processor build. Wrap #ifdef CONFIG_SMP ... #endif around it.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Convert cpu_sibling_map from a static array sized by NR_CPUS to a per_cpu
variable. This saves sizeof(cpumask_t) * NR unused cpus. Access is mostly
from startup and CPU HOTPLUG functions.
Signed-off-by: Mike Travis <travis@sgi.com>
Cc: Andi Kleen <ak@suse.de>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: "Siddha, Suresh B" <suresh.b.siddha@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch cleans up the `enable early console for SKI' patch
(471e7a4484), and
1. potentially allows the gensparse_defconfig to work again.
(there are other problems running a generic kernel on Ski)
2. fixes the `console registered twice' problem.
3. Cleans up the code by moving the `extern hpsim_cons' declaration to
a new asm/hpsim.h file.
Thanks to Jes for comments.
Signed-off-by: Peter Chubb <peterc@gelato.unsw.edu.au>
Signed-off-by: Tony Luck <tony.luck@intel.com>
The pending interrupts can be remaining at boot up time on some
platform. This will cause spurious interrupts when interrupt is
enabled for the first time. This patch clears IVR at the CPU
initialization to eliminate such spurious interrupts.
Signed-off-by: Kenji Kaneshige <kaneshige.kenji@jp.fujitsu.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
When using Ski to debug early startup, it's a bit of a pain not to
have printk.
This patch enables the simulated console very early.
It may be worth conditionalising on the command line... but this is
enough for now.
Signed-off-by: Peter Chubb <peterc@gelato.unsw.edu.au>
Signed-off-by: Tony Luck <tony.luck@intel.com>
This patch removes a few duplicate includes from arch/ia64/
Acked-by: Jes Sorensen <jes@sgi.com>
Signed-off-by: Jesper Juhl <jesper.juhl@gmail.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Parse the machvec command line option outside of the early_param()
so that ia64_mv is set before any console intialisation that
may result from early_param parsing.
Signed-off-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Since Ingo's recent scheduler rewrite which was merged as commit
0437e109e1 sched_cacheflush is unused.
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Beacuse SERIAL_PORT_DFNS is removed from include/asm-i386/serial.h and
include/asm-x86_64/serial.h. the serial8250_ports need to be probed late in
serial initializing stage. the console_init=>serial8250_console_init=>
register_console=>serial8250_console_setup will return -ENDEV, and console
ttyS0 can not be enabled at that time. need to wait till uart_add_one_port in
drivers/serial/serial_core.c to call register_console to get console ttyS0.
that is too late.
Make early_uart to use early_param, so uart console can be used earlier. Make
it to be bootconsole with CON_BOOT flag, so can use console handover feature.
and it will switch to corresponding normal serial console automatically.
new command line will be:
console=uart8250,io,0x3f8,9600n8
console=uart8250,mmio,0xff5e0000,115200n8
or
earlycon=uart8250,io,0x3f8,9600n8
earlycon=uart8250,mmio,0xff5e0000,115200n8
it will print in very early stage:
Early serial console at I/O port 0x3f8 (options '9600n8')
console [uart0] enabled
later for console it will print:
console handover: boot [uart0] -> real [ttyS0]
Signed-off-by: <yinghai.lu@sun.com>
Cc: Andi Kleen <ak@suse.de>
Cc: Bjorn Helgaas <bjorn.helgaas@hp.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Gerd Hoffmann <kraxel@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
the SMP load-balancer uses the boot-time migration-cost estimation
code to attempt to improve the quality of balancing. The reason for
this code is that the discrete priority queues do not preserve
the order of scheduling accurately, so the load-balancer skips
tasks that were running on a CPU 'recently'.
this code is fundamental fragile: the boot-time migration cost detector
doesnt really work on systems that had large L3 caches, it caused boot
delays on large systems and the whole cache-hot concept made the
balancing code pretty undeterministic as well.
(and hey, i wrote most of it, so i can say it out loud that it sucks ;-)
under CFS the same purpose of cache affinity can be achieved without
any special cache-hot special-case: tasks are sorted in the 'timeline'
tree and the SMP balancer picks tasks from the left side of the
tree, thus the most cache-cold task is balanced automatically.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Section mismatch: reference to ...
.init.text:prefill_possible_map from .text between 'setup_per_cpu_areas' and 'cpu_init'
.init.text:iosapic_override_isa_irq from .text between 'iosapic_init' and 'iosapic_remove'
Signed-off-by: Tony Luck <tony.luck@intel.com>
My patch: git commit=95235ca2c20ac0b31a8eb39e2d599bcc3e9c9a10 introduced a bug
in IA64 cpuinfo output.
Patch changed the proc_freq from 1HZ resolution to 1KHz resolution, but left
format string unchanged at " %lu.%06lu". Below is the fix.
Thanks to Bjorn for catching this.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
If a system consists of mixed processor types, kmalloc()
can be called before the per-cpu data page is initialized.
If the slab contains sufficient memory, then kmalloc() works
ok. However, if the slabs are empty, slab calls the memory
allocator. This requires per-cpu data (NODE_DATA()) & the
cpu dies.
Also noted by Russ Anderson who had a very similar patch.
Signed-off-by: Jack Steiner <steiner@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
efi_initialize_iomem_resources() is declared in both include/linux/efi.h
and arch/ia64/kernel/setup.c. This patch removes the latter.
Signed-off-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Tony Luck <tony.luck@intel.com>
The address where the ELF core header is stored is passed to the secondary
kernel as a kernel command line option. The memory area for this header is
also marked as a separate EFI memory descriptor on ia64.
The separate EFI memory descriptor is at the moment of the type
EFI_UNUSABLE_MEMORY. With such a type the secondary kernel skips over the
entire memory granule (config option, 16M or 64M) when detecting memory.
If we are lucky we will just lose some memory, but if we happen to have
data in the same granule (such as an initramfs image), then this data will
never get mapped and the kernel bombs out when trying to access it.
So this is an attempt to fix this by changing the EFI memory descriptor
type into EFI_LOADER_DATA. This type is the same type used for the kernel
data and for initramfs. In the secondary kernel we then handle the ELF
core header data the same way as we handle the initramfs image.
This patch contains the kernel changes to make this happen. Pretty
straightforward, we reserve the area in reserve_memory(). The address for
the area comes from the kernel command line and the size comes from the
specialized EFI parsing function vmcore_find_descriptor_size().
The kexec-tools-testing code for this can be found here:
http://lists.osdl.org/pipermail/fastboot/2007-February/005983.html
Signed-off-by: Magnus Damm <magnus@valinux.co.jp>
Cc: Simon Horman <horms@verge.net.au>
Cc: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Tony Luck <tony.luck@intel.com>
1. Rename saved_command_line into boot_command_line.
2. Set command_line as __initdata.
[akpm@osdl.org: move some declarations to the right place]
Signed-off-by: Alon Bar-Lev <alon.barlev@gmail.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It's not efficient to use a per-cpu variable just to store
how many physical stack register a cpu has. Ever since the
incarnation of ia64 up till upcoming Montecito processor, that
variable has "glued" to 96. Having a variable in memory means
that the kernel is burning an extra cacheline access on every
syscall and kernel exit path. Such "static" value is better
served with the instruction patching utility exists today.
Convert ia64_phys_stacked_size_p8 into dynamic insn patching.
This also has a pleasant side effect of eliminating access to
per-cpu area while psr.ic=0 in the kernel exit path. (fixable
for per-cpu DTC work, but why bother?)
There are some concerns with the default value that the instruc-
tion encoded in the kernel image. It shouldn't be concerned.
The reasons are:
(1) cpu_init() is called at CPU initialization. In there, we
find out physical stack register size from PAL and patch
two instructions in kernel exit code. The code in question
can not be executed before the patching is done.
(2) current implementation stores zero in ia64_phys_stacked_size_p8,
and that's what the current kernel exit path loads the value with.
With the new code, it is equivalent that we store reg size 96
in ia64_phys_stacked_size_p8, thus creating a better safety net.
Given (1) above can never fail, having (2) is just a bonus.
All in all, this patch allow one less memory reference in the kernel
exit path, thus reducing syscall and interrupt return latency; and
avoid polluting potential useful data in the CPU cache.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Some patches have turned up on xen-devel recently to convert strcpy()
to safer alternatives and so forth. While reviewing those patches
I noticed that the features string building could be cleaned up.
This patch uses snprintf() instead of strcpy() and direct character
pointer manipulation. It makes the features string building safe and
gets rid of the special case for features output in show_cpuinfo()
Additionally I removed the (int) cast of ARRAY_SIZE, which seems to
serve no purpose.
Signed-off-by: Aron Griffis <aron@hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Stephane thought he saw a problem here (but was just confused
by the return value from ia64_pal_get_brand_info()). But we
should be more defensive here in case an prototype PAL for
a future processor doesn't implement this PAL call.
Signed-off-by: Tony Luck <tony.luck@intel.com>
There seems to be a value in both allowing the kernel to determine
the base offset of the crashkernel automatically and allowing
users's to sepcify it.
The old behaviour on ia64, which is still the current behaviour on
most architectures is for the user to always specify the address.
Recently ia64 was changed so that it is always automatically determined.
With this patch the kernel automatically determines the offset if
the supplied value is 0, otherwise it uses the value provided.
This should probably be backed by a documentation change.
Signed-Off-By: Simon Horman <horms@verge.net.au>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Actually, on reflection I think that there is a good case for
keeping the options separate. I am thinking particularly of people
who want a very small crashdump kernel and thus don't want to compile
in kexec.
The patch below should fix things up so that all valid combinations of
KEXEC, CRASH_DUMP and VMCORE compile cleanly - VMCORE depends on
CRASH_DUMP which is why I said valid combinations. In a nutshell
it just untangles unrelated code and switches around a few defines.
Please note that it creats a new file, arch/ia64/kernel/crash_dump.c
This is in keeping with the i386 implementation.
Signed-off-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Changes and updates.
1. Remove fake rendz path and related code according to discuss with Khalid Aziz.
2. fc.i offset fix in relocate_kernel.S.
3. iospic shutdown code eoi and mask race fix from Fujitsu.
4. Warm boot hook in machine_kexec to SN SAL code from Jack Steiner.
5. Send slave to SAL slave loop patch from Jay Lan.
6. Kdump on non-recoverable MCA event patch from Jay Lan
7. Use CTL_UNNUMBERED in kdump_on_init sysctl.
Signed-off-by: Zou Nan hai <nanhai.zou@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
The check to see if the firmware drops interrupts during a
SAL_CACHE_FLUSH is done to early in the boot. SAL_CACHE_FLUSH expects
to be able to make PAL calls in virtual mode, on some cell based
machines a fault occurs causing a MCA. This patch moves the check
after mmu_context_init so the TLB and VHPT are properly setup.
Signed-off-by Troy Heber <troy.heber@hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
asm/serial.h is supposed to contain the definitions for the architecture
specific 8250 ports for the 8250 driver. It may also define BASE_BAUD,
but this is the base baud for the architecture specific ports _only_.
Therefore, nothing other than the 8250 driver should be including this
header file. In order to move towards this goal, here is a patch which
removes some of the more obvious incorrect includes of the file.
Acked-by: Paul Fulghum <paulkf@microgate.com>
Acked-by: Tony Luck <tony.luck@intel.com>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-by: David Howells <dhowells@redhat.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
screen_info.h doesn't have anything to do with the tty layer and shouldn't be
included by tty.h. This patches removes the include and modifies all users to
directly include screen_info.h. struct screen_info is mainly used to
communicate with the console drivers in drivers/video/console. Note that this
patch touches every arch and I have no way of testing it. If there is a
mistake the worst thing that will happen is a compile error.
[akpm@osdl.org: fix arm build]
[akpm@osdl.org: fix alpha build]
Signed-off-by: Jon Smirl <jonsmir@gmail.com>
Signed-off-by: Antonino Daplas <adaplas@pol.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
One more trivial, stand-alone patch from the Xen/ia64 review. Sanity
check usage of the reserved region numbers.
Signed-off-by: Alex Williamson <alex.williamson@hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Linux ia64 port tried to decode the processor family number
to something human-readable, but Intel brandnames don't change
synchronously with updates to the family number. Adopt a more
i386-like approach and just print the family number in decimal.
Add a new field "model name" that uses PAL_BRAND_INFO to find
the official name for the cpu, or on older systems, falls back
to using the well-known codenames (Merced, McKinley, Madison).
Signed-off-by: Tony Luck <tony.luck@intel.com>
Almost all users of the table addresses from the EFI system table want
physical addresses. So rather than doing the pa->va->pa conversion, just keep
physical addresses in struct efi.
This fixes a DMI bug: the efi structure contained the physical SMBIOS address
on x86 but the virtual address on ia64, so dmi_scan_machine() used ioremap()
on a virtual address on ia64.
This is essentially the same as an earlier patch by Matt Tolentino:
http://marc.theaimsgroup.com/?l=linux-kernel&m=112130292316281&w=2
except that this changes all table addresses, not just ACPI addresses.
Matt's original patch was backed out because it caused MCAs on HP sx1000
systems. That problem is resolved by the ioremap() attribute checking added
for ia64.
Signed-off-by: Bjorn Helgaas <bjorn.helgaas@hp.com>
Cc: Matt Domsch <Matt_Domsch@dell.com>
Cc: "Tolentino, Matthew E" <matthew.e.tolentino@intel.com>
Cc: "Brown, Len" <len.brown@intel.com>
Cc: Andi Kleen <ak@muc.de>
Acked-by: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Enable DMI table parsing on ia64.
Andi Kleen has a patch in his x86_64 tree which enables the use of i386
dmi_scan.c on x86_64. dmi_scan.c functions are being used by the
drivers/char/ipmi/ipmi_si_intf.c driver for autodetecting the ports or
memory spaces where the IPMI controllers may be found.
This patch adds equivalent changes for ia64 as to what is in the x86_64
tree. In addition, I reworked the DMI detection, such that on EFI-capable
systems, it uses the efi.smbios pointer to find the table, rather than
brute-force searching from 0xF0000. On non-EFI systems, it continues the
brute-force search.
My test system, an Intel S870BN4 'Tiger4', aka Dell PowerEdge 7250, with
latest BIOS, does not list the IPMI controller in the ACPI namespace, nor
does it have an ACPI SPMI table. Also note, currently shipping Dell x8xx
EM64T servers don't have these either, so DMI is the only method for
obtaining the address of the IPMI controller.
Signed-off-by: Matt Domsch <Matt_Domsch@dell.com>
Acked-by: "Luck, Tony" <tony.luck@intel.com>
Cc: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
* 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6:
[IA64] New IA64 core/thread detection patch
[IA64] Increase max node count on SN platforms
[IA64] Increase max node count on SN platforms
[IA64] Increase max node count on SN platforms
[IA64] Increase max node count on SN platforms
[IA64] Tollhouse HP: IA64 arch changes
[IA64] cleanup dig_irq_init
[IA64] MCA recovery: kernel context recovery table
IA64: Use early_parm to handle mvec_name and nomca
[IA64] move patchlist and machvec into init section
[IA64] add init declaration - nolwsys
[IA64] add init declaration - gate page functions
[IA64] add init declaration to memory initialization functions
[IA64] add init declaration to cpu initialization functions
[IA64] add __init declaration to mca functions
[IA64] Ignore disabled Local SAPIC Affinity Structure in SRAT
[IA64] sn_check_intr: use ia64_get_irr()
[IA64] fix ia64 is_hugepage_only_range
include/linux/platform.h contained nothing that was actually used except
the default_idle() prototype, and is therefore removed by this patch.
This patch does the following with the platform specific default_idle()
functions on different architectures:
- remove the unused function:
- parisc
- sparc64
- make the needlessly global function static:
- arm
- h8300
- m68k
- m68knommu
- s390
- v850
- x86_64
- add a prototype in asm/system.h:
- cris
- i386
- ia64
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Acked-by: Patrick Mochel <mochel@digitalimplant.org>
Acked-by: Kyle McMartin <kyle@parisc-linux.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I'm not sure of the worthiness of this idea, so please consider it an RFC.
Its key merits are:
* Reuse existing infrastructure
* Greatly tightens up the parsing of nomca
* Greatly simplifies the parsing of machvec
Addition cleanup (moving setup_mvec() to machvec.c) by Ken Chen.
Signed-Off-By: Horms <horms@verge.net.au>
Signed-Off-By: Tony Luck <tony.luck@intel.com>
Add init declaration to variables/functions used for memory
initialization. I don't think they would clash with memory
hotplug. If they do, please yell.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Add init declaration to cpu initialization functions.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Have a facility to account for potentially hot-pluggable CPUs. ACPI doesnt
give a determinstic method to find hot-pluggable CPUs. Hence we use 2 methods
to assist.
- BIOS can mark potentially hot-pluggable CPUs as disabled in the MADT tables.
- User can specify the number of hot-pluggable CPUs via parameter
additional_cpus=X
The option is enabled only if ACPI_CONFIG_HOTPLUG_CPU=y which enables the
physical hotplug option. Without which user can still use logical onlining
and offlining of CPUs by enabling CONFIG_HOTPLUG_CPU=y
Adds more bits to cpu_possible_map for potentially hot-pluggable cpus.
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
There were two problems with enabling the PRINTK_TIME config
option:
1) The first calls to printk() occur before per-cpu data virtual
address is pinned into the TLB, so sched_clock() can fault.
2) sched_clock() is based on ar.itc, which may not be synchronized
across cpus.
Ken Chen started this patch, Tony Luck tinkered with it, and Jes
Sorensen perfected it.
Signed-off-by: Tony Luck <tony.luck@intel.com>
)
From: Ingo Molnar <mingo@elte.hu>
This is the latest version of the scheduler cache-hot-auto-tune patch.
The first problem was that detection time scaled with O(N^2), which is
unacceptable on larger SMP and NUMA systems. To solve this:
- I've added a 'domain distance' function, which is used to cache
measurement results. Each distance is only measured once. This means
that e.g. on NUMA distances of 0, 1 and 2 might be measured, on HT
distances 0 and 1, and on SMP distance 0 is measured. The code walks
the domain tree to determine the distance, so it automatically follows
whatever hierarchy an architecture sets up. This cuts down on the boot
time significantly and removes the O(N^2) limit. The only assumption
is that migration costs can be expressed as a function of domain
distance - this covers the overwhelming majority of existing systems,
and is a good guess even for more assymetric systems.
[ People hacking systems that have assymetries that break this
assumption (e.g. different CPU speeds) should experiment a bit with
the cpu_distance() function. Adding a ->migration_distance factor to
the domain structure would be one possible solution - but lets first
see the problem systems, if they exist at all. Lets not overdesign. ]
Another problem was that only a single cache-size was used for measuring
the cost of migration, and most architectures didnt set that variable
up. Furthermore, a single cache-size does not fit NUMA hierarchies with
L3 caches and does not fit HT setups, where different CPUs will often
have different 'effective cache sizes'. To solve this problem:
- Instead of relying on a single cache-size provided by the platform and
sticking to it, the code now auto-detects the 'effective migration
cost' between two measured CPUs, via iterating through a wide range of
cachesizes. The code searches for the maximum migration cost, which
occurs when the working set of the test-workload falls just below the
'effective cache size'. I.e. real-life optimized search is done for
the maximum migration cost, between two real CPUs.
This, amongst other things, has the positive effect hat if e.g. two
CPUs share a L2/L3 cache, a different (and accurate) migration cost
will be found than between two CPUs on the same system that dont share
any caches.
(The reliable measurement of migration costs is tricky - see the source
for details.)
Furthermore i've added various boot-time options to override/tune
migration behavior.
Firstly, there's a blanket override for autodetection:
migration_cost=1000,2000,3000
will override the depth 0/1/2 values with 1msec/2msec/3msec values.
Secondly, there's a global factor that can be used to increase (or
decrease) the autodetected values:
migration_factor=120
will increase the autodetected values by 20%. This option is useful to
tune things in a workload-dependent way - e.g. if a workload is
cache-insensitive then CPU utilization can be maximized by specifying
migration_factor=0.
I've tested the autodetection code quite extensively on x86, on 3
P3/Xeon/2MB, and the autodetected values look pretty good:
Dual Celeron (128K L2 cache):
---------------------
migration cost matrix (max_cache_size: 131072, cpu: 467 MHz):
---------------------
[00] [01]
[00]: - 1.7(1)
[01]: 1.7(1) -
---------------------
cacheflush times [2]: 0.0 (0) 1.7 (1784008)
---------------------
Here the slow memory subsystem dominates system performance, and even
though caches are small, the migration cost is 1.7 msecs.
Dual HT P4 (512K L2 cache):
---------------------
migration cost matrix (max_cache_size: 524288, cpu: 2379 MHz):
---------------------
[00] [01] [02] [03]
[00]: - 0.4(1) 0.0(0) 0.4(1)
[01]: 0.4(1) - 0.4(1) 0.0(0)
[02]: 0.0(0) 0.4(1) - 0.4(1)
[03]: 0.4(1) 0.0(0) 0.4(1) -
---------------------
cacheflush times [2]: 0.0 (33900) 0.4 (448514)
---------------------
Here it can be seen that there is no migration cost between two HT
siblings (CPU#0/2 and CPU#1/3 are separate physical CPUs). A fast memory
system makes inter-physical-CPU migration pretty cheap: 0.4 msecs.
8-way P3/Xeon [2MB L2 cache]:
---------------------
migration cost matrix (max_cache_size: 2097152, cpu: 700 MHz):
---------------------
[00] [01] [02] [03] [04] [05] [06] [07]
[00]: - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[01]: 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[02]: 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[03]: 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[04]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1)
[05]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1)
[06]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1)
[07]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) -
---------------------
cacheflush times [2]: 0.0 (0) 19.2 (19281756)
---------------------
This one has huge caches and a relatively slow memory subsystem - so the
migration cost is 19 msecs.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Cc: <wilder@us.ibm.com>
Signed-off-by: John Hawkes <hawkes@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add per-arch sched_cacheflush() which is a write-back cacheflush used by
the migration-cost calibration code at bootup time.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
arch/ia64/kernel/setup.c: In function `show_cpuinfo':
arch/ia64/kernel/setup.c:576: warning: long unsigned int format, different type arg (arg 12)
arch/ia64/kernel/setup.c:576: warning: long unsigned int format, different type arg (arg 13)
Introduced by 95235ca2c2
Signed-off-by: Tony Luck <tony.luck@intel.com>
What is the value shown in "cpu MHz" of /proc/cpuinfo when CPUs are capable of
changing frequency?
Today the answer is: It depends.
On i386:
SMP kernel - It is always the boot frequency
UP kernel - Scales with the frequency change and shows that was last set.
On x86_64:
There is one single variable cpu_khz that gets written by all the CPUs. So,
the frequency set by last CPU will be seen on /proc/cpuinfo of all the
CPUs in the system. What you see also depends on whether you have constant_tsc
capable CPU or not.
On ia64:
It is always boot time frequency of a particular CPU that gets displayed.
The patch below changes this to:
Show the last known frequency of the particular CPU, when cpufreq is present. If
cpu doesnot support changing of frequency through cpufreq, then boot frequency
will be shown. The patch affects i386, x86_64 and ia64 architectures.
Signed-off-by: Venkatesh Pallipadi<venkatesh.pallipadi@intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
