linux/arch/x86_64/mm/numa.c

430 lines
11 KiB
C
Raw Normal View History

/*
* Generic VM initialization for x86-64 NUMA setups.
* Copyright 2002,2003 Andi Kleen, SuSE Labs.
*/
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/string.h>
#include <linux/init.h>
#include <linux/bootmem.h>
#include <linux/mmzone.h>
#include <linux/ctype.h>
#include <linux/module.h>
#include <linux/nodemask.h>
#include <asm/e820.h>
#include <asm/proto.h>
#include <asm/dma.h>
#include <asm/numa.h>
#include <asm/acpi.h>
#ifndef Dprintk
#define Dprintk(x...)
#endif
struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
bootmem_data_t plat_node_bdata[MAX_NUMNODES];
struct memnode memnode;
unsigned char cpu_to_node[NR_CPUS] __read_mostly = {
[0 ... NR_CPUS-1] = NUMA_NO_NODE
};
unsigned char apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
};
cpumask_t node_to_cpumask[MAX_NUMNODES] __read_mostly;
int numa_off __initdata;
/*
* Given a shift value, try to populate memnodemap[]
* Returns :
* 1 if OK
* 0 if memnodmap[] too small (of shift too small)
* -1 if node overlap or lost ram (shift too big)
*/
static int __init
populate_memnodemap(const struct bootnode *nodes, int numnodes, int shift)
{
int i;
int res = -1;
unsigned long addr, end;
if (shift >= 64)
return -1;
memset(memnodemap, 0xff, sizeof(memnodemap));
for (i = 0; i < numnodes; i++) {
addr = nodes[i].start;
end = nodes[i].end;
if (addr >= end)
continue;
if ((end >> shift) >= NODEMAPSIZE)
return 0;
do {
if (memnodemap[addr >> shift] != 0xff)
return -1;
memnodemap[addr >> shift] = i;
addr += (1UL << shift);
} while (addr < end);
res = 1;
}
return res;
}
int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
{
int shift = 20;
while (populate_memnodemap(nodes, numnodes, shift + 1) >= 0)
shift++;
printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
shift);
if (populate_memnodemap(nodes, numnodes, shift) != 1) {
printk(KERN_INFO
"Your memory is not aligned you need to rebuild your kernel "
"with a bigger NODEMAPSIZE shift=%d\n",
shift);
return -1;
}
return shift;
}
#ifdef CONFIG_SPARSEMEM
int early_pfn_to_nid(unsigned long pfn)
{
return phys_to_nid(pfn << PAGE_SHIFT);
}
#endif
static void * __init
early_node_mem(int nodeid, unsigned long start, unsigned long end,
unsigned long size)
{
unsigned long mem = find_e820_area(start, end, size);
void *ptr;
if (mem != -1L)
return __va(mem);
ptr = __alloc_bootmem_nopanic(size,
SMP_CACHE_BYTES, __pa(MAX_DMA_ADDRESS));
if (ptr == 0) {
printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
size, nodeid);
return NULL;
}
return ptr;
}
/* Initialize bootmem allocator for a node */
void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
{
unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size, bootmap_start;
unsigned long nodedata_phys;
void *bootmap;
const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
start = round_up(start, ZONE_ALIGN);
printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid, start, end);
start_pfn = start >> PAGE_SHIFT;
end_pfn = end >> PAGE_SHIFT;
node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size);
if (node_data[nodeid] == NULL)
return;
nodedata_phys = __pa(node_data[nodeid]);
memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
NODE_DATA(nodeid)->node_start_pfn = start_pfn;
NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
/* Find a place for the bootmem map */
bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
bootmap = early_node_mem(nodeid, bootmap_start, end,
bootmap_pages<<PAGE_SHIFT);
if (bootmap == NULL) {
if (nodedata_phys < start || nodedata_phys >= end)
free_bootmem((unsigned long)node_data[nodeid],pgdat_size);
node_data[nodeid] = NULL;
return;
}
bootmap_start = __pa(bootmap);
Dprintk("bootmap start %lu pages %lu\n", bootmap_start, bootmap_pages);
bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
bootmap_start >> PAGE_SHIFT,
start_pfn, end_pfn);
free_bootmem_with_active_regions(nodeid, end);
reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, bootmap_pages<<PAGE_SHIFT);
[PATCH] x86_64: Reserve SRAT hotadd memory on x86-64 From: Keith Mannthey, Andi Kleen Implement memory hotadd without sparsemem. The memory in the SRAT hotadd area is just preserved instead and can be activated later. There are a few restrictions: - Only one continuous hotadd area allowed per node The main problem is dealing with the many buggy SRAT tables that are out there. The strategy here is to reject anything suspicious. Originally from Keith Mannthey, with several hacks and changes by AK and also contributions from Andrew Morton [ TBD: Problems pointed out by KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>: 1) Goto's rebuild_zonelist patch will not work if CONFIG_MEMORY_HOTPLUG=n. Rebuilding zonelist is necessary when the system has just memory < 4G at boot, and hot add memory > 4G. because x86_64 has DMA32, ZONE_NORAML is not included into zonelist at boot time if system doesn't have memory >4G at boot. [AK: should just force the higher zones at boot time when SRAT tells us] 2) zone and node's spanned_pages and present_pages are not incremented. They should be. For example, our server (ia64/Fujitsu PrimeQuest) can equip memory from 4G to 1T(maybe 2T in future), and SRAT will *always* say we have possible 1T +memory. (Microsoft requires "write all possible memory in SRAT") When we reserve memmap for possible 1T memory, Linux will not work well in +minimum 4G configuraion ;) [AK: needs limiting to 5-10% of max memory] ] Signed-off-by: Andi Kleen <ak@suse.de> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-07 17:49:18 +00:00
#ifdef CONFIG_ACPI_NUMA
srat_reserve_add_area(nodeid);
#endif
node_set_online(nodeid);
}
/* Initialize final allocator for a zone */
void __init setup_node_zones(int nodeid)
{
unsigned long start_pfn, end_pfn, memmapsize, limit;
[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
start_pfn = node_start_pfn(nodeid);
end_pfn = node_end_pfn(nodeid);
Dprintk(KERN_INFO "Setting up memmap for node %d %lx-%lx\n",
[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
nodeid, start_pfn, end_pfn);
/* Try to allocate mem_map at end to not fill up precious <4GB
memory. */
memmapsize = sizeof(struct page) * (end_pfn-start_pfn);
limit = end_pfn << PAGE_SHIFT;
#ifdef CONFIG_FLAT_NODE_MEM_MAP
NODE_DATA(nodeid)->node_mem_map =
__alloc_bootmem_core(NODE_DATA(nodeid)->bdata,
memmapsize, SMP_CACHE_BYTES,
round_down(limit - memmapsize, PAGE_SIZE),
limit);
#endif
}
void __init numa_init_array(void)
{
int rr, i;
/* There are unfortunately some poorly designed mainboards around
that only connect memory to a single CPU. This breaks the 1:1 cpu->node
mapping. To avoid this fill in the mapping for all possible
CPUs, as the number of CPUs is not known yet.
We round robin the existing nodes. */
rr = first_node(node_online_map);
for (i = 0; i < NR_CPUS; i++) {
if (cpu_to_node[i] != NUMA_NO_NODE)
continue;
numa_set_node(i, rr);
rr = next_node(rr, node_online_map);
if (rr == MAX_NUMNODES)
rr = first_node(node_online_map);
}
}
#ifdef CONFIG_NUMA_EMU
int numa_fake __initdata = 0;
/* Numa emulation */
static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
{
int i;
struct bootnode nodes[MAX_NUMNODES];
unsigned long sz = ((end_pfn - start_pfn)<<PAGE_SHIFT) / numa_fake;
/* Kludge needed for the hash function */
if (hweight64(sz) > 1) {
unsigned long x = 1;
while ((x << 1) < sz)
x <<= 1;
if (x < sz/2)
printk(KERN_ERR "Numa emulation unbalanced. Complain to maintainer\n");
sz = x;
}
memset(&nodes,0,sizeof(nodes));
for (i = 0; i < numa_fake; i++) {
nodes[i].start = (start_pfn<<PAGE_SHIFT) + i*sz;
if (i == numa_fake-1)
sz = (end_pfn<<PAGE_SHIFT) - nodes[i].start;
nodes[i].end = nodes[i].start + sz;
printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n",
i,
nodes[i].start, nodes[i].end,
(nodes[i].end - nodes[i].start) >> 20);
node_set_online(i);
}
memnode_shift = compute_hash_shift(nodes, numa_fake);
if (memnode_shift < 0) {
memnode_shift = 0;
printk(KERN_ERR "No NUMA hash function found. Emulation disabled.\n");
return -1;
}
for_each_online_node(i) {
e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
nodes[i].end >> PAGE_SHIFT);
setup_node_bootmem(i, nodes[i].start, nodes[i].end);
}
numa_init_array();
return 0;
}
#endif
void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
{
int i;
#ifdef CONFIG_NUMA_EMU
if (numa_fake && !numa_emulation(start_pfn, end_pfn))
return;
#endif
#ifdef CONFIG_ACPI_NUMA
if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
end_pfn << PAGE_SHIFT))
return;
#endif
#ifdef CONFIG_K8_NUMA
if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
return;
#endif
printk(KERN_INFO "%s\n",
numa_off ? "NUMA turned off" : "No NUMA configuration found");
printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
start_pfn << PAGE_SHIFT,
end_pfn << PAGE_SHIFT);
/* setup dummy node covering all memory */
memnode_shift = 63;
memnodemap[0] = 0;
nodes_clear(node_online_map);
node_set_online(0);
for (i = 0; i < NR_CPUS; i++)
numa_set_node(i, 0);
node_to_cpumask[0] = cpumask_of_cpu(0);
e820_register_active_regions(0, start_pfn, end_pfn);
setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
}
__cpuinit void numa_add_cpu(int cpu)
{
set_bit(cpu, &node_to_cpumask[cpu_to_node(cpu)]);
}
void __cpuinit numa_set_node(int cpu, int node)
{
cpu_pda(cpu)->nodenumber = node;
cpu_to_node[cpu] = node;
}
unsigned long __init numa_free_all_bootmem(void)
{
int i;
unsigned long pages = 0;
for_each_online_node(i) {
pages += free_all_bootmem_node(NODE_DATA(i));
}
return pages;
}
#ifdef CONFIG_SPARSEMEM
static void __init arch_sparse_init(void)
{
int i;
for_each_online_node(i)
memory_present(i, node_start_pfn(i), node_end_pfn(i));
sparse_init();
}
#else
#define arch_sparse_init() do {} while (0)
#endif
void __init paging_init(void)
{
int i;
unsigned long max_zone_pfns[MAX_NR_ZONES];
memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
max_zone_pfns[ZONE_NORMAL] = end_pfn;
arch_sparse_init();
for_each_online_node(i) {
setup_node_zones(i);
}
free_area_init_nodes(max_zone_pfns);
}
static __init int numa_setup(char *opt)
{
if (!opt)
return -EINVAL;
if (!strncmp(opt,"off",3))
numa_off = 1;
#ifdef CONFIG_NUMA_EMU
if(!strncmp(opt, "fake=", 5)) {
numa_fake = simple_strtoul(opt+5,NULL,0); ;
if (numa_fake >= MAX_NUMNODES)
numa_fake = MAX_NUMNODES;
}
#endif
#ifdef CONFIG_ACPI_NUMA
if (!strncmp(opt,"noacpi",6))
acpi_numa = -1;
[PATCH] x86_64: Reserve SRAT hotadd memory on x86-64 From: Keith Mannthey, Andi Kleen Implement memory hotadd without sparsemem. The memory in the SRAT hotadd area is just preserved instead and can be activated later. There are a few restrictions: - Only one continuous hotadd area allowed per node The main problem is dealing with the many buggy SRAT tables that are out there. The strategy here is to reject anything suspicious. Originally from Keith Mannthey, with several hacks and changes by AK and also contributions from Andrew Morton [ TBD: Problems pointed out by KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>: 1) Goto's rebuild_zonelist patch will not work if CONFIG_MEMORY_HOTPLUG=n. Rebuilding zonelist is necessary when the system has just memory < 4G at boot, and hot add memory > 4G. because x86_64 has DMA32, ZONE_NORAML is not included into zonelist at boot time if system doesn't have memory >4G at boot. [AK: should just force the higher zones at boot time when SRAT tells us] 2) zone and node's spanned_pages and present_pages are not incremented. They should be. For example, our server (ia64/Fujitsu PrimeQuest) can equip memory from 4G to 1T(maybe 2T in future), and SRAT will *always* say we have possible 1T +memory. (Microsoft requires "write all possible memory in SRAT") When we reserve memmap for possible 1T memory, Linux will not work well in +minimum 4G configuraion ;) [AK: needs limiting to 5-10% of max memory] ] Signed-off-by: Andi Kleen <ak@suse.de> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-04-07 17:49:18 +00:00
if (!strncmp(opt,"hotadd=", 7))
hotadd_percent = simple_strtoul(opt+7, NULL, 10);
#endif
return 0;
}
early_param("numa", numa_setup);
/*
* Setup early cpu_to_node.
*
* Populate cpu_to_node[] only if x86_cpu_to_apicid[],
* and apicid_to_node[] tables have valid entries for a CPU.
* This means we skip cpu_to_node[] initialisation for NUMA
* emulation and faking node case (when running a kernel compiled
* for NUMA on a non NUMA box), which is OK as cpu_to_node[]
* is already initialized in a round robin manner at numa_init_array,
* prior to this call, and this initialization is good enough
* for the fake NUMA cases.
*/
void __init init_cpu_to_node(void)
{
int i;
for (i = 0; i < NR_CPUS; i++) {
u8 apicid = x86_cpu_to_apicid[i];
if (apicid == BAD_APICID)
continue;
if (apicid_to_node[apicid] == NUMA_NO_NODE)
continue;
numa_set_node(i,apicid_to_node[apicid]);
}
}
EXPORT_SYMBOL(cpu_to_node);
EXPORT_SYMBOL(node_to_cpumask);
EXPORT_SYMBOL(memnode);
EXPORT_SYMBOL(node_data);
#ifdef CONFIG_DISCONTIGMEM
/*
* Functions to convert PFNs from/to per node page addresses.
* These are out of line because they are quite big.
* They could be all tuned by pre caching more state.
* Should do that.
*/
int pfn_valid(unsigned long pfn)
{
unsigned nid;
if (pfn >= num_physpages)
return 0;
nid = pfn_to_nid(pfn);
if (nid == 0xff)
return 0;
return pfn >= node_start_pfn(nid) && (pfn) < node_end_pfn(nid);
}
EXPORT_SYMBOL(pfn_valid);
#endif