x86_64: prepare shared mm/numa.c
authorThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:16:14 +0000 (11:16 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:16:14 +0000 (11:16 +0200)
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86_64/mm/Makefile
arch/x86_64/mm/numa.c [deleted file]
arch/x86_64/mm/numa_64.c [new file with mode: 0644]

index a7fedc9cbb0526893a86a91f59f30bee7796d0f2..921b756ed7c01607ee6976d4de2730133f195909 100644 (file)
@@ -4,7 +4,7 @@
 
 obj-y   := init_64.o fault.o ioremap_64.o extable.o pageattr.o mmap_64.o
 obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o
-obj-$(CONFIG_NUMA) += numa.o
+obj-$(CONFIG_NUMA) += numa_64.o
 obj-$(CONFIG_K8_NUMA) += k8topology.o
 obj-$(CONFIG_ACPI_NUMA) += srat_64.o
 
diff --git a/arch/x86_64/mm/numa.c b/arch/x86_64/mm/numa.c
deleted file mode 100644 (file)
index 6da2355..0000000
+++ /dev/null
@@ -1,648 +0,0 @@
-/* 
- * 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;
-unsigned long __initdata nodemap_addr;
-unsigned long __initdata nodemap_size;
-
-
-/*
- * 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;
-
-       memset(memnodemap, 0xff, memnodemapsize);
-       for (i = 0; i < numnodes; i++) {
-               addr = nodes[i].start;
-               end = nodes[i].end;
-               if (addr >= end)
-                       continue;
-               if ((end >> shift) >= memnodemapsize)
-                       return 0;
-               do {
-                       if (memnodemap[addr >> shift] != 0xff)
-                               return -1;
-                       memnodemap[addr >> shift] = i;
-                       addr += (1UL << shift);
-               } while (addr < end);
-               res = 1;
-       } 
-       return res;
-}
-
-static int __init allocate_cachealigned_memnodemap(void)
-{
-       unsigned long pad, pad_addr;
-
-       memnodemap = memnode.embedded_map;
-       if (memnodemapsize <= 48)
-               return 0;
-
-       pad = L1_CACHE_BYTES - 1;
-       pad_addr = 0x8000;
-       nodemap_size = pad + memnodemapsize;
-       nodemap_addr = find_e820_area(pad_addr, end_pfn<<PAGE_SHIFT,
-                                     nodemap_size);
-       if (nodemap_addr == -1UL) {
-               printk(KERN_ERR
-                      "NUMA: Unable to allocate Memory to Node hash map\n");
-               nodemap_addr = nodemap_size = 0;
-               return -1;
-       }
-       pad_addr = (nodemap_addr + pad) & ~pad;
-       memnodemap = phys_to_virt(pad_addr);
-
-       printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
-              nodemap_addr, nodemap_addr + nodemap_size);
-       return 0;
-}
-
-/*
- * The LSB of all start and end addresses in the node map is the value of the
- * maximum possible shift.
- */
-static int __init
-extract_lsb_from_nodes (const struct bootnode *nodes, int numnodes)
-{
-       int i, nodes_used = 0;
-       unsigned long start, end;
-       unsigned long bitfield = 0, memtop = 0;
-
-       for (i = 0; i < numnodes; i++) {
-               start = nodes[i].start;
-               end = nodes[i].end;
-               if (start >= end)
-                       continue;
-               bitfield |= start;
-               nodes_used++;
-               if (end > memtop)
-                       memtop = end;
-       }
-       if (nodes_used <= 1)
-               i = 63;
-       else
-               i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
-       memnodemapsize = (memtop >> i)+1;
-       return i;
-}
-
-int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
-{
-       int shift;
-
-       shift = extract_lsb_from_nodes(nodes, numnodes);
-       if (allocate_cachealigned_memnodemap())
-               return -1;
-       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);
-#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;
-
-       start_pfn = node_start_pfn(nodeid);
-       end_pfn = node_end_pfn(nodeid);
-
-       Dprintk(KERN_INFO "Setting up memmap for node %d %lx-%lx\n",
-               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
-/* Numa emulation */
-char *cmdline __initdata;
-
-/*
- * Setups up nid to range from addr to addr + size.  If the end boundary is
- * greater than max_addr, then max_addr is used instead.  The return value is 0
- * if there is additional memory left for allocation past addr and -1 otherwise.
- * addr is adjusted to be at the end of the node.
- */
-static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
-                                  u64 size, u64 max_addr)
-{
-       int ret = 0;
-       nodes[nid].start = *addr;
-       *addr += size;
-       if (*addr >= max_addr) {
-               *addr = max_addr;
-               ret = -1;
-       }
-       nodes[nid].end = *addr;
-       node_set(nid, node_possible_map);
-       printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
-              nodes[nid].start, nodes[nid].end,
-              (nodes[nid].end - nodes[nid].start) >> 20);
-       return ret;
-}
-
-/*
- * Splits num_nodes nodes up equally starting at node_start.  The return value
- * is the number of nodes split up and addr is adjusted to be at the end of the
- * last node allocated.
- */
-static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
-                                     u64 max_addr, int node_start,
-                                     int num_nodes)
-{
-       unsigned int big;
-       u64 size;
-       int i;
-
-       if (num_nodes <= 0)
-               return -1;
-       if (num_nodes > MAX_NUMNODES)
-               num_nodes = MAX_NUMNODES;
-       size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
-              num_nodes;
-       /*
-        * Calculate the number of big nodes that can be allocated as a result
-        * of consolidating the leftovers.
-        */
-       big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
-             FAKE_NODE_MIN_SIZE;
-
-       /* Round down to nearest FAKE_NODE_MIN_SIZE. */
-       size &= FAKE_NODE_MIN_HASH_MASK;
-       if (!size) {
-               printk(KERN_ERR "Not enough memory for each node.  "
-                      "NUMA emulation disabled.\n");
-               return -1;
-       }
-
-       for (i = node_start; i < num_nodes + node_start; i++) {
-               u64 end = *addr + size;
-               if (i < big)
-                       end += FAKE_NODE_MIN_SIZE;
-               /*
-                * The final node can have the remaining system RAM.  Other
-                * nodes receive roughly the same amount of available pages.
-                */
-               if (i == num_nodes + node_start - 1)
-                       end = max_addr;
-               else
-                       while (end - *addr - e820_hole_size(*addr, end) <
-                              size) {
-                               end += FAKE_NODE_MIN_SIZE;
-                               if (end > max_addr) {
-                                       end = max_addr;
-                                       break;
-                               }
-                       }
-               if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
-                       break;
-       }
-       return i - node_start + 1;
-}
-
-/*
- * Splits the remaining system RAM into chunks of size.  The remaining memory is
- * always assigned to a final node and can be asymmetric.  Returns the number of
- * nodes split.
- */
-static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
-                                     u64 max_addr, int node_start, u64 size)
-{
-       int i = node_start;
-       size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
-       while (!setup_node_range(i++, nodes, addr, size, max_addr))
-               ;
-       return i - node_start;
-}
-
-/*
- * Sets up the system RAM area from start_pfn to end_pfn according to the
- * numa=fake command-line option.
- */
-static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
-{
-       struct bootnode nodes[MAX_NUMNODES];
-       u64 addr = start_pfn << PAGE_SHIFT;
-       u64 max_addr = end_pfn << PAGE_SHIFT;
-       int num_nodes = 0;
-       int coeff_flag;
-       int coeff = -1;
-       int num = 0;
-       u64 size;
-       int i;
-
-       memset(&nodes, 0, sizeof(nodes));
-       /*
-        * If the numa=fake command-line is just a single number N, split the
-        * system RAM into N fake nodes.
-        */
-       if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
-               num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0,
-                                               simple_strtol(cmdline, NULL, 0));
-               if (num_nodes < 0)
-                       return num_nodes;
-               goto out;
-       }
-
-       /* Parse the command line. */
-       for (coeff_flag = 0; ; cmdline++) {
-               if (*cmdline && isdigit(*cmdline)) {
-                       num = num * 10 + *cmdline - '0';
-                       continue;
-               }
-               if (*cmdline == '*') {
-                       if (num > 0)
-                               coeff = num;
-                       coeff_flag = 1;
-               }
-               if (!*cmdline || *cmdline == ',') {
-                       if (!coeff_flag)
-                               coeff = 1;
-                       /*
-                        * Round down to the nearest FAKE_NODE_MIN_SIZE.
-                        * Command-line coefficients are in megabytes.
-                        */
-                       size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
-                       if (size)
-                               for (i = 0; i < coeff; i++, num_nodes++)
-                                       if (setup_node_range(num_nodes, nodes,
-                                               &addr, size, max_addr) < 0)
-                                               goto done;
-                       if (!*cmdline)
-                               break;
-                       coeff_flag = 0;
-                       coeff = -1;
-               }
-               num = 0;
-       }
-done:
-       if (!num_nodes)
-               return -1;
-       /* Fill remainder of system RAM, if appropriate. */
-       if (addr < max_addr) {
-               if (coeff_flag && coeff < 0) {
-                       /* Split remaining nodes into num-sized chunks */
-                       num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
-                                                        num_nodes, num);
-                       goto out;
-               }
-               switch (*(cmdline - 1)) {
-               case '*':
-                       /* Split remaining nodes into coeff chunks */
-                       if (coeff <= 0)
-                               break;
-                       num_nodes += split_nodes_equally(nodes, &addr, max_addr,
-                                                        num_nodes, coeff);
-                       break;
-               case ',':
-                       /* Do not allocate remaining system RAM */
-                       break;
-               default:
-                       /* Give one final node */
-                       setup_node_range(num_nodes, nodes, &addr,
-                                        max_addr - addr, max_addr);
-                       num_nodes++;
-               }
-       }
-out:
-       memnode_shift = compute_hash_shift(nodes, num_nodes);
-       if (memnode_shift < 0) {
-               memnode_shift = 0;
-               printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
-                      "disabled.\n");
-               return -1;
-       }
-
-       /*
-        * We need to vacate all active ranges that may have been registered by
-        * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
-        * true.  NUMA emulation has succeeded so we will not scan ACPI nodes.
-        */
-       remove_all_active_ranges();
-#ifdef CONFIG_ACPI_NUMA
-       acpi_numa = -1;
-#endif
-       for_each_node_mask(i, node_possible_map) {
-               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);
-       }
-       acpi_fake_nodes(nodes, num_nodes);
-       numa_init_array();
-       return 0;
-}
-#endif /* CONFIG_NUMA_EMU */
-
-void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
-{ 
-       int i;
-
-       nodes_clear(node_possible_map);
-
-#ifdef CONFIG_NUMA_EMU
-       if (cmdline && !numa_emulation(start_pfn, end_pfn))
-               return;
-       nodes_clear(node_possible_map);
-#endif
-
-#ifdef CONFIG_ACPI_NUMA
-       if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
-                                         end_pfn << PAGE_SHIFT))
-               return;
-       nodes_clear(node_possible_map);
-#endif
-
-#ifdef CONFIG_K8_NUMA
-       if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
-               return;
-       nodes_clear(node_possible_map);
-#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 = memnode.embedded_map;
-       memnodemap[0] = 0;
-       nodes_clear(node_online_map);
-       node_set_online(0);
-       node_set(0, node_possible_map);
-       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;
-} 
-
-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;
-
-       sparse_memory_present_with_active_regions(MAX_NUMNODES);
-       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))
-               cmdline = opt + 5;
-#endif
-#ifdef CONFIG_ACPI_NUMA
-       if (!strncmp(opt,"noacpi",6))
-               acpi_numa = -1;
-       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
diff --git a/arch/x86_64/mm/numa_64.c b/arch/x86_64/mm/numa_64.c
new file mode 100644 (file)
index 0000000..6da2355
--- /dev/null
@@ -0,0 +1,648 @@
+/* 
+ * 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;
+unsigned long __initdata nodemap_addr;
+unsigned long __initdata nodemap_size;
+
+
+/*
+ * 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;
+
+       memset(memnodemap, 0xff, memnodemapsize);
+       for (i = 0; i < numnodes; i++) {
+               addr = nodes[i].start;
+               end = nodes[i].end;
+               if (addr >= end)
+                       continue;
+               if ((end >> shift) >= memnodemapsize)
+                       return 0;
+               do {
+                       if (memnodemap[addr >> shift] != 0xff)
+                               return -1;
+                       memnodemap[addr >> shift] = i;
+                       addr += (1UL << shift);
+               } while (addr < end);
+               res = 1;
+       } 
+       return res;
+}
+
+static int __init allocate_cachealigned_memnodemap(void)
+{
+       unsigned long pad, pad_addr;
+
+       memnodemap = memnode.embedded_map;
+       if (memnodemapsize <= 48)
+               return 0;
+
+       pad = L1_CACHE_BYTES - 1;
+       pad_addr = 0x8000;
+       nodemap_size = pad + memnodemapsize;
+       nodemap_addr = find_e820_area(pad_addr, end_pfn<<PAGE_SHIFT,
+                                     nodemap_size);
+       if (nodemap_addr == -1UL) {
+               printk(KERN_ERR
+                      "NUMA: Unable to allocate Memory to Node hash map\n");
+               nodemap_addr = nodemap_size = 0;
+               return -1;
+       }
+       pad_addr = (nodemap_addr + pad) & ~pad;
+       memnodemap = phys_to_virt(pad_addr);
+
+       printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
+              nodemap_addr, nodemap_addr + nodemap_size);
+       return 0;
+}
+
+/*
+ * The LSB of all start and end addresses in the node map is the value of the
+ * maximum possible shift.
+ */
+static int __init
+extract_lsb_from_nodes (const struct bootnode *nodes, int numnodes)
+{
+       int i, nodes_used = 0;
+       unsigned long start, end;
+       unsigned long bitfield = 0, memtop = 0;
+
+       for (i = 0; i < numnodes; i++) {
+               start = nodes[i].start;
+               end = nodes[i].end;
+               if (start >= end)
+                       continue;
+               bitfield |= start;
+               nodes_used++;
+               if (end > memtop)
+                       memtop = end;
+       }
+       if (nodes_used <= 1)
+               i = 63;
+       else
+               i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
+       memnodemapsize = (memtop >> i)+1;
+       return i;
+}
+
+int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
+{
+       int shift;
+
+       shift = extract_lsb_from_nodes(nodes, numnodes);
+       if (allocate_cachealigned_memnodemap())
+               return -1;
+       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);
+#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;
+
+       start_pfn = node_start_pfn(nodeid);
+       end_pfn = node_end_pfn(nodeid);
+
+       Dprintk(KERN_INFO "Setting up memmap for node %d %lx-%lx\n",
+               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
+/* Numa emulation */
+char *cmdline __initdata;
+
+/*
+ * Setups up nid to range from addr to addr + size.  If the end boundary is
+ * greater than max_addr, then max_addr is used instead.  The return value is 0
+ * if there is additional memory left for allocation past addr and -1 otherwise.
+ * addr is adjusted to be at the end of the node.
+ */
+static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
+                                  u64 size, u64 max_addr)
+{
+       int ret = 0;
+       nodes[nid].start = *addr;
+       *addr += size;
+       if (*addr >= max_addr) {
+               *addr = max_addr;
+               ret = -1;
+       }
+       nodes[nid].end = *addr;
+       node_set(nid, node_possible_map);
+       printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
+              nodes[nid].start, nodes[nid].end,
+              (nodes[nid].end - nodes[nid].start) >> 20);
+       return ret;
+}
+
+/*
+ * Splits num_nodes nodes up equally starting at node_start.  The return value
+ * is the number of nodes split up and addr is adjusted to be at the end of the
+ * last node allocated.
+ */
+static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
+                                     u64 max_addr, int node_start,
+                                     int num_nodes)
+{
+       unsigned int big;
+       u64 size;
+       int i;
+
+       if (num_nodes <= 0)
+               return -1;
+       if (num_nodes > MAX_NUMNODES)
+               num_nodes = MAX_NUMNODES;
+       size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
+              num_nodes;
+       /*
+        * Calculate the number of big nodes that can be allocated as a result
+        * of consolidating the leftovers.
+        */
+       big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
+             FAKE_NODE_MIN_SIZE;
+
+       /* Round down to nearest FAKE_NODE_MIN_SIZE. */
+       size &= FAKE_NODE_MIN_HASH_MASK;
+       if (!size) {
+               printk(KERN_ERR "Not enough memory for each node.  "
+                      "NUMA emulation disabled.\n");
+               return -1;
+       }
+
+       for (i = node_start; i < num_nodes + node_start; i++) {
+               u64 end = *addr + size;
+               if (i < big)
+                       end += FAKE_NODE_MIN_SIZE;
+               /*
+                * The final node can have the remaining system RAM.  Other
+                * nodes receive roughly the same amount of available pages.
+                */
+               if (i == num_nodes + node_start - 1)
+                       end = max_addr;
+               else
+                       while (end - *addr - e820_hole_size(*addr, end) <
+                              size) {
+                               end += FAKE_NODE_MIN_SIZE;
+                               if (end > max_addr) {
+                                       end = max_addr;
+                                       break;
+                               }
+                       }
+               if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
+                       break;
+       }
+       return i - node_start + 1;
+}
+
+/*
+ * Splits the remaining system RAM into chunks of size.  The remaining memory is
+ * always assigned to a final node and can be asymmetric.  Returns the number of
+ * nodes split.
+ */
+static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
+                                     u64 max_addr, int node_start, u64 size)
+{
+       int i = node_start;
+       size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
+       while (!setup_node_range(i++, nodes, addr, size, max_addr))
+               ;
+       return i - node_start;
+}
+
+/*
+ * Sets up the system RAM area from start_pfn to end_pfn according to the
+ * numa=fake command-line option.
+ */
+static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
+{
+       struct bootnode nodes[MAX_NUMNODES];
+       u64 addr = start_pfn << PAGE_SHIFT;
+       u64 max_addr = end_pfn << PAGE_SHIFT;
+       int num_nodes = 0;
+       int coeff_flag;
+       int coeff = -1;
+       int num = 0;
+       u64 size;
+       int i;
+
+       memset(&nodes, 0, sizeof(nodes));
+       /*
+        * If the numa=fake command-line is just a single number N, split the
+        * system RAM into N fake nodes.
+        */
+       if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
+               num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0,
+                                               simple_strtol(cmdline, NULL, 0));
+               if (num_nodes < 0)
+                       return num_nodes;
+               goto out;
+       }
+
+       /* Parse the command line. */
+       for (coeff_flag = 0; ; cmdline++) {
+               if (*cmdline && isdigit(*cmdline)) {
+                       num = num * 10 + *cmdline - '0';
+                       continue;
+               }
+               if (*cmdline == '*') {
+                       if (num > 0)
+                               coeff = num;
+                       coeff_flag = 1;
+               }
+               if (!*cmdline || *cmdline == ',') {
+                       if (!coeff_flag)
+                               coeff = 1;
+                       /*
+                        * Round down to the nearest FAKE_NODE_MIN_SIZE.
+                        * Command-line coefficients are in megabytes.
+                        */
+                       size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
+                       if (size)
+                               for (i = 0; i < coeff; i++, num_nodes++)
+                                       if (setup_node_range(num_nodes, nodes,
+                                               &addr, size, max_addr) < 0)
+                                               goto done;
+                       if (!*cmdline)
+                               break;
+                       coeff_flag = 0;
+                       coeff = -1;
+               }
+               num = 0;
+       }
+done:
+       if (!num_nodes)
+               return -1;
+       /* Fill remainder of system RAM, if appropriate. */
+       if (addr < max_addr) {
+               if (coeff_flag && coeff < 0) {
+                       /* Split remaining nodes into num-sized chunks */
+                       num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
+                                                        num_nodes, num);
+                       goto out;
+               }
+               switch (*(cmdline - 1)) {
+               case '*':
+                       /* Split remaining nodes into coeff chunks */
+                       if (coeff <= 0)
+                               break;
+                       num_nodes += split_nodes_equally(nodes, &addr, max_addr,
+                                                        num_nodes, coeff);
+                       break;
+               case ',':
+                       /* Do not allocate remaining system RAM */
+                       break;
+               default:
+                       /* Give one final node */
+                       setup_node_range(num_nodes, nodes, &addr,
+                                        max_addr - addr, max_addr);
+                       num_nodes++;
+               }
+       }
+out:
+       memnode_shift = compute_hash_shift(nodes, num_nodes);
+       if (memnode_shift < 0) {
+               memnode_shift = 0;
+               printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
+                      "disabled.\n");
+               return -1;
+       }
+
+       /*
+        * We need to vacate all active ranges that may have been registered by
+        * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
+        * true.  NUMA emulation has succeeded so we will not scan ACPI nodes.
+        */
+       remove_all_active_ranges();
+#ifdef CONFIG_ACPI_NUMA
+       acpi_numa = -1;
+#endif
+       for_each_node_mask(i, node_possible_map) {
+               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);
+       }
+       acpi_fake_nodes(nodes, num_nodes);
+       numa_init_array();
+       return 0;
+}
+#endif /* CONFIG_NUMA_EMU */
+
+void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
+{ 
+       int i;
+
+       nodes_clear(node_possible_map);
+
+#ifdef CONFIG_NUMA_EMU
+       if (cmdline && !numa_emulation(start_pfn, end_pfn))
+               return;
+       nodes_clear(node_possible_map);
+#endif
+
+#ifdef CONFIG_ACPI_NUMA
+       if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
+                                         end_pfn << PAGE_SHIFT))
+               return;
+       nodes_clear(node_possible_map);
+#endif
+
+#ifdef CONFIG_K8_NUMA
+       if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
+               return;
+       nodes_clear(node_possible_map);
+#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 = memnode.embedded_map;
+       memnodemap[0] = 0;
+       nodes_clear(node_online_map);
+       node_set_online(0);
+       node_set(0, node_possible_map);
+       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;
+} 
+
+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;
+
+       sparse_memory_present_with_active_regions(MAX_NUMNODES);
+       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))
+               cmdline = opt + 5;
+#endif
+#ifdef CONFIG_ACPI_NUMA
+       if (!strncmp(opt,"noacpi",6))
+               acpi_numa = -1;
+       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