Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Simple NUMA memory policy for the Linux kernel. |
| 3 | * |
| 4 | * Copyright 2003,2004 Andi Kleen, SuSE Labs. |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 5 | * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6 | * Subject to the GNU Public License, version 2. |
| 7 | * |
| 8 | * NUMA policy allows the user to give hints in which node(s) memory should |
| 9 | * be allocated. |
| 10 | * |
| 11 | * Support four policies per VMA and per process: |
| 12 | * |
| 13 | * The VMA policy has priority over the process policy for a page fault. |
| 14 | * |
| 15 | * interleave Allocate memory interleaved over a set of nodes, |
| 16 | * with normal fallback if it fails. |
| 17 | * For VMA based allocations this interleaves based on the |
| 18 | * offset into the backing object or offset into the mapping |
| 19 | * for anonymous memory. For process policy an process counter |
| 20 | * is used. |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 21 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 22 | * bind Only allocate memory on a specific set of nodes, |
| 23 | * no fallback. |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 24 | * FIXME: memory is allocated starting with the first node |
| 25 | * to the last. It would be better if bind would truly restrict |
| 26 | * the allocation to memory nodes instead |
| 27 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 28 | * preferred Try a specific node first before normal fallback. |
| 29 | * As a special case node -1 here means do the allocation |
| 30 | * on the local CPU. This is normally identical to default, |
| 31 | * but useful to set in a VMA when you have a non default |
| 32 | * process policy. |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 33 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 34 | * default Allocate on the local node first, or when on a VMA |
| 35 | * use the process policy. This is what Linux always did |
| 36 | * in a NUMA aware kernel and still does by, ahem, default. |
| 37 | * |
| 38 | * The process policy is applied for most non interrupt memory allocations |
| 39 | * in that process' context. Interrupts ignore the policies and always |
| 40 | * try to allocate on the local CPU. The VMA policy is only applied for memory |
| 41 | * allocations for a VMA in the VM. |
| 42 | * |
| 43 | * Currently there are a few corner cases in swapping where the policy |
| 44 | * is not applied, but the majority should be handled. When process policy |
| 45 | * is used it is not remembered over swap outs/swap ins. |
| 46 | * |
| 47 | * Only the highest zone in the zone hierarchy gets policied. Allocations |
| 48 | * requesting a lower zone just use default policy. This implies that |
| 49 | * on systems with highmem kernel lowmem allocation don't get policied. |
| 50 | * Same with GFP_DMA allocations. |
| 51 | * |
| 52 | * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between |
| 53 | * all users and remembered even when nobody has memory mapped. |
| 54 | */ |
| 55 | |
| 56 | /* Notebook: |
| 57 | fix mmap readahead to honour policy and enable policy for any page cache |
| 58 | object |
| 59 | statistics for bigpages |
| 60 | global policy for page cache? currently it uses process policy. Requires |
| 61 | first item above. |
| 62 | handle mremap for shared memory (currently ignored for the policy) |
| 63 | grows down? |
| 64 | make bind policy root only? It can trigger oom much faster and the |
| 65 | kernel is not always grateful with that. |
| 66 | could replace all the switch()es with a mempolicy_ops structure. |
| 67 | */ |
| 68 | |
| 69 | #include <linux/mempolicy.h> |
| 70 | #include <linux/mm.h> |
| 71 | #include <linux/highmem.h> |
| 72 | #include <linux/hugetlb.h> |
| 73 | #include <linux/kernel.h> |
| 74 | #include <linux/sched.h> |
| 75 | #include <linux/mm.h> |
| 76 | #include <linux/nodemask.h> |
| 77 | #include <linux/cpuset.h> |
| 78 | #include <linux/gfp.h> |
| 79 | #include <linux/slab.h> |
| 80 | #include <linux/string.h> |
| 81 | #include <linux/module.h> |
| 82 | #include <linux/interrupt.h> |
| 83 | #include <linux/init.h> |
| 84 | #include <linux/compat.h> |
| 85 | #include <linux/mempolicy.h> |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 86 | #include <linux/swap.h> |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 87 | #include <linux/seq_file.h> |
| 88 | #include <linux/proc_fs.h> |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 89 | #include <linux/migrate.h> |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 90 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 91 | #include <asm/tlbflush.h> |
| 92 | #include <asm/uaccess.h> |
| 93 | |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 94 | /* Internal flags */ |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 95 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 96 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 97 | #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */ |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 98 | |
Pekka Enberg | fcc234f | 2006-03-22 00:08:13 -0800 | [diff] [blame] | 99 | static struct kmem_cache *policy_cache; |
| 100 | static struct kmem_cache *sn_cache; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 101 | |
| 102 | #define PDprintk(fmt...) |
| 103 | |
| 104 | /* Highest zone. An specific allocation for a zone below that is not |
| 105 | policied. */ |
Christoph Lameter | 4be38e3 | 2006-01-06 00:11:17 -0800 | [diff] [blame] | 106 | int policy_zone = ZONE_DMA; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 107 | |
Andi Kleen | d42c699 | 2005-07-06 19:56:03 +0200 | [diff] [blame] | 108 | struct mempolicy default_policy = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 109 | .refcnt = ATOMIC_INIT(1), /* never free it */ |
| 110 | .policy = MPOL_DEFAULT, |
| 111 | }; |
| 112 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 113 | /* Do sanity checking on a policy */ |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 114 | static int mpol_check_policy(int mode, nodemask_t *nodes) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 115 | { |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 116 | int empty = nodes_empty(*nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 117 | |
| 118 | switch (mode) { |
| 119 | case MPOL_DEFAULT: |
| 120 | if (!empty) |
| 121 | return -EINVAL; |
| 122 | break; |
| 123 | case MPOL_BIND: |
| 124 | case MPOL_INTERLEAVE: |
| 125 | /* Preferred will only use the first bit, but allow |
| 126 | more for now. */ |
| 127 | if (empty) |
| 128 | return -EINVAL; |
| 129 | break; |
| 130 | } |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 131 | return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 132 | } |
Andi Kleen | dd942ae | 2006-02-17 01:39:16 +0100 | [diff] [blame] | 133 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 134 | /* Generate a custom zonelist for the BIND policy. */ |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 135 | static struct zonelist *bind_zonelist(nodemask_t *nodes) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 136 | { |
| 137 | struct zonelist *zl; |
Andi Kleen | dd942ae | 2006-02-17 01:39:16 +0100 | [diff] [blame] | 138 | int num, max, nd, k; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 139 | |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 140 | max = 1 + MAX_NR_ZONES * nodes_weight(*nodes); |
Andi Kleen | dd942ae | 2006-02-17 01:39:16 +0100 | [diff] [blame] | 141 | zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 142 | if (!zl) |
| 143 | return NULL; |
| 144 | num = 0; |
Andi Kleen | dd942ae | 2006-02-17 01:39:16 +0100 | [diff] [blame] | 145 | /* First put in the highest zones from all nodes, then all the next |
| 146 | lower zones etc. Avoid empty zones because the memory allocator |
| 147 | doesn't like them. If you implement node hot removal you |
| 148 | have to fix that. */ |
| 149 | for (k = policy_zone; k >= 0; k--) { |
| 150 | for_each_node_mask(nd, *nodes) { |
| 151 | struct zone *z = &NODE_DATA(nd)->node_zones[k]; |
| 152 | if (z->present_pages > 0) |
| 153 | zl->zones[num++] = z; |
| 154 | } |
| 155 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 156 | zl->zones[num] = NULL; |
| 157 | return zl; |
| 158 | } |
| 159 | |
| 160 | /* Create a new policy */ |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 161 | static struct mempolicy *mpol_new(int mode, nodemask_t *nodes) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 162 | { |
| 163 | struct mempolicy *policy; |
| 164 | |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 165 | PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 166 | if (mode == MPOL_DEFAULT) |
| 167 | return NULL; |
| 168 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
| 169 | if (!policy) |
| 170 | return ERR_PTR(-ENOMEM); |
| 171 | atomic_set(&policy->refcnt, 1); |
| 172 | switch (mode) { |
| 173 | case MPOL_INTERLEAVE: |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 174 | policy->v.nodes = *nodes; |
Andi Kleen | 8f493d7 | 2006-01-03 00:07:28 +0100 | [diff] [blame] | 175 | if (nodes_weight(*nodes) == 0) { |
| 176 | kmem_cache_free(policy_cache, policy); |
| 177 | return ERR_PTR(-EINVAL); |
| 178 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 179 | break; |
| 180 | case MPOL_PREFERRED: |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 181 | policy->v.preferred_node = first_node(*nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 182 | if (policy->v.preferred_node >= MAX_NUMNODES) |
| 183 | policy->v.preferred_node = -1; |
| 184 | break; |
| 185 | case MPOL_BIND: |
| 186 | policy->v.zonelist = bind_zonelist(nodes); |
| 187 | if (policy->v.zonelist == NULL) { |
| 188 | kmem_cache_free(policy_cache, policy); |
| 189 | return ERR_PTR(-ENOMEM); |
| 190 | } |
| 191 | break; |
| 192 | } |
| 193 | policy->policy = mode; |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 194 | policy->cpuset_mems_allowed = cpuset_mems_allowed(current); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 195 | return policy; |
| 196 | } |
| 197 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 198 | static void gather_stats(struct page *, void *, int pte_dirty); |
Christoph Lameter | fc30128 | 2006-01-18 17:42:29 -0800 | [diff] [blame] | 199 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
| 200 | unsigned long flags); |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 201 | |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 202 | /* Scan through pages checking if pages follow certain conditions. */ |
Nick Piggin | b581003 | 2005-10-29 18:16:12 -0700 | [diff] [blame] | 203 | static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd, |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 204 | unsigned long addr, unsigned long end, |
| 205 | const nodemask_t *nodes, unsigned long flags, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 206 | void *private) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 207 | { |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 208 | pte_t *orig_pte; |
| 209 | pte_t *pte; |
Hugh Dickins | 705e87c | 2005-10-29 18:16:27 -0700 | [diff] [blame] | 210 | spinlock_t *ptl; |
Hugh Dickins | 941150a | 2005-06-21 17:15:06 -0700 | [diff] [blame] | 211 | |
Hugh Dickins | 705e87c | 2005-10-29 18:16:27 -0700 | [diff] [blame] | 212 | orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 213 | do { |
Linus Torvalds | 6aab341 | 2005-11-28 14:34:23 -0800 | [diff] [blame] | 214 | struct page *page; |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 215 | unsigned int nid; |
| 216 | |
| 217 | if (!pte_present(*pte)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 218 | continue; |
Linus Torvalds | 6aab341 | 2005-11-28 14:34:23 -0800 | [diff] [blame] | 219 | page = vm_normal_page(vma, addr, *pte); |
| 220 | if (!page) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 221 | continue; |
Nick Piggin | 053837f | 2006-01-18 17:42:27 -0800 | [diff] [blame] | 222 | /* |
| 223 | * The check for PageReserved here is important to avoid |
| 224 | * handling zero pages and other pages that may have been |
| 225 | * marked special by the system. |
| 226 | * |
| 227 | * If the PageReserved would not be checked here then f.e. |
| 228 | * the location of the zero page could have an influence |
| 229 | * on MPOL_MF_STRICT, zero pages would be counted for |
| 230 | * the per node stats, and there would be useless attempts |
| 231 | * to put zero pages on the migration list. |
| 232 | */ |
Christoph Lameter | f4598c8 | 2006-01-12 01:05:20 -0800 | [diff] [blame] | 233 | if (PageReserved(page)) |
| 234 | continue; |
Linus Torvalds | 6aab341 | 2005-11-28 14:34:23 -0800 | [diff] [blame] | 235 | nid = page_to_nid(page); |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 236 | if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT)) |
| 237 | continue; |
| 238 | |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 239 | if (flags & MPOL_MF_STATS) |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 240 | gather_stats(page, private, pte_dirty(*pte)); |
Nick Piggin | 053837f | 2006-01-18 17:42:27 -0800 | [diff] [blame] | 241 | else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) |
Christoph Lameter | fc30128 | 2006-01-18 17:42:29 -0800 | [diff] [blame] | 242 | migrate_page_add(page, private, flags); |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 243 | else |
| 244 | break; |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 245 | } while (pte++, addr += PAGE_SIZE, addr != end); |
Hugh Dickins | 705e87c | 2005-10-29 18:16:27 -0700 | [diff] [blame] | 246 | pte_unmap_unlock(orig_pte, ptl); |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 247 | return addr != end; |
| 248 | } |
| 249 | |
Nick Piggin | b581003 | 2005-10-29 18:16:12 -0700 | [diff] [blame] | 250 | static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud, |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 251 | unsigned long addr, unsigned long end, |
| 252 | const nodemask_t *nodes, unsigned long flags, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 253 | void *private) |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 254 | { |
| 255 | pmd_t *pmd; |
| 256 | unsigned long next; |
| 257 | |
| 258 | pmd = pmd_offset(pud, addr); |
| 259 | do { |
| 260 | next = pmd_addr_end(addr, end); |
| 261 | if (pmd_none_or_clear_bad(pmd)) |
| 262 | continue; |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 263 | if (check_pte_range(vma, pmd, addr, next, nodes, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 264 | flags, private)) |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 265 | return -EIO; |
| 266 | } while (pmd++, addr = next, addr != end); |
| 267 | return 0; |
| 268 | } |
| 269 | |
Nick Piggin | b581003 | 2005-10-29 18:16:12 -0700 | [diff] [blame] | 270 | static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd, |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 271 | unsigned long addr, unsigned long end, |
| 272 | const nodemask_t *nodes, unsigned long flags, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 273 | void *private) |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 274 | { |
| 275 | pud_t *pud; |
| 276 | unsigned long next; |
| 277 | |
| 278 | pud = pud_offset(pgd, addr); |
| 279 | do { |
| 280 | next = pud_addr_end(addr, end); |
| 281 | if (pud_none_or_clear_bad(pud)) |
| 282 | continue; |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 283 | if (check_pmd_range(vma, pud, addr, next, nodes, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 284 | flags, private)) |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 285 | return -EIO; |
| 286 | } while (pud++, addr = next, addr != end); |
| 287 | return 0; |
| 288 | } |
| 289 | |
Nick Piggin | b581003 | 2005-10-29 18:16:12 -0700 | [diff] [blame] | 290 | static inline int check_pgd_range(struct vm_area_struct *vma, |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 291 | unsigned long addr, unsigned long end, |
| 292 | const nodemask_t *nodes, unsigned long flags, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 293 | void *private) |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 294 | { |
| 295 | pgd_t *pgd; |
| 296 | unsigned long next; |
| 297 | |
Nick Piggin | b581003 | 2005-10-29 18:16:12 -0700 | [diff] [blame] | 298 | pgd = pgd_offset(vma->vm_mm, addr); |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 299 | do { |
| 300 | next = pgd_addr_end(addr, end); |
| 301 | if (pgd_none_or_clear_bad(pgd)) |
| 302 | continue; |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 303 | if (check_pud_range(vma, pgd, addr, next, nodes, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 304 | flags, private)) |
Hugh Dickins | 91612e0 | 2005-06-21 17:15:07 -0700 | [diff] [blame] | 305 | return -EIO; |
| 306 | } while (pgd++, addr = next, addr != end); |
| 307 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 308 | } |
| 309 | |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 310 | /* Check if a vma is migratable */ |
| 311 | static inline int vma_migratable(struct vm_area_struct *vma) |
| 312 | { |
| 313 | if (vma->vm_flags & ( |
Christoph Lameter | f4598c8 | 2006-01-12 01:05:20 -0800 | [diff] [blame] | 314 | VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED)) |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 315 | return 0; |
| 316 | return 1; |
| 317 | } |
| 318 | |
| 319 | /* |
| 320 | * Check if all pages in a range are on a set of nodes. |
| 321 | * If pagelist != NULL then isolate pages from the LRU and |
| 322 | * put them on the pagelist. |
| 323 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 324 | static struct vm_area_struct * |
| 325 | check_range(struct mm_struct *mm, unsigned long start, unsigned long end, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 326 | const nodemask_t *nodes, unsigned long flags, void *private) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 327 | { |
| 328 | int err; |
| 329 | struct vm_area_struct *first, *vma, *prev; |
| 330 | |
Christoph Lameter | 90036ee | 2006-03-16 23:03:59 -0800 | [diff] [blame] | 331 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
Christoph Lameter | 90036ee | 2006-03-16 23:03:59 -0800 | [diff] [blame] | 332 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 333 | err = migrate_prep(); |
| 334 | if (err) |
| 335 | return ERR_PTR(err); |
Christoph Lameter | 90036ee | 2006-03-16 23:03:59 -0800 | [diff] [blame] | 336 | } |
Nick Piggin | 053837f | 2006-01-18 17:42:27 -0800 | [diff] [blame] | 337 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 338 | first = find_vma(mm, start); |
| 339 | if (!first) |
| 340 | return ERR_PTR(-EFAULT); |
| 341 | prev = NULL; |
| 342 | for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) { |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 343 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { |
| 344 | if (!vma->vm_next && vma->vm_end < end) |
| 345 | return ERR_PTR(-EFAULT); |
| 346 | if (prev && prev->vm_end < vma->vm_start) |
| 347 | return ERR_PTR(-EFAULT); |
| 348 | } |
| 349 | if (!is_vm_hugetlb_page(vma) && |
| 350 | ((flags & MPOL_MF_STRICT) || |
| 351 | ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) && |
| 352 | vma_migratable(vma)))) { |
Andi Kleen | 5b952b3 | 2005-09-13 01:25:08 -0700 | [diff] [blame] | 353 | unsigned long endvma = vma->vm_end; |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 354 | |
Andi Kleen | 5b952b3 | 2005-09-13 01:25:08 -0700 | [diff] [blame] | 355 | if (endvma > end) |
| 356 | endvma = end; |
| 357 | if (vma->vm_start > start) |
| 358 | start = vma->vm_start; |
Christoph Lameter | dc9aa5b | 2006-01-08 01:00:50 -0800 | [diff] [blame] | 359 | err = check_pgd_range(vma, start, endvma, nodes, |
Christoph Lameter | 38e3586 | 2006-01-08 01:01:01 -0800 | [diff] [blame] | 360 | flags, private); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 361 | if (err) { |
| 362 | first = ERR_PTR(err); |
| 363 | break; |
| 364 | } |
| 365 | } |
| 366 | prev = vma; |
| 367 | } |
| 368 | return first; |
| 369 | } |
| 370 | |
| 371 | /* Apply policy to a single VMA */ |
| 372 | static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new) |
| 373 | { |
| 374 | int err = 0; |
| 375 | struct mempolicy *old = vma->vm_policy; |
| 376 | |
| 377 | PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", |
| 378 | vma->vm_start, vma->vm_end, vma->vm_pgoff, |
| 379 | vma->vm_ops, vma->vm_file, |
| 380 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); |
| 381 | |
| 382 | if (vma->vm_ops && vma->vm_ops->set_policy) |
| 383 | err = vma->vm_ops->set_policy(vma, new); |
| 384 | if (!err) { |
| 385 | mpol_get(new); |
| 386 | vma->vm_policy = new; |
| 387 | mpol_free(old); |
| 388 | } |
| 389 | return err; |
| 390 | } |
| 391 | |
| 392 | /* Step 2: apply policy to a range and do splits. */ |
| 393 | static int mbind_range(struct vm_area_struct *vma, unsigned long start, |
| 394 | unsigned long end, struct mempolicy *new) |
| 395 | { |
| 396 | struct vm_area_struct *next; |
| 397 | int err; |
| 398 | |
| 399 | err = 0; |
| 400 | for (; vma && vma->vm_start < end; vma = next) { |
| 401 | next = vma->vm_next; |
| 402 | if (vma->vm_start < start) |
| 403 | err = split_vma(vma->vm_mm, vma, start, 1); |
| 404 | if (!err && vma->vm_end > end) |
| 405 | err = split_vma(vma->vm_mm, vma, end, 0); |
| 406 | if (!err) |
| 407 | err = policy_vma(vma, new); |
| 408 | if (err) |
| 409 | break; |
| 410 | } |
| 411 | return err; |
| 412 | } |
| 413 | |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 414 | static int contextualize_policy(int mode, nodemask_t *nodes) |
| 415 | { |
| 416 | if (!nodes) |
| 417 | return 0; |
| 418 | |
Paul Jackson | cf2a473 | 2006-01-08 01:01:54 -0800 | [diff] [blame] | 419 | cpuset_update_task_memory_state(); |
Paul Jackson | 5966514 | 2006-01-08 01:01:47 -0800 | [diff] [blame] | 420 | if (!cpuset_nodes_subset_current_mems_allowed(*nodes)) |
| 421 | return -EINVAL; |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 422 | return mpol_check_policy(mode, nodes); |
| 423 | } |
| 424 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 425 | /* Set the process memory policy */ |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 426 | long do_set_mempolicy(int mode, nodemask_t *nodes) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 427 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 428 | struct mempolicy *new; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 429 | |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 430 | if (contextualize_policy(mode, nodes)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 431 | return -EINVAL; |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 432 | new = mpol_new(mode, nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 433 | if (IS_ERR(new)) |
| 434 | return PTR_ERR(new); |
| 435 | mpol_free(current->mempolicy); |
| 436 | current->mempolicy = new; |
| 437 | if (new && new->policy == MPOL_INTERLEAVE) |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 438 | current->il_next = first_node(new->v.nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 439 | return 0; |
| 440 | } |
| 441 | |
| 442 | /* Fill a zone bitmap for a policy */ |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 443 | static void get_zonemask(struct mempolicy *p, nodemask_t *nodes) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 444 | { |
| 445 | int i; |
| 446 | |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 447 | nodes_clear(*nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 448 | switch (p->policy) { |
| 449 | case MPOL_BIND: |
| 450 | for (i = 0; p->v.zonelist->zones[i]; i++) |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 451 | node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id, |
| 452 | *nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 453 | break; |
| 454 | case MPOL_DEFAULT: |
| 455 | break; |
| 456 | case MPOL_INTERLEAVE: |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 457 | *nodes = p->v.nodes; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 458 | break; |
| 459 | case MPOL_PREFERRED: |
| 460 | /* or use current node instead of online map? */ |
| 461 | if (p->v.preferred_node < 0) |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 462 | *nodes = node_online_map; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 463 | else |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 464 | node_set(p->v.preferred_node, *nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 465 | break; |
| 466 | default: |
| 467 | BUG(); |
| 468 | } |
| 469 | } |
| 470 | |
| 471 | static int lookup_node(struct mm_struct *mm, unsigned long addr) |
| 472 | { |
| 473 | struct page *p; |
| 474 | int err; |
| 475 | |
| 476 | err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL); |
| 477 | if (err >= 0) { |
| 478 | err = page_to_nid(p); |
| 479 | put_page(p); |
| 480 | } |
| 481 | return err; |
| 482 | } |
| 483 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 484 | /* Retrieve NUMA policy */ |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 485 | long do_get_mempolicy(int *policy, nodemask_t *nmask, |
| 486 | unsigned long addr, unsigned long flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 487 | { |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 488 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 489 | struct mm_struct *mm = current->mm; |
| 490 | struct vm_area_struct *vma = NULL; |
| 491 | struct mempolicy *pol = current->mempolicy; |
| 492 | |
Paul Jackson | cf2a473 | 2006-01-08 01:01:54 -0800 | [diff] [blame] | 493 | cpuset_update_task_memory_state(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 494 | if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR)) |
| 495 | return -EINVAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 496 | if (flags & MPOL_F_ADDR) { |
| 497 | down_read(&mm->mmap_sem); |
| 498 | vma = find_vma_intersection(mm, addr, addr+1); |
| 499 | if (!vma) { |
| 500 | up_read(&mm->mmap_sem); |
| 501 | return -EFAULT; |
| 502 | } |
| 503 | if (vma->vm_ops && vma->vm_ops->get_policy) |
| 504 | pol = vma->vm_ops->get_policy(vma, addr); |
| 505 | else |
| 506 | pol = vma->vm_policy; |
| 507 | } else if (addr) |
| 508 | return -EINVAL; |
| 509 | |
| 510 | if (!pol) |
| 511 | pol = &default_policy; |
| 512 | |
| 513 | if (flags & MPOL_F_NODE) { |
| 514 | if (flags & MPOL_F_ADDR) { |
| 515 | err = lookup_node(mm, addr); |
| 516 | if (err < 0) |
| 517 | goto out; |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 518 | *policy = err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 519 | } else if (pol == current->mempolicy && |
| 520 | pol->policy == MPOL_INTERLEAVE) { |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 521 | *policy = current->il_next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 522 | } else { |
| 523 | err = -EINVAL; |
| 524 | goto out; |
| 525 | } |
| 526 | } else |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 527 | *policy = pol->policy; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 528 | |
| 529 | if (vma) { |
| 530 | up_read(¤t->mm->mmap_sem); |
| 531 | vma = NULL; |
| 532 | } |
| 533 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 534 | err = 0; |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 535 | if (nmask) |
| 536 | get_zonemask(pol, nmask); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 537 | |
| 538 | out: |
| 539 | if (vma) |
| 540 | up_read(¤t->mm->mmap_sem); |
| 541 | return err; |
| 542 | } |
| 543 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 544 | #ifdef CONFIG_MIGRATION |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 545 | /* |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 546 | * page migration |
| 547 | */ |
Christoph Lameter | fc30128 | 2006-01-18 17:42:29 -0800 | [diff] [blame] | 548 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
| 549 | unsigned long flags) |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 550 | { |
| 551 | /* |
Christoph Lameter | fc30128 | 2006-01-18 17:42:29 -0800 | [diff] [blame] | 552 | * Avoid migrating a page that is shared with others. |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 553 | */ |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 554 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) |
| 555 | isolate_lru_page(page, pagelist); |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 556 | } |
| 557 | |
| 558 | /* |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 559 | * Migrate pages from one node to a target node. |
| 560 | * Returns error or the number of pages not migrated. |
| 561 | */ |
| 562 | int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags) |
| 563 | { |
| 564 | nodemask_t nmask; |
| 565 | LIST_HEAD(pagelist); |
| 566 | int err = 0; |
| 567 | |
| 568 | nodes_clear(nmask); |
| 569 | node_set(source, nmask); |
| 570 | |
| 571 | check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask, |
| 572 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); |
| 573 | |
| 574 | if (!list_empty(&pagelist)) { |
| 575 | err = migrate_pages_to(&pagelist, NULL, dest); |
| 576 | if (!list_empty(&pagelist)) |
| 577 | putback_lru_pages(&pagelist); |
| 578 | } |
| 579 | return err; |
| 580 | } |
| 581 | |
| 582 | /* |
| 583 | * Move pages between the two nodesets so as to preserve the physical |
| 584 | * layout as much as possible. |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 585 | * |
| 586 | * Returns the number of page that could not be moved. |
| 587 | */ |
| 588 | int do_migrate_pages(struct mm_struct *mm, |
| 589 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) |
| 590 | { |
| 591 | LIST_HEAD(pagelist); |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 592 | int busy = 0; |
| 593 | int err = 0; |
| 594 | nodemask_t tmp; |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 595 | |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 596 | down_read(&mm->mmap_sem); |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 597 | |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 598 | /* |
| 599 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' |
| 600 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' |
| 601 | * bit in 'tmp', and return that <source, dest> pair for migration. |
| 602 | * The pair of nodemasks 'to' and 'from' define the map. |
| 603 | * |
| 604 | * If no pair of bits is found that way, fallback to picking some |
| 605 | * pair of 'source' and 'dest' bits that are not the same. If the |
| 606 | * 'source' and 'dest' bits are the same, this represents a node |
| 607 | * that will be migrating to itself, so no pages need move. |
| 608 | * |
| 609 | * If no bits are left in 'tmp', or if all remaining bits left |
| 610 | * in 'tmp' correspond to the same bit in 'to', return false |
| 611 | * (nothing left to migrate). |
| 612 | * |
| 613 | * This lets us pick a pair of nodes to migrate between, such that |
| 614 | * if possible the dest node is not already occupied by some other |
| 615 | * source node, minimizing the risk of overloading the memory on a |
| 616 | * node that would happen if we migrated incoming memory to a node |
| 617 | * before migrating outgoing memory source that same node. |
| 618 | * |
| 619 | * A single scan of tmp is sufficient. As we go, we remember the |
| 620 | * most recent <s, d> pair that moved (s != d). If we find a pair |
| 621 | * that not only moved, but what's better, moved to an empty slot |
| 622 | * (d is not set in tmp), then we break out then, with that pair. |
| 623 | * Otherwise when we finish scannng from_tmp, we at least have the |
| 624 | * most recent <s, d> pair that moved. If we get all the way through |
| 625 | * the scan of tmp without finding any node that moved, much less |
| 626 | * moved to an empty node, then there is nothing left worth migrating. |
| 627 | */ |
Christoph Lameter | d498471 | 2006-01-08 01:00:55 -0800 | [diff] [blame] | 628 | |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 629 | tmp = *from_nodes; |
| 630 | while (!nodes_empty(tmp)) { |
| 631 | int s,d; |
| 632 | int source = -1; |
| 633 | int dest = 0; |
| 634 | |
| 635 | for_each_node_mask(s, tmp) { |
| 636 | d = node_remap(s, *from_nodes, *to_nodes); |
| 637 | if (s == d) |
| 638 | continue; |
| 639 | |
| 640 | source = s; /* Node moved. Memorize */ |
| 641 | dest = d; |
| 642 | |
| 643 | /* dest not in remaining from nodes? */ |
| 644 | if (!node_isset(dest, tmp)) |
| 645 | break; |
| 646 | } |
| 647 | if (source == -1) |
| 648 | break; |
| 649 | |
| 650 | node_clear(source, tmp); |
| 651 | err = migrate_to_node(mm, source, dest, flags); |
| 652 | if (err > 0) |
| 653 | busy += err; |
| 654 | if (err < 0) |
| 655 | break; |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 656 | } |
Christoph Lameter | d498471 | 2006-01-08 01:00:55 -0800 | [diff] [blame] | 657 | |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 658 | up_read(&mm->mmap_sem); |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 659 | if (err < 0) |
| 660 | return err; |
| 661 | return busy; |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 662 | |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 663 | } |
| 664 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 665 | #else |
| 666 | |
| 667 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
| 668 | unsigned long flags) |
| 669 | { |
| 670 | } |
| 671 | |
| 672 | int do_migrate_pages(struct mm_struct *mm, |
| 673 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) |
| 674 | { |
| 675 | return -ENOSYS; |
| 676 | } |
| 677 | #endif |
| 678 | |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 679 | long do_mbind(unsigned long start, unsigned long len, |
| 680 | unsigned long mode, nodemask_t *nmask, unsigned long flags) |
| 681 | { |
| 682 | struct vm_area_struct *vma; |
| 683 | struct mm_struct *mm = current->mm; |
| 684 | struct mempolicy *new; |
| 685 | unsigned long end; |
| 686 | int err; |
| 687 | LIST_HEAD(pagelist); |
| 688 | |
| 689 | if ((flags & ~(unsigned long)(MPOL_MF_STRICT | |
| 690 | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) |
| 691 | || mode > MPOL_MAX) |
| 692 | return -EINVAL; |
Christoph Lameter | 74c0024 | 2006-03-14 19:50:21 -0800 | [diff] [blame] | 693 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 694 | return -EPERM; |
| 695 | |
| 696 | if (start & ~PAGE_MASK) |
| 697 | return -EINVAL; |
| 698 | |
| 699 | if (mode == MPOL_DEFAULT) |
| 700 | flags &= ~MPOL_MF_STRICT; |
| 701 | |
| 702 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; |
| 703 | end = start + len; |
| 704 | |
| 705 | if (end < start) |
| 706 | return -EINVAL; |
| 707 | if (end == start) |
| 708 | return 0; |
| 709 | |
| 710 | if (mpol_check_policy(mode, nmask)) |
| 711 | return -EINVAL; |
| 712 | |
| 713 | new = mpol_new(mode, nmask); |
| 714 | if (IS_ERR(new)) |
| 715 | return PTR_ERR(new); |
| 716 | |
| 717 | /* |
| 718 | * If we are using the default policy then operation |
| 719 | * on discontinuous address spaces is okay after all |
| 720 | */ |
| 721 | if (!new) |
| 722 | flags |= MPOL_MF_DISCONTIG_OK; |
| 723 | |
| 724 | PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len, |
| 725 | mode,nodes_addr(nodes)[0]); |
| 726 | |
| 727 | down_write(&mm->mmap_sem); |
| 728 | vma = check_range(mm, start, end, nmask, |
| 729 | flags | MPOL_MF_INVERT, &pagelist); |
| 730 | |
| 731 | err = PTR_ERR(vma); |
| 732 | if (!IS_ERR(vma)) { |
| 733 | int nr_failed = 0; |
| 734 | |
| 735 | err = mbind_range(vma, start, end, new); |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 736 | |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 737 | if (!list_empty(&pagelist)) |
Christoph Lameter | 7e2ab15 | 2006-02-01 03:05:40 -0800 | [diff] [blame] | 738 | nr_failed = migrate_pages_to(&pagelist, vma, -1); |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 739 | |
| 740 | if (!err && nr_failed && (flags & MPOL_MF_STRICT)) |
| 741 | err = -EIO; |
| 742 | } |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame^] | 743 | |
Christoph Lameter | 6ce3c4c | 2006-01-08 01:01:04 -0800 | [diff] [blame] | 744 | if (!list_empty(&pagelist)) |
| 745 | putback_lru_pages(&pagelist); |
| 746 | |
| 747 | up_write(&mm->mmap_sem); |
| 748 | mpol_free(new); |
| 749 | return err; |
| 750 | } |
| 751 | |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 752 | /* |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 753 | * User space interface with variable sized bitmaps for nodelists. |
| 754 | */ |
| 755 | |
| 756 | /* Copy a node mask from user space. */ |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 757 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 758 | unsigned long maxnode) |
| 759 | { |
| 760 | unsigned long k; |
| 761 | unsigned long nlongs; |
| 762 | unsigned long endmask; |
| 763 | |
| 764 | --maxnode; |
| 765 | nodes_clear(*nodes); |
| 766 | if (maxnode == 0 || !nmask) |
| 767 | return 0; |
Andi Kleen | a9c930b | 2006-02-20 18:27:59 -0800 | [diff] [blame] | 768 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
Chris Wright | 636f13c | 2006-02-17 13:59:36 -0800 | [diff] [blame] | 769 | return -EINVAL; |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 770 | |
| 771 | nlongs = BITS_TO_LONGS(maxnode); |
| 772 | if ((maxnode % BITS_PER_LONG) == 0) |
| 773 | endmask = ~0UL; |
| 774 | else |
| 775 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; |
| 776 | |
| 777 | /* When the user specified more nodes than supported just check |
| 778 | if the non supported part is all zero. */ |
| 779 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { |
| 780 | if (nlongs > PAGE_SIZE/sizeof(long)) |
| 781 | return -EINVAL; |
| 782 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { |
| 783 | unsigned long t; |
| 784 | if (get_user(t, nmask + k)) |
| 785 | return -EFAULT; |
| 786 | if (k == nlongs - 1) { |
| 787 | if (t & endmask) |
| 788 | return -EINVAL; |
| 789 | } else if (t) |
| 790 | return -EINVAL; |
| 791 | } |
| 792 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); |
| 793 | endmask = ~0UL; |
| 794 | } |
| 795 | |
| 796 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) |
| 797 | return -EFAULT; |
| 798 | nodes_addr(*nodes)[nlongs-1] &= endmask; |
| 799 | return 0; |
| 800 | } |
| 801 | |
| 802 | /* Copy a kernel node mask to user space */ |
| 803 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, |
| 804 | nodemask_t *nodes) |
| 805 | { |
| 806 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; |
| 807 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); |
| 808 | |
| 809 | if (copy > nbytes) { |
| 810 | if (copy > PAGE_SIZE) |
| 811 | return -EINVAL; |
| 812 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) |
| 813 | return -EFAULT; |
| 814 | copy = nbytes; |
| 815 | } |
| 816 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; |
| 817 | } |
| 818 | |
| 819 | asmlinkage long sys_mbind(unsigned long start, unsigned long len, |
| 820 | unsigned long mode, |
| 821 | unsigned long __user *nmask, unsigned long maxnode, |
| 822 | unsigned flags) |
| 823 | { |
| 824 | nodemask_t nodes; |
| 825 | int err; |
| 826 | |
| 827 | err = get_nodes(&nodes, nmask, maxnode); |
| 828 | if (err) |
| 829 | return err; |
| 830 | return do_mbind(start, len, mode, &nodes, flags); |
| 831 | } |
| 832 | |
| 833 | /* Set the process memory policy */ |
| 834 | asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask, |
| 835 | unsigned long maxnode) |
| 836 | { |
| 837 | int err; |
| 838 | nodemask_t nodes; |
| 839 | |
| 840 | if (mode < 0 || mode > MPOL_MAX) |
| 841 | return -EINVAL; |
| 842 | err = get_nodes(&nodes, nmask, maxnode); |
| 843 | if (err) |
| 844 | return err; |
| 845 | return do_set_mempolicy(mode, &nodes); |
| 846 | } |
| 847 | |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 848 | asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode, |
| 849 | const unsigned long __user *old_nodes, |
| 850 | const unsigned long __user *new_nodes) |
| 851 | { |
| 852 | struct mm_struct *mm; |
| 853 | struct task_struct *task; |
| 854 | nodemask_t old; |
| 855 | nodemask_t new; |
| 856 | nodemask_t task_nodes; |
| 857 | int err; |
| 858 | |
| 859 | err = get_nodes(&old, old_nodes, maxnode); |
| 860 | if (err) |
| 861 | return err; |
| 862 | |
| 863 | err = get_nodes(&new, new_nodes, maxnode); |
| 864 | if (err) |
| 865 | return err; |
| 866 | |
| 867 | /* Find the mm_struct */ |
| 868 | read_lock(&tasklist_lock); |
| 869 | task = pid ? find_task_by_pid(pid) : current; |
| 870 | if (!task) { |
| 871 | read_unlock(&tasklist_lock); |
| 872 | return -ESRCH; |
| 873 | } |
| 874 | mm = get_task_mm(task); |
| 875 | read_unlock(&tasklist_lock); |
| 876 | |
| 877 | if (!mm) |
| 878 | return -EINVAL; |
| 879 | |
| 880 | /* |
| 881 | * Check if this process has the right to modify the specified |
| 882 | * process. The right exists if the process has administrative |
| 883 | * capabilities, superuser priviledges or the same |
| 884 | * userid as the target process. |
| 885 | */ |
| 886 | if ((current->euid != task->suid) && (current->euid != task->uid) && |
| 887 | (current->uid != task->suid) && (current->uid != task->uid) && |
Christoph Lameter | 74c0024 | 2006-03-14 19:50:21 -0800 | [diff] [blame] | 888 | !capable(CAP_SYS_NICE)) { |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 889 | err = -EPERM; |
| 890 | goto out; |
| 891 | } |
| 892 | |
| 893 | task_nodes = cpuset_mems_allowed(task); |
| 894 | /* Is the user allowed to access the target nodes? */ |
Christoph Lameter | 74c0024 | 2006-03-14 19:50:21 -0800 | [diff] [blame] | 895 | if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) { |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 896 | err = -EPERM; |
| 897 | goto out; |
| 898 | } |
| 899 | |
Christoph Lameter | 511030b | 2006-02-28 16:58:57 -0800 | [diff] [blame] | 900 | err = do_migrate_pages(mm, &old, &new, |
Christoph Lameter | 74c0024 | 2006-03-14 19:50:21 -0800 | [diff] [blame] | 901 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); |
Christoph Lameter | 3974388 | 2006-01-08 01:00:51 -0800 | [diff] [blame] | 902 | out: |
| 903 | mmput(mm); |
| 904 | return err; |
| 905 | } |
| 906 | |
| 907 | |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 908 | /* Retrieve NUMA policy */ |
| 909 | asmlinkage long sys_get_mempolicy(int __user *policy, |
| 910 | unsigned long __user *nmask, |
| 911 | unsigned long maxnode, |
| 912 | unsigned long addr, unsigned long flags) |
| 913 | { |
| 914 | int err, pval; |
| 915 | nodemask_t nodes; |
| 916 | |
| 917 | if (nmask != NULL && maxnode < MAX_NUMNODES) |
| 918 | return -EINVAL; |
| 919 | |
| 920 | err = do_get_mempolicy(&pval, &nodes, addr, flags); |
| 921 | |
| 922 | if (err) |
| 923 | return err; |
| 924 | |
| 925 | if (policy && put_user(pval, policy)) |
| 926 | return -EFAULT; |
| 927 | |
| 928 | if (nmask) |
| 929 | err = copy_nodes_to_user(nmask, maxnode, &nodes); |
| 930 | |
| 931 | return err; |
| 932 | } |
| 933 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 934 | #ifdef CONFIG_COMPAT |
| 935 | |
| 936 | asmlinkage long compat_sys_get_mempolicy(int __user *policy, |
| 937 | compat_ulong_t __user *nmask, |
| 938 | compat_ulong_t maxnode, |
| 939 | compat_ulong_t addr, compat_ulong_t flags) |
| 940 | { |
| 941 | long err; |
| 942 | unsigned long __user *nm = NULL; |
| 943 | unsigned long nr_bits, alloc_size; |
| 944 | DECLARE_BITMAP(bm, MAX_NUMNODES); |
| 945 | |
| 946 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); |
| 947 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 948 | |
| 949 | if (nmask) |
| 950 | nm = compat_alloc_user_space(alloc_size); |
| 951 | |
| 952 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); |
| 953 | |
| 954 | if (!err && nmask) { |
| 955 | err = copy_from_user(bm, nm, alloc_size); |
| 956 | /* ensure entire bitmap is zeroed */ |
| 957 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); |
| 958 | err |= compat_put_bitmap(nmask, bm, nr_bits); |
| 959 | } |
| 960 | |
| 961 | return err; |
| 962 | } |
| 963 | |
| 964 | asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask, |
| 965 | compat_ulong_t maxnode) |
| 966 | { |
| 967 | long err = 0; |
| 968 | unsigned long __user *nm = NULL; |
| 969 | unsigned long nr_bits, alloc_size; |
| 970 | DECLARE_BITMAP(bm, MAX_NUMNODES); |
| 971 | |
| 972 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); |
| 973 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 974 | |
| 975 | if (nmask) { |
| 976 | err = compat_get_bitmap(bm, nmask, nr_bits); |
| 977 | nm = compat_alloc_user_space(alloc_size); |
| 978 | err |= copy_to_user(nm, bm, alloc_size); |
| 979 | } |
| 980 | |
| 981 | if (err) |
| 982 | return -EFAULT; |
| 983 | |
| 984 | return sys_set_mempolicy(mode, nm, nr_bits+1); |
| 985 | } |
| 986 | |
| 987 | asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len, |
| 988 | compat_ulong_t mode, compat_ulong_t __user *nmask, |
| 989 | compat_ulong_t maxnode, compat_ulong_t flags) |
| 990 | { |
| 991 | long err = 0; |
| 992 | unsigned long __user *nm = NULL; |
| 993 | unsigned long nr_bits, alloc_size; |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 994 | nodemask_t bm; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 995 | |
| 996 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); |
| 997 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 998 | |
| 999 | if (nmask) { |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1000 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1001 | nm = compat_alloc_user_space(alloc_size); |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1002 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1003 | } |
| 1004 | |
| 1005 | if (err) |
| 1006 | return -EFAULT; |
| 1007 | |
| 1008 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); |
| 1009 | } |
| 1010 | |
| 1011 | #endif |
| 1012 | |
| 1013 | /* Return effective policy for a VMA */ |
Christoph Lameter | 48fce34 | 2006-01-08 01:01:03 -0800 | [diff] [blame] | 1014 | static struct mempolicy * get_vma_policy(struct task_struct *task, |
| 1015 | struct vm_area_struct *vma, unsigned long addr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1016 | { |
Christoph Lameter | 6e21c8f | 2005-09-03 15:54:45 -0700 | [diff] [blame] | 1017 | struct mempolicy *pol = task->mempolicy; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1018 | |
| 1019 | if (vma) { |
| 1020 | if (vma->vm_ops && vma->vm_ops->get_policy) |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 1021 | pol = vma->vm_ops->get_policy(vma, addr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1022 | else if (vma->vm_policy && |
| 1023 | vma->vm_policy->policy != MPOL_DEFAULT) |
| 1024 | pol = vma->vm_policy; |
| 1025 | } |
| 1026 | if (!pol) |
| 1027 | pol = &default_policy; |
| 1028 | return pol; |
| 1029 | } |
| 1030 | |
| 1031 | /* Return a zonelist representing a mempolicy */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 1032 | static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1033 | { |
| 1034 | int nd; |
| 1035 | |
| 1036 | switch (policy->policy) { |
| 1037 | case MPOL_PREFERRED: |
| 1038 | nd = policy->v.preferred_node; |
| 1039 | if (nd < 0) |
| 1040 | nd = numa_node_id(); |
| 1041 | break; |
| 1042 | case MPOL_BIND: |
| 1043 | /* Lower zones don't get a policy applied */ |
| 1044 | /* Careful: current->mems_allowed might have moved */ |
Al Viro | af4ca45 | 2005-10-21 02:55:38 -0400 | [diff] [blame] | 1045 | if (gfp_zone(gfp) >= policy_zone) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1046 | if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist)) |
| 1047 | return policy->v.zonelist; |
| 1048 | /*FALL THROUGH*/ |
| 1049 | case MPOL_INTERLEAVE: /* should not happen */ |
| 1050 | case MPOL_DEFAULT: |
| 1051 | nd = numa_node_id(); |
| 1052 | break; |
| 1053 | default: |
| 1054 | nd = 0; |
| 1055 | BUG(); |
| 1056 | } |
Al Viro | af4ca45 | 2005-10-21 02:55:38 -0400 | [diff] [blame] | 1057 | return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1058 | } |
| 1059 | |
| 1060 | /* Do dynamic interleaving for a process */ |
| 1061 | static unsigned interleave_nodes(struct mempolicy *policy) |
| 1062 | { |
| 1063 | unsigned nid, next; |
| 1064 | struct task_struct *me = current; |
| 1065 | |
| 1066 | nid = me->il_next; |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1067 | next = next_node(nid, policy->v.nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1068 | if (next >= MAX_NUMNODES) |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1069 | next = first_node(policy->v.nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1070 | me->il_next = next; |
| 1071 | return nid; |
| 1072 | } |
| 1073 | |
Christoph Lameter | dc85da1 | 2006-01-18 17:42:36 -0800 | [diff] [blame] | 1074 | /* |
| 1075 | * Depending on the memory policy provide a node from which to allocate the |
| 1076 | * next slab entry. |
| 1077 | */ |
| 1078 | unsigned slab_node(struct mempolicy *policy) |
| 1079 | { |
Christoph Lameter | dc85da1 | 2006-01-18 17:42:36 -0800 | [diff] [blame] | 1080 | switch (policy->policy) { |
| 1081 | case MPOL_INTERLEAVE: |
| 1082 | return interleave_nodes(policy); |
| 1083 | |
| 1084 | case MPOL_BIND: |
| 1085 | /* |
| 1086 | * Follow bind policy behavior and start allocation at the |
| 1087 | * first node. |
| 1088 | */ |
| 1089 | return policy->v.zonelist->zones[0]->zone_pgdat->node_id; |
| 1090 | |
| 1091 | case MPOL_PREFERRED: |
| 1092 | if (policy->v.preferred_node >= 0) |
| 1093 | return policy->v.preferred_node; |
| 1094 | /* Fall through */ |
| 1095 | |
| 1096 | default: |
| 1097 | return numa_node_id(); |
| 1098 | } |
| 1099 | } |
| 1100 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1101 | /* Do static interleaving for a VMA with known offset. */ |
| 1102 | static unsigned offset_il_node(struct mempolicy *pol, |
| 1103 | struct vm_area_struct *vma, unsigned long off) |
| 1104 | { |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1105 | unsigned nnodes = nodes_weight(pol->v.nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1106 | unsigned target = (unsigned)off % nnodes; |
| 1107 | int c; |
| 1108 | int nid = -1; |
| 1109 | |
| 1110 | c = 0; |
| 1111 | do { |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1112 | nid = next_node(nid, pol->v.nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1113 | c++; |
| 1114 | } while (c <= target); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1115 | return nid; |
| 1116 | } |
| 1117 | |
Christoph Lameter | 5da7ca8 | 2006-01-06 00:10:46 -0800 | [diff] [blame] | 1118 | /* Determine a node number for interleave */ |
| 1119 | static inline unsigned interleave_nid(struct mempolicy *pol, |
| 1120 | struct vm_area_struct *vma, unsigned long addr, int shift) |
| 1121 | { |
| 1122 | if (vma) { |
| 1123 | unsigned long off; |
| 1124 | |
| 1125 | off = vma->vm_pgoff; |
| 1126 | off += (addr - vma->vm_start) >> shift; |
| 1127 | return offset_il_node(pol, vma, off); |
| 1128 | } else |
| 1129 | return interleave_nodes(pol); |
| 1130 | } |
| 1131 | |
Chen, Kenneth W | 00ac59a | 2006-02-03 21:51:14 +0100 | [diff] [blame] | 1132 | #ifdef CONFIG_HUGETLBFS |
Christoph Lameter | 5da7ca8 | 2006-01-06 00:10:46 -0800 | [diff] [blame] | 1133 | /* Return a zonelist suitable for a huge page allocation. */ |
| 1134 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr) |
| 1135 | { |
| 1136 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
| 1137 | |
| 1138 | if (pol->policy == MPOL_INTERLEAVE) { |
| 1139 | unsigned nid; |
| 1140 | |
| 1141 | nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT); |
| 1142 | return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER); |
| 1143 | } |
| 1144 | return zonelist_policy(GFP_HIGHUSER, pol); |
| 1145 | } |
Chen, Kenneth W | 00ac59a | 2006-02-03 21:51:14 +0100 | [diff] [blame] | 1146 | #endif |
Christoph Lameter | 5da7ca8 | 2006-01-06 00:10:46 -0800 | [diff] [blame] | 1147 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1148 | /* Allocate a page in interleaved policy. |
| 1149 | Own path because it needs to do special accounting. */ |
Andi Kleen | 662f3a0 | 2005-10-29 18:15:49 -0700 | [diff] [blame] | 1150 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
| 1151 | unsigned nid) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1152 | { |
| 1153 | struct zonelist *zl; |
| 1154 | struct page *page; |
| 1155 | |
Al Viro | af4ca45 | 2005-10-21 02:55:38 -0400 | [diff] [blame] | 1156 | zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1157 | page = __alloc_pages(gfp, order, zl); |
| 1158 | if (page && page_zone(page) == zl->zones[0]) { |
Christoph Lameter | e7c8d5c | 2005-06-21 17:14:47 -0700 | [diff] [blame] | 1159 | zone_pcp(zl->zones[0],get_cpu())->interleave_hit++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1160 | put_cpu(); |
| 1161 | } |
| 1162 | return page; |
| 1163 | } |
| 1164 | |
| 1165 | /** |
| 1166 | * alloc_page_vma - Allocate a page for a VMA. |
| 1167 | * |
| 1168 | * @gfp: |
| 1169 | * %GFP_USER user allocation. |
| 1170 | * %GFP_KERNEL kernel allocations, |
| 1171 | * %GFP_HIGHMEM highmem/user allocations, |
| 1172 | * %GFP_FS allocation should not call back into a file system. |
| 1173 | * %GFP_ATOMIC don't sleep. |
| 1174 | * |
| 1175 | * @vma: Pointer to VMA or NULL if not available. |
| 1176 | * @addr: Virtual Address of the allocation. Must be inside the VMA. |
| 1177 | * |
| 1178 | * This function allocates a page from the kernel page pool and applies |
| 1179 | * a NUMA policy associated with the VMA or the current process. |
| 1180 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the |
| 1181 | * mm_struct of the VMA to prevent it from going away. Should be used for |
| 1182 | * all allocations for pages that will be mapped into |
| 1183 | * user space. Returns NULL when no page can be allocated. |
| 1184 | * |
| 1185 | * Should be called with the mm_sem of the vma hold. |
| 1186 | */ |
| 1187 | struct page * |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 1188 | alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1189 | { |
Christoph Lameter | 6e21c8f | 2005-09-03 15:54:45 -0700 | [diff] [blame] | 1190 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1191 | |
Paul Jackson | cf2a473 | 2006-01-08 01:01:54 -0800 | [diff] [blame] | 1192 | cpuset_update_task_memory_state(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1193 | |
| 1194 | if (unlikely(pol->policy == MPOL_INTERLEAVE)) { |
| 1195 | unsigned nid; |
Christoph Lameter | 5da7ca8 | 2006-01-06 00:10:46 -0800 | [diff] [blame] | 1196 | |
| 1197 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1198 | return alloc_page_interleave(gfp, 0, nid); |
| 1199 | } |
| 1200 | return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol)); |
| 1201 | } |
| 1202 | |
| 1203 | /** |
| 1204 | * alloc_pages_current - Allocate pages. |
| 1205 | * |
| 1206 | * @gfp: |
| 1207 | * %GFP_USER user allocation, |
| 1208 | * %GFP_KERNEL kernel allocation, |
| 1209 | * %GFP_HIGHMEM highmem allocation, |
| 1210 | * %GFP_FS don't call back into a file system. |
| 1211 | * %GFP_ATOMIC don't sleep. |
| 1212 | * @order: Power of two of allocation size in pages. 0 is a single page. |
| 1213 | * |
| 1214 | * Allocate a page from the kernel page pool. When not in |
| 1215 | * interrupt context and apply the current process NUMA policy. |
| 1216 | * Returns NULL when no page can be allocated. |
| 1217 | * |
Paul Jackson | cf2a473 | 2006-01-08 01:01:54 -0800 | [diff] [blame] | 1218 | * Don't call cpuset_update_task_memory_state() unless |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1219 | * 1) it's ok to take cpuset_sem (can WAIT), and |
| 1220 | * 2) allocating for current task (not interrupt). |
| 1221 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 1222 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1223 | { |
| 1224 | struct mempolicy *pol = current->mempolicy; |
| 1225 | |
| 1226 | if ((gfp & __GFP_WAIT) && !in_interrupt()) |
Paul Jackson | cf2a473 | 2006-01-08 01:01:54 -0800 | [diff] [blame] | 1227 | cpuset_update_task_memory_state(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1228 | if (!pol || in_interrupt()) |
| 1229 | pol = &default_policy; |
| 1230 | if (pol->policy == MPOL_INTERLEAVE) |
| 1231 | return alloc_page_interleave(gfp, order, interleave_nodes(pol)); |
| 1232 | return __alloc_pages(gfp, order, zonelist_policy(gfp, pol)); |
| 1233 | } |
| 1234 | EXPORT_SYMBOL(alloc_pages_current); |
| 1235 | |
Paul Jackson | 4225399 | 2006-01-08 01:01:59 -0800 | [diff] [blame] | 1236 | /* |
| 1237 | * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it |
| 1238 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() |
| 1239 | * with the mems_allowed returned by cpuset_mems_allowed(). This |
| 1240 | * keeps mempolicies cpuset relative after its cpuset moves. See |
| 1241 | * further kernel/cpuset.c update_nodemask(). |
| 1242 | */ |
| 1243 | void *cpuset_being_rebound; |
| 1244 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1245 | /* Slow path of a mempolicy copy */ |
| 1246 | struct mempolicy *__mpol_copy(struct mempolicy *old) |
| 1247 | { |
| 1248 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
| 1249 | |
| 1250 | if (!new) |
| 1251 | return ERR_PTR(-ENOMEM); |
Paul Jackson | 4225399 | 2006-01-08 01:01:59 -0800 | [diff] [blame] | 1252 | if (current_cpuset_is_being_rebound()) { |
| 1253 | nodemask_t mems = cpuset_mems_allowed(current); |
| 1254 | mpol_rebind_policy(old, &mems); |
| 1255 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1256 | *new = *old; |
| 1257 | atomic_set(&new->refcnt, 1); |
| 1258 | if (new->policy == MPOL_BIND) { |
| 1259 | int sz = ksize(old->v.zonelist); |
| 1260 | new->v.zonelist = kmalloc(sz, SLAB_KERNEL); |
| 1261 | if (!new->v.zonelist) { |
| 1262 | kmem_cache_free(policy_cache, new); |
| 1263 | return ERR_PTR(-ENOMEM); |
| 1264 | } |
| 1265 | memcpy(new->v.zonelist, old->v.zonelist, sz); |
| 1266 | } |
| 1267 | return new; |
| 1268 | } |
| 1269 | |
| 1270 | /* Slow path of a mempolicy comparison */ |
| 1271 | int __mpol_equal(struct mempolicy *a, struct mempolicy *b) |
| 1272 | { |
| 1273 | if (!a || !b) |
| 1274 | return 0; |
| 1275 | if (a->policy != b->policy) |
| 1276 | return 0; |
| 1277 | switch (a->policy) { |
| 1278 | case MPOL_DEFAULT: |
| 1279 | return 1; |
| 1280 | case MPOL_INTERLEAVE: |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1281 | return nodes_equal(a->v.nodes, b->v.nodes); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1282 | case MPOL_PREFERRED: |
| 1283 | return a->v.preferred_node == b->v.preferred_node; |
| 1284 | case MPOL_BIND: { |
| 1285 | int i; |
| 1286 | for (i = 0; a->v.zonelist->zones[i]; i++) |
| 1287 | if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i]) |
| 1288 | return 0; |
| 1289 | return b->v.zonelist->zones[i] == NULL; |
| 1290 | } |
| 1291 | default: |
| 1292 | BUG(); |
| 1293 | return 0; |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | /* Slow path of a mpol destructor. */ |
| 1298 | void __mpol_free(struct mempolicy *p) |
| 1299 | { |
| 1300 | if (!atomic_dec_and_test(&p->refcnt)) |
| 1301 | return; |
| 1302 | if (p->policy == MPOL_BIND) |
| 1303 | kfree(p->v.zonelist); |
| 1304 | p->policy = MPOL_DEFAULT; |
| 1305 | kmem_cache_free(policy_cache, p); |
| 1306 | } |
| 1307 | |
| 1308 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1309 | * Shared memory backing store policy support. |
| 1310 | * |
| 1311 | * Remember policies even when nobody has shared memory mapped. |
| 1312 | * The policies are kept in Red-Black tree linked from the inode. |
| 1313 | * They are protected by the sp->lock spinlock, which should be held |
| 1314 | * for any accesses to the tree. |
| 1315 | */ |
| 1316 | |
| 1317 | /* lookup first element intersecting start-end */ |
| 1318 | /* Caller holds sp->lock */ |
| 1319 | static struct sp_node * |
| 1320 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) |
| 1321 | { |
| 1322 | struct rb_node *n = sp->root.rb_node; |
| 1323 | |
| 1324 | while (n) { |
| 1325 | struct sp_node *p = rb_entry(n, struct sp_node, nd); |
| 1326 | |
| 1327 | if (start >= p->end) |
| 1328 | n = n->rb_right; |
| 1329 | else if (end <= p->start) |
| 1330 | n = n->rb_left; |
| 1331 | else |
| 1332 | break; |
| 1333 | } |
| 1334 | if (!n) |
| 1335 | return NULL; |
| 1336 | for (;;) { |
| 1337 | struct sp_node *w = NULL; |
| 1338 | struct rb_node *prev = rb_prev(n); |
| 1339 | if (!prev) |
| 1340 | break; |
| 1341 | w = rb_entry(prev, struct sp_node, nd); |
| 1342 | if (w->end <= start) |
| 1343 | break; |
| 1344 | n = prev; |
| 1345 | } |
| 1346 | return rb_entry(n, struct sp_node, nd); |
| 1347 | } |
| 1348 | |
| 1349 | /* Insert a new shared policy into the list. */ |
| 1350 | /* Caller holds sp->lock */ |
| 1351 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) |
| 1352 | { |
| 1353 | struct rb_node **p = &sp->root.rb_node; |
| 1354 | struct rb_node *parent = NULL; |
| 1355 | struct sp_node *nd; |
| 1356 | |
| 1357 | while (*p) { |
| 1358 | parent = *p; |
| 1359 | nd = rb_entry(parent, struct sp_node, nd); |
| 1360 | if (new->start < nd->start) |
| 1361 | p = &(*p)->rb_left; |
| 1362 | else if (new->end > nd->end) |
| 1363 | p = &(*p)->rb_right; |
| 1364 | else |
| 1365 | BUG(); |
| 1366 | } |
| 1367 | rb_link_node(&new->nd, parent, p); |
| 1368 | rb_insert_color(&new->nd, &sp->root); |
| 1369 | PDprintk("inserting %lx-%lx: %d\n", new->start, new->end, |
| 1370 | new->policy ? new->policy->policy : 0); |
| 1371 | } |
| 1372 | |
| 1373 | /* Find shared policy intersecting idx */ |
| 1374 | struct mempolicy * |
| 1375 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) |
| 1376 | { |
| 1377 | struct mempolicy *pol = NULL; |
| 1378 | struct sp_node *sn; |
| 1379 | |
| 1380 | if (!sp->root.rb_node) |
| 1381 | return NULL; |
| 1382 | spin_lock(&sp->lock); |
| 1383 | sn = sp_lookup(sp, idx, idx+1); |
| 1384 | if (sn) { |
| 1385 | mpol_get(sn->policy); |
| 1386 | pol = sn->policy; |
| 1387 | } |
| 1388 | spin_unlock(&sp->lock); |
| 1389 | return pol; |
| 1390 | } |
| 1391 | |
| 1392 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) |
| 1393 | { |
| 1394 | PDprintk("deleting %lx-l%x\n", n->start, n->end); |
| 1395 | rb_erase(&n->nd, &sp->root); |
| 1396 | mpol_free(n->policy); |
| 1397 | kmem_cache_free(sn_cache, n); |
| 1398 | } |
| 1399 | |
| 1400 | struct sp_node * |
| 1401 | sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol) |
| 1402 | { |
| 1403 | struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL); |
| 1404 | |
| 1405 | if (!n) |
| 1406 | return NULL; |
| 1407 | n->start = start; |
| 1408 | n->end = end; |
| 1409 | mpol_get(pol); |
| 1410 | n->policy = pol; |
| 1411 | return n; |
| 1412 | } |
| 1413 | |
| 1414 | /* Replace a policy range. */ |
| 1415 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, |
| 1416 | unsigned long end, struct sp_node *new) |
| 1417 | { |
| 1418 | struct sp_node *n, *new2 = NULL; |
| 1419 | |
| 1420 | restart: |
| 1421 | spin_lock(&sp->lock); |
| 1422 | n = sp_lookup(sp, start, end); |
| 1423 | /* Take care of old policies in the same range. */ |
| 1424 | while (n && n->start < end) { |
| 1425 | struct rb_node *next = rb_next(&n->nd); |
| 1426 | if (n->start >= start) { |
| 1427 | if (n->end <= end) |
| 1428 | sp_delete(sp, n); |
| 1429 | else |
| 1430 | n->start = end; |
| 1431 | } else { |
| 1432 | /* Old policy spanning whole new range. */ |
| 1433 | if (n->end > end) { |
| 1434 | if (!new2) { |
| 1435 | spin_unlock(&sp->lock); |
| 1436 | new2 = sp_alloc(end, n->end, n->policy); |
| 1437 | if (!new2) |
| 1438 | return -ENOMEM; |
| 1439 | goto restart; |
| 1440 | } |
| 1441 | n->end = start; |
| 1442 | sp_insert(sp, new2); |
| 1443 | new2 = NULL; |
| 1444 | break; |
| 1445 | } else |
| 1446 | n->end = start; |
| 1447 | } |
| 1448 | if (!next) |
| 1449 | break; |
| 1450 | n = rb_entry(next, struct sp_node, nd); |
| 1451 | } |
| 1452 | if (new) |
| 1453 | sp_insert(sp, new); |
| 1454 | spin_unlock(&sp->lock); |
| 1455 | if (new2) { |
| 1456 | mpol_free(new2->policy); |
| 1457 | kmem_cache_free(sn_cache, new2); |
| 1458 | } |
| 1459 | return 0; |
| 1460 | } |
| 1461 | |
Robin Holt | 7339ff8 | 2006-01-14 13:20:48 -0800 | [diff] [blame] | 1462 | void mpol_shared_policy_init(struct shared_policy *info, int policy, |
| 1463 | nodemask_t *policy_nodes) |
| 1464 | { |
| 1465 | info->root = RB_ROOT; |
| 1466 | spin_lock_init(&info->lock); |
| 1467 | |
| 1468 | if (policy != MPOL_DEFAULT) { |
| 1469 | struct mempolicy *newpol; |
| 1470 | |
| 1471 | /* Falls back to MPOL_DEFAULT on any error */ |
| 1472 | newpol = mpol_new(policy, policy_nodes); |
| 1473 | if (!IS_ERR(newpol)) { |
| 1474 | /* Create pseudo-vma that contains just the policy */ |
| 1475 | struct vm_area_struct pvma; |
| 1476 | |
| 1477 | memset(&pvma, 0, sizeof(struct vm_area_struct)); |
| 1478 | /* Policy covers entire file */ |
| 1479 | pvma.vm_end = TASK_SIZE; |
| 1480 | mpol_set_shared_policy(info, &pvma, newpol); |
| 1481 | mpol_free(newpol); |
| 1482 | } |
| 1483 | } |
| 1484 | } |
| 1485 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1486 | int mpol_set_shared_policy(struct shared_policy *info, |
| 1487 | struct vm_area_struct *vma, struct mempolicy *npol) |
| 1488 | { |
| 1489 | int err; |
| 1490 | struct sp_node *new = NULL; |
| 1491 | unsigned long sz = vma_pages(vma); |
| 1492 | |
| 1493 | PDprintk("set_shared_policy %lx sz %lu %d %lx\n", |
| 1494 | vma->vm_pgoff, |
| 1495 | sz, npol? npol->policy : -1, |
Andi Kleen | dfcd3c0 | 2005-10-29 18:15:48 -0700 | [diff] [blame] | 1496 | npol ? nodes_addr(npol->v.nodes)[0] : -1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1497 | |
| 1498 | if (npol) { |
| 1499 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); |
| 1500 | if (!new) |
| 1501 | return -ENOMEM; |
| 1502 | } |
| 1503 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); |
| 1504 | if (err && new) |
| 1505 | kmem_cache_free(sn_cache, new); |
| 1506 | return err; |
| 1507 | } |
| 1508 | |
| 1509 | /* Free a backing policy store on inode delete. */ |
| 1510 | void mpol_free_shared_policy(struct shared_policy *p) |
| 1511 | { |
| 1512 | struct sp_node *n; |
| 1513 | struct rb_node *next; |
| 1514 | |
| 1515 | if (!p->root.rb_node) |
| 1516 | return; |
| 1517 | spin_lock(&p->lock); |
| 1518 | next = rb_first(&p->root); |
| 1519 | while (next) { |
| 1520 | n = rb_entry(next, struct sp_node, nd); |
| 1521 | next = rb_next(&n->nd); |
Andi Kleen | 90c5029 | 2005-07-27 11:43:50 -0700 | [diff] [blame] | 1522 | rb_erase(&n->nd, &p->root); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1523 | mpol_free(n->policy); |
| 1524 | kmem_cache_free(sn_cache, n); |
| 1525 | } |
| 1526 | spin_unlock(&p->lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1527 | } |
| 1528 | |
| 1529 | /* assumes fs == KERNEL_DS */ |
| 1530 | void __init numa_policy_init(void) |
| 1531 | { |
| 1532 | policy_cache = kmem_cache_create("numa_policy", |
| 1533 | sizeof(struct mempolicy), |
| 1534 | 0, SLAB_PANIC, NULL, NULL); |
| 1535 | |
| 1536 | sn_cache = kmem_cache_create("shared_policy_node", |
| 1537 | sizeof(struct sp_node), |
| 1538 | 0, SLAB_PANIC, NULL, NULL); |
| 1539 | |
| 1540 | /* Set interleaving policy for system init. This way not all |
| 1541 | the data structures allocated at system boot end up in node zero. */ |
| 1542 | |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 1543 | if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1544 | printk("numa_policy_init: interleaving failed\n"); |
| 1545 | } |
| 1546 | |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 1547 | /* Reset policy of current process to default */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1548 | void numa_default_policy(void) |
| 1549 | { |
Christoph Lameter | 8bccd85 | 2005-10-29 18:16:59 -0700 | [diff] [blame] | 1550 | do_set_mempolicy(MPOL_DEFAULT, NULL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1551 | } |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1552 | |
| 1553 | /* Migrate a policy to a different set of nodes */ |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1554 | void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask) |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1555 | { |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1556 | nodemask_t *mpolmask; |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1557 | nodemask_t tmp; |
| 1558 | |
| 1559 | if (!pol) |
| 1560 | return; |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1561 | mpolmask = &pol->cpuset_mems_allowed; |
| 1562 | if (nodes_equal(*mpolmask, *newmask)) |
| 1563 | return; |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1564 | |
| 1565 | switch (pol->policy) { |
| 1566 | case MPOL_DEFAULT: |
| 1567 | break; |
| 1568 | case MPOL_INTERLEAVE: |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1569 | nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask); |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1570 | pol->v.nodes = tmp; |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1571 | *mpolmask = *newmask; |
| 1572 | current->il_next = node_remap(current->il_next, |
| 1573 | *mpolmask, *newmask); |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1574 | break; |
| 1575 | case MPOL_PREFERRED: |
| 1576 | pol->v.preferred_node = node_remap(pol->v.preferred_node, |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1577 | *mpolmask, *newmask); |
| 1578 | *mpolmask = *newmask; |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1579 | break; |
| 1580 | case MPOL_BIND: { |
| 1581 | nodemask_t nodes; |
| 1582 | struct zone **z; |
| 1583 | struct zonelist *zonelist; |
| 1584 | |
| 1585 | nodes_clear(nodes); |
| 1586 | for (z = pol->v.zonelist->zones; *z; z++) |
| 1587 | node_set((*z)->zone_pgdat->node_id, nodes); |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1588 | nodes_remap(tmp, nodes, *mpolmask, *newmask); |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1589 | nodes = tmp; |
| 1590 | |
| 1591 | zonelist = bind_zonelist(&nodes); |
| 1592 | |
| 1593 | /* If no mem, then zonelist is NULL and we keep old zonelist. |
| 1594 | * If that old zonelist has no remaining mems_allowed nodes, |
| 1595 | * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT. |
| 1596 | */ |
| 1597 | |
| 1598 | if (zonelist) { |
| 1599 | /* Good - got mem - substitute new zonelist */ |
| 1600 | kfree(pol->v.zonelist); |
| 1601 | pol->v.zonelist = zonelist; |
| 1602 | } |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1603 | *mpolmask = *newmask; |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1604 | break; |
| 1605 | } |
| 1606 | default: |
| 1607 | BUG(); |
| 1608 | break; |
| 1609 | } |
| 1610 | } |
| 1611 | |
| 1612 | /* |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1613 | * Wrapper for mpol_rebind_policy() that just requires task |
| 1614 | * pointer, and updates task mempolicy. |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1615 | */ |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1616 | |
| 1617 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1618 | { |
Paul Jackson | 74cb215 | 2006-01-08 01:01:56 -0800 | [diff] [blame] | 1619 | mpol_rebind_policy(tsk->mempolicy, new); |
Paul Jackson | 68860ec | 2005-10-30 15:02:36 -0800 | [diff] [blame] | 1620 | } |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1621 | |
| 1622 | /* |
Paul Jackson | 4225399 | 2006-01-08 01:01:59 -0800 | [diff] [blame] | 1623 | * Rebind each vma in mm to new nodemask. |
| 1624 | * |
| 1625 | * Call holding a reference to mm. Takes mm->mmap_sem during call. |
| 1626 | */ |
| 1627 | |
| 1628 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) |
| 1629 | { |
| 1630 | struct vm_area_struct *vma; |
| 1631 | |
| 1632 | down_write(&mm->mmap_sem); |
| 1633 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
| 1634 | mpol_rebind_policy(vma->vm_policy, new); |
| 1635 | up_write(&mm->mmap_sem); |
| 1636 | } |
| 1637 | |
| 1638 | /* |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1639 | * Display pages allocated per node and memory policy via /proc. |
| 1640 | */ |
| 1641 | |
| 1642 | static const char *policy_types[] = { "default", "prefer", "bind", |
| 1643 | "interleave" }; |
| 1644 | |
| 1645 | /* |
| 1646 | * Convert a mempolicy into a string. |
| 1647 | * Returns the number of characters in buffer (if positive) |
| 1648 | * or an error (negative) |
| 1649 | */ |
| 1650 | static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) |
| 1651 | { |
| 1652 | char *p = buffer; |
| 1653 | int l; |
| 1654 | nodemask_t nodes; |
| 1655 | int mode = pol ? pol->policy : MPOL_DEFAULT; |
| 1656 | |
| 1657 | switch (mode) { |
| 1658 | case MPOL_DEFAULT: |
| 1659 | nodes_clear(nodes); |
| 1660 | break; |
| 1661 | |
| 1662 | case MPOL_PREFERRED: |
| 1663 | nodes_clear(nodes); |
| 1664 | node_set(pol->v.preferred_node, nodes); |
| 1665 | break; |
| 1666 | |
| 1667 | case MPOL_BIND: |
| 1668 | get_zonemask(pol, &nodes); |
| 1669 | break; |
| 1670 | |
| 1671 | case MPOL_INTERLEAVE: |
| 1672 | nodes = pol->v.nodes; |
| 1673 | break; |
| 1674 | |
| 1675 | default: |
| 1676 | BUG(); |
| 1677 | return -EFAULT; |
| 1678 | } |
| 1679 | |
| 1680 | l = strlen(policy_types[mode]); |
| 1681 | if (buffer + maxlen < p + l + 1) |
| 1682 | return -ENOSPC; |
| 1683 | |
| 1684 | strcpy(p, policy_types[mode]); |
| 1685 | p += l; |
| 1686 | |
| 1687 | if (!nodes_empty(nodes)) { |
| 1688 | if (buffer + maxlen < p + 2) |
| 1689 | return -ENOSPC; |
| 1690 | *p++ = '='; |
| 1691 | p += nodelist_scnprintf(p, buffer + maxlen - p, nodes); |
| 1692 | } |
| 1693 | return p - buffer; |
| 1694 | } |
| 1695 | |
| 1696 | struct numa_maps { |
| 1697 | unsigned long pages; |
| 1698 | unsigned long anon; |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1699 | unsigned long active; |
| 1700 | unsigned long writeback; |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1701 | unsigned long mapcount_max; |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1702 | unsigned long dirty; |
| 1703 | unsigned long swapcache; |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1704 | unsigned long node[MAX_NUMNODES]; |
| 1705 | }; |
| 1706 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1707 | static void gather_stats(struct page *page, void *private, int pte_dirty) |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1708 | { |
| 1709 | struct numa_maps *md = private; |
| 1710 | int count = page_mapcount(page); |
| 1711 | |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1712 | md->pages++; |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1713 | if (pte_dirty || PageDirty(page)) |
| 1714 | md->dirty++; |
| 1715 | |
| 1716 | if (PageSwapCache(page)) |
| 1717 | md->swapcache++; |
| 1718 | |
| 1719 | if (PageActive(page)) |
| 1720 | md->active++; |
| 1721 | |
| 1722 | if (PageWriteback(page)) |
| 1723 | md->writeback++; |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1724 | |
| 1725 | if (PageAnon(page)) |
| 1726 | md->anon++; |
| 1727 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1728 | if (count > md->mapcount_max) |
| 1729 | md->mapcount_max = count; |
| 1730 | |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1731 | md->node[page_to_nid(page)]++; |
| 1732 | cond_resched(); |
| 1733 | } |
| 1734 | |
Andrew Morton | 7f709ed | 2006-03-07 21:55:22 -0800 | [diff] [blame] | 1735 | #ifdef CONFIG_HUGETLB_PAGE |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1736 | static void check_huge_range(struct vm_area_struct *vma, |
| 1737 | unsigned long start, unsigned long end, |
| 1738 | struct numa_maps *md) |
| 1739 | { |
| 1740 | unsigned long addr; |
| 1741 | struct page *page; |
| 1742 | |
| 1743 | for (addr = start; addr < end; addr += HPAGE_SIZE) { |
| 1744 | pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK); |
| 1745 | pte_t pte; |
| 1746 | |
| 1747 | if (!ptep) |
| 1748 | continue; |
| 1749 | |
| 1750 | pte = *ptep; |
| 1751 | if (pte_none(pte)) |
| 1752 | continue; |
| 1753 | |
| 1754 | page = pte_page(pte); |
| 1755 | if (!page) |
| 1756 | continue; |
| 1757 | |
| 1758 | gather_stats(page, md, pte_dirty(*ptep)); |
| 1759 | } |
| 1760 | } |
Andrew Morton | 7f709ed | 2006-03-07 21:55:22 -0800 | [diff] [blame] | 1761 | #else |
| 1762 | static inline void check_huge_range(struct vm_area_struct *vma, |
| 1763 | unsigned long start, unsigned long end, |
| 1764 | struct numa_maps *md) |
| 1765 | { |
| 1766 | } |
| 1767 | #endif |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1768 | |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1769 | int show_numa_map(struct seq_file *m, void *v) |
| 1770 | { |
| 1771 | struct task_struct *task = m->private; |
| 1772 | struct vm_area_struct *vma = v; |
| 1773 | struct numa_maps *md; |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1774 | struct file *file = vma->vm_file; |
| 1775 | struct mm_struct *mm = vma->vm_mm; |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1776 | int n; |
| 1777 | char buffer[50]; |
| 1778 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1779 | if (!mm) |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1780 | return 0; |
| 1781 | |
| 1782 | md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL); |
| 1783 | if (!md) |
| 1784 | return 0; |
| 1785 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1786 | mpol_to_str(buffer, sizeof(buffer), |
| 1787 | get_vma_policy(task, vma, vma->vm_start)); |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1788 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1789 | seq_printf(m, "%08lx %s", vma->vm_start, buffer); |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1790 | |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1791 | if (file) { |
| 1792 | seq_printf(m, " file="); |
| 1793 | seq_path(m, file->f_vfsmnt, file->f_dentry, "\n\t= "); |
| 1794 | } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) { |
| 1795 | seq_printf(m, " heap"); |
| 1796 | } else if (vma->vm_start <= mm->start_stack && |
| 1797 | vma->vm_end >= mm->start_stack) { |
| 1798 | seq_printf(m, " stack"); |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1799 | } |
Christoph Lameter | 397874d | 2006-03-06 15:42:53 -0800 | [diff] [blame] | 1800 | |
| 1801 | if (is_vm_hugetlb_page(vma)) { |
| 1802 | check_huge_range(vma, vma->vm_start, vma->vm_end, md); |
| 1803 | seq_printf(m, " huge"); |
| 1804 | } else { |
| 1805 | check_pgd_range(vma, vma->vm_start, vma->vm_end, |
| 1806 | &node_online_map, MPOL_MF_STATS, md); |
| 1807 | } |
| 1808 | |
| 1809 | if (!md->pages) |
| 1810 | goto out; |
| 1811 | |
| 1812 | if (md->anon) |
| 1813 | seq_printf(m," anon=%lu",md->anon); |
| 1814 | |
| 1815 | if (md->dirty) |
| 1816 | seq_printf(m," dirty=%lu",md->dirty); |
| 1817 | |
| 1818 | if (md->pages != md->anon && md->pages != md->dirty) |
| 1819 | seq_printf(m, " mapped=%lu", md->pages); |
| 1820 | |
| 1821 | if (md->mapcount_max > 1) |
| 1822 | seq_printf(m, " mapmax=%lu", md->mapcount_max); |
| 1823 | |
| 1824 | if (md->swapcache) |
| 1825 | seq_printf(m," swapcache=%lu", md->swapcache); |
| 1826 | |
| 1827 | if (md->active < md->pages && !is_vm_hugetlb_page(vma)) |
| 1828 | seq_printf(m," active=%lu", md->active); |
| 1829 | |
| 1830 | if (md->writeback) |
| 1831 | seq_printf(m," writeback=%lu", md->writeback); |
| 1832 | |
| 1833 | for_each_online_node(n) |
| 1834 | if (md->node[n]) |
| 1835 | seq_printf(m, " N%d=%lu", n, md->node[n]); |
| 1836 | out: |
| 1837 | seq_putc(m, '\n'); |
Christoph Lameter | 1a75a6c | 2006-01-08 01:01:02 -0800 | [diff] [blame] | 1838 | kfree(md); |
| 1839 | |
| 1840 | if (m->count < m->size) |
| 1841 | m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0; |
| 1842 | return 0; |
| 1843 | } |
| 1844 | |