Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Simple NUMA memory policy for the Linux kernel. | |
3 | * | |
4 | * Copyright 2003,2004 Andi Kleen, SuSE Labs. | |
8bccd85f | 5 | * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc. |
1da177e4 LT |
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. | |
8bccd85f | 21 | * |
1da177e4 LT |
22 | * bind Only allocate memory on a specific set of nodes, |
23 | * no fallback. | |
8bccd85f CL |
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 | * | |
1da177e4 LT |
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. | |
8bccd85f | 33 | * |
1da177e4 LT |
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> | |
dc9aa5b9 | 86 | #include <linux/swap.h> |
1a75a6c8 CL |
87 | #include <linux/seq_file.h> |
88 | #include <linux/proc_fs.h> | |
dc9aa5b9 | 89 | |
1da177e4 LT |
90 | #include <asm/tlbflush.h> |
91 | #include <asm/uaccess.h> | |
92 | ||
38e35860 | 93 | /* Internal flags */ |
dc9aa5b9 | 94 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ |
38e35860 | 95 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ |
1a75a6c8 | 96 | #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */ |
dc9aa5b9 | 97 | |
7e2ab150 CL |
98 | /* The number of pages to migrate per call to migrate_pages() */ |
99 | #define MIGRATE_CHUNK_SIZE 256 | |
100 | ||
1da177e4 LT |
101 | static kmem_cache_t *policy_cache; |
102 | static kmem_cache_t *sn_cache; | |
103 | ||
104 | #define PDprintk(fmt...) | |
105 | ||
106 | /* Highest zone. An specific allocation for a zone below that is not | |
107 | policied. */ | |
4be38e35 | 108 | int policy_zone = ZONE_DMA; |
1da177e4 | 109 | |
d42c6997 | 110 | struct mempolicy default_policy = { |
1da177e4 LT |
111 | .refcnt = ATOMIC_INIT(1), /* never free it */ |
112 | .policy = MPOL_DEFAULT, | |
113 | }; | |
114 | ||
1da177e4 | 115 | /* Do sanity checking on a policy */ |
dfcd3c0d | 116 | static int mpol_check_policy(int mode, nodemask_t *nodes) |
1da177e4 | 117 | { |
dfcd3c0d | 118 | int empty = nodes_empty(*nodes); |
1da177e4 LT |
119 | |
120 | switch (mode) { | |
121 | case MPOL_DEFAULT: | |
122 | if (!empty) | |
123 | return -EINVAL; | |
124 | break; | |
125 | case MPOL_BIND: | |
126 | case MPOL_INTERLEAVE: | |
127 | /* Preferred will only use the first bit, but allow | |
128 | more for now. */ | |
129 | if (empty) | |
130 | return -EINVAL; | |
131 | break; | |
132 | } | |
dfcd3c0d | 133 | return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL; |
1da177e4 | 134 | } |
dd942ae3 | 135 | |
1da177e4 | 136 | /* Generate a custom zonelist for the BIND policy. */ |
dfcd3c0d | 137 | static struct zonelist *bind_zonelist(nodemask_t *nodes) |
1da177e4 LT |
138 | { |
139 | struct zonelist *zl; | |
dd942ae3 | 140 | int num, max, nd, k; |
1da177e4 | 141 | |
dfcd3c0d | 142 | max = 1 + MAX_NR_ZONES * nodes_weight(*nodes); |
dd942ae3 | 143 | zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL); |
1da177e4 LT |
144 | if (!zl) |
145 | return NULL; | |
146 | num = 0; | |
dd942ae3 AK |
147 | /* First put in the highest zones from all nodes, then all the next |
148 | lower zones etc. Avoid empty zones because the memory allocator | |
149 | doesn't like them. If you implement node hot removal you | |
150 | have to fix that. */ | |
151 | for (k = policy_zone; k >= 0; k--) { | |
152 | for_each_node_mask(nd, *nodes) { | |
153 | struct zone *z = &NODE_DATA(nd)->node_zones[k]; | |
154 | if (z->present_pages > 0) | |
155 | zl->zones[num++] = z; | |
156 | } | |
157 | } | |
1da177e4 LT |
158 | zl->zones[num] = NULL; |
159 | return zl; | |
160 | } | |
161 | ||
162 | /* Create a new policy */ | |
dfcd3c0d | 163 | static struct mempolicy *mpol_new(int mode, nodemask_t *nodes) |
1da177e4 LT |
164 | { |
165 | struct mempolicy *policy; | |
166 | ||
dfcd3c0d | 167 | PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]); |
1da177e4 LT |
168 | if (mode == MPOL_DEFAULT) |
169 | return NULL; | |
170 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
171 | if (!policy) | |
172 | return ERR_PTR(-ENOMEM); | |
173 | atomic_set(&policy->refcnt, 1); | |
174 | switch (mode) { | |
175 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 176 | policy->v.nodes = *nodes; |
8f493d79 AK |
177 | if (nodes_weight(*nodes) == 0) { |
178 | kmem_cache_free(policy_cache, policy); | |
179 | return ERR_PTR(-EINVAL); | |
180 | } | |
1da177e4 LT |
181 | break; |
182 | case MPOL_PREFERRED: | |
dfcd3c0d | 183 | policy->v.preferred_node = first_node(*nodes); |
1da177e4 LT |
184 | if (policy->v.preferred_node >= MAX_NUMNODES) |
185 | policy->v.preferred_node = -1; | |
186 | break; | |
187 | case MPOL_BIND: | |
188 | policy->v.zonelist = bind_zonelist(nodes); | |
189 | if (policy->v.zonelist == NULL) { | |
190 | kmem_cache_free(policy_cache, policy); | |
191 | return ERR_PTR(-ENOMEM); | |
192 | } | |
193 | break; | |
194 | } | |
195 | policy->policy = mode; | |
74cb2155 | 196 | policy->cpuset_mems_allowed = cpuset_mems_allowed(current); |
1da177e4 LT |
197 | return policy; |
198 | } | |
199 | ||
397874df | 200 | static void gather_stats(struct page *, void *, int pte_dirty); |
fc301289 CL |
201 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
202 | unsigned long flags); | |
1a75a6c8 | 203 | |
38e35860 | 204 | /* Scan through pages checking if pages follow certain conditions. */ |
b5810039 | 205 | static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd, |
dc9aa5b9 CL |
206 | unsigned long addr, unsigned long end, |
207 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 208 | void *private) |
1da177e4 | 209 | { |
91612e0d HD |
210 | pte_t *orig_pte; |
211 | pte_t *pte; | |
705e87c0 | 212 | spinlock_t *ptl; |
941150a3 | 213 | |
705e87c0 | 214 | orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
91612e0d | 215 | do { |
6aab341e | 216 | struct page *page; |
91612e0d HD |
217 | unsigned int nid; |
218 | ||
219 | if (!pte_present(*pte)) | |
1da177e4 | 220 | continue; |
6aab341e LT |
221 | page = vm_normal_page(vma, addr, *pte); |
222 | if (!page) | |
1da177e4 | 223 | continue; |
053837fc NP |
224 | /* |
225 | * The check for PageReserved here is important to avoid | |
226 | * handling zero pages and other pages that may have been | |
227 | * marked special by the system. | |
228 | * | |
229 | * If the PageReserved would not be checked here then f.e. | |
230 | * the location of the zero page could have an influence | |
231 | * on MPOL_MF_STRICT, zero pages would be counted for | |
232 | * the per node stats, and there would be useless attempts | |
233 | * to put zero pages on the migration list. | |
234 | */ | |
f4598c8b CL |
235 | if (PageReserved(page)) |
236 | continue; | |
6aab341e | 237 | nid = page_to_nid(page); |
38e35860 CL |
238 | if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT)) |
239 | continue; | |
240 | ||
1a75a6c8 | 241 | if (flags & MPOL_MF_STATS) |
397874df | 242 | gather_stats(page, private, pte_dirty(*pte)); |
053837fc | 243 | else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) |
fc301289 | 244 | migrate_page_add(page, private, flags); |
38e35860 CL |
245 | else |
246 | break; | |
91612e0d | 247 | } while (pte++, addr += PAGE_SIZE, addr != end); |
705e87c0 | 248 | pte_unmap_unlock(orig_pte, ptl); |
91612e0d HD |
249 | return addr != end; |
250 | } | |
251 | ||
b5810039 | 252 | static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud, |
dc9aa5b9 CL |
253 | unsigned long addr, unsigned long end, |
254 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 255 | void *private) |
91612e0d HD |
256 | { |
257 | pmd_t *pmd; | |
258 | unsigned long next; | |
259 | ||
260 | pmd = pmd_offset(pud, addr); | |
261 | do { | |
262 | next = pmd_addr_end(addr, end); | |
263 | if (pmd_none_or_clear_bad(pmd)) | |
264 | continue; | |
dc9aa5b9 | 265 | if (check_pte_range(vma, pmd, addr, next, nodes, |
38e35860 | 266 | flags, private)) |
91612e0d HD |
267 | return -EIO; |
268 | } while (pmd++, addr = next, addr != end); | |
269 | return 0; | |
270 | } | |
271 | ||
b5810039 | 272 | static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd, |
dc9aa5b9 CL |
273 | unsigned long addr, unsigned long end, |
274 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 275 | void *private) |
91612e0d HD |
276 | { |
277 | pud_t *pud; | |
278 | unsigned long next; | |
279 | ||
280 | pud = pud_offset(pgd, addr); | |
281 | do { | |
282 | next = pud_addr_end(addr, end); | |
283 | if (pud_none_or_clear_bad(pud)) | |
284 | continue; | |
dc9aa5b9 | 285 | if (check_pmd_range(vma, pud, addr, next, nodes, |
38e35860 | 286 | flags, private)) |
91612e0d HD |
287 | return -EIO; |
288 | } while (pud++, addr = next, addr != end); | |
289 | return 0; | |
290 | } | |
291 | ||
b5810039 | 292 | static inline int check_pgd_range(struct vm_area_struct *vma, |
dc9aa5b9 CL |
293 | unsigned long addr, unsigned long end, |
294 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 295 | void *private) |
91612e0d HD |
296 | { |
297 | pgd_t *pgd; | |
298 | unsigned long next; | |
299 | ||
b5810039 | 300 | pgd = pgd_offset(vma->vm_mm, addr); |
91612e0d HD |
301 | do { |
302 | next = pgd_addr_end(addr, end); | |
303 | if (pgd_none_or_clear_bad(pgd)) | |
304 | continue; | |
dc9aa5b9 | 305 | if (check_pud_range(vma, pgd, addr, next, nodes, |
38e35860 | 306 | flags, private)) |
91612e0d HD |
307 | return -EIO; |
308 | } while (pgd++, addr = next, addr != end); | |
309 | return 0; | |
1da177e4 LT |
310 | } |
311 | ||
dc9aa5b9 CL |
312 | /* Check if a vma is migratable */ |
313 | static inline int vma_migratable(struct vm_area_struct *vma) | |
314 | { | |
315 | if (vma->vm_flags & ( | |
f4598c8b | 316 | VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED)) |
dc9aa5b9 CL |
317 | return 0; |
318 | return 1; | |
319 | } | |
320 | ||
321 | /* | |
322 | * Check if all pages in a range are on a set of nodes. | |
323 | * If pagelist != NULL then isolate pages from the LRU and | |
324 | * put them on the pagelist. | |
325 | */ | |
1da177e4 LT |
326 | static struct vm_area_struct * |
327 | check_range(struct mm_struct *mm, unsigned long start, unsigned long end, | |
38e35860 | 328 | const nodemask_t *nodes, unsigned long flags, void *private) |
1da177e4 LT |
329 | { |
330 | int err; | |
331 | struct vm_area_struct *first, *vma, *prev; | |
332 | ||
053837fc NP |
333 | /* Clear the LRU lists so pages can be isolated */ |
334 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
335 | lru_add_drain_all(); | |
336 | ||
1da177e4 LT |
337 | first = find_vma(mm, start); |
338 | if (!first) | |
339 | return ERR_PTR(-EFAULT); | |
340 | prev = NULL; | |
341 | for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) { | |
dc9aa5b9 CL |
342 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { |
343 | if (!vma->vm_next && vma->vm_end < end) | |
344 | return ERR_PTR(-EFAULT); | |
345 | if (prev && prev->vm_end < vma->vm_start) | |
346 | return ERR_PTR(-EFAULT); | |
347 | } | |
348 | if (!is_vm_hugetlb_page(vma) && | |
349 | ((flags & MPOL_MF_STRICT) || | |
350 | ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) && | |
351 | vma_migratable(vma)))) { | |
5b952b3c | 352 | unsigned long endvma = vma->vm_end; |
dc9aa5b9 | 353 | |
5b952b3c AK |
354 | if (endvma > end) |
355 | endvma = end; | |
356 | if (vma->vm_start > start) | |
357 | start = vma->vm_start; | |
dc9aa5b9 | 358 | err = check_pgd_range(vma, start, endvma, nodes, |
38e35860 | 359 | flags, private); |
1da177e4 LT |
360 | if (err) { |
361 | first = ERR_PTR(err); | |
362 | break; | |
363 | } | |
364 | } | |
365 | prev = vma; | |
366 | } | |
367 | return first; | |
368 | } | |
369 | ||
370 | /* Apply policy to a single VMA */ | |
371 | static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new) | |
372 | { | |
373 | int err = 0; | |
374 | struct mempolicy *old = vma->vm_policy; | |
375 | ||
376 | PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", | |
377 | vma->vm_start, vma->vm_end, vma->vm_pgoff, | |
378 | vma->vm_ops, vma->vm_file, | |
379 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
380 | ||
381 | if (vma->vm_ops && vma->vm_ops->set_policy) | |
382 | err = vma->vm_ops->set_policy(vma, new); | |
383 | if (!err) { | |
384 | mpol_get(new); | |
385 | vma->vm_policy = new; | |
386 | mpol_free(old); | |
387 | } | |
388 | return err; | |
389 | } | |
390 | ||
391 | /* Step 2: apply policy to a range and do splits. */ | |
392 | static int mbind_range(struct vm_area_struct *vma, unsigned long start, | |
393 | unsigned long end, struct mempolicy *new) | |
394 | { | |
395 | struct vm_area_struct *next; | |
396 | int err; | |
397 | ||
398 | err = 0; | |
399 | for (; vma && vma->vm_start < end; vma = next) { | |
400 | next = vma->vm_next; | |
401 | if (vma->vm_start < start) | |
402 | err = split_vma(vma->vm_mm, vma, start, 1); | |
403 | if (!err && vma->vm_end > end) | |
404 | err = split_vma(vma->vm_mm, vma, end, 0); | |
405 | if (!err) | |
406 | err = policy_vma(vma, new); | |
407 | if (err) | |
408 | break; | |
409 | } | |
410 | return err; | |
411 | } | |
412 | ||
8bccd85f CL |
413 | static int contextualize_policy(int mode, nodemask_t *nodes) |
414 | { | |
415 | if (!nodes) | |
416 | return 0; | |
417 | ||
cf2a473c | 418 | cpuset_update_task_memory_state(); |
5966514d PJ |
419 | if (!cpuset_nodes_subset_current_mems_allowed(*nodes)) |
420 | return -EINVAL; | |
8bccd85f CL |
421 | return mpol_check_policy(mode, nodes); |
422 | } | |
423 | ||
1da177e4 | 424 | /* Set the process memory policy */ |
8bccd85f | 425 | long do_set_mempolicy(int mode, nodemask_t *nodes) |
1da177e4 | 426 | { |
1da177e4 | 427 | struct mempolicy *new; |
1da177e4 | 428 | |
8bccd85f | 429 | if (contextualize_policy(mode, nodes)) |
1da177e4 | 430 | return -EINVAL; |
8bccd85f | 431 | new = mpol_new(mode, nodes); |
1da177e4 LT |
432 | if (IS_ERR(new)) |
433 | return PTR_ERR(new); | |
434 | mpol_free(current->mempolicy); | |
435 | current->mempolicy = new; | |
436 | if (new && new->policy == MPOL_INTERLEAVE) | |
dfcd3c0d | 437 | current->il_next = first_node(new->v.nodes); |
1da177e4 LT |
438 | return 0; |
439 | } | |
440 | ||
441 | /* Fill a zone bitmap for a policy */ | |
dfcd3c0d | 442 | static void get_zonemask(struct mempolicy *p, nodemask_t *nodes) |
1da177e4 LT |
443 | { |
444 | int i; | |
445 | ||
dfcd3c0d | 446 | nodes_clear(*nodes); |
1da177e4 LT |
447 | switch (p->policy) { |
448 | case MPOL_BIND: | |
449 | for (i = 0; p->v.zonelist->zones[i]; i++) | |
8bccd85f CL |
450 | node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id, |
451 | *nodes); | |
1da177e4 LT |
452 | break; |
453 | case MPOL_DEFAULT: | |
454 | break; | |
455 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 456 | *nodes = p->v.nodes; |
1da177e4 LT |
457 | break; |
458 | case MPOL_PREFERRED: | |
459 | /* or use current node instead of online map? */ | |
460 | if (p->v.preferred_node < 0) | |
dfcd3c0d | 461 | *nodes = node_online_map; |
1da177e4 | 462 | else |
dfcd3c0d | 463 | node_set(p->v.preferred_node, *nodes); |
1da177e4 LT |
464 | break; |
465 | default: | |
466 | BUG(); | |
467 | } | |
468 | } | |
469 | ||
470 | static int lookup_node(struct mm_struct *mm, unsigned long addr) | |
471 | { | |
472 | struct page *p; | |
473 | int err; | |
474 | ||
475 | err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL); | |
476 | if (err >= 0) { | |
477 | err = page_to_nid(p); | |
478 | put_page(p); | |
479 | } | |
480 | return err; | |
481 | } | |
482 | ||
1da177e4 | 483 | /* Retrieve NUMA policy */ |
8bccd85f CL |
484 | long do_get_mempolicy(int *policy, nodemask_t *nmask, |
485 | unsigned long addr, unsigned long flags) | |
1da177e4 | 486 | { |
8bccd85f | 487 | int err; |
1da177e4 LT |
488 | struct mm_struct *mm = current->mm; |
489 | struct vm_area_struct *vma = NULL; | |
490 | struct mempolicy *pol = current->mempolicy; | |
491 | ||
cf2a473c | 492 | cpuset_update_task_memory_state(); |
1da177e4 LT |
493 | if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR)) |
494 | return -EINVAL; | |
1da177e4 LT |
495 | if (flags & MPOL_F_ADDR) { |
496 | down_read(&mm->mmap_sem); | |
497 | vma = find_vma_intersection(mm, addr, addr+1); | |
498 | if (!vma) { | |
499 | up_read(&mm->mmap_sem); | |
500 | return -EFAULT; | |
501 | } | |
502 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
503 | pol = vma->vm_ops->get_policy(vma, addr); | |
504 | else | |
505 | pol = vma->vm_policy; | |
506 | } else if (addr) | |
507 | return -EINVAL; | |
508 | ||
509 | if (!pol) | |
510 | pol = &default_policy; | |
511 | ||
512 | if (flags & MPOL_F_NODE) { | |
513 | if (flags & MPOL_F_ADDR) { | |
514 | err = lookup_node(mm, addr); | |
515 | if (err < 0) | |
516 | goto out; | |
8bccd85f | 517 | *policy = err; |
1da177e4 LT |
518 | } else if (pol == current->mempolicy && |
519 | pol->policy == MPOL_INTERLEAVE) { | |
8bccd85f | 520 | *policy = current->il_next; |
1da177e4 LT |
521 | } else { |
522 | err = -EINVAL; | |
523 | goto out; | |
524 | } | |
525 | } else | |
8bccd85f | 526 | *policy = pol->policy; |
1da177e4 LT |
527 | |
528 | if (vma) { | |
529 | up_read(¤t->mm->mmap_sem); | |
530 | vma = NULL; | |
531 | } | |
532 | ||
1da177e4 | 533 | err = 0; |
8bccd85f CL |
534 | if (nmask) |
535 | get_zonemask(pol, nmask); | |
1da177e4 LT |
536 | |
537 | out: | |
538 | if (vma) | |
539 | up_read(¤t->mm->mmap_sem); | |
540 | return err; | |
541 | } | |
542 | ||
6ce3c4c0 CL |
543 | /* |
544 | * page migration | |
545 | */ | |
546 | ||
fc301289 CL |
547 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
548 | unsigned long flags) | |
6ce3c4c0 CL |
549 | { |
550 | /* | |
fc301289 | 551 | * Avoid migrating a page that is shared with others. |
6ce3c4c0 | 552 | */ |
fc301289 | 553 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) { |
053837fc | 554 | if (isolate_lru_page(page)) |
1e275d40 | 555 | list_add_tail(&page->lru, pagelist); |
6ce3c4c0 CL |
556 | } |
557 | } | |
558 | ||
7e2ab150 CL |
559 | /* |
560 | * Migrate the list 'pagelist' of pages to a certain destination. | |
561 | * | |
562 | * Specify destination with either non-NULL vma or dest_node >= 0 | |
563 | * Return the number of pages not migrated or error code | |
564 | */ | |
565 | static int migrate_pages_to(struct list_head *pagelist, | |
566 | struct vm_area_struct *vma, int dest) | |
6ce3c4c0 | 567 | { |
7e2ab150 | 568 | LIST_HEAD(newlist); |
6ce3c4c0 CL |
569 | LIST_HEAD(moved); |
570 | LIST_HEAD(failed); | |
7e2ab150 | 571 | int err = 0; |
1e275d40 | 572 | unsigned long offset = 0; |
7e2ab150 CL |
573 | int nr_pages; |
574 | struct page *page; | |
575 | struct list_head *p; | |
576 | ||
577 | redo: | |
578 | nr_pages = 0; | |
579 | list_for_each(p, pagelist) { | |
1e275d40 CL |
580 | if (vma) { |
581 | /* | |
582 | * The address passed to alloc_page_vma is used to | |
583 | * generate the proper interleave behavior. We fake | |
584 | * the address here by an increasing offset in order | |
585 | * to get the proper distribution of pages. | |
586 | * | |
587 | * No decision has been made as to which page | |
588 | * a certain old page is moved to so we cannot | |
589 | * specify the correct address. | |
590 | */ | |
591 | page = alloc_page_vma(GFP_HIGHUSER, vma, | |
592 | offset + vma->vm_start); | |
593 | offset += PAGE_SIZE; | |
594 | } | |
7e2ab150 CL |
595 | else |
596 | page = alloc_pages_node(dest, GFP_HIGHUSER, 0); | |
597 | ||
598 | if (!page) { | |
599 | err = -ENOMEM; | |
600 | goto out; | |
601 | } | |
1e275d40 | 602 | list_add_tail(&page->lru, &newlist); |
7e2ab150 | 603 | nr_pages++; |
fcab6f35 | 604 | if (nr_pages > MIGRATE_CHUNK_SIZE) |
7e2ab150 CL |
605 | break; |
606 | } | |
607 | err = migrate_pages(pagelist, &newlist, &moved, &failed); | |
608 | ||
609 | putback_lru_pages(&moved); /* Call release pages instead ?? */ | |
610 | ||
611 | if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist)) | |
612 | goto redo; | |
613 | out: | |
614 | /* Return leftover allocated pages */ | |
615 | while (!list_empty(&newlist)) { | |
616 | page = list_entry(newlist.next, struct page, lru); | |
617 | list_del(&page->lru); | |
618 | __free_page(page); | |
619 | } | |
620 | list_splice(&failed, pagelist); | |
621 | if (err < 0) | |
622 | return err; | |
623 | ||
624 | /* Calculate number of leftover pages */ | |
625 | nr_pages = 0; | |
626 | list_for_each(p, pagelist) | |
627 | nr_pages++; | |
628 | return nr_pages; | |
629 | } | |
6ce3c4c0 | 630 | |
7e2ab150 CL |
631 | /* |
632 | * Migrate pages from one node to a target node. | |
633 | * Returns error or the number of pages not migrated. | |
634 | */ | |
635 | int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags) | |
636 | { | |
637 | nodemask_t nmask; | |
638 | LIST_HEAD(pagelist); | |
639 | int err = 0; | |
640 | ||
641 | nodes_clear(nmask); | |
642 | node_set(source, nmask); | |
6ce3c4c0 | 643 | |
7e2ab150 CL |
644 | check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask, |
645 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); | |
646 | ||
647 | if (!list_empty(&pagelist)) { | |
648 | err = migrate_pages_to(&pagelist, NULL, dest); | |
649 | if (!list_empty(&pagelist)) | |
650 | putback_lru_pages(&pagelist); | |
651 | } | |
652 | return err; | |
6ce3c4c0 CL |
653 | } |
654 | ||
39743889 | 655 | /* |
7e2ab150 CL |
656 | * Move pages between the two nodesets so as to preserve the physical |
657 | * layout as much as possible. | |
39743889 CL |
658 | * |
659 | * Returns the number of page that could not be moved. | |
660 | */ | |
661 | int do_migrate_pages(struct mm_struct *mm, | |
662 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
663 | { | |
664 | LIST_HEAD(pagelist); | |
7e2ab150 CL |
665 | int busy = 0; |
666 | int err = 0; | |
667 | nodemask_t tmp; | |
39743889 | 668 | |
7e2ab150 | 669 | down_read(&mm->mmap_sem); |
39743889 | 670 | |
7e2ab150 CL |
671 | /* |
672 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' | |
673 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' | |
674 | * bit in 'tmp', and return that <source, dest> pair for migration. | |
675 | * The pair of nodemasks 'to' and 'from' define the map. | |
676 | * | |
677 | * If no pair of bits is found that way, fallback to picking some | |
678 | * pair of 'source' and 'dest' bits that are not the same. If the | |
679 | * 'source' and 'dest' bits are the same, this represents a node | |
680 | * that will be migrating to itself, so no pages need move. | |
681 | * | |
682 | * If no bits are left in 'tmp', or if all remaining bits left | |
683 | * in 'tmp' correspond to the same bit in 'to', return false | |
684 | * (nothing left to migrate). | |
685 | * | |
686 | * This lets us pick a pair of nodes to migrate between, such that | |
687 | * if possible the dest node is not already occupied by some other | |
688 | * source node, minimizing the risk of overloading the memory on a | |
689 | * node that would happen if we migrated incoming memory to a node | |
690 | * before migrating outgoing memory source that same node. | |
691 | * | |
692 | * A single scan of tmp is sufficient. As we go, we remember the | |
693 | * most recent <s, d> pair that moved (s != d). If we find a pair | |
694 | * that not only moved, but what's better, moved to an empty slot | |
695 | * (d is not set in tmp), then we break out then, with that pair. | |
696 | * Otherwise when we finish scannng from_tmp, we at least have the | |
697 | * most recent <s, d> pair that moved. If we get all the way through | |
698 | * the scan of tmp without finding any node that moved, much less | |
699 | * moved to an empty node, then there is nothing left worth migrating. | |
700 | */ | |
d4984711 | 701 | |
7e2ab150 CL |
702 | tmp = *from_nodes; |
703 | while (!nodes_empty(tmp)) { | |
704 | int s,d; | |
705 | int source = -1; | |
706 | int dest = 0; | |
707 | ||
708 | for_each_node_mask(s, tmp) { | |
709 | d = node_remap(s, *from_nodes, *to_nodes); | |
710 | if (s == d) | |
711 | continue; | |
712 | ||
713 | source = s; /* Node moved. Memorize */ | |
714 | dest = d; | |
715 | ||
716 | /* dest not in remaining from nodes? */ | |
717 | if (!node_isset(dest, tmp)) | |
718 | break; | |
719 | } | |
720 | if (source == -1) | |
721 | break; | |
722 | ||
723 | node_clear(source, tmp); | |
724 | err = migrate_to_node(mm, source, dest, flags); | |
725 | if (err > 0) | |
726 | busy += err; | |
727 | if (err < 0) | |
728 | break; | |
39743889 | 729 | } |
d4984711 | 730 | |
39743889 | 731 | up_read(&mm->mmap_sem); |
7e2ab150 CL |
732 | if (err < 0) |
733 | return err; | |
734 | return busy; | |
39743889 CL |
735 | } |
736 | ||
6ce3c4c0 CL |
737 | long do_mbind(unsigned long start, unsigned long len, |
738 | unsigned long mode, nodemask_t *nmask, unsigned long flags) | |
739 | { | |
740 | struct vm_area_struct *vma; | |
741 | struct mm_struct *mm = current->mm; | |
742 | struct mempolicy *new; | |
743 | unsigned long end; | |
744 | int err; | |
745 | LIST_HEAD(pagelist); | |
746 | ||
747 | if ((flags & ~(unsigned long)(MPOL_MF_STRICT | | |
748 | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
749 | || mode > MPOL_MAX) | |
750 | return -EINVAL; | |
751 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE)) | |
752 | return -EPERM; | |
753 | ||
754 | if (start & ~PAGE_MASK) | |
755 | return -EINVAL; | |
756 | ||
757 | if (mode == MPOL_DEFAULT) | |
758 | flags &= ~MPOL_MF_STRICT; | |
759 | ||
760 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; | |
761 | end = start + len; | |
762 | ||
763 | if (end < start) | |
764 | return -EINVAL; | |
765 | if (end == start) | |
766 | return 0; | |
767 | ||
768 | if (mpol_check_policy(mode, nmask)) | |
769 | return -EINVAL; | |
770 | ||
771 | new = mpol_new(mode, nmask); | |
772 | if (IS_ERR(new)) | |
773 | return PTR_ERR(new); | |
774 | ||
775 | /* | |
776 | * If we are using the default policy then operation | |
777 | * on discontinuous address spaces is okay after all | |
778 | */ | |
779 | if (!new) | |
780 | flags |= MPOL_MF_DISCONTIG_OK; | |
781 | ||
782 | PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len, | |
783 | mode,nodes_addr(nodes)[0]); | |
784 | ||
785 | down_write(&mm->mmap_sem); | |
786 | vma = check_range(mm, start, end, nmask, | |
787 | flags | MPOL_MF_INVERT, &pagelist); | |
788 | ||
789 | err = PTR_ERR(vma); | |
790 | if (!IS_ERR(vma)) { | |
791 | int nr_failed = 0; | |
792 | ||
793 | err = mbind_range(vma, start, end, new); | |
7e2ab150 | 794 | |
6ce3c4c0 | 795 | if (!list_empty(&pagelist)) |
7e2ab150 | 796 | nr_failed = migrate_pages_to(&pagelist, vma, -1); |
6ce3c4c0 CL |
797 | |
798 | if (!err && nr_failed && (flags & MPOL_MF_STRICT)) | |
799 | err = -EIO; | |
800 | } | |
801 | if (!list_empty(&pagelist)) | |
802 | putback_lru_pages(&pagelist); | |
803 | ||
804 | up_write(&mm->mmap_sem); | |
805 | mpol_free(new); | |
806 | return err; | |
807 | } | |
808 | ||
8bccd85f CL |
809 | /* |
810 | * User space interface with variable sized bitmaps for nodelists. | |
811 | */ | |
812 | ||
813 | /* Copy a node mask from user space. */ | |
39743889 | 814 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
8bccd85f CL |
815 | unsigned long maxnode) |
816 | { | |
817 | unsigned long k; | |
818 | unsigned long nlongs; | |
819 | unsigned long endmask; | |
820 | ||
821 | --maxnode; | |
822 | nodes_clear(*nodes); | |
823 | if (maxnode == 0 || !nmask) | |
824 | return 0; | |
a9c930ba | 825 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
636f13c1 | 826 | return -EINVAL; |
8bccd85f CL |
827 | |
828 | nlongs = BITS_TO_LONGS(maxnode); | |
829 | if ((maxnode % BITS_PER_LONG) == 0) | |
830 | endmask = ~0UL; | |
831 | else | |
832 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
833 | ||
834 | /* When the user specified more nodes than supported just check | |
835 | if the non supported part is all zero. */ | |
836 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
837 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
838 | return -EINVAL; | |
839 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
840 | unsigned long t; | |
841 | if (get_user(t, nmask + k)) | |
842 | return -EFAULT; | |
843 | if (k == nlongs - 1) { | |
844 | if (t & endmask) | |
845 | return -EINVAL; | |
846 | } else if (t) | |
847 | return -EINVAL; | |
848 | } | |
849 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
850 | endmask = ~0UL; | |
851 | } | |
852 | ||
853 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
854 | return -EFAULT; | |
855 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
856 | return 0; | |
857 | } | |
858 | ||
859 | /* Copy a kernel node mask to user space */ | |
860 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
861 | nodemask_t *nodes) | |
862 | { | |
863 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
864 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
865 | ||
866 | if (copy > nbytes) { | |
867 | if (copy > PAGE_SIZE) | |
868 | return -EINVAL; | |
869 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
870 | return -EFAULT; | |
871 | copy = nbytes; | |
872 | } | |
873 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
874 | } | |
875 | ||
876 | asmlinkage long sys_mbind(unsigned long start, unsigned long len, | |
877 | unsigned long mode, | |
878 | unsigned long __user *nmask, unsigned long maxnode, | |
879 | unsigned flags) | |
880 | { | |
881 | nodemask_t nodes; | |
882 | int err; | |
883 | ||
884 | err = get_nodes(&nodes, nmask, maxnode); | |
885 | if (err) | |
886 | return err; | |
887 | return do_mbind(start, len, mode, &nodes, flags); | |
888 | } | |
889 | ||
890 | /* Set the process memory policy */ | |
891 | asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask, | |
892 | unsigned long maxnode) | |
893 | { | |
894 | int err; | |
895 | nodemask_t nodes; | |
896 | ||
897 | if (mode < 0 || mode > MPOL_MAX) | |
898 | return -EINVAL; | |
899 | err = get_nodes(&nodes, nmask, maxnode); | |
900 | if (err) | |
901 | return err; | |
902 | return do_set_mempolicy(mode, &nodes); | |
903 | } | |
904 | ||
39743889 CL |
905 | asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode, |
906 | const unsigned long __user *old_nodes, | |
907 | const unsigned long __user *new_nodes) | |
908 | { | |
909 | struct mm_struct *mm; | |
910 | struct task_struct *task; | |
911 | nodemask_t old; | |
912 | nodemask_t new; | |
913 | nodemask_t task_nodes; | |
914 | int err; | |
915 | ||
916 | err = get_nodes(&old, old_nodes, maxnode); | |
917 | if (err) | |
918 | return err; | |
919 | ||
920 | err = get_nodes(&new, new_nodes, maxnode); | |
921 | if (err) | |
922 | return err; | |
923 | ||
924 | /* Find the mm_struct */ | |
925 | read_lock(&tasklist_lock); | |
926 | task = pid ? find_task_by_pid(pid) : current; | |
927 | if (!task) { | |
928 | read_unlock(&tasklist_lock); | |
929 | return -ESRCH; | |
930 | } | |
931 | mm = get_task_mm(task); | |
932 | read_unlock(&tasklist_lock); | |
933 | ||
934 | if (!mm) | |
935 | return -EINVAL; | |
936 | ||
937 | /* | |
938 | * Check if this process has the right to modify the specified | |
939 | * process. The right exists if the process has administrative | |
940 | * capabilities, superuser priviledges or the same | |
941 | * userid as the target process. | |
942 | */ | |
943 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
944 | (current->uid != task->suid) && (current->uid != task->uid) && | |
945 | !capable(CAP_SYS_ADMIN)) { | |
946 | err = -EPERM; | |
947 | goto out; | |
948 | } | |
949 | ||
950 | task_nodes = cpuset_mems_allowed(task); | |
951 | /* Is the user allowed to access the target nodes? */ | |
952 | if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) { | |
953 | err = -EPERM; | |
954 | goto out; | |
955 | } | |
956 | ||
511030bc CL |
957 | err = do_migrate_pages(mm, &old, &new, |
958 | capable(CAP_SYS_ADMIN) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); | |
39743889 CL |
959 | out: |
960 | mmput(mm); | |
961 | return err; | |
962 | } | |
963 | ||
964 | ||
8bccd85f CL |
965 | /* Retrieve NUMA policy */ |
966 | asmlinkage long sys_get_mempolicy(int __user *policy, | |
967 | unsigned long __user *nmask, | |
968 | unsigned long maxnode, | |
969 | unsigned long addr, unsigned long flags) | |
970 | { | |
971 | int err, pval; | |
972 | nodemask_t nodes; | |
973 | ||
974 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
975 | return -EINVAL; | |
976 | ||
977 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
978 | ||
979 | if (err) | |
980 | return err; | |
981 | ||
982 | if (policy && put_user(pval, policy)) | |
983 | return -EFAULT; | |
984 | ||
985 | if (nmask) | |
986 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
987 | ||
988 | return err; | |
989 | } | |
990 | ||
1da177e4 LT |
991 | #ifdef CONFIG_COMPAT |
992 | ||
993 | asmlinkage long compat_sys_get_mempolicy(int __user *policy, | |
994 | compat_ulong_t __user *nmask, | |
995 | compat_ulong_t maxnode, | |
996 | compat_ulong_t addr, compat_ulong_t flags) | |
997 | { | |
998 | long err; | |
999 | unsigned long __user *nm = NULL; | |
1000 | unsigned long nr_bits, alloc_size; | |
1001 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1002 | ||
1003 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1004 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1005 | ||
1006 | if (nmask) | |
1007 | nm = compat_alloc_user_space(alloc_size); | |
1008 | ||
1009 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
1010 | ||
1011 | if (!err && nmask) { | |
1012 | err = copy_from_user(bm, nm, alloc_size); | |
1013 | /* ensure entire bitmap is zeroed */ | |
1014 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
1015 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
1016 | } | |
1017 | ||
1018 | return err; | |
1019 | } | |
1020 | ||
1021 | asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask, | |
1022 | compat_ulong_t maxnode) | |
1023 | { | |
1024 | long err = 0; | |
1025 | unsigned long __user *nm = NULL; | |
1026 | unsigned long nr_bits, alloc_size; | |
1027 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1028 | ||
1029 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1030 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1031 | ||
1032 | if (nmask) { | |
1033 | err = compat_get_bitmap(bm, nmask, nr_bits); | |
1034 | nm = compat_alloc_user_space(alloc_size); | |
1035 | err |= copy_to_user(nm, bm, alloc_size); | |
1036 | } | |
1037 | ||
1038 | if (err) | |
1039 | return -EFAULT; | |
1040 | ||
1041 | return sys_set_mempolicy(mode, nm, nr_bits+1); | |
1042 | } | |
1043 | ||
1044 | asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len, | |
1045 | compat_ulong_t mode, compat_ulong_t __user *nmask, | |
1046 | compat_ulong_t maxnode, compat_ulong_t flags) | |
1047 | { | |
1048 | long err = 0; | |
1049 | unsigned long __user *nm = NULL; | |
1050 | unsigned long nr_bits, alloc_size; | |
dfcd3c0d | 1051 | nodemask_t bm; |
1da177e4 LT |
1052 | |
1053 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1054 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1055 | ||
1056 | if (nmask) { | |
dfcd3c0d | 1057 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); |
1da177e4 | 1058 | nm = compat_alloc_user_space(alloc_size); |
dfcd3c0d | 1059 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); |
1da177e4 LT |
1060 | } |
1061 | ||
1062 | if (err) | |
1063 | return -EFAULT; | |
1064 | ||
1065 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); | |
1066 | } | |
1067 | ||
1068 | #endif | |
1069 | ||
1070 | /* Return effective policy for a VMA */ | |
48fce342 CL |
1071 | static struct mempolicy * get_vma_policy(struct task_struct *task, |
1072 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 1073 | { |
6e21c8f1 | 1074 | struct mempolicy *pol = task->mempolicy; |
1da177e4 LT |
1075 | |
1076 | if (vma) { | |
1077 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
8bccd85f | 1078 | pol = vma->vm_ops->get_policy(vma, addr); |
1da177e4 LT |
1079 | else if (vma->vm_policy && |
1080 | vma->vm_policy->policy != MPOL_DEFAULT) | |
1081 | pol = vma->vm_policy; | |
1082 | } | |
1083 | if (!pol) | |
1084 | pol = &default_policy; | |
1085 | return pol; | |
1086 | } | |
1087 | ||
1088 | /* Return a zonelist representing a mempolicy */ | |
dd0fc66f | 1089 | static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy) |
1da177e4 LT |
1090 | { |
1091 | int nd; | |
1092 | ||
1093 | switch (policy->policy) { | |
1094 | case MPOL_PREFERRED: | |
1095 | nd = policy->v.preferred_node; | |
1096 | if (nd < 0) | |
1097 | nd = numa_node_id(); | |
1098 | break; | |
1099 | case MPOL_BIND: | |
1100 | /* Lower zones don't get a policy applied */ | |
1101 | /* Careful: current->mems_allowed might have moved */ | |
af4ca457 | 1102 | if (gfp_zone(gfp) >= policy_zone) |
1da177e4 LT |
1103 | if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist)) |
1104 | return policy->v.zonelist; | |
1105 | /*FALL THROUGH*/ | |
1106 | case MPOL_INTERLEAVE: /* should not happen */ | |
1107 | case MPOL_DEFAULT: | |
1108 | nd = numa_node_id(); | |
1109 | break; | |
1110 | default: | |
1111 | nd = 0; | |
1112 | BUG(); | |
1113 | } | |
af4ca457 | 1114 | return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp); |
1da177e4 LT |
1115 | } |
1116 | ||
1117 | /* Do dynamic interleaving for a process */ | |
1118 | static unsigned interleave_nodes(struct mempolicy *policy) | |
1119 | { | |
1120 | unsigned nid, next; | |
1121 | struct task_struct *me = current; | |
1122 | ||
1123 | nid = me->il_next; | |
dfcd3c0d | 1124 | next = next_node(nid, policy->v.nodes); |
1da177e4 | 1125 | if (next >= MAX_NUMNODES) |
dfcd3c0d | 1126 | next = first_node(policy->v.nodes); |
1da177e4 LT |
1127 | me->il_next = next; |
1128 | return nid; | |
1129 | } | |
1130 | ||
dc85da15 CL |
1131 | /* |
1132 | * Depending on the memory policy provide a node from which to allocate the | |
1133 | * next slab entry. | |
1134 | */ | |
1135 | unsigned slab_node(struct mempolicy *policy) | |
1136 | { | |
dc85da15 CL |
1137 | switch (policy->policy) { |
1138 | case MPOL_INTERLEAVE: | |
1139 | return interleave_nodes(policy); | |
1140 | ||
1141 | case MPOL_BIND: | |
1142 | /* | |
1143 | * Follow bind policy behavior and start allocation at the | |
1144 | * first node. | |
1145 | */ | |
1146 | return policy->v.zonelist->zones[0]->zone_pgdat->node_id; | |
1147 | ||
1148 | case MPOL_PREFERRED: | |
1149 | if (policy->v.preferred_node >= 0) | |
1150 | return policy->v.preferred_node; | |
1151 | /* Fall through */ | |
1152 | ||
1153 | default: | |
1154 | return numa_node_id(); | |
1155 | } | |
1156 | } | |
1157 | ||
1da177e4 LT |
1158 | /* Do static interleaving for a VMA with known offset. */ |
1159 | static unsigned offset_il_node(struct mempolicy *pol, | |
1160 | struct vm_area_struct *vma, unsigned long off) | |
1161 | { | |
dfcd3c0d | 1162 | unsigned nnodes = nodes_weight(pol->v.nodes); |
1da177e4 LT |
1163 | unsigned target = (unsigned)off % nnodes; |
1164 | int c; | |
1165 | int nid = -1; | |
1166 | ||
1167 | c = 0; | |
1168 | do { | |
dfcd3c0d | 1169 | nid = next_node(nid, pol->v.nodes); |
1da177e4 LT |
1170 | c++; |
1171 | } while (c <= target); | |
1da177e4 LT |
1172 | return nid; |
1173 | } | |
1174 | ||
5da7ca86 CL |
1175 | /* Determine a node number for interleave */ |
1176 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1177 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1178 | { | |
1179 | if (vma) { | |
1180 | unsigned long off; | |
1181 | ||
1182 | off = vma->vm_pgoff; | |
1183 | off += (addr - vma->vm_start) >> shift; | |
1184 | return offset_il_node(pol, vma, off); | |
1185 | } else | |
1186 | return interleave_nodes(pol); | |
1187 | } | |
1188 | ||
00ac59ad | 1189 | #ifdef CONFIG_HUGETLBFS |
5da7ca86 CL |
1190 | /* Return a zonelist suitable for a huge page allocation. */ |
1191 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr) | |
1192 | { | |
1193 | struct mempolicy *pol = get_vma_policy(current, vma, addr); | |
1194 | ||
1195 | if (pol->policy == MPOL_INTERLEAVE) { | |
1196 | unsigned nid; | |
1197 | ||
1198 | nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT); | |
1199 | return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER); | |
1200 | } | |
1201 | return zonelist_policy(GFP_HIGHUSER, pol); | |
1202 | } | |
00ac59ad | 1203 | #endif |
5da7ca86 | 1204 | |
1da177e4 LT |
1205 | /* Allocate a page in interleaved policy. |
1206 | Own path because it needs to do special accounting. */ | |
662f3a0b AK |
1207 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
1208 | unsigned nid) | |
1da177e4 LT |
1209 | { |
1210 | struct zonelist *zl; | |
1211 | struct page *page; | |
1212 | ||
af4ca457 | 1213 | zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp); |
1da177e4 LT |
1214 | page = __alloc_pages(gfp, order, zl); |
1215 | if (page && page_zone(page) == zl->zones[0]) { | |
e7c8d5c9 | 1216 | zone_pcp(zl->zones[0],get_cpu())->interleave_hit++; |
1da177e4 LT |
1217 | put_cpu(); |
1218 | } | |
1219 | return page; | |
1220 | } | |
1221 | ||
1222 | /** | |
1223 | * alloc_page_vma - Allocate a page for a VMA. | |
1224 | * | |
1225 | * @gfp: | |
1226 | * %GFP_USER user allocation. | |
1227 | * %GFP_KERNEL kernel allocations, | |
1228 | * %GFP_HIGHMEM highmem/user allocations, | |
1229 | * %GFP_FS allocation should not call back into a file system. | |
1230 | * %GFP_ATOMIC don't sleep. | |
1231 | * | |
1232 | * @vma: Pointer to VMA or NULL if not available. | |
1233 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
1234 | * | |
1235 | * This function allocates a page from the kernel page pool and applies | |
1236 | * a NUMA policy associated with the VMA or the current process. | |
1237 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1238 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
1239 | * all allocations for pages that will be mapped into | |
1240 | * user space. Returns NULL when no page can be allocated. | |
1241 | * | |
1242 | * Should be called with the mm_sem of the vma hold. | |
1243 | */ | |
1244 | struct page * | |
dd0fc66f | 1245 | alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr) |
1da177e4 | 1246 | { |
6e21c8f1 | 1247 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
1da177e4 | 1248 | |
cf2a473c | 1249 | cpuset_update_task_memory_state(); |
1da177e4 LT |
1250 | |
1251 | if (unlikely(pol->policy == MPOL_INTERLEAVE)) { | |
1252 | unsigned nid; | |
5da7ca86 CL |
1253 | |
1254 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT); | |
1da177e4 LT |
1255 | return alloc_page_interleave(gfp, 0, nid); |
1256 | } | |
1257 | return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol)); | |
1258 | } | |
1259 | ||
1260 | /** | |
1261 | * alloc_pages_current - Allocate pages. | |
1262 | * | |
1263 | * @gfp: | |
1264 | * %GFP_USER user allocation, | |
1265 | * %GFP_KERNEL kernel allocation, | |
1266 | * %GFP_HIGHMEM highmem allocation, | |
1267 | * %GFP_FS don't call back into a file system. | |
1268 | * %GFP_ATOMIC don't sleep. | |
1269 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
1270 | * | |
1271 | * Allocate a page from the kernel page pool. When not in | |
1272 | * interrupt context and apply the current process NUMA policy. | |
1273 | * Returns NULL when no page can be allocated. | |
1274 | * | |
cf2a473c | 1275 | * Don't call cpuset_update_task_memory_state() unless |
1da177e4 LT |
1276 | * 1) it's ok to take cpuset_sem (can WAIT), and |
1277 | * 2) allocating for current task (not interrupt). | |
1278 | */ | |
dd0fc66f | 1279 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
1da177e4 LT |
1280 | { |
1281 | struct mempolicy *pol = current->mempolicy; | |
1282 | ||
1283 | if ((gfp & __GFP_WAIT) && !in_interrupt()) | |
cf2a473c | 1284 | cpuset_update_task_memory_state(); |
1da177e4 LT |
1285 | if (!pol || in_interrupt()) |
1286 | pol = &default_policy; | |
1287 | if (pol->policy == MPOL_INTERLEAVE) | |
1288 | return alloc_page_interleave(gfp, order, interleave_nodes(pol)); | |
1289 | return __alloc_pages(gfp, order, zonelist_policy(gfp, pol)); | |
1290 | } | |
1291 | EXPORT_SYMBOL(alloc_pages_current); | |
1292 | ||
4225399a PJ |
1293 | /* |
1294 | * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it | |
1295 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() | |
1296 | * with the mems_allowed returned by cpuset_mems_allowed(). This | |
1297 | * keeps mempolicies cpuset relative after its cpuset moves. See | |
1298 | * further kernel/cpuset.c update_nodemask(). | |
1299 | */ | |
1300 | void *cpuset_being_rebound; | |
1301 | ||
1da177e4 LT |
1302 | /* Slow path of a mempolicy copy */ |
1303 | struct mempolicy *__mpol_copy(struct mempolicy *old) | |
1304 | { | |
1305 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
1306 | ||
1307 | if (!new) | |
1308 | return ERR_PTR(-ENOMEM); | |
4225399a PJ |
1309 | if (current_cpuset_is_being_rebound()) { |
1310 | nodemask_t mems = cpuset_mems_allowed(current); | |
1311 | mpol_rebind_policy(old, &mems); | |
1312 | } | |
1da177e4 LT |
1313 | *new = *old; |
1314 | atomic_set(&new->refcnt, 1); | |
1315 | if (new->policy == MPOL_BIND) { | |
1316 | int sz = ksize(old->v.zonelist); | |
1317 | new->v.zonelist = kmalloc(sz, SLAB_KERNEL); | |
1318 | if (!new->v.zonelist) { | |
1319 | kmem_cache_free(policy_cache, new); | |
1320 | return ERR_PTR(-ENOMEM); | |
1321 | } | |
1322 | memcpy(new->v.zonelist, old->v.zonelist, sz); | |
1323 | } | |
1324 | return new; | |
1325 | } | |
1326 | ||
1327 | /* Slow path of a mempolicy comparison */ | |
1328 | int __mpol_equal(struct mempolicy *a, struct mempolicy *b) | |
1329 | { | |
1330 | if (!a || !b) | |
1331 | return 0; | |
1332 | if (a->policy != b->policy) | |
1333 | return 0; | |
1334 | switch (a->policy) { | |
1335 | case MPOL_DEFAULT: | |
1336 | return 1; | |
1337 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 1338 | return nodes_equal(a->v.nodes, b->v.nodes); |
1da177e4 LT |
1339 | case MPOL_PREFERRED: |
1340 | return a->v.preferred_node == b->v.preferred_node; | |
1341 | case MPOL_BIND: { | |
1342 | int i; | |
1343 | for (i = 0; a->v.zonelist->zones[i]; i++) | |
1344 | if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i]) | |
1345 | return 0; | |
1346 | return b->v.zonelist->zones[i] == NULL; | |
1347 | } | |
1348 | default: | |
1349 | BUG(); | |
1350 | return 0; | |
1351 | } | |
1352 | } | |
1353 | ||
1354 | /* Slow path of a mpol destructor. */ | |
1355 | void __mpol_free(struct mempolicy *p) | |
1356 | { | |
1357 | if (!atomic_dec_and_test(&p->refcnt)) | |
1358 | return; | |
1359 | if (p->policy == MPOL_BIND) | |
1360 | kfree(p->v.zonelist); | |
1361 | p->policy = MPOL_DEFAULT; | |
1362 | kmem_cache_free(policy_cache, p); | |
1363 | } | |
1364 | ||
1da177e4 LT |
1365 | /* |
1366 | * Shared memory backing store policy support. | |
1367 | * | |
1368 | * Remember policies even when nobody has shared memory mapped. | |
1369 | * The policies are kept in Red-Black tree linked from the inode. | |
1370 | * They are protected by the sp->lock spinlock, which should be held | |
1371 | * for any accesses to the tree. | |
1372 | */ | |
1373 | ||
1374 | /* lookup first element intersecting start-end */ | |
1375 | /* Caller holds sp->lock */ | |
1376 | static struct sp_node * | |
1377 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
1378 | { | |
1379 | struct rb_node *n = sp->root.rb_node; | |
1380 | ||
1381 | while (n) { | |
1382 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
1383 | ||
1384 | if (start >= p->end) | |
1385 | n = n->rb_right; | |
1386 | else if (end <= p->start) | |
1387 | n = n->rb_left; | |
1388 | else | |
1389 | break; | |
1390 | } | |
1391 | if (!n) | |
1392 | return NULL; | |
1393 | for (;;) { | |
1394 | struct sp_node *w = NULL; | |
1395 | struct rb_node *prev = rb_prev(n); | |
1396 | if (!prev) | |
1397 | break; | |
1398 | w = rb_entry(prev, struct sp_node, nd); | |
1399 | if (w->end <= start) | |
1400 | break; | |
1401 | n = prev; | |
1402 | } | |
1403 | return rb_entry(n, struct sp_node, nd); | |
1404 | } | |
1405 | ||
1406 | /* Insert a new shared policy into the list. */ | |
1407 | /* Caller holds sp->lock */ | |
1408 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) | |
1409 | { | |
1410 | struct rb_node **p = &sp->root.rb_node; | |
1411 | struct rb_node *parent = NULL; | |
1412 | struct sp_node *nd; | |
1413 | ||
1414 | while (*p) { | |
1415 | parent = *p; | |
1416 | nd = rb_entry(parent, struct sp_node, nd); | |
1417 | if (new->start < nd->start) | |
1418 | p = &(*p)->rb_left; | |
1419 | else if (new->end > nd->end) | |
1420 | p = &(*p)->rb_right; | |
1421 | else | |
1422 | BUG(); | |
1423 | } | |
1424 | rb_link_node(&new->nd, parent, p); | |
1425 | rb_insert_color(&new->nd, &sp->root); | |
1426 | PDprintk("inserting %lx-%lx: %d\n", new->start, new->end, | |
1427 | new->policy ? new->policy->policy : 0); | |
1428 | } | |
1429 | ||
1430 | /* Find shared policy intersecting idx */ | |
1431 | struct mempolicy * | |
1432 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
1433 | { | |
1434 | struct mempolicy *pol = NULL; | |
1435 | struct sp_node *sn; | |
1436 | ||
1437 | if (!sp->root.rb_node) | |
1438 | return NULL; | |
1439 | spin_lock(&sp->lock); | |
1440 | sn = sp_lookup(sp, idx, idx+1); | |
1441 | if (sn) { | |
1442 | mpol_get(sn->policy); | |
1443 | pol = sn->policy; | |
1444 | } | |
1445 | spin_unlock(&sp->lock); | |
1446 | return pol; | |
1447 | } | |
1448 | ||
1449 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) | |
1450 | { | |
1451 | PDprintk("deleting %lx-l%x\n", n->start, n->end); | |
1452 | rb_erase(&n->nd, &sp->root); | |
1453 | mpol_free(n->policy); | |
1454 | kmem_cache_free(sn_cache, n); | |
1455 | } | |
1456 | ||
1457 | struct sp_node * | |
1458 | sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol) | |
1459 | { | |
1460 | struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
1461 | ||
1462 | if (!n) | |
1463 | return NULL; | |
1464 | n->start = start; | |
1465 | n->end = end; | |
1466 | mpol_get(pol); | |
1467 | n->policy = pol; | |
1468 | return n; | |
1469 | } | |
1470 | ||
1471 | /* Replace a policy range. */ | |
1472 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
1473 | unsigned long end, struct sp_node *new) | |
1474 | { | |
1475 | struct sp_node *n, *new2 = NULL; | |
1476 | ||
1477 | restart: | |
1478 | spin_lock(&sp->lock); | |
1479 | n = sp_lookup(sp, start, end); | |
1480 | /* Take care of old policies in the same range. */ | |
1481 | while (n && n->start < end) { | |
1482 | struct rb_node *next = rb_next(&n->nd); | |
1483 | if (n->start >= start) { | |
1484 | if (n->end <= end) | |
1485 | sp_delete(sp, n); | |
1486 | else | |
1487 | n->start = end; | |
1488 | } else { | |
1489 | /* Old policy spanning whole new range. */ | |
1490 | if (n->end > end) { | |
1491 | if (!new2) { | |
1492 | spin_unlock(&sp->lock); | |
1493 | new2 = sp_alloc(end, n->end, n->policy); | |
1494 | if (!new2) | |
1495 | return -ENOMEM; | |
1496 | goto restart; | |
1497 | } | |
1498 | n->end = start; | |
1499 | sp_insert(sp, new2); | |
1500 | new2 = NULL; | |
1501 | break; | |
1502 | } else | |
1503 | n->end = start; | |
1504 | } | |
1505 | if (!next) | |
1506 | break; | |
1507 | n = rb_entry(next, struct sp_node, nd); | |
1508 | } | |
1509 | if (new) | |
1510 | sp_insert(sp, new); | |
1511 | spin_unlock(&sp->lock); | |
1512 | if (new2) { | |
1513 | mpol_free(new2->policy); | |
1514 | kmem_cache_free(sn_cache, new2); | |
1515 | } | |
1516 | return 0; | |
1517 | } | |
1518 | ||
7339ff83 RH |
1519 | void mpol_shared_policy_init(struct shared_policy *info, int policy, |
1520 | nodemask_t *policy_nodes) | |
1521 | { | |
1522 | info->root = RB_ROOT; | |
1523 | spin_lock_init(&info->lock); | |
1524 | ||
1525 | if (policy != MPOL_DEFAULT) { | |
1526 | struct mempolicy *newpol; | |
1527 | ||
1528 | /* Falls back to MPOL_DEFAULT on any error */ | |
1529 | newpol = mpol_new(policy, policy_nodes); | |
1530 | if (!IS_ERR(newpol)) { | |
1531 | /* Create pseudo-vma that contains just the policy */ | |
1532 | struct vm_area_struct pvma; | |
1533 | ||
1534 | memset(&pvma, 0, sizeof(struct vm_area_struct)); | |
1535 | /* Policy covers entire file */ | |
1536 | pvma.vm_end = TASK_SIZE; | |
1537 | mpol_set_shared_policy(info, &pvma, newpol); | |
1538 | mpol_free(newpol); | |
1539 | } | |
1540 | } | |
1541 | } | |
1542 | ||
1da177e4 LT |
1543 | int mpol_set_shared_policy(struct shared_policy *info, |
1544 | struct vm_area_struct *vma, struct mempolicy *npol) | |
1545 | { | |
1546 | int err; | |
1547 | struct sp_node *new = NULL; | |
1548 | unsigned long sz = vma_pages(vma); | |
1549 | ||
1550 | PDprintk("set_shared_policy %lx sz %lu %d %lx\n", | |
1551 | vma->vm_pgoff, | |
1552 | sz, npol? npol->policy : -1, | |
dfcd3c0d | 1553 | npol ? nodes_addr(npol->v.nodes)[0] : -1); |
1da177e4 LT |
1554 | |
1555 | if (npol) { | |
1556 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
1557 | if (!new) | |
1558 | return -ENOMEM; | |
1559 | } | |
1560 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
1561 | if (err && new) | |
1562 | kmem_cache_free(sn_cache, new); | |
1563 | return err; | |
1564 | } | |
1565 | ||
1566 | /* Free a backing policy store on inode delete. */ | |
1567 | void mpol_free_shared_policy(struct shared_policy *p) | |
1568 | { | |
1569 | struct sp_node *n; | |
1570 | struct rb_node *next; | |
1571 | ||
1572 | if (!p->root.rb_node) | |
1573 | return; | |
1574 | spin_lock(&p->lock); | |
1575 | next = rb_first(&p->root); | |
1576 | while (next) { | |
1577 | n = rb_entry(next, struct sp_node, nd); | |
1578 | next = rb_next(&n->nd); | |
90c5029e | 1579 | rb_erase(&n->nd, &p->root); |
1da177e4 LT |
1580 | mpol_free(n->policy); |
1581 | kmem_cache_free(sn_cache, n); | |
1582 | } | |
1583 | spin_unlock(&p->lock); | |
1da177e4 LT |
1584 | } |
1585 | ||
1586 | /* assumes fs == KERNEL_DS */ | |
1587 | void __init numa_policy_init(void) | |
1588 | { | |
1589 | policy_cache = kmem_cache_create("numa_policy", | |
1590 | sizeof(struct mempolicy), | |
1591 | 0, SLAB_PANIC, NULL, NULL); | |
1592 | ||
1593 | sn_cache = kmem_cache_create("shared_policy_node", | |
1594 | sizeof(struct sp_node), | |
1595 | 0, SLAB_PANIC, NULL, NULL); | |
1596 | ||
1597 | /* Set interleaving policy for system init. This way not all | |
1598 | the data structures allocated at system boot end up in node zero. */ | |
1599 | ||
8bccd85f | 1600 | if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map)) |
1da177e4 LT |
1601 | printk("numa_policy_init: interleaving failed\n"); |
1602 | } | |
1603 | ||
8bccd85f | 1604 | /* Reset policy of current process to default */ |
1da177e4 LT |
1605 | void numa_default_policy(void) |
1606 | { | |
8bccd85f | 1607 | do_set_mempolicy(MPOL_DEFAULT, NULL); |
1da177e4 | 1608 | } |
68860ec1 PJ |
1609 | |
1610 | /* Migrate a policy to a different set of nodes */ | |
74cb2155 | 1611 | void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask) |
68860ec1 | 1612 | { |
74cb2155 | 1613 | nodemask_t *mpolmask; |
68860ec1 PJ |
1614 | nodemask_t tmp; |
1615 | ||
1616 | if (!pol) | |
1617 | return; | |
74cb2155 PJ |
1618 | mpolmask = &pol->cpuset_mems_allowed; |
1619 | if (nodes_equal(*mpolmask, *newmask)) | |
1620 | return; | |
68860ec1 PJ |
1621 | |
1622 | switch (pol->policy) { | |
1623 | case MPOL_DEFAULT: | |
1624 | break; | |
1625 | case MPOL_INTERLEAVE: | |
74cb2155 | 1626 | nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask); |
68860ec1 | 1627 | pol->v.nodes = tmp; |
74cb2155 PJ |
1628 | *mpolmask = *newmask; |
1629 | current->il_next = node_remap(current->il_next, | |
1630 | *mpolmask, *newmask); | |
68860ec1 PJ |
1631 | break; |
1632 | case MPOL_PREFERRED: | |
1633 | pol->v.preferred_node = node_remap(pol->v.preferred_node, | |
74cb2155 PJ |
1634 | *mpolmask, *newmask); |
1635 | *mpolmask = *newmask; | |
68860ec1 PJ |
1636 | break; |
1637 | case MPOL_BIND: { | |
1638 | nodemask_t nodes; | |
1639 | struct zone **z; | |
1640 | struct zonelist *zonelist; | |
1641 | ||
1642 | nodes_clear(nodes); | |
1643 | for (z = pol->v.zonelist->zones; *z; z++) | |
1644 | node_set((*z)->zone_pgdat->node_id, nodes); | |
74cb2155 | 1645 | nodes_remap(tmp, nodes, *mpolmask, *newmask); |
68860ec1 PJ |
1646 | nodes = tmp; |
1647 | ||
1648 | zonelist = bind_zonelist(&nodes); | |
1649 | ||
1650 | /* If no mem, then zonelist is NULL and we keep old zonelist. | |
1651 | * If that old zonelist has no remaining mems_allowed nodes, | |
1652 | * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT. | |
1653 | */ | |
1654 | ||
1655 | if (zonelist) { | |
1656 | /* Good - got mem - substitute new zonelist */ | |
1657 | kfree(pol->v.zonelist); | |
1658 | pol->v.zonelist = zonelist; | |
1659 | } | |
74cb2155 | 1660 | *mpolmask = *newmask; |
68860ec1 PJ |
1661 | break; |
1662 | } | |
1663 | default: | |
1664 | BUG(); | |
1665 | break; | |
1666 | } | |
1667 | } | |
1668 | ||
1669 | /* | |
74cb2155 PJ |
1670 | * Wrapper for mpol_rebind_policy() that just requires task |
1671 | * pointer, and updates task mempolicy. | |
68860ec1 | 1672 | */ |
74cb2155 PJ |
1673 | |
1674 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) | |
68860ec1 | 1675 | { |
74cb2155 | 1676 | mpol_rebind_policy(tsk->mempolicy, new); |
68860ec1 | 1677 | } |
1a75a6c8 | 1678 | |
4225399a PJ |
1679 | /* |
1680 | * Rebind each vma in mm to new nodemask. | |
1681 | * | |
1682 | * Call holding a reference to mm. Takes mm->mmap_sem during call. | |
1683 | */ | |
1684 | ||
1685 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) | |
1686 | { | |
1687 | struct vm_area_struct *vma; | |
1688 | ||
1689 | down_write(&mm->mmap_sem); | |
1690 | for (vma = mm->mmap; vma; vma = vma->vm_next) | |
1691 | mpol_rebind_policy(vma->vm_policy, new); | |
1692 | up_write(&mm->mmap_sem); | |
1693 | } | |
1694 | ||
1a75a6c8 CL |
1695 | /* |
1696 | * Display pages allocated per node and memory policy via /proc. | |
1697 | */ | |
1698 | ||
1699 | static const char *policy_types[] = { "default", "prefer", "bind", | |
1700 | "interleave" }; | |
1701 | ||
1702 | /* | |
1703 | * Convert a mempolicy into a string. | |
1704 | * Returns the number of characters in buffer (if positive) | |
1705 | * or an error (negative) | |
1706 | */ | |
1707 | static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) | |
1708 | { | |
1709 | char *p = buffer; | |
1710 | int l; | |
1711 | nodemask_t nodes; | |
1712 | int mode = pol ? pol->policy : MPOL_DEFAULT; | |
1713 | ||
1714 | switch (mode) { | |
1715 | case MPOL_DEFAULT: | |
1716 | nodes_clear(nodes); | |
1717 | break; | |
1718 | ||
1719 | case MPOL_PREFERRED: | |
1720 | nodes_clear(nodes); | |
1721 | node_set(pol->v.preferred_node, nodes); | |
1722 | break; | |
1723 | ||
1724 | case MPOL_BIND: | |
1725 | get_zonemask(pol, &nodes); | |
1726 | break; | |
1727 | ||
1728 | case MPOL_INTERLEAVE: | |
1729 | nodes = pol->v.nodes; | |
1730 | break; | |
1731 | ||
1732 | default: | |
1733 | BUG(); | |
1734 | return -EFAULT; | |
1735 | } | |
1736 | ||
1737 | l = strlen(policy_types[mode]); | |
1738 | if (buffer + maxlen < p + l + 1) | |
1739 | return -ENOSPC; | |
1740 | ||
1741 | strcpy(p, policy_types[mode]); | |
1742 | p += l; | |
1743 | ||
1744 | if (!nodes_empty(nodes)) { | |
1745 | if (buffer + maxlen < p + 2) | |
1746 | return -ENOSPC; | |
1747 | *p++ = '='; | |
1748 | p += nodelist_scnprintf(p, buffer + maxlen - p, nodes); | |
1749 | } | |
1750 | return p - buffer; | |
1751 | } | |
1752 | ||
1753 | struct numa_maps { | |
1754 | unsigned long pages; | |
1755 | unsigned long anon; | |
397874df CL |
1756 | unsigned long active; |
1757 | unsigned long writeback; | |
1a75a6c8 | 1758 | unsigned long mapcount_max; |
397874df CL |
1759 | unsigned long dirty; |
1760 | unsigned long swapcache; | |
1a75a6c8 CL |
1761 | unsigned long node[MAX_NUMNODES]; |
1762 | }; | |
1763 | ||
397874df | 1764 | static void gather_stats(struct page *page, void *private, int pte_dirty) |
1a75a6c8 CL |
1765 | { |
1766 | struct numa_maps *md = private; | |
1767 | int count = page_mapcount(page); | |
1768 | ||
397874df CL |
1769 | md->pages++; |
1770 | if (pte_dirty || PageDirty(page)) | |
1771 | md->dirty++; | |
1a75a6c8 | 1772 | |
397874df CL |
1773 | if (PageSwapCache(page)) |
1774 | md->swapcache++; | |
1a75a6c8 | 1775 | |
397874df CL |
1776 | if (PageActive(page)) |
1777 | md->active++; | |
1778 | ||
1779 | if (PageWriteback(page)) | |
1780 | md->writeback++; | |
1a75a6c8 CL |
1781 | |
1782 | if (PageAnon(page)) | |
1783 | md->anon++; | |
1784 | ||
397874df CL |
1785 | if (count > md->mapcount_max) |
1786 | md->mapcount_max = count; | |
1787 | ||
1a75a6c8 CL |
1788 | md->node[page_to_nid(page)]++; |
1789 | cond_resched(); | |
1790 | } | |
1791 | ||
7f709ed0 | 1792 | #ifdef CONFIG_HUGETLB_PAGE |
397874df CL |
1793 | static void check_huge_range(struct vm_area_struct *vma, |
1794 | unsigned long start, unsigned long end, | |
1795 | struct numa_maps *md) | |
1796 | { | |
1797 | unsigned long addr; | |
1798 | struct page *page; | |
1799 | ||
1800 | for (addr = start; addr < end; addr += HPAGE_SIZE) { | |
1801 | pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK); | |
1802 | pte_t pte; | |
1803 | ||
1804 | if (!ptep) | |
1805 | continue; | |
1806 | ||
1807 | pte = *ptep; | |
1808 | if (pte_none(pte)) | |
1809 | continue; | |
1810 | ||
1811 | page = pte_page(pte); | |
1812 | if (!page) | |
1813 | continue; | |
1814 | ||
1815 | gather_stats(page, md, pte_dirty(*ptep)); | |
1816 | } | |
1817 | } | |
7f709ed0 AM |
1818 | #else |
1819 | static inline void check_huge_range(struct vm_area_struct *vma, | |
1820 | unsigned long start, unsigned long end, | |
1821 | struct numa_maps *md) | |
1822 | { | |
1823 | } | |
1824 | #endif | |
397874df | 1825 | |
1a75a6c8 CL |
1826 | int show_numa_map(struct seq_file *m, void *v) |
1827 | { | |
1828 | struct task_struct *task = m->private; | |
1829 | struct vm_area_struct *vma = v; | |
1830 | struct numa_maps *md; | |
397874df CL |
1831 | struct file *file = vma->vm_file; |
1832 | struct mm_struct *mm = vma->vm_mm; | |
1a75a6c8 CL |
1833 | int n; |
1834 | char buffer[50]; | |
1835 | ||
397874df | 1836 | if (!mm) |
1a75a6c8 CL |
1837 | return 0; |
1838 | ||
1839 | md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL); | |
1840 | if (!md) | |
1841 | return 0; | |
1842 | ||
397874df CL |
1843 | mpol_to_str(buffer, sizeof(buffer), |
1844 | get_vma_policy(task, vma, vma->vm_start)); | |
1845 | ||
1846 | seq_printf(m, "%08lx %s", vma->vm_start, buffer); | |
1847 | ||
1848 | if (file) { | |
1849 | seq_printf(m, " file="); | |
1850 | seq_path(m, file->f_vfsmnt, file->f_dentry, "\n\t= "); | |
1851 | } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) { | |
1852 | seq_printf(m, " heap"); | |
1853 | } else if (vma->vm_start <= mm->start_stack && | |
1854 | vma->vm_end >= mm->start_stack) { | |
1855 | seq_printf(m, " stack"); | |
1856 | } | |
1857 | ||
1858 | if (is_vm_hugetlb_page(vma)) { | |
1859 | check_huge_range(vma, vma->vm_start, vma->vm_end, md); | |
1860 | seq_printf(m, " huge"); | |
1861 | } else { | |
a57ebfdb | 1862 | check_pgd_range(vma, vma->vm_start, vma->vm_end, |
397874df CL |
1863 | &node_online_map, MPOL_MF_STATS, md); |
1864 | } | |
1865 | ||
1866 | if (!md->pages) | |
1867 | goto out; | |
1a75a6c8 | 1868 | |
397874df CL |
1869 | if (md->anon) |
1870 | seq_printf(m," anon=%lu",md->anon); | |
1a75a6c8 | 1871 | |
397874df CL |
1872 | if (md->dirty) |
1873 | seq_printf(m," dirty=%lu",md->dirty); | |
1a75a6c8 | 1874 | |
397874df CL |
1875 | if (md->pages != md->anon && md->pages != md->dirty) |
1876 | seq_printf(m, " mapped=%lu", md->pages); | |
1a75a6c8 | 1877 | |
397874df CL |
1878 | if (md->mapcount_max > 1) |
1879 | seq_printf(m, " mapmax=%lu", md->mapcount_max); | |
1a75a6c8 | 1880 | |
397874df CL |
1881 | if (md->swapcache) |
1882 | seq_printf(m," swapcache=%lu", md->swapcache); | |
1883 | ||
1884 | if (md->active < md->pages && !is_vm_hugetlb_page(vma)) | |
1885 | seq_printf(m," active=%lu", md->active); | |
1886 | ||
1887 | if (md->writeback) | |
1888 | seq_printf(m," writeback=%lu", md->writeback); | |
1889 | ||
1890 | for_each_online_node(n) | |
1891 | if (md->node[n]) | |
1892 | seq_printf(m, " N%d=%lu", n, md->node[n]); | |
1893 | out: | |
1894 | seq_putc(m, '\n'); | |
1a75a6c8 CL |
1895 | kfree(md); |
1896 | ||
1897 | if (m->count < m->size) | |
1898 | m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0; | |
1899 | return 0; | |
1900 | } | |
1901 |