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> | |
b20a3503 | 89 | #include <linux/migrate.h> |
95a402c3 | 90 | #include <linux/rmap.h> |
86c3a764 | 91 | #include <linux/security.h> |
dc9aa5b9 | 92 | |
1da177e4 LT |
93 | #include <asm/tlbflush.h> |
94 | #include <asm/uaccess.h> | |
95 | ||
38e35860 | 96 | /* Internal flags */ |
dc9aa5b9 | 97 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ |
38e35860 | 98 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ |
1a75a6c8 | 99 | #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */ |
dc9aa5b9 | 100 | |
fcc234f8 PE |
101 | static struct kmem_cache *policy_cache; |
102 | static struct kmem_cache *sn_cache; | |
1da177e4 LT |
103 | |
104 | #define PDprintk(fmt...) | |
105 | ||
106 | /* Highest zone. An specific allocation for a zone below that is not | |
107 | policied. */ | |
2f6726e5 | 108 | enum zone_type 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; | |
2f6726e5 CL |
140 | int num, max, nd; |
141 | enum zone_type k; | |
1da177e4 | 142 | |
dfcd3c0d | 143 | max = 1 + MAX_NR_ZONES * nodes_weight(*nodes); |
dd942ae3 | 144 | zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL); |
1da177e4 LT |
145 | if (!zl) |
146 | return NULL; | |
147 | num = 0; | |
dd942ae3 AK |
148 | /* First put in the highest zones from all nodes, then all the next |
149 | lower zones etc. Avoid empty zones because the memory allocator | |
150 | doesn't like them. If you implement node hot removal you | |
151 | have to fix that. */ | |
2f6726e5 CL |
152 | k = policy_zone; |
153 | while (1) { | |
dd942ae3 AK |
154 | for_each_node_mask(nd, *nodes) { |
155 | struct zone *z = &NODE_DATA(nd)->node_zones[k]; | |
156 | if (z->present_pages > 0) | |
157 | zl->zones[num++] = z; | |
158 | } | |
2f6726e5 CL |
159 | if (k == 0) |
160 | break; | |
161 | k--; | |
dd942ae3 | 162 | } |
1da177e4 LT |
163 | zl->zones[num] = NULL; |
164 | return zl; | |
165 | } | |
166 | ||
167 | /* Create a new policy */ | |
dfcd3c0d | 168 | static struct mempolicy *mpol_new(int mode, nodemask_t *nodes) |
1da177e4 LT |
169 | { |
170 | struct mempolicy *policy; | |
171 | ||
dfcd3c0d | 172 | PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]); |
1da177e4 LT |
173 | if (mode == MPOL_DEFAULT) |
174 | return NULL; | |
175 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
176 | if (!policy) | |
177 | return ERR_PTR(-ENOMEM); | |
178 | atomic_set(&policy->refcnt, 1); | |
179 | switch (mode) { | |
180 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 181 | policy->v.nodes = *nodes; |
8f493d79 AK |
182 | if (nodes_weight(*nodes) == 0) { |
183 | kmem_cache_free(policy_cache, policy); | |
184 | return ERR_PTR(-EINVAL); | |
185 | } | |
1da177e4 LT |
186 | break; |
187 | case MPOL_PREFERRED: | |
dfcd3c0d | 188 | policy->v.preferred_node = first_node(*nodes); |
1da177e4 LT |
189 | if (policy->v.preferred_node >= MAX_NUMNODES) |
190 | policy->v.preferred_node = -1; | |
191 | break; | |
192 | case MPOL_BIND: | |
193 | policy->v.zonelist = bind_zonelist(nodes); | |
194 | if (policy->v.zonelist == NULL) { | |
195 | kmem_cache_free(policy_cache, policy); | |
196 | return ERR_PTR(-ENOMEM); | |
197 | } | |
198 | break; | |
199 | } | |
200 | policy->policy = mode; | |
74cb2155 | 201 | policy->cpuset_mems_allowed = cpuset_mems_allowed(current); |
1da177e4 LT |
202 | return policy; |
203 | } | |
204 | ||
397874df | 205 | static void gather_stats(struct page *, void *, int pte_dirty); |
fc301289 CL |
206 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
207 | unsigned long flags); | |
1a75a6c8 | 208 | |
38e35860 | 209 | /* Scan through pages checking if pages follow certain conditions. */ |
b5810039 | 210 | static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd, |
dc9aa5b9 CL |
211 | unsigned long addr, unsigned long end, |
212 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 213 | void *private) |
1da177e4 | 214 | { |
91612e0d HD |
215 | pte_t *orig_pte; |
216 | pte_t *pte; | |
705e87c0 | 217 | spinlock_t *ptl; |
941150a3 | 218 | |
705e87c0 | 219 | orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
91612e0d | 220 | do { |
6aab341e | 221 | struct page *page; |
91612e0d HD |
222 | unsigned int nid; |
223 | ||
224 | if (!pte_present(*pte)) | |
1da177e4 | 225 | continue; |
6aab341e LT |
226 | page = vm_normal_page(vma, addr, *pte); |
227 | if (!page) | |
1da177e4 | 228 | continue; |
053837fc NP |
229 | /* |
230 | * The check for PageReserved here is important to avoid | |
231 | * handling zero pages and other pages that may have been | |
232 | * marked special by the system. | |
233 | * | |
234 | * If the PageReserved would not be checked here then f.e. | |
235 | * the location of the zero page could have an influence | |
236 | * on MPOL_MF_STRICT, zero pages would be counted for | |
237 | * the per node stats, and there would be useless attempts | |
238 | * to put zero pages on the migration list. | |
239 | */ | |
f4598c8b CL |
240 | if (PageReserved(page)) |
241 | continue; | |
6aab341e | 242 | nid = page_to_nid(page); |
38e35860 CL |
243 | if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT)) |
244 | continue; | |
245 | ||
1a75a6c8 | 246 | if (flags & MPOL_MF_STATS) |
397874df | 247 | gather_stats(page, private, pte_dirty(*pte)); |
053837fc | 248 | else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) |
fc301289 | 249 | migrate_page_add(page, private, flags); |
38e35860 CL |
250 | else |
251 | break; | |
91612e0d | 252 | } while (pte++, addr += PAGE_SIZE, addr != end); |
705e87c0 | 253 | pte_unmap_unlock(orig_pte, ptl); |
91612e0d HD |
254 | return addr != end; |
255 | } | |
256 | ||
b5810039 | 257 | static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud, |
dc9aa5b9 CL |
258 | unsigned long addr, unsigned long end, |
259 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 260 | void *private) |
91612e0d HD |
261 | { |
262 | pmd_t *pmd; | |
263 | unsigned long next; | |
264 | ||
265 | pmd = pmd_offset(pud, addr); | |
266 | do { | |
267 | next = pmd_addr_end(addr, end); | |
268 | if (pmd_none_or_clear_bad(pmd)) | |
269 | continue; | |
dc9aa5b9 | 270 | if (check_pte_range(vma, pmd, addr, next, nodes, |
38e35860 | 271 | flags, private)) |
91612e0d HD |
272 | return -EIO; |
273 | } while (pmd++, addr = next, addr != end); | |
274 | return 0; | |
275 | } | |
276 | ||
b5810039 | 277 | static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd, |
dc9aa5b9 CL |
278 | unsigned long addr, unsigned long end, |
279 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 280 | void *private) |
91612e0d HD |
281 | { |
282 | pud_t *pud; | |
283 | unsigned long next; | |
284 | ||
285 | pud = pud_offset(pgd, addr); | |
286 | do { | |
287 | next = pud_addr_end(addr, end); | |
288 | if (pud_none_or_clear_bad(pud)) | |
289 | continue; | |
dc9aa5b9 | 290 | if (check_pmd_range(vma, pud, addr, next, nodes, |
38e35860 | 291 | flags, private)) |
91612e0d HD |
292 | return -EIO; |
293 | } while (pud++, addr = next, addr != end); | |
294 | return 0; | |
295 | } | |
296 | ||
b5810039 | 297 | static inline int check_pgd_range(struct vm_area_struct *vma, |
dc9aa5b9 CL |
298 | unsigned long addr, unsigned long end, |
299 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 300 | void *private) |
91612e0d HD |
301 | { |
302 | pgd_t *pgd; | |
303 | unsigned long next; | |
304 | ||
b5810039 | 305 | pgd = pgd_offset(vma->vm_mm, addr); |
91612e0d HD |
306 | do { |
307 | next = pgd_addr_end(addr, end); | |
308 | if (pgd_none_or_clear_bad(pgd)) | |
309 | continue; | |
dc9aa5b9 | 310 | if (check_pud_range(vma, pgd, addr, next, nodes, |
38e35860 | 311 | flags, private)) |
91612e0d HD |
312 | return -EIO; |
313 | } while (pgd++, addr = next, addr != end); | |
314 | return 0; | |
1da177e4 LT |
315 | } |
316 | ||
dc9aa5b9 CL |
317 | /* Check if a vma is migratable */ |
318 | static inline int vma_migratable(struct vm_area_struct *vma) | |
319 | { | |
320 | if (vma->vm_flags & ( | |
f4598c8b | 321 | VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED)) |
dc9aa5b9 CL |
322 | return 0; |
323 | return 1; | |
324 | } | |
325 | ||
326 | /* | |
327 | * Check if all pages in a range are on a set of nodes. | |
328 | * If pagelist != NULL then isolate pages from the LRU and | |
329 | * put them on the pagelist. | |
330 | */ | |
1da177e4 LT |
331 | static struct vm_area_struct * |
332 | check_range(struct mm_struct *mm, unsigned long start, unsigned long end, | |
38e35860 | 333 | const nodemask_t *nodes, unsigned long flags, void *private) |
1da177e4 LT |
334 | { |
335 | int err; | |
336 | struct vm_area_struct *first, *vma, *prev; | |
337 | ||
90036ee5 | 338 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
90036ee5 | 339 | |
b20a3503 CL |
340 | err = migrate_prep(); |
341 | if (err) | |
342 | return ERR_PTR(err); | |
90036ee5 | 343 | } |
053837fc | 344 | |
1da177e4 LT |
345 | first = find_vma(mm, start); |
346 | if (!first) | |
347 | return ERR_PTR(-EFAULT); | |
348 | prev = NULL; | |
349 | for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) { | |
dc9aa5b9 CL |
350 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { |
351 | if (!vma->vm_next && vma->vm_end < end) | |
352 | return ERR_PTR(-EFAULT); | |
353 | if (prev && prev->vm_end < vma->vm_start) | |
354 | return ERR_PTR(-EFAULT); | |
355 | } | |
356 | if (!is_vm_hugetlb_page(vma) && | |
357 | ((flags & MPOL_MF_STRICT) || | |
358 | ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) && | |
359 | vma_migratable(vma)))) { | |
5b952b3c | 360 | unsigned long endvma = vma->vm_end; |
dc9aa5b9 | 361 | |
5b952b3c AK |
362 | if (endvma > end) |
363 | endvma = end; | |
364 | if (vma->vm_start > start) | |
365 | start = vma->vm_start; | |
dc9aa5b9 | 366 | err = check_pgd_range(vma, start, endvma, nodes, |
38e35860 | 367 | flags, private); |
1da177e4 LT |
368 | if (err) { |
369 | first = ERR_PTR(err); | |
370 | break; | |
371 | } | |
372 | } | |
373 | prev = vma; | |
374 | } | |
375 | return first; | |
376 | } | |
377 | ||
378 | /* Apply policy to a single VMA */ | |
379 | static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new) | |
380 | { | |
381 | int err = 0; | |
382 | struct mempolicy *old = vma->vm_policy; | |
383 | ||
384 | PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", | |
385 | vma->vm_start, vma->vm_end, vma->vm_pgoff, | |
386 | vma->vm_ops, vma->vm_file, | |
387 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
388 | ||
389 | if (vma->vm_ops && vma->vm_ops->set_policy) | |
390 | err = vma->vm_ops->set_policy(vma, new); | |
391 | if (!err) { | |
392 | mpol_get(new); | |
393 | vma->vm_policy = new; | |
394 | mpol_free(old); | |
395 | } | |
396 | return err; | |
397 | } | |
398 | ||
399 | /* Step 2: apply policy to a range and do splits. */ | |
400 | static int mbind_range(struct vm_area_struct *vma, unsigned long start, | |
401 | unsigned long end, struct mempolicy *new) | |
402 | { | |
403 | struct vm_area_struct *next; | |
404 | int err; | |
405 | ||
406 | err = 0; | |
407 | for (; vma && vma->vm_start < end; vma = next) { | |
408 | next = vma->vm_next; | |
409 | if (vma->vm_start < start) | |
410 | err = split_vma(vma->vm_mm, vma, start, 1); | |
411 | if (!err && vma->vm_end > end) | |
412 | err = split_vma(vma->vm_mm, vma, end, 0); | |
413 | if (!err) | |
414 | err = policy_vma(vma, new); | |
415 | if (err) | |
416 | break; | |
417 | } | |
418 | return err; | |
419 | } | |
420 | ||
8bccd85f CL |
421 | static int contextualize_policy(int mode, nodemask_t *nodes) |
422 | { | |
423 | if (!nodes) | |
424 | return 0; | |
425 | ||
cf2a473c | 426 | cpuset_update_task_memory_state(); |
5966514d PJ |
427 | if (!cpuset_nodes_subset_current_mems_allowed(*nodes)) |
428 | return -EINVAL; | |
8bccd85f CL |
429 | return mpol_check_policy(mode, nodes); |
430 | } | |
431 | ||
c61afb18 PJ |
432 | |
433 | /* | |
434 | * Update task->flags PF_MEMPOLICY bit: set iff non-default | |
435 | * mempolicy. Allows more rapid checking of this (combined perhaps | |
436 | * with other PF_* flag bits) on memory allocation hot code paths. | |
437 | * | |
438 | * If called from outside this file, the task 'p' should -only- be | |
439 | * a newly forked child not yet visible on the task list, because | |
440 | * manipulating the task flags of a visible task is not safe. | |
441 | * | |
442 | * The above limitation is why this routine has the funny name | |
443 | * mpol_fix_fork_child_flag(). | |
444 | * | |
445 | * It is also safe to call this with a task pointer of current, | |
446 | * which the static wrapper mpol_set_task_struct_flag() does, | |
447 | * for use within this file. | |
448 | */ | |
449 | ||
450 | void mpol_fix_fork_child_flag(struct task_struct *p) | |
451 | { | |
452 | if (p->mempolicy) | |
453 | p->flags |= PF_MEMPOLICY; | |
454 | else | |
455 | p->flags &= ~PF_MEMPOLICY; | |
456 | } | |
457 | ||
458 | static void mpol_set_task_struct_flag(void) | |
459 | { | |
460 | mpol_fix_fork_child_flag(current); | |
461 | } | |
462 | ||
1da177e4 | 463 | /* Set the process memory policy */ |
8bccd85f | 464 | long do_set_mempolicy(int mode, nodemask_t *nodes) |
1da177e4 | 465 | { |
1da177e4 | 466 | struct mempolicy *new; |
1da177e4 | 467 | |
8bccd85f | 468 | if (contextualize_policy(mode, nodes)) |
1da177e4 | 469 | return -EINVAL; |
8bccd85f | 470 | new = mpol_new(mode, nodes); |
1da177e4 LT |
471 | if (IS_ERR(new)) |
472 | return PTR_ERR(new); | |
473 | mpol_free(current->mempolicy); | |
474 | current->mempolicy = new; | |
c61afb18 | 475 | mpol_set_task_struct_flag(); |
1da177e4 | 476 | if (new && new->policy == MPOL_INTERLEAVE) |
dfcd3c0d | 477 | current->il_next = first_node(new->v.nodes); |
1da177e4 LT |
478 | return 0; |
479 | } | |
480 | ||
481 | /* Fill a zone bitmap for a policy */ | |
dfcd3c0d | 482 | static void get_zonemask(struct mempolicy *p, nodemask_t *nodes) |
1da177e4 LT |
483 | { |
484 | int i; | |
485 | ||
dfcd3c0d | 486 | nodes_clear(*nodes); |
1da177e4 LT |
487 | switch (p->policy) { |
488 | case MPOL_BIND: | |
489 | for (i = 0; p->v.zonelist->zones[i]; i++) | |
89fa3024 | 490 | node_set(zone_to_nid(p->v.zonelist->zones[i]), |
8bccd85f | 491 | *nodes); |
1da177e4 LT |
492 | break; |
493 | case MPOL_DEFAULT: | |
494 | break; | |
495 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 496 | *nodes = p->v.nodes; |
1da177e4 LT |
497 | break; |
498 | case MPOL_PREFERRED: | |
499 | /* or use current node instead of online map? */ | |
500 | if (p->v.preferred_node < 0) | |
dfcd3c0d | 501 | *nodes = node_online_map; |
1da177e4 | 502 | else |
dfcd3c0d | 503 | node_set(p->v.preferred_node, *nodes); |
1da177e4 LT |
504 | break; |
505 | default: | |
506 | BUG(); | |
507 | } | |
508 | } | |
509 | ||
510 | static int lookup_node(struct mm_struct *mm, unsigned long addr) | |
511 | { | |
512 | struct page *p; | |
513 | int err; | |
514 | ||
515 | err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL); | |
516 | if (err >= 0) { | |
517 | err = page_to_nid(p); | |
518 | put_page(p); | |
519 | } | |
520 | return err; | |
521 | } | |
522 | ||
1da177e4 | 523 | /* Retrieve NUMA policy */ |
8bccd85f CL |
524 | long do_get_mempolicy(int *policy, nodemask_t *nmask, |
525 | unsigned long addr, unsigned long flags) | |
1da177e4 | 526 | { |
8bccd85f | 527 | int err; |
1da177e4 LT |
528 | struct mm_struct *mm = current->mm; |
529 | struct vm_area_struct *vma = NULL; | |
530 | struct mempolicy *pol = current->mempolicy; | |
531 | ||
cf2a473c | 532 | cpuset_update_task_memory_state(); |
1da177e4 LT |
533 | if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR)) |
534 | return -EINVAL; | |
1da177e4 LT |
535 | if (flags & MPOL_F_ADDR) { |
536 | down_read(&mm->mmap_sem); | |
537 | vma = find_vma_intersection(mm, addr, addr+1); | |
538 | if (!vma) { | |
539 | up_read(&mm->mmap_sem); | |
540 | return -EFAULT; | |
541 | } | |
542 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
543 | pol = vma->vm_ops->get_policy(vma, addr); | |
544 | else | |
545 | pol = vma->vm_policy; | |
546 | } else if (addr) | |
547 | return -EINVAL; | |
548 | ||
549 | if (!pol) | |
550 | pol = &default_policy; | |
551 | ||
552 | if (flags & MPOL_F_NODE) { | |
553 | if (flags & MPOL_F_ADDR) { | |
554 | err = lookup_node(mm, addr); | |
555 | if (err < 0) | |
556 | goto out; | |
8bccd85f | 557 | *policy = err; |
1da177e4 LT |
558 | } else if (pol == current->mempolicy && |
559 | pol->policy == MPOL_INTERLEAVE) { | |
8bccd85f | 560 | *policy = current->il_next; |
1da177e4 LT |
561 | } else { |
562 | err = -EINVAL; | |
563 | goto out; | |
564 | } | |
565 | } else | |
8bccd85f | 566 | *policy = pol->policy; |
1da177e4 LT |
567 | |
568 | if (vma) { | |
569 | up_read(¤t->mm->mmap_sem); | |
570 | vma = NULL; | |
571 | } | |
572 | ||
1da177e4 | 573 | err = 0; |
8bccd85f CL |
574 | if (nmask) |
575 | get_zonemask(pol, nmask); | |
1da177e4 LT |
576 | |
577 | out: | |
578 | if (vma) | |
579 | up_read(¤t->mm->mmap_sem); | |
580 | return err; | |
581 | } | |
582 | ||
b20a3503 | 583 | #ifdef CONFIG_MIGRATION |
6ce3c4c0 CL |
584 | /* |
585 | * page migration | |
586 | */ | |
fc301289 CL |
587 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
588 | unsigned long flags) | |
6ce3c4c0 CL |
589 | { |
590 | /* | |
fc301289 | 591 | * Avoid migrating a page that is shared with others. |
6ce3c4c0 | 592 | */ |
b20a3503 CL |
593 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) |
594 | isolate_lru_page(page, pagelist); | |
7e2ab150 | 595 | } |
6ce3c4c0 | 596 | |
742755a1 | 597 | static struct page *new_node_page(struct page *page, unsigned long node, int **x) |
95a402c3 CL |
598 | { |
599 | return alloc_pages_node(node, GFP_HIGHUSER, 0); | |
600 | } | |
601 | ||
7e2ab150 CL |
602 | /* |
603 | * Migrate pages from one node to a target node. | |
604 | * Returns error or the number of pages not migrated. | |
605 | */ | |
606 | int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags) | |
607 | { | |
608 | nodemask_t nmask; | |
609 | LIST_HEAD(pagelist); | |
610 | int err = 0; | |
611 | ||
612 | nodes_clear(nmask); | |
613 | node_set(source, nmask); | |
6ce3c4c0 | 614 | |
7e2ab150 CL |
615 | check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask, |
616 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); | |
617 | ||
aaa994b3 | 618 | if (!list_empty(&pagelist)) |
95a402c3 CL |
619 | err = migrate_pages(&pagelist, new_node_page, dest); |
620 | ||
7e2ab150 | 621 | return err; |
6ce3c4c0 CL |
622 | } |
623 | ||
39743889 | 624 | /* |
7e2ab150 CL |
625 | * Move pages between the two nodesets so as to preserve the physical |
626 | * layout as much as possible. | |
39743889 CL |
627 | * |
628 | * Returns the number of page that could not be moved. | |
629 | */ | |
630 | int do_migrate_pages(struct mm_struct *mm, | |
631 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
632 | { | |
633 | LIST_HEAD(pagelist); | |
7e2ab150 CL |
634 | int busy = 0; |
635 | int err = 0; | |
636 | nodemask_t tmp; | |
39743889 | 637 | |
7e2ab150 | 638 | down_read(&mm->mmap_sem); |
39743889 | 639 | |
7b2259b3 CL |
640 | err = migrate_vmas(mm, from_nodes, to_nodes, flags); |
641 | if (err) | |
642 | goto out; | |
643 | ||
7e2ab150 CL |
644 | /* |
645 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' | |
646 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' | |
647 | * bit in 'tmp', and return that <source, dest> pair for migration. | |
648 | * The pair of nodemasks 'to' and 'from' define the map. | |
649 | * | |
650 | * If no pair of bits is found that way, fallback to picking some | |
651 | * pair of 'source' and 'dest' bits that are not the same. If the | |
652 | * 'source' and 'dest' bits are the same, this represents a node | |
653 | * that will be migrating to itself, so no pages need move. | |
654 | * | |
655 | * If no bits are left in 'tmp', or if all remaining bits left | |
656 | * in 'tmp' correspond to the same bit in 'to', return false | |
657 | * (nothing left to migrate). | |
658 | * | |
659 | * This lets us pick a pair of nodes to migrate between, such that | |
660 | * if possible the dest node is not already occupied by some other | |
661 | * source node, minimizing the risk of overloading the memory on a | |
662 | * node that would happen if we migrated incoming memory to a node | |
663 | * before migrating outgoing memory source that same node. | |
664 | * | |
665 | * A single scan of tmp is sufficient. As we go, we remember the | |
666 | * most recent <s, d> pair that moved (s != d). If we find a pair | |
667 | * that not only moved, but what's better, moved to an empty slot | |
668 | * (d is not set in tmp), then we break out then, with that pair. | |
669 | * Otherwise when we finish scannng from_tmp, we at least have the | |
670 | * most recent <s, d> pair that moved. If we get all the way through | |
671 | * the scan of tmp without finding any node that moved, much less | |
672 | * moved to an empty node, then there is nothing left worth migrating. | |
673 | */ | |
d4984711 | 674 | |
7e2ab150 CL |
675 | tmp = *from_nodes; |
676 | while (!nodes_empty(tmp)) { | |
677 | int s,d; | |
678 | int source = -1; | |
679 | int dest = 0; | |
680 | ||
681 | for_each_node_mask(s, tmp) { | |
682 | d = node_remap(s, *from_nodes, *to_nodes); | |
683 | if (s == d) | |
684 | continue; | |
685 | ||
686 | source = s; /* Node moved. Memorize */ | |
687 | dest = d; | |
688 | ||
689 | /* dest not in remaining from nodes? */ | |
690 | if (!node_isset(dest, tmp)) | |
691 | break; | |
692 | } | |
693 | if (source == -1) | |
694 | break; | |
695 | ||
696 | node_clear(source, tmp); | |
697 | err = migrate_to_node(mm, source, dest, flags); | |
698 | if (err > 0) | |
699 | busy += err; | |
700 | if (err < 0) | |
701 | break; | |
39743889 | 702 | } |
7b2259b3 | 703 | out: |
39743889 | 704 | up_read(&mm->mmap_sem); |
7e2ab150 CL |
705 | if (err < 0) |
706 | return err; | |
707 | return busy; | |
b20a3503 CL |
708 | |
709 | } | |
710 | ||
742755a1 | 711 | static struct page *new_vma_page(struct page *page, unsigned long private, int **x) |
95a402c3 CL |
712 | { |
713 | struct vm_area_struct *vma = (struct vm_area_struct *)private; | |
714 | ||
715 | return alloc_page_vma(GFP_HIGHUSER, vma, page_address_in_vma(page, vma)); | |
716 | } | |
b20a3503 CL |
717 | #else |
718 | ||
719 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
720 | unsigned long flags) | |
721 | { | |
39743889 CL |
722 | } |
723 | ||
b20a3503 CL |
724 | int do_migrate_pages(struct mm_struct *mm, |
725 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
726 | { | |
727 | return -ENOSYS; | |
728 | } | |
95a402c3 CL |
729 | |
730 | static struct page *new_vma_page(struct page *page, unsigned long private) | |
731 | { | |
732 | return NULL; | |
733 | } | |
b20a3503 CL |
734 | #endif |
735 | ||
6ce3c4c0 CL |
736 | long do_mbind(unsigned long start, unsigned long len, |
737 | unsigned long mode, nodemask_t *nmask, unsigned long flags) | |
738 | { | |
739 | struct vm_area_struct *vma; | |
740 | struct mm_struct *mm = current->mm; | |
741 | struct mempolicy *new; | |
742 | unsigned long end; | |
743 | int err; | |
744 | LIST_HEAD(pagelist); | |
745 | ||
746 | if ((flags & ~(unsigned long)(MPOL_MF_STRICT | | |
747 | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
748 | || mode > MPOL_MAX) | |
749 | return -EINVAL; | |
74c00241 | 750 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
6ce3c4c0 CL |
751 | return -EPERM; |
752 | ||
753 | if (start & ~PAGE_MASK) | |
754 | return -EINVAL; | |
755 | ||
756 | if (mode == MPOL_DEFAULT) | |
757 | flags &= ~MPOL_MF_STRICT; | |
758 | ||
759 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; | |
760 | end = start + len; | |
761 | ||
762 | if (end < start) | |
763 | return -EINVAL; | |
764 | if (end == start) | |
765 | return 0; | |
766 | ||
767 | if (mpol_check_policy(mode, nmask)) | |
768 | return -EINVAL; | |
769 | ||
770 | new = mpol_new(mode, nmask); | |
771 | if (IS_ERR(new)) | |
772 | return PTR_ERR(new); | |
773 | ||
774 | /* | |
775 | * If we are using the default policy then operation | |
776 | * on discontinuous address spaces is okay after all | |
777 | */ | |
778 | if (!new) | |
779 | flags |= MPOL_MF_DISCONTIG_OK; | |
780 | ||
781 | PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len, | |
782 | mode,nodes_addr(nodes)[0]); | |
783 | ||
784 | down_write(&mm->mmap_sem); | |
785 | vma = check_range(mm, start, end, nmask, | |
786 | flags | MPOL_MF_INVERT, &pagelist); | |
787 | ||
788 | err = PTR_ERR(vma); | |
789 | if (!IS_ERR(vma)) { | |
790 | int nr_failed = 0; | |
791 | ||
792 | err = mbind_range(vma, start, end, new); | |
7e2ab150 | 793 | |
6ce3c4c0 | 794 | if (!list_empty(&pagelist)) |
95a402c3 CL |
795 | nr_failed = migrate_pages(&pagelist, new_vma_page, |
796 | (unsigned long)vma); | |
6ce3c4c0 CL |
797 | |
798 | if (!err && nr_failed && (flags & MPOL_MF_STRICT)) | |
799 | err = -EIO; | |
800 | } | |
b20a3503 | 801 | |
6ce3c4c0 CL |
802 | up_write(&mm->mmap_sem); |
803 | mpol_free(new); | |
804 | return err; | |
805 | } | |
806 | ||
8bccd85f CL |
807 | /* |
808 | * User space interface with variable sized bitmaps for nodelists. | |
809 | */ | |
810 | ||
811 | /* Copy a node mask from user space. */ | |
39743889 | 812 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
8bccd85f CL |
813 | unsigned long maxnode) |
814 | { | |
815 | unsigned long k; | |
816 | unsigned long nlongs; | |
817 | unsigned long endmask; | |
818 | ||
819 | --maxnode; | |
820 | nodes_clear(*nodes); | |
821 | if (maxnode == 0 || !nmask) | |
822 | return 0; | |
a9c930ba | 823 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
636f13c1 | 824 | return -EINVAL; |
8bccd85f CL |
825 | |
826 | nlongs = BITS_TO_LONGS(maxnode); | |
827 | if ((maxnode % BITS_PER_LONG) == 0) | |
828 | endmask = ~0UL; | |
829 | else | |
830 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
831 | ||
832 | /* When the user specified more nodes than supported just check | |
833 | if the non supported part is all zero. */ | |
834 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
835 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
836 | return -EINVAL; | |
837 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
838 | unsigned long t; | |
839 | if (get_user(t, nmask + k)) | |
840 | return -EFAULT; | |
841 | if (k == nlongs - 1) { | |
842 | if (t & endmask) | |
843 | return -EINVAL; | |
844 | } else if (t) | |
845 | return -EINVAL; | |
846 | } | |
847 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
848 | endmask = ~0UL; | |
849 | } | |
850 | ||
851 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
852 | return -EFAULT; | |
853 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
854 | return 0; | |
855 | } | |
856 | ||
857 | /* Copy a kernel node mask to user space */ | |
858 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
859 | nodemask_t *nodes) | |
860 | { | |
861 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
862 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
863 | ||
864 | if (copy > nbytes) { | |
865 | if (copy > PAGE_SIZE) | |
866 | return -EINVAL; | |
867 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
868 | return -EFAULT; | |
869 | copy = nbytes; | |
870 | } | |
871 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
872 | } | |
873 | ||
874 | asmlinkage long sys_mbind(unsigned long start, unsigned long len, | |
875 | unsigned long mode, | |
876 | unsigned long __user *nmask, unsigned long maxnode, | |
877 | unsigned flags) | |
878 | { | |
879 | nodemask_t nodes; | |
880 | int err; | |
881 | ||
882 | err = get_nodes(&nodes, nmask, maxnode); | |
883 | if (err) | |
884 | return err; | |
885 | return do_mbind(start, len, mode, &nodes, flags); | |
886 | } | |
887 | ||
888 | /* Set the process memory policy */ | |
889 | asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask, | |
890 | unsigned long maxnode) | |
891 | { | |
892 | int err; | |
893 | nodemask_t nodes; | |
894 | ||
895 | if (mode < 0 || mode > MPOL_MAX) | |
896 | return -EINVAL; | |
897 | err = get_nodes(&nodes, nmask, maxnode); | |
898 | if (err) | |
899 | return err; | |
900 | return do_set_mempolicy(mode, &nodes); | |
901 | } | |
902 | ||
39743889 CL |
903 | asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode, |
904 | const unsigned long __user *old_nodes, | |
905 | const unsigned long __user *new_nodes) | |
906 | { | |
907 | struct mm_struct *mm; | |
908 | struct task_struct *task; | |
909 | nodemask_t old; | |
910 | nodemask_t new; | |
911 | nodemask_t task_nodes; | |
912 | int err; | |
913 | ||
914 | err = get_nodes(&old, old_nodes, maxnode); | |
915 | if (err) | |
916 | return err; | |
917 | ||
918 | err = get_nodes(&new, new_nodes, maxnode); | |
919 | if (err) | |
920 | return err; | |
921 | ||
922 | /* Find the mm_struct */ | |
923 | read_lock(&tasklist_lock); | |
924 | task = pid ? find_task_by_pid(pid) : current; | |
925 | if (!task) { | |
926 | read_unlock(&tasklist_lock); | |
927 | return -ESRCH; | |
928 | } | |
929 | mm = get_task_mm(task); | |
930 | read_unlock(&tasklist_lock); | |
931 | ||
932 | if (!mm) | |
933 | return -EINVAL; | |
934 | ||
935 | /* | |
936 | * Check if this process has the right to modify the specified | |
937 | * process. The right exists if the process has administrative | |
7f927fcc | 938 | * capabilities, superuser privileges or the same |
39743889 CL |
939 | * userid as the target process. |
940 | */ | |
941 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
942 | (current->uid != task->suid) && (current->uid != task->uid) && | |
74c00241 | 943 | !capable(CAP_SYS_NICE)) { |
39743889 CL |
944 | err = -EPERM; |
945 | goto out; | |
946 | } | |
947 | ||
948 | task_nodes = cpuset_mems_allowed(task); | |
949 | /* Is the user allowed to access the target nodes? */ | |
74c00241 | 950 | if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) { |
39743889 CL |
951 | err = -EPERM; |
952 | goto out; | |
953 | } | |
954 | ||
86c3a764 DQ |
955 | err = security_task_movememory(task); |
956 | if (err) | |
957 | goto out; | |
958 | ||
511030bc | 959 | err = do_migrate_pages(mm, &old, &new, |
74c00241 | 960 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); |
39743889 CL |
961 | out: |
962 | mmput(mm); | |
963 | return err; | |
964 | } | |
965 | ||
966 | ||
8bccd85f CL |
967 | /* Retrieve NUMA policy */ |
968 | asmlinkage long sys_get_mempolicy(int __user *policy, | |
969 | unsigned long __user *nmask, | |
970 | unsigned long maxnode, | |
971 | unsigned long addr, unsigned long flags) | |
972 | { | |
973 | int err, pval; | |
974 | nodemask_t nodes; | |
975 | ||
976 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
977 | return -EINVAL; | |
978 | ||
979 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
980 | ||
981 | if (err) | |
982 | return err; | |
983 | ||
984 | if (policy && put_user(pval, policy)) | |
985 | return -EFAULT; | |
986 | ||
987 | if (nmask) | |
988 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
989 | ||
990 | return err; | |
991 | } | |
992 | ||
1da177e4 LT |
993 | #ifdef CONFIG_COMPAT |
994 | ||
995 | asmlinkage long compat_sys_get_mempolicy(int __user *policy, | |
996 | compat_ulong_t __user *nmask, | |
997 | compat_ulong_t maxnode, | |
998 | compat_ulong_t addr, compat_ulong_t flags) | |
999 | { | |
1000 | long err; | |
1001 | unsigned long __user *nm = NULL; | |
1002 | unsigned long nr_bits, alloc_size; | |
1003 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1004 | ||
1005 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1006 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1007 | ||
1008 | if (nmask) | |
1009 | nm = compat_alloc_user_space(alloc_size); | |
1010 | ||
1011 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
1012 | ||
1013 | if (!err && nmask) { | |
1014 | err = copy_from_user(bm, nm, alloc_size); | |
1015 | /* ensure entire bitmap is zeroed */ | |
1016 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
1017 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
1018 | } | |
1019 | ||
1020 | return err; | |
1021 | } | |
1022 | ||
1023 | asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask, | |
1024 | compat_ulong_t maxnode) | |
1025 | { | |
1026 | long err = 0; | |
1027 | unsigned long __user *nm = NULL; | |
1028 | unsigned long nr_bits, alloc_size; | |
1029 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1030 | ||
1031 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1032 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1033 | ||
1034 | if (nmask) { | |
1035 | err = compat_get_bitmap(bm, nmask, nr_bits); | |
1036 | nm = compat_alloc_user_space(alloc_size); | |
1037 | err |= copy_to_user(nm, bm, alloc_size); | |
1038 | } | |
1039 | ||
1040 | if (err) | |
1041 | return -EFAULT; | |
1042 | ||
1043 | return sys_set_mempolicy(mode, nm, nr_bits+1); | |
1044 | } | |
1045 | ||
1046 | asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len, | |
1047 | compat_ulong_t mode, compat_ulong_t __user *nmask, | |
1048 | compat_ulong_t maxnode, compat_ulong_t flags) | |
1049 | { | |
1050 | long err = 0; | |
1051 | unsigned long __user *nm = NULL; | |
1052 | unsigned long nr_bits, alloc_size; | |
dfcd3c0d | 1053 | nodemask_t bm; |
1da177e4 LT |
1054 | |
1055 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1056 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1057 | ||
1058 | if (nmask) { | |
dfcd3c0d | 1059 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); |
1da177e4 | 1060 | nm = compat_alloc_user_space(alloc_size); |
dfcd3c0d | 1061 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); |
1da177e4 LT |
1062 | } |
1063 | ||
1064 | if (err) | |
1065 | return -EFAULT; | |
1066 | ||
1067 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); | |
1068 | } | |
1069 | ||
1070 | #endif | |
1071 | ||
1072 | /* Return effective policy for a VMA */ | |
48fce342 CL |
1073 | static struct mempolicy * get_vma_policy(struct task_struct *task, |
1074 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 1075 | { |
6e21c8f1 | 1076 | struct mempolicy *pol = task->mempolicy; |
1da177e4 LT |
1077 | |
1078 | if (vma) { | |
1079 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
8bccd85f | 1080 | pol = vma->vm_ops->get_policy(vma, addr); |
1da177e4 LT |
1081 | else if (vma->vm_policy && |
1082 | vma->vm_policy->policy != MPOL_DEFAULT) | |
1083 | pol = vma->vm_policy; | |
1084 | } | |
1085 | if (!pol) | |
1086 | pol = &default_policy; | |
1087 | return pol; | |
1088 | } | |
1089 | ||
1090 | /* Return a zonelist representing a mempolicy */ | |
dd0fc66f | 1091 | static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy) |
1da177e4 LT |
1092 | { |
1093 | int nd; | |
1094 | ||
1095 | switch (policy->policy) { | |
1096 | case MPOL_PREFERRED: | |
1097 | nd = policy->v.preferred_node; | |
1098 | if (nd < 0) | |
1099 | nd = numa_node_id(); | |
1100 | break; | |
1101 | case MPOL_BIND: | |
1102 | /* Lower zones don't get a policy applied */ | |
1103 | /* Careful: current->mems_allowed might have moved */ | |
19655d34 | 1104 | if (gfp_zone(gfp) >= policy_zone) |
1da177e4 LT |
1105 | if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist)) |
1106 | return policy->v.zonelist; | |
1107 | /*FALL THROUGH*/ | |
1108 | case MPOL_INTERLEAVE: /* should not happen */ | |
1109 | case MPOL_DEFAULT: | |
1110 | nd = numa_node_id(); | |
1111 | break; | |
1112 | default: | |
1113 | nd = 0; | |
1114 | BUG(); | |
1115 | } | |
af4ca457 | 1116 | return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp); |
1da177e4 LT |
1117 | } |
1118 | ||
1119 | /* Do dynamic interleaving for a process */ | |
1120 | static unsigned interleave_nodes(struct mempolicy *policy) | |
1121 | { | |
1122 | unsigned nid, next; | |
1123 | struct task_struct *me = current; | |
1124 | ||
1125 | nid = me->il_next; | |
dfcd3c0d | 1126 | next = next_node(nid, policy->v.nodes); |
1da177e4 | 1127 | if (next >= MAX_NUMNODES) |
dfcd3c0d | 1128 | next = first_node(policy->v.nodes); |
1da177e4 LT |
1129 | me->il_next = next; |
1130 | return nid; | |
1131 | } | |
1132 | ||
dc85da15 CL |
1133 | /* |
1134 | * Depending on the memory policy provide a node from which to allocate the | |
1135 | * next slab entry. | |
1136 | */ | |
1137 | unsigned slab_node(struct mempolicy *policy) | |
1138 | { | |
dc85da15 CL |
1139 | switch (policy->policy) { |
1140 | case MPOL_INTERLEAVE: | |
1141 | return interleave_nodes(policy); | |
1142 | ||
1143 | case MPOL_BIND: | |
1144 | /* | |
1145 | * Follow bind policy behavior and start allocation at the | |
1146 | * first node. | |
1147 | */ | |
89fa3024 | 1148 | return zone_to_nid(policy->v.zonelist->zones[0]); |
dc85da15 CL |
1149 | |
1150 | case MPOL_PREFERRED: | |
1151 | if (policy->v.preferred_node >= 0) | |
1152 | return policy->v.preferred_node; | |
1153 | /* Fall through */ | |
1154 | ||
1155 | default: | |
1156 | return numa_node_id(); | |
1157 | } | |
1158 | } | |
1159 | ||
1da177e4 LT |
1160 | /* Do static interleaving for a VMA with known offset. */ |
1161 | static unsigned offset_il_node(struct mempolicy *pol, | |
1162 | struct vm_area_struct *vma, unsigned long off) | |
1163 | { | |
dfcd3c0d | 1164 | unsigned nnodes = nodes_weight(pol->v.nodes); |
1da177e4 LT |
1165 | unsigned target = (unsigned)off % nnodes; |
1166 | int c; | |
1167 | int nid = -1; | |
1168 | ||
1169 | c = 0; | |
1170 | do { | |
dfcd3c0d | 1171 | nid = next_node(nid, pol->v.nodes); |
1da177e4 LT |
1172 | c++; |
1173 | } while (c <= target); | |
1da177e4 LT |
1174 | return nid; |
1175 | } | |
1176 | ||
5da7ca86 CL |
1177 | /* Determine a node number for interleave */ |
1178 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1179 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1180 | { | |
1181 | if (vma) { | |
1182 | unsigned long off; | |
1183 | ||
3b98b087 NA |
1184 | /* |
1185 | * for small pages, there is no difference between | |
1186 | * shift and PAGE_SHIFT, so the bit-shift is safe. | |
1187 | * for huge pages, since vm_pgoff is in units of small | |
1188 | * pages, we need to shift off the always 0 bits to get | |
1189 | * a useful offset. | |
1190 | */ | |
1191 | BUG_ON(shift < PAGE_SHIFT); | |
1192 | off = vma->vm_pgoff >> (shift - PAGE_SHIFT); | |
5da7ca86 CL |
1193 | off += (addr - vma->vm_start) >> shift; |
1194 | return offset_il_node(pol, vma, off); | |
1195 | } else | |
1196 | return interleave_nodes(pol); | |
1197 | } | |
1198 | ||
00ac59ad | 1199 | #ifdef CONFIG_HUGETLBFS |
5da7ca86 CL |
1200 | /* Return a zonelist suitable for a huge page allocation. */ |
1201 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr) | |
1202 | { | |
1203 | struct mempolicy *pol = get_vma_policy(current, vma, addr); | |
1204 | ||
1205 | if (pol->policy == MPOL_INTERLEAVE) { | |
1206 | unsigned nid; | |
1207 | ||
1208 | nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT); | |
1209 | return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER); | |
1210 | } | |
1211 | return zonelist_policy(GFP_HIGHUSER, pol); | |
1212 | } | |
00ac59ad | 1213 | #endif |
5da7ca86 | 1214 | |
1da177e4 LT |
1215 | /* Allocate a page in interleaved policy. |
1216 | Own path because it needs to do special accounting. */ | |
662f3a0b AK |
1217 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
1218 | unsigned nid) | |
1da177e4 LT |
1219 | { |
1220 | struct zonelist *zl; | |
1221 | struct page *page; | |
1222 | ||
af4ca457 | 1223 | zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp); |
1da177e4 | 1224 | page = __alloc_pages(gfp, order, zl); |
ca889e6c CL |
1225 | if (page && page_zone(page) == zl->zones[0]) |
1226 | inc_zone_page_state(page, NUMA_INTERLEAVE_HIT); | |
1da177e4 LT |
1227 | return page; |
1228 | } | |
1229 | ||
1230 | /** | |
1231 | * alloc_page_vma - Allocate a page for a VMA. | |
1232 | * | |
1233 | * @gfp: | |
1234 | * %GFP_USER user allocation. | |
1235 | * %GFP_KERNEL kernel allocations, | |
1236 | * %GFP_HIGHMEM highmem/user allocations, | |
1237 | * %GFP_FS allocation should not call back into a file system. | |
1238 | * %GFP_ATOMIC don't sleep. | |
1239 | * | |
1240 | * @vma: Pointer to VMA or NULL if not available. | |
1241 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
1242 | * | |
1243 | * This function allocates a page from the kernel page pool and applies | |
1244 | * a NUMA policy associated with the VMA or the current process. | |
1245 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1246 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
1247 | * all allocations for pages that will be mapped into | |
1248 | * user space. Returns NULL when no page can be allocated. | |
1249 | * | |
1250 | * Should be called with the mm_sem of the vma hold. | |
1251 | */ | |
1252 | struct page * | |
dd0fc66f | 1253 | alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr) |
1da177e4 | 1254 | { |
6e21c8f1 | 1255 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
1da177e4 | 1256 | |
cf2a473c | 1257 | cpuset_update_task_memory_state(); |
1da177e4 LT |
1258 | |
1259 | if (unlikely(pol->policy == MPOL_INTERLEAVE)) { | |
1260 | unsigned nid; | |
5da7ca86 CL |
1261 | |
1262 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT); | |
1da177e4 LT |
1263 | return alloc_page_interleave(gfp, 0, nid); |
1264 | } | |
1265 | return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol)); | |
1266 | } | |
1267 | ||
1268 | /** | |
1269 | * alloc_pages_current - Allocate pages. | |
1270 | * | |
1271 | * @gfp: | |
1272 | * %GFP_USER user allocation, | |
1273 | * %GFP_KERNEL kernel allocation, | |
1274 | * %GFP_HIGHMEM highmem allocation, | |
1275 | * %GFP_FS don't call back into a file system. | |
1276 | * %GFP_ATOMIC don't sleep. | |
1277 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
1278 | * | |
1279 | * Allocate a page from the kernel page pool. When not in | |
1280 | * interrupt context and apply the current process NUMA policy. | |
1281 | * Returns NULL when no page can be allocated. | |
1282 | * | |
cf2a473c | 1283 | * Don't call cpuset_update_task_memory_state() unless |
1da177e4 LT |
1284 | * 1) it's ok to take cpuset_sem (can WAIT), and |
1285 | * 2) allocating for current task (not interrupt). | |
1286 | */ | |
dd0fc66f | 1287 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
1da177e4 LT |
1288 | { |
1289 | struct mempolicy *pol = current->mempolicy; | |
1290 | ||
1291 | if ((gfp & __GFP_WAIT) && !in_interrupt()) | |
cf2a473c | 1292 | cpuset_update_task_memory_state(); |
9b819d20 | 1293 | if (!pol || in_interrupt() || (gfp & __GFP_THISNODE)) |
1da177e4 LT |
1294 | pol = &default_policy; |
1295 | if (pol->policy == MPOL_INTERLEAVE) | |
1296 | return alloc_page_interleave(gfp, order, interleave_nodes(pol)); | |
1297 | return __alloc_pages(gfp, order, zonelist_policy(gfp, pol)); | |
1298 | } | |
1299 | EXPORT_SYMBOL(alloc_pages_current); | |
1300 | ||
4225399a PJ |
1301 | /* |
1302 | * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it | |
1303 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() | |
1304 | * with the mems_allowed returned by cpuset_mems_allowed(). This | |
1305 | * keeps mempolicies cpuset relative after its cpuset moves. See | |
1306 | * further kernel/cpuset.c update_nodemask(). | |
1307 | */ | |
1308 | void *cpuset_being_rebound; | |
1309 | ||
1da177e4 LT |
1310 | /* Slow path of a mempolicy copy */ |
1311 | struct mempolicy *__mpol_copy(struct mempolicy *old) | |
1312 | { | |
1313 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
1314 | ||
1315 | if (!new) | |
1316 | return ERR_PTR(-ENOMEM); | |
4225399a PJ |
1317 | if (current_cpuset_is_being_rebound()) { |
1318 | nodemask_t mems = cpuset_mems_allowed(current); | |
1319 | mpol_rebind_policy(old, &mems); | |
1320 | } | |
1da177e4 LT |
1321 | *new = *old; |
1322 | atomic_set(&new->refcnt, 1); | |
1323 | if (new->policy == MPOL_BIND) { | |
1324 | int sz = ksize(old->v.zonelist); | |
1325 | new->v.zonelist = kmalloc(sz, SLAB_KERNEL); | |
1326 | if (!new->v.zonelist) { | |
1327 | kmem_cache_free(policy_cache, new); | |
1328 | return ERR_PTR(-ENOMEM); | |
1329 | } | |
1330 | memcpy(new->v.zonelist, old->v.zonelist, sz); | |
1331 | } | |
1332 | return new; | |
1333 | } | |
1334 | ||
1335 | /* Slow path of a mempolicy comparison */ | |
1336 | int __mpol_equal(struct mempolicy *a, struct mempolicy *b) | |
1337 | { | |
1338 | if (!a || !b) | |
1339 | return 0; | |
1340 | if (a->policy != b->policy) | |
1341 | return 0; | |
1342 | switch (a->policy) { | |
1343 | case MPOL_DEFAULT: | |
1344 | return 1; | |
1345 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 1346 | return nodes_equal(a->v.nodes, b->v.nodes); |
1da177e4 LT |
1347 | case MPOL_PREFERRED: |
1348 | return a->v.preferred_node == b->v.preferred_node; | |
1349 | case MPOL_BIND: { | |
1350 | int i; | |
1351 | for (i = 0; a->v.zonelist->zones[i]; i++) | |
1352 | if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i]) | |
1353 | return 0; | |
1354 | return b->v.zonelist->zones[i] == NULL; | |
1355 | } | |
1356 | default: | |
1357 | BUG(); | |
1358 | return 0; | |
1359 | } | |
1360 | } | |
1361 | ||
1362 | /* Slow path of a mpol destructor. */ | |
1363 | void __mpol_free(struct mempolicy *p) | |
1364 | { | |
1365 | if (!atomic_dec_and_test(&p->refcnt)) | |
1366 | return; | |
1367 | if (p->policy == MPOL_BIND) | |
1368 | kfree(p->v.zonelist); | |
1369 | p->policy = MPOL_DEFAULT; | |
1370 | kmem_cache_free(policy_cache, p); | |
1371 | } | |
1372 | ||
1da177e4 LT |
1373 | /* |
1374 | * Shared memory backing store policy support. | |
1375 | * | |
1376 | * Remember policies even when nobody has shared memory mapped. | |
1377 | * The policies are kept in Red-Black tree linked from the inode. | |
1378 | * They are protected by the sp->lock spinlock, which should be held | |
1379 | * for any accesses to the tree. | |
1380 | */ | |
1381 | ||
1382 | /* lookup first element intersecting start-end */ | |
1383 | /* Caller holds sp->lock */ | |
1384 | static struct sp_node * | |
1385 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
1386 | { | |
1387 | struct rb_node *n = sp->root.rb_node; | |
1388 | ||
1389 | while (n) { | |
1390 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
1391 | ||
1392 | if (start >= p->end) | |
1393 | n = n->rb_right; | |
1394 | else if (end <= p->start) | |
1395 | n = n->rb_left; | |
1396 | else | |
1397 | break; | |
1398 | } | |
1399 | if (!n) | |
1400 | return NULL; | |
1401 | for (;;) { | |
1402 | struct sp_node *w = NULL; | |
1403 | struct rb_node *prev = rb_prev(n); | |
1404 | if (!prev) | |
1405 | break; | |
1406 | w = rb_entry(prev, struct sp_node, nd); | |
1407 | if (w->end <= start) | |
1408 | break; | |
1409 | n = prev; | |
1410 | } | |
1411 | return rb_entry(n, struct sp_node, nd); | |
1412 | } | |
1413 | ||
1414 | /* Insert a new shared policy into the list. */ | |
1415 | /* Caller holds sp->lock */ | |
1416 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) | |
1417 | { | |
1418 | struct rb_node **p = &sp->root.rb_node; | |
1419 | struct rb_node *parent = NULL; | |
1420 | struct sp_node *nd; | |
1421 | ||
1422 | while (*p) { | |
1423 | parent = *p; | |
1424 | nd = rb_entry(parent, struct sp_node, nd); | |
1425 | if (new->start < nd->start) | |
1426 | p = &(*p)->rb_left; | |
1427 | else if (new->end > nd->end) | |
1428 | p = &(*p)->rb_right; | |
1429 | else | |
1430 | BUG(); | |
1431 | } | |
1432 | rb_link_node(&new->nd, parent, p); | |
1433 | rb_insert_color(&new->nd, &sp->root); | |
1434 | PDprintk("inserting %lx-%lx: %d\n", new->start, new->end, | |
1435 | new->policy ? new->policy->policy : 0); | |
1436 | } | |
1437 | ||
1438 | /* Find shared policy intersecting idx */ | |
1439 | struct mempolicy * | |
1440 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
1441 | { | |
1442 | struct mempolicy *pol = NULL; | |
1443 | struct sp_node *sn; | |
1444 | ||
1445 | if (!sp->root.rb_node) | |
1446 | return NULL; | |
1447 | spin_lock(&sp->lock); | |
1448 | sn = sp_lookup(sp, idx, idx+1); | |
1449 | if (sn) { | |
1450 | mpol_get(sn->policy); | |
1451 | pol = sn->policy; | |
1452 | } | |
1453 | spin_unlock(&sp->lock); | |
1454 | return pol; | |
1455 | } | |
1456 | ||
1457 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) | |
1458 | { | |
1459 | PDprintk("deleting %lx-l%x\n", n->start, n->end); | |
1460 | rb_erase(&n->nd, &sp->root); | |
1461 | mpol_free(n->policy); | |
1462 | kmem_cache_free(sn_cache, n); | |
1463 | } | |
1464 | ||
1465 | struct sp_node * | |
1466 | sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol) | |
1467 | { | |
1468 | struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
1469 | ||
1470 | if (!n) | |
1471 | return NULL; | |
1472 | n->start = start; | |
1473 | n->end = end; | |
1474 | mpol_get(pol); | |
1475 | n->policy = pol; | |
1476 | return n; | |
1477 | } | |
1478 | ||
1479 | /* Replace a policy range. */ | |
1480 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
1481 | unsigned long end, struct sp_node *new) | |
1482 | { | |
1483 | struct sp_node *n, *new2 = NULL; | |
1484 | ||
1485 | restart: | |
1486 | spin_lock(&sp->lock); | |
1487 | n = sp_lookup(sp, start, end); | |
1488 | /* Take care of old policies in the same range. */ | |
1489 | while (n && n->start < end) { | |
1490 | struct rb_node *next = rb_next(&n->nd); | |
1491 | if (n->start >= start) { | |
1492 | if (n->end <= end) | |
1493 | sp_delete(sp, n); | |
1494 | else | |
1495 | n->start = end; | |
1496 | } else { | |
1497 | /* Old policy spanning whole new range. */ | |
1498 | if (n->end > end) { | |
1499 | if (!new2) { | |
1500 | spin_unlock(&sp->lock); | |
1501 | new2 = sp_alloc(end, n->end, n->policy); | |
1502 | if (!new2) | |
1503 | return -ENOMEM; | |
1504 | goto restart; | |
1505 | } | |
1506 | n->end = start; | |
1507 | sp_insert(sp, new2); | |
1508 | new2 = NULL; | |
1509 | break; | |
1510 | } else | |
1511 | n->end = start; | |
1512 | } | |
1513 | if (!next) | |
1514 | break; | |
1515 | n = rb_entry(next, struct sp_node, nd); | |
1516 | } | |
1517 | if (new) | |
1518 | sp_insert(sp, new); | |
1519 | spin_unlock(&sp->lock); | |
1520 | if (new2) { | |
1521 | mpol_free(new2->policy); | |
1522 | kmem_cache_free(sn_cache, new2); | |
1523 | } | |
1524 | return 0; | |
1525 | } | |
1526 | ||
7339ff83 RH |
1527 | void mpol_shared_policy_init(struct shared_policy *info, int policy, |
1528 | nodemask_t *policy_nodes) | |
1529 | { | |
1530 | info->root = RB_ROOT; | |
1531 | spin_lock_init(&info->lock); | |
1532 | ||
1533 | if (policy != MPOL_DEFAULT) { | |
1534 | struct mempolicy *newpol; | |
1535 | ||
1536 | /* Falls back to MPOL_DEFAULT on any error */ | |
1537 | newpol = mpol_new(policy, policy_nodes); | |
1538 | if (!IS_ERR(newpol)) { | |
1539 | /* Create pseudo-vma that contains just the policy */ | |
1540 | struct vm_area_struct pvma; | |
1541 | ||
1542 | memset(&pvma, 0, sizeof(struct vm_area_struct)); | |
1543 | /* Policy covers entire file */ | |
1544 | pvma.vm_end = TASK_SIZE; | |
1545 | mpol_set_shared_policy(info, &pvma, newpol); | |
1546 | mpol_free(newpol); | |
1547 | } | |
1548 | } | |
1549 | } | |
1550 | ||
1da177e4 LT |
1551 | int mpol_set_shared_policy(struct shared_policy *info, |
1552 | struct vm_area_struct *vma, struct mempolicy *npol) | |
1553 | { | |
1554 | int err; | |
1555 | struct sp_node *new = NULL; | |
1556 | unsigned long sz = vma_pages(vma); | |
1557 | ||
1558 | PDprintk("set_shared_policy %lx sz %lu %d %lx\n", | |
1559 | vma->vm_pgoff, | |
1560 | sz, npol? npol->policy : -1, | |
dfcd3c0d | 1561 | npol ? nodes_addr(npol->v.nodes)[0] : -1); |
1da177e4 LT |
1562 | |
1563 | if (npol) { | |
1564 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
1565 | if (!new) | |
1566 | return -ENOMEM; | |
1567 | } | |
1568 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
1569 | if (err && new) | |
1570 | kmem_cache_free(sn_cache, new); | |
1571 | return err; | |
1572 | } | |
1573 | ||
1574 | /* Free a backing policy store on inode delete. */ | |
1575 | void mpol_free_shared_policy(struct shared_policy *p) | |
1576 | { | |
1577 | struct sp_node *n; | |
1578 | struct rb_node *next; | |
1579 | ||
1580 | if (!p->root.rb_node) | |
1581 | return; | |
1582 | spin_lock(&p->lock); | |
1583 | next = rb_first(&p->root); | |
1584 | while (next) { | |
1585 | n = rb_entry(next, struct sp_node, nd); | |
1586 | next = rb_next(&n->nd); | |
90c5029e | 1587 | rb_erase(&n->nd, &p->root); |
1da177e4 LT |
1588 | mpol_free(n->policy); |
1589 | kmem_cache_free(sn_cache, n); | |
1590 | } | |
1591 | spin_unlock(&p->lock); | |
1da177e4 LT |
1592 | } |
1593 | ||
1594 | /* assumes fs == KERNEL_DS */ | |
1595 | void __init numa_policy_init(void) | |
1596 | { | |
1597 | policy_cache = kmem_cache_create("numa_policy", | |
1598 | sizeof(struct mempolicy), | |
1599 | 0, SLAB_PANIC, NULL, NULL); | |
1600 | ||
1601 | sn_cache = kmem_cache_create("shared_policy_node", | |
1602 | sizeof(struct sp_node), | |
1603 | 0, SLAB_PANIC, NULL, NULL); | |
1604 | ||
1605 | /* Set interleaving policy for system init. This way not all | |
1606 | the data structures allocated at system boot end up in node zero. */ | |
1607 | ||
8bccd85f | 1608 | if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map)) |
1da177e4 LT |
1609 | printk("numa_policy_init: interleaving failed\n"); |
1610 | } | |
1611 | ||
8bccd85f | 1612 | /* Reset policy of current process to default */ |
1da177e4 LT |
1613 | void numa_default_policy(void) |
1614 | { | |
8bccd85f | 1615 | do_set_mempolicy(MPOL_DEFAULT, NULL); |
1da177e4 | 1616 | } |
68860ec1 PJ |
1617 | |
1618 | /* Migrate a policy to a different set of nodes */ | |
74cb2155 | 1619 | void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask) |
68860ec1 | 1620 | { |
74cb2155 | 1621 | nodemask_t *mpolmask; |
68860ec1 PJ |
1622 | nodemask_t tmp; |
1623 | ||
1624 | if (!pol) | |
1625 | return; | |
74cb2155 PJ |
1626 | mpolmask = &pol->cpuset_mems_allowed; |
1627 | if (nodes_equal(*mpolmask, *newmask)) | |
1628 | return; | |
68860ec1 PJ |
1629 | |
1630 | switch (pol->policy) { | |
1631 | case MPOL_DEFAULT: | |
1632 | break; | |
1633 | case MPOL_INTERLEAVE: | |
74cb2155 | 1634 | nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask); |
68860ec1 | 1635 | pol->v.nodes = tmp; |
74cb2155 PJ |
1636 | *mpolmask = *newmask; |
1637 | current->il_next = node_remap(current->il_next, | |
1638 | *mpolmask, *newmask); | |
68860ec1 PJ |
1639 | break; |
1640 | case MPOL_PREFERRED: | |
1641 | pol->v.preferred_node = node_remap(pol->v.preferred_node, | |
74cb2155 PJ |
1642 | *mpolmask, *newmask); |
1643 | *mpolmask = *newmask; | |
68860ec1 PJ |
1644 | break; |
1645 | case MPOL_BIND: { | |
1646 | nodemask_t nodes; | |
1647 | struct zone **z; | |
1648 | struct zonelist *zonelist; | |
1649 | ||
1650 | nodes_clear(nodes); | |
1651 | for (z = pol->v.zonelist->zones; *z; z++) | |
89fa3024 | 1652 | node_set(zone_to_nid(*z), nodes); |
74cb2155 | 1653 | nodes_remap(tmp, nodes, *mpolmask, *newmask); |
68860ec1 PJ |
1654 | nodes = tmp; |
1655 | ||
1656 | zonelist = bind_zonelist(&nodes); | |
1657 | ||
1658 | /* If no mem, then zonelist is NULL and we keep old zonelist. | |
1659 | * If that old zonelist has no remaining mems_allowed nodes, | |
1660 | * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT. | |
1661 | */ | |
1662 | ||
1663 | if (zonelist) { | |
1664 | /* Good - got mem - substitute new zonelist */ | |
1665 | kfree(pol->v.zonelist); | |
1666 | pol->v.zonelist = zonelist; | |
1667 | } | |
74cb2155 | 1668 | *mpolmask = *newmask; |
68860ec1 PJ |
1669 | break; |
1670 | } | |
1671 | default: | |
1672 | BUG(); | |
1673 | break; | |
1674 | } | |
1675 | } | |
1676 | ||
1677 | /* | |
74cb2155 PJ |
1678 | * Wrapper for mpol_rebind_policy() that just requires task |
1679 | * pointer, and updates task mempolicy. | |
68860ec1 | 1680 | */ |
74cb2155 PJ |
1681 | |
1682 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) | |
68860ec1 | 1683 | { |
74cb2155 | 1684 | mpol_rebind_policy(tsk->mempolicy, new); |
68860ec1 | 1685 | } |
1a75a6c8 | 1686 | |
4225399a PJ |
1687 | /* |
1688 | * Rebind each vma in mm to new nodemask. | |
1689 | * | |
1690 | * Call holding a reference to mm. Takes mm->mmap_sem during call. | |
1691 | */ | |
1692 | ||
1693 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) | |
1694 | { | |
1695 | struct vm_area_struct *vma; | |
1696 | ||
1697 | down_write(&mm->mmap_sem); | |
1698 | for (vma = mm->mmap; vma; vma = vma->vm_next) | |
1699 | mpol_rebind_policy(vma->vm_policy, new); | |
1700 | up_write(&mm->mmap_sem); | |
1701 | } | |
1702 | ||
1a75a6c8 CL |
1703 | /* |
1704 | * Display pages allocated per node and memory policy via /proc. | |
1705 | */ | |
1706 | ||
1707 | static const char *policy_types[] = { "default", "prefer", "bind", | |
1708 | "interleave" }; | |
1709 | ||
1710 | /* | |
1711 | * Convert a mempolicy into a string. | |
1712 | * Returns the number of characters in buffer (if positive) | |
1713 | * or an error (negative) | |
1714 | */ | |
1715 | static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) | |
1716 | { | |
1717 | char *p = buffer; | |
1718 | int l; | |
1719 | nodemask_t nodes; | |
1720 | int mode = pol ? pol->policy : MPOL_DEFAULT; | |
1721 | ||
1722 | switch (mode) { | |
1723 | case MPOL_DEFAULT: | |
1724 | nodes_clear(nodes); | |
1725 | break; | |
1726 | ||
1727 | case MPOL_PREFERRED: | |
1728 | nodes_clear(nodes); | |
1729 | node_set(pol->v.preferred_node, nodes); | |
1730 | break; | |
1731 | ||
1732 | case MPOL_BIND: | |
1733 | get_zonemask(pol, &nodes); | |
1734 | break; | |
1735 | ||
1736 | case MPOL_INTERLEAVE: | |
1737 | nodes = pol->v.nodes; | |
1738 | break; | |
1739 | ||
1740 | default: | |
1741 | BUG(); | |
1742 | return -EFAULT; | |
1743 | } | |
1744 | ||
1745 | l = strlen(policy_types[mode]); | |
1746 | if (buffer + maxlen < p + l + 1) | |
1747 | return -ENOSPC; | |
1748 | ||
1749 | strcpy(p, policy_types[mode]); | |
1750 | p += l; | |
1751 | ||
1752 | if (!nodes_empty(nodes)) { | |
1753 | if (buffer + maxlen < p + 2) | |
1754 | return -ENOSPC; | |
1755 | *p++ = '='; | |
1756 | p += nodelist_scnprintf(p, buffer + maxlen - p, nodes); | |
1757 | } | |
1758 | return p - buffer; | |
1759 | } | |
1760 | ||
1761 | struct numa_maps { | |
1762 | unsigned long pages; | |
1763 | unsigned long anon; | |
397874df CL |
1764 | unsigned long active; |
1765 | unsigned long writeback; | |
1a75a6c8 | 1766 | unsigned long mapcount_max; |
397874df CL |
1767 | unsigned long dirty; |
1768 | unsigned long swapcache; | |
1a75a6c8 CL |
1769 | unsigned long node[MAX_NUMNODES]; |
1770 | }; | |
1771 | ||
397874df | 1772 | static void gather_stats(struct page *page, void *private, int pte_dirty) |
1a75a6c8 CL |
1773 | { |
1774 | struct numa_maps *md = private; | |
1775 | int count = page_mapcount(page); | |
1776 | ||
397874df CL |
1777 | md->pages++; |
1778 | if (pte_dirty || PageDirty(page)) | |
1779 | md->dirty++; | |
1a75a6c8 | 1780 | |
397874df CL |
1781 | if (PageSwapCache(page)) |
1782 | md->swapcache++; | |
1a75a6c8 | 1783 | |
397874df CL |
1784 | if (PageActive(page)) |
1785 | md->active++; | |
1786 | ||
1787 | if (PageWriteback(page)) | |
1788 | md->writeback++; | |
1a75a6c8 CL |
1789 | |
1790 | if (PageAnon(page)) | |
1791 | md->anon++; | |
1792 | ||
397874df CL |
1793 | if (count > md->mapcount_max) |
1794 | md->mapcount_max = count; | |
1795 | ||
1a75a6c8 | 1796 | md->node[page_to_nid(page)]++; |
1a75a6c8 CL |
1797 | } |
1798 | ||
7f709ed0 | 1799 | #ifdef CONFIG_HUGETLB_PAGE |
397874df CL |
1800 | static void check_huge_range(struct vm_area_struct *vma, |
1801 | unsigned long start, unsigned long end, | |
1802 | struct numa_maps *md) | |
1803 | { | |
1804 | unsigned long addr; | |
1805 | struct page *page; | |
1806 | ||
1807 | for (addr = start; addr < end; addr += HPAGE_SIZE) { | |
1808 | pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK); | |
1809 | pte_t pte; | |
1810 | ||
1811 | if (!ptep) | |
1812 | continue; | |
1813 | ||
1814 | pte = *ptep; | |
1815 | if (pte_none(pte)) | |
1816 | continue; | |
1817 | ||
1818 | page = pte_page(pte); | |
1819 | if (!page) | |
1820 | continue; | |
1821 | ||
1822 | gather_stats(page, md, pte_dirty(*ptep)); | |
1823 | } | |
1824 | } | |
7f709ed0 AM |
1825 | #else |
1826 | static inline void check_huge_range(struct vm_area_struct *vma, | |
1827 | unsigned long start, unsigned long end, | |
1828 | struct numa_maps *md) | |
1829 | { | |
1830 | } | |
1831 | #endif | |
397874df | 1832 | |
1a75a6c8 CL |
1833 | int show_numa_map(struct seq_file *m, void *v) |
1834 | { | |
99f89551 | 1835 | struct proc_maps_private *priv = m->private; |
1a75a6c8 CL |
1836 | struct vm_area_struct *vma = v; |
1837 | struct numa_maps *md; | |
397874df CL |
1838 | struct file *file = vma->vm_file; |
1839 | struct mm_struct *mm = vma->vm_mm; | |
1a75a6c8 CL |
1840 | int n; |
1841 | char buffer[50]; | |
1842 | ||
397874df | 1843 | if (!mm) |
1a75a6c8 CL |
1844 | return 0; |
1845 | ||
1846 | md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL); | |
1847 | if (!md) | |
1848 | return 0; | |
1849 | ||
397874df | 1850 | mpol_to_str(buffer, sizeof(buffer), |
99f89551 | 1851 | get_vma_policy(priv->task, vma, vma->vm_start)); |
397874df CL |
1852 | |
1853 | seq_printf(m, "%08lx %s", vma->vm_start, buffer); | |
1854 | ||
1855 | if (file) { | |
1856 | seq_printf(m, " file="); | |
1857 | seq_path(m, file->f_vfsmnt, file->f_dentry, "\n\t= "); | |
1858 | } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) { | |
1859 | seq_printf(m, " heap"); | |
1860 | } else if (vma->vm_start <= mm->start_stack && | |
1861 | vma->vm_end >= mm->start_stack) { | |
1862 | seq_printf(m, " stack"); | |
1863 | } | |
1864 | ||
1865 | if (is_vm_hugetlb_page(vma)) { | |
1866 | check_huge_range(vma, vma->vm_start, vma->vm_end, md); | |
1867 | seq_printf(m, " huge"); | |
1868 | } else { | |
a57ebfdb | 1869 | check_pgd_range(vma, vma->vm_start, vma->vm_end, |
397874df CL |
1870 | &node_online_map, MPOL_MF_STATS, md); |
1871 | } | |
1872 | ||
1873 | if (!md->pages) | |
1874 | goto out; | |
1a75a6c8 | 1875 | |
397874df CL |
1876 | if (md->anon) |
1877 | seq_printf(m," anon=%lu",md->anon); | |
1a75a6c8 | 1878 | |
397874df CL |
1879 | if (md->dirty) |
1880 | seq_printf(m," dirty=%lu",md->dirty); | |
1a75a6c8 | 1881 | |
397874df CL |
1882 | if (md->pages != md->anon && md->pages != md->dirty) |
1883 | seq_printf(m, " mapped=%lu", md->pages); | |
1a75a6c8 | 1884 | |
397874df CL |
1885 | if (md->mapcount_max > 1) |
1886 | seq_printf(m, " mapmax=%lu", md->mapcount_max); | |
1a75a6c8 | 1887 | |
397874df CL |
1888 | if (md->swapcache) |
1889 | seq_printf(m," swapcache=%lu", md->swapcache); | |
1890 | ||
1891 | if (md->active < md->pages && !is_vm_hugetlb_page(vma)) | |
1892 | seq_printf(m," active=%lu", md->active); | |
1893 | ||
1894 | if (md->writeback) | |
1895 | seq_printf(m," writeback=%lu", md->writeback); | |
1896 | ||
1897 | for_each_online_node(n) | |
1898 | if (md->node[n]) | |
1899 | seq_printf(m, " N%d=%lu", n, md->node[n]); | |
1900 | out: | |
1901 | seq_putc(m, '\n'); | |
1a75a6c8 CL |
1902 | kfree(md); |
1903 | ||
1904 | if (m->count < m->size) | |
99f89551 | 1905 | m->version = (vma != priv->tail_vma) ? vma->vm_start : 0; |
1a75a6c8 CL |
1906 | return 0; |
1907 | } | |
1908 |