]>
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> | |
1da177e4 LT |
75 | #include <linux/nodemask.h> |
76 | #include <linux/cpuset.h> | |
77 | #include <linux/gfp.h> | |
78 | #include <linux/slab.h> | |
79 | #include <linux/string.h> | |
80 | #include <linux/module.h> | |
b488893a | 81 | #include <linux/nsproxy.h> |
1da177e4 LT |
82 | #include <linux/interrupt.h> |
83 | #include <linux/init.h> | |
84 | #include <linux/compat.h> | |
dc9aa5b9 | 85 | #include <linux/swap.h> |
1a75a6c8 CL |
86 | #include <linux/seq_file.h> |
87 | #include <linux/proc_fs.h> | |
b20a3503 | 88 | #include <linux/migrate.h> |
95a402c3 | 89 | #include <linux/rmap.h> |
86c3a764 | 90 | #include <linux/security.h> |
dbcb0f19 | 91 | #include <linux/syscalls.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 | 103 | |
1da177e4 LT |
104 | /* Highest zone. An specific allocation for a zone below that is not |
105 | policied. */ | |
6267276f | 106 | enum zone_type policy_zone = 0; |
1da177e4 | 107 | |
d42c6997 | 108 | struct mempolicy default_policy = { |
1da177e4 LT |
109 | .refcnt = ATOMIC_INIT(1), /* never free it */ |
110 | .policy = MPOL_DEFAULT, | |
111 | }; | |
112 | ||
dbcb0f19 AB |
113 | static void mpol_rebind_policy(struct mempolicy *pol, |
114 | const nodemask_t *newmask); | |
115 | ||
19770b32 MG |
116 | /* Check that the nodemask contains at least one populated zone */ |
117 | static int is_valid_nodemask(nodemask_t *nodemask) | |
1da177e4 | 118 | { |
19770b32 | 119 | int nd, k; |
1da177e4 | 120 | |
19770b32 MG |
121 | /* Check that there is something useful in this mask */ |
122 | k = policy_zone; | |
123 | ||
124 | for_each_node_mask(nd, *nodemask) { | |
125 | struct zone *z; | |
126 | ||
127 | for (k = 0; k <= policy_zone; k++) { | |
128 | z = &NODE_DATA(nd)->node_zones[k]; | |
129 | if (z->present_pages > 0) | |
130 | return 1; | |
dd942ae3 | 131 | } |
8af5e2eb | 132 | } |
19770b32 MG |
133 | |
134 | return 0; | |
1da177e4 LT |
135 | } |
136 | ||
f5b087b5 DR |
137 | static inline int mpol_store_user_nodemask(const struct mempolicy *pol) |
138 | { | |
139 | return pol->flags & MPOL_F_STATIC_NODES; | |
140 | } | |
141 | ||
1da177e4 | 142 | /* Create a new policy */ |
028fec41 DR |
143 | static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags, |
144 | nodemask_t *nodes) | |
1da177e4 LT |
145 | { |
146 | struct mempolicy *policy; | |
f5b087b5 | 147 | nodemask_t cpuset_context_nmask; |
1da177e4 | 148 | |
028fec41 DR |
149 | pr_debug("setting mode %d flags %d nodes[0] %lx\n", |
150 | mode, flags, nodes ? nodes_addr(*nodes)[0] : -1); | |
140d5a49 | 151 | |
1da177e4 | 152 | if (mode == MPOL_DEFAULT) |
f5b087b5 DR |
153 | return (nodes && nodes_weight(*nodes)) ? ERR_PTR(-EINVAL) : |
154 | NULL; | |
1da177e4 LT |
155 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
156 | if (!policy) | |
157 | return ERR_PTR(-ENOMEM); | |
158 | atomic_set(&policy->refcnt, 1); | |
f5b087b5 DR |
159 | cpuset_update_task_memory_state(); |
160 | nodes_and(cpuset_context_nmask, *nodes, cpuset_current_mems_allowed); | |
1da177e4 LT |
161 | switch (mode) { |
162 | case MPOL_INTERLEAVE: | |
f5b087b5 DR |
163 | if (nodes_empty(*nodes) || nodes_empty(cpuset_context_nmask)) |
164 | goto free; | |
165 | policy->v.nodes = cpuset_context_nmask; | |
1da177e4 LT |
166 | break; |
167 | case MPOL_PREFERRED: | |
f5b087b5 | 168 | policy->v.preferred_node = first_node(cpuset_context_nmask); |
1da177e4 | 169 | if (policy->v.preferred_node >= MAX_NUMNODES) |
f5b087b5 | 170 | goto free; |
1da177e4 LT |
171 | break; |
172 | case MPOL_BIND: | |
f5b087b5 DR |
173 | if (!is_valid_nodemask(&cpuset_context_nmask)) |
174 | goto free; | |
175 | policy->v.nodes = cpuset_context_nmask; | |
1da177e4 | 176 | break; |
a3b51e01 DR |
177 | default: |
178 | BUG(); | |
1da177e4 LT |
179 | } |
180 | policy->policy = mode; | |
028fec41 | 181 | policy->flags = flags; |
f5b087b5 DR |
182 | if (mpol_store_user_nodemask(policy)) |
183 | policy->w.user_nodemask = *nodes; | |
184 | else | |
185 | policy->w.cpuset_mems_allowed = cpuset_mems_allowed(current); | |
1da177e4 | 186 | return policy; |
f5b087b5 DR |
187 | |
188 | free: | |
189 | kmem_cache_free(policy_cache, policy); | |
190 | return ERR_PTR(-EINVAL); | |
1da177e4 LT |
191 | } |
192 | ||
397874df | 193 | static void gather_stats(struct page *, void *, int pte_dirty); |
fc301289 CL |
194 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
195 | unsigned long flags); | |
1a75a6c8 | 196 | |
38e35860 | 197 | /* Scan through pages checking if pages follow certain conditions. */ |
b5810039 | 198 | static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd, |
dc9aa5b9 CL |
199 | unsigned long addr, unsigned long end, |
200 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 201 | void *private) |
1da177e4 | 202 | { |
91612e0d HD |
203 | pte_t *orig_pte; |
204 | pte_t *pte; | |
705e87c0 | 205 | spinlock_t *ptl; |
941150a3 | 206 | |
705e87c0 | 207 | orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
91612e0d | 208 | do { |
6aab341e | 209 | struct page *page; |
25ba77c1 | 210 | int nid; |
91612e0d HD |
211 | |
212 | if (!pte_present(*pte)) | |
1da177e4 | 213 | continue; |
6aab341e LT |
214 | page = vm_normal_page(vma, addr, *pte); |
215 | if (!page) | |
1da177e4 | 216 | continue; |
053837fc NP |
217 | /* |
218 | * The check for PageReserved here is important to avoid | |
219 | * handling zero pages and other pages that may have been | |
220 | * marked special by the system. | |
221 | * | |
222 | * If the PageReserved would not be checked here then f.e. | |
223 | * the location of the zero page could have an influence | |
224 | * on MPOL_MF_STRICT, zero pages would be counted for | |
225 | * the per node stats, and there would be useless attempts | |
226 | * to put zero pages on the migration list. | |
227 | */ | |
f4598c8b CL |
228 | if (PageReserved(page)) |
229 | continue; | |
6aab341e | 230 | nid = page_to_nid(page); |
38e35860 CL |
231 | if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT)) |
232 | continue; | |
233 | ||
1a75a6c8 | 234 | if (flags & MPOL_MF_STATS) |
397874df | 235 | gather_stats(page, private, pte_dirty(*pte)); |
053837fc | 236 | else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) |
fc301289 | 237 | migrate_page_add(page, private, flags); |
38e35860 CL |
238 | else |
239 | break; | |
91612e0d | 240 | } while (pte++, addr += PAGE_SIZE, addr != end); |
705e87c0 | 241 | pte_unmap_unlock(orig_pte, ptl); |
91612e0d HD |
242 | return addr != end; |
243 | } | |
244 | ||
b5810039 | 245 | static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud, |
dc9aa5b9 CL |
246 | unsigned long addr, unsigned long end, |
247 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 248 | void *private) |
91612e0d HD |
249 | { |
250 | pmd_t *pmd; | |
251 | unsigned long next; | |
252 | ||
253 | pmd = pmd_offset(pud, addr); | |
254 | do { | |
255 | next = pmd_addr_end(addr, end); | |
256 | if (pmd_none_or_clear_bad(pmd)) | |
257 | continue; | |
dc9aa5b9 | 258 | if (check_pte_range(vma, pmd, addr, next, nodes, |
38e35860 | 259 | flags, private)) |
91612e0d HD |
260 | return -EIO; |
261 | } while (pmd++, addr = next, addr != end); | |
262 | return 0; | |
263 | } | |
264 | ||
b5810039 | 265 | static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd, |
dc9aa5b9 CL |
266 | unsigned long addr, unsigned long end, |
267 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 268 | void *private) |
91612e0d HD |
269 | { |
270 | pud_t *pud; | |
271 | unsigned long next; | |
272 | ||
273 | pud = pud_offset(pgd, addr); | |
274 | do { | |
275 | next = pud_addr_end(addr, end); | |
276 | if (pud_none_or_clear_bad(pud)) | |
277 | continue; | |
dc9aa5b9 | 278 | if (check_pmd_range(vma, pud, addr, next, nodes, |
38e35860 | 279 | flags, private)) |
91612e0d HD |
280 | return -EIO; |
281 | } while (pud++, addr = next, addr != end); | |
282 | return 0; | |
283 | } | |
284 | ||
b5810039 | 285 | static inline int check_pgd_range(struct vm_area_struct *vma, |
dc9aa5b9 CL |
286 | unsigned long addr, unsigned long end, |
287 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 288 | void *private) |
91612e0d HD |
289 | { |
290 | pgd_t *pgd; | |
291 | unsigned long next; | |
292 | ||
b5810039 | 293 | pgd = pgd_offset(vma->vm_mm, addr); |
91612e0d HD |
294 | do { |
295 | next = pgd_addr_end(addr, end); | |
296 | if (pgd_none_or_clear_bad(pgd)) | |
297 | continue; | |
dc9aa5b9 | 298 | if (check_pud_range(vma, pgd, addr, next, nodes, |
38e35860 | 299 | flags, private)) |
91612e0d HD |
300 | return -EIO; |
301 | } while (pgd++, addr = next, addr != end); | |
302 | return 0; | |
1da177e4 LT |
303 | } |
304 | ||
dc9aa5b9 CL |
305 | /* |
306 | * Check if all pages in a range are on a set of nodes. | |
307 | * If pagelist != NULL then isolate pages from the LRU and | |
308 | * put them on the pagelist. | |
309 | */ | |
1da177e4 LT |
310 | static struct vm_area_struct * |
311 | check_range(struct mm_struct *mm, unsigned long start, unsigned long end, | |
38e35860 | 312 | const nodemask_t *nodes, unsigned long flags, void *private) |
1da177e4 LT |
313 | { |
314 | int err; | |
315 | struct vm_area_struct *first, *vma, *prev; | |
316 | ||
90036ee5 | 317 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
90036ee5 | 318 | |
b20a3503 CL |
319 | err = migrate_prep(); |
320 | if (err) | |
321 | return ERR_PTR(err); | |
90036ee5 | 322 | } |
053837fc | 323 | |
1da177e4 LT |
324 | first = find_vma(mm, start); |
325 | if (!first) | |
326 | return ERR_PTR(-EFAULT); | |
327 | prev = NULL; | |
328 | for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) { | |
dc9aa5b9 CL |
329 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { |
330 | if (!vma->vm_next && vma->vm_end < end) | |
331 | return ERR_PTR(-EFAULT); | |
332 | if (prev && prev->vm_end < vma->vm_start) | |
333 | return ERR_PTR(-EFAULT); | |
334 | } | |
335 | if (!is_vm_hugetlb_page(vma) && | |
336 | ((flags & MPOL_MF_STRICT) || | |
337 | ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) && | |
338 | vma_migratable(vma)))) { | |
5b952b3c | 339 | unsigned long endvma = vma->vm_end; |
dc9aa5b9 | 340 | |
5b952b3c AK |
341 | if (endvma > end) |
342 | endvma = end; | |
343 | if (vma->vm_start > start) | |
344 | start = vma->vm_start; | |
dc9aa5b9 | 345 | err = check_pgd_range(vma, start, endvma, nodes, |
38e35860 | 346 | flags, private); |
1da177e4 LT |
347 | if (err) { |
348 | first = ERR_PTR(err); | |
349 | break; | |
350 | } | |
351 | } | |
352 | prev = vma; | |
353 | } | |
354 | return first; | |
355 | } | |
356 | ||
357 | /* Apply policy to a single VMA */ | |
358 | static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new) | |
359 | { | |
360 | int err = 0; | |
361 | struct mempolicy *old = vma->vm_policy; | |
362 | ||
140d5a49 | 363 | pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", |
1da177e4 LT |
364 | vma->vm_start, vma->vm_end, vma->vm_pgoff, |
365 | vma->vm_ops, vma->vm_file, | |
366 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
367 | ||
368 | if (vma->vm_ops && vma->vm_ops->set_policy) | |
369 | err = vma->vm_ops->set_policy(vma, new); | |
370 | if (!err) { | |
371 | mpol_get(new); | |
372 | vma->vm_policy = new; | |
373 | mpol_free(old); | |
374 | } | |
375 | return err; | |
376 | } | |
377 | ||
378 | /* Step 2: apply policy to a range and do splits. */ | |
379 | static int mbind_range(struct vm_area_struct *vma, unsigned long start, | |
380 | unsigned long end, struct mempolicy *new) | |
381 | { | |
382 | struct vm_area_struct *next; | |
383 | int err; | |
384 | ||
385 | err = 0; | |
386 | for (; vma && vma->vm_start < end; vma = next) { | |
387 | next = vma->vm_next; | |
388 | if (vma->vm_start < start) | |
389 | err = split_vma(vma->vm_mm, vma, start, 1); | |
390 | if (!err && vma->vm_end > end) | |
391 | err = split_vma(vma->vm_mm, vma, end, 0); | |
392 | if (!err) | |
393 | err = policy_vma(vma, new); | |
394 | if (err) | |
395 | break; | |
396 | } | |
397 | return err; | |
398 | } | |
399 | ||
c61afb18 PJ |
400 | /* |
401 | * Update task->flags PF_MEMPOLICY bit: set iff non-default | |
402 | * mempolicy. Allows more rapid checking of this (combined perhaps | |
403 | * with other PF_* flag bits) on memory allocation hot code paths. | |
404 | * | |
405 | * If called from outside this file, the task 'p' should -only- be | |
406 | * a newly forked child not yet visible on the task list, because | |
407 | * manipulating the task flags of a visible task is not safe. | |
408 | * | |
409 | * The above limitation is why this routine has the funny name | |
410 | * mpol_fix_fork_child_flag(). | |
411 | * | |
412 | * It is also safe to call this with a task pointer of current, | |
413 | * which the static wrapper mpol_set_task_struct_flag() does, | |
414 | * for use within this file. | |
415 | */ | |
416 | ||
417 | void mpol_fix_fork_child_flag(struct task_struct *p) | |
418 | { | |
419 | if (p->mempolicy) | |
420 | p->flags |= PF_MEMPOLICY; | |
421 | else | |
422 | p->flags &= ~PF_MEMPOLICY; | |
423 | } | |
424 | ||
425 | static void mpol_set_task_struct_flag(void) | |
426 | { | |
427 | mpol_fix_fork_child_flag(current); | |
428 | } | |
429 | ||
1da177e4 | 430 | /* Set the process memory policy */ |
028fec41 DR |
431 | static long do_set_mempolicy(unsigned short mode, unsigned short flags, |
432 | nodemask_t *nodes) | |
1da177e4 | 433 | { |
1da177e4 | 434 | struct mempolicy *new; |
1da177e4 | 435 | |
028fec41 | 436 | new = mpol_new(mode, flags, nodes); |
1da177e4 LT |
437 | if (IS_ERR(new)) |
438 | return PTR_ERR(new); | |
439 | mpol_free(current->mempolicy); | |
440 | current->mempolicy = new; | |
c61afb18 | 441 | mpol_set_task_struct_flag(); |
f5b087b5 DR |
442 | if (new && new->policy == MPOL_INTERLEAVE && |
443 | nodes_weight(new->v.nodes)) | |
dfcd3c0d | 444 | current->il_next = first_node(new->v.nodes); |
1da177e4 LT |
445 | return 0; |
446 | } | |
447 | ||
448 | /* Fill a zone bitmap for a policy */ | |
dfcd3c0d | 449 | static void get_zonemask(struct mempolicy *p, nodemask_t *nodes) |
1da177e4 | 450 | { |
dfcd3c0d | 451 | nodes_clear(*nodes); |
1da177e4 | 452 | switch (p->policy) { |
1da177e4 LT |
453 | case MPOL_DEFAULT: |
454 | break; | |
19770b32 MG |
455 | case MPOL_BIND: |
456 | /* Fall through */ | |
1da177e4 | 457 | case MPOL_INTERLEAVE: |
dfcd3c0d | 458 | *nodes = p->v.nodes; |
1da177e4 LT |
459 | break; |
460 | case MPOL_PREFERRED: | |
56bbd65d | 461 | /* or use current node instead of memory_map? */ |
1da177e4 | 462 | if (p->v.preferred_node < 0) |
56bbd65d | 463 | *nodes = node_states[N_HIGH_MEMORY]; |
1da177e4 | 464 | else |
dfcd3c0d | 465 | node_set(p->v.preferred_node, *nodes); |
1da177e4 LT |
466 | break; |
467 | default: | |
468 | BUG(); | |
469 | } | |
470 | } | |
471 | ||
472 | static int lookup_node(struct mm_struct *mm, unsigned long addr) | |
473 | { | |
474 | struct page *p; | |
475 | int err; | |
476 | ||
477 | err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL); | |
478 | if (err >= 0) { | |
479 | err = page_to_nid(p); | |
480 | put_page(p); | |
481 | } | |
482 | return err; | |
483 | } | |
484 | ||
1da177e4 | 485 | /* Retrieve NUMA policy */ |
dbcb0f19 AB |
486 | static long do_get_mempolicy(int *policy, nodemask_t *nmask, |
487 | unsigned long addr, unsigned long flags) | |
1da177e4 | 488 | { |
8bccd85f | 489 | int err; |
1da177e4 LT |
490 | struct mm_struct *mm = current->mm; |
491 | struct vm_area_struct *vma = NULL; | |
492 | struct mempolicy *pol = current->mempolicy; | |
493 | ||
cf2a473c | 494 | cpuset_update_task_memory_state(); |
754af6f5 LS |
495 | if (flags & |
496 | ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED)) | |
1da177e4 | 497 | return -EINVAL; |
754af6f5 LS |
498 | |
499 | if (flags & MPOL_F_MEMS_ALLOWED) { | |
500 | if (flags & (MPOL_F_NODE|MPOL_F_ADDR)) | |
501 | return -EINVAL; | |
502 | *policy = 0; /* just so it's initialized */ | |
503 | *nmask = cpuset_current_mems_allowed; | |
504 | return 0; | |
505 | } | |
506 | ||
1da177e4 LT |
507 | if (flags & MPOL_F_ADDR) { |
508 | down_read(&mm->mmap_sem); | |
509 | vma = find_vma_intersection(mm, addr, addr+1); | |
510 | if (!vma) { | |
511 | up_read(&mm->mmap_sem); | |
512 | return -EFAULT; | |
513 | } | |
514 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
515 | pol = vma->vm_ops->get_policy(vma, addr); | |
516 | else | |
517 | pol = vma->vm_policy; | |
518 | } else if (addr) | |
519 | return -EINVAL; | |
520 | ||
521 | if (!pol) | |
522 | pol = &default_policy; | |
523 | ||
524 | if (flags & MPOL_F_NODE) { | |
525 | if (flags & MPOL_F_ADDR) { | |
526 | err = lookup_node(mm, addr); | |
527 | if (err < 0) | |
528 | goto out; | |
8bccd85f | 529 | *policy = err; |
1da177e4 LT |
530 | } else if (pol == current->mempolicy && |
531 | pol->policy == MPOL_INTERLEAVE) { | |
8bccd85f | 532 | *policy = current->il_next; |
1da177e4 LT |
533 | } else { |
534 | err = -EINVAL; | |
535 | goto out; | |
536 | } | |
537 | } else | |
028fec41 | 538 | *policy = pol->policy | pol->flags; |
1da177e4 LT |
539 | |
540 | if (vma) { | |
541 | up_read(¤t->mm->mmap_sem); | |
542 | vma = NULL; | |
543 | } | |
544 | ||
1da177e4 | 545 | err = 0; |
8bccd85f CL |
546 | if (nmask) |
547 | get_zonemask(pol, nmask); | |
1da177e4 LT |
548 | |
549 | out: | |
550 | if (vma) | |
551 | up_read(¤t->mm->mmap_sem); | |
552 | return err; | |
553 | } | |
554 | ||
b20a3503 | 555 | #ifdef CONFIG_MIGRATION |
6ce3c4c0 CL |
556 | /* |
557 | * page migration | |
558 | */ | |
fc301289 CL |
559 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
560 | unsigned long flags) | |
6ce3c4c0 CL |
561 | { |
562 | /* | |
fc301289 | 563 | * Avoid migrating a page that is shared with others. |
6ce3c4c0 | 564 | */ |
b20a3503 CL |
565 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) |
566 | isolate_lru_page(page, pagelist); | |
7e2ab150 | 567 | } |
6ce3c4c0 | 568 | |
742755a1 | 569 | static struct page *new_node_page(struct page *page, unsigned long node, int **x) |
95a402c3 | 570 | { |
769848c0 | 571 | return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0); |
95a402c3 CL |
572 | } |
573 | ||
7e2ab150 CL |
574 | /* |
575 | * Migrate pages from one node to a target node. | |
576 | * Returns error or the number of pages not migrated. | |
577 | */ | |
dbcb0f19 AB |
578 | static int migrate_to_node(struct mm_struct *mm, int source, int dest, |
579 | int flags) | |
7e2ab150 CL |
580 | { |
581 | nodemask_t nmask; | |
582 | LIST_HEAD(pagelist); | |
583 | int err = 0; | |
584 | ||
585 | nodes_clear(nmask); | |
586 | node_set(source, nmask); | |
6ce3c4c0 | 587 | |
7e2ab150 CL |
588 | check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask, |
589 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); | |
590 | ||
aaa994b3 | 591 | if (!list_empty(&pagelist)) |
95a402c3 CL |
592 | err = migrate_pages(&pagelist, new_node_page, dest); |
593 | ||
7e2ab150 | 594 | return err; |
6ce3c4c0 CL |
595 | } |
596 | ||
39743889 | 597 | /* |
7e2ab150 CL |
598 | * Move pages between the two nodesets so as to preserve the physical |
599 | * layout as much as possible. | |
39743889 CL |
600 | * |
601 | * Returns the number of page that could not be moved. | |
602 | */ | |
603 | int do_migrate_pages(struct mm_struct *mm, | |
604 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
605 | { | |
606 | LIST_HEAD(pagelist); | |
7e2ab150 CL |
607 | int busy = 0; |
608 | int err = 0; | |
609 | nodemask_t tmp; | |
39743889 | 610 | |
7e2ab150 | 611 | down_read(&mm->mmap_sem); |
39743889 | 612 | |
7b2259b3 CL |
613 | err = migrate_vmas(mm, from_nodes, to_nodes, flags); |
614 | if (err) | |
615 | goto out; | |
616 | ||
7e2ab150 CL |
617 | /* |
618 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' | |
619 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' | |
620 | * bit in 'tmp', and return that <source, dest> pair for migration. | |
621 | * The pair of nodemasks 'to' and 'from' define the map. | |
622 | * | |
623 | * If no pair of bits is found that way, fallback to picking some | |
624 | * pair of 'source' and 'dest' bits that are not the same. If the | |
625 | * 'source' and 'dest' bits are the same, this represents a node | |
626 | * that will be migrating to itself, so no pages need move. | |
627 | * | |
628 | * If no bits are left in 'tmp', or if all remaining bits left | |
629 | * in 'tmp' correspond to the same bit in 'to', return false | |
630 | * (nothing left to migrate). | |
631 | * | |
632 | * This lets us pick a pair of nodes to migrate between, such that | |
633 | * if possible the dest node is not already occupied by some other | |
634 | * source node, minimizing the risk of overloading the memory on a | |
635 | * node that would happen if we migrated incoming memory to a node | |
636 | * before migrating outgoing memory source that same node. | |
637 | * | |
638 | * A single scan of tmp is sufficient. As we go, we remember the | |
639 | * most recent <s, d> pair that moved (s != d). If we find a pair | |
640 | * that not only moved, but what's better, moved to an empty slot | |
641 | * (d is not set in tmp), then we break out then, with that pair. | |
642 | * Otherwise when we finish scannng from_tmp, we at least have the | |
643 | * most recent <s, d> pair that moved. If we get all the way through | |
644 | * the scan of tmp without finding any node that moved, much less | |
645 | * moved to an empty node, then there is nothing left worth migrating. | |
646 | */ | |
d4984711 | 647 | |
7e2ab150 CL |
648 | tmp = *from_nodes; |
649 | while (!nodes_empty(tmp)) { | |
650 | int s,d; | |
651 | int source = -1; | |
652 | int dest = 0; | |
653 | ||
654 | for_each_node_mask(s, tmp) { | |
655 | d = node_remap(s, *from_nodes, *to_nodes); | |
656 | if (s == d) | |
657 | continue; | |
658 | ||
659 | source = s; /* Node moved. Memorize */ | |
660 | dest = d; | |
661 | ||
662 | /* dest not in remaining from nodes? */ | |
663 | if (!node_isset(dest, tmp)) | |
664 | break; | |
665 | } | |
666 | if (source == -1) | |
667 | break; | |
668 | ||
669 | node_clear(source, tmp); | |
670 | err = migrate_to_node(mm, source, dest, flags); | |
671 | if (err > 0) | |
672 | busy += err; | |
673 | if (err < 0) | |
674 | break; | |
39743889 | 675 | } |
7b2259b3 | 676 | out: |
39743889 | 677 | up_read(&mm->mmap_sem); |
7e2ab150 CL |
678 | if (err < 0) |
679 | return err; | |
680 | return busy; | |
b20a3503 CL |
681 | |
682 | } | |
683 | ||
3ad33b24 LS |
684 | /* |
685 | * Allocate a new page for page migration based on vma policy. | |
686 | * Start assuming that page is mapped by vma pointed to by @private. | |
687 | * Search forward from there, if not. N.B., this assumes that the | |
688 | * list of pages handed to migrate_pages()--which is how we get here-- | |
689 | * is in virtual address order. | |
690 | */ | |
742755a1 | 691 | static struct page *new_vma_page(struct page *page, unsigned long private, int **x) |
95a402c3 CL |
692 | { |
693 | struct vm_area_struct *vma = (struct vm_area_struct *)private; | |
3ad33b24 | 694 | unsigned long uninitialized_var(address); |
95a402c3 | 695 | |
3ad33b24 LS |
696 | while (vma) { |
697 | address = page_address_in_vma(page, vma); | |
698 | if (address != -EFAULT) | |
699 | break; | |
700 | vma = vma->vm_next; | |
701 | } | |
702 | ||
703 | /* | |
704 | * if !vma, alloc_page_vma() will use task or system default policy | |
705 | */ | |
706 | return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); | |
95a402c3 | 707 | } |
b20a3503 CL |
708 | #else |
709 | ||
710 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
711 | unsigned long flags) | |
712 | { | |
39743889 CL |
713 | } |
714 | ||
b20a3503 CL |
715 | int do_migrate_pages(struct mm_struct *mm, |
716 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
717 | { | |
718 | return -ENOSYS; | |
719 | } | |
95a402c3 | 720 | |
69939749 | 721 | static struct page *new_vma_page(struct page *page, unsigned long private, int **x) |
95a402c3 CL |
722 | { |
723 | return NULL; | |
724 | } | |
b20a3503 CL |
725 | #endif |
726 | ||
dbcb0f19 | 727 | static long do_mbind(unsigned long start, unsigned long len, |
028fec41 DR |
728 | unsigned short mode, unsigned short mode_flags, |
729 | nodemask_t *nmask, unsigned long flags) | |
6ce3c4c0 CL |
730 | { |
731 | struct vm_area_struct *vma; | |
732 | struct mm_struct *mm = current->mm; | |
733 | struct mempolicy *new; | |
734 | unsigned long end; | |
735 | int err; | |
736 | LIST_HEAD(pagelist); | |
737 | ||
a3b51e01 DR |
738 | if (flags & ~(unsigned long)(MPOL_MF_STRICT | |
739 | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
6ce3c4c0 | 740 | return -EINVAL; |
74c00241 | 741 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
6ce3c4c0 CL |
742 | return -EPERM; |
743 | ||
744 | if (start & ~PAGE_MASK) | |
745 | return -EINVAL; | |
746 | ||
747 | if (mode == MPOL_DEFAULT) | |
748 | flags &= ~MPOL_MF_STRICT; | |
749 | ||
750 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; | |
751 | end = start + len; | |
752 | ||
753 | if (end < start) | |
754 | return -EINVAL; | |
755 | if (end == start) | |
756 | return 0; | |
757 | ||
028fec41 | 758 | new = mpol_new(mode, mode_flags, nmask); |
6ce3c4c0 CL |
759 | if (IS_ERR(new)) |
760 | return PTR_ERR(new); | |
761 | ||
762 | /* | |
763 | * If we are using the default policy then operation | |
764 | * on discontinuous address spaces is okay after all | |
765 | */ | |
766 | if (!new) | |
767 | flags |= MPOL_MF_DISCONTIG_OK; | |
768 | ||
028fec41 DR |
769 | pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n", |
770 | start, start + len, mode, mode_flags, | |
771 | nmask ? nodes_addr(*nmask)[0] : -1); | |
6ce3c4c0 CL |
772 | |
773 | down_write(&mm->mmap_sem); | |
774 | vma = check_range(mm, start, end, nmask, | |
775 | flags | MPOL_MF_INVERT, &pagelist); | |
776 | ||
777 | err = PTR_ERR(vma); | |
778 | if (!IS_ERR(vma)) { | |
779 | int nr_failed = 0; | |
780 | ||
781 | err = mbind_range(vma, start, end, new); | |
7e2ab150 | 782 | |
6ce3c4c0 | 783 | if (!list_empty(&pagelist)) |
95a402c3 CL |
784 | nr_failed = migrate_pages(&pagelist, new_vma_page, |
785 | (unsigned long)vma); | |
6ce3c4c0 CL |
786 | |
787 | if (!err && nr_failed && (flags & MPOL_MF_STRICT)) | |
788 | err = -EIO; | |
789 | } | |
b20a3503 | 790 | |
6ce3c4c0 CL |
791 | up_write(&mm->mmap_sem); |
792 | mpol_free(new); | |
793 | return err; | |
794 | } | |
795 | ||
8bccd85f CL |
796 | /* |
797 | * User space interface with variable sized bitmaps for nodelists. | |
798 | */ | |
799 | ||
800 | /* Copy a node mask from user space. */ | |
39743889 | 801 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
8bccd85f CL |
802 | unsigned long maxnode) |
803 | { | |
804 | unsigned long k; | |
805 | unsigned long nlongs; | |
806 | unsigned long endmask; | |
807 | ||
808 | --maxnode; | |
809 | nodes_clear(*nodes); | |
810 | if (maxnode == 0 || !nmask) | |
811 | return 0; | |
a9c930ba | 812 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
636f13c1 | 813 | return -EINVAL; |
8bccd85f CL |
814 | |
815 | nlongs = BITS_TO_LONGS(maxnode); | |
816 | if ((maxnode % BITS_PER_LONG) == 0) | |
817 | endmask = ~0UL; | |
818 | else | |
819 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
820 | ||
821 | /* When the user specified more nodes than supported just check | |
822 | if the non supported part is all zero. */ | |
823 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
824 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
825 | return -EINVAL; | |
826 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
827 | unsigned long t; | |
828 | if (get_user(t, nmask + k)) | |
829 | return -EFAULT; | |
830 | if (k == nlongs - 1) { | |
831 | if (t & endmask) | |
832 | return -EINVAL; | |
833 | } else if (t) | |
834 | return -EINVAL; | |
835 | } | |
836 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
837 | endmask = ~0UL; | |
838 | } | |
839 | ||
840 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
841 | return -EFAULT; | |
842 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
843 | return 0; | |
844 | } | |
845 | ||
846 | /* Copy a kernel node mask to user space */ | |
847 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
848 | nodemask_t *nodes) | |
849 | { | |
850 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
851 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
852 | ||
853 | if (copy > nbytes) { | |
854 | if (copy > PAGE_SIZE) | |
855 | return -EINVAL; | |
856 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
857 | return -EFAULT; | |
858 | copy = nbytes; | |
859 | } | |
860 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
861 | } | |
862 | ||
863 | asmlinkage long sys_mbind(unsigned long start, unsigned long len, | |
864 | unsigned long mode, | |
865 | unsigned long __user *nmask, unsigned long maxnode, | |
866 | unsigned flags) | |
867 | { | |
868 | nodemask_t nodes; | |
869 | int err; | |
028fec41 | 870 | unsigned short mode_flags; |
8bccd85f | 871 | |
028fec41 DR |
872 | mode_flags = mode & MPOL_MODE_FLAGS; |
873 | mode &= ~MPOL_MODE_FLAGS; | |
a3b51e01 DR |
874 | if (mode >= MPOL_MAX) |
875 | return -EINVAL; | |
8bccd85f CL |
876 | err = get_nodes(&nodes, nmask, maxnode); |
877 | if (err) | |
878 | return err; | |
028fec41 | 879 | return do_mbind(start, len, mode, mode_flags, &nodes, flags); |
8bccd85f CL |
880 | } |
881 | ||
882 | /* Set the process memory policy */ | |
883 | asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask, | |
884 | unsigned long maxnode) | |
885 | { | |
886 | int err; | |
887 | nodemask_t nodes; | |
028fec41 | 888 | unsigned short flags; |
8bccd85f | 889 | |
028fec41 DR |
890 | flags = mode & MPOL_MODE_FLAGS; |
891 | mode &= ~MPOL_MODE_FLAGS; | |
892 | if ((unsigned int)mode >= MPOL_MAX) | |
8bccd85f CL |
893 | return -EINVAL; |
894 | err = get_nodes(&nodes, nmask, maxnode); | |
895 | if (err) | |
896 | return err; | |
028fec41 | 897 | return do_set_mempolicy(mode, flags, &nodes); |
8bccd85f CL |
898 | } |
899 | ||
39743889 CL |
900 | asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode, |
901 | const unsigned long __user *old_nodes, | |
902 | const unsigned long __user *new_nodes) | |
903 | { | |
904 | struct mm_struct *mm; | |
905 | struct task_struct *task; | |
906 | nodemask_t old; | |
907 | nodemask_t new; | |
908 | nodemask_t task_nodes; | |
909 | int err; | |
910 | ||
911 | err = get_nodes(&old, old_nodes, maxnode); | |
912 | if (err) | |
913 | return err; | |
914 | ||
915 | err = get_nodes(&new, new_nodes, maxnode); | |
916 | if (err) | |
917 | return err; | |
918 | ||
919 | /* Find the mm_struct */ | |
920 | read_lock(&tasklist_lock); | |
228ebcbe | 921 | task = pid ? find_task_by_vpid(pid) : current; |
39743889 CL |
922 | if (!task) { |
923 | read_unlock(&tasklist_lock); | |
924 | return -ESRCH; | |
925 | } | |
926 | mm = get_task_mm(task); | |
927 | read_unlock(&tasklist_lock); | |
928 | ||
929 | if (!mm) | |
930 | return -EINVAL; | |
931 | ||
932 | /* | |
933 | * Check if this process has the right to modify the specified | |
934 | * process. The right exists if the process has administrative | |
7f927fcc | 935 | * capabilities, superuser privileges or the same |
39743889 CL |
936 | * userid as the target process. |
937 | */ | |
938 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
939 | (current->uid != task->suid) && (current->uid != task->uid) && | |
74c00241 | 940 | !capable(CAP_SYS_NICE)) { |
39743889 CL |
941 | err = -EPERM; |
942 | goto out; | |
943 | } | |
944 | ||
945 | task_nodes = cpuset_mems_allowed(task); | |
946 | /* Is the user allowed to access the target nodes? */ | |
74c00241 | 947 | if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) { |
39743889 CL |
948 | err = -EPERM; |
949 | goto out; | |
950 | } | |
951 | ||
37b07e41 | 952 | if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) { |
3b42d28b CL |
953 | err = -EINVAL; |
954 | goto out; | |
955 | } | |
956 | ||
86c3a764 DQ |
957 | err = security_task_movememory(task); |
958 | if (err) | |
959 | goto out; | |
960 | ||
511030bc | 961 | err = do_migrate_pages(mm, &old, &new, |
74c00241 | 962 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); |
39743889 CL |
963 | out: |
964 | mmput(mm); | |
965 | return err; | |
966 | } | |
967 | ||
968 | ||
8bccd85f CL |
969 | /* Retrieve NUMA policy */ |
970 | asmlinkage long sys_get_mempolicy(int __user *policy, | |
971 | unsigned long __user *nmask, | |
972 | unsigned long maxnode, | |
973 | unsigned long addr, unsigned long flags) | |
974 | { | |
dbcb0f19 AB |
975 | int err; |
976 | int uninitialized_var(pval); | |
8bccd85f CL |
977 | nodemask_t nodes; |
978 | ||
979 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
980 | return -EINVAL; | |
981 | ||
982 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
983 | ||
984 | if (err) | |
985 | return err; | |
986 | ||
987 | if (policy && put_user(pval, policy)) | |
988 | return -EFAULT; | |
989 | ||
990 | if (nmask) | |
991 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
992 | ||
993 | return err; | |
994 | } | |
995 | ||
1da177e4 LT |
996 | #ifdef CONFIG_COMPAT |
997 | ||
998 | asmlinkage long compat_sys_get_mempolicy(int __user *policy, | |
999 | compat_ulong_t __user *nmask, | |
1000 | compat_ulong_t maxnode, | |
1001 | compat_ulong_t addr, compat_ulong_t flags) | |
1002 | { | |
1003 | long err; | |
1004 | unsigned long __user *nm = NULL; | |
1005 | unsigned long nr_bits, alloc_size; | |
1006 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1007 | ||
1008 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1009 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1010 | ||
1011 | if (nmask) | |
1012 | nm = compat_alloc_user_space(alloc_size); | |
1013 | ||
1014 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
1015 | ||
1016 | if (!err && nmask) { | |
1017 | err = copy_from_user(bm, nm, alloc_size); | |
1018 | /* ensure entire bitmap is zeroed */ | |
1019 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
1020 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
1021 | } | |
1022 | ||
1023 | return err; | |
1024 | } | |
1025 | ||
1026 | asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask, | |
1027 | compat_ulong_t maxnode) | |
1028 | { | |
1029 | long err = 0; | |
1030 | unsigned long __user *nm = NULL; | |
1031 | unsigned long nr_bits, alloc_size; | |
1032 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1033 | ||
1034 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1035 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1036 | ||
1037 | if (nmask) { | |
1038 | err = compat_get_bitmap(bm, nmask, nr_bits); | |
1039 | nm = compat_alloc_user_space(alloc_size); | |
1040 | err |= copy_to_user(nm, bm, alloc_size); | |
1041 | } | |
1042 | ||
1043 | if (err) | |
1044 | return -EFAULT; | |
1045 | ||
1046 | return sys_set_mempolicy(mode, nm, nr_bits+1); | |
1047 | } | |
1048 | ||
1049 | asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len, | |
1050 | compat_ulong_t mode, compat_ulong_t __user *nmask, | |
1051 | compat_ulong_t maxnode, compat_ulong_t flags) | |
1052 | { | |
1053 | long err = 0; | |
1054 | unsigned long __user *nm = NULL; | |
1055 | unsigned long nr_bits, alloc_size; | |
dfcd3c0d | 1056 | nodemask_t bm; |
1da177e4 LT |
1057 | |
1058 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1059 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1060 | ||
1061 | if (nmask) { | |
dfcd3c0d | 1062 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); |
1da177e4 | 1063 | nm = compat_alloc_user_space(alloc_size); |
dfcd3c0d | 1064 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); |
1da177e4 LT |
1065 | } |
1066 | ||
1067 | if (err) | |
1068 | return -EFAULT; | |
1069 | ||
1070 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); | |
1071 | } | |
1072 | ||
1073 | #endif | |
1074 | ||
480eccf9 LS |
1075 | /* |
1076 | * get_vma_policy(@task, @vma, @addr) | |
1077 | * @task - task for fallback if vma policy == default | |
1078 | * @vma - virtual memory area whose policy is sought | |
1079 | * @addr - address in @vma for shared policy lookup | |
1080 | * | |
1081 | * Returns effective policy for a VMA at specified address. | |
1082 | * Falls back to @task or system default policy, as necessary. | |
1083 | * Returned policy has extra reference count if shared, vma, | |
1084 | * or some other task's policy [show_numa_maps() can pass | |
1085 | * @task != current]. It is the caller's responsibility to | |
1086 | * free the reference in these cases. | |
1087 | */ | |
48fce342 CL |
1088 | static struct mempolicy * get_vma_policy(struct task_struct *task, |
1089 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 1090 | { |
6e21c8f1 | 1091 | struct mempolicy *pol = task->mempolicy; |
480eccf9 | 1092 | int shared_pol = 0; |
1da177e4 LT |
1093 | |
1094 | if (vma) { | |
480eccf9 | 1095 | if (vma->vm_ops && vma->vm_ops->get_policy) { |
8bccd85f | 1096 | pol = vma->vm_ops->get_policy(vma, addr); |
480eccf9 LS |
1097 | shared_pol = 1; /* if pol non-NULL, add ref below */ |
1098 | } else if (vma->vm_policy && | |
1da177e4 LT |
1099 | vma->vm_policy->policy != MPOL_DEFAULT) |
1100 | pol = vma->vm_policy; | |
1101 | } | |
1102 | if (!pol) | |
1103 | pol = &default_policy; | |
480eccf9 LS |
1104 | else if (!shared_pol && pol != current->mempolicy) |
1105 | mpol_get(pol); /* vma or other task's policy */ | |
1da177e4 LT |
1106 | return pol; |
1107 | } | |
1108 | ||
19770b32 MG |
1109 | /* Return a nodemask representing a mempolicy */ |
1110 | static nodemask_t *nodemask_policy(gfp_t gfp, struct mempolicy *policy) | |
1111 | { | |
1112 | /* Lower zones don't get a nodemask applied for MPOL_BIND */ | |
1113 | if (unlikely(policy->policy == MPOL_BIND) && | |
1114 | gfp_zone(gfp) >= policy_zone && | |
1115 | cpuset_nodemask_valid_mems_allowed(&policy->v.nodes)) | |
1116 | return &policy->v.nodes; | |
1117 | ||
1118 | return NULL; | |
1119 | } | |
1120 | ||
1da177e4 | 1121 | /* Return a zonelist representing a mempolicy */ |
dd0fc66f | 1122 | static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy) |
1da177e4 LT |
1123 | { |
1124 | int nd; | |
1125 | ||
1126 | switch (policy->policy) { | |
1127 | case MPOL_PREFERRED: | |
1128 | nd = policy->v.preferred_node; | |
1129 | if (nd < 0) | |
1130 | nd = numa_node_id(); | |
1131 | break; | |
1132 | case MPOL_BIND: | |
19770b32 MG |
1133 | /* |
1134 | * Normally, MPOL_BIND allocations node-local are node-local | |
1135 | * within the allowed nodemask. However, if __GFP_THISNODE is | |
1136 | * set and the current node is part of the mask, we use the | |
1137 | * the zonelist for the first node in the mask instead. | |
1138 | */ | |
1139 | nd = numa_node_id(); | |
1140 | if (unlikely(gfp & __GFP_THISNODE) && | |
1141 | unlikely(!node_isset(nd, policy->v.nodes))) | |
1142 | nd = first_node(policy->v.nodes); | |
1143 | break; | |
1da177e4 LT |
1144 | case MPOL_INTERLEAVE: /* should not happen */ |
1145 | case MPOL_DEFAULT: | |
1146 | nd = numa_node_id(); | |
1147 | break; | |
1148 | default: | |
1149 | nd = 0; | |
1150 | BUG(); | |
1151 | } | |
0e88460d | 1152 | return node_zonelist(nd, gfp); |
1da177e4 LT |
1153 | } |
1154 | ||
1155 | /* Do dynamic interleaving for a process */ | |
1156 | static unsigned interleave_nodes(struct mempolicy *policy) | |
1157 | { | |
1158 | unsigned nid, next; | |
1159 | struct task_struct *me = current; | |
1160 | ||
1161 | nid = me->il_next; | |
dfcd3c0d | 1162 | next = next_node(nid, policy->v.nodes); |
1da177e4 | 1163 | if (next >= MAX_NUMNODES) |
dfcd3c0d | 1164 | next = first_node(policy->v.nodes); |
f5b087b5 DR |
1165 | if (next < MAX_NUMNODES) |
1166 | me->il_next = next; | |
1da177e4 LT |
1167 | return nid; |
1168 | } | |
1169 | ||
dc85da15 CL |
1170 | /* |
1171 | * Depending on the memory policy provide a node from which to allocate the | |
1172 | * next slab entry. | |
1173 | */ | |
1174 | unsigned slab_node(struct mempolicy *policy) | |
1175 | { | |
a3b51e01 | 1176 | unsigned short pol = policy ? policy->policy : MPOL_DEFAULT; |
765c4507 CL |
1177 | |
1178 | switch (pol) { | |
dc85da15 CL |
1179 | case MPOL_INTERLEAVE: |
1180 | return interleave_nodes(policy); | |
1181 | ||
dd1a239f | 1182 | case MPOL_BIND: { |
dc85da15 CL |
1183 | /* |
1184 | * Follow bind policy behavior and start allocation at the | |
1185 | * first node. | |
1186 | */ | |
19770b32 MG |
1187 | struct zonelist *zonelist; |
1188 | struct zone *zone; | |
1189 | enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL); | |
1190 | zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0]; | |
1191 | (void)first_zones_zonelist(zonelist, highest_zoneidx, | |
1192 | &policy->v.nodes, | |
1193 | &zone); | |
1194 | return zone->node; | |
dd1a239f | 1195 | } |
dc85da15 CL |
1196 | |
1197 | case MPOL_PREFERRED: | |
1198 | if (policy->v.preferred_node >= 0) | |
1199 | return policy->v.preferred_node; | |
1200 | /* Fall through */ | |
1201 | ||
1202 | default: | |
1203 | return numa_node_id(); | |
1204 | } | |
1205 | } | |
1206 | ||
1da177e4 LT |
1207 | /* Do static interleaving for a VMA with known offset. */ |
1208 | static unsigned offset_il_node(struct mempolicy *pol, | |
1209 | struct vm_area_struct *vma, unsigned long off) | |
1210 | { | |
dfcd3c0d | 1211 | unsigned nnodes = nodes_weight(pol->v.nodes); |
f5b087b5 | 1212 | unsigned target; |
1da177e4 LT |
1213 | int c; |
1214 | int nid = -1; | |
1215 | ||
f5b087b5 DR |
1216 | if (!nnodes) |
1217 | return numa_node_id(); | |
1218 | target = (unsigned int)off % nnodes; | |
1da177e4 LT |
1219 | c = 0; |
1220 | do { | |
dfcd3c0d | 1221 | nid = next_node(nid, pol->v.nodes); |
1da177e4 LT |
1222 | c++; |
1223 | } while (c <= target); | |
1da177e4 LT |
1224 | return nid; |
1225 | } | |
1226 | ||
5da7ca86 CL |
1227 | /* Determine a node number for interleave */ |
1228 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1229 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1230 | { | |
1231 | if (vma) { | |
1232 | unsigned long off; | |
1233 | ||
3b98b087 NA |
1234 | /* |
1235 | * for small pages, there is no difference between | |
1236 | * shift and PAGE_SHIFT, so the bit-shift is safe. | |
1237 | * for huge pages, since vm_pgoff is in units of small | |
1238 | * pages, we need to shift off the always 0 bits to get | |
1239 | * a useful offset. | |
1240 | */ | |
1241 | BUG_ON(shift < PAGE_SHIFT); | |
1242 | off = vma->vm_pgoff >> (shift - PAGE_SHIFT); | |
5da7ca86 CL |
1243 | off += (addr - vma->vm_start) >> shift; |
1244 | return offset_il_node(pol, vma, off); | |
1245 | } else | |
1246 | return interleave_nodes(pol); | |
1247 | } | |
1248 | ||
00ac59ad | 1249 | #ifdef CONFIG_HUGETLBFS |
480eccf9 LS |
1250 | /* |
1251 | * huge_zonelist(@vma, @addr, @gfp_flags, @mpol) | |
1252 | * @vma = virtual memory area whose policy is sought | |
1253 | * @addr = address in @vma for shared policy lookup and interleave policy | |
1254 | * @gfp_flags = for requested zone | |
19770b32 MG |
1255 | * @mpol = pointer to mempolicy pointer for reference counted mempolicy |
1256 | * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask | |
480eccf9 LS |
1257 | * |
1258 | * Returns a zonelist suitable for a huge page allocation. | |
19770b32 MG |
1259 | * If the effective policy is 'BIND, returns pointer to local node's zonelist, |
1260 | * and a pointer to the mempolicy's @nodemask for filtering the zonelist. | |
480eccf9 | 1261 | * If it is also a policy for which get_vma_policy() returns an extra |
19770b32 | 1262 | * reference, we must hold that reference until after the allocation. |
480eccf9 | 1263 | * In that case, return policy via @mpol so hugetlb allocation can drop |
19770b32 | 1264 | * the reference. For non-'BIND referenced policies, we can/do drop the |
480eccf9 LS |
1265 | * reference here, so the caller doesn't need to know about the special case |
1266 | * for default and current task policy. | |
1267 | */ | |
396faf03 | 1268 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr, |
19770b32 MG |
1269 | gfp_t gfp_flags, struct mempolicy **mpol, |
1270 | nodemask_t **nodemask) | |
5da7ca86 CL |
1271 | { |
1272 | struct mempolicy *pol = get_vma_policy(current, vma, addr); | |
480eccf9 | 1273 | struct zonelist *zl; |
5da7ca86 | 1274 | |
480eccf9 | 1275 | *mpol = NULL; /* probably no unref needed */ |
19770b32 MG |
1276 | *nodemask = NULL; /* assume !MPOL_BIND */ |
1277 | if (pol->policy == MPOL_BIND) { | |
1278 | *nodemask = &pol->v.nodes; | |
1279 | } else if (pol->policy == MPOL_INTERLEAVE) { | |
5da7ca86 CL |
1280 | unsigned nid; |
1281 | ||
1282 | nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT); | |
69682d85 LS |
1283 | if (unlikely(pol != &default_policy && |
1284 | pol != current->mempolicy)) | |
1285 | __mpol_free(pol); /* finished with pol */ | |
0e88460d | 1286 | return node_zonelist(nid, gfp_flags); |
5da7ca86 | 1287 | } |
480eccf9 LS |
1288 | |
1289 | zl = zonelist_policy(GFP_HIGHUSER, pol); | |
1290 | if (unlikely(pol != &default_policy && pol != current->mempolicy)) { | |
1291 | if (pol->policy != MPOL_BIND) | |
1292 | __mpol_free(pol); /* finished with pol */ | |
1293 | else | |
1294 | *mpol = pol; /* unref needed after allocation */ | |
1295 | } | |
1296 | return zl; | |
5da7ca86 | 1297 | } |
00ac59ad | 1298 | #endif |
5da7ca86 | 1299 | |
1da177e4 LT |
1300 | /* Allocate a page in interleaved policy. |
1301 | Own path because it needs to do special accounting. */ | |
662f3a0b AK |
1302 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
1303 | unsigned nid) | |
1da177e4 LT |
1304 | { |
1305 | struct zonelist *zl; | |
1306 | struct page *page; | |
1307 | ||
0e88460d | 1308 | zl = node_zonelist(nid, gfp); |
1da177e4 | 1309 | page = __alloc_pages(gfp, order, zl); |
dd1a239f | 1310 | if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0])) |
ca889e6c | 1311 | inc_zone_page_state(page, NUMA_INTERLEAVE_HIT); |
1da177e4 LT |
1312 | return page; |
1313 | } | |
1314 | ||
1315 | /** | |
1316 | * alloc_page_vma - Allocate a page for a VMA. | |
1317 | * | |
1318 | * @gfp: | |
1319 | * %GFP_USER user allocation. | |
1320 | * %GFP_KERNEL kernel allocations, | |
1321 | * %GFP_HIGHMEM highmem/user allocations, | |
1322 | * %GFP_FS allocation should not call back into a file system. | |
1323 | * %GFP_ATOMIC don't sleep. | |
1324 | * | |
1325 | * @vma: Pointer to VMA or NULL if not available. | |
1326 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
1327 | * | |
1328 | * This function allocates a page from the kernel page pool and applies | |
1329 | * a NUMA policy associated with the VMA or the current process. | |
1330 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1331 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
1332 | * all allocations for pages that will be mapped into | |
1333 | * user space. Returns NULL when no page can be allocated. | |
1334 | * | |
1335 | * Should be called with the mm_sem of the vma hold. | |
1336 | */ | |
1337 | struct page * | |
dd0fc66f | 1338 | alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr) |
1da177e4 | 1339 | { |
6e21c8f1 | 1340 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
480eccf9 | 1341 | struct zonelist *zl; |
1da177e4 | 1342 | |
cf2a473c | 1343 | cpuset_update_task_memory_state(); |
1da177e4 LT |
1344 | |
1345 | if (unlikely(pol->policy == MPOL_INTERLEAVE)) { | |
1346 | unsigned nid; | |
5da7ca86 CL |
1347 | |
1348 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT); | |
69682d85 LS |
1349 | if (unlikely(pol != &default_policy && |
1350 | pol != current->mempolicy)) | |
1351 | __mpol_free(pol); /* finished with pol */ | |
1da177e4 LT |
1352 | return alloc_page_interleave(gfp, 0, nid); |
1353 | } | |
480eccf9 LS |
1354 | zl = zonelist_policy(gfp, pol); |
1355 | if (pol != &default_policy && pol != current->mempolicy) { | |
1356 | /* | |
1357 | * slow path: ref counted policy -- shared or vma | |
1358 | */ | |
19770b32 MG |
1359 | struct page *page = __alloc_pages_nodemask(gfp, 0, |
1360 | zl, nodemask_policy(gfp, pol)); | |
480eccf9 LS |
1361 | __mpol_free(pol); |
1362 | return page; | |
1363 | } | |
1364 | /* | |
1365 | * fast path: default or task policy | |
1366 | */ | |
19770b32 | 1367 | return __alloc_pages_nodemask(gfp, 0, zl, nodemask_policy(gfp, pol)); |
1da177e4 LT |
1368 | } |
1369 | ||
1370 | /** | |
1371 | * alloc_pages_current - Allocate pages. | |
1372 | * | |
1373 | * @gfp: | |
1374 | * %GFP_USER user allocation, | |
1375 | * %GFP_KERNEL kernel allocation, | |
1376 | * %GFP_HIGHMEM highmem allocation, | |
1377 | * %GFP_FS don't call back into a file system. | |
1378 | * %GFP_ATOMIC don't sleep. | |
1379 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
1380 | * | |
1381 | * Allocate a page from the kernel page pool. When not in | |
1382 | * interrupt context and apply the current process NUMA policy. | |
1383 | * Returns NULL when no page can be allocated. | |
1384 | * | |
cf2a473c | 1385 | * Don't call cpuset_update_task_memory_state() unless |
1da177e4 LT |
1386 | * 1) it's ok to take cpuset_sem (can WAIT), and |
1387 | * 2) allocating for current task (not interrupt). | |
1388 | */ | |
dd0fc66f | 1389 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
1da177e4 LT |
1390 | { |
1391 | struct mempolicy *pol = current->mempolicy; | |
1392 | ||
1393 | if ((gfp & __GFP_WAIT) && !in_interrupt()) | |
cf2a473c | 1394 | cpuset_update_task_memory_state(); |
9b819d20 | 1395 | if (!pol || in_interrupt() || (gfp & __GFP_THISNODE)) |
1da177e4 LT |
1396 | pol = &default_policy; |
1397 | if (pol->policy == MPOL_INTERLEAVE) | |
1398 | return alloc_page_interleave(gfp, order, interleave_nodes(pol)); | |
19770b32 MG |
1399 | return __alloc_pages_nodemask(gfp, order, |
1400 | zonelist_policy(gfp, pol), nodemask_policy(gfp, pol)); | |
1da177e4 LT |
1401 | } |
1402 | EXPORT_SYMBOL(alloc_pages_current); | |
1403 | ||
4225399a PJ |
1404 | /* |
1405 | * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it | |
1406 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() | |
1407 | * with the mems_allowed returned by cpuset_mems_allowed(). This | |
1408 | * keeps mempolicies cpuset relative after its cpuset moves. See | |
1409 | * further kernel/cpuset.c update_nodemask(). | |
1410 | */ | |
4225399a | 1411 | |
1da177e4 LT |
1412 | /* Slow path of a mempolicy copy */ |
1413 | struct mempolicy *__mpol_copy(struct mempolicy *old) | |
1414 | { | |
1415 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
1416 | ||
1417 | if (!new) | |
1418 | return ERR_PTR(-ENOMEM); | |
4225399a PJ |
1419 | if (current_cpuset_is_being_rebound()) { |
1420 | nodemask_t mems = cpuset_mems_allowed(current); | |
1421 | mpol_rebind_policy(old, &mems); | |
1422 | } | |
1da177e4 LT |
1423 | *new = *old; |
1424 | atomic_set(&new->refcnt, 1); | |
1da177e4 LT |
1425 | return new; |
1426 | } | |
1427 | ||
f5b087b5 DR |
1428 | static int mpol_match_intent(const struct mempolicy *a, |
1429 | const struct mempolicy *b) | |
1430 | { | |
1431 | if (a->flags != b->flags) | |
1432 | return 0; | |
1433 | if (!mpol_store_user_nodemask(a)) | |
1434 | return 1; | |
1435 | return nodes_equal(a->w.user_nodemask, b->w.user_nodemask); | |
1436 | } | |
1437 | ||
1da177e4 LT |
1438 | /* Slow path of a mempolicy comparison */ |
1439 | int __mpol_equal(struct mempolicy *a, struct mempolicy *b) | |
1440 | { | |
1441 | if (!a || !b) | |
1442 | return 0; | |
1443 | if (a->policy != b->policy) | |
1444 | return 0; | |
f5b087b5 DR |
1445 | if (a->policy != MPOL_DEFAULT && !mpol_match_intent(a, b)) |
1446 | return 0; | |
1da177e4 LT |
1447 | switch (a->policy) { |
1448 | case MPOL_DEFAULT: | |
1449 | return 1; | |
19770b32 MG |
1450 | case MPOL_BIND: |
1451 | /* Fall through */ | |
1da177e4 | 1452 | case MPOL_INTERLEAVE: |
dfcd3c0d | 1453 | return nodes_equal(a->v.nodes, b->v.nodes); |
1da177e4 LT |
1454 | case MPOL_PREFERRED: |
1455 | return a->v.preferred_node == b->v.preferred_node; | |
1da177e4 LT |
1456 | default: |
1457 | BUG(); | |
1458 | return 0; | |
1459 | } | |
1460 | } | |
1461 | ||
1462 | /* Slow path of a mpol destructor. */ | |
1463 | void __mpol_free(struct mempolicy *p) | |
1464 | { | |
1465 | if (!atomic_dec_and_test(&p->refcnt)) | |
1466 | return; | |
1da177e4 LT |
1467 | p->policy = MPOL_DEFAULT; |
1468 | kmem_cache_free(policy_cache, p); | |
1469 | } | |
1470 | ||
1da177e4 LT |
1471 | /* |
1472 | * Shared memory backing store policy support. | |
1473 | * | |
1474 | * Remember policies even when nobody has shared memory mapped. | |
1475 | * The policies are kept in Red-Black tree linked from the inode. | |
1476 | * They are protected by the sp->lock spinlock, which should be held | |
1477 | * for any accesses to the tree. | |
1478 | */ | |
1479 | ||
1480 | /* lookup first element intersecting start-end */ | |
1481 | /* Caller holds sp->lock */ | |
1482 | static struct sp_node * | |
1483 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
1484 | { | |
1485 | struct rb_node *n = sp->root.rb_node; | |
1486 | ||
1487 | while (n) { | |
1488 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
1489 | ||
1490 | if (start >= p->end) | |
1491 | n = n->rb_right; | |
1492 | else if (end <= p->start) | |
1493 | n = n->rb_left; | |
1494 | else | |
1495 | break; | |
1496 | } | |
1497 | if (!n) | |
1498 | return NULL; | |
1499 | for (;;) { | |
1500 | struct sp_node *w = NULL; | |
1501 | struct rb_node *prev = rb_prev(n); | |
1502 | if (!prev) | |
1503 | break; | |
1504 | w = rb_entry(prev, struct sp_node, nd); | |
1505 | if (w->end <= start) | |
1506 | break; | |
1507 | n = prev; | |
1508 | } | |
1509 | return rb_entry(n, struct sp_node, nd); | |
1510 | } | |
1511 | ||
1512 | /* Insert a new shared policy into the list. */ | |
1513 | /* Caller holds sp->lock */ | |
1514 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) | |
1515 | { | |
1516 | struct rb_node **p = &sp->root.rb_node; | |
1517 | struct rb_node *parent = NULL; | |
1518 | struct sp_node *nd; | |
1519 | ||
1520 | while (*p) { | |
1521 | parent = *p; | |
1522 | nd = rb_entry(parent, struct sp_node, nd); | |
1523 | if (new->start < nd->start) | |
1524 | p = &(*p)->rb_left; | |
1525 | else if (new->end > nd->end) | |
1526 | p = &(*p)->rb_right; | |
1527 | else | |
1528 | BUG(); | |
1529 | } | |
1530 | rb_link_node(&new->nd, parent, p); | |
1531 | rb_insert_color(&new->nd, &sp->root); | |
140d5a49 | 1532 | pr_debug("inserting %lx-%lx: %d\n", new->start, new->end, |
1da177e4 LT |
1533 | new->policy ? new->policy->policy : 0); |
1534 | } | |
1535 | ||
1536 | /* Find shared policy intersecting idx */ | |
1537 | struct mempolicy * | |
1538 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
1539 | { | |
1540 | struct mempolicy *pol = NULL; | |
1541 | struct sp_node *sn; | |
1542 | ||
1543 | if (!sp->root.rb_node) | |
1544 | return NULL; | |
1545 | spin_lock(&sp->lock); | |
1546 | sn = sp_lookup(sp, idx, idx+1); | |
1547 | if (sn) { | |
1548 | mpol_get(sn->policy); | |
1549 | pol = sn->policy; | |
1550 | } | |
1551 | spin_unlock(&sp->lock); | |
1552 | return pol; | |
1553 | } | |
1554 | ||
1555 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) | |
1556 | { | |
140d5a49 | 1557 | pr_debug("deleting %lx-l%lx\n", n->start, n->end); |
1da177e4 LT |
1558 | rb_erase(&n->nd, &sp->root); |
1559 | mpol_free(n->policy); | |
1560 | kmem_cache_free(sn_cache, n); | |
1561 | } | |
1562 | ||
dbcb0f19 AB |
1563 | static struct sp_node *sp_alloc(unsigned long start, unsigned long end, |
1564 | struct mempolicy *pol) | |
1da177e4 LT |
1565 | { |
1566 | struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
1567 | ||
1568 | if (!n) | |
1569 | return NULL; | |
1570 | n->start = start; | |
1571 | n->end = end; | |
1572 | mpol_get(pol); | |
1573 | n->policy = pol; | |
1574 | return n; | |
1575 | } | |
1576 | ||
1577 | /* Replace a policy range. */ | |
1578 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
1579 | unsigned long end, struct sp_node *new) | |
1580 | { | |
1581 | struct sp_node *n, *new2 = NULL; | |
1582 | ||
1583 | restart: | |
1584 | spin_lock(&sp->lock); | |
1585 | n = sp_lookup(sp, start, end); | |
1586 | /* Take care of old policies in the same range. */ | |
1587 | while (n && n->start < end) { | |
1588 | struct rb_node *next = rb_next(&n->nd); | |
1589 | if (n->start >= start) { | |
1590 | if (n->end <= end) | |
1591 | sp_delete(sp, n); | |
1592 | else | |
1593 | n->start = end; | |
1594 | } else { | |
1595 | /* Old policy spanning whole new range. */ | |
1596 | if (n->end > end) { | |
1597 | if (!new2) { | |
1598 | spin_unlock(&sp->lock); | |
1599 | new2 = sp_alloc(end, n->end, n->policy); | |
1600 | if (!new2) | |
1601 | return -ENOMEM; | |
1602 | goto restart; | |
1603 | } | |
1604 | n->end = start; | |
1605 | sp_insert(sp, new2); | |
1606 | new2 = NULL; | |
1607 | break; | |
1608 | } else | |
1609 | n->end = start; | |
1610 | } | |
1611 | if (!next) | |
1612 | break; | |
1613 | n = rb_entry(next, struct sp_node, nd); | |
1614 | } | |
1615 | if (new) | |
1616 | sp_insert(sp, new); | |
1617 | spin_unlock(&sp->lock); | |
1618 | if (new2) { | |
1619 | mpol_free(new2->policy); | |
1620 | kmem_cache_free(sn_cache, new2); | |
1621 | } | |
1622 | return 0; | |
1623 | } | |
1624 | ||
a3b51e01 | 1625 | void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy, |
028fec41 | 1626 | unsigned short flags, nodemask_t *policy_nodes) |
7339ff83 RH |
1627 | { |
1628 | info->root = RB_ROOT; | |
1629 | spin_lock_init(&info->lock); | |
1630 | ||
1631 | if (policy != MPOL_DEFAULT) { | |
1632 | struct mempolicy *newpol; | |
1633 | ||
1634 | /* Falls back to MPOL_DEFAULT on any error */ | |
028fec41 | 1635 | newpol = mpol_new(policy, flags, policy_nodes); |
7339ff83 RH |
1636 | if (!IS_ERR(newpol)) { |
1637 | /* Create pseudo-vma that contains just the policy */ | |
1638 | struct vm_area_struct pvma; | |
1639 | ||
1640 | memset(&pvma, 0, sizeof(struct vm_area_struct)); | |
1641 | /* Policy covers entire file */ | |
1642 | pvma.vm_end = TASK_SIZE; | |
1643 | mpol_set_shared_policy(info, &pvma, newpol); | |
1644 | mpol_free(newpol); | |
1645 | } | |
1646 | } | |
1647 | } | |
1648 | ||
1da177e4 LT |
1649 | int mpol_set_shared_policy(struct shared_policy *info, |
1650 | struct vm_area_struct *vma, struct mempolicy *npol) | |
1651 | { | |
1652 | int err; | |
1653 | struct sp_node *new = NULL; | |
1654 | unsigned long sz = vma_pages(vma); | |
1655 | ||
028fec41 | 1656 | pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n", |
1da177e4 | 1657 | vma->vm_pgoff, |
028fec41 DR |
1658 | sz, npol ? npol->policy : -1, |
1659 | npol ? npol->flags : -1, | |
140d5a49 | 1660 | npol ? nodes_addr(npol->v.nodes)[0] : -1); |
1da177e4 LT |
1661 | |
1662 | if (npol) { | |
1663 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
1664 | if (!new) | |
1665 | return -ENOMEM; | |
1666 | } | |
1667 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
1668 | if (err && new) | |
1669 | kmem_cache_free(sn_cache, new); | |
1670 | return err; | |
1671 | } | |
1672 | ||
1673 | /* Free a backing policy store on inode delete. */ | |
1674 | void mpol_free_shared_policy(struct shared_policy *p) | |
1675 | { | |
1676 | struct sp_node *n; | |
1677 | struct rb_node *next; | |
1678 | ||
1679 | if (!p->root.rb_node) | |
1680 | return; | |
1681 | spin_lock(&p->lock); | |
1682 | next = rb_first(&p->root); | |
1683 | while (next) { | |
1684 | n = rb_entry(next, struct sp_node, nd); | |
1685 | next = rb_next(&n->nd); | |
90c5029e | 1686 | rb_erase(&n->nd, &p->root); |
1da177e4 LT |
1687 | mpol_free(n->policy); |
1688 | kmem_cache_free(sn_cache, n); | |
1689 | } | |
1690 | spin_unlock(&p->lock); | |
1da177e4 LT |
1691 | } |
1692 | ||
1693 | /* assumes fs == KERNEL_DS */ | |
1694 | void __init numa_policy_init(void) | |
1695 | { | |
b71636e2 PM |
1696 | nodemask_t interleave_nodes; |
1697 | unsigned long largest = 0; | |
1698 | int nid, prefer = 0; | |
1699 | ||
1da177e4 LT |
1700 | policy_cache = kmem_cache_create("numa_policy", |
1701 | sizeof(struct mempolicy), | |
20c2df83 | 1702 | 0, SLAB_PANIC, NULL); |
1da177e4 LT |
1703 | |
1704 | sn_cache = kmem_cache_create("shared_policy_node", | |
1705 | sizeof(struct sp_node), | |
20c2df83 | 1706 | 0, SLAB_PANIC, NULL); |
1da177e4 | 1707 | |
b71636e2 PM |
1708 | /* |
1709 | * Set interleaving policy for system init. Interleaving is only | |
1710 | * enabled across suitably sized nodes (default is >= 16MB), or | |
1711 | * fall back to the largest node if they're all smaller. | |
1712 | */ | |
1713 | nodes_clear(interleave_nodes); | |
56bbd65d | 1714 | for_each_node_state(nid, N_HIGH_MEMORY) { |
b71636e2 PM |
1715 | unsigned long total_pages = node_present_pages(nid); |
1716 | ||
1717 | /* Preserve the largest node */ | |
1718 | if (largest < total_pages) { | |
1719 | largest = total_pages; | |
1720 | prefer = nid; | |
1721 | } | |
1722 | ||
1723 | /* Interleave this node? */ | |
1724 | if ((total_pages << PAGE_SHIFT) >= (16 << 20)) | |
1725 | node_set(nid, interleave_nodes); | |
1726 | } | |
1727 | ||
1728 | /* All too small, use the largest */ | |
1729 | if (unlikely(nodes_empty(interleave_nodes))) | |
1730 | node_set(prefer, interleave_nodes); | |
1da177e4 | 1731 | |
028fec41 | 1732 | if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes)) |
1da177e4 LT |
1733 | printk("numa_policy_init: interleaving failed\n"); |
1734 | } | |
1735 | ||
8bccd85f | 1736 | /* Reset policy of current process to default */ |
1da177e4 LT |
1737 | void numa_default_policy(void) |
1738 | { | |
028fec41 | 1739 | do_set_mempolicy(MPOL_DEFAULT, 0, NULL); |
1da177e4 | 1740 | } |
68860ec1 PJ |
1741 | |
1742 | /* Migrate a policy to a different set of nodes */ | |
dbcb0f19 AB |
1743 | static void mpol_rebind_policy(struct mempolicy *pol, |
1744 | const nodemask_t *newmask) | |
68860ec1 PJ |
1745 | { |
1746 | nodemask_t tmp; | |
f5b087b5 | 1747 | int static_nodes; |
68860ec1 PJ |
1748 | |
1749 | if (!pol) | |
1750 | return; | |
f5b087b5 DR |
1751 | static_nodes = pol->flags & MPOL_F_STATIC_NODES; |
1752 | if (!mpol_store_user_nodemask(pol) && | |
1753 | nodes_equal(pol->w.cpuset_mems_allowed, *newmask)) | |
74cb2155 | 1754 | return; |
68860ec1 PJ |
1755 | |
1756 | switch (pol->policy) { | |
1757 | case MPOL_DEFAULT: | |
1758 | break; | |
19770b32 MG |
1759 | case MPOL_BIND: |
1760 | /* Fall through */ | |
68860ec1 | 1761 | case MPOL_INTERLEAVE: |
f5b087b5 DR |
1762 | if (static_nodes) |
1763 | nodes_and(tmp, pol->w.user_nodemask, *newmask); | |
1764 | else { | |
1765 | nodes_remap(tmp, pol->v.nodes, | |
1766 | pol->w.cpuset_mems_allowed, *newmask); | |
1767 | pol->w.cpuset_mems_allowed = *newmask; | |
1768 | } | |
68860ec1 | 1769 | pol->v.nodes = tmp; |
f5b087b5 DR |
1770 | if (!node_isset(current->il_next, tmp)) { |
1771 | current->il_next = next_node(current->il_next, tmp); | |
1772 | if (current->il_next >= MAX_NUMNODES) | |
1773 | current->il_next = first_node(tmp); | |
1774 | if (current->il_next >= MAX_NUMNODES) | |
1775 | current->il_next = numa_node_id(); | |
1776 | } | |
68860ec1 PJ |
1777 | break; |
1778 | case MPOL_PREFERRED: | |
f5b087b5 DR |
1779 | if (static_nodes) { |
1780 | int node = first_node(pol->w.user_nodemask); | |
1781 | ||
1782 | if (node_isset(node, *newmask)) | |
1783 | pol->v.preferred_node = node; | |
1784 | else | |
1785 | pol->v.preferred_node = -1; | |
1786 | } else { | |
1787 | pol->v.preferred_node = node_remap(pol->v.preferred_node, | |
1788 | pol->w.cpuset_mems_allowed, *newmask); | |
1789 | pol->w.cpuset_mems_allowed = *newmask; | |
1790 | } | |
68860ec1 | 1791 | break; |
68860ec1 PJ |
1792 | default: |
1793 | BUG(); | |
1794 | break; | |
1795 | } | |
1796 | } | |
1797 | ||
1798 | /* | |
74cb2155 PJ |
1799 | * Wrapper for mpol_rebind_policy() that just requires task |
1800 | * pointer, and updates task mempolicy. | |
68860ec1 | 1801 | */ |
74cb2155 PJ |
1802 | |
1803 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) | |
68860ec1 | 1804 | { |
74cb2155 | 1805 | mpol_rebind_policy(tsk->mempolicy, new); |
68860ec1 | 1806 | } |
1a75a6c8 | 1807 | |
4225399a PJ |
1808 | /* |
1809 | * Rebind each vma in mm to new nodemask. | |
1810 | * | |
1811 | * Call holding a reference to mm. Takes mm->mmap_sem during call. | |
1812 | */ | |
1813 | ||
1814 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) | |
1815 | { | |
1816 | struct vm_area_struct *vma; | |
1817 | ||
1818 | down_write(&mm->mmap_sem); | |
1819 | for (vma = mm->mmap; vma; vma = vma->vm_next) | |
1820 | mpol_rebind_policy(vma->vm_policy, new); | |
1821 | up_write(&mm->mmap_sem); | |
1822 | } | |
1823 | ||
1a75a6c8 CL |
1824 | /* |
1825 | * Display pages allocated per node and memory policy via /proc. | |
1826 | */ | |
1827 | ||
15ad7cdc HD |
1828 | static const char * const policy_types[] = |
1829 | { "default", "prefer", "bind", "interleave" }; | |
1a75a6c8 CL |
1830 | |
1831 | /* | |
1832 | * Convert a mempolicy into a string. | |
1833 | * Returns the number of characters in buffer (if positive) | |
1834 | * or an error (negative) | |
1835 | */ | |
1836 | static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) | |
1837 | { | |
1838 | char *p = buffer; | |
1839 | int l; | |
1840 | nodemask_t nodes; | |
a3b51e01 | 1841 | unsigned short mode = pol ? pol->policy : MPOL_DEFAULT; |
f5b087b5 | 1842 | unsigned short flags = pol ? pol->flags : 0; |
1a75a6c8 CL |
1843 | |
1844 | switch (mode) { | |
1845 | case MPOL_DEFAULT: | |
1846 | nodes_clear(nodes); | |
1847 | break; | |
1848 | ||
1849 | case MPOL_PREFERRED: | |
1850 | nodes_clear(nodes); | |
1851 | node_set(pol->v.preferred_node, nodes); | |
1852 | break; | |
1853 | ||
1854 | case MPOL_BIND: | |
19770b32 | 1855 | /* Fall through */ |
1a75a6c8 CL |
1856 | case MPOL_INTERLEAVE: |
1857 | nodes = pol->v.nodes; | |
1858 | break; | |
1859 | ||
1860 | default: | |
1861 | BUG(); | |
1862 | return -EFAULT; | |
1863 | } | |
1864 | ||
1865 | l = strlen(policy_types[mode]); | |
1866 | if (buffer + maxlen < p + l + 1) | |
1867 | return -ENOSPC; | |
1868 | ||
1869 | strcpy(p, policy_types[mode]); | |
1870 | p += l; | |
1871 | ||
f5b087b5 DR |
1872 | if (flags) { |
1873 | int need_bar = 0; | |
1874 | ||
1875 | if (buffer + maxlen < p + 2) | |
1876 | return -ENOSPC; | |
1877 | *p++ = '='; | |
1878 | ||
1879 | if (flags & MPOL_F_STATIC_NODES) | |
1880 | p += sprintf(p, "%sstatic", need_bar++ ? "|" : ""); | |
1881 | } | |
1882 | ||
1a75a6c8 CL |
1883 | if (!nodes_empty(nodes)) { |
1884 | if (buffer + maxlen < p + 2) | |
1885 | return -ENOSPC; | |
1886 | *p++ = '='; | |
1887 | p += nodelist_scnprintf(p, buffer + maxlen - p, nodes); | |
1888 | } | |
1889 | return p - buffer; | |
1890 | } | |
1891 | ||
1892 | struct numa_maps { | |
1893 | unsigned long pages; | |
1894 | unsigned long anon; | |
397874df CL |
1895 | unsigned long active; |
1896 | unsigned long writeback; | |
1a75a6c8 | 1897 | unsigned long mapcount_max; |
397874df CL |
1898 | unsigned long dirty; |
1899 | unsigned long swapcache; | |
1a75a6c8 CL |
1900 | unsigned long node[MAX_NUMNODES]; |
1901 | }; | |
1902 | ||
397874df | 1903 | static void gather_stats(struct page *page, void *private, int pte_dirty) |
1a75a6c8 CL |
1904 | { |
1905 | struct numa_maps *md = private; | |
1906 | int count = page_mapcount(page); | |
1907 | ||
397874df CL |
1908 | md->pages++; |
1909 | if (pte_dirty || PageDirty(page)) | |
1910 | md->dirty++; | |
1a75a6c8 | 1911 | |
397874df CL |
1912 | if (PageSwapCache(page)) |
1913 | md->swapcache++; | |
1a75a6c8 | 1914 | |
397874df CL |
1915 | if (PageActive(page)) |
1916 | md->active++; | |
1917 | ||
1918 | if (PageWriteback(page)) | |
1919 | md->writeback++; | |
1a75a6c8 CL |
1920 | |
1921 | if (PageAnon(page)) | |
1922 | md->anon++; | |
1923 | ||
397874df CL |
1924 | if (count > md->mapcount_max) |
1925 | md->mapcount_max = count; | |
1926 | ||
1a75a6c8 | 1927 | md->node[page_to_nid(page)]++; |
1a75a6c8 CL |
1928 | } |
1929 | ||
7f709ed0 | 1930 | #ifdef CONFIG_HUGETLB_PAGE |
397874df CL |
1931 | static void check_huge_range(struct vm_area_struct *vma, |
1932 | unsigned long start, unsigned long end, | |
1933 | struct numa_maps *md) | |
1934 | { | |
1935 | unsigned long addr; | |
1936 | struct page *page; | |
1937 | ||
1938 | for (addr = start; addr < end; addr += HPAGE_SIZE) { | |
1939 | pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK); | |
1940 | pte_t pte; | |
1941 | ||
1942 | if (!ptep) | |
1943 | continue; | |
1944 | ||
1945 | pte = *ptep; | |
1946 | if (pte_none(pte)) | |
1947 | continue; | |
1948 | ||
1949 | page = pte_page(pte); | |
1950 | if (!page) | |
1951 | continue; | |
1952 | ||
1953 | gather_stats(page, md, pte_dirty(*ptep)); | |
1954 | } | |
1955 | } | |
7f709ed0 AM |
1956 | #else |
1957 | static inline void check_huge_range(struct vm_area_struct *vma, | |
1958 | unsigned long start, unsigned long end, | |
1959 | struct numa_maps *md) | |
1960 | { | |
1961 | } | |
1962 | #endif | |
397874df | 1963 | |
1a75a6c8 CL |
1964 | int show_numa_map(struct seq_file *m, void *v) |
1965 | { | |
99f89551 | 1966 | struct proc_maps_private *priv = m->private; |
1a75a6c8 CL |
1967 | struct vm_area_struct *vma = v; |
1968 | struct numa_maps *md; | |
397874df CL |
1969 | struct file *file = vma->vm_file; |
1970 | struct mm_struct *mm = vma->vm_mm; | |
480eccf9 | 1971 | struct mempolicy *pol; |
1a75a6c8 CL |
1972 | int n; |
1973 | char buffer[50]; | |
1974 | ||
397874df | 1975 | if (!mm) |
1a75a6c8 CL |
1976 | return 0; |
1977 | ||
1978 | md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL); | |
1979 | if (!md) | |
1980 | return 0; | |
1981 | ||
480eccf9 LS |
1982 | pol = get_vma_policy(priv->task, vma, vma->vm_start); |
1983 | mpol_to_str(buffer, sizeof(buffer), pol); | |
1984 | /* | |
1985 | * unref shared or other task's mempolicy | |
1986 | */ | |
1987 | if (pol != &default_policy && pol != current->mempolicy) | |
1988 | __mpol_free(pol); | |
397874df CL |
1989 | |
1990 | seq_printf(m, "%08lx %s", vma->vm_start, buffer); | |
1991 | ||
1992 | if (file) { | |
1993 | seq_printf(m, " file="); | |
c32c2f63 | 1994 | seq_path(m, &file->f_path, "\n\t= "); |
397874df CL |
1995 | } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) { |
1996 | seq_printf(m, " heap"); | |
1997 | } else if (vma->vm_start <= mm->start_stack && | |
1998 | vma->vm_end >= mm->start_stack) { | |
1999 | seq_printf(m, " stack"); | |
2000 | } | |
2001 | ||
2002 | if (is_vm_hugetlb_page(vma)) { | |
2003 | check_huge_range(vma, vma->vm_start, vma->vm_end, md); | |
2004 | seq_printf(m, " huge"); | |
2005 | } else { | |
a57ebfdb | 2006 | check_pgd_range(vma, vma->vm_start, vma->vm_end, |
56bbd65d | 2007 | &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md); |
397874df CL |
2008 | } |
2009 | ||
2010 | if (!md->pages) | |
2011 | goto out; | |
1a75a6c8 | 2012 | |
397874df CL |
2013 | if (md->anon) |
2014 | seq_printf(m," anon=%lu",md->anon); | |
1a75a6c8 | 2015 | |
397874df CL |
2016 | if (md->dirty) |
2017 | seq_printf(m," dirty=%lu",md->dirty); | |
1a75a6c8 | 2018 | |
397874df CL |
2019 | if (md->pages != md->anon && md->pages != md->dirty) |
2020 | seq_printf(m, " mapped=%lu", md->pages); | |
1a75a6c8 | 2021 | |
397874df CL |
2022 | if (md->mapcount_max > 1) |
2023 | seq_printf(m, " mapmax=%lu", md->mapcount_max); | |
1a75a6c8 | 2024 | |
397874df CL |
2025 | if (md->swapcache) |
2026 | seq_printf(m," swapcache=%lu", md->swapcache); | |
2027 | ||
2028 | if (md->active < md->pages && !is_vm_hugetlb_page(vma)) | |
2029 | seq_printf(m," active=%lu", md->active); | |
2030 | ||
2031 | if (md->writeback) | |
2032 | seq_printf(m," writeback=%lu", md->writeback); | |
2033 | ||
56bbd65d | 2034 | for_each_node_state(n, N_HIGH_MEMORY) |
397874df CL |
2035 | if (md->node[n]) |
2036 | seq_printf(m, " N%d=%lu", n, md->node[n]); | |
2037 | out: | |
2038 | seq_putc(m, '\n'); | |
1a75a6c8 CL |
2039 | kfree(md); |
2040 | ||
2041 | if (m->count < m->size) | |
99f89551 | 2042 | m->version = (vma != priv->tail_vma) ? vma->vm_start : 0; |
1a75a6c8 CL |
2043 | return 0; |
2044 | } |