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 | ||
1da177e4 | 116 | /* Do sanity checking on a policy */ |
dfcd3c0d | 117 | static int mpol_check_policy(int mode, nodemask_t *nodes) |
1da177e4 | 118 | { |
31f1de46 KM |
119 | int was_empty, is_empty; |
120 | ||
121 | if (!nodes) | |
122 | return 0; | |
123 | ||
124 | /* | |
125 | * "Contextualize" the in-coming nodemast for cpusets: | |
126 | * Remember whether in-coming nodemask was empty, If not, | |
127 | * restrict the nodes to the allowed nodes in the cpuset. | |
128 | * This is guaranteed to be a subset of nodes with memory. | |
129 | */ | |
130 | cpuset_update_task_memory_state(); | |
131 | is_empty = was_empty = nodes_empty(*nodes); | |
132 | if (!was_empty) { | |
133 | nodes_and(*nodes, *nodes, cpuset_current_mems_allowed); | |
134 | is_empty = nodes_empty(*nodes); /* after "contextualization" */ | |
135 | } | |
1da177e4 LT |
136 | |
137 | switch (mode) { | |
138 | case MPOL_DEFAULT: | |
31f1de46 KM |
139 | /* |
140 | * require caller to specify an empty nodemask | |
141 | * before "contextualization" | |
142 | */ | |
143 | if (!was_empty) | |
1da177e4 LT |
144 | return -EINVAL; |
145 | break; | |
146 | case MPOL_BIND: | |
147 | case MPOL_INTERLEAVE: | |
31f1de46 KM |
148 | /* |
149 | * require at least 1 valid node after "contextualization" | |
150 | */ | |
151 | if (is_empty) | |
152 | return -EINVAL; | |
153 | break; | |
154 | case MPOL_PREFERRED: | |
155 | /* | |
156 | * Did caller specify invalid nodes? | |
157 | * Don't silently accept this as "local allocation". | |
158 | */ | |
159 | if (!was_empty && is_empty) | |
1da177e4 LT |
160 | return -EINVAL; |
161 | break; | |
162 | } | |
31f1de46 | 163 | return 0; |
1da177e4 | 164 | } |
dd942ae3 | 165 | |
1da177e4 | 166 | /* Generate a custom zonelist for the BIND policy. */ |
dfcd3c0d | 167 | static struct zonelist *bind_zonelist(nodemask_t *nodes) |
1da177e4 LT |
168 | { |
169 | struct zonelist *zl; | |
2f6726e5 CL |
170 | int num, max, nd; |
171 | enum zone_type k; | |
1da177e4 | 172 | |
dfcd3c0d | 173 | max = 1 + MAX_NR_ZONES * nodes_weight(*nodes); |
9276b1bc | 174 | max++; /* space for zlcache_ptr (see mmzone.h) */ |
dd942ae3 | 175 | zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL); |
1da177e4 | 176 | if (!zl) |
8af5e2eb | 177 | return ERR_PTR(-ENOMEM); |
9276b1bc | 178 | zl->zlcache_ptr = NULL; |
1da177e4 | 179 | num = 0; |
dd942ae3 AK |
180 | /* First put in the highest zones from all nodes, then all the next |
181 | lower zones etc. Avoid empty zones because the memory allocator | |
182 | doesn't like them. If you implement node hot removal you | |
183 | have to fix that. */ | |
b377fd39 | 184 | k = MAX_NR_ZONES - 1; |
2f6726e5 | 185 | while (1) { |
dd942ae3 AK |
186 | for_each_node_mask(nd, *nodes) { |
187 | struct zone *z = &NODE_DATA(nd)->node_zones[k]; | |
188 | if (z->present_pages > 0) | |
dd1a239f | 189 | zoneref_set_zone(z, &zl->_zonerefs[num++]); |
dd942ae3 | 190 | } |
2f6726e5 CL |
191 | if (k == 0) |
192 | break; | |
193 | k--; | |
dd942ae3 | 194 | } |
8af5e2eb KH |
195 | if (num == 0) { |
196 | kfree(zl); | |
197 | return ERR_PTR(-EINVAL); | |
198 | } | |
dd1a239f MG |
199 | zl->_zonerefs[num].zone = NULL; |
200 | zl->_zonerefs[num].zone_idx = 0; | |
1da177e4 LT |
201 | return zl; |
202 | } | |
203 | ||
204 | /* Create a new policy */ | |
dfcd3c0d | 205 | static struct mempolicy *mpol_new(int mode, nodemask_t *nodes) |
1da177e4 LT |
206 | { |
207 | struct mempolicy *policy; | |
208 | ||
140d5a49 PM |
209 | pr_debug("setting mode %d nodes[0] %lx\n", |
210 | mode, nodes ? nodes_addr(*nodes)[0] : -1); | |
211 | ||
1da177e4 LT |
212 | if (mode == MPOL_DEFAULT) |
213 | return NULL; | |
214 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
215 | if (!policy) | |
216 | return ERR_PTR(-ENOMEM); | |
217 | atomic_set(&policy->refcnt, 1); | |
218 | switch (mode) { | |
219 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 220 | policy->v.nodes = *nodes; |
6eaf806a | 221 | if (nodes_weight(policy->v.nodes) == 0) { |
8f493d79 AK |
222 | kmem_cache_free(policy_cache, policy); |
223 | return ERR_PTR(-EINVAL); | |
224 | } | |
1da177e4 LT |
225 | break; |
226 | case MPOL_PREFERRED: | |
dfcd3c0d | 227 | policy->v.preferred_node = first_node(*nodes); |
1da177e4 LT |
228 | if (policy->v.preferred_node >= MAX_NUMNODES) |
229 | policy->v.preferred_node = -1; | |
230 | break; | |
231 | case MPOL_BIND: | |
232 | policy->v.zonelist = bind_zonelist(nodes); | |
8af5e2eb KH |
233 | if (IS_ERR(policy->v.zonelist)) { |
234 | void *error_code = policy->v.zonelist; | |
1da177e4 | 235 | kmem_cache_free(policy_cache, policy); |
8af5e2eb | 236 | return error_code; |
1da177e4 LT |
237 | } |
238 | break; | |
239 | } | |
240 | policy->policy = mode; | |
74cb2155 | 241 | policy->cpuset_mems_allowed = cpuset_mems_allowed(current); |
1da177e4 LT |
242 | return policy; |
243 | } | |
244 | ||
397874df | 245 | static void gather_stats(struct page *, void *, int pte_dirty); |
fc301289 CL |
246 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
247 | unsigned long flags); | |
1a75a6c8 | 248 | |
38e35860 | 249 | /* Scan through pages checking if pages follow certain conditions. */ |
b5810039 | 250 | static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd, |
dc9aa5b9 CL |
251 | unsigned long addr, unsigned long end, |
252 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 253 | void *private) |
1da177e4 | 254 | { |
91612e0d HD |
255 | pte_t *orig_pte; |
256 | pte_t *pte; | |
705e87c0 | 257 | spinlock_t *ptl; |
941150a3 | 258 | |
705e87c0 | 259 | orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
91612e0d | 260 | do { |
6aab341e | 261 | struct page *page; |
25ba77c1 | 262 | int nid; |
91612e0d HD |
263 | |
264 | if (!pte_present(*pte)) | |
1da177e4 | 265 | continue; |
6aab341e LT |
266 | page = vm_normal_page(vma, addr, *pte); |
267 | if (!page) | |
1da177e4 | 268 | continue; |
053837fc NP |
269 | /* |
270 | * The check for PageReserved here is important to avoid | |
271 | * handling zero pages and other pages that may have been | |
272 | * marked special by the system. | |
273 | * | |
274 | * If the PageReserved would not be checked here then f.e. | |
275 | * the location of the zero page could have an influence | |
276 | * on MPOL_MF_STRICT, zero pages would be counted for | |
277 | * the per node stats, and there would be useless attempts | |
278 | * to put zero pages on the migration list. | |
279 | */ | |
f4598c8b CL |
280 | if (PageReserved(page)) |
281 | continue; | |
6aab341e | 282 | nid = page_to_nid(page); |
38e35860 CL |
283 | if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT)) |
284 | continue; | |
285 | ||
1a75a6c8 | 286 | if (flags & MPOL_MF_STATS) |
397874df | 287 | gather_stats(page, private, pte_dirty(*pte)); |
053837fc | 288 | else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) |
fc301289 | 289 | migrate_page_add(page, private, flags); |
38e35860 CL |
290 | else |
291 | break; | |
91612e0d | 292 | } while (pte++, addr += PAGE_SIZE, addr != end); |
705e87c0 | 293 | pte_unmap_unlock(orig_pte, ptl); |
91612e0d HD |
294 | return addr != end; |
295 | } | |
296 | ||
b5810039 | 297 | static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud, |
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 | pmd_t *pmd; | |
303 | unsigned long next; | |
304 | ||
305 | pmd = pmd_offset(pud, addr); | |
306 | do { | |
307 | next = pmd_addr_end(addr, end); | |
308 | if (pmd_none_or_clear_bad(pmd)) | |
309 | continue; | |
dc9aa5b9 | 310 | if (check_pte_range(vma, pmd, addr, next, nodes, |
38e35860 | 311 | flags, private)) |
91612e0d HD |
312 | return -EIO; |
313 | } while (pmd++, addr = next, addr != end); | |
314 | return 0; | |
315 | } | |
316 | ||
b5810039 | 317 | static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd, |
dc9aa5b9 CL |
318 | unsigned long addr, unsigned long end, |
319 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 320 | void *private) |
91612e0d HD |
321 | { |
322 | pud_t *pud; | |
323 | unsigned long next; | |
324 | ||
325 | pud = pud_offset(pgd, addr); | |
326 | do { | |
327 | next = pud_addr_end(addr, end); | |
328 | if (pud_none_or_clear_bad(pud)) | |
329 | continue; | |
dc9aa5b9 | 330 | if (check_pmd_range(vma, pud, addr, next, nodes, |
38e35860 | 331 | flags, private)) |
91612e0d HD |
332 | return -EIO; |
333 | } while (pud++, addr = next, addr != end); | |
334 | return 0; | |
335 | } | |
336 | ||
b5810039 | 337 | static inline int check_pgd_range(struct vm_area_struct *vma, |
dc9aa5b9 CL |
338 | unsigned long addr, unsigned long end, |
339 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 340 | void *private) |
91612e0d HD |
341 | { |
342 | pgd_t *pgd; | |
343 | unsigned long next; | |
344 | ||
b5810039 | 345 | pgd = pgd_offset(vma->vm_mm, addr); |
91612e0d HD |
346 | do { |
347 | next = pgd_addr_end(addr, end); | |
348 | if (pgd_none_or_clear_bad(pgd)) | |
349 | continue; | |
dc9aa5b9 | 350 | if (check_pud_range(vma, pgd, addr, next, nodes, |
38e35860 | 351 | flags, private)) |
91612e0d HD |
352 | return -EIO; |
353 | } while (pgd++, addr = next, addr != end); | |
354 | return 0; | |
1da177e4 LT |
355 | } |
356 | ||
dc9aa5b9 CL |
357 | /* |
358 | * Check if all pages in a range are on a set of nodes. | |
359 | * If pagelist != NULL then isolate pages from the LRU and | |
360 | * put them on the pagelist. | |
361 | */ | |
1da177e4 LT |
362 | static struct vm_area_struct * |
363 | check_range(struct mm_struct *mm, unsigned long start, unsigned long end, | |
38e35860 | 364 | const nodemask_t *nodes, unsigned long flags, void *private) |
1da177e4 LT |
365 | { |
366 | int err; | |
367 | struct vm_area_struct *first, *vma, *prev; | |
368 | ||
90036ee5 | 369 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
90036ee5 | 370 | |
b20a3503 CL |
371 | err = migrate_prep(); |
372 | if (err) | |
373 | return ERR_PTR(err); | |
90036ee5 | 374 | } |
053837fc | 375 | |
1da177e4 LT |
376 | first = find_vma(mm, start); |
377 | if (!first) | |
378 | return ERR_PTR(-EFAULT); | |
379 | prev = NULL; | |
380 | for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) { | |
dc9aa5b9 CL |
381 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { |
382 | if (!vma->vm_next && vma->vm_end < end) | |
383 | return ERR_PTR(-EFAULT); | |
384 | if (prev && prev->vm_end < vma->vm_start) | |
385 | return ERR_PTR(-EFAULT); | |
386 | } | |
387 | if (!is_vm_hugetlb_page(vma) && | |
388 | ((flags & MPOL_MF_STRICT) || | |
389 | ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) && | |
390 | vma_migratable(vma)))) { | |
5b952b3c | 391 | unsigned long endvma = vma->vm_end; |
dc9aa5b9 | 392 | |
5b952b3c AK |
393 | if (endvma > end) |
394 | endvma = end; | |
395 | if (vma->vm_start > start) | |
396 | start = vma->vm_start; | |
dc9aa5b9 | 397 | err = check_pgd_range(vma, start, endvma, nodes, |
38e35860 | 398 | flags, private); |
1da177e4 LT |
399 | if (err) { |
400 | first = ERR_PTR(err); | |
401 | break; | |
402 | } | |
403 | } | |
404 | prev = vma; | |
405 | } | |
406 | return first; | |
407 | } | |
408 | ||
409 | /* Apply policy to a single VMA */ | |
410 | static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new) | |
411 | { | |
412 | int err = 0; | |
413 | struct mempolicy *old = vma->vm_policy; | |
414 | ||
140d5a49 | 415 | pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", |
1da177e4 LT |
416 | vma->vm_start, vma->vm_end, vma->vm_pgoff, |
417 | vma->vm_ops, vma->vm_file, | |
418 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
419 | ||
420 | if (vma->vm_ops && vma->vm_ops->set_policy) | |
421 | err = vma->vm_ops->set_policy(vma, new); | |
422 | if (!err) { | |
423 | mpol_get(new); | |
424 | vma->vm_policy = new; | |
425 | mpol_free(old); | |
426 | } | |
427 | return err; | |
428 | } | |
429 | ||
430 | /* Step 2: apply policy to a range and do splits. */ | |
431 | static int mbind_range(struct vm_area_struct *vma, unsigned long start, | |
432 | unsigned long end, struct mempolicy *new) | |
433 | { | |
434 | struct vm_area_struct *next; | |
435 | int err; | |
436 | ||
437 | err = 0; | |
438 | for (; vma && vma->vm_start < end; vma = next) { | |
439 | next = vma->vm_next; | |
440 | if (vma->vm_start < start) | |
441 | err = split_vma(vma->vm_mm, vma, start, 1); | |
442 | if (!err && vma->vm_end > end) | |
443 | err = split_vma(vma->vm_mm, vma, end, 0); | |
444 | if (!err) | |
445 | err = policy_vma(vma, new); | |
446 | if (err) | |
447 | break; | |
448 | } | |
449 | return err; | |
450 | } | |
451 | ||
c61afb18 PJ |
452 | /* |
453 | * Update task->flags PF_MEMPOLICY bit: set iff non-default | |
454 | * mempolicy. Allows more rapid checking of this (combined perhaps | |
455 | * with other PF_* flag bits) on memory allocation hot code paths. | |
456 | * | |
457 | * If called from outside this file, the task 'p' should -only- be | |
458 | * a newly forked child not yet visible on the task list, because | |
459 | * manipulating the task flags of a visible task is not safe. | |
460 | * | |
461 | * The above limitation is why this routine has the funny name | |
462 | * mpol_fix_fork_child_flag(). | |
463 | * | |
464 | * It is also safe to call this with a task pointer of current, | |
465 | * which the static wrapper mpol_set_task_struct_flag() does, | |
466 | * for use within this file. | |
467 | */ | |
468 | ||
469 | void mpol_fix_fork_child_flag(struct task_struct *p) | |
470 | { | |
471 | if (p->mempolicy) | |
472 | p->flags |= PF_MEMPOLICY; | |
473 | else | |
474 | p->flags &= ~PF_MEMPOLICY; | |
475 | } | |
476 | ||
477 | static void mpol_set_task_struct_flag(void) | |
478 | { | |
479 | mpol_fix_fork_child_flag(current); | |
480 | } | |
481 | ||
1da177e4 | 482 | /* Set the process memory policy */ |
dbcb0f19 | 483 | static long do_set_mempolicy(int mode, nodemask_t *nodes) |
1da177e4 | 484 | { |
1da177e4 | 485 | struct mempolicy *new; |
1da177e4 | 486 | |
31f1de46 | 487 | if (mpol_check_policy(mode, nodes)) |
1da177e4 | 488 | return -EINVAL; |
8bccd85f | 489 | new = mpol_new(mode, nodes); |
1da177e4 LT |
490 | if (IS_ERR(new)) |
491 | return PTR_ERR(new); | |
492 | mpol_free(current->mempolicy); | |
493 | current->mempolicy = new; | |
c61afb18 | 494 | mpol_set_task_struct_flag(); |
1da177e4 | 495 | if (new && new->policy == MPOL_INTERLEAVE) |
dfcd3c0d | 496 | current->il_next = first_node(new->v.nodes); |
1da177e4 LT |
497 | return 0; |
498 | } | |
499 | ||
500 | /* Fill a zone bitmap for a policy */ | |
dfcd3c0d | 501 | static void get_zonemask(struct mempolicy *p, nodemask_t *nodes) |
1da177e4 LT |
502 | { |
503 | int i; | |
504 | ||
dfcd3c0d | 505 | nodes_clear(*nodes); |
1da177e4 LT |
506 | switch (p->policy) { |
507 | case MPOL_BIND: | |
dd1a239f MG |
508 | for (i = 0; p->v.zonelist->_zonerefs[i].zone; i++) { |
509 | struct zoneref *zref; | |
510 | zref = &p->v.zonelist->_zonerefs[i]; | |
511 | node_set(zonelist_node_idx(zref), *nodes); | |
512 | } | |
1da177e4 LT |
513 | break; |
514 | case MPOL_DEFAULT: | |
515 | break; | |
516 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 517 | *nodes = p->v.nodes; |
1da177e4 LT |
518 | break; |
519 | case MPOL_PREFERRED: | |
56bbd65d | 520 | /* or use current node instead of memory_map? */ |
1da177e4 | 521 | if (p->v.preferred_node < 0) |
56bbd65d | 522 | *nodes = node_states[N_HIGH_MEMORY]; |
1da177e4 | 523 | else |
dfcd3c0d | 524 | node_set(p->v.preferred_node, *nodes); |
1da177e4 LT |
525 | break; |
526 | default: | |
527 | BUG(); | |
528 | } | |
529 | } | |
530 | ||
531 | static int lookup_node(struct mm_struct *mm, unsigned long addr) | |
532 | { | |
533 | struct page *p; | |
534 | int err; | |
535 | ||
536 | err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL); | |
537 | if (err >= 0) { | |
538 | err = page_to_nid(p); | |
539 | put_page(p); | |
540 | } | |
541 | return err; | |
542 | } | |
543 | ||
1da177e4 | 544 | /* Retrieve NUMA policy */ |
dbcb0f19 AB |
545 | static long do_get_mempolicy(int *policy, nodemask_t *nmask, |
546 | unsigned long addr, unsigned long flags) | |
1da177e4 | 547 | { |
8bccd85f | 548 | int err; |
1da177e4 LT |
549 | struct mm_struct *mm = current->mm; |
550 | struct vm_area_struct *vma = NULL; | |
551 | struct mempolicy *pol = current->mempolicy; | |
552 | ||
cf2a473c | 553 | cpuset_update_task_memory_state(); |
754af6f5 LS |
554 | if (flags & |
555 | ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED)) | |
1da177e4 | 556 | return -EINVAL; |
754af6f5 LS |
557 | |
558 | if (flags & MPOL_F_MEMS_ALLOWED) { | |
559 | if (flags & (MPOL_F_NODE|MPOL_F_ADDR)) | |
560 | return -EINVAL; | |
561 | *policy = 0; /* just so it's initialized */ | |
562 | *nmask = cpuset_current_mems_allowed; | |
563 | return 0; | |
564 | } | |
565 | ||
1da177e4 LT |
566 | if (flags & MPOL_F_ADDR) { |
567 | down_read(&mm->mmap_sem); | |
568 | vma = find_vma_intersection(mm, addr, addr+1); | |
569 | if (!vma) { | |
570 | up_read(&mm->mmap_sem); | |
571 | return -EFAULT; | |
572 | } | |
573 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
574 | pol = vma->vm_ops->get_policy(vma, addr); | |
575 | else | |
576 | pol = vma->vm_policy; | |
577 | } else if (addr) | |
578 | return -EINVAL; | |
579 | ||
580 | if (!pol) | |
581 | pol = &default_policy; | |
582 | ||
583 | if (flags & MPOL_F_NODE) { | |
584 | if (flags & MPOL_F_ADDR) { | |
585 | err = lookup_node(mm, addr); | |
586 | if (err < 0) | |
587 | goto out; | |
8bccd85f | 588 | *policy = err; |
1da177e4 LT |
589 | } else if (pol == current->mempolicy && |
590 | pol->policy == MPOL_INTERLEAVE) { | |
8bccd85f | 591 | *policy = current->il_next; |
1da177e4 LT |
592 | } else { |
593 | err = -EINVAL; | |
594 | goto out; | |
595 | } | |
596 | } else | |
8bccd85f | 597 | *policy = pol->policy; |
1da177e4 LT |
598 | |
599 | if (vma) { | |
600 | up_read(¤t->mm->mmap_sem); | |
601 | vma = NULL; | |
602 | } | |
603 | ||
1da177e4 | 604 | err = 0; |
8bccd85f CL |
605 | if (nmask) |
606 | get_zonemask(pol, nmask); | |
1da177e4 LT |
607 | |
608 | out: | |
609 | if (vma) | |
610 | up_read(¤t->mm->mmap_sem); | |
611 | return err; | |
612 | } | |
613 | ||
b20a3503 | 614 | #ifdef CONFIG_MIGRATION |
6ce3c4c0 CL |
615 | /* |
616 | * page migration | |
617 | */ | |
fc301289 CL |
618 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
619 | unsigned long flags) | |
6ce3c4c0 CL |
620 | { |
621 | /* | |
fc301289 | 622 | * Avoid migrating a page that is shared with others. |
6ce3c4c0 | 623 | */ |
b20a3503 CL |
624 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) |
625 | isolate_lru_page(page, pagelist); | |
7e2ab150 | 626 | } |
6ce3c4c0 | 627 | |
742755a1 | 628 | static struct page *new_node_page(struct page *page, unsigned long node, int **x) |
95a402c3 | 629 | { |
769848c0 | 630 | return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0); |
95a402c3 CL |
631 | } |
632 | ||
7e2ab150 CL |
633 | /* |
634 | * Migrate pages from one node to a target node. | |
635 | * Returns error or the number of pages not migrated. | |
636 | */ | |
dbcb0f19 AB |
637 | static int migrate_to_node(struct mm_struct *mm, int source, int dest, |
638 | int flags) | |
7e2ab150 CL |
639 | { |
640 | nodemask_t nmask; | |
641 | LIST_HEAD(pagelist); | |
642 | int err = 0; | |
643 | ||
644 | nodes_clear(nmask); | |
645 | node_set(source, nmask); | |
6ce3c4c0 | 646 | |
7e2ab150 CL |
647 | check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask, |
648 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); | |
649 | ||
aaa994b3 | 650 | if (!list_empty(&pagelist)) |
95a402c3 CL |
651 | err = migrate_pages(&pagelist, new_node_page, dest); |
652 | ||
7e2ab150 | 653 | return err; |
6ce3c4c0 CL |
654 | } |
655 | ||
39743889 | 656 | /* |
7e2ab150 CL |
657 | * Move pages between the two nodesets so as to preserve the physical |
658 | * layout as much as possible. | |
39743889 CL |
659 | * |
660 | * Returns the number of page that could not be moved. | |
661 | */ | |
662 | int do_migrate_pages(struct mm_struct *mm, | |
663 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
664 | { | |
665 | LIST_HEAD(pagelist); | |
7e2ab150 CL |
666 | int busy = 0; |
667 | int err = 0; | |
668 | nodemask_t tmp; | |
39743889 | 669 | |
7e2ab150 | 670 | down_read(&mm->mmap_sem); |
39743889 | 671 | |
7b2259b3 CL |
672 | err = migrate_vmas(mm, from_nodes, to_nodes, flags); |
673 | if (err) | |
674 | goto out; | |
675 | ||
7e2ab150 CL |
676 | /* |
677 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' | |
678 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' | |
679 | * bit in 'tmp', and return that <source, dest> pair for migration. | |
680 | * The pair of nodemasks 'to' and 'from' define the map. | |
681 | * | |
682 | * If no pair of bits is found that way, fallback to picking some | |
683 | * pair of 'source' and 'dest' bits that are not the same. If the | |
684 | * 'source' and 'dest' bits are the same, this represents a node | |
685 | * that will be migrating to itself, so no pages need move. | |
686 | * | |
687 | * If no bits are left in 'tmp', or if all remaining bits left | |
688 | * in 'tmp' correspond to the same bit in 'to', return false | |
689 | * (nothing left to migrate). | |
690 | * | |
691 | * This lets us pick a pair of nodes to migrate between, such that | |
692 | * if possible the dest node is not already occupied by some other | |
693 | * source node, minimizing the risk of overloading the memory on a | |
694 | * node that would happen if we migrated incoming memory to a node | |
695 | * before migrating outgoing memory source that same node. | |
696 | * | |
697 | * A single scan of tmp is sufficient. As we go, we remember the | |
698 | * most recent <s, d> pair that moved (s != d). If we find a pair | |
699 | * that not only moved, but what's better, moved to an empty slot | |
700 | * (d is not set in tmp), then we break out then, with that pair. | |
701 | * Otherwise when we finish scannng from_tmp, we at least have the | |
702 | * most recent <s, d> pair that moved. If we get all the way through | |
703 | * the scan of tmp without finding any node that moved, much less | |
704 | * moved to an empty node, then there is nothing left worth migrating. | |
705 | */ | |
d4984711 | 706 | |
7e2ab150 CL |
707 | tmp = *from_nodes; |
708 | while (!nodes_empty(tmp)) { | |
709 | int s,d; | |
710 | int source = -1; | |
711 | int dest = 0; | |
712 | ||
713 | for_each_node_mask(s, tmp) { | |
714 | d = node_remap(s, *from_nodes, *to_nodes); | |
715 | if (s == d) | |
716 | continue; | |
717 | ||
718 | source = s; /* Node moved. Memorize */ | |
719 | dest = d; | |
720 | ||
721 | /* dest not in remaining from nodes? */ | |
722 | if (!node_isset(dest, tmp)) | |
723 | break; | |
724 | } | |
725 | if (source == -1) | |
726 | break; | |
727 | ||
728 | node_clear(source, tmp); | |
729 | err = migrate_to_node(mm, source, dest, flags); | |
730 | if (err > 0) | |
731 | busy += err; | |
732 | if (err < 0) | |
733 | break; | |
39743889 | 734 | } |
7b2259b3 | 735 | out: |
39743889 | 736 | up_read(&mm->mmap_sem); |
7e2ab150 CL |
737 | if (err < 0) |
738 | return err; | |
739 | return busy; | |
b20a3503 CL |
740 | |
741 | } | |
742 | ||
3ad33b24 LS |
743 | /* |
744 | * Allocate a new page for page migration based on vma policy. | |
745 | * Start assuming that page is mapped by vma pointed to by @private. | |
746 | * Search forward from there, if not. N.B., this assumes that the | |
747 | * list of pages handed to migrate_pages()--which is how we get here-- | |
748 | * is in virtual address order. | |
749 | */ | |
742755a1 | 750 | static struct page *new_vma_page(struct page *page, unsigned long private, int **x) |
95a402c3 CL |
751 | { |
752 | struct vm_area_struct *vma = (struct vm_area_struct *)private; | |
3ad33b24 | 753 | unsigned long uninitialized_var(address); |
95a402c3 | 754 | |
3ad33b24 LS |
755 | while (vma) { |
756 | address = page_address_in_vma(page, vma); | |
757 | if (address != -EFAULT) | |
758 | break; | |
759 | vma = vma->vm_next; | |
760 | } | |
761 | ||
762 | /* | |
763 | * if !vma, alloc_page_vma() will use task or system default policy | |
764 | */ | |
765 | return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); | |
95a402c3 | 766 | } |
b20a3503 CL |
767 | #else |
768 | ||
769 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
770 | unsigned long flags) | |
771 | { | |
39743889 CL |
772 | } |
773 | ||
b20a3503 CL |
774 | int do_migrate_pages(struct mm_struct *mm, |
775 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
776 | { | |
777 | return -ENOSYS; | |
778 | } | |
95a402c3 | 779 | |
69939749 | 780 | static struct page *new_vma_page(struct page *page, unsigned long private, int **x) |
95a402c3 CL |
781 | { |
782 | return NULL; | |
783 | } | |
b20a3503 CL |
784 | #endif |
785 | ||
dbcb0f19 AB |
786 | static long do_mbind(unsigned long start, unsigned long len, |
787 | unsigned long mode, nodemask_t *nmask, | |
788 | unsigned long flags) | |
6ce3c4c0 CL |
789 | { |
790 | struct vm_area_struct *vma; | |
791 | struct mm_struct *mm = current->mm; | |
792 | struct mempolicy *new; | |
793 | unsigned long end; | |
794 | int err; | |
795 | LIST_HEAD(pagelist); | |
796 | ||
797 | if ((flags & ~(unsigned long)(MPOL_MF_STRICT | | |
798 | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
799 | || mode > MPOL_MAX) | |
800 | return -EINVAL; | |
74c00241 | 801 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
6ce3c4c0 CL |
802 | return -EPERM; |
803 | ||
804 | if (start & ~PAGE_MASK) | |
805 | return -EINVAL; | |
806 | ||
807 | if (mode == MPOL_DEFAULT) | |
808 | flags &= ~MPOL_MF_STRICT; | |
809 | ||
810 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; | |
811 | end = start + len; | |
812 | ||
813 | if (end < start) | |
814 | return -EINVAL; | |
815 | if (end == start) | |
816 | return 0; | |
817 | ||
818 | if (mpol_check_policy(mode, nmask)) | |
819 | return -EINVAL; | |
820 | ||
821 | new = mpol_new(mode, nmask); | |
822 | if (IS_ERR(new)) | |
823 | return PTR_ERR(new); | |
824 | ||
825 | /* | |
826 | * If we are using the default policy then operation | |
827 | * on discontinuous address spaces is okay after all | |
828 | */ | |
829 | if (!new) | |
830 | flags |= MPOL_MF_DISCONTIG_OK; | |
831 | ||
140d5a49 PM |
832 | pr_debug("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len, |
833 | mode, nmask ? nodes_addr(*nmask)[0] : -1); | |
6ce3c4c0 CL |
834 | |
835 | down_write(&mm->mmap_sem); | |
836 | vma = check_range(mm, start, end, nmask, | |
837 | flags | MPOL_MF_INVERT, &pagelist); | |
838 | ||
839 | err = PTR_ERR(vma); | |
840 | if (!IS_ERR(vma)) { | |
841 | int nr_failed = 0; | |
842 | ||
843 | err = mbind_range(vma, start, end, new); | |
7e2ab150 | 844 | |
6ce3c4c0 | 845 | if (!list_empty(&pagelist)) |
95a402c3 CL |
846 | nr_failed = migrate_pages(&pagelist, new_vma_page, |
847 | (unsigned long)vma); | |
6ce3c4c0 CL |
848 | |
849 | if (!err && nr_failed && (flags & MPOL_MF_STRICT)) | |
850 | err = -EIO; | |
851 | } | |
b20a3503 | 852 | |
6ce3c4c0 CL |
853 | up_write(&mm->mmap_sem); |
854 | mpol_free(new); | |
855 | return err; | |
856 | } | |
857 | ||
8bccd85f CL |
858 | /* |
859 | * User space interface with variable sized bitmaps for nodelists. | |
860 | */ | |
861 | ||
862 | /* Copy a node mask from user space. */ | |
39743889 | 863 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
8bccd85f CL |
864 | unsigned long maxnode) |
865 | { | |
866 | unsigned long k; | |
867 | unsigned long nlongs; | |
868 | unsigned long endmask; | |
869 | ||
870 | --maxnode; | |
871 | nodes_clear(*nodes); | |
872 | if (maxnode == 0 || !nmask) | |
873 | return 0; | |
a9c930ba | 874 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
636f13c1 | 875 | return -EINVAL; |
8bccd85f CL |
876 | |
877 | nlongs = BITS_TO_LONGS(maxnode); | |
878 | if ((maxnode % BITS_PER_LONG) == 0) | |
879 | endmask = ~0UL; | |
880 | else | |
881 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
882 | ||
883 | /* When the user specified more nodes than supported just check | |
884 | if the non supported part is all zero. */ | |
885 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
886 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
887 | return -EINVAL; | |
888 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
889 | unsigned long t; | |
890 | if (get_user(t, nmask + k)) | |
891 | return -EFAULT; | |
892 | if (k == nlongs - 1) { | |
893 | if (t & endmask) | |
894 | return -EINVAL; | |
895 | } else if (t) | |
896 | return -EINVAL; | |
897 | } | |
898 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
899 | endmask = ~0UL; | |
900 | } | |
901 | ||
902 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
903 | return -EFAULT; | |
904 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
905 | return 0; | |
906 | } | |
907 | ||
908 | /* Copy a kernel node mask to user space */ | |
909 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
910 | nodemask_t *nodes) | |
911 | { | |
912 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
913 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
914 | ||
915 | if (copy > nbytes) { | |
916 | if (copy > PAGE_SIZE) | |
917 | return -EINVAL; | |
918 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
919 | return -EFAULT; | |
920 | copy = nbytes; | |
921 | } | |
922 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
923 | } | |
924 | ||
925 | asmlinkage long sys_mbind(unsigned long start, unsigned long len, | |
926 | unsigned long mode, | |
927 | unsigned long __user *nmask, unsigned long maxnode, | |
928 | unsigned flags) | |
929 | { | |
930 | nodemask_t nodes; | |
931 | int err; | |
932 | ||
933 | err = get_nodes(&nodes, nmask, maxnode); | |
934 | if (err) | |
935 | return err; | |
936 | return do_mbind(start, len, mode, &nodes, flags); | |
937 | } | |
938 | ||
939 | /* Set the process memory policy */ | |
940 | asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask, | |
941 | unsigned long maxnode) | |
942 | { | |
943 | int err; | |
944 | nodemask_t nodes; | |
945 | ||
946 | if (mode < 0 || mode > MPOL_MAX) | |
947 | return -EINVAL; | |
948 | err = get_nodes(&nodes, nmask, maxnode); | |
949 | if (err) | |
950 | return err; | |
951 | return do_set_mempolicy(mode, &nodes); | |
952 | } | |
953 | ||
39743889 CL |
954 | asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode, |
955 | const unsigned long __user *old_nodes, | |
956 | const unsigned long __user *new_nodes) | |
957 | { | |
958 | struct mm_struct *mm; | |
959 | struct task_struct *task; | |
960 | nodemask_t old; | |
961 | nodemask_t new; | |
962 | nodemask_t task_nodes; | |
963 | int err; | |
964 | ||
965 | err = get_nodes(&old, old_nodes, maxnode); | |
966 | if (err) | |
967 | return err; | |
968 | ||
969 | err = get_nodes(&new, new_nodes, maxnode); | |
970 | if (err) | |
971 | return err; | |
972 | ||
973 | /* Find the mm_struct */ | |
974 | read_lock(&tasklist_lock); | |
228ebcbe | 975 | task = pid ? find_task_by_vpid(pid) : current; |
39743889 CL |
976 | if (!task) { |
977 | read_unlock(&tasklist_lock); | |
978 | return -ESRCH; | |
979 | } | |
980 | mm = get_task_mm(task); | |
981 | read_unlock(&tasklist_lock); | |
982 | ||
983 | if (!mm) | |
984 | return -EINVAL; | |
985 | ||
986 | /* | |
987 | * Check if this process has the right to modify the specified | |
988 | * process. The right exists if the process has administrative | |
7f927fcc | 989 | * capabilities, superuser privileges or the same |
39743889 CL |
990 | * userid as the target process. |
991 | */ | |
992 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
993 | (current->uid != task->suid) && (current->uid != task->uid) && | |
74c00241 | 994 | !capable(CAP_SYS_NICE)) { |
39743889 CL |
995 | err = -EPERM; |
996 | goto out; | |
997 | } | |
998 | ||
999 | task_nodes = cpuset_mems_allowed(task); | |
1000 | /* Is the user allowed to access the target nodes? */ | |
74c00241 | 1001 | if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) { |
39743889 CL |
1002 | err = -EPERM; |
1003 | goto out; | |
1004 | } | |
1005 | ||
37b07e41 | 1006 | if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) { |
3b42d28b CL |
1007 | err = -EINVAL; |
1008 | goto out; | |
1009 | } | |
1010 | ||
86c3a764 DQ |
1011 | err = security_task_movememory(task); |
1012 | if (err) | |
1013 | goto out; | |
1014 | ||
511030bc | 1015 | err = do_migrate_pages(mm, &old, &new, |
74c00241 | 1016 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); |
39743889 CL |
1017 | out: |
1018 | mmput(mm); | |
1019 | return err; | |
1020 | } | |
1021 | ||
1022 | ||
8bccd85f CL |
1023 | /* Retrieve NUMA policy */ |
1024 | asmlinkage long sys_get_mempolicy(int __user *policy, | |
1025 | unsigned long __user *nmask, | |
1026 | unsigned long maxnode, | |
1027 | unsigned long addr, unsigned long flags) | |
1028 | { | |
dbcb0f19 AB |
1029 | int err; |
1030 | int uninitialized_var(pval); | |
8bccd85f CL |
1031 | nodemask_t nodes; |
1032 | ||
1033 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
1034 | return -EINVAL; | |
1035 | ||
1036 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
1037 | ||
1038 | if (err) | |
1039 | return err; | |
1040 | ||
1041 | if (policy && put_user(pval, policy)) | |
1042 | return -EFAULT; | |
1043 | ||
1044 | if (nmask) | |
1045 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
1046 | ||
1047 | return err; | |
1048 | } | |
1049 | ||
1da177e4 LT |
1050 | #ifdef CONFIG_COMPAT |
1051 | ||
1052 | asmlinkage long compat_sys_get_mempolicy(int __user *policy, | |
1053 | compat_ulong_t __user *nmask, | |
1054 | compat_ulong_t maxnode, | |
1055 | compat_ulong_t addr, compat_ulong_t flags) | |
1056 | { | |
1057 | long err; | |
1058 | unsigned long __user *nm = NULL; | |
1059 | unsigned long nr_bits, alloc_size; | |
1060 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1061 | ||
1062 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1063 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1064 | ||
1065 | if (nmask) | |
1066 | nm = compat_alloc_user_space(alloc_size); | |
1067 | ||
1068 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
1069 | ||
1070 | if (!err && nmask) { | |
1071 | err = copy_from_user(bm, nm, alloc_size); | |
1072 | /* ensure entire bitmap is zeroed */ | |
1073 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
1074 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
1075 | } | |
1076 | ||
1077 | return err; | |
1078 | } | |
1079 | ||
1080 | asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask, | |
1081 | compat_ulong_t maxnode) | |
1082 | { | |
1083 | long err = 0; | |
1084 | unsigned long __user *nm = NULL; | |
1085 | unsigned long nr_bits, alloc_size; | |
1086 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1087 | ||
1088 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1089 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1090 | ||
1091 | if (nmask) { | |
1092 | err = compat_get_bitmap(bm, nmask, nr_bits); | |
1093 | nm = compat_alloc_user_space(alloc_size); | |
1094 | err |= copy_to_user(nm, bm, alloc_size); | |
1095 | } | |
1096 | ||
1097 | if (err) | |
1098 | return -EFAULT; | |
1099 | ||
1100 | return sys_set_mempolicy(mode, nm, nr_bits+1); | |
1101 | } | |
1102 | ||
1103 | asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len, | |
1104 | compat_ulong_t mode, compat_ulong_t __user *nmask, | |
1105 | compat_ulong_t maxnode, compat_ulong_t flags) | |
1106 | { | |
1107 | long err = 0; | |
1108 | unsigned long __user *nm = NULL; | |
1109 | unsigned long nr_bits, alloc_size; | |
dfcd3c0d | 1110 | nodemask_t bm; |
1da177e4 LT |
1111 | |
1112 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1113 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1114 | ||
1115 | if (nmask) { | |
dfcd3c0d | 1116 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); |
1da177e4 | 1117 | nm = compat_alloc_user_space(alloc_size); |
dfcd3c0d | 1118 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); |
1da177e4 LT |
1119 | } |
1120 | ||
1121 | if (err) | |
1122 | return -EFAULT; | |
1123 | ||
1124 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); | |
1125 | } | |
1126 | ||
1127 | #endif | |
1128 | ||
480eccf9 LS |
1129 | /* |
1130 | * get_vma_policy(@task, @vma, @addr) | |
1131 | * @task - task for fallback if vma policy == default | |
1132 | * @vma - virtual memory area whose policy is sought | |
1133 | * @addr - address in @vma for shared policy lookup | |
1134 | * | |
1135 | * Returns effective policy for a VMA at specified address. | |
1136 | * Falls back to @task or system default policy, as necessary. | |
1137 | * Returned policy has extra reference count if shared, vma, | |
1138 | * or some other task's policy [show_numa_maps() can pass | |
1139 | * @task != current]. It is the caller's responsibility to | |
1140 | * free the reference in these cases. | |
1141 | */ | |
48fce342 CL |
1142 | static struct mempolicy * get_vma_policy(struct task_struct *task, |
1143 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 1144 | { |
6e21c8f1 | 1145 | struct mempolicy *pol = task->mempolicy; |
480eccf9 | 1146 | int shared_pol = 0; |
1da177e4 LT |
1147 | |
1148 | if (vma) { | |
480eccf9 | 1149 | if (vma->vm_ops && vma->vm_ops->get_policy) { |
8bccd85f | 1150 | pol = vma->vm_ops->get_policy(vma, addr); |
480eccf9 LS |
1151 | shared_pol = 1; /* if pol non-NULL, add ref below */ |
1152 | } else if (vma->vm_policy && | |
1da177e4 LT |
1153 | vma->vm_policy->policy != MPOL_DEFAULT) |
1154 | pol = vma->vm_policy; | |
1155 | } | |
1156 | if (!pol) | |
1157 | pol = &default_policy; | |
480eccf9 LS |
1158 | else if (!shared_pol && pol != current->mempolicy) |
1159 | mpol_get(pol); /* vma or other task's policy */ | |
1da177e4 LT |
1160 | return pol; |
1161 | } | |
1162 | ||
1163 | /* Return a zonelist representing a mempolicy */ | |
dd0fc66f | 1164 | static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy) |
1da177e4 LT |
1165 | { |
1166 | int nd; | |
1167 | ||
1168 | switch (policy->policy) { | |
1169 | case MPOL_PREFERRED: | |
1170 | nd = policy->v.preferred_node; | |
1171 | if (nd < 0) | |
1172 | nd = numa_node_id(); | |
1173 | break; | |
1174 | case MPOL_BIND: | |
1175 | /* Lower zones don't get a policy applied */ | |
1176 | /* Careful: current->mems_allowed might have moved */ | |
19655d34 | 1177 | if (gfp_zone(gfp) >= policy_zone) |
1da177e4 LT |
1178 | if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist)) |
1179 | return policy->v.zonelist; | |
1180 | /*FALL THROUGH*/ | |
1181 | case MPOL_INTERLEAVE: /* should not happen */ | |
1182 | case MPOL_DEFAULT: | |
1183 | nd = numa_node_id(); | |
1184 | break; | |
1185 | default: | |
1186 | nd = 0; | |
1187 | BUG(); | |
1188 | } | |
0e88460d | 1189 | return node_zonelist(nd, gfp); |
1da177e4 LT |
1190 | } |
1191 | ||
1192 | /* Do dynamic interleaving for a process */ | |
1193 | static unsigned interleave_nodes(struct mempolicy *policy) | |
1194 | { | |
1195 | unsigned nid, next; | |
1196 | struct task_struct *me = current; | |
1197 | ||
1198 | nid = me->il_next; | |
dfcd3c0d | 1199 | next = next_node(nid, policy->v.nodes); |
1da177e4 | 1200 | if (next >= MAX_NUMNODES) |
dfcd3c0d | 1201 | next = first_node(policy->v.nodes); |
1da177e4 LT |
1202 | me->il_next = next; |
1203 | return nid; | |
1204 | } | |
1205 | ||
dc85da15 CL |
1206 | /* |
1207 | * Depending on the memory policy provide a node from which to allocate the | |
1208 | * next slab entry. | |
1209 | */ | |
1210 | unsigned slab_node(struct mempolicy *policy) | |
1211 | { | |
765c4507 CL |
1212 | int pol = policy ? policy->policy : MPOL_DEFAULT; |
1213 | ||
1214 | switch (pol) { | |
dc85da15 CL |
1215 | case MPOL_INTERLEAVE: |
1216 | return interleave_nodes(policy); | |
1217 | ||
dd1a239f | 1218 | case MPOL_BIND: { |
dc85da15 CL |
1219 | /* |
1220 | * Follow bind policy behavior and start allocation at the | |
1221 | * first node. | |
1222 | */ | |
dd1a239f MG |
1223 | return zonelist_node_idx(policy->v.zonelist->_zonerefs); |
1224 | } | |
dc85da15 CL |
1225 | |
1226 | case MPOL_PREFERRED: | |
1227 | if (policy->v.preferred_node >= 0) | |
1228 | return policy->v.preferred_node; | |
1229 | /* Fall through */ | |
1230 | ||
1231 | default: | |
1232 | return numa_node_id(); | |
1233 | } | |
1234 | } | |
1235 | ||
1da177e4 LT |
1236 | /* Do static interleaving for a VMA with known offset. */ |
1237 | static unsigned offset_il_node(struct mempolicy *pol, | |
1238 | struct vm_area_struct *vma, unsigned long off) | |
1239 | { | |
dfcd3c0d | 1240 | unsigned nnodes = nodes_weight(pol->v.nodes); |
1da177e4 LT |
1241 | unsigned target = (unsigned)off % nnodes; |
1242 | int c; | |
1243 | int nid = -1; | |
1244 | ||
1245 | c = 0; | |
1246 | do { | |
dfcd3c0d | 1247 | nid = next_node(nid, pol->v.nodes); |
1da177e4 LT |
1248 | c++; |
1249 | } while (c <= target); | |
1da177e4 LT |
1250 | return nid; |
1251 | } | |
1252 | ||
5da7ca86 CL |
1253 | /* Determine a node number for interleave */ |
1254 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1255 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1256 | { | |
1257 | if (vma) { | |
1258 | unsigned long off; | |
1259 | ||
3b98b087 NA |
1260 | /* |
1261 | * for small pages, there is no difference between | |
1262 | * shift and PAGE_SHIFT, so the bit-shift is safe. | |
1263 | * for huge pages, since vm_pgoff is in units of small | |
1264 | * pages, we need to shift off the always 0 bits to get | |
1265 | * a useful offset. | |
1266 | */ | |
1267 | BUG_ON(shift < PAGE_SHIFT); | |
1268 | off = vma->vm_pgoff >> (shift - PAGE_SHIFT); | |
5da7ca86 CL |
1269 | off += (addr - vma->vm_start) >> shift; |
1270 | return offset_il_node(pol, vma, off); | |
1271 | } else | |
1272 | return interleave_nodes(pol); | |
1273 | } | |
1274 | ||
00ac59ad | 1275 | #ifdef CONFIG_HUGETLBFS |
480eccf9 LS |
1276 | /* |
1277 | * huge_zonelist(@vma, @addr, @gfp_flags, @mpol) | |
1278 | * @vma = virtual memory area whose policy is sought | |
1279 | * @addr = address in @vma for shared policy lookup and interleave policy | |
1280 | * @gfp_flags = for requested zone | |
1281 | * @mpol = pointer to mempolicy pointer for reference counted 'BIND policy | |
1282 | * | |
1283 | * Returns a zonelist suitable for a huge page allocation. | |
1284 | * If the effective policy is 'BIND, returns pointer to policy's zonelist. | |
1285 | * If it is also a policy for which get_vma_policy() returns an extra | |
1286 | * reference, we must hold that reference until after allocation. | |
1287 | * In that case, return policy via @mpol so hugetlb allocation can drop | |
1288 | * the reference. For non-'BIND referenced policies, we can/do drop the | |
1289 | * reference here, so the caller doesn't need to know about the special case | |
1290 | * for default and current task policy. | |
1291 | */ | |
396faf03 | 1292 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr, |
480eccf9 | 1293 | gfp_t gfp_flags, struct mempolicy **mpol) |
5da7ca86 CL |
1294 | { |
1295 | struct mempolicy *pol = get_vma_policy(current, vma, addr); | |
480eccf9 | 1296 | struct zonelist *zl; |
5da7ca86 | 1297 | |
480eccf9 | 1298 | *mpol = NULL; /* probably no unref needed */ |
5da7ca86 CL |
1299 | if (pol->policy == MPOL_INTERLEAVE) { |
1300 | unsigned nid; | |
1301 | ||
1302 | nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT); | |
69682d85 LS |
1303 | if (unlikely(pol != &default_policy && |
1304 | pol != current->mempolicy)) | |
1305 | __mpol_free(pol); /* finished with pol */ | |
0e88460d | 1306 | return node_zonelist(nid, gfp_flags); |
5da7ca86 | 1307 | } |
480eccf9 LS |
1308 | |
1309 | zl = zonelist_policy(GFP_HIGHUSER, pol); | |
1310 | if (unlikely(pol != &default_policy && pol != current->mempolicy)) { | |
1311 | if (pol->policy != MPOL_BIND) | |
1312 | __mpol_free(pol); /* finished with pol */ | |
1313 | else | |
1314 | *mpol = pol; /* unref needed after allocation */ | |
1315 | } | |
1316 | return zl; | |
5da7ca86 | 1317 | } |
00ac59ad | 1318 | #endif |
5da7ca86 | 1319 | |
1da177e4 LT |
1320 | /* Allocate a page in interleaved policy. |
1321 | Own path because it needs to do special accounting. */ | |
662f3a0b AK |
1322 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
1323 | unsigned nid) | |
1da177e4 LT |
1324 | { |
1325 | struct zonelist *zl; | |
1326 | struct page *page; | |
1327 | ||
0e88460d | 1328 | zl = node_zonelist(nid, gfp); |
1da177e4 | 1329 | page = __alloc_pages(gfp, order, zl); |
dd1a239f | 1330 | if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0])) |
ca889e6c | 1331 | inc_zone_page_state(page, NUMA_INTERLEAVE_HIT); |
1da177e4 LT |
1332 | return page; |
1333 | } | |
1334 | ||
1335 | /** | |
1336 | * alloc_page_vma - Allocate a page for a VMA. | |
1337 | * | |
1338 | * @gfp: | |
1339 | * %GFP_USER user allocation. | |
1340 | * %GFP_KERNEL kernel allocations, | |
1341 | * %GFP_HIGHMEM highmem/user allocations, | |
1342 | * %GFP_FS allocation should not call back into a file system. | |
1343 | * %GFP_ATOMIC don't sleep. | |
1344 | * | |
1345 | * @vma: Pointer to VMA or NULL if not available. | |
1346 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
1347 | * | |
1348 | * This function allocates a page from the kernel page pool and applies | |
1349 | * a NUMA policy associated with the VMA or the current process. | |
1350 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1351 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
1352 | * all allocations for pages that will be mapped into | |
1353 | * user space. Returns NULL when no page can be allocated. | |
1354 | * | |
1355 | * Should be called with the mm_sem of the vma hold. | |
1356 | */ | |
1357 | struct page * | |
dd0fc66f | 1358 | alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr) |
1da177e4 | 1359 | { |
6e21c8f1 | 1360 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
480eccf9 | 1361 | struct zonelist *zl; |
1da177e4 | 1362 | |
cf2a473c | 1363 | cpuset_update_task_memory_state(); |
1da177e4 LT |
1364 | |
1365 | if (unlikely(pol->policy == MPOL_INTERLEAVE)) { | |
1366 | unsigned nid; | |
5da7ca86 CL |
1367 | |
1368 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT); | |
69682d85 LS |
1369 | if (unlikely(pol != &default_policy && |
1370 | pol != current->mempolicy)) | |
1371 | __mpol_free(pol); /* finished with pol */ | |
1da177e4 LT |
1372 | return alloc_page_interleave(gfp, 0, nid); |
1373 | } | |
480eccf9 LS |
1374 | zl = zonelist_policy(gfp, pol); |
1375 | if (pol != &default_policy && pol != current->mempolicy) { | |
1376 | /* | |
1377 | * slow path: ref counted policy -- shared or vma | |
1378 | */ | |
1379 | struct page *page = __alloc_pages(gfp, 0, zl); | |
1380 | __mpol_free(pol); | |
1381 | return page; | |
1382 | } | |
1383 | /* | |
1384 | * fast path: default or task policy | |
1385 | */ | |
1386 | return __alloc_pages(gfp, 0, zl); | |
1da177e4 LT |
1387 | } |
1388 | ||
1389 | /** | |
1390 | * alloc_pages_current - Allocate pages. | |
1391 | * | |
1392 | * @gfp: | |
1393 | * %GFP_USER user allocation, | |
1394 | * %GFP_KERNEL kernel allocation, | |
1395 | * %GFP_HIGHMEM highmem allocation, | |
1396 | * %GFP_FS don't call back into a file system. | |
1397 | * %GFP_ATOMIC don't sleep. | |
1398 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
1399 | * | |
1400 | * Allocate a page from the kernel page pool. When not in | |
1401 | * interrupt context and apply the current process NUMA policy. | |
1402 | * Returns NULL when no page can be allocated. | |
1403 | * | |
cf2a473c | 1404 | * Don't call cpuset_update_task_memory_state() unless |
1da177e4 LT |
1405 | * 1) it's ok to take cpuset_sem (can WAIT), and |
1406 | * 2) allocating for current task (not interrupt). | |
1407 | */ | |
dd0fc66f | 1408 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
1da177e4 LT |
1409 | { |
1410 | struct mempolicy *pol = current->mempolicy; | |
1411 | ||
1412 | if ((gfp & __GFP_WAIT) && !in_interrupt()) | |
cf2a473c | 1413 | cpuset_update_task_memory_state(); |
9b819d20 | 1414 | if (!pol || in_interrupt() || (gfp & __GFP_THISNODE)) |
1da177e4 LT |
1415 | pol = &default_policy; |
1416 | if (pol->policy == MPOL_INTERLEAVE) | |
1417 | return alloc_page_interleave(gfp, order, interleave_nodes(pol)); | |
1418 | return __alloc_pages(gfp, order, zonelist_policy(gfp, pol)); | |
1419 | } | |
1420 | EXPORT_SYMBOL(alloc_pages_current); | |
1421 | ||
4225399a PJ |
1422 | /* |
1423 | * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it | |
1424 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() | |
1425 | * with the mems_allowed returned by cpuset_mems_allowed(). This | |
1426 | * keeps mempolicies cpuset relative after its cpuset moves. See | |
1427 | * further kernel/cpuset.c update_nodemask(). | |
1428 | */ | |
4225399a | 1429 | |
1da177e4 LT |
1430 | /* Slow path of a mempolicy copy */ |
1431 | struct mempolicy *__mpol_copy(struct mempolicy *old) | |
1432 | { | |
1433 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
1434 | ||
1435 | if (!new) | |
1436 | return ERR_PTR(-ENOMEM); | |
4225399a PJ |
1437 | if (current_cpuset_is_being_rebound()) { |
1438 | nodemask_t mems = cpuset_mems_allowed(current); | |
1439 | mpol_rebind_policy(old, &mems); | |
1440 | } | |
1da177e4 LT |
1441 | *new = *old; |
1442 | atomic_set(&new->refcnt, 1); | |
1443 | if (new->policy == MPOL_BIND) { | |
1444 | int sz = ksize(old->v.zonelist); | |
e94b1766 | 1445 | new->v.zonelist = kmemdup(old->v.zonelist, sz, GFP_KERNEL); |
1da177e4 LT |
1446 | if (!new->v.zonelist) { |
1447 | kmem_cache_free(policy_cache, new); | |
1448 | return ERR_PTR(-ENOMEM); | |
1449 | } | |
1da177e4 LT |
1450 | } |
1451 | return new; | |
1452 | } | |
1453 | ||
1454 | /* Slow path of a mempolicy comparison */ | |
1455 | int __mpol_equal(struct mempolicy *a, struct mempolicy *b) | |
1456 | { | |
1457 | if (!a || !b) | |
1458 | return 0; | |
1459 | if (a->policy != b->policy) | |
1460 | return 0; | |
1461 | switch (a->policy) { | |
1462 | case MPOL_DEFAULT: | |
1463 | return 1; | |
1464 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 1465 | return nodes_equal(a->v.nodes, b->v.nodes); |
1da177e4 LT |
1466 | case MPOL_PREFERRED: |
1467 | return a->v.preferred_node == b->v.preferred_node; | |
1468 | case MPOL_BIND: { | |
1469 | int i; | |
dd1a239f MG |
1470 | for (i = 0; a->v.zonelist->_zonerefs[i].zone; i++) { |
1471 | struct zone *za, *zb; | |
1472 | za = zonelist_zone(&a->v.zonelist->_zonerefs[i]); | |
1473 | zb = zonelist_zone(&b->v.zonelist->_zonerefs[i]); | |
1474 | if (za != zb) | |
1da177e4 | 1475 | return 0; |
dd1a239f MG |
1476 | } |
1477 | return b->v.zonelist->_zonerefs[i].zone == NULL; | |
1da177e4 LT |
1478 | } |
1479 | default: | |
1480 | BUG(); | |
1481 | return 0; | |
1482 | } | |
1483 | } | |
1484 | ||
1485 | /* Slow path of a mpol destructor. */ | |
1486 | void __mpol_free(struct mempolicy *p) | |
1487 | { | |
1488 | if (!atomic_dec_and_test(&p->refcnt)) | |
1489 | return; | |
1490 | if (p->policy == MPOL_BIND) | |
1491 | kfree(p->v.zonelist); | |
1492 | p->policy = MPOL_DEFAULT; | |
1493 | kmem_cache_free(policy_cache, p); | |
1494 | } | |
1495 | ||
1da177e4 LT |
1496 | /* |
1497 | * Shared memory backing store policy support. | |
1498 | * | |
1499 | * Remember policies even when nobody has shared memory mapped. | |
1500 | * The policies are kept in Red-Black tree linked from the inode. | |
1501 | * They are protected by the sp->lock spinlock, which should be held | |
1502 | * for any accesses to the tree. | |
1503 | */ | |
1504 | ||
1505 | /* lookup first element intersecting start-end */ | |
1506 | /* Caller holds sp->lock */ | |
1507 | static struct sp_node * | |
1508 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
1509 | { | |
1510 | struct rb_node *n = sp->root.rb_node; | |
1511 | ||
1512 | while (n) { | |
1513 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
1514 | ||
1515 | if (start >= p->end) | |
1516 | n = n->rb_right; | |
1517 | else if (end <= p->start) | |
1518 | n = n->rb_left; | |
1519 | else | |
1520 | break; | |
1521 | } | |
1522 | if (!n) | |
1523 | return NULL; | |
1524 | for (;;) { | |
1525 | struct sp_node *w = NULL; | |
1526 | struct rb_node *prev = rb_prev(n); | |
1527 | if (!prev) | |
1528 | break; | |
1529 | w = rb_entry(prev, struct sp_node, nd); | |
1530 | if (w->end <= start) | |
1531 | break; | |
1532 | n = prev; | |
1533 | } | |
1534 | return rb_entry(n, struct sp_node, nd); | |
1535 | } | |
1536 | ||
1537 | /* Insert a new shared policy into the list. */ | |
1538 | /* Caller holds sp->lock */ | |
1539 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) | |
1540 | { | |
1541 | struct rb_node **p = &sp->root.rb_node; | |
1542 | struct rb_node *parent = NULL; | |
1543 | struct sp_node *nd; | |
1544 | ||
1545 | while (*p) { | |
1546 | parent = *p; | |
1547 | nd = rb_entry(parent, struct sp_node, nd); | |
1548 | if (new->start < nd->start) | |
1549 | p = &(*p)->rb_left; | |
1550 | else if (new->end > nd->end) | |
1551 | p = &(*p)->rb_right; | |
1552 | else | |
1553 | BUG(); | |
1554 | } | |
1555 | rb_link_node(&new->nd, parent, p); | |
1556 | rb_insert_color(&new->nd, &sp->root); | |
140d5a49 | 1557 | pr_debug("inserting %lx-%lx: %d\n", new->start, new->end, |
1da177e4 LT |
1558 | new->policy ? new->policy->policy : 0); |
1559 | } | |
1560 | ||
1561 | /* Find shared policy intersecting idx */ | |
1562 | struct mempolicy * | |
1563 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
1564 | { | |
1565 | struct mempolicy *pol = NULL; | |
1566 | struct sp_node *sn; | |
1567 | ||
1568 | if (!sp->root.rb_node) | |
1569 | return NULL; | |
1570 | spin_lock(&sp->lock); | |
1571 | sn = sp_lookup(sp, idx, idx+1); | |
1572 | if (sn) { | |
1573 | mpol_get(sn->policy); | |
1574 | pol = sn->policy; | |
1575 | } | |
1576 | spin_unlock(&sp->lock); | |
1577 | return pol; | |
1578 | } | |
1579 | ||
1580 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) | |
1581 | { | |
140d5a49 | 1582 | pr_debug("deleting %lx-l%lx\n", n->start, n->end); |
1da177e4 LT |
1583 | rb_erase(&n->nd, &sp->root); |
1584 | mpol_free(n->policy); | |
1585 | kmem_cache_free(sn_cache, n); | |
1586 | } | |
1587 | ||
dbcb0f19 AB |
1588 | static struct sp_node *sp_alloc(unsigned long start, unsigned long end, |
1589 | struct mempolicy *pol) | |
1da177e4 LT |
1590 | { |
1591 | struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
1592 | ||
1593 | if (!n) | |
1594 | return NULL; | |
1595 | n->start = start; | |
1596 | n->end = end; | |
1597 | mpol_get(pol); | |
1598 | n->policy = pol; | |
1599 | return n; | |
1600 | } | |
1601 | ||
1602 | /* Replace a policy range. */ | |
1603 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
1604 | unsigned long end, struct sp_node *new) | |
1605 | { | |
1606 | struct sp_node *n, *new2 = NULL; | |
1607 | ||
1608 | restart: | |
1609 | spin_lock(&sp->lock); | |
1610 | n = sp_lookup(sp, start, end); | |
1611 | /* Take care of old policies in the same range. */ | |
1612 | while (n && n->start < end) { | |
1613 | struct rb_node *next = rb_next(&n->nd); | |
1614 | if (n->start >= start) { | |
1615 | if (n->end <= end) | |
1616 | sp_delete(sp, n); | |
1617 | else | |
1618 | n->start = end; | |
1619 | } else { | |
1620 | /* Old policy spanning whole new range. */ | |
1621 | if (n->end > end) { | |
1622 | if (!new2) { | |
1623 | spin_unlock(&sp->lock); | |
1624 | new2 = sp_alloc(end, n->end, n->policy); | |
1625 | if (!new2) | |
1626 | return -ENOMEM; | |
1627 | goto restart; | |
1628 | } | |
1629 | n->end = start; | |
1630 | sp_insert(sp, new2); | |
1631 | new2 = NULL; | |
1632 | break; | |
1633 | } else | |
1634 | n->end = start; | |
1635 | } | |
1636 | if (!next) | |
1637 | break; | |
1638 | n = rb_entry(next, struct sp_node, nd); | |
1639 | } | |
1640 | if (new) | |
1641 | sp_insert(sp, new); | |
1642 | spin_unlock(&sp->lock); | |
1643 | if (new2) { | |
1644 | mpol_free(new2->policy); | |
1645 | kmem_cache_free(sn_cache, new2); | |
1646 | } | |
1647 | return 0; | |
1648 | } | |
1649 | ||
7339ff83 RH |
1650 | void mpol_shared_policy_init(struct shared_policy *info, int policy, |
1651 | nodemask_t *policy_nodes) | |
1652 | { | |
1653 | info->root = RB_ROOT; | |
1654 | spin_lock_init(&info->lock); | |
1655 | ||
1656 | if (policy != MPOL_DEFAULT) { | |
1657 | struct mempolicy *newpol; | |
1658 | ||
1659 | /* Falls back to MPOL_DEFAULT on any error */ | |
1660 | newpol = mpol_new(policy, policy_nodes); | |
1661 | if (!IS_ERR(newpol)) { | |
1662 | /* Create pseudo-vma that contains just the policy */ | |
1663 | struct vm_area_struct pvma; | |
1664 | ||
1665 | memset(&pvma, 0, sizeof(struct vm_area_struct)); | |
1666 | /* Policy covers entire file */ | |
1667 | pvma.vm_end = TASK_SIZE; | |
1668 | mpol_set_shared_policy(info, &pvma, newpol); | |
1669 | mpol_free(newpol); | |
1670 | } | |
1671 | } | |
1672 | } | |
1673 | ||
1da177e4 LT |
1674 | int mpol_set_shared_policy(struct shared_policy *info, |
1675 | struct vm_area_struct *vma, struct mempolicy *npol) | |
1676 | { | |
1677 | int err; | |
1678 | struct sp_node *new = NULL; | |
1679 | unsigned long sz = vma_pages(vma); | |
1680 | ||
140d5a49 | 1681 | pr_debug("set_shared_policy %lx sz %lu %d %lx\n", |
1da177e4 LT |
1682 | vma->vm_pgoff, |
1683 | sz, npol? npol->policy : -1, | |
140d5a49 | 1684 | npol ? nodes_addr(npol->v.nodes)[0] : -1); |
1da177e4 LT |
1685 | |
1686 | if (npol) { | |
1687 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
1688 | if (!new) | |
1689 | return -ENOMEM; | |
1690 | } | |
1691 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
1692 | if (err && new) | |
1693 | kmem_cache_free(sn_cache, new); | |
1694 | return err; | |
1695 | } | |
1696 | ||
1697 | /* Free a backing policy store on inode delete. */ | |
1698 | void mpol_free_shared_policy(struct shared_policy *p) | |
1699 | { | |
1700 | struct sp_node *n; | |
1701 | struct rb_node *next; | |
1702 | ||
1703 | if (!p->root.rb_node) | |
1704 | return; | |
1705 | spin_lock(&p->lock); | |
1706 | next = rb_first(&p->root); | |
1707 | while (next) { | |
1708 | n = rb_entry(next, struct sp_node, nd); | |
1709 | next = rb_next(&n->nd); | |
90c5029e | 1710 | rb_erase(&n->nd, &p->root); |
1da177e4 LT |
1711 | mpol_free(n->policy); |
1712 | kmem_cache_free(sn_cache, n); | |
1713 | } | |
1714 | spin_unlock(&p->lock); | |
1da177e4 LT |
1715 | } |
1716 | ||
1717 | /* assumes fs == KERNEL_DS */ | |
1718 | void __init numa_policy_init(void) | |
1719 | { | |
b71636e2 PM |
1720 | nodemask_t interleave_nodes; |
1721 | unsigned long largest = 0; | |
1722 | int nid, prefer = 0; | |
1723 | ||
1da177e4 LT |
1724 | policy_cache = kmem_cache_create("numa_policy", |
1725 | sizeof(struct mempolicy), | |
20c2df83 | 1726 | 0, SLAB_PANIC, NULL); |
1da177e4 LT |
1727 | |
1728 | sn_cache = kmem_cache_create("shared_policy_node", | |
1729 | sizeof(struct sp_node), | |
20c2df83 | 1730 | 0, SLAB_PANIC, NULL); |
1da177e4 | 1731 | |
b71636e2 PM |
1732 | /* |
1733 | * Set interleaving policy for system init. Interleaving is only | |
1734 | * enabled across suitably sized nodes (default is >= 16MB), or | |
1735 | * fall back to the largest node if they're all smaller. | |
1736 | */ | |
1737 | nodes_clear(interleave_nodes); | |
56bbd65d | 1738 | for_each_node_state(nid, N_HIGH_MEMORY) { |
b71636e2 PM |
1739 | unsigned long total_pages = node_present_pages(nid); |
1740 | ||
1741 | /* Preserve the largest node */ | |
1742 | if (largest < total_pages) { | |
1743 | largest = total_pages; | |
1744 | prefer = nid; | |
1745 | } | |
1746 | ||
1747 | /* Interleave this node? */ | |
1748 | if ((total_pages << PAGE_SHIFT) >= (16 << 20)) | |
1749 | node_set(nid, interleave_nodes); | |
1750 | } | |
1751 | ||
1752 | /* All too small, use the largest */ | |
1753 | if (unlikely(nodes_empty(interleave_nodes))) | |
1754 | node_set(prefer, interleave_nodes); | |
1da177e4 | 1755 | |
b71636e2 | 1756 | if (do_set_mempolicy(MPOL_INTERLEAVE, &interleave_nodes)) |
1da177e4 LT |
1757 | printk("numa_policy_init: interleaving failed\n"); |
1758 | } | |
1759 | ||
8bccd85f | 1760 | /* Reset policy of current process to default */ |
1da177e4 LT |
1761 | void numa_default_policy(void) |
1762 | { | |
8bccd85f | 1763 | do_set_mempolicy(MPOL_DEFAULT, NULL); |
1da177e4 | 1764 | } |
68860ec1 PJ |
1765 | |
1766 | /* Migrate a policy to a different set of nodes */ | |
dbcb0f19 AB |
1767 | static void mpol_rebind_policy(struct mempolicy *pol, |
1768 | const nodemask_t *newmask) | |
68860ec1 | 1769 | { |
74cb2155 | 1770 | nodemask_t *mpolmask; |
68860ec1 PJ |
1771 | nodemask_t tmp; |
1772 | ||
1773 | if (!pol) | |
1774 | return; | |
74cb2155 PJ |
1775 | mpolmask = &pol->cpuset_mems_allowed; |
1776 | if (nodes_equal(*mpolmask, *newmask)) | |
1777 | return; | |
68860ec1 PJ |
1778 | |
1779 | switch (pol->policy) { | |
1780 | case MPOL_DEFAULT: | |
1781 | break; | |
1782 | case MPOL_INTERLEAVE: | |
74cb2155 | 1783 | nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask); |
68860ec1 | 1784 | pol->v.nodes = tmp; |
74cb2155 PJ |
1785 | *mpolmask = *newmask; |
1786 | current->il_next = node_remap(current->il_next, | |
1787 | *mpolmask, *newmask); | |
68860ec1 PJ |
1788 | break; |
1789 | case MPOL_PREFERRED: | |
1790 | pol->v.preferred_node = node_remap(pol->v.preferred_node, | |
74cb2155 PJ |
1791 | *mpolmask, *newmask); |
1792 | *mpolmask = *newmask; | |
68860ec1 PJ |
1793 | break; |
1794 | case MPOL_BIND: { | |
1795 | nodemask_t nodes; | |
dd1a239f | 1796 | struct zoneref *z; |
68860ec1 PJ |
1797 | struct zonelist *zonelist; |
1798 | ||
1799 | nodes_clear(nodes); | |
dd1a239f MG |
1800 | for (z = pol->v.zonelist->_zonerefs; z->zone; z++) |
1801 | node_set(zonelist_node_idx(z), nodes); | |
74cb2155 | 1802 | nodes_remap(tmp, nodes, *mpolmask, *newmask); |
68860ec1 PJ |
1803 | nodes = tmp; |
1804 | ||
1805 | zonelist = bind_zonelist(&nodes); | |
1806 | ||
1807 | /* If no mem, then zonelist is NULL and we keep old zonelist. | |
1808 | * If that old zonelist has no remaining mems_allowed nodes, | |
1809 | * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT. | |
1810 | */ | |
1811 | ||
8af5e2eb | 1812 | if (!IS_ERR(zonelist)) { |
68860ec1 PJ |
1813 | /* Good - got mem - substitute new zonelist */ |
1814 | kfree(pol->v.zonelist); | |
1815 | pol->v.zonelist = zonelist; | |
1816 | } | |
74cb2155 | 1817 | *mpolmask = *newmask; |
68860ec1 PJ |
1818 | break; |
1819 | } | |
1820 | default: | |
1821 | BUG(); | |
1822 | break; | |
1823 | } | |
1824 | } | |
1825 | ||
1826 | /* | |
74cb2155 PJ |
1827 | * Wrapper for mpol_rebind_policy() that just requires task |
1828 | * pointer, and updates task mempolicy. | |
68860ec1 | 1829 | */ |
74cb2155 PJ |
1830 | |
1831 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) | |
68860ec1 | 1832 | { |
74cb2155 | 1833 | mpol_rebind_policy(tsk->mempolicy, new); |
68860ec1 | 1834 | } |
1a75a6c8 | 1835 | |
4225399a PJ |
1836 | /* |
1837 | * Rebind each vma in mm to new nodemask. | |
1838 | * | |
1839 | * Call holding a reference to mm. Takes mm->mmap_sem during call. | |
1840 | */ | |
1841 | ||
1842 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) | |
1843 | { | |
1844 | struct vm_area_struct *vma; | |
1845 | ||
1846 | down_write(&mm->mmap_sem); | |
1847 | for (vma = mm->mmap; vma; vma = vma->vm_next) | |
1848 | mpol_rebind_policy(vma->vm_policy, new); | |
1849 | up_write(&mm->mmap_sem); | |
1850 | } | |
1851 | ||
1a75a6c8 CL |
1852 | /* |
1853 | * Display pages allocated per node and memory policy via /proc. | |
1854 | */ | |
1855 | ||
15ad7cdc HD |
1856 | static const char * const policy_types[] = |
1857 | { "default", "prefer", "bind", "interleave" }; | |
1a75a6c8 CL |
1858 | |
1859 | /* | |
1860 | * Convert a mempolicy into a string. | |
1861 | * Returns the number of characters in buffer (if positive) | |
1862 | * or an error (negative) | |
1863 | */ | |
1864 | static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) | |
1865 | { | |
1866 | char *p = buffer; | |
1867 | int l; | |
1868 | nodemask_t nodes; | |
1869 | int mode = pol ? pol->policy : MPOL_DEFAULT; | |
1870 | ||
1871 | switch (mode) { | |
1872 | case MPOL_DEFAULT: | |
1873 | nodes_clear(nodes); | |
1874 | break; | |
1875 | ||
1876 | case MPOL_PREFERRED: | |
1877 | nodes_clear(nodes); | |
1878 | node_set(pol->v.preferred_node, nodes); | |
1879 | break; | |
1880 | ||
1881 | case MPOL_BIND: | |
1882 | get_zonemask(pol, &nodes); | |
1883 | break; | |
1884 | ||
1885 | case MPOL_INTERLEAVE: | |
1886 | nodes = pol->v.nodes; | |
1887 | break; | |
1888 | ||
1889 | default: | |
1890 | BUG(); | |
1891 | return -EFAULT; | |
1892 | } | |
1893 | ||
1894 | l = strlen(policy_types[mode]); | |
1895 | if (buffer + maxlen < p + l + 1) | |
1896 | return -ENOSPC; | |
1897 | ||
1898 | strcpy(p, policy_types[mode]); | |
1899 | p += l; | |
1900 | ||
1901 | if (!nodes_empty(nodes)) { | |
1902 | if (buffer + maxlen < p + 2) | |
1903 | return -ENOSPC; | |
1904 | *p++ = '='; | |
1905 | p += nodelist_scnprintf(p, buffer + maxlen - p, nodes); | |
1906 | } | |
1907 | return p - buffer; | |
1908 | } | |
1909 | ||
1910 | struct numa_maps { | |
1911 | unsigned long pages; | |
1912 | unsigned long anon; | |
397874df CL |
1913 | unsigned long active; |
1914 | unsigned long writeback; | |
1a75a6c8 | 1915 | unsigned long mapcount_max; |
397874df CL |
1916 | unsigned long dirty; |
1917 | unsigned long swapcache; | |
1a75a6c8 CL |
1918 | unsigned long node[MAX_NUMNODES]; |
1919 | }; | |
1920 | ||
397874df | 1921 | static void gather_stats(struct page *page, void *private, int pte_dirty) |
1a75a6c8 CL |
1922 | { |
1923 | struct numa_maps *md = private; | |
1924 | int count = page_mapcount(page); | |
1925 | ||
397874df CL |
1926 | md->pages++; |
1927 | if (pte_dirty || PageDirty(page)) | |
1928 | md->dirty++; | |
1a75a6c8 | 1929 | |
397874df CL |
1930 | if (PageSwapCache(page)) |
1931 | md->swapcache++; | |
1a75a6c8 | 1932 | |
397874df CL |
1933 | if (PageActive(page)) |
1934 | md->active++; | |
1935 | ||
1936 | if (PageWriteback(page)) | |
1937 | md->writeback++; | |
1a75a6c8 CL |
1938 | |
1939 | if (PageAnon(page)) | |
1940 | md->anon++; | |
1941 | ||
397874df CL |
1942 | if (count > md->mapcount_max) |
1943 | md->mapcount_max = count; | |
1944 | ||
1a75a6c8 | 1945 | md->node[page_to_nid(page)]++; |
1a75a6c8 CL |
1946 | } |
1947 | ||
7f709ed0 | 1948 | #ifdef CONFIG_HUGETLB_PAGE |
397874df CL |
1949 | static void check_huge_range(struct vm_area_struct *vma, |
1950 | unsigned long start, unsigned long end, | |
1951 | struct numa_maps *md) | |
1952 | { | |
1953 | unsigned long addr; | |
1954 | struct page *page; | |
1955 | ||
1956 | for (addr = start; addr < end; addr += HPAGE_SIZE) { | |
1957 | pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK); | |
1958 | pte_t pte; | |
1959 | ||
1960 | if (!ptep) | |
1961 | continue; | |
1962 | ||
1963 | pte = *ptep; | |
1964 | if (pte_none(pte)) | |
1965 | continue; | |
1966 | ||
1967 | page = pte_page(pte); | |
1968 | if (!page) | |
1969 | continue; | |
1970 | ||
1971 | gather_stats(page, md, pte_dirty(*ptep)); | |
1972 | } | |
1973 | } | |
7f709ed0 AM |
1974 | #else |
1975 | static inline void check_huge_range(struct vm_area_struct *vma, | |
1976 | unsigned long start, unsigned long end, | |
1977 | struct numa_maps *md) | |
1978 | { | |
1979 | } | |
1980 | #endif | |
397874df | 1981 | |
1a75a6c8 CL |
1982 | int show_numa_map(struct seq_file *m, void *v) |
1983 | { | |
99f89551 | 1984 | struct proc_maps_private *priv = m->private; |
1a75a6c8 CL |
1985 | struct vm_area_struct *vma = v; |
1986 | struct numa_maps *md; | |
397874df CL |
1987 | struct file *file = vma->vm_file; |
1988 | struct mm_struct *mm = vma->vm_mm; | |
480eccf9 | 1989 | struct mempolicy *pol; |
1a75a6c8 CL |
1990 | int n; |
1991 | char buffer[50]; | |
1992 | ||
397874df | 1993 | if (!mm) |
1a75a6c8 CL |
1994 | return 0; |
1995 | ||
1996 | md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL); | |
1997 | if (!md) | |
1998 | return 0; | |
1999 | ||
480eccf9 LS |
2000 | pol = get_vma_policy(priv->task, vma, vma->vm_start); |
2001 | mpol_to_str(buffer, sizeof(buffer), pol); | |
2002 | /* | |
2003 | * unref shared or other task's mempolicy | |
2004 | */ | |
2005 | if (pol != &default_policy && pol != current->mempolicy) | |
2006 | __mpol_free(pol); | |
397874df CL |
2007 | |
2008 | seq_printf(m, "%08lx %s", vma->vm_start, buffer); | |
2009 | ||
2010 | if (file) { | |
2011 | seq_printf(m, " file="); | |
c32c2f63 | 2012 | seq_path(m, &file->f_path, "\n\t= "); |
397874df CL |
2013 | } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) { |
2014 | seq_printf(m, " heap"); | |
2015 | } else if (vma->vm_start <= mm->start_stack && | |
2016 | vma->vm_end >= mm->start_stack) { | |
2017 | seq_printf(m, " stack"); | |
2018 | } | |
2019 | ||
2020 | if (is_vm_hugetlb_page(vma)) { | |
2021 | check_huge_range(vma, vma->vm_start, vma->vm_end, md); | |
2022 | seq_printf(m, " huge"); | |
2023 | } else { | |
a57ebfdb | 2024 | check_pgd_range(vma, vma->vm_start, vma->vm_end, |
56bbd65d | 2025 | &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md); |
397874df CL |
2026 | } |
2027 | ||
2028 | if (!md->pages) | |
2029 | goto out; | |
1a75a6c8 | 2030 | |
397874df CL |
2031 | if (md->anon) |
2032 | seq_printf(m," anon=%lu",md->anon); | |
1a75a6c8 | 2033 | |
397874df CL |
2034 | if (md->dirty) |
2035 | seq_printf(m," dirty=%lu",md->dirty); | |
1a75a6c8 | 2036 | |
397874df CL |
2037 | if (md->pages != md->anon && md->pages != md->dirty) |
2038 | seq_printf(m, " mapped=%lu", md->pages); | |
1a75a6c8 | 2039 | |
397874df CL |
2040 | if (md->mapcount_max > 1) |
2041 | seq_printf(m, " mapmax=%lu", md->mapcount_max); | |
1a75a6c8 | 2042 | |
397874df CL |
2043 | if (md->swapcache) |
2044 | seq_printf(m," swapcache=%lu", md->swapcache); | |
2045 | ||
2046 | if (md->active < md->pages && !is_vm_hugetlb_page(vma)) | |
2047 | seq_printf(m," active=%lu", md->active); | |
2048 | ||
2049 | if (md->writeback) | |
2050 | seq_printf(m," writeback=%lu", md->writeback); | |
2051 | ||
56bbd65d | 2052 | for_each_node_state(n, N_HIGH_MEMORY) |
397874df CL |
2053 | if (md->node[n]) |
2054 | seq_printf(m, " N%d=%lu", n, md->node[n]); | |
2055 | out: | |
2056 | seq_putc(m, '\n'); | |
1a75a6c8 CL |
2057 | kfree(md); |
2058 | ||
2059 | if (m->count < m->size) | |
99f89551 | 2060 | m->version = (vma != priv->tail_vma) ? vma->vm_start : 0; |
1a75a6c8 CL |
2061 | return 0; |
2062 | } |