]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Simple NUMA memory policy for the Linux kernel. | |
3 | * | |
4 | * Copyright 2003,2004 Andi Kleen, SuSE Labs. | |
8bccd85f | 5 | * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc. |
1da177e4 LT |
6 | * Subject to the GNU Public License, version 2. |
7 | * | |
8 | * NUMA policy allows the user to give hints in which node(s) memory should | |
9 | * be allocated. | |
10 | * | |
11 | * Support four policies per VMA and per process: | |
12 | * | |
13 | * The VMA policy has priority over the process policy for a page fault. | |
14 | * | |
15 | * interleave Allocate memory interleaved over a set of nodes, | |
16 | * with normal fallback if it fails. | |
17 | * For VMA based allocations this interleaves based on the | |
18 | * offset into the backing object or offset into the mapping | |
19 | * for anonymous memory. For process policy an process counter | |
20 | * is used. | |
8bccd85f | 21 | * |
1da177e4 LT |
22 | * bind Only allocate memory on a specific set of nodes, |
23 | * no fallback. | |
8bccd85f CL |
24 | * FIXME: memory is allocated starting with the first node |
25 | * to the last. It would be better if bind would truly restrict | |
26 | * the allocation to memory nodes instead | |
27 | * | |
1da177e4 LT |
28 | * preferred Try a specific node first before normal fallback. |
29 | * As a special case node -1 here means do the allocation | |
30 | * on the local CPU. This is normally identical to default, | |
31 | * but useful to set in a VMA when you have a non default | |
32 | * process policy. | |
8bccd85f | 33 | * |
1da177e4 LT |
34 | * default Allocate on the local node first, or when on a VMA |
35 | * use the process policy. This is what Linux always did | |
36 | * in a NUMA aware kernel and still does by, ahem, default. | |
37 | * | |
38 | * The process policy is applied for most non interrupt memory allocations | |
39 | * in that process' context. Interrupts ignore the policies and always | |
40 | * try to allocate on the local CPU. The VMA policy is only applied for memory | |
41 | * allocations for a VMA in the VM. | |
42 | * | |
43 | * Currently there are a few corner cases in swapping where the policy | |
44 | * is not applied, but the majority should be handled. When process policy | |
45 | * is used it is not remembered over swap outs/swap ins. | |
46 | * | |
47 | * Only the highest zone in the zone hierarchy gets policied. Allocations | |
48 | * requesting a lower zone just use default policy. This implies that | |
49 | * on systems with highmem kernel lowmem allocation don't get policied. | |
50 | * Same with GFP_DMA allocations. | |
51 | * | |
52 | * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between | |
53 | * all users and remembered even when nobody has memory mapped. | |
54 | */ | |
55 | ||
56 | /* Notebook: | |
57 | fix mmap readahead to honour policy and enable policy for any page cache | |
58 | object | |
59 | statistics for bigpages | |
60 | global policy for page cache? currently it uses process policy. Requires | |
61 | first item above. | |
62 | handle mremap for shared memory (currently ignored for the policy) | |
63 | grows down? | |
64 | make bind policy root only? It can trigger oom much faster and the | |
65 | kernel is not always grateful with that. | |
66 | could replace all the switch()es with a mempolicy_ops structure. | |
67 | */ | |
68 | ||
69 | #include <linux/mempolicy.h> | |
70 | #include <linux/mm.h> | |
71 | #include <linux/highmem.h> | |
72 | #include <linux/hugetlb.h> | |
73 | #include <linux/kernel.h> | |
74 | #include <linux/sched.h> | |
75 | #include <linux/mm.h> | |
76 | #include <linux/nodemask.h> | |
77 | #include <linux/cpuset.h> | |
78 | #include <linux/gfp.h> | |
79 | #include <linux/slab.h> | |
80 | #include <linux/string.h> | |
81 | #include <linux/module.h> | |
82 | #include <linux/interrupt.h> | |
83 | #include <linux/init.h> | |
84 | #include <linux/compat.h> | |
85 | #include <linux/mempolicy.h> | |
dc9aa5b9 | 86 | #include <linux/swap.h> |
1a75a6c8 CL |
87 | #include <linux/seq_file.h> |
88 | #include <linux/proc_fs.h> | |
dc9aa5b9 | 89 | |
1da177e4 LT |
90 | #include <asm/tlbflush.h> |
91 | #include <asm/uaccess.h> | |
92 | ||
38e35860 | 93 | /* Internal flags */ |
dc9aa5b9 | 94 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ |
38e35860 | 95 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ |
1a75a6c8 | 96 | #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */ |
dc9aa5b9 | 97 | |
1da177e4 LT |
98 | static kmem_cache_t *policy_cache; |
99 | static kmem_cache_t *sn_cache; | |
100 | ||
101 | #define PDprintk(fmt...) | |
102 | ||
103 | /* Highest zone. An specific allocation for a zone below that is not | |
104 | policied. */ | |
4be38e35 | 105 | int policy_zone = ZONE_DMA; |
1da177e4 | 106 | |
d42c6997 | 107 | struct mempolicy default_policy = { |
1da177e4 LT |
108 | .refcnt = ATOMIC_INIT(1), /* never free it */ |
109 | .policy = MPOL_DEFAULT, | |
110 | }; | |
111 | ||
1da177e4 | 112 | /* Do sanity checking on a policy */ |
dfcd3c0d | 113 | static int mpol_check_policy(int mode, nodemask_t *nodes) |
1da177e4 | 114 | { |
dfcd3c0d | 115 | int empty = nodes_empty(*nodes); |
1da177e4 LT |
116 | |
117 | switch (mode) { | |
118 | case MPOL_DEFAULT: | |
119 | if (!empty) | |
120 | return -EINVAL; | |
121 | break; | |
122 | case MPOL_BIND: | |
123 | case MPOL_INTERLEAVE: | |
124 | /* Preferred will only use the first bit, but allow | |
125 | more for now. */ | |
126 | if (empty) | |
127 | return -EINVAL; | |
128 | break; | |
129 | } | |
dfcd3c0d | 130 | return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL; |
1da177e4 | 131 | } |
1da177e4 | 132 | /* Generate a custom zonelist for the BIND policy. */ |
dfcd3c0d | 133 | static struct zonelist *bind_zonelist(nodemask_t *nodes) |
1da177e4 LT |
134 | { |
135 | struct zonelist *zl; | |
136 | int num, max, nd; | |
137 | ||
dfcd3c0d | 138 | max = 1 + MAX_NR_ZONES * nodes_weight(*nodes); |
1da177e4 LT |
139 | zl = kmalloc(sizeof(void *) * max, GFP_KERNEL); |
140 | if (!zl) | |
141 | return NULL; | |
142 | num = 0; | |
4be38e35 CL |
143 | for_each_node_mask(nd, *nodes) |
144 | zl->zones[num++] = &NODE_DATA(nd)->node_zones[policy_zone]; | |
1da177e4 LT |
145 | zl->zones[num] = NULL; |
146 | return zl; | |
147 | } | |
148 | ||
149 | /* Create a new policy */ | |
dfcd3c0d | 150 | static struct mempolicy *mpol_new(int mode, nodemask_t *nodes) |
1da177e4 LT |
151 | { |
152 | struct mempolicy *policy; | |
153 | ||
dfcd3c0d | 154 | PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]); |
1da177e4 LT |
155 | if (mode == MPOL_DEFAULT) |
156 | return NULL; | |
157 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
158 | if (!policy) | |
159 | return ERR_PTR(-ENOMEM); | |
160 | atomic_set(&policy->refcnt, 1); | |
161 | switch (mode) { | |
162 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 163 | policy->v.nodes = *nodes; |
8f493d79 AK |
164 | if (nodes_weight(*nodes) == 0) { |
165 | kmem_cache_free(policy_cache, policy); | |
166 | return ERR_PTR(-EINVAL); | |
167 | } | |
1da177e4 LT |
168 | break; |
169 | case MPOL_PREFERRED: | |
dfcd3c0d | 170 | policy->v.preferred_node = first_node(*nodes); |
1da177e4 LT |
171 | if (policy->v.preferred_node >= MAX_NUMNODES) |
172 | policy->v.preferred_node = -1; | |
173 | break; | |
174 | case MPOL_BIND: | |
175 | policy->v.zonelist = bind_zonelist(nodes); | |
176 | if (policy->v.zonelist == NULL) { | |
177 | kmem_cache_free(policy_cache, policy); | |
178 | return ERR_PTR(-ENOMEM); | |
179 | } | |
180 | break; | |
181 | } | |
182 | policy->policy = mode; | |
183 | return policy; | |
184 | } | |
185 | ||
dc9aa5b9 CL |
186 | /* Check if we are the only process mapping the page in question */ |
187 | static inline int single_mm_mapping(struct mm_struct *mm, | |
188 | struct address_space *mapping) | |
189 | { | |
190 | struct vm_area_struct *vma; | |
191 | struct prio_tree_iter iter; | |
192 | int rc = 1; | |
193 | ||
194 | spin_lock(&mapping->i_mmap_lock); | |
195 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX) | |
196 | if (mm != vma->vm_mm) { | |
197 | rc = 0; | |
198 | goto out; | |
199 | } | |
200 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) | |
201 | if (mm != vma->vm_mm) { | |
202 | rc = 0; | |
203 | goto out; | |
204 | } | |
205 | out: | |
206 | spin_unlock(&mapping->i_mmap_lock); | |
207 | return rc; | |
208 | } | |
209 | ||
210 | /* | |
211 | * Add a page to be migrated to the pagelist | |
212 | */ | |
213 | static void migrate_page_add(struct vm_area_struct *vma, | |
214 | struct page *page, struct list_head *pagelist, unsigned long flags) | |
215 | { | |
216 | /* | |
217 | * Avoid migrating a page that is shared by others and not writable. | |
218 | */ | |
219 | if ((flags & MPOL_MF_MOVE_ALL) || !page->mapping || PageAnon(page) || | |
220 | mapping_writably_mapped(page->mapping) || | |
221 | single_mm_mapping(vma->vm_mm, page->mapping)) { | |
222 | int rc = isolate_lru_page(page); | |
223 | ||
224 | if (rc == 1) | |
225 | list_add(&page->lru, pagelist); | |
226 | /* | |
227 | * If the isolate attempt was not successful then we just | |
228 | * encountered an unswappable page. Something must be wrong. | |
229 | */ | |
230 | WARN_ON(rc == 0); | |
231 | } | |
232 | } | |
233 | ||
1a75a6c8 CL |
234 | static void gather_stats(struct page *, void *); |
235 | ||
38e35860 | 236 | /* Scan through pages checking if pages follow certain conditions. */ |
b5810039 | 237 | static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd, |
dc9aa5b9 CL |
238 | unsigned long addr, unsigned long end, |
239 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 240 | void *private) |
1da177e4 | 241 | { |
91612e0d HD |
242 | pte_t *orig_pte; |
243 | pte_t *pte; | |
705e87c0 | 244 | spinlock_t *ptl; |
941150a3 | 245 | |
705e87c0 | 246 | orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
91612e0d | 247 | do { |
6aab341e | 248 | struct page *page; |
91612e0d HD |
249 | unsigned int nid; |
250 | ||
251 | if (!pte_present(*pte)) | |
1da177e4 | 252 | continue; |
6aab341e LT |
253 | page = vm_normal_page(vma, addr, *pte); |
254 | if (!page) | |
1da177e4 | 255 | continue; |
6aab341e | 256 | nid = page_to_nid(page); |
38e35860 CL |
257 | if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT)) |
258 | continue; | |
259 | ||
1a75a6c8 CL |
260 | if (flags & MPOL_MF_STATS) |
261 | gather_stats(page, private); | |
132beacf CL |
262 | else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
263 | spin_unlock(ptl); | |
38e35860 | 264 | migrate_page_add(vma, page, private, flags); |
132beacf CL |
265 | spin_lock(ptl); |
266 | } | |
38e35860 CL |
267 | else |
268 | break; | |
91612e0d | 269 | } while (pte++, addr += PAGE_SIZE, addr != end); |
705e87c0 | 270 | pte_unmap_unlock(orig_pte, ptl); |
91612e0d HD |
271 | return addr != end; |
272 | } | |
273 | ||
b5810039 | 274 | static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud, |
dc9aa5b9 CL |
275 | unsigned long addr, unsigned long end, |
276 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 277 | void *private) |
91612e0d HD |
278 | { |
279 | pmd_t *pmd; | |
280 | unsigned long next; | |
281 | ||
282 | pmd = pmd_offset(pud, addr); | |
283 | do { | |
284 | next = pmd_addr_end(addr, end); | |
285 | if (pmd_none_or_clear_bad(pmd)) | |
286 | continue; | |
dc9aa5b9 | 287 | if (check_pte_range(vma, pmd, addr, next, nodes, |
38e35860 | 288 | flags, private)) |
91612e0d HD |
289 | return -EIO; |
290 | } while (pmd++, addr = next, addr != end); | |
291 | return 0; | |
292 | } | |
293 | ||
b5810039 | 294 | static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd, |
dc9aa5b9 CL |
295 | unsigned long addr, unsigned long end, |
296 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 297 | void *private) |
91612e0d HD |
298 | { |
299 | pud_t *pud; | |
300 | unsigned long next; | |
301 | ||
302 | pud = pud_offset(pgd, addr); | |
303 | do { | |
304 | next = pud_addr_end(addr, end); | |
305 | if (pud_none_or_clear_bad(pud)) | |
306 | continue; | |
dc9aa5b9 | 307 | if (check_pmd_range(vma, pud, addr, next, nodes, |
38e35860 | 308 | flags, private)) |
91612e0d HD |
309 | return -EIO; |
310 | } while (pud++, addr = next, addr != end); | |
311 | return 0; | |
312 | } | |
313 | ||
b5810039 | 314 | static inline int check_pgd_range(struct vm_area_struct *vma, |
dc9aa5b9 CL |
315 | unsigned long addr, unsigned long end, |
316 | const nodemask_t *nodes, unsigned long flags, | |
38e35860 | 317 | void *private) |
91612e0d HD |
318 | { |
319 | pgd_t *pgd; | |
320 | unsigned long next; | |
321 | ||
b5810039 | 322 | pgd = pgd_offset(vma->vm_mm, addr); |
91612e0d HD |
323 | do { |
324 | next = pgd_addr_end(addr, end); | |
325 | if (pgd_none_or_clear_bad(pgd)) | |
326 | continue; | |
dc9aa5b9 | 327 | if (check_pud_range(vma, pgd, addr, next, nodes, |
38e35860 | 328 | flags, private)) |
91612e0d HD |
329 | return -EIO; |
330 | } while (pgd++, addr = next, addr != end); | |
331 | return 0; | |
1da177e4 LT |
332 | } |
333 | ||
dc9aa5b9 CL |
334 | /* Check if a vma is migratable */ |
335 | static inline int vma_migratable(struct vm_area_struct *vma) | |
336 | { | |
337 | if (vma->vm_flags & ( | |
338 | VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP)) | |
339 | return 0; | |
340 | return 1; | |
341 | } | |
342 | ||
343 | /* | |
344 | * Check if all pages in a range are on a set of nodes. | |
345 | * If pagelist != NULL then isolate pages from the LRU and | |
346 | * put them on the pagelist. | |
347 | */ | |
1da177e4 LT |
348 | static struct vm_area_struct * |
349 | check_range(struct mm_struct *mm, unsigned long start, unsigned long end, | |
38e35860 | 350 | const nodemask_t *nodes, unsigned long flags, void *private) |
1da177e4 LT |
351 | { |
352 | int err; | |
353 | struct vm_area_struct *first, *vma, *prev; | |
354 | ||
355 | first = find_vma(mm, start); | |
356 | if (!first) | |
357 | return ERR_PTR(-EFAULT); | |
358 | prev = NULL; | |
359 | for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) { | |
dc9aa5b9 CL |
360 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { |
361 | if (!vma->vm_next && vma->vm_end < end) | |
362 | return ERR_PTR(-EFAULT); | |
363 | if (prev && prev->vm_end < vma->vm_start) | |
364 | return ERR_PTR(-EFAULT); | |
365 | } | |
366 | if (!is_vm_hugetlb_page(vma) && | |
367 | ((flags & MPOL_MF_STRICT) || | |
368 | ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) && | |
369 | vma_migratable(vma)))) { | |
5b952b3c | 370 | unsigned long endvma = vma->vm_end; |
dc9aa5b9 | 371 | |
5b952b3c AK |
372 | if (endvma > end) |
373 | endvma = end; | |
374 | if (vma->vm_start > start) | |
375 | start = vma->vm_start; | |
dc9aa5b9 | 376 | err = check_pgd_range(vma, start, endvma, nodes, |
38e35860 | 377 | flags, private); |
1da177e4 LT |
378 | if (err) { |
379 | first = ERR_PTR(err); | |
380 | break; | |
381 | } | |
382 | } | |
383 | prev = vma; | |
384 | } | |
385 | return first; | |
386 | } | |
387 | ||
388 | /* Apply policy to a single VMA */ | |
389 | static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new) | |
390 | { | |
391 | int err = 0; | |
392 | struct mempolicy *old = vma->vm_policy; | |
393 | ||
394 | PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", | |
395 | vma->vm_start, vma->vm_end, vma->vm_pgoff, | |
396 | vma->vm_ops, vma->vm_file, | |
397 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
398 | ||
399 | if (vma->vm_ops && vma->vm_ops->set_policy) | |
400 | err = vma->vm_ops->set_policy(vma, new); | |
401 | if (!err) { | |
402 | mpol_get(new); | |
403 | vma->vm_policy = new; | |
404 | mpol_free(old); | |
405 | } | |
406 | return err; | |
407 | } | |
408 | ||
409 | /* Step 2: apply policy to a range and do splits. */ | |
410 | static int mbind_range(struct vm_area_struct *vma, unsigned long start, | |
411 | unsigned long end, struct mempolicy *new) | |
412 | { | |
413 | struct vm_area_struct *next; | |
414 | int err; | |
415 | ||
416 | err = 0; | |
417 | for (; vma && vma->vm_start < end; vma = next) { | |
418 | next = vma->vm_next; | |
419 | if (vma->vm_start < start) | |
420 | err = split_vma(vma->vm_mm, vma, start, 1); | |
421 | if (!err && vma->vm_end > end) | |
422 | err = split_vma(vma->vm_mm, vma, end, 0); | |
423 | if (!err) | |
424 | err = policy_vma(vma, new); | |
425 | if (err) | |
426 | break; | |
427 | } | |
428 | return err; | |
429 | } | |
430 | ||
8bccd85f CL |
431 | static int contextualize_policy(int mode, nodemask_t *nodes) |
432 | { | |
433 | if (!nodes) | |
434 | return 0; | |
435 | ||
436 | /* Update current mems_allowed */ | |
437 | cpuset_update_current_mems_allowed(); | |
438 | /* Ignore nodes not set in current->mems_allowed */ | |
439 | cpuset_restrict_to_mems_allowed(nodes->bits); | |
440 | return mpol_check_policy(mode, nodes); | |
441 | } | |
442 | ||
d4984711 CL |
443 | static int swap_pages(struct list_head *pagelist) |
444 | { | |
445 | LIST_HEAD(moved); | |
446 | LIST_HEAD(failed); | |
447 | int n; | |
448 | ||
449 | n = migrate_pages(pagelist, NULL, &moved, &failed); | |
450 | putback_lru_pages(&failed); | |
451 | putback_lru_pages(&moved); | |
452 | ||
453 | return n; | |
454 | } | |
455 | ||
8bccd85f CL |
456 | long do_mbind(unsigned long start, unsigned long len, |
457 | unsigned long mode, nodemask_t *nmask, unsigned long flags) | |
1da177e4 LT |
458 | { |
459 | struct vm_area_struct *vma; | |
460 | struct mm_struct *mm = current->mm; | |
461 | struct mempolicy *new; | |
462 | unsigned long end; | |
1da177e4 | 463 | int err; |
dc9aa5b9 | 464 | LIST_HEAD(pagelist); |
1da177e4 | 465 | |
38e35860 CL |
466 | if ((flags & ~(unsigned long)(MPOL_MF_STRICT | |
467 | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
dc9aa5b9 | 468 | || mode > MPOL_MAX) |
1da177e4 | 469 | return -EINVAL; |
dc9aa5b9 CL |
470 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE)) |
471 | return -EPERM; | |
472 | ||
1da177e4 LT |
473 | if (start & ~PAGE_MASK) |
474 | return -EINVAL; | |
dc9aa5b9 | 475 | |
1da177e4 LT |
476 | if (mode == MPOL_DEFAULT) |
477 | flags &= ~MPOL_MF_STRICT; | |
dc9aa5b9 | 478 | |
1da177e4 LT |
479 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; |
480 | end = start + len; | |
dc9aa5b9 | 481 | |
1da177e4 LT |
482 | if (end < start) |
483 | return -EINVAL; | |
484 | if (end == start) | |
485 | return 0; | |
dc9aa5b9 | 486 | |
5fcbb230 | 487 | if (mpol_check_policy(mode, nmask)) |
8bccd85f | 488 | return -EINVAL; |
dc9aa5b9 | 489 | |
8bccd85f | 490 | new = mpol_new(mode, nmask); |
1da177e4 LT |
491 | if (IS_ERR(new)) |
492 | return PTR_ERR(new); | |
493 | ||
dc9aa5b9 CL |
494 | /* |
495 | * If we are using the default policy then operation | |
496 | * on discontinuous address spaces is okay after all | |
497 | */ | |
498 | if (!new) | |
499 | flags |= MPOL_MF_DISCONTIG_OK; | |
500 | ||
1da177e4 | 501 | PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len, |
dfcd3c0d | 502 | mode,nodes_addr(nodes)[0]); |
1da177e4 LT |
503 | |
504 | down_write(&mm->mmap_sem); | |
38e35860 CL |
505 | vma = check_range(mm, start, end, nmask, |
506 | flags | MPOL_MF_INVERT, &pagelist); | |
507 | ||
1da177e4 | 508 | err = PTR_ERR(vma); |
dc9aa5b9 | 509 | if (!IS_ERR(vma)) { |
d4984711 CL |
510 | int nr_failed = 0; |
511 | ||
1da177e4 | 512 | err = mbind_range(vma, start, end, new); |
dc9aa5b9 | 513 | if (!list_empty(&pagelist)) |
d4984711 CL |
514 | nr_failed = swap_pages(&pagelist); |
515 | ||
516 | if (!err && nr_failed && (flags & MPOL_MF_STRICT)) | |
dc9aa5b9 CL |
517 | err = -EIO; |
518 | } | |
519 | if (!list_empty(&pagelist)) | |
520 | putback_lru_pages(&pagelist); | |
521 | ||
1da177e4 LT |
522 | up_write(&mm->mmap_sem); |
523 | mpol_free(new); | |
524 | return err; | |
525 | } | |
526 | ||
527 | /* Set the process memory policy */ | |
8bccd85f | 528 | long do_set_mempolicy(int mode, nodemask_t *nodes) |
1da177e4 | 529 | { |
1da177e4 | 530 | struct mempolicy *new; |
1da177e4 | 531 | |
8bccd85f | 532 | if (contextualize_policy(mode, nodes)) |
1da177e4 | 533 | return -EINVAL; |
8bccd85f | 534 | new = mpol_new(mode, nodes); |
1da177e4 LT |
535 | if (IS_ERR(new)) |
536 | return PTR_ERR(new); | |
537 | mpol_free(current->mempolicy); | |
538 | current->mempolicy = new; | |
539 | if (new && new->policy == MPOL_INTERLEAVE) | |
dfcd3c0d | 540 | current->il_next = first_node(new->v.nodes); |
1da177e4 LT |
541 | return 0; |
542 | } | |
543 | ||
544 | /* Fill a zone bitmap for a policy */ | |
dfcd3c0d | 545 | static void get_zonemask(struct mempolicy *p, nodemask_t *nodes) |
1da177e4 LT |
546 | { |
547 | int i; | |
548 | ||
dfcd3c0d | 549 | nodes_clear(*nodes); |
1da177e4 LT |
550 | switch (p->policy) { |
551 | case MPOL_BIND: | |
552 | for (i = 0; p->v.zonelist->zones[i]; i++) | |
8bccd85f CL |
553 | node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id, |
554 | *nodes); | |
1da177e4 LT |
555 | break; |
556 | case MPOL_DEFAULT: | |
557 | break; | |
558 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 559 | *nodes = p->v.nodes; |
1da177e4 LT |
560 | break; |
561 | case MPOL_PREFERRED: | |
562 | /* or use current node instead of online map? */ | |
563 | if (p->v.preferred_node < 0) | |
dfcd3c0d | 564 | *nodes = node_online_map; |
1da177e4 | 565 | else |
dfcd3c0d | 566 | node_set(p->v.preferred_node, *nodes); |
1da177e4 LT |
567 | break; |
568 | default: | |
569 | BUG(); | |
570 | } | |
571 | } | |
572 | ||
573 | static int lookup_node(struct mm_struct *mm, unsigned long addr) | |
574 | { | |
575 | struct page *p; | |
576 | int err; | |
577 | ||
578 | err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL); | |
579 | if (err >= 0) { | |
580 | err = page_to_nid(p); | |
581 | put_page(p); | |
582 | } | |
583 | return err; | |
584 | } | |
585 | ||
1da177e4 | 586 | /* Retrieve NUMA policy */ |
8bccd85f CL |
587 | long do_get_mempolicy(int *policy, nodemask_t *nmask, |
588 | unsigned long addr, unsigned long flags) | |
1da177e4 | 589 | { |
8bccd85f | 590 | int err; |
1da177e4 LT |
591 | struct mm_struct *mm = current->mm; |
592 | struct vm_area_struct *vma = NULL; | |
593 | struct mempolicy *pol = current->mempolicy; | |
594 | ||
68860ec1 | 595 | cpuset_update_current_mems_allowed(); |
1da177e4 LT |
596 | if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR)) |
597 | return -EINVAL; | |
1da177e4 LT |
598 | if (flags & MPOL_F_ADDR) { |
599 | down_read(&mm->mmap_sem); | |
600 | vma = find_vma_intersection(mm, addr, addr+1); | |
601 | if (!vma) { | |
602 | up_read(&mm->mmap_sem); | |
603 | return -EFAULT; | |
604 | } | |
605 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
606 | pol = vma->vm_ops->get_policy(vma, addr); | |
607 | else | |
608 | pol = vma->vm_policy; | |
609 | } else if (addr) | |
610 | return -EINVAL; | |
611 | ||
612 | if (!pol) | |
613 | pol = &default_policy; | |
614 | ||
615 | if (flags & MPOL_F_NODE) { | |
616 | if (flags & MPOL_F_ADDR) { | |
617 | err = lookup_node(mm, addr); | |
618 | if (err < 0) | |
619 | goto out; | |
8bccd85f | 620 | *policy = err; |
1da177e4 LT |
621 | } else if (pol == current->mempolicy && |
622 | pol->policy == MPOL_INTERLEAVE) { | |
8bccd85f | 623 | *policy = current->il_next; |
1da177e4 LT |
624 | } else { |
625 | err = -EINVAL; | |
626 | goto out; | |
627 | } | |
628 | } else | |
8bccd85f | 629 | *policy = pol->policy; |
1da177e4 LT |
630 | |
631 | if (vma) { | |
632 | up_read(¤t->mm->mmap_sem); | |
633 | vma = NULL; | |
634 | } | |
635 | ||
1da177e4 | 636 | err = 0; |
8bccd85f CL |
637 | if (nmask) |
638 | get_zonemask(pol, nmask); | |
1da177e4 LT |
639 | |
640 | out: | |
641 | if (vma) | |
642 | up_read(¤t->mm->mmap_sem); | |
643 | return err; | |
644 | } | |
645 | ||
39743889 CL |
646 | /* |
647 | * For now migrate_pages simply swaps out the pages from nodes that are in | |
648 | * the source set but not in the target set. In the future, we would | |
649 | * want a function that moves pages between the two nodesets in such | |
650 | * a way as to preserve the physical layout as much as possible. | |
651 | * | |
652 | * Returns the number of page that could not be moved. | |
653 | */ | |
654 | int do_migrate_pages(struct mm_struct *mm, | |
655 | const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags) | |
656 | { | |
657 | LIST_HEAD(pagelist); | |
658 | int count = 0; | |
659 | nodemask_t nodes; | |
660 | ||
661 | nodes_andnot(nodes, *from_nodes, *to_nodes); | |
39743889 CL |
662 | |
663 | down_read(&mm->mmap_sem); | |
664 | check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nodes, | |
665 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); | |
d4984711 | 666 | |
39743889 | 667 | if (!list_empty(&pagelist)) { |
d4984711 CL |
668 | count = swap_pages(&pagelist); |
669 | putback_lru_pages(&pagelist); | |
39743889 | 670 | } |
d4984711 | 671 | |
39743889 CL |
672 | up_read(&mm->mmap_sem); |
673 | return count; | |
674 | } | |
675 | ||
8bccd85f CL |
676 | /* |
677 | * User space interface with variable sized bitmaps for nodelists. | |
678 | */ | |
679 | ||
680 | /* Copy a node mask from user space. */ | |
39743889 | 681 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
8bccd85f CL |
682 | unsigned long maxnode) |
683 | { | |
684 | unsigned long k; | |
685 | unsigned long nlongs; | |
686 | unsigned long endmask; | |
687 | ||
688 | --maxnode; | |
689 | nodes_clear(*nodes); | |
690 | if (maxnode == 0 || !nmask) | |
691 | return 0; | |
692 | ||
693 | nlongs = BITS_TO_LONGS(maxnode); | |
694 | if ((maxnode % BITS_PER_LONG) == 0) | |
695 | endmask = ~0UL; | |
696 | else | |
697 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
698 | ||
699 | /* When the user specified more nodes than supported just check | |
700 | if the non supported part is all zero. */ | |
701 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
702 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
703 | return -EINVAL; | |
704 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
705 | unsigned long t; | |
706 | if (get_user(t, nmask + k)) | |
707 | return -EFAULT; | |
708 | if (k == nlongs - 1) { | |
709 | if (t & endmask) | |
710 | return -EINVAL; | |
711 | } else if (t) | |
712 | return -EINVAL; | |
713 | } | |
714 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
715 | endmask = ~0UL; | |
716 | } | |
717 | ||
718 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
719 | return -EFAULT; | |
720 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
721 | return 0; | |
722 | } | |
723 | ||
724 | /* Copy a kernel node mask to user space */ | |
725 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
726 | nodemask_t *nodes) | |
727 | { | |
728 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
729 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
730 | ||
731 | if (copy > nbytes) { | |
732 | if (copy > PAGE_SIZE) | |
733 | return -EINVAL; | |
734 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
735 | return -EFAULT; | |
736 | copy = nbytes; | |
737 | } | |
738 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
739 | } | |
740 | ||
741 | asmlinkage long sys_mbind(unsigned long start, unsigned long len, | |
742 | unsigned long mode, | |
743 | unsigned long __user *nmask, unsigned long maxnode, | |
744 | unsigned flags) | |
745 | { | |
746 | nodemask_t nodes; | |
747 | int err; | |
748 | ||
749 | err = get_nodes(&nodes, nmask, maxnode); | |
750 | if (err) | |
751 | return err; | |
752 | return do_mbind(start, len, mode, &nodes, flags); | |
753 | } | |
754 | ||
755 | /* Set the process memory policy */ | |
756 | asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask, | |
757 | unsigned long maxnode) | |
758 | { | |
759 | int err; | |
760 | nodemask_t nodes; | |
761 | ||
762 | if (mode < 0 || mode > MPOL_MAX) | |
763 | return -EINVAL; | |
764 | err = get_nodes(&nodes, nmask, maxnode); | |
765 | if (err) | |
766 | return err; | |
767 | return do_set_mempolicy(mode, &nodes); | |
768 | } | |
769 | ||
39743889 CL |
770 | /* Macro needed until Paul implements this function in kernel/cpusets.c */ |
771 | #define cpuset_mems_allowed(task) node_online_map | |
772 | ||
773 | asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode, | |
774 | const unsigned long __user *old_nodes, | |
775 | const unsigned long __user *new_nodes) | |
776 | { | |
777 | struct mm_struct *mm; | |
778 | struct task_struct *task; | |
779 | nodemask_t old; | |
780 | nodemask_t new; | |
781 | nodemask_t task_nodes; | |
782 | int err; | |
783 | ||
784 | err = get_nodes(&old, old_nodes, maxnode); | |
785 | if (err) | |
786 | return err; | |
787 | ||
788 | err = get_nodes(&new, new_nodes, maxnode); | |
789 | if (err) | |
790 | return err; | |
791 | ||
792 | /* Find the mm_struct */ | |
793 | read_lock(&tasklist_lock); | |
794 | task = pid ? find_task_by_pid(pid) : current; | |
795 | if (!task) { | |
796 | read_unlock(&tasklist_lock); | |
797 | return -ESRCH; | |
798 | } | |
799 | mm = get_task_mm(task); | |
800 | read_unlock(&tasklist_lock); | |
801 | ||
802 | if (!mm) | |
803 | return -EINVAL; | |
804 | ||
805 | /* | |
806 | * Check if this process has the right to modify the specified | |
807 | * process. The right exists if the process has administrative | |
808 | * capabilities, superuser priviledges or the same | |
809 | * userid as the target process. | |
810 | */ | |
811 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
812 | (current->uid != task->suid) && (current->uid != task->uid) && | |
813 | !capable(CAP_SYS_ADMIN)) { | |
814 | err = -EPERM; | |
815 | goto out; | |
816 | } | |
817 | ||
818 | task_nodes = cpuset_mems_allowed(task); | |
819 | /* Is the user allowed to access the target nodes? */ | |
820 | if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) { | |
821 | err = -EPERM; | |
822 | goto out; | |
823 | } | |
824 | ||
825 | err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE); | |
826 | out: | |
827 | mmput(mm); | |
828 | return err; | |
829 | } | |
830 | ||
831 | ||
8bccd85f CL |
832 | /* Retrieve NUMA policy */ |
833 | asmlinkage long sys_get_mempolicy(int __user *policy, | |
834 | unsigned long __user *nmask, | |
835 | unsigned long maxnode, | |
836 | unsigned long addr, unsigned long flags) | |
837 | { | |
838 | int err, pval; | |
839 | nodemask_t nodes; | |
840 | ||
841 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
842 | return -EINVAL; | |
843 | ||
844 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
845 | ||
846 | if (err) | |
847 | return err; | |
848 | ||
849 | if (policy && put_user(pval, policy)) | |
850 | return -EFAULT; | |
851 | ||
852 | if (nmask) | |
853 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
854 | ||
855 | return err; | |
856 | } | |
857 | ||
1da177e4 LT |
858 | #ifdef CONFIG_COMPAT |
859 | ||
860 | asmlinkage long compat_sys_get_mempolicy(int __user *policy, | |
861 | compat_ulong_t __user *nmask, | |
862 | compat_ulong_t maxnode, | |
863 | compat_ulong_t addr, compat_ulong_t flags) | |
864 | { | |
865 | long err; | |
866 | unsigned long __user *nm = NULL; | |
867 | unsigned long nr_bits, alloc_size; | |
868 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
869 | ||
870 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
871 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
872 | ||
873 | if (nmask) | |
874 | nm = compat_alloc_user_space(alloc_size); | |
875 | ||
876 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
877 | ||
878 | if (!err && nmask) { | |
879 | err = copy_from_user(bm, nm, alloc_size); | |
880 | /* ensure entire bitmap is zeroed */ | |
881 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
882 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
883 | } | |
884 | ||
885 | return err; | |
886 | } | |
887 | ||
888 | asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask, | |
889 | compat_ulong_t maxnode) | |
890 | { | |
891 | long err = 0; | |
892 | unsigned long __user *nm = NULL; | |
893 | unsigned long nr_bits, alloc_size; | |
894 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
895 | ||
896 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
897 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
898 | ||
899 | if (nmask) { | |
900 | err = compat_get_bitmap(bm, nmask, nr_bits); | |
901 | nm = compat_alloc_user_space(alloc_size); | |
902 | err |= copy_to_user(nm, bm, alloc_size); | |
903 | } | |
904 | ||
905 | if (err) | |
906 | return -EFAULT; | |
907 | ||
908 | return sys_set_mempolicy(mode, nm, nr_bits+1); | |
909 | } | |
910 | ||
911 | asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len, | |
912 | compat_ulong_t mode, compat_ulong_t __user *nmask, | |
913 | compat_ulong_t maxnode, compat_ulong_t flags) | |
914 | { | |
915 | long err = 0; | |
916 | unsigned long __user *nm = NULL; | |
917 | unsigned long nr_bits, alloc_size; | |
dfcd3c0d | 918 | nodemask_t bm; |
1da177e4 LT |
919 | |
920 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
921 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
922 | ||
923 | if (nmask) { | |
dfcd3c0d | 924 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); |
1da177e4 | 925 | nm = compat_alloc_user_space(alloc_size); |
dfcd3c0d | 926 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); |
1da177e4 LT |
927 | } |
928 | ||
929 | if (err) | |
930 | return -EFAULT; | |
931 | ||
932 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); | |
933 | } | |
934 | ||
935 | #endif | |
936 | ||
937 | /* Return effective policy for a VMA */ | |
48fce342 CL |
938 | static struct mempolicy * get_vma_policy(struct task_struct *task, |
939 | struct vm_area_struct *vma, unsigned long addr) | |
1da177e4 | 940 | { |
6e21c8f1 | 941 | struct mempolicy *pol = task->mempolicy; |
1da177e4 LT |
942 | |
943 | if (vma) { | |
944 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
8bccd85f | 945 | pol = vma->vm_ops->get_policy(vma, addr); |
1da177e4 LT |
946 | else if (vma->vm_policy && |
947 | vma->vm_policy->policy != MPOL_DEFAULT) | |
948 | pol = vma->vm_policy; | |
949 | } | |
950 | if (!pol) | |
951 | pol = &default_policy; | |
952 | return pol; | |
953 | } | |
954 | ||
955 | /* Return a zonelist representing a mempolicy */ | |
dd0fc66f | 956 | static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy) |
1da177e4 LT |
957 | { |
958 | int nd; | |
959 | ||
960 | switch (policy->policy) { | |
961 | case MPOL_PREFERRED: | |
962 | nd = policy->v.preferred_node; | |
963 | if (nd < 0) | |
964 | nd = numa_node_id(); | |
965 | break; | |
966 | case MPOL_BIND: | |
967 | /* Lower zones don't get a policy applied */ | |
968 | /* Careful: current->mems_allowed might have moved */ | |
af4ca457 | 969 | if (gfp_zone(gfp) >= policy_zone) |
1da177e4 LT |
970 | if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist)) |
971 | return policy->v.zonelist; | |
972 | /*FALL THROUGH*/ | |
973 | case MPOL_INTERLEAVE: /* should not happen */ | |
974 | case MPOL_DEFAULT: | |
975 | nd = numa_node_id(); | |
976 | break; | |
977 | default: | |
978 | nd = 0; | |
979 | BUG(); | |
980 | } | |
af4ca457 | 981 | return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp); |
1da177e4 LT |
982 | } |
983 | ||
984 | /* Do dynamic interleaving for a process */ | |
985 | static unsigned interleave_nodes(struct mempolicy *policy) | |
986 | { | |
987 | unsigned nid, next; | |
988 | struct task_struct *me = current; | |
989 | ||
990 | nid = me->il_next; | |
dfcd3c0d | 991 | next = next_node(nid, policy->v.nodes); |
1da177e4 | 992 | if (next >= MAX_NUMNODES) |
dfcd3c0d | 993 | next = first_node(policy->v.nodes); |
1da177e4 LT |
994 | me->il_next = next; |
995 | return nid; | |
996 | } | |
997 | ||
998 | /* Do static interleaving for a VMA with known offset. */ | |
999 | static unsigned offset_il_node(struct mempolicy *pol, | |
1000 | struct vm_area_struct *vma, unsigned long off) | |
1001 | { | |
dfcd3c0d | 1002 | unsigned nnodes = nodes_weight(pol->v.nodes); |
1da177e4 LT |
1003 | unsigned target = (unsigned)off % nnodes; |
1004 | int c; | |
1005 | int nid = -1; | |
1006 | ||
1007 | c = 0; | |
1008 | do { | |
dfcd3c0d | 1009 | nid = next_node(nid, pol->v.nodes); |
1da177e4 LT |
1010 | c++; |
1011 | } while (c <= target); | |
1da177e4 LT |
1012 | return nid; |
1013 | } | |
1014 | ||
5da7ca86 CL |
1015 | /* Determine a node number for interleave */ |
1016 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1017 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1018 | { | |
1019 | if (vma) { | |
1020 | unsigned long off; | |
1021 | ||
1022 | off = vma->vm_pgoff; | |
1023 | off += (addr - vma->vm_start) >> shift; | |
1024 | return offset_il_node(pol, vma, off); | |
1025 | } else | |
1026 | return interleave_nodes(pol); | |
1027 | } | |
1028 | ||
1029 | /* Return a zonelist suitable for a huge page allocation. */ | |
1030 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr) | |
1031 | { | |
1032 | struct mempolicy *pol = get_vma_policy(current, vma, addr); | |
1033 | ||
1034 | if (pol->policy == MPOL_INTERLEAVE) { | |
1035 | unsigned nid; | |
1036 | ||
1037 | nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT); | |
1038 | return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER); | |
1039 | } | |
1040 | return zonelist_policy(GFP_HIGHUSER, pol); | |
1041 | } | |
1042 | ||
1da177e4 LT |
1043 | /* Allocate a page in interleaved policy. |
1044 | Own path because it needs to do special accounting. */ | |
662f3a0b AK |
1045 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
1046 | unsigned nid) | |
1da177e4 LT |
1047 | { |
1048 | struct zonelist *zl; | |
1049 | struct page *page; | |
1050 | ||
af4ca457 | 1051 | zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp); |
1da177e4 LT |
1052 | page = __alloc_pages(gfp, order, zl); |
1053 | if (page && page_zone(page) == zl->zones[0]) { | |
e7c8d5c9 | 1054 | zone_pcp(zl->zones[0],get_cpu())->interleave_hit++; |
1da177e4 LT |
1055 | put_cpu(); |
1056 | } | |
1057 | return page; | |
1058 | } | |
1059 | ||
1060 | /** | |
1061 | * alloc_page_vma - Allocate a page for a VMA. | |
1062 | * | |
1063 | * @gfp: | |
1064 | * %GFP_USER user allocation. | |
1065 | * %GFP_KERNEL kernel allocations, | |
1066 | * %GFP_HIGHMEM highmem/user allocations, | |
1067 | * %GFP_FS allocation should not call back into a file system. | |
1068 | * %GFP_ATOMIC don't sleep. | |
1069 | * | |
1070 | * @vma: Pointer to VMA or NULL if not available. | |
1071 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
1072 | * | |
1073 | * This function allocates a page from the kernel page pool and applies | |
1074 | * a NUMA policy associated with the VMA or the current process. | |
1075 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1076 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
1077 | * all allocations for pages that will be mapped into | |
1078 | * user space. Returns NULL when no page can be allocated. | |
1079 | * | |
1080 | * Should be called with the mm_sem of the vma hold. | |
1081 | */ | |
1082 | struct page * | |
dd0fc66f | 1083 | alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr) |
1da177e4 | 1084 | { |
6e21c8f1 | 1085 | struct mempolicy *pol = get_vma_policy(current, vma, addr); |
1da177e4 LT |
1086 | |
1087 | cpuset_update_current_mems_allowed(); | |
1088 | ||
1089 | if (unlikely(pol->policy == MPOL_INTERLEAVE)) { | |
1090 | unsigned nid; | |
5da7ca86 CL |
1091 | |
1092 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT); | |
1da177e4 LT |
1093 | return alloc_page_interleave(gfp, 0, nid); |
1094 | } | |
1095 | return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol)); | |
1096 | } | |
1097 | ||
1098 | /** | |
1099 | * alloc_pages_current - Allocate pages. | |
1100 | * | |
1101 | * @gfp: | |
1102 | * %GFP_USER user allocation, | |
1103 | * %GFP_KERNEL kernel allocation, | |
1104 | * %GFP_HIGHMEM highmem allocation, | |
1105 | * %GFP_FS don't call back into a file system. | |
1106 | * %GFP_ATOMIC don't sleep. | |
1107 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
1108 | * | |
1109 | * Allocate a page from the kernel page pool. When not in | |
1110 | * interrupt context and apply the current process NUMA policy. | |
1111 | * Returns NULL when no page can be allocated. | |
1112 | * | |
1113 | * Don't call cpuset_update_current_mems_allowed() unless | |
1114 | * 1) it's ok to take cpuset_sem (can WAIT), and | |
1115 | * 2) allocating for current task (not interrupt). | |
1116 | */ | |
dd0fc66f | 1117 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
1da177e4 LT |
1118 | { |
1119 | struct mempolicy *pol = current->mempolicy; | |
1120 | ||
1121 | if ((gfp & __GFP_WAIT) && !in_interrupt()) | |
1122 | cpuset_update_current_mems_allowed(); | |
1123 | if (!pol || in_interrupt()) | |
1124 | pol = &default_policy; | |
1125 | if (pol->policy == MPOL_INTERLEAVE) | |
1126 | return alloc_page_interleave(gfp, order, interleave_nodes(pol)); | |
1127 | return __alloc_pages(gfp, order, zonelist_policy(gfp, pol)); | |
1128 | } | |
1129 | EXPORT_SYMBOL(alloc_pages_current); | |
1130 | ||
1131 | /* Slow path of a mempolicy copy */ | |
1132 | struct mempolicy *__mpol_copy(struct mempolicy *old) | |
1133 | { | |
1134 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
1135 | ||
1136 | if (!new) | |
1137 | return ERR_PTR(-ENOMEM); | |
1138 | *new = *old; | |
1139 | atomic_set(&new->refcnt, 1); | |
1140 | if (new->policy == MPOL_BIND) { | |
1141 | int sz = ksize(old->v.zonelist); | |
1142 | new->v.zonelist = kmalloc(sz, SLAB_KERNEL); | |
1143 | if (!new->v.zonelist) { | |
1144 | kmem_cache_free(policy_cache, new); | |
1145 | return ERR_PTR(-ENOMEM); | |
1146 | } | |
1147 | memcpy(new->v.zonelist, old->v.zonelist, sz); | |
1148 | } | |
1149 | return new; | |
1150 | } | |
1151 | ||
1152 | /* Slow path of a mempolicy comparison */ | |
1153 | int __mpol_equal(struct mempolicy *a, struct mempolicy *b) | |
1154 | { | |
1155 | if (!a || !b) | |
1156 | return 0; | |
1157 | if (a->policy != b->policy) | |
1158 | return 0; | |
1159 | switch (a->policy) { | |
1160 | case MPOL_DEFAULT: | |
1161 | return 1; | |
1162 | case MPOL_INTERLEAVE: | |
dfcd3c0d | 1163 | return nodes_equal(a->v.nodes, b->v.nodes); |
1da177e4 LT |
1164 | case MPOL_PREFERRED: |
1165 | return a->v.preferred_node == b->v.preferred_node; | |
1166 | case MPOL_BIND: { | |
1167 | int i; | |
1168 | for (i = 0; a->v.zonelist->zones[i]; i++) | |
1169 | if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i]) | |
1170 | return 0; | |
1171 | return b->v.zonelist->zones[i] == NULL; | |
1172 | } | |
1173 | default: | |
1174 | BUG(); | |
1175 | return 0; | |
1176 | } | |
1177 | } | |
1178 | ||
1179 | /* Slow path of a mpol destructor. */ | |
1180 | void __mpol_free(struct mempolicy *p) | |
1181 | { | |
1182 | if (!atomic_dec_and_test(&p->refcnt)) | |
1183 | return; | |
1184 | if (p->policy == MPOL_BIND) | |
1185 | kfree(p->v.zonelist); | |
1186 | p->policy = MPOL_DEFAULT; | |
1187 | kmem_cache_free(policy_cache, p); | |
1188 | } | |
1189 | ||
1da177e4 LT |
1190 | /* |
1191 | * Shared memory backing store policy support. | |
1192 | * | |
1193 | * Remember policies even when nobody has shared memory mapped. | |
1194 | * The policies are kept in Red-Black tree linked from the inode. | |
1195 | * They are protected by the sp->lock spinlock, which should be held | |
1196 | * for any accesses to the tree. | |
1197 | */ | |
1198 | ||
1199 | /* lookup first element intersecting start-end */ | |
1200 | /* Caller holds sp->lock */ | |
1201 | static struct sp_node * | |
1202 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
1203 | { | |
1204 | struct rb_node *n = sp->root.rb_node; | |
1205 | ||
1206 | while (n) { | |
1207 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
1208 | ||
1209 | if (start >= p->end) | |
1210 | n = n->rb_right; | |
1211 | else if (end <= p->start) | |
1212 | n = n->rb_left; | |
1213 | else | |
1214 | break; | |
1215 | } | |
1216 | if (!n) | |
1217 | return NULL; | |
1218 | for (;;) { | |
1219 | struct sp_node *w = NULL; | |
1220 | struct rb_node *prev = rb_prev(n); | |
1221 | if (!prev) | |
1222 | break; | |
1223 | w = rb_entry(prev, struct sp_node, nd); | |
1224 | if (w->end <= start) | |
1225 | break; | |
1226 | n = prev; | |
1227 | } | |
1228 | return rb_entry(n, struct sp_node, nd); | |
1229 | } | |
1230 | ||
1231 | /* Insert a new shared policy into the list. */ | |
1232 | /* Caller holds sp->lock */ | |
1233 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) | |
1234 | { | |
1235 | struct rb_node **p = &sp->root.rb_node; | |
1236 | struct rb_node *parent = NULL; | |
1237 | struct sp_node *nd; | |
1238 | ||
1239 | while (*p) { | |
1240 | parent = *p; | |
1241 | nd = rb_entry(parent, struct sp_node, nd); | |
1242 | if (new->start < nd->start) | |
1243 | p = &(*p)->rb_left; | |
1244 | else if (new->end > nd->end) | |
1245 | p = &(*p)->rb_right; | |
1246 | else | |
1247 | BUG(); | |
1248 | } | |
1249 | rb_link_node(&new->nd, parent, p); | |
1250 | rb_insert_color(&new->nd, &sp->root); | |
1251 | PDprintk("inserting %lx-%lx: %d\n", new->start, new->end, | |
1252 | new->policy ? new->policy->policy : 0); | |
1253 | } | |
1254 | ||
1255 | /* Find shared policy intersecting idx */ | |
1256 | struct mempolicy * | |
1257 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
1258 | { | |
1259 | struct mempolicy *pol = NULL; | |
1260 | struct sp_node *sn; | |
1261 | ||
1262 | if (!sp->root.rb_node) | |
1263 | return NULL; | |
1264 | spin_lock(&sp->lock); | |
1265 | sn = sp_lookup(sp, idx, idx+1); | |
1266 | if (sn) { | |
1267 | mpol_get(sn->policy); | |
1268 | pol = sn->policy; | |
1269 | } | |
1270 | spin_unlock(&sp->lock); | |
1271 | return pol; | |
1272 | } | |
1273 | ||
1274 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) | |
1275 | { | |
1276 | PDprintk("deleting %lx-l%x\n", n->start, n->end); | |
1277 | rb_erase(&n->nd, &sp->root); | |
1278 | mpol_free(n->policy); | |
1279 | kmem_cache_free(sn_cache, n); | |
1280 | } | |
1281 | ||
1282 | struct sp_node * | |
1283 | sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol) | |
1284 | { | |
1285 | struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
1286 | ||
1287 | if (!n) | |
1288 | return NULL; | |
1289 | n->start = start; | |
1290 | n->end = end; | |
1291 | mpol_get(pol); | |
1292 | n->policy = pol; | |
1293 | return n; | |
1294 | } | |
1295 | ||
1296 | /* Replace a policy range. */ | |
1297 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
1298 | unsigned long end, struct sp_node *new) | |
1299 | { | |
1300 | struct sp_node *n, *new2 = NULL; | |
1301 | ||
1302 | restart: | |
1303 | spin_lock(&sp->lock); | |
1304 | n = sp_lookup(sp, start, end); | |
1305 | /* Take care of old policies in the same range. */ | |
1306 | while (n && n->start < end) { | |
1307 | struct rb_node *next = rb_next(&n->nd); | |
1308 | if (n->start >= start) { | |
1309 | if (n->end <= end) | |
1310 | sp_delete(sp, n); | |
1311 | else | |
1312 | n->start = end; | |
1313 | } else { | |
1314 | /* Old policy spanning whole new range. */ | |
1315 | if (n->end > end) { | |
1316 | if (!new2) { | |
1317 | spin_unlock(&sp->lock); | |
1318 | new2 = sp_alloc(end, n->end, n->policy); | |
1319 | if (!new2) | |
1320 | return -ENOMEM; | |
1321 | goto restart; | |
1322 | } | |
1323 | n->end = start; | |
1324 | sp_insert(sp, new2); | |
1325 | new2 = NULL; | |
1326 | break; | |
1327 | } else | |
1328 | n->end = start; | |
1329 | } | |
1330 | if (!next) | |
1331 | break; | |
1332 | n = rb_entry(next, struct sp_node, nd); | |
1333 | } | |
1334 | if (new) | |
1335 | sp_insert(sp, new); | |
1336 | spin_unlock(&sp->lock); | |
1337 | if (new2) { | |
1338 | mpol_free(new2->policy); | |
1339 | kmem_cache_free(sn_cache, new2); | |
1340 | } | |
1341 | return 0; | |
1342 | } | |
1343 | ||
1344 | int mpol_set_shared_policy(struct shared_policy *info, | |
1345 | struct vm_area_struct *vma, struct mempolicy *npol) | |
1346 | { | |
1347 | int err; | |
1348 | struct sp_node *new = NULL; | |
1349 | unsigned long sz = vma_pages(vma); | |
1350 | ||
1351 | PDprintk("set_shared_policy %lx sz %lu %d %lx\n", | |
1352 | vma->vm_pgoff, | |
1353 | sz, npol? npol->policy : -1, | |
dfcd3c0d | 1354 | npol ? nodes_addr(npol->v.nodes)[0] : -1); |
1da177e4 LT |
1355 | |
1356 | if (npol) { | |
1357 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
1358 | if (!new) | |
1359 | return -ENOMEM; | |
1360 | } | |
1361 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
1362 | if (err && new) | |
1363 | kmem_cache_free(sn_cache, new); | |
1364 | return err; | |
1365 | } | |
1366 | ||
1367 | /* Free a backing policy store on inode delete. */ | |
1368 | void mpol_free_shared_policy(struct shared_policy *p) | |
1369 | { | |
1370 | struct sp_node *n; | |
1371 | struct rb_node *next; | |
1372 | ||
1373 | if (!p->root.rb_node) | |
1374 | return; | |
1375 | spin_lock(&p->lock); | |
1376 | next = rb_first(&p->root); | |
1377 | while (next) { | |
1378 | n = rb_entry(next, struct sp_node, nd); | |
1379 | next = rb_next(&n->nd); | |
90c5029e | 1380 | rb_erase(&n->nd, &p->root); |
1da177e4 LT |
1381 | mpol_free(n->policy); |
1382 | kmem_cache_free(sn_cache, n); | |
1383 | } | |
1384 | spin_unlock(&p->lock); | |
1da177e4 LT |
1385 | } |
1386 | ||
1387 | /* assumes fs == KERNEL_DS */ | |
1388 | void __init numa_policy_init(void) | |
1389 | { | |
1390 | policy_cache = kmem_cache_create("numa_policy", | |
1391 | sizeof(struct mempolicy), | |
1392 | 0, SLAB_PANIC, NULL, NULL); | |
1393 | ||
1394 | sn_cache = kmem_cache_create("shared_policy_node", | |
1395 | sizeof(struct sp_node), | |
1396 | 0, SLAB_PANIC, NULL, NULL); | |
1397 | ||
1398 | /* Set interleaving policy for system init. This way not all | |
1399 | the data structures allocated at system boot end up in node zero. */ | |
1400 | ||
8bccd85f | 1401 | if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map)) |
1da177e4 LT |
1402 | printk("numa_policy_init: interleaving failed\n"); |
1403 | } | |
1404 | ||
8bccd85f | 1405 | /* Reset policy of current process to default */ |
1da177e4 LT |
1406 | void numa_default_policy(void) |
1407 | { | |
8bccd85f | 1408 | do_set_mempolicy(MPOL_DEFAULT, NULL); |
1da177e4 | 1409 | } |
68860ec1 PJ |
1410 | |
1411 | /* Migrate a policy to a different set of nodes */ | |
1412 | static void rebind_policy(struct mempolicy *pol, const nodemask_t *old, | |
1413 | const nodemask_t *new) | |
1414 | { | |
1415 | nodemask_t tmp; | |
1416 | ||
1417 | if (!pol) | |
1418 | return; | |
1419 | ||
1420 | switch (pol->policy) { | |
1421 | case MPOL_DEFAULT: | |
1422 | break; | |
1423 | case MPOL_INTERLEAVE: | |
1424 | nodes_remap(tmp, pol->v.nodes, *old, *new); | |
1425 | pol->v.nodes = tmp; | |
1426 | current->il_next = node_remap(current->il_next, *old, *new); | |
1427 | break; | |
1428 | case MPOL_PREFERRED: | |
1429 | pol->v.preferred_node = node_remap(pol->v.preferred_node, | |
1430 | *old, *new); | |
1431 | break; | |
1432 | case MPOL_BIND: { | |
1433 | nodemask_t nodes; | |
1434 | struct zone **z; | |
1435 | struct zonelist *zonelist; | |
1436 | ||
1437 | nodes_clear(nodes); | |
1438 | for (z = pol->v.zonelist->zones; *z; z++) | |
1439 | node_set((*z)->zone_pgdat->node_id, nodes); | |
1440 | nodes_remap(tmp, nodes, *old, *new); | |
1441 | nodes = tmp; | |
1442 | ||
1443 | zonelist = bind_zonelist(&nodes); | |
1444 | ||
1445 | /* If no mem, then zonelist is NULL and we keep old zonelist. | |
1446 | * If that old zonelist has no remaining mems_allowed nodes, | |
1447 | * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT. | |
1448 | */ | |
1449 | ||
1450 | if (zonelist) { | |
1451 | /* Good - got mem - substitute new zonelist */ | |
1452 | kfree(pol->v.zonelist); | |
1453 | pol->v.zonelist = zonelist; | |
1454 | } | |
1455 | break; | |
1456 | } | |
1457 | default: | |
1458 | BUG(); | |
1459 | break; | |
1460 | } | |
1461 | } | |
1462 | ||
1463 | /* | |
1464 | * Someone moved this task to different nodes. Fixup mempolicies. | |
1465 | * | |
1466 | * TODO - fixup current->mm->vma and shmfs/tmpfs/hugetlbfs policies as well, | |
1467 | * once we have a cpuset mechanism to mark which cpuset subtree is migrating. | |
1468 | */ | |
1469 | void numa_policy_rebind(const nodemask_t *old, const nodemask_t *new) | |
1470 | { | |
1471 | rebind_policy(current->mempolicy, old, new); | |
1472 | } | |
1a75a6c8 CL |
1473 | |
1474 | /* | |
1475 | * Display pages allocated per node and memory policy via /proc. | |
1476 | */ | |
1477 | ||
1478 | static const char *policy_types[] = { "default", "prefer", "bind", | |
1479 | "interleave" }; | |
1480 | ||
1481 | /* | |
1482 | * Convert a mempolicy into a string. | |
1483 | * Returns the number of characters in buffer (if positive) | |
1484 | * or an error (negative) | |
1485 | */ | |
1486 | static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) | |
1487 | { | |
1488 | char *p = buffer; | |
1489 | int l; | |
1490 | nodemask_t nodes; | |
1491 | int mode = pol ? pol->policy : MPOL_DEFAULT; | |
1492 | ||
1493 | switch (mode) { | |
1494 | case MPOL_DEFAULT: | |
1495 | nodes_clear(nodes); | |
1496 | break; | |
1497 | ||
1498 | case MPOL_PREFERRED: | |
1499 | nodes_clear(nodes); | |
1500 | node_set(pol->v.preferred_node, nodes); | |
1501 | break; | |
1502 | ||
1503 | case MPOL_BIND: | |
1504 | get_zonemask(pol, &nodes); | |
1505 | break; | |
1506 | ||
1507 | case MPOL_INTERLEAVE: | |
1508 | nodes = pol->v.nodes; | |
1509 | break; | |
1510 | ||
1511 | default: | |
1512 | BUG(); | |
1513 | return -EFAULT; | |
1514 | } | |
1515 | ||
1516 | l = strlen(policy_types[mode]); | |
1517 | if (buffer + maxlen < p + l + 1) | |
1518 | return -ENOSPC; | |
1519 | ||
1520 | strcpy(p, policy_types[mode]); | |
1521 | p += l; | |
1522 | ||
1523 | if (!nodes_empty(nodes)) { | |
1524 | if (buffer + maxlen < p + 2) | |
1525 | return -ENOSPC; | |
1526 | *p++ = '='; | |
1527 | p += nodelist_scnprintf(p, buffer + maxlen - p, nodes); | |
1528 | } | |
1529 | return p - buffer; | |
1530 | } | |
1531 | ||
1532 | struct numa_maps { | |
1533 | unsigned long pages; | |
1534 | unsigned long anon; | |
1535 | unsigned long mapped; | |
1536 | unsigned long mapcount_max; | |
1537 | unsigned long node[MAX_NUMNODES]; | |
1538 | }; | |
1539 | ||
1540 | static void gather_stats(struct page *page, void *private) | |
1541 | { | |
1542 | struct numa_maps *md = private; | |
1543 | int count = page_mapcount(page); | |
1544 | ||
1545 | if (count) | |
1546 | md->mapped++; | |
1547 | ||
1548 | if (count > md->mapcount_max) | |
1549 | md->mapcount_max = count; | |
1550 | ||
1551 | md->pages++; | |
1552 | ||
1553 | if (PageAnon(page)) | |
1554 | md->anon++; | |
1555 | ||
1556 | md->node[page_to_nid(page)]++; | |
1557 | cond_resched(); | |
1558 | } | |
1559 | ||
1560 | int show_numa_map(struct seq_file *m, void *v) | |
1561 | { | |
1562 | struct task_struct *task = m->private; | |
1563 | struct vm_area_struct *vma = v; | |
1564 | struct numa_maps *md; | |
1565 | int n; | |
1566 | char buffer[50]; | |
1567 | ||
1568 | if (!vma->vm_mm) | |
1569 | return 0; | |
1570 | ||
1571 | md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL); | |
1572 | if (!md) | |
1573 | return 0; | |
1574 | ||
1575 | check_pgd_range(vma, vma->vm_start, vma->vm_end, | |
1576 | &node_online_map, MPOL_MF_STATS, md); | |
1577 | ||
1578 | if (md->pages) { | |
1579 | mpol_to_str(buffer, sizeof(buffer), | |
1580 | get_vma_policy(task, vma, vma->vm_start)); | |
1581 | ||
1582 | seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu", | |
1583 | vma->vm_start, buffer, md->pages, | |
1584 | md->mapped, md->mapcount_max); | |
1585 | ||
1586 | if (md->anon) | |
1587 | seq_printf(m," anon=%lu",md->anon); | |
1588 | ||
1589 | for_each_online_node(n) | |
1590 | if (md->node[n]) | |
1591 | seq_printf(m, " N%d=%lu", n, md->node[n]); | |
1592 | ||
1593 | seq_putc(m, '\n'); | |
1594 | } | |
1595 | kfree(md); | |
1596 | ||
1597 | if (m->count < m->size) | |
1598 | m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0; | |
1599 | return 0; | |
1600 | } | |
1601 |