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