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