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