]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Simple NUMA memory policy for the Linux kernel. | |
3 | * | |
4 | * Copyright 2003,2004 Andi Kleen, SuSE Labs. | |
5 | * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc. | |
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. | |
21 | * | |
22 | * bind Only allocate memory on a specific set of nodes, | |
23 | * no fallback. | |
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 | * | |
28 | * preferred Try a specific node first before normal fallback. | |
29 | * As a special case NUMA_NO_NODE 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. | |
33 | * | |
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 | */ | |
67 | ||
68 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
69 | ||
70 | #include <linux/mempolicy.h> | |
71 | #include <linux/mm.h> | |
72 | #include <linux/highmem.h> | |
73 | #include <linux/hugetlb.h> | |
74 | #include <linux/kernel.h> | |
75 | #include <linux/sched.h> | |
76 | #include <linux/nodemask.h> | |
77 | #include <linux/cpuset.h> | |
78 | #include <linux/slab.h> | |
79 | #include <linux/string.h> | |
80 | #include <linux/export.h> | |
81 | #include <linux/nsproxy.h> | |
82 | #include <linux/interrupt.h> | |
83 | #include <linux/init.h> | |
84 | #include <linux/compat.h> | |
85 | #include <linux/swap.h> | |
86 | #include <linux/seq_file.h> | |
87 | #include <linux/proc_fs.h> | |
88 | #include <linux/migrate.h> | |
89 | #include <linux/ksm.h> | |
90 | #include <linux/rmap.h> | |
91 | #include <linux/security.h> | |
92 | #include <linux/syscalls.h> | |
93 | #include <linux/ctype.h> | |
94 | #include <linux/mm_inline.h> | |
95 | #include <linux/mmu_notifier.h> | |
96 | #include <linux/printk.h> | |
97 | ||
98 | #include <asm/tlbflush.h> | |
99 | #include <asm/uaccess.h> | |
100 | ||
101 | #include "internal.h" | |
102 | ||
103 | /* Internal flags */ | |
104 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ | |
105 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ | |
106 | ||
107 | static struct kmem_cache *policy_cache; | |
108 | static struct kmem_cache *sn_cache; | |
109 | ||
110 | /* Highest zone. An specific allocation for a zone below that is not | |
111 | policied. */ | |
112 | enum zone_type policy_zone = 0; | |
113 | ||
114 | /* | |
115 | * run-time system-wide default policy => local allocation | |
116 | */ | |
117 | static struct mempolicy default_policy = { | |
118 | .refcnt = ATOMIC_INIT(1), /* never free it */ | |
119 | .mode = MPOL_PREFERRED, | |
120 | .flags = MPOL_F_LOCAL, | |
121 | }; | |
122 | ||
123 | static struct mempolicy preferred_node_policy[MAX_NUMNODES]; | |
124 | ||
125 | struct mempolicy *get_task_policy(struct task_struct *p) | |
126 | { | |
127 | struct mempolicy *pol = p->mempolicy; | |
128 | int node; | |
129 | ||
130 | if (pol) | |
131 | return pol; | |
132 | ||
133 | node = numa_node_id(); | |
134 | if (node != NUMA_NO_NODE) { | |
135 | pol = &preferred_node_policy[node]; | |
136 | /* preferred_node_policy is not initialised early in boot */ | |
137 | if (pol->mode) | |
138 | return pol; | |
139 | } | |
140 | ||
141 | return &default_policy; | |
142 | } | |
143 | ||
144 | static const struct mempolicy_operations { | |
145 | int (*create)(struct mempolicy *pol, const nodemask_t *nodes); | |
146 | /* | |
147 | * If read-side task has no lock to protect task->mempolicy, write-side | |
148 | * task will rebind the task->mempolicy by two step. The first step is | |
149 | * setting all the newly nodes, and the second step is cleaning all the | |
150 | * disallowed nodes. In this way, we can avoid finding no node to alloc | |
151 | * page. | |
152 | * If we have a lock to protect task->mempolicy in read-side, we do | |
153 | * rebind directly. | |
154 | * | |
155 | * step: | |
156 | * MPOL_REBIND_ONCE - do rebind work at once | |
157 | * MPOL_REBIND_STEP1 - set all the newly nodes | |
158 | * MPOL_REBIND_STEP2 - clean all the disallowed nodes | |
159 | */ | |
160 | void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes, | |
161 | enum mpol_rebind_step step); | |
162 | } mpol_ops[MPOL_MAX]; | |
163 | ||
164 | static inline int mpol_store_user_nodemask(const struct mempolicy *pol) | |
165 | { | |
166 | return pol->flags & MPOL_MODE_FLAGS; | |
167 | } | |
168 | ||
169 | static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig, | |
170 | const nodemask_t *rel) | |
171 | { | |
172 | nodemask_t tmp; | |
173 | nodes_fold(tmp, *orig, nodes_weight(*rel)); | |
174 | nodes_onto(*ret, tmp, *rel); | |
175 | } | |
176 | ||
177 | static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes) | |
178 | { | |
179 | if (nodes_empty(*nodes)) | |
180 | return -EINVAL; | |
181 | pol->v.nodes = *nodes; | |
182 | return 0; | |
183 | } | |
184 | ||
185 | static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes) | |
186 | { | |
187 | if (!nodes) | |
188 | pol->flags |= MPOL_F_LOCAL; /* local allocation */ | |
189 | else if (nodes_empty(*nodes)) | |
190 | return -EINVAL; /* no allowed nodes */ | |
191 | else | |
192 | pol->v.preferred_node = first_node(*nodes); | |
193 | return 0; | |
194 | } | |
195 | ||
196 | static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes) | |
197 | { | |
198 | if (nodes_empty(*nodes)) | |
199 | return -EINVAL; | |
200 | pol->v.nodes = *nodes; | |
201 | return 0; | |
202 | } | |
203 | ||
204 | /* | |
205 | * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if | |
206 | * any, for the new policy. mpol_new() has already validated the nodes | |
207 | * parameter with respect to the policy mode and flags. But, we need to | |
208 | * handle an empty nodemask with MPOL_PREFERRED here. | |
209 | * | |
210 | * Must be called holding task's alloc_lock to protect task's mems_allowed | |
211 | * and mempolicy. May also be called holding the mmap_semaphore for write. | |
212 | */ | |
213 | static int mpol_set_nodemask(struct mempolicy *pol, | |
214 | const nodemask_t *nodes, struct nodemask_scratch *nsc) | |
215 | { | |
216 | int ret; | |
217 | ||
218 | /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */ | |
219 | if (pol == NULL) | |
220 | return 0; | |
221 | /* Check N_MEMORY */ | |
222 | nodes_and(nsc->mask1, | |
223 | cpuset_current_mems_allowed, node_states[N_MEMORY]); | |
224 | ||
225 | VM_BUG_ON(!nodes); | |
226 | if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes)) | |
227 | nodes = NULL; /* explicit local allocation */ | |
228 | else { | |
229 | if (pol->flags & MPOL_F_RELATIVE_NODES) | |
230 | mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1); | |
231 | else | |
232 | nodes_and(nsc->mask2, *nodes, nsc->mask1); | |
233 | ||
234 | if (mpol_store_user_nodemask(pol)) | |
235 | pol->w.user_nodemask = *nodes; | |
236 | else | |
237 | pol->w.cpuset_mems_allowed = | |
238 | cpuset_current_mems_allowed; | |
239 | } | |
240 | ||
241 | if (nodes) | |
242 | ret = mpol_ops[pol->mode].create(pol, &nsc->mask2); | |
243 | else | |
244 | ret = mpol_ops[pol->mode].create(pol, NULL); | |
245 | return ret; | |
246 | } | |
247 | ||
248 | /* | |
249 | * This function just creates a new policy, does some check and simple | |
250 | * initialization. You must invoke mpol_set_nodemask() to set nodes. | |
251 | */ | |
252 | static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags, | |
253 | nodemask_t *nodes) | |
254 | { | |
255 | struct mempolicy *policy; | |
256 | ||
257 | pr_debug("setting mode %d flags %d nodes[0] %lx\n", | |
258 | mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE); | |
259 | ||
260 | if (mode == MPOL_DEFAULT) { | |
261 | if (nodes && !nodes_empty(*nodes)) | |
262 | return ERR_PTR(-EINVAL); | |
263 | return NULL; | |
264 | } | |
265 | VM_BUG_ON(!nodes); | |
266 | ||
267 | /* | |
268 | * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or | |
269 | * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation). | |
270 | * All other modes require a valid pointer to a non-empty nodemask. | |
271 | */ | |
272 | if (mode == MPOL_PREFERRED) { | |
273 | if (nodes_empty(*nodes)) { | |
274 | if (((flags & MPOL_F_STATIC_NODES) || | |
275 | (flags & MPOL_F_RELATIVE_NODES))) | |
276 | return ERR_PTR(-EINVAL); | |
277 | } | |
278 | } else if (mode == MPOL_LOCAL) { | |
279 | if (!nodes_empty(*nodes)) | |
280 | return ERR_PTR(-EINVAL); | |
281 | mode = MPOL_PREFERRED; | |
282 | } else if (nodes_empty(*nodes)) | |
283 | return ERR_PTR(-EINVAL); | |
284 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
285 | if (!policy) | |
286 | return ERR_PTR(-ENOMEM); | |
287 | atomic_set(&policy->refcnt, 1); | |
288 | policy->mode = mode; | |
289 | policy->flags = flags; | |
290 | ||
291 | return policy; | |
292 | } | |
293 | ||
294 | /* Slow path of a mpol destructor. */ | |
295 | void __mpol_put(struct mempolicy *p) | |
296 | { | |
297 | if (!atomic_dec_and_test(&p->refcnt)) | |
298 | return; | |
299 | kmem_cache_free(policy_cache, p); | |
300 | } | |
301 | ||
302 | static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes, | |
303 | enum mpol_rebind_step step) | |
304 | { | |
305 | } | |
306 | ||
307 | /* | |
308 | * step: | |
309 | * MPOL_REBIND_ONCE - do rebind work at once | |
310 | * MPOL_REBIND_STEP1 - set all the newly nodes | |
311 | * MPOL_REBIND_STEP2 - clean all the disallowed nodes | |
312 | */ | |
313 | static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes, | |
314 | enum mpol_rebind_step step) | |
315 | { | |
316 | nodemask_t tmp; | |
317 | ||
318 | if (pol->flags & MPOL_F_STATIC_NODES) | |
319 | nodes_and(tmp, pol->w.user_nodemask, *nodes); | |
320 | else if (pol->flags & MPOL_F_RELATIVE_NODES) | |
321 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); | |
322 | else { | |
323 | /* | |
324 | * if step == 1, we use ->w.cpuset_mems_allowed to cache the | |
325 | * result | |
326 | */ | |
327 | if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) { | |
328 | nodes_remap(tmp, pol->v.nodes, | |
329 | pol->w.cpuset_mems_allowed, *nodes); | |
330 | pol->w.cpuset_mems_allowed = step ? tmp : *nodes; | |
331 | } else if (step == MPOL_REBIND_STEP2) { | |
332 | tmp = pol->w.cpuset_mems_allowed; | |
333 | pol->w.cpuset_mems_allowed = *nodes; | |
334 | } else | |
335 | BUG(); | |
336 | } | |
337 | ||
338 | if (nodes_empty(tmp)) | |
339 | tmp = *nodes; | |
340 | ||
341 | if (step == MPOL_REBIND_STEP1) | |
342 | nodes_or(pol->v.nodes, pol->v.nodes, tmp); | |
343 | else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2) | |
344 | pol->v.nodes = tmp; | |
345 | else | |
346 | BUG(); | |
347 | ||
348 | if (!node_isset(current->il_next, tmp)) { | |
349 | current->il_next = next_node_in(current->il_next, tmp); | |
350 | if (current->il_next >= MAX_NUMNODES) | |
351 | current->il_next = numa_node_id(); | |
352 | } | |
353 | } | |
354 | ||
355 | static void mpol_rebind_preferred(struct mempolicy *pol, | |
356 | const nodemask_t *nodes, | |
357 | enum mpol_rebind_step step) | |
358 | { | |
359 | nodemask_t tmp; | |
360 | ||
361 | if (pol->flags & MPOL_F_STATIC_NODES) { | |
362 | int node = first_node(pol->w.user_nodemask); | |
363 | ||
364 | if (node_isset(node, *nodes)) { | |
365 | pol->v.preferred_node = node; | |
366 | pol->flags &= ~MPOL_F_LOCAL; | |
367 | } else | |
368 | pol->flags |= MPOL_F_LOCAL; | |
369 | } else if (pol->flags & MPOL_F_RELATIVE_NODES) { | |
370 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); | |
371 | pol->v.preferred_node = first_node(tmp); | |
372 | } else if (!(pol->flags & MPOL_F_LOCAL)) { | |
373 | pol->v.preferred_node = node_remap(pol->v.preferred_node, | |
374 | pol->w.cpuset_mems_allowed, | |
375 | *nodes); | |
376 | pol->w.cpuset_mems_allowed = *nodes; | |
377 | } | |
378 | } | |
379 | ||
380 | /* | |
381 | * mpol_rebind_policy - Migrate a policy to a different set of nodes | |
382 | * | |
383 | * If read-side task has no lock to protect task->mempolicy, write-side | |
384 | * task will rebind the task->mempolicy by two step. The first step is | |
385 | * setting all the newly nodes, and the second step is cleaning all the | |
386 | * disallowed nodes. In this way, we can avoid finding no node to alloc | |
387 | * page. | |
388 | * If we have a lock to protect task->mempolicy in read-side, we do | |
389 | * rebind directly. | |
390 | * | |
391 | * step: | |
392 | * MPOL_REBIND_ONCE - do rebind work at once | |
393 | * MPOL_REBIND_STEP1 - set all the newly nodes | |
394 | * MPOL_REBIND_STEP2 - clean all the disallowed nodes | |
395 | */ | |
396 | static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask, | |
397 | enum mpol_rebind_step step) | |
398 | { | |
399 | if (!pol) | |
400 | return; | |
401 | if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE && | |
402 | nodes_equal(pol->w.cpuset_mems_allowed, *newmask)) | |
403 | return; | |
404 | ||
405 | if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING)) | |
406 | return; | |
407 | ||
408 | if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING)) | |
409 | BUG(); | |
410 | ||
411 | if (step == MPOL_REBIND_STEP1) | |
412 | pol->flags |= MPOL_F_REBINDING; | |
413 | else if (step == MPOL_REBIND_STEP2) | |
414 | pol->flags &= ~MPOL_F_REBINDING; | |
415 | else if (step >= MPOL_REBIND_NSTEP) | |
416 | BUG(); | |
417 | ||
418 | mpol_ops[pol->mode].rebind(pol, newmask, step); | |
419 | } | |
420 | ||
421 | /* | |
422 | * Wrapper for mpol_rebind_policy() that just requires task | |
423 | * pointer, and updates task mempolicy. | |
424 | * | |
425 | * Called with task's alloc_lock held. | |
426 | */ | |
427 | ||
428 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new, | |
429 | enum mpol_rebind_step step) | |
430 | { | |
431 | mpol_rebind_policy(tsk->mempolicy, new, step); | |
432 | } | |
433 | ||
434 | /* | |
435 | * Rebind each vma in mm to new nodemask. | |
436 | * | |
437 | * Call holding a reference to mm. Takes mm->mmap_sem during call. | |
438 | */ | |
439 | ||
440 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) | |
441 | { | |
442 | struct vm_area_struct *vma; | |
443 | ||
444 | down_write(&mm->mmap_sem); | |
445 | for (vma = mm->mmap; vma; vma = vma->vm_next) | |
446 | mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE); | |
447 | up_write(&mm->mmap_sem); | |
448 | } | |
449 | ||
450 | static const struct mempolicy_operations mpol_ops[MPOL_MAX] = { | |
451 | [MPOL_DEFAULT] = { | |
452 | .rebind = mpol_rebind_default, | |
453 | }, | |
454 | [MPOL_INTERLEAVE] = { | |
455 | .create = mpol_new_interleave, | |
456 | .rebind = mpol_rebind_nodemask, | |
457 | }, | |
458 | [MPOL_PREFERRED] = { | |
459 | .create = mpol_new_preferred, | |
460 | .rebind = mpol_rebind_preferred, | |
461 | }, | |
462 | [MPOL_BIND] = { | |
463 | .create = mpol_new_bind, | |
464 | .rebind = mpol_rebind_nodemask, | |
465 | }, | |
466 | }; | |
467 | ||
468 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
469 | unsigned long flags); | |
470 | ||
471 | struct queue_pages { | |
472 | struct list_head *pagelist; | |
473 | unsigned long flags; | |
474 | nodemask_t *nmask; | |
475 | struct vm_area_struct *prev; | |
476 | }; | |
477 | ||
478 | /* | |
479 | * Scan through pages checking if pages follow certain conditions, | |
480 | * and move them to the pagelist if they do. | |
481 | */ | |
482 | static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr, | |
483 | unsigned long end, struct mm_walk *walk) | |
484 | { | |
485 | struct vm_area_struct *vma = walk->vma; | |
486 | struct page *page; | |
487 | struct queue_pages *qp = walk->private; | |
488 | unsigned long flags = qp->flags; | |
489 | int nid, ret; | |
490 | pte_t *pte; | |
491 | spinlock_t *ptl; | |
492 | ||
493 | if (pmd_trans_huge(*pmd)) { | |
494 | ptl = pmd_lock(walk->mm, pmd); | |
495 | if (pmd_trans_huge(*pmd)) { | |
496 | page = pmd_page(*pmd); | |
497 | if (is_huge_zero_page(page)) { | |
498 | spin_unlock(ptl); | |
499 | split_huge_pmd(vma, pmd, addr); | |
500 | } else { | |
501 | get_page(page); | |
502 | spin_unlock(ptl); | |
503 | lock_page(page); | |
504 | ret = split_huge_page(page); | |
505 | unlock_page(page); | |
506 | put_page(page); | |
507 | if (ret) | |
508 | return 0; | |
509 | } | |
510 | } else { | |
511 | spin_unlock(ptl); | |
512 | } | |
513 | } | |
514 | ||
515 | if (pmd_trans_unstable(pmd)) | |
516 | return 0; | |
517 | retry: | |
518 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); | |
519 | for (; addr != end; pte++, addr += PAGE_SIZE) { | |
520 | if (!pte_present(*pte)) | |
521 | continue; | |
522 | page = vm_normal_page(vma, addr, *pte); | |
523 | if (!page) | |
524 | continue; | |
525 | /* | |
526 | * vm_normal_page() filters out zero pages, but there might | |
527 | * still be PageReserved pages to skip, perhaps in a VDSO. | |
528 | */ | |
529 | if (PageReserved(page)) | |
530 | continue; | |
531 | nid = page_to_nid(page); | |
532 | if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT)) | |
533 | continue; | |
534 | if (PageTransCompound(page)) { | |
535 | get_page(page); | |
536 | pte_unmap_unlock(pte, ptl); | |
537 | lock_page(page); | |
538 | ret = split_huge_page(page); | |
539 | unlock_page(page); | |
540 | put_page(page); | |
541 | /* Failed to split -- skip. */ | |
542 | if (ret) { | |
543 | pte = pte_offset_map_lock(walk->mm, pmd, | |
544 | addr, &ptl); | |
545 | continue; | |
546 | } | |
547 | goto retry; | |
548 | } | |
549 | ||
550 | migrate_page_add(page, qp->pagelist, flags); | |
551 | } | |
552 | pte_unmap_unlock(pte - 1, ptl); | |
553 | cond_resched(); | |
554 | return 0; | |
555 | } | |
556 | ||
557 | static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask, | |
558 | unsigned long addr, unsigned long end, | |
559 | struct mm_walk *walk) | |
560 | { | |
561 | #ifdef CONFIG_HUGETLB_PAGE | |
562 | struct queue_pages *qp = walk->private; | |
563 | unsigned long flags = qp->flags; | |
564 | int nid; | |
565 | struct page *page; | |
566 | spinlock_t *ptl; | |
567 | pte_t entry; | |
568 | ||
569 | ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); | |
570 | entry = huge_ptep_get(pte); | |
571 | if (!pte_present(entry)) | |
572 | goto unlock; | |
573 | page = pte_page(entry); | |
574 | nid = page_to_nid(page); | |
575 | if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT)) | |
576 | goto unlock; | |
577 | /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */ | |
578 | if (flags & (MPOL_MF_MOVE_ALL) || | |
579 | (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) | |
580 | isolate_huge_page(page, qp->pagelist); | |
581 | unlock: | |
582 | spin_unlock(ptl); | |
583 | #else | |
584 | BUG(); | |
585 | #endif | |
586 | return 0; | |
587 | } | |
588 | ||
589 | #ifdef CONFIG_NUMA_BALANCING | |
590 | /* | |
591 | * This is used to mark a range of virtual addresses to be inaccessible. | |
592 | * These are later cleared by a NUMA hinting fault. Depending on these | |
593 | * faults, pages may be migrated for better NUMA placement. | |
594 | * | |
595 | * This is assuming that NUMA faults are handled using PROT_NONE. If | |
596 | * an architecture makes a different choice, it will need further | |
597 | * changes to the core. | |
598 | */ | |
599 | unsigned long change_prot_numa(struct vm_area_struct *vma, | |
600 | unsigned long addr, unsigned long end) | |
601 | { | |
602 | int nr_updated; | |
603 | ||
604 | nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1); | |
605 | if (nr_updated) | |
606 | count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated); | |
607 | ||
608 | return nr_updated; | |
609 | } | |
610 | #else | |
611 | static unsigned long change_prot_numa(struct vm_area_struct *vma, | |
612 | unsigned long addr, unsigned long end) | |
613 | { | |
614 | return 0; | |
615 | } | |
616 | #endif /* CONFIG_NUMA_BALANCING */ | |
617 | ||
618 | static int queue_pages_test_walk(unsigned long start, unsigned long end, | |
619 | struct mm_walk *walk) | |
620 | { | |
621 | struct vm_area_struct *vma = walk->vma; | |
622 | struct queue_pages *qp = walk->private; | |
623 | unsigned long endvma = vma->vm_end; | |
624 | unsigned long flags = qp->flags; | |
625 | ||
626 | if (!vma_migratable(vma)) | |
627 | return 1; | |
628 | ||
629 | if (endvma > end) | |
630 | endvma = end; | |
631 | if (vma->vm_start > start) | |
632 | start = vma->vm_start; | |
633 | ||
634 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { | |
635 | if (!vma->vm_next && vma->vm_end < end) | |
636 | return -EFAULT; | |
637 | if (qp->prev && qp->prev->vm_end < vma->vm_start) | |
638 | return -EFAULT; | |
639 | } | |
640 | ||
641 | qp->prev = vma; | |
642 | ||
643 | if (flags & MPOL_MF_LAZY) { | |
644 | /* Similar to task_numa_work, skip inaccessible VMAs */ | |
645 | if (!is_vm_hugetlb_page(vma) && | |
646 | (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) && | |
647 | !(vma->vm_flags & VM_MIXEDMAP)) | |
648 | change_prot_numa(vma, start, endvma); | |
649 | return 1; | |
650 | } | |
651 | ||
652 | /* queue pages from current vma */ | |
653 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
654 | return 0; | |
655 | return 1; | |
656 | } | |
657 | ||
658 | /* | |
659 | * Walk through page tables and collect pages to be migrated. | |
660 | * | |
661 | * If pages found in a given range are on a set of nodes (determined by | |
662 | * @nodes and @flags,) it's isolated and queued to the pagelist which is | |
663 | * passed via @private.) | |
664 | */ | |
665 | static int | |
666 | queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end, | |
667 | nodemask_t *nodes, unsigned long flags, | |
668 | struct list_head *pagelist) | |
669 | { | |
670 | struct queue_pages qp = { | |
671 | .pagelist = pagelist, | |
672 | .flags = flags, | |
673 | .nmask = nodes, | |
674 | .prev = NULL, | |
675 | }; | |
676 | struct mm_walk queue_pages_walk = { | |
677 | .hugetlb_entry = queue_pages_hugetlb, | |
678 | .pmd_entry = queue_pages_pte_range, | |
679 | .test_walk = queue_pages_test_walk, | |
680 | .mm = mm, | |
681 | .private = &qp, | |
682 | }; | |
683 | ||
684 | return walk_page_range(start, end, &queue_pages_walk); | |
685 | } | |
686 | ||
687 | /* | |
688 | * Apply policy to a single VMA | |
689 | * This must be called with the mmap_sem held for writing. | |
690 | */ | |
691 | static int vma_replace_policy(struct vm_area_struct *vma, | |
692 | struct mempolicy *pol) | |
693 | { | |
694 | int err; | |
695 | struct mempolicy *old; | |
696 | struct mempolicy *new; | |
697 | ||
698 | pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", | |
699 | vma->vm_start, vma->vm_end, vma->vm_pgoff, | |
700 | vma->vm_ops, vma->vm_file, | |
701 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
702 | ||
703 | new = mpol_dup(pol); | |
704 | if (IS_ERR(new)) | |
705 | return PTR_ERR(new); | |
706 | ||
707 | if (vma->vm_ops && vma->vm_ops->set_policy) { | |
708 | err = vma->vm_ops->set_policy(vma, new); | |
709 | if (err) | |
710 | goto err_out; | |
711 | } | |
712 | ||
713 | old = vma->vm_policy; | |
714 | vma->vm_policy = new; /* protected by mmap_sem */ | |
715 | mpol_put(old); | |
716 | ||
717 | return 0; | |
718 | err_out: | |
719 | mpol_put(new); | |
720 | return err; | |
721 | } | |
722 | ||
723 | /* Step 2: apply policy to a range and do splits. */ | |
724 | static int mbind_range(struct mm_struct *mm, unsigned long start, | |
725 | unsigned long end, struct mempolicy *new_pol) | |
726 | { | |
727 | struct vm_area_struct *next; | |
728 | struct vm_area_struct *prev; | |
729 | struct vm_area_struct *vma; | |
730 | int err = 0; | |
731 | pgoff_t pgoff; | |
732 | unsigned long vmstart; | |
733 | unsigned long vmend; | |
734 | ||
735 | vma = find_vma(mm, start); | |
736 | if (!vma || vma->vm_start > start) | |
737 | return -EFAULT; | |
738 | ||
739 | prev = vma->vm_prev; | |
740 | if (start > vma->vm_start) | |
741 | prev = vma; | |
742 | ||
743 | for (; vma && vma->vm_start < end; prev = vma, vma = next) { | |
744 | next = vma->vm_next; | |
745 | vmstart = max(start, vma->vm_start); | |
746 | vmend = min(end, vma->vm_end); | |
747 | ||
748 | if (mpol_equal(vma_policy(vma), new_pol)) | |
749 | continue; | |
750 | ||
751 | pgoff = vma->vm_pgoff + | |
752 | ((vmstart - vma->vm_start) >> PAGE_SHIFT); | |
753 | prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags, | |
754 | vma->anon_vma, vma->vm_file, pgoff, | |
755 | new_pol, vma->vm_userfaultfd_ctx); | |
756 | if (prev) { | |
757 | vma = prev; | |
758 | next = vma->vm_next; | |
759 | if (mpol_equal(vma_policy(vma), new_pol)) | |
760 | continue; | |
761 | /* vma_merge() joined vma && vma->next, case 8 */ | |
762 | goto replace; | |
763 | } | |
764 | if (vma->vm_start != vmstart) { | |
765 | err = split_vma(vma->vm_mm, vma, vmstart, 1); | |
766 | if (err) | |
767 | goto out; | |
768 | } | |
769 | if (vma->vm_end != vmend) { | |
770 | err = split_vma(vma->vm_mm, vma, vmend, 0); | |
771 | if (err) | |
772 | goto out; | |
773 | } | |
774 | replace: | |
775 | err = vma_replace_policy(vma, new_pol); | |
776 | if (err) | |
777 | goto out; | |
778 | } | |
779 | ||
780 | out: | |
781 | return err; | |
782 | } | |
783 | ||
784 | /* Set the process memory policy */ | |
785 | static long do_set_mempolicy(unsigned short mode, unsigned short flags, | |
786 | nodemask_t *nodes) | |
787 | { | |
788 | struct mempolicy *new, *old; | |
789 | NODEMASK_SCRATCH(scratch); | |
790 | int ret; | |
791 | ||
792 | if (!scratch) | |
793 | return -ENOMEM; | |
794 | ||
795 | new = mpol_new(mode, flags, nodes); | |
796 | if (IS_ERR(new)) { | |
797 | ret = PTR_ERR(new); | |
798 | goto out; | |
799 | } | |
800 | ||
801 | task_lock(current); | |
802 | ret = mpol_set_nodemask(new, nodes, scratch); | |
803 | if (ret) { | |
804 | task_unlock(current); | |
805 | mpol_put(new); | |
806 | goto out; | |
807 | } | |
808 | old = current->mempolicy; | |
809 | current->mempolicy = new; | |
810 | if (new && new->mode == MPOL_INTERLEAVE && | |
811 | nodes_weight(new->v.nodes)) | |
812 | current->il_next = first_node(new->v.nodes); | |
813 | task_unlock(current); | |
814 | mpol_put(old); | |
815 | ret = 0; | |
816 | out: | |
817 | NODEMASK_SCRATCH_FREE(scratch); | |
818 | return ret; | |
819 | } | |
820 | ||
821 | /* | |
822 | * Return nodemask for policy for get_mempolicy() query | |
823 | * | |
824 | * Called with task's alloc_lock held | |
825 | */ | |
826 | static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes) | |
827 | { | |
828 | nodes_clear(*nodes); | |
829 | if (p == &default_policy) | |
830 | return; | |
831 | ||
832 | switch (p->mode) { | |
833 | case MPOL_BIND: | |
834 | /* Fall through */ | |
835 | case MPOL_INTERLEAVE: | |
836 | *nodes = p->v.nodes; | |
837 | break; | |
838 | case MPOL_PREFERRED: | |
839 | if (!(p->flags & MPOL_F_LOCAL)) | |
840 | node_set(p->v.preferred_node, *nodes); | |
841 | /* else return empty node mask for local allocation */ | |
842 | break; | |
843 | default: | |
844 | BUG(); | |
845 | } | |
846 | } | |
847 | ||
848 | static int lookup_node(unsigned long addr) | |
849 | { | |
850 | struct page *p; | |
851 | int err; | |
852 | ||
853 | err = get_user_pages(addr & PAGE_MASK, 1, 0, 0, &p, NULL); | |
854 | if (err >= 0) { | |
855 | err = page_to_nid(p); | |
856 | put_page(p); | |
857 | } | |
858 | return err; | |
859 | } | |
860 | ||
861 | /* Retrieve NUMA policy */ | |
862 | static long do_get_mempolicy(int *policy, nodemask_t *nmask, | |
863 | unsigned long addr, unsigned long flags) | |
864 | { | |
865 | int err; | |
866 | struct mm_struct *mm = current->mm; | |
867 | struct vm_area_struct *vma = NULL; | |
868 | struct mempolicy *pol = current->mempolicy; | |
869 | ||
870 | if (flags & | |
871 | ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED)) | |
872 | return -EINVAL; | |
873 | ||
874 | if (flags & MPOL_F_MEMS_ALLOWED) { | |
875 | if (flags & (MPOL_F_NODE|MPOL_F_ADDR)) | |
876 | return -EINVAL; | |
877 | *policy = 0; /* just so it's initialized */ | |
878 | task_lock(current); | |
879 | *nmask = cpuset_current_mems_allowed; | |
880 | task_unlock(current); | |
881 | return 0; | |
882 | } | |
883 | ||
884 | if (flags & MPOL_F_ADDR) { | |
885 | /* | |
886 | * Do NOT fall back to task policy if the | |
887 | * vma/shared policy at addr is NULL. We | |
888 | * want to return MPOL_DEFAULT in this case. | |
889 | */ | |
890 | down_read(&mm->mmap_sem); | |
891 | vma = find_vma_intersection(mm, addr, addr+1); | |
892 | if (!vma) { | |
893 | up_read(&mm->mmap_sem); | |
894 | return -EFAULT; | |
895 | } | |
896 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
897 | pol = vma->vm_ops->get_policy(vma, addr); | |
898 | else | |
899 | pol = vma->vm_policy; | |
900 | } else if (addr) | |
901 | return -EINVAL; | |
902 | ||
903 | if (!pol) | |
904 | pol = &default_policy; /* indicates default behavior */ | |
905 | ||
906 | if (flags & MPOL_F_NODE) { | |
907 | if (flags & MPOL_F_ADDR) { | |
908 | err = lookup_node(addr); | |
909 | if (err < 0) | |
910 | goto out; | |
911 | *policy = err; | |
912 | } else if (pol == current->mempolicy && | |
913 | pol->mode == MPOL_INTERLEAVE) { | |
914 | *policy = current->il_next; | |
915 | } else { | |
916 | err = -EINVAL; | |
917 | goto out; | |
918 | } | |
919 | } else { | |
920 | *policy = pol == &default_policy ? MPOL_DEFAULT : | |
921 | pol->mode; | |
922 | /* | |
923 | * Internal mempolicy flags must be masked off before exposing | |
924 | * the policy to userspace. | |
925 | */ | |
926 | *policy |= (pol->flags & MPOL_MODE_FLAGS); | |
927 | } | |
928 | ||
929 | if (vma) { | |
930 | up_read(¤t->mm->mmap_sem); | |
931 | vma = NULL; | |
932 | } | |
933 | ||
934 | err = 0; | |
935 | if (nmask) { | |
936 | if (mpol_store_user_nodemask(pol)) { | |
937 | *nmask = pol->w.user_nodemask; | |
938 | } else { | |
939 | task_lock(current); | |
940 | get_policy_nodemask(pol, nmask); | |
941 | task_unlock(current); | |
942 | } | |
943 | } | |
944 | ||
945 | out: | |
946 | mpol_cond_put(pol); | |
947 | if (vma) | |
948 | up_read(¤t->mm->mmap_sem); | |
949 | return err; | |
950 | } | |
951 | ||
952 | #ifdef CONFIG_MIGRATION | |
953 | /* | |
954 | * page migration | |
955 | */ | |
956 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
957 | unsigned long flags) | |
958 | { | |
959 | /* | |
960 | * Avoid migrating a page that is shared with others. | |
961 | */ | |
962 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) { | |
963 | if (!isolate_lru_page(page)) { | |
964 | list_add_tail(&page->lru, pagelist); | |
965 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
966 | page_is_file_cache(page)); | |
967 | } | |
968 | } | |
969 | } | |
970 | ||
971 | static struct page *new_node_page(struct page *page, unsigned long node, int **x) | |
972 | { | |
973 | if (PageHuge(page)) | |
974 | return alloc_huge_page_node(page_hstate(compound_head(page)), | |
975 | node); | |
976 | else | |
977 | return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE | | |
978 | __GFP_THISNODE, 0); | |
979 | } | |
980 | ||
981 | /* | |
982 | * Migrate pages from one node to a target node. | |
983 | * Returns error or the number of pages not migrated. | |
984 | */ | |
985 | static int migrate_to_node(struct mm_struct *mm, int source, int dest, | |
986 | int flags) | |
987 | { | |
988 | nodemask_t nmask; | |
989 | LIST_HEAD(pagelist); | |
990 | int err = 0; | |
991 | ||
992 | nodes_clear(nmask); | |
993 | node_set(source, nmask); | |
994 | ||
995 | /* | |
996 | * This does not "check" the range but isolates all pages that | |
997 | * need migration. Between passing in the full user address | |
998 | * space range and MPOL_MF_DISCONTIG_OK, this call can not fail. | |
999 | */ | |
1000 | VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))); | |
1001 | queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask, | |
1002 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); | |
1003 | ||
1004 | if (!list_empty(&pagelist)) { | |
1005 | err = migrate_pages(&pagelist, new_node_page, NULL, dest, | |
1006 | MIGRATE_SYNC, MR_SYSCALL); | |
1007 | if (err) | |
1008 | putback_movable_pages(&pagelist); | |
1009 | } | |
1010 | ||
1011 | return err; | |
1012 | } | |
1013 | ||
1014 | /* | |
1015 | * Move pages between the two nodesets so as to preserve the physical | |
1016 | * layout as much as possible. | |
1017 | * | |
1018 | * Returns the number of page that could not be moved. | |
1019 | */ | |
1020 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, | |
1021 | const nodemask_t *to, int flags) | |
1022 | { | |
1023 | int busy = 0; | |
1024 | int err; | |
1025 | nodemask_t tmp; | |
1026 | ||
1027 | err = migrate_prep(); | |
1028 | if (err) | |
1029 | return err; | |
1030 | ||
1031 | down_read(&mm->mmap_sem); | |
1032 | ||
1033 | /* | |
1034 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' | |
1035 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' | |
1036 | * bit in 'tmp', and return that <source, dest> pair for migration. | |
1037 | * The pair of nodemasks 'to' and 'from' define the map. | |
1038 | * | |
1039 | * If no pair of bits is found that way, fallback to picking some | |
1040 | * pair of 'source' and 'dest' bits that are not the same. If the | |
1041 | * 'source' and 'dest' bits are the same, this represents a node | |
1042 | * that will be migrating to itself, so no pages need move. | |
1043 | * | |
1044 | * If no bits are left in 'tmp', or if all remaining bits left | |
1045 | * in 'tmp' correspond to the same bit in 'to', return false | |
1046 | * (nothing left to migrate). | |
1047 | * | |
1048 | * This lets us pick a pair of nodes to migrate between, such that | |
1049 | * if possible the dest node is not already occupied by some other | |
1050 | * source node, minimizing the risk of overloading the memory on a | |
1051 | * node that would happen if we migrated incoming memory to a node | |
1052 | * before migrating outgoing memory source that same node. | |
1053 | * | |
1054 | * A single scan of tmp is sufficient. As we go, we remember the | |
1055 | * most recent <s, d> pair that moved (s != d). If we find a pair | |
1056 | * that not only moved, but what's better, moved to an empty slot | |
1057 | * (d is not set in tmp), then we break out then, with that pair. | |
1058 | * Otherwise when we finish scanning from_tmp, we at least have the | |
1059 | * most recent <s, d> pair that moved. If we get all the way through | |
1060 | * the scan of tmp without finding any node that moved, much less | |
1061 | * moved to an empty node, then there is nothing left worth migrating. | |
1062 | */ | |
1063 | ||
1064 | tmp = *from; | |
1065 | while (!nodes_empty(tmp)) { | |
1066 | int s,d; | |
1067 | int source = NUMA_NO_NODE; | |
1068 | int dest = 0; | |
1069 | ||
1070 | for_each_node_mask(s, tmp) { | |
1071 | ||
1072 | /* | |
1073 | * do_migrate_pages() tries to maintain the relative | |
1074 | * node relationship of the pages established between | |
1075 | * threads and memory areas. | |
1076 | * | |
1077 | * However if the number of source nodes is not equal to | |
1078 | * the number of destination nodes we can not preserve | |
1079 | * this node relative relationship. In that case, skip | |
1080 | * copying memory from a node that is in the destination | |
1081 | * mask. | |
1082 | * | |
1083 | * Example: [2,3,4] -> [3,4,5] moves everything. | |
1084 | * [0-7] - > [3,4,5] moves only 0,1,2,6,7. | |
1085 | */ | |
1086 | ||
1087 | if ((nodes_weight(*from) != nodes_weight(*to)) && | |
1088 | (node_isset(s, *to))) | |
1089 | continue; | |
1090 | ||
1091 | d = node_remap(s, *from, *to); | |
1092 | if (s == d) | |
1093 | continue; | |
1094 | ||
1095 | source = s; /* Node moved. Memorize */ | |
1096 | dest = d; | |
1097 | ||
1098 | /* dest not in remaining from nodes? */ | |
1099 | if (!node_isset(dest, tmp)) | |
1100 | break; | |
1101 | } | |
1102 | if (source == NUMA_NO_NODE) | |
1103 | break; | |
1104 | ||
1105 | node_clear(source, tmp); | |
1106 | err = migrate_to_node(mm, source, dest, flags); | |
1107 | if (err > 0) | |
1108 | busy += err; | |
1109 | if (err < 0) | |
1110 | break; | |
1111 | } | |
1112 | up_read(&mm->mmap_sem); | |
1113 | if (err < 0) | |
1114 | return err; | |
1115 | return busy; | |
1116 | ||
1117 | } | |
1118 | ||
1119 | /* | |
1120 | * Allocate a new page for page migration based on vma policy. | |
1121 | * Start by assuming the page is mapped by the same vma as contains @start. | |
1122 | * Search forward from there, if not. N.B., this assumes that the | |
1123 | * list of pages handed to migrate_pages()--which is how we get here-- | |
1124 | * is in virtual address order. | |
1125 | */ | |
1126 | static struct page *new_page(struct page *page, unsigned long start, int **x) | |
1127 | { | |
1128 | struct vm_area_struct *vma; | |
1129 | unsigned long uninitialized_var(address); | |
1130 | ||
1131 | vma = find_vma(current->mm, start); | |
1132 | while (vma) { | |
1133 | address = page_address_in_vma(page, vma); | |
1134 | if (address != -EFAULT) | |
1135 | break; | |
1136 | vma = vma->vm_next; | |
1137 | } | |
1138 | ||
1139 | if (PageHuge(page)) { | |
1140 | BUG_ON(!vma); | |
1141 | return alloc_huge_page_noerr(vma, address, 1); | |
1142 | } | |
1143 | /* | |
1144 | * if !vma, alloc_page_vma() will use task or system default policy | |
1145 | */ | |
1146 | return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); | |
1147 | } | |
1148 | #else | |
1149 | ||
1150 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
1151 | unsigned long flags) | |
1152 | { | |
1153 | } | |
1154 | ||
1155 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, | |
1156 | const nodemask_t *to, int flags) | |
1157 | { | |
1158 | return -ENOSYS; | |
1159 | } | |
1160 | ||
1161 | static struct page *new_page(struct page *page, unsigned long start, int **x) | |
1162 | { | |
1163 | return NULL; | |
1164 | } | |
1165 | #endif | |
1166 | ||
1167 | static long do_mbind(unsigned long start, unsigned long len, | |
1168 | unsigned short mode, unsigned short mode_flags, | |
1169 | nodemask_t *nmask, unsigned long flags) | |
1170 | { | |
1171 | struct mm_struct *mm = current->mm; | |
1172 | struct mempolicy *new; | |
1173 | unsigned long end; | |
1174 | int err; | |
1175 | LIST_HEAD(pagelist); | |
1176 | ||
1177 | if (flags & ~(unsigned long)MPOL_MF_VALID) | |
1178 | return -EINVAL; | |
1179 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1180 | return -EPERM; | |
1181 | ||
1182 | if (start & ~PAGE_MASK) | |
1183 | return -EINVAL; | |
1184 | ||
1185 | if (mode == MPOL_DEFAULT) | |
1186 | flags &= ~MPOL_MF_STRICT; | |
1187 | ||
1188 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; | |
1189 | end = start + len; | |
1190 | ||
1191 | if (end < start) | |
1192 | return -EINVAL; | |
1193 | if (end == start) | |
1194 | return 0; | |
1195 | ||
1196 | new = mpol_new(mode, mode_flags, nmask); | |
1197 | if (IS_ERR(new)) | |
1198 | return PTR_ERR(new); | |
1199 | ||
1200 | if (flags & MPOL_MF_LAZY) | |
1201 | new->flags |= MPOL_F_MOF; | |
1202 | ||
1203 | /* | |
1204 | * If we are using the default policy then operation | |
1205 | * on discontinuous address spaces is okay after all | |
1206 | */ | |
1207 | if (!new) | |
1208 | flags |= MPOL_MF_DISCONTIG_OK; | |
1209 | ||
1210 | pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n", | |
1211 | start, start + len, mode, mode_flags, | |
1212 | nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE); | |
1213 | ||
1214 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { | |
1215 | ||
1216 | err = migrate_prep(); | |
1217 | if (err) | |
1218 | goto mpol_out; | |
1219 | } | |
1220 | { | |
1221 | NODEMASK_SCRATCH(scratch); | |
1222 | if (scratch) { | |
1223 | down_write(&mm->mmap_sem); | |
1224 | task_lock(current); | |
1225 | err = mpol_set_nodemask(new, nmask, scratch); | |
1226 | task_unlock(current); | |
1227 | if (err) | |
1228 | up_write(&mm->mmap_sem); | |
1229 | } else | |
1230 | err = -ENOMEM; | |
1231 | NODEMASK_SCRATCH_FREE(scratch); | |
1232 | } | |
1233 | if (err) | |
1234 | goto mpol_out; | |
1235 | ||
1236 | err = queue_pages_range(mm, start, end, nmask, | |
1237 | flags | MPOL_MF_INVERT, &pagelist); | |
1238 | if (!err) | |
1239 | err = mbind_range(mm, start, end, new); | |
1240 | ||
1241 | if (!err) { | |
1242 | int nr_failed = 0; | |
1243 | ||
1244 | if (!list_empty(&pagelist)) { | |
1245 | WARN_ON_ONCE(flags & MPOL_MF_LAZY); | |
1246 | nr_failed = migrate_pages(&pagelist, new_page, NULL, | |
1247 | start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND); | |
1248 | if (nr_failed) | |
1249 | putback_movable_pages(&pagelist); | |
1250 | } | |
1251 | ||
1252 | if (nr_failed && (flags & MPOL_MF_STRICT)) | |
1253 | err = -EIO; | |
1254 | } else | |
1255 | putback_movable_pages(&pagelist); | |
1256 | ||
1257 | up_write(&mm->mmap_sem); | |
1258 | mpol_out: | |
1259 | mpol_put(new); | |
1260 | return err; | |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * User space interface with variable sized bitmaps for nodelists. | |
1265 | */ | |
1266 | ||
1267 | /* Copy a node mask from user space. */ | |
1268 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, | |
1269 | unsigned long maxnode) | |
1270 | { | |
1271 | unsigned long k; | |
1272 | unsigned long nlongs; | |
1273 | unsigned long endmask; | |
1274 | ||
1275 | --maxnode; | |
1276 | nodes_clear(*nodes); | |
1277 | if (maxnode == 0 || !nmask) | |
1278 | return 0; | |
1279 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) | |
1280 | return -EINVAL; | |
1281 | ||
1282 | nlongs = BITS_TO_LONGS(maxnode); | |
1283 | if ((maxnode % BITS_PER_LONG) == 0) | |
1284 | endmask = ~0UL; | |
1285 | else | |
1286 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
1287 | ||
1288 | /* When the user specified more nodes than supported just check | |
1289 | if the non supported part is all zero. */ | |
1290 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
1291 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
1292 | return -EINVAL; | |
1293 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
1294 | unsigned long t; | |
1295 | if (get_user(t, nmask + k)) | |
1296 | return -EFAULT; | |
1297 | if (k == nlongs - 1) { | |
1298 | if (t & endmask) | |
1299 | return -EINVAL; | |
1300 | } else if (t) | |
1301 | return -EINVAL; | |
1302 | } | |
1303 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
1304 | endmask = ~0UL; | |
1305 | } | |
1306 | ||
1307 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
1308 | return -EFAULT; | |
1309 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
1310 | return 0; | |
1311 | } | |
1312 | ||
1313 | /* Copy a kernel node mask to user space */ | |
1314 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
1315 | nodemask_t *nodes) | |
1316 | { | |
1317 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
1318 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
1319 | ||
1320 | if (copy > nbytes) { | |
1321 | if (copy > PAGE_SIZE) | |
1322 | return -EINVAL; | |
1323 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
1324 | return -EFAULT; | |
1325 | copy = nbytes; | |
1326 | } | |
1327 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
1328 | } | |
1329 | ||
1330 | SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len, | |
1331 | unsigned long, mode, const unsigned long __user *, nmask, | |
1332 | unsigned long, maxnode, unsigned, flags) | |
1333 | { | |
1334 | nodemask_t nodes; | |
1335 | int err; | |
1336 | unsigned short mode_flags; | |
1337 | ||
1338 | mode_flags = mode & MPOL_MODE_FLAGS; | |
1339 | mode &= ~MPOL_MODE_FLAGS; | |
1340 | if (mode >= MPOL_MAX) | |
1341 | return -EINVAL; | |
1342 | if ((mode_flags & MPOL_F_STATIC_NODES) && | |
1343 | (mode_flags & MPOL_F_RELATIVE_NODES)) | |
1344 | return -EINVAL; | |
1345 | err = get_nodes(&nodes, nmask, maxnode); | |
1346 | if (err) | |
1347 | return err; | |
1348 | return do_mbind(start, len, mode, mode_flags, &nodes, flags); | |
1349 | } | |
1350 | ||
1351 | /* Set the process memory policy */ | |
1352 | SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask, | |
1353 | unsigned long, maxnode) | |
1354 | { | |
1355 | int err; | |
1356 | nodemask_t nodes; | |
1357 | unsigned short flags; | |
1358 | ||
1359 | flags = mode & MPOL_MODE_FLAGS; | |
1360 | mode &= ~MPOL_MODE_FLAGS; | |
1361 | if ((unsigned int)mode >= MPOL_MAX) | |
1362 | return -EINVAL; | |
1363 | if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES)) | |
1364 | return -EINVAL; | |
1365 | err = get_nodes(&nodes, nmask, maxnode); | |
1366 | if (err) | |
1367 | return err; | |
1368 | return do_set_mempolicy(mode, flags, &nodes); | |
1369 | } | |
1370 | ||
1371 | SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode, | |
1372 | const unsigned long __user *, old_nodes, | |
1373 | const unsigned long __user *, new_nodes) | |
1374 | { | |
1375 | const struct cred *cred = current_cred(), *tcred; | |
1376 | struct mm_struct *mm = NULL; | |
1377 | struct task_struct *task; | |
1378 | nodemask_t task_nodes; | |
1379 | int err; | |
1380 | nodemask_t *old; | |
1381 | nodemask_t *new; | |
1382 | NODEMASK_SCRATCH(scratch); | |
1383 | ||
1384 | if (!scratch) | |
1385 | return -ENOMEM; | |
1386 | ||
1387 | old = &scratch->mask1; | |
1388 | new = &scratch->mask2; | |
1389 | ||
1390 | err = get_nodes(old, old_nodes, maxnode); | |
1391 | if (err) | |
1392 | goto out; | |
1393 | ||
1394 | err = get_nodes(new, new_nodes, maxnode); | |
1395 | if (err) | |
1396 | goto out; | |
1397 | ||
1398 | /* Find the mm_struct */ | |
1399 | rcu_read_lock(); | |
1400 | task = pid ? find_task_by_vpid(pid) : current; | |
1401 | if (!task) { | |
1402 | rcu_read_unlock(); | |
1403 | err = -ESRCH; | |
1404 | goto out; | |
1405 | } | |
1406 | get_task_struct(task); | |
1407 | ||
1408 | err = -EINVAL; | |
1409 | ||
1410 | /* | |
1411 | * Check if this process has the right to modify the specified | |
1412 | * process. The right exists if the process has administrative | |
1413 | * capabilities, superuser privileges or the same | |
1414 | * userid as the target process. | |
1415 | */ | |
1416 | tcred = __task_cred(task); | |
1417 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && | |
1418 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
1419 | !capable(CAP_SYS_NICE)) { | |
1420 | rcu_read_unlock(); | |
1421 | err = -EPERM; | |
1422 | goto out_put; | |
1423 | } | |
1424 | rcu_read_unlock(); | |
1425 | ||
1426 | task_nodes = cpuset_mems_allowed(task); | |
1427 | /* Is the user allowed to access the target nodes? */ | |
1428 | if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) { | |
1429 | err = -EPERM; | |
1430 | goto out_put; | |
1431 | } | |
1432 | ||
1433 | if (!nodes_subset(*new, node_states[N_MEMORY])) { | |
1434 | err = -EINVAL; | |
1435 | goto out_put; | |
1436 | } | |
1437 | ||
1438 | err = security_task_movememory(task); | |
1439 | if (err) | |
1440 | goto out_put; | |
1441 | ||
1442 | mm = get_task_mm(task); | |
1443 | put_task_struct(task); | |
1444 | ||
1445 | if (!mm) { | |
1446 | err = -EINVAL; | |
1447 | goto out; | |
1448 | } | |
1449 | ||
1450 | err = do_migrate_pages(mm, old, new, | |
1451 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); | |
1452 | ||
1453 | mmput(mm); | |
1454 | out: | |
1455 | NODEMASK_SCRATCH_FREE(scratch); | |
1456 | ||
1457 | return err; | |
1458 | ||
1459 | out_put: | |
1460 | put_task_struct(task); | |
1461 | goto out; | |
1462 | ||
1463 | } | |
1464 | ||
1465 | ||
1466 | /* Retrieve NUMA policy */ | |
1467 | SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, | |
1468 | unsigned long __user *, nmask, unsigned long, maxnode, | |
1469 | unsigned long, addr, unsigned long, flags) | |
1470 | { | |
1471 | int err; | |
1472 | int uninitialized_var(pval); | |
1473 | nodemask_t nodes; | |
1474 | ||
1475 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
1476 | return -EINVAL; | |
1477 | ||
1478 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
1479 | ||
1480 | if (err) | |
1481 | return err; | |
1482 | ||
1483 | if (policy && put_user(pval, policy)) | |
1484 | return -EFAULT; | |
1485 | ||
1486 | if (nmask) | |
1487 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
1488 | ||
1489 | return err; | |
1490 | } | |
1491 | ||
1492 | #ifdef CONFIG_COMPAT | |
1493 | ||
1494 | COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, | |
1495 | compat_ulong_t __user *, nmask, | |
1496 | compat_ulong_t, maxnode, | |
1497 | compat_ulong_t, addr, compat_ulong_t, flags) | |
1498 | { | |
1499 | long err; | |
1500 | unsigned long __user *nm = NULL; | |
1501 | unsigned long nr_bits, alloc_size; | |
1502 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1503 | ||
1504 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1505 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1506 | ||
1507 | if (nmask) | |
1508 | nm = compat_alloc_user_space(alloc_size); | |
1509 | ||
1510 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
1511 | ||
1512 | if (!err && nmask) { | |
1513 | unsigned long copy_size; | |
1514 | copy_size = min_t(unsigned long, sizeof(bm), alloc_size); | |
1515 | err = copy_from_user(bm, nm, copy_size); | |
1516 | /* ensure entire bitmap is zeroed */ | |
1517 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
1518 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
1519 | } | |
1520 | ||
1521 | return err; | |
1522 | } | |
1523 | ||
1524 | COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask, | |
1525 | compat_ulong_t, maxnode) | |
1526 | { | |
1527 | long err = 0; | |
1528 | unsigned long __user *nm = NULL; | |
1529 | unsigned long nr_bits, alloc_size; | |
1530 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1531 | ||
1532 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1533 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1534 | ||
1535 | if (nmask) { | |
1536 | err = compat_get_bitmap(bm, nmask, nr_bits); | |
1537 | nm = compat_alloc_user_space(alloc_size); | |
1538 | err |= copy_to_user(nm, bm, alloc_size); | |
1539 | } | |
1540 | ||
1541 | if (err) | |
1542 | return -EFAULT; | |
1543 | ||
1544 | return sys_set_mempolicy(mode, nm, nr_bits+1); | |
1545 | } | |
1546 | ||
1547 | COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len, | |
1548 | compat_ulong_t, mode, compat_ulong_t __user *, nmask, | |
1549 | compat_ulong_t, maxnode, compat_ulong_t, flags) | |
1550 | { | |
1551 | long err = 0; | |
1552 | unsigned long __user *nm = NULL; | |
1553 | unsigned long nr_bits, alloc_size; | |
1554 | nodemask_t bm; | |
1555 | ||
1556 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1557 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1558 | ||
1559 | if (nmask) { | |
1560 | err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits); | |
1561 | nm = compat_alloc_user_space(alloc_size); | |
1562 | err |= copy_to_user(nm, nodes_addr(bm), alloc_size); | |
1563 | } | |
1564 | ||
1565 | if (err) | |
1566 | return -EFAULT; | |
1567 | ||
1568 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); | |
1569 | } | |
1570 | ||
1571 | #endif | |
1572 | ||
1573 | struct mempolicy *__get_vma_policy(struct vm_area_struct *vma, | |
1574 | unsigned long addr) | |
1575 | { | |
1576 | struct mempolicy *pol = NULL; | |
1577 | ||
1578 | if (vma) { | |
1579 | if (vma->vm_ops && vma->vm_ops->get_policy) { | |
1580 | pol = vma->vm_ops->get_policy(vma, addr); | |
1581 | } else if (vma->vm_policy) { | |
1582 | pol = vma->vm_policy; | |
1583 | ||
1584 | /* | |
1585 | * shmem_alloc_page() passes MPOL_F_SHARED policy with | |
1586 | * a pseudo vma whose vma->vm_ops=NULL. Take a reference | |
1587 | * count on these policies which will be dropped by | |
1588 | * mpol_cond_put() later | |
1589 | */ | |
1590 | if (mpol_needs_cond_ref(pol)) | |
1591 | mpol_get(pol); | |
1592 | } | |
1593 | } | |
1594 | ||
1595 | return pol; | |
1596 | } | |
1597 | ||
1598 | /* | |
1599 | * get_vma_policy(@vma, @addr) | |
1600 | * @vma: virtual memory area whose policy is sought | |
1601 | * @addr: address in @vma for shared policy lookup | |
1602 | * | |
1603 | * Returns effective policy for a VMA at specified address. | |
1604 | * Falls back to current->mempolicy or system default policy, as necessary. | |
1605 | * Shared policies [those marked as MPOL_F_SHARED] require an extra reference | |
1606 | * count--added by the get_policy() vm_op, as appropriate--to protect against | |
1607 | * freeing by another task. It is the caller's responsibility to free the | |
1608 | * extra reference for shared policies. | |
1609 | */ | |
1610 | static struct mempolicy *get_vma_policy(struct vm_area_struct *vma, | |
1611 | unsigned long addr) | |
1612 | { | |
1613 | struct mempolicy *pol = __get_vma_policy(vma, addr); | |
1614 | ||
1615 | if (!pol) | |
1616 | pol = get_task_policy(current); | |
1617 | ||
1618 | return pol; | |
1619 | } | |
1620 | ||
1621 | bool vma_policy_mof(struct vm_area_struct *vma) | |
1622 | { | |
1623 | struct mempolicy *pol; | |
1624 | ||
1625 | if (vma->vm_ops && vma->vm_ops->get_policy) { | |
1626 | bool ret = false; | |
1627 | ||
1628 | pol = vma->vm_ops->get_policy(vma, vma->vm_start); | |
1629 | if (pol && (pol->flags & MPOL_F_MOF)) | |
1630 | ret = true; | |
1631 | mpol_cond_put(pol); | |
1632 | ||
1633 | return ret; | |
1634 | } | |
1635 | ||
1636 | pol = vma->vm_policy; | |
1637 | if (!pol) | |
1638 | pol = get_task_policy(current); | |
1639 | ||
1640 | return pol->flags & MPOL_F_MOF; | |
1641 | } | |
1642 | ||
1643 | static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone) | |
1644 | { | |
1645 | enum zone_type dynamic_policy_zone = policy_zone; | |
1646 | ||
1647 | BUG_ON(dynamic_policy_zone == ZONE_MOVABLE); | |
1648 | ||
1649 | /* | |
1650 | * if policy->v.nodes has movable memory only, | |
1651 | * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only. | |
1652 | * | |
1653 | * policy->v.nodes is intersect with node_states[N_MEMORY]. | |
1654 | * so if the following test faile, it implies | |
1655 | * policy->v.nodes has movable memory only. | |
1656 | */ | |
1657 | if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY])) | |
1658 | dynamic_policy_zone = ZONE_MOVABLE; | |
1659 | ||
1660 | return zone >= dynamic_policy_zone; | |
1661 | } | |
1662 | ||
1663 | /* | |
1664 | * Return a nodemask representing a mempolicy for filtering nodes for | |
1665 | * page allocation | |
1666 | */ | |
1667 | static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy) | |
1668 | { | |
1669 | /* Lower zones don't get a nodemask applied for MPOL_BIND */ | |
1670 | if (unlikely(policy->mode == MPOL_BIND) && | |
1671 | apply_policy_zone(policy, gfp_zone(gfp)) && | |
1672 | cpuset_nodemask_valid_mems_allowed(&policy->v.nodes)) | |
1673 | return &policy->v.nodes; | |
1674 | ||
1675 | return NULL; | |
1676 | } | |
1677 | ||
1678 | /* Return a zonelist indicated by gfp for node representing a mempolicy */ | |
1679 | static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy, | |
1680 | int nd) | |
1681 | { | |
1682 | switch (policy->mode) { | |
1683 | case MPOL_PREFERRED: | |
1684 | if (!(policy->flags & MPOL_F_LOCAL)) | |
1685 | nd = policy->v.preferred_node; | |
1686 | break; | |
1687 | case MPOL_BIND: | |
1688 | /* | |
1689 | * Normally, MPOL_BIND allocations are node-local within the | |
1690 | * allowed nodemask. However, if __GFP_THISNODE is set and the | |
1691 | * current node isn't part of the mask, we use the zonelist for | |
1692 | * the first node in the mask instead. | |
1693 | */ | |
1694 | if (unlikely(gfp & __GFP_THISNODE) && | |
1695 | unlikely(!node_isset(nd, policy->v.nodes))) | |
1696 | nd = first_node(policy->v.nodes); | |
1697 | break; | |
1698 | default: | |
1699 | BUG(); | |
1700 | } | |
1701 | return node_zonelist(nd, gfp); | |
1702 | } | |
1703 | ||
1704 | /* Do dynamic interleaving for a process */ | |
1705 | static unsigned interleave_nodes(struct mempolicy *policy) | |
1706 | { | |
1707 | unsigned nid, next; | |
1708 | struct task_struct *me = current; | |
1709 | ||
1710 | nid = me->il_next; | |
1711 | next = next_node_in(nid, policy->v.nodes); | |
1712 | if (next < MAX_NUMNODES) | |
1713 | me->il_next = next; | |
1714 | return nid; | |
1715 | } | |
1716 | ||
1717 | /* | |
1718 | * Depending on the memory policy provide a node from which to allocate the | |
1719 | * next slab entry. | |
1720 | */ | |
1721 | unsigned int mempolicy_slab_node(void) | |
1722 | { | |
1723 | struct mempolicy *policy; | |
1724 | int node = numa_mem_id(); | |
1725 | ||
1726 | if (in_interrupt()) | |
1727 | return node; | |
1728 | ||
1729 | policy = current->mempolicy; | |
1730 | if (!policy || policy->flags & MPOL_F_LOCAL) | |
1731 | return node; | |
1732 | ||
1733 | switch (policy->mode) { | |
1734 | case MPOL_PREFERRED: | |
1735 | /* | |
1736 | * handled MPOL_F_LOCAL above | |
1737 | */ | |
1738 | return policy->v.preferred_node; | |
1739 | ||
1740 | case MPOL_INTERLEAVE: | |
1741 | return interleave_nodes(policy); | |
1742 | ||
1743 | case MPOL_BIND: { | |
1744 | struct zoneref *z; | |
1745 | ||
1746 | /* | |
1747 | * Follow bind policy behavior and start allocation at the | |
1748 | * first node. | |
1749 | */ | |
1750 | struct zonelist *zonelist; | |
1751 | enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL); | |
1752 | zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK]; | |
1753 | z = first_zones_zonelist(zonelist, highest_zoneidx, | |
1754 | &policy->v.nodes); | |
1755 | return z->zone ? z->zone->node : node; | |
1756 | } | |
1757 | ||
1758 | default: | |
1759 | BUG(); | |
1760 | } | |
1761 | } | |
1762 | ||
1763 | /* | |
1764 | * Do static interleaving for a VMA with known offset @n. Returns the n'th | |
1765 | * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the | |
1766 | * number of present nodes. | |
1767 | */ | |
1768 | static unsigned offset_il_node(struct mempolicy *pol, | |
1769 | struct vm_area_struct *vma, unsigned long n) | |
1770 | { | |
1771 | unsigned nnodes = nodes_weight(pol->v.nodes); | |
1772 | unsigned target; | |
1773 | int i; | |
1774 | int nid; | |
1775 | ||
1776 | if (!nnodes) | |
1777 | return numa_node_id(); | |
1778 | target = (unsigned int)n % nnodes; | |
1779 | nid = first_node(pol->v.nodes); | |
1780 | for (i = 0; i < target; i++) | |
1781 | nid = next_node(nid, pol->v.nodes); | |
1782 | return nid; | |
1783 | } | |
1784 | ||
1785 | /* Determine a node number for interleave */ | |
1786 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1787 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1788 | { | |
1789 | if (vma) { | |
1790 | unsigned long off; | |
1791 | ||
1792 | /* | |
1793 | * for small pages, there is no difference between | |
1794 | * shift and PAGE_SHIFT, so the bit-shift is safe. | |
1795 | * for huge pages, since vm_pgoff is in units of small | |
1796 | * pages, we need to shift off the always 0 bits to get | |
1797 | * a useful offset. | |
1798 | */ | |
1799 | BUG_ON(shift < PAGE_SHIFT); | |
1800 | off = vma->vm_pgoff >> (shift - PAGE_SHIFT); | |
1801 | off += (addr - vma->vm_start) >> shift; | |
1802 | return offset_il_node(pol, vma, off); | |
1803 | } else | |
1804 | return interleave_nodes(pol); | |
1805 | } | |
1806 | ||
1807 | #ifdef CONFIG_HUGETLBFS | |
1808 | /* | |
1809 | * huge_zonelist(@vma, @addr, @gfp_flags, @mpol) | |
1810 | * @vma: virtual memory area whose policy is sought | |
1811 | * @addr: address in @vma for shared policy lookup and interleave policy | |
1812 | * @gfp_flags: for requested zone | |
1813 | * @mpol: pointer to mempolicy pointer for reference counted mempolicy | |
1814 | * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask | |
1815 | * | |
1816 | * Returns a zonelist suitable for a huge page allocation and a pointer | |
1817 | * to the struct mempolicy for conditional unref after allocation. | |
1818 | * If the effective policy is 'BIND, returns a pointer to the mempolicy's | |
1819 | * @nodemask for filtering the zonelist. | |
1820 | * | |
1821 | * Must be protected by read_mems_allowed_begin() | |
1822 | */ | |
1823 | struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr, | |
1824 | gfp_t gfp_flags, struct mempolicy **mpol, | |
1825 | nodemask_t **nodemask) | |
1826 | { | |
1827 | struct zonelist *zl; | |
1828 | ||
1829 | *mpol = get_vma_policy(vma, addr); | |
1830 | *nodemask = NULL; /* assume !MPOL_BIND */ | |
1831 | ||
1832 | if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) { | |
1833 | zl = node_zonelist(interleave_nid(*mpol, vma, addr, | |
1834 | huge_page_shift(hstate_vma(vma))), gfp_flags); | |
1835 | } else { | |
1836 | zl = policy_zonelist(gfp_flags, *mpol, numa_node_id()); | |
1837 | if ((*mpol)->mode == MPOL_BIND) | |
1838 | *nodemask = &(*mpol)->v.nodes; | |
1839 | } | |
1840 | return zl; | |
1841 | } | |
1842 | ||
1843 | /* | |
1844 | * init_nodemask_of_mempolicy | |
1845 | * | |
1846 | * If the current task's mempolicy is "default" [NULL], return 'false' | |
1847 | * to indicate default policy. Otherwise, extract the policy nodemask | |
1848 | * for 'bind' or 'interleave' policy into the argument nodemask, or | |
1849 | * initialize the argument nodemask to contain the single node for | |
1850 | * 'preferred' or 'local' policy and return 'true' to indicate presence | |
1851 | * of non-default mempolicy. | |
1852 | * | |
1853 | * We don't bother with reference counting the mempolicy [mpol_get/put] | |
1854 | * because the current task is examining it's own mempolicy and a task's | |
1855 | * mempolicy is only ever changed by the task itself. | |
1856 | * | |
1857 | * N.B., it is the caller's responsibility to free a returned nodemask. | |
1858 | */ | |
1859 | bool init_nodemask_of_mempolicy(nodemask_t *mask) | |
1860 | { | |
1861 | struct mempolicy *mempolicy; | |
1862 | int nid; | |
1863 | ||
1864 | if (!(mask && current->mempolicy)) | |
1865 | return false; | |
1866 | ||
1867 | task_lock(current); | |
1868 | mempolicy = current->mempolicy; | |
1869 | switch (mempolicy->mode) { | |
1870 | case MPOL_PREFERRED: | |
1871 | if (mempolicy->flags & MPOL_F_LOCAL) | |
1872 | nid = numa_node_id(); | |
1873 | else | |
1874 | nid = mempolicy->v.preferred_node; | |
1875 | init_nodemask_of_node(mask, nid); | |
1876 | break; | |
1877 | ||
1878 | case MPOL_BIND: | |
1879 | /* Fall through */ | |
1880 | case MPOL_INTERLEAVE: | |
1881 | *mask = mempolicy->v.nodes; | |
1882 | break; | |
1883 | ||
1884 | default: | |
1885 | BUG(); | |
1886 | } | |
1887 | task_unlock(current); | |
1888 | ||
1889 | return true; | |
1890 | } | |
1891 | #endif | |
1892 | ||
1893 | /* | |
1894 | * mempolicy_nodemask_intersects | |
1895 | * | |
1896 | * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default | |
1897 | * policy. Otherwise, check for intersection between mask and the policy | |
1898 | * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local' | |
1899 | * policy, always return true since it may allocate elsewhere on fallback. | |
1900 | * | |
1901 | * Takes task_lock(tsk) to prevent freeing of its mempolicy. | |
1902 | */ | |
1903 | bool mempolicy_nodemask_intersects(struct task_struct *tsk, | |
1904 | const nodemask_t *mask) | |
1905 | { | |
1906 | struct mempolicy *mempolicy; | |
1907 | bool ret = true; | |
1908 | ||
1909 | if (!mask) | |
1910 | return ret; | |
1911 | task_lock(tsk); | |
1912 | mempolicy = tsk->mempolicy; | |
1913 | if (!mempolicy) | |
1914 | goto out; | |
1915 | ||
1916 | switch (mempolicy->mode) { | |
1917 | case MPOL_PREFERRED: | |
1918 | /* | |
1919 | * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to | |
1920 | * allocate from, they may fallback to other nodes when oom. | |
1921 | * Thus, it's possible for tsk to have allocated memory from | |
1922 | * nodes in mask. | |
1923 | */ | |
1924 | break; | |
1925 | case MPOL_BIND: | |
1926 | case MPOL_INTERLEAVE: | |
1927 | ret = nodes_intersects(mempolicy->v.nodes, *mask); | |
1928 | break; | |
1929 | default: | |
1930 | BUG(); | |
1931 | } | |
1932 | out: | |
1933 | task_unlock(tsk); | |
1934 | return ret; | |
1935 | } | |
1936 | ||
1937 | /* Allocate a page in interleaved policy. | |
1938 | Own path because it needs to do special accounting. */ | |
1939 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, | |
1940 | unsigned nid) | |
1941 | { | |
1942 | struct zonelist *zl; | |
1943 | struct page *page; | |
1944 | ||
1945 | zl = node_zonelist(nid, gfp); | |
1946 | page = __alloc_pages(gfp, order, zl); | |
1947 | if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0])) | |
1948 | inc_zone_page_state(page, NUMA_INTERLEAVE_HIT); | |
1949 | return page; | |
1950 | } | |
1951 | ||
1952 | /** | |
1953 | * alloc_pages_vma - Allocate a page for a VMA. | |
1954 | * | |
1955 | * @gfp: | |
1956 | * %GFP_USER user allocation. | |
1957 | * %GFP_KERNEL kernel allocations, | |
1958 | * %GFP_HIGHMEM highmem/user allocations, | |
1959 | * %GFP_FS allocation should not call back into a file system. | |
1960 | * %GFP_ATOMIC don't sleep. | |
1961 | * | |
1962 | * @order:Order of the GFP allocation. | |
1963 | * @vma: Pointer to VMA or NULL if not available. | |
1964 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
1965 | * @node: Which node to prefer for allocation (modulo policy). | |
1966 | * @hugepage: for hugepages try only the preferred node if possible | |
1967 | * | |
1968 | * This function allocates a page from the kernel page pool and applies | |
1969 | * a NUMA policy associated with the VMA or the current process. | |
1970 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1971 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
1972 | * all allocations for pages that will be mapped into user space. Returns | |
1973 | * NULL when no page can be allocated. | |
1974 | */ | |
1975 | struct page * | |
1976 | alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma, | |
1977 | unsigned long addr, int node, bool hugepage) | |
1978 | { | |
1979 | struct mempolicy *pol; | |
1980 | struct page *page; | |
1981 | unsigned int cpuset_mems_cookie; | |
1982 | struct zonelist *zl; | |
1983 | nodemask_t *nmask; | |
1984 | ||
1985 | retry_cpuset: | |
1986 | pol = get_vma_policy(vma, addr); | |
1987 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1988 | ||
1989 | if (pol->mode == MPOL_INTERLEAVE) { | |
1990 | unsigned nid; | |
1991 | ||
1992 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order); | |
1993 | mpol_cond_put(pol); | |
1994 | page = alloc_page_interleave(gfp, order, nid); | |
1995 | goto out; | |
1996 | } | |
1997 | ||
1998 | if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) { | |
1999 | int hpage_node = node; | |
2000 | ||
2001 | /* | |
2002 | * For hugepage allocation and non-interleave policy which | |
2003 | * allows the current node (or other explicitly preferred | |
2004 | * node) we only try to allocate from the current/preferred | |
2005 | * node and don't fall back to other nodes, as the cost of | |
2006 | * remote accesses would likely offset THP benefits. | |
2007 | * | |
2008 | * If the policy is interleave, or does not allow the current | |
2009 | * node in its nodemask, we allocate the standard way. | |
2010 | */ | |
2011 | if (pol->mode == MPOL_PREFERRED && | |
2012 | !(pol->flags & MPOL_F_LOCAL)) | |
2013 | hpage_node = pol->v.preferred_node; | |
2014 | ||
2015 | nmask = policy_nodemask(gfp, pol); | |
2016 | if (!nmask || node_isset(hpage_node, *nmask)) { | |
2017 | mpol_cond_put(pol); | |
2018 | page = __alloc_pages_node(hpage_node, | |
2019 | gfp | __GFP_THISNODE, order); | |
2020 | goto out; | |
2021 | } | |
2022 | } | |
2023 | ||
2024 | nmask = policy_nodemask(gfp, pol); | |
2025 | zl = policy_zonelist(gfp, pol, node); | |
2026 | mpol_cond_put(pol); | |
2027 | page = __alloc_pages_nodemask(gfp, order, zl, nmask); | |
2028 | out: | |
2029 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) | |
2030 | goto retry_cpuset; | |
2031 | return page; | |
2032 | } | |
2033 | ||
2034 | /** | |
2035 | * alloc_pages_current - Allocate pages. | |
2036 | * | |
2037 | * @gfp: | |
2038 | * %GFP_USER user allocation, | |
2039 | * %GFP_KERNEL kernel allocation, | |
2040 | * %GFP_HIGHMEM highmem allocation, | |
2041 | * %GFP_FS don't call back into a file system. | |
2042 | * %GFP_ATOMIC don't sleep. | |
2043 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
2044 | * | |
2045 | * Allocate a page from the kernel page pool. When not in | |
2046 | * interrupt context and apply the current process NUMA policy. | |
2047 | * Returns NULL when no page can be allocated. | |
2048 | * | |
2049 | * Don't call cpuset_update_task_memory_state() unless | |
2050 | * 1) it's ok to take cpuset_sem (can WAIT), and | |
2051 | * 2) allocating for current task (not interrupt). | |
2052 | */ | |
2053 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) | |
2054 | { | |
2055 | struct mempolicy *pol = &default_policy; | |
2056 | struct page *page; | |
2057 | unsigned int cpuset_mems_cookie; | |
2058 | ||
2059 | if (!in_interrupt() && !(gfp & __GFP_THISNODE)) | |
2060 | pol = get_task_policy(current); | |
2061 | ||
2062 | retry_cpuset: | |
2063 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
2064 | ||
2065 | /* | |
2066 | * No reference counting needed for current->mempolicy | |
2067 | * nor system default_policy | |
2068 | */ | |
2069 | if (pol->mode == MPOL_INTERLEAVE) | |
2070 | page = alloc_page_interleave(gfp, order, interleave_nodes(pol)); | |
2071 | else | |
2072 | page = __alloc_pages_nodemask(gfp, order, | |
2073 | policy_zonelist(gfp, pol, numa_node_id()), | |
2074 | policy_nodemask(gfp, pol)); | |
2075 | ||
2076 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) | |
2077 | goto retry_cpuset; | |
2078 | ||
2079 | return page; | |
2080 | } | |
2081 | EXPORT_SYMBOL(alloc_pages_current); | |
2082 | ||
2083 | int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst) | |
2084 | { | |
2085 | struct mempolicy *pol = mpol_dup(vma_policy(src)); | |
2086 | ||
2087 | if (IS_ERR(pol)) | |
2088 | return PTR_ERR(pol); | |
2089 | dst->vm_policy = pol; | |
2090 | return 0; | |
2091 | } | |
2092 | ||
2093 | /* | |
2094 | * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it | |
2095 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() | |
2096 | * with the mems_allowed returned by cpuset_mems_allowed(). This | |
2097 | * keeps mempolicies cpuset relative after its cpuset moves. See | |
2098 | * further kernel/cpuset.c update_nodemask(). | |
2099 | * | |
2100 | * current's mempolicy may be rebinded by the other task(the task that changes | |
2101 | * cpuset's mems), so we needn't do rebind work for current task. | |
2102 | */ | |
2103 | ||
2104 | /* Slow path of a mempolicy duplicate */ | |
2105 | struct mempolicy *__mpol_dup(struct mempolicy *old) | |
2106 | { | |
2107 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
2108 | ||
2109 | if (!new) | |
2110 | return ERR_PTR(-ENOMEM); | |
2111 | ||
2112 | /* task's mempolicy is protected by alloc_lock */ | |
2113 | if (old == current->mempolicy) { | |
2114 | task_lock(current); | |
2115 | *new = *old; | |
2116 | task_unlock(current); | |
2117 | } else | |
2118 | *new = *old; | |
2119 | ||
2120 | if (current_cpuset_is_being_rebound()) { | |
2121 | nodemask_t mems = cpuset_mems_allowed(current); | |
2122 | if (new->flags & MPOL_F_REBINDING) | |
2123 | mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2); | |
2124 | else | |
2125 | mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE); | |
2126 | } | |
2127 | atomic_set(&new->refcnt, 1); | |
2128 | return new; | |
2129 | } | |
2130 | ||
2131 | /* Slow path of a mempolicy comparison */ | |
2132 | bool __mpol_equal(struct mempolicy *a, struct mempolicy *b) | |
2133 | { | |
2134 | if (!a || !b) | |
2135 | return false; | |
2136 | if (a->mode != b->mode) | |
2137 | return false; | |
2138 | if (a->flags != b->flags) | |
2139 | return false; | |
2140 | if (mpol_store_user_nodemask(a)) | |
2141 | if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask)) | |
2142 | return false; | |
2143 | ||
2144 | switch (a->mode) { | |
2145 | case MPOL_BIND: | |
2146 | /* Fall through */ | |
2147 | case MPOL_INTERLEAVE: | |
2148 | return !!nodes_equal(a->v.nodes, b->v.nodes); | |
2149 | case MPOL_PREFERRED: | |
2150 | return a->v.preferred_node == b->v.preferred_node; | |
2151 | default: | |
2152 | BUG(); | |
2153 | return false; | |
2154 | } | |
2155 | } | |
2156 | ||
2157 | /* | |
2158 | * Shared memory backing store policy support. | |
2159 | * | |
2160 | * Remember policies even when nobody has shared memory mapped. | |
2161 | * The policies are kept in Red-Black tree linked from the inode. | |
2162 | * They are protected by the sp->lock rwlock, which should be held | |
2163 | * for any accesses to the tree. | |
2164 | */ | |
2165 | ||
2166 | /* | |
2167 | * lookup first element intersecting start-end. Caller holds sp->lock for | |
2168 | * reading or for writing | |
2169 | */ | |
2170 | static struct sp_node * | |
2171 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
2172 | { | |
2173 | struct rb_node *n = sp->root.rb_node; | |
2174 | ||
2175 | while (n) { | |
2176 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
2177 | ||
2178 | if (start >= p->end) | |
2179 | n = n->rb_right; | |
2180 | else if (end <= p->start) | |
2181 | n = n->rb_left; | |
2182 | else | |
2183 | break; | |
2184 | } | |
2185 | if (!n) | |
2186 | return NULL; | |
2187 | for (;;) { | |
2188 | struct sp_node *w = NULL; | |
2189 | struct rb_node *prev = rb_prev(n); | |
2190 | if (!prev) | |
2191 | break; | |
2192 | w = rb_entry(prev, struct sp_node, nd); | |
2193 | if (w->end <= start) | |
2194 | break; | |
2195 | n = prev; | |
2196 | } | |
2197 | return rb_entry(n, struct sp_node, nd); | |
2198 | } | |
2199 | ||
2200 | /* | |
2201 | * Insert a new shared policy into the list. Caller holds sp->lock for | |
2202 | * writing. | |
2203 | */ | |
2204 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) | |
2205 | { | |
2206 | struct rb_node **p = &sp->root.rb_node; | |
2207 | struct rb_node *parent = NULL; | |
2208 | struct sp_node *nd; | |
2209 | ||
2210 | while (*p) { | |
2211 | parent = *p; | |
2212 | nd = rb_entry(parent, struct sp_node, nd); | |
2213 | if (new->start < nd->start) | |
2214 | p = &(*p)->rb_left; | |
2215 | else if (new->end > nd->end) | |
2216 | p = &(*p)->rb_right; | |
2217 | else | |
2218 | BUG(); | |
2219 | } | |
2220 | rb_link_node(&new->nd, parent, p); | |
2221 | rb_insert_color(&new->nd, &sp->root); | |
2222 | pr_debug("inserting %lx-%lx: %d\n", new->start, new->end, | |
2223 | new->policy ? new->policy->mode : 0); | |
2224 | } | |
2225 | ||
2226 | /* Find shared policy intersecting idx */ | |
2227 | struct mempolicy * | |
2228 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
2229 | { | |
2230 | struct mempolicy *pol = NULL; | |
2231 | struct sp_node *sn; | |
2232 | ||
2233 | if (!sp->root.rb_node) | |
2234 | return NULL; | |
2235 | read_lock(&sp->lock); | |
2236 | sn = sp_lookup(sp, idx, idx+1); | |
2237 | if (sn) { | |
2238 | mpol_get(sn->policy); | |
2239 | pol = sn->policy; | |
2240 | } | |
2241 | read_unlock(&sp->lock); | |
2242 | return pol; | |
2243 | } | |
2244 | ||
2245 | static void sp_free(struct sp_node *n) | |
2246 | { | |
2247 | mpol_put(n->policy); | |
2248 | kmem_cache_free(sn_cache, n); | |
2249 | } | |
2250 | ||
2251 | /** | |
2252 | * mpol_misplaced - check whether current page node is valid in policy | |
2253 | * | |
2254 | * @page: page to be checked | |
2255 | * @vma: vm area where page mapped | |
2256 | * @addr: virtual address where page mapped | |
2257 | * | |
2258 | * Lookup current policy node id for vma,addr and "compare to" page's | |
2259 | * node id. | |
2260 | * | |
2261 | * Returns: | |
2262 | * -1 - not misplaced, page is in the right node | |
2263 | * node - node id where the page should be | |
2264 | * | |
2265 | * Policy determination "mimics" alloc_page_vma(). | |
2266 | * Called from fault path where we know the vma and faulting address. | |
2267 | */ | |
2268 | int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr) | |
2269 | { | |
2270 | struct mempolicy *pol; | |
2271 | struct zoneref *z; | |
2272 | int curnid = page_to_nid(page); | |
2273 | unsigned long pgoff; | |
2274 | int thiscpu = raw_smp_processor_id(); | |
2275 | int thisnid = cpu_to_node(thiscpu); | |
2276 | int polnid = -1; | |
2277 | int ret = -1; | |
2278 | ||
2279 | BUG_ON(!vma); | |
2280 | ||
2281 | pol = get_vma_policy(vma, addr); | |
2282 | if (!(pol->flags & MPOL_F_MOF)) | |
2283 | goto out; | |
2284 | ||
2285 | switch (pol->mode) { | |
2286 | case MPOL_INTERLEAVE: | |
2287 | BUG_ON(addr >= vma->vm_end); | |
2288 | BUG_ON(addr < vma->vm_start); | |
2289 | ||
2290 | pgoff = vma->vm_pgoff; | |
2291 | pgoff += (addr - vma->vm_start) >> PAGE_SHIFT; | |
2292 | polnid = offset_il_node(pol, vma, pgoff); | |
2293 | break; | |
2294 | ||
2295 | case MPOL_PREFERRED: | |
2296 | if (pol->flags & MPOL_F_LOCAL) | |
2297 | polnid = numa_node_id(); | |
2298 | else | |
2299 | polnid = pol->v.preferred_node; | |
2300 | break; | |
2301 | ||
2302 | case MPOL_BIND: | |
2303 | ||
2304 | /* | |
2305 | * allows binding to multiple nodes. | |
2306 | * use current page if in policy nodemask, | |
2307 | * else select nearest allowed node, if any. | |
2308 | * If no allowed nodes, use current [!misplaced]. | |
2309 | */ | |
2310 | if (node_isset(curnid, pol->v.nodes)) | |
2311 | goto out; | |
2312 | z = first_zones_zonelist( | |
2313 | node_zonelist(numa_node_id(), GFP_HIGHUSER), | |
2314 | gfp_zone(GFP_HIGHUSER), | |
2315 | &pol->v.nodes); | |
2316 | polnid = z->zone->node; | |
2317 | break; | |
2318 | ||
2319 | default: | |
2320 | BUG(); | |
2321 | } | |
2322 | ||
2323 | /* Migrate the page towards the node whose CPU is referencing it */ | |
2324 | if (pol->flags & MPOL_F_MORON) { | |
2325 | polnid = thisnid; | |
2326 | ||
2327 | if (!should_numa_migrate_memory(current, page, curnid, thiscpu)) | |
2328 | goto out; | |
2329 | } | |
2330 | ||
2331 | if (curnid != polnid) | |
2332 | ret = polnid; | |
2333 | out: | |
2334 | mpol_cond_put(pol); | |
2335 | ||
2336 | return ret; | |
2337 | } | |
2338 | ||
2339 | /* | |
2340 | * Drop the (possibly final) reference to task->mempolicy. It needs to be | |
2341 | * dropped after task->mempolicy is set to NULL so that any allocation done as | |
2342 | * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed | |
2343 | * policy. | |
2344 | */ | |
2345 | void mpol_put_task_policy(struct task_struct *task) | |
2346 | { | |
2347 | struct mempolicy *pol; | |
2348 | ||
2349 | task_lock(task); | |
2350 | pol = task->mempolicy; | |
2351 | task->mempolicy = NULL; | |
2352 | task_unlock(task); | |
2353 | mpol_put(pol); | |
2354 | } | |
2355 | ||
2356 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) | |
2357 | { | |
2358 | pr_debug("deleting %lx-l%lx\n", n->start, n->end); | |
2359 | rb_erase(&n->nd, &sp->root); | |
2360 | sp_free(n); | |
2361 | } | |
2362 | ||
2363 | static void sp_node_init(struct sp_node *node, unsigned long start, | |
2364 | unsigned long end, struct mempolicy *pol) | |
2365 | { | |
2366 | node->start = start; | |
2367 | node->end = end; | |
2368 | node->policy = pol; | |
2369 | } | |
2370 | ||
2371 | static struct sp_node *sp_alloc(unsigned long start, unsigned long end, | |
2372 | struct mempolicy *pol) | |
2373 | { | |
2374 | struct sp_node *n; | |
2375 | struct mempolicy *newpol; | |
2376 | ||
2377 | n = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
2378 | if (!n) | |
2379 | return NULL; | |
2380 | ||
2381 | newpol = mpol_dup(pol); | |
2382 | if (IS_ERR(newpol)) { | |
2383 | kmem_cache_free(sn_cache, n); | |
2384 | return NULL; | |
2385 | } | |
2386 | newpol->flags |= MPOL_F_SHARED; | |
2387 | sp_node_init(n, start, end, newpol); | |
2388 | ||
2389 | return n; | |
2390 | } | |
2391 | ||
2392 | /* Replace a policy range. */ | |
2393 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
2394 | unsigned long end, struct sp_node *new) | |
2395 | { | |
2396 | struct sp_node *n; | |
2397 | struct sp_node *n_new = NULL; | |
2398 | struct mempolicy *mpol_new = NULL; | |
2399 | int ret = 0; | |
2400 | ||
2401 | restart: | |
2402 | write_lock(&sp->lock); | |
2403 | n = sp_lookup(sp, start, end); | |
2404 | /* Take care of old policies in the same range. */ | |
2405 | while (n && n->start < end) { | |
2406 | struct rb_node *next = rb_next(&n->nd); | |
2407 | if (n->start >= start) { | |
2408 | if (n->end <= end) | |
2409 | sp_delete(sp, n); | |
2410 | else | |
2411 | n->start = end; | |
2412 | } else { | |
2413 | /* Old policy spanning whole new range. */ | |
2414 | if (n->end > end) { | |
2415 | if (!n_new) | |
2416 | goto alloc_new; | |
2417 | ||
2418 | *mpol_new = *n->policy; | |
2419 | atomic_set(&mpol_new->refcnt, 1); | |
2420 | sp_node_init(n_new, end, n->end, mpol_new); | |
2421 | n->end = start; | |
2422 | sp_insert(sp, n_new); | |
2423 | n_new = NULL; | |
2424 | mpol_new = NULL; | |
2425 | break; | |
2426 | } else | |
2427 | n->end = start; | |
2428 | } | |
2429 | if (!next) | |
2430 | break; | |
2431 | n = rb_entry(next, struct sp_node, nd); | |
2432 | } | |
2433 | if (new) | |
2434 | sp_insert(sp, new); | |
2435 | write_unlock(&sp->lock); | |
2436 | ret = 0; | |
2437 | ||
2438 | err_out: | |
2439 | if (mpol_new) | |
2440 | mpol_put(mpol_new); | |
2441 | if (n_new) | |
2442 | kmem_cache_free(sn_cache, n_new); | |
2443 | ||
2444 | return ret; | |
2445 | ||
2446 | alloc_new: | |
2447 | write_unlock(&sp->lock); | |
2448 | ret = -ENOMEM; | |
2449 | n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
2450 | if (!n_new) | |
2451 | goto err_out; | |
2452 | mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
2453 | if (!mpol_new) | |
2454 | goto err_out; | |
2455 | goto restart; | |
2456 | } | |
2457 | ||
2458 | /** | |
2459 | * mpol_shared_policy_init - initialize shared policy for inode | |
2460 | * @sp: pointer to inode shared policy | |
2461 | * @mpol: struct mempolicy to install | |
2462 | * | |
2463 | * Install non-NULL @mpol in inode's shared policy rb-tree. | |
2464 | * On entry, the current task has a reference on a non-NULL @mpol. | |
2465 | * This must be released on exit. | |
2466 | * This is called at get_inode() calls and we can use GFP_KERNEL. | |
2467 | */ | |
2468 | void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol) | |
2469 | { | |
2470 | int ret; | |
2471 | ||
2472 | sp->root = RB_ROOT; /* empty tree == default mempolicy */ | |
2473 | rwlock_init(&sp->lock); | |
2474 | ||
2475 | if (mpol) { | |
2476 | struct vm_area_struct pvma; | |
2477 | struct mempolicy *new; | |
2478 | NODEMASK_SCRATCH(scratch); | |
2479 | ||
2480 | if (!scratch) | |
2481 | goto put_mpol; | |
2482 | /* contextualize the tmpfs mount point mempolicy */ | |
2483 | new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask); | |
2484 | if (IS_ERR(new)) | |
2485 | goto free_scratch; /* no valid nodemask intersection */ | |
2486 | ||
2487 | task_lock(current); | |
2488 | ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch); | |
2489 | task_unlock(current); | |
2490 | if (ret) | |
2491 | goto put_new; | |
2492 | ||
2493 | /* Create pseudo-vma that contains just the policy */ | |
2494 | memset(&pvma, 0, sizeof(struct vm_area_struct)); | |
2495 | pvma.vm_end = TASK_SIZE; /* policy covers entire file */ | |
2496 | mpol_set_shared_policy(sp, &pvma, new); /* adds ref */ | |
2497 | ||
2498 | put_new: | |
2499 | mpol_put(new); /* drop initial ref */ | |
2500 | free_scratch: | |
2501 | NODEMASK_SCRATCH_FREE(scratch); | |
2502 | put_mpol: | |
2503 | mpol_put(mpol); /* drop our incoming ref on sb mpol */ | |
2504 | } | |
2505 | } | |
2506 | ||
2507 | int mpol_set_shared_policy(struct shared_policy *info, | |
2508 | struct vm_area_struct *vma, struct mempolicy *npol) | |
2509 | { | |
2510 | int err; | |
2511 | struct sp_node *new = NULL; | |
2512 | unsigned long sz = vma_pages(vma); | |
2513 | ||
2514 | pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n", | |
2515 | vma->vm_pgoff, | |
2516 | sz, npol ? npol->mode : -1, | |
2517 | npol ? npol->flags : -1, | |
2518 | npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE); | |
2519 | ||
2520 | if (npol) { | |
2521 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
2522 | if (!new) | |
2523 | return -ENOMEM; | |
2524 | } | |
2525 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
2526 | if (err && new) | |
2527 | sp_free(new); | |
2528 | return err; | |
2529 | } | |
2530 | ||
2531 | /* Free a backing policy store on inode delete. */ | |
2532 | void mpol_free_shared_policy(struct shared_policy *p) | |
2533 | { | |
2534 | struct sp_node *n; | |
2535 | struct rb_node *next; | |
2536 | ||
2537 | if (!p->root.rb_node) | |
2538 | return; | |
2539 | write_lock(&p->lock); | |
2540 | next = rb_first(&p->root); | |
2541 | while (next) { | |
2542 | n = rb_entry(next, struct sp_node, nd); | |
2543 | next = rb_next(&n->nd); | |
2544 | sp_delete(p, n); | |
2545 | } | |
2546 | write_unlock(&p->lock); | |
2547 | } | |
2548 | ||
2549 | #ifdef CONFIG_NUMA_BALANCING | |
2550 | static int __initdata numabalancing_override; | |
2551 | ||
2552 | static void __init check_numabalancing_enable(void) | |
2553 | { | |
2554 | bool numabalancing_default = false; | |
2555 | ||
2556 | if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED)) | |
2557 | numabalancing_default = true; | |
2558 | ||
2559 | /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */ | |
2560 | if (numabalancing_override) | |
2561 | set_numabalancing_state(numabalancing_override == 1); | |
2562 | ||
2563 | if (num_online_nodes() > 1 && !numabalancing_override) { | |
2564 | pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n", | |
2565 | numabalancing_default ? "Enabling" : "Disabling"); | |
2566 | set_numabalancing_state(numabalancing_default); | |
2567 | } | |
2568 | } | |
2569 | ||
2570 | static int __init setup_numabalancing(char *str) | |
2571 | { | |
2572 | int ret = 0; | |
2573 | if (!str) | |
2574 | goto out; | |
2575 | ||
2576 | if (!strcmp(str, "enable")) { | |
2577 | numabalancing_override = 1; | |
2578 | ret = 1; | |
2579 | } else if (!strcmp(str, "disable")) { | |
2580 | numabalancing_override = -1; | |
2581 | ret = 1; | |
2582 | } | |
2583 | out: | |
2584 | if (!ret) | |
2585 | pr_warn("Unable to parse numa_balancing=\n"); | |
2586 | ||
2587 | return ret; | |
2588 | } | |
2589 | __setup("numa_balancing=", setup_numabalancing); | |
2590 | #else | |
2591 | static inline void __init check_numabalancing_enable(void) | |
2592 | { | |
2593 | } | |
2594 | #endif /* CONFIG_NUMA_BALANCING */ | |
2595 | ||
2596 | /* assumes fs == KERNEL_DS */ | |
2597 | void __init numa_policy_init(void) | |
2598 | { | |
2599 | nodemask_t interleave_nodes; | |
2600 | unsigned long largest = 0; | |
2601 | int nid, prefer = 0; | |
2602 | ||
2603 | policy_cache = kmem_cache_create("numa_policy", | |
2604 | sizeof(struct mempolicy), | |
2605 | 0, SLAB_PANIC, NULL); | |
2606 | ||
2607 | sn_cache = kmem_cache_create("shared_policy_node", | |
2608 | sizeof(struct sp_node), | |
2609 | 0, SLAB_PANIC, NULL); | |
2610 | ||
2611 | for_each_node(nid) { | |
2612 | preferred_node_policy[nid] = (struct mempolicy) { | |
2613 | .refcnt = ATOMIC_INIT(1), | |
2614 | .mode = MPOL_PREFERRED, | |
2615 | .flags = MPOL_F_MOF | MPOL_F_MORON, | |
2616 | .v = { .preferred_node = nid, }, | |
2617 | }; | |
2618 | } | |
2619 | ||
2620 | /* | |
2621 | * Set interleaving policy for system init. Interleaving is only | |
2622 | * enabled across suitably sized nodes (default is >= 16MB), or | |
2623 | * fall back to the largest node if they're all smaller. | |
2624 | */ | |
2625 | nodes_clear(interleave_nodes); | |
2626 | for_each_node_state(nid, N_MEMORY) { | |
2627 | unsigned long total_pages = node_present_pages(nid); | |
2628 | ||
2629 | /* Preserve the largest node */ | |
2630 | if (largest < total_pages) { | |
2631 | largest = total_pages; | |
2632 | prefer = nid; | |
2633 | } | |
2634 | ||
2635 | /* Interleave this node? */ | |
2636 | if ((total_pages << PAGE_SHIFT) >= (16 << 20)) | |
2637 | node_set(nid, interleave_nodes); | |
2638 | } | |
2639 | ||
2640 | /* All too small, use the largest */ | |
2641 | if (unlikely(nodes_empty(interleave_nodes))) | |
2642 | node_set(prefer, interleave_nodes); | |
2643 | ||
2644 | if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes)) | |
2645 | pr_err("%s: interleaving failed\n", __func__); | |
2646 | ||
2647 | check_numabalancing_enable(); | |
2648 | } | |
2649 | ||
2650 | /* Reset policy of current process to default */ | |
2651 | void numa_default_policy(void) | |
2652 | { | |
2653 | do_set_mempolicy(MPOL_DEFAULT, 0, NULL); | |
2654 | } | |
2655 | ||
2656 | /* | |
2657 | * Parse and format mempolicy from/to strings | |
2658 | */ | |
2659 | ||
2660 | /* | |
2661 | * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag. | |
2662 | */ | |
2663 | static const char * const policy_modes[] = | |
2664 | { | |
2665 | [MPOL_DEFAULT] = "default", | |
2666 | [MPOL_PREFERRED] = "prefer", | |
2667 | [MPOL_BIND] = "bind", | |
2668 | [MPOL_INTERLEAVE] = "interleave", | |
2669 | [MPOL_LOCAL] = "local", | |
2670 | }; | |
2671 | ||
2672 | ||
2673 | #ifdef CONFIG_TMPFS | |
2674 | /** | |
2675 | * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option. | |
2676 | * @str: string containing mempolicy to parse | |
2677 | * @mpol: pointer to struct mempolicy pointer, returned on success. | |
2678 | * | |
2679 | * Format of input: | |
2680 | * <mode>[=<flags>][:<nodelist>] | |
2681 | * | |
2682 | * On success, returns 0, else 1 | |
2683 | */ | |
2684 | int mpol_parse_str(char *str, struct mempolicy **mpol) | |
2685 | { | |
2686 | struct mempolicy *new = NULL; | |
2687 | unsigned short mode; | |
2688 | unsigned short mode_flags; | |
2689 | nodemask_t nodes; | |
2690 | char *nodelist = strchr(str, ':'); | |
2691 | char *flags = strchr(str, '='); | |
2692 | int err = 1; | |
2693 | ||
2694 | if (nodelist) { | |
2695 | /* NUL-terminate mode or flags string */ | |
2696 | *nodelist++ = '\0'; | |
2697 | if (nodelist_parse(nodelist, nodes)) | |
2698 | goto out; | |
2699 | if (!nodes_subset(nodes, node_states[N_MEMORY])) | |
2700 | goto out; | |
2701 | } else | |
2702 | nodes_clear(nodes); | |
2703 | ||
2704 | if (flags) | |
2705 | *flags++ = '\0'; /* terminate mode string */ | |
2706 | ||
2707 | for (mode = 0; mode < MPOL_MAX; mode++) { | |
2708 | if (!strcmp(str, policy_modes[mode])) { | |
2709 | break; | |
2710 | } | |
2711 | } | |
2712 | if (mode >= MPOL_MAX) | |
2713 | goto out; | |
2714 | ||
2715 | switch (mode) { | |
2716 | case MPOL_PREFERRED: | |
2717 | /* | |
2718 | * Insist on a nodelist of one node only | |
2719 | */ | |
2720 | if (nodelist) { | |
2721 | char *rest = nodelist; | |
2722 | while (isdigit(*rest)) | |
2723 | rest++; | |
2724 | if (*rest) | |
2725 | goto out; | |
2726 | } | |
2727 | break; | |
2728 | case MPOL_INTERLEAVE: | |
2729 | /* | |
2730 | * Default to online nodes with memory if no nodelist | |
2731 | */ | |
2732 | if (!nodelist) | |
2733 | nodes = node_states[N_MEMORY]; | |
2734 | break; | |
2735 | case MPOL_LOCAL: | |
2736 | /* | |
2737 | * Don't allow a nodelist; mpol_new() checks flags | |
2738 | */ | |
2739 | if (nodelist) | |
2740 | goto out; | |
2741 | mode = MPOL_PREFERRED; | |
2742 | break; | |
2743 | case MPOL_DEFAULT: | |
2744 | /* | |
2745 | * Insist on a empty nodelist | |
2746 | */ | |
2747 | if (!nodelist) | |
2748 | err = 0; | |
2749 | goto out; | |
2750 | case MPOL_BIND: | |
2751 | /* | |
2752 | * Insist on a nodelist | |
2753 | */ | |
2754 | if (!nodelist) | |
2755 | goto out; | |
2756 | } | |
2757 | ||
2758 | mode_flags = 0; | |
2759 | if (flags) { | |
2760 | /* | |
2761 | * Currently, we only support two mutually exclusive | |
2762 | * mode flags. | |
2763 | */ | |
2764 | if (!strcmp(flags, "static")) | |
2765 | mode_flags |= MPOL_F_STATIC_NODES; | |
2766 | else if (!strcmp(flags, "relative")) | |
2767 | mode_flags |= MPOL_F_RELATIVE_NODES; | |
2768 | else | |
2769 | goto out; | |
2770 | } | |
2771 | ||
2772 | new = mpol_new(mode, mode_flags, &nodes); | |
2773 | if (IS_ERR(new)) | |
2774 | goto out; | |
2775 | ||
2776 | /* | |
2777 | * Save nodes for mpol_to_str() to show the tmpfs mount options | |
2778 | * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo. | |
2779 | */ | |
2780 | if (mode != MPOL_PREFERRED) | |
2781 | new->v.nodes = nodes; | |
2782 | else if (nodelist) | |
2783 | new->v.preferred_node = first_node(nodes); | |
2784 | else | |
2785 | new->flags |= MPOL_F_LOCAL; | |
2786 | ||
2787 | /* | |
2788 | * Save nodes for contextualization: this will be used to "clone" | |
2789 | * the mempolicy in a specific context [cpuset] at a later time. | |
2790 | */ | |
2791 | new->w.user_nodemask = nodes; | |
2792 | ||
2793 | err = 0; | |
2794 | ||
2795 | out: | |
2796 | /* Restore string for error message */ | |
2797 | if (nodelist) | |
2798 | *--nodelist = ':'; | |
2799 | if (flags) | |
2800 | *--flags = '='; | |
2801 | if (!err) | |
2802 | *mpol = new; | |
2803 | return err; | |
2804 | } | |
2805 | #endif /* CONFIG_TMPFS */ | |
2806 | ||
2807 | /** | |
2808 | * mpol_to_str - format a mempolicy structure for printing | |
2809 | * @buffer: to contain formatted mempolicy string | |
2810 | * @maxlen: length of @buffer | |
2811 | * @pol: pointer to mempolicy to be formatted | |
2812 | * | |
2813 | * Convert @pol into a string. If @buffer is too short, truncate the string. | |
2814 | * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the | |
2815 | * longest flag, "relative", and to display at least a few node ids. | |
2816 | */ | |
2817 | void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) | |
2818 | { | |
2819 | char *p = buffer; | |
2820 | nodemask_t nodes = NODE_MASK_NONE; | |
2821 | unsigned short mode = MPOL_DEFAULT; | |
2822 | unsigned short flags = 0; | |
2823 | ||
2824 | if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) { | |
2825 | mode = pol->mode; | |
2826 | flags = pol->flags; | |
2827 | } | |
2828 | ||
2829 | switch (mode) { | |
2830 | case MPOL_DEFAULT: | |
2831 | break; | |
2832 | case MPOL_PREFERRED: | |
2833 | if (flags & MPOL_F_LOCAL) | |
2834 | mode = MPOL_LOCAL; | |
2835 | else | |
2836 | node_set(pol->v.preferred_node, nodes); | |
2837 | break; | |
2838 | case MPOL_BIND: | |
2839 | case MPOL_INTERLEAVE: | |
2840 | nodes = pol->v.nodes; | |
2841 | break; | |
2842 | default: | |
2843 | WARN_ON_ONCE(1); | |
2844 | snprintf(p, maxlen, "unknown"); | |
2845 | return; | |
2846 | } | |
2847 | ||
2848 | p += snprintf(p, maxlen, "%s", policy_modes[mode]); | |
2849 | ||
2850 | if (flags & MPOL_MODE_FLAGS) { | |
2851 | p += snprintf(p, buffer + maxlen - p, "="); | |
2852 | ||
2853 | /* | |
2854 | * Currently, the only defined flags are mutually exclusive | |
2855 | */ | |
2856 | if (flags & MPOL_F_STATIC_NODES) | |
2857 | p += snprintf(p, buffer + maxlen - p, "static"); | |
2858 | else if (flags & MPOL_F_RELATIVE_NODES) | |
2859 | p += snprintf(p, buffer + maxlen - p, "relative"); | |
2860 | } | |
2861 | ||
2862 | if (!nodes_empty(nodes)) | |
2863 | p += scnprintf(p, buffer + maxlen - p, ":%*pbl", | |
2864 | nodemask_pr_args(&nodes)); | |
2865 | } |