Linux 3.10.23
[linux/fpc-iii.git] / mm / mempolicy.c
blob4baf12e534d19031d28ddc5eba5335f72de54099
1 /*
2 * Simple NUMA memory policy for the Linux kernel.
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.
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
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.
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
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.
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.
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.
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.
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.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
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.
68 #include <linux/mempolicy.h>
69 #include <linux/mm.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
93 #include <linux/mmu_notifier.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
97 #include <linux/random.h>
99 #include "internal.h"
101 /* Internal flags */
102 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
105 static struct kmem_cache *policy_cache;
106 static struct kmem_cache *sn_cache;
108 /* Highest zone. An specific allocation for a zone below that is not
109 policied. */
110 enum zone_type policy_zone = 0;
113 * run-time system-wide default policy => local allocation
115 static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
121 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
123 static struct mempolicy *get_task_policy(struct task_struct *p)
125 struct mempolicy *pol = p->mempolicy;
126 int node;
128 if (!pol) {
129 node = numa_node_id();
130 if (node != NUMA_NO_NODE)
131 pol = &preferred_node_policy[node];
133 /* preferred_node_policy is not initialised early in boot */
134 if (!pol->mode)
135 pol = NULL;
138 return pol;
141 static const struct mempolicy_operations {
142 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
144 * If read-side task has no lock to protect task->mempolicy, write-side
145 * task will rebind the task->mempolicy by two step. The first step is
146 * setting all the newly nodes, and the second step is cleaning all the
147 * disallowed nodes. In this way, we can avoid finding no node to alloc
148 * page.
149 * If we have a lock to protect task->mempolicy in read-side, we do
150 * rebind directly.
152 * step:
153 * MPOL_REBIND_ONCE - do rebind work at once
154 * MPOL_REBIND_STEP1 - set all the newly nodes
155 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
157 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
158 enum mpol_rebind_step step);
159 } mpol_ops[MPOL_MAX];
161 /* Check that the nodemask contains at least one populated zone */
162 static int is_valid_nodemask(const nodemask_t *nodemask)
164 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
167 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
169 return pol->flags & MPOL_MODE_FLAGS;
172 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
173 const nodemask_t *rel)
175 nodemask_t tmp;
176 nodes_fold(tmp, *orig, nodes_weight(*rel));
177 nodes_onto(*ret, tmp, *rel);
180 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
182 if (nodes_empty(*nodes))
183 return -EINVAL;
184 pol->v.nodes = *nodes;
185 return 0;
188 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
190 if (!nodes)
191 pol->flags |= MPOL_F_LOCAL; /* local allocation */
192 else if (nodes_empty(*nodes))
193 return -EINVAL; /* no allowed nodes */
194 else
195 pol->v.preferred_node = first_node(*nodes);
196 return 0;
199 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
201 if (!is_valid_nodemask(nodes))
202 return -EINVAL;
203 pol->v.nodes = *nodes;
204 return 0;
208 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
209 * any, for the new policy. mpol_new() has already validated the nodes
210 * parameter with respect to the policy mode and flags. But, we need to
211 * handle an empty nodemask with MPOL_PREFERRED here.
213 * Must be called holding task's alloc_lock to protect task's mems_allowed
214 * and mempolicy. May also be called holding the mmap_semaphore for write.
216 static int mpol_set_nodemask(struct mempolicy *pol,
217 const nodemask_t *nodes, struct nodemask_scratch *nsc)
219 int ret;
221 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
222 if (pol == NULL)
223 return 0;
224 /* Check N_MEMORY */
225 nodes_and(nsc->mask1,
226 cpuset_current_mems_allowed, node_states[N_MEMORY]);
228 VM_BUG_ON(!nodes);
229 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
230 nodes = NULL; /* explicit local allocation */
231 else {
232 if (pol->flags & MPOL_F_RELATIVE_NODES)
233 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
234 else
235 nodes_and(nsc->mask2, *nodes, nsc->mask1);
237 if (mpol_store_user_nodemask(pol))
238 pol->w.user_nodemask = *nodes;
239 else
240 pol->w.cpuset_mems_allowed =
241 cpuset_current_mems_allowed;
244 if (nodes)
245 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
246 else
247 ret = mpol_ops[pol->mode].create(pol, NULL);
248 return ret;
252 * This function just creates a new policy, does some check and simple
253 * initialization. You must invoke mpol_set_nodemask() to set nodes.
255 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
256 nodemask_t *nodes)
258 struct mempolicy *policy;
260 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
261 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
263 if (mode == MPOL_DEFAULT) {
264 if (nodes && !nodes_empty(*nodes))
265 return ERR_PTR(-EINVAL);
266 return NULL;
268 VM_BUG_ON(!nodes);
271 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
272 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
273 * All other modes require a valid pointer to a non-empty nodemask.
275 if (mode == MPOL_PREFERRED) {
276 if (nodes_empty(*nodes)) {
277 if (((flags & MPOL_F_STATIC_NODES) ||
278 (flags & MPOL_F_RELATIVE_NODES)))
279 return ERR_PTR(-EINVAL);
281 } else if (mode == MPOL_LOCAL) {
282 if (!nodes_empty(*nodes))
283 return ERR_PTR(-EINVAL);
284 mode = MPOL_PREFERRED;
285 } else if (nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
288 if (!policy)
289 return ERR_PTR(-ENOMEM);
290 atomic_set(&policy->refcnt, 1);
291 policy->mode = mode;
292 policy->flags = flags;
294 return policy;
297 /* Slow path of a mpol destructor. */
298 void __mpol_put(struct mempolicy *p)
300 if (!atomic_dec_and_test(&p->refcnt))
301 return;
302 kmem_cache_free(policy_cache, p);
305 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
306 enum mpol_rebind_step step)
311 * step:
312 * MPOL_REBIND_ONCE - do rebind work at once
313 * MPOL_REBIND_STEP1 - set all the newly nodes
314 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
316 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
317 enum mpol_rebind_step step)
319 nodemask_t tmp;
321 if (pol->flags & MPOL_F_STATIC_NODES)
322 nodes_and(tmp, pol->w.user_nodemask, *nodes);
323 else if (pol->flags & MPOL_F_RELATIVE_NODES)
324 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
325 else {
327 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
328 * result
330 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
331 nodes_remap(tmp, pol->v.nodes,
332 pol->w.cpuset_mems_allowed, *nodes);
333 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
334 } else if (step == MPOL_REBIND_STEP2) {
335 tmp = pol->w.cpuset_mems_allowed;
336 pol->w.cpuset_mems_allowed = *nodes;
337 } else
338 BUG();
341 if (nodes_empty(tmp))
342 tmp = *nodes;
344 if (step == MPOL_REBIND_STEP1)
345 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
346 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
347 pol->v.nodes = tmp;
348 else
349 BUG();
351 if (!node_isset(current->il_next, tmp)) {
352 current->il_next = next_node(current->il_next, tmp);
353 if (current->il_next >= MAX_NUMNODES)
354 current->il_next = first_node(tmp);
355 if (current->il_next >= MAX_NUMNODES)
356 current->il_next = numa_node_id();
360 static void mpol_rebind_preferred(struct mempolicy *pol,
361 const nodemask_t *nodes,
362 enum mpol_rebind_step step)
364 nodemask_t tmp;
366 if (pol->flags & MPOL_F_STATIC_NODES) {
367 int node = first_node(pol->w.user_nodemask);
369 if (node_isset(node, *nodes)) {
370 pol->v.preferred_node = node;
371 pol->flags &= ~MPOL_F_LOCAL;
372 } else
373 pol->flags |= MPOL_F_LOCAL;
374 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
375 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
376 pol->v.preferred_node = first_node(tmp);
377 } else if (!(pol->flags & MPOL_F_LOCAL)) {
378 pol->v.preferred_node = node_remap(pol->v.preferred_node,
379 pol->w.cpuset_mems_allowed,
380 *nodes);
381 pol->w.cpuset_mems_allowed = *nodes;
386 * mpol_rebind_policy - Migrate a policy to a different set of nodes
388 * If read-side task has no lock to protect task->mempolicy, write-side
389 * task will rebind the task->mempolicy by two step. The first step is
390 * setting all the newly nodes, and the second step is cleaning all the
391 * disallowed nodes. In this way, we can avoid finding no node to alloc
392 * page.
393 * If we have a lock to protect task->mempolicy in read-side, we do
394 * rebind directly.
396 * step:
397 * MPOL_REBIND_ONCE - do rebind work at once
398 * MPOL_REBIND_STEP1 - set all the newly nodes
399 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
401 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
402 enum mpol_rebind_step step)
404 if (!pol)
405 return;
406 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
407 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
408 return;
410 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
411 return;
413 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
414 BUG();
416 if (step == MPOL_REBIND_STEP1)
417 pol->flags |= MPOL_F_REBINDING;
418 else if (step == MPOL_REBIND_STEP2)
419 pol->flags &= ~MPOL_F_REBINDING;
420 else if (step >= MPOL_REBIND_NSTEP)
421 BUG();
423 mpol_ops[pol->mode].rebind(pol, newmask, step);
427 * Wrapper for mpol_rebind_policy() that just requires task
428 * pointer, and updates task mempolicy.
430 * Called with task's alloc_lock held.
433 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
434 enum mpol_rebind_step step)
436 mpol_rebind_policy(tsk->mempolicy, new, step);
440 * Rebind each vma in mm to new nodemask.
442 * Call holding a reference to mm. Takes mm->mmap_sem during call.
445 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
447 struct vm_area_struct *vma;
449 down_write(&mm->mmap_sem);
450 for (vma = mm->mmap; vma; vma = vma->vm_next)
451 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
452 up_write(&mm->mmap_sem);
455 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
456 [MPOL_DEFAULT] = {
457 .rebind = mpol_rebind_default,
459 [MPOL_INTERLEAVE] = {
460 .create = mpol_new_interleave,
461 .rebind = mpol_rebind_nodemask,
463 [MPOL_PREFERRED] = {
464 .create = mpol_new_preferred,
465 .rebind = mpol_rebind_preferred,
467 [MPOL_BIND] = {
468 .create = mpol_new_bind,
469 .rebind = mpol_rebind_nodemask,
473 static void migrate_page_add(struct page *page, struct list_head *pagelist,
474 unsigned long flags);
476 /* Scan through pages checking if pages follow certain conditions. */
477 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
478 unsigned long addr, unsigned long end,
479 const nodemask_t *nodes, unsigned long flags,
480 void *private)
482 pte_t *orig_pte;
483 pte_t *pte;
484 spinlock_t *ptl;
486 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
487 do {
488 struct page *page;
489 int nid;
491 if (!pte_present(*pte))
492 continue;
493 page = vm_normal_page(vma, addr, *pte);
494 if (!page)
495 continue;
497 * vm_normal_page() filters out zero pages, but there might
498 * still be PageReserved pages to skip, perhaps in a VDSO.
500 if (PageReserved(page))
501 continue;
502 nid = page_to_nid(page);
503 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
504 continue;
506 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
507 migrate_page_add(page, private, flags);
508 else
509 break;
510 } while (pte++, addr += PAGE_SIZE, addr != end);
511 pte_unmap_unlock(orig_pte, ptl);
512 return addr != end;
515 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
516 unsigned long addr, unsigned long end,
517 const nodemask_t *nodes, unsigned long flags,
518 void *private)
520 pmd_t *pmd;
521 unsigned long next;
523 pmd = pmd_offset(pud, addr);
524 do {
525 next = pmd_addr_end(addr, end);
526 split_huge_page_pmd(vma, addr, pmd);
527 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
528 continue;
529 if (check_pte_range(vma, pmd, addr, next, nodes,
530 flags, private))
531 return -EIO;
532 } while (pmd++, addr = next, addr != end);
533 return 0;
536 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
537 unsigned long addr, unsigned long end,
538 const nodemask_t *nodes, unsigned long flags,
539 void *private)
541 pud_t *pud;
542 unsigned long next;
544 pud = pud_offset(pgd, addr);
545 do {
546 next = pud_addr_end(addr, end);
547 if (pud_none_or_clear_bad(pud))
548 continue;
549 if (check_pmd_range(vma, pud, addr, next, nodes,
550 flags, private))
551 return -EIO;
552 } while (pud++, addr = next, addr != end);
553 return 0;
556 static inline int check_pgd_range(struct vm_area_struct *vma,
557 unsigned long addr, unsigned long end,
558 const nodemask_t *nodes, unsigned long flags,
559 void *private)
561 pgd_t *pgd;
562 unsigned long next;
564 pgd = pgd_offset(vma->vm_mm, addr);
565 do {
566 next = pgd_addr_end(addr, end);
567 if (pgd_none_or_clear_bad(pgd))
568 continue;
569 if (check_pud_range(vma, pgd, addr, next, nodes,
570 flags, private))
571 return -EIO;
572 } while (pgd++, addr = next, addr != end);
573 return 0;
576 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
578 * This is used to mark a range of virtual addresses to be inaccessible.
579 * These are later cleared by a NUMA hinting fault. Depending on these
580 * faults, pages may be migrated for better NUMA placement.
582 * This is assuming that NUMA faults are handled using PROT_NONE. If
583 * an architecture makes a different choice, it will need further
584 * changes to the core.
586 unsigned long change_prot_numa(struct vm_area_struct *vma,
587 unsigned long addr, unsigned long end)
589 int nr_updated;
590 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
592 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
593 if (nr_updated)
594 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
596 return nr_updated;
598 #else
599 static unsigned long change_prot_numa(struct vm_area_struct *vma,
600 unsigned long addr, unsigned long end)
602 return 0;
604 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
607 * Check if all pages in a range are on a set of nodes.
608 * If pagelist != NULL then isolate pages from the LRU and
609 * put them on the pagelist.
611 static struct vm_area_struct *
612 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
613 const nodemask_t *nodes, unsigned long flags, void *private)
615 int err;
616 struct vm_area_struct *first, *vma, *prev;
619 first = find_vma(mm, start);
620 if (!first)
621 return ERR_PTR(-EFAULT);
622 prev = NULL;
623 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
624 unsigned long endvma = vma->vm_end;
626 if (endvma > end)
627 endvma = end;
628 if (vma->vm_start > start)
629 start = vma->vm_start;
631 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
632 if (!vma->vm_next && vma->vm_end < end)
633 return ERR_PTR(-EFAULT);
634 if (prev && prev->vm_end < vma->vm_start)
635 return ERR_PTR(-EFAULT);
638 if (is_vm_hugetlb_page(vma))
639 goto next;
641 if (flags & MPOL_MF_LAZY) {
642 change_prot_numa(vma, start, endvma);
643 goto next;
646 if ((flags & MPOL_MF_STRICT) ||
647 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
648 vma_migratable(vma))) {
650 err = check_pgd_range(vma, start, endvma, nodes,
651 flags, private);
652 if (err) {
653 first = ERR_PTR(err);
654 break;
657 next:
658 prev = vma;
660 return first;
664 * Apply policy to a single VMA
665 * This must be called with the mmap_sem held for writing.
667 static int vma_replace_policy(struct vm_area_struct *vma,
668 struct mempolicy *pol)
670 int err;
671 struct mempolicy *old;
672 struct mempolicy *new;
674 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
675 vma->vm_start, vma->vm_end, vma->vm_pgoff,
676 vma->vm_ops, vma->vm_file,
677 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
679 new = mpol_dup(pol);
680 if (IS_ERR(new))
681 return PTR_ERR(new);
683 if (vma->vm_ops && vma->vm_ops->set_policy) {
684 err = vma->vm_ops->set_policy(vma, new);
685 if (err)
686 goto err_out;
689 old = vma->vm_policy;
690 vma->vm_policy = new; /* protected by mmap_sem */
691 mpol_put(old);
693 return 0;
694 err_out:
695 mpol_put(new);
696 return err;
699 /* Step 2: apply policy to a range and do splits. */
700 static int mbind_range(struct mm_struct *mm, unsigned long start,
701 unsigned long end, struct mempolicy *new_pol)
703 struct vm_area_struct *next;
704 struct vm_area_struct *prev;
705 struct vm_area_struct *vma;
706 int err = 0;
707 pgoff_t pgoff;
708 unsigned long vmstart;
709 unsigned long vmend;
711 vma = find_vma(mm, start);
712 if (!vma || vma->vm_start > start)
713 return -EFAULT;
715 prev = vma->vm_prev;
716 if (start > vma->vm_start)
717 prev = vma;
719 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
720 next = vma->vm_next;
721 vmstart = max(start, vma->vm_start);
722 vmend = min(end, vma->vm_end);
724 if (mpol_equal(vma_policy(vma), new_pol))
725 continue;
727 pgoff = vma->vm_pgoff +
728 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
729 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
730 vma->anon_vma, vma->vm_file, pgoff,
731 new_pol);
732 if (prev) {
733 vma = prev;
734 next = vma->vm_next;
735 if (mpol_equal(vma_policy(vma), new_pol))
736 continue;
737 /* vma_merge() joined vma && vma->next, case 8 */
738 goto replace;
740 if (vma->vm_start != vmstart) {
741 err = split_vma(vma->vm_mm, vma, vmstart, 1);
742 if (err)
743 goto out;
745 if (vma->vm_end != vmend) {
746 err = split_vma(vma->vm_mm, vma, vmend, 0);
747 if (err)
748 goto out;
750 replace:
751 err = vma_replace_policy(vma, new_pol);
752 if (err)
753 goto out;
756 out:
757 return err;
761 * Update task->flags PF_MEMPOLICY bit: set iff non-default
762 * mempolicy. Allows more rapid checking of this (combined perhaps
763 * with other PF_* flag bits) on memory allocation hot code paths.
765 * If called from outside this file, the task 'p' should -only- be
766 * a newly forked child not yet visible on the task list, because
767 * manipulating the task flags of a visible task is not safe.
769 * The above limitation is why this routine has the funny name
770 * mpol_fix_fork_child_flag().
772 * It is also safe to call this with a task pointer of current,
773 * which the static wrapper mpol_set_task_struct_flag() does,
774 * for use within this file.
777 void mpol_fix_fork_child_flag(struct task_struct *p)
779 if (p->mempolicy)
780 p->flags |= PF_MEMPOLICY;
781 else
782 p->flags &= ~PF_MEMPOLICY;
785 static void mpol_set_task_struct_flag(void)
787 mpol_fix_fork_child_flag(current);
790 /* Set the process memory policy */
791 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
792 nodemask_t *nodes)
794 struct mempolicy *new, *old;
795 struct mm_struct *mm = current->mm;
796 NODEMASK_SCRATCH(scratch);
797 int ret;
799 if (!scratch)
800 return -ENOMEM;
802 new = mpol_new(mode, flags, nodes);
803 if (IS_ERR(new)) {
804 ret = PTR_ERR(new);
805 goto out;
808 * prevent changing our mempolicy while show_numa_maps()
809 * is using it.
810 * Note: do_set_mempolicy() can be called at init time
811 * with no 'mm'.
813 if (mm)
814 down_write(&mm->mmap_sem);
815 task_lock(current);
816 ret = mpol_set_nodemask(new, nodes, scratch);
817 if (ret) {
818 task_unlock(current);
819 if (mm)
820 up_write(&mm->mmap_sem);
821 mpol_put(new);
822 goto out;
824 old = current->mempolicy;
825 current->mempolicy = new;
826 mpol_set_task_struct_flag();
827 if (new && new->mode == MPOL_INTERLEAVE &&
828 nodes_weight(new->v.nodes))
829 current->il_next = first_node(new->v.nodes);
830 task_unlock(current);
831 if (mm)
832 up_write(&mm->mmap_sem);
834 mpol_put(old);
835 ret = 0;
836 out:
837 NODEMASK_SCRATCH_FREE(scratch);
838 return ret;
842 * Return nodemask for policy for get_mempolicy() query
844 * Called with task's alloc_lock held
846 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
848 nodes_clear(*nodes);
849 if (p == &default_policy)
850 return;
852 switch (p->mode) {
853 case MPOL_BIND:
854 /* Fall through */
855 case MPOL_INTERLEAVE:
856 *nodes = p->v.nodes;
857 break;
858 case MPOL_PREFERRED:
859 if (!(p->flags & MPOL_F_LOCAL))
860 node_set(p->v.preferred_node, *nodes);
861 /* else return empty node mask for local allocation */
862 break;
863 default:
864 BUG();
868 static int lookup_node(struct mm_struct *mm, unsigned long addr)
870 struct page *p;
871 int err;
873 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
874 if (err >= 0) {
875 err = page_to_nid(p);
876 put_page(p);
878 return err;
881 /* Retrieve NUMA policy */
882 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
883 unsigned long addr, unsigned long flags)
885 int err;
886 struct mm_struct *mm = current->mm;
887 struct vm_area_struct *vma = NULL;
888 struct mempolicy *pol = current->mempolicy;
890 if (flags &
891 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
892 return -EINVAL;
894 if (flags & MPOL_F_MEMS_ALLOWED) {
895 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
896 return -EINVAL;
897 *policy = 0; /* just so it's initialized */
898 task_lock(current);
899 *nmask = cpuset_current_mems_allowed;
900 task_unlock(current);
901 return 0;
904 if (flags & MPOL_F_ADDR) {
906 * Do NOT fall back to task policy if the
907 * vma/shared policy at addr is NULL. We
908 * want to return MPOL_DEFAULT in this case.
910 down_read(&mm->mmap_sem);
911 vma = find_vma_intersection(mm, addr, addr+1);
912 if (!vma) {
913 up_read(&mm->mmap_sem);
914 return -EFAULT;
916 if (vma->vm_ops && vma->vm_ops->get_policy)
917 pol = vma->vm_ops->get_policy(vma, addr);
918 else
919 pol = vma->vm_policy;
920 } else if (addr)
921 return -EINVAL;
923 if (!pol)
924 pol = &default_policy; /* indicates default behavior */
926 if (flags & MPOL_F_NODE) {
927 if (flags & MPOL_F_ADDR) {
928 err = lookup_node(mm, addr);
929 if (err < 0)
930 goto out;
931 *policy = err;
932 } else if (pol == current->mempolicy &&
933 pol->mode == MPOL_INTERLEAVE) {
934 *policy = current->il_next;
935 } else {
936 err = -EINVAL;
937 goto out;
939 } else {
940 *policy = pol == &default_policy ? MPOL_DEFAULT :
941 pol->mode;
943 * Internal mempolicy flags must be masked off before exposing
944 * the policy to userspace.
946 *policy |= (pol->flags & MPOL_MODE_FLAGS);
949 if (vma) {
950 up_read(&current->mm->mmap_sem);
951 vma = NULL;
954 err = 0;
955 if (nmask) {
956 if (mpol_store_user_nodemask(pol)) {
957 *nmask = pol->w.user_nodemask;
958 } else {
959 task_lock(current);
960 get_policy_nodemask(pol, nmask);
961 task_unlock(current);
965 out:
966 mpol_cond_put(pol);
967 if (vma)
968 up_read(&current->mm->mmap_sem);
969 return err;
972 #ifdef CONFIG_MIGRATION
974 * page migration
976 static void migrate_page_add(struct page *page, struct list_head *pagelist,
977 unsigned long flags)
980 * Avoid migrating a page that is shared with others.
982 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
983 if (!isolate_lru_page(page)) {
984 list_add_tail(&page->lru, pagelist);
985 inc_zone_page_state(page, NR_ISOLATED_ANON +
986 page_is_file_cache(page));
991 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
993 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
997 * Migrate pages from one node to a target node.
998 * Returns error or the number of pages not migrated.
1000 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1001 int flags)
1003 nodemask_t nmask;
1004 LIST_HEAD(pagelist);
1005 int err = 0;
1007 nodes_clear(nmask);
1008 node_set(source, nmask);
1011 * This does not "check" the range but isolates all pages that
1012 * need migration. Between passing in the full user address
1013 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1015 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1016 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1017 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1019 if (!list_empty(&pagelist)) {
1020 err = migrate_pages(&pagelist, new_node_page, dest,
1021 MIGRATE_SYNC, MR_SYSCALL);
1022 if (err)
1023 putback_lru_pages(&pagelist);
1026 return err;
1030 * Move pages between the two nodesets so as to preserve the physical
1031 * layout as much as possible.
1033 * Returns the number of page that could not be moved.
1035 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1036 const nodemask_t *to, int flags)
1038 int busy = 0;
1039 int err;
1040 nodemask_t tmp;
1042 err = migrate_prep();
1043 if (err)
1044 return err;
1046 down_read(&mm->mmap_sem);
1048 err = migrate_vmas(mm, from, to, flags);
1049 if (err)
1050 goto out;
1053 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1054 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1055 * bit in 'tmp', and return that <source, dest> pair for migration.
1056 * The pair of nodemasks 'to' and 'from' define the map.
1058 * If no pair of bits is found that way, fallback to picking some
1059 * pair of 'source' and 'dest' bits that are not the same. If the
1060 * 'source' and 'dest' bits are the same, this represents a node
1061 * that will be migrating to itself, so no pages need move.
1063 * If no bits are left in 'tmp', or if all remaining bits left
1064 * in 'tmp' correspond to the same bit in 'to', return false
1065 * (nothing left to migrate).
1067 * This lets us pick a pair of nodes to migrate between, such that
1068 * if possible the dest node is not already occupied by some other
1069 * source node, minimizing the risk of overloading the memory on a
1070 * node that would happen if we migrated incoming memory to a node
1071 * before migrating outgoing memory source that same node.
1073 * A single scan of tmp is sufficient. As we go, we remember the
1074 * most recent <s, d> pair that moved (s != d). If we find a pair
1075 * that not only moved, but what's better, moved to an empty slot
1076 * (d is not set in tmp), then we break out then, with that pair.
1077 * Otherwise when we finish scanning from_tmp, we at least have the
1078 * most recent <s, d> pair that moved. If we get all the way through
1079 * the scan of tmp without finding any node that moved, much less
1080 * moved to an empty node, then there is nothing left worth migrating.
1083 tmp = *from;
1084 while (!nodes_empty(tmp)) {
1085 int s,d;
1086 int source = -1;
1087 int dest = 0;
1089 for_each_node_mask(s, tmp) {
1092 * do_migrate_pages() tries to maintain the relative
1093 * node relationship of the pages established between
1094 * threads and memory areas.
1096 * However if the number of source nodes is not equal to
1097 * the number of destination nodes we can not preserve
1098 * this node relative relationship. In that case, skip
1099 * copying memory from a node that is in the destination
1100 * mask.
1102 * Example: [2,3,4] -> [3,4,5] moves everything.
1103 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1106 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1107 (node_isset(s, *to)))
1108 continue;
1110 d = node_remap(s, *from, *to);
1111 if (s == d)
1112 continue;
1114 source = s; /* Node moved. Memorize */
1115 dest = d;
1117 /* dest not in remaining from nodes? */
1118 if (!node_isset(dest, tmp))
1119 break;
1121 if (source == -1)
1122 break;
1124 node_clear(source, tmp);
1125 err = migrate_to_node(mm, source, dest, flags);
1126 if (err > 0)
1127 busy += err;
1128 if (err < 0)
1129 break;
1131 out:
1132 up_read(&mm->mmap_sem);
1133 if (err < 0)
1134 return err;
1135 return busy;
1140 * Allocate a new page for page migration based on vma policy.
1141 * Start assuming that page is mapped by vma pointed to by @private.
1142 * Search forward from there, if not. N.B., this assumes that the
1143 * list of pages handed to migrate_pages()--which is how we get here--
1144 * is in virtual address order.
1146 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1148 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1149 unsigned long uninitialized_var(address);
1151 while (vma) {
1152 address = page_address_in_vma(page, vma);
1153 if (address != -EFAULT)
1154 break;
1155 vma = vma->vm_next;
1159 * if !vma, alloc_page_vma() will use task or system default policy
1161 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1163 #else
1165 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1166 unsigned long flags)
1170 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1171 const nodemask_t *to, int flags)
1173 return -ENOSYS;
1176 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1178 return NULL;
1180 #endif
1182 static long do_mbind(unsigned long start, unsigned long len,
1183 unsigned short mode, unsigned short mode_flags,
1184 nodemask_t *nmask, unsigned long flags)
1186 struct vm_area_struct *vma;
1187 struct mm_struct *mm = current->mm;
1188 struct mempolicy *new;
1189 unsigned long end;
1190 int err;
1191 LIST_HEAD(pagelist);
1193 if (flags & ~(unsigned long)MPOL_MF_VALID)
1194 return -EINVAL;
1195 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1196 return -EPERM;
1198 if (start & ~PAGE_MASK)
1199 return -EINVAL;
1201 if (mode == MPOL_DEFAULT)
1202 flags &= ~MPOL_MF_STRICT;
1204 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1205 end = start + len;
1207 if (end < start)
1208 return -EINVAL;
1209 if (end == start)
1210 return 0;
1212 new = mpol_new(mode, mode_flags, nmask);
1213 if (IS_ERR(new))
1214 return PTR_ERR(new);
1216 if (flags & MPOL_MF_LAZY)
1217 new->flags |= MPOL_F_MOF;
1220 * If we are using the default policy then operation
1221 * on discontinuous address spaces is okay after all
1223 if (!new)
1224 flags |= MPOL_MF_DISCONTIG_OK;
1226 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1227 start, start + len, mode, mode_flags,
1228 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1230 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1232 err = migrate_prep();
1233 if (err)
1234 goto mpol_out;
1237 NODEMASK_SCRATCH(scratch);
1238 if (scratch) {
1239 down_write(&mm->mmap_sem);
1240 task_lock(current);
1241 err = mpol_set_nodemask(new, nmask, scratch);
1242 task_unlock(current);
1243 if (err)
1244 up_write(&mm->mmap_sem);
1245 } else
1246 err = -ENOMEM;
1247 NODEMASK_SCRATCH_FREE(scratch);
1249 if (err)
1250 goto mpol_out;
1252 vma = check_range(mm, start, end, nmask,
1253 flags | MPOL_MF_INVERT, &pagelist);
1255 err = PTR_ERR(vma); /* maybe ... */
1256 if (!IS_ERR(vma))
1257 err = mbind_range(mm, start, end, new);
1259 if (!err) {
1260 int nr_failed = 0;
1262 if (!list_empty(&pagelist)) {
1263 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1264 nr_failed = migrate_pages(&pagelist, new_vma_page,
1265 (unsigned long)vma,
1266 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1267 if (nr_failed)
1268 putback_lru_pages(&pagelist);
1271 if (nr_failed && (flags & MPOL_MF_STRICT))
1272 err = -EIO;
1273 } else
1274 putback_lru_pages(&pagelist);
1276 up_write(&mm->mmap_sem);
1277 mpol_out:
1278 mpol_put(new);
1279 return err;
1283 * User space interface with variable sized bitmaps for nodelists.
1286 /* Copy a node mask from user space. */
1287 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1288 unsigned long maxnode)
1290 unsigned long k;
1291 unsigned long nlongs;
1292 unsigned long endmask;
1294 --maxnode;
1295 nodes_clear(*nodes);
1296 if (maxnode == 0 || !nmask)
1297 return 0;
1298 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1299 return -EINVAL;
1301 nlongs = BITS_TO_LONGS(maxnode);
1302 if ((maxnode % BITS_PER_LONG) == 0)
1303 endmask = ~0UL;
1304 else
1305 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1307 /* When the user specified more nodes than supported just check
1308 if the non supported part is all zero. */
1309 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1310 if (nlongs > PAGE_SIZE/sizeof(long))
1311 return -EINVAL;
1312 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1313 unsigned long t;
1314 if (get_user(t, nmask + k))
1315 return -EFAULT;
1316 if (k == nlongs - 1) {
1317 if (t & endmask)
1318 return -EINVAL;
1319 } else if (t)
1320 return -EINVAL;
1322 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1323 endmask = ~0UL;
1326 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1327 return -EFAULT;
1328 nodes_addr(*nodes)[nlongs-1] &= endmask;
1329 return 0;
1332 /* Copy a kernel node mask to user space */
1333 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1334 nodemask_t *nodes)
1336 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1337 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1339 if (copy > nbytes) {
1340 if (copy > PAGE_SIZE)
1341 return -EINVAL;
1342 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1343 return -EFAULT;
1344 copy = nbytes;
1346 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1349 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1350 unsigned long, mode, unsigned long __user *, nmask,
1351 unsigned long, maxnode, unsigned, flags)
1353 nodemask_t nodes;
1354 int err;
1355 unsigned short mode_flags;
1357 mode_flags = mode & MPOL_MODE_FLAGS;
1358 mode &= ~MPOL_MODE_FLAGS;
1359 if (mode >= MPOL_MAX)
1360 return -EINVAL;
1361 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1362 (mode_flags & MPOL_F_RELATIVE_NODES))
1363 return -EINVAL;
1364 err = get_nodes(&nodes, nmask, maxnode);
1365 if (err)
1366 return err;
1367 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1370 /* Set the process memory policy */
1371 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1372 unsigned long, maxnode)
1374 int err;
1375 nodemask_t nodes;
1376 unsigned short flags;
1378 flags = mode & MPOL_MODE_FLAGS;
1379 mode &= ~MPOL_MODE_FLAGS;
1380 if ((unsigned int)mode >= MPOL_MAX)
1381 return -EINVAL;
1382 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1383 return -EINVAL;
1384 err = get_nodes(&nodes, nmask, maxnode);
1385 if (err)
1386 return err;
1387 return do_set_mempolicy(mode, flags, &nodes);
1390 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1391 const unsigned long __user *, old_nodes,
1392 const unsigned long __user *, new_nodes)
1394 const struct cred *cred = current_cred(), *tcred;
1395 struct mm_struct *mm = NULL;
1396 struct task_struct *task;
1397 nodemask_t task_nodes;
1398 int err;
1399 nodemask_t *old;
1400 nodemask_t *new;
1401 NODEMASK_SCRATCH(scratch);
1403 if (!scratch)
1404 return -ENOMEM;
1406 old = &scratch->mask1;
1407 new = &scratch->mask2;
1409 err = get_nodes(old, old_nodes, maxnode);
1410 if (err)
1411 goto out;
1413 err = get_nodes(new, new_nodes, maxnode);
1414 if (err)
1415 goto out;
1417 /* Find the mm_struct */
1418 rcu_read_lock();
1419 task = pid ? find_task_by_vpid(pid) : current;
1420 if (!task) {
1421 rcu_read_unlock();
1422 err = -ESRCH;
1423 goto out;
1425 get_task_struct(task);
1427 err = -EINVAL;
1430 * Check if this process has the right to modify the specified
1431 * process. The right exists if the process has administrative
1432 * capabilities, superuser privileges or the same
1433 * userid as the target process.
1435 tcred = __task_cred(task);
1436 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1437 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1438 !capable(CAP_SYS_NICE)) {
1439 rcu_read_unlock();
1440 err = -EPERM;
1441 goto out_put;
1443 rcu_read_unlock();
1445 task_nodes = cpuset_mems_allowed(task);
1446 /* Is the user allowed to access the target nodes? */
1447 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1448 err = -EPERM;
1449 goto out_put;
1452 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1453 err = -EINVAL;
1454 goto out_put;
1457 err = security_task_movememory(task);
1458 if (err)
1459 goto out_put;
1461 mm = get_task_mm(task);
1462 put_task_struct(task);
1464 if (!mm) {
1465 err = -EINVAL;
1466 goto out;
1469 err = do_migrate_pages(mm, old, new,
1470 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1472 mmput(mm);
1473 out:
1474 NODEMASK_SCRATCH_FREE(scratch);
1476 return err;
1478 out_put:
1479 put_task_struct(task);
1480 goto out;
1485 /* Retrieve NUMA policy */
1486 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1487 unsigned long __user *, nmask, unsigned long, maxnode,
1488 unsigned long, addr, unsigned long, flags)
1490 int err;
1491 int uninitialized_var(pval);
1492 nodemask_t nodes;
1494 if (nmask != NULL && maxnode < MAX_NUMNODES)
1495 return -EINVAL;
1497 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1499 if (err)
1500 return err;
1502 if (policy && put_user(pval, policy))
1503 return -EFAULT;
1505 if (nmask)
1506 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1508 return err;
1511 #ifdef CONFIG_COMPAT
1513 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1514 compat_ulong_t __user *nmask,
1515 compat_ulong_t maxnode,
1516 compat_ulong_t addr, compat_ulong_t flags)
1518 long err;
1519 unsigned long __user *nm = NULL;
1520 unsigned long nr_bits, alloc_size;
1521 DECLARE_BITMAP(bm, MAX_NUMNODES);
1523 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1524 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1526 if (nmask)
1527 nm = compat_alloc_user_space(alloc_size);
1529 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1531 if (!err && nmask) {
1532 unsigned long copy_size;
1533 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1534 err = copy_from_user(bm, nm, copy_size);
1535 /* ensure entire bitmap is zeroed */
1536 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1537 err |= compat_put_bitmap(nmask, bm, nr_bits);
1540 return err;
1543 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1544 compat_ulong_t maxnode)
1546 long err = 0;
1547 unsigned long __user *nm = NULL;
1548 unsigned long nr_bits, alloc_size;
1549 DECLARE_BITMAP(bm, MAX_NUMNODES);
1551 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1552 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1554 if (nmask) {
1555 err = compat_get_bitmap(bm, nmask, nr_bits);
1556 nm = compat_alloc_user_space(alloc_size);
1557 err |= copy_to_user(nm, bm, alloc_size);
1560 if (err)
1561 return -EFAULT;
1563 return sys_set_mempolicy(mode, nm, nr_bits+1);
1566 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1567 compat_ulong_t mode, compat_ulong_t __user *nmask,
1568 compat_ulong_t maxnode, compat_ulong_t flags)
1570 long err = 0;
1571 unsigned long __user *nm = NULL;
1572 unsigned long nr_bits, alloc_size;
1573 nodemask_t bm;
1575 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1576 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1578 if (nmask) {
1579 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1580 nm = compat_alloc_user_space(alloc_size);
1581 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1584 if (err)
1585 return -EFAULT;
1587 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1590 #endif
1593 * get_vma_policy(@task, @vma, @addr)
1594 * @task - task for fallback if vma policy == default
1595 * @vma - virtual memory area whose policy is sought
1596 * @addr - address in @vma for shared policy lookup
1598 * Returns effective policy for a VMA at specified address.
1599 * Falls back to @task or system default policy, as necessary.
1600 * Current or other task's task mempolicy and non-shared vma policies must be
1601 * protected by task_lock(task) by the caller.
1602 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1603 * count--added by the get_policy() vm_op, as appropriate--to protect against
1604 * freeing by another task. It is the caller's responsibility to free the
1605 * extra reference for shared policies.
1607 struct mempolicy *get_vma_policy(struct task_struct *task,
1608 struct vm_area_struct *vma, unsigned long addr)
1610 struct mempolicy *pol = get_task_policy(task);
1612 if (vma) {
1613 if (vma->vm_ops && vma->vm_ops->get_policy) {
1614 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1615 addr);
1616 if (vpol)
1617 pol = vpol;
1618 } else if (vma->vm_policy) {
1619 pol = vma->vm_policy;
1622 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1623 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1624 * count on these policies which will be dropped by
1625 * mpol_cond_put() later
1627 if (mpol_needs_cond_ref(pol))
1628 mpol_get(pol);
1631 if (!pol)
1632 pol = &default_policy;
1633 return pol;
1636 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1638 enum zone_type dynamic_policy_zone = policy_zone;
1640 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1643 * if policy->v.nodes has movable memory only,
1644 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1646 * policy->v.nodes is intersect with node_states[N_MEMORY].
1647 * so if the following test faile, it implies
1648 * policy->v.nodes has movable memory only.
1650 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1651 dynamic_policy_zone = ZONE_MOVABLE;
1653 return zone >= dynamic_policy_zone;
1657 * Return a nodemask representing a mempolicy for filtering nodes for
1658 * page allocation
1660 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1662 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1663 if (unlikely(policy->mode == MPOL_BIND) &&
1664 apply_policy_zone(policy, gfp_zone(gfp)) &&
1665 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1666 return &policy->v.nodes;
1668 return NULL;
1671 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1672 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1673 int nd)
1675 switch (policy->mode) {
1676 case MPOL_PREFERRED:
1677 if (!(policy->flags & MPOL_F_LOCAL))
1678 nd = policy->v.preferred_node;
1679 break;
1680 case MPOL_BIND:
1682 * Normally, MPOL_BIND allocations are node-local within the
1683 * allowed nodemask. However, if __GFP_THISNODE is set and the
1684 * current node isn't part of the mask, we use the zonelist for
1685 * the first node in the mask instead.
1687 if (unlikely(gfp & __GFP_THISNODE) &&
1688 unlikely(!node_isset(nd, policy->v.nodes)))
1689 nd = first_node(policy->v.nodes);
1690 break;
1691 default:
1692 BUG();
1694 return node_zonelist(nd, gfp);
1697 /* Do dynamic interleaving for a process */
1698 static unsigned interleave_nodes(struct mempolicy *policy)
1700 unsigned nid, next;
1701 struct task_struct *me = current;
1703 nid = me->il_next;
1704 next = next_node(nid, policy->v.nodes);
1705 if (next >= MAX_NUMNODES)
1706 next = first_node(policy->v.nodes);
1707 if (next < MAX_NUMNODES)
1708 me->il_next = next;
1709 return nid;
1713 * Depending on the memory policy provide a node from which to allocate the
1714 * next slab entry.
1715 * @policy must be protected by freeing by the caller. If @policy is
1716 * the current task's mempolicy, this protection is implicit, as only the
1717 * task can change it's policy. The system default policy requires no
1718 * such protection.
1720 unsigned slab_node(void)
1722 struct mempolicy *policy;
1724 if (in_interrupt())
1725 return numa_node_id();
1727 policy = current->mempolicy;
1728 if (!policy || policy->flags & MPOL_F_LOCAL)
1729 return numa_node_id();
1731 switch (policy->mode) {
1732 case MPOL_PREFERRED:
1734 * handled MPOL_F_LOCAL above
1736 return policy->v.preferred_node;
1738 case MPOL_INTERLEAVE:
1739 return interleave_nodes(policy);
1741 case MPOL_BIND: {
1743 * Follow bind policy behavior and start allocation at the
1744 * first node.
1746 struct zonelist *zonelist;
1747 struct zone *zone;
1748 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1749 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1750 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1751 &policy->v.nodes,
1752 &zone);
1753 return zone ? zone->node : numa_node_id();
1756 default:
1757 BUG();
1761 /* Do static interleaving for a VMA with known offset. */
1762 static unsigned offset_il_node(struct mempolicy *pol,
1763 struct vm_area_struct *vma, unsigned long off)
1765 unsigned nnodes = nodes_weight(pol->v.nodes);
1766 unsigned target;
1767 int c;
1768 int nid = -1;
1770 if (!nnodes)
1771 return numa_node_id();
1772 target = (unsigned int)off % nnodes;
1773 c = 0;
1774 do {
1775 nid = next_node(nid, pol->v.nodes);
1776 c++;
1777 } while (c <= target);
1778 return nid;
1781 /* Determine a node number for interleave */
1782 static inline unsigned interleave_nid(struct mempolicy *pol,
1783 struct vm_area_struct *vma, unsigned long addr, int shift)
1785 if (vma) {
1786 unsigned long off;
1789 * for small pages, there is no difference between
1790 * shift and PAGE_SHIFT, so the bit-shift is safe.
1791 * for huge pages, since vm_pgoff is in units of small
1792 * pages, we need to shift off the always 0 bits to get
1793 * a useful offset.
1795 BUG_ON(shift < PAGE_SHIFT);
1796 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1797 off += (addr - vma->vm_start) >> shift;
1798 return offset_il_node(pol, vma, off);
1799 } else
1800 return interleave_nodes(pol);
1804 * Return the bit number of a random bit set in the nodemask.
1805 * (returns -1 if nodemask is empty)
1807 int node_random(const nodemask_t *maskp)
1809 int w, bit = -1;
1811 w = nodes_weight(*maskp);
1812 if (w)
1813 bit = bitmap_ord_to_pos(maskp->bits,
1814 get_random_int() % w, MAX_NUMNODES);
1815 return bit;
1818 #ifdef CONFIG_HUGETLBFS
1820 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1821 * @vma = virtual memory area whose policy is sought
1822 * @addr = address in @vma for shared policy lookup and interleave policy
1823 * @gfp_flags = for requested zone
1824 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1825 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1827 * Returns a zonelist suitable for a huge page allocation and a pointer
1828 * to the struct mempolicy for conditional unref after allocation.
1829 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1830 * @nodemask for filtering the zonelist.
1832 * Must be protected by get_mems_allowed()
1834 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1835 gfp_t gfp_flags, struct mempolicy **mpol,
1836 nodemask_t **nodemask)
1838 struct zonelist *zl;
1840 *mpol = get_vma_policy(current, vma, addr);
1841 *nodemask = NULL; /* assume !MPOL_BIND */
1843 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1844 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1845 huge_page_shift(hstate_vma(vma))), gfp_flags);
1846 } else {
1847 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1848 if ((*mpol)->mode == MPOL_BIND)
1849 *nodemask = &(*mpol)->v.nodes;
1851 return zl;
1855 * init_nodemask_of_mempolicy
1857 * If the current task's mempolicy is "default" [NULL], return 'false'
1858 * to indicate default policy. Otherwise, extract the policy nodemask
1859 * for 'bind' or 'interleave' policy into the argument nodemask, or
1860 * initialize the argument nodemask to contain the single node for
1861 * 'preferred' or 'local' policy and return 'true' to indicate presence
1862 * of non-default mempolicy.
1864 * We don't bother with reference counting the mempolicy [mpol_get/put]
1865 * because the current task is examining it's own mempolicy and a task's
1866 * mempolicy is only ever changed by the task itself.
1868 * N.B., it is the caller's responsibility to free a returned nodemask.
1870 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1872 struct mempolicy *mempolicy;
1873 int nid;
1875 if (!(mask && current->mempolicy))
1876 return false;
1878 task_lock(current);
1879 mempolicy = current->mempolicy;
1880 switch (mempolicy->mode) {
1881 case MPOL_PREFERRED:
1882 if (mempolicy->flags & MPOL_F_LOCAL)
1883 nid = numa_node_id();
1884 else
1885 nid = mempolicy->v.preferred_node;
1886 init_nodemask_of_node(mask, nid);
1887 break;
1889 case MPOL_BIND:
1890 /* Fall through */
1891 case MPOL_INTERLEAVE:
1892 *mask = mempolicy->v.nodes;
1893 break;
1895 default:
1896 BUG();
1898 task_unlock(current);
1900 return true;
1902 #endif
1905 * mempolicy_nodemask_intersects
1907 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1908 * policy. Otherwise, check for intersection between mask and the policy
1909 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1910 * policy, always return true since it may allocate elsewhere on fallback.
1912 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1914 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1915 const nodemask_t *mask)
1917 struct mempolicy *mempolicy;
1918 bool ret = true;
1920 if (!mask)
1921 return ret;
1922 task_lock(tsk);
1923 mempolicy = tsk->mempolicy;
1924 if (!mempolicy)
1925 goto out;
1927 switch (mempolicy->mode) {
1928 case MPOL_PREFERRED:
1930 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1931 * allocate from, they may fallback to other nodes when oom.
1932 * Thus, it's possible for tsk to have allocated memory from
1933 * nodes in mask.
1935 break;
1936 case MPOL_BIND:
1937 case MPOL_INTERLEAVE:
1938 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1939 break;
1940 default:
1941 BUG();
1943 out:
1944 task_unlock(tsk);
1945 return ret;
1948 /* Allocate a page in interleaved policy.
1949 Own path because it needs to do special accounting. */
1950 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1951 unsigned nid)
1953 struct zonelist *zl;
1954 struct page *page;
1956 zl = node_zonelist(nid, gfp);
1957 page = __alloc_pages(gfp, order, zl);
1958 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1959 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1960 return page;
1964 * alloc_pages_vma - Allocate a page for a VMA.
1966 * @gfp:
1967 * %GFP_USER user allocation.
1968 * %GFP_KERNEL kernel allocations,
1969 * %GFP_HIGHMEM highmem/user allocations,
1970 * %GFP_FS allocation should not call back into a file system.
1971 * %GFP_ATOMIC don't sleep.
1973 * @order:Order of the GFP allocation.
1974 * @vma: Pointer to VMA or NULL if not available.
1975 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1977 * This function allocates a page from the kernel page pool and applies
1978 * a NUMA policy associated with the VMA or the current process.
1979 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1980 * mm_struct of the VMA to prevent it from going away. Should be used for
1981 * all allocations for pages that will be mapped into
1982 * user space. Returns NULL when no page can be allocated.
1984 * Should be called with the mm_sem of the vma hold.
1986 struct page *
1987 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1988 unsigned long addr, int node)
1990 struct mempolicy *pol;
1991 struct page *page;
1992 unsigned int cpuset_mems_cookie;
1994 retry_cpuset:
1995 pol = get_vma_policy(current, vma, addr);
1996 cpuset_mems_cookie = get_mems_allowed();
1998 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1999 unsigned nid;
2001 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2002 mpol_cond_put(pol);
2003 page = alloc_page_interleave(gfp, order, nid);
2004 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2005 goto retry_cpuset;
2007 return page;
2009 page = __alloc_pages_nodemask(gfp, order,
2010 policy_zonelist(gfp, pol, node),
2011 policy_nodemask(gfp, pol));
2012 if (unlikely(mpol_needs_cond_ref(pol)))
2013 __mpol_put(pol);
2014 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2015 goto retry_cpuset;
2016 return page;
2020 * alloc_pages_current - Allocate pages.
2022 * @gfp:
2023 * %GFP_USER user allocation,
2024 * %GFP_KERNEL kernel allocation,
2025 * %GFP_HIGHMEM highmem allocation,
2026 * %GFP_FS don't call back into a file system.
2027 * %GFP_ATOMIC don't sleep.
2028 * @order: Power of two of allocation size in pages. 0 is a single page.
2030 * Allocate a page from the kernel page pool. When not in
2031 * interrupt context and apply the current process NUMA policy.
2032 * Returns NULL when no page can be allocated.
2034 * Don't call cpuset_update_task_memory_state() unless
2035 * 1) it's ok to take cpuset_sem (can WAIT), and
2036 * 2) allocating for current task (not interrupt).
2038 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2040 struct mempolicy *pol = get_task_policy(current);
2041 struct page *page;
2042 unsigned int cpuset_mems_cookie;
2044 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2045 pol = &default_policy;
2047 retry_cpuset:
2048 cpuset_mems_cookie = get_mems_allowed();
2051 * No reference counting needed for current->mempolicy
2052 * nor system default_policy
2054 if (pol->mode == MPOL_INTERLEAVE)
2055 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2056 else
2057 page = __alloc_pages_nodemask(gfp, order,
2058 policy_zonelist(gfp, pol, numa_node_id()),
2059 policy_nodemask(gfp, pol));
2061 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2062 goto retry_cpuset;
2064 return page;
2066 EXPORT_SYMBOL(alloc_pages_current);
2069 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2070 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2071 * with the mems_allowed returned by cpuset_mems_allowed(). This
2072 * keeps mempolicies cpuset relative after its cpuset moves. See
2073 * further kernel/cpuset.c update_nodemask().
2075 * current's mempolicy may be rebinded by the other task(the task that changes
2076 * cpuset's mems), so we needn't do rebind work for current task.
2079 /* Slow path of a mempolicy duplicate */
2080 struct mempolicy *__mpol_dup(struct mempolicy *old)
2082 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2084 if (!new)
2085 return ERR_PTR(-ENOMEM);
2087 /* task's mempolicy is protected by alloc_lock */
2088 if (old == current->mempolicy) {
2089 task_lock(current);
2090 *new = *old;
2091 task_unlock(current);
2092 } else
2093 *new = *old;
2095 rcu_read_lock();
2096 if (current_cpuset_is_being_rebound()) {
2097 nodemask_t mems = cpuset_mems_allowed(current);
2098 if (new->flags & MPOL_F_REBINDING)
2099 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2100 else
2101 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2103 rcu_read_unlock();
2104 atomic_set(&new->refcnt, 1);
2105 return new;
2108 /* Slow path of a mempolicy comparison */
2109 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2111 if (!a || !b)
2112 return false;
2113 if (a->mode != b->mode)
2114 return false;
2115 if (a->flags != b->flags)
2116 return false;
2117 if (mpol_store_user_nodemask(a))
2118 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2119 return false;
2121 switch (a->mode) {
2122 case MPOL_BIND:
2123 /* Fall through */
2124 case MPOL_INTERLEAVE:
2125 return !!nodes_equal(a->v.nodes, b->v.nodes);
2126 case MPOL_PREFERRED:
2127 return a->v.preferred_node == b->v.preferred_node;
2128 default:
2129 BUG();
2130 return false;
2135 * Shared memory backing store policy support.
2137 * Remember policies even when nobody has shared memory mapped.
2138 * The policies are kept in Red-Black tree linked from the inode.
2139 * They are protected by the sp->lock spinlock, which should be held
2140 * for any accesses to the tree.
2143 /* lookup first element intersecting start-end */
2144 /* Caller holds sp->lock */
2145 static struct sp_node *
2146 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2148 struct rb_node *n = sp->root.rb_node;
2150 while (n) {
2151 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2153 if (start >= p->end)
2154 n = n->rb_right;
2155 else if (end <= p->start)
2156 n = n->rb_left;
2157 else
2158 break;
2160 if (!n)
2161 return NULL;
2162 for (;;) {
2163 struct sp_node *w = NULL;
2164 struct rb_node *prev = rb_prev(n);
2165 if (!prev)
2166 break;
2167 w = rb_entry(prev, struct sp_node, nd);
2168 if (w->end <= start)
2169 break;
2170 n = prev;
2172 return rb_entry(n, struct sp_node, nd);
2175 /* Insert a new shared policy into the list. */
2176 /* Caller holds sp->lock */
2177 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2179 struct rb_node **p = &sp->root.rb_node;
2180 struct rb_node *parent = NULL;
2181 struct sp_node *nd;
2183 while (*p) {
2184 parent = *p;
2185 nd = rb_entry(parent, struct sp_node, nd);
2186 if (new->start < nd->start)
2187 p = &(*p)->rb_left;
2188 else if (new->end > nd->end)
2189 p = &(*p)->rb_right;
2190 else
2191 BUG();
2193 rb_link_node(&new->nd, parent, p);
2194 rb_insert_color(&new->nd, &sp->root);
2195 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2196 new->policy ? new->policy->mode : 0);
2199 /* Find shared policy intersecting idx */
2200 struct mempolicy *
2201 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2203 struct mempolicy *pol = NULL;
2204 struct sp_node *sn;
2206 if (!sp->root.rb_node)
2207 return NULL;
2208 spin_lock(&sp->lock);
2209 sn = sp_lookup(sp, idx, idx+1);
2210 if (sn) {
2211 mpol_get(sn->policy);
2212 pol = sn->policy;
2214 spin_unlock(&sp->lock);
2215 return pol;
2218 static void sp_free(struct sp_node *n)
2220 mpol_put(n->policy);
2221 kmem_cache_free(sn_cache, n);
2225 * mpol_misplaced - check whether current page node is valid in policy
2227 * @page - page to be checked
2228 * @vma - vm area where page mapped
2229 * @addr - virtual address where page mapped
2231 * Lookup current policy node id for vma,addr and "compare to" page's
2232 * node id.
2234 * Returns:
2235 * -1 - not misplaced, page is in the right node
2236 * node - node id where the page should be
2238 * Policy determination "mimics" alloc_page_vma().
2239 * Called from fault path where we know the vma and faulting address.
2241 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2243 struct mempolicy *pol;
2244 struct zone *zone;
2245 int curnid = page_to_nid(page);
2246 unsigned long pgoff;
2247 int polnid = -1;
2248 int ret = -1;
2250 BUG_ON(!vma);
2252 pol = get_vma_policy(current, vma, addr);
2253 if (!(pol->flags & MPOL_F_MOF))
2254 goto out;
2256 switch (pol->mode) {
2257 case MPOL_INTERLEAVE:
2258 BUG_ON(addr >= vma->vm_end);
2259 BUG_ON(addr < vma->vm_start);
2261 pgoff = vma->vm_pgoff;
2262 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2263 polnid = offset_il_node(pol, vma, pgoff);
2264 break;
2266 case MPOL_PREFERRED:
2267 if (pol->flags & MPOL_F_LOCAL)
2268 polnid = numa_node_id();
2269 else
2270 polnid = pol->v.preferred_node;
2271 break;
2273 case MPOL_BIND:
2275 * allows binding to multiple nodes.
2276 * use current page if in policy nodemask,
2277 * else select nearest allowed node, if any.
2278 * If no allowed nodes, use current [!misplaced].
2280 if (node_isset(curnid, pol->v.nodes))
2281 goto out;
2282 (void)first_zones_zonelist(
2283 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2284 gfp_zone(GFP_HIGHUSER),
2285 &pol->v.nodes, &zone);
2286 polnid = zone->node;
2287 break;
2289 default:
2290 BUG();
2293 /* Migrate the page towards the node whose CPU is referencing it */
2294 if (pol->flags & MPOL_F_MORON) {
2295 int last_nid;
2297 polnid = numa_node_id();
2300 * Multi-stage node selection is used in conjunction
2301 * with a periodic migration fault to build a temporal
2302 * task<->page relation. By using a two-stage filter we
2303 * remove short/unlikely relations.
2305 * Using P(p) ~ n_p / n_t as per frequentist
2306 * probability, we can equate a task's usage of a
2307 * particular page (n_p) per total usage of this
2308 * page (n_t) (in a given time-span) to a probability.
2310 * Our periodic faults will sample this probability and
2311 * getting the same result twice in a row, given these
2312 * samples are fully independent, is then given by
2313 * P(n)^2, provided our sample period is sufficiently
2314 * short compared to the usage pattern.
2316 * This quadric squishes small probabilities, making
2317 * it less likely we act on an unlikely task<->page
2318 * relation.
2320 last_nid = page_nid_xchg_last(page, polnid);
2321 if (last_nid != polnid)
2322 goto out;
2325 if (curnid != polnid)
2326 ret = polnid;
2327 out:
2328 mpol_cond_put(pol);
2330 return ret;
2333 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2335 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2336 rb_erase(&n->nd, &sp->root);
2337 sp_free(n);
2340 static void sp_node_init(struct sp_node *node, unsigned long start,
2341 unsigned long end, struct mempolicy *pol)
2343 node->start = start;
2344 node->end = end;
2345 node->policy = pol;
2348 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2349 struct mempolicy *pol)
2351 struct sp_node *n;
2352 struct mempolicy *newpol;
2354 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2355 if (!n)
2356 return NULL;
2358 newpol = mpol_dup(pol);
2359 if (IS_ERR(newpol)) {
2360 kmem_cache_free(sn_cache, n);
2361 return NULL;
2363 newpol->flags |= MPOL_F_SHARED;
2364 sp_node_init(n, start, end, newpol);
2366 return n;
2369 /* Replace a policy range. */
2370 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2371 unsigned long end, struct sp_node *new)
2373 struct sp_node *n;
2374 struct sp_node *n_new = NULL;
2375 struct mempolicy *mpol_new = NULL;
2376 int ret = 0;
2378 restart:
2379 spin_lock(&sp->lock);
2380 n = sp_lookup(sp, start, end);
2381 /* Take care of old policies in the same range. */
2382 while (n && n->start < end) {
2383 struct rb_node *next = rb_next(&n->nd);
2384 if (n->start >= start) {
2385 if (n->end <= end)
2386 sp_delete(sp, n);
2387 else
2388 n->start = end;
2389 } else {
2390 /* Old policy spanning whole new range. */
2391 if (n->end > end) {
2392 if (!n_new)
2393 goto alloc_new;
2395 *mpol_new = *n->policy;
2396 atomic_set(&mpol_new->refcnt, 1);
2397 sp_node_init(n_new, end, n->end, mpol_new);
2398 n->end = start;
2399 sp_insert(sp, n_new);
2400 n_new = NULL;
2401 mpol_new = NULL;
2402 break;
2403 } else
2404 n->end = start;
2406 if (!next)
2407 break;
2408 n = rb_entry(next, struct sp_node, nd);
2410 if (new)
2411 sp_insert(sp, new);
2412 spin_unlock(&sp->lock);
2413 ret = 0;
2415 err_out:
2416 if (mpol_new)
2417 mpol_put(mpol_new);
2418 if (n_new)
2419 kmem_cache_free(sn_cache, n_new);
2421 return ret;
2423 alloc_new:
2424 spin_unlock(&sp->lock);
2425 ret = -ENOMEM;
2426 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2427 if (!n_new)
2428 goto err_out;
2429 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2430 if (!mpol_new)
2431 goto err_out;
2432 goto restart;
2436 * mpol_shared_policy_init - initialize shared policy for inode
2437 * @sp: pointer to inode shared policy
2438 * @mpol: struct mempolicy to install
2440 * Install non-NULL @mpol in inode's shared policy rb-tree.
2441 * On entry, the current task has a reference on a non-NULL @mpol.
2442 * This must be released on exit.
2443 * This is called at get_inode() calls and we can use GFP_KERNEL.
2445 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2447 int ret;
2449 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2450 spin_lock_init(&sp->lock);
2452 if (mpol) {
2453 struct vm_area_struct pvma;
2454 struct mempolicy *new;
2455 NODEMASK_SCRATCH(scratch);
2457 if (!scratch)
2458 goto put_mpol;
2459 /* contextualize the tmpfs mount point mempolicy */
2460 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2461 if (IS_ERR(new))
2462 goto free_scratch; /* no valid nodemask intersection */
2464 task_lock(current);
2465 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2466 task_unlock(current);
2467 if (ret)
2468 goto put_new;
2470 /* Create pseudo-vma that contains just the policy */
2471 memset(&pvma, 0, sizeof(struct vm_area_struct));
2472 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2473 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2475 put_new:
2476 mpol_put(new); /* drop initial ref */
2477 free_scratch:
2478 NODEMASK_SCRATCH_FREE(scratch);
2479 put_mpol:
2480 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2484 int mpol_set_shared_policy(struct shared_policy *info,
2485 struct vm_area_struct *vma, struct mempolicy *npol)
2487 int err;
2488 struct sp_node *new = NULL;
2489 unsigned long sz = vma_pages(vma);
2491 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2492 vma->vm_pgoff,
2493 sz, npol ? npol->mode : -1,
2494 npol ? npol->flags : -1,
2495 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2497 if (npol) {
2498 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2499 if (!new)
2500 return -ENOMEM;
2502 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2503 if (err && new)
2504 sp_free(new);
2505 return err;
2508 /* Free a backing policy store on inode delete. */
2509 void mpol_free_shared_policy(struct shared_policy *p)
2511 struct sp_node *n;
2512 struct rb_node *next;
2514 if (!p->root.rb_node)
2515 return;
2516 spin_lock(&p->lock);
2517 next = rb_first(&p->root);
2518 while (next) {
2519 n = rb_entry(next, struct sp_node, nd);
2520 next = rb_next(&n->nd);
2521 sp_delete(p, n);
2523 spin_unlock(&p->lock);
2526 #ifdef CONFIG_NUMA_BALANCING
2527 static bool __initdata numabalancing_override;
2529 static void __init check_numabalancing_enable(void)
2531 bool numabalancing_default = false;
2533 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2534 numabalancing_default = true;
2536 if (nr_node_ids > 1 && !numabalancing_override) {
2537 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2538 "Configure with numa_balancing= or sysctl");
2539 set_numabalancing_state(numabalancing_default);
2543 static int __init setup_numabalancing(char *str)
2545 int ret = 0;
2546 if (!str)
2547 goto out;
2548 numabalancing_override = true;
2550 if (!strcmp(str, "enable")) {
2551 set_numabalancing_state(true);
2552 ret = 1;
2553 } else if (!strcmp(str, "disable")) {
2554 set_numabalancing_state(false);
2555 ret = 1;
2557 out:
2558 if (!ret)
2559 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2561 return ret;
2563 __setup("numa_balancing=", setup_numabalancing);
2564 #else
2565 static inline void __init check_numabalancing_enable(void)
2568 #endif /* CONFIG_NUMA_BALANCING */
2570 /* assumes fs == KERNEL_DS */
2571 void __init numa_policy_init(void)
2573 nodemask_t interleave_nodes;
2574 unsigned long largest = 0;
2575 int nid, prefer = 0;
2577 policy_cache = kmem_cache_create("numa_policy",
2578 sizeof(struct mempolicy),
2579 0, SLAB_PANIC, NULL);
2581 sn_cache = kmem_cache_create("shared_policy_node",
2582 sizeof(struct sp_node),
2583 0, SLAB_PANIC, NULL);
2585 for_each_node(nid) {
2586 preferred_node_policy[nid] = (struct mempolicy) {
2587 .refcnt = ATOMIC_INIT(1),
2588 .mode = MPOL_PREFERRED,
2589 .flags = MPOL_F_MOF | MPOL_F_MORON,
2590 .v = { .preferred_node = nid, },
2595 * Set interleaving policy for system init. Interleaving is only
2596 * enabled across suitably sized nodes (default is >= 16MB), or
2597 * fall back to the largest node if they're all smaller.
2599 nodes_clear(interleave_nodes);
2600 for_each_node_state(nid, N_MEMORY) {
2601 unsigned long total_pages = node_present_pages(nid);
2603 /* Preserve the largest node */
2604 if (largest < total_pages) {
2605 largest = total_pages;
2606 prefer = nid;
2609 /* Interleave this node? */
2610 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2611 node_set(nid, interleave_nodes);
2614 /* All too small, use the largest */
2615 if (unlikely(nodes_empty(interleave_nodes)))
2616 node_set(prefer, interleave_nodes);
2618 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2619 printk("numa_policy_init: interleaving failed\n");
2621 check_numabalancing_enable();
2624 /* Reset policy of current process to default */
2625 void numa_default_policy(void)
2627 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2631 * Parse and format mempolicy from/to strings
2635 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2637 static const char * const policy_modes[] =
2639 [MPOL_DEFAULT] = "default",
2640 [MPOL_PREFERRED] = "prefer",
2641 [MPOL_BIND] = "bind",
2642 [MPOL_INTERLEAVE] = "interleave",
2643 [MPOL_LOCAL] = "local",
2647 #ifdef CONFIG_TMPFS
2649 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2650 * @str: string containing mempolicy to parse
2651 * @mpol: pointer to struct mempolicy pointer, returned on success.
2653 * Format of input:
2654 * <mode>[=<flags>][:<nodelist>]
2656 * On success, returns 0, else 1
2658 int mpol_parse_str(char *str, struct mempolicy **mpol)
2660 struct mempolicy *new = NULL;
2661 unsigned short mode;
2662 unsigned short mode_flags;
2663 nodemask_t nodes;
2664 char *nodelist = strchr(str, ':');
2665 char *flags = strchr(str, '=');
2666 int err = 1;
2668 if (nodelist) {
2669 /* NUL-terminate mode or flags string */
2670 *nodelist++ = '\0';
2671 if (nodelist_parse(nodelist, nodes))
2672 goto out;
2673 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2674 goto out;
2675 } else
2676 nodes_clear(nodes);
2678 if (flags)
2679 *flags++ = '\0'; /* terminate mode string */
2681 for (mode = 0; mode < MPOL_MAX; mode++) {
2682 if (!strcmp(str, policy_modes[mode])) {
2683 break;
2686 if (mode >= MPOL_MAX)
2687 goto out;
2689 switch (mode) {
2690 case MPOL_PREFERRED:
2692 * Insist on a nodelist of one node only
2694 if (nodelist) {
2695 char *rest = nodelist;
2696 while (isdigit(*rest))
2697 rest++;
2698 if (*rest)
2699 goto out;
2701 break;
2702 case MPOL_INTERLEAVE:
2704 * Default to online nodes with memory if no nodelist
2706 if (!nodelist)
2707 nodes = node_states[N_MEMORY];
2708 break;
2709 case MPOL_LOCAL:
2711 * Don't allow a nodelist; mpol_new() checks flags
2713 if (nodelist)
2714 goto out;
2715 mode = MPOL_PREFERRED;
2716 break;
2717 case MPOL_DEFAULT:
2719 * Insist on a empty nodelist
2721 if (!nodelist)
2722 err = 0;
2723 goto out;
2724 case MPOL_BIND:
2726 * Insist on a nodelist
2728 if (!nodelist)
2729 goto out;
2732 mode_flags = 0;
2733 if (flags) {
2735 * Currently, we only support two mutually exclusive
2736 * mode flags.
2738 if (!strcmp(flags, "static"))
2739 mode_flags |= MPOL_F_STATIC_NODES;
2740 else if (!strcmp(flags, "relative"))
2741 mode_flags |= MPOL_F_RELATIVE_NODES;
2742 else
2743 goto out;
2746 new = mpol_new(mode, mode_flags, &nodes);
2747 if (IS_ERR(new))
2748 goto out;
2751 * Save nodes for mpol_to_str() to show the tmpfs mount options
2752 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2754 if (mode != MPOL_PREFERRED)
2755 new->v.nodes = nodes;
2756 else if (nodelist)
2757 new->v.preferred_node = first_node(nodes);
2758 else
2759 new->flags |= MPOL_F_LOCAL;
2762 * Save nodes for contextualization: this will be used to "clone"
2763 * the mempolicy in a specific context [cpuset] at a later time.
2765 new->w.user_nodemask = nodes;
2767 err = 0;
2769 out:
2770 /* Restore string for error message */
2771 if (nodelist)
2772 *--nodelist = ':';
2773 if (flags)
2774 *--flags = '=';
2775 if (!err)
2776 *mpol = new;
2777 return err;
2779 #endif /* CONFIG_TMPFS */
2782 * mpol_to_str - format a mempolicy structure for printing
2783 * @buffer: to contain formatted mempolicy string
2784 * @maxlen: length of @buffer
2785 * @pol: pointer to mempolicy to be formatted
2787 * Convert a mempolicy into a string.
2788 * Returns the number of characters in buffer (if positive)
2789 * or an error (negative)
2791 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2793 char *p = buffer;
2794 int l;
2795 nodemask_t nodes;
2796 unsigned short mode;
2797 unsigned short flags = pol ? pol->flags : 0;
2800 * Sanity check: room for longest mode, flag and some nodes
2802 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2804 if (!pol || pol == &default_policy)
2805 mode = MPOL_DEFAULT;
2806 else
2807 mode = pol->mode;
2809 switch (mode) {
2810 case MPOL_DEFAULT:
2811 nodes_clear(nodes);
2812 break;
2814 case MPOL_PREFERRED:
2815 nodes_clear(nodes);
2816 if (flags & MPOL_F_LOCAL)
2817 mode = MPOL_LOCAL;
2818 else
2819 node_set(pol->v.preferred_node, nodes);
2820 break;
2822 case MPOL_BIND:
2823 /* Fall through */
2824 case MPOL_INTERLEAVE:
2825 nodes = pol->v.nodes;
2826 break;
2828 default:
2829 return -EINVAL;
2832 l = strlen(policy_modes[mode]);
2833 if (buffer + maxlen < p + l + 1)
2834 return -ENOSPC;
2836 strcpy(p, policy_modes[mode]);
2837 p += l;
2839 if (flags & MPOL_MODE_FLAGS) {
2840 if (buffer + maxlen < p + 2)
2841 return -ENOSPC;
2842 *p++ = '=';
2845 * Currently, the only defined flags are mutually exclusive
2847 if (flags & MPOL_F_STATIC_NODES)
2848 p += snprintf(p, buffer + maxlen - p, "static");
2849 else if (flags & MPOL_F_RELATIVE_NODES)
2850 p += snprintf(p, buffer + maxlen - p, "relative");
2853 if (!nodes_empty(nodes)) {
2854 if (buffer + maxlen < p + 2)
2855 return -ENOSPC;
2856 *p++ = ':';
2857 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2859 return p - buffer;