Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux/fpc-iii.git] / mm / mempolicy.c
blobae3c8f3595d4ff522f0427b05ace2f7e041da8ad
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;
127 if (!pol) {
128 int 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
134 * boot
136 if (!pol->mode)
137 pol = NULL;
141 return pol;
144 static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
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.
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
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
162 } mpol_ops[MPOL_MAX];
164 /* Check that the nodemask contains at least one populated zone */
165 static int is_valid_nodemask(const nodemask_t *nodemask)
167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
170 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
172 return pol->flags & MPOL_MODE_FLAGS;
175 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
178 nodemask_t tmp;
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
183 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
185 if (nodes_empty(*nodes))
186 return -EINVAL;
187 pol->v.nodes = *nodes;
188 return 0;
191 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
193 if (!nodes)
194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
197 else
198 pol->v.preferred_node = first_node(*nodes);
199 return 0;
202 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
204 if (!is_valid_nodemask(nodes))
205 return -EINVAL;
206 pol->v.nodes = *nodes;
207 return 0;
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
219 static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
222 int ret;
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
225 if (pol == NULL)
226 return 0;
227 /* Check N_MEMORY */
228 nodes_and(nsc->mask1,
229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
231 VM_BUG_ON(!nodes);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
234 else {
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
237 else
238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
242 else
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
247 if (nodes)
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
249 else
250 ret = mpol_ops[pol->mode].create(pol, NULL);
251 return ret;
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
259 nodemask_t *nodes)
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
269 return NULL;
271 VM_BUG_ON(!nodes);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
291 if (!policy)
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
294 policy->mode = mode;
295 policy->flags = flags;
297 return policy;
300 /* Slow path of a mpol destructor. */
301 void __mpol_put(struct mempolicy *p)
303 if (!atomic_dec_and_test(&p->refcnt))
304 return;
305 kmem_cache_free(policy_cache, p);
308 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
314 * step:
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
319 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
322 nodemask_t tmp;
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
328 else {
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
331 * result
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
340 } else
341 BUG();
344 if (nodes_empty(tmp))
345 tmp = *nodes;
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
350 pol->v.nodes = tmp;
351 else
352 BUG();
354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
363 static void mpol_rebind_preferred(struct mempolicy *pol,
364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
367 nodemask_t tmp;
369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
372 if (node_isset(node, *nodes)) {
373 pol->v.preferred_node = node;
374 pol->flags &= ~MPOL_F_LOCAL;
375 } else
376 pol->flags |= MPOL_F_LOCAL;
377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
383 *nodes);
384 pol->w.cpuset_mems_allowed = *nodes;
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
395 * page.
396 * If we have a lock to protect task->mempolicy in read-side, we do
397 * rebind directly.
399 * step:
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
404 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
407 if (!pol)
408 return;
409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
411 return;
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
414 return;
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
417 BUG();
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
424 BUG();
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
433 * Called with task's alloc_lock held.
436 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
439 mpol_rebind_policy(tsk->mempolicy, new, step);
443 * Rebind each vma in mm to new nodemask.
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
448 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
450 struct vm_area_struct *vma;
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
455 up_write(&mm->mmap_sem);
458 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
459 [MPOL_DEFAULT] = {
460 .rebind = mpol_rebind_default,
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
466 [MPOL_PREFERRED] = {
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
470 [MPOL_BIND] = {
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
476 static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
483 static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
486 void *private)
488 pte_t *orig_pte;
489 pte_t *pte;
490 spinlock_t *ptl;
492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
493 do {
494 struct page *page;
495 int nid;
497 if (!pte_present(*pte))
498 continue;
499 page = vm_normal_page(vma, addr, *pte);
500 if (!page)
501 continue;
503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
506 if (PageReserved(page))
507 continue;
508 nid = page_to_nid(page);
509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
510 continue;
512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
513 migrate_page_add(page, private, flags);
514 else
515 break;
516 } while (pte++, addr += PAGE_SIZE, addr != end);
517 pte_unmap_unlock(orig_pte, ptl);
518 return addr != end;
521 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
523 void *private)
525 #ifdef CONFIG_HUGETLB_PAGE
526 int nid;
527 struct page *page;
528 spinlock_t *ptl;
530 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
531 page = pte_page(huge_ptep_get((pte_t *)pmd));
532 nid = page_to_nid(page);
533 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
534 goto unlock;
535 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
536 if (flags & (MPOL_MF_MOVE_ALL) ||
537 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
538 isolate_huge_page(page, private);
539 unlock:
540 spin_unlock(ptl);
541 #else
542 BUG();
543 #endif
546 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
547 unsigned long addr, unsigned long end,
548 const nodemask_t *nodes, unsigned long flags,
549 void *private)
551 pmd_t *pmd;
552 unsigned long next;
554 pmd = pmd_offset(pud, addr);
555 do {
556 next = pmd_addr_end(addr, end);
557 if (!pmd_present(*pmd))
558 continue;
559 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
560 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
561 flags, private);
562 continue;
564 split_huge_page_pmd(vma, addr, pmd);
565 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
566 continue;
567 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
568 flags, private))
569 return -EIO;
570 } while (pmd++, addr = next, addr != end);
571 return 0;
574 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
575 unsigned long addr, unsigned long end,
576 const nodemask_t *nodes, unsigned long flags,
577 void *private)
579 pud_t *pud;
580 unsigned long next;
582 pud = pud_offset(pgd, addr);
583 do {
584 next = pud_addr_end(addr, end);
585 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
586 continue;
587 if (pud_none_or_clear_bad(pud))
588 continue;
589 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
590 flags, private))
591 return -EIO;
592 } while (pud++, addr = next, addr != end);
593 return 0;
596 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
597 unsigned long addr, unsigned long end,
598 const nodemask_t *nodes, unsigned long flags,
599 void *private)
601 pgd_t *pgd;
602 unsigned long next;
604 pgd = pgd_offset(vma->vm_mm, addr);
605 do {
606 next = pgd_addr_end(addr, end);
607 if (pgd_none_or_clear_bad(pgd))
608 continue;
609 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
610 flags, private))
611 return -EIO;
612 } while (pgd++, addr = next, addr != end);
613 return 0;
616 #ifdef CONFIG_NUMA_BALANCING
618 * This is used to mark a range of virtual addresses to be inaccessible.
619 * These are later cleared by a NUMA hinting fault. Depending on these
620 * faults, pages may be migrated for better NUMA placement.
622 * This is assuming that NUMA faults are handled using PROT_NONE. If
623 * an architecture makes a different choice, it will need further
624 * changes to the core.
626 unsigned long change_prot_numa(struct vm_area_struct *vma,
627 unsigned long addr, unsigned long end)
629 int nr_updated;
631 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
632 if (nr_updated)
633 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
635 return nr_updated;
637 #else
638 static unsigned long change_prot_numa(struct vm_area_struct *vma,
639 unsigned long addr, unsigned long end)
641 return 0;
643 #endif /* CONFIG_NUMA_BALANCING */
646 * Walk through page tables and collect pages to be migrated.
648 * If pages found in a given range are on a set of nodes (determined by
649 * @nodes and @flags,) it's isolated and queued to the pagelist which is
650 * passed via @private.)
652 static struct vm_area_struct *
653 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
654 const nodemask_t *nodes, unsigned long flags, void *private)
656 int err;
657 struct vm_area_struct *first, *vma, *prev;
660 first = find_vma(mm, start);
661 if (!first)
662 return ERR_PTR(-EFAULT);
663 prev = NULL;
664 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
665 unsigned long endvma = vma->vm_end;
667 if (endvma > end)
668 endvma = end;
669 if (vma->vm_start > start)
670 start = vma->vm_start;
672 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
673 if (!vma->vm_next && vma->vm_end < end)
674 return ERR_PTR(-EFAULT);
675 if (prev && prev->vm_end < vma->vm_start)
676 return ERR_PTR(-EFAULT);
679 if (flags & MPOL_MF_LAZY) {
680 change_prot_numa(vma, start, endvma);
681 goto next;
684 if ((flags & MPOL_MF_STRICT) ||
685 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
686 vma_migratable(vma))) {
688 err = queue_pages_pgd_range(vma, start, endvma, nodes,
689 flags, private);
690 if (err) {
691 first = ERR_PTR(err);
692 break;
695 next:
696 prev = vma;
698 return first;
702 * Apply policy to a single VMA
703 * This must be called with the mmap_sem held for writing.
705 static int vma_replace_policy(struct vm_area_struct *vma,
706 struct mempolicy *pol)
708 int err;
709 struct mempolicy *old;
710 struct mempolicy *new;
712 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
713 vma->vm_start, vma->vm_end, vma->vm_pgoff,
714 vma->vm_ops, vma->vm_file,
715 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
717 new = mpol_dup(pol);
718 if (IS_ERR(new))
719 return PTR_ERR(new);
721 if (vma->vm_ops && vma->vm_ops->set_policy) {
722 err = vma->vm_ops->set_policy(vma, new);
723 if (err)
724 goto err_out;
727 old = vma->vm_policy;
728 vma->vm_policy = new; /* protected by mmap_sem */
729 mpol_put(old);
731 return 0;
732 err_out:
733 mpol_put(new);
734 return err;
737 /* Step 2: apply policy to a range and do splits. */
738 static int mbind_range(struct mm_struct *mm, unsigned long start,
739 unsigned long end, struct mempolicy *new_pol)
741 struct vm_area_struct *next;
742 struct vm_area_struct *prev;
743 struct vm_area_struct *vma;
744 int err = 0;
745 pgoff_t pgoff;
746 unsigned long vmstart;
747 unsigned long vmend;
749 vma = find_vma(mm, start);
750 if (!vma || vma->vm_start > start)
751 return -EFAULT;
753 prev = vma->vm_prev;
754 if (start > vma->vm_start)
755 prev = vma;
757 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
758 next = vma->vm_next;
759 vmstart = max(start, vma->vm_start);
760 vmend = min(end, vma->vm_end);
762 if (mpol_equal(vma_policy(vma), new_pol))
763 continue;
765 pgoff = vma->vm_pgoff +
766 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
767 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
768 vma->anon_vma, vma->vm_file, pgoff,
769 new_pol);
770 if (prev) {
771 vma = prev;
772 next = vma->vm_next;
773 if (mpol_equal(vma_policy(vma), new_pol))
774 continue;
775 /* vma_merge() joined vma && vma->next, case 8 */
776 goto replace;
778 if (vma->vm_start != vmstart) {
779 err = split_vma(vma->vm_mm, vma, vmstart, 1);
780 if (err)
781 goto out;
783 if (vma->vm_end != vmend) {
784 err = split_vma(vma->vm_mm, vma, vmend, 0);
785 if (err)
786 goto out;
788 replace:
789 err = vma_replace_policy(vma, new_pol);
790 if (err)
791 goto out;
794 out:
795 return err;
799 * Update task->flags PF_MEMPOLICY bit: set iff non-default
800 * mempolicy. Allows more rapid checking of this (combined perhaps
801 * with other PF_* flag bits) on memory allocation hot code paths.
803 * If called from outside this file, the task 'p' should -only- be
804 * a newly forked child not yet visible on the task list, because
805 * manipulating the task flags of a visible task is not safe.
807 * The above limitation is why this routine has the funny name
808 * mpol_fix_fork_child_flag().
810 * It is also safe to call this with a task pointer of current,
811 * which the static wrapper mpol_set_task_struct_flag() does,
812 * for use within this file.
815 void mpol_fix_fork_child_flag(struct task_struct *p)
817 if (p->mempolicy)
818 p->flags |= PF_MEMPOLICY;
819 else
820 p->flags &= ~PF_MEMPOLICY;
823 static void mpol_set_task_struct_flag(void)
825 mpol_fix_fork_child_flag(current);
828 /* Set the process memory policy */
829 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
830 nodemask_t *nodes)
832 struct mempolicy *new, *old;
833 struct mm_struct *mm = current->mm;
834 NODEMASK_SCRATCH(scratch);
835 int ret;
837 if (!scratch)
838 return -ENOMEM;
840 new = mpol_new(mode, flags, nodes);
841 if (IS_ERR(new)) {
842 ret = PTR_ERR(new);
843 goto out;
846 * prevent changing our mempolicy while show_numa_maps()
847 * is using it.
848 * Note: do_set_mempolicy() can be called at init time
849 * with no 'mm'.
851 if (mm)
852 down_write(&mm->mmap_sem);
853 task_lock(current);
854 ret = mpol_set_nodemask(new, nodes, scratch);
855 if (ret) {
856 task_unlock(current);
857 if (mm)
858 up_write(&mm->mmap_sem);
859 mpol_put(new);
860 goto out;
862 old = current->mempolicy;
863 current->mempolicy = new;
864 mpol_set_task_struct_flag();
865 if (new && new->mode == MPOL_INTERLEAVE &&
866 nodes_weight(new->v.nodes))
867 current->il_next = first_node(new->v.nodes);
868 task_unlock(current);
869 if (mm)
870 up_write(&mm->mmap_sem);
872 mpol_put(old);
873 ret = 0;
874 out:
875 NODEMASK_SCRATCH_FREE(scratch);
876 return ret;
880 * Return nodemask for policy for get_mempolicy() query
882 * Called with task's alloc_lock held
884 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
886 nodes_clear(*nodes);
887 if (p == &default_policy)
888 return;
890 switch (p->mode) {
891 case MPOL_BIND:
892 /* Fall through */
893 case MPOL_INTERLEAVE:
894 *nodes = p->v.nodes;
895 break;
896 case MPOL_PREFERRED:
897 if (!(p->flags & MPOL_F_LOCAL))
898 node_set(p->v.preferred_node, *nodes);
899 /* else return empty node mask for local allocation */
900 break;
901 default:
902 BUG();
906 static int lookup_node(struct mm_struct *mm, unsigned long addr)
908 struct page *p;
909 int err;
911 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
912 if (err >= 0) {
913 err = page_to_nid(p);
914 put_page(p);
916 return err;
919 /* Retrieve NUMA policy */
920 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
921 unsigned long addr, unsigned long flags)
923 int err;
924 struct mm_struct *mm = current->mm;
925 struct vm_area_struct *vma = NULL;
926 struct mempolicy *pol = current->mempolicy;
928 if (flags &
929 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
930 return -EINVAL;
932 if (flags & MPOL_F_MEMS_ALLOWED) {
933 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
934 return -EINVAL;
935 *policy = 0; /* just so it's initialized */
936 task_lock(current);
937 *nmask = cpuset_current_mems_allowed;
938 task_unlock(current);
939 return 0;
942 if (flags & MPOL_F_ADDR) {
944 * Do NOT fall back to task policy if the
945 * vma/shared policy at addr is NULL. We
946 * want to return MPOL_DEFAULT in this case.
948 down_read(&mm->mmap_sem);
949 vma = find_vma_intersection(mm, addr, addr+1);
950 if (!vma) {
951 up_read(&mm->mmap_sem);
952 return -EFAULT;
954 if (vma->vm_ops && vma->vm_ops->get_policy)
955 pol = vma->vm_ops->get_policy(vma, addr);
956 else
957 pol = vma->vm_policy;
958 } else if (addr)
959 return -EINVAL;
961 if (!pol)
962 pol = &default_policy; /* indicates default behavior */
964 if (flags & MPOL_F_NODE) {
965 if (flags & MPOL_F_ADDR) {
966 err = lookup_node(mm, addr);
967 if (err < 0)
968 goto out;
969 *policy = err;
970 } else if (pol == current->mempolicy &&
971 pol->mode == MPOL_INTERLEAVE) {
972 *policy = current->il_next;
973 } else {
974 err = -EINVAL;
975 goto out;
977 } else {
978 *policy = pol == &default_policy ? MPOL_DEFAULT :
979 pol->mode;
981 * Internal mempolicy flags must be masked off before exposing
982 * the policy to userspace.
984 *policy |= (pol->flags & MPOL_MODE_FLAGS);
987 if (vma) {
988 up_read(&current->mm->mmap_sem);
989 vma = NULL;
992 err = 0;
993 if (nmask) {
994 if (mpol_store_user_nodemask(pol)) {
995 *nmask = pol->w.user_nodemask;
996 } else {
997 task_lock(current);
998 get_policy_nodemask(pol, nmask);
999 task_unlock(current);
1003 out:
1004 mpol_cond_put(pol);
1005 if (vma)
1006 up_read(&current->mm->mmap_sem);
1007 return err;
1010 #ifdef CONFIG_MIGRATION
1012 * page migration
1014 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1015 unsigned long flags)
1018 * Avoid migrating a page that is shared with others.
1020 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1021 if (!isolate_lru_page(page)) {
1022 list_add_tail(&page->lru, pagelist);
1023 inc_zone_page_state(page, NR_ISOLATED_ANON +
1024 page_is_file_cache(page));
1029 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1031 if (PageHuge(page))
1032 return alloc_huge_page_node(page_hstate(compound_head(page)),
1033 node);
1034 else
1035 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1039 * Migrate pages from one node to a target node.
1040 * Returns error or the number of pages not migrated.
1042 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1043 int flags)
1045 nodemask_t nmask;
1046 LIST_HEAD(pagelist);
1047 int err = 0;
1049 nodes_clear(nmask);
1050 node_set(source, nmask);
1053 * This does not "check" the range but isolates all pages that
1054 * need migration. Between passing in the full user address
1055 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1057 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1058 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1059 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1061 if (!list_empty(&pagelist)) {
1062 err = migrate_pages(&pagelist, new_node_page, dest,
1063 MIGRATE_SYNC, MR_SYSCALL);
1064 if (err)
1065 putback_movable_pages(&pagelist);
1068 return err;
1072 * Move pages between the two nodesets so as to preserve the physical
1073 * layout as much as possible.
1075 * Returns the number of page that could not be moved.
1077 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1078 const nodemask_t *to, int flags)
1080 int busy = 0;
1081 int err;
1082 nodemask_t tmp;
1084 err = migrate_prep();
1085 if (err)
1086 return err;
1088 down_read(&mm->mmap_sem);
1090 err = migrate_vmas(mm, from, to, flags);
1091 if (err)
1092 goto out;
1095 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1096 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1097 * bit in 'tmp', and return that <source, dest> pair for migration.
1098 * The pair of nodemasks 'to' and 'from' define the map.
1100 * If no pair of bits is found that way, fallback to picking some
1101 * pair of 'source' and 'dest' bits that are not the same. If the
1102 * 'source' and 'dest' bits are the same, this represents a node
1103 * that will be migrating to itself, so no pages need move.
1105 * If no bits are left in 'tmp', or if all remaining bits left
1106 * in 'tmp' correspond to the same bit in 'to', return false
1107 * (nothing left to migrate).
1109 * This lets us pick a pair of nodes to migrate between, such that
1110 * if possible the dest node is not already occupied by some other
1111 * source node, minimizing the risk of overloading the memory on a
1112 * node that would happen if we migrated incoming memory to a node
1113 * before migrating outgoing memory source that same node.
1115 * A single scan of tmp is sufficient. As we go, we remember the
1116 * most recent <s, d> pair that moved (s != d). If we find a pair
1117 * that not only moved, but what's better, moved to an empty slot
1118 * (d is not set in tmp), then we break out then, with that pair.
1119 * Otherwise when we finish scanning from_tmp, we at least have the
1120 * most recent <s, d> pair that moved. If we get all the way through
1121 * the scan of tmp without finding any node that moved, much less
1122 * moved to an empty node, then there is nothing left worth migrating.
1125 tmp = *from;
1126 while (!nodes_empty(tmp)) {
1127 int s,d;
1128 int source = NUMA_NO_NODE;
1129 int dest = 0;
1131 for_each_node_mask(s, tmp) {
1134 * do_migrate_pages() tries to maintain the relative
1135 * node relationship of the pages established between
1136 * threads and memory areas.
1138 * However if the number of source nodes is not equal to
1139 * the number of destination nodes we can not preserve
1140 * this node relative relationship. In that case, skip
1141 * copying memory from a node that is in the destination
1142 * mask.
1144 * Example: [2,3,4] -> [3,4,5] moves everything.
1145 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1148 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1149 (node_isset(s, *to)))
1150 continue;
1152 d = node_remap(s, *from, *to);
1153 if (s == d)
1154 continue;
1156 source = s; /* Node moved. Memorize */
1157 dest = d;
1159 /* dest not in remaining from nodes? */
1160 if (!node_isset(dest, tmp))
1161 break;
1163 if (source == NUMA_NO_NODE)
1164 break;
1166 node_clear(source, tmp);
1167 err = migrate_to_node(mm, source, dest, flags);
1168 if (err > 0)
1169 busy += err;
1170 if (err < 0)
1171 break;
1173 out:
1174 up_read(&mm->mmap_sem);
1175 if (err < 0)
1176 return err;
1177 return busy;
1182 * Allocate a new page for page migration based on vma policy.
1183 * Start assuming that page is mapped by vma pointed to by @private.
1184 * Search forward from there, if not. N.B., this assumes that the
1185 * list of pages handed to migrate_pages()--which is how we get here--
1186 * is in virtual address order.
1188 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1190 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1191 unsigned long uninitialized_var(address);
1193 while (vma) {
1194 address = page_address_in_vma(page, vma);
1195 if (address != -EFAULT)
1196 break;
1197 vma = vma->vm_next;
1200 if (PageHuge(page)) {
1201 BUG_ON(!vma);
1202 return alloc_huge_page_noerr(vma, address, 1);
1205 * if !vma, alloc_page_vma() will use task or system default policy
1207 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1209 #else
1211 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1212 unsigned long flags)
1216 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1217 const nodemask_t *to, int flags)
1219 return -ENOSYS;
1222 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1224 return NULL;
1226 #endif
1228 static long do_mbind(unsigned long start, unsigned long len,
1229 unsigned short mode, unsigned short mode_flags,
1230 nodemask_t *nmask, unsigned long flags)
1232 struct vm_area_struct *vma;
1233 struct mm_struct *mm = current->mm;
1234 struct mempolicy *new;
1235 unsigned long end;
1236 int err;
1237 LIST_HEAD(pagelist);
1239 if (flags & ~(unsigned long)MPOL_MF_VALID)
1240 return -EINVAL;
1241 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1242 return -EPERM;
1244 if (start & ~PAGE_MASK)
1245 return -EINVAL;
1247 if (mode == MPOL_DEFAULT)
1248 flags &= ~MPOL_MF_STRICT;
1250 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1251 end = start + len;
1253 if (end < start)
1254 return -EINVAL;
1255 if (end == start)
1256 return 0;
1258 new = mpol_new(mode, mode_flags, nmask);
1259 if (IS_ERR(new))
1260 return PTR_ERR(new);
1262 if (flags & MPOL_MF_LAZY)
1263 new->flags |= MPOL_F_MOF;
1266 * If we are using the default policy then operation
1267 * on discontinuous address spaces is okay after all
1269 if (!new)
1270 flags |= MPOL_MF_DISCONTIG_OK;
1272 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1273 start, start + len, mode, mode_flags,
1274 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1276 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1278 err = migrate_prep();
1279 if (err)
1280 goto mpol_out;
1283 NODEMASK_SCRATCH(scratch);
1284 if (scratch) {
1285 down_write(&mm->mmap_sem);
1286 task_lock(current);
1287 err = mpol_set_nodemask(new, nmask, scratch);
1288 task_unlock(current);
1289 if (err)
1290 up_write(&mm->mmap_sem);
1291 } else
1292 err = -ENOMEM;
1293 NODEMASK_SCRATCH_FREE(scratch);
1295 if (err)
1296 goto mpol_out;
1298 vma = queue_pages_range(mm, start, end, nmask,
1299 flags | MPOL_MF_INVERT, &pagelist);
1301 err = PTR_ERR(vma); /* maybe ... */
1302 if (!IS_ERR(vma))
1303 err = mbind_range(mm, start, end, new);
1305 if (!err) {
1306 int nr_failed = 0;
1308 if (!list_empty(&pagelist)) {
1309 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1310 nr_failed = migrate_pages(&pagelist, new_vma_page,
1311 (unsigned long)vma,
1312 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1313 if (nr_failed)
1314 putback_movable_pages(&pagelist);
1317 if (nr_failed && (flags & MPOL_MF_STRICT))
1318 err = -EIO;
1319 } else
1320 putback_movable_pages(&pagelist);
1322 up_write(&mm->mmap_sem);
1323 mpol_out:
1324 mpol_put(new);
1325 return err;
1329 * User space interface with variable sized bitmaps for nodelists.
1332 /* Copy a node mask from user space. */
1333 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1334 unsigned long maxnode)
1336 unsigned long k;
1337 unsigned long nlongs;
1338 unsigned long endmask;
1340 --maxnode;
1341 nodes_clear(*nodes);
1342 if (maxnode == 0 || !nmask)
1343 return 0;
1344 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1345 return -EINVAL;
1347 nlongs = BITS_TO_LONGS(maxnode);
1348 if ((maxnode % BITS_PER_LONG) == 0)
1349 endmask = ~0UL;
1350 else
1351 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1353 /* When the user specified more nodes than supported just check
1354 if the non supported part is all zero. */
1355 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1356 if (nlongs > PAGE_SIZE/sizeof(long))
1357 return -EINVAL;
1358 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1359 unsigned long t;
1360 if (get_user(t, nmask + k))
1361 return -EFAULT;
1362 if (k == nlongs - 1) {
1363 if (t & endmask)
1364 return -EINVAL;
1365 } else if (t)
1366 return -EINVAL;
1368 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1369 endmask = ~0UL;
1372 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1373 return -EFAULT;
1374 nodes_addr(*nodes)[nlongs-1] &= endmask;
1375 return 0;
1378 /* Copy a kernel node mask to user space */
1379 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1380 nodemask_t *nodes)
1382 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1383 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1385 if (copy > nbytes) {
1386 if (copy > PAGE_SIZE)
1387 return -EINVAL;
1388 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1389 return -EFAULT;
1390 copy = nbytes;
1392 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1395 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1396 unsigned long, mode, unsigned long __user *, nmask,
1397 unsigned long, maxnode, unsigned, flags)
1399 nodemask_t nodes;
1400 int err;
1401 unsigned short mode_flags;
1403 mode_flags = mode & MPOL_MODE_FLAGS;
1404 mode &= ~MPOL_MODE_FLAGS;
1405 if (mode >= MPOL_MAX)
1406 return -EINVAL;
1407 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1408 (mode_flags & MPOL_F_RELATIVE_NODES))
1409 return -EINVAL;
1410 err = get_nodes(&nodes, nmask, maxnode);
1411 if (err)
1412 return err;
1413 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1416 /* Set the process memory policy */
1417 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1418 unsigned long, maxnode)
1420 int err;
1421 nodemask_t nodes;
1422 unsigned short flags;
1424 flags = mode & MPOL_MODE_FLAGS;
1425 mode &= ~MPOL_MODE_FLAGS;
1426 if ((unsigned int)mode >= MPOL_MAX)
1427 return -EINVAL;
1428 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1429 return -EINVAL;
1430 err = get_nodes(&nodes, nmask, maxnode);
1431 if (err)
1432 return err;
1433 return do_set_mempolicy(mode, flags, &nodes);
1436 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1437 const unsigned long __user *, old_nodes,
1438 const unsigned long __user *, new_nodes)
1440 const struct cred *cred = current_cred(), *tcred;
1441 struct mm_struct *mm = NULL;
1442 struct task_struct *task;
1443 nodemask_t task_nodes;
1444 int err;
1445 nodemask_t *old;
1446 nodemask_t *new;
1447 NODEMASK_SCRATCH(scratch);
1449 if (!scratch)
1450 return -ENOMEM;
1452 old = &scratch->mask1;
1453 new = &scratch->mask2;
1455 err = get_nodes(old, old_nodes, maxnode);
1456 if (err)
1457 goto out;
1459 err = get_nodes(new, new_nodes, maxnode);
1460 if (err)
1461 goto out;
1463 /* Find the mm_struct */
1464 rcu_read_lock();
1465 task = pid ? find_task_by_vpid(pid) : current;
1466 if (!task) {
1467 rcu_read_unlock();
1468 err = -ESRCH;
1469 goto out;
1471 get_task_struct(task);
1473 err = -EINVAL;
1476 * Check if this process has the right to modify the specified
1477 * process. The right exists if the process has administrative
1478 * capabilities, superuser privileges or the same
1479 * userid as the target process.
1481 tcred = __task_cred(task);
1482 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1483 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1484 !capable(CAP_SYS_NICE)) {
1485 rcu_read_unlock();
1486 err = -EPERM;
1487 goto out_put;
1489 rcu_read_unlock();
1491 task_nodes = cpuset_mems_allowed(task);
1492 /* Is the user allowed to access the target nodes? */
1493 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1494 err = -EPERM;
1495 goto out_put;
1498 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1499 err = -EINVAL;
1500 goto out_put;
1503 err = security_task_movememory(task);
1504 if (err)
1505 goto out_put;
1507 mm = get_task_mm(task);
1508 put_task_struct(task);
1510 if (!mm) {
1511 err = -EINVAL;
1512 goto out;
1515 err = do_migrate_pages(mm, old, new,
1516 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1518 mmput(mm);
1519 out:
1520 NODEMASK_SCRATCH_FREE(scratch);
1522 return err;
1524 out_put:
1525 put_task_struct(task);
1526 goto out;
1531 /* Retrieve NUMA policy */
1532 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1533 unsigned long __user *, nmask, unsigned long, maxnode,
1534 unsigned long, addr, unsigned long, flags)
1536 int err;
1537 int uninitialized_var(pval);
1538 nodemask_t nodes;
1540 if (nmask != NULL && maxnode < MAX_NUMNODES)
1541 return -EINVAL;
1543 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1545 if (err)
1546 return err;
1548 if (policy && put_user(pval, policy))
1549 return -EFAULT;
1551 if (nmask)
1552 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1554 return err;
1557 #ifdef CONFIG_COMPAT
1559 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1560 compat_ulong_t __user *nmask,
1561 compat_ulong_t maxnode,
1562 compat_ulong_t addr, compat_ulong_t flags)
1564 long err;
1565 unsigned long __user *nm = NULL;
1566 unsigned long nr_bits, alloc_size;
1567 DECLARE_BITMAP(bm, MAX_NUMNODES);
1569 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1570 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1572 if (nmask)
1573 nm = compat_alloc_user_space(alloc_size);
1575 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1577 if (!err && nmask) {
1578 unsigned long copy_size;
1579 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1580 err = copy_from_user(bm, nm, copy_size);
1581 /* ensure entire bitmap is zeroed */
1582 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1583 err |= compat_put_bitmap(nmask, bm, nr_bits);
1586 return err;
1589 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1590 compat_ulong_t maxnode)
1592 long err = 0;
1593 unsigned long __user *nm = NULL;
1594 unsigned long nr_bits, alloc_size;
1595 DECLARE_BITMAP(bm, MAX_NUMNODES);
1597 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1598 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1600 if (nmask) {
1601 err = compat_get_bitmap(bm, nmask, nr_bits);
1602 nm = compat_alloc_user_space(alloc_size);
1603 err |= copy_to_user(nm, bm, alloc_size);
1606 if (err)
1607 return -EFAULT;
1609 return sys_set_mempolicy(mode, nm, nr_bits+1);
1612 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1613 compat_ulong_t mode, compat_ulong_t __user *nmask,
1614 compat_ulong_t maxnode, compat_ulong_t flags)
1616 long err = 0;
1617 unsigned long __user *nm = NULL;
1618 unsigned long nr_bits, alloc_size;
1619 nodemask_t bm;
1621 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1622 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1624 if (nmask) {
1625 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1626 nm = compat_alloc_user_space(alloc_size);
1627 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1630 if (err)
1631 return -EFAULT;
1633 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1636 #endif
1639 * get_vma_policy(@task, @vma, @addr)
1640 * @task - task for fallback if vma policy == default
1641 * @vma - virtual memory area whose policy is sought
1642 * @addr - address in @vma for shared policy lookup
1644 * Returns effective policy for a VMA at specified address.
1645 * Falls back to @task or system default policy, as necessary.
1646 * Current or other task's task mempolicy and non-shared vma policies must be
1647 * protected by task_lock(task) by the caller.
1648 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1649 * count--added by the get_policy() vm_op, as appropriate--to protect against
1650 * freeing by another task. It is the caller's responsibility to free the
1651 * extra reference for shared policies.
1653 struct mempolicy *get_vma_policy(struct task_struct *task,
1654 struct vm_area_struct *vma, unsigned long addr)
1656 struct mempolicy *pol = get_task_policy(task);
1658 if (vma) {
1659 if (vma->vm_ops && vma->vm_ops->get_policy) {
1660 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1661 addr);
1662 if (vpol)
1663 pol = vpol;
1664 } else if (vma->vm_policy) {
1665 pol = vma->vm_policy;
1668 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1669 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1670 * count on these policies which will be dropped by
1671 * mpol_cond_put() later
1673 if (mpol_needs_cond_ref(pol))
1674 mpol_get(pol);
1677 if (!pol)
1678 pol = &default_policy;
1679 return pol;
1682 bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1684 struct mempolicy *pol = get_task_policy(task);
1685 if (vma) {
1686 if (vma->vm_ops && vma->vm_ops->get_policy) {
1687 bool ret = false;
1689 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1690 if (pol && (pol->flags & MPOL_F_MOF))
1691 ret = true;
1692 mpol_cond_put(pol);
1694 return ret;
1695 } else if (vma->vm_policy) {
1696 pol = vma->vm_policy;
1700 if (!pol)
1701 return default_policy.flags & MPOL_F_MOF;
1703 return pol->flags & MPOL_F_MOF;
1706 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1708 enum zone_type dynamic_policy_zone = policy_zone;
1710 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1713 * if policy->v.nodes has movable memory only,
1714 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1716 * policy->v.nodes is intersect with node_states[N_MEMORY].
1717 * so if the following test faile, it implies
1718 * policy->v.nodes has movable memory only.
1720 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1721 dynamic_policy_zone = ZONE_MOVABLE;
1723 return zone >= dynamic_policy_zone;
1727 * Return a nodemask representing a mempolicy for filtering nodes for
1728 * page allocation
1730 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1732 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1733 if (unlikely(policy->mode == MPOL_BIND) &&
1734 apply_policy_zone(policy, gfp_zone(gfp)) &&
1735 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1736 return &policy->v.nodes;
1738 return NULL;
1741 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1742 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1743 int nd)
1745 switch (policy->mode) {
1746 case MPOL_PREFERRED:
1747 if (!(policy->flags & MPOL_F_LOCAL))
1748 nd = policy->v.preferred_node;
1749 break;
1750 case MPOL_BIND:
1752 * Normally, MPOL_BIND allocations are node-local within the
1753 * allowed nodemask. However, if __GFP_THISNODE is set and the
1754 * current node isn't part of the mask, we use the zonelist for
1755 * the first node in the mask instead.
1757 if (unlikely(gfp & __GFP_THISNODE) &&
1758 unlikely(!node_isset(nd, policy->v.nodes)))
1759 nd = first_node(policy->v.nodes);
1760 break;
1761 default:
1762 BUG();
1764 return node_zonelist(nd, gfp);
1767 /* Do dynamic interleaving for a process */
1768 static unsigned interleave_nodes(struct mempolicy *policy)
1770 unsigned nid, next;
1771 struct task_struct *me = current;
1773 nid = me->il_next;
1774 next = next_node(nid, policy->v.nodes);
1775 if (next >= MAX_NUMNODES)
1776 next = first_node(policy->v.nodes);
1777 if (next < MAX_NUMNODES)
1778 me->il_next = next;
1779 return nid;
1783 * Depending on the memory policy provide a node from which to allocate the
1784 * next slab entry.
1785 * @policy must be protected by freeing by the caller. If @policy is
1786 * the current task's mempolicy, this protection is implicit, as only the
1787 * task can change it's policy. The system default policy requires no
1788 * such protection.
1790 unsigned slab_node(void)
1792 struct mempolicy *policy;
1794 if (in_interrupt())
1795 return numa_node_id();
1797 policy = current->mempolicy;
1798 if (!policy || policy->flags & MPOL_F_LOCAL)
1799 return numa_node_id();
1801 switch (policy->mode) {
1802 case MPOL_PREFERRED:
1804 * handled MPOL_F_LOCAL above
1806 return policy->v.preferred_node;
1808 case MPOL_INTERLEAVE:
1809 return interleave_nodes(policy);
1811 case MPOL_BIND: {
1813 * Follow bind policy behavior and start allocation at the
1814 * first node.
1816 struct zonelist *zonelist;
1817 struct zone *zone;
1818 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1819 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1820 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1821 &policy->v.nodes,
1822 &zone);
1823 return zone ? zone->node : numa_node_id();
1826 default:
1827 BUG();
1831 /* Do static interleaving for a VMA with known offset. */
1832 static unsigned offset_il_node(struct mempolicy *pol,
1833 struct vm_area_struct *vma, unsigned long off)
1835 unsigned nnodes = nodes_weight(pol->v.nodes);
1836 unsigned target;
1837 int c;
1838 int nid = NUMA_NO_NODE;
1840 if (!nnodes)
1841 return numa_node_id();
1842 target = (unsigned int)off % nnodes;
1843 c = 0;
1844 do {
1845 nid = next_node(nid, pol->v.nodes);
1846 c++;
1847 } while (c <= target);
1848 return nid;
1851 /* Determine a node number for interleave */
1852 static inline unsigned interleave_nid(struct mempolicy *pol,
1853 struct vm_area_struct *vma, unsigned long addr, int shift)
1855 if (vma) {
1856 unsigned long off;
1859 * for small pages, there is no difference between
1860 * shift and PAGE_SHIFT, so the bit-shift is safe.
1861 * for huge pages, since vm_pgoff is in units of small
1862 * pages, we need to shift off the always 0 bits to get
1863 * a useful offset.
1865 BUG_ON(shift < PAGE_SHIFT);
1866 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1867 off += (addr - vma->vm_start) >> shift;
1868 return offset_il_node(pol, vma, off);
1869 } else
1870 return interleave_nodes(pol);
1874 * Return the bit number of a random bit set in the nodemask.
1875 * (returns NUMA_NO_NODE if nodemask is empty)
1877 int node_random(const nodemask_t *maskp)
1879 int w, bit = NUMA_NO_NODE;
1881 w = nodes_weight(*maskp);
1882 if (w)
1883 bit = bitmap_ord_to_pos(maskp->bits,
1884 get_random_int() % w, MAX_NUMNODES);
1885 return bit;
1888 #ifdef CONFIG_HUGETLBFS
1890 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1891 * @vma = virtual memory area whose policy is sought
1892 * @addr = address in @vma for shared policy lookup and interleave policy
1893 * @gfp_flags = for requested zone
1894 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1895 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1897 * Returns a zonelist suitable for a huge page allocation and a pointer
1898 * to the struct mempolicy for conditional unref after allocation.
1899 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1900 * @nodemask for filtering the zonelist.
1902 * Must be protected by get_mems_allowed()
1904 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1905 gfp_t gfp_flags, struct mempolicy **mpol,
1906 nodemask_t **nodemask)
1908 struct zonelist *zl;
1910 *mpol = get_vma_policy(current, vma, addr);
1911 *nodemask = NULL; /* assume !MPOL_BIND */
1913 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1914 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1915 huge_page_shift(hstate_vma(vma))), gfp_flags);
1916 } else {
1917 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1918 if ((*mpol)->mode == MPOL_BIND)
1919 *nodemask = &(*mpol)->v.nodes;
1921 return zl;
1925 * init_nodemask_of_mempolicy
1927 * If the current task's mempolicy is "default" [NULL], return 'false'
1928 * to indicate default policy. Otherwise, extract the policy nodemask
1929 * for 'bind' or 'interleave' policy into the argument nodemask, or
1930 * initialize the argument nodemask to contain the single node for
1931 * 'preferred' or 'local' policy and return 'true' to indicate presence
1932 * of non-default mempolicy.
1934 * We don't bother with reference counting the mempolicy [mpol_get/put]
1935 * because the current task is examining it's own mempolicy and a task's
1936 * mempolicy is only ever changed by the task itself.
1938 * N.B., it is the caller's responsibility to free a returned nodemask.
1940 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1942 struct mempolicy *mempolicy;
1943 int nid;
1945 if (!(mask && current->mempolicy))
1946 return false;
1948 task_lock(current);
1949 mempolicy = current->mempolicy;
1950 switch (mempolicy->mode) {
1951 case MPOL_PREFERRED:
1952 if (mempolicy->flags & MPOL_F_LOCAL)
1953 nid = numa_node_id();
1954 else
1955 nid = mempolicy->v.preferred_node;
1956 init_nodemask_of_node(mask, nid);
1957 break;
1959 case MPOL_BIND:
1960 /* Fall through */
1961 case MPOL_INTERLEAVE:
1962 *mask = mempolicy->v.nodes;
1963 break;
1965 default:
1966 BUG();
1968 task_unlock(current);
1970 return true;
1972 #endif
1975 * mempolicy_nodemask_intersects
1977 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1978 * policy. Otherwise, check for intersection between mask and the policy
1979 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1980 * policy, always return true since it may allocate elsewhere on fallback.
1982 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1984 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1985 const nodemask_t *mask)
1987 struct mempolicy *mempolicy;
1988 bool ret = true;
1990 if (!mask)
1991 return ret;
1992 task_lock(tsk);
1993 mempolicy = tsk->mempolicy;
1994 if (!mempolicy)
1995 goto out;
1997 switch (mempolicy->mode) {
1998 case MPOL_PREFERRED:
2000 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2001 * allocate from, they may fallback to other nodes when oom.
2002 * Thus, it's possible for tsk to have allocated memory from
2003 * nodes in mask.
2005 break;
2006 case MPOL_BIND:
2007 case MPOL_INTERLEAVE:
2008 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2009 break;
2010 default:
2011 BUG();
2013 out:
2014 task_unlock(tsk);
2015 return ret;
2018 /* Allocate a page in interleaved policy.
2019 Own path because it needs to do special accounting. */
2020 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2021 unsigned nid)
2023 struct zonelist *zl;
2024 struct page *page;
2026 zl = node_zonelist(nid, gfp);
2027 page = __alloc_pages(gfp, order, zl);
2028 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
2029 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2030 return page;
2034 * alloc_pages_vma - Allocate a page for a VMA.
2036 * @gfp:
2037 * %GFP_USER user allocation.
2038 * %GFP_KERNEL kernel allocations,
2039 * %GFP_HIGHMEM highmem/user allocations,
2040 * %GFP_FS allocation should not call back into a file system.
2041 * %GFP_ATOMIC don't sleep.
2043 * @order:Order of the GFP allocation.
2044 * @vma: Pointer to VMA or NULL if not available.
2045 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2047 * This function allocates a page from the kernel page pool and applies
2048 * a NUMA policy associated with the VMA or the current process.
2049 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2050 * mm_struct of the VMA to prevent it from going away. Should be used for
2051 * all allocations for pages that will be mapped into
2052 * user space. Returns NULL when no page can be allocated.
2054 * Should be called with the mm_sem of the vma hold.
2056 struct page *
2057 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2058 unsigned long addr, int node)
2060 struct mempolicy *pol;
2061 struct page *page;
2062 unsigned int cpuset_mems_cookie;
2064 retry_cpuset:
2065 pol = get_vma_policy(current, vma, addr);
2066 cpuset_mems_cookie = get_mems_allowed();
2068 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2069 unsigned nid;
2071 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2072 mpol_cond_put(pol);
2073 page = alloc_page_interleave(gfp, order, nid);
2074 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2075 goto retry_cpuset;
2077 return page;
2079 page = __alloc_pages_nodemask(gfp, order,
2080 policy_zonelist(gfp, pol, node),
2081 policy_nodemask(gfp, pol));
2082 if (unlikely(mpol_needs_cond_ref(pol)))
2083 __mpol_put(pol);
2084 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2085 goto retry_cpuset;
2086 return page;
2090 * alloc_pages_current - Allocate pages.
2092 * @gfp:
2093 * %GFP_USER user allocation,
2094 * %GFP_KERNEL kernel allocation,
2095 * %GFP_HIGHMEM highmem allocation,
2096 * %GFP_FS don't call back into a file system.
2097 * %GFP_ATOMIC don't sleep.
2098 * @order: Power of two of allocation size in pages. 0 is a single page.
2100 * Allocate a page from the kernel page pool. When not in
2101 * interrupt context and apply the current process NUMA policy.
2102 * Returns NULL when no page can be allocated.
2104 * Don't call cpuset_update_task_memory_state() unless
2105 * 1) it's ok to take cpuset_sem (can WAIT), and
2106 * 2) allocating for current task (not interrupt).
2108 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2110 struct mempolicy *pol = get_task_policy(current);
2111 struct page *page;
2112 unsigned int cpuset_mems_cookie;
2114 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2115 pol = &default_policy;
2117 retry_cpuset:
2118 cpuset_mems_cookie = get_mems_allowed();
2121 * No reference counting needed for current->mempolicy
2122 * nor system default_policy
2124 if (pol->mode == MPOL_INTERLEAVE)
2125 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2126 else
2127 page = __alloc_pages_nodemask(gfp, order,
2128 policy_zonelist(gfp, pol, numa_node_id()),
2129 policy_nodemask(gfp, pol));
2131 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2132 goto retry_cpuset;
2134 return page;
2136 EXPORT_SYMBOL(alloc_pages_current);
2138 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2140 struct mempolicy *pol = mpol_dup(vma_policy(src));
2142 if (IS_ERR(pol))
2143 return PTR_ERR(pol);
2144 dst->vm_policy = pol;
2145 return 0;
2149 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2150 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2151 * with the mems_allowed returned by cpuset_mems_allowed(). This
2152 * keeps mempolicies cpuset relative after its cpuset moves. See
2153 * further kernel/cpuset.c update_nodemask().
2155 * current's mempolicy may be rebinded by the other task(the task that changes
2156 * cpuset's mems), so we needn't do rebind work for current task.
2159 /* Slow path of a mempolicy duplicate */
2160 struct mempolicy *__mpol_dup(struct mempolicy *old)
2162 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2164 if (!new)
2165 return ERR_PTR(-ENOMEM);
2167 /* task's mempolicy is protected by alloc_lock */
2168 if (old == current->mempolicy) {
2169 task_lock(current);
2170 *new = *old;
2171 task_unlock(current);
2172 } else
2173 *new = *old;
2175 rcu_read_lock();
2176 if (current_cpuset_is_being_rebound()) {
2177 nodemask_t mems = cpuset_mems_allowed(current);
2178 if (new->flags & MPOL_F_REBINDING)
2179 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2180 else
2181 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2183 rcu_read_unlock();
2184 atomic_set(&new->refcnt, 1);
2185 return new;
2188 /* Slow path of a mempolicy comparison */
2189 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2191 if (!a || !b)
2192 return false;
2193 if (a->mode != b->mode)
2194 return false;
2195 if (a->flags != b->flags)
2196 return false;
2197 if (mpol_store_user_nodemask(a))
2198 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2199 return false;
2201 switch (a->mode) {
2202 case MPOL_BIND:
2203 /* Fall through */
2204 case MPOL_INTERLEAVE:
2205 return !!nodes_equal(a->v.nodes, b->v.nodes);
2206 case MPOL_PREFERRED:
2207 return a->v.preferred_node == b->v.preferred_node;
2208 default:
2209 BUG();
2210 return false;
2215 * Shared memory backing store policy support.
2217 * Remember policies even when nobody has shared memory mapped.
2218 * The policies are kept in Red-Black tree linked from the inode.
2219 * They are protected by the sp->lock spinlock, which should be held
2220 * for any accesses to the tree.
2223 /* lookup first element intersecting start-end */
2224 /* Caller holds sp->lock */
2225 static struct sp_node *
2226 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2228 struct rb_node *n = sp->root.rb_node;
2230 while (n) {
2231 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2233 if (start >= p->end)
2234 n = n->rb_right;
2235 else if (end <= p->start)
2236 n = n->rb_left;
2237 else
2238 break;
2240 if (!n)
2241 return NULL;
2242 for (;;) {
2243 struct sp_node *w = NULL;
2244 struct rb_node *prev = rb_prev(n);
2245 if (!prev)
2246 break;
2247 w = rb_entry(prev, struct sp_node, nd);
2248 if (w->end <= start)
2249 break;
2250 n = prev;
2252 return rb_entry(n, struct sp_node, nd);
2255 /* Insert a new shared policy into the list. */
2256 /* Caller holds sp->lock */
2257 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2259 struct rb_node **p = &sp->root.rb_node;
2260 struct rb_node *parent = NULL;
2261 struct sp_node *nd;
2263 while (*p) {
2264 parent = *p;
2265 nd = rb_entry(parent, struct sp_node, nd);
2266 if (new->start < nd->start)
2267 p = &(*p)->rb_left;
2268 else if (new->end > nd->end)
2269 p = &(*p)->rb_right;
2270 else
2271 BUG();
2273 rb_link_node(&new->nd, parent, p);
2274 rb_insert_color(&new->nd, &sp->root);
2275 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2276 new->policy ? new->policy->mode : 0);
2279 /* Find shared policy intersecting idx */
2280 struct mempolicy *
2281 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2283 struct mempolicy *pol = NULL;
2284 struct sp_node *sn;
2286 if (!sp->root.rb_node)
2287 return NULL;
2288 spin_lock(&sp->lock);
2289 sn = sp_lookup(sp, idx, idx+1);
2290 if (sn) {
2291 mpol_get(sn->policy);
2292 pol = sn->policy;
2294 spin_unlock(&sp->lock);
2295 return pol;
2298 static void sp_free(struct sp_node *n)
2300 mpol_put(n->policy);
2301 kmem_cache_free(sn_cache, n);
2304 #ifdef CONFIG_NUMA_BALANCING
2305 static bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2307 /* Never defer a private fault */
2308 if (cpupid_match_pid(p, last_cpupid))
2309 return false;
2311 if (p->numa_migrate_deferred) {
2312 p->numa_migrate_deferred--;
2313 return true;
2315 return false;
2318 static inline void defer_numa_migrate(struct task_struct *p)
2320 p->numa_migrate_deferred = sysctl_numa_balancing_migrate_deferred;
2322 #else
2323 static inline bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2325 return false;
2328 static inline void defer_numa_migrate(struct task_struct *p)
2331 #endif /* CONFIG_NUMA_BALANCING */
2334 * mpol_misplaced - check whether current page node is valid in policy
2336 * @page - page to be checked
2337 * @vma - vm area where page mapped
2338 * @addr - virtual address where page mapped
2340 * Lookup current policy node id for vma,addr and "compare to" page's
2341 * node id.
2343 * Returns:
2344 * -1 - not misplaced, page is in the right node
2345 * node - node id where the page should be
2347 * Policy determination "mimics" alloc_page_vma().
2348 * Called from fault path where we know the vma and faulting address.
2350 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2352 struct mempolicy *pol;
2353 struct zone *zone;
2354 int curnid = page_to_nid(page);
2355 unsigned long pgoff;
2356 int thiscpu = raw_smp_processor_id();
2357 int thisnid = cpu_to_node(thiscpu);
2358 int polnid = -1;
2359 int ret = -1;
2361 BUG_ON(!vma);
2363 pol = get_vma_policy(current, vma, addr);
2364 if (!(pol->flags & MPOL_F_MOF))
2365 goto out;
2367 switch (pol->mode) {
2368 case MPOL_INTERLEAVE:
2369 BUG_ON(addr >= vma->vm_end);
2370 BUG_ON(addr < vma->vm_start);
2372 pgoff = vma->vm_pgoff;
2373 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2374 polnid = offset_il_node(pol, vma, pgoff);
2375 break;
2377 case MPOL_PREFERRED:
2378 if (pol->flags & MPOL_F_LOCAL)
2379 polnid = numa_node_id();
2380 else
2381 polnid = pol->v.preferred_node;
2382 break;
2384 case MPOL_BIND:
2386 * allows binding to multiple nodes.
2387 * use current page if in policy nodemask,
2388 * else select nearest allowed node, if any.
2389 * If no allowed nodes, use current [!misplaced].
2391 if (node_isset(curnid, pol->v.nodes))
2392 goto out;
2393 (void)first_zones_zonelist(
2394 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2395 gfp_zone(GFP_HIGHUSER),
2396 &pol->v.nodes, &zone);
2397 polnid = zone->node;
2398 break;
2400 default:
2401 BUG();
2404 /* Migrate the page towards the node whose CPU is referencing it */
2405 if (pol->flags & MPOL_F_MORON) {
2406 int last_cpupid;
2407 int this_cpupid;
2409 polnid = thisnid;
2410 this_cpupid = cpu_pid_to_cpupid(thiscpu, current->pid);
2413 * Multi-stage node selection is used in conjunction
2414 * with a periodic migration fault to build a temporal
2415 * task<->page relation. By using a two-stage filter we
2416 * remove short/unlikely relations.
2418 * Using P(p) ~ n_p / n_t as per frequentist
2419 * probability, we can equate a task's usage of a
2420 * particular page (n_p) per total usage of this
2421 * page (n_t) (in a given time-span) to a probability.
2423 * Our periodic faults will sample this probability and
2424 * getting the same result twice in a row, given these
2425 * samples are fully independent, is then given by
2426 * P(n)^2, provided our sample period is sufficiently
2427 * short compared to the usage pattern.
2429 * This quadric squishes small probabilities, making
2430 * it less likely we act on an unlikely task<->page
2431 * relation.
2433 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
2434 if (!cpupid_pid_unset(last_cpupid) && cpupid_to_nid(last_cpupid) != thisnid) {
2436 /* See sysctl_numa_balancing_migrate_deferred comment */
2437 if (!cpupid_match_pid(current, last_cpupid))
2438 defer_numa_migrate(current);
2440 goto out;
2444 * The quadratic filter above reduces extraneous migration
2445 * of shared pages somewhat. This code reduces it even more,
2446 * reducing the overhead of page migrations of shared pages.
2447 * This makes workloads with shared pages rely more on
2448 * "move task near its memory", and less on "move memory
2449 * towards its task", which is exactly what we want.
2451 if (numa_migrate_deferred(current, last_cpupid))
2452 goto out;
2455 if (curnid != polnid)
2456 ret = polnid;
2457 out:
2458 mpol_cond_put(pol);
2460 return ret;
2463 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2465 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2466 rb_erase(&n->nd, &sp->root);
2467 sp_free(n);
2470 static void sp_node_init(struct sp_node *node, unsigned long start,
2471 unsigned long end, struct mempolicy *pol)
2473 node->start = start;
2474 node->end = end;
2475 node->policy = pol;
2478 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2479 struct mempolicy *pol)
2481 struct sp_node *n;
2482 struct mempolicy *newpol;
2484 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2485 if (!n)
2486 return NULL;
2488 newpol = mpol_dup(pol);
2489 if (IS_ERR(newpol)) {
2490 kmem_cache_free(sn_cache, n);
2491 return NULL;
2493 newpol->flags |= MPOL_F_SHARED;
2494 sp_node_init(n, start, end, newpol);
2496 return n;
2499 /* Replace a policy range. */
2500 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2501 unsigned long end, struct sp_node *new)
2503 struct sp_node *n;
2504 struct sp_node *n_new = NULL;
2505 struct mempolicy *mpol_new = NULL;
2506 int ret = 0;
2508 restart:
2509 spin_lock(&sp->lock);
2510 n = sp_lookup(sp, start, end);
2511 /* Take care of old policies in the same range. */
2512 while (n && n->start < end) {
2513 struct rb_node *next = rb_next(&n->nd);
2514 if (n->start >= start) {
2515 if (n->end <= end)
2516 sp_delete(sp, n);
2517 else
2518 n->start = end;
2519 } else {
2520 /* Old policy spanning whole new range. */
2521 if (n->end > end) {
2522 if (!n_new)
2523 goto alloc_new;
2525 *mpol_new = *n->policy;
2526 atomic_set(&mpol_new->refcnt, 1);
2527 sp_node_init(n_new, end, n->end, mpol_new);
2528 n->end = start;
2529 sp_insert(sp, n_new);
2530 n_new = NULL;
2531 mpol_new = NULL;
2532 break;
2533 } else
2534 n->end = start;
2536 if (!next)
2537 break;
2538 n = rb_entry(next, struct sp_node, nd);
2540 if (new)
2541 sp_insert(sp, new);
2542 spin_unlock(&sp->lock);
2543 ret = 0;
2545 err_out:
2546 if (mpol_new)
2547 mpol_put(mpol_new);
2548 if (n_new)
2549 kmem_cache_free(sn_cache, n_new);
2551 return ret;
2553 alloc_new:
2554 spin_unlock(&sp->lock);
2555 ret = -ENOMEM;
2556 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2557 if (!n_new)
2558 goto err_out;
2559 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2560 if (!mpol_new)
2561 goto err_out;
2562 goto restart;
2566 * mpol_shared_policy_init - initialize shared policy for inode
2567 * @sp: pointer to inode shared policy
2568 * @mpol: struct mempolicy to install
2570 * Install non-NULL @mpol in inode's shared policy rb-tree.
2571 * On entry, the current task has a reference on a non-NULL @mpol.
2572 * This must be released on exit.
2573 * This is called at get_inode() calls and we can use GFP_KERNEL.
2575 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2577 int ret;
2579 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2580 spin_lock_init(&sp->lock);
2582 if (mpol) {
2583 struct vm_area_struct pvma;
2584 struct mempolicy *new;
2585 NODEMASK_SCRATCH(scratch);
2587 if (!scratch)
2588 goto put_mpol;
2589 /* contextualize the tmpfs mount point mempolicy */
2590 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2591 if (IS_ERR(new))
2592 goto free_scratch; /* no valid nodemask intersection */
2594 task_lock(current);
2595 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2596 task_unlock(current);
2597 if (ret)
2598 goto put_new;
2600 /* Create pseudo-vma that contains just the policy */
2601 memset(&pvma, 0, sizeof(struct vm_area_struct));
2602 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2603 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2605 put_new:
2606 mpol_put(new); /* drop initial ref */
2607 free_scratch:
2608 NODEMASK_SCRATCH_FREE(scratch);
2609 put_mpol:
2610 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2614 int mpol_set_shared_policy(struct shared_policy *info,
2615 struct vm_area_struct *vma, struct mempolicy *npol)
2617 int err;
2618 struct sp_node *new = NULL;
2619 unsigned long sz = vma_pages(vma);
2621 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2622 vma->vm_pgoff,
2623 sz, npol ? npol->mode : -1,
2624 npol ? npol->flags : -1,
2625 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2627 if (npol) {
2628 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2629 if (!new)
2630 return -ENOMEM;
2632 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2633 if (err && new)
2634 sp_free(new);
2635 return err;
2638 /* Free a backing policy store on inode delete. */
2639 void mpol_free_shared_policy(struct shared_policy *p)
2641 struct sp_node *n;
2642 struct rb_node *next;
2644 if (!p->root.rb_node)
2645 return;
2646 spin_lock(&p->lock);
2647 next = rb_first(&p->root);
2648 while (next) {
2649 n = rb_entry(next, struct sp_node, nd);
2650 next = rb_next(&n->nd);
2651 sp_delete(p, n);
2653 spin_unlock(&p->lock);
2656 #ifdef CONFIG_NUMA_BALANCING
2657 static int __initdata numabalancing_override;
2659 static void __init check_numabalancing_enable(void)
2661 bool numabalancing_default = false;
2663 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2664 numabalancing_default = true;
2666 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2667 if (numabalancing_override)
2668 set_numabalancing_state(numabalancing_override == 1);
2670 if (nr_node_ids > 1 && !numabalancing_override) {
2671 pr_info("%s automatic NUMA balancing. "
2672 "Configure with numa_balancing= or the "
2673 "kernel.numa_balancing sysctl",
2674 numabalancing_default ? "Enabling" : "Disabling");
2675 set_numabalancing_state(numabalancing_default);
2679 static int __init setup_numabalancing(char *str)
2681 int ret = 0;
2682 if (!str)
2683 goto out;
2685 if (!strcmp(str, "enable")) {
2686 numabalancing_override = 1;
2687 ret = 1;
2688 } else if (!strcmp(str, "disable")) {
2689 numabalancing_override = -1;
2690 ret = 1;
2692 out:
2693 if (!ret)
2694 pr_warn("Unable to parse numa_balancing=\n");
2696 return ret;
2698 __setup("numa_balancing=", setup_numabalancing);
2699 #else
2700 static inline void __init check_numabalancing_enable(void)
2703 #endif /* CONFIG_NUMA_BALANCING */
2705 /* assumes fs == KERNEL_DS */
2706 void __init numa_policy_init(void)
2708 nodemask_t interleave_nodes;
2709 unsigned long largest = 0;
2710 int nid, prefer = 0;
2712 policy_cache = kmem_cache_create("numa_policy",
2713 sizeof(struct mempolicy),
2714 0, SLAB_PANIC, NULL);
2716 sn_cache = kmem_cache_create("shared_policy_node",
2717 sizeof(struct sp_node),
2718 0, SLAB_PANIC, NULL);
2720 for_each_node(nid) {
2721 preferred_node_policy[nid] = (struct mempolicy) {
2722 .refcnt = ATOMIC_INIT(1),
2723 .mode = MPOL_PREFERRED,
2724 .flags = MPOL_F_MOF | MPOL_F_MORON,
2725 .v = { .preferred_node = nid, },
2730 * Set interleaving policy for system init. Interleaving is only
2731 * enabled across suitably sized nodes (default is >= 16MB), or
2732 * fall back to the largest node if they're all smaller.
2734 nodes_clear(interleave_nodes);
2735 for_each_node_state(nid, N_MEMORY) {
2736 unsigned long total_pages = node_present_pages(nid);
2738 /* Preserve the largest node */
2739 if (largest < total_pages) {
2740 largest = total_pages;
2741 prefer = nid;
2744 /* Interleave this node? */
2745 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2746 node_set(nid, interleave_nodes);
2749 /* All too small, use the largest */
2750 if (unlikely(nodes_empty(interleave_nodes)))
2751 node_set(prefer, interleave_nodes);
2753 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2754 printk("numa_policy_init: interleaving failed\n");
2756 check_numabalancing_enable();
2759 /* Reset policy of current process to default */
2760 void numa_default_policy(void)
2762 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2766 * Parse and format mempolicy from/to strings
2770 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2772 static const char * const policy_modes[] =
2774 [MPOL_DEFAULT] = "default",
2775 [MPOL_PREFERRED] = "prefer",
2776 [MPOL_BIND] = "bind",
2777 [MPOL_INTERLEAVE] = "interleave",
2778 [MPOL_LOCAL] = "local",
2782 #ifdef CONFIG_TMPFS
2784 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2785 * @str: string containing mempolicy to parse
2786 * @mpol: pointer to struct mempolicy pointer, returned on success.
2788 * Format of input:
2789 * <mode>[=<flags>][:<nodelist>]
2791 * On success, returns 0, else 1
2793 int mpol_parse_str(char *str, struct mempolicy **mpol)
2795 struct mempolicy *new = NULL;
2796 unsigned short mode;
2797 unsigned short mode_flags;
2798 nodemask_t nodes;
2799 char *nodelist = strchr(str, ':');
2800 char *flags = strchr(str, '=');
2801 int err = 1;
2803 if (nodelist) {
2804 /* NUL-terminate mode or flags string */
2805 *nodelist++ = '\0';
2806 if (nodelist_parse(nodelist, nodes))
2807 goto out;
2808 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2809 goto out;
2810 } else
2811 nodes_clear(nodes);
2813 if (flags)
2814 *flags++ = '\0'; /* terminate mode string */
2816 for (mode = 0; mode < MPOL_MAX; mode++) {
2817 if (!strcmp(str, policy_modes[mode])) {
2818 break;
2821 if (mode >= MPOL_MAX)
2822 goto out;
2824 switch (mode) {
2825 case MPOL_PREFERRED:
2827 * Insist on a nodelist of one node only
2829 if (nodelist) {
2830 char *rest = nodelist;
2831 while (isdigit(*rest))
2832 rest++;
2833 if (*rest)
2834 goto out;
2836 break;
2837 case MPOL_INTERLEAVE:
2839 * Default to online nodes with memory if no nodelist
2841 if (!nodelist)
2842 nodes = node_states[N_MEMORY];
2843 break;
2844 case MPOL_LOCAL:
2846 * Don't allow a nodelist; mpol_new() checks flags
2848 if (nodelist)
2849 goto out;
2850 mode = MPOL_PREFERRED;
2851 break;
2852 case MPOL_DEFAULT:
2854 * Insist on a empty nodelist
2856 if (!nodelist)
2857 err = 0;
2858 goto out;
2859 case MPOL_BIND:
2861 * Insist on a nodelist
2863 if (!nodelist)
2864 goto out;
2867 mode_flags = 0;
2868 if (flags) {
2870 * Currently, we only support two mutually exclusive
2871 * mode flags.
2873 if (!strcmp(flags, "static"))
2874 mode_flags |= MPOL_F_STATIC_NODES;
2875 else if (!strcmp(flags, "relative"))
2876 mode_flags |= MPOL_F_RELATIVE_NODES;
2877 else
2878 goto out;
2881 new = mpol_new(mode, mode_flags, &nodes);
2882 if (IS_ERR(new))
2883 goto out;
2886 * Save nodes for mpol_to_str() to show the tmpfs mount options
2887 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2889 if (mode != MPOL_PREFERRED)
2890 new->v.nodes = nodes;
2891 else if (nodelist)
2892 new->v.preferred_node = first_node(nodes);
2893 else
2894 new->flags |= MPOL_F_LOCAL;
2897 * Save nodes for contextualization: this will be used to "clone"
2898 * the mempolicy in a specific context [cpuset] at a later time.
2900 new->w.user_nodemask = nodes;
2902 err = 0;
2904 out:
2905 /* Restore string for error message */
2906 if (nodelist)
2907 *--nodelist = ':';
2908 if (flags)
2909 *--flags = '=';
2910 if (!err)
2911 *mpol = new;
2912 return err;
2914 #endif /* CONFIG_TMPFS */
2917 * mpol_to_str - format a mempolicy structure for printing
2918 * @buffer: to contain formatted mempolicy string
2919 * @maxlen: length of @buffer
2920 * @pol: pointer to mempolicy to be formatted
2922 * Convert @pol into a string. If @buffer is too short, truncate the string.
2923 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2924 * longest flag, "relative", and to display at least a few node ids.
2926 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2928 char *p = buffer;
2929 nodemask_t nodes = NODE_MASK_NONE;
2930 unsigned short mode = MPOL_DEFAULT;
2931 unsigned short flags = 0;
2933 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2934 mode = pol->mode;
2935 flags = pol->flags;
2938 switch (mode) {
2939 case MPOL_DEFAULT:
2940 break;
2941 case MPOL_PREFERRED:
2942 if (flags & MPOL_F_LOCAL)
2943 mode = MPOL_LOCAL;
2944 else
2945 node_set(pol->v.preferred_node, nodes);
2946 break;
2947 case MPOL_BIND:
2948 case MPOL_INTERLEAVE:
2949 nodes = pol->v.nodes;
2950 break;
2951 default:
2952 WARN_ON_ONCE(1);
2953 snprintf(p, maxlen, "unknown");
2954 return;
2957 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2959 if (flags & MPOL_MODE_FLAGS) {
2960 p += snprintf(p, buffer + maxlen - p, "=");
2963 * Currently, the only defined flags are mutually exclusive
2965 if (flags & MPOL_F_STATIC_NODES)
2966 p += snprintf(p, buffer + maxlen - p, "static");
2967 else if (flags & MPOL_F_RELATIVE_NODES)
2968 p += snprintf(p, buffer + maxlen - p, "relative");
2971 if (!nodes_empty(nodes)) {
2972 p += snprintf(p, buffer + maxlen - p, ":");
2973 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);