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
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
22 * bind Only allocate memory on a specific set of nodes,
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 node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
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.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
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>
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>
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
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 const struct mempolicy_operations {
122 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
124 * If read-side task has no lock to protect task->mempolicy, write-side
125 * task will rebind the task->mempolicy by two step. The first step is
126 * setting all the newly nodes, and the second step is cleaning all the
127 * disallowed nodes. In this way, we can avoid finding no node to alloc
129 * If we have a lock to protect task->mempolicy in read-side, we do
133 * MPOL_REBIND_ONCE - do rebind work at once
134 * MPOL_REBIND_STEP1 - set all the newly nodes
135 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
137 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
138 enum mpol_rebind_step step);
139 } mpol_ops[MPOL_MAX];
141 /* Check that the nodemask contains at least one populated zone */
142 static int is_valid_nodemask(const nodemask_t *nodemask)
146 for_each_node_mask(nd, *nodemask) {
149 for (k = 0; k <= policy_zone; k++) {
150 z = &NODE_DATA(nd)->node_zones[k];
151 if (z->present_pages > 0)
159 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
161 return pol->flags & MPOL_MODE_FLAGS;
164 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
165 const nodemask_t *rel)
168 nodes_fold(tmp, *orig, nodes_weight(*rel));
169 nodes_onto(*ret, tmp, *rel);
172 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
174 if (nodes_empty(*nodes))
176 pol->v.nodes = *nodes;
180 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
183 pol->flags |= MPOL_F_LOCAL; /* local allocation */
184 else if (nodes_empty(*nodes))
185 return -EINVAL; /* no allowed nodes */
187 pol->v.preferred_node = first_node(*nodes);
191 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
193 if (!is_valid_nodemask(nodes))
195 pol->v.nodes = *nodes;
200 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
201 * any, for the new policy. mpol_new() has already validated the nodes
202 * parameter with respect to the policy mode and flags. But, we need to
203 * handle an empty nodemask with MPOL_PREFERRED here.
205 * Must be called holding task's alloc_lock to protect task's mems_allowed
206 * and mempolicy. May also be called holding the mmap_semaphore for write.
208 static int mpol_set_nodemask(struct mempolicy *pol,
209 const nodemask_t *nodes, struct nodemask_scratch *nsc)
213 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
216 /* Check N_HIGH_MEMORY */
217 nodes_and(nsc->mask1,
218 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
221 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
222 nodes = NULL; /* explicit local allocation */
224 if (pol->flags & MPOL_F_RELATIVE_NODES)
225 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
227 nodes_and(nsc->mask2, *nodes, nsc->mask1);
229 if (mpol_store_user_nodemask(pol))
230 pol->w.user_nodemask = *nodes;
232 pol->w.cpuset_mems_allowed =
233 cpuset_current_mems_allowed;
237 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
239 ret = mpol_ops[pol->mode].create(pol, NULL);
244 * This function just creates a new policy, does some check and simple
245 * initialization. You must invoke mpol_set_nodemask() to set nodes.
247 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
250 struct mempolicy *policy;
252 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
253 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
255 if (mode == MPOL_DEFAULT) {
256 if (nodes && !nodes_empty(*nodes))
257 return ERR_PTR(-EINVAL);
263 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
264 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
265 * All other modes require a valid pointer to a non-empty nodemask.
267 if (mode == MPOL_PREFERRED) {
268 if (nodes_empty(*nodes)) {
269 if (((flags & MPOL_F_STATIC_NODES) ||
270 (flags & MPOL_F_RELATIVE_NODES)))
271 return ERR_PTR(-EINVAL);
273 } else if (mode == MPOL_LOCAL) {
274 if (!nodes_empty(*nodes))
275 return ERR_PTR(-EINVAL);
276 mode = MPOL_PREFERRED;
277 } else if (nodes_empty(*nodes))
278 return ERR_PTR(-EINVAL);
279 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
281 return ERR_PTR(-ENOMEM);
282 atomic_set(&policy->refcnt, 1);
284 policy->flags = flags;
289 /* Slow path of a mpol destructor. */
290 void __mpol_put(struct mempolicy *p)
292 if (!atomic_dec_and_test(&p->refcnt))
294 kmem_cache_free(policy_cache, p);
297 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
298 enum mpol_rebind_step step)
304 * MPOL_REBIND_ONCE - do rebind work at once
305 * MPOL_REBIND_STEP1 - set all the newly nodes
306 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
308 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
313 if (pol->flags & MPOL_F_STATIC_NODES)
314 nodes_and(tmp, pol->w.user_nodemask, *nodes);
315 else if (pol->flags & MPOL_F_RELATIVE_NODES)
316 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
319 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
322 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
323 nodes_remap(tmp, pol->v.nodes,
324 pol->w.cpuset_mems_allowed, *nodes);
325 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
326 } else if (step == MPOL_REBIND_STEP2) {
327 tmp = pol->w.cpuset_mems_allowed;
328 pol->w.cpuset_mems_allowed = *nodes;
333 if (nodes_empty(tmp))
336 if (step == MPOL_REBIND_STEP1)
337 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
338 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
343 if (!node_isset(current->il_next, tmp)) {
344 current->il_next = next_node(current->il_next, tmp);
345 if (current->il_next >= MAX_NUMNODES)
346 current->il_next = first_node(tmp);
347 if (current->il_next >= MAX_NUMNODES)
348 current->il_next = numa_node_id();
352 static void mpol_rebind_preferred(struct mempolicy *pol,
353 const nodemask_t *nodes,
354 enum mpol_rebind_step step)
358 if (pol->flags & MPOL_F_STATIC_NODES) {
359 int node = first_node(pol->w.user_nodemask);
361 if (node_isset(node, *nodes)) {
362 pol->v.preferred_node = node;
363 pol->flags &= ~MPOL_F_LOCAL;
365 pol->flags |= MPOL_F_LOCAL;
366 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
367 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
368 pol->v.preferred_node = first_node(tmp);
369 } else if (!(pol->flags & MPOL_F_LOCAL)) {
370 pol->v.preferred_node = node_remap(pol->v.preferred_node,
371 pol->w.cpuset_mems_allowed,
373 pol->w.cpuset_mems_allowed = *nodes;
378 * mpol_rebind_policy - Migrate a policy to a different set of nodes
380 * If read-side task has no lock to protect task->mempolicy, write-side
381 * task will rebind the task->mempolicy by two step. The first step is
382 * setting all the newly nodes, and the second step is cleaning all the
383 * disallowed nodes. In this way, we can avoid finding no node to alloc
385 * If we have a lock to protect task->mempolicy in read-side, we do
389 * MPOL_REBIND_ONCE - do rebind work at once
390 * MPOL_REBIND_STEP1 - set all the newly nodes
391 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
393 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
394 enum mpol_rebind_step step)
398 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
399 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
402 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
405 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
408 if (step == MPOL_REBIND_STEP1)
409 pol->flags |= MPOL_F_REBINDING;
410 else if (step == MPOL_REBIND_STEP2)
411 pol->flags &= ~MPOL_F_REBINDING;
412 else if (step >= MPOL_REBIND_NSTEP)
415 mpol_ops[pol->mode].rebind(pol, newmask, step);
419 * Wrapper for mpol_rebind_policy() that just requires task
420 * pointer, and updates task mempolicy.
422 * Called with task's alloc_lock held.
425 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
426 enum mpol_rebind_step step)
428 mpol_rebind_policy(tsk->mempolicy, new, step);
432 * Rebind each vma in mm to new nodemask.
434 * Call holding a reference to mm. Takes mm->mmap_sem during call.
437 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
439 struct vm_area_struct *vma;
441 down_write(&mm->mmap_sem);
442 for (vma = mm->mmap; vma; vma = vma->vm_next)
443 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
444 up_write(&mm->mmap_sem);
447 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
449 .rebind = mpol_rebind_default,
451 [MPOL_INTERLEAVE] = {
452 .create = mpol_new_interleave,
453 .rebind = mpol_rebind_nodemask,
456 .create = mpol_new_preferred,
457 .rebind = mpol_rebind_preferred,
460 .create = mpol_new_bind,
461 .rebind = mpol_rebind_nodemask,
465 static void migrate_page_add(struct page *page, struct list_head *pagelist,
466 unsigned long flags);
468 /* Scan through pages checking if pages follow certain conditions. */
469 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
470 unsigned long addr, unsigned long end,
471 const nodemask_t *nodes, unsigned long flags,
478 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
483 if (!pte_present(*pte))
485 page = vm_normal_page(vma, addr, *pte);
489 * vm_normal_page() filters out zero pages, but there might
490 * still be PageReserved pages to skip, perhaps in a VDSO.
491 * And we cannot move PageKsm pages sensibly or safely yet.
493 if (PageReserved(page) || PageKsm(page))
495 nid = page_to_nid(page);
496 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
499 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
500 migrate_page_add(page, private, flags);
503 } while (pte++, addr += PAGE_SIZE, addr != end);
504 pte_unmap_unlock(orig_pte, ptl);
508 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
509 unsigned long addr, unsigned long end,
510 const nodemask_t *nodes, unsigned long flags,
516 pmd = pmd_offset(pud, addr);
518 next = pmd_addr_end(addr, end);
519 split_huge_page_pmd(vma->vm_mm, pmd);
520 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
522 if (check_pte_range(vma, pmd, addr, next, nodes,
525 } while (pmd++, addr = next, addr != end);
529 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
530 unsigned long addr, unsigned long end,
531 const nodemask_t *nodes, unsigned long flags,
537 pud = pud_offset(pgd, addr);
539 next = pud_addr_end(addr, end);
540 if (pud_none_or_clear_bad(pud))
542 if (check_pmd_range(vma, pud, addr, next, nodes,
545 } while (pud++, addr = next, addr != end);
549 static inline int check_pgd_range(struct vm_area_struct *vma,
550 unsigned long addr, unsigned long end,
551 const nodemask_t *nodes, unsigned long flags,
557 pgd = pgd_offset(vma->vm_mm, addr);
559 next = pgd_addr_end(addr, end);
560 if (pgd_none_or_clear_bad(pgd))
562 if (check_pud_range(vma, pgd, addr, next, nodes,
565 } while (pgd++, addr = next, addr != end);
569 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
571 * This is used to mark a range of virtual addresses to be inaccessible.
572 * These are later cleared by a NUMA hinting fault. Depending on these
573 * faults, pages may be migrated for better NUMA placement.
575 * This is assuming that NUMA faults are handled using PROT_NONE. If
576 * an architecture makes a different choice, it will need further
577 * changes to the core.
579 unsigned long change_prot_numa(struct vm_area_struct *vma,
580 unsigned long addr, unsigned long end)
583 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
585 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
587 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
592 static unsigned long change_prot_numa(struct vm_area_struct *vma,
593 unsigned long addr, unsigned long end)
597 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
600 * Check if all pages in a range are on a set of nodes.
601 * If pagelist != NULL then isolate pages from the LRU and
602 * put them on the pagelist.
604 static struct vm_area_struct *
605 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
606 const nodemask_t *nodes, unsigned long flags, void *private)
609 struct vm_area_struct *first, *vma, *prev;
612 first = find_vma(mm, start);
614 return ERR_PTR(-EFAULT);
616 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
617 unsigned long endvma = vma->vm_end;
621 if (vma->vm_start > start)
622 start = vma->vm_start;
624 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
625 if (!vma->vm_next && vma->vm_end < end)
626 return ERR_PTR(-EFAULT);
627 if (prev && prev->vm_end < vma->vm_start)
628 return ERR_PTR(-EFAULT);
631 if (is_vm_hugetlb_page(vma))
634 if (flags & MPOL_MF_LAZY) {
635 change_prot_numa(vma, start, endvma);
639 if ((flags & MPOL_MF_STRICT) ||
640 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
641 vma_migratable(vma))) {
643 err = check_pgd_range(vma, start, endvma, nodes,
646 first = ERR_PTR(err);
657 * Apply policy to a single VMA
658 * This must be called with the mmap_sem held for writing.
660 static int vma_replace_policy(struct vm_area_struct *vma,
661 struct mempolicy *pol)
664 struct mempolicy *old;
665 struct mempolicy *new;
667 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
668 vma->vm_start, vma->vm_end, vma->vm_pgoff,
669 vma->vm_ops, vma->vm_file,
670 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
676 if (vma->vm_ops && vma->vm_ops->set_policy) {
677 err = vma->vm_ops->set_policy(vma, new);
682 old = vma->vm_policy;
683 vma->vm_policy = new; /* protected by mmap_sem */
692 /* Step 2: apply policy to a range and do splits. */
693 static int mbind_range(struct mm_struct *mm, unsigned long start,
694 unsigned long end, struct mempolicy *new_pol)
696 struct vm_area_struct *next;
697 struct vm_area_struct *prev;
698 struct vm_area_struct *vma;
701 unsigned long vmstart;
704 vma = find_vma(mm, start);
705 if (!vma || vma->vm_start > start)
709 if (start > vma->vm_start)
712 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
714 vmstart = max(start, vma->vm_start);
715 vmend = min(end, vma->vm_end);
717 if (mpol_equal(vma_policy(vma), new_pol))
720 pgoff = vma->vm_pgoff +
721 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
722 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
723 vma->anon_vma, vma->vm_file, pgoff,
730 if (vma->vm_start != vmstart) {
731 err = split_vma(vma->vm_mm, vma, vmstart, 1);
735 if (vma->vm_end != vmend) {
736 err = split_vma(vma->vm_mm, vma, vmend, 0);
740 err = vma_replace_policy(vma, new_pol);
750 * Update task->flags PF_MEMPOLICY bit: set iff non-default
751 * mempolicy. Allows more rapid checking of this (combined perhaps
752 * with other PF_* flag bits) on memory allocation hot code paths.
754 * If called from outside this file, the task 'p' should -only- be
755 * a newly forked child not yet visible on the task list, because
756 * manipulating the task flags of a visible task is not safe.
758 * The above limitation is why this routine has the funny name
759 * mpol_fix_fork_child_flag().
761 * It is also safe to call this with a task pointer of current,
762 * which the static wrapper mpol_set_task_struct_flag() does,
763 * for use within this file.
766 void mpol_fix_fork_child_flag(struct task_struct *p)
769 p->flags |= PF_MEMPOLICY;
771 p->flags &= ~PF_MEMPOLICY;
774 static void mpol_set_task_struct_flag(void)
776 mpol_fix_fork_child_flag(current);
779 /* Set the process memory policy */
780 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
783 struct mempolicy *new, *old;
784 struct mm_struct *mm = current->mm;
785 NODEMASK_SCRATCH(scratch);
791 new = mpol_new(mode, flags, nodes);
797 * prevent changing our mempolicy while show_numa_maps()
799 * Note: do_set_mempolicy() can be called at init time
803 down_write(&mm->mmap_sem);
805 ret = mpol_set_nodemask(new, nodes, scratch);
807 task_unlock(current);
809 up_write(&mm->mmap_sem);
813 old = current->mempolicy;
814 current->mempolicy = new;
815 mpol_set_task_struct_flag();
816 if (new && new->mode == MPOL_INTERLEAVE &&
817 nodes_weight(new->v.nodes))
818 current->il_next = first_node(new->v.nodes);
819 task_unlock(current);
821 up_write(&mm->mmap_sem);
826 NODEMASK_SCRATCH_FREE(scratch);
831 * Return nodemask for policy for get_mempolicy() query
833 * Called with task's alloc_lock held
835 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
838 if (p == &default_policy)
844 case MPOL_INTERLEAVE:
848 if (!(p->flags & MPOL_F_LOCAL))
849 node_set(p->v.preferred_node, *nodes);
850 /* else return empty node mask for local allocation */
857 static int lookup_node(struct mm_struct *mm, unsigned long addr)
862 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
864 err = page_to_nid(p);
870 /* Retrieve NUMA policy */
871 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
872 unsigned long addr, unsigned long flags)
875 struct mm_struct *mm = current->mm;
876 struct vm_area_struct *vma = NULL;
877 struct mempolicy *pol = current->mempolicy;
880 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
883 if (flags & MPOL_F_MEMS_ALLOWED) {
884 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
886 *policy = 0; /* just so it's initialized */
888 *nmask = cpuset_current_mems_allowed;
889 task_unlock(current);
893 if (flags & MPOL_F_ADDR) {
895 * Do NOT fall back to task policy if the
896 * vma/shared policy at addr is NULL. We
897 * want to return MPOL_DEFAULT in this case.
899 down_read(&mm->mmap_sem);
900 vma = find_vma_intersection(mm, addr, addr+1);
902 up_read(&mm->mmap_sem);
905 if (vma->vm_ops && vma->vm_ops->get_policy)
906 pol = vma->vm_ops->get_policy(vma, addr);
908 pol = vma->vm_policy;
913 pol = &default_policy; /* indicates default behavior */
915 if (flags & MPOL_F_NODE) {
916 if (flags & MPOL_F_ADDR) {
917 err = lookup_node(mm, addr);
921 } else if (pol == current->mempolicy &&
922 pol->mode == MPOL_INTERLEAVE) {
923 *policy = current->il_next;
929 *policy = pol == &default_policy ? MPOL_DEFAULT :
932 * Internal mempolicy flags must be masked off before exposing
933 * the policy to userspace.
935 *policy |= (pol->flags & MPOL_MODE_FLAGS);
939 up_read(¤t->mm->mmap_sem);
945 if (mpol_store_user_nodemask(pol)) {
946 *nmask = pol->w.user_nodemask;
949 get_policy_nodemask(pol, nmask);
950 task_unlock(current);
957 up_read(¤t->mm->mmap_sem);
961 #ifdef CONFIG_MIGRATION
965 static void migrate_page_add(struct page *page, struct list_head *pagelist,
969 * Avoid migrating a page that is shared with others.
971 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
972 if (!isolate_lru_page(page)) {
973 list_add_tail(&page->lru, pagelist);
974 inc_zone_page_state(page, NR_ISOLATED_ANON +
975 page_is_file_cache(page));
980 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
982 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
986 * Migrate pages from one node to a target node.
987 * Returns error or the number of pages not migrated.
989 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
997 node_set(source, nmask);
1000 * This does not "check" the range but isolates all pages that
1001 * need migration. Between passing in the full user address
1002 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1004 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1005 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1006 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1008 if (!list_empty(&pagelist)) {
1009 err = migrate_pages(&pagelist, new_node_page, dest,
1010 false, MIGRATE_SYNC,
1013 putback_lru_pages(&pagelist);
1020 * Move pages between the two nodesets so as to preserve the physical
1021 * layout as much as possible.
1023 * Returns the number of page that could not be moved.
1025 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1026 const nodemask_t *to, int flags)
1032 err = migrate_prep();
1036 down_read(&mm->mmap_sem);
1038 err = migrate_vmas(mm, from, to, flags);
1043 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1044 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1045 * bit in 'tmp', and return that <source, dest> pair for migration.
1046 * The pair of nodemasks 'to' and 'from' define the map.
1048 * If no pair of bits is found that way, fallback to picking some
1049 * pair of 'source' and 'dest' bits that are not the same. If the
1050 * 'source' and 'dest' bits are the same, this represents a node
1051 * that will be migrating to itself, so no pages need move.
1053 * If no bits are left in 'tmp', or if all remaining bits left
1054 * in 'tmp' correspond to the same bit in 'to', return false
1055 * (nothing left to migrate).
1057 * This lets us pick a pair of nodes to migrate between, such that
1058 * if possible the dest node is not already occupied by some other
1059 * source node, minimizing the risk of overloading the memory on a
1060 * node that would happen if we migrated incoming memory to a node
1061 * before migrating outgoing memory source that same node.
1063 * A single scan of tmp is sufficient. As we go, we remember the
1064 * most recent <s, d> pair that moved (s != d). If we find a pair
1065 * that not only moved, but what's better, moved to an empty slot
1066 * (d is not set in tmp), then we break out then, with that pair.
1067 * Otherwise when we finish scanning from_tmp, we at least have the
1068 * most recent <s, d> pair that moved. If we get all the way through
1069 * the scan of tmp without finding any node that moved, much less
1070 * moved to an empty node, then there is nothing left worth migrating.
1074 while (!nodes_empty(tmp)) {
1079 for_each_node_mask(s, tmp) {
1082 * do_migrate_pages() tries to maintain the relative
1083 * node relationship of the pages established between
1084 * threads and memory areas.
1086 * However if the number of source nodes is not equal to
1087 * the number of destination nodes we can not preserve
1088 * this node relative relationship. In that case, skip
1089 * copying memory from a node that is in the destination
1092 * Example: [2,3,4] -> [3,4,5] moves everything.
1093 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1096 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1097 (node_isset(s, *to)))
1100 d = node_remap(s, *from, *to);
1104 source = s; /* Node moved. Memorize */
1107 /* dest not in remaining from nodes? */
1108 if (!node_isset(dest, tmp))
1114 node_clear(source, tmp);
1115 err = migrate_to_node(mm, source, dest, flags);
1122 up_read(&mm->mmap_sem);
1130 * Allocate a new page for page migration based on vma policy.
1131 * Start assuming that page is mapped by vma pointed to by @private.
1132 * Search forward from there, if not. N.B., this assumes that the
1133 * list of pages handed to migrate_pages()--which is how we get here--
1134 * is in virtual address order.
1136 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1138 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1139 unsigned long uninitialized_var(address);
1142 address = page_address_in_vma(page, vma);
1143 if (address != -EFAULT)
1149 * if !vma, alloc_page_vma() will use task or system default policy
1151 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1155 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1156 unsigned long flags)
1160 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1161 const nodemask_t *to, int flags)
1166 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1172 static long do_mbind(unsigned long start, unsigned long len,
1173 unsigned short mode, unsigned short mode_flags,
1174 nodemask_t *nmask, unsigned long flags)
1176 struct vm_area_struct *vma;
1177 struct mm_struct *mm = current->mm;
1178 struct mempolicy *new;
1181 LIST_HEAD(pagelist);
1183 if (flags & ~(unsigned long)MPOL_MF_VALID)
1185 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1188 if (start & ~PAGE_MASK)
1191 if (mode == MPOL_DEFAULT)
1192 flags &= ~MPOL_MF_STRICT;
1194 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1202 new = mpol_new(mode, mode_flags, nmask);
1204 return PTR_ERR(new);
1206 if (flags & MPOL_MF_LAZY)
1207 new->flags |= MPOL_F_MOF;
1210 * If we are using the default policy then operation
1211 * on discontinuous address spaces is okay after all
1214 flags |= MPOL_MF_DISCONTIG_OK;
1216 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1217 start, start + len, mode, mode_flags,
1218 nmask ? nodes_addr(*nmask)[0] : -1);
1220 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1222 err = migrate_prep();
1227 NODEMASK_SCRATCH(scratch);
1229 down_write(&mm->mmap_sem);
1231 err = mpol_set_nodemask(new, nmask, scratch);
1232 task_unlock(current);
1234 up_write(&mm->mmap_sem);
1237 NODEMASK_SCRATCH_FREE(scratch);
1242 vma = check_range(mm, start, end, nmask,
1243 flags | MPOL_MF_INVERT, &pagelist);
1245 err = PTR_ERR(vma); /* maybe ... */
1247 err = mbind_range(mm, start, end, new);
1252 if (!list_empty(&pagelist)) {
1253 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1254 nr_failed = migrate_pages(&pagelist, new_vma_page,
1256 false, MIGRATE_SYNC,
1257 MR_MEMPOLICY_MBIND);
1259 putback_lru_pages(&pagelist);
1262 if (nr_failed && (flags & MPOL_MF_STRICT))
1265 putback_lru_pages(&pagelist);
1267 up_write(&mm->mmap_sem);
1274 * User space interface with variable sized bitmaps for nodelists.
1277 /* Copy a node mask from user space. */
1278 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1279 unsigned long maxnode)
1282 unsigned long nlongs;
1283 unsigned long endmask;
1286 nodes_clear(*nodes);
1287 if (maxnode == 0 || !nmask)
1289 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1292 nlongs = BITS_TO_LONGS(maxnode);
1293 if ((maxnode % BITS_PER_LONG) == 0)
1296 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1298 /* When the user specified more nodes than supported just check
1299 if the non supported part is all zero. */
1300 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1301 if (nlongs > PAGE_SIZE/sizeof(long))
1303 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1305 if (get_user(t, nmask + k))
1307 if (k == nlongs - 1) {
1313 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1317 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1319 nodes_addr(*nodes)[nlongs-1] &= endmask;
1323 /* Copy a kernel node mask to user space */
1324 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1327 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1328 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1330 if (copy > nbytes) {
1331 if (copy > PAGE_SIZE)
1333 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1337 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1340 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1341 unsigned long, mode, unsigned long __user *, nmask,
1342 unsigned long, maxnode, unsigned, flags)
1346 unsigned short mode_flags;
1348 mode_flags = mode & MPOL_MODE_FLAGS;
1349 mode &= ~MPOL_MODE_FLAGS;
1350 if (mode >= MPOL_MAX)
1352 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1353 (mode_flags & MPOL_F_RELATIVE_NODES))
1355 err = get_nodes(&nodes, nmask, maxnode);
1358 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1361 /* Set the process memory policy */
1362 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1363 unsigned long, maxnode)
1367 unsigned short flags;
1369 flags = mode & MPOL_MODE_FLAGS;
1370 mode &= ~MPOL_MODE_FLAGS;
1371 if ((unsigned int)mode >= MPOL_MAX)
1373 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1375 err = get_nodes(&nodes, nmask, maxnode);
1378 return do_set_mempolicy(mode, flags, &nodes);
1381 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1382 const unsigned long __user *, old_nodes,
1383 const unsigned long __user *, new_nodes)
1385 const struct cred *cred = current_cred(), *tcred;
1386 struct mm_struct *mm = NULL;
1387 struct task_struct *task;
1388 nodemask_t task_nodes;
1392 NODEMASK_SCRATCH(scratch);
1397 old = &scratch->mask1;
1398 new = &scratch->mask2;
1400 err = get_nodes(old, old_nodes, maxnode);
1404 err = get_nodes(new, new_nodes, maxnode);
1408 /* Find the mm_struct */
1410 task = pid ? find_task_by_vpid(pid) : current;
1416 get_task_struct(task);
1421 * Check if this process has the right to modify the specified
1422 * process. The right exists if the process has administrative
1423 * capabilities, superuser privileges or the same
1424 * userid as the target process.
1426 tcred = __task_cred(task);
1427 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1428 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1429 !capable(CAP_SYS_NICE)) {
1436 task_nodes = cpuset_mems_allowed(task);
1437 /* Is the user allowed to access the target nodes? */
1438 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1443 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1448 err = security_task_movememory(task);
1452 mm = get_task_mm(task);
1453 put_task_struct(task);
1460 err = do_migrate_pages(mm, old, new,
1461 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1465 NODEMASK_SCRATCH_FREE(scratch);
1470 put_task_struct(task);
1476 /* Retrieve NUMA policy */
1477 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1478 unsigned long __user *, nmask, unsigned long, maxnode,
1479 unsigned long, addr, unsigned long, flags)
1482 int uninitialized_var(pval);
1485 if (nmask != NULL && maxnode < MAX_NUMNODES)
1488 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1493 if (policy && put_user(pval, policy))
1497 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1502 #ifdef CONFIG_COMPAT
1504 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1505 compat_ulong_t __user *nmask,
1506 compat_ulong_t maxnode,
1507 compat_ulong_t addr, compat_ulong_t flags)
1510 unsigned long __user *nm = NULL;
1511 unsigned long nr_bits, alloc_size;
1512 DECLARE_BITMAP(bm, MAX_NUMNODES);
1514 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1515 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1518 nm = compat_alloc_user_space(alloc_size);
1520 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1522 if (!err && nmask) {
1523 unsigned long copy_size;
1524 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1525 err = copy_from_user(bm, nm, copy_size);
1526 /* ensure entire bitmap is zeroed */
1527 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1528 err |= compat_put_bitmap(nmask, bm, nr_bits);
1534 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1535 compat_ulong_t maxnode)
1538 unsigned long __user *nm = NULL;
1539 unsigned long nr_bits, alloc_size;
1540 DECLARE_BITMAP(bm, MAX_NUMNODES);
1542 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1543 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1546 err = compat_get_bitmap(bm, nmask, nr_bits);
1547 nm = compat_alloc_user_space(alloc_size);
1548 err |= copy_to_user(nm, bm, alloc_size);
1554 return sys_set_mempolicy(mode, nm, nr_bits+1);
1557 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1558 compat_ulong_t mode, compat_ulong_t __user *nmask,
1559 compat_ulong_t maxnode, compat_ulong_t flags)
1562 unsigned long __user *nm = NULL;
1563 unsigned long nr_bits, alloc_size;
1566 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1567 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1570 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1571 nm = compat_alloc_user_space(alloc_size);
1572 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1578 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1584 * get_vma_policy(@task, @vma, @addr)
1585 * @task - task for fallback if vma policy == default
1586 * @vma - virtual memory area whose policy is sought
1587 * @addr - address in @vma for shared policy lookup
1589 * Returns effective policy for a VMA at specified address.
1590 * Falls back to @task or system default policy, as necessary.
1591 * Current or other task's task mempolicy and non-shared vma policies must be
1592 * protected by task_lock(task) by the caller.
1593 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1594 * count--added by the get_policy() vm_op, as appropriate--to protect against
1595 * freeing by another task. It is the caller's responsibility to free the
1596 * extra reference for shared policies.
1598 struct mempolicy *get_vma_policy(struct task_struct *task,
1599 struct vm_area_struct *vma, unsigned long addr)
1601 struct mempolicy *pol = task->mempolicy;
1604 if (vma->vm_ops && vma->vm_ops->get_policy) {
1605 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1609 } else if (vma->vm_policy) {
1610 pol = vma->vm_policy;
1613 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1614 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1615 * count on these policies which will be dropped by
1616 * mpol_cond_put() later
1618 if (mpol_needs_cond_ref(pol))
1623 pol = &default_policy;
1628 * Return a nodemask representing a mempolicy for filtering nodes for
1631 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1633 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1634 if (unlikely(policy->mode == MPOL_BIND) &&
1635 gfp_zone(gfp) >= policy_zone &&
1636 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1637 return &policy->v.nodes;
1642 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1643 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1646 switch (policy->mode) {
1647 case MPOL_PREFERRED:
1648 if (!(policy->flags & MPOL_F_LOCAL))
1649 nd = policy->v.preferred_node;
1653 * Normally, MPOL_BIND allocations are node-local within the
1654 * allowed nodemask. However, if __GFP_THISNODE is set and the
1655 * current node isn't part of the mask, we use the zonelist for
1656 * the first node in the mask instead.
1658 if (unlikely(gfp & __GFP_THISNODE) &&
1659 unlikely(!node_isset(nd, policy->v.nodes)))
1660 nd = first_node(policy->v.nodes);
1665 return node_zonelist(nd, gfp);
1668 /* Do dynamic interleaving for a process */
1669 static unsigned interleave_nodes(struct mempolicy *policy)
1672 struct task_struct *me = current;
1675 next = next_node(nid, policy->v.nodes);
1676 if (next >= MAX_NUMNODES)
1677 next = first_node(policy->v.nodes);
1678 if (next < MAX_NUMNODES)
1684 * Depending on the memory policy provide a node from which to allocate the
1686 * @policy must be protected by freeing by the caller. If @policy is
1687 * the current task's mempolicy, this protection is implicit, as only the
1688 * task can change it's policy. The system default policy requires no
1691 unsigned slab_node(void)
1693 struct mempolicy *policy;
1696 return numa_node_id();
1698 policy = current->mempolicy;
1699 if (!policy || policy->flags & MPOL_F_LOCAL)
1700 return numa_node_id();
1702 switch (policy->mode) {
1703 case MPOL_PREFERRED:
1705 * handled MPOL_F_LOCAL above
1707 return policy->v.preferred_node;
1709 case MPOL_INTERLEAVE:
1710 return interleave_nodes(policy);
1714 * Follow bind policy behavior and start allocation at the
1717 struct zonelist *zonelist;
1719 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1720 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1721 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1724 return zone ? zone->node : numa_node_id();
1732 /* Do static interleaving for a VMA with known offset. */
1733 static unsigned offset_il_node(struct mempolicy *pol,
1734 struct vm_area_struct *vma, unsigned long off)
1736 unsigned nnodes = nodes_weight(pol->v.nodes);
1742 return numa_node_id();
1743 target = (unsigned int)off % nnodes;
1746 nid = next_node(nid, pol->v.nodes);
1748 } while (c <= target);
1752 /* Determine a node number for interleave */
1753 static inline unsigned interleave_nid(struct mempolicy *pol,
1754 struct vm_area_struct *vma, unsigned long addr, int shift)
1760 * for small pages, there is no difference between
1761 * shift and PAGE_SHIFT, so the bit-shift is safe.
1762 * for huge pages, since vm_pgoff is in units of small
1763 * pages, we need to shift off the always 0 bits to get
1766 BUG_ON(shift < PAGE_SHIFT);
1767 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1768 off += (addr - vma->vm_start) >> shift;
1769 return offset_il_node(pol, vma, off);
1771 return interleave_nodes(pol);
1775 * Return the bit number of a random bit set in the nodemask.
1776 * (returns -1 if nodemask is empty)
1778 int node_random(const nodemask_t *maskp)
1782 w = nodes_weight(*maskp);
1784 bit = bitmap_ord_to_pos(maskp->bits,
1785 get_random_int() % w, MAX_NUMNODES);
1789 #ifdef CONFIG_HUGETLBFS
1791 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1792 * @vma = virtual memory area whose policy is sought
1793 * @addr = address in @vma for shared policy lookup and interleave policy
1794 * @gfp_flags = for requested zone
1795 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1796 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1798 * Returns a zonelist suitable for a huge page allocation and a pointer
1799 * to the struct mempolicy for conditional unref after allocation.
1800 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1801 * @nodemask for filtering the zonelist.
1803 * Must be protected by get_mems_allowed()
1805 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1806 gfp_t gfp_flags, struct mempolicy **mpol,
1807 nodemask_t **nodemask)
1809 struct zonelist *zl;
1811 *mpol = get_vma_policy(current, vma, addr);
1812 *nodemask = NULL; /* assume !MPOL_BIND */
1814 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1815 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1816 huge_page_shift(hstate_vma(vma))), gfp_flags);
1818 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1819 if ((*mpol)->mode == MPOL_BIND)
1820 *nodemask = &(*mpol)->v.nodes;
1826 * init_nodemask_of_mempolicy
1828 * If the current task's mempolicy is "default" [NULL], return 'false'
1829 * to indicate default policy. Otherwise, extract the policy nodemask
1830 * for 'bind' or 'interleave' policy into the argument nodemask, or
1831 * initialize the argument nodemask to contain the single node for
1832 * 'preferred' or 'local' policy and return 'true' to indicate presence
1833 * of non-default mempolicy.
1835 * We don't bother with reference counting the mempolicy [mpol_get/put]
1836 * because the current task is examining it's own mempolicy and a task's
1837 * mempolicy is only ever changed by the task itself.
1839 * N.B., it is the caller's responsibility to free a returned nodemask.
1841 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1843 struct mempolicy *mempolicy;
1846 if (!(mask && current->mempolicy))
1850 mempolicy = current->mempolicy;
1851 switch (mempolicy->mode) {
1852 case MPOL_PREFERRED:
1853 if (mempolicy->flags & MPOL_F_LOCAL)
1854 nid = numa_node_id();
1856 nid = mempolicy->v.preferred_node;
1857 init_nodemask_of_node(mask, nid);
1862 case MPOL_INTERLEAVE:
1863 *mask = mempolicy->v.nodes;
1869 task_unlock(current);
1876 * mempolicy_nodemask_intersects
1878 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1879 * policy. Otherwise, check for intersection between mask and the policy
1880 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1881 * policy, always return true since it may allocate elsewhere on fallback.
1883 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1885 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1886 const nodemask_t *mask)
1888 struct mempolicy *mempolicy;
1894 mempolicy = tsk->mempolicy;
1898 switch (mempolicy->mode) {
1899 case MPOL_PREFERRED:
1901 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1902 * allocate from, they may fallback to other nodes when oom.
1903 * Thus, it's possible for tsk to have allocated memory from
1908 case MPOL_INTERLEAVE:
1909 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1919 /* Allocate a page in interleaved policy.
1920 Own path because it needs to do special accounting. */
1921 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1924 struct zonelist *zl;
1927 zl = node_zonelist(nid, gfp);
1928 page = __alloc_pages(gfp, order, zl);
1929 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1930 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1935 * alloc_pages_vma - Allocate a page for a VMA.
1938 * %GFP_USER user allocation.
1939 * %GFP_KERNEL kernel allocations,
1940 * %GFP_HIGHMEM highmem/user allocations,
1941 * %GFP_FS allocation should not call back into a file system.
1942 * %GFP_ATOMIC don't sleep.
1944 * @order:Order of the GFP allocation.
1945 * @vma: Pointer to VMA or NULL if not available.
1946 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1948 * This function allocates a page from the kernel page pool and applies
1949 * a NUMA policy associated with the VMA or the current process.
1950 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1951 * mm_struct of the VMA to prevent it from going away. Should be used for
1952 * all allocations for pages that will be mapped into
1953 * user space. Returns NULL when no page can be allocated.
1955 * Should be called with the mm_sem of the vma hold.
1958 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1959 unsigned long addr, int node)
1961 struct mempolicy *pol;
1962 struct zonelist *zl;
1964 unsigned int cpuset_mems_cookie;
1967 pol = get_vma_policy(current, vma, addr);
1968 cpuset_mems_cookie = get_mems_allowed();
1970 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1973 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1975 page = alloc_page_interleave(gfp, order, nid);
1976 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1981 zl = policy_zonelist(gfp, pol, node);
1982 if (unlikely(mpol_needs_cond_ref(pol))) {
1984 * slow path: ref counted shared policy
1986 struct page *page = __alloc_pages_nodemask(gfp, order,
1987 zl, policy_nodemask(gfp, pol));
1989 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1994 * fast path: default or task policy
1996 page = __alloc_pages_nodemask(gfp, order, zl,
1997 policy_nodemask(gfp, pol));
1998 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2004 * alloc_pages_current - Allocate pages.
2007 * %GFP_USER user allocation,
2008 * %GFP_KERNEL kernel allocation,
2009 * %GFP_HIGHMEM highmem allocation,
2010 * %GFP_FS don't call back into a file system.
2011 * %GFP_ATOMIC don't sleep.
2012 * @order: Power of two of allocation size in pages. 0 is a single page.
2014 * Allocate a page from the kernel page pool. When not in
2015 * interrupt context and apply the current process NUMA policy.
2016 * Returns NULL when no page can be allocated.
2018 * Don't call cpuset_update_task_memory_state() unless
2019 * 1) it's ok to take cpuset_sem (can WAIT), and
2020 * 2) allocating for current task (not interrupt).
2022 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2024 struct mempolicy *pol = current->mempolicy;
2026 unsigned int cpuset_mems_cookie;
2028 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2029 pol = &default_policy;
2032 cpuset_mems_cookie = get_mems_allowed();
2035 * No reference counting needed for current->mempolicy
2036 * nor system default_policy
2038 if (pol->mode == MPOL_INTERLEAVE)
2039 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2041 page = __alloc_pages_nodemask(gfp, order,
2042 policy_zonelist(gfp, pol, numa_node_id()),
2043 policy_nodemask(gfp, pol));
2045 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2050 EXPORT_SYMBOL(alloc_pages_current);
2053 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2054 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2055 * with the mems_allowed returned by cpuset_mems_allowed(). This
2056 * keeps mempolicies cpuset relative after its cpuset moves. See
2057 * further kernel/cpuset.c update_nodemask().
2059 * current's mempolicy may be rebinded by the other task(the task that changes
2060 * cpuset's mems), so we needn't do rebind work for current task.
2063 /* Slow path of a mempolicy duplicate */
2064 struct mempolicy *__mpol_dup(struct mempolicy *old)
2066 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2069 return ERR_PTR(-ENOMEM);
2071 /* task's mempolicy is protected by alloc_lock */
2072 if (old == current->mempolicy) {
2075 task_unlock(current);
2080 if (current_cpuset_is_being_rebound()) {
2081 nodemask_t mems = cpuset_mems_allowed(current);
2082 if (new->flags & MPOL_F_REBINDING)
2083 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2085 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2088 atomic_set(&new->refcnt, 1);
2093 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2094 * eliminate the * MPOL_F_* flags that require conditional ref and
2095 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2096 * after return. Use the returned value.
2098 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2099 * policy lookup, even if the policy needs/has extra ref on lookup.
2100 * shmem_readahead needs this.
2102 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2103 struct mempolicy *frompol)
2105 if (!mpol_needs_cond_ref(frompol))
2109 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2110 __mpol_put(frompol);
2114 /* Slow path of a mempolicy comparison */
2115 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2119 if (a->mode != b->mode)
2121 if (a->flags != b->flags)
2123 if (mpol_store_user_nodemask(a))
2124 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2130 case MPOL_INTERLEAVE:
2131 return !!nodes_equal(a->v.nodes, b->v.nodes);
2132 case MPOL_PREFERRED:
2133 return a->v.preferred_node == b->v.preferred_node;
2141 * Shared memory backing store policy support.
2143 * Remember policies even when nobody has shared memory mapped.
2144 * The policies are kept in Red-Black tree linked from the inode.
2145 * They are protected by the sp->lock spinlock, which should be held
2146 * for any accesses to the tree.
2149 /* lookup first element intersecting start-end */
2150 /* Caller holds sp->mutex */
2151 static struct sp_node *
2152 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2154 struct rb_node *n = sp->root.rb_node;
2157 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2159 if (start >= p->end)
2161 else if (end <= p->start)
2169 struct sp_node *w = NULL;
2170 struct rb_node *prev = rb_prev(n);
2173 w = rb_entry(prev, struct sp_node, nd);
2174 if (w->end <= start)
2178 return rb_entry(n, struct sp_node, nd);
2181 /* Insert a new shared policy into the list. */
2182 /* Caller holds sp->lock */
2183 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2185 struct rb_node **p = &sp->root.rb_node;
2186 struct rb_node *parent = NULL;
2191 nd = rb_entry(parent, struct sp_node, nd);
2192 if (new->start < nd->start)
2194 else if (new->end > nd->end)
2195 p = &(*p)->rb_right;
2199 rb_link_node(&new->nd, parent, p);
2200 rb_insert_color(&new->nd, &sp->root);
2201 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2202 new->policy ? new->policy->mode : 0);
2205 /* Find shared policy intersecting idx */
2207 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2209 struct mempolicy *pol = NULL;
2212 if (!sp->root.rb_node)
2214 mutex_lock(&sp->mutex);
2215 sn = sp_lookup(sp, idx, idx+1);
2217 mpol_get(sn->policy);
2220 mutex_unlock(&sp->mutex);
2224 static void sp_free(struct sp_node *n)
2226 mpol_put(n->policy);
2227 kmem_cache_free(sn_cache, n);
2231 * mpol_misplaced - check whether current page node is valid in policy
2233 * @page - page to be checked
2234 * @vma - vm area where page mapped
2235 * @addr - virtual address where page mapped
2237 * Lookup current policy node id for vma,addr and "compare to" page's
2241 * -1 - not misplaced, page is in the right node
2242 * node - node id where the page should be
2244 * Policy determination "mimics" alloc_page_vma().
2245 * Called from fault path where we know the vma and faulting address.
2247 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2249 struct mempolicy *pol;
2251 int curnid = page_to_nid(page);
2252 unsigned long pgoff;
2258 pol = get_vma_policy(current, vma, addr);
2259 if (!(pol->flags & MPOL_F_MOF))
2262 switch (pol->mode) {
2263 case MPOL_INTERLEAVE:
2264 BUG_ON(addr >= vma->vm_end);
2265 BUG_ON(addr < vma->vm_start);
2267 pgoff = vma->vm_pgoff;
2268 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2269 polnid = offset_il_node(pol, vma, pgoff);
2272 case MPOL_PREFERRED:
2273 if (pol->flags & MPOL_F_LOCAL)
2274 polnid = numa_node_id();
2276 polnid = pol->v.preferred_node;
2281 * allows binding to multiple nodes.
2282 * use current page if in policy nodemask,
2283 * else select nearest allowed node, if any.
2284 * If no allowed nodes, use current [!misplaced].
2286 if (node_isset(curnid, pol->v.nodes))
2288 (void)first_zones_zonelist(
2289 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2290 gfp_zone(GFP_HIGHUSER),
2291 &pol->v.nodes, &zone);
2292 polnid = zone->node;
2298 if (curnid != polnid)
2306 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2308 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2309 rb_erase(&n->nd, &sp->root);
2313 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2314 struct mempolicy *pol)
2317 struct mempolicy *newpol;
2319 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2323 newpol = mpol_dup(pol);
2324 if (IS_ERR(newpol)) {
2325 kmem_cache_free(sn_cache, n);
2328 newpol->flags |= MPOL_F_SHARED;
2337 /* Replace a policy range. */
2338 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2339 unsigned long end, struct sp_node *new)
2344 mutex_lock(&sp->mutex);
2345 n = sp_lookup(sp, start, end);
2346 /* Take care of old policies in the same range. */
2347 while (n && n->start < end) {
2348 struct rb_node *next = rb_next(&n->nd);
2349 if (n->start >= start) {
2355 /* Old policy spanning whole new range. */
2357 struct sp_node *new2;
2358 new2 = sp_alloc(end, n->end, n->policy);
2364 sp_insert(sp, new2);
2371 n = rb_entry(next, struct sp_node, nd);
2376 mutex_unlock(&sp->mutex);
2381 * mpol_shared_policy_init - initialize shared policy for inode
2382 * @sp: pointer to inode shared policy
2383 * @mpol: struct mempolicy to install
2385 * Install non-NULL @mpol in inode's shared policy rb-tree.
2386 * On entry, the current task has a reference on a non-NULL @mpol.
2387 * This must be released on exit.
2388 * This is called at get_inode() calls and we can use GFP_KERNEL.
2390 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2394 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2395 mutex_init(&sp->mutex);
2398 struct vm_area_struct pvma;
2399 struct mempolicy *new;
2400 NODEMASK_SCRATCH(scratch);
2404 /* contextualize the tmpfs mount point mempolicy */
2405 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2407 goto free_scratch; /* no valid nodemask intersection */
2410 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2411 task_unlock(current);
2415 /* Create pseudo-vma that contains just the policy */
2416 memset(&pvma, 0, sizeof(struct vm_area_struct));
2417 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2418 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2421 mpol_put(new); /* drop initial ref */
2423 NODEMASK_SCRATCH_FREE(scratch);
2425 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2429 int mpol_set_shared_policy(struct shared_policy *info,
2430 struct vm_area_struct *vma, struct mempolicy *npol)
2433 struct sp_node *new = NULL;
2434 unsigned long sz = vma_pages(vma);
2436 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2438 sz, npol ? npol->mode : -1,
2439 npol ? npol->flags : -1,
2440 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2443 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2447 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2453 /* Free a backing policy store on inode delete. */
2454 void mpol_free_shared_policy(struct shared_policy *p)
2457 struct rb_node *next;
2459 if (!p->root.rb_node)
2461 mutex_lock(&p->mutex);
2462 next = rb_first(&p->root);
2464 n = rb_entry(next, struct sp_node, nd);
2465 next = rb_next(&n->nd);
2468 mutex_unlock(&p->mutex);
2471 /* assumes fs == KERNEL_DS */
2472 void __init numa_policy_init(void)
2474 nodemask_t interleave_nodes;
2475 unsigned long largest = 0;
2476 int nid, prefer = 0;
2478 policy_cache = kmem_cache_create("numa_policy",
2479 sizeof(struct mempolicy),
2480 0, SLAB_PANIC, NULL);
2482 sn_cache = kmem_cache_create("shared_policy_node",
2483 sizeof(struct sp_node),
2484 0, SLAB_PANIC, NULL);
2487 * Set interleaving policy for system init. Interleaving is only
2488 * enabled across suitably sized nodes (default is >= 16MB), or
2489 * fall back to the largest node if they're all smaller.
2491 nodes_clear(interleave_nodes);
2492 for_each_node_state(nid, N_HIGH_MEMORY) {
2493 unsigned long total_pages = node_present_pages(nid);
2495 /* Preserve the largest node */
2496 if (largest < total_pages) {
2497 largest = total_pages;
2501 /* Interleave this node? */
2502 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2503 node_set(nid, interleave_nodes);
2506 /* All too small, use the largest */
2507 if (unlikely(nodes_empty(interleave_nodes)))
2508 node_set(prefer, interleave_nodes);
2510 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2511 printk("numa_policy_init: interleaving failed\n");
2514 /* Reset policy of current process to default */
2515 void numa_default_policy(void)
2517 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2521 * Parse and format mempolicy from/to strings
2525 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2526 * Used only for mpol_parse_str() and mpol_to_str()
2528 static const char * const policy_modes[] =
2530 [MPOL_DEFAULT] = "default",
2531 [MPOL_PREFERRED] = "prefer",
2532 [MPOL_BIND] = "bind",
2533 [MPOL_INTERLEAVE] = "interleave",
2534 [MPOL_LOCAL] = "local",
2540 * mpol_parse_str - parse string to mempolicy
2541 * @str: string containing mempolicy to parse
2542 * @mpol: pointer to struct mempolicy pointer, returned on success.
2543 * @no_context: flag whether to "contextualize" the mempolicy
2546 * <mode>[=<flags>][:<nodelist>]
2548 * if @no_context is true, save the input nodemask in w.user_nodemask in
2549 * the returned mempolicy. This will be used to "clone" the mempolicy in
2550 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2551 * mount option. Note that if 'static' or 'relative' mode flags were
2552 * specified, the input nodemask will already have been saved. Saving
2553 * it again is redundant, but safe.
2555 * On success, returns 0, else 1
2557 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2559 struct mempolicy *new = NULL;
2560 unsigned short mode;
2561 unsigned short uninitialized_var(mode_flags);
2563 char *nodelist = strchr(str, ':');
2564 char *flags = strchr(str, '=');
2568 /* NUL-terminate mode or flags string */
2570 if (nodelist_parse(nodelist, nodes))
2572 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2578 *flags++ = '\0'; /* terminate mode string */
2580 for (mode = 0; mode < MPOL_MAX; mode++) {
2581 if (!strcmp(str, policy_modes[mode])) {
2585 if (mode >= MPOL_MAX)
2589 case MPOL_PREFERRED:
2591 * Insist on a nodelist of one node only
2594 char *rest = nodelist;
2595 while (isdigit(*rest))
2601 case MPOL_INTERLEAVE:
2603 * Default to online nodes with memory if no nodelist
2606 nodes = node_states[N_HIGH_MEMORY];
2610 * Don't allow a nodelist; mpol_new() checks flags
2614 mode = MPOL_PREFERRED;
2618 * Insist on a empty nodelist
2625 * Insist on a nodelist
2634 * Currently, we only support two mutually exclusive
2637 if (!strcmp(flags, "static"))
2638 mode_flags |= MPOL_F_STATIC_NODES;
2639 else if (!strcmp(flags, "relative"))
2640 mode_flags |= MPOL_F_RELATIVE_NODES;
2645 new = mpol_new(mode, mode_flags, &nodes);
2650 /* save for contextualization */
2651 new->w.user_nodemask = nodes;
2654 NODEMASK_SCRATCH(scratch);
2657 ret = mpol_set_nodemask(new, &nodes, scratch);
2658 task_unlock(current);
2661 NODEMASK_SCRATCH_FREE(scratch);
2670 /* Restore string for error message */
2679 #endif /* CONFIG_TMPFS */
2682 * mpol_to_str - format a mempolicy structure for printing
2683 * @buffer: to contain formatted mempolicy string
2684 * @maxlen: length of @buffer
2685 * @pol: pointer to mempolicy to be formatted
2686 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2688 * Convert a mempolicy into a string.
2689 * Returns the number of characters in buffer (if positive)
2690 * or an error (negative)
2692 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2697 unsigned short mode;
2698 unsigned short flags = pol ? pol->flags : 0;
2701 * Sanity check: room for longest mode, flag and some nodes
2703 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2705 if (!pol || pol == &default_policy)
2706 mode = MPOL_DEFAULT;
2715 case MPOL_PREFERRED:
2717 if (flags & MPOL_F_LOCAL)
2718 mode = MPOL_LOCAL; /* pseudo-policy */
2720 node_set(pol->v.preferred_node, nodes);
2725 case MPOL_INTERLEAVE:
2727 nodes = pol->w.user_nodemask;
2729 nodes = pol->v.nodes;
2736 l = strlen(policy_modes[mode]);
2737 if (buffer + maxlen < p + l + 1)
2740 strcpy(p, policy_modes[mode]);
2743 if (flags & MPOL_MODE_FLAGS) {
2744 if (buffer + maxlen < p + 2)
2749 * Currently, the only defined flags are mutually exclusive
2751 if (flags & MPOL_F_STATIC_NODES)
2752 p += snprintf(p, buffer + maxlen - p, "static");
2753 else if (flags & MPOL_F_RELATIVE_NODES)
2754 p += snprintf(p, buffer + maxlen - p, "relative");
2757 if (!nodes_empty(nodes)) {
2758 if (buffer + maxlen < p + 2)
2761 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);