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>
94 #include <asm/tlbflush.h>
95 #include <asm/uaccess.h>
96 #include <linux/random.h>
101 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
104 static struct kmem_cache *policy_cache;
105 static struct kmem_cache *sn_cache;
107 /* Highest zone. An specific allocation for a zone below that is not
109 enum zone_type policy_zone = 0;
112 * run-time system-wide default policy => local allocation
114 static struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
120 static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
128 * If we have a lock to protect task->mempolicy in read-side, we do
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
138 } mpol_ops[MPOL_MAX];
140 /* Check that the nodemask contains at least one populated zone */
141 static int is_valid_nodemask(const nodemask_t *nodemask)
145 for_each_node_mask(nd, *nodemask) {
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
158 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
160 return pol->flags & MPOL_MODE_FLAGS;
163 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
171 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
173 if (nodes_empty(*nodes))
175 pol->v.nodes = *nodes;
179 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
186 pol->v.preferred_node = first_node(*nodes);
190 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
192 if (!is_valid_nodemask(nodes))
194 pol->v.nodes = *nodes;
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
207 static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
238 ret = mpol_ops[pol->mode].create(pol, NULL);
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
246 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
249 struct mempolicy *policy;
251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 return NULL; /* simply delete any existing policy */
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
279 policy->flags = flags;
284 /* Slow path of a mpol destructor. */
285 void __mpol_put(struct mempolicy *p)
287 if (!atomic_dec_and_test(&p->refcnt))
289 kmem_cache_free(policy_cache, p);
292 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
303 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
328 if (nodes_empty(tmp))
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
347 static void mpol_rebind_preferred(struct mempolicy *pol,
348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
356 if (node_isset(node, *nodes)) {
357 pol->v.preferred_node = node;
358 pol->flags &= ~MPOL_F_LOCAL;
360 pol->flags |= MPOL_F_LOCAL;
361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
368 pol->w.cpuset_mems_allowed = *nodes;
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
380 * If we have a lock to protect task->mempolicy in read-side, we do
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
388 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
393 if (!mpol_store_user_nodemask(pol) && step == 0 &&
394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
417 * Called with task's alloc_lock held.
420 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
423 mpol_rebind_policy(tsk->mempolicy, new, step);
427 * Rebind each vma in mm to new nodemask.
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
432 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
434 struct vm_area_struct *vma;
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
439 up_write(&mm->mmap_sem);
442 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
444 .rebind = mpol_rebind_default,
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
460 static void migrate_page_add(struct page *page, struct list_head *pagelist,
461 unsigned long flags);
463 /* Scan through pages checking if pages follow certain conditions. */
464 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
465 unsigned long addr, unsigned long end,
466 const nodemask_t *nodes, unsigned long flags,
473 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
478 if (!pte_present(*pte))
480 page = vm_normal_page(vma, addr, *pte);
484 * vm_normal_page() filters out zero pages, but there might
485 * still be PageReserved pages to skip, perhaps in a VDSO.
486 * And we cannot move PageKsm pages sensibly or safely yet.
488 if (PageReserved(page) || PageKsm(page))
490 nid = page_to_nid(page);
491 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
495 migrate_page_add(page, private, flags);
498 } while (pte++, addr += PAGE_SIZE, addr != end);
499 pte_unmap_unlock(orig_pte, ptl);
503 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
504 unsigned long addr, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags,
511 pmd = pmd_offset(pud, addr);
513 next = pmd_addr_end(addr, end);
514 split_huge_page_pmd(vma->vm_mm, pmd);
515 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
517 if (check_pte_range(vma, pmd, addr, next, nodes,
520 } while (pmd++, addr = next, addr != end);
524 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
525 unsigned long addr, unsigned long end,
526 const nodemask_t *nodes, unsigned long flags,
532 pud = pud_offset(pgd, addr);
534 next = pud_addr_end(addr, end);
535 if (pud_none_or_clear_bad(pud))
537 if (check_pmd_range(vma, pud, addr, next, nodes,
540 } while (pud++, addr = next, addr != end);
544 static inline int check_pgd_range(struct vm_area_struct *vma,
545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
552 pgd = pgd_offset(vma->vm_mm, addr);
554 next = pgd_addr_end(addr, end);
555 if (pgd_none_or_clear_bad(pgd))
557 if (check_pud_range(vma, pgd, addr, next, nodes,
560 } while (pgd++, addr = next, addr != end);
565 * Check if all pages in a range are on a set of nodes.
566 * If pagelist != NULL then isolate pages from the LRU and
567 * put them on the pagelist.
570 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
571 const nodemask_t *nodes, unsigned long flags, void *private)
574 struct vm_area_struct *vma, *prev;
576 vma = find_vma(mm, start);
580 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
581 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
582 if (!vma->vm_next && vma->vm_end < end)
584 if (prev && prev->vm_end < vma->vm_start)
587 if (!is_vm_hugetlb_page(vma) &&
588 ((flags & MPOL_MF_STRICT) ||
589 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
590 vma_migratable(vma)))) {
591 unsigned long endvma = vma->vm_end;
595 if (vma->vm_start > start)
596 start = vma->vm_start;
597 err = check_pgd_range(vma, start, endvma, nodes,
608 * Apply policy to a single VMA
609 * This must be called with the mmap_sem held for writing.
611 static int vma_replace_policy(struct vm_area_struct *vma,
612 struct mempolicy *pol)
615 struct mempolicy *old;
616 struct mempolicy *new;
618 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
619 vma->vm_start, vma->vm_end, vma->vm_pgoff,
620 vma->vm_ops, vma->vm_file,
621 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
627 if (vma->vm_ops && vma->vm_ops->set_policy) {
628 err = vma->vm_ops->set_policy(vma, new);
633 old = vma->vm_policy;
634 vma->vm_policy = new; /* protected by mmap_sem */
643 /* Step 2: apply policy to a range and do splits. */
644 static int mbind_range(struct mm_struct *mm, unsigned long start,
645 unsigned long end, struct mempolicy *new_pol)
647 struct vm_area_struct *next;
648 struct vm_area_struct *prev;
649 struct vm_area_struct *vma;
652 unsigned long vmstart;
655 vma = find_vma_prev(mm, start, &prev);
656 if (!vma || vma->vm_start > start)
659 if (start > vma->vm_start)
662 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
664 vmstart = max(start, vma->vm_start);
665 vmend = min(end, vma->vm_end);
667 if (mpol_equal(vma_policy(vma), new_pol))
670 pgoff = vma->vm_pgoff +
671 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
672 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
673 vma->anon_vma, vma->vm_file, pgoff,
680 if (vma->vm_start != vmstart) {
681 err = split_vma(vma->vm_mm, vma, vmstart, 1);
685 if (vma->vm_end != vmend) {
686 err = split_vma(vma->vm_mm, vma, vmend, 0);
690 err = vma_replace_policy(vma, new_pol);
700 * Update task->flags PF_MEMPOLICY bit: set iff non-default
701 * mempolicy. Allows more rapid checking of this (combined perhaps
702 * with other PF_* flag bits) on memory allocation hot code paths.
704 * If called from outside this file, the task 'p' should -only- be
705 * a newly forked child not yet visible on the task list, because
706 * manipulating the task flags of a visible task is not safe.
708 * The above limitation is why this routine has the funny name
709 * mpol_fix_fork_child_flag().
711 * It is also safe to call this with a task pointer of current,
712 * which the static wrapper mpol_set_task_struct_flag() does,
713 * for use within this file.
716 void mpol_fix_fork_child_flag(struct task_struct *p)
719 p->flags |= PF_MEMPOLICY;
721 p->flags &= ~PF_MEMPOLICY;
724 static void mpol_set_task_struct_flag(void)
726 mpol_fix_fork_child_flag(current);
729 /* Set the process memory policy */
730 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
733 struct mempolicy *new, *old;
734 struct mm_struct *mm = current->mm;
735 NODEMASK_SCRATCH(scratch);
741 new = mpol_new(mode, flags, nodes);
747 * prevent changing our mempolicy while show_numa_maps()
749 * Note: do_set_mempolicy() can be called at init time
753 down_write(&mm->mmap_sem);
755 ret = mpol_set_nodemask(new, nodes, scratch);
757 task_unlock(current);
759 up_write(&mm->mmap_sem);
763 old = current->mempolicy;
764 current->mempolicy = new;
765 mpol_set_task_struct_flag();
766 if (new && new->mode == MPOL_INTERLEAVE &&
767 nodes_weight(new->v.nodes))
768 current->il_next = first_node(new->v.nodes);
769 task_unlock(current);
771 up_write(&mm->mmap_sem);
776 NODEMASK_SCRATCH_FREE(scratch);
781 * Return nodemask for policy for get_mempolicy() query
783 * Called with task's alloc_lock held
785 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
788 if (p == &default_policy)
794 case MPOL_INTERLEAVE:
798 if (!(p->flags & MPOL_F_LOCAL))
799 node_set(p->v.preferred_node, *nodes);
800 /* else return empty node mask for local allocation */
807 static int lookup_node(struct mm_struct *mm, unsigned long addr)
812 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
814 err = page_to_nid(p);
820 /* Retrieve NUMA policy */
821 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
822 unsigned long addr, unsigned long flags)
825 struct mm_struct *mm = current->mm;
826 struct vm_area_struct *vma = NULL;
827 struct mempolicy *pol = current->mempolicy;
830 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
833 if (flags & MPOL_F_MEMS_ALLOWED) {
834 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
836 *policy = 0; /* just so it's initialized */
838 *nmask = cpuset_current_mems_allowed;
839 task_unlock(current);
843 if (flags & MPOL_F_ADDR) {
845 * Do NOT fall back to task policy if the
846 * vma/shared policy at addr is NULL. We
847 * want to return MPOL_DEFAULT in this case.
849 down_read(&mm->mmap_sem);
850 vma = find_vma_intersection(mm, addr, addr+1);
852 up_read(&mm->mmap_sem);
855 if (vma->vm_ops && vma->vm_ops->get_policy)
856 pol = vma->vm_ops->get_policy(vma, addr);
858 pol = vma->vm_policy;
863 pol = &default_policy; /* indicates default behavior */
865 if (flags & MPOL_F_NODE) {
866 if (flags & MPOL_F_ADDR) {
867 err = lookup_node(mm, addr);
871 } else if (pol == current->mempolicy &&
872 pol->mode == MPOL_INTERLEAVE) {
873 *policy = current->il_next;
879 *policy = pol == &default_policy ? MPOL_DEFAULT :
882 * Internal mempolicy flags must be masked off before exposing
883 * the policy to userspace.
885 *policy |= (pol->flags & MPOL_MODE_FLAGS);
890 if (mpol_store_user_nodemask(pol)) {
891 *nmask = pol->w.user_nodemask;
894 get_policy_nodemask(pol, nmask);
895 task_unlock(current);
902 up_read(¤t->mm->mmap_sem);
906 #ifdef CONFIG_MIGRATION
910 static void migrate_page_add(struct page *page, struct list_head *pagelist,
914 * Avoid migrating a page that is shared with others.
916 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
917 if (!isolate_lru_page(page)) {
918 list_add_tail(&page->lru, pagelist);
919 inc_zone_page_state(page, NR_ISOLATED_ANON +
920 page_is_file_cache(page));
925 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
927 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
931 * Migrate pages from one node to a target node.
932 * Returns error or the number of pages not migrated.
934 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
942 node_set(source, nmask);
945 * This does not "check" the range but isolates all pages that
946 * need migration. Between passing in the full user address
947 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
949 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
950 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
951 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
953 if (!list_empty(&pagelist)) {
954 err = migrate_pages(&pagelist, new_node_page, dest,
955 false, MIGRATE_SYNC);
957 putback_lru_pages(&pagelist);
964 * Move pages between the two nodesets so as to preserve the physical
965 * layout as much as possible.
967 * Returns the number of page that could not be moved.
969 int do_migrate_pages(struct mm_struct *mm,
970 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
976 err = migrate_prep();
980 down_read(&mm->mmap_sem);
982 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
987 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
988 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
989 * bit in 'tmp', and return that <source, dest> pair for migration.
990 * The pair of nodemasks 'to' and 'from' define the map.
992 * If no pair of bits is found that way, fallback to picking some
993 * pair of 'source' and 'dest' bits that are not the same. If the
994 * 'source' and 'dest' bits are the same, this represents a node
995 * that will be migrating to itself, so no pages need move.
997 * If no bits are left in 'tmp', or if all remaining bits left
998 * in 'tmp' correspond to the same bit in 'to', return false
999 * (nothing left to migrate).
1001 * This lets us pick a pair of nodes to migrate between, such that
1002 * if possible the dest node is not already occupied by some other
1003 * source node, minimizing the risk of overloading the memory on a
1004 * node that would happen if we migrated incoming memory to a node
1005 * before migrating outgoing memory source that same node.
1007 * A single scan of tmp is sufficient. As we go, we remember the
1008 * most recent <s, d> pair that moved (s != d). If we find a pair
1009 * that not only moved, but what's better, moved to an empty slot
1010 * (d is not set in tmp), then we break out then, with that pair.
1011 * Otherwise when we finish scanning from_tmp, we at least have the
1012 * most recent <s, d> pair that moved. If we get all the way through
1013 * the scan of tmp without finding any node that moved, much less
1014 * moved to an empty node, then there is nothing left worth migrating.
1018 while (!nodes_empty(tmp)) {
1023 for_each_node_mask(s, tmp) {
1024 d = node_remap(s, *from_nodes, *to_nodes);
1028 source = s; /* Node moved. Memorize */
1031 /* dest not in remaining from nodes? */
1032 if (!node_isset(dest, tmp))
1038 node_clear(source, tmp);
1039 err = migrate_to_node(mm, source, dest, flags);
1046 up_read(&mm->mmap_sem);
1054 * Allocate a new page for page migration based on vma policy.
1055 * Start by assuming the page is mapped by the same vma as contains @start.
1056 * Search forward from there, if not. N.B., this assumes that the
1057 * list of pages handed to migrate_pages()--which is how we get here--
1058 * is in virtual address order.
1060 static struct page *new_page(struct page *page, unsigned long start, int **x)
1062 struct vm_area_struct *vma;
1063 unsigned long uninitialized_var(address);
1065 vma = find_vma(current->mm, start);
1067 address = page_address_in_vma(page, vma);
1068 if (address != -EFAULT)
1074 * if !vma, alloc_page_vma() will use task or system default policy
1076 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1080 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1081 unsigned long flags)
1085 int do_migrate_pages(struct mm_struct *mm,
1086 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1091 static struct page *new_page(struct page *page, unsigned long start, int **x)
1097 static long do_mbind(unsigned long start, unsigned long len,
1098 unsigned short mode, unsigned short mode_flags,
1099 nodemask_t *nmask, unsigned long flags)
1101 struct mm_struct *mm = current->mm;
1102 struct mempolicy *new;
1105 LIST_HEAD(pagelist);
1107 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1108 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1110 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1113 if (start & ~PAGE_MASK)
1116 if (mode == MPOL_DEFAULT)
1117 flags &= ~MPOL_MF_STRICT;
1119 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1127 new = mpol_new(mode, mode_flags, nmask);
1129 return PTR_ERR(new);
1132 * If we are using the default policy then operation
1133 * on discontinuous address spaces is okay after all
1136 flags |= MPOL_MF_DISCONTIG_OK;
1138 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1139 start, start + len, mode, mode_flags,
1140 nmask ? nodes_addr(*nmask)[0] : -1);
1142 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1144 err = migrate_prep();
1149 NODEMASK_SCRATCH(scratch);
1151 down_write(&mm->mmap_sem);
1153 err = mpol_set_nodemask(new, nmask, scratch);
1154 task_unlock(current);
1156 up_write(&mm->mmap_sem);
1159 NODEMASK_SCRATCH_FREE(scratch);
1164 err = check_range(mm, start, end, nmask,
1165 flags | MPOL_MF_INVERT, &pagelist);
1169 err = mbind_range(mm, start, end, new);
1171 if (!list_empty(&pagelist)) {
1172 nr_failed = migrate_pages(&pagelist, new_page,
1173 start, false, true);
1175 putback_lru_pages(&pagelist);
1178 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1181 putback_lru_pages(&pagelist);
1183 up_write(&mm->mmap_sem);
1190 * User space interface with variable sized bitmaps for nodelists.
1193 /* Copy a node mask from user space. */
1194 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1195 unsigned long maxnode)
1198 unsigned long nlongs;
1199 unsigned long endmask;
1202 nodes_clear(*nodes);
1203 if (maxnode == 0 || !nmask)
1205 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1208 nlongs = BITS_TO_LONGS(maxnode);
1209 if ((maxnode % BITS_PER_LONG) == 0)
1212 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1214 /* When the user specified more nodes than supported just check
1215 if the non supported part is all zero. */
1216 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1217 if (nlongs > PAGE_SIZE/sizeof(long))
1219 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1221 if (get_user(t, nmask + k))
1223 if (k == nlongs - 1) {
1229 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1233 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1235 nodes_addr(*nodes)[nlongs-1] &= endmask;
1239 /* Copy a kernel node mask to user space */
1240 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1243 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1244 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1246 if (copy > nbytes) {
1247 if (copy > PAGE_SIZE)
1249 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1253 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1256 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1257 unsigned long, mode, unsigned long __user *, nmask,
1258 unsigned long, maxnode, unsigned, flags)
1262 unsigned short mode_flags;
1264 mode_flags = mode & MPOL_MODE_FLAGS;
1265 mode &= ~MPOL_MODE_FLAGS;
1266 if (mode >= MPOL_MAX)
1268 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1269 (mode_flags & MPOL_F_RELATIVE_NODES))
1271 err = get_nodes(&nodes, nmask, maxnode);
1274 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1277 /* Set the process memory policy */
1278 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1279 unsigned long, maxnode)
1283 unsigned short flags;
1285 flags = mode & MPOL_MODE_FLAGS;
1286 mode &= ~MPOL_MODE_FLAGS;
1287 if ((unsigned int)mode >= MPOL_MAX)
1289 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1291 err = get_nodes(&nodes, nmask, maxnode);
1294 return do_set_mempolicy(mode, flags, &nodes);
1297 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1298 const unsigned long __user *, old_nodes,
1299 const unsigned long __user *, new_nodes)
1301 const struct cred *cred = current_cred(), *tcred;
1302 struct mm_struct *mm = NULL;
1303 struct task_struct *task;
1304 nodemask_t task_nodes;
1308 NODEMASK_SCRATCH(scratch);
1313 old = &scratch->mask1;
1314 new = &scratch->mask2;
1316 err = get_nodes(old, old_nodes, maxnode);
1320 err = get_nodes(new, new_nodes, maxnode);
1324 /* Find the mm_struct */
1326 task = pid ? find_task_by_vpid(pid) : current;
1332 get_task_struct(task);
1337 * Check if this process has the right to modify the specified
1338 * process. The right exists if the process has administrative
1339 * capabilities, superuser privileges or the same
1340 * userid as the target process.
1342 tcred = __task_cred(task);
1343 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1344 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1345 !capable(CAP_SYS_NICE)) {
1352 task_nodes = cpuset_mems_allowed(task);
1353 /* Is the user allowed to access the target nodes? */
1354 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1359 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1364 err = security_task_movememory(task);
1368 mm = get_task_mm(task);
1369 put_task_struct(task);
1376 err = do_migrate_pages(mm, old, new,
1377 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1381 NODEMASK_SCRATCH_FREE(scratch);
1386 put_task_struct(task);
1392 /* Retrieve NUMA policy */
1393 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1394 unsigned long __user *, nmask, unsigned long, maxnode,
1395 unsigned long, addr, unsigned long, flags)
1398 int uninitialized_var(pval);
1401 if (nmask != NULL && maxnode < MAX_NUMNODES)
1404 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1409 if (policy && put_user(pval, policy))
1413 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1418 #ifdef CONFIG_COMPAT
1420 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1421 compat_ulong_t __user *nmask,
1422 compat_ulong_t maxnode,
1423 compat_ulong_t addr, compat_ulong_t flags)
1426 unsigned long __user *nm = NULL;
1427 unsigned long nr_bits, alloc_size;
1428 DECLARE_BITMAP(bm, MAX_NUMNODES);
1430 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1431 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1434 nm = compat_alloc_user_space(alloc_size);
1436 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1438 if (!err && nmask) {
1439 unsigned long copy_size;
1440 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1441 err = copy_from_user(bm, nm, copy_size);
1442 /* ensure entire bitmap is zeroed */
1443 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1444 err |= compat_put_bitmap(nmask, bm, nr_bits);
1450 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1451 compat_ulong_t maxnode)
1453 unsigned long __user *nm = NULL;
1454 unsigned long nr_bits, alloc_size;
1455 DECLARE_BITMAP(bm, MAX_NUMNODES);
1457 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1458 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1461 if (compat_get_bitmap(bm, nmask, nr_bits))
1463 nm = compat_alloc_user_space(alloc_size);
1464 if (copy_to_user(nm, bm, alloc_size))
1468 return sys_set_mempolicy(mode, nm, nr_bits+1);
1471 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1472 compat_ulong_t mode, compat_ulong_t __user *nmask,
1473 compat_ulong_t maxnode, compat_ulong_t flags)
1475 unsigned long __user *nm = NULL;
1476 unsigned long nr_bits, alloc_size;
1479 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1480 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1483 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1485 nm = compat_alloc_user_space(alloc_size);
1486 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1490 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1496 * get_vma_policy(@task, @vma, @addr)
1497 * @task - task for fallback if vma policy == default
1498 * @vma - virtual memory area whose policy is sought
1499 * @addr - address in @vma for shared policy lookup
1501 * Returns effective policy for a VMA at specified address.
1502 * Falls back to @task or system default policy, as necessary.
1503 * Current or other task's task mempolicy and non-shared vma policies
1504 * are protected by the task's mmap_sem, which must be held for read by
1506 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1507 * count--added by the get_policy() vm_op, as appropriate--to protect against
1508 * freeing by another task. It is the caller's responsibility to free the
1509 * extra reference for shared policies.
1511 struct mempolicy *get_vma_policy(struct task_struct *task,
1512 struct vm_area_struct *vma, unsigned long addr)
1514 struct mempolicy *pol = task->mempolicy;
1517 if (vma->vm_ops && vma->vm_ops->get_policy) {
1518 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1522 } else if (vma->vm_policy) {
1523 pol = vma->vm_policy;
1526 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1527 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1528 * count on these policies which will be dropped by
1529 * mpol_cond_put() later
1531 if (mpol_needs_cond_ref(pol))
1536 pol = &default_policy;
1541 * Return a nodemask representing a mempolicy for filtering nodes for
1544 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1546 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1547 if (unlikely(policy->mode == MPOL_BIND) &&
1548 gfp_zone(gfp) >= policy_zone &&
1549 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1550 return &policy->v.nodes;
1555 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1556 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1559 switch (policy->mode) {
1560 case MPOL_PREFERRED:
1561 if (!(policy->flags & MPOL_F_LOCAL))
1562 nd = policy->v.preferred_node;
1566 * Normally, MPOL_BIND allocations are node-local within the
1567 * allowed nodemask. However, if __GFP_THISNODE is set and the
1568 * current node isn't part of the mask, we use the zonelist for
1569 * the first node in the mask instead.
1571 if (unlikely(gfp & __GFP_THISNODE) &&
1572 unlikely(!node_isset(nd, policy->v.nodes)))
1573 nd = first_node(policy->v.nodes);
1578 return node_zonelist(nd, gfp);
1581 /* Do dynamic interleaving for a process */
1582 static unsigned interleave_nodes(struct mempolicy *policy)
1585 struct task_struct *me = current;
1588 next = next_node(nid, policy->v.nodes);
1589 if (next >= MAX_NUMNODES)
1590 next = first_node(policy->v.nodes);
1591 if (next < MAX_NUMNODES)
1597 * Depending on the memory policy provide a node from which to allocate the
1599 * @policy must be protected by freeing by the caller. If @policy is
1600 * the current task's mempolicy, this protection is implicit, as only the
1601 * task can change it's policy. The system default policy requires no
1604 unsigned slab_node(void)
1606 struct mempolicy *policy;
1609 return numa_node_id();
1611 policy = current->mempolicy;
1612 if (!policy || policy->flags & MPOL_F_LOCAL)
1613 return numa_node_id();
1615 switch (policy->mode) {
1616 case MPOL_PREFERRED:
1618 * handled MPOL_F_LOCAL above
1620 return policy->v.preferred_node;
1622 case MPOL_INTERLEAVE:
1623 return interleave_nodes(policy);
1627 * Follow bind policy behavior and start allocation at the
1630 struct zonelist *zonelist;
1632 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1633 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1634 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1637 return zone ? zone->node : numa_node_id();
1645 /* Do static interleaving for a VMA with known offset. */
1646 static unsigned offset_il_node(struct mempolicy *pol,
1647 struct vm_area_struct *vma, unsigned long off)
1649 unsigned nnodes = nodes_weight(pol->v.nodes);
1655 return numa_node_id();
1656 target = (unsigned int)off % nnodes;
1659 nid = next_node(nid, pol->v.nodes);
1661 } while (c <= target);
1665 /* Determine a node number for interleave */
1666 static inline unsigned interleave_nid(struct mempolicy *pol,
1667 struct vm_area_struct *vma, unsigned long addr, int shift)
1673 * for small pages, there is no difference between
1674 * shift and PAGE_SHIFT, so the bit-shift is safe.
1675 * for huge pages, since vm_pgoff is in units of small
1676 * pages, we need to shift off the always 0 bits to get
1679 BUG_ON(shift < PAGE_SHIFT);
1680 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1681 off += (addr - vma->vm_start) >> shift;
1682 return offset_il_node(pol, vma, off);
1684 return interleave_nodes(pol);
1688 * Return the bit number of a random bit set in the nodemask.
1689 * (returns -1 if nodemask is empty)
1691 int node_random(const nodemask_t *maskp)
1695 w = nodes_weight(*maskp);
1697 bit = bitmap_ord_to_pos(maskp->bits,
1698 get_random_int() % w, MAX_NUMNODES);
1702 #ifdef CONFIG_HUGETLBFS
1704 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1705 * @vma = virtual memory area whose policy is sought
1706 * @addr = address in @vma for shared policy lookup and interleave policy
1707 * @gfp_flags = for requested zone
1708 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1709 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1711 * Returns a zonelist suitable for a huge page allocation and a pointer
1712 * to the struct mempolicy for conditional unref after allocation.
1713 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1714 * @nodemask for filtering the zonelist.
1716 * Must be protected by get_mems_allowed()
1718 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1719 gfp_t gfp_flags, struct mempolicy **mpol,
1720 nodemask_t **nodemask)
1722 struct zonelist *zl;
1724 *mpol = get_vma_policy(current, vma, addr);
1725 *nodemask = NULL; /* assume !MPOL_BIND */
1727 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1728 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1729 huge_page_shift(hstate_vma(vma))), gfp_flags);
1731 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1732 if ((*mpol)->mode == MPOL_BIND)
1733 *nodemask = &(*mpol)->v.nodes;
1739 * init_nodemask_of_mempolicy
1741 * If the current task's mempolicy is "default" [NULL], return 'false'
1742 * to indicate default policy. Otherwise, extract the policy nodemask
1743 * for 'bind' or 'interleave' policy into the argument nodemask, or
1744 * initialize the argument nodemask to contain the single node for
1745 * 'preferred' or 'local' policy and return 'true' to indicate presence
1746 * of non-default mempolicy.
1748 * We don't bother with reference counting the mempolicy [mpol_get/put]
1749 * because the current task is examining it's own mempolicy and a task's
1750 * mempolicy is only ever changed by the task itself.
1752 * N.B., it is the caller's responsibility to free a returned nodemask.
1754 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1756 struct mempolicy *mempolicy;
1759 if (!(mask && current->mempolicy))
1763 mempolicy = current->mempolicy;
1764 switch (mempolicy->mode) {
1765 case MPOL_PREFERRED:
1766 if (mempolicy->flags & MPOL_F_LOCAL)
1767 nid = numa_node_id();
1769 nid = mempolicy->v.preferred_node;
1770 init_nodemask_of_node(mask, nid);
1775 case MPOL_INTERLEAVE:
1776 *mask = mempolicy->v.nodes;
1782 task_unlock(current);
1789 * mempolicy_nodemask_intersects
1791 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1792 * policy. Otherwise, check for intersection between mask and the policy
1793 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1794 * policy, always return true since it may allocate elsewhere on fallback.
1796 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1798 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1799 const nodemask_t *mask)
1801 struct mempolicy *mempolicy;
1807 mempolicy = tsk->mempolicy;
1811 switch (mempolicy->mode) {
1812 case MPOL_PREFERRED:
1814 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1815 * allocate from, they may fallback to other nodes when oom.
1816 * Thus, it's possible for tsk to have allocated memory from
1821 case MPOL_INTERLEAVE:
1822 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1832 /* Allocate a page in interleaved policy.
1833 Own path because it needs to do special accounting. */
1834 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1837 struct zonelist *zl;
1840 zl = node_zonelist(nid, gfp);
1841 page = __alloc_pages(gfp, order, zl);
1842 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1843 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1848 * alloc_pages_vma - Allocate a page for a VMA.
1851 * %GFP_USER user allocation.
1852 * %GFP_KERNEL kernel allocations,
1853 * %GFP_HIGHMEM highmem/user allocations,
1854 * %GFP_FS allocation should not call back into a file system.
1855 * %GFP_ATOMIC don't sleep.
1857 * @order:Order of the GFP allocation.
1858 * @vma: Pointer to VMA or NULL if not available.
1859 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1861 * This function allocates a page from the kernel page pool and applies
1862 * a NUMA policy associated with the VMA or the current process.
1863 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1864 * mm_struct of the VMA to prevent it from going away. Should be used for
1865 * all allocations for pages that will be mapped into
1866 * user space. Returns NULL when no page can be allocated.
1868 * Should be called with the mm_sem of the vma hold.
1871 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1872 unsigned long addr, int node)
1874 struct mempolicy *pol;
1875 struct zonelist *zl;
1877 unsigned int cpuset_mems_cookie;
1880 pol = get_vma_policy(current, vma, addr);
1881 cpuset_mems_cookie = get_mems_allowed();
1883 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1886 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1888 page = alloc_page_interleave(gfp, order, nid);
1889 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1894 zl = policy_zonelist(gfp, pol, node);
1895 if (unlikely(mpol_needs_cond_ref(pol))) {
1897 * slow path: ref counted shared policy
1899 struct page *page = __alloc_pages_nodemask(gfp, order,
1900 zl, policy_nodemask(gfp, pol));
1902 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1907 * fast path: default or task policy
1909 page = __alloc_pages_nodemask(gfp, order, zl,
1910 policy_nodemask(gfp, pol));
1911 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1917 * alloc_pages_current - Allocate pages.
1920 * %GFP_USER user allocation,
1921 * %GFP_KERNEL kernel allocation,
1922 * %GFP_HIGHMEM highmem allocation,
1923 * %GFP_FS don't call back into a file system.
1924 * %GFP_ATOMIC don't sleep.
1925 * @order: Power of two of allocation size in pages. 0 is a single page.
1927 * Allocate a page from the kernel page pool. When not in
1928 * interrupt context and apply the current process NUMA policy.
1929 * Returns NULL when no page can be allocated.
1931 * Don't call cpuset_update_task_memory_state() unless
1932 * 1) it's ok to take cpuset_sem (can WAIT), and
1933 * 2) allocating for current task (not interrupt).
1935 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1937 struct mempolicy *pol = current->mempolicy;
1939 unsigned int cpuset_mems_cookie;
1941 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1942 pol = &default_policy;
1945 cpuset_mems_cookie = get_mems_allowed();
1948 * No reference counting needed for current->mempolicy
1949 * nor system default_policy
1951 if (pol->mode == MPOL_INTERLEAVE)
1952 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1954 page = __alloc_pages_nodemask(gfp, order,
1955 policy_zonelist(gfp, pol, numa_node_id()),
1956 policy_nodemask(gfp, pol));
1958 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1963 EXPORT_SYMBOL(alloc_pages_current);
1966 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1967 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1968 * with the mems_allowed returned by cpuset_mems_allowed(). This
1969 * keeps mempolicies cpuset relative after its cpuset moves. See
1970 * further kernel/cpuset.c update_nodemask().
1972 * current's mempolicy may be rebinded by the other task(the task that changes
1973 * cpuset's mems), so we needn't do rebind work for current task.
1976 /* Slow path of a mempolicy duplicate */
1977 struct mempolicy *__mpol_dup(struct mempolicy *old)
1979 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1982 return ERR_PTR(-ENOMEM);
1984 /* task's mempolicy is protected by alloc_lock */
1985 if (old == current->mempolicy) {
1988 task_unlock(current);
1992 if (current_cpuset_is_being_rebound()) {
1993 nodemask_t mems = cpuset_mems_allowed(current);
1994 if (new->flags & MPOL_F_REBINDING)
1995 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1997 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1999 atomic_set(&new->refcnt, 1);
2003 /* Slow path of a mempolicy comparison */
2004 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2008 if (a->mode != b->mode)
2010 if (a->flags != b->flags)
2012 if (mpol_store_user_nodemask(a))
2013 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2019 case MPOL_INTERLEAVE:
2020 return nodes_equal(a->v.nodes, b->v.nodes);
2021 case MPOL_PREFERRED:
2022 /* a's ->flags is the same as b's */
2023 if (a->flags & MPOL_F_LOCAL)
2025 return a->v.preferred_node == b->v.preferred_node;
2033 * Shared memory backing store policy support.
2035 * Remember policies even when nobody has shared memory mapped.
2036 * The policies are kept in Red-Black tree linked from the inode.
2037 * They are protected by the sp->lock spinlock, which should be held
2038 * for any accesses to the tree.
2041 /* lookup first element intersecting start-end */
2042 /* Caller holds sp->mutex */
2043 static struct sp_node *
2044 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2046 struct rb_node *n = sp->root.rb_node;
2049 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2051 if (start >= p->end)
2053 else if (end <= p->start)
2061 struct sp_node *w = NULL;
2062 struct rb_node *prev = rb_prev(n);
2065 w = rb_entry(prev, struct sp_node, nd);
2066 if (w->end <= start)
2070 return rb_entry(n, struct sp_node, nd);
2073 /* Insert a new shared policy into the list. */
2074 /* Caller holds sp->lock */
2075 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2077 struct rb_node **p = &sp->root.rb_node;
2078 struct rb_node *parent = NULL;
2083 nd = rb_entry(parent, struct sp_node, nd);
2084 if (new->start < nd->start)
2086 else if (new->end > nd->end)
2087 p = &(*p)->rb_right;
2091 rb_link_node(&new->nd, parent, p);
2092 rb_insert_color(&new->nd, &sp->root);
2093 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2094 new->policy ? new->policy->mode : 0);
2097 /* Find shared policy intersecting idx */
2099 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2101 struct mempolicy *pol = NULL;
2104 if (!sp->root.rb_node)
2106 mutex_lock(&sp->mutex);
2107 sn = sp_lookup(sp, idx, idx+1);
2109 mpol_get(sn->policy);
2112 mutex_unlock(&sp->mutex);
2116 static void sp_free(struct sp_node *n)
2118 mpol_put(n->policy);
2119 kmem_cache_free(sn_cache, n);
2122 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2124 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2125 rb_erase(&n->nd, &sp->root);
2129 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2130 struct mempolicy *pol)
2133 struct mempolicy *newpol;
2135 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2139 newpol = mpol_dup(pol);
2140 if (IS_ERR(newpol)) {
2141 kmem_cache_free(sn_cache, n);
2144 newpol->flags |= MPOL_F_SHARED;
2153 /* Replace a policy range. */
2154 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2155 unsigned long end, struct sp_node *new)
2160 mutex_lock(&sp->mutex);
2161 n = sp_lookup(sp, start, end);
2162 /* Take care of old policies in the same range. */
2163 while (n && n->start < end) {
2164 struct rb_node *next = rb_next(&n->nd);
2165 if (n->start >= start) {
2171 /* Old policy spanning whole new range. */
2173 struct sp_node *new2;
2174 new2 = sp_alloc(end, n->end, n->policy);
2180 sp_insert(sp, new2);
2187 n = rb_entry(next, struct sp_node, nd);
2192 mutex_unlock(&sp->mutex);
2197 * mpol_shared_policy_init - initialize shared policy for inode
2198 * @sp: pointer to inode shared policy
2199 * @mpol: struct mempolicy to install
2201 * Install non-NULL @mpol in inode's shared policy rb-tree.
2202 * On entry, the current task has a reference on a non-NULL @mpol.
2203 * This must be released on exit.
2204 * This is called at get_inode() calls and we can use GFP_KERNEL.
2206 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2210 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2211 mutex_init(&sp->mutex);
2214 struct vm_area_struct pvma;
2215 struct mempolicy *new;
2216 NODEMASK_SCRATCH(scratch);
2220 /* contextualize the tmpfs mount point mempolicy */
2221 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2223 goto free_scratch; /* no valid nodemask intersection */
2226 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2227 task_unlock(current);
2231 /* Create pseudo-vma that contains just the policy */
2232 memset(&pvma, 0, sizeof(struct vm_area_struct));
2233 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2234 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2237 mpol_put(new); /* drop initial ref */
2239 NODEMASK_SCRATCH_FREE(scratch);
2241 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2245 int mpol_set_shared_policy(struct shared_policy *info,
2246 struct vm_area_struct *vma, struct mempolicy *npol)
2249 struct sp_node *new = NULL;
2250 unsigned long sz = vma_pages(vma);
2252 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2254 sz, npol ? npol->mode : -1,
2255 npol ? npol->flags : -1,
2256 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2259 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2263 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2269 /* Free a backing policy store on inode delete. */
2270 void mpol_free_shared_policy(struct shared_policy *p)
2273 struct rb_node *next;
2275 if (!p->root.rb_node)
2277 mutex_lock(&p->mutex);
2278 next = rb_first(&p->root);
2280 n = rb_entry(next, struct sp_node, nd);
2281 next = rb_next(&n->nd);
2284 mutex_unlock(&p->mutex);
2287 /* assumes fs == KERNEL_DS */
2288 void __init numa_policy_init(void)
2290 nodemask_t interleave_nodes;
2291 unsigned long largest = 0;
2292 int nid, prefer = 0;
2294 policy_cache = kmem_cache_create("numa_policy",
2295 sizeof(struct mempolicy),
2296 0, SLAB_PANIC, NULL);
2298 sn_cache = kmem_cache_create("shared_policy_node",
2299 sizeof(struct sp_node),
2300 0, SLAB_PANIC, NULL);
2303 * Set interleaving policy for system init. Interleaving is only
2304 * enabled across suitably sized nodes (default is >= 16MB), or
2305 * fall back to the largest node if they're all smaller.
2307 nodes_clear(interleave_nodes);
2308 for_each_node_state(nid, N_HIGH_MEMORY) {
2309 unsigned long total_pages = node_present_pages(nid);
2311 /* Preserve the largest node */
2312 if (largest < total_pages) {
2313 largest = total_pages;
2317 /* Interleave this node? */
2318 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2319 node_set(nid, interleave_nodes);
2322 /* All too small, use the largest */
2323 if (unlikely(nodes_empty(interleave_nodes)))
2324 node_set(prefer, interleave_nodes);
2326 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2327 printk("numa_policy_init: interleaving failed\n");
2330 /* Reset policy of current process to default */
2331 void numa_default_policy(void)
2333 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2337 * Parse and format mempolicy from/to strings
2341 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2343 #define MPOL_LOCAL MPOL_MAX
2344 static const char * const policy_modes[] =
2346 [MPOL_DEFAULT] = "default",
2347 [MPOL_PREFERRED] = "prefer",
2348 [MPOL_BIND] = "bind",
2349 [MPOL_INTERLEAVE] = "interleave",
2350 [MPOL_LOCAL] = "local"
2356 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2357 * @str: string containing mempolicy to parse
2358 * @mpol: pointer to struct mempolicy pointer, returned on success.
2359 * @unused: redundant argument, to be removed later.
2362 * <mode>[=<flags>][:<nodelist>]
2364 * On success, returns 0, else 1
2366 int mpol_parse_str(char *str, struct mempolicy **mpol, int unused)
2368 struct mempolicy *new = NULL;
2369 unsigned short mode;
2370 unsigned short mode_flags;
2372 char *nodelist = strchr(str, ':');
2373 char *flags = strchr(str, '=');
2377 /* NUL-terminate mode or flags string */
2379 if (nodelist_parse(nodelist, nodes))
2381 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2387 *flags++ = '\0'; /* terminate mode string */
2389 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2390 if (!strcmp(str, policy_modes[mode])) {
2394 if (mode > MPOL_LOCAL)
2398 case MPOL_PREFERRED:
2400 * Insist on a nodelist of one node only
2403 char *rest = nodelist;
2404 while (isdigit(*rest))
2410 case MPOL_INTERLEAVE:
2412 * Default to online nodes with memory if no nodelist
2415 nodes = node_states[N_HIGH_MEMORY];
2419 * Don't allow a nodelist; mpol_new() checks flags
2423 mode = MPOL_PREFERRED;
2427 * Insist on a empty nodelist
2434 * Insist on a nodelist
2443 * Currently, we only support two mutually exclusive
2446 if (!strcmp(flags, "static"))
2447 mode_flags |= MPOL_F_STATIC_NODES;
2448 else if (!strcmp(flags, "relative"))
2449 mode_flags |= MPOL_F_RELATIVE_NODES;
2454 new = mpol_new(mode, mode_flags, &nodes);
2459 * Save nodes for mpol_to_str() to show the tmpfs mount options
2460 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2462 if (mode != MPOL_PREFERRED)
2463 new->v.nodes = nodes;
2465 new->v.preferred_node = first_node(nodes);
2467 new->flags |= MPOL_F_LOCAL;
2470 * Save nodes for contextualization: this will be used to "clone"
2471 * the mempolicy in a specific context [cpuset] at a later time.
2473 new->w.user_nodemask = nodes;
2478 /* Restore string for error message */
2487 #endif /* CONFIG_TMPFS */
2490 * mpol_to_str - format a mempolicy structure for printing
2491 * @buffer: to contain formatted mempolicy string
2492 * @maxlen: length of @buffer
2493 * @pol: pointer to mempolicy to be formatted
2494 * @unused: redundant argument, to be removed later.
2496 * Convert a mempolicy into a string.
2497 * Returns the number of characters in buffer (if positive)
2498 * or an error (negative)
2500 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int unused)
2505 unsigned short mode;
2506 unsigned short flags = pol ? pol->flags : 0;
2509 * Sanity check: room for longest mode, flag and some nodes
2511 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2513 if (!pol || pol == &default_policy)
2514 mode = MPOL_DEFAULT;
2523 case MPOL_PREFERRED:
2525 if (flags & MPOL_F_LOCAL)
2528 node_set(pol->v.preferred_node, nodes);
2533 case MPOL_INTERLEAVE:
2534 nodes = pol->v.nodes;
2541 l = strlen(policy_modes[mode]);
2542 if (buffer + maxlen < p + l + 1)
2545 strcpy(p, policy_modes[mode]);
2548 if (flags & MPOL_MODE_FLAGS) {
2549 if (buffer + maxlen < p + 2)
2554 * Currently, the only defined flags are mutually exclusive
2556 if (flags & MPOL_F_STATIC_NODES)
2557 p += snprintf(p, buffer + maxlen - p, "static");
2558 else if (flags & MPOL_F_RELATIVE_NODES)
2559 p += snprintf(p, buffer + maxlen - p, "relative");
2562 if (!nodes_empty(nodes)) {
2563 if (buffer + maxlen < p + 2)
2566 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);