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);
889 up_read(¤t->mm->mmap_sem);
895 if (mpol_store_user_nodemask(pol)) {
896 *nmask = pol->w.user_nodemask;
899 get_policy_nodemask(pol, nmask);
900 task_unlock(current);
907 up_read(¤t->mm->mmap_sem);
911 #ifdef CONFIG_MIGRATION
915 static void migrate_page_add(struct page *page, struct list_head *pagelist,
919 * Avoid migrating a page that is shared with others.
921 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
922 if (!isolate_lru_page(page)) {
923 list_add_tail(&page->lru, pagelist);
924 inc_zone_page_state(page, NR_ISOLATED_ANON +
925 page_is_file_cache(page));
930 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
932 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
936 * Migrate pages from one node to a target node.
937 * Returns error or the number of pages not migrated.
939 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
947 node_set(source, nmask);
950 * This does not "check" the range but isolates all pages that
951 * need migration. Between passing in the full user address
952 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
954 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
955 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
956 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
958 if (!list_empty(&pagelist)) {
959 err = migrate_pages(&pagelist, new_node_page, dest,
960 false, MIGRATE_SYNC);
962 putback_lru_pages(&pagelist);
969 * Move pages between the two nodesets so as to preserve the physical
970 * layout as much as possible.
972 * Returns the number of page that could not be moved.
974 int do_migrate_pages(struct mm_struct *mm,
975 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
981 err = migrate_prep();
985 down_read(&mm->mmap_sem);
987 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
992 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
993 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
994 * bit in 'tmp', and return that <source, dest> pair for migration.
995 * The pair of nodemasks 'to' and 'from' define the map.
997 * If no pair of bits is found that way, fallback to picking some
998 * pair of 'source' and 'dest' bits that are not the same. If the
999 * 'source' and 'dest' bits are the same, this represents a node
1000 * that will be migrating to itself, so no pages need move.
1002 * If no bits are left in 'tmp', or if all remaining bits left
1003 * in 'tmp' correspond to the same bit in 'to', return false
1004 * (nothing left to migrate).
1006 * This lets us pick a pair of nodes to migrate between, such that
1007 * if possible the dest node is not already occupied by some other
1008 * source node, minimizing the risk of overloading the memory on a
1009 * node that would happen if we migrated incoming memory to a node
1010 * before migrating outgoing memory source that same node.
1012 * A single scan of tmp is sufficient. As we go, we remember the
1013 * most recent <s, d> pair that moved (s != d). If we find a pair
1014 * that not only moved, but what's better, moved to an empty slot
1015 * (d is not set in tmp), then we break out then, with that pair.
1016 * Otherwise when we finish scanning from_tmp, we at least have the
1017 * most recent <s, d> pair that moved. If we get all the way through
1018 * the scan of tmp without finding any node that moved, much less
1019 * moved to an empty node, then there is nothing left worth migrating.
1023 while (!nodes_empty(tmp)) {
1028 for_each_node_mask(s, tmp) {
1029 d = node_remap(s, *from_nodes, *to_nodes);
1033 source = s; /* Node moved. Memorize */
1036 /* dest not in remaining from nodes? */
1037 if (!node_isset(dest, tmp))
1043 node_clear(source, tmp);
1044 err = migrate_to_node(mm, source, dest, flags);
1051 up_read(&mm->mmap_sem);
1059 * Allocate a new page for page migration based on vma policy.
1060 * Start by assuming the page is mapped by the same vma as contains @start.
1061 * Search forward from there, if not. N.B., this assumes that the
1062 * list of pages handed to migrate_pages()--which is how we get here--
1063 * is in virtual address order.
1065 static struct page *new_page(struct page *page, unsigned long start, int **x)
1067 struct vm_area_struct *vma;
1068 unsigned long uninitialized_var(address);
1070 vma = find_vma(current->mm, start);
1072 address = page_address_in_vma(page, vma);
1073 if (address != -EFAULT)
1079 * if !vma, alloc_page_vma() will use task or system default policy
1081 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1085 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1086 unsigned long flags)
1090 int do_migrate_pages(struct mm_struct *mm,
1091 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1096 static struct page *new_page(struct page *page, unsigned long start, int **x)
1102 static long do_mbind(unsigned long start, unsigned long len,
1103 unsigned short mode, unsigned short mode_flags,
1104 nodemask_t *nmask, unsigned long flags)
1106 struct mm_struct *mm = current->mm;
1107 struct mempolicy *new;
1110 LIST_HEAD(pagelist);
1112 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1113 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1115 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1118 if (start & ~PAGE_MASK)
1121 if (mode == MPOL_DEFAULT)
1122 flags &= ~MPOL_MF_STRICT;
1124 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1132 new = mpol_new(mode, mode_flags, nmask);
1134 return PTR_ERR(new);
1137 * If we are using the default policy then operation
1138 * on discontinuous address spaces is okay after all
1141 flags |= MPOL_MF_DISCONTIG_OK;
1143 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1144 start, start + len, mode, mode_flags,
1145 nmask ? nodes_addr(*nmask)[0] : -1);
1147 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1149 err = migrate_prep();
1154 NODEMASK_SCRATCH(scratch);
1156 down_write(&mm->mmap_sem);
1158 err = mpol_set_nodemask(new, nmask, scratch);
1159 task_unlock(current);
1161 up_write(&mm->mmap_sem);
1164 NODEMASK_SCRATCH_FREE(scratch);
1169 err = check_range(mm, start, end, nmask,
1170 flags | MPOL_MF_INVERT, &pagelist);
1174 err = mbind_range(mm, start, end, new);
1176 if (!list_empty(&pagelist)) {
1177 nr_failed = migrate_pages(&pagelist, new_page,
1178 start, false, true);
1180 putback_lru_pages(&pagelist);
1183 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1186 putback_lru_pages(&pagelist);
1188 up_write(&mm->mmap_sem);
1195 * User space interface with variable sized bitmaps for nodelists.
1198 /* Copy a node mask from user space. */
1199 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1200 unsigned long maxnode)
1203 unsigned long nlongs;
1204 unsigned long endmask;
1207 nodes_clear(*nodes);
1208 if (maxnode == 0 || !nmask)
1210 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1213 nlongs = BITS_TO_LONGS(maxnode);
1214 if ((maxnode % BITS_PER_LONG) == 0)
1217 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1219 /* When the user specified more nodes than supported just check
1220 if the non supported part is all zero. */
1221 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1222 if (nlongs > PAGE_SIZE/sizeof(long))
1224 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1226 if (get_user(t, nmask + k))
1228 if (k == nlongs - 1) {
1234 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1238 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1240 nodes_addr(*nodes)[nlongs-1] &= endmask;
1244 /* Copy a kernel node mask to user space */
1245 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1248 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1249 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1251 if (copy > nbytes) {
1252 if (copy > PAGE_SIZE)
1254 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1258 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1261 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1262 unsigned long, mode, unsigned long __user *, nmask,
1263 unsigned long, maxnode, unsigned, flags)
1267 unsigned short mode_flags;
1269 mode_flags = mode & MPOL_MODE_FLAGS;
1270 mode &= ~MPOL_MODE_FLAGS;
1271 if (mode >= MPOL_MAX)
1273 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1274 (mode_flags & MPOL_F_RELATIVE_NODES))
1276 err = get_nodes(&nodes, nmask, maxnode);
1279 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1282 /* Set the process memory policy */
1283 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1284 unsigned long, maxnode)
1288 unsigned short flags;
1290 flags = mode & MPOL_MODE_FLAGS;
1291 mode &= ~MPOL_MODE_FLAGS;
1292 if ((unsigned int)mode >= MPOL_MAX)
1294 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1296 err = get_nodes(&nodes, nmask, maxnode);
1299 return do_set_mempolicy(mode, flags, &nodes);
1302 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1303 const unsigned long __user *, old_nodes,
1304 const unsigned long __user *, new_nodes)
1306 const struct cred *cred = current_cred(), *tcred;
1307 struct mm_struct *mm = NULL;
1308 struct task_struct *task;
1309 nodemask_t task_nodes;
1313 NODEMASK_SCRATCH(scratch);
1318 old = &scratch->mask1;
1319 new = &scratch->mask2;
1321 err = get_nodes(old, old_nodes, maxnode);
1325 err = get_nodes(new, new_nodes, maxnode);
1329 /* Find the mm_struct */
1331 task = pid ? find_task_by_vpid(pid) : current;
1337 get_task_struct(task);
1342 * Check if this process has the right to modify the specified
1343 * process. The right exists if the process has administrative
1344 * capabilities, superuser privileges or the same
1345 * userid as the target process.
1347 tcred = __task_cred(task);
1348 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1349 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1350 !capable(CAP_SYS_NICE)) {
1357 task_nodes = cpuset_mems_allowed(task);
1358 /* Is the user allowed to access the target nodes? */
1359 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1364 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1369 err = security_task_movememory(task);
1373 mm = get_task_mm(task);
1374 put_task_struct(task);
1381 err = do_migrate_pages(mm, old, new,
1382 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1386 NODEMASK_SCRATCH_FREE(scratch);
1391 put_task_struct(task);
1397 /* Retrieve NUMA policy */
1398 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1399 unsigned long __user *, nmask, unsigned long, maxnode,
1400 unsigned long, addr, unsigned long, flags)
1403 int uninitialized_var(pval);
1406 if (nmask != NULL && maxnode < MAX_NUMNODES)
1409 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1414 if (policy && put_user(pval, policy))
1418 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1423 #ifdef CONFIG_COMPAT
1425 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1426 compat_ulong_t __user *nmask,
1427 compat_ulong_t maxnode,
1428 compat_ulong_t addr, compat_ulong_t flags)
1431 unsigned long __user *nm = NULL;
1432 unsigned long nr_bits, alloc_size;
1433 DECLARE_BITMAP(bm, MAX_NUMNODES);
1435 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1436 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1439 nm = compat_alloc_user_space(alloc_size);
1441 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1443 if (!err && nmask) {
1444 unsigned long copy_size;
1445 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1446 err = copy_from_user(bm, nm, copy_size);
1447 /* ensure entire bitmap is zeroed */
1448 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1449 err |= compat_put_bitmap(nmask, bm, nr_bits);
1455 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1456 compat_ulong_t maxnode)
1458 unsigned long __user *nm = NULL;
1459 unsigned long nr_bits, alloc_size;
1460 DECLARE_BITMAP(bm, MAX_NUMNODES);
1462 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1463 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1466 if (compat_get_bitmap(bm, nmask, nr_bits))
1468 nm = compat_alloc_user_space(alloc_size);
1469 if (copy_to_user(nm, bm, alloc_size))
1473 return sys_set_mempolicy(mode, nm, nr_bits+1);
1476 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1477 compat_ulong_t mode, compat_ulong_t __user *nmask,
1478 compat_ulong_t maxnode, compat_ulong_t flags)
1480 unsigned long __user *nm = NULL;
1481 unsigned long nr_bits, alloc_size;
1484 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1485 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1488 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1490 nm = compat_alloc_user_space(alloc_size);
1491 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1495 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1501 * get_vma_policy(@task, @vma, @addr)
1502 * @task - task for fallback if vma policy == default
1503 * @vma - virtual memory area whose policy is sought
1504 * @addr - address in @vma for shared policy lookup
1506 * Returns effective policy for a VMA at specified address.
1507 * Falls back to @task or system default policy, as necessary.
1508 * Current or other task's task mempolicy and non-shared vma policies
1509 * are protected by the task's mmap_sem, which must be held for read by
1511 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1512 * count--added by the get_policy() vm_op, as appropriate--to protect against
1513 * freeing by another task. It is the caller's responsibility to free the
1514 * extra reference for shared policies.
1516 struct mempolicy *get_vma_policy(struct task_struct *task,
1517 struct vm_area_struct *vma, unsigned long addr)
1519 struct mempolicy *pol = task->mempolicy;
1522 if (vma->vm_ops && vma->vm_ops->get_policy) {
1523 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1527 } else if (vma->vm_policy) {
1528 pol = vma->vm_policy;
1531 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1532 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1533 * count on these policies which will be dropped by
1534 * mpol_cond_put() later
1536 if (mpol_needs_cond_ref(pol))
1541 pol = &default_policy;
1546 * Return a nodemask representing a mempolicy for filtering nodes for
1549 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1551 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1552 if (unlikely(policy->mode == MPOL_BIND) &&
1553 gfp_zone(gfp) >= policy_zone &&
1554 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1555 return &policy->v.nodes;
1560 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1561 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1564 switch (policy->mode) {
1565 case MPOL_PREFERRED:
1566 if (!(policy->flags & MPOL_F_LOCAL))
1567 nd = policy->v.preferred_node;
1571 * Normally, MPOL_BIND allocations are node-local within the
1572 * allowed nodemask. However, if __GFP_THISNODE is set and the
1573 * current node isn't part of the mask, we use the zonelist for
1574 * the first node in the mask instead.
1576 if (unlikely(gfp & __GFP_THISNODE) &&
1577 unlikely(!node_isset(nd, policy->v.nodes)))
1578 nd = first_node(policy->v.nodes);
1583 return node_zonelist(nd, gfp);
1586 /* Do dynamic interleaving for a process */
1587 static unsigned interleave_nodes(struct mempolicy *policy)
1590 struct task_struct *me = current;
1593 next = next_node(nid, policy->v.nodes);
1594 if (next >= MAX_NUMNODES)
1595 next = first_node(policy->v.nodes);
1596 if (next < MAX_NUMNODES)
1602 * Depending on the memory policy provide a node from which to allocate the
1604 * @policy must be protected by freeing by the caller. If @policy is
1605 * the current task's mempolicy, this protection is implicit, as only the
1606 * task can change it's policy. The system default policy requires no
1609 unsigned slab_node(void)
1611 struct mempolicy *policy;
1614 return numa_node_id();
1616 policy = current->mempolicy;
1617 if (!policy || policy->flags & MPOL_F_LOCAL)
1618 return numa_node_id();
1620 switch (policy->mode) {
1621 case MPOL_PREFERRED:
1623 * handled MPOL_F_LOCAL above
1625 return policy->v.preferred_node;
1627 case MPOL_INTERLEAVE:
1628 return interleave_nodes(policy);
1632 * Follow bind policy behavior and start allocation at the
1635 struct zonelist *zonelist;
1637 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1638 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1639 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1642 return zone ? zone->node : numa_node_id();
1650 /* Do static interleaving for a VMA with known offset. */
1651 static unsigned offset_il_node(struct mempolicy *pol,
1652 struct vm_area_struct *vma, unsigned long off)
1654 unsigned nnodes = nodes_weight(pol->v.nodes);
1660 return numa_node_id();
1661 target = (unsigned int)off % nnodes;
1664 nid = next_node(nid, pol->v.nodes);
1666 } while (c <= target);
1670 /* Determine a node number for interleave */
1671 static inline unsigned interleave_nid(struct mempolicy *pol,
1672 struct vm_area_struct *vma, unsigned long addr, int shift)
1678 * for small pages, there is no difference between
1679 * shift and PAGE_SHIFT, so the bit-shift is safe.
1680 * for huge pages, since vm_pgoff is in units of small
1681 * pages, we need to shift off the always 0 bits to get
1684 BUG_ON(shift < PAGE_SHIFT);
1685 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1686 off += (addr - vma->vm_start) >> shift;
1687 return offset_il_node(pol, vma, off);
1689 return interleave_nodes(pol);
1693 * Return the bit number of a random bit set in the nodemask.
1694 * (returns -1 if nodemask is empty)
1696 int node_random(const nodemask_t *maskp)
1700 w = nodes_weight(*maskp);
1702 bit = bitmap_ord_to_pos(maskp->bits,
1703 get_random_int() % w, MAX_NUMNODES);
1707 #ifdef CONFIG_HUGETLBFS
1709 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1710 * @vma = virtual memory area whose policy is sought
1711 * @addr = address in @vma for shared policy lookup and interleave policy
1712 * @gfp_flags = for requested zone
1713 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1714 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1716 * Returns a zonelist suitable for a huge page allocation and a pointer
1717 * to the struct mempolicy for conditional unref after allocation.
1718 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1719 * @nodemask for filtering the zonelist.
1721 * Must be protected by get_mems_allowed()
1723 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1724 gfp_t gfp_flags, struct mempolicy **mpol,
1725 nodemask_t **nodemask)
1727 struct zonelist *zl;
1729 *mpol = get_vma_policy(current, vma, addr);
1730 *nodemask = NULL; /* assume !MPOL_BIND */
1732 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1733 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1734 huge_page_shift(hstate_vma(vma))), gfp_flags);
1736 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1737 if ((*mpol)->mode == MPOL_BIND)
1738 *nodemask = &(*mpol)->v.nodes;
1744 * init_nodemask_of_mempolicy
1746 * If the current task's mempolicy is "default" [NULL], return 'false'
1747 * to indicate default policy. Otherwise, extract the policy nodemask
1748 * for 'bind' or 'interleave' policy into the argument nodemask, or
1749 * initialize the argument nodemask to contain the single node for
1750 * 'preferred' or 'local' policy and return 'true' to indicate presence
1751 * of non-default mempolicy.
1753 * We don't bother with reference counting the mempolicy [mpol_get/put]
1754 * because the current task is examining it's own mempolicy and a task's
1755 * mempolicy is only ever changed by the task itself.
1757 * N.B., it is the caller's responsibility to free a returned nodemask.
1759 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1761 struct mempolicy *mempolicy;
1764 if (!(mask && current->mempolicy))
1768 mempolicy = current->mempolicy;
1769 switch (mempolicy->mode) {
1770 case MPOL_PREFERRED:
1771 if (mempolicy->flags & MPOL_F_LOCAL)
1772 nid = numa_node_id();
1774 nid = mempolicy->v.preferred_node;
1775 init_nodemask_of_node(mask, nid);
1780 case MPOL_INTERLEAVE:
1781 *mask = mempolicy->v.nodes;
1787 task_unlock(current);
1794 * mempolicy_nodemask_intersects
1796 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1797 * policy. Otherwise, check for intersection between mask and the policy
1798 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1799 * policy, always return true since it may allocate elsewhere on fallback.
1801 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1803 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1804 const nodemask_t *mask)
1806 struct mempolicy *mempolicy;
1812 mempolicy = tsk->mempolicy;
1816 switch (mempolicy->mode) {
1817 case MPOL_PREFERRED:
1819 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1820 * allocate from, they may fallback to other nodes when oom.
1821 * Thus, it's possible for tsk to have allocated memory from
1826 case MPOL_INTERLEAVE:
1827 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1837 /* Allocate a page in interleaved policy.
1838 Own path because it needs to do special accounting. */
1839 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1842 struct zonelist *zl;
1845 zl = node_zonelist(nid, gfp);
1846 page = __alloc_pages(gfp, order, zl);
1847 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1848 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1853 * alloc_pages_vma - Allocate a page for a VMA.
1856 * %GFP_USER user allocation.
1857 * %GFP_KERNEL kernel allocations,
1858 * %GFP_HIGHMEM highmem/user allocations,
1859 * %GFP_FS allocation should not call back into a file system.
1860 * %GFP_ATOMIC don't sleep.
1862 * @order:Order of the GFP allocation.
1863 * @vma: Pointer to VMA or NULL if not available.
1864 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1866 * This function allocates a page from the kernel page pool and applies
1867 * a NUMA policy associated with the VMA or the current process.
1868 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1869 * mm_struct of the VMA to prevent it from going away. Should be used for
1870 * all allocations for pages that will be mapped into
1871 * user space. Returns NULL when no page can be allocated.
1873 * Should be called with the mm_sem of the vma hold.
1876 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1877 unsigned long addr, int node)
1879 struct mempolicy *pol;
1880 struct zonelist *zl;
1882 unsigned int cpuset_mems_cookie;
1885 pol = get_vma_policy(current, vma, addr);
1886 cpuset_mems_cookie = get_mems_allowed();
1888 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1891 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1893 page = alloc_page_interleave(gfp, order, nid);
1894 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1899 zl = policy_zonelist(gfp, pol, node);
1900 if (unlikely(mpol_needs_cond_ref(pol))) {
1902 * slow path: ref counted shared policy
1904 struct page *page = __alloc_pages_nodemask(gfp, order,
1905 zl, policy_nodemask(gfp, pol));
1907 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1912 * fast path: default or task policy
1914 page = __alloc_pages_nodemask(gfp, order, zl,
1915 policy_nodemask(gfp, pol));
1916 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1922 * alloc_pages_current - Allocate pages.
1925 * %GFP_USER user allocation,
1926 * %GFP_KERNEL kernel allocation,
1927 * %GFP_HIGHMEM highmem allocation,
1928 * %GFP_FS don't call back into a file system.
1929 * %GFP_ATOMIC don't sleep.
1930 * @order: Power of two of allocation size in pages. 0 is a single page.
1932 * Allocate a page from the kernel page pool. When not in
1933 * interrupt context and apply the current process NUMA policy.
1934 * Returns NULL when no page can be allocated.
1936 * Don't call cpuset_update_task_memory_state() unless
1937 * 1) it's ok to take cpuset_sem (can WAIT), and
1938 * 2) allocating for current task (not interrupt).
1940 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1942 struct mempolicy *pol = current->mempolicy;
1944 unsigned int cpuset_mems_cookie;
1946 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1947 pol = &default_policy;
1950 cpuset_mems_cookie = get_mems_allowed();
1953 * No reference counting needed for current->mempolicy
1954 * nor system default_policy
1956 if (pol->mode == MPOL_INTERLEAVE)
1957 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1959 page = __alloc_pages_nodemask(gfp, order,
1960 policy_zonelist(gfp, pol, numa_node_id()),
1961 policy_nodemask(gfp, pol));
1963 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1968 EXPORT_SYMBOL(alloc_pages_current);
1971 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1972 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1973 * with the mems_allowed returned by cpuset_mems_allowed(). This
1974 * keeps mempolicies cpuset relative after its cpuset moves. See
1975 * further kernel/cpuset.c update_nodemask().
1977 * current's mempolicy may be rebinded by the other task(the task that changes
1978 * cpuset's mems), so we needn't do rebind work for current task.
1981 /* Slow path of a mempolicy duplicate */
1982 struct mempolicy *__mpol_dup(struct mempolicy *old)
1984 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1987 return ERR_PTR(-ENOMEM);
1989 /* task's mempolicy is protected by alloc_lock */
1990 if (old == current->mempolicy) {
1993 task_unlock(current);
1997 if (current_cpuset_is_being_rebound()) {
1998 nodemask_t mems = cpuset_mems_allowed(current);
1999 if (new->flags & MPOL_F_REBINDING)
2000 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2002 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2004 atomic_set(&new->refcnt, 1);
2008 /* Slow path of a mempolicy comparison */
2009 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2013 if (a->mode != b->mode)
2015 if (a->flags != b->flags)
2017 if (mpol_store_user_nodemask(a))
2018 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2024 case MPOL_INTERLEAVE:
2025 return nodes_equal(a->v.nodes, b->v.nodes);
2026 case MPOL_PREFERRED:
2027 return a->v.preferred_node == b->v.preferred_node;
2035 * Shared memory backing store policy support.
2037 * Remember policies even when nobody has shared memory mapped.
2038 * The policies are kept in Red-Black tree linked from the inode.
2039 * They are protected by the sp->lock spinlock, which should be held
2040 * for any accesses to the tree.
2043 /* lookup first element intersecting start-end */
2044 /* Caller holds sp->mutex */
2045 static struct sp_node *
2046 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2048 struct rb_node *n = sp->root.rb_node;
2051 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2053 if (start >= p->end)
2055 else if (end <= p->start)
2063 struct sp_node *w = NULL;
2064 struct rb_node *prev = rb_prev(n);
2067 w = rb_entry(prev, struct sp_node, nd);
2068 if (w->end <= start)
2072 return rb_entry(n, struct sp_node, nd);
2075 /* Insert a new shared policy into the list. */
2076 /* Caller holds sp->lock */
2077 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2079 struct rb_node **p = &sp->root.rb_node;
2080 struct rb_node *parent = NULL;
2085 nd = rb_entry(parent, struct sp_node, nd);
2086 if (new->start < nd->start)
2088 else if (new->end > nd->end)
2089 p = &(*p)->rb_right;
2093 rb_link_node(&new->nd, parent, p);
2094 rb_insert_color(&new->nd, &sp->root);
2095 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2096 new->policy ? new->policy->mode : 0);
2099 /* Find shared policy intersecting idx */
2101 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2103 struct mempolicy *pol = NULL;
2106 if (!sp->root.rb_node)
2108 mutex_lock(&sp->mutex);
2109 sn = sp_lookup(sp, idx, idx+1);
2111 mpol_get(sn->policy);
2114 mutex_unlock(&sp->mutex);
2118 static void sp_free(struct sp_node *n)
2120 mpol_put(n->policy);
2121 kmem_cache_free(sn_cache, n);
2124 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2126 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2127 rb_erase(&n->nd, &sp->root);
2131 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2132 struct mempolicy *pol)
2135 struct mempolicy *newpol;
2137 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2141 newpol = mpol_dup(pol);
2142 if (IS_ERR(newpol)) {
2143 kmem_cache_free(sn_cache, n);
2146 newpol->flags |= MPOL_F_SHARED;
2155 /* Replace a policy range. */
2156 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2157 unsigned long end, struct sp_node *new)
2162 mutex_lock(&sp->mutex);
2163 n = sp_lookup(sp, start, end);
2164 /* Take care of old policies in the same range. */
2165 while (n && n->start < end) {
2166 struct rb_node *next = rb_next(&n->nd);
2167 if (n->start >= start) {
2173 /* Old policy spanning whole new range. */
2175 struct sp_node *new2;
2176 new2 = sp_alloc(end, n->end, n->policy);
2182 sp_insert(sp, new2);
2189 n = rb_entry(next, struct sp_node, nd);
2194 mutex_unlock(&sp->mutex);
2199 * mpol_shared_policy_init - initialize shared policy for inode
2200 * @sp: pointer to inode shared policy
2201 * @mpol: struct mempolicy to install
2203 * Install non-NULL @mpol in inode's shared policy rb-tree.
2204 * On entry, the current task has a reference on a non-NULL @mpol.
2205 * This must be released on exit.
2206 * This is called at get_inode() calls and we can use GFP_KERNEL.
2208 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2212 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2213 mutex_init(&sp->mutex);
2216 struct vm_area_struct pvma;
2217 struct mempolicy *new;
2218 NODEMASK_SCRATCH(scratch);
2222 /* contextualize the tmpfs mount point mempolicy */
2223 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2225 goto free_scratch; /* no valid nodemask intersection */
2228 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2229 task_unlock(current);
2233 /* Create pseudo-vma that contains just the policy */
2234 memset(&pvma, 0, sizeof(struct vm_area_struct));
2235 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2236 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2239 mpol_put(new); /* drop initial ref */
2241 NODEMASK_SCRATCH_FREE(scratch);
2243 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2247 int mpol_set_shared_policy(struct shared_policy *info,
2248 struct vm_area_struct *vma, struct mempolicy *npol)
2251 struct sp_node *new = NULL;
2252 unsigned long sz = vma_pages(vma);
2254 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2256 sz, npol ? npol->mode : -1,
2257 npol ? npol->flags : -1,
2258 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2261 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2265 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2271 /* Free a backing policy store on inode delete. */
2272 void mpol_free_shared_policy(struct shared_policy *p)
2275 struct rb_node *next;
2277 if (!p->root.rb_node)
2279 mutex_lock(&p->mutex);
2280 next = rb_first(&p->root);
2282 n = rb_entry(next, struct sp_node, nd);
2283 next = rb_next(&n->nd);
2286 mutex_unlock(&p->mutex);
2289 /* assumes fs == KERNEL_DS */
2290 void __init numa_policy_init(void)
2292 nodemask_t interleave_nodes;
2293 unsigned long largest = 0;
2294 int nid, prefer = 0;
2296 policy_cache = kmem_cache_create("numa_policy",
2297 sizeof(struct mempolicy),
2298 0, SLAB_PANIC, NULL);
2300 sn_cache = kmem_cache_create("shared_policy_node",
2301 sizeof(struct sp_node),
2302 0, SLAB_PANIC, NULL);
2305 * Set interleaving policy for system init. Interleaving is only
2306 * enabled across suitably sized nodes (default is >= 16MB), or
2307 * fall back to the largest node if they're all smaller.
2309 nodes_clear(interleave_nodes);
2310 for_each_node_state(nid, N_HIGH_MEMORY) {
2311 unsigned long total_pages = node_present_pages(nid);
2313 /* Preserve the largest node */
2314 if (largest < total_pages) {
2315 largest = total_pages;
2319 /* Interleave this node? */
2320 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2321 node_set(nid, interleave_nodes);
2324 /* All too small, use the largest */
2325 if (unlikely(nodes_empty(interleave_nodes)))
2326 node_set(prefer, interleave_nodes);
2328 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2329 printk("numa_policy_init: interleaving failed\n");
2332 /* Reset policy of current process to default */
2333 void numa_default_policy(void)
2335 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2339 * Parse and format mempolicy from/to strings
2343 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2345 #define MPOL_LOCAL MPOL_MAX
2346 static const char * const policy_modes[] =
2348 [MPOL_DEFAULT] = "default",
2349 [MPOL_PREFERRED] = "prefer",
2350 [MPOL_BIND] = "bind",
2351 [MPOL_INTERLEAVE] = "interleave",
2352 [MPOL_LOCAL] = "local"
2358 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2359 * @str: string containing mempolicy to parse
2360 * @mpol: pointer to struct mempolicy pointer, returned on success.
2361 * @unused: redundant argument, to be removed later.
2364 * <mode>[=<flags>][:<nodelist>]
2366 * On success, returns 0, else 1
2368 int mpol_parse_str(char *str, struct mempolicy **mpol, int unused)
2370 struct mempolicy *new = NULL;
2371 unsigned short mode;
2372 unsigned short mode_flags;
2374 char *nodelist = strchr(str, ':');
2375 char *flags = strchr(str, '=');
2379 /* NUL-terminate mode or flags string */
2381 if (nodelist_parse(nodelist, nodes))
2383 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2389 *flags++ = '\0'; /* terminate mode string */
2391 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2392 if (!strcmp(str, policy_modes[mode])) {
2396 if (mode > MPOL_LOCAL)
2400 case MPOL_PREFERRED:
2402 * Insist on a nodelist of one node only
2405 char *rest = nodelist;
2406 while (isdigit(*rest))
2412 case MPOL_INTERLEAVE:
2414 * Default to online nodes with memory if no nodelist
2417 nodes = node_states[N_HIGH_MEMORY];
2421 * Don't allow a nodelist; mpol_new() checks flags
2425 mode = MPOL_PREFERRED;
2429 * Insist on a empty nodelist
2436 * Insist on a nodelist
2445 * Currently, we only support two mutually exclusive
2448 if (!strcmp(flags, "static"))
2449 mode_flags |= MPOL_F_STATIC_NODES;
2450 else if (!strcmp(flags, "relative"))
2451 mode_flags |= MPOL_F_RELATIVE_NODES;
2456 new = mpol_new(mode, mode_flags, &nodes);
2461 * Save nodes for mpol_to_str() to show the tmpfs mount options
2462 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2464 if (mode != MPOL_PREFERRED)
2465 new->v.nodes = nodes;
2467 new->v.preferred_node = first_node(nodes);
2469 new->flags |= MPOL_F_LOCAL;
2472 * Save nodes for contextualization: this will be used to "clone"
2473 * the mempolicy in a specific context [cpuset] at a later time.
2475 new->w.user_nodemask = nodes;
2480 /* Restore string for error message */
2489 #endif /* CONFIG_TMPFS */
2492 * mpol_to_str - format a mempolicy structure for printing
2493 * @buffer: to contain formatted mempolicy string
2494 * @maxlen: length of @buffer
2495 * @pol: pointer to mempolicy to be formatted
2496 * @unused: redundant argument, to be removed later.
2498 * Convert a mempolicy into a string.
2499 * Returns the number of characters in buffer (if positive)
2500 * or an error (negative)
2502 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int unused)
2507 unsigned short mode;
2508 unsigned short flags = pol ? pol->flags : 0;
2511 * Sanity check: room for longest mode, flag and some nodes
2513 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2515 if (!pol || pol == &default_policy)
2516 mode = MPOL_DEFAULT;
2525 case MPOL_PREFERRED:
2527 if (flags & MPOL_F_LOCAL)
2530 node_set(pol->v.preferred_node, nodes);
2535 case MPOL_INTERLEAVE:
2536 nodes = pol->v.nodes;
2543 l = strlen(policy_modes[mode]);
2544 if (buffer + maxlen < p + l + 1)
2547 strcpy(p, policy_modes[mode]);
2550 if (flags & MPOL_MODE_FLAGS) {
2551 if (buffer + maxlen < p + 2)
2556 * Currently, the only defined flags are mutually exclusive
2558 if (flags & MPOL_F_STATIC_NODES)
2559 p += snprintf(p, buffer + maxlen - p, "static");
2560 else if (flags & MPOL_F_RELATIVE_NODES)
2561 p += snprintf(p, buffer + maxlen - p, "relative");
2564 if (!nodes_empty(nodes)) {
2565 if (buffer + maxlen < p + 2)
2568 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);