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.
569 static struct vm_area_struct *
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 *first, *vma, *prev;
577 first = find_vma(mm, start);
579 return ERR_PTR(-EFAULT);
581 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
582 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
583 if (!vma->vm_next && vma->vm_end < end)
584 return ERR_PTR(-EFAULT);
585 if (prev && prev->vm_end < vma->vm_start)
586 return ERR_PTR(-EFAULT);
588 if (!is_vm_hugetlb_page(vma) &&
589 ((flags & MPOL_MF_STRICT) ||
590 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
591 vma_migratable(vma)))) {
592 unsigned long endvma = vma->vm_end;
596 if (vma->vm_start > start)
597 start = vma->vm_start;
598 err = check_pgd_range(vma, start, endvma, nodes,
601 first = ERR_PTR(err);
611 * Apply policy to a single VMA
612 * This must be called with the mmap_sem held for writing.
614 static int vma_replace_policy(struct vm_area_struct *vma,
615 struct mempolicy *pol)
618 struct mempolicy *old;
619 struct mempolicy *new;
621 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
622 vma->vm_start, vma->vm_end, vma->vm_pgoff,
623 vma->vm_ops, vma->vm_file,
624 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
630 if (vma->vm_ops && vma->vm_ops->set_policy) {
631 err = vma->vm_ops->set_policy(vma, new);
636 old = vma->vm_policy;
637 vma->vm_policy = new; /* protected by mmap_sem */
646 /* Step 2: apply policy to a range and do splits. */
647 static int mbind_range(struct mm_struct *mm, unsigned long start,
648 unsigned long end, struct mempolicy *new_pol)
650 struct vm_area_struct *next;
651 struct vm_area_struct *prev;
652 struct vm_area_struct *vma;
655 unsigned long vmstart;
658 vma = find_vma_prev(mm, start, &prev);
659 if (!vma || vma->vm_start > start)
662 if (start > vma->vm_start)
665 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
667 vmstart = max(start, vma->vm_start);
668 vmend = min(end, vma->vm_end);
670 if (mpol_equal(vma_policy(vma), new_pol))
673 pgoff = vma->vm_pgoff +
674 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
675 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
676 vma->anon_vma, vma->vm_file, pgoff,
683 if (vma->vm_start != vmstart) {
684 err = split_vma(vma->vm_mm, vma, vmstart, 1);
688 if (vma->vm_end != vmend) {
689 err = split_vma(vma->vm_mm, vma, vmend, 0);
693 err = vma_replace_policy(vma, new_pol);
703 * Update task->flags PF_MEMPOLICY bit: set iff non-default
704 * mempolicy. Allows more rapid checking of this (combined perhaps
705 * with other PF_* flag bits) on memory allocation hot code paths.
707 * If called from outside this file, the task 'p' should -only- be
708 * a newly forked child not yet visible on the task list, because
709 * manipulating the task flags of a visible task is not safe.
711 * The above limitation is why this routine has the funny name
712 * mpol_fix_fork_child_flag().
714 * It is also safe to call this with a task pointer of current,
715 * which the static wrapper mpol_set_task_struct_flag() does,
716 * for use within this file.
719 void mpol_fix_fork_child_flag(struct task_struct *p)
722 p->flags |= PF_MEMPOLICY;
724 p->flags &= ~PF_MEMPOLICY;
727 static void mpol_set_task_struct_flag(void)
729 mpol_fix_fork_child_flag(current);
732 /* Set the process memory policy */
733 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
736 struct mempolicy *new, *old;
737 struct mm_struct *mm = current->mm;
738 NODEMASK_SCRATCH(scratch);
744 new = mpol_new(mode, flags, nodes);
750 * prevent changing our mempolicy while show_numa_maps()
752 * Note: do_set_mempolicy() can be called at init time
756 down_write(&mm->mmap_sem);
758 ret = mpol_set_nodemask(new, nodes, scratch);
760 task_unlock(current);
762 up_write(&mm->mmap_sem);
766 old = current->mempolicy;
767 current->mempolicy = new;
768 mpol_set_task_struct_flag();
769 if (new && new->mode == MPOL_INTERLEAVE &&
770 nodes_weight(new->v.nodes))
771 current->il_next = first_node(new->v.nodes);
772 task_unlock(current);
774 up_write(&mm->mmap_sem);
779 NODEMASK_SCRATCH_FREE(scratch);
784 * Return nodemask for policy for get_mempolicy() query
786 * Called with task's alloc_lock held
788 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
791 if (p == &default_policy)
797 case MPOL_INTERLEAVE:
801 if (!(p->flags & MPOL_F_LOCAL))
802 node_set(p->v.preferred_node, *nodes);
803 /* else return empty node mask for local allocation */
810 static int lookup_node(struct mm_struct *mm, unsigned long addr)
815 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
817 err = page_to_nid(p);
823 /* Retrieve NUMA policy */
824 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
825 unsigned long addr, unsigned long flags)
828 struct mm_struct *mm = current->mm;
829 struct vm_area_struct *vma = NULL;
830 struct mempolicy *pol = current->mempolicy;
833 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
836 if (flags & MPOL_F_MEMS_ALLOWED) {
837 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
839 *policy = 0; /* just so it's initialized */
841 *nmask = cpuset_current_mems_allowed;
842 task_unlock(current);
846 if (flags & MPOL_F_ADDR) {
848 * Do NOT fall back to task policy if the
849 * vma/shared policy at addr is NULL. We
850 * want to return MPOL_DEFAULT in this case.
852 down_read(&mm->mmap_sem);
853 vma = find_vma_intersection(mm, addr, addr+1);
855 up_read(&mm->mmap_sem);
858 if (vma->vm_ops && vma->vm_ops->get_policy)
859 pol = vma->vm_ops->get_policy(vma, addr);
861 pol = vma->vm_policy;
866 pol = &default_policy; /* indicates default behavior */
868 if (flags & MPOL_F_NODE) {
869 if (flags & MPOL_F_ADDR) {
870 err = lookup_node(mm, addr);
874 } else if (pol == current->mempolicy &&
875 pol->mode == MPOL_INTERLEAVE) {
876 *policy = current->il_next;
882 *policy = pol == &default_policy ? MPOL_DEFAULT :
885 * Internal mempolicy flags must be masked off before exposing
886 * the policy to userspace.
888 *policy |= (pol->flags & MPOL_MODE_FLAGS);
892 up_read(¤t->mm->mmap_sem);
898 if (mpol_store_user_nodemask(pol)) {
899 *nmask = pol->w.user_nodemask;
902 get_policy_nodemask(pol, nmask);
903 task_unlock(current);
910 up_read(¤t->mm->mmap_sem);
914 #ifdef CONFIG_MIGRATION
918 static void migrate_page_add(struct page *page, struct list_head *pagelist,
922 * Avoid migrating a page that is shared with others.
924 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
925 if (!isolate_lru_page(page)) {
926 list_add_tail(&page->lru, pagelist);
927 inc_zone_page_state(page, NR_ISOLATED_ANON +
928 page_is_file_cache(page));
933 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
935 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
939 * Migrate pages from one node to a target node.
940 * Returns error or the number of pages not migrated.
942 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
948 struct vm_area_struct *vma;
951 node_set(source, nmask);
953 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
954 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 assuming that page is mapped by vma pointed to by @private.
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_vma_page(struct page *page, unsigned long private, int **x)
1067 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1068 unsigned long uninitialized_var(address);
1071 address = page_address_in_vma(page, vma);
1072 if (address != -EFAULT)
1078 * if !vma, alloc_page_vma() will use task or system default policy
1080 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1084 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1085 unsigned long flags)
1089 int do_migrate_pages(struct mm_struct *mm,
1090 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1095 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1101 static long do_mbind(unsigned long start, unsigned long len,
1102 unsigned short mode, unsigned short mode_flags,
1103 nodemask_t *nmask, unsigned long flags)
1105 struct vm_area_struct *vma;
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 vma = check_range(mm, start, end, nmask,
1170 flags | MPOL_MF_INVERT, &pagelist);
1176 err = mbind_range(mm, start, end, new);
1178 if (!list_empty(&pagelist)) {
1179 nr_failed = migrate_pages(&pagelist, new_vma_page,
1183 putback_lru_pages(&pagelist);
1186 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1189 putback_lru_pages(&pagelist);
1191 up_write(&mm->mmap_sem);
1198 * User space interface with variable sized bitmaps for nodelists.
1201 /* Copy a node mask from user space. */
1202 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1203 unsigned long maxnode)
1206 unsigned long nlongs;
1207 unsigned long endmask;
1210 nodes_clear(*nodes);
1211 if (maxnode == 0 || !nmask)
1213 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1216 nlongs = BITS_TO_LONGS(maxnode);
1217 if ((maxnode % BITS_PER_LONG) == 0)
1220 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1222 /* When the user specified more nodes than supported just check
1223 if the non supported part is all zero. */
1224 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1225 if (nlongs > PAGE_SIZE/sizeof(long))
1227 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1229 if (get_user(t, nmask + k))
1231 if (k == nlongs - 1) {
1237 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1241 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1243 nodes_addr(*nodes)[nlongs-1] &= endmask;
1247 /* Copy a kernel node mask to user space */
1248 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1251 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1252 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1254 if (copy > nbytes) {
1255 if (copy > PAGE_SIZE)
1257 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1261 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1264 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1265 unsigned long, mode, unsigned long __user *, nmask,
1266 unsigned long, maxnode, unsigned, flags)
1270 unsigned short mode_flags;
1272 mode_flags = mode & MPOL_MODE_FLAGS;
1273 mode &= ~MPOL_MODE_FLAGS;
1274 if (mode >= MPOL_MAX)
1276 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1277 (mode_flags & MPOL_F_RELATIVE_NODES))
1279 err = get_nodes(&nodes, nmask, maxnode);
1282 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1285 /* Set the process memory policy */
1286 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1287 unsigned long, maxnode)
1291 unsigned short flags;
1293 flags = mode & MPOL_MODE_FLAGS;
1294 mode &= ~MPOL_MODE_FLAGS;
1295 if ((unsigned int)mode >= MPOL_MAX)
1297 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1299 err = get_nodes(&nodes, nmask, maxnode);
1302 return do_set_mempolicy(mode, flags, &nodes);
1305 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1306 const unsigned long __user *, old_nodes,
1307 const unsigned long __user *, new_nodes)
1309 const struct cred *cred = current_cred(), *tcred;
1310 struct mm_struct *mm = NULL;
1311 struct task_struct *task;
1312 nodemask_t task_nodes;
1316 NODEMASK_SCRATCH(scratch);
1321 old = &scratch->mask1;
1322 new = &scratch->mask2;
1324 err = get_nodes(old, old_nodes, maxnode);
1328 err = get_nodes(new, new_nodes, maxnode);
1332 /* Find the mm_struct */
1334 task = pid ? find_task_by_vpid(pid) : current;
1340 mm = get_task_mm(task);
1348 * Check if this process has the right to modify the specified
1349 * process. The right exists if the process has administrative
1350 * capabilities, superuser privileges or the same
1351 * userid as the target process.
1354 tcred = __task_cred(task);
1355 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1356 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1357 !capable(CAP_SYS_NICE)) {
1364 task_nodes = cpuset_mems_allowed(task);
1365 /* Is the user allowed to access the target nodes? */
1366 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1371 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1376 err = security_task_movememory(task);
1380 err = do_migrate_pages(mm, old, new,
1381 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1385 NODEMASK_SCRATCH_FREE(scratch);
1391 /* Retrieve NUMA policy */
1392 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1393 unsigned long __user *, nmask, unsigned long, maxnode,
1394 unsigned long, addr, unsigned long, flags)
1397 int uninitialized_var(pval);
1400 if (nmask != NULL && maxnode < MAX_NUMNODES)
1403 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1408 if (policy && put_user(pval, policy))
1412 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1417 #ifdef CONFIG_COMPAT
1419 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1420 compat_ulong_t __user *nmask,
1421 compat_ulong_t maxnode,
1422 compat_ulong_t addr, compat_ulong_t flags)
1425 unsigned long __user *nm = NULL;
1426 unsigned long nr_bits, alloc_size;
1427 DECLARE_BITMAP(bm, MAX_NUMNODES);
1429 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1430 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1433 nm = compat_alloc_user_space(alloc_size);
1435 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1437 if (!err && nmask) {
1438 unsigned long copy_size;
1439 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1440 err = copy_from_user(bm, nm, copy_size);
1441 /* ensure entire bitmap is zeroed */
1442 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1443 err |= compat_put_bitmap(nmask, bm, nr_bits);
1449 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1450 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 err = compat_get_bitmap(bm, nmask, nr_bits);
1462 nm = compat_alloc_user_space(alloc_size);
1463 err |= copy_to_user(nm, bm, alloc_size);
1469 return sys_set_mempolicy(mode, nm, nr_bits+1);
1472 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1473 compat_ulong_t mode, compat_ulong_t __user *nmask,
1474 compat_ulong_t maxnode, compat_ulong_t flags)
1477 unsigned long __user *nm = NULL;
1478 unsigned long nr_bits, alloc_size;
1481 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1482 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1485 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1486 nm = compat_alloc_user_space(alloc_size);
1487 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1493 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1499 * get_vma_policy(@task, @vma, @addr)
1500 * @task - task for fallback if vma policy == default
1501 * @vma - virtual memory area whose policy is sought
1502 * @addr - address in @vma for shared policy lookup
1504 * Returns effective policy for a VMA at specified address.
1505 * Falls back to @task or system default policy, as necessary.
1506 * Current or other task's task mempolicy and non-shared vma policies
1507 * are protected by the task's mmap_sem, which must be held for read by
1509 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1510 * count--added by the get_policy() vm_op, as appropriate--to protect against
1511 * freeing by another task. It is the caller's responsibility to free the
1512 * extra reference for shared policies.
1514 struct mempolicy *get_vma_policy(struct task_struct *task,
1515 struct vm_area_struct *vma, unsigned long addr)
1517 struct mempolicy *pol = task->mempolicy;
1520 if (vma->vm_ops && vma->vm_ops->get_policy) {
1521 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1525 } else if (vma->vm_policy) {
1526 pol = vma->vm_policy;
1529 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1530 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1531 * count on these policies which will be dropped by
1532 * mpol_cond_put() later
1534 if (mpol_needs_cond_ref(pol))
1539 pol = &default_policy;
1544 * Return a nodemask representing a mempolicy for filtering nodes for
1547 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1549 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1550 if (unlikely(policy->mode == MPOL_BIND) &&
1551 gfp_zone(gfp) >= policy_zone &&
1552 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1553 return &policy->v.nodes;
1558 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1559 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1562 switch (policy->mode) {
1563 case MPOL_PREFERRED:
1564 if (!(policy->flags & MPOL_F_LOCAL))
1565 nd = policy->v.preferred_node;
1569 * Normally, MPOL_BIND allocations are node-local within the
1570 * allowed nodemask. However, if __GFP_THISNODE is set and the
1571 * current node isn't part of the mask, we use the zonelist for
1572 * the first node in the mask instead.
1574 if (unlikely(gfp & __GFP_THISNODE) &&
1575 unlikely(!node_isset(nd, policy->v.nodes)))
1576 nd = first_node(policy->v.nodes);
1581 return node_zonelist(nd, gfp);
1584 /* Do dynamic interleaving for a process */
1585 static unsigned interleave_nodes(struct mempolicy *policy)
1588 struct task_struct *me = current;
1591 next = next_node(nid, policy->v.nodes);
1592 if (next >= MAX_NUMNODES)
1593 next = first_node(policy->v.nodes);
1594 if (next < MAX_NUMNODES)
1600 * Depending on the memory policy provide a node from which to allocate the
1602 * @policy must be protected by freeing by the caller. If @policy is
1603 * the current task's mempolicy, this protection is implicit, as only the
1604 * task can change it's policy. The system default policy requires no
1607 unsigned slab_node(struct mempolicy *policy)
1609 if (!policy || policy->flags & MPOL_F_LOCAL)
1610 return numa_node_id();
1612 switch (policy->mode) {
1613 case MPOL_PREFERRED:
1615 * handled MPOL_F_LOCAL above
1617 return policy->v.preferred_node;
1619 case MPOL_INTERLEAVE:
1620 return interleave_nodes(policy);
1624 * Follow bind policy behavior and start allocation at the
1627 struct zonelist *zonelist;
1629 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1630 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1631 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1634 return zone ? zone->node : numa_node_id();
1642 /* Do static interleaving for a VMA with known offset. */
1643 static unsigned offset_il_node(struct mempolicy *pol,
1644 struct vm_area_struct *vma, unsigned long off)
1646 unsigned nnodes = nodes_weight(pol->v.nodes);
1652 return numa_node_id();
1653 target = (unsigned int)off % nnodes;
1656 nid = next_node(nid, pol->v.nodes);
1658 } while (c <= target);
1662 /* Determine a node number for interleave */
1663 static inline unsigned interleave_nid(struct mempolicy *pol,
1664 struct vm_area_struct *vma, unsigned long addr, int shift)
1670 * for small pages, there is no difference between
1671 * shift and PAGE_SHIFT, so the bit-shift is safe.
1672 * for huge pages, since vm_pgoff is in units of small
1673 * pages, we need to shift off the always 0 bits to get
1676 BUG_ON(shift < PAGE_SHIFT);
1677 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1678 off += (addr - vma->vm_start) >> shift;
1679 return offset_il_node(pol, vma, off);
1681 return interleave_nodes(pol);
1685 * Return the bit number of a random bit set in the nodemask.
1686 * (returns -1 if nodemask is empty)
1688 int node_random(const nodemask_t *maskp)
1692 w = nodes_weight(*maskp);
1694 bit = bitmap_ord_to_pos(maskp->bits,
1695 get_random_int() % w, MAX_NUMNODES);
1699 #ifdef CONFIG_HUGETLBFS
1701 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1702 * @vma = virtual memory area whose policy is sought
1703 * @addr = address in @vma for shared policy lookup and interleave policy
1704 * @gfp_flags = for requested zone
1705 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1706 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1708 * Returns a zonelist suitable for a huge page allocation and a pointer
1709 * to the struct mempolicy for conditional unref after allocation.
1710 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1711 * @nodemask for filtering the zonelist.
1713 * Must be protected by get_mems_allowed()
1715 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1716 gfp_t gfp_flags, struct mempolicy **mpol,
1717 nodemask_t **nodemask)
1719 struct zonelist *zl;
1721 *mpol = get_vma_policy(current, vma, addr);
1722 *nodemask = NULL; /* assume !MPOL_BIND */
1724 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1725 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1726 huge_page_shift(hstate_vma(vma))), gfp_flags);
1728 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1729 if ((*mpol)->mode == MPOL_BIND)
1730 *nodemask = &(*mpol)->v.nodes;
1736 * init_nodemask_of_mempolicy
1738 * If the current task's mempolicy is "default" [NULL], return 'false'
1739 * to indicate default policy. Otherwise, extract the policy nodemask
1740 * for 'bind' or 'interleave' policy into the argument nodemask, or
1741 * initialize the argument nodemask to contain the single node for
1742 * 'preferred' or 'local' policy and return 'true' to indicate presence
1743 * of non-default mempolicy.
1745 * We don't bother with reference counting the mempolicy [mpol_get/put]
1746 * because the current task is examining it's own mempolicy and a task's
1747 * mempolicy is only ever changed by the task itself.
1749 * N.B., it is the caller's responsibility to free a returned nodemask.
1751 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1753 struct mempolicy *mempolicy;
1756 if (!(mask && current->mempolicy))
1760 mempolicy = current->mempolicy;
1761 switch (mempolicy->mode) {
1762 case MPOL_PREFERRED:
1763 if (mempolicy->flags & MPOL_F_LOCAL)
1764 nid = numa_node_id();
1766 nid = mempolicy->v.preferred_node;
1767 init_nodemask_of_node(mask, nid);
1772 case MPOL_INTERLEAVE:
1773 *mask = mempolicy->v.nodes;
1779 task_unlock(current);
1786 * mempolicy_nodemask_intersects
1788 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1789 * policy. Otherwise, check for intersection between mask and the policy
1790 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1791 * policy, always return true since it may allocate elsewhere on fallback.
1793 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1795 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1796 const nodemask_t *mask)
1798 struct mempolicy *mempolicy;
1804 mempolicy = tsk->mempolicy;
1808 switch (mempolicy->mode) {
1809 case MPOL_PREFERRED:
1811 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1812 * allocate from, they may fallback to other nodes when oom.
1813 * Thus, it's possible for tsk to have allocated memory from
1818 case MPOL_INTERLEAVE:
1819 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1829 /* Allocate a page in interleaved policy.
1830 Own path because it needs to do special accounting. */
1831 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1834 struct zonelist *zl;
1837 zl = node_zonelist(nid, gfp);
1838 page = __alloc_pages(gfp, order, zl);
1839 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1840 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1845 * alloc_pages_vma - Allocate a page for a VMA.
1848 * %GFP_USER user allocation.
1849 * %GFP_KERNEL kernel allocations,
1850 * %GFP_HIGHMEM highmem/user allocations,
1851 * %GFP_FS allocation should not call back into a file system.
1852 * %GFP_ATOMIC don't sleep.
1854 * @order:Order of the GFP allocation.
1855 * @vma: Pointer to VMA or NULL if not available.
1856 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1858 * This function allocates a page from the kernel page pool and applies
1859 * a NUMA policy associated with the VMA or the current process.
1860 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1861 * mm_struct of the VMA to prevent it from going away. Should be used for
1862 * all allocations for pages that will be mapped into
1863 * user space. Returns NULL when no page can be allocated.
1865 * Should be called with the mm_sem of the vma hold.
1868 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1869 unsigned long addr, int node)
1871 struct mempolicy *pol;
1872 struct zonelist *zl;
1874 unsigned int cpuset_mems_cookie;
1877 pol = get_vma_policy(current, vma, addr);
1878 cpuset_mems_cookie = get_mems_allowed();
1880 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1883 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1885 page = alloc_page_interleave(gfp, order, nid);
1886 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1891 zl = policy_zonelist(gfp, pol, node);
1892 if (unlikely(mpol_needs_cond_ref(pol))) {
1894 * slow path: ref counted shared policy
1896 struct page *page = __alloc_pages_nodemask(gfp, order,
1897 zl, policy_nodemask(gfp, pol));
1899 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1904 * fast path: default or task policy
1906 page = __alloc_pages_nodemask(gfp, order, zl,
1907 policy_nodemask(gfp, pol));
1908 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1914 * alloc_pages_current - Allocate pages.
1917 * %GFP_USER user allocation,
1918 * %GFP_KERNEL kernel allocation,
1919 * %GFP_HIGHMEM highmem allocation,
1920 * %GFP_FS don't call back into a file system.
1921 * %GFP_ATOMIC don't sleep.
1922 * @order: Power of two of allocation size in pages. 0 is a single page.
1924 * Allocate a page from the kernel page pool. When not in
1925 * interrupt context and apply the current process NUMA policy.
1926 * Returns NULL when no page can be allocated.
1928 * Don't call cpuset_update_task_memory_state() unless
1929 * 1) it's ok to take cpuset_sem (can WAIT), and
1930 * 2) allocating for current task (not interrupt).
1932 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1934 struct mempolicy *pol = current->mempolicy;
1936 unsigned int cpuset_mems_cookie;
1938 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1939 pol = &default_policy;
1942 cpuset_mems_cookie = get_mems_allowed();
1945 * No reference counting needed for current->mempolicy
1946 * nor system default_policy
1948 if (pol->mode == MPOL_INTERLEAVE)
1949 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1951 page = __alloc_pages_nodemask(gfp, order,
1952 policy_zonelist(gfp, pol, numa_node_id()),
1953 policy_nodemask(gfp, pol));
1955 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1960 EXPORT_SYMBOL(alloc_pages_current);
1963 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1964 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1965 * with the mems_allowed returned by cpuset_mems_allowed(). This
1966 * keeps mempolicies cpuset relative after its cpuset moves. See
1967 * further kernel/cpuset.c update_nodemask().
1969 * current's mempolicy may be rebinded by the other task(the task that changes
1970 * cpuset's mems), so we needn't do rebind work for current task.
1973 /* Slow path of a mempolicy duplicate */
1974 struct mempolicy *__mpol_dup(struct mempolicy *old)
1976 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1979 return ERR_PTR(-ENOMEM);
1981 /* task's mempolicy is protected by alloc_lock */
1982 if (old == current->mempolicy) {
1985 task_unlock(current);
1990 if (current_cpuset_is_being_rebound()) {
1991 nodemask_t mems = cpuset_mems_allowed(current);
1992 if (new->flags & MPOL_F_REBINDING)
1993 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1995 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1998 atomic_set(&new->refcnt, 1);
2003 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2004 * eliminate the * MPOL_F_* flags that require conditional ref and
2005 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2006 * after return. Use the returned value.
2008 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2009 * policy lookup, even if the policy needs/has extra ref on lookup.
2010 * shmem_readahead needs this.
2012 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2013 struct mempolicy *frompol)
2015 if (!mpol_needs_cond_ref(frompol))
2019 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2020 __mpol_put(frompol);
2024 /* Slow path of a mempolicy comparison */
2025 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2029 if (a->mode != b->mode)
2031 if (a->flags != b->flags)
2033 if (mpol_store_user_nodemask(a))
2034 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2040 case MPOL_INTERLEAVE:
2041 return nodes_equal(a->v.nodes, b->v.nodes);
2042 case MPOL_PREFERRED:
2043 return a->v.preferred_node == b->v.preferred_node;
2051 * Shared memory backing store policy support.
2053 * Remember policies even when nobody has shared memory mapped.
2054 * The policies are kept in Red-Black tree linked from the inode.
2055 * They are protected by the sp->lock spinlock, which should be held
2056 * for any accesses to the tree.
2059 /* lookup first element intersecting start-end */
2060 /* Caller holds sp->mutex */
2061 static struct sp_node *
2062 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2064 struct rb_node *n = sp->root.rb_node;
2067 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2069 if (start >= p->end)
2071 else if (end <= p->start)
2079 struct sp_node *w = NULL;
2080 struct rb_node *prev = rb_prev(n);
2083 w = rb_entry(prev, struct sp_node, nd);
2084 if (w->end <= start)
2088 return rb_entry(n, struct sp_node, nd);
2091 /* Insert a new shared policy into the list. */
2092 /* Caller holds sp->lock */
2093 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2095 struct rb_node **p = &sp->root.rb_node;
2096 struct rb_node *parent = NULL;
2101 nd = rb_entry(parent, struct sp_node, nd);
2102 if (new->start < nd->start)
2104 else if (new->end > nd->end)
2105 p = &(*p)->rb_right;
2109 rb_link_node(&new->nd, parent, p);
2110 rb_insert_color(&new->nd, &sp->root);
2111 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2112 new->policy ? new->policy->mode : 0);
2115 /* Find shared policy intersecting idx */
2117 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2119 struct mempolicy *pol = NULL;
2122 if (!sp->root.rb_node)
2124 mutex_lock(&sp->mutex);
2125 sn = sp_lookup(sp, idx, idx+1);
2127 mpol_get(sn->policy);
2130 mutex_unlock(&sp->mutex);
2134 static void sp_free(struct sp_node *n)
2136 mpol_put(n->policy);
2137 kmem_cache_free(sn_cache, n);
2140 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2142 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2143 rb_erase(&n->nd, &sp->root);
2147 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2148 struct mempolicy *pol)
2151 struct mempolicy *newpol;
2153 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2157 newpol = mpol_dup(pol);
2158 if (IS_ERR(newpol)) {
2159 kmem_cache_free(sn_cache, n);
2162 newpol->flags |= MPOL_F_SHARED;
2171 /* Replace a policy range. */
2172 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2173 unsigned long end, struct sp_node *new)
2178 mutex_lock(&sp->mutex);
2179 n = sp_lookup(sp, start, end);
2180 /* Take care of old policies in the same range. */
2181 while (n && n->start < end) {
2182 struct rb_node *next = rb_next(&n->nd);
2183 if (n->start >= start) {
2189 /* Old policy spanning whole new range. */
2191 struct sp_node *new2;
2192 new2 = sp_alloc(end, n->end, n->policy);
2198 sp_insert(sp, new2);
2205 n = rb_entry(next, struct sp_node, nd);
2210 mutex_unlock(&sp->mutex);
2215 * mpol_shared_policy_init - initialize shared policy for inode
2216 * @sp: pointer to inode shared policy
2217 * @mpol: struct mempolicy to install
2219 * Install non-NULL @mpol in inode's shared policy rb-tree.
2220 * On entry, the current task has a reference on a non-NULL @mpol.
2221 * This must be released on exit.
2222 * This is called at get_inode() calls and we can use GFP_KERNEL.
2224 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2228 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2229 mutex_init(&sp->mutex);
2232 struct vm_area_struct pvma;
2233 struct mempolicy *new;
2234 NODEMASK_SCRATCH(scratch);
2238 /* contextualize the tmpfs mount point mempolicy */
2239 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2241 goto free_scratch; /* no valid nodemask intersection */
2244 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2245 task_unlock(current);
2249 /* Create pseudo-vma that contains just the policy */
2250 memset(&pvma, 0, sizeof(struct vm_area_struct));
2251 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2252 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2255 mpol_put(new); /* drop initial ref */
2257 NODEMASK_SCRATCH_FREE(scratch);
2259 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2263 int mpol_set_shared_policy(struct shared_policy *info,
2264 struct vm_area_struct *vma, struct mempolicy *npol)
2267 struct sp_node *new = NULL;
2268 unsigned long sz = vma_pages(vma);
2270 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2272 sz, npol ? npol->mode : -1,
2273 npol ? npol->flags : -1,
2274 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2277 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2281 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2287 /* Free a backing policy store on inode delete. */
2288 void mpol_free_shared_policy(struct shared_policy *p)
2291 struct rb_node *next;
2293 if (!p->root.rb_node)
2295 mutex_lock(&p->mutex);
2296 next = rb_first(&p->root);
2298 n = rb_entry(next, struct sp_node, nd);
2299 next = rb_next(&n->nd);
2302 mutex_unlock(&p->mutex);
2305 /* assumes fs == KERNEL_DS */
2306 void __init numa_policy_init(void)
2308 nodemask_t interleave_nodes;
2309 unsigned long largest = 0;
2310 int nid, prefer = 0;
2312 policy_cache = kmem_cache_create("numa_policy",
2313 sizeof(struct mempolicy),
2314 0, SLAB_PANIC, NULL);
2316 sn_cache = kmem_cache_create("shared_policy_node",
2317 sizeof(struct sp_node),
2318 0, SLAB_PANIC, NULL);
2321 * Set interleaving policy for system init. Interleaving is only
2322 * enabled across suitably sized nodes (default is >= 16MB), or
2323 * fall back to the largest node if they're all smaller.
2325 nodes_clear(interleave_nodes);
2326 for_each_node_state(nid, N_HIGH_MEMORY) {
2327 unsigned long total_pages = node_present_pages(nid);
2329 /* Preserve the largest node */
2330 if (largest < total_pages) {
2331 largest = total_pages;
2335 /* Interleave this node? */
2336 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2337 node_set(nid, interleave_nodes);
2340 /* All too small, use the largest */
2341 if (unlikely(nodes_empty(interleave_nodes)))
2342 node_set(prefer, interleave_nodes);
2344 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2345 printk("numa_policy_init: interleaving failed\n");
2348 /* Reset policy of current process to default */
2349 void numa_default_policy(void)
2351 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2355 * Parse and format mempolicy from/to strings
2359 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2360 * Used only for mpol_parse_str() and mpol_to_str()
2362 #define MPOL_LOCAL MPOL_MAX
2363 static const char * const policy_modes[] =
2365 [MPOL_DEFAULT] = "default",
2366 [MPOL_PREFERRED] = "prefer",
2367 [MPOL_BIND] = "bind",
2368 [MPOL_INTERLEAVE] = "interleave",
2369 [MPOL_LOCAL] = "local"
2375 * mpol_parse_str - parse string to mempolicy
2376 * @str: string containing mempolicy to parse
2377 * @mpol: pointer to struct mempolicy pointer, returned on success.
2378 * @no_context: flag whether to "contextualize" the mempolicy
2381 * <mode>[=<flags>][:<nodelist>]
2383 * if @no_context is true, save the input nodemask in w.user_nodemask in
2384 * the returned mempolicy. This will be used to "clone" the mempolicy in
2385 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2386 * mount option. Note that if 'static' or 'relative' mode flags were
2387 * specified, the input nodemask will already have been saved. Saving
2388 * it again is redundant, but safe.
2390 * On success, returns 0, else 1
2392 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2394 struct mempolicy *new = NULL;
2395 unsigned short mode;
2396 unsigned short uninitialized_var(mode_flags);
2398 char *nodelist = strchr(str, ':');
2399 char *flags = strchr(str, '=');
2403 /* NUL-terminate mode or flags string */
2405 if (nodelist_parse(nodelist, nodes))
2407 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2413 *flags++ = '\0'; /* terminate mode string */
2415 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2416 if (!strcmp(str, policy_modes[mode])) {
2420 if (mode > MPOL_LOCAL)
2424 case MPOL_PREFERRED:
2426 * Insist on a nodelist of one node only
2429 char *rest = nodelist;
2430 while (isdigit(*rest))
2436 case MPOL_INTERLEAVE:
2438 * Default to online nodes with memory if no nodelist
2441 nodes = node_states[N_HIGH_MEMORY];
2445 * Don't allow a nodelist; mpol_new() checks flags
2449 mode = MPOL_PREFERRED;
2453 * Insist on a empty nodelist
2460 * Insist on a nodelist
2469 * Currently, we only support two mutually exclusive
2472 if (!strcmp(flags, "static"))
2473 mode_flags |= MPOL_F_STATIC_NODES;
2474 else if (!strcmp(flags, "relative"))
2475 mode_flags |= MPOL_F_RELATIVE_NODES;
2480 new = mpol_new(mode, mode_flags, &nodes);
2485 /* save for contextualization */
2486 new->w.user_nodemask = nodes;
2489 NODEMASK_SCRATCH(scratch);
2492 ret = mpol_set_nodemask(new, &nodes, scratch);
2493 task_unlock(current);
2496 NODEMASK_SCRATCH_FREE(scratch);
2505 /* Restore string for error message */
2514 #endif /* CONFIG_TMPFS */
2517 * mpol_to_str - format a mempolicy structure for printing
2518 * @buffer: to contain formatted mempolicy string
2519 * @maxlen: length of @buffer
2520 * @pol: pointer to mempolicy to be formatted
2521 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2523 * Convert a mempolicy into a string.
2524 * Returns the number of characters in buffer (if positive)
2525 * or an error (negative)
2527 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2532 unsigned short mode;
2533 unsigned short flags = pol ? pol->flags : 0;
2536 * Sanity check: room for longest mode, flag and some nodes
2538 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2540 if (!pol || pol == &default_policy)
2541 mode = MPOL_DEFAULT;
2550 case MPOL_PREFERRED:
2552 if (flags & MPOL_F_LOCAL)
2553 mode = MPOL_LOCAL; /* pseudo-policy */
2555 node_set(pol->v.preferred_node, nodes);
2560 case MPOL_INTERLEAVE:
2562 nodes = pol->w.user_nodemask;
2564 nodes = pol->v.nodes;
2571 l = strlen(policy_modes[mode]);
2572 if (buffer + maxlen < p + l + 1)
2575 strcpy(p, policy_modes[mode]);
2578 if (flags & MPOL_MODE_FLAGS) {
2579 if (buffer + maxlen < p + 2)
2584 * Currently, the only defined flags are mutually exclusive
2586 if (flags & MPOL_F_STATIC_NODES)
2587 p += snprintf(p, buffer + maxlen - p, "static");
2588 else if (flags & MPOL_F_RELATIVE_NODES)
2589 p += snprintf(p, buffer + maxlen - p, "relative");
2592 if (!nodes_empty(nodes)) {
2593 if (buffer + maxlen < p + 2)
2596 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);