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 == MPOL_REBIND_ONCE &&
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);
610 /* Apply policy to a single VMA */
611 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
614 struct mempolicy *old = vma->vm_policy;
616 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
617 vma->vm_start, vma->vm_end, vma->vm_pgoff,
618 vma->vm_ops, vma->vm_file,
619 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
621 if (vma->vm_ops && vma->vm_ops->set_policy)
622 err = vma->vm_ops->set_policy(vma, new);
625 vma->vm_policy = new;
631 /* Step 2: apply policy to a range and do splits. */
632 static int mbind_range(struct mm_struct *mm, unsigned long start,
633 unsigned long end, struct mempolicy *new_pol)
635 struct vm_area_struct *next;
636 struct vm_area_struct *prev;
637 struct vm_area_struct *vma;
640 unsigned long vmstart;
643 vma = find_vma(mm, start);
644 if (!vma || vma->vm_start > start)
648 if (start > vma->vm_start)
651 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
653 vmstart = max(start, vma->vm_start);
654 vmend = min(end, vma->vm_end);
656 if (mpol_equal(vma_policy(vma), new_pol))
659 pgoff = vma->vm_pgoff +
660 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
661 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
662 vma->anon_vma, vma->vm_file, pgoff,
669 if (vma->vm_start != vmstart) {
670 err = split_vma(vma->vm_mm, vma, vmstart, 1);
674 if (vma->vm_end != vmend) {
675 err = split_vma(vma->vm_mm, vma, vmend, 0);
679 err = policy_vma(vma, new_pol);
689 * Update task->flags PF_MEMPOLICY bit: set iff non-default
690 * mempolicy. Allows more rapid checking of this (combined perhaps
691 * with other PF_* flag bits) on memory allocation hot code paths.
693 * If called from outside this file, the task 'p' should -only- be
694 * a newly forked child not yet visible on the task list, because
695 * manipulating the task flags of a visible task is not safe.
697 * The above limitation is why this routine has the funny name
698 * mpol_fix_fork_child_flag().
700 * It is also safe to call this with a task pointer of current,
701 * which the static wrapper mpol_set_task_struct_flag() does,
702 * for use within this file.
705 void mpol_fix_fork_child_flag(struct task_struct *p)
708 p->flags |= PF_MEMPOLICY;
710 p->flags &= ~PF_MEMPOLICY;
713 static void mpol_set_task_struct_flag(void)
715 mpol_fix_fork_child_flag(current);
718 /* Set the process memory policy */
719 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
722 struct mempolicy *new, *old;
723 struct mm_struct *mm = current->mm;
724 NODEMASK_SCRATCH(scratch);
730 new = mpol_new(mode, flags, nodes);
736 * prevent changing our mempolicy while show_numa_maps()
738 * Note: do_set_mempolicy() can be called at init time
742 down_write(&mm->mmap_sem);
744 ret = mpol_set_nodemask(new, nodes, scratch);
746 task_unlock(current);
748 up_write(&mm->mmap_sem);
752 old = current->mempolicy;
753 current->mempolicy = new;
754 mpol_set_task_struct_flag();
755 if (new && new->mode == MPOL_INTERLEAVE &&
756 nodes_weight(new->v.nodes))
757 current->il_next = first_node(new->v.nodes);
758 task_unlock(current);
760 up_write(&mm->mmap_sem);
765 NODEMASK_SCRATCH_FREE(scratch);
770 * Return nodemask for policy for get_mempolicy() query
772 * Called with task's alloc_lock held
774 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
777 if (p == &default_policy)
783 case MPOL_INTERLEAVE:
787 if (!(p->flags & MPOL_F_LOCAL))
788 node_set(p->v.preferred_node, *nodes);
789 /* else return empty node mask for local allocation */
796 static int lookup_node(struct mm_struct *mm, unsigned long addr)
801 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
803 err = page_to_nid(p);
809 /* Retrieve NUMA policy */
810 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
811 unsigned long addr, unsigned long flags)
814 struct mm_struct *mm = current->mm;
815 struct vm_area_struct *vma = NULL;
816 struct mempolicy *pol = current->mempolicy;
819 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
822 if (flags & MPOL_F_MEMS_ALLOWED) {
823 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
825 *policy = 0; /* just so it's initialized */
827 *nmask = cpuset_current_mems_allowed;
828 task_unlock(current);
832 if (flags & MPOL_F_ADDR) {
834 * Do NOT fall back to task policy if the
835 * vma/shared policy at addr is NULL. We
836 * want to return MPOL_DEFAULT in this case.
838 down_read(&mm->mmap_sem);
839 vma = find_vma_intersection(mm, addr, addr+1);
841 up_read(&mm->mmap_sem);
844 if (vma->vm_ops && vma->vm_ops->get_policy)
845 pol = vma->vm_ops->get_policy(vma, addr);
847 pol = vma->vm_policy;
852 pol = &default_policy; /* indicates default behavior */
854 if (flags & MPOL_F_NODE) {
855 if (flags & MPOL_F_ADDR) {
856 err = lookup_node(mm, addr);
860 } else if (pol == current->mempolicy &&
861 pol->mode == MPOL_INTERLEAVE) {
862 *policy = current->il_next;
868 *policy = pol == &default_policy ? MPOL_DEFAULT :
871 * Internal mempolicy flags must be masked off before exposing
872 * the policy to userspace.
874 *policy |= (pol->flags & MPOL_MODE_FLAGS);
878 up_read(¤t->mm->mmap_sem);
884 if (mpol_store_user_nodemask(pol)) {
885 *nmask = pol->w.user_nodemask;
888 get_policy_nodemask(pol, nmask);
889 task_unlock(current);
896 up_read(¤t->mm->mmap_sem);
900 #ifdef CONFIG_MIGRATION
904 static void migrate_page_add(struct page *page, struct list_head *pagelist,
908 * Avoid migrating a page that is shared with others.
910 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
911 if (!isolate_lru_page(page)) {
912 list_add_tail(&page->lru, pagelist);
913 inc_zone_page_state(page, NR_ISOLATED_ANON +
914 page_is_file_cache(page));
919 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
921 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
925 * Migrate pages from one node to a target node.
926 * Returns error or the number of pages not migrated.
928 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
934 struct vm_area_struct *vma;
937 node_set(source, nmask);
939 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
940 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
944 if (!list_empty(&pagelist)) {
945 err = migrate_pages(&pagelist, new_node_page, dest,
946 false, MIGRATE_SYNC);
948 putback_lru_pages(&pagelist);
955 * Move pages between the two nodesets so as to preserve the physical
956 * layout as much as possible.
958 * Returns the number of page that could not be moved.
960 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
961 const nodemask_t *to, int flags)
967 err = migrate_prep();
971 down_read(&mm->mmap_sem);
973 err = migrate_vmas(mm, from, to, flags);
978 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
979 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
980 * bit in 'tmp', and return that <source, dest> pair for migration.
981 * The pair of nodemasks 'to' and 'from' define the map.
983 * If no pair of bits is found that way, fallback to picking some
984 * pair of 'source' and 'dest' bits that are not the same. If the
985 * 'source' and 'dest' bits are the same, this represents a node
986 * that will be migrating to itself, so no pages need move.
988 * If no bits are left in 'tmp', or if all remaining bits left
989 * in 'tmp' correspond to the same bit in 'to', return false
990 * (nothing left to migrate).
992 * This lets us pick a pair of nodes to migrate between, such that
993 * if possible the dest node is not already occupied by some other
994 * source node, minimizing the risk of overloading the memory on a
995 * node that would happen if we migrated incoming memory to a node
996 * before migrating outgoing memory source that same node.
998 * A single scan of tmp is sufficient. As we go, we remember the
999 * most recent <s, d> pair that moved (s != d). If we find a pair
1000 * that not only moved, but what's better, moved to an empty slot
1001 * (d is not set in tmp), then we break out then, with that pair.
1002 * Otherwise when we finish scanning from_tmp, we at least have the
1003 * most recent <s, d> pair that moved. If we get all the way through
1004 * the scan of tmp without finding any node that moved, much less
1005 * moved to an empty node, then there is nothing left worth migrating.
1009 while (!nodes_empty(tmp)) {
1014 for_each_node_mask(s, tmp) {
1017 * do_migrate_pages() tries to maintain the relative
1018 * node relationship of the pages established between
1019 * threads and memory areas.
1021 * However if the number of source nodes is not equal to
1022 * the number of destination nodes we can not preserve
1023 * this node relative relationship. In that case, skip
1024 * copying memory from a node that is in the destination
1027 * Example: [2,3,4] -> [3,4,5] moves everything.
1028 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1031 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1032 (node_isset(s, *to)))
1035 d = node_remap(s, *from, *to);
1039 source = s; /* Node moved. Memorize */
1042 /* dest not in remaining from nodes? */
1043 if (!node_isset(dest, tmp))
1049 node_clear(source, tmp);
1050 err = migrate_to_node(mm, source, dest, flags);
1057 up_read(&mm->mmap_sem);
1065 * Allocate a new page for page migration based on vma policy.
1066 * Start assuming that page is mapped by vma pointed to by @private.
1067 * Search forward from there, if not. N.B., this assumes that the
1068 * list of pages handed to migrate_pages()--which is how we get here--
1069 * is in virtual address order.
1071 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1073 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1074 unsigned long uninitialized_var(address);
1077 address = page_address_in_vma(page, vma);
1078 if (address != -EFAULT)
1084 * if !vma, alloc_page_vma() will use task or system default policy
1086 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1090 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1091 unsigned long flags)
1095 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1096 const nodemask_t *to, int flags)
1101 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1107 static long do_mbind(unsigned long start, unsigned long len,
1108 unsigned short mode, unsigned short mode_flags,
1109 nodemask_t *nmask, unsigned long flags)
1111 struct vm_area_struct *vma;
1112 struct mm_struct *mm = current->mm;
1113 struct mempolicy *new;
1116 LIST_HEAD(pagelist);
1118 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1119 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1121 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1124 if (start & ~PAGE_MASK)
1127 if (mode == MPOL_DEFAULT)
1128 flags &= ~MPOL_MF_STRICT;
1130 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1138 new = mpol_new(mode, mode_flags, nmask);
1140 return PTR_ERR(new);
1143 * If we are using the default policy then operation
1144 * on discontinuous address spaces is okay after all
1147 flags |= MPOL_MF_DISCONTIG_OK;
1149 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1150 start, start + len, mode, mode_flags,
1151 nmask ? nodes_addr(*nmask)[0] : -1);
1153 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1155 err = migrate_prep();
1160 NODEMASK_SCRATCH(scratch);
1162 down_write(&mm->mmap_sem);
1164 err = mpol_set_nodemask(new, nmask, scratch);
1165 task_unlock(current);
1167 up_write(&mm->mmap_sem);
1170 NODEMASK_SCRATCH_FREE(scratch);
1175 vma = check_range(mm, start, end, nmask,
1176 flags | MPOL_MF_INVERT, &pagelist);
1182 err = mbind_range(mm, start, end, new);
1184 if (!list_empty(&pagelist)) {
1185 nr_failed = migrate_pages(&pagelist, new_vma_page,
1187 false, MIGRATE_SYNC);
1189 putback_lru_pages(&pagelist);
1192 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1195 putback_lru_pages(&pagelist);
1197 up_write(&mm->mmap_sem);
1204 * User space interface with variable sized bitmaps for nodelists.
1207 /* Copy a node mask from user space. */
1208 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1209 unsigned long maxnode)
1212 unsigned long nlongs;
1213 unsigned long endmask;
1216 nodes_clear(*nodes);
1217 if (maxnode == 0 || !nmask)
1219 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1222 nlongs = BITS_TO_LONGS(maxnode);
1223 if ((maxnode % BITS_PER_LONG) == 0)
1226 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1228 /* When the user specified more nodes than supported just check
1229 if the non supported part is all zero. */
1230 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1231 if (nlongs > PAGE_SIZE/sizeof(long))
1233 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1235 if (get_user(t, nmask + k))
1237 if (k == nlongs - 1) {
1243 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1247 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1249 nodes_addr(*nodes)[nlongs-1] &= endmask;
1253 /* Copy a kernel node mask to user space */
1254 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1257 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1258 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1260 if (copy > nbytes) {
1261 if (copy > PAGE_SIZE)
1263 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1267 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1270 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1271 unsigned long, mode, unsigned long __user *, nmask,
1272 unsigned long, maxnode, unsigned, flags)
1276 unsigned short mode_flags;
1278 mode_flags = mode & MPOL_MODE_FLAGS;
1279 mode &= ~MPOL_MODE_FLAGS;
1280 if (mode >= MPOL_MAX)
1282 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1283 (mode_flags & MPOL_F_RELATIVE_NODES))
1285 err = get_nodes(&nodes, nmask, maxnode);
1288 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1291 /* Set the process memory policy */
1292 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1293 unsigned long, maxnode)
1297 unsigned short flags;
1299 flags = mode & MPOL_MODE_FLAGS;
1300 mode &= ~MPOL_MODE_FLAGS;
1301 if ((unsigned int)mode >= MPOL_MAX)
1303 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1305 err = get_nodes(&nodes, nmask, maxnode);
1308 return do_set_mempolicy(mode, flags, &nodes);
1311 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1312 const unsigned long __user *, old_nodes,
1313 const unsigned long __user *, new_nodes)
1315 const struct cred *cred = current_cred(), *tcred;
1316 struct mm_struct *mm = NULL;
1317 struct task_struct *task;
1318 nodemask_t task_nodes;
1322 NODEMASK_SCRATCH(scratch);
1327 old = &scratch->mask1;
1328 new = &scratch->mask2;
1330 err = get_nodes(old, old_nodes, maxnode);
1334 err = get_nodes(new, new_nodes, maxnode);
1338 /* Find the mm_struct */
1340 task = pid ? find_task_by_vpid(pid) : current;
1346 get_task_struct(task);
1351 * Check if this process has the right to modify the specified
1352 * process. The right exists if the process has administrative
1353 * capabilities, superuser privileges or the same
1354 * userid as the target process.
1356 tcred = __task_cred(task);
1357 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1358 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1359 !capable(CAP_SYS_NICE)) {
1366 task_nodes = cpuset_mems_allowed(task);
1367 /* Is the user allowed to access the target nodes? */
1368 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1373 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1378 err = security_task_movememory(task);
1382 mm = get_task_mm(task);
1383 put_task_struct(task);
1390 err = do_migrate_pages(mm, old, new,
1391 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1395 NODEMASK_SCRATCH_FREE(scratch);
1400 put_task_struct(task);
1406 /* Retrieve NUMA policy */
1407 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1408 unsigned long __user *, nmask, unsigned long, maxnode,
1409 unsigned long, addr, unsigned long, flags)
1412 int uninitialized_var(pval);
1415 if (nmask != NULL && maxnode < MAX_NUMNODES)
1418 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1423 if (policy && put_user(pval, policy))
1427 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1432 #ifdef CONFIG_COMPAT
1434 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1435 compat_ulong_t __user *nmask,
1436 compat_ulong_t maxnode,
1437 compat_ulong_t addr, compat_ulong_t flags)
1440 unsigned long __user *nm = NULL;
1441 unsigned long nr_bits, alloc_size;
1442 DECLARE_BITMAP(bm, MAX_NUMNODES);
1444 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1445 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1448 nm = compat_alloc_user_space(alloc_size);
1450 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1452 if (!err && nmask) {
1453 unsigned long copy_size;
1454 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1455 err = copy_from_user(bm, nm, copy_size);
1456 /* ensure entire bitmap is zeroed */
1457 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1458 err |= compat_put_bitmap(nmask, bm, nr_bits);
1464 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1465 compat_ulong_t maxnode)
1468 unsigned long __user *nm = NULL;
1469 unsigned long nr_bits, alloc_size;
1470 DECLARE_BITMAP(bm, MAX_NUMNODES);
1472 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1473 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1476 err = compat_get_bitmap(bm, nmask, nr_bits);
1477 nm = compat_alloc_user_space(alloc_size);
1478 err |= copy_to_user(nm, bm, alloc_size);
1484 return sys_set_mempolicy(mode, nm, nr_bits+1);
1487 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1488 compat_ulong_t mode, compat_ulong_t __user *nmask,
1489 compat_ulong_t maxnode, compat_ulong_t flags)
1492 unsigned long __user *nm = NULL;
1493 unsigned long nr_bits, alloc_size;
1496 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1497 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1500 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1501 nm = compat_alloc_user_space(alloc_size);
1502 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1508 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1514 * get_vma_policy(@task, @vma, @addr)
1515 * @task - task for fallback if vma policy == default
1516 * @vma - virtual memory area whose policy is sought
1517 * @addr - address in @vma for shared policy lookup
1519 * Returns effective policy for a VMA at specified address.
1520 * Falls back to @task or system default policy, as necessary.
1521 * Current or other task's task mempolicy and non-shared vma policies
1522 * are protected by the task's mmap_sem, which must be held for read by
1524 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1525 * count--added by the get_policy() vm_op, as appropriate--to protect against
1526 * freeing by another task. It is the caller's responsibility to free the
1527 * extra reference for shared policies.
1529 struct mempolicy *get_vma_policy(struct task_struct *task,
1530 struct vm_area_struct *vma, unsigned long addr)
1532 struct mempolicy *pol = task->mempolicy;
1535 if (vma->vm_ops && vma->vm_ops->get_policy) {
1536 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1540 } else if (vma->vm_policy)
1541 pol = vma->vm_policy;
1544 pol = &default_policy;
1549 * Return a nodemask representing a mempolicy for filtering nodes for
1552 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1554 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1555 if (unlikely(policy->mode == MPOL_BIND) &&
1556 gfp_zone(gfp) >= policy_zone &&
1557 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1558 return &policy->v.nodes;
1563 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1564 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1567 switch (policy->mode) {
1568 case MPOL_PREFERRED:
1569 if (!(policy->flags & MPOL_F_LOCAL))
1570 nd = policy->v.preferred_node;
1574 * Normally, MPOL_BIND allocations are node-local within the
1575 * allowed nodemask. However, if __GFP_THISNODE is set and the
1576 * current node isn't part of the mask, we use the zonelist for
1577 * the first node in the mask instead.
1579 if (unlikely(gfp & __GFP_THISNODE) &&
1580 unlikely(!node_isset(nd, policy->v.nodes)))
1581 nd = first_node(policy->v.nodes);
1586 return node_zonelist(nd, gfp);
1589 /* Do dynamic interleaving for a process */
1590 static unsigned interleave_nodes(struct mempolicy *policy)
1593 struct task_struct *me = current;
1596 next = next_node(nid, policy->v.nodes);
1597 if (next >= MAX_NUMNODES)
1598 next = first_node(policy->v.nodes);
1599 if (next < MAX_NUMNODES)
1605 * Depending on the memory policy provide a node from which to allocate the
1607 * @policy must be protected by freeing by the caller. If @policy is
1608 * the current task's mempolicy, this protection is implicit, as only the
1609 * task can change it's policy. The system default policy requires no
1612 unsigned slab_node(void)
1614 struct mempolicy *policy;
1617 return numa_node_id();
1619 policy = current->mempolicy;
1620 if (!policy || policy->flags & MPOL_F_LOCAL)
1621 return numa_node_id();
1623 switch (policy->mode) {
1624 case MPOL_PREFERRED:
1626 * handled MPOL_F_LOCAL above
1628 return policy->v.preferred_node;
1630 case MPOL_INTERLEAVE:
1631 return interleave_nodes(policy);
1635 * Follow bind policy behavior and start allocation at the
1638 struct zonelist *zonelist;
1640 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1641 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1642 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1645 return zone ? zone->node : numa_node_id();
1653 /* Do static interleaving for a VMA with known offset. */
1654 static unsigned offset_il_node(struct mempolicy *pol,
1655 struct vm_area_struct *vma, unsigned long off)
1657 unsigned nnodes = nodes_weight(pol->v.nodes);
1663 return numa_node_id();
1664 target = (unsigned int)off % nnodes;
1667 nid = next_node(nid, pol->v.nodes);
1669 } while (c <= target);
1673 /* Determine a node number for interleave */
1674 static inline unsigned interleave_nid(struct mempolicy *pol,
1675 struct vm_area_struct *vma, unsigned long addr, int shift)
1681 * for small pages, there is no difference between
1682 * shift and PAGE_SHIFT, so the bit-shift is safe.
1683 * for huge pages, since vm_pgoff is in units of small
1684 * pages, we need to shift off the always 0 bits to get
1687 BUG_ON(shift < PAGE_SHIFT);
1688 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1689 off += (addr - vma->vm_start) >> shift;
1690 return offset_il_node(pol, vma, off);
1692 return interleave_nodes(pol);
1696 * Return the bit number of a random bit set in the nodemask.
1697 * (returns -1 if nodemask is empty)
1699 int node_random(const nodemask_t *maskp)
1703 w = nodes_weight(*maskp);
1705 bit = bitmap_ord_to_pos(maskp->bits,
1706 get_random_int() % w, MAX_NUMNODES);
1710 #ifdef CONFIG_HUGETLBFS
1712 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1713 * @vma = virtual memory area whose policy is sought
1714 * @addr = address in @vma for shared policy lookup and interleave policy
1715 * @gfp_flags = for requested zone
1716 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1717 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1719 * Returns a zonelist suitable for a huge page allocation and a pointer
1720 * to the struct mempolicy for conditional unref after allocation.
1721 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1722 * @nodemask for filtering the zonelist.
1724 * Must be protected by get_mems_allowed()
1726 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1727 gfp_t gfp_flags, struct mempolicy **mpol,
1728 nodemask_t **nodemask)
1730 struct zonelist *zl;
1732 *mpol = get_vma_policy(current, vma, addr);
1733 *nodemask = NULL; /* assume !MPOL_BIND */
1735 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1736 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1737 huge_page_shift(hstate_vma(vma))), gfp_flags);
1739 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1740 if ((*mpol)->mode == MPOL_BIND)
1741 *nodemask = &(*mpol)->v.nodes;
1747 * init_nodemask_of_mempolicy
1749 * If the current task's mempolicy is "default" [NULL], return 'false'
1750 * to indicate default policy. Otherwise, extract the policy nodemask
1751 * for 'bind' or 'interleave' policy into the argument nodemask, or
1752 * initialize the argument nodemask to contain the single node for
1753 * 'preferred' or 'local' policy and return 'true' to indicate presence
1754 * of non-default mempolicy.
1756 * We don't bother with reference counting the mempolicy [mpol_get/put]
1757 * because the current task is examining it's own mempolicy and a task's
1758 * mempolicy is only ever changed by the task itself.
1760 * N.B., it is the caller's responsibility to free a returned nodemask.
1762 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1764 struct mempolicy *mempolicy;
1767 if (!(mask && current->mempolicy))
1771 mempolicy = current->mempolicy;
1772 switch (mempolicy->mode) {
1773 case MPOL_PREFERRED:
1774 if (mempolicy->flags & MPOL_F_LOCAL)
1775 nid = numa_node_id();
1777 nid = mempolicy->v.preferred_node;
1778 init_nodemask_of_node(mask, nid);
1783 case MPOL_INTERLEAVE:
1784 *mask = mempolicy->v.nodes;
1790 task_unlock(current);
1797 * mempolicy_nodemask_intersects
1799 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1800 * policy. Otherwise, check for intersection between mask and the policy
1801 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1802 * policy, always return true since it may allocate elsewhere on fallback.
1804 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1806 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1807 const nodemask_t *mask)
1809 struct mempolicy *mempolicy;
1815 mempolicy = tsk->mempolicy;
1819 switch (mempolicy->mode) {
1820 case MPOL_PREFERRED:
1822 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1823 * allocate from, they may fallback to other nodes when oom.
1824 * Thus, it's possible for tsk to have allocated memory from
1829 case MPOL_INTERLEAVE:
1830 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1840 /* Allocate a page in interleaved policy.
1841 Own path because it needs to do special accounting. */
1842 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1845 struct zonelist *zl;
1848 zl = node_zonelist(nid, gfp);
1849 page = __alloc_pages(gfp, order, zl);
1850 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1851 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1856 * alloc_pages_vma - Allocate a page for a VMA.
1859 * %GFP_USER user allocation.
1860 * %GFP_KERNEL kernel allocations,
1861 * %GFP_HIGHMEM highmem/user allocations,
1862 * %GFP_FS allocation should not call back into a file system.
1863 * %GFP_ATOMIC don't sleep.
1865 * @order:Order of the GFP allocation.
1866 * @vma: Pointer to VMA or NULL if not available.
1867 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1869 * This function allocates a page from the kernel page pool and applies
1870 * a NUMA policy associated with the VMA or the current process.
1871 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1872 * mm_struct of the VMA to prevent it from going away. Should be used for
1873 * all allocations for pages that will be mapped into
1874 * user space. Returns NULL when no page can be allocated.
1876 * Should be called with the mm_sem of the vma hold.
1879 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1880 unsigned long addr, int node)
1882 struct mempolicy *pol;
1883 struct zonelist *zl;
1885 unsigned int cpuset_mems_cookie;
1888 pol = get_vma_policy(current, vma, addr);
1889 cpuset_mems_cookie = get_mems_allowed();
1891 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1894 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1896 page = alloc_page_interleave(gfp, order, nid);
1897 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1902 zl = policy_zonelist(gfp, pol, node);
1903 if (unlikely(mpol_needs_cond_ref(pol))) {
1905 * slow path: ref counted shared policy
1907 struct page *page = __alloc_pages_nodemask(gfp, order,
1908 zl, policy_nodemask(gfp, pol));
1910 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1915 * fast path: default or task policy
1917 page = __alloc_pages_nodemask(gfp, order, zl,
1918 policy_nodemask(gfp, pol));
1919 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1925 * alloc_pages_current - Allocate pages.
1928 * %GFP_USER user allocation,
1929 * %GFP_KERNEL kernel allocation,
1930 * %GFP_HIGHMEM highmem allocation,
1931 * %GFP_FS don't call back into a file system.
1932 * %GFP_ATOMIC don't sleep.
1933 * @order: Power of two of allocation size in pages. 0 is a single page.
1935 * Allocate a page from the kernel page pool. When not in
1936 * interrupt context and apply the current process NUMA policy.
1937 * Returns NULL when no page can be allocated.
1939 * Don't call cpuset_update_task_memory_state() unless
1940 * 1) it's ok to take cpuset_sem (can WAIT), and
1941 * 2) allocating for current task (not interrupt).
1943 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1945 struct mempolicy *pol = current->mempolicy;
1947 unsigned int cpuset_mems_cookie;
1949 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1950 pol = &default_policy;
1953 cpuset_mems_cookie = get_mems_allowed();
1956 * No reference counting needed for current->mempolicy
1957 * nor system default_policy
1959 if (pol->mode == MPOL_INTERLEAVE)
1960 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1962 page = __alloc_pages_nodemask(gfp, order,
1963 policy_zonelist(gfp, pol, numa_node_id()),
1964 policy_nodemask(gfp, pol));
1966 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1971 EXPORT_SYMBOL(alloc_pages_current);
1974 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1975 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1976 * with the mems_allowed returned by cpuset_mems_allowed(). This
1977 * keeps mempolicies cpuset relative after its cpuset moves. See
1978 * further kernel/cpuset.c update_nodemask().
1980 * current's mempolicy may be rebinded by the other task(the task that changes
1981 * cpuset's mems), so we needn't do rebind work for current task.
1984 /* Slow path of a mempolicy duplicate */
1985 struct mempolicy *__mpol_dup(struct mempolicy *old)
1987 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1990 return ERR_PTR(-ENOMEM);
1992 /* task's mempolicy is protected by alloc_lock */
1993 if (old == current->mempolicy) {
1996 task_unlock(current);
2001 if (current_cpuset_is_being_rebound()) {
2002 nodemask_t mems = cpuset_mems_allowed(current);
2003 if (new->flags & MPOL_F_REBINDING)
2004 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2006 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2009 atomic_set(&new->refcnt, 1);
2014 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2015 * eliminate the * MPOL_F_* flags that require conditional ref and
2016 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2017 * after return. Use the returned value.
2019 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2020 * policy lookup, even if the policy needs/has extra ref on lookup.
2021 * shmem_readahead needs this.
2023 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2024 struct mempolicy *frompol)
2026 if (!mpol_needs_cond_ref(frompol))
2030 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2031 __mpol_put(frompol);
2035 /* Slow path of a mempolicy comparison */
2036 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2040 if (a->mode != b->mode)
2042 if (a->flags != b->flags)
2044 if (mpol_store_user_nodemask(a))
2045 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2051 case MPOL_INTERLEAVE:
2052 return !!nodes_equal(a->v.nodes, b->v.nodes);
2053 case MPOL_PREFERRED:
2054 return a->v.preferred_node == b->v.preferred_node;
2062 * Shared memory backing store policy support.
2064 * Remember policies even when nobody has shared memory mapped.
2065 * The policies are kept in Red-Black tree linked from the inode.
2066 * They are protected by the sp->lock spinlock, which should be held
2067 * for any accesses to the tree.
2070 /* lookup first element intersecting start-end */
2071 /* Caller holds sp->lock */
2072 static struct sp_node *
2073 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2075 struct rb_node *n = sp->root.rb_node;
2078 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2080 if (start >= p->end)
2082 else if (end <= p->start)
2090 struct sp_node *w = NULL;
2091 struct rb_node *prev = rb_prev(n);
2094 w = rb_entry(prev, struct sp_node, nd);
2095 if (w->end <= start)
2099 return rb_entry(n, struct sp_node, nd);
2102 /* Insert a new shared policy into the list. */
2103 /* Caller holds sp->lock */
2104 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2106 struct rb_node **p = &sp->root.rb_node;
2107 struct rb_node *parent = NULL;
2112 nd = rb_entry(parent, struct sp_node, nd);
2113 if (new->start < nd->start)
2115 else if (new->end > nd->end)
2116 p = &(*p)->rb_right;
2120 rb_link_node(&new->nd, parent, p);
2121 rb_insert_color(&new->nd, &sp->root);
2122 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2123 new->policy ? new->policy->mode : 0);
2126 /* Find shared policy intersecting idx */
2128 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2130 struct mempolicy *pol = NULL;
2133 if (!sp->root.rb_node)
2135 spin_lock(&sp->lock);
2136 sn = sp_lookup(sp, idx, idx+1);
2138 mpol_get(sn->policy);
2141 spin_unlock(&sp->lock);
2145 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2147 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2148 rb_erase(&n->nd, &sp->root);
2149 mpol_put(n->policy);
2150 kmem_cache_free(sn_cache, n);
2153 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2154 struct mempolicy *pol)
2156 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2163 pol->flags |= MPOL_F_SHARED; /* for unref */
2168 /* Replace a policy range. */
2169 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2170 unsigned long end, struct sp_node *new)
2172 struct sp_node *n, *new2 = NULL;
2175 spin_lock(&sp->lock);
2176 n = sp_lookup(sp, start, end);
2177 /* Take care of old policies in the same range. */
2178 while (n && n->start < end) {
2179 struct rb_node *next = rb_next(&n->nd);
2180 if (n->start >= start) {
2186 /* Old policy spanning whole new range. */
2189 spin_unlock(&sp->lock);
2190 new2 = sp_alloc(end, n->end, n->policy);
2196 sp_insert(sp, new2);
2204 n = rb_entry(next, struct sp_node, nd);
2208 spin_unlock(&sp->lock);
2210 mpol_put(new2->policy);
2211 kmem_cache_free(sn_cache, new2);
2217 * mpol_shared_policy_init - initialize shared policy for inode
2218 * @sp: pointer to inode shared policy
2219 * @mpol: struct mempolicy to install
2221 * Install non-NULL @mpol in inode's shared policy rb-tree.
2222 * On entry, the current task has a reference on a non-NULL @mpol.
2223 * This must be released on exit.
2224 * This is called at get_inode() calls and we can use GFP_KERNEL.
2226 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2230 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2231 spin_lock_init(&sp->lock);
2234 struct vm_area_struct pvma;
2235 struct mempolicy *new;
2236 NODEMASK_SCRATCH(scratch);
2240 /* contextualize the tmpfs mount point mempolicy */
2241 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2243 goto free_scratch; /* no valid nodemask intersection */
2246 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2247 task_unlock(current);
2251 /* Create pseudo-vma that contains just the policy */
2252 memset(&pvma, 0, sizeof(struct vm_area_struct));
2253 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2254 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2257 mpol_put(new); /* drop initial ref */
2259 NODEMASK_SCRATCH_FREE(scratch);
2261 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2265 int mpol_set_shared_policy(struct shared_policy *info,
2266 struct vm_area_struct *vma, struct mempolicy *npol)
2269 struct sp_node *new = NULL;
2270 unsigned long sz = vma_pages(vma);
2272 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2274 sz, npol ? npol->mode : -1,
2275 npol ? npol->flags : -1,
2276 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2279 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2283 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2285 kmem_cache_free(sn_cache, new);
2289 /* Free a backing policy store on inode delete. */
2290 void mpol_free_shared_policy(struct shared_policy *p)
2293 struct rb_node *next;
2295 if (!p->root.rb_node)
2297 spin_lock(&p->lock);
2298 next = rb_first(&p->root);
2300 n = rb_entry(next, struct sp_node, nd);
2301 next = rb_next(&n->nd);
2302 rb_erase(&n->nd, &p->root);
2303 mpol_put(n->policy);
2304 kmem_cache_free(sn_cache, n);
2306 spin_unlock(&p->lock);
2309 /* assumes fs == KERNEL_DS */
2310 void __init numa_policy_init(void)
2312 nodemask_t interleave_nodes;
2313 unsigned long largest = 0;
2314 int nid, prefer = 0;
2316 policy_cache = kmem_cache_create("numa_policy",
2317 sizeof(struct mempolicy),
2318 0, SLAB_PANIC, NULL);
2320 sn_cache = kmem_cache_create("shared_policy_node",
2321 sizeof(struct sp_node),
2322 0, SLAB_PANIC, NULL);
2325 * Set interleaving policy for system init. Interleaving is only
2326 * enabled across suitably sized nodes (default is >= 16MB), or
2327 * fall back to the largest node if they're all smaller.
2329 nodes_clear(interleave_nodes);
2330 for_each_node_state(nid, N_HIGH_MEMORY) {
2331 unsigned long total_pages = node_present_pages(nid);
2333 /* Preserve the largest node */
2334 if (largest < total_pages) {
2335 largest = total_pages;
2339 /* Interleave this node? */
2340 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2341 node_set(nid, interleave_nodes);
2344 /* All too small, use the largest */
2345 if (unlikely(nodes_empty(interleave_nodes)))
2346 node_set(prefer, interleave_nodes);
2348 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2349 printk("numa_policy_init: interleaving failed\n");
2352 /* Reset policy of current process to default */
2353 void numa_default_policy(void)
2355 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2359 * Parse and format mempolicy from/to strings
2363 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2364 * Used only for mpol_parse_str() and mpol_to_str()
2366 #define MPOL_LOCAL MPOL_MAX
2367 static const char * const policy_modes[] =
2369 [MPOL_DEFAULT] = "default",
2370 [MPOL_PREFERRED] = "prefer",
2371 [MPOL_BIND] = "bind",
2372 [MPOL_INTERLEAVE] = "interleave",
2373 [MPOL_LOCAL] = "local"
2379 * mpol_parse_str - parse string to mempolicy
2380 * @str: string containing mempolicy to parse
2381 * @mpol: pointer to struct mempolicy pointer, returned on success.
2382 * @no_context: flag whether to "contextualize" the mempolicy
2385 * <mode>[=<flags>][:<nodelist>]
2387 * if @no_context is true, save the input nodemask in w.user_nodemask in
2388 * the returned mempolicy. This will be used to "clone" the mempolicy in
2389 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2390 * mount option. Note that if 'static' or 'relative' mode flags were
2391 * specified, the input nodemask will already have been saved. Saving
2392 * it again is redundant, but safe.
2394 * On success, returns 0, else 1
2396 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2398 struct mempolicy *new = NULL;
2399 unsigned short mode;
2400 unsigned short uninitialized_var(mode_flags);
2402 char *nodelist = strchr(str, ':');
2403 char *flags = strchr(str, '=');
2407 /* NUL-terminate mode or flags string */
2409 if (nodelist_parse(nodelist, nodes))
2411 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2417 *flags++ = '\0'; /* terminate mode string */
2419 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2420 if (!strcmp(str, policy_modes[mode])) {
2424 if (mode > MPOL_LOCAL)
2428 case MPOL_PREFERRED:
2430 * Insist on a nodelist of one node only
2433 char *rest = nodelist;
2434 while (isdigit(*rest))
2440 case MPOL_INTERLEAVE:
2442 * Default to online nodes with memory if no nodelist
2445 nodes = node_states[N_HIGH_MEMORY];
2449 * Don't allow a nodelist; mpol_new() checks flags
2453 mode = MPOL_PREFERRED;
2457 * Insist on a empty nodelist
2464 * Insist on a nodelist
2473 * Currently, we only support two mutually exclusive
2476 if (!strcmp(flags, "static"))
2477 mode_flags |= MPOL_F_STATIC_NODES;
2478 else if (!strcmp(flags, "relative"))
2479 mode_flags |= MPOL_F_RELATIVE_NODES;
2484 new = mpol_new(mode, mode_flags, &nodes);
2489 /* save for contextualization */
2490 new->w.user_nodemask = nodes;
2493 NODEMASK_SCRATCH(scratch);
2496 ret = mpol_set_nodemask(new, &nodes, scratch);
2497 task_unlock(current);
2500 NODEMASK_SCRATCH_FREE(scratch);
2509 /* Restore string for error message */
2518 #endif /* CONFIG_TMPFS */
2521 * mpol_to_str - format a mempolicy structure for printing
2522 * @buffer: to contain formatted mempolicy string
2523 * @maxlen: length of @buffer
2524 * @pol: pointer to mempolicy to be formatted
2525 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2527 * Convert a mempolicy into a string.
2528 * Returns the number of characters in buffer (if positive)
2529 * or an error (negative)
2531 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2536 unsigned short mode;
2537 unsigned short flags = pol ? pol->flags : 0;
2540 * Sanity check: room for longest mode, flag and some nodes
2542 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2544 if (!pol || pol == &default_policy)
2545 mode = MPOL_DEFAULT;
2554 case MPOL_PREFERRED:
2556 if (flags & MPOL_F_LOCAL)
2557 mode = MPOL_LOCAL; /* pseudo-policy */
2559 node_set(pol->v.preferred_node, nodes);
2564 case MPOL_INTERLEAVE:
2566 nodes = pol->w.user_nodemask;
2568 nodes = pol->v.nodes;
2575 l = strlen(policy_modes[mode]);
2576 if (buffer + maxlen < p + l + 1)
2579 strcpy(p, policy_modes[mode]);
2582 if (flags & MPOL_MODE_FLAGS) {
2583 if (buffer + maxlen < p + 2)
2588 * Currently, the only defined flags are mutually exclusive
2590 if (flags & MPOL_F_STATIC_NODES)
2591 p += snprintf(p, buffer + maxlen - p, "static");
2592 else if (flags & MPOL_F_RELATIVE_NODES)
2593 p += snprintf(p, buffer + maxlen - p, "relative");
2596 if (!nodes_empty(nodes)) {
2597 if (buffer + maxlen < p + 2)
2600 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);