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 || mode == MPOL_NOOP) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
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 (mode == MPOL_LOCAL) {
273 if (!nodes_empty(*nodes))
274 return ERR_PTR(-EINVAL);
275 mode = MPOL_PREFERRED;
276 } else if (nodes_empty(*nodes))
277 return ERR_PTR(-EINVAL);
278 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
280 return ERR_PTR(-ENOMEM);
281 atomic_set(&policy->refcnt, 1);
283 policy->flags = flags;
288 /* Slow path of a mpol destructor. */
289 void __mpol_put(struct mempolicy *p)
291 if (!atomic_dec_and_test(&p->refcnt))
293 kmem_cache_free(policy_cache, p);
296 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
297 enum mpol_rebind_step step)
303 * MPOL_REBIND_ONCE - do rebind work at once
304 * MPOL_REBIND_STEP1 - set all the newly nodes
305 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
307 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
308 enum mpol_rebind_step step)
312 if (pol->flags & MPOL_F_STATIC_NODES)
313 nodes_and(tmp, pol->w.user_nodemask, *nodes);
314 else if (pol->flags & MPOL_F_RELATIVE_NODES)
315 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
318 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
321 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
322 nodes_remap(tmp, pol->v.nodes,
323 pol->w.cpuset_mems_allowed, *nodes);
324 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
325 } else if (step == MPOL_REBIND_STEP2) {
326 tmp = pol->w.cpuset_mems_allowed;
327 pol->w.cpuset_mems_allowed = *nodes;
332 if (nodes_empty(tmp))
335 if (step == MPOL_REBIND_STEP1)
336 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
337 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
342 if (!node_isset(current->il_next, tmp)) {
343 current->il_next = next_node(current->il_next, tmp);
344 if (current->il_next >= MAX_NUMNODES)
345 current->il_next = first_node(tmp);
346 if (current->il_next >= MAX_NUMNODES)
347 current->il_next = numa_node_id();
351 static void mpol_rebind_preferred(struct mempolicy *pol,
352 const nodemask_t *nodes,
353 enum mpol_rebind_step step)
357 if (pol->flags & MPOL_F_STATIC_NODES) {
358 int node = first_node(pol->w.user_nodemask);
360 if (node_isset(node, *nodes)) {
361 pol->v.preferred_node = node;
362 pol->flags &= ~MPOL_F_LOCAL;
364 pol->flags |= MPOL_F_LOCAL;
365 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
366 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
367 pol->v.preferred_node = first_node(tmp);
368 } else if (!(pol->flags & MPOL_F_LOCAL)) {
369 pol->v.preferred_node = node_remap(pol->v.preferred_node,
370 pol->w.cpuset_mems_allowed,
372 pol->w.cpuset_mems_allowed = *nodes;
377 * mpol_rebind_policy - Migrate a policy to a different set of nodes
379 * If read-side task has no lock to protect task->mempolicy, write-side
380 * task will rebind the task->mempolicy by two step. The first step is
381 * setting all the newly nodes, and the second step is cleaning all the
382 * disallowed nodes. In this way, we can avoid finding no node to alloc
384 * If we have a lock to protect task->mempolicy in read-side, we do
388 * MPOL_REBIND_ONCE - do rebind work at once
389 * MPOL_REBIND_STEP1 - set all the newly nodes
390 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
392 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
393 enum mpol_rebind_step step)
397 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
398 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
401 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
404 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
407 if (step == MPOL_REBIND_STEP1)
408 pol->flags |= MPOL_F_REBINDING;
409 else if (step == MPOL_REBIND_STEP2)
410 pol->flags &= ~MPOL_F_REBINDING;
411 else if (step >= MPOL_REBIND_NSTEP)
414 mpol_ops[pol->mode].rebind(pol, newmask, step);
418 * Wrapper for mpol_rebind_policy() that just requires task
419 * pointer, and updates task mempolicy.
421 * Called with task's alloc_lock held.
424 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
425 enum mpol_rebind_step step)
427 mpol_rebind_policy(tsk->mempolicy, new, step);
431 * Rebind each vma in mm to new nodemask.
433 * Call holding a reference to mm. Takes mm->mmap_sem during call.
436 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
438 struct vm_area_struct *vma;
440 down_write(&mm->mmap_sem);
441 for (vma = mm->mmap; vma; vma = vma->vm_next)
442 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
443 up_write(&mm->mmap_sem);
446 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
448 .rebind = mpol_rebind_default,
450 [MPOL_INTERLEAVE] = {
451 .create = mpol_new_interleave,
452 .rebind = mpol_rebind_nodemask,
455 .create = mpol_new_preferred,
456 .rebind = mpol_rebind_preferred,
459 .create = mpol_new_bind,
460 .rebind = mpol_rebind_nodemask,
464 static void migrate_page_add(struct page *page, struct list_head *pagelist,
465 unsigned long flags);
467 /* Scan through pages checking if pages follow certain conditions. */
468 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
469 unsigned long addr, unsigned long end,
470 const nodemask_t *nodes, unsigned long flags,
477 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
482 if (!pte_present(*pte))
484 page = vm_normal_page(vma, addr, *pte);
488 * vm_normal_page() filters out zero pages, but there might
489 * still be PageReserved pages to skip, perhaps in a VDSO.
490 * And we cannot move PageKsm pages sensibly or safely yet.
492 if (PageReserved(page) || PageKsm(page))
494 nid = page_to_nid(page);
495 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
498 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
499 migrate_page_add(page, private, flags);
502 } while (pte++, addr += PAGE_SIZE, addr != end);
503 pte_unmap_unlock(orig_pte, ptl);
507 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
508 unsigned long addr, unsigned long end,
509 const nodemask_t *nodes, unsigned long flags,
515 pmd = pmd_offset(pud, addr);
517 next = pmd_addr_end(addr, end);
518 split_huge_page_pmd(vma->vm_mm, pmd);
519 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
521 if (check_pte_range(vma, pmd, addr, next, nodes,
524 } while (pmd++, addr = next, addr != end);
528 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
529 unsigned long addr, unsigned long end,
530 const nodemask_t *nodes, unsigned long flags,
536 pud = pud_offset(pgd, addr);
538 next = pud_addr_end(addr, end);
539 if (pud_none_or_clear_bad(pud))
541 if (check_pmd_range(vma, pud, addr, next, nodes,
544 } while (pud++, addr = next, addr != end);
548 static inline int check_pgd_range(struct vm_area_struct *vma,
549 unsigned long addr, unsigned long end,
550 const nodemask_t *nodes, unsigned long flags,
556 pgd = pgd_offset(vma->vm_mm, addr);
558 next = pgd_addr_end(addr, end);
559 if (pgd_none_or_clear_bad(pgd))
561 if (check_pud_range(vma, pgd, addr, next, nodes,
564 } while (pgd++, addr = next, addr != end);
569 * Check if all pages in a range are on a set of nodes.
570 * If pagelist != NULL then isolate pages from the LRU and
571 * put them on the pagelist.
573 static struct vm_area_struct *
574 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
575 const nodemask_t *nodes, unsigned long flags, void *private)
578 struct vm_area_struct *first, *vma, *prev;
581 first = find_vma(mm, start);
583 return ERR_PTR(-EFAULT);
585 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
586 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
587 if (!vma->vm_next && vma->vm_end < end)
588 return ERR_PTR(-EFAULT);
589 if (prev && prev->vm_end < vma->vm_start)
590 return ERR_PTR(-EFAULT);
592 if (!is_vm_hugetlb_page(vma) &&
593 ((flags & MPOL_MF_STRICT) ||
594 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
595 vma_migratable(vma)))) {
596 unsigned long endvma = vma->vm_end;
600 if (vma->vm_start > start)
601 start = vma->vm_start;
602 err = check_pgd_range(vma, start, endvma, nodes,
605 first = ERR_PTR(err);
615 * Apply policy to a single VMA
616 * This must be called with the mmap_sem held for writing.
618 static int vma_replace_policy(struct vm_area_struct *vma,
619 struct mempolicy *pol)
622 struct mempolicy *old;
623 struct mempolicy *new;
625 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
626 vma->vm_start, vma->vm_end, vma->vm_pgoff,
627 vma->vm_ops, vma->vm_file,
628 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
634 if (vma->vm_ops && vma->vm_ops->set_policy) {
635 err = vma->vm_ops->set_policy(vma, new);
640 old = vma->vm_policy;
641 vma->vm_policy = new; /* protected by mmap_sem */
650 /* Step 2: apply policy to a range and do splits. */
651 static int mbind_range(struct mm_struct *mm, unsigned long start,
652 unsigned long end, struct mempolicy *new_pol)
654 struct vm_area_struct *next;
655 struct vm_area_struct *prev;
656 struct vm_area_struct *vma;
659 unsigned long vmstart;
662 vma = find_vma(mm, start);
663 if (!vma || vma->vm_start > start)
667 if (start > vma->vm_start)
670 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
672 vmstart = max(start, vma->vm_start);
673 vmend = min(end, vma->vm_end);
675 if (mpol_equal(vma_policy(vma), new_pol))
678 pgoff = vma->vm_pgoff +
679 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
680 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
681 vma->anon_vma, vma->vm_file, pgoff,
688 if (vma->vm_start != vmstart) {
689 err = split_vma(vma->vm_mm, vma, vmstart, 1);
693 if (vma->vm_end != vmend) {
694 err = split_vma(vma->vm_mm, vma, vmend, 0);
698 err = vma_replace_policy(vma, new_pol);
708 * Update task->flags PF_MEMPOLICY bit: set iff non-default
709 * mempolicy. Allows more rapid checking of this (combined perhaps
710 * with other PF_* flag bits) on memory allocation hot code paths.
712 * If called from outside this file, the task 'p' should -only- be
713 * a newly forked child not yet visible on the task list, because
714 * manipulating the task flags of a visible task is not safe.
716 * The above limitation is why this routine has the funny name
717 * mpol_fix_fork_child_flag().
719 * It is also safe to call this with a task pointer of current,
720 * which the static wrapper mpol_set_task_struct_flag() does,
721 * for use within this file.
724 void mpol_fix_fork_child_flag(struct task_struct *p)
727 p->flags |= PF_MEMPOLICY;
729 p->flags &= ~PF_MEMPOLICY;
732 static void mpol_set_task_struct_flag(void)
734 mpol_fix_fork_child_flag(current);
737 /* Set the process memory policy */
738 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
741 struct mempolicy *new, *old;
742 struct mm_struct *mm = current->mm;
743 NODEMASK_SCRATCH(scratch);
749 new = mpol_new(mode, flags, nodes);
755 * prevent changing our mempolicy while show_numa_maps()
757 * Note: do_set_mempolicy() can be called at init time
761 down_write(&mm->mmap_sem);
763 ret = mpol_set_nodemask(new, nodes, scratch);
765 task_unlock(current);
767 up_write(&mm->mmap_sem);
771 old = current->mempolicy;
772 current->mempolicy = new;
773 mpol_set_task_struct_flag();
774 if (new && new->mode == MPOL_INTERLEAVE &&
775 nodes_weight(new->v.nodes))
776 current->il_next = first_node(new->v.nodes);
777 task_unlock(current);
779 up_write(&mm->mmap_sem);
784 NODEMASK_SCRATCH_FREE(scratch);
789 * Return nodemask for policy for get_mempolicy() query
791 * Called with task's alloc_lock held
793 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
796 if (p == &default_policy)
802 case MPOL_INTERLEAVE:
806 if (!(p->flags & MPOL_F_LOCAL))
807 node_set(p->v.preferred_node, *nodes);
808 /* else return empty node mask for local allocation */
815 static int lookup_node(struct mm_struct *mm, unsigned long addr)
820 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
822 err = page_to_nid(p);
828 /* Retrieve NUMA policy */
829 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
830 unsigned long addr, unsigned long flags)
833 struct mm_struct *mm = current->mm;
834 struct vm_area_struct *vma = NULL;
835 struct mempolicy *pol = current->mempolicy;
838 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
841 if (flags & MPOL_F_MEMS_ALLOWED) {
842 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
844 *policy = 0; /* just so it's initialized */
846 *nmask = cpuset_current_mems_allowed;
847 task_unlock(current);
851 if (flags & MPOL_F_ADDR) {
853 * Do NOT fall back to task policy if the
854 * vma/shared policy at addr is NULL. We
855 * want to return MPOL_DEFAULT in this case.
857 down_read(&mm->mmap_sem);
858 vma = find_vma_intersection(mm, addr, addr+1);
860 up_read(&mm->mmap_sem);
863 if (vma->vm_ops && vma->vm_ops->get_policy)
864 pol = vma->vm_ops->get_policy(vma, addr);
866 pol = vma->vm_policy;
871 pol = &default_policy; /* indicates default behavior */
873 if (flags & MPOL_F_NODE) {
874 if (flags & MPOL_F_ADDR) {
875 err = lookup_node(mm, addr);
879 } else if (pol == current->mempolicy &&
880 pol->mode == MPOL_INTERLEAVE) {
881 *policy = current->il_next;
887 *policy = pol == &default_policy ? MPOL_DEFAULT :
890 * Internal mempolicy flags must be masked off before exposing
891 * the policy to userspace.
893 *policy |= (pol->flags & MPOL_MODE_FLAGS);
897 up_read(¤t->mm->mmap_sem);
903 if (mpol_store_user_nodemask(pol)) {
904 *nmask = pol->w.user_nodemask;
907 get_policy_nodemask(pol, nmask);
908 task_unlock(current);
915 up_read(¤t->mm->mmap_sem);
919 #ifdef CONFIG_MIGRATION
923 static void migrate_page_add(struct page *page, struct list_head *pagelist,
927 * Avoid migrating a page that is shared with others.
929 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
930 if (!isolate_lru_page(page)) {
931 list_add_tail(&page->lru, pagelist);
932 inc_zone_page_state(page, NR_ISOLATED_ANON +
933 page_is_file_cache(page));
938 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
940 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
944 * Migrate pages from one node to a target node.
945 * Returns error or the number of pages not migrated.
947 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
955 node_set(source, nmask);
958 * This does not "check" the range but isolates all pages that
959 * need migration. Between passing in the full user address
960 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
962 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
963 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
964 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
966 if (!list_empty(&pagelist)) {
967 err = migrate_pages(&pagelist, new_node_page, dest,
971 putback_lru_pages(&pagelist);
978 * Move pages between the two nodesets so as to preserve the physical
979 * layout as much as possible.
981 * Returns the number of page that could not be moved.
983 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
984 const nodemask_t *to, int flags)
990 err = migrate_prep();
994 down_read(&mm->mmap_sem);
996 err = migrate_vmas(mm, from, to, flags);
1001 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1002 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1003 * bit in 'tmp', and return that <source, dest> pair for migration.
1004 * The pair of nodemasks 'to' and 'from' define the map.
1006 * If no pair of bits is found that way, fallback to picking some
1007 * pair of 'source' and 'dest' bits that are not the same. If the
1008 * 'source' and 'dest' bits are the same, this represents a node
1009 * that will be migrating to itself, so no pages need move.
1011 * If no bits are left in 'tmp', or if all remaining bits left
1012 * in 'tmp' correspond to the same bit in 'to', return false
1013 * (nothing left to migrate).
1015 * This lets us pick a pair of nodes to migrate between, such that
1016 * if possible the dest node is not already occupied by some other
1017 * source node, minimizing the risk of overloading the memory on a
1018 * node that would happen if we migrated incoming memory to a node
1019 * before migrating outgoing memory source that same node.
1021 * A single scan of tmp is sufficient. As we go, we remember the
1022 * most recent <s, d> pair that moved (s != d). If we find a pair
1023 * that not only moved, but what's better, moved to an empty slot
1024 * (d is not set in tmp), then we break out then, with that pair.
1025 * Otherwise when we finish scanning from_tmp, we at least have the
1026 * most recent <s, d> pair that moved. If we get all the way through
1027 * the scan of tmp without finding any node that moved, much less
1028 * moved to an empty node, then there is nothing left worth migrating.
1032 while (!nodes_empty(tmp)) {
1037 for_each_node_mask(s, tmp) {
1040 * do_migrate_pages() tries to maintain the relative
1041 * node relationship of the pages established between
1042 * threads and memory areas.
1044 * However if the number of source nodes is not equal to
1045 * the number of destination nodes we can not preserve
1046 * this node relative relationship. In that case, skip
1047 * copying memory from a node that is in the destination
1050 * Example: [2,3,4] -> [3,4,5] moves everything.
1051 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1054 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1055 (node_isset(s, *to)))
1058 d = node_remap(s, *from, *to);
1062 source = s; /* Node moved. Memorize */
1065 /* dest not in remaining from nodes? */
1066 if (!node_isset(dest, tmp))
1072 node_clear(source, tmp);
1073 err = migrate_to_node(mm, source, dest, flags);
1080 up_read(&mm->mmap_sem);
1088 * Allocate a new page for page migration based on vma policy.
1089 * Start assuming that page is mapped by vma pointed to by @private.
1090 * Search forward from there, if not. N.B., this assumes that the
1091 * list of pages handed to migrate_pages()--which is how we get here--
1092 * is in virtual address order.
1094 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1096 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1097 unsigned long uninitialized_var(address);
1100 address = page_address_in_vma(page, vma);
1101 if (address != -EFAULT)
1107 * if !vma, alloc_page_vma() will use task or system default policy
1109 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1113 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1114 unsigned long flags)
1118 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1119 const nodemask_t *to, int flags)
1124 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1130 static long do_mbind(unsigned long start, unsigned long len,
1131 unsigned short mode, unsigned short mode_flags,
1132 nodemask_t *nmask, unsigned long flags)
1134 struct vm_area_struct *vma;
1135 struct mm_struct *mm = current->mm;
1136 struct mempolicy *new;
1139 LIST_HEAD(pagelist);
1141 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1142 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1144 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1147 if (start & ~PAGE_MASK)
1150 if (mode == MPOL_DEFAULT || mode == MPOL_NOOP)
1151 flags &= ~MPOL_MF_STRICT;
1153 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1161 new = mpol_new(mode, mode_flags, nmask);
1163 return PTR_ERR(new);
1166 * If we are using the default policy then operation
1167 * on discontinuous address spaces is okay after all
1170 flags |= MPOL_MF_DISCONTIG_OK;
1172 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1173 start, start + len, mode, mode_flags,
1174 nmask ? nodes_addr(*nmask)[0] : -1);
1176 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1178 err = migrate_prep();
1183 NODEMASK_SCRATCH(scratch);
1185 down_write(&mm->mmap_sem);
1187 err = mpol_set_nodemask(new, nmask, scratch);
1188 task_unlock(current);
1190 up_write(&mm->mmap_sem);
1193 NODEMASK_SCRATCH_FREE(scratch);
1198 vma = check_range(mm, start, end, nmask,
1199 flags | MPOL_MF_INVERT, &pagelist);
1205 err = mbind_range(mm, start, end, new);
1207 if (!list_empty(&pagelist)) {
1208 nr_failed = migrate_pages(&pagelist, new_vma_page,
1210 false, MIGRATE_SYNC,
1211 MR_MEMPOLICY_MBIND);
1213 putback_lru_pages(&pagelist);
1216 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1219 putback_lru_pages(&pagelist);
1221 up_write(&mm->mmap_sem);
1228 * User space interface with variable sized bitmaps for nodelists.
1231 /* Copy a node mask from user space. */
1232 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1233 unsigned long maxnode)
1236 unsigned long nlongs;
1237 unsigned long endmask;
1240 nodes_clear(*nodes);
1241 if (maxnode == 0 || !nmask)
1243 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1246 nlongs = BITS_TO_LONGS(maxnode);
1247 if ((maxnode % BITS_PER_LONG) == 0)
1250 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1252 /* When the user specified more nodes than supported just check
1253 if the non supported part is all zero. */
1254 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1255 if (nlongs > PAGE_SIZE/sizeof(long))
1257 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1259 if (get_user(t, nmask + k))
1261 if (k == nlongs - 1) {
1267 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1271 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1273 nodes_addr(*nodes)[nlongs-1] &= endmask;
1277 /* Copy a kernel node mask to user space */
1278 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1281 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1282 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1284 if (copy > nbytes) {
1285 if (copy > PAGE_SIZE)
1287 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1291 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1294 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1295 unsigned long, mode, unsigned long __user *, nmask,
1296 unsigned long, maxnode, unsigned, flags)
1300 unsigned short mode_flags;
1302 mode_flags = mode & MPOL_MODE_FLAGS;
1303 mode &= ~MPOL_MODE_FLAGS;
1304 if (mode >= MPOL_MAX)
1306 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1307 (mode_flags & MPOL_F_RELATIVE_NODES))
1309 err = get_nodes(&nodes, nmask, maxnode);
1312 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1315 /* Set the process memory policy */
1316 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1317 unsigned long, maxnode)
1321 unsigned short flags;
1323 flags = mode & MPOL_MODE_FLAGS;
1324 mode &= ~MPOL_MODE_FLAGS;
1325 if ((unsigned int)mode >= MPOL_MAX)
1327 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1329 err = get_nodes(&nodes, nmask, maxnode);
1332 return do_set_mempolicy(mode, flags, &nodes);
1335 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1336 const unsigned long __user *, old_nodes,
1337 const unsigned long __user *, new_nodes)
1339 const struct cred *cred = current_cred(), *tcred;
1340 struct mm_struct *mm = NULL;
1341 struct task_struct *task;
1342 nodemask_t task_nodes;
1346 NODEMASK_SCRATCH(scratch);
1351 old = &scratch->mask1;
1352 new = &scratch->mask2;
1354 err = get_nodes(old, old_nodes, maxnode);
1358 err = get_nodes(new, new_nodes, maxnode);
1362 /* Find the mm_struct */
1364 task = pid ? find_task_by_vpid(pid) : current;
1370 get_task_struct(task);
1375 * Check if this process has the right to modify the specified
1376 * process. The right exists if the process has administrative
1377 * capabilities, superuser privileges or the same
1378 * userid as the target process.
1380 tcred = __task_cred(task);
1381 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1382 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1383 !capable(CAP_SYS_NICE)) {
1390 task_nodes = cpuset_mems_allowed(task);
1391 /* Is the user allowed to access the target nodes? */
1392 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1397 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1402 err = security_task_movememory(task);
1406 mm = get_task_mm(task);
1407 put_task_struct(task);
1414 err = do_migrate_pages(mm, old, new,
1415 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1419 NODEMASK_SCRATCH_FREE(scratch);
1424 put_task_struct(task);
1430 /* Retrieve NUMA policy */
1431 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1432 unsigned long __user *, nmask, unsigned long, maxnode,
1433 unsigned long, addr, unsigned long, flags)
1436 int uninitialized_var(pval);
1439 if (nmask != NULL && maxnode < MAX_NUMNODES)
1442 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1447 if (policy && put_user(pval, policy))
1451 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1456 #ifdef CONFIG_COMPAT
1458 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1459 compat_ulong_t __user *nmask,
1460 compat_ulong_t maxnode,
1461 compat_ulong_t addr, compat_ulong_t flags)
1464 unsigned long __user *nm = NULL;
1465 unsigned long nr_bits, alloc_size;
1466 DECLARE_BITMAP(bm, MAX_NUMNODES);
1468 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1469 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1472 nm = compat_alloc_user_space(alloc_size);
1474 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1476 if (!err && nmask) {
1477 unsigned long copy_size;
1478 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1479 err = copy_from_user(bm, nm, copy_size);
1480 /* ensure entire bitmap is zeroed */
1481 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1482 err |= compat_put_bitmap(nmask, bm, nr_bits);
1488 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1489 compat_ulong_t maxnode)
1492 unsigned long __user *nm = NULL;
1493 unsigned long nr_bits, alloc_size;
1494 DECLARE_BITMAP(bm, MAX_NUMNODES);
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(bm, nmask, nr_bits);
1501 nm = compat_alloc_user_space(alloc_size);
1502 err |= copy_to_user(nm, bm, alloc_size);
1508 return sys_set_mempolicy(mode, nm, nr_bits+1);
1511 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1512 compat_ulong_t mode, compat_ulong_t __user *nmask,
1513 compat_ulong_t maxnode, compat_ulong_t flags)
1516 unsigned long __user *nm = NULL;
1517 unsigned long nr_bits, alloc_size;
1520 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1521 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1524 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1525 nm = compat_alloc_user_space(alloc_size);
1526 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1532 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1538 * get_vma_policy(@task, @vma, @addr)
1539 * @task - task for fallback if vma policy == default
1540 * @vma - virtual memory area whose policy is sought
1541 * @addr - address in @vma for shared policy lookup
1543 * Returns effective policy for a VMA at specified address.
1544 * Falls back to @task or system default policy, as necessary.
1545 * Current or other task's task mempolicy and non-shared vma policies must be
1546 * protected by task_lock(task) by the caller.
1547 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1548 * count--added by the get_policy() vm_op, as appropriate--to protect against
1549 * freeing by another task. It is the caller's responsibility to free the
1550 * extra reference for shared policies.
1552 struct mempolicy *get_vma_policy(struct task_struct *task,
1553 struct vm_area_struct *vma, unsigned long addr)
1555 struct mempolicy *pol = task->mempolicy;
1558 if (vma->vm_ops && vma->vm_ops->get_policy) {
1559 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1563 } else if (vma->vm_policy) {
1564 pol = vma->vm_policy;
1567 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1568 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1569 * count on these policies which will be dropped by
1570 * mpol_cond_put() later
1572 if (mpol_needs_cond_ref(pol))
1577 pol = &default_policy;
1582 * Return a nodemask representing a mempolicy for filtering nodes for
1585 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1587 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1588 if (unlikely(policy->mode == MPOL_BIND) &&
1589 gfp_zone(gfp) >= policy_zone &&
1590 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1591 return &policy->v.nodes;
1596 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1597 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1600 switch (policy->mode) {
1601 case MPOL_PREFERRED:
1602 if (!(policy->flags & MPOL_F_LOCAL))
1603 nd = policy->v.preferred_node;
1607 * Normally, MPOL_BIND allocations are node-local within the
1608 * allowed nodemask. However, if __GFP_THISNODE is set and the
1609 * current node isn't part of the mask, we use the zonelist for
1610 * the first node in the mask instead.
1612 if (unlikely(gfp & __GFP_THISNODE) &&
1613 unlikely(!node_isset(nd, policy->v.nodes)))
1614 nd = first_node(policy->v.nodes);
1619 return node_zonelist(nd, gfp);
1622 /* Do dynamic interleaving for a process */
1623 static unsigned interleave_nodes(struct mempolicy *policy)
1626 struct task_struct *me = current;
1629 next = next_node(nid, policy->v.nodes);
1630 if (next >= MAX_NUMNODES)
1631 next = first_node(policy->v.nodes);
1632 if (next < MAX_NUMNODES)
1638 * Depending on the memory policy provide a node from which to allocate the
1640 * @policy must be protected by freeing by the caller. If @policy is
1641 * the current task's mempolicy, this protection is implicit, as only the
1642 * task can change it's policy. The system default policy requires no
1645 unsigned slab_node(void)
1647 struct mempolicy *policy;
1650 return numa_node_id();
1652 policy = current->mempolicy;
1653 if (!policy || policy->flags & MPOL_F_LOCAL)
1654 return numa_node_id();
1656 switch (policy->mode) {
1657 case MPOL_PREFERRED:
1659 * handled MPOL_F_LOCAL above
1661 return policy->v.preferred_node;
1663 case MPOL_INTERLEAVE:
1664 return interleave_nodes(policy);
1668 * Follow bind policy behavior and start allocation at the
1671 struct zonelist *zonelist;
1673 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1674 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1675 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1678 return zone ? zone->node : numa_node_id();
1686 /* Do static interleaving for a VMA with known offset. */
1687 static unsigned offset_il_node(struct mempolicy *pol,
1688 struct vm_area_struct *vma, unsigned long off)
1690 unsigned nnodes = nodes_weight(pol->v.nodes);
1696 return numa_node_id();
1697 target = (unsigned int)off % nnodes;
1700 nid = next_node(nid, pol->v.nodes);
1702 } while (c <= target);
1706 /* Determine a node number for interleave */
1707 static inline unsigned interleave_nid(struct mempolicy *pol,
1708 struct vm_area_struct *vma, unsigned long addr, int shift)
1714 * for small pages, there is no difference between
1715 * shift and PAGE_SHIFT, so the bit-shift is safe.
1716 * for huge pages, since vm_pgoff is in units of small
1717 * pages, we need to shift off the always 0 bits to get
1720 BUG_ON(shift < PAGE_SHIFT);
1721 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1722 off += (addr - vma->vm_start) >> shift;
1723 return offset_il_node(pol, vma, off);
1725 return interleave_nodes(pol);
1729 * Return the bit number of a random bit set in the nodemask.
1730 * (returns -1 if nodemask is empty)
1732 int node_random(const nodemask_t *maskp)
1736 w = nodes_weight(*maskp);
1738 bit = bitmap_ord_to_pos(maskp->bits,
1739 get_random_int() % w, MAX_NUMNODES);
1743 #ifdef CONFIG_HUGETLBFS
1745 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1746 * @vma = virtual memory area whose policy is sought
1747 * @addr = address in @vma for shared policy lookup and interleave policy
1748 * @gfp_flags = for requested zone
1749 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1750 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1752 * Returns a zonelist suitable for a huge page allocation and a pointer
1753 * to the struct mempolicy for conditional unref after allocation.
1754 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1755 * @nodemask for filtering the zonelist.
1757 * Must be protected by get_mems_allowed()
1759 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1760 gfp_t gfp_flags, struct mempolicy **mpol,
1761 nodemask_t **nodemask)
1763 struct zonelist *zl;
1765 *mpol = get_vma_policy(current, vma, addr);
1766 *nodemask = NULL; /* assume !MPOL_BIND */
1768 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1769 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1770 huge_page_shift(hstate_vma(vma))), gfp_flags);
1772 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1773 if ((*mpol)->mode == MPOL_BIND)
1774 *nodemask = &(*mpol)->v.nodes;
1780 * init_nodemask_of_mempolicy
1782 * If the current task's mempolicy is "default" [NULL], return 'false'
1783 * to indicate default policy. Otherwise, extract the policy nodemask
1784 * for 'bind' or 'interleave' policy into the argument nodemask, or
1785 * initialize the argument nodemask to contain the single node for
1786 * 'preferred' or 'local' policy and return 'true' to indicate presence
1787 * of non-default mempolicy.
1789 * We don't bother with reference counting the mempolicy [mpol_get/put]
1790 * because the current task is examining it's own mempolicy and a task's
1791 * mempolicy is only ever changed by the task itself.
1793 * N.B., it is the caller's responsibility to free a returned nodemask.
1795 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1797 struct mempolicy *mempolicy;
1800 if (!(mask && current->mempolicy))
1804 mempolicy = current->mempolicy;
1805 switch (mempolicy->mode) {
1806 case MPOL_PREFERRED:
1807 if (mempolicy->flags & MPOL_F_LOCAL)
1808 nid = numa_node_id();
1810 nid = mempolicy->v.preferred_node;
1811 init_nodemask_of_node(mask, nid);
1816 case MPOL_INTERLEAVE:
1817 *mask = mempolicy->v.nodes;
1823 task_unlock(current);
1830 * mempolicy_nodemask_intersects
1832 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1833 * policy. Otherwise, check for intersection between mask and the policy
1834 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1835 * policy, always return true since it may allocate elsewhere on fallback.
1837 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1839 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1840 const nodemask_t *mask)
1842 struct mempolicy *mempolicy;
1848 mempolicy = tsk->mempolicy;
1852 switch (mempolicy->mode) {
1853 case MPOL_PREFERRED:
1855 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1856 * allocate from, they may fallback to other nodes when oom.
1857 * Thus, it's possible for tsk to have allocated memory from
1862 case MPOL_INTERLEAVE:
1863 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1873 /* Allocate a page in interleaved policy.
1874 Own path because it needs to do special accounting. */
1875 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1878 struct zonelist *zl;
1881 zl = node_zonelist(nid, gfp);
1882 page = __alloc_pages(gfp, order, zl);
1883 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1884 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1889 * alloc_pages_vma - Allocate a page for a VMA.
1892 * %GFP_USER user allocation.
1893 * %GFP_KERNEL kernel allocations,
1894 * %GFP_HIGHMEM highmem/user allocations,
1895 * %GFP_FS allocation should not call back into a file system.
1896 * %GFP_ATOMIC don't sleep.
1898 * @order:Order of the GFP allocation.
1899 * @vma: Pointer to VMA or NULL if not available.
1900 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1902 * This function allocates a page from the kernel page pool and applies
1903 * a NUMA policy associated with the VMA or the current process.
1904 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1905 * mm_struct of the VMA to prevent it from going away. Should be used for
1906 * all allocations for pages that will be mapped into
1907 * user space. Returns NULL when no page can be allocated.
1909 * Should be called with the mm_sem of the vma hold.
1912 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1913 unsigned long addr, int node)
1915 struct mempolicy *pol;
1916 struct zonelist *zl;
1918 unsigned int cpuset_mems_cookie;
1921 pol = get_vma_policy(current, vma, addr);
1922 cpuset_mems_cookie = get_mems_allowed();
1924 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1927 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1929 page = alloc_page_interleave(gfp, order, nid);
1930 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1935 zl = policy_zonelist(gfp, pol, node);
1936 if (unlikely(mpol_needs_cond_ref(pol))) {
1938 * slow path: ref counted shared policy
1940 struct page *page = __alloc_pages_nodemask(gfp, order,
1941 zl, policy_nodemask(gfp, pol));
1943 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1948 * fast path: default or task policy
1950 page = __alloc_pages_nodemask(gfp, order, zl,
1951 policy_nodemask(gfp, pol));
1952 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1958 * alloc_pages_current - Allocate pages.
1961 * %GFP_USER user allocation,
1962 * %GFP_KERNEL kernel allocation,
1963 * %GFP_HIGHMEM highmem allocation,
1964 * %GFP_FS don't call back into a file system.
1965 * %GFP_ATOMIC don't sleep.
1966 * @order: Power of two of allocation size in pages. 0 is a single page.
1968 * Allocate a page from the kernel page pool. When not in
1969 * interrupt context and apply the current process NUMA policy.
1970 * Returns NULL when no page can be allocated.
1972 * Don't call cpuset_update_task_memory_state() unless
1973 * 1) it's ok to take cpuset_sem (can WAIT), and
1974 * 2) allocating for current task (not interrupt).
1976 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1978 struct mempolicy *pol = current->mempolicy;
1980 unsigned int cpuset_mems_cookie;
1982 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1983 pol = &default_policy;
1986 cpuset_mems_cookie = get_mems_allowed();
1989 * No reference counting needed for current->mempolicy
1990 * nor system default_policy
1992 if (pol->mode == MPOL_INTERLEAVE)
1993 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1995 page = __alloc_pages_nodemask(gfp, order,
1996 policy_zonelist(gfp, pol, numa_node_id()),
1997 policy_nodemask(gfp, pol));
1999 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2004 EXPORT_SYMBOL(alloc_pages_current);
2007 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2008 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2009 * with the mems_allowed returned by cpuset_mems_allowed(). This
2010 * keeps mempolicies cpuset relative after its cpuset moves. See
2011 * further kernel/cpuset.c update_nodemask().
2013 * current's mempolicy may be rebinded by the other task(the task that changes
2014 * cpuset's mems), so we needn't do rebind work for current task.
2017 /* Slow path of a mempolicy duplicate */
2018 struct mempolicy *__mpol_dup(struct mempolicy *old)
2020 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2023 return ERR_PTR(-ENOMEM);
2025 /* task's mempolicy is protected by alloc_lock */
2026 if (old == current->mempolicy) {
2029 task_unlock(current);
2034 if (current_cpuset_is_being_rebound()) {
2035 nodemask_t mems = cpuset_mems_allowed(current);
2036 if (new->flags & MPOL_F_REBINDING)
2037 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2039 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2042 atomic_set(&new->refcnt, 1);
2047 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2048 * eliminate the * MPOL_F_* flags that require conditional ref and
2049 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2050 * after return. Use the returned value.
2052 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2053 * policy lookup, even if the policy needs/has extra ref on lookup.
2054 * shmem_readahead needs this.
2056 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2057 struct mempolicy *frompol)
2059 if (!mpol_needs_cond_ref(frompol))
2063 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2064 __mpol_put(frompol);
2068 /* Slow path of a mempolicy comparison */
2069 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2073 if (a->mode != b->mode)
2075 if (a->flags != b->flags)
2077 if (mpol_store_user_nodemask(a))
2078 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2084 case MPOL_INTERLEAVE:
2085 return !!nodes_equal(a->v.nodes, b->v.nodes);
2086 case MPOL_PREFERRED:
2087 return a->v.preferred_node == b->v.preferred_node;
2095 * Shared memory backing store policy support.
2097 * Remember policies even when nobody has shared memory mapped.
2098 * The policies are kept in Red-Black tree linked from the inode.
2099 * They are protected by the sp->lock spinlock, which should be held
2100 * for any accesses to the tree.
2103 /* lookup first element intersecting start-end */
2104 /* Caller holds sp->mutex */
2105 static struct sp_node *
2106 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2108 struct rb_node *n = sp->root.rb_node;
2111 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2113 if (start >= p->end)
2115 else if (end <= p->start)
2123 struct sp_node *w = NULL;
2124 struct rb_node *prev = rb_prev(n);
2127 w = rb_entry(prev, struct sp_node, nd);
2128 if (w->end <= start)
2132 return rb_entry(n, struct sp_node, nd);
2135 /* Insert a new shared policy into the list. */
2136 /* Caller holds sp->lock */
2137 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2139 struct rb_node **p = &sp->root.rb_node;
2140 struct rb_node *parent = NULL;
2145 nd = rb_entry(parent, struct sp_node, nd);
2146 if (new->start < nd->start)
2148 else if (new->end > nd->end)
2149 p = &(*p)->rb_right;
2153 rb_link_node(&new->nd, parent, p);
2154 rb_insert_color(&new->nd, &sp->root);
2155 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2156 new->policy ? new->policy->mode : 0);
2159 /* Find shared policy intersecting idx */
2161 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2163 struct mempolicy *pol = NULL;
2166 if (!sp->root.rb_node)
2168 mutex_lock(&sp->mutex);
2169 sn = sp_lookup(sp, idx, idx+1);
2171 mpol_get(sn->policy);
2174 mutex_unlock(&sp->mutex);
2178 static void sp_free(struct sp_node *n)
2180 mpol_put(n->policy);
2181 kmem_cache_free(sn_cache, n);
2184 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2186 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2187 rb_erase(&n->nd, &sp->root);
2191 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2192 struct mempolicy *pol)
2195 struct mempolicy *newpol;
2197 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2201 newpol = mpol_dup(pol);
2202 if (IS_ERR(newpol)) {
2203 kmem_cache_free(sn_cache, n);
2206 newpol->flags |= MPOL_F_SHARED;
2215 /* Replace a policy range. */
2216 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2217 unsigned long end, struct sp_node *new)
2222 mutex_lock(&sp->mutex);
2223 n = sp_lookup(sp, start, end);
2224 /* Take care of old policies in the same range. */
2225 while (n && n->start < end) {
2226 struct rb_node *next = rb_next(&n->nd);
2227 if (n->start >= start) {
2233 /* Old policy spanning whole new range. */
2235 struct sp_node *new2;
2236 new2 = sp_alloc(end, n->end, n->policy);
2242 sp_insert(sp, new2);
2249 n = rb_entry(next, struct sp_node, nd);
2254 mutex_unlock(&sp->mutex);
2259 * mpol_shared_policy_init - initialize shared policy for inode
2260 * @sp: pointer to inode shared policy
2261 * @mpol: struct mempolicy to install
2263 * Install non-NULL @mpol in inode's shared policy rb-tree.
2264 * On entry, the current task has a reference on a non-NULL @mpol.
2265 * This must be released on exit.
2266 * This is called at get_inode() calls and we can use GFP_KERNEL.
2268 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2272 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2273 mutex_init(&sp->mutex);
2276 struct vm_area_struct pvma;
2277 struct mempolicy *new;
2278 NODEMASK_SCRATCH(scratch);
2282 /* contextualize the tmpfs mount point mempolicy */
2283 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2285 goto free_scratch; /* no valid nodemask intersection */
2288 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2289 task_unlock(current);
2293 /* Create pseudo-vma that contains just the policy */
2294 memset(&pvma, 0, sizeof(struct vm_area_struct));
2295 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2296 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2299 mpol_put(new); /* drop initial ref */
2301 NODEMASK_SCRATCH_FREE(scratch);
2303 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2307 int mpol_set_shared_policy(struct shared_policy *info,
2308 struct vm_area_struct *vma, struct mempolicy *npol)
2311 struct sp_node *new = NULL;
2312 unsigned long sz = vma_pages(vma);
2314 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2316 sz, npol ? npol->mode : -1,
2317 npol ? npol->flags : -1,
2318 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2321 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2325 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2331 /* Free a backing policy store on inode delete. */
2332 void mpol_free_shared_policy(struct shared_policy *p)
2335 struct rb_node *next;
2337 if (!p->root.rb_node)
2339 mutex_lock(&p->mutex);
2340 next = rb_first(&p->root);
2342 n = rb_entry(next, struct sp_node, nd);
2343 next = rb_next(&n->nd);
2346 mutex_unlock(&p->mutex);
2349 /* assumes fs == KERNEL_DS */
2350 void __init numa_policy_init(void)
2352 nodemask_t interleave_nodes;
2353 unsigned long largest = 0;
2354 int nid, prefer = 0;
2356 policy_cache = kmem_cache_create("numa_policy",
2357 sizeof(struct mempolicy),
2358 0, SLAB_PANIC, NULL);
2360 sn_cache = kmem_cache_create("shared_policy_node",
2361 sizeof(struct sp_node),
2362 0, SLAB_PANIC, NULL);
2365 * Set interleaving policy for system init. Interleaving is only
2366 * enabled across suitably sized nodes (default is >= 16MB), or
2367 * fall back to the largest node if they're all smaller.
2369 nodes_clear(interleave_nodes);
2370 for_each_node_state(nid, N_HIGH_MEMORY) {
2371 unsigned long total_pages = node_present_pages(nid);
2373 /* Preserve the largest node */
2374 if (largest < total_pages) {
2375 largest = total_pages;
2379 /* Interleave this node? */
2380 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2381 node_set(nid, interleave_nodes);
2384 /* All too small, use the largest */
2385 if (unlikely(nodes_empty(interleave_nodes)))
2386 node_set(prefer, interleave_nodes);
2388 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2389 printk("numa_policy_init: interleaving failed\n");
2392 /* Reset policy of current process to default */
2393 void numa_default_policy(void)
2395 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2399 * Parse and format mempolicy from/to strings
2403 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2404 * Used only for mpol_parse_str() and mpol_to_str()
2406 static const char * const policy_modes[] =
2408 [MPOL_DEFAULT] = "default",
2409 [MPOL_PREFERRED] = "prefer",
2410 [MPOL_BIND] = "bind",
2411 [MPOL_INTERLEAVE] = "interleave",
2412 [MPOL_LOCAL] = "local",
2413 [MPOL_NOOP] = "noop", /* should not actually be used */
2419 * mpol_parse_str - parse string to mempolicy
2420 * @str: string containing mempolicy to parse
2421 * @mpol: pointer to struct mempolicy pointer, returned on success.
2422 * @no_context: flag whether to "contextualize" the mempolicy
2425 * <mode>[=<flags>][:<nodelist>]
2427 * if @no_context is true, save the input nodemask in w.user_nodemask in
2428 * the returned mempolicy. This will be used to "clone" the mempolicy in
2429 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2430 * mount option. Note that if 'static' or 'relative' mode flags were
2431 * specified, the input nodemask will already have been saved. Saving
2432 * it again is redundant, but safe.
2434 * On success, returns 0, else 1
2436 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2438 struct mempolicy *new = NULL;
2439 unsigned short mode;
2440 unsigned short uninitialized_var(mode_flags);
2442 char *nodelist = strchr(str, ':');
2443 char *flags = strchr(str, '=');
2447 /* NUL-terminate mode or flags string */
2449 if (nodelist_parse(nodelist, nodes))
2451 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2457 *flags++ = '\0'; /* terminate mode string */
2459 for (mode = 0; mode < MPOL_MAX; mode++) {
2460 if (!strcmp(str, policy_modes[mode])) {
2464 if (mode >= MPOL_MAX || mode == MPOL_NOOP)
2468 case MPOL_PREFERRED:
2470 * Insist on a nodelist of one node only
2473 char *rest = nodelist;
2474 while (isdigit(*rest))
2480 case MPOL_INTERLEAVE:
2482 * Default to online nodes with memory if no nodelist
2485 nodes = node_states[N_HIGH_MEMORY];
2489 * Don't allow a nodelist; mpol_new() checks flags
2493 mode = MPOL_PREFERRED;
2497 * Insist on a empty nodelist
2504 * Insist on a nodelist
2513 * Currently, we only support two mutually exclusive
2516 if (!strcmp(flags, "static"))
2517 mode_flags |= MPOL_F_STATIC_NODES;
2518 else if (!strcmp(flags, "relative"))
2519 mode_flags |= MPOL_F_RELATIVE_NODES;
2524 new = mpol_new(mode, mode_flags, &nodes);
2529 /* save for contextualization */
2530 new->w.user_nodemask = nodes;
2533 NODEMASK_SCRATCH(scratch);
2536 ret = mpol_set_nodemask(new, &nodes, scratch);
2537 task_unlock(current);
2540 NODEMASK_SCRATCH_FREE(scratch);
2549 /* Restore string for error message */
2558 #endif /* CONFIG_TMPFS */
2561 * mpol_to_str - format a mempolicy structure for printing
2562 * @buffer: to contain formatted mempolicy string
2563 * @maxlen: length of @buffer
2564 * @pol: pointer to mempolicy to be formatted
2565 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2567 * Convert a mempolicy into a string.
2568 * Returns the number of characters in buffer (if positive)
2569 * or an error (negative)
2571 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2576 unsigned short mode;
2577 unsigned short flags = pol ? pol->flags : 0;
2580 * Sanity check: room for longest mode, flag and some nodes
2582 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2584 if (!pol || pol == &default_policy)
2585 mode = MPOL_DEFAULT;
2594 case MPOL_PREFERRED:
2596 if (flags & MPOL_F_LOCAL)
2597 mode = MPOL_LOCAL; /* pseudo-policy */
2599 node_set(pol->v.preferred_node, nodes);
2604 case MPOL_INTERLEAVE:
2606 nodes = pol->w.user_nodemask;
2608 nodes = pol->v.nodes;
2615 l = strlen(policy_modes[mode]);
2616 if (buffer + maxlen < p + l + 1)
2619 strcpy(p, policy_modes[mode]);
2622 if (flags & MPOL_MODE_FLAGS) {
2623 if (buffer + maxlen < p + 2)
2628 * Currently, the only defined flags are mutually exclusive
2630 if (flags & MPOL_F_STATIC_NODES)
2631 p += snprintf(p, buffer + maxlen - p, "static");
2632 else if (flags & MPOL_F_RELATIVE_NODES)
2633 p += snprintf(p, buffer + maxlen - p, "relative");
2636 if (!nodes_empty(nodes)) {
2637 if (buffer + maxlen < p + 2)
2640 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);