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/gfp.h>
77 #include <linux/slab.h>
78 #include <linux/string.h>
79 #include <linux/module.h>
80 #include <linux/nsproxy.h>
81 #include <linux/interrupt.h>
82 #include <linux/init.h>
83 #include <linux/compat.h>
84 #include <linux/swap.h>
85 #include <linux/seq_file.h>
86 #include <linux/proc_fs.h>
87 #include <linux/migrate.h>
88 #include <linux/ksm.h>
89 #include <linux/rmap.h>
90 #include <linux/security.h>
91 #include <linux/syscalls.h>
92 #include <linux/ctype.h>
93 #include <linux/mm_inline.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.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 */
103 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
105 static struct kmem_cache *policy_cache;
106 static struct kmem_cache *sn_cache;
108 /* Highest zone. An specific allocation for a zone below that is not
110 enum zone_type policy_zone = 0;
113 * run-time system-wide default policy => local allocation
115 struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
121 static const struct mempolicy_operations {
122 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
123 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
124 } mpol_ops[MPOL_MAX];
126 /* Check that the nodemask contains at least one populated zone */
127 static int is_valid_nodemask(const nodemask_t *nodemask)
131 /* Check that there is something useful in this mask */
134 for_each_node_mask(nd, *nodemask) {
137 for (k = 0; k <= policy_zone; k++) {
138 z = &NODE_DATA(nd)->node_zones[k];
139 if (z->present_pages > 0)
147 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
149 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
152 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
153 const nodemask_t *rel)
156 nodes_fold(tmp, *orig, nodes_weight(*rel));
157 nodes_onto(*ret, tmp, *rel);
160 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
162 if (nodes_empty(*nodes))
164 pol->v.nodes = *nodes;
168 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
171 pol->flags |= MPOL_F_LOCAL; /* local allocation */
172 else if (nodes_empty(*nodes))
173 return -EINVAL; /* no allowed nodes */
175 pol->v.preferred_node = first_node(*nodes);
179 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
181 if (!is_valid_nodemask(nodes))
183 pol->v.nodes = *nodes;
188 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
189 * any, for the new policy. mpol_new() has already validated the nodes
190 * parameter with respect to the policy mode and flags. But, we need to
191 * handle an empty nodemask with MPOL_PREFERRED here.
193 * Must be called holding task's alloc_lock to protect task's mems_allowed
194 * and mempolicy. May also be called holding the mmap_semaphore for write.
196 static int mpol_set_nodemask(struct mempolicy *pol,
197 const nodemask_t *nodes, struct nodemask_scratch *nsc)
201 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
204 /* Check N_HIGH_MEMORY */
205 nodes_and(nsc->mask1,
206 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
209 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
210 nodes = NULL; /* explicit local allocation */
212 if (pol->flags & MPOL_F_RELATIVE_NODES)
213 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
215 nodes_and(nsc->mask2, *nodes, nsc->mask1);
217 if (mpol_store_user_nodemask(pol))
218 pol->w.user_nodemask = *nodes;
220 pol->w.cpuset_mems_allowed =
221 cpuset_current_mems_allowed;
225 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
227 ret = mpol_ops[pol->mode].create(pol, NULL);
232 * This function just creates a new policy, does some check and simple
233 * initialization. You must invoke mpol_set_nodemask() to set nodes.
235 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
238 struct mempolicy *policy;
240 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
241 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
243 if (mode == MPOL_DEFAULT) {
244 if (nodes && !nodes_empty(*nodes))
245 return ERR_PTR(-EINVAL);
246 return NULL; /* simply delete any existing policy */
251 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
252 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
253 * All other modes require a valid pointer to a non-empty nodemask.
255 if (mode == MPOL_PREFERRED) {
256 if (nodes_empty(*nodes)) {
257 if (((flags & MPOL_F_STATIC_NODES) ||
258 (flags & MPOL_F_RELATIVE_NODES)))
259 return ERR_PTR(-EINVAL);
261 } else if (nodes_empty(*nodes))
262 return ERR_PTR(-EINVAL);
263 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
265 return ERR_PTR(-ENOMEM);
266 atomic_set(&policy->refcnt, 1);
268 policy->flags = flags;
273 /* Slow path of a mpol destructor. */
274 void __mpol_put(struct mempolicy *p)
276 if (!atomic_dec_and_test(&p->refcnt))
278 kmem_cache_free(policy_cache, p);
281 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
285 static void mpol_rebind_nodemask(struct mempolicy *pol,
286 const nodemask_t *nodes)
290 if (pol->flags & MPOL_F_STATIC_NODES)
291 nodes_and(tmp, pol->w.user_nodemask, *nodes);
292 else if (pol->flags & MPOL_F_RELATIVE_NODES)
293 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
295 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
297 pol->w.cpuset_mems_allowed = *nodes;
301 if (!node_isset(current->il_next, tmp)) {
302 current->il_next = next_node(current->il_next, tmp);
303 if (current->il_next >= MAX_NUMNODES)
304 current->il_next = first_node(tmp);
305 if (current->il_next >= MAX_NUMNODES)
306 current->il_next = numa_node_id();
310 static void mpol_rebind_preferred(struct mempolicy *pol,
311 const nodemask_t *nodes)
315 if (pol->flags & MPOL_F_STATIC_NODES) {
316 int node = first_node(pol->w.user_nodemask);
318 if (node_isset(node, *nodes)) {
319 pol->v.preferred_node = node;
320 pol->flags &= ~MPOL_F_LOCAL;
322 pol->flags |= MPOL_F_LOCAL;
323 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
324 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
325 pol->v.preferred_node = first_node(tmp);
326 } else if (!(pol->flags & MPOL_F_LOCAL)) {
327 pol->v.preferred_node = node_remap(pol->v.preferred_node,
328 pol->w.cpuset_mems_allowed,
330 pol->w.cpuset_mems_allowed = *nodes;
334 /* Migrate a policy to a different set of nodes */
335 static void mpol_rebind_policy(struct mempolicy *pol,
336 const nodemask_t *newmask)
340 if (!mpol_store_user_nodemask(pol) &&
341 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
343 mpol_ops[pol->mode].rebind(pol, newmask);
347 * Wrapper for mpol_rebind_policy() that just requires task
348 * pointer, and updates task mempolicy.
350 * Called with task's alloc_lock held.
353 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
355 mpol_rebind_policy(tsk->mempolicy, new);
359 * Rebind each vma in mm to new nodemask.
361 * Call holding a reference to mm. Takes mm->mmap_sem during call.
364 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
366 struct vm_area_struct *vma;
368 down_write(&mm->mmap_sem);
369 for (vma = mm->mmap; vma; vma = vma->vm_next)
370 mpol_rebind_policy(vma->vm_policy, new);
371 up_write(&mm->mmap_sem);
374 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
376 .rebind = mpol_rebind_default,
378 [MPOL_INTERLEAVE] = {
379 .create = mpol_new_interleave,
380 .rebind = mpol_rebind_nodemask,
383 .create = mpol_new_preferred,
384 .rebind = mpol_rebind_preferred,
387 .create = mpol_new_bind,
388 .rebind = mpol_rebind_nodemask,
392 static void gather_stats(struct page *, void *, int pte_dirty);
393 static void migrate_page_add(struct page *page, struct list_head *pagelist,
394 unsigned long flags);
396 /* Scan through pages checking if pages follow certain conditions. */
397 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
398 unsigned long addr, unsigned long end,
399 const nodemask_t *nodes, unsigned long flags,
406 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
411 if (!pte_present(*pte))
413 page = vm_normal_page(vma, addr, *pte);
417 * vm_normal_page() filters out zero pages, but there might
418 * still be PageReserved pages to skip, perhaps in a VDSO.
419 * And we cannot move PageKsm pages sensibly or safely yet.
421 if (PageReserved(page) || PageKsm(page))
423 nid = page_to_nid(page);
424 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
427 if (flags & MPOL_MF_STATS)
428 gather_stats(page, private, pte_dirty(*pte));
429 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
430 migrate_page_add(page, private, flags);
433 } while (pte++, addr += PAGE_SIZE, addr != end);
434 pte_unmap_unlock(orig_pte, ptl);
438 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
439 unsigned long addr, unsigned long end,
440 const nodemask_t *nodes, unsigned long flags,
446 pmd = pmd_offset(pud, addr);
448 next = pmd_addr_end(addr, end);
449 if (pmd_none_or_clear_bad(pmd))
451 if (check_pte_range(vma, pmd, addr, next, nodes,
454 } while (pmd++, addr = next, addr != end);
458 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
459 unsigned long addr, unsigned long end,
460 const nodemask_t *nodes, unsigned long flags,
466 pud = pud_offset(pgd, addr);
468 next = pud_addr_end(addr, end);
469 if (pud_none_or_clear_bad(pud))
471 if (check_pmd_range(vma, pud, addr, next, nodes,
474 } while (pud++, addr = next, addr != end);
478 static inline int check_pgd_range(struct vm_area_struct *vma,
479 unsigned long addr, unsigned long end,
480 const nodemask_t *nodes, unsigned long flags,
486 pgd = pgd_offset(vma->vm_mm, addr);
488 next = pgd_addr_end(addr, end);
489 if (pgd_none_or_clear_bad(pgd))
491 if (check_pud_range(vma, pgd, addr, next, nodes,
494 } while (pgd++, addr = next, addr != end);
499 * Check if all pages in a range are on a set of nodes.
500 * If pagelist != NULL then isolate pages from the LRU and
501 * put them on the pagelist.
503 static struct vm_area_struct *
504 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags, void *private)
508 struct vm_area_struct *first, *vma, *prev;
511 first = find_vma(mm, start);
513 return ERR_PTR(-EFAULT);
515 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
516 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
517 if (!vma->vm_next && vma->vm_end < end)
518 return ERR_PTR(-EFAULT);
519 if (prev && prev->vm_end < vma->vm_start)
520 return ERR_PTR(-EFAULT);
522 if (!is_vm_hugetlb_page(vma) &&
523 ((flags & MPOL_MF_STRICT) ||
524 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
525 vma_migratable(vma)))) {
526 unsigned long endvma = vma->vm_end;
530 if (vma->vm_start > start)
531 start = vma->vm_start;
532 err = check_pgd_range(vma, start, endvma, nodes,
535 first = ERR_PTR(err);
544 /* Apply policy to a single VMA */
545 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
548 struct mempolicy *old = vma->vm_policy;
550 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
551 vma->vm_start, vma->vm_end, vma->vm_pgoff,
552 vma->vm_ops, vma->vm_file,
553 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
555 if (vma->vm_ops && vma->vm_ops->set_policy)
556 err = vma->vm_ops->set_policy(vma, new);
559 vma->vm_policy = new;
565 /* Step 2: apply policy to a range and do splits. */
566 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
567 unsigned long end, struct mempolicy *new)
569 struct vm_area_struct *next;
573 for (; vma && vma->vm_start < end; vma = next) {
575 if (vma->vm_start < start)
576 err = split_vma(vma->vm_mm, vma, start, 1);
577 if (!err && vma->vm_end > end)
578 err = split_vma(vma->vm_mm, vma, end, 0);
580 err = policy_vma(vma, new);
588 * Update task->flags PF_MEMPOLICY bit: set iff non-default
589 * mempolicy. Allows more rapid checking of this (combined perhaps
590 * with other PF_* flag bits) on memory allocation hot code paths.
592 * If called from outside this file, the task 'p' should -only- be
593 * a newly forked child not yet visible on the task list, because
594 * manipulating the task flags of a visible task is not safe.
596 * The above limitation is why this routine has the funny name
597 * mpol_fix_fork_child_flag().
599 * It is also safe to call this with a task pointer of current,
600 * which the static wrapper mpol_set_task_struct_flag() does,
601 * for use within this file.
604 void mpol_fix_fork_child_flag(struct task_struct *p)
607 p->flags |= PF_MEMPOLICY;
609 p->flags &= ~PF_MEMPOLICY;
612 static void mpol_set_task_struct_flag(void)
614 mpol_fix_fork_child_flag(current);
617 /* Set the process memory policy */
618 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
621 struct mempolicy *new, *old;
622 struct mm_struct *mm = current->mm;
623 NODEMASK_SCRATCH(scratch);
629 new = mpol_new(mode, flags, nodes);
635 * prevent changing our mempolicy while show_numa_maps()
637 * Note: do_set_mempolicy() can be called at init time
641 down_write(&mm->mmap_sem);
643 ret = mpol_set_nodemask(new, nodes, scratch);
645 task_unlock(current);
647 up_write(&mm->mmap_sem);
651 old = current->mempolicy;
652 current->mempolicy = new;
653 mpol_set_task_struct_flag();
654 if (new && new->mode == MPOL_INTERLEAVE &&
655 nodes_weight(new->v.nodes))
656 current->il_next = first_node(new->v.nodes);
657 task_unlock(current);
659 up_write(&mm->mmap_sem);
664 NODEMASK_SCRATCH_FREE(scratch);
669 * Return nodemask for policy for get_mempolicy() query
671 * Called with task's alloc_lock held
673 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
676 if (p == &default_policy)
682 case MPOL_INTERLEAVE:
686 if (!(p->flags & MPOL_F_LOCAL))
687 node_set(p->v.preferred_node, *nodes);
688 /* else return empty node mask for local allocation */
695 static int lookup_node(struct mm_struct *mm, unsigned long addr)
700 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
702 err = page_to_nid(p);
708 /* Retrieve NUMA policy */
709 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
710 unsigned long addr, unsigned long flags)
713 struct mm_struct *mm = current->mm;
714 struct vm_area_struct *vma = NULL;
715 struct mempolicy *pol = current->mempolicy;
718 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
721 if (flags & MPOL_F_MEMS_ALLOWED) {
722 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
724 *policy = 0; /* just so it's initialized */
726 *nmask = cpuset_current_mems_allowed;
727 task_unlock(current);
731 if (flags & MPOL_F_ADDR) {
733 * Do NOT fall back to task policy if the
734 * vma/shared policy at addr is NULL. We
735 * want to return MPOL_DEFAULT in this case.
737 down_read(&mm->mmap_sem);
738 vma = find_vma_intersection(mm, addr, addr+1);
740 up_read(&mm->mmap_sem);
743 if (vma->vm_ops && vma->vm_ops->get_policy)
744 pol = vma->vm_ops->get_policy(vma, addr);
746 pol = vma->vm_policy;
751 pol = &default_policy; /* indicates default behavior */
753 if (flags & MPOL_F_NODE) {
754 if (flags & MPOL_F_ADDR) {
755 err = lookup_node(mm, addr);
759 } else if (pol == current->mempolicy &&
760 pol->mode == MPOL_INTERLEAVE) {
761 *policy = current->il_next;
767 *policy = pol == &default_policy ? MPOL_DEFAULT :
770 * Internal mempolicy flags must be masked off before exposing
771 * the policy to userspace.
773 *policy |= (pol->flags & MPOL_MODE_FLAGS);
777 up_read(¤t->mm->mmap_sem);
784 get_policy_nodemask(pol, nmask);
785 task_unlock(current);
791 up_read(¤t->mm->mmap_sem);
795 #ifdef CONFIG_MIGRATION
799 static void migrate_page_add(struct page *page, struct list_head *pagelist,
803 * Avoid migrating a page that is shared with others.
805 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
806 if (!isolate_lru_page(page)) {
807 list_add_tail(&page->lru, pagelist);
808 inc_zone_page_state(page, NR_ISOLATED_ANON +
809 page_is_file_cache(page));
814 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
816 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
820 * Migrate pages from one node to a target node.
821 * Returns error or the number of pages not migrated.
823 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
831 node_set(source, nmask);
833 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
834 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
836 if (!list_empty(&pagelist))
837 err = migrate_pages(&pagelist, new_node_page, dest, 0);
843 * Move pages between the two nodesets so as to preserve the physical
844 * layout as much as possible.
846 * Returns the number of page that could not be moved.
848 int do_migrate_pages(struct mm_struct *mm,
849 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
855 err = migrate_prep();
859 down_read(&mm->mmap_sem);
861 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
866 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
867 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
868 * bit in 'tmp', and return that <source, dest> pair for migration.
869 * The pair of nodemasks 'to' and 'from' define the map.
871 * If no pair of bits is found that way, fallback to picking some
872 * pair of 'source' and 'dest' bits that are not the same. If the
873 * 'source' and 'dest' bits are the same, this represents a node
874 * that will be migrating to itself, so no pages need move.
876 * If no bits are left in 'tmp', or if all remaining bits left
877 * in 'tmp' correspond to the same bit in 'to', return false
878 * (nothing left to migrate).
880 * This lets us pick a pair of nodes to migrate between, such that
881 * if possible the dest node is not already occupied by some other
882 * source node, minimizing the risk of overloading the memory on a
883 * node that would happen if we migrated incoming memory to a node
884 * before migrating outgoing memory source that same node.
886 * A single scan of tmp is sufficient. As we go, we remember the
887 * most recent <s, d> pair that moved (s != d). If we find a pair
888 * that not only moved, but what's better, moved to an empty slot
889 * (d is not set in tmp), then we break out then, with that pair.
890 * Otherwise when we finish scannng from_tmp, we at least have the
891 * most recent <s, d> pair that moved. If we get all the way through
892 * the scan of tmp without finding any node that moved, much less
893 * moved to an empty node, then there is nothing left worth migrating.
897 while (!nodes_empty(tmp)) {
902 for_each_node_mask(s, tmp) {
903 d = node_remap(s, *from_nodes, *to_nodes);
907 source = s; /* Node moved. Memorize */
910 /* dest not in remaining from nodes? */
911 if (!node_isset(dest, tmp))
917 node_clear(source, tmp);
918 err = migrate_to_node(mm, source, dest, flags);
925 up_read(&mm->mmap_sem);
933 * Allocate a new page for page migration based on vma policy.
934 * Start assuming that page is mapped by vma pointed to by @private.
935 * Search forward from there, if not. N.B., this assumes that the
936 * list of pages handed to migrate_pages()--which is how we get here--
937 * is in virtual address order.
939 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
941 struct vm_area_struct *vma = (struct vm_area_struct *)private;
942 unsigned long uninitialized_var(address);
945 address = page_address_in_vma(page, vma);
946 if (address != -EFAULT)
952 * if !vma, alloc_page_vma() will use task or system default policy
954 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
958 static void migrate_page_add(struct page *page, struct list_head *pagelist,
963 int do_migrate_pages(struct mm_struct *mm,
964 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
969 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
975 static long do_mbind(unsigned long start, unsigned long len,
976 unsigned short mode, unsigned short mode_flags,
977 nodemask_t *nmask, unsigned long flags)
979 struct vm_area_struct *vma;
980 struct mm_struct *mm = current->mm;
981 struct mempolicy *new;
986 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
987 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
989 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
992 if (start & ~PAGE_MASK)
995 if (mode == MPOL_DEFAULT)
996 flags &= ~MPOL_MF_STRICT;
998 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1006 new = mpol_new(mode, mode_flags, nmask);
1008 return PTR_ERR(new);
1011 * If we are using the default policy then operation
1012 * on discontinuous address spaces is okay after all
1015 flags |= MPOL_MF_DISCONTIG_OK;
1017 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1018 start, start + len, mode, mode_flags,
1019 nmask ? nodes_addr(*nmask)[0] : -1);
1021 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1023 err = migrate_prep();
1028 NODEMASK_SCRATCH(scratch);
1030 down_write(&mm->mmap_sem);
1032 err = mpol_set_nodemask(new, nmask, scratch);
1033 task_unlock(current);
1035 up_write(&mm->mmap_sem);
1038 NODEMASK_SCRATCH_FREE(scratch);
1043 vma = check_range(mm, start, end, nmask,
1044 flags | MPOL_MF_INVERT, &pagelist);
1050 err = mbind_range(vma, start, end, new);
1052 if (!list_empty(&pagelist))
1053 nr_failed = migrate_pages(&pagelist, new_vma_page,
1054 (unsigned long)vma, 0);
1056 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1059 putback_lru_pages(&pagelist);
1061 up_write(&mm->mmap_sem);
1068 * User space interface with variable sized bitmaps for nodelists.
1071 /* Copy a node mask from user space. */
1072 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1073 unsigned long maxnode)
1076 unsigned long nlongs;
1077 unsigned long endmask;
1080 nodes_clear(*nodes);
1081 if (maxnode == 0 || !nmask)
1083 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1086 nlongs = BITS_TO_LONGS(maxnode);
1087 if ((maxnode % BITS_PER_LONG) == 0)
1090 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1092 /* When the user specified more nodes than supported just check
1093 if the non supported part is all zero. */
1094 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1095 if (nlongs > PAGE_SIZE/sizeof(long))
1097 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1099 if (get_user(t, nmask + k))
1101 if (k == nlongs - 1) {
1107 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1111 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1113 nodes_addr(*nodes)[nlongs-1] &= endmask;
1117 /* Copy a kernel node mask to user space */
1118 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1121 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1122 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1124 if (copy > nbytes) {
1125 if (copy > PAGE_SIZE)
1127 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1131 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1134 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1135 unsigned long, mode, unsigned long __user *, nmask,
1136 unsigned long, maxnode, unsigned, flags)
1140 unsigned short mode_flags;
1142 mode_flags = mode & MPOL_MODE_FLAGS;
1143 mode &= ~MPOL_MODE_FLAGS;
1144 if (mode >= MPOL_MAX)
1146 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1147 (mode_flags & MPOL_F_RELATIVE_NODES))
1149 err = get_nodes(&nodes, nmask, maxnode);
1152 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1155 /* Set the process memory policy */
1156 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1157 unsigned long, maxnode)
1161 unsigned short flags;
1163 flags = mode & MPOL_MODE_FLAGS;
1164 mode &= ~MPOL_MODE_FLAGS;
1165 if ((unsigned int)mode >= MPOL_MAX)
1167 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1169 err = get_nodes(&nodes, nmask, maxnode);
1172 return do_set_mempolicy(mode, flags, &nodes);
1175 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1176 const unsigned long __user *, old_nodes,
1177 const unsigned long __user *, new_nodes)
1179 const struct cred *cred = current_cred(), *tcred;
1180 struct mm_struct *mm;
1181 struct task_struct *task;
1184 nodemask_t task_nodes;
1187 err = get_nodes(&old, old_nodes, maxnode);
1191 err = get_nodes(&new, new_nodes, maxnode);
1195 /* Find the mm_struct */
1196 read_lock(&tasklist_lock);
1197 task = pid ? find_task_by_vpid(pid) : current;
1199 read_unlock(&tasklist_lock);
1202 mm = get_task_mm(task);
1203 read_unlock(&tasklist_lock);
1209 * Check if this process has the right to modify the specified
1210 * process. The right exists if the process has administrative
1211 * capabilities, superuser privileges or the same
1212 * userid as the target process.
1215 tcred = __task_cred(task);
1216 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1217 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1218 !capable(CAP_SYS_NICE)) {
1225 task_nodes = cpuset_mems_allowed(task);
1226 /* Is the user allowed to access the target nodes? */
1227 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1232 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1237 err = security_task_movememory(task);
1241 err = do_migrate_pages(mm, &old, &new,
1242 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1249 /* Retrieve NUMA policy */
1250 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1251 unsigned long __user *, nmask, unsigned long, maxnode,
1252 unsigned long, addr, unsigned long, flags)
1255 int uninitialized_var(pval);
1258 if (nmask != NULL && maxnode < MAX_NUMNODES)
1261 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1266 if (policy && put_user(pval, policy))
1270 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1275 #ifdef CONFIG_COMPAT
1277 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1278 compat_ulong_t __user *nmask,
1279 compat_ulong_t maxnode,
1280 compat_ulong_t addr, compat_ulong_t flags)
1283 unsigned long __user *nm = NULL;
1284 unsigned long nr_bits, alloc_size;
1285 DECLARE_BITMAP(bm, MAX_NUMNODES);
1287 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1288 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1291 nm = compat_alloc_user_space(alloc_size);
1293 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1295 if (!err && nmask) {
1296 err = copy_from_user(bm, nm, alloc_size);
1297 /* ensure entire bitmap is zeroed */
1298 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1299 err |= compat_put_bitmap(nmask, bm, nr_bits);
1305 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1306 compat_ulong_t maxnode)
1309 unsigned long __user *nm = NULL;
1310 unsigned long nr_bits, alloc_size;
1311 DECLARE_BITMAP(bm, MAX_NUMNODES);
1313 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1314 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1317 err = compat_get_bitmap(bm, nmask, nr_bits);
1318 nm = compat_alloc_user_space(alloc_size);
1319 err |= copy_to_user(nm, bm, alloc_size);
1325 return sys_set_mempolicy(mode, nm, nr_bits+1);
1328 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1329 compat_ulong_t mode, compat_ulong_t __user *nmask,
1330 compat_ulong_t maxnode, compat_ulong_t flags)
1333 unsigned long __user *nm = NULL;
1334 unsigned long nr_bits, alloc_size;
1337 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1338 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1341 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1342 nm = compat_alloc_user_space(alloc_size);
1343 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1349 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1355 * get_vma_policy(@task, @vma, @addr)
1356 * @task - task for fallback if vma policy == default
1357 * @vma - virtual memory area whose policy is sought
1358 * @addr - address in @vma for shared policy lookup
1360 * Returns effective policy for a VMA at specified address.
1361 * Falls back to @task or system default policy, as necessary.
1362 * Current or other task's task mempolicy and non-shared vma policies
1363 * are protected by the task's mmap_sem, which must be held for read by
1365 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1366 * count--added by the get_policy() vm_op, as appropriate--to protect against
1367 * freeing by another task. It is the caller's responsibility to free the
1368 * extra reference for shared policies.
1370 static struct mempolicy *get_vma_policy(struct task_struct *task,
1371 struct vm_area_struct *vma, unsigned long addr)
1373 struct mempolicy *pol = task->mempolicy;
1376 if (vma->vm_ops && vma->vm_ops->get_policy) {
1377 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1381 } else if (vma->vm_policy)
1382 pol = vma->vm_policy;
1385 pol = &default_policy;
1390 * Return a nodemask representing a mempolicy for filtering nodes for
1393 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1395 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1396 if (unlikely(policy->mode == MPOL_BIND) &&
1397 gfp_zone(gfp) >= policy_zone &&
1398 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1399 return &policy->v.nodes;
1404 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1405 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1407 int nd = numa_node_id();
1409 switch (policy->mode) {
1410 case MPOL_PREFERRED:
1411 if (!(policy->flags & MPOL_F_LOCAL))
1412 nd = policy->v.preferred_node;
1416 * Normally, MPOL_BIND allocations are node-local within the
1417 * allowed nodemask. However, if __GFP_THISNODE is set and the
1418 * current node is part of the mask, we use the zonelist for
1419 * the first node in the mask instead.
1421 if (unlikely(gfp & __GFP_THISNODE) &&
1422 unlikely(!node_isset(nd, policy->v.nodes)))
1423 nd = first_node(policy->v.nodes);
1425 case MPOL_INTERLEAVE: /* should not happen */
1430 return node_zonelist(nd, gfp);
1433 /* Do dynamic interleaving for a process */
1434 static unsigned interleave_nodes(struct mempolicy *policy)
1437 struct task_struct *me = current;
1440 next = next_node(nid, policy->v.nodes);
1441 if (next >= MAX_NUMNODES)
1442 next = first_node(policy->v.nodes);
1443 if (next < MAX_NUMNODES)
1449 * Depending on the memory policy provide a node from which to allocate the
1451 * @policy must be protected by freeing by the caller. If @policy is
1452 * the current task's mempolicy, this protection is implicit, as only the
1453 * task can change it's policy. The system default policy requires no
1456 unsigned slab_node(struct mempolicy *policy)
1458 if (!policy || policy->flags & MPOL_F_LOCAL)
1459 return numa_node_id();
1461 switch (policy->mode) {
1462 case MPOL_PREFERRED:
1464 * handled MPOL_F_LOCAL above
1466 return policy->v.preferred_node;
1468 case MPOL_INTERLEAVE:
1469 return interleave_nodes(policy);
1473 * Follow bind policy behavior and start allocation at the
1476 struct zonelist *zonelist;
1478 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1479 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1480 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1491 /* Do static interleaving for a VMA with known offset. */
1492 static unsigned offset_il_node(struct mempolicy *pol,
1493 struct vm_area_struct *vma, unsigned long off)
1495 unsigned nnodes = nodes_weight(pol->v.nodes);
1501 return numa_node_id();
1502 target = (unsigned int)off % nnodes;
1505 nid = next_node(nid, pol->v.nodes);
1507 } while (c <= target);
1511 /* Determine a node number for interleave */
1512 static inline unsigned interleave_nid(struct mempolicy *pol,
1513 struct vm_area_struct *vma, unsigned long addr, int shift)
1519 * for small pages, there is no difference between
1520 * shift and PAGE_SHIFT, so the bit-shift is safe.
1521 * for huge pages, since vm_pgoff is in units of small
1522 * pages, we need to shift off the always 0 bits to get
1525 BUG_ON(shift < PAGE_SHIFT);
1526 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1527 off += (addr - vma->vm_start) >> shift;
1528 return offset_il_node(pol, vma, off);
1530 return interleave_nodes(pol);
1533 #ifdef CONFIG_HUGETLBFS
1535 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1536 * @vma = virtual memory area whose policy is sought
1537 * @addr = address in @vma for shared policy lookup and interleave policy
1538 * @gfp_flags = for requested zone
1539 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1540 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1542 * Returns a zonelist suitable for a huge page allocation and a pointer
1543 * to the struct mempolicy for conditional unref after allocation.
1544 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1545 * @nodemask for filtering the zonelist.
1547 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1548 gfp_t gfp_flags, struct mempolicy **mpol,
1549 nodemask_t **nodemask)
1551 struct zonelist *zl;
1553 *mpol = get_vma_policy(current, vma, addr);
1554 *nodemask = NULL; /* assume !MPOL_BIND */
1556 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1557 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1558 huge_page_shift(hstate_vma(vma))), gfp_flags);
1560 zl = policy_zonelist(gfp_flags, *mpol);
1561 if ((*mpol)->mode == MPOL_BIND)
1562 *nodemask = &(*mpol)->v.nodes;
1568 * init_nodemask_of_mempolicy
1570 * If the current task's mempolicy is "default" [NULL], return 'false'
1571 * to indicate default policy. Otherwise, extract the policy nodemask
1572 * for 'bind' or 'interleave' policy into the argument nodemask, or
1573 * initialize the argument nodemask to contain the single node for
1574 * 'preferred' or 'local' policy and return 'true' to indicate presence
1575 * of non-default mempolicy.
1577 * We don't bother with reference counting the mempolicy [mpol_get/put]
1578 * because the current task is examining it's own mempolicy and a task's
1579 * mempolicy is only ever changed by the task itself.
1581 * N.B., it is the caller's responsibility to free a returned nodemask.
1583 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1585 struct mempolicy *mempolicy;
1588 if (!(mask && current->mempolicy))
1591 mempolicy = current->mempolicy;
1592 switch (mempolicy->mode) {
1593 case MPOL_PREFERRED:
1594 if (mempolicy->flags & MPOL_F_LOCAL)
1595 nid = numa_node_id();
1597 nid = mempolicy->v.preferred_node;
1598 init_nodemask_of_node(mask, nid);
1603 case MPOL_INTERLEAVE:
1604 *mask = mempolicy->v.nodes;
1615 /* Allocate a page in interleaved policy.
1616 Own path because it needs to do special accounting. */
1617 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1620 struct zonelist *zl;
1623 zl = node_zonelist(nid, gfp);
1624 page = __alloc_pages(gfp, order, zl);
1625 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1626 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1631 * alloc_page_vma - Allocate a page for a VMA.
1634 * %GFP_USER user allocation.
1635 * %GFP_KERNEL kernel allocations,
1636 * %GFP_HIGHMEM highmem/user allocations,
1637 * %GFP_FS allocation should not call back into a file system.
1638 * %GFP_ATOMIC don't sleep.
1640 * @vma: Pointer to VMA or NULL if not available.
1641 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1643 * This function allocates a page from the kernel page pool and applies
1644 * a NUMA policy associated with the VMA or the current process.
1645 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1646 * mm_struct of the VMA to prevent it from going away. Should be used for
1647 * all allocations for pages that will be mapped into
1648 * user space. Returns NULL when no page can be allocated.
1650 * Should be called with the mm_sem of the vma hold.
1653 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1655 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1656 struct zonelist *zl;
1658 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1661 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1663 return alloc_page_interleave(gfp, 0, nid);
1665 zl = policy_zonelist(gfp, pol);
1666 if (unlikely(mpol_needs_cond_ref(pol))) {
1668 * slow path: ref counted shared policy
1670 struct page *page = __alloc_pages_nodemask(gfp, 0,
1671 zl, policy_nodemask(gfp, pol));
1676 * fast path: default or task policy
1678 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1682 * alloc_pages_current - Allocate pages.
1685 * %GFP_USER user allocation,
1686 * %GFP_KERNEL kernel allocation,
1687 * %GFP_HIGHMEM highmem allocation,
1688 * %GFP_FS don't call back into a file system.
1689 * %GFP_ATOMIC don't sleep.
1690 * @order: Power of two of allocation size in pages. 0 is a single page.
1692 * Allocate a page from the kernel page pool. When not in
1693 * interrupt context and apply the current process NUMA policy.
1694 * Returns NULL when no page can be allocated.
1696 * Don't call cpuset_update_task_memory_state() unless
1697 * 1) it's ok to take cpuset_sem (can WAIT), and
1698 * 2) allocating for current task (not interrupt).
1700 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1702 struct mempolicy *pol = current->mempolicy;
1704 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1705 pol = &default_policy;
1708 * No reference counting needed for current->mempolicy
1709 * nor system default_policy
1711 if (pol->mode == MPOL_INTERLEAVE)
1712 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1713 return __alloc_pages_nodemask(gfp, order,
1714 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1716 EXPORT_SYMBOL(alloc_pages_current);
1719 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1720 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1721 * with the mems_allowed returned by cpuset_mems_allowed(). This
1722 * keeps mempolicies cpuset relative after its cpuset moves. See
1723 * further kernel/cpuset.c update_nodemask().
1726 /* Slow path of a mempolicy duplicate */
1727 struct mempolicy *__mpol_dup(struct mempolicy *old)
1729 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1732 return ERR_PTR(-ENOMEM);
1734 if (current_cpuset_is_being_rebound()) {
1735 nodemask_t mems = cpuset_mems_allowed(current);
1736 mpol_rebind_policy(old, &mems);
1740 atomic_set(&new->refcnt, 1);
1745 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1746 * eliminate the * MPOL_F_* flags that require conditional ref and
1747 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1748 * after return. Use the returned value.
1750 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1751 * policy lookup, even if the policy needs/has extra ref on lookup.
1752 * shmem_readahead needs this.
1754 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1755 struct mempolicy *frompol)
1757 if (!mpol_needs_cond_ref(frompol))
1761 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1762 __mpol_put(frompol);
1766 static int mpol_match_intent(const struct mempolicy *a,
1767 const struct mempolicy *b)
1769 if (a->flags != b->flags)
1771 if (!mpol_store_user_nodemask(a))
1773 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1776 /* Slow path of a mempolicy comparison */
1777 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1781 if (a->mode != b->mode)
1783 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
1788 case MPOL_INTERLEAVE:
1789 return nodes_equal(a->v.nodes, b->v.nodes);
1790 case MPOL_PREFERRED:
1791 return a->v.preferred_node == b->v.preferred_node &&
1792 a->flags == b->flags;
1800 * Shared memory backing store policy support.
1802 * Remember policies even when nobody has shared memory mapped.
1803 * The policies are kept in Red-Black tree linked from the inode.
1804 * They are protected by the sp->lock spinlock, which should be held
1805 * for any accesses to the tree.
1808 /* lookup first element intersecting start-end */
1809 /* Caller holds sp->lock */
1810 static struct sp_node *
1811 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1813 struct rb_node *n = sp->root.rb_node;
1816 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1818 if (start >= p->end)
1820 else if (end <= p->start)
1828 struct sp_node *w = NULL;
1829 struct rb_node *prev = rb_prev(n);
1832 w = rb_entry(prev, struct sp_node, nd);
1833 if (w->end <= start)
1837 return rb_entry(n, struct sp_node, nd);
1840 /* Insert a new shared policy into the list. */
1841 /* Caller holds sp->lock */
1842 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1844 struct rb_node **p = &sp->root.rb_node;
1845 struct rb_node *parent = NULL;
1850 nd = rb_entry(parent, struct sp_node, nd);
1851 if (new->start < nd->start)
1853 else if (new->end > nd->end)
1854 p = &(*p)->rb_right;
1858 rb_link_node(&new->nd, parent, p);
1859 rb_insert_color(&new->nd, &sp->root);
1860 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1861 new->policy ? new->policy->mode : 0);
1864 /* Find shared policy intersecting idx */
1866 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1868 struct mempolicy *pol = NULL;
1871 if (!sp->root.rb_node)
1873 spin_lock(&sp->lock);
1874 sn = sp_lookup(sp, idx, idx+1);
1876 mpol_get(sn->policy);
1879 spin_unlock(&sp->lock);
1883 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1885 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1886 rb_erase(&n->nd, &sp->root);
1887 mpol_put(n->policy);
1888 kmem_cache_free(sn_cache, n);
1891 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1892 struct mempolicy *pol)
1894 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1901 pol->flags |= MPOL_F_SHARED; /* for unref */
1906 /* Replace a policy range. */
1907 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1908 unsigned long end, struct sp_node *new)
1910 struct sp_node *n, *new2 = NULL;
1913 spin_lock(&sp->lock);
1914 n = sp_lookup(sp, start, end);
1915 /* Take care of old policies in the same range. */
1916 while (n && n->start < end) {
1917 struct rb_node *next = rb_next(&n->nd);
1918 if (n->start >= start) {
1924 /* Old policy spanning whole new range. */
1927 spin_unlock(&sp->lock);
1928 new2 = sp_alloc(end, n->end, n->policy);
1934 sp_insert(sp, new2);
1942 n = rb_entry(next, struct sp_node, nd);
1946 spin_unlock(&sp->lock);
1948 mpol_put(new2->policy);
1949 kmem_cache_free(sn_cache, new2);
1955 * mpol_shared_policy_init - initialize shared policy for inode
1956 * @sp: pointer to inode shared policy
1957 * @mpol: struct mempolicy to install
1959 * Install non-NULL @mpol in inode's shared policy rb-tree.
1960 * On entry, the current task has a reference on a non-NULL @mpol.
1961 * This must be released on exit.
1962 * This is called at get_inode() calls and we can use GFP_KERNEL.
1964 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
1968 sp->root = RB_ROOT; /* empty tree == default mempolicy */
1969 spin_lock_init(&sp->lock);
1972 struct vm_area_struct pvma;
1973 struct mempolicy *new;
1974 NODEMASK_SCRATCH(scratch);
1978 /* contextualize the tmpfs mount point mempolicy */
1979 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
1981 mpol_put(mpol); /* drop our ref on sb mpol */
1982 NODEMASK_SCRATCH_FREE(scratch);
1983 return; /* no valid nodemask intersection */
1987 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
1988 task_unlock(current);
1989 mpol_put(mpol); /* drop our ref on sb mpol */
1991 NODEMASK_SCRATCH_FREE(scratch);
1996 /* Create pseudo-vma that contains just the policy */
1997 memset(&pvma, 0, sizeof(struct vm_area_struct));
1998 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
1999 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2000 mpol_put(new); /* drop initial ref */
2001 NODEMASK_SCRATCH_FREE(scratch);
2005 int mpol_set_shared_policy(struct shared_policy *info,
2006 struct vm_area_struct *vma, struct mempolicy *npol)
2009 struct sp_node *new = NULL;
2010 unsigned long sz = vma_pages(vma);
2012 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2014 sz, npol ? npol->mode : -1,
2015 npol ? npol->flags : -1,
2016 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2019 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2023 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2025 kmem_cache_free(sn_cache, new);
2029 /* Free a backing policy store on inode delete. */
2030 void mpol_free_shared_policy(struct shared_policy *p)
2033 struct rb_node *next;
2035 if (!p->root.rb_node)
2037 spin_lock(&p->lock);
2038 next = rb_first(&p->root);
2040 n = rb_entry(next, struct sp_node, nd);
2041 next = rb_next(&n->nd);
2042 rb_erase(&n->nd, &p->root);
2043 mpol_put(n->policy);
2044 kmem_cache_free(sn_cache, n);
2046 spin_unlock(&p->lock);
2049 /* assumes fs == KERNEL_DS */
2050 void __init numa_policy_init(void)
2052 nodemask_t interleave_nodes;
2053 unsigned long largest = 0;
2054 int nid, prefer = 0;
2056 policy_cache = kmem_cache_create("numa_policy",
2057 sizeof(struct mempolicy),
2058 0, SLAB_PANIC, NULL);
2060 sn_cache = kmem_cache_create("shared_policy_node",
2061 sizeof(struct sp_node),
2062 0, SLAB_PANIC, NULL);
2065 * Set interleaving policy for system init. Interleaving is only
2066 * enabled across suitably sized nodes (default is >= 16MB), or
2067 * fall back to the largest node if they're all smaller.
2069 nodes_clear(interleave_nodes);
2070 for_each_node_state(nid, N_HIGH_MEMORY) {
2071 unsigned long total_pages = node_present_pages(nid);
2073 /* Preserve the largest node */
2074 if (largest < total_pages) {
2075 largest = total_pages;
2079 /* Interleave this node? */
2080 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2081 node_set(nid, interleave_nodes);
2084 /* All too small, use the largest */
2085 if (unlikely(nodes_empty(interleave_nodes)))
2086 node_set(prefer, interleave_nodes);
2088 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2089 printk("numa_policy_init: interleaving failed\n");
2092 /* Reset policy of current process to default */
2093 void numa_default_policy(void)
2095 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2099 * Parse and format mempolicy from/to strings
2103 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2104 * Used only for mpol_parse_str() and mpol_to_str()
2106 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
2107 static const char * const policy_types[] =
2108 { "default", "prefer", "bind", "interleave", "local" };
2113 * mpol_parse_str - parse string to mempolicy
2114 * @str: string containing mempolicy to parse
2115 * @mpol: pointer to struct mempolicy pointer, returned on success.
2116 * @no_context: flag whether to "contextualize" the mempolicy
2119 * <mode>[=<flags>][:<nodelist>]
2121 * if @no_context is true, save the input nodemask in w.user_nodemask in
2122 * the returned mempolicy. This will be used to "clone" the mempolicy in
2123 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2124 * mount option. Note that if 'static' or 'relative' mode flags were
2125 * specified, the input nodemask will already have been saved. Saving
2126 * it again is redundant, but safe.
2128 * On success, returns 0, else 1
2130 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2132 struct mempolicy *new = NULL;
2133 unsigned short uninitialized_var(mode);
2134 unsigned short uninitialized_var(mode_flags);
2136 char *nodelist = strchr(str, ':');
2137 char *flags = strchr(str, '=');
2142 /* NUL-terminate mode or flags string */
2144 if (nodelist_parse(nodelist, nodes))
2146 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2152 *flags++ = '\0'; /* terminate mode string */
2154 for (i = 0; i <= MPOL_LOCAL; i++) {
2155 if (!strcmp(str, policy_types[i])) {
2164 case MPOL_PREFERRED:
2166 * Insist on a nodelist of one node only
2169 char *rest = nodelist;
2170 while (isdigit(*rest))
2176 case MPOL_INTERLEAVE:
2178 * Default to online nodes with memory if no nodelist
2181 nodes = node_states[N_HIGH_MEMORY];
2186 * Don't allow a nodelist; mpol_new() checks flags
2190 mode = MPOL_PREFERRED;
2194 * case MPOL_BIND: mpol_new() enforces non-empty nodemask.
2195 * case MPOL_DEFAULT: mpol_new() enforces empty nodemask, ignores flags.
2202 * Currently, we only support two mutually exclusive
2205 if (!strcmp(flags, "static"))
2206 mode_flags |= MPOL_F_STATIC_NODES;
2207 else if (!strcmp(flags, "relative"))
2208 mode_flags |= MPOL_F_RELATIVE_NODES;
2213 new = mpol_new(mode, mode_flags, &nodes);
2218 NODEMASK_SCRATCH(scratch);
2221 ret = mpol_set_nodemask(new, &nodes, scratch);
2222 task_unlock(current);
2225 NODEMASK_SCRATCH_FREE(scratch);
2229 } else if (no_context) {
2230 /* save for contextualization */
2231 new->w.user_nodemask = nodes;
2236 /* Restore string for error message */
2245 #endif /* CONFIG_TMPFS */
2248 * mpol_to_str - format a mempolicy structure for printing
2249 * @buffer: to contain formatted mempolicy string
2250 * @maxlen: length of @buffer
2251 * @pol: pointer to mempolicy to be formatted
2252 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2254 * Convert a mempolicy into a string.
2255 * Returns the number of characters in buffer (if positive)
2256 * or an error (negative)
2258 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2263 unsigned short mode;
2264 unsigned short flags = pol ? pol->flags : 0;
2267 * Sanity check: room for longest mode, flag and some nodes
2269 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2271 if (!pol || pol == &default_policy)
2272 mode = MPOL_DEFAULT;
2281 case MPOL_PREFERRED:
2283 if (flags & MPOL_F_LOCAL)
2284 mode = MPOL_LOCAL; /* pseudo-policy */
2286 node_set(pol->v.preferred_node, nodes);
2291 case MPOL_INTERLEAVE:
2293 nodes = pol->w.user_nodemask;
2295 nodes = pol->v.nodes;
2302 l = strlen(policy_types[mode]);
2303 if (buffer + maxlen < p + l + 1)
2306 strcpy(p, policy_types[mode]);
2309 if (flags & MPOL_MODE_FLAGS) {
2310 if (buffer + maxlen < p + 2)
2315 * Currently, the only defined flags are mutually exclusive
2317 if (flags & MPOL_F_STATIC_NODES)
2318 p += snprintf(p, buffer + maxlen - p, "static");
2319 else if (flags & MPOL_F_RELATIVE_NODES)
2320 p += snprintf(p, buffer + maxlen - p, "relative");
2323 if (!nodes_empty(nodes)) {
2324 if (buffer + maxlen < p + 2)
2327 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2333 unsigned long pages;
2335 unsigned long active;
2336 unsigned long writeback;
2337 unsigned long mapcount_max;
2338 unsigned long dirty;
2339 unsigned long swapcache;
2340 unsigned long node[MAX_NUMNODES];
2343 static void gather_stats(struct page *page, void *private, int pte_dirty)
2345 struct numa_maps *md = private;
2346 int count = page_mapcount(page);
2349 if (pte_dirty || PageDirty(page))
2352 if (PageSwapCache(page))
2355 if (PageActive(page) || PageUnevictable(page))
2358 if (PageWriteback(page))
2364 if (count > md->mapcount_max)
2365 md->mapcount_max = count;
2367 md->node[page_to_nid(page)]++;
2370 #ifdef CONFIG_HUGETLB_PAGE
2371 static void check_huge_range(struct vm_area_struct *vma,
2372 unsigned long start, unsigned long end,
2373 struct numa_maps *md)
2377 struct hstate *h = hstate_vma(vma);
2378 unsigned long sz = huge_page_size(h);
2380 for (addr = start; addr < end; addr += sz) {
2381 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2382 addr & huge_page_mask(h));
2392 page = pte_page(pte);
2396 gather_stats(page, md, pte_dirty(*ptep));
2400 static inline void check_huge_range(struct vm_area_struct *vma,
2401 unsigned long start, unsigned long end,
2402 struct numa_maps *md)
2408 * Display pages allocated per node and memory policy via /proc.
2410 int show_numa_map(struct seq_file *m, void *v)
2412 struct proc_maps_private *priv = m->private;
2413 struct vm_area_struct *vma = v;
2414 struct numa_maps *md;
2415 struct file *file = vma->vm_file;
2416 struct mm_struct *mm = vma->vm_mm;
2417 struct mempolicy *pol;
2424 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2428 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2429 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2432 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2435 seq_printf(m, " file=");
2436 seq_path(m, &file->f_path, "\n\t= ");
2437 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2438 seq_printf(m, " heap");
2439 } else if (vma->vm_start <= mm->start_stack &&
2440 vma->vm_end >= mm->start_stack) {
2441 seq_printf(m, " stack");
2444 if (is_vm_hugetlb_page(vma)) {
2445 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2446 seq_printf(m, " huge");
2448 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2449 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2456 seq_printf(m," anon=%lu",md->anon);
2459 seq_printf(m," dirty=%lu",md->dirty);
2461 if (md->pages != md->anon && md->pages != md->dirty)
2462 seq_printf(m, " mapped=%lu", md->pages);
2464 if (md->mapcount_max > 1)
2465 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2468 seq_printf(m," swapcache=%lu", md->swapcache);
2470 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2471 seq_printf(m," active=%lu", md->active);
2474 seq_printf(m," writeback=%lu", md->writeback);
2476 for_each_node_state(n, N_HIGH_MEMORY)
2478 seq_printf(m, " N%d=%lu", n, md->node[n]);
2483 if (m->count < m->size)
2484 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;