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/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
92 #include <asm/tlbflush.h>
93 #include <asm/uaccess.h>
96 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
97 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
98 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
100 static struct kmem_cache *policy_cache;
101 static struct kmem_cache *sn_cache;
103 /* Highest zone. An specific allocation for a zone below that is not
105 enum zone_type policy_zone = 0;
107 struct mempolicy default_policy = {
108 .refcnt = ATOMIC_INIT(1), /* never free it */
109 .policy = MPOL_DEFAULT,
112 static const struct mempolicy_operations {
113 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
114 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
115 } mpol_ops[MPOL_MAX];
117 /* Check that the nodemask contains at least one populated zone */
118 static int is_valid_nodemask(const nodemask_t *nodemask)
122 /* Check that there is something useful in this mask */
125 for_each_node_mask(nd, *nodemask) {
128 for (k = 0; k <= policy_zone; k++) {
129 z = &NODE_DATA(nd)->node_zones[k];
130 if (z->present_pages > 0)
138 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
140 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
143 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
144 const nodemask_t *rel)
147 nodes_fold(tmp, *orig, nodes_weight(*rel));
148 nodes_onto(*ret, tmp, *rel);
151 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
153 if (nodes_empty(*nodes))
155 pol->v.nodes = *nodes;
159 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
162 pol->v.preferred_node = -1; /* local allocation */
163 else if (nodes_empty(*nodes))
164 return -EINVAL; /* no allowed nodes */
166 pol->v.preferred_node = first_node(*nodes);
170 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
172 if (!is_valid_nodemask(nodes))
174 pol->v.nodes = *nodes;
178 /* Create a new policy */
179 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
182 struct mempolicy *policy;
183 nodemask_t cpuset_context_nmask;
187 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
188 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
190 if (mode == MPOL_DEFAULT)
192 if (!nodes || nodes_empty(*nodes)) {
193 if (mode != MPOL_PREFERRED)
194 return ERR_PTR(-EINVAL);
195 localalloc = 1; /* special case: no mode flags */
197 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
199 return ERR_PTR(-ENOMEM);
200 atomic_set(&policy->refcnt, 1);
201 policy->policy = mode;
204 policy->flags = flags;
205 cpuset_update_task_memory_state();
206 if (flags & MPOL_F_RELATIVE_NODES)
207 mpol_relative_nodemask(&cpuset_context_nmask, nodes,
208 &cpuset_current_mems_allowed);
210 nodes_and(cpuset_context_nmask, *nodes,
211 cpuset_current_mems_allowed);
212 if (mpol_store_user_nodemask(policy))
213 policy->w.user_nodemask = *nodes;
215 policy->w.cpuset_mems_allowed =
216 cpuset_mems_allowed(current);
219 ret = mpol_ops[mode].create(policy,
220 localalloc ? NULL : &cpuset_context_nmask);
222 kmem_cache_free(policy_cache, policy);
228 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
232 static void mpol_rebind_nodemask(struct mempolicy *pol,
233 const nodemask_t *nodes)
237 if (pol->flags & MPOL_F_STATIC_NODES)
238 nodes_and(tmp, pol->w.user_nodemask, *nodes);
239 else if (pol->flags & MPOL_F_RELATIVE_NODES)
240 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
242 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
244 pol->w.cpuset_mems_allowed = *nodes;
248 if (!node_isset(current->il_next, tmp)) {
249 current->il_next = next_node(current->il_next, tmp);
250 if (current->il_next >= MAX_NUMNODES)
251 current->il_next = first_node(tmp);
252 if (current->il_next >= MAX_NUMNODES)
253 current->il_next = numa_node_id();
257 static void mpol_rebind_preferred(struct mempolicy *pol,
258 const nodemask_t *nodes)
263 * check 'STATIC_NODES first, as preferred_node == -1 may be
264 * a temporary, "fallback" state for this policy.
266 if (pol->flags & MPOL_F_STATIC_NODES) {
267 int node = first_node(pol->w.user_nodemask);
269 if (node_isset(node, *nodes))
270 pol->v.preferred_node = node;
272 pol->v.preferred_node = -1;
273 } else if (pol->v.preferred_node == -1) {
274 return; /* no remap required for explicit local alloc */
275 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
276 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
277 pol->v.preferred_node = first_node(tmp);
279 pol->v.preferred_node = node_remap(pol->v.preferred_node,
280 pol->w.cpuset_mems_allowed,
282 pol->w.cpuset_mems_allowed = *nodes;
286 /* Migrate a policy to a different set of nodes */
287 static void mpol_rebind_policy(struct mempolicy *pol,
288 const nodemask_t *newmask)
292 if (!mpol_store_user_nodemask(pol) &&
293 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
295 mpol_ops[pol->policy].rebind(pol, newmask);
299 * Wrapper for mpol_rebind_policy() that just requires task
300 * pointer, and updates task mempolicy.
303 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
305 mpol_rebind_policy(tsk->mempolicy, new);
309 * Rebind each vma in mm to new nodemask.
311 * Call holding a reference to mm. Takes mm->mmap_sem during call.
314 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
316 struct vm_area_struct *vma;
318 down_write(&mm->mmap_sem);
319 for (vma = mm->mmap; vma; vma = vma->vm_next)
320 mpol_rebind_policy(vma->vm_policy, new);
321 up_write(&mm->mmap_sem);
324 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
326 .rebind = mpol_rebind_default,
328 [MPOL_INTERLEAVE] = {
329 .create = mpol_new_interleave,
330 .rebind = mpol_rebind_nodemask,
333 .create = mpol_new_preferred,
334 .rebind = mpol_rebind_preferred,
337 .create = mpol_new_bind,
338 .rebind = mpol_rebind_nodemask,
342 static void gather_stats(struct page *, void *, int pte_dirty);
343 static void migrate_page_add(struct page *page, struct list_head *pagelist,
344 unsigned long flags);
346 /* Scan through pages checking if pages follow certain conditions. */
347 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
348 unsigned long addr, unsigned long end,
349 const nodemask_t *nodes, unsigned long flags,
356 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
361 if (!pte_present(*pte))
363 page = vm_normal_page(vma, addr, *pte);
367 * The check for PageReserved here is important to avoid
368 * handling zero pages and other pages that may have been
369 * marked special by the system.
371 * If the PageReserved would not be checked here then f.e.
372 * the location of the zero page could have an influence
373 * on MPOL_MF_STRICT, zero pages would be counted for
374 * the per node stats, and there would be useless attempts
375 * to put zero pages on the migration list.
377 if (PageReserved(page))
379 nid = page_to_nid(page);
380 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
383 if (flags & MPOL_MF_STATS)
384 gather_stats(page, private, pte_dirty(*pte));
385 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
386 migrate_page_add(page, private, flags);
389 } while (pte++, addr += PAGE_SIZE, addr != end);
390 pte_unmap_unlock(orig_pte, ptl);
394 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
395 unsigned long addr, unsigned long end,
396 const nodemask_t *nodes, unsigned long flags,
402 pmd = pmd_offset(pud, addr);
404 next = pmd_addr_end(addr, end);
405 if (pmd_none_or_clear_bad(pmd))
407 if (check_pte_range(vma, pmd, addr, next, nodes,
410 } while (pmd++, addr = next, addr != end);
414 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
415 unsigned long addr, unsigned long end,
416 const nodemask_t *nodes, unsigned long flags,
422 pud = pud_offset(pgd, addr);
424 next = pud_addr_end(addr, end);
425 if (pud_none_or_clear_bad(pud))
427 if (check_pmd_range(vma, pud, addr, next, nodes,
430 } while (pud++, addr = next, addr != end);
434 static inline int check_pgd_range(struct vm_area_struct *vma,
435 unsigned long addr, unsigned long end,
436 const nodemask_t *nodes, unsigned long flags,
442 pgd = pgd_offset(vma->vm_mm, addr);
444 next = pgd_addr_end(addr, end);
445 if (pgd_none_or_clear_bad(pgd))
447 if (check_pud_range(vma, pgd, addr, next, nodes,
450 } while (pgd++, addr = next, addr != end);
455 * Check if all pages in a range are on a set of nodes.
456 * If pagelist != NULL then isolate pages from the LRU and
457 * put them on the pagelist.
459 static struct vm_area_struct *
460 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
461 const nodemask_t *nodes, unsigned long flags, void *private)
464 struct vm_area_struct *first, *vma, *prev;
466 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
468 err = migrate_prep();
473 first = find_vma(mm, start);
475 return ERR_PTR(-EFAULT);
477 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
478 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
479 if (!vma->vm_next && vma->vm_end < end)
480 return ERR_PTR(-EFAULT);
481 if (prev && prev->vm_end < vma->vm_start)
482 return ERR_PTR(-EFAULT);
484 if (!is_vm_hugetlb_page(vma) &&
485 ((flags & MPOL_MF_STRICT) ||
486 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
487 vma_migratable(vma)))) {
488 unsigned long endvma = vma->vm_end;
492 if (vma->vm_start > start)
493 start = vma->vm_start;
494 err = check_pgd_range(vma, start, endvma, nodes,
497 first = ERR_PTR(err);
506 /* Apply policy to a single VMA */
507 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
510 struct mempolicy *old = vma->vm_policy;
512 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
513 vma->vm_start, vma->vm_end, vma->vm_pgoff,
514 vma->vm_ops, vma->vm_file,
515 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
517 if (vma->vm_ops && vma->vm_ops->set_policy)
518 err = vma->vm_ops->set_policy(vma, new);
521 vma->vm_policy = new;
527 /* Step 2: apply policy to a range and do splits. */
528 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
529 unsigned long end, struct mempolicy *new)
531 struct vm_area_struct *next;
535 for (; vma && vma->vm_start < end; vma = next) {
537 if (vma->vm_start < start)
538 err = split_vma(vma->vm_mm, vma, start, 1);
539 if (!err && vma->vm_end > end)
540 err = split_vma(vma->vm_mm, vma, end, 0);
542 err = policy_vma(vma, new);
550 * Update task->flags PF_MEMPOLICY bit: set iff non-default
551 * mempolicy. Allows more rapid checking of this (combined perhaps
552 * with other PF_* flag bits) on memory allocation hot code paths.
554 * If called from outside this file, the task 'p' should -only- be
555 * a newly forked child not yet visible on the task list, because
556 * manipulating the task flags of a visible task is not safe.
558 * The above limitation is why this routine has the funny name
559 * mpol_fix_fork_child_flag().
561 * It is also safe to call this with a task pointer of current,
562 * which the static wrapper mpol_set_task_struct_flag() does,
563 * for use within this file.
566 void mpol_fix_fork_child_flag(struct task_struct *p)
569 p->flags |= PF_MEMPOLICY;
571 p->flags &= ~PF_MEMPOLICY;
574 static void mpol_set_task_struct_flag(void)
576 mpol_fix_fork_child_flag(current);
579 /* Set the process memory policy */
580 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
583 struct mempolicy *new;
585 new = mpol_new(mode, flags, nodes);
588 mpol_free(current->mempolicy);
589 current->mempolicy = new;
590 mpol_set_task_struct_flag();
591 if (new && new->policy == MPOL_INTERLEAVE &&
592 nodes_weight(new->v.nodes))
593 current->il_next = first_node(new->v.nodes);
597 /* Fill a zone bitmap for a policy */
598 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
606 case MPOL_INTERLEAVE:
610 /* or use current node instead of memory_map? */
611 if (p->v.preferred_node < 0)
612 *nodes = node_states[N_HIGH_MEMORY];
614 node_set(p->v.preferred_node, *nodes);
621 static int lookup_node(struct mm_struct *mm, unsigned long addr)
626 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
628 err = page_to_nid(p);
634 /* Retrieve NUMA policy */
635 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
636 unsigned long addr, unsigned long flags)
639 struct mm_struct *mm = current->mm;
640 struct vm_area_struct *vma = NULL;
641 struct mempolicy *pol = current->mempolicy;
643 cpuset_update_task_memory_state();
645 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
648 if (flags & MPOL_F_MEMS_ALLOWED) {
649 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
651 *policy = 0; /* just so it's initialized */
652 *nmask = cpuset_current_mems_allowed;
656 if (flags & MPOL_F_ADDR) {
657 down_read(&mm->mmap_sem);
658 vma = find_vma_intersection(mm, addr, addr+1);
660 up_read(&mm->mmap_sem);
663 if (vma->vm_ops && vma->vm_ops->get_policy)
664 pol = vma->vm_ops->get_policy(vma, addr);
666 pol = vma->vm_policy;
671 pol = &default_policy;
673 if (flags & MPOL_F_NODE) {
674 if (flags & MPOL_F_ADDR) {
675 err = lookup_node(mm, addr);
679 } else if (pol == current->mempolicy &&
680 pol->policy == MPOL_INTERLEAVE) {
681 *policy = current->il_next;
687 *policy = pol->policy | pol->flags;
690 up_read(¤t->mm->mmap_sem);
696 get_zonemask(pol, nmask);
700 up_read(¤t->mm->mmap_sem);
704 #ifdef CONFIG_MIGRATION
708 static void migrate_page_add(struct page *page, struct list_head *pagelist,
712 * Avoid migrating a page that is shared with others.
714 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
715 isolate_lru_page(page, pagelist);
718 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
720 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
724 * Migrate pages from one node to a target node.
725 * Returns error or the number of pages not migrated.
727 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
735 node_set(source, nmask);
737 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
738 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
740 if (!list_empty(&pagelist))
741 err = migrate_pages(&pagelist, new_node_page, dest);
747 * Move pages between the two nodesets so as to preserve the physical
748 * layout as much as possible.
750 * Returns the number of page that could not be moved.
752 int do_migrate_pages(struct mm_struct *mm,
753 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
760 down_read(&mm->mmap_sem);
762 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
767 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
768 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
769 * bit in 'tmp', and return that <source, dest> pair for migration.
770 * The pair of nodemasks 'to' and 'from' define the map.
772 * If no pair of bits is found that way, fallback to picking some
773 * pair of 'source' and 'dest' bits that are not the same. If the
774 * 'source' and 'dest' bits are the same, this represents a node
775 * that will be migrating to itself, so no pages need move.
777 * If no bits are left in 'tmp', or if all remaining bits left
778 * in 'tmp' correspond to the same bit in 'to', return false
779 * (nothing left to migrate).
781 * This lets us pick a pair of nodes to migrate between, such that
782 * if possible the dest node is not already occupied by some other
783 * source node, minimizing the risk of overloading the memory on a
784 * node that would happen if we migrated incoming memory to a node
785 * before migrating outgoing memory source that same node.
787 * A single scan of tmp is sufficient. As we go, we remember the
788 * most recent <s, d> pair that moved (s != d). If we find a pair
789 * that not only moved, but what's better, moved to an empty slot
790 * (d is not set in tmp), then we break out then, with that pair.
791 * Otherwise when we finish scannng from_tmp, we at least have the
792 * most recent <s, d> pair that moved. If we get all the way through
793 * the scan of tmp without finding any node that moved, much less
794 * moved to an empty node, then there is nothing left worth migrating.
798 while (!nodes_empty(tmp)) {
803 for_each_node_mask(s, tmp) {
804 d = node_remap(s, *from_nodes, *to_nodes);
808 source = s; /* Node moved. Memorize */
811 /* dest not in remaining from nodes? */
812 if (!node_isset(dest, tmp))
818 node_clear(source, tmp);
819 err = migrate_to_node(mm, source, dest, flags);
826 up_read(&mm->mmap_sem);
834 * Allocate a new page for page migration based on vma policy.
835 * Start assuming that page is mapped by vma pointed to by @private.
836 * Search forward from there, if not. N.B., this assumes that the
837 * list of pages handed to migrate_pages()--which is how we get here--
838 * is in virtual address order.
840 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
842 struct vm_area_struct *vma = (struct vm_area_struct *)private;
843 unsigned long uninitialized_var(address);
846 address = page_address_in_vma(page, vma);
847 if (address != -EFAULT)
853 * if !vma, alloc_page_vma() will use task or system default policy
855 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
859 static void migrate_page_add(struct page *page, struct list_head *pagelist,
864 int do_migrate_pages(struct mm_struct *mm,
865 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
870 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
876 static long do_mbind(unsigned long start, unsigned long len,
877 unsigned short mode, unsigned short mode_flags,
878 nodemask_t *nmask, unsigned long flags)
880 struct vm_area_struct *vma;
881 struct mm_struct *mm = current->mm;
882 struct mempolicy *new;
887 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
888 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
890 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
893 if (start & ~PAGE_MASK)
896 if (mode == MPOL_DEFAULT)
897 flags &= ~MPOL_MF_STRICT;
899 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
907 new = mpol_new(mode, mode_flags, nmask);
912 * If we are using the default policy then operation
913 * on discontinuous address spaces is okay after all
916 flags |= MPOL_MF_DISCONTIG_OK;
918 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
919 start, start + len, mode, mode_flags,
920 nmask ? nodes_addr(*nmask)[0] : -1);
922 down_write(&mm->mmap_sem);
923 vma = check_range(mm, start, end, nmask,
924 flags | MPOL_MF_INVERT, &pagelist);
930 err = mbind_range(vma, start, end, new);
932 if (!list_empty(&pagelist))
933 nr_failed = migrate_pages(&pagelist, new_vma_page,
936 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
940 up_write(&mm->mmap_sem);
946 * User space interface with variable sized bitmaps for nodelists.
949 /* Copy a node mask from user space. */
950 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
951 unsigned long maxnode)
954 unsigned long nlongs;
955 unsigned long endmask;
959 if (maxnode == 0 || !nmask)
961 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
964 nlongs = BITS_TO_LONGS(maxnode);
965 if ((maxnode % BITS_PER_LONG) == 0)
968 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
970 /* When the user specified more nodes than supported just check
971 if the non supported part is all zero. */
972 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
973 if (nlongs > PAGE_SIZE/sizeof(long))
975 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
977 if (get_user(t, nmask + k))
979 if (k == nlongs - 1) {
985 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
989 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
991 nodes_addr(*nodes)[nlongs-1] &= endmask;
995 /* Copy a kernel node mask to user space */
996 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
999 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1000 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1002 if (copy > nbytes) {
1003 if (copy > PAGE_SIZE)
1005 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1009 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1012 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
1014 unsigned long __user *nmask, unsigned long maxnode,
1019 unsigned short mode_flags;
1021 mode_flags = mode & MPOL_MODE_FLAGS;
1022 mode &= ~MPOL_MODE_FLAGS;
1023 if (mode >= MPOL_MAX)
1025 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1026 (mode_flags & MPOL_F_RELATIVE_NODES))
1028 err = get_nodes(&nodes, nmask, maxnode);
1031 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1034 /* Set the process memory policy */
1035 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
1036 unsigned long maxnode)
1040 unsigned short flags;
1042 flags = mode & MPOL_MODE_FLAGS;
1043 mode &= ~MPOL_MODE_FLAGS;
1044 if ((unsigned int)mode >= MPOL_MAX)
1046 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1048 err = get_nodes(&nodes, nmask, maxnode);
1051 return do_set_mempolicy(mode, flags, &nodes);
1054 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
1055 const unsigned long __user *old_nodes,
1056 const unsigned long __user *new_nodes)
1058 struct mm_struct *mm;
1059 struct task_struct *task;
1062 nodemask_t task_nodes;
1065 err = get_nodes(&old, old_nodes, maxnode);
1069 err = get_nodes(&new, new_nodes, maxnode);
1073 /* Find the mm_struct */
1074 read_lock(&tasklist_lock);
1075 task = pid ? find_task_by_vpid(pid) : current;
1077 read_unlock(&tasklist_lock);
1080 mm = get_task_mm(task);
1081 read_unlock(&tasklist_lock);
1087 * Check if this process has the right to modify the specified
1088 * process. The right exists if the process has administrative
1089 * capabilities, superuser privileges or the same
1090 * userid as the target process.
1092 if ((current->euid != task->suid) && (current->euid != task->uid) &&
1093 (current->uid != task->suid) && (current->uid != task->uid) &&
1094 !capable(CAP_SYS_NICE)) {
1099 task_nodes = cpuset_mems_allowed(task);
1100 /* Is the user allowed to access the target nodes? */
1101 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1106 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1111 err = security_task_movememory(task);
1115 err = do_migrate_pages(mm, &old, &new,
1116 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1123 /* Retrieve NUMA policy */
1124 asmlinkage long sys_get_mempolicy(int __user *policy,
1125 unsigned long __user *nmask,
1126 unsigned long maxnode,
1127 unsigned long addr, unsigned long flags)
1130 int uninitialized_var(pval);
1133 if (nmask != NULL && maxnode < MAX_NUMNODES)
1136 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1141 if (policy && put_user(pval, policy))
1145 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1150 #ifdef CONFIG_COMPAT
1152 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1153 compat_ulong_t __user *nmask,
1154 compat_ulong_t maxnode,
1155 compat_ulong_t addr, compat_ulong_t flags)
1158 unsigned long __user *nm = NULL;
1159 unsigned long nr_bits, alloc_size;
1160 DECLARE_BITMAP(bm, MAX_NUMNODES);
1162 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1163 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1166 nm = compat_alloc_user_space(alloc_size);
1168 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1170 if (!err && nmask) {
1171 err = copy_from_user(bm, nm, alloc_size);
1172 /* ensure entire bitmap is zeroed */
1173 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1174 err |= compat_put_bitmap(nmask, bm, nr_bits);
1180 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1181 compat_ulong_t maxnode)
1184 unsigned long __user *nm = NULL;
1185 unsigned long nr_bits, alloc_size;
1186 DECLARE_BITMAP(bm, MAX_NUMNODES);
1188 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1189 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1192 err = compat_get_bitmap(bm, nmask, nr_bits);
1193 nm = compat_alloc_user_space(alloc_size);
1194 err |= copy_to_user(nm, bm, alloc_size);
1200 return sys_set_mempolicy(mode, nm, nr_bits+1);
1203 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1204 compat_ulong_t mode, compat_ulong_t __user *nmask,
1205 compat_ulong_t maxnode, compat_ulong_t flags)
1208 unsigned long __user *nm = NULL;
1209 unsigned long nr_bits, alloc_size;
1212 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1213 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1216 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1217 nm = compat_alloc_user_space(alloc_size);
1218 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1224 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1230 * get_vma_policy(@task, @vma, @addr)
1231 * @task - task for fallback if vma policy == default
1232 * @vma - virtual memory area whose policy is sought
1233 * @addr - address in @vma for shared policy lookup
1235 * Returns effective policy for a VMA at specified address.
1236 * Falls back to @task or system default policy, as necessary.
1237 * Returned policy has extra reference count if shared, vma,
1238 * or some other task's policy [show_numa_maps() can pass
1239 * @task != current]. It is the caller's responsibility to
1240 * free the reference in these cases.
1242 static struct mempolicy * get_vma_policy(struct task_struct *task,
1243 struct vm_area_struct *vma, unsigned long addr)
1245 struct mempolicy *pol = task->mempolicy;
1249 if (vma->vm_ops && vma->vm_ops->get_policy) {
1250 pol = vma->vm_ops->get_policy(vma, addr);
1251 shared_pol = 1; /* if pol non-NULL, add ref below */
1252 } else if (vma->vm_policy &&
1253 vma->vm_policy->policy != MPOL_DEFAULT)
1254 pol = vma->vm_policy;
1257 pol = &default_policy;
1258 else if (!shared_pol && pol != current->mempolicy)
1259 mpol_get(pol); /* vma or other task's policy */
1263 /* Return a nodemask representing a mempolicy */
1264 static nodemask_t *nodemask_policy(gfp_t gfp, struct mempolicy *policy)
1266 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1267 if (unlikely(policy->policy == MPOL_BIND) &&
1268 gfp_zone(gfp) >= policy_zone &&
1269 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1270 return &policy->v.nodes;
1275 /* Return a zonelist representing a mempolicy */
1276 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1280 switch (policy->policy) {
1281 case MPOL_PREFERRED:
1282 nd = policy->v.preferred_node;
1284 nd = numa_node_id();
1288 * Normally, MPOL_BIND allocations node-local are node-local
1289 * within the allowed nodemask. However, if __GFP_THISNODE is
1290 * set and the current node is part of the mask, we use the
1291 * the zonelist for the first node in the mask instead.
1293 nd = numa_node_id();
1294 if (unlikely(gfp & __GFP_THISNODE) &&
1295 unlikely(!node_isset(nd, policy->v.nodes)))
1296 nd = first_node(policy->v.nodes);
1298 case MPOL_INTERLEAVE: /* should not happen */
1300 nd = numa_node_id();
1306 return node_zonelist(nd, gfp);
1309 /* Do dynamic interleaving for a process */
1310 static unsigned interleave_nodes(struct mempolicy *policy)
1313 struct task_struct *me = current;
1316 next = next_node(nid, policy->v.nodes);
1317 if (next >= MAX_NUMNODES)
1318 next = first_node(policy->v.nodes);
1319 if (next < MAX_NUMNODES)
1325 * Depending on the memory policy provide a node from which to allocate the
1328 unsigned slab_node(struct mempolicy *policy)
1330 unsigned short pol = policy ? policy->policy : MPOL_DEFAULT;
1333 case MPOL_INTERLEAVE:
1334 return interleave_nodes(policy);
1338 * Follow bind policy behavior and start allocation at the
1341 struct zonelist *zonelist;
1343 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1344 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1345 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1351 case MPOL_PREFERRED:
1352 if (policy->v.preferred_node >= 0)
1353 return policy->v.preferred_node;
1357 return numa_node_id();
1361 /* Do static interleaving for a VMA with known offset. */
1362 static unsigned offset_il_node(struct mempolicy *pol,
1363 struct vm_area_struct *vma, unsigned long off)
1365 unsigned nnodes = nodes_weight(pol->v.nodes);
1371 return numa_node_id();
1372 target = (unsigned int)off % nnodes;
1375 nid = next_node(nid, pol->v.nodes);
1377 } while (c <= target);
1381 /* Determine a node number for interleave */
1382 static inline unsigned interleave_nid(struct mempolicy *pol,
1383 struct vm_area_struct *vma, unsigned long addr, int shift)
1389 * for small pages, there is no difference between
1390 * shift and PAGE_SHIFT, so the bit-shift is safe.
1391 * for huge pages, since vm_pgoff is in units of small
1392 * pages, we need to shift off the always 0 bits to get
1395 BUG_ON(shift < PAGE_SHIFT);
1396 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1397 off += (addr - vma->vm_start) >> shift;
1398 return offset_il_node(pol, vma, off);
1400 return interleave_nodes(pol);
1403 #ifdef CONFIG_HUGETLBFS
1405 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1406 * @vma = virtual memory area whose policy is sought
1407 * @addr = address in @vma for shared policy lookup and interleave policy
1408 * @gfp_flags = for requested zone
1409 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1410 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1412 * Returns a zonelist suitable for a huge page allocation.
1413 * If the effective policy is 'BIND, returns pointer to local node's zonelist,
1414 * and a pointer to the mempolicy's @nodemask for filtering the zonelist.
1415 * If it is also a policy for which get_vma_policy() returns an extra
1416 * reference, we must hold that reference until after the allocation.
1417 * In that case, return policy via @mpol so hugetlb allocation can drop
1418 * the reference. For non-'BIND referenced policies, we can/do drop the
1419 * reference here, so the caller doesn't need to know about the special case
1420 * for default and current task policy.
1422 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1423 gfp_t gfp_flags, struct mempolicy **mpol,
1424 nodemask_t **nodemask)
1426 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1427 struct zonelist *zl;
1429 *mpol = NULL; /* probably no unref needed */
1430 *nodemask = NULL; /* assume !MPOL_BIND */
1431 if (pol->policy == MPOL_BIND) {
1432 *nodemask = &pol->v.nodes;
1433 } else if (pol->policy == MPOL_INTERLEAVE) {
1436 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1437 if (unlikely(pol != &default_policy &&
1438 pol != current->mempolicy))
1439 __mpol_free(pol); /* finished with pol */
1440 return node_zonelist(nid, gfp_flags);
1443 zl = zonelist_policy(GFP_HIGHUSER, pol);
1444 if (unlikely(pol != &default_policy && pol != current->mempolicy)) {
1445 if (pol->policy != MPOL_BIND)
1446 __mpol_free(pol); /* finished with pol */
1448 *mpol = pol; /* unref needed after allocation */
1454 /* Allocate a page in interleaved policy.
1455 Own path because it needs to do special accounting. */
1456 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1459 struct zonelist *zl;
1462 zl = node_zonelist(nid, gfp);
1463 page = __alloc_pages(gfp, order, zl);
1464 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1465 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1470 * alloc_page_vma - Allocate a page for a VMA.
1473 * %GFP_USER user allocation.
1474 * %GFP_KERNEL kernel allocations,
1475 * %GFP_HIGHMEM highmem/user allocations,
1476 * %GFP_FS allocation should not call back into a file system.
1477 * %GFP_ATOMIC don't sleep.
1479 * @vma: Pointer to VMA or NULL if not available.
1480 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1482 * This function allocates a page from the kernel page pool and applies
1483 * a NUMA policy associated with the VMA or the current process.
1484 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1485 * mm_struct of the VMA to prevent it from going away. Should be used for
1486 * all allocations for pages that will be mapped into
1487 * user space. Returns NULL when no page can be allocated.
1489 * Should be called with the mm_sem of the vma hold.
1492 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1494 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1495 struct zonelist *zl;
1497 cpuset_update_task_memory_state();
1499 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1502 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1503 if (unlikely(pol != &default_policy &&
1504 pol != current->mempolicy))
1505 __mpol_free(pol); /* finished with pol */
1506 return alloc_page_interleave(gfp, 0, nid);
1508 zl = zonelist_policy(gfp, pol);
1509 if (pol != &default_policy && pol != current->mempolicy) {
1511 * slow path: ref counted policy -- shared or vma
1513 struct page *page = __alloc_pages_nodemask(gfp, 0,
1514 zl, nodemask_policy(gfp, pol));
1519 * fast path: default or task policy
1521 return __alloc_pages_nodemask(gfp, 0, zl, nodemask_policy(gfp, pol));
1525 * alloc_pages_current - Allocate pages.
1528 * %GFP_USER user allocation,
1529 * %GFP_KERNEL kernel allocation,
1530 * %GFP_HIGHMEM highmem allocation,
1531 * %GFP_FS don't call back into a file system.
1532 * %GFP_ATOMIC don't sleep.
1533 * @order: Power of two of allocation size in pages. 0 is a single page.
1535 * Allocate a page from the kernel page pool. When not in
1536 * interrupt context and apply the current process NUMA policy.
1537 * Returns NULL when no page can be allocated.
1539 * Don't call cpuset_update_task_memory_state() unless
1540 * 1) it's ok to take cpuset_sem (can WAIT), and
1541 * 2) allocating for current task (not interrupt).
1543 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1545 struct mempolicy *pol = current->mempolicy;
1547 if ((gfp & __GFP_WAIT) && !in_interrupt())
1548 cpuset_update_task_memory_state();
1549 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1550 pol = &default_policy;
1551 if (pol->policy == MPOL_INTERLEAVE)
1552 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1553 return __alloc_pages_nodemask(gfp, order,
1554 zonelist_policy(gfp, pol), nodemask_policy(gfp, pol));
1556 EXPORT_SYMBOL(alloc_pages_current);
1559 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1560 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1561 * with the mems_allowed returned by cpuset_mems_allowed(). This
1562 * keeps mempolicies cpuset relative after its cpuset moves. See
1563 * further kernel/cpuset.c update_nodemask().
1566 /* Slow path of a mempolicy copy */
1567 struct mempolicy *__mpol_copy(struct mempolicy *old)
1569 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1572 return ERR_PTR(-ENOMEM);
1573 if (current_cpuset_is_being_rebound()) {
1574 nodemask_t mems = cpuset_mems_allowed(current);
1575 mpol_rebind_policy(old, &mems);
1578 atomic_set(&new->refcnt, 1);
1582 static int mpol_match_intent(const struct mempolicy *a,
1583 const struct mempolicy *b)
1585 if (a->flags != b->flags)
1587 if (!mpol_store_user_nodemask(a))
1589 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1592 /* Slow path of a mempolicy comparison */
1593 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1597 if (a->policy != b->policy)
1599 if (a->policy != MPOL_DEFAULT && !mpol_match_intent(a, b))
1601 switch (a->policy) {
1606 case MPOL_INTERLEAVE:
1607 return nodes_equal(a->v.nodes, b->v.nodes);
1608 case MPOL_PREFERRED:
1609 return a->v.preferred_node == b->v.preferred_node;
1616 /* Slow path of a mpol destructor. */
1617 void __mpol_free(struct mempolicy *p)
1619 if (!atomic_dec_and_test(&p->refcnt))
1621 p->policy = MPOL_DEFAULT;
1622 kmem_cache_free(policy_cache, p);
1626 * Shared memory backing store policy support.
1628 * Remember policies even when nobody has shared memory mapped.
1629 * The policies are kept in Red-Black tree linked from the inode.
1630 * They are protected by the sp->lock spinlock, which should be held
1631 * for any accesses to the tree.
1634 /* lookup first element intersecting start-end */
1635 /* Caller holds sp->lock */
1636 static struct sp_node *
1637 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1639 struct rb_node *n = sp->root.rb_node;
1642 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1644 if (start >= p->end)
1646 else if (end <= p->start)
1654 struct sp_node *w = NULL;
1655 struct rb_node *prev = rb_prev(n);
1658 w = rb_entry(prev, struct sp_node, nd);
1659 if (w->end <= start)
1663 return rb_entry(n, struct sp_node, nd);
1666 /* Insert a new shared policy into the list. */
1667 /* Caller holds sp->lock */
1668 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1670 struct rb_node **p = &sp->root.rb_node;
1671 struct rb_node *parent = NULL;
1676 nd = rb_entry(parent, struct sp_node, nd);
1677 if (new->start < nd->start)
1679 else if (new->end > nd->end)
1680 p = &(*p)->rb_right;
1684 rb_link_node(&new->nd, parent, p);
1685 rb_insert_color(&new->nd, &sp->root);
1686 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1687 new->policy ? new->policy->policy : 0);
1690 /* Find shared policy intersecting idx */
1692 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1694 struct mempolicy *pol = NULL;
1697 if (!sp->root.rb_node)
1699 spin_lock(&sp->lock);
1700 sn = sp_lookup(sp, idx, idx+1);
1702 mpol_get(sn->policy);
1705 spin_unlock(&sp->lock);
1709 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1711 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1712 rb_erase(&n->nd, &sp->root);
1713 mpol_free(n->policy);
1714 kmem_cache_free(sn_cache, n);
1717 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1718 struct mempolicy *pol)
1720 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1731 /* Replace a policy range. */
1732 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1733 unsigned long end, struct sp_node *new)
1735 struct sp_node *n, *new2 = NULL;
1738 spin_lock(&sp->lock);
1739 n = sp_lookup(sp, start, end);
1740 /* Take care of old policies in the same range. */
1741 while (n && n->start < end) {
1742 struct rb_node *next = rb_next(&n->nd);
1743 if (n->start >= start) {
1749 /* Old policy spanning whole new range. */
1752 spin_unlock(&sp->lock);
1753 new2 = sp_alloc(end, n->end, n->policy);
1759 sp_insert(sp, new2);
1767 n = rb_entry(next, struct sp_node, nd);
1771 spin_unlock(&sp->lock);
1773 mpol_free(new2->policy);
1774 kmem_cache_free(sn_cache, new2);
1779 void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
1780 unsigned short flags, nodemask_t *policy_nodes)
1782 info->root = RB_ROOT;
1783 spin_lock_init(&info->lock);
1785 if (policy != MPOL_DEFAULT) {
1786 struct mempolicy *newpol;
1788 /* Falls back to MPOL_DEFAULT on any error */
1789 newpol = mpol_new(policy, flags, policy_nodes);
1790 if (!IS_ERR(newpol)) {
1791 /* Create pseudo-vma that contains just the policy */
1792 struct vm_area_struct pvma;
1794 memset(&pvma, 0, sizeof(struct vm_area_struct));
1795 /* Policy covers entire file */
1796 pvma.vm_end = TASK_SIZE;
1797 mpol_set_shared_policy(info, &pvma, newpol);
1803 int mpol_set_shared_policy(struct shared_policy *info,
1804 struct vm_area_struct *vma, struct mempolicy *npol)
1807 struct sp_node *new = NULL;
1808 unsigned long sz = vma_pages(vma);
1810 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1812 sz, npol ? npol->policy : -1,
1813 npol ? npol->flags : -1,
1814 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1817 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1821 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1823 kmem_cache_free(sn_cache, new);
1827 /* Free a backing policy store on inode delete. */
1828 void mpol_free_shared_policy(struct shared_policy *p)
1831 struct rb_node *next;
1833 if (!p->root.rb_node)
1835 spin_lock(&p->lock);
1836 next = rb_first(&p->root);
1838 n = rb_entry(next, struct sp_node, nd);
1839 next = rb_next(&n->nd);
1840 rb_erase(&n->nd, &p->root);
1841 mpol_free(n->policy);
1842 kmem_cache_free(sn_cache, n);
1844 spin_unlock(&p->lock);
1847 /* assumes fs == KERNEL_DS */
1848 void __init numa_policy_init(void)
1850 nodemask_t interleave_nodes;
1851 unsigned long largest = 0;
1852 int nid, prefer = 0;
1854 policy_cache = kmem_cache_create("numa_policy",
1855 sizeof(struct mempolicy),
1856 0, SLAB_PANIC, NULL);
1858 sn_cache = kmem_cache_create("shared_policy_node",
1859 sizeof(struct sp_node),
1860 0, SLAB_PANIC, NULL);
1863 * Set interleaving policy for system init. Interleaving is only
1864 * enabled across suitably sized nodes (default is >= 16MB), or
1865 * fall back to the largest node if they're all smaller.
1867 nodes_clear(interleave_nodes);
1868 for_each_node_state(nid, N_HIGH_MEMORY) {
1869 unsigned long total_pages = node_present_pages(nid);
1871 /* Preserve the largest node */
1872 if (largest < total_pages) {
1873 largest = total_pages;
1877 /* Interleave this node? */
1878 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1879 node_set(nid, interleave_nodes);
1882 /* All too small, use the largest */
1883 if (unlikely(nodes_empty(interleave_nodes)))
1884 node_set(prefer, interleave_nodes);
1886 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
1887 printk("numa_policy_init: interleaving failed\n");
1890 /* Reset policy of current process to default */
1891 void numa_default_policy(void)
1893 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1897 * Display pages allocated per node and memory policy via /proc.
1899 static const char * const policy_types[] =
1900 { "default", "prefer", "bind", "interleave" };
1903 * Convert a mempolicy into a string.
1904 * Returns the number of characters in buffer (if positive)
1905 * or an error (negative)
1907 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1912 unsigned short mode = pol ? pol->policy : MPOL_DEFAULT;
1913 unsigned short flags = pol ? pol->flags : 0;
1920 case MPOL_PREFERRED:
1922 node_set(pol->v.preferred_node, nodes);
1927 case MPOL_INTERLEAVE:
1928 nodes = pol->v.nodes;
1936 l = strlen(policy_types[mode]);
1937 if (buffer + maxlen < p + l + 1)
1940 strcpy(p, policy_types[mode]);
1946 if (buffer + maxlen < p + 2)
1950 if (flags & MPOL_F_STATIC_NODES)
1951 p += sprintf(p, "%sstatic", need_bar++ ? "|" : "");
1952 if (flags & MPOL_F_RELATIVE_NODES)
1953 p += sprintf(p, "%srelative", need_bar++ ? "|" : "");
1956 if (!nodes_empty(nodes)) {
1957 if (buffer + maxlen < p + 2)
1960 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1966 unsigned long pages;
1968 unsigned long active;
1969 unsigned long writeback;
1970 unsigned long mapcount_max;
1971 unsigned long dirty;
1972 unsigned long swapcache;
1973 unsigned long node[MAX_NUMNODES];
1976 static void gather_stats(struct page *page, void *private, int pte_dirty)
1978 struct numa_maps *md = private;
1979 int count = page_mapcount(page);
1982 if (pte_dirty || PageDirty(page))
1985 if (PageSwapCache(page))
1988 if (PageActive(page))
1991 if (PageWriteback(page))
1997 if (count > md->mapcount_max)
1998 md->mapcount_max = count;
2000 md->node[page_to_nid(page)]++;
2003 #ifdef CONFIG_HUGETLB_PAGE
2004 static void check_huge_range(struct vm_area_struct *vma,
2005 unsigned long start, unsigned long end,
2006 struct numa_maps *md)
2011 for (addr = start; addr < end; addr += HPAGE_SIZE) {
2012 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
2022 page = pte_page(pte);
2026 gather_stats(page, md, pte_dirty(*ptep));
2030 static inline void check_huge_range(struct vm_area_struct *vma,
2031 unsigned long start, unsigned long end,
2032 struct numa_maps *md)
2037 int show_numa_map(struct seq_file *m, void *v)
2039 struct proc_maps_private *priv = m->private;
2040 struct vm_area_struct *vma = v;
2041 struct numa_maps *md;
2042 struct file *file = vma->vm_file;
2043 struct mm_struct *mm = vma->vm_mm;
2044 struct mempolicy *pol;
2051 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2055 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2056 mpol_to_str(buffer, sizeof(buffer), pol);
2058 * unref shared or other task's mempolicy
2060 if (pol != &default_policy && pol != current->mempolicy)
2063 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2066 seq_printf(m, " file=");
2067 seq_path(m, &file->f_path, "\n\t= ");
2068 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2069 seq_printf(m, " heap");
2070 } else if (vma->vm_start <= mm->start_stack &&
2071 vma->vm_end >= mm->start_stack) {
2072 seq_printf(m, " stack");
2075 if (is_vm_hugetlb_page(vma)) {
2076 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2077 seq_printf(m, " huge");
2079 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2080 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2087 seq_printf(m," anon=%lu",md->anon);
2090 seq_printf(m," dirty=%lu",md->dirty);
2092 if (md->pages != md->anon && md->pages != md->dirty)
2093 seq_printf(m, " mapped=%lu", md->pages);
2095 if (md->mapcount_max > 1)
2096 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2099 seq_printf(m," swapcache=%lu", md->swapcache);
2101 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2102 seq_printf(m," active=%lu", md->active);
2105 seq_printf(m," writeback=%lu", md->writeback);
2107 for_each_node_state(n, N_HIGH_MEMORY)
2109 seq_printf(m, " N%d=%lu", n, md->node[n]);
2114 if (m->count < m->size)
2115 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;