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;
108 * run-time system-wide default policy => local allocation
110 struct mempolicy default_policy = {
111 .refcnt = ATOMIC_INIT(1), /* never free it */
112 .mode = MPOL_PREFERRED,
113 .v = { .preferred_node = -1 },
116 static const struct mempolicy_operations {
117 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
118 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
119 } mpol_ops[MPOL_MAX];
121 /* Check that the nodemask contains at least one populated zone */
122 static int is_valid_nodemask(const nodemask_t *nodemask)
126 /* Check that there is something useful in this mask */
129 for_each_node_mask(nd, *nodemask) {
132 for (k = 0; k <= policy_zone; k++) {
133 z = &NODE_DATA(nd)->node_zones[k];
134 if (z->present_pages > 0)
142 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
144 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
147 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
148 const nodemask_t *rel)
151 nodes_fold(tmp, *orig, nodes_weight(*rel));
152 nodes_onto(*ret, tmp, *rel);
155 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
157 if (nodes_empty(*nodes))
159 pol->v.nodes = *nodes;
163 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
166 pol->v.preferred_node = -1; /* local allocation */
167 else if (nodes_empty(*nodes))
168 return -EINVAL; /* no allowed nodes */
170 pol->v.preferred_node = first_node(*nodes);
174 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
176 if (!is_valid_nodemask(nodes))
178 pol->v.nodes = *nodes;
182 /* Create a new policy */
183 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
186 struct mempolicy *policy;
187 nodemask_t cpuset_context_nmask;
190 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
191 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
193 if (mode == MPOL_DEFAULT) {
194 if (nodes && !nodes_empty(*nodes))
195 return ERR_PTR(-EINVAL);
196 return NULL; /* simply delete any existing policy */
201 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
202 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
203 * All other modes require a valid pointer to a non-empty nodemask.
205 if (mode == MPOL_PREFERRED) {
206 if (nodes_empty(*nodes)) {
207 if (((flags & MPOL_F_STATIC_NODES) ||
208 (flags & MPOL_F_RELATIVE_NODES)))
209 return ERR_PTR(-EINVAL);
210 nodes = NULL; /* flag local alloc */
212 } else if (nodes_empty(*nodes))
213 return ERR_PTR(-EINVAL);
214 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
216 return ERR_PTR(-ENOMEM);
217 atomic_set(&policy->refcnt, 1);
219 policy->flags = flags;
223 * cpuset related setup doesn't apply to local allocation
225 cpuset_update_task_memory_state();
226 if (flags & MPOL_F_RELATIVE_NODES)
227 mpol_relative_nodemask(&cpuset_context_nmask, nodes,
228 &cpuset_current_mems_allowed);
230 nodes_and(cpuset_context_nmask, *nodes,
231 cpuset_current_mems_allowed);
232 if (mpol_store_user_nodemask(policy))
233 policy->w.user_nodemask = *nodes;
235 policy->w.cpuset_mems_allowed =
236 cpuset_mems_allowed(current);
239 ret = mpol_ops[mode].create(policy,
240 nodes ? &cpuset_context_nmask : NULL);
242 kmem_cache_free(policy_cache, policy);
248 /* Slow path of a mpol destructor. */
249 void __mpol_put(struct mempolicy *p)
251 if (!atomic_dec_and_test(&p->refcnt))
253 kmem_cache_free(policy_cache, p);
256 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
260 static void mpol_rebind_nodemask(struct mempolicy *pol,
261 const nodemask_t *nodes)
265 if (pol->flags & MPOL_F_STATIC_NODES)
266 nodes_and(tmp, pol->w.user_nodemask, *nodes);
267 else if (pol->flags & MPOL_F_RELATIVE_NODES)
268 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
270 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
272 pol->w.cpuset_mems_allowed = *nodes;
276 if (!node_isset(current->il_next, tmp)) {
277 current->il_next = next_node(current->il_next, tmp);
278 if (current->il_next >= MAX_NUMNODES)
279 current->il_next = first_node(tmp);
280 if (current->il_next >= MAX_NUMNODES)
281 current->il_next = numa_node_id();
285 static void mpol_rebind_preferred(struct mempolicy *pol,
286 const nodemask_t *nodes)
290 if (pol->flags & MPOL_F_STATIC_NODES) {
291 int node = first_node(pol->w.user_nodemask);
293 if (node_isset(node, *nodes))
294 pol->v.preferred_node = node;
296 pol->v.preferred_node = -1;
297 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
298 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
299 pol->v.preferred_node = first_node(tmp);
300 } else if (pol->v.preferred_node != -1) {
301 pol->v.preferred_node = node_remap(pol->v.preferred_node,
302 pol->w.cpuset_mems_allowed,
304 pol->w.cpuset_mems_allowed = *nodes;
308 /* Migrate a policy to a different set of nodes */
309 static void mpol_rebind_policy(struct mempolicy *pol,
310 const nodemask_t *newmask)
314 if (!mpol_store_user_nodemask(pol) &&
315 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
317 mpol_ops[pol->mode].rebind(pol, newmask);
321 * Wrapper for mpol_rebind_policy() that just requires task
322 * pointer, and updates task mempolicy.
325 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
327 mpol_rebind_policy(tsk->mempolicy, new);
331 * Rebind each vma in mm to new nodemask.
333 * Call holding a reference to mm. Takes mm->mmap_sem during call.
336 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
338 struct vm_area_struct *vma;
340 down_write(&mm->mmap_sem);
341 for (vma = mm->mmap; vma; vma = vma->vm_next)
342 mpol_rebind_policy(vma->vm_policy, new);
343 up_write(&mm->mmap_sem);
346 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
348 .rebind = mpol_rebind_default,
350 [MPOL_INTERLEAVE] = {
351 .create = mpol_new_interleave,
352 .rebind = mpol_rebind_nodemask,
355 .create = mpol_new_preferred,
356 .rebind = mpol_rebind_preferred,
359 .create = mpol_new_bind,
360 .rebind = mpol_rebind_nodemask,
364 static void gather_stats(struct page *, void *, int pte_dirty);
365 static void migrate_page_add(struct page *page, struct list_head *pagelist,
366 unsigned long flags);
368 /* Scan through pages checking if pages follow certain conditions. */
369 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
370 unsigned long addr, unsigned long end,
371 const nodemask_t *nodes, unsigned long flags,
378 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
383 if (!pte_present(*pte))
385 page = vm_normal_page(vma, addr, *pte);
389 * The check for PageReserved here is important to avoid
390 * handling zero pages and other pages that may have been
391 * marked special by the system.
393 * If the PageReserved would not be checked here then f.e.
394 * the location of the zero page could have an influence
395 * on MPOL_MF_STRICT, zero pages would be counted for
396 * the per node stats, and there would be useless attempts
397 * to put zero pages on the migration list.
399 if (PageReserved(page))
401 nid = page_to_nid(page);
402 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
405 if (flags & MPOL_MF_STATS)
406 gather_stats(page, private, pte_dirty(*pte));
407 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
408 migrate_page_add(page, private, flags);
411 } while (pte++, addr += PAGE_SIZE, addr != end);
412 pte_unmap_unlock(orig_pte, ptl);
416 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
417 unsigned long addr, unsigned long end,
418 const nodemask_t *nodes, unsigned long flags,
424 pmd = pmd_offset(pud, addr);
426 next = pmd_addr_end(addr, end);
427 if (pmd_none_or_clear_bad(pmd))
429 if (check_pte_range(vma, pmd, addr, next, nodes,
432 } while (pmd++, addr = next, addr != end);
436 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
437 unsigned long addr, unsigned long end,
438 const nodemask_t *nodes, unsigned long flags,
444 pud = pud_offset(pgd, addr);
446 next = pud_addr_end(addr, end);
447 if (pud_none_or_clear_bad(pud))
449 if (check_pmd_range(vma, pud, addr, next, nodes,
452 } while (pud++, addr = next, addr != end);
456 static inline int check_pgd_range(struct vm_area_struct *vma,
457 unsigned long addr, unsigned long end,
458 const nodemask_t *nodes, unsigned long flags,
464 pgd = pgd_offset(vma->vm_mm, addr);
466 next = pgd_addr_end(addr, end);
467 if (pgd_none_or_clear_bad(pgd))
469 if (check_pud_range(vma, pgd, addr, next, nodes,
472 } while (pgd++, addr = next, addr != end);
477 * Check if all pages in a range are on a set of nodes.
478 * If pagelist != NULL then isolate pages from the LRU and
479 * put them on the pagelist.
481 static struct vm_area_struct *
482 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
483 const nodemask_t *nodes, unsigned long flags, void *private)
486 struct vm_area_struct *first, *vma, *prev;
488 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
490 err = migrate_prep();
495 first = find_vma(mm, start);
497 return ERR_PTR(-EFAULT);
499 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
500 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
501 if (!vma->vm_next && vma->vm_end < end)
502 return ERR_PTR(-EFAULT);
503 if (prev && prev->vm_end < vma->vm_start)
504 return ERR_PTR(-EFAULT);
506 if (!is_vm_hugetlb_page(vma) &&
507 ((flags & MPOL_MF_STRICT) ||
508 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
509 vma_migratable(vma)))) {
510 unsigned long endvma = vma->vm_end;
514 if (vma->vm_start > start)
515 start = vma->vm_start;
516 err = check_pgd_range(vma, start, endvma, nodes,
519 first = ERR_PTR(err);
528 /* Apply policy to a single VMA */
529 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
532 struct mempolicy *old = vma->vm_policy;
534 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
535 vma->vm_start, vma->vm_end, vma->vm_pgoff,
536 vma->vm_ops, vma->vm_file,
537 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
539 if (vma->vm_ops && vma->vm_ops->set_policy)
540 err = vma->vm_ops->set_policy(vma, new);
543 vma->vm_policy = new;
549 /* Step 2: apply policy to a range and do splits. */
550 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
551 unsigned long end, struct mempolicy *new)
553 struct vm_area_struct *next;
557 for (; vma && vma->vm_start < end; vma = next) {
559 if (vma->vm_start < start)
560 err = split_vma(vma->vm_mm, vma, start, 1);
561 if (!err && vma->vm_end > end)
562 err = split_vma(vma->vm_mm, vma, end, 0);
564 err = policy_vma(vma, new);
572 * Update task->flags PF_MEMPOLICY bit: set iff non-default
573 * mempolicy. Allows more rapid checking of this (combined perhaps
574 * with other PF_* flag bits) on memory allocation hot code paths.
576 * If called from outside this file, the task 'p' should -only- be
577 * a newly forked child not yet visible on the task list, because
578 * manipulating the task flags of a visible task is not safe.
580 * The above limitation is why this routine has the funny name
581 * mpol_fix_fork_child_flag().
583 * It is also safe to call this with a task pointer of current,
584 * which the static wrapper mpol_set_task_struct_flag() does,
585 * for use within this file.
588 void mpol_fix_fork_child_flag(struct task_struct *p)
591 p->flags |= PF_MEMPOLICY;
593 p->flags &= ~PF_MEMPOLICY;
596 static void mpol_set_task_struct_flag(void)
598 mpol_fix_fork_child_flag(current);
601 /* Set the process memory policy */
602 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
605 struct mempolicy *new;
606 struct mm_struct *mm = current->mm;
608 new = mpol_new(mode, flags, nodes);
613 * prevent changing our mempolicy while show_numa_maps()
615 * Note: do_set_mempolicy() can be called at init time
619 down_write(&mm->mmap_sem);
620 mpol_put(current->mempolicy);
621 current->mempolicy = new;
622 mpol_set_task_struct_flag();
623 if (new && new->mode == MPOL_INTERLEAVE &&
624 nodes_weight(new->v.nodes))
625 current->il_next = first_node(new->v.nodes);
627 up_write(&mm->mmap_sem);
633 * Return nodemask for policy for get_mempolicy() query
635 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
638 if (p == &default_policy)
644 case MPOL_INTERLEAVE:
648 if (p->v.preferred_node >= 0)
649 node_set(p->v.preferred_node, *nodes);
650 /* else return empty node mask for local allocation */
657 static int lookup_node(struct mm_struct *mm, unsigned long addr)
662 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
664 err = page_to_nid(p);
670 /* Retrieve NUMA policy */
671 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
672 unsigned long addr, unsigned long flags)
675 struct mm_struct *mm = current->mm;
676 struct vm_area_struct *vma = NULL;
677 struct mempolicy *pol = current->mempolicy;
679 cpuset_update_task_memory_state();
681 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
684 if (flags & MPOL_F_MEMS_ALLOWED) {
685 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
687 *policy = 0; /* just so it's initialized */
688 *nmask = cpuset_current_mems_allowed;
692 if (flags & MPOL_F_ADDR) {
694 * Do NOT fall back to task policy if the
695 * vma/shared policy at addr is NULL. We
696 * want to return MPOL_DEFAULT in this case.
698 down_read(&mm->mmap_sem);
699 vma = find_vma_intersection(mm, addr, addr+1);
701 up_read(&mm->mmap_sem);
704 if (vma->vm_ops && vma->vm_ops->get_policy)
705 pol = vma->vm_ops->get_policy(vma, addr);
707 pol = vma->vm_policy;
712 pol = &default_policy; /* indicates default behavior */
714 if (flags & MPOL_F_NODE) {
715 if (flags & MPOL_F_ADDR) {
716 err = lookup_node(mm, addr);
720 } else if (pol == current->mempolicy &&
721 pol->mode == MPOL_INTERLEAVE) {
722 *policy = current->il_next;
728 *policy = pol == &default_policy ? MPOL_DEFAULT :
730 *policy |= pol->flags;
734 up_read(¤t->mm->mmap_sem);
740 get_policy_nodemask(pol, nmask);
745 up_read(¤t->mm->mmap_sem);
749 #ifdef CONFIG_MIGRATION
753 static void migrate_page_add(struct page *page, struct list_head *pagelist,
757 * Avoid migrating a page that is shared with others.
759 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
760 isolate_lru_page(page, pagelist);
763 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
765 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
769 * Migrate pages from one node to a target node.
770 * Returns error or the number of pages not migrated.
772 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
780 node_set(source, nmask);
782 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
783 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
785 if (!list_empty(&pagelist))
786 err = migrate_pages(&pagelist, new_node_page, dest);
792 * Move pages between the two nodesets so as to preserve the physical
793 * layout as much as possible.
795 * Returns the number of page that could not be moved.
797 int do_migrate_pages(struct mm_struct *mm,
798 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
805 down_read(&mm->mmap_sem);
807 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
812 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
813 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
814 * bit in 'tmp', and return that <source, dest> pair for migration.
815 * The pair of nodemasks 'to' and 'from' define the map.
817 * If no pair of bits is found that way, fallback to picking some
818 * pair of 'source' and 'dest' bits that are not the same. If the
819 * 'source' and 'dest' bits are the same, this represents a node
820 * that will be migrating to itself, so no pages need move.
822 * If no bits are left in 'tmp', or if all remaining bits left
823 * in 'tmp' correspond to the same bit in 'to', return false
824 * (nothing left to migrate).
826 * This lets us pick a pair of nodes to migrate between, such that
827 * if possible the dest node is not already occupied by some other
828 * source node, minimizing the risk of overloading the memory on a
829 * node that would happen if we migrated incoming memory to a node
830 * before migrating outgoing memory source that same node.
832 * A single scan of tmp is sufficient. As we go, we remember the
833 * most recent <s, d> pair that moved (s != d). If we find a pair
834 * that not only moved, but what's better, moved to an empty slot
835 * (d is not set in tmp), then we break out then, with that pair.
836 * Otherwise when we finish scannng from_tmp, we at least have the
837 * most recent <s, d> pair that moved. If we get all the way through
838 * the scan of tmp without finding any node that moved, much less
839 * moved to an empty node, then there is nothing left worth migrating.
843 while (!nodes_empty(tmp)) {
848 for_each_node_mask(s, tmp) {
849 d = node_remap(s, *from_nodes, *to_nodes);
853 source = s; /* Node moved. Memorize */
856 /* dest not in remaining from nodes? */
857 if (!node_isset(dest, tmp))
863 node_clear(source, tmp);
864 err = migrate_to_node(mm, source, dest, flags);
871 up_read(&mm->mmap_sem);
879 * Allocate a new page for page migration based on vma policy.
880 * Start assuming that page is mapped by vma pointed to by @private.
881 * Search forward from there, if not. N.B., this assumes that the
882 * list of pages handed to migrate_pages()--which is how we get here--
883 * is in virtual address order.
885 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
887 struct vm_area_struct *vma = (struct vm_area_struct *)private;
888 unsigned long uninitialized_var(address);
891 address = page_address_in_vma(page, vma);
892 if (address != -EFAULT)
898 * if !vma, alloc_page_vma() will use task or system default policy
900 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
904 static void migrate_page_add(struct page *page, struct list_head *pagelist,
909 int do_migrate_pages(struct mm_struct *mm,
910 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
915 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
921 static long do_mbind(unsigned long start, unsigned long len,
922 unsigned short mode, unsigned short mode_flags,
923 nodemask_t *nmask, unsigned long flags)
925 struct vm_area_struct *vma;
926 struct mm_struct *mm = current->mm;
927 struct mempolicy *new;
932 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
933 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
935 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
938 if (start & ~PAGE_MASK)
941 if (mode == MPOL_DEFAULT)
942 flags &= ~MPOL_MF_STRICT;
944 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
952 new = mpol_new(mode, mode_flags, nmask);
957 * If we are using the default policy then operation
958 * on discontinuous address spaces is okay after all
961 flags |= MPOL_MF_DISCONTIG_OK;
963 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
964 start, start + len, mode, mode_flags,
965 nmask ? nodes_addr(*nmask)[0] : -1);
967 down_write(&mm->mmap_sem);
968 vma = check_range(mm, start, end, nmask,
969 flags | MPOL_MF_INVERT, &pagelist);
975 err = mbind_range(vma, start, end, new);
977 if (!list_empty(&pagelist))
978 nr_failed = migrate_pages(&pagelist, new_vma_page,
981 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
985 up_write(&mm->mmap_sem);
991 * User space interface with variable sized bitmaps for nodelists.
994 /* Copy a node mask from user space. */
995 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
996 unsigned long maxnode)
999 unsigned long nlongs;
1000 unsigned long endmask;
1003 nodes_clear(*nodes);
1004 if (maxnode == 0 || !nmask)
1006 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1009 nlongs = BITS_TO_LONGS(maxnode);
1010 if ((maxnode % BITS_PER_LONG) == 0)
1013 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1015 /* When the user specified more nodes than supported just check
1016 if the non supported part is all zero. */
1017 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1018 if (nlongs > PAGE_SIZE/sizeof(long))
1020 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1022 if (get_user(t, nmask + k))
1024 if (k == nlongs - 1) {
1030 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1034 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1036 nodes_addr(*nodes)[nlongs-1] &= endmask;
1040 /* Copy a kernel node mask to user space */
1041 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1044 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1045 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1047 if (copy > nbytes) {
1048 if (copy > PAGE_SIZE)
1050 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1054 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1057 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
1059 unsigned long __user *nmask, unsigned long maxnode,
1064 unsigned short mode_flags;
1066 mode_flags = mode & MPOL_MODE_FLAGS;
1067 mode &= ~MPOL_MODE_FLAGS;
1068 if (mode >= MPOL_MAX)
1070 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1071 (mode_flags & MPOL_F_RELATIVE_NODES))
1073 err = get_nodes(&nodes, nmask, maxnode);
1076 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1079 /* Set the process memory policy */
1080 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
1081 unsigned long maxnode)
1085 unsigned short flags;
1087 flags = mode & MPOL_MODE_FLAGS;
1088 mode &= ~MPOL_MODE_FLAGS;
1089 if ((unsigned int)mode >= MPOL_MAX)
1091 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1093 err = get_nodes(&nodes, nmask, maxnode);
1096 return do_set_mempolicy(mode, flags, &nodes);
1099 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
1100 const unsigned long __user *old_nodes,
1101 const unsigned long __user *new_nodes)
1103 struct mm_struct *mm;
1104 struct task_struct *task;
1107 nodemask_t task_nodes;
1110 err = get_nodes(&old, old_nodes, maxnode);
1114 err = get_nodes(&new, new_nodes, maxnode);
1118 /* Find the mm_struct */
1119 read_lock(&tasklist_lock);
1120 task = pid ? find_task_by_vpid(pid) : current;
1122 read_unlock(&tasklist_lock);
1125 mm = get_task_mm(task);
1126 read_unlock(&tasklist_lock);
1132 * Check if this process has the right to modify the specified
1133 * process. The right exists if the process has administrative
1134 * capabilities, superuser privileges or the same
1135 * userid as the target process.
1137 if ((current->euid != task->suid) && (current->euid != task->uid) &&
1138 (current->uid != task->suid) && (current->uid != task->uid) &&
1139 !capable(CAP_SYS_NICE)) {
1144 task_nodes = cpuset_mems_allowed(task);
1145 /* Is the user allowed to access the target nodes? */
1146 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1151 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1156 err = security_task_movememory(task);
1160 err = do_migrate_pages(mm, &old, &new,
1161 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1168 /* Retrieve NUMA policy */
1169 asmlinkage long sys_get_mempolicy(int __user *policy,
1170 unsigned long __user *nmask,
1171 unsigned long maxnode,
1172 unsigned long addr, unsigned long flags)
1175 int uninitialized_var(pval);
1178 if (nmask != NULL && maxnode < MAX_NUMNODES)
1181 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1186 if (policy && put_user(pval, policy))
1190 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1195 #ifdef CONFIG_COMPAT
1197 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1198 compat_ulong_t __user *nmask,
1199 compat_ulong_t maxnode,
1200 compat_ulong_t addr, compat_ulong_t flags)
1203 unsigned long __user *nm = NULL;
1204 unsigned long nr_bits, alloc_size;
1205 DECLARE_BITMAP(bm, MAX_NUMNODES);
1207 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1208 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1211 nm = compat_alloc_user_space(alloc_size);
1213 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1215 if (!err && nmask) {
1216 err = copy_from_user(bm, nm, alloc_size);
1217 /* ensure entire bitmap is zeroed */
1218 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1219 err |= compat_put_bitmap(nmask, bm, nr_bits);
1225 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1226 compat_ulong_t maxnode)
1229 unsigned long __user *nm = NULL;
1230 unsigned long nr_bits, alloc_size;
1231 DECLARE_BITMAP(bm, MAX_NUMNODES);
1233 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1234 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1237 err = compat_get_bitmap(bm, nmask, nr_bits);
1238 nm = compat_alloc_user_space(alloc_size);
1239 err |= copy_to_user(nm, bm, alloc_size);
1245 return sys_set_mempolicy(mode, nm, nr_bits+1);
1248 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1249 compat_ulong_t mode, compat_ulong_t __user *nmask,
1250 compat_ulong_t maxnode, compat_ulong_t flags)
1253 unsigned long __user *nm = NULL;
1254 unsigned long nr_bits, alloc_size;
1257 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1258 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1261 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1262 nm = compat_alloc_user_space(alloc_size);
1263 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1269 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1275 * get_vma_policy(@task, @vma, @addr)
1276 * @task - task for fallback if vma policy == default
1277 * @vma - virtual memory area whose policy is sought
1278 * @addr - address in @vma for shared policy lookup
1280 * Returns effective policy for a VMA at specified address.
1281 * Falls back to @task or system default policy, as necessary.
1282 * Current or other task's task mempolicy and non-shared vma policies
1283 * are protected by the task's mmap_sem, which must be held for read by
1285 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1286 * count--added by the get_policy() vm_op, as appropriate--to protect against
1287 * freeing by another task. It is the caller's responsibility to free the
1288 * extra reference for shared policies.
1290 static struct mempolicy *get_vma_policy(struct task_struct *task,
1291 struct vm_area_struct *vma, unsigned long addr)
1293 struct mempolicy *pol = task->mempolicy;
1296 if (vma->vm_ops && vma->vm_ops->get_policy) {
1297 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1301 } else if (vma->vm_policy)
1302 pol = vma->vm_policy;
1305 pol = &default_policy;
1310 * Return a nodemask representing a mempolicy for filtering nodes for
1313 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1315 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1316 if (unlikely(policy->mode == MPOL_BIND) &&
1317 gfp_zone(gfp) >= policy_zone &&
1318 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1319 return &policy->v.nodes;
1324 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1325 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1329 switch (policy->mode) {
1330 case MPOL_PREFERRED:
1331 nd = policy->v.preferred_node;
1333 nd = numa_node_id();
1337 * Normally, MPOL_BIND allocations are node-local within the
1338 * allowed nodemask. However, if __GFP_THISNODE is set and the
1339 * current node is part of the mask, we use the zonelist for
1340 * the first node in the mask instead.
1342 nd = numa_node_id();
1343 if (unlikely(gfp & __GFP_THISNODE) &&
1344 unlikely(!node_isset(nd, policy->v.nodes)))
1345 nd = first_node(policy->v.nodes);
1347 case MPOL_INTERLEAVE: /* should not happen */
1348 nd = numa_node_id();
1354 return node_zonelist(nd, gfp);
1357 /* Do dynamic interleaving for a process */
1358 static unsigned interleave_nodes(struct mempolicy *policy)
1361 struct task_struct *me = current;
1364 next = next_node(nid, policy->v.nodes);
1365 if (next >= MAX_NUMNODES)
1366 next = first_node(policy->v.nodes);
1367 if (next < MAX_NUMNODES)
1373 * Depending on the memory policy provide a node from which to allocate the
1375 * @policy must be protected by freeing by the caller. If @policy is
1376 * the current task's mempolicy, this protection is implicit, as only the
1377 * task can change it's policy. The system default policy requires no
1380 unsigned slab_node(struct mempolicy *policy)
1383 return numa_node_id();
1385 switch (policy->mode) {
1386 case MPOL_PREFERRED:
1387 if (unlikely(policy->v.preferred_node >= 0))
1388 return policy->v.preferred_node;
1389 return numa_node_id();
1391 case MPOL_INTERLEAVE:
1392 return interleave_nodes(policy);
1396 * Follow bind policy behavior and start allocation at the
1399 struct zonelist *zonelist;
1401 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1402 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1403 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1414 /* Do static interleaving for a VMA with known offset. */
1415 static unsigned offset_il_node(struct mempolicy *pol,
1416 struct vm_area_struct *vma, unsigned long off)
1418 unsigned nnodes = nodes_weight(pol->v.nodes);
1424 return numa_node_id();
1425 target = (unsigned int)off % nnodes;
1428 nid = next_node(nid, pol->v.nodes);
1430 } while (c <= target);
1434 /* Determine a node number for interleave */
1435 static inline unsigned interleave_nid(struct mempolicy *pol,
1436 struct vm_area_struct *vma, unsigned long addr, int shift)
1442 * for small pages, there is no difference between
1443 * shift and PAGE_SHIFT, so the bit-shift is safe.
1444 * for huge pages, since vm_pgoff is in units of small
1445 * pages, we need to shift off the always 0 bits to get
1448 BUG_ON(shift < PAGE_SHIFT);
1449 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1450 off += (addr - vma->vm_start) >> shift;
1451 return offset_il_node(pol, vma, off);
1453 return interleave_nodes(pol);
1456 #ifdef CONFIG_HUGETLBFS
1458 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1459 * @vma = virtual memory area whose policy is sought
1460 * @addr = address in @vma for shared policy lookup and interleave policy
1461 * @gfp_flags = for requested zone
1462 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1463 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1465 * Returns a zonelist suitable for a huge page allocation and a pointer
1466 * to the struct mempolicy for conditional unref after allocation.
1467 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1468 * @nodemask for filtering the zonelist.
1470 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1471 gfp_t gfp_flags, struct mempolicy **mpol,
1472 nodemask_t **nodemask)
1474 struct zonelist *zl;
1476 *mpol = get_vma_policy(current, vma, addr);
1477 *nodemask = NULL; /* assume !MPOL_BIND */
1479 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1480 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1481 HPAGE_SHIFT), gfp_flags);
1483 zl = policy_zonelist(gfp_flags, *mpol);
1484 if ((*mpol)->mode == MPOL_BIND)
1485 *nodemask = &(*mpol)->v.nodes;
1491 /* Allocate a page in interleaved policy.
1492 Own path because it needs to do special accounting. */
1493 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1496 struct zonelist *zl;
1499 zl = node_zonelist(nid, gfp);
1500 page = __alloc_pages(gfp, order, zl);
1501 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1502 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1507 * alloc_page_vma - Allocate a page for a VMA.
1510 * %GFP_USER user allocation.
1511 * %GFP_KERNEL kernel allocations,
1512 * %GFP_HIGHMEM highmem/user allocations,
1513 * %GFP_FS allocation should not call back into a file system.
1514 * %GFP_ATOMIC don't sleep.
1516 * @vma: Pointer to VMA or NULL if not available.
1517 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1519 * This function allocates a page from the kernel page pool and applies
1520 * a NUMA policy associated with the VMA or the current process.
1521 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1522 * mm_struct of the VMA to prevent it from going away. Should be used for
1523 * all allocations for pages that will be mapped into
1524 * user space. Returns NULL when no page can be allocated.
1526 * Should be called with the mm_sem of the vma hold.
1529 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1531 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1532 struct zonelist *zl;
1534 cpuset_update_task_memory_state();
1536 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1539 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1541 return alloc_page_interleave(gfp, 0, nid);
1543 zl = policy_zonelist(gfp, pol);
1544 if (unlikely(mpol_needs_cond_ref(pol))) {
1546 * slow path: ref counted shared policy
1548 struct page *page = __alloc_pages_nodemask(gfp, 0,
1549 zl, policy_nodemask(gfp, pol));
1554 * fast path: default or task policy
1556 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1560 * alloc_pages_current - Allocate pages.
1563 * %GFP_USER user allocation,
1564 * %GFP_KERNEL kernel allocation,
1565 * %GFP_HIGHMEM highmem allocation,
1566 * %GFP_FS don't call back into a file system.
1567 * %GFP_ATOMIC don't sleep.
1568 * @order: Power of two of allocation size in pages. 0 is a single page.
1570 * Allocate a page from the kernel page pool. When not in
1571 * interrupt context and apply the current process NUMA policy.
1572 * Returns NULL when no page can be allocated.
1574 * Don't call cpuset_update_task_memory_state() unless
1575 * 1) it's ok to take cpuset_sem (can WAIT), and
1576 * 2) allocating for current task (not interrupt).
1578 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1580 struct mempolicy *pol = current->mempolicy;
1582 if ((gfp & __GFP_WAIT) && !in_interrupt())
1583 cpuset_update_task_memory_state();
1584 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1585 pol = &default_policy;
1588 * No reference counting needed for current->mempolicy
1589 * nor system default_policy
1591 if (pol->mode == MPOL_INTERLEAVE)
1592 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1593 return __alloc_pages_nodemask(gfp, order,
1594 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1596 EXPORT_SYMBOL(alloc_pages_current);
1599 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1600 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1601 * with the mems_allowed returned by cpuset_mems_allowed(). This
1602 * keeps mempolicies cpuset relative after its cpuset moves. See
1603 * further kernel/cpuset.c update_nodemask().
1606 /* Slow path of a mempolicy duplicate */
1607 struct mempolicy *__mpol_dup(struct mempolicy *old)
1609 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1612 return ERR_PTR(-ENOMEM);
1613 if (current_cpuset_is_being_rebound()) {
1614 nodemask_t mems = cpuset_mems_allowed(current);
1615 mpol_rebind_policy(old, &mems);
1618 atomic_set(&new->refcnt, 1);
1623 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1624 * eliminate the * MPOL_F_* flags that require conditional ref and
1625 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1626 * after return. Use the returned value.
1628 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1629 * policy lookup, even if the policy needs/has extra ref on lookup.
1630 * shmem_readahead needs this.
1632 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1633 struct mempolicy *frompol)
1635 if (!mpol_needs_cond_ref(frompol))
1639 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1640 __mpol_put(frompol);
1644 static int mpol_match_intent(const struct mempolicy *a,
1645 const struct mempolicy *b)
1647 if (a->flags != b->flags)
1649 if (!mpol_store_user_nodemask(a))
1651 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1654 /* Slow path of a mempolicy comparison */
1655 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1659 if (a->mode != b->mode)
1661 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
1666 case MPOL_INTERLEAVE:
1667 return nodes_equal(a->v.nodes, b->v.nodes);
1668 case MPOL_PREFERRED:
1669 return a->v.preferred_node == b->v.preferred_node;
1677 * Shared memory backing store policy support.
1679 * Remember policies even when nobody has shared memory mapped.
1680 * The policies are kept in Red-Black tree linked from the inode.
1681 * They are protected by the sp->lock spinlock, which should be held
1682 * for any accesses to the tree.
1685 /* lookup first element intersecting start-end */
1686 /* Caller holds sp->lock */
1687 static struct sp_node *
1688 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1690 struct rb_node *n = sp->root.rb_node;
1693 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1695 if (start >= p->end)
1697 else if (end <= p->start)
1705 struct sp_node *w = NULL;
1706 struct rb_node *prev = rb_prev(n);
1709 w = rb_entry(prev, struct sp_node, nd);
1710 if (w->end <= start)
1714 return rb_entry(n, struct sp_node, nd);
1717 /* Insert a new shared policy into the list. */
1718 /* Caller holds sp->lock */
1719 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1721 struct rb_node **p = &sp->root.rb_node;
1722 struct rb_node *parent = NULL;
1727 nd = rb_entry(parent, struct sp_node, nd);
1728 if (new->start < nd->start)
1730 else if (new->end > nd->end)
1731 p = &(*p)->rb_right;
1735 rb_link_node(&new->nd, parent, p);
1736 rb_insert_color(&new->nd, &sp->root);
1737 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1738 new->policy ? new->policy->mode : 0);
1741 /* Find shared policy intersecting idx */
1743 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1745 struct mempolicy *pol = NULL;
1748 if (!sp->root.rb_node)
1750 spin_lock(&sp->lock);
1751 sn = sp_lookup(sp, idx, idx+1);
1753 mpol_get(sn->policy);
1756 spin_unlock(&sp->lock);
1760 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1762 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1763 rb_erase(&n->nd, &sp->root);
1764 mpol_put(n->policy);
1765 kmem_cache_free(sn_cache, n);
1768 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1769 struct mempolicy *pol)
1771 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1778 pol->flags |= MPOL_F_SHARED; /* for unref */
1783 /* Replace a policy range. */
1784 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1785 unsigned long end, struct sp_node *new)
1787 struct sp_node *n, *new2 = NULL;
1790 spin_lock(&sp->lock);
1791 n = sp_lookup(sp, start, end);
1792 /* Take care of old policies in the same range. */
1793 while (n && n->start < end) {
1794 struct rb_node *next = rb_next(&n->nd);
1795 if (n->start >= start) {
1801 /* Old policy spanning whole new range. */
1804 spin_unlock(&sp->lock);
1805 new2 = sp_alloc(end, n->end, n->policy);
1811 sp_insert(sp, new2);
1819 n = rb_entry(next, struct sp_node, nd);
1823 spin_unlock(&sp->lock);
1825 mpol_put(new2->policy);
1826 kmem_cache_free(sn_cache, new2);
1831 void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
1832 unsigned short flags, nodemask_t *policy_nodes)
1834 info->root = RB_ROOT;
1835 spin_lock_init(&info->lock);
1837 if (policy != MPOL_DEFAULT) {
1838 struct mempolicy *newpol;
1840 /* Falls back to NULL policy [MPOL_DEFAULT] on any error */
1841 newpol = mpol_new(policy, flags, policy_nodes);
1842 if (!IS_ERR(newpol)) {
1843 /* Create pseudo-vma that contains just the policy */
1844 struct vm_area_struct pvma;
1846 memset(&pvma, 0, sizeof(struct vm_area_struct));
1847 /* Policy covers entire file */
1848 pvma.vm_end = TASK_SIZE;
1849 mpol_set_shared_policy(info, &pvma, newpol);
1855 int mpol_set_shared_policy(struct shared_policy *info,
1856 struct vm_area_struct *vma, struct mempolicy *npol)
1859 struct sp_node *new = NULL;
1860 unsigned long sz = vma_pages(vma);
1862 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1864 sz, npol ? npol->mode : -1,
1865 npol ? npol->flags : -1,
1866 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1869 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1873 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1875 kmem_cache_free(sn_cache, new);
1879 /* Free a backing policy store on inode delete. */
1880 void mpol_free_shared_policy(struct shared_policy *p)
1883 struct rb_node *next;
1885 if (!p->root.rb_node)
1887 spin_lock(&p->lock);
1888 next = rb_first(&p->root);
1890 n = rb_entry(next, struct sp_node, nd);
1891 next = rb_next(&n->nd);
1892 rb_erase(&n->nd, &p->root);
1893 mpol_put(n->policy);
1894 kmem_cache_free(sn_cache, n);
1896 spin_unlock(&p->lock);
1899 /* assumes fs == KERNEL_DS */
1900 void __init numa_policy_init(void)
1902 nodemask_t interleave_nodes;
1903 unsigned long largest = 0;
1904 int nid, prefer = 0;
1906 policy_cache = kmem_cache_create("numa_policy",
1907 sizeof(struct mempolicy),
1908 0, SLAB_PANIC, NULL);
1910 sn_cache = kmem_cache_create("shared_policy_node",
1911 sizeof(struct sp_node),
1912 0, SLAB_PANIC, NULL);
1915 * Set interleaving policy for system init. Interleaving is only
1916 * enabled across suitably sized nodes (default is >= 16MB), or
1917 * fall back to the largest node if they're all smaller.
1919 nodes_clear(interleave_nodes);
1920 for_each_node_state(nid, N_HIGH_MEMORY) {
1921 unsigned long total_pages = node_present_pages(nid);
1923 /* Preserve the largest node */
1924 if (largest < total_pages) {
1925 largest = total_pages;
1929 /* Interleave this node? */
1930 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1931 node_set(nid, interleave_nodes);
1934 /* All too small, use the largest */
1935 if (unlikely(nodes_empty(interleave_nodes)))
1936 node_set(prefer, interleave_nodes);
1938 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
1939 printk("numa_policy_init: interleaving failed\n");
1942 /* Reset policy of current process to default */
1943 void numa_default_policy(void)
1945 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1949 * "local" is pseudo-policy: MPOL_PREFERRED with preferred_node == -1
1950 * Used only for mpol_to_str()
1952 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
1953 static const char * const policy_types[] =
1954 { "default", "prefer", "bind", "interleave", "local" };
1957 * Convert a mempolicy into a string.
1958 * Returns the number of characters in buffer (if positive)
1959 * or an error (negative)
1961 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1967 unsigned short mode;
1968 unsigned short flags = pol ? pol->flags : 0;
1970 if (!pol || pol == &default_policy)
1971 mode = MPOL_DEFAULT;
1980 case MPOL_PREFERRED:
1982 nid = pol->v.preferred_node;
1984 mode = MPOL_LOCAL; /* pseudo-policy */
1986 node_set(nid, nodes);
1991 case MPOL_INTERLEAVE:
1992 nodes = pol->v.nodes;
2000 l = strlen(policy_types[mode]);
2001 if (buffer + maxlen < p + l + 1)
2004 strcpy(p, policy_types[mode]);
2010 if (buffer + maxlen < p + 2)
2014 if (flags & MPOL_F_STATIC_NODES)
2015 p += sprintf(p, "%sstatic", need_bar++ ? "|" : "");
2016 if (flags & MPOL_F_RELATIVE_NODES)
2017 p += sprintf(p, "%srelative", need_bar++ ? "|" : "");
2020 if (!nodes_empty(nodes)) {
2021 if (buffer + maxlen < p + 2)
2024 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2030 unsigned long pages;
2032 unsigned long active;
2033 unsigned long writeback;
2034 unsigned long mapcount_max;
2035 unsigned long dirty;
2036 unsigned long swapcache;
2037 unsigned long node[MAX_NUMNODES];
2040 static void gather_stats(struct page *page, void *private, int pte_dirty)
2042 struct numa_maps *md = private;
2043 int count = page_mapcount(page);
2046 if (pte_dirty || PageDirty(page))
2049 if (PageSwapCache(page))
2052 if (PageActive(page))
2055 if (PageWriteback(page))
2061 if (count > md->mapcount_max)
2062 md->mapcount_max = count;
2064 md->node[page_to_nid(page)]++;
2067 #ifdef CONFIG_HUGETLB_PAGE
2068 static void check_huge_range(struct vm_area_struct *vma,
2069 unsigned long start, unsigned long end,
2070 struct numa_maps *md)
2075 for (addr = start; addr < end; addr += HPAGE_SIZE) {
2076 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
2086 page = pte_page(pte);
2090 gather_stats(page, md, pte_dirty(*ptep));
2094 static inline void check_huge_range(struct vm_area_struct *vma,
2095 unsigned long start, unsigned long end,
2096 struct numa_maps *md)
2102 * Display pages allocated per node and memory policy via /proc.
2104 int show_numa_map(struct seq_file *m, void *v)
2106 struct proc_maps_private *priv = m->private;
2107 struct vm_area_struct *vma = v;
2108 struct numa_maps *md;
2109 struct file *file = vma->vm_file;
2110 struct mm_struct *mm = vma->vm_mm;
2111 struct mempolicy *pol;
2118 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2122 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2123 mpol_to_str(buffer, sizeof(buffer), pol);
2126 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2129 seq_printf(m, " file=");
2130 seq_path(m, &file->f_path, "\n\t= ");
2131 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2132 seq_printf(m, " heap");
2133 } else if (vma->vm_start <= mm->start_stack &&
2134 vma->vm_end >= mm->start_stack) {
2135 seq_printf(m, " stack");
2138 if (is_vm_hugetlb_page(vma)) {
2139 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2140 seq_printf(m, " huge");
2142 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2143 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2150 seq_printf(m," anon=%lu",md->anon);
2153 seq_printf(m," dirty=%lu",md->dirty);
2155 if (md->pages != md->anon && md->pages != md->dirty)
2156 seq_printf(m, " mapped=%lu", md->pages);
2158 if (md->mapcount_max > 1)
2159 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2162 seq_printf(m," swapcache=%lu", md->swapcache);
2164 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2165 seq_printf(m," active=%lu", md->active);
2168 seq_printf(m," writeback=%lu", md->writeback);
2170 for_each_node_state(n, N_HIGH_MEMORY)
2172 seq_printf(m, " N%d=%lu", n, md->node[n]);
2177 if (m->count < m->size)
2178 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;