2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/module.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
94 #include <asm/tlbflush.h>
95 #include <asm/uaccess.h>
100 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
101 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
102 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
104 static struct kmem_cache *policy_cache;
105 static struct kmem_cache *sn_cache;
107 /* Highest zone. An specific allocation for a zone below that is not
109 enum zone_type policy_zone = 0;
112 * run-time system-wide default policy => local allocation
114 struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
120 static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
122 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
123 } mpol_ops[MPOL_MAX];
125 /* Check that the nodemask contains at least one populated zone */
126 static int is_valid_nodemask(const nodemask_t *nodemask)
130 for_each_node_mask(nd, *nodemask) {
133 for (k = 0; k <= policy_zone; k++) {
134 z = &NODE_DATA(nd)->node_zones[k];
135 if (z->present_pages > 0)
143 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
145 return pol->flags & MPOL_MODE_FLAGS;
148 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
149 const nodemask_t *rel)
152 nodes_fold(tmp, *orig, nodes_weight(*rel));
153 nodes_onto(*ret, tmp, *rel);
156 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
158 if (nodes_empty(*nodes))
160 pol->v.nodes = *nodes;
164 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
167 pol->flags |= MPOL_F_LOCAL; /* local allocation */
168 else if (nodes_empty(*nodes))
169 return -EINVAL; /* no allowed nodes */
171 pol->v.preferred_node = first_node(*nodes);
175 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
177 if (!is_valid_nodemask(nodes))
179 pol->v.nodes = *nodes;
184 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
185 * any, for the new policy. mpol_new() has already validated the nodes
186 * parameter with respect to the policy mode and flags. But, we need to
187 * handle an empty nodemask with MPOL_PREFERRED here.
189 * Must be called holding task's alloc_lock to protect task's mems_allowed
190 * and mempolicy. May also be called holding the mmap_semaphore for write.
192 static int mpol_set_nodemask(struct mempolicy *pol,
193 const nodemask_t *nodes, struct nodemask_scratch *nsc)
197 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
200 /* Check N_HIGH_MEMORY */
201 nodes_and(nsc->mask1,
202 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
205 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
206 nodes = NULL; /* explicit local allocation */
208 if (pol->flags & MPOL_F_RELATIVE_NODES)
209 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
211 nodes_and(nsc->mask2, *nodes, nsc->mask1);
213 if (mpol_store_user_nodemask(pol))
214 pol->w.user_nodemask = *nodes;
216 pol->w.cpuset_mems_allowed =
217 cpuset_current_mems_allowed;
221 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
223 ret = mpol_ops[pol->mode].create(pol, NULL);
228 * This function just creates a new policy, does some check and simple
229 * initialization. You must invoke mpol_set_nodemask() to set nodes.
231 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
234 struct mempolicy *policy;
236 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
237 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
239 if (mode == MPOL_DEFAULT) {
240 if (nodes && !nodes_empty(*nodes))
241 return ERR_PTR(-EINVAL);
242 return NULL; /* simply delete any existing policy */
247 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
248 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
249 * All other modes require a valid pointer to a non-empty nodemask.
251 if (mode == MPOL_PREFERRED) {
252 if (nodes_empty(*nodes)) {
253 if (((flags & MPOL_F_STATIC_NODES) ||
254 (flags & MPOL_F_RELATIVE_NODES)))
255 return ERR_PTR(-EINVAL);
257 } else if (nodes_empty(*nodes))
258 return ERR_PTR(-EINVAL);
259 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
261 return ERR_PTR(-ENOMEM);
262 atomic_set(&policy->refcnt, 1);
264 policy->flags = flags;
269 /* Slow path of a mpol destructor. */
270 void __mpol_put(struct mempolicy *p)
272 if (!atomic_dec_and_test(&p->refcnt))
274 kmem_cache_free(policy_cache, p);
277 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
281 static void mpol_rebind_nodemask(struct mempolicy *pol,
282 const nodemask_t *nodes)
286 if (pol->flags & MPOL_F_STATIC_NODES)
287 nodes_and(tmp, pol->w.user_nodemask, *nodes);
288 else if (pol->flags & MPOL_F_RELATIVE_NODES)
289 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
291 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
293 pol->w.cpuset_mems_allowed = *nodes;
297 if (!node_isset(current->il_next, tmp)) {
298 current->il_next = next_node(current->il_next, tmp);
299 if (current->il_next >= MAX_NUMNODES)
300 current->il_next = first_node(tmp);
301 if (current->il_next >= MAX_NUMNODES)
302 current->il_next = numa_node_id();
306 static void mpol_rebind_preferred(struct mempolicy *pol,
307 const nodemask_t *nodes)
311 if (pol->flags & MPOL_F_STATIC_NODES) {
312 int node = first_node(pol->w.user_nodemask);
314 if (node_isset(node, *nodes)) {
315 pol->v.preferred_node = node;
316 pol->flags &= ~MPOL_F_LOCAL;
318 pol->flags |= MPOL_F_LOCAL;
319 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
320 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
321 pol->v.preferred_node = first_node(tmp);
322 } else if (!(pol->flags & MPOL_F_LOCAL)) {
323 pol->v.preferred_node = node_remap(pol->v.preferred_node,
324 pol->w.cpuset_mems_allowed,
326 pol->w.cpuset_mems_allowed = *nodes;
330 /* Migrate a policy to a different set of nodes */
331 static void mpol_rebind_policy(struct mempolicy *pol,
332 const nodemask_t *newmask)
336 if (!mpol_store_user_nodemask(pol) &&
337 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
339 mpol_ops[pol->mode].rebind(pol, newmask);
343 * Wrapper for mpol_rebind_policy() that just requires task
344 * pointer, and updates task mempolicy.
346 * Called with task's alloc_lock held.
349 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
351 mpol_rebind_policy(tsk->mempolicy, new);
355 * Rebind each vma in mm to new nodemask.
357 * Call holding a reference to mm. Takes mm->mmap_sem during call.
360 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
362 struct vm_area_struct *vma;
364 down_write(&mm->mmap_sem);
365 for (vma = mm->mmap; vma; vma = vma->vm_next)
366 mpol_rebind_policy(vma->vm_policy, new);
367 up_write(&mm->mmap_sem);
370 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
372 .rebind = mpol_rebind_default,
374 [MPOL_INTERLEAVE] = {
375 .create = mpol_new_interleave,
376 .rebind = mpol_rebind_nodemask,
379 .create = mpol_new_preferred,
380 .rebind = mpol_rebind_preferred,
383 .create = mpol_new_bind,
384 .rebind = mpol_rebind_nodemask,
388 static void gather_stats(struct page *, void *, int pte_dirty);
389 static void migrate_page_add(struct page *page, struct list_head *pagelist,
390 unsigned long flags);
392 /* Scan through pages checking if pages follow certain conditions. */
393 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
394 unsigned long addr, unsigned long end,
395 const nodemask_t *nodes, unsigned long flags,
402 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
407 if (!pte_present(*pte))
409 page = vm_normal_page(vma, addr, *pte);
413 * vm_normal_page() filters out zero pages, but there might
414 * still be PageReserved pages to skip, perhaps in a VDSO.
415 * And we cannot move PageKsm pages sensibly or safely yet.
417 if (PageReserved(page) || PageKsm(page))
419 nid = page_to_nid(page);
420 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
423 if (flags & MPOL_MF_STATS)
424 gather_stats(page, private, pte_dirty(*pte));
425 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
426 migrate_page_add(page, private, flags);
429 } while (pte++, addr += PAGE_SIZE, addr != end);
430 pte_unmap_unlock(orig_pte, ptl);
434 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
435 unsigned long addr, unsigned long end,
436 const nodemask_t *nodes, unsigned long flags,
442 pmd = pmd_offset(pud, addr);
444 next = pmd_addr_end(addr, end);
445 if (pmd_none_or_clear_bad(pmd))
447 if (check_pte_range(vma, pmd, addr, next, nodes,
450 } while (pmd++, addr = next, addr != end);
454 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
455 unsigned long addr, unsigned long end,
456 const nodemask_t *nodes, unsigned long flags,
462 pud = pud_offset(pgd, addr);
464 next = pud_addr_end(addr, end);
465 if (pud_none_or_clear_bad(pud))
467 if (check_pmd_range(vma, pud, addr, next, nodes,
470 } while (pud++, addr = next, addr != end);
474 static inline int check_pgd_range(struct vm_area_struct *vma,
475 unsigned long addr, unsigned long end,
476 const nodemask_t *nodes, unsigned long flags,
482 pgd = pgd_offset(vma->vm_mm, addr);
484 next = pgd_addr_end(addr, end);
485 if (pgd_none_or_clear_bad(pgd))
487 if (check_pud_range(vma, pgd, addr, next, nodes,
490 } while (pgd++, addr = next, addr != end);
495 * Check if all pages in a range are on a set of nodes.
496 * If pagelist != NULL then isolate pages from the LRU and
497 * put them on the pagelist.
499 static struct vm_area_struct *
500 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
501 const nodemask_t *nodes, unsigned long flags, void *private)
504 struct vm_area_struct *first, *vma, *prev;
507 first = find_vma(mm, start);
509 return ERR_PTR(-EFAULT);
511 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
512 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
513 if (!vma->vm_next && vma->vm_end < end)
514 return ERR_PTR(-EFAULT);
515 if (prev && prev->vm_end < vma->vm_start)
516 return ERR_PTR(-EFAULT);
518 if (!is_vm_hugetlb_page(vma) &&
519 ((flags & MPOL_MF_STRICT) ||
520 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
521 vma_migratable(vma)))) {
522 unsigned long endvma = vma->vm_end;
526 if (vma->vm_start > start)
527 start = vma->vm_start;
528 err = check_pgd_range(vma, start, endvma, nodes,
531 first = ERR_PTR(err);
540 /* Apply policy to a single VMA */
541 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
544 struct mempolicy *old = vma->vm_policy;
546 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
547 vma->vm_start, vma->vm_end, vma->vm_pgoff,
548 vma->vm_ops, vma->vm_file,
549 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
551 if (vma->vm_ops && vma->vm_ops->set_policy)
552 err = vma->vm_ops->set_policy(vma, new);
555 vma->vm_policy = new;
561 /* Step 2: apply policy to a range and do splits. */
562 static int mbind_range(struct mm_struct *mm, unsigned long start,
563 unsigned long end, struct mempolicy *new_pol)
565 struct vm_area_struct *next;
566 struct vm_area_struct *prev;
567 struct vm_area_struct *vma;
570 unsigned long vmstart;
573 vma = find_vma_prev(mm, start, &prev);
574 if (!vma || vma->vm_start > start)
577 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
579 vmstart = max(start, vma->vm_start);
580 vmend = min(end, vma->vm_end);
582 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
583 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
584 vma->anon_vma, vma->vm_file, pgoff, new_pol);
590 if (vma->vm_start != vmstart) {
591 err = split_vma(vma->vm_mm, vma, vmstart, 1);
595 if (vma->vm_end != vmend) {
596 err = split_vma(vma->vm_mm, vma, vmend, 0);
600 err = policy_vma(vma, new_pol);
610 * Update task->flags PF_MEMPOLICY bit: set iff non-default
611 * mempolicy. Allows more rapid checking of this (combined perhaps
612 * with other PF_* flag bits) on memory allocation hot code paths.
614 * If called from outside this file, the task 'p' should -only- be
615 * a newly forked child not yet visible on the task list, because
616 * manipulating the task flags of a visible task is not safe.
618 * The above limitation is why this routine has the funny name
619 * mpol_fix_fork_child_flag().
621 * It is also safe to call this with a task pointer of current,
622 * which the static wrapper mpol_set_task_struct_flag() does,
623 * for use within this file.
626 void mpol_fix_fork_child_flag(struct task_struct *p)
629 p->flags |= PF_MEMPOLICY;
631 p->flags &= ~PF_MEMPOLICY;
634 static void mpol_set_task_struct_flag(void)
636 mpol_fix_fork_child_flag(current);
639 /* Set the process memory policy */
640 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
643 struct mempolicy *new, *old;
644 struct mm_struct *mm = current->mm;
645 NODEMASK_SCRATCH(scratch);
651 new = mpol_new(mode, flags, nodes);
657 * prevent changing our mempolicy while show_numa_maps()
659 * Note: do_set_mempolicy() can be called at init time
663 down_write(&mm->mmap_sem);
665 ret = mpol_set_nodemask(new, nodes, scratch);
667 task_unlock(current);
669 up_write(&mm->mmap_sem);
673 old = current->mempolicy;
674 current->mempolicy = new;
675 mpol_set_task_struct_flag();
676 if (new && new->mode == MPOL_INTERLEAVE &&
677 nodes_weight(new->v.nodes))
678 current->il_next = first_node(new->v.nodes);
679 task_unlock(current);
681 up_write(&mm->mmap_sem);
686 NODEMASK_SCRATCH_FREE(scratch);
691 * Return nodemask for policy for get_mempolicy() query
693 * Called with task's alloc_lock held
695 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
698 if (p == &default_policy)
704 case MPOL_INTERLEAVE:
708 if (!(p->flags & MPOL_F_LOCAL))
709 node_set(p->v.preferred_node, *nodes);
710 /* else return empty node mask for local allocation */
717 static int lookup_node(struct mm_struct *mm, unsigned long addr)
722 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
724 err = page_to_nid(p);
730 /* Retrieve NUMA policy */
731 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
732 unsigned long addr, unsigned long flags)
735 struct mm_struct *mm = current->mm;
736 struct vm_area_struct *vma = NULL;
737 struct mempolicy *pol = current->mempolicy;
740 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
743 if (flags & MPOL_F_MEMS_ALLOWED) {
744 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
746 *policy = 0; /* just so it's initialized */
748 *nmask = cpuset_current_mems_allowed;
749 task_unlock(current);
753 if (flags & MPOL_F_ADDR) {
755 * Do NOT fall back to task policy if the
756 * vma/shared policy at addr is NULL. We
757 * want to return MPOL_DEFAULT in this case.
759 down_read(&mm->mmap_sem);
760 vma = find_vma_intersection(mm, addr, addr+1);
762 up_read(&mm->mmap_sem);
765 if (vma->vm_ops && vma->vm_ops->get_policy)
766 pol = vma->vm_ops->get_policy(vma, addr);
768 pol = vma->vm_policy;
773 pol = &default_policy; /* indicates default behavior */
775 if (flags & MPOL_F_NODE) {
776 if (flags & MPOL_F_ADDR) {
777 err = lookup_node(mm, addr);
781 } else if (pol == current->mempolicy &&
782 pol->mode == MPOL_INTERLEAVE) {
783 *policy = current->il_next;
789 *policy = pol == &default_policy ? MPOL_DEFAULT :
792 * Internal mempolicy flags must be masked off before exposing
793 * the policy to userspace.
795 *policy |= (pol->flags & MPOL_MODE_FLAGS);
799 up_read(¤t->mm->mmap_sem);
805 if (mpol_store_user_nodemask(pol)) {
806 *nmask = pol->w.user_nodemask;
809 get_policy_nodemask(pol, nmask);
810 task_unlock(current);
817 up_read(¤t->mm->mmap_sem);
821 #ifdef CONFIG_MIGRATION
825 static void migrate_page_add(struct page *page, struct list_head *pagelist,
829 * Avoid migrating a page that is shared with others.
831 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
832 if (!isolate_lru_page(page)) {
833 list_add_tail(&page->lru, pagelist);
834 inc_zone_page_state(page, NR_ISOLATED_ANON +
835 page_is_file_cache(page));
840 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
842 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
846 * Migrate pages from one node to a target node.
847 * Returns error or the number of pages not migrated.
849 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
857 node_set(source, nmask);
859 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
860 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
862 if (!list_empty(&pagelist))
863 err = migrate_pages(&pagelist, new_node_page, dest, 0);
869 * Move pages between the two nodesets so as to preserve the physical
870 * layout as much as possible.
872 * Returns the number of page that could not be moved.
874 int do_migrate_pages(struct mm_struct *mm,
875 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
881 err = migrate_prep();
885 down_read(&mm->mmap_sem);
887 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
892 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
893 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
894 * bit in 'tmp', and return that <source, dest> pair for migration.
895 * The pair of nodemasks 'to' and 'from' define the map.
897 * If no pair of bits is found that way, fallback to picking some
898 * pair of 'source' and 'dest' bits that are not the same. If the
899 * 'source' and 'dest' bits are the same, this represents a node
900 * that will be migrating to itself, so no pages need move.
902 * If no bits are left in 'tmp', or if all remaining bits left
903 * in 'tmp' correspond to the same bit in 'to', return false
904 * (nothing left to migrate).
906 * This lets us pick a pair of nodes to migrate between, such that
907 * if possible the dest node is not already occupied by some other
908 * source node, minimizing the risk of overloading the memory on a
909 * node that would happen if we migrated incoming memory to a node
910 * before migrating outgoing memory source that same node.
912 * A single scan of tmp is sufficient. As we go, we remember the
913 * most recent <s, d> pair that moved (s != d). If we find a pair
914 * that not only moved, but what's better, moved to an empty slot
915 * (d is not set in tmp), then we break out then, with that pair.
916 * Otherwise when we finish scannng from_tmp, we at least have the
917 * most recent <s, d> pair that moved. If we get all the way through
918 * the scan of tmp without finding any node that moved, much less
919 * moved to an empty node, then there is nothing left worth migrating.
923 while (!nodes_empty(tmp)) {
928 for_each_node_mask(s, tmp) {
929 d = node_remap(s, *from_nodes, *to_nodes);
933 source = s; /* Node moved. Memorize */
936 /* dest not in remaining from nodes? */
937 if (!node_isset(dest, tmp))
943 node_clear(source, tmp);
944 err = migrate_to_node(mm, source, dest, flags);
951 up_read(&mm->mmap_sem);
959 * Allocate a new page for page migration based on vma policy.
960 * Start assuming that page is mapped by vma pointed to by @private.
961 * Search forward from there, if not. N.B., this assumes that the
962 * list of pages handed to migrate_pages()--which is how we get here--
963 * is in virtual address order.
965 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
967 struct vm_area_struct *vma = (struct vm_area_struct *)private;
968 unsigned long uninitialized_var(address);
971 address = page_address_in_vma(page, vma);
972 if (address != -EFAULT)
978 * if !vma, alloc_page_vma() will use task or system default policy
980 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
984 static void migrate_page_add(struct page *page, struct list_head *pagelist,
989 int do_migrate_pages(struct mm_struct *mm,
990 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
995 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1001 static long do_mbind(unsigned long start, unsigned long len,
1002 unsigned short mode, unsigned short mode_flags,
1003 nodemask_t *nmask, unsigned long flags)
1005 struct vm_area_struct *vma;
1006 struct mm_struct *mm = current->mm;
1007 struct mempolicy *new;
1010 LIST_HEAD(pagelist);
1012 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1013 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1015 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1018 if (start & ~PAGE_MASK)
1021 if (mode == MPOL_DEFAULT)
1022 flags &= ~MPOL_MF_STRICT;
1024 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1032 new = mpol_new(mode, mode_flags, nmask);
1034 return PTR_ERR(new);
1037 * If we are using the default policy then operation
1038 * on discontinuous address spaces is okay after all
1041 flags |= MPOL_MF_DISCONTIG_OK;
1043 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1044 start, start + len, mode, mode_flags,
1045 nmask ? nodes_addr(*nmask)[0] : -1);
1047 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1049 err = migrate_prep();
1054 NODEMASK_SCRATCH(scratch);
1056 down_write(&mm->mmap_sem);
1058 err = mpol_set_nodemask(new, nmask, scratch);
1059 task_unlock(current);
1061 up_write(&mm->mmap_sem);
1064 NODEMASK_SCRATCH_FREE(scratch);
1069 vma = check_range(mm, start, end, nmask,
1070 flags | MPOL_MF_INVERT, &pagelist);
1076 err = mbind_range(mm, start, end, new);
1078 if (!list_empty(&pagelist))
1079 nr_failed = migrate_pages(&pagelist, new_vma_page,
1080 (unsigned long)vma, 0);
1082 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1085 putback_lru_pages(&pagelist);
1087 up_write(&mm->mmap_sem);
1094 * User space interface with variable sized bitmaps for nodelists.
1097 /* Copy a node mask from user space. */
1098 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1099 unsigned long maxnode)
1102 unsigned long nlongs;
1103 unsigned long endmask;
1106 nodes_clear(*nodes);
1107 if (maxnode == 0 || !nmask)
1109 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1112 nlongs = BITS_TO_LONGS(maxnode);
1113 if ((maxnode % BITS_PER_LONG) == 0)
1116 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1118 /* When the user specified more nodes than supported just check
1119 if the non supported part is all zero. */
1120 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1121 if (nlongs > PAGE_SIZE/sizeof(long))
1123 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1125 if (get_user(t, nmask + k))
1127 if (k == nlongs - 1) {
1133 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1137 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1139 nodes_addr(*nodes)[nlongs-1] &= endmask;
1143 /* Copy a kernel node mask to user space */
1144 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1147 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1148 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1150 if (copy > nbytes) {
1151 if (copy > PAGE_SIZE)
1153 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1157 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1160 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1161 unsigned long, mode, unsigned long __user *, nmask,
1162 unsigned long, maxnode, unsigned, flags)
1166 unsigned short mode_flags;
1168 mode_flags = mode & MPOL_MODE_FLAGS;
1169 mode &= ~MPOL_MODE_FLAGS;
1170 if (mode >= MPOL_MAX)
1172 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1173 (mode_flags & MPOL_F_RELATIVE_NODES))
1175 err = get_nodes(&nodes, nmask, maxnode);
1178 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1181 /* Set the process memory policy */
1182 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1183 unsigned long, maxnode)
1187 unsigned short flags;
1189 flags = mode & MPOL_MODE_FLAGS;
1190 mode &= ~MPOL_MODE_FLAGS;
1191 if ((unsigned int)mode >= MPOL_MAX)
1193 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1195 err = get_nodes(&nodes, nmask, maxnode);
1198 return do_set_mempolicy(mode, flags, &nodes);
1201 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1202 const unsigned long __user *, old_nodes,
1203 const unsigned long __user *, new_nodes)
1205 const struct cred *cred = current_cred(), *tcred;
1206 struct mm_struct *mm;
1207 struct task_struct *task;
1210 nodemask_t task_nodes;
1213 err = get_nodes(&old, old_nodes, maxnode);
1217 err = get_nodes(&new, new_nodes, maxnode);
1221 /* Find the mm_struct */
1222 read_lock(&tasklist_lock);
1223 task = pid ? find_task_by_vpid(pid) : current;
1225 read_unlock(&tasklist_lock);
1228 mm = get_task_mm(task);
1229 read_unlock(&tasklist_lock);
1235 * Check if this process has the right to modify the specified
1236 * process. The right exists if the process has administrative
1237 * capabilities, superuser privileges or the same
1238 * userid as the target process.
1241 tcred = __task_cred(task);
1242 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1243 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1244 !capable(CAP_SYS_NICE)) {
1251 task_nodes = cpuset_mems_allowed(task);
1252 /* Is the user allowed to access the target nodes? */
1253 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1258 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1263 err = security_task_movememory(task);
1267 err = do_migrate_pages(mm, &old, &new,
1268 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1275 /* Retrieve NUMA policy */
1276 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1277 unsigned long __user *, nmask, unsigned long, maxnode,
1278 unsigned long, addr, unsigned long, flags)
1281 int uninitialized_var(pval);
1284 if (nmask != NULL && maxnode < MAX_NUMNODES)
1287 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1292 if (policy && put_user(pval, policy))
1296 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1301 #ifdef CONFIG_COMPAT
1303 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1304 compat_ulong_t __user *nmask,
1305 compat_ulong_t maxnode,
1306 compat_ulong_t addr, compat_ulong_t flags)
1309 unsigned long __user *nm = NULL;
1310 unsigned long nr_bits, alloc_size;
1311 DECLARE_BITMAP(bm, MAX_NUMNODES);
1313 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1314 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1317 nm = compat_alloc_user_space(alloc_size);
1319 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1321 if (!err && nmask) {
1322 err = copy_from_user(bm, nm, alloc_size);
1323 /* ensure entire bitmap is zeroed */
1324 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1325 err |= compat_put_bitmap(nmask, bm, nr_bits);
1331 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1332 compat_ulong_t maxnode)
1335 unsigned long __user *nm = NULL;
1336 unsigned long nr_bits, alloc_size;
1337 DECLARE_BITMAP(bm, MAX_NUMNODES);
1339 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1340 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1343 err = compat_get_bitmap(bm, nmask, nr_bits);
1344 nm = compat_alloc_user_space(alloc_size);
1345 err |= copy_to_user(nm, bm, alloc_size);
1351 return sys_set_mempolicy(mode, nm, nr_bits+1);
1354 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1355 compat_ulong_t mode, compat_ulong_t __user *nmask,
1356 compat_ulong_t maxnode, compat_ulong_t flags)
1359 unsigned long __user *nm = NULL;
1360 unsigned long nr_bits, alloc_size;
1363 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1364 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1367 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1368 nm = compat_alloc_user_space(alloc_size);
1369 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1375 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1381 * get_vma_policy(@task, @vma, @addr)
1382 * @task - task for fallback if vma policy == default
1383 * @vma - virtual memory area whose policy is sought
1384 * @addr - address in @vma for shared policy lookup
1386 * Returns effective policy for a VMA at specified address.
1387 * Falls back to @task or system default policy, as necessary.
1388 * Current or other task's task mempolicy and non-shared vma policies
1389 * are protected by the task's mmap_sem, which must be held for read by
1391 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1392 * count--added by the get_policy() vm_op, as appropriate--to protect against
1393 * freeing by another task. It is the caller's responsibility to free the
1394 * extra reference for shared policies.
1396 static struct mempolicy *get_vma_policy(struct task_struct *task,
1397 struct vm_area_struct *vma, unsigned long addr)
1399 struct mempolicy *pol = task->mempolicy;
1402 if (vma->vm_ops && vma->vm_ops->get_policy) {
1403 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1407 } else if (vma->vm_policy)
1408 pol = vma->vm_policy;
1411 pol = &default_policy;
1416 * Return a nodemask representing a mempolicy for filtering nodes for
1419 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1421 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1422 if (unlikely(policy->mode == MPOL_BIND) &&
1423 gfp_zone(gfp) >= policy_zone &&
1424 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1425 return &policy->v.nodes;
1430 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1431 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1433 int nd = numa_node_id();
1435 switch (policy->mode) {
1436 case MPOL_PREFERRED:
1437 if (!(policy->flags & MPOL_F_LOCAL))
1438 nd = policy->v.preferred_node;
1442 * Normally, MPOL_BIND allocations are node-local within the
1443 * allowed nodemask. However, if __GFP_THISNODE is set and the
1444 * current node isn't part of the mask, we use the zonelist for
1445 * the first node in the mask instead.
1447 if (unlikely(gfp & __GFP_THISNODE) &&
1448 unlikely(!node_isset(nd, policy->v.nodes)))
1449 nd = first_node(policy->v.nodes);
1454 return node_zonelist(nd, gfp);
1457 /* Do dynamic interleaving for a process */
1458 static unsigned interleave_nodes(struct mempolicy *policy)
1461 struct task_struct *me = current;
1464 next = next_node(nid, policy->v.nodes);
1465 if (next >= MAX_NUMNODES)
1466 next = first_node(policy->v.nodes);
1467 if (next < MAX_NUMNODES)
1473 * Depending on the memory policy provide a node from which to allocate the
1475 * @policy must be protected by freeing by the caller. If @policy is
1476 * the current task's mempolicy, this protection is implicit, as only the
1477 * task can change it's policy. The system default policy requires no
1480 unsigned slab_node(struct mempolicy *policy)
1482 if (!policy || policy->flags & MPOL_F_LOCAL)
1483 return numa_node_id();
1485 switch (policy->mode) {
1486 case MPOL_PREFERRED:
1488 * handled MPOL_F_LOCAL above
1490 return policy->v.preferred_node;
1492 case MPOL_INTERLEAVE:
1493 return interleave_nodes(policy);
1497 * Follow bind policy behavior and start allocation at the
1500 struct zonelist *zonelist;
1502 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1503 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1504 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1515 /* Do static interleaving for a VMA with known offset. */
1516 static unsigned offset_il_node(struct mempolicy *pol,
1517 struct vm_area_struct *vma, unsigned long off)
1519 unsigned nnodes = nodes_weight(pol->v.nodes);
1525 return numa_node_id();
1526 target = (unsigned int)off % nnodes;
1529 nid = next_node(nid, pol->v.nodes);
1531 } while (c <= target);
1535 /* Determine a node number for interleave */
1536 static inline unsigned interleave_nid(struct mempolicy *pol,
1537 struct vm_area_struct *vma, unsigned long addr, int shift)
1543 * for small pages, there is no difference between
1544 * shift and PAGE_SHIFT, so the bit-shift is safe.
1545 * for huge pages, since vm_pgoff is in units of small
1546 * pages, we need to shift off the always 0 bits to get
1549 BUG_ON(shift < PAGE_SHIFT);
1550 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1551 off += (addr - vma->vm_start) >> shift;
1552 return offset_il_node(pol, vma, off);
1554 return interleave_nodes(pol);
1557 #ifdef CONFIG_HUGETLBFS
1559 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1560 * @vma = virtual memory area whose policy is sought
1561 * @addr = address in @vma for shared policy lookup and interleave policy
1562 * @gfp_flags = for requested zone
1563 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1564 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1566 * Returns a zonelist suitable for a huge page allocation and a pointer
1567 * to the struct mempolicy for conditional unref after allocation.
1568 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1569 * @nodemask for filtering the zonelist.
1571 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1572 gfp_t gfp_flags, struct mempolicy **mpol,
1573 nodemask_t **nodemask)
1575 struct zonelist *zl;
1577 *mpol = get_vma_policy(current, vma, addr);
1578 *nodemask = NULL; /* assume !MPOL_BIND */
1580 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1581 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1582 huge_page_shift(hstate_vma(vma))), gfp_flags);
1584 zl = policy_zonelist(gfp_flags, *mpol);
1585 if ((*mpol)->mode == MPOL_BIND)
1586 *nodemask = &(*mpol)->v.nodes;
1592 * init_nodemask_of_mempolicy
1594 * If the current task's mempolicy is "default" [NULL], return 'false'
1595 * to indicate default policy. Otherwise, extract the policy nodemask
1596 * for 'bind' or 'interleave' policy into the argument nodemask, or
1597 * initialize the argument nodemask to contain the single node for
1598 * 'preferred' or 'local' policy and return 'true' to indicate presence
1599 * of non-default mempolicy.
1601 * We don't bother with reference counting the mempolicy [mpol_get/put]
1602 * because the current task is examining it's own mempolicy and a task's
1603 * mempolicy is only ever changed by the task itself.
1605 * N.B., it is the caller's responsibility to free a returned nodemask.
1607 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1609 struct mempolicy *mempolicy;
1612 if (!(mask && current->mempolicy))
1615 mempolicy = current->mempolicy;
1616 switch (mempolicy->mode) {
1617 case MPOL_PREFERRED:
1618 if (mempolicy->flags & MPOL_F_LOCAL)
1619 nid = numa_node_id();
1621 nid = mempolicy->v.preferred_node;
1622 init_nodemask_of_node(mask, nid);
1627 case MPOL_INTERLEAVE:
1628 *mask = mempolicy->v.nodes;
1639 /* Allocate a page in interleaved policy.
1640 Own path because it needs to do special accounting. */
1641 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1644 struct zonelist *zl;
1647 zl = node_zonelist(nid, gfp);
1648 page = __alloc_pages(gfp, order, zl);
1649 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1650 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1655 * alloc_page_vma - Allocate a page for a VMA.
1658 * %GFP_USER user allocation.
1659 * %GFP_KERNEL kernel allocations,
1660 * %GFP_HIGHMEM highmem/user allocations,
1661 * %GFP_FS allocation should not call back into a file system.
1662 * %GFP_ATOMIC don't sleep.
1664 * @vma: Pointer to VMA or NULL if not available.
1665 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1667 * This function allocates a page from the kernel page pool and applies
1668 * a NUMA policy associated with the VMA or the current process.
1669 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1670 * mm_struct of the VMA to prevent it from going away. Should be used for
1671 * all allocations for pages that will be mapped into
1672 * user space. Returns NULL when no page can be allocated.
1674 * Should be called with the mm_sem of the vma hold.
1677 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1679 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1680 struct zonelist *zl;
1682 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1685 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1687 return alloc_page_interleave(gfp, 0, nid);
1689 zl = policy_zonelist(gfp, pol);
1690 if (unlikely(mpol_needs_cond_ref(pol))) {
1692 * slow path: ref counted shared policy
1694 struct page *page = __alloc_pages_nodemask(gfp, 0,
1695 zl, policy_nodemask(gfp, pol));
1700 * fast path: default or task policy
1702 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1706 * alloc_pages_current - Allocate pages.
1709 * %GFP_USER user allocation,
1710 * %GFP_KERNEL kernel allocation,
1711 * %GFP_HIGHMEM highmem allocation,
1712 * %GFP_FS don't call back into a file system.
1713 * %GFP_ATOMIC don't sleep.
1714 * @order: Power of two of allocation size in pages. 0 is a single page.
1716 * Allocate a page from the kernel page pool. When not in
1717 * interrupt context and apply the current process NUMA policy.
1718 * Returns NULL when no page can be allocated.
1720 * Don't call cpuset_update_task_memory_state() unless
1721 * 1) it's ok to take cpuset_sem (can WAIT), and
1722 * 2) allocating for current task (not interrupt).
1724 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1726 struct mempolicy *pol = current->mempolicy;
1728 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1729 pol = &default_policy;
1732 * No reference counting needed for current->mempolicy
1733 * nor system default_policy
1735 if (pol->mode == MPOL_INTERLEAVE)
1736 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1737 return __alloc_pages_nodemask(gfp, order,
1738 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1740 EXPORT_SYMBOL(alloc_pages_current);
1743 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1744 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1745 * with the mems_allowed returned by cpuset_mems_allowed(). This
1746 * keeps mempolicies cpuset relative after its cpuset moves. See
1747 * further kernel/cpuset.c update_nodemask().
1750 /* Slow path of a mempolicy duplicate */
1751 struct mempolicy *__mpol_dup(struct mempolicy *old)
1753 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1756 return ERR_PTR(-ENOMEM);
1758 if (current_cpuset_is_being_rebound()) {
1759 nodemask_t mems = cpuset_mems_allowed(current);
1760 mpol_rebind_policy(old, &mems);
1764 atomic_set(&new->refcnt, 1);
1769 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1770 * eliminate the * MPOL_F_* flags that require conditional ref and
1771 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1772 * after return. Use the returned value.
1774 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1775 * policy lookup, even if the policy needs/has extra ref on lookup.
1776 * shmem_readahead needs this.
1778 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1779 struct mempolicy *frompol)
1781 if (!mpol_needs_cond_ref(frompol))
1785 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1786 __mpol_put(frompol);
1790 static int mpol_match_intent(const struct mempolicy *a,
1791 const struct mempolicy *b)
1793 if (a->flags != b->flags)
1795 if (!mpol_store_user_nodemask(a))
1797 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1800 /* Slow path of a mempolicy comparison */
1801 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1805 if (a->mode != b->mode)
1807 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
1812 case MPOL_INTERLEAVE:
1813 return nodes_equal(a->v.nodes, b->v.nodes);
1814 case MPOL_PREFERRED:
1815 return a->v.preferred_node == b->v.preferred_node &&
1816 a->flags == b->flags;
1824 * Shared memory backing store policy support.
1826 * Remember policies even when nobody has shared memory mapped.
1827 * The policies are kept in Red-Black tree linked from the inode.
1828 * They are protected by the sp->lock spinlock, which should be held
1829 * for any accesses to the tree.
1832 /* lookup first element intersecting start-end */
1833 /* Caller holds sp->lock */
1834 static struct sp_node *
1835 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1837 struct rb_node *n = sp->root.rb_node;
1840 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1842 if (start >= p->end)
1844 else if (end <= p->start)
1852 struct sp_node *w = NULL;
1853 struct rb_node *prev = rb_prev(n);
1856 w = rb_entry(prev, struct sp_node, nd);
1857 if (w->end <= start)
1861 return rb_entry(n, struct sp_node, nd);
1864 /* Insert a new shared policy into the list. */
1865 /* Caller holds sp->lock */
1866 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1868 struct rb_node **p = &sp->root.rb_node;
1869 struct rb_node *parent = NULL;
1874 nd = rb_entry(parent, struct sp_node, nd);
1875 if (new->start < nd->start)
1877 else if (new->end > nd->end)
1878 p = &(*p)->rb_right;
1882 rb_link_node(&new->nd, parent, p);
1883 rb_insert_color(&new->nd, &sp->root);
1884 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1885 new->policy ? new->policy->mode : 0);
1888 /* Find shared policy intersecting idx */
1890 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1892 struct mempolicy *pol = NULL;
1895 if (!sp->root.rb_node)
1897 spin_lock(&sp->lock);
1898 sn = sp_lookup(sp, idx, idx+1);
1900 mpol_get(sn->policy);
1903 spin_unlock(&sp->lock);
1907 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1909 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1910 rb_erase(&n->nd, &sp->root);
1911 mpol_put(n->policy);
1912 kmem_cache_free(sn_cache, n);
1915 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1916 struct mempolicy *pol)
1918 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1925 pol->flags |= MPOL_F_SHARED; /* for unref */
1930 /* Replace a policy range. */
1931 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1932 unsigned long end, struct sp_node *new)
1934 struct sp_node *n, *new2 = NULL;
1937 spin_lock(&sp->lock);
1938 n = sp_lookup(sp, start, end);
1939 /* Take care of old policies in the same range. */
1940 while (n && n->start < end) {
1941 struct rb_node *next = rb_next(&n->nd);
1942 if (n->start >= start) {
1948 /* Old policy spanning whole new range. */
1951 spin_unlock(&sp->lock);
1952 new2 = sp_alloc(end, n->end, n->policy);
1958 sp_insert(sp, new2);
1966 n = rb_entry(next, struct sp_node, nd);
1970 spin_unlock(&sp->lock);
1972 mpol_put(new2->policy);
1973 kmem_cache_free(sn_cache, new2);
1979 * mpol_shared_policy_init - initialize shared policy for inode
1980 * @sp: pointer to inode shared policy
1981 * @mpol: struct mempolicy to install
1983 * Install non-NULL @mpol in inode's shared policy rb-tree.
1984 * On entry, the current task has a reference on a non-NULL @mpol.
1985 * This must be released on exit.
1986 * This is called at get_inode() calls and we can use GFP_KERNEL.
1988 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
1992 sp->root = RB_ROOT; /* empty tree == default mempolicy */
1993 spin_lock_init(&sp->lock);
1996 struct vm_area_struct pvma;
1997 struct mempolicy *new;
1998 NODEMASK_SCRATCH(scratch);
2002 /* contextualize the tmpfs mount point mempolicy */
2003 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2005 mpol_put(mpol); /* drop our ref on sb mpol */
2006 NODEMASK_SCRATCH_FREE(scratch);
2007 return; /* no valid nodemask intersection */
2011 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2012 task_unlock(current);
2013 mpol_put(mpol); /* drop our ref on sb mpol */
2015 NODEMASK_SCRATCH_FREE(scratch);
2020 /* Create pseudo-vma that contains just the policy */
2021 memset(&pvma, 0, sizeof(struct vm_area_struct));
2022 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2023 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2024 mpol_put(new); /* drop initial ref */
2025 NODEMASK_SCRATCH_FREE(scratch);
2029 int mpol_set_shared_policy(struct shared_policy *info,
2030 struct vm_area_struct *vma, struct mempolicy *npol)
2033 struct sp_node *new = NULL;
2034 unsigned long sz = vma_pages(vma);
2036 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2038 sz, npol ? npol->mode : -1,
2039 npol ? npol->flags : -1,
2040 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2043 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2047 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2049 kmem_cache_free(sn_cache, new);
2053 /* Free a backing policy store on inode delete. */
2054 void mpol_free_shared_policy(struct shared_policy *p)
2057 struct rb_node *next;
2059 if (!p->root.rb_node)
2061 spin_lock(&p->lock);
2062 next = rb_first(&p->root);
2064 n = rb_entry(next, struct sp_node, nd);
2065 next = rb_next(&n->nd);
2066 rb_erase(&n->nd, &p->root);
2067 mpol_put(n->policy);
2068 kmem_cache_free(sn_cache, n);
2070 spin_unlock(&p->lock);
2073 /* assumes fs == KERNEL_DS */
2074 void __init numa_policy_init(void)
2076 nodemask_t interleave_nodes;
2077 unsigned long largest = 0;
2078 int nid, prefer = 0;
2080 policy_cache = kmem_cache_create("numa_policy",
2081 sizeof(struct mempolicy),
2082 0, SLAB_PANIC, NULL);
2084 sn_cache = kmem_cache_create("shared_policy_node",
2085 sizeof(struct sp_node),
2086 0, SLAB_PANIC, NULL);
2089 * Set interleaving policy for system init. Interleaving is only
2090 * enabled across suitably sized nodes (default is >= 16MB), or
2091 * fall back to the largest node if they're all smaller.
2093 nodes_clear(interleave_nodes);
2094 for_each_node_state(nid, N_HIGH_MEMORY) {
2095 unsigned long total_pages = node_present_pages(nid);
2097 /* Preserve the largest node */
2098 if (largest < total_pages) {
2099 largest = total_pages;
2103 /* Interleave this node? */
2104 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2105 node_set(nid, interleave_nodes);
2108 /* All too small, use the largest */
2109 if (unlikely(nodes_empty(interleave_nodes)))
2110 node_set(prefer, interleave_nodes);
2112 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2113 printk("numa_policy_init: interleaving failed\n");
2116 /* Reset policy of current process to default */
2117 void numa_default_policy(void)
2119 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2123 * Parse and format mempolicy from/to strings
2127 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2128 * Used only for mpol_parse_str() and mpol_to_str()
2130 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
2131 static const char * const policy_types[] =
2132 { "default", "prefer", "bind", "interleave", "local" };
2137 * mpol_parse_str - parse string to mempolicy
2138 * @str: string containing mempolicy to parse
2139 * @mpol: pointer to struct mempolicy pointer, returned on success.
2140 * @no_context: flag whether to "contextualize" the mempolicy
2143 * <mode>[=<flags>][:<nodelist>]
2145 * if @no_context is true, save the input nodemask in w.user_nodemask in
2146 * the returned mempolicy. This will be used to "clone" the mempolicy in
2147 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2148 * mount option. Note that if 'static' or 'relative' mode flags were
2149 * specified, the input nodemask will already have been saved. Saving
2150 * it again is redundant, but safe.
2152 * On success, returns 0, else 1
2154 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2156 struct mempolicy *new = NULL;
2157 unsigned short uninitialized_var(mode);
2158 unsigned short uninitialized_var(mode_flags);
2160 char *nodelist = strchr(str, ':');
2161 char *flags = strchr(str, '=');
2166 /* NUL-terminate mode or flags string */
2168 if (nodelist_parse(nodelist, nodes))
2170 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2176 *flags++ = '\0'; /* terminate mode string */
2178 for (i = 0; i <= MPOL_LOCAL; i++) {
2179 if (!strcmp(str, policy_types[i])) {
2188 case MPOL_PREFERRED:
2190 * Insist on a nodelist of one node only
2193 char *rest = nodelist;
2194 while (isdigit(*rest))
2200 case MPOL_INTERLEAVE:
2202 * Default to online nodes with memory if no nodelist
2205 nodes = node_states[N_HIGH_MEMORY];
2209 * Don't allow a nodelist; mpol_new() checks flags
2213 mode = MPOL_PREFERRED;
2217 * Insist on a empty nodelist
2224 * Insist on a nodelist
2233 * Currently, we only support two mutually exclusive
2236 if (!strcmp(flags, "static"))
2237 mode_flags |= MPOL_F_STATIC_NODES;
2238 else if (!strcmp(flags, "relative"))
2239 mode_flags |= MPOL_F_RELATIVE_NODES;
2244 new = mpol_new(mode, mode_flags, &nodes);
2250 NODEMASK_SCRATCH(scratch);
2253 ret = mpol_set_nodemask(new, &nodes, scratch);
2254 task_unlock(current);
2257 NODEMASK_SCRATCH_FREE(scratch);
2265 /* save for contextualization */
2266 new->w.user_nodemask = nodes;
2270 /* Restore string for error message */
2279 #endif /* CONFIG_TMPFS */
2282 * mpol_to_str - format a mempolicy structure for printing
2283 * @buffer: to contain formatted mempolicy string
2284 * @maxlen: length of @buffer
2285 * @pol: pointer to mempolicy to be formatted
2286 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2288 * Convert a mempolicy into a string.
2289 * Returns the number of characters in buffer (if positive)
2290 * or an error (negative)
2292 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2297 unsigned short mode;
2298 unsigned short flags = pol ? pol->flags : 0;
2301 * Sanity check: room for longest mode, flag and some nodes
2303 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2305 if (!pol || pol == &default_policy)
2306 mode = MPOL_DEFAULT;
2315 case MPOL_PREFERRED:
2317 if (flags & MPOL_F_LOCAL)
2318 mode = MPOL_LOCAL; /* pseudo-policy */
2320 node_set(pol->v.preferred_node, nodes);
2325 case MPOL_INTERLEAVE:
2327 nodes = pol->w.user_nodemask;
2329 nodes = pol->v.nodes;
2336 l = strlen(policy_types[mode]);
2337 if (buffer + maxlen < p + l + 1)
2340 strcpy(p, policy_types[mode]);
2343 if (flags & MPOL_MODE_FLAGS) {
2344 if (buffer + maxlen < p + 2)
2349 * Currently, the only defined flags are mutually exclusive
2351 if (flags & MPOL_F_STATIC_NODES)
2352 p += snprintf(p, buffer + maxlen - p, "static");
2353 else if (flags & MPOL_F_RELATIVE_NODES)
2354 p += snprintf(p, buffer + maxlen - p, "relative");
2357 if (!nodes_empty(nodes)) {
2358 if (buffer + maxlen < p + 2)
2361 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2367 unsigned long pages;
2369 unsigned long active;
2370 unsigned long writeback;
2371 unsigned long mapcount_max;
2372 unsigned long dirty;
2373 unsigned long swapcache;
2374 unsigned long node[MAX_NUMNODES];
2377 static void gather_stats(struct page *page, void *private, int pte_dirty)
2379 struct numa_maps *md = private;
2380 int count = page_mapcount(page);
2383 if (pte_dirty || PageDirty(page))
2386 if (PageSwapCache(page))
2389 if (PageActive(page) || PageUnevictable(page))
2392 if (PageWriteback(page))
2398 if (count > md->mapcount_max)
2399 md->mapcount_max = count;
2401 md->node[page_to_nid(page)]++;
2404 #ifdef CONFIG_HUGETLB_PAGE
2405 static void check_huge_range(struct vm_area_struct *vma,
2406 unsigned long start, unsigned long end,
2407 struct numa_maps *md)
2411 struct hstate *h = hstate_vma(vma);
2412 unsigned long sz = huge_page_size(h);
2414 for (addr = start; addr < end; addr += sz) {
2415 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2416 addr & huge_page_mask(h));
2426 page = pte_page(pte);
2430 gather_stats(page, md, pte_dirty(*ptep));
2434 static inline void check_huge_range(struct vm_area_struct *vma,
2435 unsigned long start, unsigned long end,
2436 struct numa_maps *md)
2442 * Display pages allocated per node and memory policy via /proc.
2444 int show_numa_map(struct seq_file *m, void *v)
2446 struct proc_maps_private *priv = m->private;
2447 struct vm_area_struct *vma = v;
2448 struct numa_maps *md;
2449 struct file *file = vma->vm_file;
2450 struct mm_struct *mm = vma->vm_mm;
2451 struct mempolicy *pol;
2458 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2462 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2463 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2466 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2469 seq_printf(m, " file=");
2470 seq_path(m, &file->f_path, "\n\t= ");
2471 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2472 seq_printf(m, " heap");
2473 } else if (vma->vm_start <= mm->start_stack &&
2474 vma->vm_end >= mm->start_stack) {
2475 seq_printf(m, " stack");
2478 if (is_vm_hugetlb_page(vma)) {
2479 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2480 seq_printf(m, " huge");
2482 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2483 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2490 seq_printf(m," anon=%lu",md->anon);
2493 seq_printf(m," dirty=%lu",md->dirty);
2495 if (md->pages != md->anon && md->pages != md->dirty)
2496 seq_printf(m, " mapped=%lu", md->pages);
2498 if (md->mapcount_max > 1)
2499 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2502 seq_printf(m," swapcache=%lu", md->swapcache);
2504 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2505 seq_printf(m," active=%lu", md->active);
2508 seq_printf(m," writeback=%lu", md->writeback);
2510 for_each_node_state(n, N_HIGH_MEMORY)
2512 seq_printf(m, " N%d=%lu", n, md->node[n]);
2517 if (m->count < m->size)
2518 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;