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.
66 could replace all the switch()es with a mempolicy_ops structure.
69 #include <linux/mempolicy.h>
71 #include <linux/highmem.h>
72 #include <linux/hugetlb.h>
73 #include <linux/kernel.h>
74 #include <linux/sched.h>
75 #include <linux/nodemask.h>
76 #include <linux/cpuset.h>
77 #include <linux/gfp.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/module.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/rmap.h>
90 #include <linux/security.h>
91 #include <linux/syscalls.h>
93 #include <asm/tlbflush.h>
94 #include <asm/uaccess.h>
97 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
98 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
99 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
101 static struct kmem_cache *policy_cache;
102 static struct kmem_cache *sn_cache;
104 /* Highest zone. An specific allocation for a zone below that is not
106 enum zone_type policy_zone = 0;
108 struct mempolicy default_policy = {
109 .refcnt = ATOMIC_INIT(1), /* never free it */
110 .policy = MPOL_DEFAULT,
113 /* Check that the nodemask contains at least one populated zone */
114 static int is_valid_nodemask(nodemask_t *nodemask)
118 /* Check that there is something useful in this mask */
121 for_each_node_mask(nd, *nodemask) {
124 for (k = 0; k <= policy_zone; k++) {
125 z = &NODE_DATA(nd)->node_zones[k];
126 if (z->present_pages > 0)
134 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
136 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
139 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
140 const nodemask_t *rel)
143 nodes_fold(tmp, *orig, nodes_weight(*rel));
144 nodes_onto(*ret, tmp, *rel);
147 /* Create a new policy */
148 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
151 struct mempolicy *policy;
152 nodemask_t cpuset_context_nmask;
154 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
155 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
157 if (mode == MPOL_DEFAULT)
158 return (nodes && nodes_weight(*nodes)) ? ERR_PTR(-EINVAL) :
160 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
162 return ERR_PTR(-ENOMEM);
163 atomic_set(&policy->refcnt, 1);
164 cpuset_update_task_memory_state();
165 if (flags & MPOL_F_RELATIVE_NODES)
166 mpol_relative_nodemask(&cpuset_context_nmask, nodes,
167 &cpuset_current_mems_allowed);
169 nodes_and(cpuset_context_nmask, *nodes,
170 cpuset_current_mems_allowed);
172 case MPOL_INTERLEAVE:
173 if (nodes_empty(*nodes) || nodes_empty(cpuset_context_nmask))
175 policy->v.nodes = cpuset_context_nmask;
178 policy->v.preferred_node = first_node(cpuset_context_nmask);
179 if (policy->v.preferred_node >= MAX_NUMNODES)
183 if (!is_valid_nodemask(&cpuset_context_nmask))
185 policy->v.nodes = cpuset_context_nmask;
190 policy->policy = mode;
191 policy->flags = flags;
192 if (mpol_store_user_nodemask(policy))
193 policy->w.user_nodemask = *nodes;
195 policy->w.cpuset_mems_allowed = cpuset_mems_allowed(current);
199 kmem_cache_free(policy_cache, policy);
200 return ERR_PTR(-EINVAL);
203 /* Migrate a policy to a different set of nodes */
204 static void mpol_rebind_policy(struct mempolicy *pol,
205 const nodemask_t *newmask)
213 static_nodes = pol->flags & MPOL_F_STATIC_NODES;
214 relative_nodes = pol->flags & MPOL_F_RELATIVE_NODES;
215 if (!mpol_store_user_nodemask(pol) &&
216 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
219 switch (pol->policy) {
224 case MPOL_INTERLEAVE:
226 nodes_and(tmp, pol->w.user_nodemask, *newmask);
227 else if (relative_nodes)
228 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask,
231 nodes_remap(tmp, pol->v.nodes,
232 pol->w.cpuset_mems_allowed, *newmask);
233 pol->w.cpuset_mems_allowed = *newmask;
236 if (!node_isset(current->il_next, tmp)) {
237 current->il_next = next_node(current->il_next, tmp);
238 if (current->il_next >= MAX_NUMNODES)
239 current->il_next = first_node(tmp);
240 if (current->il_next >= MAX_NUMNODES)
241 current->il_next = numa_node_id();
246 int node = first_node(pol->w.user_nodemask);
248 if (node_isset(node, *newmask))
249 pol->v.preferred_node = node;
251 pol->v.preferred_node = -1;
252 } else if (relative_nodes) {
253 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask,
255 pol->v.preferred_node = first_node(tmp);
257 pol->v.preferred_node = node_remap(pol->v.preferred_node,
258 pol->w.cpuset_mems_allowed, *newmask);
259 pol->w.cpuset_mems_allowed = *newmask;
269 * Wrapper for mpol_rebind_policy() that just requires task
270 * pointer, and updates task mempolicy.
273 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
275 mpol_rebind_policy(tsk->mempolicy, new);
279 * Rebind each vma in mm to new nodemask.
281 * Call holding a reference to mm. Takes mm->mmap_sem during call.
284 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
286 struct vm_area_struct *vma;
288 down_write(&mm->mmap_sem);
289 for (vma = mm->mmap; vma; vma = vma->vm_next)
290 mpol_rebind_policy(vma->vm_policy, new);
291 up_write(&mm->mmap_sem);
294 static void gather_stats(struct page *, void *, int pte_dirty);
295 static void migrate_page_add(struct page *page, struct list_head *pagelist,
296 unsigned long flags);
298 /* Scan through pages checking if pages follow certain conditions. */
299 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
300 unsigned long addr, unsigned long end,
301 const nodemask_t *nodes, unsigned long flags,
308 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
313 if (!pte_present(*pte))
315 page = vm_normal_page(vma, addr, *pte);
319 * The check for PageReserved here is important to avoid
320 * handling zero pages and other pages that may have been
321 * marked special by the system.
323 * If the PageReserved would not be checked here then f.e.
324 * the location of the zero page could have an influence
325 * on MPOL_MF_STRICT, zero pages would be counted for
326 * the per node stats, and there would be useless attempts
327 * to put zero pages on the migration list.
329 if (PageReserved(page))
331 nid = page_to_nid(page);
332 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
335 if (flags & MPOL_MF_STATS)
336 gather_stats(page, private, pte_dirty(*pte));
337 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
338 migrate_page_add(page, private, flags);
341 } while (pte++, addr += PAGE_SIZE, addr != end);
342 pte_unmap_unlock(orig_pte, ptl);
346 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
347 unsigned long addr, unsigned long end,
348 const nodemask_t *nodes, unsigned long flags,
354 pmd = pmd_offset(pud, addr);
356 next = pmd_addr_end(addr, end);
357 if (pmd_none_or_clear_bad(pmd))
359 if (check_pte_range(vma, pmd, addr, next, nodes,
362 } while (pmd++, addr = next, addr != end);
366 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
367 unsigned long addr, unsigned long end,
368 const nodemask_t *nodes, unsigned long flags,
374 pud = pud_offset(pgd, addr);
376 next = pud_addr_end(addr, end);
377 if (pud_none_or_clear_bad(pud))
379 if (check_pmd_range(vma, pud, addr, next, nodes,
382 } while (pud++, addr = next, addr != end);
386 static inline int check_pgd_range(struct vm_area_struct *vma,
387 unsigned long addr, unsigned long end,
388 const nodemask_t *nodes, unsigned long flags,
394 pgd = pgd_offset(vma->vm_mm, addr);
396 next = pgd_addr_end(addr, end);
397 if (pgd_none_or_clear_bad(pgd))
399 if (check_pud_range(vma, pgd, addr, next, nodes,
402 } while (pgd++, addr = next, addr != end);
407 * Check if all pages in a range are on a set of nodes.
408 * If pagelist != NULL then isolate pages from the LRU and
409 * put them on the pagelist.
411 static struct vm_area_struct *
412 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
413 const nodemask_t *nodes, unsigned long flags, void *private)
416 struct vm_area_struct *first, *vma, *prev;
418 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
420 err = migrate_prep();
425 first = find_vma(mm, start);
427 return ERR_PTR(-EFAULT);
429 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
430 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
431 if (!vma->vm_next && vma->vm_end < end)
432 return ERR_PTR(-EFAULT);
433 if (prev && prev->vm_end < vma->vm_start)
434 return ERR_PTR(-EFAULT);
436 if (!is_vm_hugetlb_page(vma) &&
437 ((flags & MPOL_MF_STRICT) ||
438 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
439 vma_migratable(vma)))) {
440 unsigned long endvma = vma->vm_end;
444 if (vma->vm_start > start)
445 start = vma->vm_start;
446 err = check_pgd_range(vma, start, endvma, nodes,
449 first = ERR_PTR(err);
458 /* Apply policy to a single VMA */
459 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
462 struct mempolicy *old = vma->vm_policy;
464 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
465 vma->vm_start, vma->vm_end, vma->vm_pgoff,
466 vma->vm_ops, vma->vm_file,
467 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
469 if (vma->vm_ops && vma->vm_ops->set_policy)
470 err = vma->vm_ops->set_policy(vma, new);
473 vma->vm_policy = new;
479 /* Step 2: apply policy to a range and do splits. */
480 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
481 unsigned long end, struct mempolicy *new)
483 struct vm_area_struct *next;
487 for (; vma && vma->vm_start < end; vma = next) {
489 if (vma->vm_start < start)
490 err = split_vma(vma->vm_mm, vma, start, 1);
491 if (!err && vma->vm_end > end)
492 err = split_vma(vma->vm_mm, vma, end, 0);
494 err = policy_vma(vma, new);
502 * Update task->flags PF_MEMPOLICY bit: set iff non-default
503 * mempolicy. Allows more rapid checking of this (combined perhaps
504 * with other PF_* flag bits) on memory allocation hot code paths.
506 * If called from outside this file, the task 'p' should -only- be
507 * a newly forked child not yet visible on the task list, because
508 * manipulating the task flags of a visible task is not safe.
510 * The above limitation is why this routine has the funny name
511 * mpol_fix_fork_child_flag().
513 * It is also safe to call this with a task pointer of current,
514 * which the static wrapper mpol_set_task_struct_flag() does,
515 * for use within this file.
518 void mpol_fix_fork_child_flag(struct task_struct *p)
521 p->flags |= PF_MEMPOLICY;
523 p->flags &= ~PF_MEMPOLICY;
526 static void mpol_set_task_struct_flag(void)
528 mpol_fix_fork_child_flag(current);
531 /* Set the process memory policy */
532 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
535 struct mempolicy *new;
537 new = mpol_new(mode, flags, nodes);
540 mpol_free(current->mempolicy);
541 current->mempolicy = new;
542 mpol_set_task_struct_flag();
543 if (new && new->policy == MPOL_INTERLEAVE &&
544 nodes_weight(new->v.nodes))
545 current->il_next = first_node(new->v.nodes);
549 /* Fill a zone bitmap for a policy */
550 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
558 case MPOL_INTERLEAVE:
562 /* or use current node instead of memory_map? */
563 if (p->v.preferred_node < 0)
564 *nodes = node_states[N_HIGH_MEMORY];
566 node_set(p->v.preferred_node, *nodes);
573 static int lookup_node(struct mm_struct *mm, unsigned long addr)
578 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
580 err = page_to_nid(p);
586 /* Retrieve NUMA policy */
587 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
588 unsigned long addr, unsigned long flags)
591 struct mm_struct *mm = current->mm;
592 struct vm_area_struct *vma = NULL;
593 struct mempolicy *pol = current->mempolicy;
595 cpuset_update_task_memory_state();
597 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
600 if (flags & MPOL_F_MEMS_ALLOWED) {
601 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
603 *policy = 0; /* just so it's initialized */
604 *nmask = cpuset_current_mems_allowed;
608 if (flags & MPOL_F_ADDR) {
609 down_read(&mm->mmap_sem);
610 vma = find_vma_intersection(mm, addr, addr+1);
612 up_read(&mm->mmap_sem);
615 if (vma->vm_ops && vma->vm_ops->get_policy)
616 pol = vma->vm_ops->get_policy(vma, addr);
618 pol = vma->vm_policy;
623 pol = &default_policy;
625 if (flags & MPOL_F_NODE) {
626 if (flags & MPOL_F_ADDR) {
627 err = lookup_node(mm, addr);
631 } else if (pol == current->mempolicy &&
632 pol->policy == MPOL_INTERLEAVE) {
633 *policy = current->il_next;
639 *policy = pol->policy | pol->flags;
642 up_read(¤t->mm->mmap_sem);
648 get_zonemask(pol, nmask);
652 up_read(¤t->mm->mmap_sem);
656 #ifdef CONFIG_MIGRATION
660 static void migrate_page_add(struct page *page, struct list_head *pagelist,
664 * Avoid migrating a page that is shared with others.
666 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
667 isolate_lru_page(page, pagelist);
670 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
672 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
676 * Migrate pages from one node to a target node.
677 * Returns error or the number of pages not migrated.
679 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
687 node_set(source, nmask);
689 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
690 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
692 if (!list_empty(&pagelist))
693 err = migrate_pages(&pagelist, new_node_page, dest);
699 * Move pages between the two nodesets so as to preserve the physical
700 * layout as much as possible.
702 * Returns the number of page that could not be moved.
704 int do_migrate_pages(struct mm_struct *mm,
705 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
712 down_read(&mm->mmap_sem);
714 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
719 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
720 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
721 * bit in 'tmp', and return that <source, dest> pair for migration.
722 * The pair of nodemasks 'to' and 'from' define the map.
724 * If no pair of bits is found that way, fallback to picking some
725 * pair of 'source' and 'dest' bits that are not the same. If the
726 * 'source' and 'dest' bits are the same, this represents a node
727 * that will be migrating to itself, so no pages need move.
729 * If no bits are left in 'tmp', or if all remaining bits left
730 * in 'tmp' correspond to the same bit in 'to', return false
731 * (nothing left to migrate).
733 * This lets us pick a pair of nodes to migrate between, such that
734 * if possible the dest node is not already occupied by some other
735 * source node, minimizing the risk of overloading the memory on a
736 * node that would happen if we migrated incoming memory to a node
737 * before migrating outgoing memory source that same node.
739 * A single scan of tmp is sufficient. As we go, we remember the
740 * most recent <s, d> pair that moved (s != d). If we find a pair
741 * that not only moved, but what's better, moved to an empty slot
742 * (d is not set in tmp), then we break out then, with that pair.
743 * Otherwise when we finish scannng from_tmp, we at least have the
744 * most recent <s, d> pair that moved. If we get all the way through
745 * the scan of tmp without finding any node that moved, much less
746 * moved to an empty node, then there is nothing left worth migrating.
750 while (!nodes_empty(tmp)) {
755 for_each_node_mask(s, tmp) {
756 d = node_remap(s, *from_nodes, *to_nodes);
760 source = s; /* Node moved. Memorize */
763 /* dest not in remaining from nodes? */
764 if (!node_isset(dest, tmp))
770 node_clear(source, tmp);
771 err = migrate_to_node(mm, source, dest, flags);
778 up_read(&mm->mmap_sem);
786 * Allocate a new page for page migration based on vma policy.
787 * Start assuming that page is mapped by vma pointed to by @private.
788 * Search forward from there, if not. N.B., this assumes that the
789 * list of pages handed to migrate_pages()--which is how we get here--
790 * is in virtual address order.
792 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
794 struct vm_area_struct *vma = (struct vm_area_struct *)private;
795 unsigned long uninitialized_var(address);
798 address = page_address_in_vma(page, vma);
799 if (address != -EFAULT)
805 * if !vma, alloc_page_vma() will use task or system default policy
807 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
811 static void migrate_page_add(struct page *page, struct list_head *pagelist,
816 int do_migrate_pages(struct mm_struct *mm,
817 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
822 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
828 static long do_mbind(unsigned long start, unsigned long len,
829 unsigned short mode, unsigned short mode_flags,
830 nodemask_t *nmask, unsigned long flags)
832 struct vm_area_struct *vma;
833 struct mm_struct *mm = current->mm;
834 struct mempolicy *new;
839 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
840 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
842 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
845 if (start & ~PAGE_MASK)
848 if (mode == MPOL_DEFAULT)
849 flags &= ~MPOL_MF_STRICT;
851 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
859 new = mpol_new(mode, mode_flags, nmask);
864 * If we are using the default policy then operation
865 * on discontinuous address spaces is okay after all
868 flags |= MPOL_MF_DISCONTIG_OK;
870 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
871 start, start + len, mode, mode_flags,
872 nmask ? nodes_addr(*nmask)[0] : -1);
874 down_write(&mm->mmap_sem);
875 vma = check_range(mm, start, end, nmask,
876 flags | MPOL_MF_INVERT, &pagelist);
882 err = mbind_range(vma, start, end, new);
884 if (!list_empty(&pagelist))
885 nr_failed = migrate_pages(&pagelist, new_vma_page,
888 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
892 up_write(&mm->mmap_sem);
898 * User space interface with variable sized bitmaps for nodelists.
901 /* Copy a node mask from user space. */
902 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
903 unsigned long maxnode)
906 unsigned long nlongs;
907 unsigned long endmask;
911 if (maxnode == 0 || !nmask)
913 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
916 nlongs = BITS_TO_LONGS(maxnode);
917 if ((maxnode % BITS_PER_LONG) == 0)
920 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
922 /* When the user specified more nodes than supported just check
923 if the non supported part is all zero. */
924 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
925 if (nlongs > PAGE_SIZE/sizeof(long))
927 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
929 if (get_user(t, nmask + k))
931 if (k == nlongs - 1) {
937 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
941 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
943 nodes_addr(*nodes)[nlongs-1] &= endmask;
947 /* Copy a kernel node mask to user space */
948 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
951 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
952 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
955 if (copy > PAGE_SIZE)
957 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
961 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
964 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
966 unsigned long __user *nmask, unsigned long maxnode,
971 unsigned short mode_flags;
973 mode_flags = mode & MPOL_MODE_FLAGS;
974 mode &= ~MPOL_MODE_FLAGS;
975 if (mode >= MPOL_MAX)
977 if ((mode_flags & MPOL_F_STATIC_NODES) &&
978 (mode_flags & MPOL_F_RELATIVE_NODES))
980 err = get_nodes(&nodes, nmask, maxnode);
983 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
986 /* Set the process memory policy */
987 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
988 unsigned long maxnode)
992 unsigned short flags;
994 flags = mode & MPOL_MODE_FLAGS;
995 mode &= ~MPOL_MODE_FLAGS;
996 if ((unsigned int)mode >= MPOL_MAX)
998 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1000 err = get_nodes(&nodes, nmask, maxnode);
1003 return do_set_mempolicy(mode, flags, &nodes);
1006 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
1007 const unsigned long __user *old_nodes,
1008 const unsigned long __user *new_nodes)
1010 struct mm_struct *mm;
1011 struct task_struct *task;
1014 nodemask_t task_nodes;
1017 err = get_nodes(&old, old_nodes, maxnode);
1021 err = get_nodes(&new, new_nodes, maxnode);
1025 /* Find the mm_struct */
1026 read_lock(&tasklist_lock);
1027 task = pid ? find_task_by_vpid(pid) : current;
1029 read_unlock(&tasklist_lock);
1032 mm = get_task_mm(task);
1033 read_unlock(&tasklist_lock);
1039 * Check if this process has the right to modify the specified
1040 * process. The right exists if the process has administrative
1041 * capabilities, superuser privileges or the same
1042 * userid as the target process.
1044 if ((current->euid != task->suid) && (current->euid != task->uid) &&
1045 (current->uid != task->suid) && (current->uid != task->uid) &&
1046 !capable(CAP_SYS_NICE)) {
1051 task_nodes = cpuset_mems_allowed(task);
1052 /* Is the user allowed to access the target nodes? */
1053 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1058 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1063 err = security_task_movememory(task);
1067 err = do_migrate_pages(mm, &old, &new,
1068 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1075 /* Retrieve NUMA policy */
1076 asmlinkage long sys_get_mempolicy(int __user *policy,
1077 unsigned long __user *nmask,
1078 unsigned long maxnode,
1079 unsigned long addr, unsigned long flags)
1082 int uninitialized_var(pval);
1085 if (nmask != NULL && maxnode < MAX_NUMNODES)
1088 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1093 if (policy && put_user(pval, policy))
1097 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1102 #ifdef CONFIG_COMPAT
1104 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1105 compat_ulong_t __user *nmask,
1106 compat_ulong_t maxnode,
1107 compat_ulong_t addr, compat_ulong_t flags)
1110 unsigned long __user *nm = NULL;
1111 unsigned long nr_bits, alloc_size;
1112 DECLARE_BITMAP(bm, MAX_NUMNODES);
1114 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1115 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1118 nm = compat_alloc_user_space(alloc_size);
1120 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1122 if (!err && nmask) {
1123 err = copy_from_user(bm, nm, alloc_size);
1124 /* ensure entire bitmap is zeroed */
1125 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1126 err |= compat_put_bitmap(nmask, bm, nr_bits);
1132 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1133 compat_ulong_t maxnode)
1136 unsigned long __user *nm = NULL;
1137 unsigned long nr_bits, alloc_size;
1138 DECLARE_BITMAP(bm, MAX_NUMNODES);
1140 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1141 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1144 err = compat_get_bitmap(bm, nmask, nr_bits);
1145 nm = compat_alloc_user_space(alloc_size);
1146 err |= copy_to_user(nm, bm, alloc_size);
1152 return sys_set_mempolicy(mode, nm, nr_bits+1);
1155 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1156 compat_ulong_t mode, compat_ulong_t __user *nmask,
1157 compat_ulong_t maxnode, compat_ulong_t flags)
1160 unsigned long __user *nm = NULL;
1161 unsigned long nr_bits, alloc_size;
1164 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1165 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1168 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1169 nm = compat_alloc_user_space(alloc_size);
1170 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1176 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1182 * get_vma_policy(@task, @vma, @addr)
1183 * @task - task for fallback if vma policy == default
1184 * @vma - virtual memory area whose policy is sought
1185 * @addr - address in @vma for shared policy lookup
1187 * Returns effective policy for a VMA at specified address.
1188 * Falls back to @task or system default policy, as necessary.
1189 * Returned policy has extra reference count if shared, vma,
1190 * or some other task's policy [show_numa_maps() can pass
1191 * @task != current]. It is the caller's responsibility to
1192 * free the reference in these cases.
1194 static struct mempolicy * get_vma_policy(struct task_struct *task,
1195 struct vm_area_struct *vma, unsigned long addr)
1197 struct mempolicy *pol = task->mempolicy;
1201 if (vma->vm_ops && vma->vm_ops->get_policy) {
1202 pol = vma->vm_ops->get_policy(vma, addr);
1203 shared_pol = 1; /* if pol non-NULL, add ref below */
1204 } else if (vma->vm_policy &&
1205 vma->vm_policy->policy != MPOL_DEFAULT)
1206 pol = vma->vm_policy;
1209 pol = &default_policy;
1210 else if (!shared_pol && pol != current->mempolicy)
1211 mpol_get(pol); /* vma or other task's policy */
1215 /* Return a nodemask representing a mempolicy */
1216 static nodemask_t *nodemask_policy(gfp_t gfp, struct mempolicy *policy)
1218 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1219 if (unlikely(policy->policy == MPOL_BIND) &&
1220 gfp_zone(gfp) >= policy_zone &&
1221 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1222 return &policy->v.nodes;
1227 /* Return a zonelist representing a mempolicy */
1228 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1232 switch (policy->policy) {
1233 case MPOL_PREFERRED:
1234 nd = policy->v.preferred_node;
1236 nd = numa_node_id();
1240 * Normally, MPOL_BIND allocations node-local are node-local
1241 * within the allowed nodemask. However, if __GFP_THISNODE is
1242 * set and the current node is part of the mask, we use the
1243 * the zonelist for the first node in the mask instead.
1245 nd = numa_node_id();
1246 if (unlikely(gfp & __GFP_THISNODE) &&
1247 unlikely(!node_isset(nd, policy->v.nodes)))
1248 nd = first_node(policy->v.nodes);
1250 case MPOL_INTERLEAVE: /* should not happen */
1252 nd = numa_node_id();
1258 return node_zonelist(nd, gfp);
1261 /* Do dynamic interleaving for a process */
1262 static unsigned interleave_nodes(struct mempolicy *policy)
1265 struct task_struct *me = current;
1268 next = next_node(nid, policy->v.nodes);
1269 if (next >= MAX_NUMNODES)
1270 next = first_node(policy->v.nodes);
1271 if (next < MAX_NUMNODES)
1277 * Depending on the memory policy provide a node from which to allocate the
1280 unsigned slab_node(struct mempolicy *policy)
1282 unsigned short pol = policy ? policy->policy : MPOL_DEFAULT;
1285 case MPOL_INTERLEAVE:
1286 return interleave_nodes(policy);
1290 * Follow bind policy behavior and start allocation at the
1293 struct zonelist *zonelist;
1295 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1296 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1297 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1303 case MPOL_PREFERRED:
1304 if (policy->v.preferred_node >= 0)
1305 return policy->v.preferred_node;
1309 return numa_node_id();
1313 /* Do static interleaving for a VMA with known offset. */
1314 static unsigned offset_il_node(struct mempolicy *pol,
1315 struct vm_area_struct *vma, unsigned long off)
1317 unsigned nnodes = nodes_weight(pol->v.nodes);
1323 return numa_node_id();
1324 target = (unsigned int)off % nnodes;
1327 nid = next_node(nid, pol->v.nodes);
1329 } while (c <= target);
1333 /* Determine a node number for interleave */
1334 static inline unsigned interleave_nid(struct mempolicy *pol,
1335 struct vm_area_struct *vma, unsigned long addr, int shift)
1341 * for small pages, there is no difference between
1342 * shift and PAGE_SHIFT, so the bit-shift is safe.
1343 * for huge pages, since vm_pgoff is in units of small
1344 * pages, we need to shift off the always 0 bits to get
1347 BUG_ON(shift < PAGE_SHIFT);
1348 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1349 off += (addr - vma->vm_start) >> shift;
1350 return offset_il_node(pol, vma, off);
1352 return interleave_nodes(pol);
1355 #ifdef CONFIG_HUGETLBFS
1357 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1358 * @vma = virtual memory area whose policy is sought
1359 * @addr = address in @vma for shared policy lookup and interleave policy
1360 * @gfp_flags = for requested zone
1361 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1362 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1364 * Returns a zonelist suitable for a huge page allocation.
1365 * If the effective policy is 'BIND, returns pointer to local node's zonelist,
1366 * and a pointer to the mempolicy's @nodemask for filtering the zonelist.
1367 * If it is also a policy for which get_vma_policy() returns an extra
1368 * reference, we must hold that reference until after the allocation.
1369 * In that case, return policy via @mpol so hugetlb allocation can drop
1370 * the reference. For non-'BIND referenced policies, we can/do drop the
1371 * reference here, so the caller doesn't need to know about the special case
1372 * for default and current task policy.
1374 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1375 gfp_t gfp_flags, struct mempolicy **mpol,
1376 nodemask_t **nodemask)
1378 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1379 struct zonelist *zl;
1381 *mpol = NULL; /* probably no unref needed */
1382 *nodemask = NULL; /* assume !MPOL_BIND */
1383 if (pol->policy == MPOL_BIND) {
1384 *nodemask = &pol->v.nodes;
1385 } else if (pol->policy == MPOL_INTERLEAVE) {
1388 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1389 if (unlikely(pol != &default_policy &&
1390 pol != current->mempolicy))
1391 __mpol_free(pol); /* finished with pol */
1392 return node_zonelist(nid, gfp_flags);
1395 zl = zonelist_policy(GFP_HIGHUSER, pol);
1396 if (unlikely(pol != &default_policy && pol != current->mempolicy)) {
1397 if (pol->policy != MPOL_BIND)
1398 __mpol_free(pol); /* finished with pol */
1400 *mpol = pol; /* unref needed after allocation */
1406 /* Allocate a page in interleaved policy.
1407 Own path because it needs to do special accounting. */
1408 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1411 struct zonelist *zl;
1414 zl = node_zonelist(nid, gfp);
1415 page = __alloc_pages(gfp, order, zl);
1416 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1417 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1422 * alloc_page_vma - Allocate a page for a VMA.
1425 * %GFP_USER user allocation.
1426 * %GFP_KERNEL kernel allocations,
1427 * %GFP_HIGHMEM highmem/user allocations,
1428 * %GFP_FS allocation should not call back into a file system.
1429 * %GFP_ATOMIC don't sleep.
1431 * @vma: Pointer to VMA or NULL if not available.
1432 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1434 * This function allocates a page from the kernel page pool and applies
1435 * a NUMA policy associated with the VMA or the current process.
1436 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1437 * mm_struct of the VMA to prevent it from going away. Should be used for
1438 * all allocations for pages that will be mapped into
1439 * user space. Returns NULL when no page can be allocated.
1441 * Should be called with the mm_sem of the vma hold.
1444 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1446 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1447 struct zonelist *zl;
1449 cpuset_update_task_memory_state();
1451 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1454 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1455 if (unlikely(pol != &default_policy &&
1456 pol != current->mempolicy))
1457 __mpol_free(pol); /* finished with pol */
1458 return alloc_page_interleave(gfp, 0, nid);
1460 zl = zonelist_policy(gfp, pol);
1461 if (pol != &default_policy && pol != current->mempolicy) {
1463 * slow path: ref counted policy -- shared or vma
1465 struct page *page = __alloc_pages_nodemask(gfp, 0,
1466 zl, nodemask_policy(gfp, pol));
1471 * fast path: default or task policy
1473 return __alloc_pages_nodemask(gfp, 0, zl, nodemask_policy(gfp, pol));
1477 * alloc_pages_current - Allocate pages.
1480 * %GFP_USER user allocation,
1481 * %GFP_KERNEL kernel allocation,
1482 * %GFP_HIGHMEM highmem allocation,
1483 * %GFP_FS don't call back into a file system.
1484 * %GFP_ATOMIC don't sleep.
1485 * @order: Power of two of allocation size in pages. 0 is a single page.
1487 * Allocate a page from the kernel page pool. When not in
1488 * interrupt context and apply the current process NUMA policy.
1489 * Returns NULL when no page can be allocated.
1491 * Don't call cpuset_update_task_memory_state() unless
1492 * 1) it's ok to take cpuset_sem (can WAIT), and
1493 * 2) allocating for current task (not interrupt).
1495 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1497 struct mempolicy *pol = current->mempolicy;
1499 if ((gfp & __GFP_WAIT) && !in_interrupt())
1500 cpuset_update_task_memory_state();
1501 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1502 pol = &default_policy;
1503 if (pol->policy == MPOL_INTERLEAVE)
1504 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1505 return __alloc_pages_nodemask(gfp, order,
1506 zonelist_policy(gfp, pol), nodemask_policy(gfp, pol));
1508 EXPORT_SYMBOL(alloc_pages_current);
1511 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1512 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1513 * with the mems_allowed returned by cpuset_mems_allowed(). This
1514 * keeps mempolicies cpuset relative after its cpuset moves. See
1515 * further kernel/cpuset.c update_nodemask().
1518 /* Slow path of a mempolicy copy */
1519 struct mempolicy *__mpol_copy(struct mempolicy *old)
1521 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1524 return ERR_PTR(-ENOMEM);
1525 if (current_cpuset_is_being_rebound()) {
1526 nodemask_t mems = cpuset_mems_allowed(current);
1527 mpol_rebind_policy(old, &mems);
1530 atomic_set(&new->refcnt, 1);
1534 static int mpol_match_intent(const struct mempolicy *a,
1535 const struct mempolicy *b)
1537 if (a->flags != b->flags)
1539 if (!mpol_store_user_nodemask(a))
1541 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1544 /* Slow path of a mempolicy comparison */
1545 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1549 if (a->policy != b->policy)
1551 if (a->policy != MPOL_DEFAULT && !mpol_match_intent(a, b))
1553 switch (a->policy) {
1558 case MPOL_INTERLEAVE:
1559 return nodes_equal(a->v.nodes, b->v.nodes);
1560 case MPOL_PREFERRED:
1561 return a->v.preferred_node == b->v.preferred_node;
1568 /* Slow path of a mpol destructor. */
1569 void __mpol_free(struct mempolicy *p)
1571 if (!atomic_dec_and_test(&p->refcnt))
1573 p->policy = MPOL_DEFAULT;
1574 kmem_cache_free(policy_cache, p);
1578 * Shared memory backing store policy support.
1580 * Remember policies even when nobody has shared memory mapped.
1581 * The policies are kept in Red-Black tree linked from the inode.
1582 * They are protected by the sp->lock spinlock, which should be held
1583 * for any accesses to the tree.
1586 /* lookup first element intersecting start-end */
1587 /* Caller holds sp->lock */
1588 static struct sp_node *
1589 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1591 struct rb_node *n = sp->root.rb_node;
1594 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1596 if (start >= p->end)
1598 else if (end <= p->start)
1606 struct sp_node *w = NULL;
1607 struct rb_node *prev = rb_prev(n);
1610 w = rb_entry(prev, struct sp_node, nd);
1611 if (w->end <= start)
1615 return rb_entry(n, struct sp_node, nd);
1618 /* Insert a new shared policy into the list. */
1619 /* Caller holds sp->lock */
1620 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1622 struct rb_node **p = &sp->root.rb_node;
1623 struct rb_node *parent = NULL;
1628 nd = rb_entry(parent, struct sp_node, nd);
1629 if (new->start < nd->start)
1631 else if (new->end > nd->end)
1632 p = &(*p)->rb_right;
1636 rb_link_node(&new->nd, parent, p);
1637 rb_insert_color(&new->nd, &sp->root);
1638 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1639 new->policy ? new->policy->policy : 0);
1642 /* Find shared policy intersecting idx */
1644 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1646 struct mempolicy *pol = NULL;
1649 if (!sp->root.rb_node)
1651 spin_lock(&sp->lock);
1652 sn = sp_lookup(sp, idx, idx+1);
1654 mpol_get(sn->policy);
1657 spin_unlock(&sp->lock);
1661 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1663 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1664 rb_erase(&n->nd, &sp->root);
1665 mpol_free(n->policy);
1666 kmem_cache_free(sn_cache, n);
1669 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1670 struct mempolicy *pol)
1672 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1683 /* Replace a policy range. */
1684 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1685 unsigned long end, struct sp_node *new)
1687 struct sp_node *n, *new2 = NULL;
1690 spin_lock(&sp->lock);
1691 n = sp_lookup(sp, start, end);
1692 /* Take care of old policies in the same range. */
1693 while (n && n->start < end) {
1694 struct rb_node *next = rb_next(&n->nd);
1695 if (n->start >= start) {
1701 /* Old policy spanning whole new range. */
1704 spin_unlock(&sp->lock);
1705 new2 = sp_alloc(end, n->end, n->policy);
1711 sp_insert(sp, new2);
1719 n = rb_entry(next, struct sp_node, nd);
1723 spin_unlock(&sp->lock);
1725 mpol_free(new2->policy);
1726 kmem_cache_free(sn_cache, new2);
1731 void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
1732 unsigned short flags, nodemask_t *policy_nodes)
1734 info->root = RB_ROOT;
1735 spin_lock_init(&info->lock);
1737 if (policy != MPOL_DEFAULT) {
1738 struct mempolicy *newpol;
1740 /* Falls back to MPOL_DEFAULT on any error */
1741 newpol = mpol_new(policy, flags, policy_nodes);
1742 if (!IS_ERR(newpol)) {
1743 /* Create pseudo-vma that contains just the policy */
1744 struct vm_area_struct pvma;
1746 memset(&pvma, 0, sizeof(struct vm_area_struct));
1747 /* Policy covers entire file */
1748 pvma.vm_end = TASK_SIZE;
1749 mpol_set_shared_policy(info, &pvma, newpol);
1755 int mpol_set_shared_policy(struct shared_policy *info,
1756 struct vm_area_struct *vma, struct mempolicy *npol)
1759 struct sp_node *new = NULL;
1760 unsigned long sz = vma_pages(vma);
1762 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1764 sz, npol ? npol->policy : -1,
1765 npol ? npol->flags : -1,
1766 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1769 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1773 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1775 kmem_cache_free(sn_cache, new);
1779 /* Free a backing policy store on inode delete. */
1780 void mpol_free_shared_policy(struct shared_policy *p)
1783 struct rb_node *next;
1785 if (!p->root.rb_node)
1787 spin_lock(&p->lock);
1788 next = rb_first(&p->root);
1790 n = rb_entry(next, struct sp_node, nd);
1791 next = rb_next(&n->nd);
1792 rb_erase(&n->nd, &p->root);
1793 mpol_free(n->policy);
1794 kmem_cache_free(sn_cache, n);
1796 spin_unlock(&p->lock);
1799 /* assumes fs == KERNEL_DS */
1800 void __init numa_policy_init(void)
1802 nodemask_t interleave_nodes;
1803 unsigned long largest = 0;
1804 int nid, prefer = 0;
1806 policy_cache = kmem_cache_create("numa_policy",
1807 sizeof(struct mempolicy),
1808 0, SLAB_PANIC, NULL);
1810 sn_cache = kmem_cache_create("shared_policy_node",
1811 sizeof(struct sp_node),
1812 0, SLAB_PANIC, NULL);
1815 * Set interleaving policy for system init. Interleaving is only
1816 * enabled across suitably sized nodes (default is >= 16MB), or
1817 * fall back to the largest node if they're all smaller.
1819 nodes_clear(interleave_nodes);
1820 for_each_node_state(nid, N_HIGH_MEMORY) {
1821 unsigned long total_pages = node_present_pages(nid);
1823 /* Preserve the largest node */
1824 if (largest < total_pages) {
1825 largest = total_pages;
1829 /* Interleave this node? */
1830 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1831 node_set(nid, interleave_nodes);
1834 /* All too small, use the largest */
1835 if (unlikely(nodes_empty(interleave_nodes)))
1836 node_set(prefer, interleave_nodes);
1838 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
1839 printk("numa_policy_init: interleaving failed\n");
1842 /* Reset policy of current process to default */
1843 void numa_default_policy(void)
1845 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1849 * Display pages allocated per node and memory policy via /proc.
1852 static const char * const policy_types[] =
1853 { "default", "prefer", "bind", "interleave" };
1856 * Convert a mempolicy into a string.
1857 * Returns the number of characters in buffer (if positive)
1858 * or an error (negative)
1860 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1865 unsigned short mode = pol ? pol->policy : MPOL_DEFAULT;
1866 unsigned short flags = pol ? pol->flags : 0;
1873 case MPOL_PREFERRED:
1875 node_set(pol->v.preferred_node, nodes);
1880 case MPOL_INTERLEAVE:
1881 nodes = pol->v.nodes;
1889 l = strlen(policy_types[mode]);
1890 if (buffer + maxlen < p + l + 1)
1893 strcpy(p, policy_types[mode]);
1899 if (buffer + maxlen < p + 2)
1903 if (flags & MPOL_F_STATIC_NODES)
1904 p += sprintf(p, "%sstatic", need_bar++ ? "|" : "");
1905 if (flags & MPOL_F_RELATIVE_NODES)
1906 p += sprintf(p, "%srelative", need_bar++ ? "|" : "");
1909 if (!nodes_empty(nodes)) {
1910 if (buffer + maxlen < p + 2)
1913 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1919 unsigned long pages;
1921 unsigned long active;
1922 unsigned long writeback;
1923 unsigned long mapcount_max;
1924 unsigned long dirty;
1925 unsigned long swapcache;
1926 unsigned long node[MAX_NUMNODES];
1929 static void gather_stats(struct page *page, void *private, int pte_dirty)
1931 struct numa_maps *md = private;
1932 int count = page_mapcount(page);
1935 if (pte_dirty || PageDirty(page))
1938 if (PageSwapCache(page))
1941 if (PageActive(page))
1944 if (PageWriteback(page))
1950 if (count > md->mapcount_max)
1951 md->mapcount_max = count;
1953 md->node[page_to_nid(page)]++;
1956 #ifdef CONFIG_HUGETLB_PAGE
1957 static void check_huge_range(struct vm_area_struct *vma,
1958 unsigned long start, unsigned long end,
1959 struct numa_maps *md)
1964 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1965 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1975 page = pte_page(pte);
1979 gather_stats(page, md, pte_dirty(*ptep));
1983 static inline void check_huge_range(struct vm_area_struct *vma,
1984 unsigned long start, unsigned long end,
1985 struct numa_maps *md)
1990 int show_numa_map(struct seq_file *m, void *v)
1992 struct proc_maps_private *priv = m->private;
1993 struct vm_area_struct *vma = v;
1994 struct numa_maps *md;
1995 struct file *file = vma->vm_file;
1996 struct mm_struct *mm = vma->vm_mm;
1997 struct mempolicy *pol;
2004 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2008 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2009 mpol_to_str(buffer, sizeof(buffer), pol);
2011 * unref shared or other task's mempolicy
2013 if (pol != &default_policy && pol != current->mempolicy)
2016 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2019 seq_printf(m, " file=");
2020 seq_path(m, &file->f_path, "\n\t= ");
2021 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2022 seq_printf(m, " heap");
2023 } else if (vma->vm_start <= mm->start_stack &&
2024 vma->vm_end >= mm->start_stack) {
2025 seq_printf(m, " stack");
2028 if (is_vm_hugetlb_page(vma)) {
2029 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2030 seq_printf(m, " huge");
2032 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2033 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2040 seq_printf(m," anon=%lu",md->anon);
2043 seq_printf(m," dirty=%lu",md->dirty);
2045 if (md->pages != md->anon && md->pages != md->dirty)
2046 seq_printf(m, " mapped=%lu", md->pages);
2048 if (md->mapcount_max > 1)
2049 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2052 seq_printf(m," swapcache=%lu", md->swapcache);
2054 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2055 seq_printf(m," active=%lu", md->active);
2058 seq_printf(m," writeback=%lu", md->writeback);
2060 for_each_node_state(n, N_HIGH_MEMORY)
2062 seq_printf(m, " N%d=%lu", n, md->node[n]);
2067 if (m->count < m->size)
2068 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;