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/interrupt.h>
82 #include <linux/init.h>
83 #include <linux/compat.h>
84 #include <linux/swap.h>
85 #include <linux/seq_file.h>
86 #include <linux/proc_fs.h>
87 #include <linux/migrate.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
91 #include <asm/tlbflush.h>
92 #include <asm/uaccess.h>
95 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
96 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
97 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
99 static struct kmem_cache *policy_cache;
100 static struct kmem_cache *sn_cache;
102 /* Highest zone. An specific allocation for a zone below that is not
104 enum zone_type policy_zone = 0;
106 struct mempolicy default_policy = {
107 .refcnt = ATOMIC_INIT(1), /* never free it */
108 .policy = MPOL_DEFAULT,
111 /* Do sanity checking on a policy */
112 static int mpol_check_policy(int mode, nodemask_t *nodes)
114 int empty = nodes_empty(*nodes);
122 case MPOL_INTERLEAVE:
123 /* Preferred will only use the first bit, but allow
129 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
132 /* Generate a custom zonelist for the BIND policy. */
133 static struct zonelist *bind_zonelist(nodemask_t *nodes)
139 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
140 max++; /* space for zlcache_ptr (see mmzone.h) */
141 zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL);
143 return ERR_PTR(-ENOMEM);
144 zl->zlcache_ptr = NULL;
146 /* First put in the highest zones from all nodes, then all the next
147 lower zones etc. Avoid empty zones because the memory allocator
148 doesn't like them. If you implement node hot removal you
150 k = MAX_NR_ZONES - 1;
152 for_each_node_mask(nd, *nodes) {
153 struct zone *z = &NODE_DATA(nd)->node_zones[k];
154 if (z->present_pages > 0)
155 zl->zones[num++] = z;
163 return ERR_PTR(-EINVAL);
165 zl->zones[num] = NULL;
169 /* Create a new policy */
170 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
172 struct mempolicy *policy;
174 pr_debug("setting mode %d nodes[0] %lx\n",
175 mode, nodes ? nodes_addr(*nodes)[0] : -1);
177 if (mode == MPOL_DEFAULT)
179 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
181 return ERR_PTR(-ENOMEM);
182 atomic_set(&policy->refcnt, 1);
184 case MPOL_INTERLEAVE:
185 policy->v.nodes = *nodes;
186 if (nodes_weight(*nodes) == 0) {
187 kmem_cache_free(policy_cache, policy);
188 return ERR_PTR(-EINVAL);
192 policy->v.preferred_node = first_node(*nodes);
193 if (policy->v.preferred_node >= MAX_NUMNODES)
194 policy->v.preferred_node = -1;
197 policy->v.zonelist = bind_zonelist(nodes);
198 if (IS_ERR(policy->v.zonelist)) {
199 void *error_code = policy->v.zonelist;
200 kmem_cache_free(policy_cache, policy);
205 policy->policy = mode;
206 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
210 static void gather_stats(struct page *, void *, int pte_dirty);
211 static void migrate_page_add(struct page *page, struct list_head *pagelist,
212 unsigned long flags);
214 /* Scan through pages checking if pages follow certain conditions. */
215 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
216 unsigned long addr, unsigned long end,
217 const nodemask_t *nodes, unsigned long flags,
224 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
229 if (!pte_present(*pte))
231 page = vm_normal_page(vma, addr, *pte);
235 * The check for PageReserved here is important to avoid
236 * handling zero pages and other pages that may have been
237 * marked special by the system.
239 * If the PageReserved would not be checked here then f.e.
240 * the location of the zero page could have an influence
241 * on MPOL_MF_STRICT, zero pages would be counted for
242 * the per node stats, and there would be useless attempts
243 * to put zero pages on the migration list.
245 if (PageReserved(page))
247 nid = page_to_nid(page);
248 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
251 if (flags & MPOL_MF_STATS)
252 gather_stats(page, private, pte_dirty(*pte));
253 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
254 migrate_page_add(page, private, flags);
257 } while (pte++, addr += PAGE_SIZE, addr != end);
258 pte_unmap_unlock(orig_pte, ptl);
262 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
263 unsigned long addr, unsigned long end,
264 const nodemask_t *nodes, unsigned long flags,
270 pmd = pmd_offset(pud, addr);
272 next = pmd_addr_end(addr, end);
273 if (pmd_none_or_clear_bad(pmd))
275 if (check_pte_range(vma, pmd, addr, next, nodes,
278 } while (pmd++, addr = next, addr != end);
282 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
283 unsigned long addr, unsigned long end,
284 const nodemask_t *nodes, unsigned long flags,
290 pud = pud_offset(pgd, addr);
292 next = pud_addr_end(addr, end);
293 if (pud_none_or_clear_bad(pud))
295 if (check_pmd_range(vma, pud, addr, next, nodes,
298 } while (pud++, addr = next, addr != end);
302 static inline int check_pgd_range(struct vm_area_struct *vma,
303 unsigned long addr, unsigned long end,
304 const nodemask_t *nodes, unsigned long flags,
310 pgd = pgd_offset(vma->vm_mm, addr);
312 next = pgd_addr_end(addr, end);
313 if (pgd_none_or_clear_bad(pgd))
315 if (check_pud_range(vma, pgd, addr, next, nodes,
318 } while (pgd++, addr = next, addr != end);
323 * Check if all pages in a range are on a set of nodes.
324 * If pagelist != NULL then isolate pages from the LRU and
325 * put them on the pagelist.
327 static struct vm_area_struct *
328 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
329 const nodemask_t *nodes, unsigned long flags, void *private)
332 struct vm_area_struct *first, *vma, *prev;
334 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
336 err = migrate_prep();
341 first = find_vma(mm, start);
343 return ERR_PTR(-EFAULT);
345 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
346 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
347 if (!vma->vm_next && vma->vm_end < end)
348 return ERR_PTR(-EFAULT);
349 if (prev && prev->vm_end < vma->vm_start)
350 return ERR_PTR(-EFAULT);
352 if (!is_vm_hugetlb_page(vma) &&
353 ((flags & MPOL_MF_STRICT) ||
354 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
355 vma_migratable(vma)))) {
356 unsigned long endvma = vma->vm_end;
360 if (vma->vm_start > start)
361 start = vma->vm_start;
362 err = check_pgd_range(vma, start, endvma, nodes,
365 first = ERR_PTR(err);
374 /* Apply policy to a single VMA */
375 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
378 struct mempolicy *old = vma->vm_policy;
380 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
381 vma->vm_start, vma->vm_end, vma->vm_pgoff,
382 vma->vm_ops, vma->vm_file,
383 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
385 if (vma->vm_ops && vma->vm_ops->set_policy)
386 err = vma->vm_ops->set_policy(vma, new);
389 vma->vm_policy = new;
395 /* Step 2: apply policy to a range and do splits. */
396 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
397 unsigned long end, struct mempolicy *new)
399 struct vm_area_struct *next;
403 for (; vma && vma->vm_start < end; vma = next) {
405 if (vma->vm_start < start)
406 err = split_vma(vma->vm_mm, vma, start, 1);
407 if (!err && vma->vm_end > end)
408 err = split_vma(vma->vm_mm, vma, end, 0);
410 err = policy_vma(vma, new);
417 static int contextualize_policy(int mode, nodemask_t *nodes)
422 cpuset_update_task_memory_state();
423 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
425 return mpol_check_policy(mode, nodes);
430 * Update task->flags PF_MEMPOLICY bit: set iff non-default
431 * mempolicy. Allows more rapid checking of this (combined perhaps
432 * with other PF_* flag bits) on memory allocation hot code paths.
434 * If called from outside this file, the task 'p' should -only- be
435 * a newly forked child not yet visible on the task list, because
436 * manipulating the task flags of a visible task is not safe.
438 * The above limitation is why this routine has the funny name
439 * mpol_fix_fork_child_flag().
441 * It is also safe to call this with a task pointer of current,
442 * which the static wrapper mpol_set_task_struct_flag() does,
443 * for use within this file.
446 void mpol_fix_fork_child_flag(struct task_struct *p)
449 p->flags |= PF_MEMPOLICY;
451 p->flags &= ~PF_MEMPOLICY;
454 static void mpol_set_task_struct_flag(void)
456 mpol_fix_fork_child_flag(current);
459 /* Set the process memory policy */
460 long do_set_mempolicy(int mode, nodemask_t *nodes)
462 struct mempolicy *new;
464 if (contextualize_policy(mode, nodes))
466 new = mpol_new(mode, nodes);
469 mpol_free(current->mempolicy);
470 current->mempolicy = new;
471 mpol_set_task_struct_flag();
472 if (new && new->policy == MPOL_INTERLEAVE)
473 current->il_next = first_node(new->v.nodes);
477 /* Fill a zone bitmap for a policy */
478 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
485 for (i = 0; p->v.zonelist->zones[i]; i++)
486 node_set(zone_to_nid(p->v.zonelist->zones[i]),
491 case MPOL_INTERLEAVE:
495 /* or use current node instead of online map? */
496 if (p->v.preferred_node < 0)
497 *nodes = node_online_map;
499 node_set(p->v.preferred_node, *nodes);
506 static int lookup_node(struct mm_struct *mm, unsigned long addr)
511 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
513 err = page_to_nid(p);
519 /* Retrieve NUMA policy */
520 long do_get_mempolicy(int *policy, nodemask_t *nmask,
521 unsigned long addr, unsigned long flags)
524 struct mm_struct *mm = current->mm;
525 struct vm_area_struct *vma = NULL;
526 struct mempolicy *pol = current->mempolicy;
528 cpuset_update_task_memory_state();
529 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
531 if (flags & MPOL_F_ADDR) {
532 down_read(&mm->mmap_sem);
533 vma = find_vma_intersection(mm, addr, addr+1);
535 up_read(&mm->mmap_sem);
538 if (vma->vm_ops && vma->vm_ops->get_policy)
539 pol = vma->vm_ops->get_policy(vma, addr);
541 pol = vma->vm_policy;
546 pol = &default_policy;
548 if (flags & MPOL_F_NODE) {
549 if (flags & MPOL_F_ADDR) {
550 err = lookup_node(mm, addr);
554 } else if (pol == current->mempolicy &&
555 pol->policy == MPOL_INTERLEAVE) {
556 *policy = current->il_next;
562 *policy = pol->policy;
565 up_read(¤t->mm->mmap_sem);
571 get_zonemask(pol, nmask);
575 up_read(¤t->mm->mmap_sem);
579 #ifdef CONFIG_MIGRATION
583 static void migrate_page_add(struct page *page, struct list_head *pagelist,
587 * Avoid migrating a page that is shared with others.
589 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
590 isolate_lru_page(page, pagelist);
593 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
595 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
599 * Migrate pages from one node to a target node.
600 * Returns error or the number of pages not migrated.
602 int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags)
609 node_set(source, nmask);
611 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
612 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
614 if (!list_empty(&pagelist))
615 err = migrate_pages(&pagelist, new_node_page, dest);
621 * Move pages between the two nodesets so as to preserve the physical
622 * layout as much as possible.
624 * Returns the number of page that could not be moved.
626 int do_migrate_pages(struct mm_struct *mm,
627 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
634 down_read(&mm->mmap_sem);
636 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
641 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
642 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
643 * bit in 'tmp', and return that <source, dest> pair for migration.
644 * The pair of nodemasks 'to' and 'from' define the map.
646 * If no pair of bits is found that way, fallback to picking some
647 * pair of 'source' and 'dest' bits that are not the same. If the
648 * 'source' and 'dest' bits are the same, this represents a node
649 * that will be migrating to itself, so no pages need move.
651 * If no bits are left in 'tmp', or if all remaining bits left
652 * in 'tmp' correspond to the same bit in 'to', return false
653 * (nothing left to migrate).
655 * This lets us pick a pair of nodes to migrate between, such that
656 * if possible the dest node is not already occupied by some other
657 * source node, minimizing the risk of overloading the memory on a
658 * node that would happen if we migrated incoming memory to a node
659 * before migrating outgoing memory source that same node.
661 * A single scan of tmp is sufficient. As we go, we remember the
662 * most recent <s, d> pair that moved (s != d). If we find a pair
663 * that not only moved, but what's better, moved to an empty slot
664 * (d is not set in tmp), then we break out then, with that pair.
665 * Otherwise when we finish scannng from_tmp, we at least have the
666 * most recent <s, d> pair that moved. If we get all the way through
667 * the scan of tmp without finding any node that moved, much less
668 * moved to an empty node, then there is nothing left worth migrating.
672 while (!nodes_empty(tmp)) {
677 for_each_node_mask(s, tmp) {
678 d = node_remap(s, *from_nodes, *to_nodes);
682 source = s; /* Node moved. Memorize */
685 /* dest not in remaining from nodes? */
686 if (!node_isset(dest, tmp))
692 node_clear(source, tmp);
693 err = migrate_to_node(mm, source, dest, flags);
700 up_read(&mm->mmap_sem);
707 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
709 struct vm_area_struct *vma = (struct vm_area_struct *)private;
711 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
712 page_address_in_vma(page, vma));
716 static void migrate_page_add(struct page *page, struct list_head *pagelist,
721 int do_migrate_pages(struct mm_struct *mm,
722 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
727 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
733 long do_mbind(unsigned long start, unsigned long len,
734 unsigned long mode, nodemask_t *nmask, unsigned long flags)
736 struct vm_area_struct *vma;
737 struct mm_struct *mm = current->mm;
738 struct mempolicy *new;
743 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
744 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
747 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
750 if (start & ~PAGE_MASK)
753 if (mode == MPOL_DEFAULT)
754 flags &= ~MPOL_MF_STRICT;
756 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
764 if (mpol_check_policy(mode, nmask))
767 new = mpol_new(mode, nmask);
772 * If we are using the default policy then operation
773 * on discontinuous address spaces is okay after all
776 flags |= MPOL_MF_DISCONTIG_OK;
778 pr_debug("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
779 mode, nmask ? nodes_addr(*nmask)[0] : -1);
781 down_write(&mm->mmap_sem);
782 vma = check_range(mm, start, end, nmask,
783 flags | MPOL_MF_INVERT, &pagelist);
789 err = mbind_range(vma, start, end, new);
791 if (!list_empty(&pagelist))
792 nr_failed = migrate_pages(&pagelist, new_vma_page,
795 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
799 up_write(&mm->mmap_sem);
805 * User space interface with variable sized bitmaps for nodelists.
808 /* Copy a node mask from user space. */
809 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
810 unsigned long maxnode)
813 unsigned long nlongs;
814 unsigned long endmask;
818 if (maxnode == 0 || !nmask)
820 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
823 nlongs = BITS_TO_LONGS(maxnode);
824 if ((maxnode % BITS_PER_LONG) == 0)
827 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
829 /* When the user specified more nodes than supported just check
830 if the non supported part is all zero. */
831 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
832 if (nlongs > PAGE_SIZE/sizeof(long))
834 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
836 if (get_user(t, nmask + k))
838 if (k == nlongs - 1) {
844 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
848 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
850 nodes_addr(*nodes)[nlongs-1] &= endmask;
854 /* Copy a kernel node mask to user space */
855 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
858 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
859 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
862 if (copy > PAGE_SIZE)
864 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
868 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
871 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
873 unsigned long __user *nmask, unsigned long maxnode,
879 err = get_nodes(&nodes, nmask, maxnode);
882 #ifdef CONFIG_CPUSETS
883 /* Restrict the nodes to the allowed nodes in the cpuset */
884 nodes_and(nodes, nodes, current->mems_allowed);
886 return do_mbind(start, len, mode, &nodes, flags);
889 /* Set the process memory policy */
890 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
891 unsigned long maxnode)
896 if (mode < 0 || mode > MPOL_MAX)
898 err = get_nodes(&nodes, nmask, maxnode);
901 return do_set_mempolicy(mode, &nodes);
904 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
905 const unsigned long __user *old_nodes,
906 const unsigned long __user *new_nodes)
908 struct mm_struct *mm;
909 struct task_struct *task;
912 nodemask_t task_nodes;
915 err = get_nodes(&old, old_nodes, maxnode);
919 err = get_nodes(&new, new_nodes, maxnode);
923 /* Find the mm_struct */
924 read_lock(&tasklist_lock);
925 task = pid ? find_task_by_pid(pid) : current;
927 read_unlock(&tasklist_lock);
930 mm = get_task_mm(task);
931 read_unlock(&tasklist_lock);
937 * Check if this process has the right to modify the specified
938 * process. The right exists if the process has administrative
939 * capabilities, superuser privileges or the same
940 * userid as the target process.
942 if ((current->euid != task->suid) && (current->euid != task->uid) &&
943 (current->uid != task->suid) && (current->uid != task->uid) &&
944 !capable(CAP_SYS_NICE)) {
949 task_nodes = cpuset_mems_allowed(task);
950 /* Is the user allowed to access the target nodes? */
951 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
956 if (!nodes_subset(new, node_online_map)) {
961 err = security_task_movememory(task);
965 err = do_migrate_pages(mm, &old, &new,
966 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
973 /* Retrieve NUMA policy */
974 asmlinkage long sys_get_mempolicy(int __user *policy,
975 unsigned long __user *nmask,
976 unsigned long maxnode,
977 unsigned long addr, unsigned long flags)
982 if (nmask != NULL && maxnode < MAX_NUMNODES)
985 err = do_get_mempolicy(&pval, &nodes, addr, flags);
990 if (policy && put_user(pval, policy))
994 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1001 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1002 compat_ulong_t __user *nmask,
1003 compat_ulong_t maxnode,
1004 compat_ulong_t addr, compat_ulong_t flags)
1007 unsigned long __user *nm = NULL;
1008 unsigned long nr_bits, alloc_size;
1009 DECLARE_BITMAP(bm, MAX_NUMNODES);
1011 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1012 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1015 nm = compat_alloc_user_space(alloc_size);
1017 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1019 if (!err && nmask) {
1020 err = copy_from_user(bm, nm, alloc_size);
1021 /* ensure entire bitmap is zeroed */
1022 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1023 err |= compat_put_bitmap(nmask, bm, nr_bits);
1029 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1030 compat_ulong_t maxnode)
1033 unsigned long __user *nm = NULL;
1034 unsigned long nr_bits, alloc_size;
1035 DECLARE_BITMAP(bm, MAX_NUMNODES);
1037 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1038 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1041 err = compat_get_bitmap(bm, nmask, nr_bits);
1042 nm = compat_alloc_user_space(alloc_size);
1043 err |= copy_to_user(nm, bm, alloc_size);
1049 return sys_set_mempolicy(mode, nm, nr_bits+1);
1052 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1053 compat_ulong_t mode, compat_ulong_t __user *nmask,
1054 compat_ulong_t maxnode, compat_ulong_t flags)
1057 unsigned long __user *nm = NULL;
1058 unsigned long nr_bits, alloc_size;
1061 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1062 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1065 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1066 nm = compat_alloc_user_space(alloc_size);
1067 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1073 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1079 * get_vma_policy(@task, @vma, @addr)
1080 * @task - task for fallback if vma policy == default
1081 * @vma - virtual memory area whose policy is sought
1082 * @addr - address in @vma for shared policy lookup
1084 * Returns effective policy for a VMA at specified address.
1085 * Falls back to @task or system default policy, as necessary.
1086 * Returned policy has extra reference count if shared, vma,
1087 * or some other task's policy [show_numa_maps() can pass
1088 * @task != current]. It is the caller's responsibility to
1089 * free the reference in these cases.
1091 static struct mempolicy * get_vma_policy(struct task_struct *task,
1092 struct vm_area_struct *vma, unsigned long addr)
1094 struct mempolicy *pol = task->mempolicy;
1098 if (vma->vm_ops && vma->vm_ops->get_policy) {
1099 pol = vma->vm_ops->get_policy(vma, addr);
1100 shared_pol = 1; /* if pol non-NULL, add ref below */
1101 } else if (vma->vm_policy &&
1102 vma->vm_policy->policy != MPOL_DEFAULT)
1103 pol = vma->vm_policy;
1106 pol = &default_policy;
1107 else if (!shared_pol && pol != current->mempolicy)
1108 mpol_get(pol); /* vma or other task's policy */
1112 /* Return a zonelist representing a mempolicy */
1113 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1117 switch (policy->policy) {
1118 case MPOL_PREFERRED:
1119 nd = policy->v.preferred_node;
1121 nd = numa_node_id();
1124 /* Lower zones don't get a policy applied */
1125 /* Careful: current->mems_allowed might have moved */
1126 if (gfp_zone(gfp) >= policy_zone)
1127 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
1128 return policy->v.zonelist;
1130 case MPOL_INTERLEAVE: /* should not happen */
1132 nd = numa_node_id();
1138 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
1141 /* Do dynamic interleaving for a process */
1142 static unsigned interleave_nodes(struct mempolicy *policy)
1145 struct task_struct *me = current;
1148 next = next_node(nid, policy->v.nodes);
1149 if (next >= MAX_NUMNODES)
1150 next = first_node(policy->v.nodes);
1156 * Depending on the memory policy provide a node from which to allocate the
1159 unsigned slab_node(struct mempolicy *policy)
1161 int pol = policy ? policy->policy : MPOL_DEFAULT;
1164 case MPOL_INTERLEAVE:
1165 return interleave_nodes(policy);
1169 * Follow bind policy behavior and start allocation at the
1172 return zone_to_nid(policy->v.zonelist->zones[0]);
1174 case MPOL_PREFERRED:
1175 if (policy->v.preferred_node >= 0)
1176 return policy->v.preferred_node;
1180 return numa_node_id();
1184 /* Do static interleaving for a VMA with known offset. */
1185 static unsigned offset_il_node(struct mempolicy *pol,
1186 struct vm_area_struct *vma, unsigned long off)
1188 unsigned nnodes = nodes_weight(pol->v.nodes);
1189 unsigned target = (unsigned)off % nnodes;
1195 nid = next_node(nid, pol->v.nodes);
1197 } while (c <= target);
1201 /* Determine a node number for interleave */
1202 static inline unsigned interleave_nid(struct mempolicy *pol,
1203 struct vm_area_struct *vma, unsigned long addr, int shift)
1209 * for small pages, there is no difference between
1210 * shift and PAGE_SHIFT, so the bit-shift is safe.
1211 * for huge pages, since vm_pgoff is in units of small
1212 * pages, we need to shift off the always 0 bits to get
1215 BUG_ON(shift < PAGE_SHIFT);
1216 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1217 off += (addr - vma->vm_start) >> shift;
1218 return offset_il_node(pol, vma, off);
1220 return interleave_nodes(pol);
1223 #ifdef CONFIG_HUGETLBFS
1225 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1226 * @vma = virtual memory area whose policy is sought
1227 * @addr = address in @vma for shared policy lookup and interleave policy
1228 * @gfp_flags = for requested zone
1229 * @mpol = pointer to mempolicy pointer for reference counted 'BIND policy
1231 * Returns a zonelist suitable for a huge page allocation.
1232 * If the effective policy is 'BIND, returns pointer to policy's zonelist.
1233 * If it is also a policy for which get_vma_policy() returns an extra
1234 * reference, we must hold that reference until after allocation.
1235 * In that case, return policy via @mpol so hugetlb allocation can drop
1236 * the reference. For non-'BIND referenced policies, we can/do drop the
1237 * reference here, so the caller doesn't need to know about the special case
1238 * for default and current task policy.
1240 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1241 gfp_t gfp_flags, struct mempolicy **mpol)
1243 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1244 struct zonelist *zl;
1246 *mpol = NULL; /* probably no unref needed */
1247 if (pol->policy == MPOL_INTERLEAVE) {
1250 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1251 __mpol_free(pol); /* finished with pol */
1252 return NODE_DATA(nid)->node_zonelists + gfp_zone(gfp_flags);
1255 zl = zonelist_policy(GFP_HIGHUSER, pol);
1256 if (unlikely(pol != &default_policy && pol != current->mempolicy)) {
1257 if (pol->policy != MPOL_BIND)
1258 __mpol_free(pol); /* finished with pol */
1260 *mpol = pol; /* unref needed after allocation */
1266 /* Allocate a page in interleaved policy.
1267 Own path because it needs to do special accounting. */
1268 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1271 struct zonelist *zl;
1274 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1275 page = __alloc_pages(gfp, order, zl);
1276 if (page && page_zone(page) == zl->zones[0])
1277 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1282 * alloc_page_vma - Allocate a page for a VMA.
1285 * %GFP_USER user allocation.
1286 * %GFP_KERNEL kernel allocations,
1287 * %GFP_HIGHMEM highmem/user allocations,
1288 * %GFP_FS allocation should not call back into a file system.
1289 * %GFP_ATOMIC don't sleep.
1291 * @vma: Pointer to VMA or NULL if not available.
1292 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1294 * This function allocates a page from the kernel page pool and applies
1295 * a NUMA policy associated with the VMA or the current process.
1296 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1297 * mm_struct of the VMA to prevent it from going away. Should be used for
1298 * all allocations for pages that will be mapped into
1299 * user space. Returns NULL when no page can be allocated.
1301 * Should be called with the mm_sem of the vma hold.
1304 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1306 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1307 struct zonelist *zl;
1309 cpuset_update_task_memory_state();
1311 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1314 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1315 return alloc_page_interleave(gfp, 0, nid);
1317 zl = zonelist_policy(gfp, pol);
1318 if (pol != &default_policy && pol != current->mempolicy) {
1320 * slow path: ref counted policy -- shared or vma
1322 struct page *page = __alloc_pages(gfp, 0, zl);
1327 * fast path: default or task policy
1329 return __alloc_pages(gfp, 0, zl);
1333 * alloc_pages_current - Allocate pages.
1336 * %GFP_USER user allocation,
1337 * %GFP_KERNEL kernel allocation,
1338 * %GFP_HIGHMEM highmem allocation,
1339 * %GFP_FS don't call back into a file system.
1340 * %GFP_ATOMIC don't sleep.
1341 * @order: Power of two of allocation size in pages. 0 is a single page.
1343 * Allocate a page from the kernel page pool. When not in
1344 * interrupt context and apply the current process NUMA policy.
1345 * Returns NULL when no page can be allocated.
1347 * Don't call cpuset_update_task_memory_state() unless
1348 * 1) it's ok to take cpuset_sem (can WAIT), and
1349 * 2) allocating for current task (not interrupt).
1351 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1353 struct mempolicy *pol = current->mempolicy;
1355 if ((gfp & __GFP_WAIT) && !in_interrupt())
1356 cpuset_update_task_memory_state();
1357 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1358 pol = &default_policy;
1359 if (pol->policy == MPOL_INTERLEAVE)
1360 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1361 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1363 EXPORT_SYMBOL(alloc_pages_current);
1366 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1367 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1368 * with the mems_allowed returned by cpuset_mems_allowed(). This
1369 * keeps mempolicies cpuset relative after its cpuset moves. See
1370 * further kernel/cpuset.c update_nodemask().
1372 void *cpuset_being_rebound;
1374 /* Slow path of a mempolicy copy */
1375 struct mempolicy *__mpol_copy(struct mempolicy *old)
1377 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1380 return ERR_PTR(-ENOMEM);
1381 if (current_cpuset_is_being_rebound()) {
1382 nodemask_t mems = cpuset_mems_allowed(current);
1383 mpol_rebind_policy(old, &mems);
1386 atomic_set(&new->refcnt, 1);
1387 if (new->policy == MPOL_BIND) {
1388 int sz = ksize(old->v.zonelist);
1389 new->v.zonelist = kmemdup(old->v.zonelist, sz, GFP_KERNEL);
1390 if (!new->v.zonelist) {
1391 kmem_cache_free(policy_cache, new);
1392 return ERR_PTR(-ENOMEM);
1398 /* Slow path of a mempolicy comparison */
1399 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1403 if (a->policy != b->policy)
1405 switch (a->policy) {
1408 case MPOL_INTERLEAVE:
1409 return nodes_equal(a->v.nodes, b->v.nodes);
1410 case MPOL_PREFERRED:
1411 return a->v.preferred_node == b->v.preferred_node;
1414 for (i = 0; a->v.zonelist->zones[i]; i++)
1415 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1417 return b->v.zonelist->zones[i] == NULL;
1425 /* Slow path of a mpol destructor. */
1426 void __mpol_free(struct mempolicy *p)
1428 if (!atomic_dec_and_test(&p->refcnt))
1430 if (p->policy == MPOL_BIND)
1431 kfree(p->v.zonelist);
1432 p->policy = MPOL_DEFAULT;
1433 kmem_cache_free(policy_cache, p);
1437 * Shared memory backing store policy support.
1439 * Remember policies even when nobody has shared memory mapped.
1440 * The policies are kept in Red-Black tree linked from the inode.
1441 * They are protected by the sp->lock spinlock, which should be held
1442 * for any accesses to the tree.
1445 /* lookup first element intersecting start-end */
1446 /* Caller holds sp->lock */
1447 static struct sp_node *
1448 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1450 struct rb_node *n = sp->root.rb_node;
1453 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1455 if (start >= p->end)
1457 else if (end <= p->start)
1465 struct sp_node *w = NULL;
1466 struct rb_node *prev = rb_prev(n);
1469 w = rb_entry(prev, struct sp_node, nd);
1470 if (w->end <= start)
1474 return rb_entry(n, struct sp_node, nd);
1477 /* Insert a new shared policy into the list. */
1478 /* Caller holds sp->lock */
1479 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1481 struct rb_node **p = &sp->root.rb_node;
1482 struct rb_node *parent = NULL;
1487 nd = rb_entry(parent, struct sp_node, nd);
1488 if (new->start < nd->start)
1490 else if (new->end > nd->end)
1491 p = &(*p)->rb_right;
1495 rb_link_node(&new->nd, parent, p);
1496 rb_insert_color(&new->nd, &sp->root);
1497 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1498 new->policy ? new->policy->policy : 0);
1501 /* Find shared policy intersecting idx */
1503 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1505 struct mempolicy *pol = NULL;
1508 if (!sp->root.rb_node)
1510 spin_lock(&sp->lock);
1511 sn = sp_lookup(sp, idx, idx+1);
1513 mpol_get(sn->policy);
1516 spin_unlock(&sp->lock);
1520 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1522 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1523 rb_erase(&n->nd, &sp->root);
1524 mpol_free(n->policy);
1525 kmem_cache_free(sn_cache, n);
1529 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1531 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1542 /* Replace a policy range. */
1543 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1544 unsigned long end, struct sp_node *new)
1546 struct sp_node *n, *new2 = NULL;
1549 spin_lock(&sp->lock);
1550 n = sp_lookup(sp, start, end);
1551 /* Take care of old policies in the same range. */
1552 while (n && n->start < end) {
1553 struct rb_node *next = rb_next(&n->nd);
1554 if (n->start >= start) {
1560 /* Old policy spanning whole new range. */
1563 spin_unlock(&sp->lock);
1564 new2 = sp_alloc(end, n->end, n->policy);
1570 sp_insert(sp, new2);
1578 n = rb_entry(next, struct sp_node, nd);
1582 spin_unlock(&sp->lock);
1584 mpol_free(new2->policy);
1585 kmem_cache_free(sn_cache, new2);
1590 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1591 nodemask_t *policy_nodes)
1593 info->root = RB_ROOT;
1594 spin_lock_init(&info->lock);
1596 if (policy != MPOL_DEFAULT) {
1597 struct mempolicy *newpol;
1599 /* Falls back to MPOL_DEFAULT on any error */
1600 newpol = mpol_new(policy, policy_nodes);
1601 if (!IS_ERR(newpol)) {
1602 /* Create pseudo-vma that contains just the policy */
1603 struct vm_area_struct pvma;
1605 memset(&pvma, 0, sizeof(struct vm_area_struct));
1606 /* Policy covers entire file */
1607 pvma.vm_end = TASK_SIZE;
1608 mpol_set_shared_policy(info, &pvma, newpol);
1614 int mpol_set_shared_policy(struct shared_policy *info,
1615 struct vm_area_struct *vma, struct mempolicy *npol)
1618 struct sp_node *new = NULL;
1619 unsigned long sz = vma_pages(vma);
1621 pr_debug("set_shared_policy %lx sz %lu %d %lx\n",
1623 sz, npol? npol->policy : -1,
1624 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1627 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1631 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1633 kmem_cache_free(sn_cache, new);
1637 /* Free a backing policy store on inode delete. */
1638 void mpol_free_shared_policy(struct shared_policy *p)
1641 struct rb_node *next;
1643 if (!p->root.rb_node)
1645 spin_lock(&p->lock);
1646 next = rb_first(&p->root);
1648 n = rb_entry(next, struct sp_node, nd);
1649 next = rb_next(&n->nd);
1650 rb_erase(&n->nd, &p->root);
1651 mpol_free(n->policy);
1652 kmem_cache_free(sn_cache, n);
1654 spin_unlock(&p->lock);
1657 /* assumes fs == KERNEL_DS */
1658 void __init numa_policy_init(void)
1660 nodemask_t interleave_nodes;
1661 unsigned long largest = 0;
1662 int nid, prefer = 0;
1664 policy_cache = kmem_cache_create("numa_policy",
1665 sizeof(struct mempolicy),
1666 0, SLAB_PANIC, NULL);
1668 sn_cache = kmem_cache_create("shared_policy_node",
1669 sizeof(struct sp_node),
1670 0, SLAB_PANIC, NULL);
1673 * Set interleaving policy for system init. Interleaving is only
1674 * enabled across suitably sized nodes (default is >= 16MB), or
1675 * fall back to the largest node if they're all smaller.
1677 nodes_clear(interleave_nodes);
1678 for_each_online_node(nid) {
1679 unsigned long total_pages = node_present_pages(nid);
1681 /* Preserve the largest node */
1682 if (largest < total_pages) {
1683 largest = total_pages;
1687 /* Interleave this node? */
1688 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1689 node_set(nid, interleave_nodes);
1692 /* All too small, use the largest */
1693 if (unlikely(nodes_empty(interleave_nodes)))
1694 node_set(prefer, interleave_nodes);
1696 if (do_set_mempolicy(MPOL_INTERLEAVE, &interleave_nodes))
1697 printk("numa_policy_init: interleaving failed\n");
1700 /* Reset policy of current process to default */
1701 void numa_default_policy(void)
1703 do_set_mempolicy(MPOL_DEFAULT, NULL);
1706 /* Migrate a policy to a different set of nodes */
1707 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1709 nodemask_t *mpolmask;
1714 mpolmask = &pol->cpuset_mems_allowed;
1715 if (nodes_equal(*mpolmask, *newmask))
1718 switch (pol->policy) {
1721 case MPOL_INTERLEAVE:
1722 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1724 *mpolmask = *newmask;
1725 current->il_next = node_remap(current->il_next,
1726 *mpolmask, *newmask);
1728 case MPOL_PREFERRED:
1729 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1730 *mpolmask, *newmask);
1731 *mpolmask = *newmask;
1736 struct zonelist *zonelist;
1739 for (z = pol->v.zonelist->zones; *z; z++)
1740 node_set(zone_to_nid(*z), nodes);
1741 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1744 zonelist = bind_zonelist(&nodes);
1746 /* If no mem, then zonelist is NULL and we keep old zonelist.
1747 * If that old zonelist has no remaining mems_allowed nodes,
1748 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1751 if (!IS_ERR(zonelist)) {
1752 /* Good - got mem - substitute new zonelist */
1753 kfree(pol->v.zonelist);
1754 pol->v.zonelist = zonelist;
1756 *mpolmask = *newmask;
1766 * Wrapper for mpol_rebind_policy() that just requires task
1767 * pointer, and updates task mempolicy.
1770 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1772 mpol_rebind_policy(tsk->mempolicy, new);
1776 * Rebind each vma in mm to new nodemask.
1778 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1781 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1783 struct vm_area_struct *vma;
1785 down_write(&mm->mmap_sem);
1786 for (vma = mm->mmap; vma; vma = vma->vm_next)
1787 mpol_rebind_policy(vma->vm_policy, new);
1788 up_write(&mm->mmap_sem);
1792 * Display pages allocated per node and memory policy via /proc.
1795 static const char * const policy_types[] =
1796 { "default", "prefer", "bind", "interleave" };
1799 * Convert a mempolicy into a string.
1800 * Returns the number of characters in buffer (if positive)
1801 * or an error (negative)
1803 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1808 int mode = pol ? pol->policy : MPOL_DEFAULT;
1815 case MPOL_PREFERRED:
1817 node_set(pol->v.preferred_node, nodes);
1821 get_zonemask(pol, &nodes);
1824 case MPOL_INTERLEAVE:
1825 nodes = pol->v.nodes;
1833 l = strlen(policy_types[mode]);
1834 if (buffer + maxlen < p + l + 1)
1837 strcpy(p, policy_types[mode]);
1840 if (!nodes_empty(nodes)) {
1841 if (buffer + maxlen < p + 2)
1844 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1850 unsigned long pages;
1852 unsigned long active;
1853 unsigned long writeback;
1854 unsigned long mapcount_max;
1855 unsigned long dirty;
1856 unsigned long swapcache;
1857 unsigned long node[MAX_NUMNODES];
1860 static void gather_stats(struct page *page, void *private, int pte_dirty)
1862 struct numa_maps *md = private;
1863 int count = page_mapcount(page);
1866 if (pte_dirty || PageDirty(page))
1869 if (PageSwapCache(page))
1872 if (PageActive(page))
1875 if (PageWriteback(page))
1881 if (count > md->mapcount_max)
1882 md->mapcount_max = count;
1884 md->node[page_to_nid(page)]++;
1887 #ifdef CONFIG_HUGETLB_PAGE
1888 static void check_huge_range(struct vm_area_struct *vma,
1889 unsigned long start, unsigned long end,
1890 struct numa_maps *md)
1895 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1896 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1906 page = pte_page(pte);
1910 gather_stats(page, md, pte_dirty(*ptep));
1914 static inline void check_huge_range(struct vm_area_struct *vma,
1915 unsigned long start, unsigned long end,
1916 struct numa_maps *md)
1921 int show_numa_map(struct seq_file *m, void *v)
1923 struct proc_maps_private *priv = m->private;
1924 struct vm_area_struct *vma = v;
1925 struct numa_maps *md;
1926 struct file *file = vma->vm_file;
1927 struct mm_struct *mm = vma->vm_mm;
1928 struct mempolicy *pol;
1935 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1939 pol = get_vma_policy(priv->task, vma, vma->vm_start);
1940 mpol_to_str(buffer, sizeof(buffer), pol);
1942 * unref shared or other task's mempolicy
1944 if (pol != &default_policy && pol != current->mempolicy)
1947 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1950 seq_printf(m, " file=");
1951 seq_path(m, file->f_path.mnt, file->f_path.dentry, "\n\t= ");
1952 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1953 seq_printf(m, " heap");
1954 } else if (vma->vm_start <= mm->start_stack &&
1955 vma->vm_end >= mm->start_stack) {
1956 seq_printf(m, " stack");
1959 if (is_vm_hugetlb_page(vma)) {
1960 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
1961 seq_printf(m, " huge");
1963 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1964 &node_online_map, MPOL_MF_STATS, md);
1971 seq_printf(m," anon=%lu",md->anon);
1974 seq_printf(m," dirty=%lu",md->dirty);
1976 if (md->pages != md->anon && md->pages != md->dirty)
1977 seq_printf(m, " mapped=%lu", md->pages);
1979 if (md->mapcount_max > 1)
1980 seq_printf(m, " mapmax=%lu", md->mapcount_max);
1983 seq_printf(m," swapcache=%lu", md->swapcache);
1985 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
1986 seq_printf(m," active=%lu", md->active);
1989 seq_printf(m," writeback=%lu", md->writeback);
1991 for_each_online_node(n)
1993 seq_printf(m, " N%d=%lu", n, md->node[n]);
1998 if (m->count < m->size)
1999 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;