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 static void mpol_rebind_policy(struct mempolicy *pol,
114 const nodemask_t *newmask);
116 /* Do sanity checking on a policy */
117 static int mpol_check_policy(unsigned short mode, nodemask_t *nodes)
119 int was_empty, is_empty;
125 * "Contextualize" the in-coming nodemast for cpusets:
126 * Remember whether in-coming nodemask was empty, If not,
127 * restrict the nodes to the allowed nodes in the cpuset.
128 * This is guaranteed to be a subset of nodes with memory.
130 cpuset_update_task_memory_state();
131 is_empty = was_empty = nodes_empty(*nodes);
133 nodes_and(*nodes, *nodes, cpuset_current_mems_allowed);
134 is_empty = nodes_empty(*nodes); /* after "contextualization" */
140 * require caller to specify an empty nodemask
141 * before "contextualization"
147 case MPOL_INTERLEAVE:
149 * require at least 1 valid node after "contextualization"
156 * Did caller specify invalid nodes?
157 * Don't silently accept this as "local allocation".
159 if (!was_empty && is_empty)
168 /* Check that the nodemask contains at least one populated zone */
169 static int is_valid_nodemask(nodemask_t *nodemask)
173 /* Check that there is something useful in this mask */
176 for_each_node_mask(nd, *nodemask) {
179 for (k = 0; k <= policy_zone; k++) {
180 z = &NODE_DATA(nd)->node_zones[k];
181 if (z->present_pages > 0)
189 /* Create a new policy */
190 static struct mempolicy *mpol_new(unsigned short mode, nodemask_t *nodes)
192 struct mempolicy *policy;
194 pr_debug("setting mode %d nodes[0] %lx\n",
195 mode, nodes ? nodes_addr(*nodes)[0] : -1);
197 if (mode == MPOL_DEFAULT)
199 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
201 return ERR_PTR(-ENOMEM);
202 atomic_set(&policy->refcnt, 1);
204 case MPOL_INTERLEAVE:
205 policy->v.nodes = *nodes;
206 if (nodes_weight(policy->v.nodes) == 0) {
207 kmem_cache_free(policy_cache, policy);
208 return ERR_PTR(-EINVAL);
212 policy->v.preferred_node = first_node(*nodes);
213 if (policy->v.preferred_node >= MAX_NUMNODES)
214 policy->v.preferred_node = -1;
217 if (!is_valid_nodemask(nodes)) {
218 kmem_cache_free(policy_cache, policy);
219 return ERR_PTR(-EINVAL);
221 policy->v.nodes = *nodes;
226 policy->policy = mode;
227 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
231 static void gather_stats(struct page *, void *, int pte_dirty);
232 static void migrate_page_add(struct page *page, struct list_head *pagelist,
233 unsigned long flags);
235 /* Scan through pages checking if pages follow certain conditions. */
236 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
237 unsigned long addr, unsigned long end,
238 const nodemask_t *nodes, unsigned long flags,
245 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
250 if (!pte_present(*pte))
252 page = vm_normal_page(vma, addr, *pte);
256 * The check for PageReserved here is important to avoid
257 * handling zero pages and other pages that may have been
258 * marked special by the system.
260 * If the PageReserved would not be checked here then f.e.
261 * the location of the zero page could have an influence
262 * on MPOL_MF_STRICT, zero pages would be counted for
263 * the per node stats, and there would be useless attempts
264 * to put zero pages on the migration list.
266 if (PageReserved(page))
268 nid = page_to_nid(page);
269 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
272 if (flags & MPOL_MF_STATS)
273 gather_stats(page, private, pte_dirty(*pte));
274 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
275 migrate_page_add(page, private, flags);
278 } while (pte++, addr += PAGE_SIZE, addr != end);
279 pte_unmap_unlock(orig_pte, ptl);
283 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
284 unsigned long addr, unsigned long end,
285 const nodemask_t *nodes, unsigned long flags,
291 pmd = pmd_offset(pud, addr);
293 next = pmd_addr_end(addr, end);
294 if (pmd_none_or_clear_bad(pmd))
296 if (check_pte_range(vma, pmd, addr, next, nodes,
299 } while (pmd++, addr = next, addr != end);
303 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
304 unsigned long addr, unsigned long end,
305 const nodemask_t *nodes, unsigned long flags,
311 pud = pud_offset(pgd, addr);
313 next = pud_addr_end(addr, end);
314 if (pud_none_or_clear_bad(pud))
316 if (check_pmd_range(vma, pud, addr, next, nodes,
319 } while (pud++, addr = next, addr != end);
323 static inline int check_pgd_range(struct vm_area_struct *vma,
324 unsigned long addr, unsigned long end,
325 const nodemask_t *nodes, unsigned long flags,
331 pgd = pgd_offset(vma->vm_mm, addr);
333 next = pgd_addr_end(addr, end);
334 if (pgd_none_or_clear_bad(pgd))
336 if (check_pud_range(vma, pgd, addr, next, nodes,
339 } while (pgd++, addr = next, addr != end);
344 * Check if all pages in a range are on a set of nodes.
345 * If pagelist != NULL then isolate pages from the LRU and
346 * put them on the pagelist.
348 static struct vm_area_struct *
349 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
350 const nodemask_t *nodes, unsigned long flags, void *private)
353 struct vm_area_struct *first, *vma, *prev;
355 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
357 err = migrate_prep();
362 first = find_vma(mm, start);
364 return ERR_PTR(-EFAULT);
366 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
367 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
368 if (!vma->vm_next && vma->vm_end < end)
369 return ERR_PTR(-EFAULT);
370 if (prev && prev->vm_end < vma->vm_start)
371 return ERR_PTR(-EFAULT);
373 if (!is_vm_hugetlb_page(vma) &&
374 ((flags & MPOL_MF_STRICT) ||
375 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
376 vma_migratable(vma)))) {
377 unsigned long endvma = vma->vm_end;
381 if (vma->vm_start > start)
382 start = vma->vm_start;
383 err = check_pgd_range(vma, start, endvma, nodes,
386 first = ERR_PTR(err);
395 /* Apply policy to a single VMA */
396 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
399 struct mempolicy *old = vma->vm_policy;
401 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
402 vma->vm_start, vma->vm_end, vma->vm_pgoff,
403 vma->vm_ops, vma->vm_file,
404 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
406 if (vma->vm_ops && vma->vm_ops->set_policy)
407 err = vma->vm_ops->set_policy(vma, new);
410 vma->vm_policy = new;
416 /* Step 2: apply policy to a range and do splits. */
417 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
418 unsigned long end, struct mempolicy *new)
420 struct vm_area_struct *next;
424 for (; vma && vma->vm_start < end; vma = next) {
426 if (vma->vm_start < start)
427 err = split_vma(vma->vm_mm, vma, start, 1);
428 if (!err && vma->vm_end > end)
429 err = split_vma(vma->vm_mm, vma, end, 0);
431 err = policy_vma(vma, new);
439 * Update task->flags PF_MEMPOLICY bit: set iff non-default
440 * mempolicy. Allows more rapid checking of this (combined perhaps
441 * with other PF_* flag bits) on memory allocation hot code paths.
443 * If called from outside this file, the task 'p' should -only- be
444 * a newly forked child not yet visible on the task list, because
445 * manipulating the task flags of a visible task is not safe.
447 * The above limitation is why this routine has the funny name
448 * mpol_fix_fork_child_flag().
450 * It is also safe to call this with a task pointer of current,
451 * which the static wrapper mpol_set_task_struct_flag() does,
452 * for use within this file.
455 void mpol_fix_fork_child_flag(struct task_struct *p)
458 p->flags |= PF_MEMPOLICY;
460 p->flags &= ~PF_MEMPOLICY;
463 static void mpol_set_task_struct_flag(void)
465 mpol_fix_fork_child_flag(current);
468 /* Set the process memory policy */
469 static long do_set_mempolicy(unsigned short mode, nodemask_t *nodes)
471 struct mempolicy *new;
473 if (mpol_check_policy(mode, nodes))
475 new = mpol_new(mode, nodes);
478 mpol_free(current->mempolicy);
479 current->mempolicy = new;
480 mpol_set_task_struct_flag();
481 if (new && new->policy == MPOL_INTERLEAVE)
482 current->il_next = first_node(new->v.nodes);
486 /* Fill a zone bitmap for a policy */
487 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
495 case MPOL_INTERLEAVE:
499 /* or use current node instead of memory_map? */
500 if (p->v.preferred_node < 0)
501 *nodes = node_states[N_HIGH_MEMORY];
503 node_set(p->v.preferred_node, *nodes);
510 static int lookup_node(struct mm_struct *mm, unsigned long addr)
515 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
517 err = page_to_nid(p);
523 /* Retrieve NUMA policy */
524 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
525 unsigned long addr, unsigned long flags)
528 struct mm_struct *mm = current->mm;
529 struct vm_area_struct *vma = NULL;
530 struct mempolicy *pol = current->mempolicy;
532 cpuset_update_task_memory_state();
534 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
537 if (flags & MPOL_F_MEMS_ALLOWED) {
538 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
540 *policy = 0; /* just so it's initialized */
541 *nmask = cpuset_current_mems_allowed;
545 if (flags & MPOL_F_ADDR) {
546 down_read(&mm->mmap_sem);
547 vma = find_vma_intersection(mm, addr, addr+1);
549 up_read(&mm->mmap_sem);
552 if (vma->vm_ops && vma->vm_ops->get_policy)
553 pol = vma->vm_ops->get_policy(vma, addr);
555 pol = vma->vm_policy;
560 pol = &default_policy;
562 if (flags & MPOL_F_NODE) {
563 if (flags & MPOL_F_ADDR) {
564 err = lookup_node(mm, addr);
568 } else if (pol == current->mempolicy &&
569 pol->policy == MPOL_INTERLEAVE) {
570 *policy = current->il_next;
576 *policy = pol->policy;
579 up_read(¤t->mm->mmap_sem);
585 get_zonemask(pol, nmask);
589 up_read(¤t->mm->mmap_sem);
593 #ifdef CONFIG_MIGRATION
597 static void migrate_page_add(struct page *page, struct list_head *pagelist,
601 * Avoid migrating a page that is shared with others.
603 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
604 isolate_lru_page(page, pagelist);
607 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
609 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
613 * Migrate pages from one node to a target node.
614 * Returns error or the number of pages not migrated.
616 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
624 node_set(source, nmask);
626 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
627 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
629 if (!list_empty(&pagelist))
630 err = migrate_pages(&pagelist, new_node_page, dest);
636 * Move pages between the two nodesets so as to preserve the physical
637 * layout as much as possible.
639 * Returns the number of page that could not be moved.
641 int do_migrate_pages(struct mm_struct *mm,
642 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
649 down_read(&mm->mmap_sem);
651 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
656 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
657 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
658 * bit in 'tmp', and return that <source, dest> pair for migration.
659 * The pair of nodemasks 'to' and 'from' define the map.
661 * If no pair of bits is found that way, fallback to picking some
662 * pair of 'source' and 'dest' bits that are not the same. If the
663 * 'source' and 'dest' bits are the same, this represents a node
664 * that will be migrating to itself, so no pages need move.
666 * If no bits are left in 'tmp', or if all remaining bits left
667 * in 'tmp' correspond to the same bit in 'to', return false
668 * (nothing left to migrate).
670 * This lets us pick a pair of nodes to migrate between, such that
671 * if possible the dest node is not already occupied by some other
672 * source node, minimizing the risk of overloading the memory on a
673 * node that would happen if we migrated incoming memory to a node
674 * before migrating outgoing memory source that same node.
676 * A single scan of tmp is sufficient. As we go, we remember the
677 * most recent <s, d> pair that moved (s != d). If we find a pair
678 * that not only moved, but what's better, moved to an empty slot
679 * (d is not set in tmp), then we break out then, with that pair.
680 * Otherwise when we finish scannng from_tmp, we at least have the
681 * most recent <s, d> pair that moved. If we get all the way through
682 * the scan of tmp without finding any node that moved, much less
683 * moved to an empty node, then there is nothing left worth migrating.
687 while (!nodes_empty(tmp)) {
692 for_each_node_mask(s, tmp) {
693 d = node_remap(s, *from_nodes, *to_nodes);
697 source = s; /* Node moved. Memorize */
700 /* dest not in remaining from nodes? */
701 if (!node_isset(dest, tmp))
707 node_clear(source, tmp);
708 err = migrate_to_node(mm, source, dest, flags);
715 up_read(&mm->mmap_sem);
723 * Allocate a new page for page migration based on vma policy.
724 * Start assuming that page is mapped by vma pointed to by @private.
725 * Search forward from there, if not. N.B., this assumes that the
726 * list of pages handed to migrate_pages()--which is how we get here--
727 * is in virtual address order.
729 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
731 struct vm_area_struct *vma = (struct vm_area_struct *)private;
732 unsigned long uninitialized_var(address);
735 address = page_address_in_vma(page, vma);
736 if (address != -EFAULT)
742 * if !vma, alloc_page_vma() will use task or system default policy
744 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
748 static void migrate_page_add(struct page *page, struct list_head *pagelist,
753 int do_migrate_pages(struct mm_struct *mm,
754 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
759 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
765 static long do_mbind(unsigned long start, unsigned long len,
766 unsigned short mode, nodemask_t *nmask,
769 struct vm_area_struct *vma;
770 struct mm_struct *mm = current->mm;
771 struct mempolicy *new;
776 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
777 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
779 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
782 if (start & ~PAGE_MASK)
785 if (mode == MPOL_DEFAULT)
786 flags &= ~MPOL_MF_STRICT;
788 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
796 if (mpol_check_policy(mode, nmask))
799 new = mpol_new(mode, nmask);
804 * If we are using the default policy then operation
805 * on discontinuous address spaces is okay after all
808 flags |= MPOL_MF_DISCONTIG_OK;
810 pr_debug("mbind %lx-%lx mode:%d nodes:%lx\n", start, start + len,
811 mode, nmask ? nodes_addr(*nmask)[0] : -1);
813 down_write(&mm->mmap_sem);
814 vma = check_range(mm, start, end, nmask,
815 flags | MPOL_MF_INVERT, &pagelist);
821 err = mbind_range(vma, start, end, new);
823 if (!list_empty(&pagelist))
824 nr_failed = migrate_pages(&pagelist, new_vma_page,
827 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
831 up_write(&mm->mmap_sem);
837 * User space interface with variable sized bitmaps for nodelists.
840 /* Copy a node mask from user space. */
841 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
842 unsigned long maxnode)
845 unsigned long nlongs;
846 unsigned long endmask;
850 if (maxnode == 0 || !nmask)
852 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
855 nlongs = BITS_TO_LONGS(maxnode);
856 if ((maxnode % BITS_PER_LONG) == 0)
859 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
861 /* When the user specified more nodes than supported just check
862 if the non supported part is all zero. */
863 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
864 if (nlongs > PAGE_SIZE/sizeof(long))
866 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
868 if (get_user(t, nmask + k))
870 if (k == nlongs - 1) {
876 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
880 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
882 nodes_addr(*nodes)[nlongs-1] &= endmask;
886 /* Copy a kernel node mask to user space */
887 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
890 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
891 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
894 if (copy > PAGE_SIZE)
896 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
900 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
903 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
905 unsigned long __user *nmask, unsigned long maxnode,
911 if (mode >= MPOL_MAX)
913 err = get_nodes(&nodes, nmask, maxnode);
916 return do_mbind(start, len, mode, &nodes, flags);
919 /* Set the process memory policy */
920 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
921 unsigned long maxnode)
926 if (mode < 0 || mode >= MPOL_MAX)
928 err = get_nodes(&nodes, nmask, maxnode);
931 return do_set_mempolicy(mode, &nodes);
934 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
935 const unsigned long __user *old_nodes,
936 const unsigned long __user *new_nodes)
938 struct mm_struct *mm;
939 struct task_struct *task;
942 nodemask_t task_nodes;
945 err = get_nodes(&old, old_nodes, maxnode);
949 err = get_nodes(&new, new_nodes, maxnode);
953 /* Find the mm_struct */
954 read_lock(&tasklist_lock);
955 task = pid ? find_task_by_vpid(pid) : current;
957 read_unlock(&tasklist_lock);
960 mm = get_task_mm(task);
961 read_unlock(&tasklist_lock);
967 * Check if this process has the right to modify the specified
968 * process. The right exists if the process has administrative
969 * capabilities, superuser privileges or the same
970 * userid as the target process.
972 if ((current->euid != task->suid) && (current->euid != task->uid) &&
973 (current->uid != task->suid) && (current->uid != task->uid) &&
974 !capable(CAP_SYS_NICE)) {
979 task_nodes = cpuset_mems_allowed(task);
980 /* Is the user allowed to access the target nodes? */
981 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
986 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
991 err = security_task_movememory(task);
995 err = do_migrate_pages(mm, &old, &new,
996 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1003 /* Retrieve NUMA policy */
1004 asmlinkage long sys_get_mempolicy(int __user *policy,
1005 unsigned long __user *nmask,
1006 unsigned long maxnode,
1007 unsigned long addr, unsigned long flags)
1010 int uninitialized_var(pval);
1013 if (nmask != NULL && maxnode < MAX_NUMNODES)
1016 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1021 if (policy && put_user(pval, policy))
1025 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1030 #ifdef CONFIG_COMPAT
1032 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1033 compat_ulong_t __user *nmask,
1034 compat_ulong_t maxnode,
1035 compat_ulong_t addr, compat_ulong_t flags)
1038 unsigned long __user *nm = NULL;
1039 unsigned long nr_bits, alloc_size;
1040 DECLARE_BITMAP(bm, MAX_NUMNODES);
1042 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1043 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1046 nm = compat_alloc_user_space(alloc_size);
1048 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1050 if (!err && nmask) {
1051 err = copy_from_user(bm, nm, alloc_size);
1052 /* ensure entire bitmap is zeroed */
1053 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1054 err |= compat_put_bitmap(nmask, bm, nr_bits);
1060 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1061 compat_ulong_t maxnode)
1064 unsigned long __user *nm = NULL;
1065 unsigned long nr_bits, alloc_size;
1066 DECLARE_BITMAP(bm, MAX_NUMNODES);
1068 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1069 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1072 err = compat_get_bitmap(bm, nmask, nr_bits);
1073 nm = compat_alloc_user_space(alloc_size);
1074 err |= copy_to_user(nm, bm, alloc_size);
1080 return sys_set_mempolicy(mode, nm, nr_bits+1);
1083 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1084 compat_ulong_t mode, compat_ulong_t __user *nmask,
1085 compat_ulong_t maxnode, compat_ulong_t flags)
1088 unsigned long __user *nm = NULL;
1089 unsigned long nr_bits, alloc_size;
1092 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1093 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1096 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1097 nm = compat_alloc_user_space(alloc_size);
1098 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1104 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1110 * get_vma_policy(@task, @vma, @addr)
1111 * @task - task for fallback if vma policy == default
1112 * @vma - virtual memory area whose policy is sought
1113 * @addr - address in @vma for shared policy lookup
1115 * Returns effective policy for a VMA at specified address.
1116 * Falls back to @task or system default policy, as necessary.
1117 * Returned policy has extra reference count if shared, vma,
1118 * or some other task's policy [show_numa_maps() can pass
1119 * @task != current]. It is the caller's responsibility to
1120 * free the reference in these cases.
1122 static struct mempolicy * get_vma_policy(struct task_struct *task,
1123 struct vm_area_struct *vma, unsigned long addr)
1125 struct mempolicy *pol = task->mempolicy;
1129 if (vma->vm_ops && vma->vm_ops->get_policy) {
1130 pol = vma->vm_ops->get_policy(vma, addr);
1131 shared_pol = 1; /* if pol non-NULL, add ref below */
1132 } else if (vma->vm_policy &&
1133 vma->vm_policy->policy != MPOL_DEFAULT)
1134 pol = vma->vm_policy;
1137 pol = &default_policy;
1138 else if (!shared_pol && pol != current->mempolicy)
1139 mpol_get(pol); /* vma or other task's policy */
1143 /* Return a nodemask representing a mempolicy */
1144 static nodemask_t *nodemask_policy(gfp_t gfp, struct mempolicy *policy)
1146 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1147 if (unlikely(policy->policy == MPOL_BIND) &&
1148 gfp_zone(gfp) >= policy_zone &&
1149 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1150 return &policy->v.nodes;
1155 /* Return a zonelist representing a mempolicy */
1156 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1160 switch (policy->policy) {
1161 case MPOL_PREFERRED:
1162 nd = policy->v.preferred_node;
1164 nd = numa_node_id();
1168 * Normally, MPOL_BIND allocations node-local are node-local
1169 * within the allowed nodemask. However, if __GFP_THISNODE is
1170 * set and the current node is part of the mask, we use the
1171 * the zonelist for the first node in the mask instead.
1173 nd = numa_node_id();
1174 if (unlikely(gfp & __GFP_THISNODE) &&
1175 unlikely(!node_isset(nd, policy->v.nodes)))
1176 nd = first_node(policy->v.nodes);
1178 case MPOL_INTERLEAVE: /* should not happen */
1180 nd = numa_node_id();
1186 return node_zonelist(nd, gfp);
1189 /* Do dynamic interleaving for a process */
1190 static unsigned interleave_nodes(struct mempolicy *policy)
1193 struct task_struct *me = current;
1196 next = next_node(nid, policy->v.nodes);
1197 if (next >= MAX_NUMNODES)
1198 next = first_node(policy->v.nodes);
1204 * Depending on the memory policy provide a node from which to allocate the
1207 unsigned slab_node(struct mempolicy *policy)
1209 unsigned short pol = policy ? policy->policy : MPOL_DEFAULT;
1212 case MPOL_INTERLEAVE:
1213 return interleave_nodes(policy);
1217 * Follow bind policy behavior and start allocation at the
1220 struct zonelist *zonelist;
1222 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1223 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1224 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1230 case MPOL_PREFERRED:
1231 if (policy->v.preferred_node >= 0)
1232 return policy->v.preferred_node;
1236 return numa_node_id();
1240 /* Do static interleaving for a VMA with known offset. */
1241 static unsigned offset_il_node(struct mempolicy *pol,
1242 struct vm_area_struct *vma, unsigned long off)
1244 unsigned nnodes = nodes_weight(pol->v.nodes);
1245 unsigned target = (unsigned)off % nnodes;
1251 nid = next_node(nid, pol->v.nodes);
1253 } while (c <= target);
1257 /* Determine a node number for interleave */
1258 static inline unsigned interleave_nid(struct mempolicy *pol,
1259 struct vm_area_struct *vma, unsigned long addr, int shift)
1265 * for small pages, there is no difference between
1266 * shift and PAGE_SHIFT, so the bit-shift is safe.
1267 * for huge pages, since vm_pgoff is in units of small
1268 * pages, we need to shift off the always 0 bits to get
1271 BUG_ON(shift < PAGE_SHIFT);
1272 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1273 off += (addr - vma->vm_start) >> shift;
1274 return offset_il_node(pol, vma, off);
1276 return interleave_nodes(pol);
1279 #ifdef CONFIG_HUGETLBFS
1281 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1282 * @vma = virtual memory area whose policy is sought
1283 * @addr = address in @vma for shared policy lookup and interleave policy
1284 * @gfp_flags = for requested zone
1285 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1286 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1288 * Returns a zonelist suitable for a huge page allocation.
1289 * If the effective policy is 'BIND, returns pointer to local node's zonelist,
1290 * and a pointer to the mempolicy's @nodemask for filtering the zonelist.
1291 * If it is also a policy for which get_vma_policy() returns an extra
1292 * reference, we must hold that reference until after the allocation.
1293 * In that case, return policy via @mpol so hugetlb allocation can drop
1294 * the reference. For non-'BIND referenced policies, we can/do drop the
1295 * reference here, so the caller doesn't need to know about the special case
1296 * for default and current task policy.
1298 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1299 gfp_t gfp_flags, struct mempolicy **mpol,
1300 nodemask_t **nodemask)
1302 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1303 struct zonelist *zl;
1305 *mpol = NULL; /* probably no unref needed */
1306 *nodemask = NULL; /* assume !MPOL_BIND */
1307 if (pol->policy == MPOL_BIND) {
1308 *nodemask = &pol->v.nodes;
1309 } else if (pol->policy == MPOL_INTERLEAVE) {
1312 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1313 if (unlikely(pol != &default_policy &&
1314 pol != current->mempolicy))
1315 __mpol_free(pol); /* finished with pol */
1316 return node_zonelist(nid, gfp_flags);
1319 zl = zonelist_policy(GFP_HIGHUSER, pol);
1320 if (unlikely(pol != &default_policy && pol != current->mempolicy)) {
1321 if (pol->policy != MPOL_BIND)
1322 __mpol_free(pol); /* finished with pol */
1324 *mpol = pol; /* unref needed after allocation */
1330 /* Allocate a page in interleaved policy.
1331 Own path because it needs to do special accounting. */
1332 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1335 struct zonelist *zl;
1338 zl = node_zonelist(nid, gfp);
1339 page = __alloc_pages(gfp, order, zl);
1340 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1341 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1346 * alloc_page_vma - Allocate a page for a VMA.
1349 * %GFP_USER user allocation.
1350 * %GFP_KERNEL kernel allocations,
1351 * %GFP_HIGHMEM highmem/user allocations,
1352 * %GFP_FS allocation should not call back into a file system.
1353 * %GFP_ATOMIC don't sleep.
1355 * @vma: Pointer to VMA or NULL if not available.
1356 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1358 * This function allocates a page from the kernel page pool and applies
1359 * a NUMA policy associated with the VMA or the current process.
1360 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1361 * mm_struct of the VMA to prevent it from going away. Should be used for
1362 * all allocations for pages that will be mapped into
1363 * user space. Returns NULL when no page can be allocated.
1365 * Should be called with the mm_sem of the vma hold.
1368 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1370 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1371 struct zonelist *zl;
1373 cpuset_update_task_memory_state();
1375 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1378 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1379 if (unlikely(pol != &default_policy &&
1380 pol != current->mempolicy))
1381 __mpol_free(pol); /* finished with pol */
1382 return alloc_page_interleave(gfp, 0, nid);
1384 zl = zonelist_policy(gfp, pol);
1385 if (pol != &default_policy && pol != current->mempolicy) {
1387 * slow path: ref counted policy -- shared or vma
1389 struct page *page = __alloc_pages_nodemask(gfp, 0,
1390 zl, nodemask_policy(gfp, pol));
1395 * fast path: default or task policy
1397 return __alloc_pages_nodemask(gfp, 0, zl, nodemask_policy(gfp, pol));
1401 * alloc_pages_current - Allocate pages.
1404 * %GFP_USER user allocation,
1405 * %GFP_KERNEL kernel allocation,
1406 * %GFP_HIGHMEM highmem allocation,
1407 * %GFP_FS don't call back into a file system.
1408 * %GFP_ATOMIC don't sleep.
1409 * @order: Power of two of allocation size in pages. 0 is a single page.
1411 * Allocate a page from the kernel page pool. When not in
1412 * interrupt context and apply the current process NUMA policy.
1413 * Returns NULL when no page can be allocated.
1415 * Don't call cpuset_update_task_memory_state() unless
1416 * 1) it's ok to take cpuset_sem (can WAIT), and
1417 * 2) allocating for current task (not interrupt).
1419 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1421 struct mempolicy *pol = current->mempolicy;
1423 if ((gfp & __GFP_WAIT) && !in_interrupt())
1424 cpuset_update_task_memory_state();
1425 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1426 pol = &default_policy;
1427 if (pol->policy == MPOL_INTERLEAVE)
1428 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1429 return __alloc_pages_nodemask(gfp, order,
1430 zonelist_policy(gfp, pol), nodemask_policy(gfp, pol));
1432 EXPORT_SYMBOL(alloc_pages_current);
1435 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1436 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1437 * with the mems_allowed returned by cpuset_mems_allowed(). This
1438 * keeps mempolicies cpuset relative after its cpuset moves. See
1439 * further kernel/cpuset.c update_nodemask().
1442 /* Slow path of a mempolicy copy */
1443 struct mempolicy *__mpol_copy(struct mempolicy *old)
1445 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1448 return ERR_PTR(-ENOMEM);
1449 if (current_cpuset_is_being_rebound()) {
1450 nodemask_t mems = cpuset_mems_allowed(current);
1451 mpol_rebind_policy(old, &mems);
1454 atomic_set(&new->refcnt, 1);
1458 /* Slow path of a mempolicy comparison */
1459 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1463 if (a->policy != b->policy)
1465 switch (a->policy) {
1470 case MPOL_INTERLEAVE:
1471 return nodes_equal(a->v.nodes, b->v.nodes);
1472 case MPOL_PREFERRED:
1473 return a->v.preferred_node == b->v.preferred_node;
1480 /* Slow path of a mpol destructor. */
1481 void __mpol_free(struct mempolicy *p)
1483 if (!atomic_dec_and_test(&p->refcnt))
1485 p->policy = MPOL_DEFAULT;
1486 kmem_cache_free(policy_cache, p);
1490 * Shared memory backing store policy support.
1492 * Remember policies even when nobody has shared memory mapped.
1493 * The policies are kept in Red-Black tree linked from the inode.
1494 * They are protected by the sp->lock spinlock, which should be held
1495 * for any accesses to the tree.
1498 /* lookup first element intersecting start-end */
1499 /* Caller holds sp->lock */
1500 static struct sp_node *
1501 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1503 struct rb_node *n = sp->root.rb_node;
1506 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1508 if (start >= p->end)
1510 else if (end <= p->start)
1518 struct sp_node *w = NULL;
1519 struct rb_node *prev = rb_prev(n);
1522 w = rb_entry(prev, struct sp_node, nd);
1523 if (w->end <= start)
1527 return rb_entry(n, struct sp_node, nd);
1530 /* Insert a new shared policy into the list. */
1531 /* Caller holds sp->lock */
1532 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1534 struct rb_node **p = &sp->root.rb_node;
1535 struct rb_node *parent = NULL;
1540 nd = rb_entry(parent, struct sp_node, nd);
1541 if (new->start < nd->start)
1543 else if (new->end > nd->end)
1544 p = &(*p)->rb_right;
1548 rb_link_node(&new->nd, parent, p);
1549 rb_insert_color(&new->nd, &sp->root);
1550 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1551 new->policy ? new->policy->policy : 0);
1554 /* Find shared policy intersecting idx */
1556 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1558 struct mempolicy *pol = NULL;
1561 if (!sp->root.rb_node)
1563 spin_lock(&sp->lock);
1564 sn = sp_lookup(sp, idx, idx+1);
1566 mpol_get(sn->policy);
1569 spin_unlock(&sp->lock);
1573 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1575 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1576 rb_erase(&n->nd, &sp->root);
1577 mpol_free(n->policy);
1578 kmem_cache_free(sn_cache, n);
1581 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1582 struct mempolicy *pol)
1584 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1595 /* Replace a policy range. */
1596 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1597 unsigned long end, struct sp_node *new)
1599 struct sp_node *n, *new2 = NULL;
1602 spin_lock(&sp->lock);
1603 n = sp_lookup(sp, start, end);
1604 /* Take care of old policies in the same range. */
1605 while (n && n->start < end) {
1606 struct rb_node *next = rb_next(&n->nd);
1607 if (n->start >= start) {
1613 /* Old policy spanning whole new range. */
1616 spin_unlock(&sp->lock);
1617 new2 = sp_alloc(end, n->end, n->policy);
1623 sp_insert(sp, new2);
1631 n = rb_entry(next, struct sp_node, nd);
1635 spin_unlock(&sp->lock);
1637 mpol_free(new2->policy);
1638 kmem_cache_free(sn_cache, new2);
1643 void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
1644 nodemask_t *policy_nodes)
1646 info->root = RB_ROOT;
1647 spin_lock_init(&info->lock);
1649 if (policy != MPOL_DEFAULT) {
1650 struct mempolicy *newpol;
1652 /* Falls back to MPOL_DEFAULT on any error */
1653 newpol = mpol_new(policy, policy_nodes);
1654 if (!IS_ERR(newpol)) {
1655 /* Create pseudo-vma that contains just the policy */
1656 struct vm_area_struct pvma;
1658 memset(&pvma, 0, sizeof(struct vm_area_struct));
1659 /* Policy covers entire file */
1660 pvma.vm_end = TASK_SIZE;
1661 mpol_set_shared_policy(info, &pvma, newpol);
1667 int mpol_set_shared_policy(struct shared_policy *info,
1668 struct vm_area_struct *vma, struct mempolicy *npol)
1671 struct sp_node *new = NULL;
1672 unsigned long sz = vma_pages(vma);
1674 pr_debug("set_shared_policy %lx sz %lu %d %lx\n",
1676 sz, npol? npol->policy : -1,
1677 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1680 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1684 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1686 kmem_cache_free(sn_cache, new);
1690 /* Free a backing policy store on inode delete. */
1691 void mpol_free_shared_policy(struct shared_policy *p)
1694 struct rb_node *next;
1696 if (!p->root.rb_node)
1698 spin_lock(&p->lock);
1699 next = rb_first(&p->root);
1701 n = rb_entry(next, struct sp_node, nd);
1702 next = rb_next(&n->nd);
1703 rb_erase(&n->nd, &p->root);
1704 mpol_free(n->policy);
1705 kmem_cache_free(sn_cache, n);
1707 spin_unlock(&p->lock);
1710 /* assumes fs == KERNEL_DS */
1711 void __init numa_policy_init(void)
1713 nodemask_t interleave_nodes;
1714 unsigned long largest = 0;
1715 int nid, prefer = 0;
1717 policy_cache = kmem_cache_create("numa_policy",
1718 sizeof(struct mempolicy),
1719 0, SLAB_PANIC, NULL);
1721 sn_cache = kmem_cache_create("shared_policy_node",
1722 sizeof(struct sp_node),
1723 0, SLAB_PANIC, NULL);
1726 * Set interleaving policy for system init. Interleaving is only
1727 * enabled across suitably sized nodes (default is >= 16MB), or
1728 * fall back to the largest node if they're all smaller.
1730 nodes_clear(interleave_nodes);
1731 for_each_node_state(nid, N_HIGH_MEMORY) {
1732 unsigned long total_pages = node_present_pages(nid);
1734 /* Preserve the largest node */
1735 if (largest < total_pages) {
1736 largest = total_pages;
1740 /* Interleave this node? */
1741 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1742 node_set(nid, interleave_nodes);
1745 /* All too small, use the largest */
1746 if (unlikely(nodes_empty(interleave_nodes)))
1747 node_set(prefer, interleave_nodes);
1749 if (do_set_mempolicy(MPOL_INTERLEAVE, &interleave_nodes))
1750 printk("numa_policy_init: interleaving failed\n");
1753 /* Reset policy of current process to default */
1754 void numa_default_policy(void)
1756 do_set_mempolicy(MPOL_DEFAULT, NULL);
1759 /* Migrate a policy to a different set of nodes */
1760 static void mpol_rebind_policy(struct mempolicy *pol,
1761 const nodemask_t *newmask)
1763 nodemask_t *mpolmask;
1768 mpolmask = &pol->cpuset_mems_allowed;
1769 if (nodes_equal(*mpolmask, *newmask))
1772 switch (pol->policy) {
1777 case MPOL_INTERLEAVE:
1778 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1780 *mpolmask = *newmask;
1781 current->il_next = node_remap(current->il_next,
1782 *mpolmask, *newmask);
1784 case MPOL_PREFERRED:
1785 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1786 *mpolmask, *newmask);
1787 *mpolmask = *newmask;
1796 * Wrapper for mpol_rebind_policy() that just requires task
1797 * pointer, and updates task mempolicy.
1800 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1802 mpol_rebind_policy(tsk->mempolicy, new);
1806 * Rebind each vma in mm to new nodemask.
1808 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1811 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1813 struct vm_area_struct *vma;
1815 down_write(&mm->mmap_sem);
1816 for (vma = mm->mmap; vma; vma = vma->vm_next)
1817 mpol_rebind_policy(vma->vm_policy, new);
1818 up_write(&mm->mmap_sem);
1822 * Display pages allocated per node and memory policy via /proc.
1825 static const char * const policy_types[] =
1826 { "default", "prefer", "bind", "interleave" };
1829 * Convert a mempolicy into a string.
1830 * Returns the number of characters in buffer (if positive)
1831 * or an error (negative)
1833 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1838 unsigned short mode = pol ? pol->policy : MPOL_DEFAULT;
1845 case MPOL_PREFERRED:
1847 node_set(pol->v.preferred_node, nodes);
1852 case MPOL_INTERLEAVE:
1853 nodes = pol->v.nodes;
1861 l = strlen(policy_types[mode]);
1862 if (buffer + maxlen < p + l + 1)
1865 strcpy(p, policy_types[mode]);
1868 if (!nodes_empty(nodes)) {
1869 if (buffer + maxlen < p + 2)
1872 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1878 unsigned long pages;
1880 unsigned long active;
1881 unsigned long writeback;
1882 unsigned long mapcount_max;
1883 unsigned long dirty;
1884 unsigned long swapcache;
1885 unsigned long node[MAX_NUMNODES];
1888 static void gather_stats(struct page *page, void *private, int pte_dirty)
1890 struct numa_maps *md = private;
1891 int count = page_mapcount(page);
1894 if (pte_dirty || PageDirty(page))
1897 if (PageSwapCache(page))
1900 if (PageActive(page))
1903 if (PageWriteback(page))
1909 if (count > md->mapcount_max)
1910 md->mapcount_max = count;
1912 md->node[page_to_nid(page)]++;
1915 #ifdef CONFIG_HUGETLB_PAGE
1916 static void check_huge_range(struct vm_area_struct *vma,
1917 unsigned long start, unsigned long end,
1918 struct numa_maps *md)
1923 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1924 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1934 page = pte_page(pte);
1938 gather_stats(page, md, pte_dirty(*ptep));
1942 static inline void check_huge_range(struct vm_area_struct *vma,
1943 unsigned long start, unsigned long end,
1944 struct numa_maps *md)
1949 int show_numa_map(struct seq_file *m, void *v)
1951 struct proc_maps_private *priv = m->private;
1952 struct vm_area_struct *vma = v;
1953 struct numa_maps *md;
1954 struct file *file = vma->vm_file;
1955 struct mm_struct *mm = vma->vm_mm;
1956 struct mempolicy *pol;
1963 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1967 pol = get_vma_policy(priv->task, vma, vma->vm_start);
1968 mpol_to_str(buffer, sizeof(buffer), pol);
1970 * unref shared or other task's mempolicy
1972 if (pol != &default_policy && pol != current->mempolicy)
1975 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1978 seq_printf(m, " file=");
1979 seq_path(m, &file->f_path, "\n\t= ");
1980 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1981 seq_printf(m, " heap");
1982 } else if (vma->vm_start <= mm->start_stack &&
1983 vma->vm_end >= mm->start_stack) {
1984 seq_printf(m, " stack");
1987 if (is_vm_hugetlb_page(vma)) {
1988 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
1989 seq_printf(m, " huge");
1991 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1992 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
1999 seq_printf(m," anon=%lu",md->anon);
2002 seq_printf(m," dirty=%lu",md->dirty);
2004 if (md->pages != md->anon && md->pages != md->dirty)
2005 seq_printf(m, " mapped=%lu", md->pages);
2007 if (md->mapcount_max > 1)
2008 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2011 seq_printf(m," swapcache=%lu", md->swapcache);
2013 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2014 seq_printf(m," active=%lu", md->active);
2017 seq_printf(m," writeback=%lu", md->writeback);
2019 for_each_node_state(n, N_HIGH_MEMORY)
2021 seq_printf(m, " N%d=%lu", n, md->node[n]);
2026 if (m->count < m->size)
2027 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;