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 /* Check that the nodemask contains at least one populated zone */
117 static int is_valid_nodemask(nodemask_t *nodemask)
121 /* Check that there is something useful in this mask */
124 for_each_node_mask(nd, *nodemask) {
127 for (k = 0; k <= policy_zone; k++) {
128 z = &NODE_DATA(nd)->node_zones[k];
129 if (z->present_pages > 0)
137 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
139 return pol->flags & MPOL_F_STATIC_NODES;
142 /* Create a new policy */
143 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
146 struct mempolicy *policy;
147 nodemask_t cpuset_context_nmask;
149 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
150 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
152 if (mode == MPOL_DEFAULT)
153 return (nodes && nodes_weight(*nodes)) ? ERR_PTR(-EINVAL) :
155 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
157 return ERR_PTR(-ENOMEM);
158 atomic_set(&policy->refcnt, 1);
159 cpuset_update_task_memory_state();
160 nodes_and(cpuset_context_nmask, *nodes, cpuset_current_mems_allowed);
162 case MPOL_INTERLEAVE:
163 if (nodes_empty(*nodes) || nodes_empty(cpuset_context_nmask))
165 policy->v.nodes = cpuset_context_nmask;
168 policy->v.preferred_node = first_node(cpuset_context_nmask);
169 if (policy->v.preferred_node >= MAX_NUMNODES)
173 if (!is_valid_nodemask(&cpuset_context_nmask))
175 policy->v.nodes = cpuset_context_nmask;
180 policy->policy = mode;
181 policy->flags = flags;
182 if (mpol_store_user_nodemask(policy))
183 policy->w.user_nodemask = *nodes;
185 policy->w.cpuset_mems_allowed = cpuset_mems_allowed(current);
189 kmem_cache_free(policy_cache, policy);
190 return ERR_PTR(-EINVAL);
193 static void gather_stats(struct page *, void *, int pte_dirty);
194 static void migrate_page_add(struct page *page, struct list_head *pagelist,
195 unsigned long flags);
197 /* Scan through pages checking if pages follow certain conditions. */
198 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
199 unsigned long addr, unsigned long end,
200 const nodemask_t *nodes, unsigned long flags,
207 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
212 if (!pte_present(*pte))
214 page = vm_normal_page(vma, addr, *pte);
218 * The check for PageReserved here is important to avoid
219 * handling zero pages and other pages that may have been
220 * marked special by the system.
222 * If the PageReserved would not be checked here then f.e.
223 * the location of the zero page could have an influence
224 * on MPOL_MF_STRICT, zero pages would be counted for
225 * the per node stats, and there would be useless attempts
226 * to put zero pages on the migration list.
228 if (PageReserved(page))
230 nid = page_to_nid(page);
231 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
234 if (flags & MPOL_MF_STATS)
235 gather_stats(page, private, pte_dirty(*pte));
236 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
237 migrate_page_add(page, private, flags);
240 } while (pte++, addr += PAGE_SIZE, addr != end);
241 pte_unmap_unlock(orig_pte, ptl);
245 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
246 unsigned long addr, unsigned long end,
247 const nodemask_t *nodes, unsigned long flags,
253 pmd = pmd_offset(pud, addr);
255 next = pmd_addr_end(addr, end);
256 if (pmd_none_or_clear_bad(pmd))
258 if (check_pte_range(vma, pmd, addr, next, nodes,
261 } while (pmd++, addr = next, addr != end);
265 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
266 unsigned long addr, unsigned long end,
267 const nodemask_t *nodes, unsigned long flags,
273 pud = pud_offset(pgd, addr);
275 next = pud_addr_end(addr, end);
276 if (pud_none_or_clear_bad(pud))
278 if (check_pmd_range(vma, pud, addr, next, nodes,
281 } while (pud++, addr = next, addr != end);
285 static inline int check_pgd_range(struct vm_area_struct *vma,
286 unsigned long addr, unsigned long end,
287 const nodemask_t *nodes, unsigned long flags,
293 pgd = pgd_offset(vma->vm_mm, addr);
295 next = pgd_addr_end(addr, end);
296 if (pgd_none_or_clear_bad(pgd))
298 if (check_pud_range(vma, pgd, addr, next, nodes,
301 } while (pgd++, addr = next, addr != end);
306 * Check if all pages in a range are on a set of nodes.
307 * If pagelist != NULL then isolate pages from the LRU and
308 * put them on the pagelist.
310 static struct vm_area_struct *
311 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
312 const nodemask_t *nodes, unsigned long flags, void *private)
315 struct vm_area_struct *first, *vma, *prev;
317 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
319 err = migrate_prep();
324 first = find_vma(mm, start);
326 return ERR_PTR(-EFAULT);
328 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
329 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
330 if (!vma->vm_next && vma->vm_end < end)
331 return ERR_PTR(-EFAULT);
332 if (prev && prev->vm_end < vma->vm_start)
333 return ERR_PTR(-EFAULT);
335 if (!is_vm_hugetlb_page(vma) &&
336 ((flags & MPOL_MF_STRICT) ||
337 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
338 vma_migratable(vma)))) {
339 unsigned long endvma = vma->vm_end;
343 if (vma->vm_start > start)
344 start = vma->vm_start;
345 err = check_pgd_range(vma, start, endvma, nodes,
348 first = ERR_PTR(err);
357 /* Apply policy to a single VMA */
358 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
361 struct mempolicy *old = vma->vm_policy;
363 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
364 vma->vm_start, vma->vm_end, vma->vm_pgoff,
365 vma->vm_ops, vma->vm_file,
366 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
368 if (vma->vm_ops && vma->vm_ops->set_policy)
369 err = vma->vm_ops->set_policy(vma, new);
372 vma->vm_policy = new;
378 /* Step 2: apply policy to a range and do splits. */
379 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
380 unsigned long end, struct mempolicy *new)
382 struct vm_area_struct *next;
386 for (; vma && vma->vm_start < end; vma = next) {
388 if (vma->vm_start < start)
389 err = split_vma(vma->vm_mm, vma, start, 1);
390 if (!err && vma->vm_end > end)
391 err = split_vma(vma->vm_mm, vma, end, 0);
393 err = policy_vma(vma, new);
401 * Update task->flags PF_MEMPOLICY bit: set iff non-default
402 * mempolicy. Allows more rapid checking of this (combined perhaps
403 * with other PF_* flag bits) on memory allocation hot code paths.
405 * If called from outside this file, the task 'p' should -only- be
406 * a newly forked child not yet visible on the task list, because
407 * manipulating the task flags of a visible task is not safe.
409 * The above limitation is why this routine has the funny name
410 * mpol_fix_fork_child_flag().
412 * It is also safe to call this with a task pointer of current,
413 * which the static wrapper mpol_set_task_struct_flag() does,
414 * for use within this file.
417 void mpol_fix_fork_child_flag(struct task_struct *p)
420 p->flags |= PF_MEMPOLICY;
422 p->flags &= ~PF_MEMPOLICY;
425 static void mpol_set_task_struct_flag(void)
427 mpol_fix_fork_child_flag(current);
430 /* Set the process memory policy */
431 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
434 struct mempolicy *new;
436 new = mpol_new(mode, flags, nodes);
439 mpol_free(current->mempolicy);
440 current->mempolicy = new;
441 mpol_set_task_struct_flag();
442 if (new && new->policy == MPOL_INTERLEAVE &&
443 nodes_weight(new->v.nodes))
444 current->il_next = first_node(new->v.nodes);
448 /* Fill a zone bitmap for a policy */
449 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
457 case MPOL_INTERLEAVE:
461 /* or use current node instead of memory_map? */
462 if (p->v.preferred_node < 0)
463 *nodes = node_states[N_HIGH_MEMORY];
465 node_set(p->v.preferred_node, *nodes);
472 static int lookup_node(struct mm_struct *mm, unsigned long addr)
477 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
479 err = page_to_nid(p);
485 /* Retrieve NUMA policy */
486 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
487 unsigned long addr, unsigned long flags)
490 struct mm_struct *mm = current->mm;
491 struct vm_area_struct *vma = NULL;
492 struct mempolicy *pol = current->mempolicy;
494 cpuset_update_task_memory_state();
496 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
499 if (flags & MPOL_F_MEMS_ALLOWED) {
500 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
502 *policy = 0; /* just so it's initialized */
503 *nmask = cpuset_current_mems_allowed;
507 if (flags & MPOL_F_ADDR) {
508 down_read(&mm->mmap_sem);
509 vma = find_vma_intersection(mm, addr, addr+1);
511 up_read(&mm->mmap_sem);
514 if (vma->vm_ops && vma->vm_ops->get_policy)
515 pol = vma->vm_ops->get_policy(vma, addr);
517 pol = vma->vm_policy;
522 pol = &default_policy;
524 if (flags & MPOL_F_NODE) {
525 if (flags & MPOL_F_ADDR) {
526 err = lookup_node(mm, addr);
530 } else if (pol == current->mempolicy &&
531 pol->policy == MPOL_INTERLEAVE) {
532 *policy = current->il_next;
538 *policy = pol->policy | pol->flags;
541 up_read(¤t->mm->mmap_sem);
547 get_zonemask(pol, nmask);
551 up_read(¤t->mm->mmap_sem);
555 #ifdef CONFIG_MIGRATION
559 static void migrate_page_add(struct page *page, struct list_head *pagelist,
563 * Avoid migrating a page that is shared with others.
565 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
566 isolate_lru_page(page, pagelist);
569 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
571 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
575 * Migrate pages from one node to a target node.
576 * Returns error or the number of pages not migrated.
578 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
586 node_set(source, nmask);
588 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
589 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
591 if (!list_empty(&pagelist))
592 err = migrate_pages(&pagelist, new_node_page, dest);
598 * Move pages between the two nodesets so as to preserve the physical
599 * layout as much as possible.
601 * Returns the number of page that could not be moved.
603 int do_migrate_pages(struct mm_struct *mm,
604 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
611 down_read(&mm->mmap_sem);
613 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
618 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
619 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
620 * bit in 'tmp', and return that <source, dest> pair for migration.
621 * The pair of nodemasks 'to' and 'from' define the map.
623 * If no pair of bits is found that way, fallback to picking some
624 * pair of 'source' and 'dest' bits that are not the same. If the
625 * 'source' and 'dest' bits are the same, this represents a node
626 * that will be migrating to itself, so no pages need move.
628 * If no bits are left in 'tmp', or if all remaining bits left
629 * in 'tmp' correspond to the same bit in 'to', return false
630 * (nothing left to migrate).
632 * This lets us pick a pair of nodes to migrate between, such that
633 * if possible the dest node is not already occupied by some other
634 * source node, minimizing the risk of overloading the memory on a
635 * node that would happen if we migrated incoming memory to a node
636 * before migrating outgoing memory source that same node.
638 * A single scan of tmp is sufficient. As we go, we remember the
639 * most recent <s, d> pair that moved (s != d). If we find a pair
640 * that not only moved, but what's better, moved to an empty slot
641 * (d is not set in tmp), then we break out then, with that pair.
642 * Otherwise when we finish scannng from_tmp, we at least have the
643 * most recent <s, d> pair that moved. If we get all the way through
644 * the scan of tmp without finding any node that moved, much less
645 * moved to an empty node, then there is nothing left worth migrating.
649 while (!nodes_empty(tmp)) {
654 for_each_node_mask(s, tmp) {
655 d = node_remap(s, *from_nodes, *to_nodes);
659 source = s; /* Node moved. Memorize */
662 /* dest not in remaining from nodes? */
663 if (!node_isset(dest, tmp))
669 node_clear(source, tmp);
670 err = migrate_to_node(mm, source, dest, flags);
677 up_read(&mm->mmap_sem);
685 * Allocate a new page for page migration based on vma policy.
686 * Start assuming that page is mapped by vma pointed to by @private.
687 * Search forward from there, if not. N.B., this assumes that the
688 * list of pages handed to migrate_pages()--which is how we get here--
689 * is in virtual address order.
691 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
693 struct vm_area_struct *vma = (struct vm_area_struct *)private;
694 unsigned long uninitialized_var(address);
697 address = page_address_in_vma(page, vma);
698 if (address != -EFAULT)
704 * if !vma, alloc_page_vma() will use task or system default policy
706 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
710 static void migrate_page_add(struct page *page, struct list_head *pagelist,
715 int do_migrate_pages(struct mm_struct *mm,
716 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
721 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
727 static long do_mbind(unsigned long start, unsigned long len,
728 unsigned short mode, unsigned short mode_flags,
729 nodemask_t *nmask, unsigned long flags)
731 struct vm_area_struct *vma;
732 struct mm_struct *mm = current->mm;
733 struct mempolicy *new;
738 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
739 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
741 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
744 if (start & ~PAGE_MASK)
747 if (mode == MPOL_DEFAULT)
748 flags &= ~MPOL_MF_STRICT;
750 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
758 new = mpol_new(mode, mode_flags, nmask);
763 * If we are using the default policy then operation
764 * on discontinuous address spaces is okay after all
767 flags |= MPOL_MF_DISCONTIG_OK;
769 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
770 start, start + len, mode, mode_flags,
771 nmask ? nodes_addr(*nmask)[0] : -1);
773 down_write(&mm->mmap_sem);
774 vma = check_range(mm, start, end, nmask,
775 flags | MPOL_MF_INVERT, &pagelist);
781 err = mbind_range(vma, start, end, new);
783 if (!list_empty(&pagelist))
784 nr_failed = migrate_pages(&pagelist, new_vma_page,
787 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
791 up_write(&mm->mmap_sem);
797 * User space interface with variable sized bitmaps for nodelists.
800 /* Copy a node mask from user space. */
801 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
802 unsigned long maxnode)
805 unsigned long nlongs;
806 unsigned long endmask;
810 if (maxnode == 0 || !nmask)
812 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
815 nlongs = BITS_TO_LONGS(maxnode);
816 if ((maxnode % BITS_PER_LONG) == 0)
819 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
821 /* When the user specified more nodes than supported just check
822 if the non supported part is all zero. */
823 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
824 if (nlongs > PAGE_SIZE/sizeof(long))
826 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
828 if (get_user(t, nmask + k))
830 if (k == nlongs - 1) {
836 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
840 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
842 nodes_addr(*nodes)[nlongs-1] &= endmask;
846 /* Copy a kernel node mask to user space */
847 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
850 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
851 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
854 if (copy > PAGE_SIZE)
856 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
860 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
863 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
865 unsigned long __user *nmask, unsigned long maxnode,
870 unsigned short mode_flags;
872 mode_flags = mode & MPOL_MODE_FLAGS;
873 mode &= ~MPOL_MODE_FLAGS;
874 if (mode >= MPOL_MAX)
876 err = get_nodes(&nodes, nmask, maxnode);
879 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
882 /* Set the process memory policy */
883 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
884 unsigned long maxnode)
888 unsigned short flags;
890 flags = mode & MPOL_MODE_FLAGS;
891 mode &= ~MPOL_MODE_FLAGS;
892 if ((unsigned int)mode >= MPOL_MAX)
894 err = get_nodes(&nodes, nmask, maxnode);
897 return do_set_mempolicy(mode, flags, &nodes);
900 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
901 const unsigned long __user *old_nodes,
902 const unsigned long __user *new_nodes)
904 struct mm_struct *mm;
905 struct task_struct *task;
908 nodemask_t task_nodes;
911 err = get_nodes(&old, old_nodes, maxnode);
915 err = get_nodes(&new, new_nodes, maxnode);
919 /* Find the mm_struct */
920 read_lock(&tasklist_lock);
921 task = pid ? find_task_by_vpid(pid) : current;
923 read_unlock(&tasklist_lock);
926 mm = get_task_mm(task);
927 read_unlock(&tasklist_lock);
933 * Check if this process has the right to modify the specified
934 * process. The right exists if the process has administrative
935 * capabilities, superuser privileges or the same
936 * userid as the target process.
938 if ((current->euid != task->suid) && (current->euid != task->uid) &&
939 (current->uid != task->suid) && (current->uid != task->uid) &&
940 !capable(CAP_SYS_NICE)) {
945 task_nodes = cpuset_mems_allowed(task);
946 /* Is the user allowed to access the target nodes? */
947 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
952 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
957 err = security_task_movememory(task);
961 err = do_migrate_pages(mm, &old, &new,
962 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
969 /* Retrieve NUMA policy */
970 asmlinkage long sys_get_mempolicy(int __user *policy,
971 unsigned long __user *nmask,
972 unsigned long maxnode,
973 unsigned long addr, unsigned long flags)
976 int uninitialized_var(pval);
979 if (nmask != NULL && maxnode < MAX_NUMNODES)
982 err = do_get_mempolicy(&pval, &nodes, addr, flags);
987 if (policy && put_user(pval, policy))
991 err = copy_nodes_to_user(nmask, maxnode, &nodes);
998 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
999 compat_ulong_t __user *nmask,
1000 compat_ulong_t maxnode,
1001 compat_ulong_t addr, compat_ulong_t flags)
1004 unsigned long __user *nm = NULL;
1005 unsigned long nr_bits, alloc_size;
1006 DECLARE_BITMAP(bm, MAX_NUMNODES);
1008 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1009 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1012 nm = compat_alloc_user_space(alloc_size);
1014 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1016 if (!err && nmask) {
1017 err = copy_from_user(bm, nm, alloc_size);
1018 /* ensure entire bitmap is zeroed */
1019 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1020 err |= compat_put_bitmap(nmask, bm, nr_bits);
1026 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1027 compat_ulong_t maxnode)
1030 unsigned long __user *nm = NULL;
1031 unsigned long nr_bits, alloc_size;
1032 DECLARE_BITMAP(bm, MAX_NUMNODES);
1034 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1035 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1038 err = compat_get_bitmap(bm, nmask, nr_bits);
1039 nm = compat_alloc_user_space(alloc_size);
1040 err |= copy_to_user(nm, bm, alloc_size);
1046 return sys_set_mempolicy(mode, nm, nr_bits+1);
1049 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1050 compat_ulong_t mode, compat_ulong_t __user *nmask,
1051 compat_ulong_t maxnode, compat_ulong_t flags)
1054 unsigned long __user *nm = NULL;
1055 unsigned long nr_bits, alloc_size;
1058 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1059 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1062 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1063 nm = compat_alloc_user_space(alloc_size);
1064 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1070 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1076 * get_vma_policy(@task, @vma, @addr)
1077 * @task - task for fallback if vma policy == default
1078 * @vma - virtual memory area whose policy is sought
1079 * @addr - address in @vma for shared policy lookup
1081 * Returns effective policy for a VMA at specified address.
1082 * Falls back to @task or system default policy, as necessary.
1083 * Returned policy has extra reference count if shared, vma,
1084 * or some other task's policy [show_numa_maps() can pass
1085 * @task != current]. It is the caller's responsibility to
1086 * free the reference in these cases.
1088 static struct mempolicy * get_vma_policy(struct task_struct *task,
1089 struct vm_area_struct *vma, unsigned long addr)
1091 struct mempolicy *pol = task->mempolicy;
1095 if (vma->vm_ops && vma->vm_ops->get_policy) {
1096 pol = vma->vm_ops->get_policy(vma, addr);
1097 shared_pol = 1; /* if pol non-NULL, add ref below */
1098 } else if (vma->vm_policy &&
1099 vma->vm_policy->policy != MPOL_DEFAULT)
1100 pol = vma->vm_policy;
1103 pol = &default_policy;
1104 else if (!shared_pol && pol != current->mempolicy)
1105 mpol_get(pol); /* vma or other task's policy */
1109 /* Return a nodemask representing a mempolicy */
1110 static nodemask_t *nodemask_policy(gfp_t gfp, struct mempolicy *policy)
1112 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1113 if (unlikely(policy->policy == MPOL_BIND) &&
1114 gfp_zone(gfp) >= policy_zone &&
1115 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1116 return &policy->v.nodes;
1121 /* Return a zonelist representing a mempolicy */
1122 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1126 switch (policy->policy) {
1127 case MPOL_PREFERRED:
1128 nd = policy->v.preferred_node;
1130 nd = numa_node_id();
1134 * Normally, MPOL_BIND allocations node-local are node-local
1135 * within the allowed nodemask. However, if __GFP_THISNODE is
1136 * set and the current node is part of the mask, we use the
1137 * the zonelist for the first node in the mask instead.
1139 nd = numa_node_id();
1140 if (unlikely(gfp & __GFP_THISNODE) &&
1141 unlikely(!node_isset(nd, policy->v.nodes)))
1142 nd = first_node(policy->v.nodes);
1144 case MPOL_INTERLEAVE: /* should not happen */
1146 nd = numa_node_id();
1152 return node_zonelist(nd, gfp);
1155 /* Do dynamic interleaving for a process */
1156 static unsigned interleave_nodes(struct mempolicy *policy)
1159 struct task_struct *me = current;
1162 next = next_node(nid, policy->v.nodes);
1163 if (next >= MAX_NUMNODES)
1164 next = first_node(policy->v.nodes);
1165 if (next < MAX_NUMNODES)
1171 * Depending on the memory policy provide a node from which to allocate the
1174 unsigned slab_node(struct mempolicy *policy)
1176 unsigned short pol = policy ? policy->policy : MPOL_DEFAULT;
1179 case MPOL_INTERLEAVE:
1180 return interleave_nodes(policy);
1184 * Follow bind policy behavior and start allocation at the
1187 struct zonelist *zonelist;
1189 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1190 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1191 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1197 case MPOL_PREFERRED:
1198 if (policy->v.preferred_node >= 0)
1199 return policy->v.preferred_node;
1203 return numa_node_id();
1207 /* Do static interleaving for a VMA with known offset. */
1208 static unsigned offset_il_node(struct mempolicy *pol,
1209 struct vm_area_struct *vma, unsigned long off)
1211 unsigned nnodes = nodes_weight(pol->v.nodes);
1217 return numa_node_id();
1218 target = (unsigned int)off % nnodes;
1221 nid = next_node(nid, pol->v.nodes);
1223 } while (c <= target);
1227 /* Determine a node number for interleave */
1228 static inline unsigned interleave_nid(struct mempolicy *pol,
1229 struct vm_area_struct *vma, unsigned long addr, int shift)
1235 * for small pages, there is no difference between
1236 * shift and PAGE_SHIFT, so the bit-shift is safe.
1237 * for huge pages, since vm_pgoff is in units of small
1238 * pages, we need to shift off the always 0 bits to get
1241 BUG_ON(shift < PAGE_SHIFT);
1242 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1243 off += (addr - vma->vm_start) >> shift;
1244 return offset_il_node(pol, vma, off);
1246 return interleave_nodes(pol);
1249 #ifdef CONFIG_HUGETLBFS
1251 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1252 * @vma = virtual memory area whose policy is sought
1253 * @addr = address in @vma for shared policy lookup and interleave policy
1254 * @gfp_flags = for requested zone
1255 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1256 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1258 * Returns a zonelist suitable for a huge page allocation.
1259 * If the effective policy is 'BIND, returns pointer to local node's zonelist,
1260 * and a pointer to the mempolicy's @nodemask for filtering the zonelist.
1261 * If it is also a policy for which get_vma_policy() returns an extra
1262 * reference, we must hold that reference until after the allocation.
1263 * In that case, return policy via @mpol so hugetlb allocation can drop
1264 * the reference. For non-'BIND referenced policies, we can/do drop the
1265 * reference here, so the caller doesn't need to know about the special case
1266 * for default and current task policy.
1268 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1269 gfp_t gfp_flags, struct mempolicy **mpol,
1270 nodemask_t **nodemask)
1272 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1273 struct zonelist *zl;
1275 *mpol = NULL; /* probably no unref needed */
1276 *nodemask = NULL; /* assume !MPOL_BIND */
1277 if (pol->policy == MPOL_BIND) {
1278 *nodemask = &pol->v.nodes;
1279 } else if (pol->policy == MPOL_INTERLEAVE) {
1282 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1283 if (unlikely(pol != &default_policy &&
1284 pol != current->mempolicy))
1285 __mpol_free(pol); /* finished with pol */
1286 return node_zonelist(nid, gfp_flags);
1289 zl = zonelist_policy(GFP_HIGHUSER, pol);
1290 if (unlikely(pol != &default_policy && pol != current->mempolicy)) {
1291 if (pol->policy != MPOL_BIND)
1292 __mpol_free(pol); /* finished with pol */
1294 *mpol = pol; /* unref needed after allocation */
1300 /* Allocate a page in interleaved policy.
1301 Own path because it needs to do special accounting. */
1302 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1305 struct zonelist *zl;
1308 zl = node_zonelist(nid, gfp);
1309 page = __alloc_pages(gfp, order, zl);
1310 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1311 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1316 * alloc_page_vma - Allocate a page for a VMA.
1319 * %GFP_USER user allocation.
1320 * %GFP_KERNEL kernel allocations,
1321 * %GFP_HIGHMEM highmem/user allocations,
1322 * %GFP_FS allocation should not call back into a file system.
1323 * %GFP_ATOMIC don't sleep.
1325 * @vma: Pointer to VMA or NULL if not available.
1326 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1328 * This function allocates a page from the kernel page pool and applies
1329 * a NUMA policy associated with the VMA or the current process.
1330 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1331 * mm_struct of the VMA to prevent it from going away. Should be used for
1332 * all allocations for pages that will be mapped into
1333 * user space. Returns NULL when no page can be allocated.
1335 * Should be called with the mm_sem of the vma hold.
1338 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1340 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1341 struct zonelist *zl;
1343 cpuset_update_task_memory_state();
1345 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1348 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1349 if (unlikely(pol != &default_policy &&
1350 pol != current->mempolicy))
1351 __mpol_free(pol); /* finished with pol */
1352 return alloc_page_interleave(gfp, 0, nid);
1354 zl = zonelist_policy(gfp, pol);
1355 if (pol != &default_policy && pol != current->mempolicy) {
1357 * slow path: ref counted policy -- shared or vma
1359 struct page *page = __alloc_pages_nodemask(gfp, 0,
1360 zl, nodemask_policy(gfp, pol));
1365 * fast path: default or task policy
1367 return __alloc_pages_nodemask(gfp, 0, zl, nodemask_policy(gfp, pol));
1371 * alloc_pages_current - Allocate pages.
1374 * %GFP_USER user allocation,
1375 * %GFP_KERNEL kernel allocation,
1376 * %GFP_HIGHMEM highmem allocation,
1377 * %GFP_FS don't call back into a file system.
1378 * %GFP_ATOMIC don't sleep.
1379 * @order: Power of two of allocation size in pages. 0 is a single page.
1381 * Allocate a page from the kernel page pool. When not in
1382 * interrupt context and apply the current process NUMA policy.
1383 * Returns NULL when no page can be allocated.
1385 * Don't call cpuset_update_task_memory_state() unless
1386 * 1) it's ok to take cpuset_sem (can WAIT), and
1387 * 2) allocating for current task (not interrupt).
1389 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1391 struct mempolicy *pol = current->mempolicy;
1393 if ((gfp & __GFP_WAIT) && !in_interrupt())
1394 cpuset_update_task_memory_state();
1395 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1396 pol = &default_policy;
1397 if (pol->policy == MPOL_INTERLEAVE)
1398 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1399 return __alloc_pages_nodemask(gfp, order,
1400 zonelist_policy(gfp, pol), nodemask_policy(gfp, pol));
1402 EXPORT_SYMBOL(alloc_pages_current);
1405 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1406 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1407 * with the mems_allowed returned by cpuset_mems_allowed(). This
1408 * keeps mempolicies cpuset relative after its cpuset moves. See
1409 * further kernel/cpuset.c update_nodemask().
1412 /* Slow path of a mempolicy copy */
1413 struct mempolicy *__mpol_copy(struct mempolicy *old)
1415 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1418 return ERR_PTR(-ENOMEM);
1419 if (current_cpuset_is_being_rebound()) {
1420 nodemask_t mems = cpuset_mems_allowed(current);
1421 mpol_rebind_policy(old, &mems);
1424 atomic_set(&new->refcnt, 1);
1428 static int mpol_match_intent(const struct mempolicy *a,
1429 const struct mempolicy *b)
1431 if (a->flags != b->flags)
1433 if (!mpol_store_user_nodemask(a))
1435 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1438 /* Slow path of a mempolicy comparison */
1439 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1443 if (a->policy != b->policy)
1445 if (a->policy != MPOL_DEFAULT && !mpol_match_intent(a, b))
1447 switch (a->policy) {
1452 case MPOL_INTERLEAVE:
1453 return nodes_equal(a->v.nodes, b->v.nodes);
1454 case MPOL_PREFERRED:
1455 return a->v.preferred_node == b->v.preferred_node;
1462 /* Slow path of a mpol destructor. */
1463 void __mpol_free(struct mempolicy *p)
1465 if (!atomic_dec_and_test(&p->refcnt))
1467 p->policy = MPOL_DEFAULT;
1468 kmem_cache_free(policy_cache, p);
1472 * Shared memory backing store policy support.
1474 * Remember policies even when nobody has shared memory mapped.
1475 * The policies are kept in Red-Black tree linked from the inode.
1476 * They are protected by the sp->lock spinlock, which should be held
1477 * for any accesses to the tree.
1480 /* lookup first element intersecting start-end */
1481 /* Caller holds sp->lock */
1482 static struct sp_node *
1483 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1485 struct rb_node *n = sp->root.rb_node;
1488 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1490 if (start >= p->end)
1492 else if (end <= p->start)
1500 struct sp_node *w = NULL;
1501 struct rb_node *prev = rb_prev(n);
1504 w = rb_entry(prev, struct sp_node, nd);
1505 if (w->end <= start)
1509 return rb_entry(n, struct sp_node, nd);
1512 /* Insert a new shared policy into the list. */
1513 /* Caller holds sp->lock */
1514 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1516 struct rb_node **p = &sp->root.rb_node;
1517 struct rb_node *parent = NULL;
1522 nd = rb_entry(parent, struct sp_node, nd);
1523 if (new->start < nd->start)
1525 else if (new->end > nd->end)
1526 p = &(*p)->rb_right;
1530 rb_link_node(&new->nd, parent, p);
1531 rb_insert_color(&new->nd, &sp->root);
1532 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1533 new->policy ? new->policy->policy : 0);
1536 /* Find shared policy intersecting idx */
1538 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1540 struct mempolicy *pol = NULL;
1543 if (!sp->root.rb_node)
1545 spin_lock(&sp->lock);
1546 sn = sp_lookup(sp, idx, idx+1);
1548 mpol_get(sn->policy);
1551 spin_unlock(&sp->lock);
1555 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1557 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1558 rb_erase(&n->nd, &sp->root);
1559 mpol_free(n->policy);
1560 kmem_cache_free(sn_cache, n);
1563 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1564 struct mempolicy *pol)
1566 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1577 /* Replace a policy range. */
1578 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1579 unsigned long end, struct sp_node *new)
1581 struct sp_node *n, *new2 = NULL;
1584 spin_lock(&sp->lock);
1585 n = sp_lookup(sp, start, end);
1586 /* Take care of old policies in the same range. */
1587 while (n && n->start < end) {
1588 struct rb_node *next = rb_next(&n->nd);
1589 if (n->start >= start) {
1595 /* Old policy spanning whole new range. */
1598 spin_unlock(&sp->lock);
1599 new2 = sp_alloc(end, n->end, n->policy);
1605 sp_insert(sp, new2);
1613 n = rb_entry(next, struct sp_node, nd);
1617 spin_unlock(&sp->lock);
1619 mpol_free(new2->policy);
1620 kmem_cache_free(sn_cache, new2);
1625 void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
1626 unsigned short flags, nodemask_t *policy_nodes)
1628 info->root = RB_ROOT;
1629 spin_lock_init(&info->lock);
1631 if (policy != MPOL_DEFAULT) {
1632 struct mempolicy *newpol;
1634 /* Falls back to MPOL_DEFAULT on any error */
1635 newpol = mpol_new(policy, flags, policy_nodes);
1636 if (!IS_ERR(newpol)) {
1637 /* Create pseudo-vma that contains just the policy */
1638 struct vm_area_struct pvma;
1640 memset(&pvma, 0, sizeof(struct vm_area_struct));
1641 /* Policy covers entire file */
1642 pvma.vm_end = TASK_SIZE;
1643 mpol_set_shared_policy(info, &pvma, newpol);
1649 int mpol_set_shared_policy(struct shared_policy *info,
1650 struct vm_area_struct *vma, struct mempolicy *npol)
1653 struct sp_node *new = NULL;
1654 unsigned long sz = vma_pages(vma);
1656 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1658 sz, npol ? npol->policy : -1,
1659 npol ? npol->flags : -1,
1660 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1663 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1667 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1669 kmem_cache_free(sn_cache, new);
1673 /* Free a backing policy store on inode delete. */
1674 void mpol_free_shared_policy(struct shared_policy *p)
1677 struct rb_node *next;
1679 if (!p->root.rb_node)
1681 spin_lock(&p->lock);
1682 next = rb_first(&p->root);
1684 n = rb_entry(next, struct sp_node, nd);
1685 next = rb_next(&n->nd);
1686 rb_erase(&n->nd, &p->root);
1687 mpol_free(n->policy);
1688 kmem_cache_free(sn_cache, n);
1690 spin_unlock(&p->lock);
1693 /* assumes fs == KERNEL_DS */
1694 void __init numa_policy_init(void)
1696 nodemask_t interleave_nodes;
1697 unsigned long largest = 0;
1698 int nid, prefer = 0;
1700 policy_cache = kmem_cache_create("numa_policy",
1701 sizeof(struct mempolicy),
1702 0, SLAB_PANIC, NULL);
1704 sn_cache = kmem_cache_create("shared_policy_node",
1705 sizeof(struct sp_node),
1706 0, SLAB_PANIC, NULL);
1709 * Set interleaving policy for system init. Interleaving is only
1710 * enabled across suitably sized nodes (default is >= 16MB), or
1711 * fall back to the largest node if they're all smaller.
1713 nodes_clear(interleave_nodes);
1714 for_each_node_state(nid, N_HIGH_MEMORY) {
1715 unsigned long total_pages = node_present_pages(nid);
1717 /* Preserve the largest node */
1718 if (largest < total_pages) {
1719 largest = total_pages;
1723 /* Interleave this node? */
1724 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1725 node_set(nid, interleave_nodes);
1728 /* All too small, use the largest */
1729 if (unlikely(nodes_empty(interleave_nodes)))
1730 node_set(prefer, interleave_nodes);
1732 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
1733 printk("numa_policy_init: interleaving failed\n");
1736 /* Reset policy of current process to default */
1737 void numa_default_policy(void)
1739 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1742 /* Migrate a policy to a different set of nodes */
1743 static void mpol_rebind_policy(struct mempolicy *pol,
1744 const nodemask_t *newmask)
1751 static_nodes = pol->flags & MPOL_F_STATIC_NODES;
1752 if (!mpol_store_user_nodemask(pol) &&
1753 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
1756 switch (pol->policy) {
1761 case MPOL_INTERLEAVE:
1763 nodes_and(tmp, pol->w.user_nodemask, *newmask);
1765 nodes_remap(tmp, pol->v.nodes,
1766 pol->w.cpuset_mems_allowed, *newmask);
1767 pol->w.cpuset_mems_allowed = *newmask;
1770 if (!node_isset(current->il_next, tmp)) {
1771 current->il_next = next_node(current->il_next, tmp);
1772 if (current->il_next >= MAX_NUMNODES)
1773 current->il_next = first_node(tmp);
1774 if (current->il_next >= MAX_NUMNODES)
1775 current->il_next = numa_node_id();
1778 case MPOL_PREFERRED:
1780 int node = first_node(pol->w.user_nodemask);
1782 if (node_isset(node, *newmask))
1783 pol->v.preferred_node = node;
1785 pol->v.preferred_node = -1;
1787 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1788 pol->w.cpuset_mems_allowed, *newmask);
1789 pol->w.cpuset_mems_allowed = *newmask;
1799 * Wrapper for mpol_rebind_policy() that just requires task
1800 * pointer, and updates task mempolicy.
1803 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1805 mpol_rebind_policy(tsk->mempolicy, new);
1809 * Rebind each vma in mm to new nodemask.
1811 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1814 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1816 struct vm_area_struct *vma;
1818 down_write(&mm->mmap_sem);
1819 for (vma = mm->mmap; vma; vma = vma->vm_next)
1820 mpol_rebind_policy(vma->vm_policy, new);
1821 up_write(&mm->mmap_sem);
1825 * Display pages allocated per node and memory policy via /proc.
1828 static const char * const policy_types[] =
1829 { "default", "prefer", "bind", "interleave" };
1832 * Convert a mempolicy into a string.
1833 * Returns the number of characters in buffer (if positive)
1834 * or an error (negative)
1836 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1841 unsigned short mode = pol ? pol->policy : MPOL_DEFAULT;
1842 unsigned short flags = pol ? pol->flags : 0;
1849 case MPOL_PREFERRED:
1851 node_set(pol->v.preferred_node, nodes);
1856 case MPOL_INTERLEAVE:
1857 nodes = pol->v.nodes;
1865 l = strlen(policy_types[mode]);
1866 if (buffer + maxlen < p + l + 1)
1869 strcpy(p, policy_types[mode]);
1875 if (buffer + maxlen < p + 2)
1879 if (flags & MPOL_F_STATIC_NODES)
1880 p += sprintf(p, "%sstatic", need_bar++ ? "|" : "");
1883 if (!nodes_empty(nodes)) {
1884 if (buffer + maxlen < p + 2)
1887 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1893 unsigned long pages;
1895 unsigned long active;
1896 unsigned long writeback;
1897 unsigned long mapcount_max;
1898 unsigned long dirty;
1899 unsigned long swapcache;
1900 unsigned long node[MAX_NUMNODES];
1903 static void gather_stats(struct page *page, void *private, int pte_dirty)
1905 struct numa_maps *md = private;
1906 int count = page_mapcount(page);
1909 if (pte_dirty || PageDirty(page))
1912 if (PageSwapCache(page))
1915 if (PageActive(page))
1918 if (PageWriteback(page))
1924 if (count > md->mapcount_max)
1925 md->mapcount_max = count;
1927 md->node[page_to_nid(page)]++;
1930 #ifdef CONFIG_HUGETLB_PAGE
1931 static void check_huge_range(struct vm_area_struct *vma,
1932 unsigned long start, unsigned long end,
1933 struct numa_maps *md)
1938 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1939 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1949 page = pte_page(pte);
1953 gather_stats(page, md, pte_dirty(*ptep));
1957 static inline void check_huge_range(struct vm_area_struct *vma,
1958 unsigned long start, unsigned long end,
1959 struct numa_maps *md)
1964 int show_numa_map(struct seq_file *m, void *v)
1966 struct proc_maps_private *priv = m->private;
1967 struct vm_area_struct *vma = v;
1968 struct numa_maps *md;
1969 struct file *file = vma->vm_file;
1970 struct mm_struct *mm = vma->vm_mm;
1971 struct mempolicy *pol;
1978 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1982 pol = get_vma_policy(priv->task, vma, vma->vm_start);
1983 mpol_to_str(buffer, sizeof(buffer), pol);
1985 * unref shared or other task's mempolicy
1987 if (pol != &default_policy && pol != current->mempolicy)
1990 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1993 seq_printf(m, " file=");
1994 seq_path(m, &file->f_path, "\n\t= ");
1995 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1996 seq_printf(m, " heap");
1997 } else if (vma->vm_start <= mm->start_stack &&
1998 vma->vm_end >= mm->start_stack) {
1999 seq_printf(m, " stack");
2002 if (is_vm_hugetlb_page(vma)) {
2003 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2004 seq_printf(m, " huge");
2006 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2007 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2014 seq_printf(m," anon=%lu",md->anon);
2017 seq_printf(m," dirty=%lu",md->dirty);
2019 if (md->pages != md->anon && md->pages != md->dirty)
2020 seq_printf(m, " mapped=%lu", md->pages);
2022 if (md->mapcount_max > 1)
2023 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2026 seq_printf(m," swapcache=%lu", md->swapcache);
2028 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2029 seq_printf(m," active=%lu", md->active);
2032 seq_printf(m," writeback=%lu", md->writeback);
2034 for_each_node_state(n, N_HIGH_MEMORY)
2036 seq_printf(m, " N%d=%lu", n, md->node[n]);
2041 if (m->count < m->size)
2042 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;