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>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/gfp.h>
79 #include <linux/slab.h>
80 #include <linux/string.h>
81 #include <linux/module.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/mempolicy.h>
86 #include <linux/swap.h>
87 #include <linux/seq_file.h>
88 #include <linux/proc_fs.h>
89 #include <linux/migrate.h>
91 #include <asm/tlbflush.h>
92 #include <asm/uaccess.h>
95 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
96 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
97 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
99 static struct kmem_cache *policy_cache;
100 static struct kmem_cache *sn_cache;
102 #define PDprintk(fmt...)
104 /* Highest zone. An specific allocation for a zone below that is not
106 int policy_zone = ZONE_DMA;
108 struct mempolicy default_policy = {
109 .refcnt = ATOMIC_INIT(1), /* never free it */
110 .policy = MPOL_DEFAULT,
113 /* Do sanity checking on a policy */
114 static int mpol_check_policy(int mode, nodemask_t *nodes)
116 int empty = nodes_empty(*nodes);
124 case MPOL_INTERLEAVE:
125 /* Preferred will only use the first bit, but allow
131 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
134 /* Generate a custom zonelist for the BIND policy. */
135 static struct zonelist *bind_zonelist(nodemask_t *nodes)
140 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
141 zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL);
145 /* First put in the highest zones from all nodes, then all the next
146 lower zones etc. Avoid empty zones because the memory allocator
147 doesn't like them. If you implement node hot removal you
149 for (k = policy_zone; k >= 0; k--) {
150 for_each_node_mask(nd, *nodes) {
151 struct zone *z = &NODE_DATA(nd)->node_zones[k];
152 if (z->present_pages > 0)
153 zl->zones[num++] = z;
156 zl->zones[num] = NULL;
160 /* Create a new policy */
161 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
163 struct mempolicy *policy;
165 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
166 if (mode == MPOL_DEFAULT)
168 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
170 return ERR_PTR(-ENOMEM);
171 atomic_set(&policy->refcnt, 1);
173 case MPOL_INTERLEAVE:
174 policy->v.nodes = *nodes;
175 if (nodes_weight(*nodes) == 0) {
176 kmem_cache_free(policy_cache, policy);
177 return ERR_PTR(-EINVAL);
181 policy->v.preferred_node = first_node(*nodes);
182 if (policy->v.preferred_node >= MAX_NUMNODES)
183 policy->v.preferred_node = -1;
186 policy->v.zonelist = bind_zonelist(nodes);
187 if (policy->v.zonelist == NULL) {
188 kmem_cache_free(policy_cache, policy);
189 return ERR_PTR(-ENOMEM);
193 policy->policy = mode;
194 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
198 static void gather_stats(struct page *, void *, int pte_dirty);
199 static void migrate_page_add(struct page *page, struct list_head *pagelist,
200 unsigned long flags);
202 /* Scan through pages checking if pages follow certain conditions. */
203 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
204 unsigned long addr, unsigned long end,
205 const nodemask_t *nodes, unsigned long flags,
212 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
217 if (!pte_present(*pte))
219 page = vm_normal_page(vma, addr, *pte);
223 * The check for PageReserved here is important to avoid
224 * handling zero pages and other pages that may have been
225 * marked special by the system.
227 * If the PageReserved would not be checked here then f.e.
228 * the location of the zero page could have an influence
229 * on MPOL_MF_STRICT, zero pages would be counted for
230 * the per node stats, and there would be useless attempts
231 * to put zero pages on the migration list.
233 if (PageReserved(page))
235 nid = page_to_nid(page);
236 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
239 if (flags & MPOL_MF_STATS)
240 gather_stats(page, private, pte_dirty(*pte));
241 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
242 migrate_page_add(page, private, flags);
245 } while (pte++, addr += PAGE_SIZE, addr != end);
246 pte_unmap_unlock(orig_pte, ptl);
250 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
251 unsigned long addr, unsigned long end,
252 const nodemask_t *nodes, unsigned long flags,
258 pmd = pmd_offset(pud, addr);
260 next = pmd_addr_end(addr, end);
261 if (pmd_none_or_clear_bad(pmd))
263 if (check_pte_range(vma, pmd, addr, next, nodes,
266 } while (pmd++, addr = next, addr != end);
270 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
271 unsigned long addr, unsigned long end,
272 const nodemask_t *nodes, unsigned long flags,
278 pud = pud_offset(pgd, addr);
280 next = pud_addr_end(addr, end);
281 if (pud_none_or_clear_bad(pud))
283 if (check_pmd_range(vma, pud, addr, next, nodes,
286 } while (pud++, addr = next, addr != end);
290 static inline int check_pgd_range(struct vm_area_struct *vma,
291 unsigned long addr, unsigned long end,
292 const nodemask_t *nodes, unsigned long flags,
298 pgd = pgd_offset(vma->vm_mm, addr);
300 next = pgd_addr_end(addr, end);
301 if (pgd_none_or_clear_bad(pgd))
303 if (check_pud_range(vma, pgd, addr, next, nodes,
306 } while (pgd++, addr = next, addr != end);
310 /* Check if a vma is migratable */
311 static inline int vma_migratable(struct vm_area_struct *vma)
313 if (vma->vm_flags & (
314 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED))
320 * Check if all pages in a range are on a set of nodes.
321 * If pagelist != NULL then isolate pages from the LRU and
322 * put them on the pagelist.
324 static struct vm_area_struct *
325 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
326 const nodemask_t *nodes, unsigned long flags, void *private)
329 struct vm_area_struct *first, *vma, *prev;
331 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
333 err = migrate_prep();
338 first = find_vma(mm, start);
340 return ERR_PTR(-EFAULT);
342 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
343 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
344 if (!vma->vm_next && vma->vm_end < end)
345 return ERR_PTR(-EFAULT);
346 if (prev && prev->vm_end < vma->vm_start)
347 return ERR_PTR(-EFAULT);
349 if (!is_vm_hugetlb_page(vma) &&
350 ((flags & MPOL_MF_STRICT) ||
351 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
352 vma_migratable(vma)))) {
353 unsigned long endvma = vma->vm_end;
357 if (vma->vm_start > start)
358 start = vma->vm_start;
359 err = check_pgd_range(vma, start, endvma, nodes,
362 first = ERR_PTR(err);
371 /* Apply policy to a single VMA */
372 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
375 struct mempolicy *old = vma->vm_policy;
377 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
378 vma->vm_start, vma->vm_end, vma->vm_pgoff,
379 vma->vm_ops, vma->vm_file,
380 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
382 if (vma->vm_ops && vma->vm_ops->set_policy)
383 err = vma->vm_ops->set_policy(vma, new);
386 vma->vm_policy = new;
392 /* Step 2: apply policy to a range and do splits. */
393 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
394 unsigned long end, struct mempolicy *new)
396 struct vm_area_struct *next;
400 for (; vma && vma->vm_start < end; vma = next) {
402 if (vma->vm_start < start)
403 err = split_vma(vma->vm_mm, vma, start, 1);
404 if (!err && vma->vm_end > end)
405 err = split_vma(vma->vm_mm, vma, end, 0);
407 err = policy_vma(vma, new);
414 static int contextualize_policy(int mode, nodemask_t *nodes)
419 cpuset_update_task_memory_state();
420 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
422 return mpol_check_policy(mode, nodes);
427 * Update task->flags PF_MEMPOLICY bit: set iff non-default
428 * mempolicy. Allows more rapid checking of this (combined perhaps
429 * with other PF_* flag bits) on memory allocation hot code paths.
431 * If called from outside this file, the task 'p' should -only- be
432 * a newly forked child not yet visible on the task list, because
433 * manipulating the task flags of a visible task is not safe.
435 * The above limitation is why this routine has the funny name
436 * mpol_fix_fork_child_flag().
438 * It is also safe to call this with a task pointer of current,
439 * which the static wrapper mpol_set_task_struct_flag() does,
440 * for use within this file.
443 void mpol_fix_fork_child_flag(struct task_struct *p)
446 p->flags |= PF_MEMPOLICY;
448 p->flags &= ~PF_MEMPOLICY;
451 static void mpol_set_task_struct_flag(void)
453 mpol_fix_fork_child_flag(current);
456 /* Set the process memory policy */
457 long do_set_mempolicy(int mode, nodemask_t *nodes)
459 struct mempolicy *new;
461 if (contextualize_policy(mode, nodes))
463 new = mpol_new(mode, nodes);
466 mpol_free(current->mempolicy);
467 current->mempolicy = new;
468 mpol_set_task_struct_flag();
469 if (new && new->policy == MPOL_INTERLEAVE)
470 current->il_next = first_node(new->v.nodes);
474 /* Fill a zone bitmap for a policy */
475 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
482 for (i = 0; p->v.zonelist->zones[i]; i++)
483 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
488 case MPOL_INTERLEAVE:
492 /* or use current node instead of online map? */
493 if (p->v.preferred_node < 0)
494 *nodes = node_online_map;
496 node_set(p->v.preferred_node, *nodes);
503 static int lookup_node(struct mm_struct *mm, unsigned long addr)
508 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
510 err = page_to_nid(p);
516 /* Retrieve NUMA policy */
517 long do_get_mempolicy(int *policy, nodemask_t *nmask,
518 unsigned long addr, unsigned long flags)
521 struct mm_struct *mm = current->mm;
522 struct vm_area_struct *vma = NULL;
523 struct mempolicy *pol = current->mempolicy;
525 cpuset_update_task_memory_state();
526 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
528 if (flags & MPOL_F_ADDR) {
529 down_read(&mm->mmap_sem);
530 vma = find_vma_intersection(mm, addr, addr+1);
532 up_read(&mm->mmap_sem);
535 if (vma->vm_ops && vma->vm_ops->get_policy)
536 pol = vma->vm_ops->get_policy(vma, addr);
538 pol = vma->vm_policy;
543 pol = &default_policy;
545 if (flags & MPOL_F_NODE) {
546 if (flags & MPOL_F_ADDR) {
547 err = lookup_node(mm, addr);
551 } else if (pol == current->mempolicy &&
552 pol->policy == MPOL_INTERLEAVE) {
553 *policy = current->il_next;
559 *policy = pol->policy;
562 up_read(¤t->mm->mmap_sem);
568 get_zonemask(pol, nmask);
572 up_read(¤t->mm->mmap_sem);
576 #ifdef CONFIG_MIGRATION
580 static void migrate_page_add(struct page *page, struct list_head *pagelist,
584 * Avoid migrating a page that is shared with others.
586 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
587 isolate_lru_page(page, pagelist);
591 * Migrate pages from one node to a target node.
592 * Returns error or the number of pages not migrated.
594 int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags)
601 node_set(source, nmask);
603 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
604 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
606 if (!list_empty(&pagelist))
607 err = migrate_pages_to(&pagelist, NULL, dest);
612 * Move pages between the two nodesets so as to preserve the physical
613 * layout as much as possible.
615 * Returns the number of page that could not be moved.
617 int do_migrate_pages(struct mm_struct *mm,
618 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
625 down_read(&mm->mmap_sem);
628 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
629 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
630 * bit in 'tmp', and return that <source, dest> pair for migration.
631 * The pair of nodemasks 'to' and 'from' define the map.
633 * If no pair of bits is found that way, fallback to picking some
634 * pair of 'source' and 'dest' bits that are not the same. If the
635 * 'source' and 'dest' bits are the same, this represents a node
636 * that will be migrating to itself, so no pages need move.
638 * If no bits are left in 'tmp', or if all remaining bits left
639 * in 'tmp' correspond to the same bit in 'to', return false
640 * (nothing left to migrate).
642 * This lets us pick a pair of nodes to migrate between, such that
643 * if possible the dest node is not already occupied by some other
644 * source node, minimizing the risk of overloading the memory on a
645 * node that would happen if we migrated incoming memory to a node
646 * before migrating outgoing memory source that same node.
648 * A single scan of tmp is sufficient. As we go, we remember the
649 * most recent <s, d> pair that moved (s != d). If we find a pair
650 * that not only moved, but what's better, moved to an empty slot
651 * (d is not set in tmp), then we break out then, with that pair.
652 * Otherwise when we finish scannng from_tmp, we at least have the
653 * most recent <s, d> pair that moved. If we get all the way through
654 * the scan of tmp without finding any node that moved, much less
655 * moved to an empty node, then there is nothing left worth migrating.
659 while (!nodes_empty(tmp)) {
664 for_each_node_mask(s, tmp) {
665 d = node_remap(s, *from_nodes, *to_nodes);
669 source = s; /* Node moved. Memorize */
672 /* dest not in remaining from nodes? */
673 if (!node_isset(dest, tmp))
679 node_clear(source, tmp);
680 err = migrate_to_node(mm, source, dest, flags);
687 up_read(&mm->mmap_sem);
696 static void migrate_page_add(struct page *page, struct list_head *pagelist,
701 int do_migrate_pages(struct mm_struct *mm,
702 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
708 long do_mbind(unsigned long start, unsigned long len,
709 unsigned long mode, nodemask_t *nmask, unsigned long flags)
711 struct vm_area_struct *vma;
712 struct mm_struct *mm = current->mm;
713 struct mempolicy *new;
718 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
719 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
722 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
725 if (start & ~PAGE_MASK)
728 if (mode == MPOL_DEFAULT)
729 flags &= ~MPOL_MF_STRICT;
731 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
739 if (mpol_check_policy(mode, nmask))
742 new = mpol_new(mode, nmask);
747 * If we are using the default policy then operation
748 * on discontinuous address spaces is okay after all
751 flags |= MPOL_MF_DISCONTIG_OK;
753 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
754 mode,nodes_addr(nodes)[0]);
756 down_write(&mm->mmap_sem);
757 vma = check_range(mm, start, end, nmask,
758 flags | MPOL_MF_INVERT, &pagelist);
764 err = mbind_range(vma, start, end, new);
766 if (!list_empty(&pagelist))
767 nr_failed = migrate_pages_to(&pagelist, vma, -1);
769 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
773 up_write(&mm->mmap_sem);
779 * User space interface with variable sized bitmaps for nodelists.
782 /* Copy a node mask from user space. */
783 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
784 unsigned long maxnode)
787 unsigned long nlongs;
788 unsigned long endmask;
792 if (maxnode == 0 || !nmask)
794 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
797 nlongs = BITS_TO_LONGS(maxnode);
798 if ((maxnode % BITS_PER_LONG) == 0)
801 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
803 /* When the user specified more nodes than supported just check
804 if the non supported part is all zero. */
805 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
806 if (nlongs > PAGE_SIZE/sizeof(long))
808 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
810 if (get_user(t, nmask + k))
812 if (k == nlongs - 1) {
818 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
822 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
824 nodes_addr(*nodes)[nlongs-1] &= endmask;
828 /* Copy a kernel node mask to user space */
829 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
832 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
833 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
836 if (copy > PAGE_SIZE)
838 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
842 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
845 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
847 unsigned long __user *nmask, unsigned long maxnode,
853 err = get_nodes(&nodes, nmask, maxnode);
856 return do_mbind(start, len, mode, &nodes, flags);
859 /* Set the process memory policy */
860 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
861 unsigned long maxnode)
866 if (mode < 0 || mode > MPOL_MAX)
868 err = get_nodes(&nodes, nmask, maxnode);
871 return do_set_mempolicy(mode, &nodes);
874 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
875 const unsigned long __user *old_nodes,
876 const unsigned long __user *new_nodes)
878 struct mm_struct *mm;
879 struct task_struct *task;
882 nodemask_t task_nodes;
885 err = get_nodes(&old, old_nodes, maxnode);
889 err = get_nodes(&new, new_nodes, maxnode);
893 /* Find the mm_struct */
894 read_lock(&tasklist_lock);
895 task = pid ? find_task_by_pid(pid) : current;
897 read_unlock(&tasklist_lock);
900 mm = get_task_mm(task);
901 read_unlock(&tasklist_lock);
907 * Check if this process has the right to modify the specified
908 * process. The right exists if the process has administrative
909 * capabilities, superuser privileges or the same
910 * userid as the target process.
912 if ((current->euid != task->suid) && (current->euid != task->uid) &&
913 (current->uid != task->suid) && (current->uid != task->uid) &&
914 !capable(CAP_SYS_NICE)) {
919 task_nodes = cpuset_mems_allowed(task);
920 /* Is the user allowed to access the target nodes? */
921 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
926 err = do_migrate_pages(mm, &old, &new,
927 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
934 /* Retrieve NUMA policy */
935 asmlinkage long sys_get_mempolicy(int __user *policy,
936 unsigned long __user *nmask,
937 unsigned long maxnode,
938 unsigned long addr, unsigned long flags)
943 if (nmask != NULL && maxnode < MAX_NUMNODES)
946 err = do_get_mempolicy(&pval, &nodes, addr, flags);
951 if (policy && put_user(pval, policy))
955 err = copy_nodes_to_user(nmask, maxnode, &nodes);
962 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
963 compat_ulong_t __user *nmask,
964 compat_ulong_t maxnode,
965 compat_ulong_t addr, compat_ulong_t flags)
968 unsigned long __user *nm = NULL;
969 unsigned long nr_bits, alloc_size;
970 DECLARE_BITMAP(bm, MAX_NUMNODES);
972 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
973 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
976 nm = compat_alloc_user_space(alloc_size);
978 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
981 err = copy_from_user(bm, nm, alloc_size);
982 /* ensure entire bitmap is zeroed */
983 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
984 err |= compat_put_bitmap(nmask, bm, nr_bits);
990 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
991 compat_ulong_t maxnode)
994 unsigned long __user *nm = NULL;
995 unsigned long nr_bits, alloc_size;
996 DECLARE_BITMAP(bm, MAX_NUMNODES);
998 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
999 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1002 err = compat_get_bitmap(bm, nmask, nr_bits);
1003 nm = compat_alloc_user_space(alloc_size);
1004 err |= copy_to_user(nm, bm, alloc_size);
1010 return sys_set_mempolicy(mode, nm, nr_bits+1);
1013 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1014 compat_ulong_t mode, compat_ulong_t __user *nmask,
1015 compat_ulong_t maxnode, compat_ulong_t flags)
1018 unsigned long __user *nm = NULL;
1019 unsigned long nr_bits, alloc_size;
1022 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1023 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1026 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1027 nm = compat_alloc_user_space(alloc_size);
1028 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1034 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1039 /* Return effective policy for a VMA */
1040 static struct mempolicy * get_vma_policy(struct task_struct *task,
1041 struct vm_area_struct *vma, unsigned long addr)
1043 struct mempolicy *pol = task->mempolicy;
1046 if (vma->vm_ops && vma->vm_ops->get_policy)
1047 pol = vma->vm_ops->get_policy(vma, addr);
1048 else if (vma->vm_policy &&
1049 vma->vm_policy->policy != MPOL_DEFAULT)
1050 pol = vma->vm_policy;
1053 pol = &default_policy;
1057 /* Return a zonelist representing a mempolicy */
1058 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1062 switch (policy->policy) {
1063 case MPOL_PREFERRED:
1064 nd = policy->v.preferred_node;
1066 nd = numa_node_id();
1069 /* Lower zones don't get a policy applied */
1070 /* Careful: current->mems_allowed might have moved */
1071 if (gfp_zone(gfp) >= policy_zone)
1072 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
1073 return policy->v.zonelist;
1075 case MPOL_INTERLEAVE: /* should not happen */
1077 nd = numa_node_id();
1083 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
1086 /* Do dynamic interleaving for a process */
1087 static unsigned interleave_nodes(struct mempolicy *policy)
1090 struct task_struct *me = current;
1093 next = next_node(nid, policy->v.nodes);
1094 if (next >= MAX_NUMNODES)
1095 next = first_node(policy->v.nodes);
1101 * Depending on the memory policy provide a node from which to allocate the
1104 unsigned slab_node(struct mempolicy *policy)
1106 switch (policy->policy) {
1107 case MPOL_INTERLEAVE:
1108 return interleave_nodes(policy);
1112 * Follow bind policy behavior and start allocation at the
1115 return policy->v.zonelist->zones[0]->zone_pgdat->node_id;
1117 case MPOL_PREFERRED:
1118 if (policy->v.preferred_node >= 0)
1119 return policy->v.preferred_node;
1123 return numa_node_id();
1127 /* Do static interleaving for a VMA with known offset. */
1128 static unsigned offset_il_node(struct mempolicy *pol,
1129 struct vm_area_struct *vma, unsigned long off)
1131 unsigned nnodes = nodes_weight(pol->v.nodes);
1132 unsigned target = (unsigned)off % nnodes;
1138 nid = next_node(nid, pol->v.nodes);
1140 } while (c <= target);
1144 /* Determine a node number for interleave */
1145 static inline unsigned interleave_nid(struct mempolicy *pol,
1146 struct vm_area_struct *vma, unsigned long addr, int shift)
1151 off = vma->vm_pgoff;
1152 off += (addr - vma->vm_start) >> shift;
1153 return offset_il_node(pol, vma, off);
1155 return interleave_nodes(pol);
1158 #ifdef CONFIG_HUGETLBFS
1159 /* Return a zonelist suitable for a huge page allocation. */
1160 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1162 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1164 if (pol->policy == MPOL_INTERLEAVE) {
1167 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1168 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1170 return zonelist_policy(GFP_HIGHUSER, pol);
1174 /* Allocate a page in interleaved policy.
1175 Own path because it needs to do special accounting. */
1176 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1179 struct zonelist *zl;
1182 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1183 page = __alloc_pages(gfp, order, zl);
1184 if (page && page_zone(page) == zl->zones[0]) {
1185 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1192 * alloc_page_vma - Allocate a page for a VMA.
1195 * %GFP_USER user allocation.
1196 * %GFP_KERNEL kernel allocations,
1197 * %GFP_HIGHMEM highmem/user allocations,
1198 * %GFP_FS allocation should not call back into a file system.
1199 * %GFP_ATOMIC don't sleep.
1201 * @vma: Pointer to VMA or NULL if not available.
1202 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1204 * This function allocates a page from the kernel page pool and applies
1205 * a NUMA policy associated with the VMA or the current process.
1206 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1207 * mm_struct of the VMA to prevent it from going away. Should be used for
1208 * all allocations for pages that will be mapped into
1209 * user space. Returns NULL when no page can be allocated.
1211 * Should be called with the mm_sem of the vma hold.
1214 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1216 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1218 cpuset_update_task_memory_state();
1220 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1223 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1224 return alloc_page_interleave(gfp, 0, nid);
1226 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1230 * alloc_pages_current - Allocate pages.
1233 * %GFP_USER user allocation,
1234 * %GFP_KERNEL kernel allocation,
1235 * %GFP_HIGHMEM highmem allocation,
1236 * %GFP_FS don't call back into a file system.
1237 * %GFP_ATOMIC don't sleep.
1238 * @order: Power of two of allocation size in pages. 0 is a single page.
1240 * Allocate a page from the kernel page pool. When not in
1241 * interrupt context and apply the current process NUMA policy.
1242 * Returns NULL when no page can be allocated.
1244 * Don't call cpuset_update_task_memory_state() unless
1245 * 1) it's ok to take cpuset_sem (can WAIT), and
1246 * 2) allocating for current task (not interrupt).
1248 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1250 struct mempolicy *pol = current->mempolicy;
1252 if ((gfp & __GFP_WAIT) && !in_interrupt())
1253 cpuset_update_task_memory_state();
1254 if (!pol || in_interrupt())
1255 pol = &default_policy;
1256 if (pol->policy == MPOL_INTERLEAVE)
1257 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1258 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1260 EXPORT_SYMBOL(alloc_pages_current);
1263 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1264 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1265 * with the mems_allowed returned by cpuset_mems_allowed(). This
1266 * keeps mempolicies cpuset relative after its cpuset moves. See
1267 * further kernel/cpuset.c update_nodemask().
1269 void *cpuset_being_rebound;
1271 /* Slow path of a mempolicy copy */
1272 struct mempolicy *__mpol_copy(struct mempolicy *old)
1274 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1277 return ERR_PTR(-ENOMEM);
1278 if (current_cpuset_is_being_rebound()) {
1279 nodemask_t mems = cpuset_mems_allowed(current);
1280 mpol_rebind_policy(old, &mems);
1283 atomic_set(&new->refcnt, 1);
1284 if (new->policy == MPOL_BIND) {
1285 int sz = ksize(old->v.zonelist);
1286 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1287 if (!new->v.zonelist) {
1288 kmem_cache_free(policy_cache, new);
1289 return ERR_PTR(-ENOMEM);
1291 memcpy(new->v.zonelist, old->v.zonelist, sz);
1296 /* Slow path of a mempolicy comparison */
1297 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1301 if (a->policy != b->policy)
1303 switch (a->policy) {
1306 case MPOL_INTERLEAVE:
1307 return nodes_equal(a->v.nodes, b->v.nodes);
1308 case MPOL_PREFERRED:
1309 return a->v.preferred_node == b->v.preferred_node;
1312 for (i = 0; a->v.zonelist->zones[i]; i++)
1313 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1315 return b->v.zonelist->zones[i] == NULL;
1323 /* Slow path of a mpol destructor. */
1324 void __mpol_free(struct mempolicy *p)
1326 if (!atomic_dec_and_test(&p->refcnt))
1328 if (p->policy == MPOL_BIND)
1329 kfree(p->v.zonelist);
1330 p->policy = MPOL_DEFAULT;
1331 kmem_cache_free(policy_cache, p);
1335 * Shared memory backing store policy support.
1337 * Remember policies even when nobody has shared memory mapped.
1338 * The policies are kept in Red-Black tree linked from the inode.
1339 * They are protected by the sp->lock spinlock, which should be held
1340 * for any accesses to the tree.
1343 /* lookup first element intersecting start-end */
1344 /* Caller holds sp->lock */
1345 static struct sp_node *
1346 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1348 struct rb_node *n = sp->root.rb_node;
1351 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1353 if (start >= p->end)
1355 else if (end <= p->start)
1363 struct sp_node *w = NULL;
1364 struct rb_node *prev = rb_prev(n);
1367 w = rb_entry(prev, struct sp_node, nd);
1368 if (w->end <= start)
1372 return rb_entry(n, struct sp_node, nd);
1375 /* Insert a new shared policy into the list. */
1376 /* Caller holds sp->lock */
1377 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1379 struct rb_node **p = &sp->root.rb_node;
1380 struct rb_node *parent = NULL;
1385 nd = rb_entry(parent, struct sp_node, nd);
1386 if (new->start < nd->start)
1388 else if (new->end > nd->end)
1389 p = &(*p)->rb_right;
1393 rb_link_node(&new->nd, parent, p);
1394 rb_insert_color(&new->nd, &sp->root);
1395 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1396 new->policy ? new->policy->policy : 0);
1399 /* Find shared policy intersecting idx */
1401 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1403 struct mempolicy *pol = NULL;
1406 if (!sp->root.rb_node)
1408 spin_lock(&sp->lock);
1409 sn = sp_lookup(sp, idx, idx+1);
1411 mpol_get(sn->policy);
1414 spin_unlock(&sp->lock);
1418 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1420 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1421 rb_erase(&n->nd, &sp->root);
1422 mpol_free(n->policy);
1423 kmem_cache_free(sn_cache, n);
1427 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1429 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1440 /* Replace a policy range. */
1441 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1442 unsigned long end, struct sp_node *new)
1444 struct sp_node *n, *new2 = NULL;
1447 spin_lock(&sp->lock);
1448 n = sp_lookup(sp, start, end);
1449 /* Take care of old policies in the same range. */
1450 while (n && n->start < end) {
1451 struct rb_node *next = rb_next(&n->nd);
1452 if (n->start >= start) {
1458 /* Old policy spanning whole new range. */
1461 spin_unlock(&sp->lock);
1462 new2 = sp_alloc(end, n->end, n->policy);
1468 sp_insert(sp, new2);
1476 n = rb_entry(next, struct sp_node, nd);
1480 spin_unlock(&sp->lock);
1482 mpol_free(new2->policy);
1483 kmem_cache_free(sn_cache, new2);
1488 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1489 nodemask_t *policy_nodes)
1491 info->root = RB_ROOT;
1492 spin_lock_init(&info->lock);
1494 if (policy != MPOL_DEFAULT) {
1495 struct mempolicy *newpol;
1497 /* Falls back to MPOL_DEFAULT on any error */
1498 newpol = mpol_new(policy, policy_nodes);
1499 if (!IS_ERR(newpol)) {
1500 /* Create pseudo-vma that contains just the policy */
1501 struct vm_area_struct pvma;
1503 memset(&pvma, 0, sizeof(struct vm_area_struct));
1504 /* Policy covers entire file */
1505 pvma.vm_end = TASK_SIZE;
1506 mpol_set_shared_policy(info, &pvma, newpol);
1512 int mpol_set_shared_policy(struct shared_policy *info,
1513 struct vm_area_struct *vma, struct mempolicy *npol)
1516 struct sp_node *new = NULL;
1517 unsigned long sz = vma_pages(vma);
1519 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1521 sz, npol? npol->policy : -1,
1522 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1525 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1529 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1531 kmem_cache_free(sn_cache, new);
1535 /* Free a backing policy store on inode delete. */
1536 void mpol_free_shared_policy(struct shared_policy *p)
1539 struct rb_node *next;
1541 if (!p->root.rb_node)
1543 spin_lock(&p->lock);
1544 next = rb_first(&p->root);
1546 n = rb_entry(next, struct sp_node, nd);
1547 next = rb_next(&n->nd);
1548 rb_erase(&n->nd, &p->root);
1549 mpol_free(n->policy);
1550 kmem_cache_free(sn_cache, n);
1552 spin_unlock(&p->lock);
1555 /* assumes fs == KERNEL_DS */
1556 void __init numa_policy_init(void)
1558 policy_cache = kmem_cache_create("numa_policy",
1559 sizeof(struct mempolicy),
1560 0, SLAB_PANIC, NULL, NULL);
1562 sn_cache = kmem_cache_create("shared_policy_node",
1563 sizeof(struct sp_node),
1564 0, SLAB_PANIC, NULL, NULL);
1566 /* Set interleaving policy for system init. This way not all
1567 the data structures allocated at system boot end up in node zero. */
1569 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1570 printk("numa_policy_init: interleaving failed\n");
1573 /* Reset policy of current process to default */
1574 void numa_default_policy(void)
1576 do_set_mempolicy(MPOL_DEFAULT, NULL);
1579 /* Migrate a policy to a different set of nodes */
1580 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1582 nodemask_t *mpolmask;
1587 mpolmask = &pol->cpuset_mems_allowed;
1588 if (nodes_equal(*mpolmask, *newmask))
1591 switch (pol->policy) {
1594 case MPOL_INTERLEAVE:
1595 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1597 *mpolmask = *newmask;
1598 current->il_next = node_remap(current->il_next,
1599 *mpolmask, *newmask);
1601 case MPOL_PREFERRED:
1602 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1603 *mpolmask, *newmask);
1604 *mpolmask = *newmask;
1609 struct zonelist *zonelist;
1612 for (z = pol->v.zonelist->zones; *z; z++)
1613 node_set((*z)->zone_pgdat->node_id, nodes);
1614 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1617 zonelist = bind_zonelist(&nodes);
1619 /* If no mem, then zonelist is NULL and we keep old zonelist.
1620 * If that old zonelist has no remaining mems_allowed nodes,
1621 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1625 /* Good - got mem - substitute new zonelist */
1626 kfree(pol->v.zonelist);
1627 pol->v.zonelist = zonelist;
1629 *mpolmask = *newmask;
1639 * Wrapper for mpol_rebind_policy() that just requires task
1640 * pointer, and updates task mempolicy.
1643 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1645 mpol_rebind_policy(tsk->mempolicy, new);
1649 * Rebind each vma in mm to new nodemask.
1651 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1654 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1656 struct vm_area_struct *vma;
1658 down_write(&mm->mmap_sem);
1659 for (vma = mm->mmap; vma; vma = vma->vm_next)
1660 mpol_rebind_policy(vma->vm_policy, new);
1661 up_write(&mm->mmap_sem);
1665 * Display pages allocated per node and memory policy via /proc.
1668 static const char *policy_types[] = { "default", "prefer", "bind",
1672 * Convert a mempolicy into a string.
1673 * Returns the number of characters in buffer (if positive)
1674 * or an error (negative)
1676 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1681 int mode = pol ? pol->policy : MPOL_DEFAULT;
1688 case MPOL_PREFERRED:
1690 node_set(pol->v.preferred_node, nodes);
1694 get_zonemask(pol, &nodes);
1697 case MPOL_INTERLEAVE:
1698 nodes = pol->v.nodes;
1706 l = strlen(policy_types[mode]);
1707 if (buffer + maxlen < p + l + 1)
1710 strcpy(p, policy_types[mode]);
1713 if (!nodes_empty(nodes)) {
1714 if (buffer + maxlen < p + 2)
1717 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1723 unsigned long pages;
1725 unsigned long active;
1726 unsigned long writeback;
1727 unsigned long mapcount_max;
1728 unsigned long dirty;
1729 unsigned long swapcache;
1730 unsigned long node[MAX_NUMNODES];
1733 static void gather_stats(struct page *page, void *private, int pte_dirty)
1735 struct numa_maps *md = private;
1736 int count = page_mapcount(page);
1739 if (pte_dirty || PageDirty(page))
1742 if (PageSwapCache(page))
1745 if (PageActive(page))
1748 if (PageWriteback(page))
1754 if (count > md->mapcount_max)
1755 md->mapcount_max = count;
1757 md->node[page_to_nid(page)]++;
1760 #ifdef CONFIG_HUGETLB_PAGE
1761 static void check_huge_range(struct vm_area_struct *vma,
1762 unsigned long start, unsigned long end,
1763 struct numa_maps *md)
1768 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1769 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1779 page = pte_page(pte);
1783 gather_stats(page, md, pte_dirty(*ptep));
1787 static inline void check_huge_range(struct vm_area_struct *vma,
1788 unsigned long start, unsigned long end,
1789 struct numa_maps *md)
1794 int show_numa_map(struct seq_file *m, void *v)
1796 struct task_struct *task = m->private;
1797 struct vm_area_struct *vma = v;
1798 struct numa_maps *md;
1799 struct file *file = vma->vm_file;
1800 struct mm_struct *mm = vma->vm_mm;
1807 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1811 mpol_to_str(buffer, sizeof(buffer),
1812 get_vma_policy(task, vma, vma->vm_start));
1814 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1817 seq_printf(m, " file=");
1818 seq_path(m, file->f_vfsmnt, file->f_dentry, "\n\t= ");
1819 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1820 seq_printf(m, " heap");
1821 } else if (vma->vm_start <= mm->start_stack &&
1822 vma->vm_end >= mm->start_stack) {
1823 seq_printf(m, " stack");
1826 if (is_vm_hugetlb_page(vma)) {
1827 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
1828 seq_printf(m, " huge");
1830 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1831 &node_online_map, MPOL_MF_STATS, md);
1838 seq_printf(m," anon=%lu",md->anon);
1841 seq_printf(m," dirty=%lu",md->dirty);
1843 if (md->pages != md->anon && md->pages != md->dirty)
1844 seq_printf(m, " mapped=%lu", md->pages);
1846 if (md->mapcount_max > 1)
1847 seq_printf(m, " mapmax=%lu", md->mapcount_max);
1850 seq_printf(m," swapcache=%lu", md->swapcache);
1852 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
1853 seq_printf(m," active=%lu", md->active);
1856 seq_printf(m," writeback=%lu", md->writeback);
1858 for_each_online_node(n)
1860 seq_printf(m, " N%d=%lu", n, md->node[n]);
1865 if (m->count < m->size)
1866 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;