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>
90 #include <linux/rmap.h>
92 #include <asm/tlbflush.h>
93 #include <asm/uaccess.h>
96 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
97 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
98 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
100 static struct kmem_cache *policy_cache;
101 static struct kmem_cache *sn_cache;
103 #define PDprintk(fmt...)
105 /* Highest zone. An specific allocation for a zone below that is not
107 int policy_zone = ZONE_DMA;
109 struct mempolicy default_policy = {
110 .refcnt = ATOMIC_INIT(1), /* never free it */
111 .policy = MPOL_DEFAULT,
114 /* Do sanity checking on a policy */
115 static int mpol_check_policy(int mode, nodemask_t *nodes)
117 int empty = nodes_empty(*nodes);
125 case MPOL_INTERLEAVE:
126 /* Preferred will only use the first bit, but allow
132 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
135 /* Generate a custom zonelist for the BIND policy. */
136 static struct zonelist *bind_zonelist(nodemask_t *nodes)
141 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
142 zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL);
146 /* First put in the highest zones from all nodes, then all the next
147 lower zones etc. Avoid empty zones because the memory allocator
148 doesn't like them. If you implement node hot removal you
150 for (k = policy_zone; k >= 0; k--) {
151 for_each_node_mask(nd, *nodes) {
152 struct zone *z = &NODE_DATA(nd)->node_zones[k];
153 if (z->present_pages > 0)
154 zl->zones[num++] = z;
157 zl->zones[num] = NULL;
161 /* Create a new policy */
162 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
164 struct mempolicy *policy;
166 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
167 if (mode == MPOL_DEFAULT)
169 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
171 return ERR_PTR(-ENOMEM);
172 atomic_set(&policy->refcnt, 1);
174 case MPOL_INTERLEAVE:
175 policy->v.nodes = *nodes;
176 if (nodes_weight(*nodes) == 0) {
177 kmem_cache_free(policy_cache, policy);
178 return ERR_PTR(-EINVAL);
182 policy->v.preferred_node = first_node(*nodes);
183 if (policy->v.preferred_node >= MAX_NUMNODES)
184 policy->v.preferred_node = -1;
187 policy->v.zonelist = bind_zonelist(nodes);
188 if (policy->v.zonelist == NULL) {
189 kmem_cache_free(policy_cache, policy);
190 return ERR_PTR(-ENOMEM);
194 policy->policy = mode;
195 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
199 static void gather_stats(struct page *, void *, int pte_dirty);
200 static void migrate_page_add(struct page *page, struct list_head *pagelist,
201 unsigned long flags);
203 /* Scan through pages checking if pages follow certain conditions. */
204 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
205 unsigned long addr, unsigned long end,
206 const nodemask_t *nodes, unsigned long flags,
213 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
218 if (!pte_present(*pte))
220 page = vm_normal_page(vma, addr, *pte);
224 * The check for PageReserved here is important to avoid
225 * handling zero pages and other pages that may have been
226 * marked special by the system.
228 * If the PageReserved would not be checked here then f.e.
229 * the location of the zero page could have an influence
230 * on MPOL_MF_STRICT, zero pages would be counted for
231 * the per node stats, and there would be useless attempts
232 * to put zero pages on the migration list.
234 if (PageReserved(page))
236 nid = page_to_nid(page);
237 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
240 if (flags & MPOL_MF_STATS)
241 gather_stats(page, private, pte_dirty(*pte));
242 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
243 migrate_page_add(page, private, flags);
246 } while (pte++, addr += PAGE_SIZE, addr != end);
247 pte_unmap_unlock(orig_pte, ptl);
251 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
252 unsigned long addr, unsigned long end,
253 const nodemask_t *nodes, unsigned long flags,
259 pmd = pmd_offset(pud, addr);
261 next = pmd_addr_end(addr, end);
262 if (pmd_none_or_clear_bad(pmd))
264 if (check_pte_range(vma, pmd, addr, next, nodes,
267 } while (pmd++, addr = next, addr != end);
271 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
272 unsigned long addr, unsigned long end,
273 const nodemask_t *nodes, unsigned long flags,
279 pud = pud_offset(pgd, addr);
281 next = pud_addr_end(addr, end);
282 if (pud_none_or_clear_bad(pud))
284 if (check_pmd_range(vma, pud, addr, next, nodes,
287 } while (pud++, addr = next, addr != end);
291 static inline int check_pgd_range(struct vm_area_struct *vma,
292 unsigned long addr, unsigned long end,
293 const nodemask_t *nodes, unsigned long flags,
299 pgd = pgd_offset(vma->vm_mm, addr);
301 next = pgd_addr_end(addr, end);
302 if (pgd_none_or_clear_bad(pgd))
304 if (check_pud_range(vma, pgd, addr, next, nodes,
307 } while (pgd++, addr = next, addr != end);
311 /* Check if a vma is migratable */
312 static inline int vma_migratable(struct vm_area_struct *vma)
314 if (vma->vm_flags & (
315 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED))
321 * Check if all pages in a range are on a set of nodes.
322 * If pagelist != NULL then isolate pages from the LRU and
323 * put them on the pagelist.
325 static struct vm_area_struct *
326 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
327 const nodemask_t *nodes, unsigned long flags, void *private)
330 struct vm_area_struct *first, *vma, *prev;
332 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
334 err = migrate_prep();
339 first = find_vma(mm, start);
341 return ERR_PTR(-EFAULT);
343 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
344 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
345 if (!vma->vm_next && vma->vm_end < end)
346 return ERR_PTR(-EFAULT);
347 if (prev && prev->vm_end < vma->vm_start)
348 return ERR_PTR(-EFAULT);
350 if (!is_vm_hugetlb_page(vma) &&
351 ((flags & MPOL_MF_STRICT) ||
352 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
353 vma_migratable(vma)))) {
354 unsigned long endvma = vma->vm_end;
358 if (vma->vm_start > start)
359 start = vma->vm_start;
360 err = check_pgd_range(vma, start, endvma, nodes,
363 first = ERR_PTR(err);
372 /* Apply policy to a single VMA */
373 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
376 struct mempolicy *old = vma->vm_policy;
378 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
379 vma->vm_start, vma->vm_end, vma->vm_pgoff,
380 vma->vm_ops, vma->vm_file,
381 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
383 if (vma->vm_ops && vma->vm_ops->set_policy)
384 err = vma->vm_ops->set_policy(vma, new);
387 vma->vm_policy = new;
393 /* Step 2: apply policy to a range and do splits. */
394 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
395 unsigned long end, struct mempolicy *new)
397 struct vm_area_struct *next;
401 for (; vma && vma->vm_start < end; vma = next) {
403 if (vma->vm_start < start)
404 err = split_vma(vma->vm_mm, vma, start, 1);
405 if (!err && vma->vm_end > end)
406 err = split_vma(vma->vm_mm, vma, end, 0);
408 err = policy_vma(vma, new);
415 static int contextualize_policy(int mode, nodemask_t *nodes)
420 cpuset_update_task_memory_state();
421 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
423 return mpol_check_policy(mode, nodes);
428 * Update task->flags PF_MEMPOLICY bit: set iff non-default
429 * mempolicy. Allows more rapid checking of this (combined perhaps
430 * with other PF_* flag bits) on memory allocation hot code paths.
432 * If called from outside this file, the task 'p' should -only- be
433 * a newly forked child not yet visible on the task list, because
434 * manipulating the task flags of a visible task is not safe.
436 * The above limitation is why this routine has the funny name
437 * mpol_fix_fork_child_flag().
439 * It is also safe to call this with a task pointer of current,
440 * which the static wrapper mpol_set_task_struct_flag() does,
441 * for use within this file.
444 void mpol_fix_fork_child_flag(struct task_struct *p)
447 p->flags |= PF_MEMPOLICY;
449 p->flags &= ~PF_MEMPOLICY;
452 static void mpol_set_task_struct_flag(void)
454 mpol_fix_fork_child_flag(current);
457 /* Set the process memory policy */
458 long do_set_mempolicy(int mode, nodemask_t *nodes)
460 struct mempolicy *new;
462 if (contextualize_policy(mode, nodes))
464 new = mpol_new(mode, nodes);
467 mpol_free(current->mempolicy);
468 current->mempolicy = new;
469 mpol_set_task_struct_flag();
470 if (new && new->policy == MPOL_INTERLEAVE)
471 current->il_next = first_node(new->v.nodes);
475 /* Fill a zone bitmap for a policy */
476 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
483 for (i = 0; p->v.zonelist->zones[i]; i++)
484 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
489 case MPOL_INTERLEAVE:
493 /* or use current node instead of online map? */
494 if (p->v.preferred_node < 0)
495 *nodes = node_online_map;
497 node_set(p->v.preferred_node, *nodes);
504 static int lookup_node(struct mm_struct *mm, unsigned long addr)
509 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
511 err = page_to_nid(p);
517 /* Retrieve NUMA policy */
518 long do_get_mempolicy(int *policy, nodemask_t *nmask,
519 unsigned long addr, unsigned long flags)
522 struct mm_struct *mm = current->mm;
523 struct vm_area_struct *vma = NULL;
524 struct mempolicy *pol = current->mempolicy;
526 cpuset_update_task_memory_state();
527 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
529 if (flags & MPOL_F_ADDR) {
530 down_read(&mm->mmap_sem);
531 vma = find_vma_intersection(mm, addr, addr+1);
533 up_read(&mm->mmap_sem);
536 if (vma->vm_ops && vma->vm_ops->get_policy)
537 pol = vma->vm_ops->get_policy(vma, addr);
539 pol = vma->vm_policy;
544 pol = &default_policy;
546 if (flags & MPOL_F_NODE) {
547 if (flags & MPOL_F_ADDR) {
548 err = lookup_node(mm, addr);
552 } else if (pol == current->mempolicy &&
553 pol->policy == MPOL_INTERLEAVE) {
554 *policy = current->il_next;
560 *policy = pol->policy;
563 up_read(¤t->mm->mmap_sem);
569 get_zonemask(pol, nmask);
573 up_read(¤t->mm->mmap_sem);
577 #ifdef CONFIG_MIGRATION
581 static void migrate_page_add(struct page *page, struct list_head *pagelist,
585 * Avoid migrating a page that is shared with others.
587 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
588 isolate_lru_page(page, pagelist);
591 static struct page *new_node_page(struct page *page, unsigned long node)
593 return alloc_pages_node(node, GFP_HIGHUSER, 0);
597 * Migrate pages from one node to a target node.
598 * Returns error or the number of pages not migrated.
600 int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags)
607 node_set(source, nmask);
609 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
610 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
612 if (!list_empty(&pagelist))
613 err = migrate_pages(&pagelist, new_node_page, dest);
619 * Move pages between the two nodesets so as to preserve the physical
620 * layout as much as possible.
622 * Returns the number of page that could not be moved.
624 int do_migrate_pages(struct mm_struct *mm,
625 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
632 down_read(&mm->mmap_sem);
635 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
636 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
637 * bit in 'tmp', and return that <source, dest> pair for migration.
638 * The pair of nodemasks 'to' and 'from' define the map.
640 * If no pair of bits is found that way, fallback to picking some
641 * pair of 'source' and 'dest' bits that are not the same. If the
642 * 'source' and 'dest' bits are the same, this represents a node
643 * that will be migrating to itself, so no pages need move.
645 * If no bits are left in 'tmp', or if all remaining bits left
646 * in 'tmp' correspond to the same bit in 'to', return false
647 * (nothing left to migrate).
649 * This lets us pick a pair of nodes to migrate between, such that
650 * if possible the dest node is not already occupied by some other
651 * source node, minimizing the risk of overloading the memory on a
652 * node that would happen if we migrated incoming memory to a node
653 * before migrating outgoing memory source that same node.
655 * A single scan of tmp is sufficient. As we go, we remember the
656 * most recent <s, d> pair that moved (s != d). If we find a pair
657 * that not only moved, but what's better, moved to an empty slot
658 * (d is not set in tmp), then we break out then, with that pair.
659 * Otherwise when we finish scannng from_tmp, we at least have the
660 * most recent <s, d> pair that moved. If we get all the way through
661 * the scan of tmp without finding any node that moved, much less
662 * moved to an empty node, then there is nothing left worth migrating.
666 while (!nodes_empty(tmp)) {
671 for_each_node_mask(s, tmp) {
672 d = node_remap(s, *from_nodes, *to_nodes);
676 source = s; /* Node moved. Memorize */
679 /* dest not in remaining from nodes? */
680 if (!node_isset(dest, tmp))
686 node_clear(source, tmp);
687 err = migrate_to_node(mm, source, dest, flags);
694 up_read(&mm->mmap_sem);
701 static struct page *new_vma_page(struct page *page, unsigned long private)
703 struct vm_area_struct *vma = (struct vm_area_struct *)private;
705 return alloc_page_vma(GFP_HIGHUSER, vma, page_address_in_vma(page, vma));
709 static void migrate_page_add(struct page *page, struct list_head *pagelist,
714 int do_migrate_pages(struct mm_struct *mm,
715 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
720 static struct page *new_vma_page(struct page *page, unsigned long private)
726 long do_mbind(unsigned long start, unsigned long len,
727 unsigned long mode, nodemask_t *nmask, unsigned long flags)
729 struct vm_area_struct *vma;
730 struct mm_struct *mm = current->mm;
731 struct mempolicy *new;
736 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
737 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
740 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
743 if (start & ~PAGE_MASK)
746 if (mode == MPOL_DEFAULT)
747 flags &= ~MPOL_MF_STRICT;
749 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
757 if (mpol_check_policy(mode, nmask))
760 new = mpol_new(mode, nmask);
765 * If we are using the default policy then operation
766 * on discontinuous address spaces is okay after all
769 flags |= MPOL_MF_DISCONTIG_OK;
771 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
772 mode,nodes_addr(nodes)[0]);
774 down_write(&mm->mmap_sem);
775 vma = check_range(mm, start, end, nmask,
776 flags | MPOL_MF_INVERT, &pagelist);
782 err = mbind_range(vma, start, end, new);
784 if (!list_empty(&pagelist))
785 nr_failed = migrate_pages(&pagelist, new_vma_page,
788 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
792 up_write(&mm->mmap_sem);
798 * User space interface with variable sized bitmaps for nodelists.
801 /* Copy a node mask from user space. */
802 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
803 unsigned long maxnode)
806 unsigned long nlongs;
807 unsigned long endmask;
811 if (maxnode == 0 || !nmask)
813 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
816 nlongs = BITS_TO_LONGS(maxnode);
817 if ((maxnode % BITS_PER_LONG) == 0)
820 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
822 /* When the user specified more nodes than supported just check
823 if the non supported part is all zero. */
824 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
825 if (nlongs > PAGE_SIZE/sizeof(long))
827 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
829 if (get_user(t, nmask + k))
831 if (k == nlongs - 1) {
837 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
841 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
843 nodes_addr(*nodes)[nlongs-1] &= endmask;
847 /* Copy a kernel node mask to user space */
848 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
851 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
852 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
855 if (copy > PAGE_SIZE)
857 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
861 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
864 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
866 unsigned long __user *nmask, unsigned long maxnode,
872 err = get_nodes(&nodes, nmask, maxnode);
875 return do_mbind(start, len, mode, &nodes, flags);
878 /* Set the process memory policy */
879 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
880 unsigned long maxnode)
885 if (mode < 0 || mode > MPOL_MAX)
887 err = get_nodes(&nodes, nmask, maxnode);
890 return do_set_mempolicy(mode, &nodes);
893 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
894 const unsigned long __user *old_nodes,
895 const unsigned long __user *new_nodes)
897 struct mm_struct *mm;
898 struct task_struct *task;
901 nodemask_t task_nodes;
904 err = get_nodes(&old, old_nodes, maxnode);
908 err = get_nodes(&new, new_nodes, maxnode);
912 /* Find the mm_struct */
913 read_lock(&tasklist_lock);
914 task = pid ? find_task_by_pid(pid) : current;
916 read_unlock(&tasklist_lock);
919 mm = get_task_mm(task);
920 read_unlock(&tasklist_lock);
926 * Check if this process has the right to modify the specified
927 * process. The right exists if the process has administrative
928 * capabilities, superuser privileges or the same
929 * userid as the target process.
931 if ((current->euid != task->suid) && (current->euid != task->uid) &&
932 (current->uid != task->suid) && (current->uid != task->uid) &&
933 !capable(CAP_SYS_NICE)) {
938 task_nodes = cpuset_mems_allowed(task);
939 /* Is the user allowed to access the target nodes? */
940 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
945 err = do_migrate_pages(mm, &old, &new,
946 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
953 /* Retrieve NUMA policy */
954 asmlinkage long sys_get_mempolicy(int __user *policy,
955 unsigned long __user *nmask,
956 unsigned long maxnode,
957 unsigned long addr, unsigned long flags)
962 if (nmask != NULL && maxnode < MAX_NUMNODES)
965 err = do_get_mempolicy(&pval, &nodes, addr, flags);
970 if (policy && put_user(pval, policy))
974 err = copy_nodes_to_user(nmask, maxnode, &nodes);
981 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
982 compat_ulong_t __user *nmask,
983 compat_ulong_t maxnode,
984 compat_ulong_t addr, compat_ulong_t flags)
987 unsigned long __user *nm = NULL;
988 unsigned long nr_bits, alloc_size;
989 DECLARE_BITMAP(bm, MAX_NUMNODES);
991 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
992 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
995 nm = compat_alloc_user_space(alloc_size);
997 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1000 err = copy_from_user(bm, nm, alloc_size);
1001 /* ensure entire bitmap is zeroed */
1002 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1003 err |= compat_put_bitmap(nmask, bm, nr_bits);
1009 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1010 compat_ulong_t maxnode)
1013 unsigned long __user *nm = NULL;
1014 unsigned long nr_bits, alloc_size;
1015 DECLARE_BITMAP(bm, MAX_NUMNODES);
1017 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1018 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1021 err = compat_get_bitmap(bm, nmask, nr_bits);
1022 nm = compat_alloc_user_space(alloc_size);
1023 err |= copy_to_user(nm, bm, alloc_size);
1029 return sys_set_mempolicy(mode, nm, nr_bits+1);
1032 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1033 compat_ulong_t mode, compat_ulong_t __user *nmask,
1034 compat_ulong_t maxnode, compat_ulong_t flags)
1037 unsigned long __user *nm = NULL;
1038 unsigned long nr_bits, alloc_size;
1041 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1042 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1045 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1046 nm = compat_alloc_user_space(alloc_size);
1047 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1053 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1058 /* Return effective policy for a VMA */
1059 static struct mempolicy * get_vma_policy(struct task_struct *task,
1060 struct vm_area_struct *vma, unsigned long addr)
1062 struct mempolicy *pol = task->mempolicy;
1065 if (vma->vm_ops && vma->vm_ops->get_policy)
1066 pol = vma->vm_ops->get_policy(vma, addr);
1067 else if (vma->vm_policy &&
1068 vma->vm_policy->policy != MPOL_DEFAULT)
1069 pol = vma->vm_policy;
1072 pol = &default_policy;
1076 /* Return a zonelist representing a mempolicy */
1077 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1081 switch (policy->policy) {
1082 case MPOL_PREFERRED:
1083 nd = policy->v.preferred_node;
1085 nd = numa_node_id();
1088 /* Lower zones don't get a policy applied */
1089 /* Careful: current->mems_allowed might have moved */
1090 if (gfp_zone(gfp) >= policy_zone)
1091 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
1092 return policy->v.zonelist;
1094 case MPOL_INTERLEAVE: /* should not happen */
1096 nd = numa_node_id();
1102 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
1105 /* Do dynamic interleaving for a process */
1106 static unsigned interleave_nodes(struct mempolicy *policy)
1109 struct task_struct *me = current;
1112 next = next_node(nid, policy->v.nodes);
1113 if (next >= MAX_NUMNODES)
1114 next = first_node(policy->v.nodes);
1120 * Depending on the memory policy provide a node from which to allocate the
1123 unsigned slab_node(struct mempolicy *policy)
1125 switch (policy->policy) {
1126 case MPOL_INTERLEAVE:
1127 return interleave_nodes(policy);
1131 * Follow bind policy behavior and start allocation at the
1134 return policy->v.zonelist->zones[0]->zone_pgdat->node_id;
1136 case MPOL_PREFERRED:
1137 if (policy->v.preferred_node >= 0)
1138 return policy->v.preferred_node;
1142 return numa_node_id();
1146 /* Do static interleaving for a VMA with known offset. */
1147 static unsigned offset_il_node(struct mempolicy *pol,
1148 struct vm_area_struct *vma, unsigned long off)
1150 unsigned nnodes = nodes_weight(pol->v.nodes);
1151 unsigned target = (unsigned)off % nnodes;
1157 nid = next_node(nid, pol->v.nodes);
1159 } while (c <= target);
1163 /* Determine a node number for interleave */
1164 static inline unsigned interleave_nid(struct mempolicy *pol,
1165 struct vm_area_struct *vma, unsigned long addr, int shift)
1170 off = vma->vm_pgoff;
1171 off += (addr - vma->vm_start) >> shift;
1172 return offset_il_node(pol, vma, off);
1174 return interleave_nodes(pol);
1177 #ifdef CONFIG_HUGETLBFS
1178 /* Return a zonelist suitable for a huge page allocation. */
1179 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1181 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1183 if (pol->policy == MPOL_INTERLEAVE) {
1186 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1187 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1189 return zonelist_policy(GFP_HIGHUSER, pol);
1193 /* Allocate a page in interleaved policy.
1194 Own path because it needs to do special accounting. */
1195 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1198 struct zonelist *zl;
1201 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1202 page = __alloc_pages(gfp, order, zl);
1203 if (page && page_zone(page) == zl->zones[0]) {
1204 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1211 * alloc_page_vma - Allocate a page for a VMA.
1214 * %GFP_USER user allocation.
1215 * %GFP_KERNEL kernel allocations,
1216 * %GFP_HIGHMEM highmem/user allocations,
1217 * %GFP_FS allocation should not call back into a file system.
1218 * %GFP_ATOMIC don't sleep.
1220 * @vma: Pointer to VMA or NULL if not available.
1221 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1223 * This function allocates a page from the kernel page pool and applies
1224 * a NUMA policy associated with the VMA or the current process.
1225 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1226 * mm_struct of the VMA to prevent it from going away. Should be used for
1227 * all allocations for pages that will be mapped into
1228 * user space. Returns NULL when no page can be allocated.
1230 * Should be called with the mm_sem of the vma hold.
1233 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1235 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1237 cpuset_update_task_memory_state();
1239 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1242 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1243 return alloc_page_interleave(gfp, 0, nid);
1245 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1249 * alloc_pages_current - Allocate pages.
1252 * %GFP_USER user allocation,
1253 * %GFP_KERNEL kernel allocation,
1254 * %GFP_HIGHMEM highmem allocation,
1255 * %GFP_FS don't call back into a file system.
1256 * %GFP_ATOMIC don't sleep.
1257 * @order: Power of two of allocation size in pages. 0 is a single page.
1259 * Allocate a page from the kernel page pool. When not in
1260 * interrupt context and apply the current process NUMA policy.
1261 * Returns NULL when no page can be allocated.
1263 * Don't call cpuset_update_task_memory_state() unless
1264 * 1) it's ok to take cpuset_sem (can WAIT), and
1265 * 2) allocating for current task (not interrupt).
1267 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1269 struct mempolicy *pol = current->mempolicy;
1271 if ((gfp & __GFP_WAIT) && !in_interrupt())
1272 cpuset_update_task_memory_state();
1273 if (!pol || in_interrupt())
1274 pol = &default_policy;
1275 if (pol->policy == MPOL_INTERLEAVE)
1276 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1277 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1279 EXPORT_SYMBOL(alloc_pages_current);
1282 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1283 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1284 * with the mems_allowed returned by cpuset_mems_allowed(). This
1285 * keeps mempolicies cpuset relative after its cpuset moves. See
1286 * further kernel/cpuset.c update_nodemask().
1288 void *cpuset_being_rebound;
1290 /* Slow path of a mempolicy copy */
1291 struct mempolicy *__mpol_copy(struct mempolicy *old)
1293 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1296 return ERR_PTR(-ENOMEM);
1297 if (current_cpuset_is_being_rebound()) {
1298 nodemask_t mems = cpuset_mems_allowed(current);
1299 mpol_rebind_policy(old, &mems);
1302 atomic_set(&new->refcnt, 1);
1303 if (new->policy == MPOL_BIND) {
1304 int sz = ksize(old->v.zonelist);
1305 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1306 if (!new->v.zonelist) {
1307 kmem_cache_free(policy_cache, new);
1308 return ERR_PTR(-ENOMEM);
1310 memcpy(new->v.zonelist, old->v.zonelist, sz);
1315 /* Slow path of a mempolicy comparison */
1316 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1320 if (a->policy != b->policy)
1322 switch (a->policy) {
1325 case MPOL_INTERLEAVE:
1326 return nodes_equal(a->v.nodes, b->v.nodes);
1327 case MPOL_PREFERRED:
1328 return a->v.preferred_node == b->v.preferred_node;
1331 for (i = 0; a->v.zonelist->zones[i]; i++)
1332 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1334 return b->v.zonelist->zones[i] == NULL;
1342 /* Slow path of a mpol destructor. */
1343 void __mpol_free(struct mempolicy *p)
1345 if (!atomic_dec_and_test(&p->refcnt))
1347 if (p->policy == MPOL_BIND)
1348 kfree(p->v.zonelist);
1349 p->policy = MPOL_DEFAULT;
1350 kmem_cache_free(policy_cache, p);
1354 * Shared memory backing store policy support.
1356 * Remember policies even when nobody has shared memory mapped.
1357 * The policies are kept in Red-Black tree linked from the inode.
1358 * They are protected by the sp->lock spinlock, which should be held
1359 * for any accesses to the tree.
1362 /* lookup first element intersecting start-end */
1363 /* Caller holds sp->lock */
1364 static struct sp_node *
1365 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1367 struct rb_node *n = sp->root.rb_node;
1370 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1372 if (start >= p->end)
1374 else if (end <= p->start)
1382 struct sp_node *w = NULL;
1383 struct rb_node *prev = rb_prev(n);
1386 w = rb_entry(prev, struct sp_node, nd);
1387 if (w->end <= start)
1391 return rb_entry(n, struct sp_node, nd);
1394 /* Insert a new shared policy into the list. */
1395 /* Caller holds sp->lock */
1396 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1398 struct rb_node **p = &sp->root.rb_node;
1399 struct rb_node *parent = NULL;
1404 nd = rb_entry(parent, struct sp_node, nd);
1405 if (new->start < nd->start)
1407 else if (new->end > nd->end)
1408 p = &(*p)->rb_right;
1412 rb_link_node(&new->nd, parent, p);
1413 rb_insert_color(&new->nd, &sp->root);
1414 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1415 new->policy ? new->policy->policy : 0);
1418 /* Find shared policy intersecting idx */
1420 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1422 struct mempolicy *pol = NULL;
1425 if (!sp->root.rb_node)
1427 spin_lock(&sp->lock);
1428 sn = sp_lookup(sp, idx, idx+1);
1430 mpol_get(sn->policy);
1433 spin_unlock(&sp->lock);
1437 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1439 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1440 rb_erase(&n->nd, &sp->root);
1441 mpol_free(n->policy);
1442 kmem_cache_free(sn_cache, n);
1446 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1448 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1459 /* Replace a policy range. */
1460 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1461 unsigned long end, struct sp_node *new)
1463 struct sp_node *n, *new2 = NULL;
1466 spin_lock(&sp->lock);
1467 n = sp_lookup(sp, start, end);
1468 /* Take care of old policies in the same range. */
1469 while (n && n->start < end) {
1470 struct rb_node *next = rb_next(&n->nd);
1471 if (n->start >= start) {
1477 /* Old policy spanning whole new range. */
1480 spin_unlock(&sp->lock);
1481 new2 = sp_alloc(end, n->end, n->policy);
1487 sp_insert(sp, new2);
1495 n = rb_entry(next, struct sp_node, nd);
1499 spin_unlock(&sp->lock);
1501 mpol_free(new2->policy);
1502 kmem_cache_free(sn_cache, new2);
1507 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1508 nodemask_t *policy_nodes)
1510 info->root = RB_ROOT;
1511 spin_lock_init(&info->lock);
1513 if (policy != MPOL_DEFAULT) {
1514 struct mempolicy *newpol;
1516 /* Falls back to MPOL_DEFAULT on any error */
1517 newpol = mpol_new(policy, policy_nodes);
1518 if (!IS_ERR(newpol)) {
1519 /* Create pseudo-vma that contains just the policy */
1520 struct vm_area_struct pvma;
1522 memset(&pvma, 0, sizeof(struct vm_area_struct));
1523 /* Policy covers entire file */
1524 pvma.vm_end = TASK_SIZE;
1525 mpol_set_shared_policy(info, &pvma, newpol);
1531 int mpol_set_shared_policy(struct shared_policy *info,
1532 struct vm_area_struct *vma, struct mempolicy *npol)
1535 struct sp_node *new = NULL;
1536 unsigned long sz = vma_pages(vma);
1538 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1540 sz, npol? npol->policy : -1,
1541 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1544 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1548 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1550 kmem_cache_free(sn_cache, new);
1554 /* Free a backing policy store on inode delete. */
1555 void mpol_free_shared_policy(struct shared_policy *p)
1558 struct rb_node *next;
1560 if (!p->root.rb_node)
1562 spin_lock(&p->lock);
1563 next = rb_first(&p->root);
1565 n = rb_entry(next, struct sp_node, nd);
1566 next = rb_next(&n->nd);
1567 rb_erase(&n->nd, &p->root);
1568 mpol_free(n->policy);
1569 kmem_cache_free(sn_cache, n);
1571 spin_unlock(&p->lock);
1574 /* assumes fs == KERNEL_DS */
1575 void __init numa_policy_init(void)
1577 policy_cache = kmem_cache_create("numa_policy",
1578 sizeof(struct mempolicy),
1579 0, SLAB_PANIC, NULL, NULL);
1581 sn_cache = kmem_cache_create("shared_policy_node",
1582 sizeof(struct sp_node),
1583 0, SLAB_PANIC, NULL, NULL);
1585 /* Set interleaving policy for system init. This way not all
1586 the data structures allocated at system boot end up in node zero. */
1588 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1589 printk("numa_policy_init: interleaving failed\n");
1592 /* Reset policy of current process to default */
1593 void numa_default_policy(void)
1595 do_set_mempolicy(MPOL_DEFAULT, NULL);
1598 /* Migrate a policy to a different set of nodes */
1599 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1601 nodemask_t *mpolmask;
1606 mpolmask = &pol->cpuset_mems_allowed;
1607 if (nodes_equal(*mpolmask, *newmask))
1610 switch (pol->policy) {
1613 case MPOL_INTERLEAVE:
1614 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1616 *mpolmask = *newmask;
1617 current->il_next = node_remap(current->il_next,
1618 *mpolmask, *newmask);
1620 case MPOL_PREFERRED:
1621 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1622 *mpolmask, *newmask);
1623 *mpolmask = *newmask;
1628 struct zonelist *zonelist;
1631 for (z = pol->v.zonelist->zones; *z; z++)
1632 node_set((*z)->zone_pgdat->node_id, nodes);
1633 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1636 zonelist = bind_zonelist(&nodes);
1638 /* If no mem, then zonelist is NULL and we keep old zonelist.
1639 * If that old zonelist has no remaining mems_allowed nodes,
1640 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1644 /* Good - got mem - substitute new zonelist */
1645 kfree(pol->v.zonelist);
1646 pol->v.zonelist = zonelist;
1648 *mpolmask = *newmask;
1658 * Wrapper for mpol_rebind_policy() that just requires task
1659 * pointer, and updates task mempolicy.
1662 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1664 mpol_rebind_policy(tsk->mempolicy, new);
1668 * Rebind each vma in mm to new nodemask.
1670 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1673 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1675 struct vm_area_struct *vma;
1677 down_write(&mm->mmap_sem);
1678 for (vma = mm->mmap; vma; vma = vma->vm_next)
1679 mpol_rebind_policy(vma->vm_policy, new);
1680 up_write(&mm->mmap_sem);
1684 * Display pages allocated per node and memory policy via /proc.
1687 static const char *policy_types[] = { "default", "prefer", "bind",
1691 * Convert a mempolicy into a string.
1692 * Returns the number of characters in buffer (if positive)
1693 * or an error (negative)
1695 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1700 int mode = pol ? pol->policy : MPOL_DEFAULT;
1707 case MPOL_PREFERRED:
1709 node_set(pol->v.preferred_node, nodes);
1713 get_zonemask(pol, &nodes);
1716 case MPOL_INTERLEAVE:
1717 nodes = pol->v.nodes;
1725 l = strlen(policy_types[mode]);
1726 if (buffer + maxlen < p + l + 1)
1729 strcpy(p, policy_types[mode]);
1732 if (!nodes_empty(nodes)) {
1733 if (buffer + maxlen < p + 2)
1736 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1742 unsigned long pages;
1744 unsigned long active;
1745 unsigned long writeback;
1746 unsigned long mapcount_max;
1747 unsigned long dirty;
1748 unsigned long swapcache;
1749 unsigned long node[MAX_NUMNODES];
1752 static void gather_stats(struct page *page, void *private, int pte_dirty)
1754 struct numa_maps *md = private;
1755 int count = page_mapcount(page);
1758 if (pte_dirty || PageDirty(page))
1761 if (PageSwapCache(page))
1764 if (PageActive(page))
1767 if (PageWriteback(page))
1773 if (count > md->mapcount_max)
1774 md->mapcount_max = count;
1776 md->node[page_to_nid(page)]++;
1779 #ifdef CONFIG_HUGETLB_PAGE
1780 static void check_huge_range(struct vm_area_struct *vma,
1781 unsigned long start, unsigned long end,
1782 struct numa_maps *md)
1787 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1788 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1798 page = pte_page(pte);
1802 gather_stats(page, md, pte_dirty(*ptep));
1806 static inline void check_huge_range(struct vm_area_struct *vma,
1807 unsigned long start, unsigned long end,
1808 struct numa_maps *md)
1813 int show_numa_map(struct seq_file *m, void *v)
1815 struct task_struct *task = m->private;
1816 struct vm_area_struct *vma = v;
1817 struct numa_maps *md;
1818 struct file *file = vma->vm_file;
1819 struct mm_struct *mm = vma->vm_mm;
1826 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1830 mpol_to_str(buffer, sizeof(buffer),
1831 get_vma_policy(task, vma, vma->vm_start));
1833 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1836 seq_printf(m, " file=");
1837 seq_path(m, file->f_vfsmnt, file->f_dentry, "\n\t= ");
1838 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1839 seq_printf(m, " heap");
1840 } else if (vma->vm_start <= mm->start_stack &&
1841 vma->vm_end >= mm->start_stack) {
1842 seq_printf(m, " stack");
1845 if (is_vm_hugetlb_page(vma)) {
1846 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
1847 seq_printf(m, " huge");
1849 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1850 &node_online_map, MPOL_MF_STATS, md);
1857 seq_printf(m," anon=%lu",md->anon);
1860 seq_printf(m," dirty=%lu",md->dirty);
1862 if (md->pages != md->anon && md->pages != md->dirty)
1863 seq_printf(m, " mapped=%lu", md->pages);
1865 if (md->mapcount_max > 1)
1866 seq_printf(m, " mapmax=%lu", md->mapcount_max);
1869 seq_printf(m," swapcache=%lu", md->swapcache);
1871 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
1872 seq_printf(m," active=%lu", md->active);
1875 seq_printf(m," writeback=%lu", md->writeback);
1877 for_each_online_node(n)
1879 seq_printf(m, " N%d=%lu", n, md->node[n]);
1884 if (m->count < m->size)
1885 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;