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>
90 #include <asm/tlbflush.h>
91 #include <asm/uaccess.h>
94 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
95 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
96 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
98 /* The number of pages to migrate per call to migrate_pages() */
99 #define MIGRATE_CHUNK_SIZE 256
101 static kmem_cache_t *policy_cache;
102 static kmem_cache_t *sn_cache;
104 #define PDprintk(fmt...)
106 /* Highest zone. An specific allocation for a zone below that is not
108 int policy_zone = ZONE_DMA;
110 struct mempolicy default_policy = {
111 .refcnt = ATOMIC_INIT(1), /* never free it */
112 .policy = MPOL_DEFAULT,
115 /* Do sanity checking on a policy */
116 static int mpol_check_policy(int mode, nodemask_t *nodes)
118 int empty = nodes_empty(*nodes);
126 case MPOL_INTERLEAVE:
127 /* Preferred will only use the first bit, but allow
133 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
136 /* Generate a custom zonelist for the BIND policy. */
137 static struct zonelist *bind_zonelist(nodemask_t *nodes)
142 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
143 zl = kmalloc(sizeof(struct zone *) * max, GFP_KERNEL);
147 /* First put in the highest zones from all nodes, then all the next
148 lower zones etc. Avoid empty zones because the memory allocator
149 doesn't like them. If you implement node hot removal you
151 for (k = policy_zone; k >= 0; k--) {
152 for_each_node_mask(nd, *nodes) {
153 struct zone *z = &NODE_DATA(nd)->node_zones[k];
154 if (z->present_pages > 0)
155 zl->zones[num++] = z;
158 zl->zones[num] = NULL;
162 /* Create a new policy */
163 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
165 struct mempolicy *policy;
167 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
168 if (mode == MPOL_DEFAULT)
170 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
172 return ERR_PTR(-ENOMEM);
173 atomic_set(&policy->refcnt, 1);
175 case MPOL_INTERLEAVE:
176 policy->v.nodes = *nodes;
177 if (nodes_weight(*nodes) == 0) {
178 kmem_cache_free(policy_cache, policy);
179 return ERR_PTR(-EINVAL);
183 policy->v.preferred_node = first_node(*nodes);
184 if (policy->v.preferred_node >= MAX_NUMNODES)
185 policy->v.preferred_node = -1;
188 policy->v.zonelist = bind_zonelist(nodes);
189 if (policy->v.zonelist == NULL) {
190 kmem_cache_free(policy_cache, policy);
191 return ERR_PTR(-ENOMEM);
195 policy->policy = mode;
196 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
200 static void gather_stats(struct page *, void *);
201 static void migrate_page_add(struct page *page, struct list_head *pagelist,
202 unsigned long flags);
204 /* Scan through pages checking if pages follow certain conditions. */
205 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
206 unsigned long addr, unsigned long end,
207 const nodemask_t *nodes, unsigned long flags,
214 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
219 if (!pte_present(*pte))
221 page = vm_normal_page(vma, addr, *pte);
225 * The check for PageReserved here is important to avoid
226 * handling zero pages and other pages that may have been
227 * marked special by the system.
229 * If the PageReserved would not be checked here then f.e.
230 * the location of the zero page could have an influence
231 * on MPOL_MF_STRICT, zero pages would be counted for
232 * the per node stats, and there would be useless attempts
233 * to put zero pages on the migration list.
235 if (PageReserved(page))
237 nid = page_to_nid(page);
238 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
241 if (flags & MPOL_MF_STATS)
242 gather_stats(page, private);
243 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
244 migrate_page_add(page, private, flags);
247 } while (pte++, addr += PAGE_SIZE, addr != end);
248 pte_unmap_unlock(orig_pte, ptl);
252 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
253 unsigned long addr, unsigned long end,
254 const nodemask_t *nodes, unsigned long flags,
260 pmd = pmd_offset(pud, addr);
262 next = pmd_addr_end(addr, end);
263 if (pmd_none_or_clear_bad(pmd))
265 if (check_pte_range(vma, pmd, addr, next, nodes,
268 } while (pmd++, addr = next, addr != end);
272 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
273 unsigned long addr, unsigned long end,
274 const nodemask_t *nodes, unsigned long flags,
280 pud = pud_offset(pgd, addr);
282 next = pud_addr_end(addr, end);
283 if (pud_none_or_clear_bad(pud))
285 if (check_pmd_range(vma, pud, addr, next, nodes,
288 } while (pud++, addr = next, addr != end);
292 static inline int check_pgd_range(struct vm_area_struct *vma,
293 unsigned long addr, unsigned long end,
294 const nodemask_t *nodes, unsigned long flags,
300 pgd = pgd_offset(vma->vm_mm, addr);
302 next = pgd_addr_end(addr, end);
303 if (pgd_none_or_clear_bad(pgd))
305 if (check_pud_range(vma, pgd, addr, next, nodes,
308 } while (pgd++, addr = next, addr != end);
312 /* Check if a vma is migratable */
313 static inline int vma_migratable(struct vm_area_struct *vma)
315 if (vma->vm_flags & (
316 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED))
322 * Check if all pages in a range are on a set of nodes.
323 * If pagelist != NULL then isolate pages from the LRU and
324 * put them on the pagelist.
326 static struct vm_area_struct *
327 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
328 const nodemask_t *nodes, unsigned long flags, void *private)
331 struct vm_area_struct *first, *vma, *prev;
333 /* Clear the LRU lists so pages can be isolated */
334 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
337 first = find_vma(mm, start);
339 return ERR_PTR(-EFAULT);
341 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
342 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
343 if (!vma->vm_next && vma->vm_end < end)
344 return ERR_PTR(-EFAULT);
345 if (prev && prev->vm_end < vma->vm_start)
346 return ERR_PTR(-EFAULT);
348 if (!is_vm_hugetlb_page(vma) &&
349 ((flags & MPOL_MF_STRICT) ||
350 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
351 vma_migratable(vma)))) {
352 unsigned long endvma = vma->vm_end;
356 if (vma->vm_start > start)
357 start = vma->vm_start;
358 err = check_pgd_range(vma, start, endvma, nodes,
361 first = ERR_PTR(err);
370 /* Apply policy to a single VMA */
371 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
374 struct mempolicy *old = vma->vm_policy;
376 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
377 vma->vm_start, vma->vm_end, vma->vm_pgoff,
378 vma->vm_ops, vma->vm_file,
379 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
381 if (vma->vm_ops && vma->vm_ops->set_policy)
382 err = vma->vm_ops->set_policy(vma, new);
385 vma->vm_policy = new;
391 /* Step 2: apply policy to a range and do splits. */
392 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
393 unsigned long end, struct mempolicy *new)
395 struct vm_area_struct *next;
399 for (; vma && vma->vm_start < end; vma = next) {
401 if (vma->vm_start < start)
402 err = split_vma(vma->vm_mm, vma, start, 1);
403 if (!err && vma->vm_end > end)
404 err = split_vma(vma->vm_mm, vma, end, 0);
406 err = policy_vma(vma, new);
413 static int contextualize_policy(int mode, nodemask_t *nodes)
418 cpuset_update_task_memory_state();
419 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
421 return mpol_check_policy(mode, nodes);
424 /* Set the process memory policy */
425 long do_set_mempolicy(int mode, nodemask_t *nodes)
427 struct mempolicy *new;
429 if (contextualize_policy(mode, nodes))
431 new = mpol_new(mode, nodes);
434 mpol_free(current->mempolicy);
435 current->mempolicy = new;
436 if (new && new->policy == MPOL_INTERLEAVE)
437 current->il_next = first_node(new->v.nodes);
441 /* Fill a zone bitmap for a policy */
442 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
449 for (i = 0; p->v.zonelist->zones[i]; i++)
450 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
455 case MPOL_INTERLEAVE:
459 /* or use current node instead of online map? */
460 if (p->v.preferred_node < 0)
461 *nodes = node_online_map;
463 node_set(p->v.preferred_node, *nodes);
470 static int lookup_node(struct mm_struct *mm, unsigned long addr)
475 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
477 err = page_to_nid(p);
483 /* Retrieve NUMA policy */
484 long do_get_mempolicy(int *policy, nodemask_t *nmask,
485 unsigned long addr, unsigned long flags)
488 struct mm_struct *mm = current->mm;
489 struct vm_area_struct *vma = NULL;
490 struct mempolicy *pol = current->mempolicy;
492 cpuset_update_task_memory_state();
493 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
495 if (flags & MPOL_F_ADDR) {
496 down_read(&mm->mmap_sem);
497 vma = find_vma_intersection(mm, addr, addr+1);
499 up_read(&mm->mmap_sem);
502 if (vma->vm_ops && vma->vm_ops->get_policy)
503 pol = vma->vm_ops->get_policy(vma, addr);
505 pol = vma->vm_policy;
510 pol = &default_policy;
512 if (flags & MPOL_F_NODE) {
513 if (flags & MPOL_F_ADDR) {
514 err = lookup_node(mm, addr);
518 } else if (pol == current->mempolicy &&
519 pol->policy == MPOL_INTERLEAVE) {
520 *policy = current->il_next;
526 *policy = pol->policy;
529 up_read(¤t->mm->mmap_sem);
535 get_zonemask(pol, nmask);
539 up_read(¤t->mm->mmap_sem);
547 static void migrate_page_add(struct page *page, struct list_head *pagelist,
551 * Avoid migrating a page that is shared with others.
553 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
554 if (isolate_lru_page(page))
555 list_add(&page->lru, pagelist);
560 * Migrate the list 'pagelist' of pages to a certain destination.
562 * Specify destination with either non-NULL vma or dest_node >= 0
563 * Return the number of pages not migrated or error code
565 static int migrate_pages_to(struct list_head *pagelist,
566 struct vm_area_struct *vma, int dest)
578 list_for_each(p, pagelist) {
580 page = alloc_page_vma(GFP_HIGHUSER, vma, vma->vm_start);
582 page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
588 list_add(&page->lru, &newlist);
590 if (nr_pages > MIGRATE_CHUNK_SIZE)
593 err = migrate_pages(pagelist, &newlist, &moved, &failed);
595 putback_lru_pages(&moved); /* Call release pages instead ?? */
597 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
600 /* Return leftover allocated pages */
601 while (!list_empty(&newlist)) {
602 page = list_entry(newlist.next, struct page, lru);
603 list_del(&page->lru);
606 list_splice(&failed, pagelist);
610 /* Calculate number of leftover pages */
612 list_for_each(p, pagelist)
618 * Migrate pages from one node to a target node.
619 * Returns error or the number of pages not migrated.
621 int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags)
628 node_set(source, nmask);
630 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
631 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
633 if (!list_empty(&pagelist)) {
634 err = migrate_pages_to(&pagelist, NULL, dest);
635 if (!list_empty(&pagelist))
636 putback_lru_pages(&pagelist);
642 * Move pages between the two nodesets so as to preserve the physical
643 * layout as much as possible.
645 * Returns the number of page that could not be moved.
647 int do_migrate_pages(struct mm_struct *mm,
648 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
655 down_read(&mm->mmap_sem);
658 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
659 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
660 * bit in 'tmp', and return that <source, dest> pair for migration.
661 * The pair of nodemasks 'to' and 'from' define the map.
663 * If no pair of bits is found that way, fallback to picking some
664 * pair of 'source' and 'dest' bits that are not the same. If the
665 * 'source' and 'dest' bits are the same, this represents a node
666 * that will be migrating to itself, so no pages need move.
668 * If no bits are left in 'tmp', or if all remaining bits left
669 * in 'tmp' correspond to the same bit in 'to', return false
670 * (nothing left to migrate).
672 * This lets us pick a pair of nodes to migrate between, such that
673 * if possible the dest node is not already occupied by some other
674 * source node, minimizing the risk of overloading the memory on a
675 * node that would happen if we migrated incoming memory to a node
676 * before migrating outgoing memory source that same node.
678 * A single scan of tmp is sufficient. As we go, we remember the
679 * most recent <s, d> pair that moved (s != d). If we find a pair
680 * that not only moved, but what's better, moved to an empty slot
681 * (d is not set in tmp), then we break out then, with that pair.
682 * Otherwise when we finish scannng from_tmp, we at least have the
683 * most recent <s, d> pair that moved. If we get all the way through
684 * the scan of tmp without finding any node that moved, much less
685 * moved to an empty node, then there is nothing left worth migrating.
689 while (!nodes_empty(tmp)) {
694 for_each_node_mask(s, tmp) {
695 d = node_remap(s, *from_nodes, *to_nodes);
699 source = s; /* Node moved. Memorize */
702 /* dest not in remaining from nodes? */
703 if (!node_isset(dest, tmp))
709 node_clear(source, tmp);
710 err = migrate_to_node(mm, source, dest, flags);
717 up_read(&mm->mmap_sem);
723 long do_mbind(unsigned long start, unsigned long len,
724 unsigned long mode, nodemask_t *nmask, unsigned long flags)
726 struct vm_area_struct *vma;
727 struct mm_struct *mm = current->mm;
728 struct mempolicy *new;
733 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
734 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
737 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
740 if (start & ~PAGE_MASK)
743 if (mode == MPOL_DEFAULT)
744 flags &= ~MPOL_MF_STRICT;
746 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
754 if (mpol_check_policy(mode, nmask))
757 new = mpol_new(mode, nmask);
762 * If we are using the default policy then operation
763 * on discontinuous address spaces is okay after all
766 flags |= MPOL_MF_DISCONTIG_OK;
768 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
769 mode,nodes_addr(nodes)[0]);
771 down_write(&mm->mmap_sem);
772 vma = check_range(mm, start, end, nmask,
773 flags | MPOL_MF_INVERT, &pagelist);
779 err = mbind_range(vma, start, end, new);
781 if (!list_empty(&pagelist))
782 nr_failed = migrate_pages_to(&pagelist, vma, -1);
784 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
787 if (!list_empty(&pagelist))
788 putback_lru_pages(&pagelist);
790 up_write(&mm->mmap_sem);
796 * User space interface with variable sized bitmaps for nodelists.
799 /* Copy a node mask from user space. */
800 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
801 unsigned long maxnode)
804 unsigned long nlongs;
805 unsigned long endmask;
809 if (maxnode == 0 || !nmask)
811 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
814 nlongs = BITS_TO_LONGS(maxnode);
815 if ((maxnode % BITS_PER_LONG) == 0)
818 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
820 /* When the user specified more nodes than supported just check
821 if the non supported part is all zero. */
822 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
823 if (nlongs > PAGE_SIZE/sizeof(long))
825 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
827 if (get_user(t, nmask + k))
829 if (k == nlongs - 1) {
835 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
839 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
841 nodes_addr(*nodes)[nlongs-1] &= endmask;
845 /* Copy a kernel node mask to user space */
846 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
849 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
850 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
853 if (copy > PAGE_SIZE)
855 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
859 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
862 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
864 unsigned long __user *nmask, unsigned long maxnode,
870 err = get_nodes(&nodes, nmask, maxnode);
873 return do_mbind(start, len, mode, &nodes, flags);
876 /* Set the process memory policy */
877 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
878 unsigned long maxnode)
883 if (mode < 0 || mode > MPOL_MAX)
885 err = get_nodes(&nodes, nmask, maxnode);
888 return do_set_mempolicy(mode, &nodes);
891 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
892 const unsigned long __user *old_nodes,
893 const unsigned long __user *new_nodes)
895 struct mm_struct *mm;
896 struct task_struct *task;
899 nodemask_t task_nodes;
902 err = get_nodes(&old, old_nodes, maxnode);
906 err = get_nodes(&new, new_nodes, maxnode);
910 /* Find the mm_struct */
911 read_lock(&tasklist_lock);
912 task = pid ? find_task_by_pid(pid) : current;
914 read_unlock(&tasklist_lock);
917 mm = get_task_mm(task);
918 read_unlock(&tasklist_lock);
924 * Check if this process has the right to modify the specified
925 * process. The right exists if the process has administrative
926 * capabilities, superuser priviledges or the same
927 * userid as the target process.
929 if ((current->euid != task->suid) && (current->euid != task->uid) &&
930 (current->uid != task->suid) && (current->uid != task->uid) &&
931 !capable(CAP_SYS_ADMIN)) {
936 task_nodes = cpuset_mems_allowed(task);
937 /* Is the user allowed to access the target nodes? */
938 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
943 err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
950 /* Retrieve NUMA policy */
951 asmlinkage long sys_get_mempolicy(int __user *policy,
952 unsigned long __user *nmask,
953 unsigned long maxnode,
954 unsigned long addr, unsigned long flags)
959 if (nmask != NULL && maxnode < MAX_NUMNODES)
962 err = do_get_mempolicy(&pval, &nodes, addr, flags);
967 if (policy && put_user(pval, policy))
971 err = copy_nodes_to_user(nmask, maxnode, &nodes);
978 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
979 compat_ulong_t __user *nmask,
980 compat_ulong_t maxnode,
981 compat_ulong_t addr, compat_ulong_t flags)
984 unsigned long __user *nm = NULL;
985 unsigned long nr_bits, alloc_size;
986 DECLARE_BITMAP(bm, MAX_NUMNODES);
988 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
989 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
992 nm = compat_alloc_user_space(alloc_size);
994 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
997 err = copy_from_user(bm, nm, alloc_size);
998 /* ensure entire bitmap is zeroed */
999 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1000 err |= compat_put_bitmap(nmask, bm, nr_bits);
1006 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1007 compat_ulong_t maxnode)
1010 unsigned long __user *nm = NULL;
1011 unsigned long nr_bits, alloc_size;
1012 DECLARE_BITMAP(bm, MAX_NUMNODES);
1014 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1015 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1018 err = compat_get_bitmap(bm, nmask, nr_bits);
1019 nm = compat_alloc_user_space(alloc_size);
1020 err |= copy_to_user(nm, bm, alloc_size);
1026 return sys_set_mempolicy(mode, nm, nr_bits+1);
1029 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1030 compat_ulong_t mode, compat_ulong_t __user *nmask,
1031 compat_ulong_t maxnode, compat_ulong_t flags)
1034 unsigned long __user *nm = NULL;
1035 unsigned long nr_bits, alloc_size;
1038 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1039 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1042 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1043 nm = compat_alloc_user_space(alloc_size);
1044 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1050 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1055 /* Return effective policy for a VMA */
1056 static struct mempolicy * get_vma_policy(struct task_struct *task,
1057 struct vm_area_struct *vma, unsigned long addr)
1059 struct mempolicy *pol = task->mempolicy;
1062 if (vma->vm_ops && vma->vm_ops->get_policy)
1063 pol = vma->vm_ops->get_policy(vma, addr);
1064 else if (vma->vm_policy &&
1065 vma->vm_policy->policy != MPOL_DEFAULT)
1066 pol = vma->vm_policy;
1069 pol = &default_policy;
1073 /* Return a zonelist representing a mempolicy */
1074 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1078 switch (policy->policy) {
1079 case MPOL_PREFERRED:
1080 nd = policy->v.preferred_node;
1082 nd = numa_node_id();
1085 /* Lower zones don't get a policy applied */
1086 /* Careful: current->mems_allowed might have moved */
1087 if (gfp_zone(gfp) >= policy_zone)
1088 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
1089 return policy->v.zonelist;
1091 case MPOL_INTERLEAVE: /* should not happen */
1093 nd = numa_node_id();
1099 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
1102 /* Do dynamic interleaving for a process */
1103 static unsigned interleave_nodes(struct mempolicy *policy)
1106 struct task_struct *me = current;
1109 next = next_node(nid, policy->v.nodes);
1110 if (next >= MAX_NUMNODES)
1111 next = first_node(policy->v.nodes);
1117 * Depending on the memory policy provide a node from which to allocate the
1120 unsigned slab_node(struct mempolicy *policy)
1122 switch (policy->policy) {
1123 case MPOL_INTERLEAVE:
1124 return interleave_nodes(policy);
1128 * Follow bind policy behavior and start allocation at the
1131 return policy->v.zonelist->zones[0]->zone_pgdat->node_id;
1133 case MPOL_PREFERRED:
1134 if (policy->v.preferred_node >= 0)
1135 return policy->v.preferred_node;
1139 return numa_node_id();
1143 /* Do static interleaving for a VMA with known offset. */
1144 static unsigned offset_il_node(struct mempolicy *pol,
1145 struct vm_area_struct *vma, unsigned long off)
1147 unsigned nnodes = nodes_weight(pol->v.nodes);
1148 unsigned target = (unsigned)off % nnodes;
1154 nid = next_node(nid, pol->v.nodes);
1156 } while (c <= target);
1160 /* Determine a node number for interleave */
1161 static inline unsigned interleave_nid(struct mempolicy *pol,
1162 struct vm_area_struct *vma, unsigned long addr, int shift)
1167 off = vma->vm_pgoff;
1168 off += (addr - vma->vm_start) >> shift;
1169 return offset_il_node(pol, vma, off);
1171 return interleave_nodes(pol);
1174 #ifdef CONFIG_HUGETLBFS
1175 /* Return a zonelist suitable for a huge page allocation. */
1176 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1178 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1180 if (pol->policy == MPOL_INTERLEAVE) {
1183 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1184 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1186 return zonelist_policy(GFP_HIGHUSER, pol);
1190 /* Allocate a page in interleaved policy.
1191 Own path because it needs to do special accounting. */
1192 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1195 struct zonelist *zl;
1198 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1199 page = __alloc_pages(gfp, order, zl);
1200 if (page && page_zone(page) == zl->zones[0]) {
1201 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1208 * alloc_page_vma - Allocate a page for a VMA.
1211 * %GFP_USER user allocation.
1212 * %GFP_KERNEL kernel allocations,
1213 * %GFP_HIGHMEM highmem/user allocations,
1214 * %GFP_FS allocation should not call back into a file system.
1215 * %GFP_ATOMIC don't sleep.
1217 * @vma: Pointer to VMA or NULL if not available.
1218 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1220 * This function allocates a page from the kernel page pool and applies
1221 * a NUMA policy associated with the VMA or the current process.
1222 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1223 * mm_struct of the VMA to prevent it from going away. Should be used for
1224 * all allocations for pages that will be mapped into
1225 * user space. Returns NULL when no page can be allocated.
1227 * Should be called with the mm_sem of the vma hold.
1230 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1232 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1234 cpuset_update_task_memory_state();
1236 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1239 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1240 return alloc_page_interleave(gfp, 0, nid);
1242 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1246 * alloc_pages_current - Allocate pages.
1249 * %GFP_USER user allocation,
1250 * %GFP_KERNEL kernel allocation,
1251 * %GFP_HIGHMEM highmem allocation,
1252 * %GFP_FS don't call back into a file system.
1253 * %GFP_ATOMIC don't sleep.
1254 * @order: Power of two of allocation size in pages. 0 is a single page.
1256 * Allocate a page from the kernel page pool. When not in
1257 * interrupt context and apply the current process NUMA policy.
1258 * Returns NULL when no page can be allocated.
1260 * Don't call cpuset_update_task_memory_state() unless
1261 * 1) it's ok to take cpuset_sem (can WAIT), and
1262 * 2) allocating for current task (not interrupt).
1264 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1266 struct mempolicy *pol = current->mempolicy;
1268 if ((gfp & __GFP_WAIT) && !in_interrupt())
1269 cpuset_update_task_memory_state();
1270 if (!pol || in_interrupt())
1271 pol = &default_policy;
1272 if (pol->policy == MPOL_INTERLEAVE)
1273 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1274 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1276 EXPORT_SYMBOL(alloc_pages_current);
1279 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1280 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1281 * with the mems_allowed returned by cpuset_mems_allowed(). This
1282 * keeps mempolicies cpuset relative after its cpuset moves. See
1283 * further kernel/cpuset.c update_nodemask().
1285 void *cpuset_being_rebound;
1287 /* Slow path of a mempolicy copy */
1288 struct mempolicy *__mpol_copy(struct mempolicy *old)
1290 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1293 return ERR_PTR(-ENOMEM);
1294 if (current_cpuset_is_being_rebound()) {
1295 nodemask_t mems = cpuset_mems_allowed(current);
1296 mpol_rebind_policy(old, &mems);
1299 atomic_set(&new->refcnt, 1);
1300 if (new->policy == MPOL_BIND) {
1301 int sz = ksize(old->v.zonelist);
1302 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1303 if (!new->v.zonelist) {
1304 kmem_cache_free(policy_cache, new);
1305 return ERR_PTR(-ENOMEM);
1307 memcpy(new->v.zonelist, old->v.zonelist, sz);
1312 /* Slow path of a mempolicy comparison */
1313 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1317 if (a->policy != b->policy)
1319 switch (a->policy) {
1322 case MPOL_INTERLEAVE:
1323 return nodes_equal(a->v.nodes, b->v.nodes);
1324 case MPOL_PREFERRED:
1325 return a->v.preferred_node == b->v.preferred_node;
1328 for (i = 0; a->v.zonelist->zones[i]; i++)
1329 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1331 return b->v.zonelist->zones[i] == NULL;
1339 /* Slow path of a mpol destructor. */
1340 void __mpol_free(struct mempolicy *p)
1342 if (!atomic_dec_and_test(&p->refcnt))
1344 if (p->policy == MPOL_BIND)
1345 kfree(p->v.zonelist);
1346 p->policy = MPOL_DEFAULT;
1347 kmem_cache_free(policy_cache, p);
1351 * Shared memory backing store policy support.
1353 * Remember policies even when nobody has shared memory mapped.
1354 * The policies are kept in Red-Black tree linked from the inode.
1355 * They are protected by the sp->lock spinlock, which should be held
1356 * for any accesses to the tree.
1359 /* lookup first element intersecting start-end */
1360 /* Caller holds sp->lock */
1361 static struct sp_node *
1362 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1364 struct rb_node *n = sp->root.rb_node;
1367 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1369 if (start >= p->end)
1371 else if (end <= p->start)
1379 struct sp_node *w = NULL;
1380 struct rb_node *prev = rb_prev(n);
1383 w = rb_entry(prev, struct sp_node, nd);
1384 if (w->end <= start)
1388 return rb_entry(n, struct sp_node, nd);
1391 /* Insert a new shared policy into the list. */
1392 /* Caller holds sp->lock */
1393 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1395 struct rb_node **p = &sp->root.rb_node;
1396 struct rb_node *parent = NULL;
1401 nd = rb_entry(parent, struct sp_node, nd);
1402 if (new->start < nd->start)
1404 else if (new->end > nd->end)
1405 p = &(*p)->rb_right;
1409 rb_link_node(&new->nd, parent, p);
1410 rb_insert_color(&new->nd, &sp->root);
1411 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1412 new->policy ? new->policy->policy : 0);
1415 /* Find shared policy intersecting idx */
1417 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1419 struct mempolicy *pol = NULL;
1422 if (!sp->root.rb_node)
1424 spin_lock(&sp->lock);
1425 sn = sp_lookup(sp, idx, idx+1);
1427 mpol_get(sn->policy);
1430 spin_unlock(&sp->lock);
1434 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1436 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1437 rb_erase(&n->nd, &sp->root);
1438 mpol_free(n->policy);
1439 kmem_cache_free(sn_cache, n);
1443 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1445 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1456 /* Replace a policy range. */
1457 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1458 unsigned long end, struct sp_node *new)
1460 struct sp_node *n, *new2 = NULL;
1463 spin_lock(&sp->lock);
1464 n = sp_lookup(sp, start, end);
1465 /* Take care of old policies in the same range. */
1466 while (n && n->start < end) {
1467 struct rb_node *next = rb_next(&n->nd);
1468 if (n->start >= start) {
1474 /* Old policy spanning whole new range. */
1477 spin_unlock(&sp->lock);
1478 new2 = sp_alloc(end, n->end, n->policy);
1484 sp_insert(sp, new2);
1492 n = rb_entry(next, struct sp_node, nd);
1496 spin_unlock(&sp->lock);
1498 mpol_free(new2->policy);
1499 kmem_cache_free(sn_cache, new2);
1504 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1505 nodemask_t *policy_nodes)
1507 info->root = RB_ROOT;
1508 spin_lock_init(&info->lock);
1510 if (policy != MPOL_DEFAULT) {
1511 struct mempolicy *newpol;
1513 /* Falls back to MPOL_DEFAULT on any error */
1514 newpol = mpol_new(policy, policy_nodes);
1515 if (!IS_ERR(newpol)) {
1516 /* Create pseudo-vma that contains just the policy */
1517 struct vm_area_struct pvma;
1519 memset(&pvma, 0, sizeof(struct vm_area_struct));
1520 /* Policy covers entire file */
1521 pvma.vm_end = TASK_SIZE;
1522 mpol_set_shared_policy(info, &pvma, newpol);
1528 int mpol_set_shared_policy(struct shared_policy *info,
1529 struct vm_area_struct *vma, struct mempolicy *npol)
1532 struct sp_node *new = NULL;
1533 unsigned long sz = vma_pages(vma);
1535 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1537 sz, npol? npol->policy : -1,
1538 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1541 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1545 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1547 kmem_cache_free(sn_cache, new);
1551 /* Free a backing policy store on inode delete. */
1552 void mpol_free_shared_policy(struct shared_policy *p)
1555 struct rb_node *next;
1557 if (!p->root.rb_node)
1559 spin_lock(&p->lock);
1560 next = rb_first(&p->root);
1562 n = rb_entry(next, struct sp_node, nd);
1563 next = rb_next(&n->nd);
1564 rb_erase(&n->nd, &p->root);
1565 mpol_free(n->policy);
1566 kmem_cache_free(sn_cache, n);
1568 spin_unlock(&p->lock);
1571 /* assumes fs == KERNEL_DS */
1572 void __init numa_policy_init(void)
1574 policy_cache = kmem_cache_create("numa_policy",
1575 sizeof(struct mempolicy),
1576 0, SLAB_PANIC, NULL, NULL);
1578 sn_cache = kmem_cache_create("shared_policy_node",
1579 sizeof(struct sp_node),
1580 0, SLAB_PANIC, NULL, NULL);
1582 /* Set interleaving policy for system init. This way not all
1583 the data structures allocated at system boot end up in node zero. */
1585 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1586 printk("numa_policy_init: interleaving failed\n");
1589 /* Reset policy of current process to default */
1590 void numa_default_policy(void)
1592 do_set_mempolicy(MPOL_DEFAULT, NULL);
1595 /* Migrate a policy to a different set of nodes */
1596 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1598 nodemask_t *mpolmask;
1603 mpolmask = &pol->cpuset_mems_allowed;
1604 if (nodes_equal(*mpolmask, *newmask))
1607 switch (pol->policy) {
1610 case MPOL_INTERLEAVE:
1611 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1613 *mpolmask = *newmask;
1614 current->il_next = node_remap(current->il_next,
1615 *mpolmask, *newmask);
1617 case MPOL_PREFERRED:
1618 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1619 *mpolmask, *newmask);
1620 *mpolmask = *newmask;
1625 struct zonelist *zonelist;
1628 for (z = pol->v.zonelist->zones; *z; z++)
1629 node_set((*z)->zone_pgdat->node_id, nodes);
1630 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1633 zonelist = bind_zonelist(&nodes);
1635 /* If no mem, then zonelist is NULL and we keep old zonelist.
1636 * If that old zonelist has no remaining mems_allowed nodes,
1637 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1641 /* Good - got mem - substitute new zonelist */
1642 kfree(pol->v.zonelist);
1643 pol->v.zonelist = zonelist;
1645 *mpolmask = *newmask;
1655 * Wrapper for mpol_rebind_policy() that just requires task
1656 * pointer, and updates task mempolicy.
1659 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1661 mpol_rebind_policy(tsk->mempolicy, new);
1665 * Rebind each vma in mm to new nodemask.
1667 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1670 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1672 struct vm_area_struct *vma;
1674 down_write(&mm->mmap_sem);
1675 for (vma = mm->mmap; vma; vma = vma->vm_next)
1676 mpol_rebind_policy(vma->vm_policy, new);
1677 up_write(&mm->mmap_sem);
1681 * Display pages allocated per node and memory policy via /proc.
1684 static const char *policy_types[] = { "default", "prefer", "bind",
1688 * Convert a mempolicy into a string.
1689 * Returns the number of characters in buffer (if positive)
1690 * or an error (negative)
1692 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1697 int mode = pol ? pol->policy : MPOL_DEFAULT;
1704 case MPOL_PREFERRED:
1706 node_set(pol->v.preferred_node, nodes);
1710 get_zonemask(pol, &nodes);
1713 case MPOL_INTERLEAVE:
1714 nodes = pol->v.nodes;
1722 l = strlen(policy_types[mode]);
1723 if (buffer + maxlen < p + l + 1)
1726 strcpy(p, policy_types[mode]);
1729 if (!nodes_empty(nodes)) {
1730 if (buffer + maxlen < p + 2)
1733 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1739 unsigned long pages;
1741 unsigned long mapped;
1742 unsigned long mapcount_max;
1743 unsigned long node[MAX_NUMNODES];
1746 static void gather_stats(struct page *page, void *private)
1748 struct numa_maps *md = private;
1749 int count = page_mapcount(page);
1754 if (count > md->mapcount_max)
1755 md->mapcount_max = count;
1762 md->node[page_to_nid(page)]++;
1766 int show_numa_map(struct seq_file *m, void *v)
1768 struct task_struct *task = m->private;
1769 struct vm_area_struct *vma = v;
1770 struct numa_maps *md;
1777 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1781 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1782 &node_online_map, MPOL_MF_STATS, md);
1785 mpol_to_str(buffer, sizeof(buffer),
1786 get_vma_policy(task, vma, vma->vm_start));
1788 seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1789 vma->vm_start, buffer, md->pages,
1790 md->mapped, md->mapcount_max);
1793 seq_printf(m," anon=%lu",md->anon);
1795 for_each_online_node(n)
1797 seq_printf(m, " N%d=%lu", n, md->node[n]);
1803 if (m->count < m->size)
1804 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;