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;
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(void *) * max, GFP_KERNEL);
146 for_each_node_mask(nd, *nodes)
147 zl->zones[num++] = &NODE_DATA(nd)->node_zones[policy_zone];
148 zl->zones[num] = NULL;
152 /* Create a new policy */
153 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
155 struct mempolicy *policy;
157 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
158 if (mode == MPOL_DEFAULT)
160 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
162 return ERR_PTR(-ENOMEM);
163 atomic_set(&policy->refcnt, 1);
165 case MPOL_INTERLEAVE:
166 policy->v.nodes = *nodes;
167 if (nodes_weight(*nodes) == 0) {
168 kmem_cache_free(policy_cache, policy);
169 return ERR_PTR(-EINVAL);
173 policy->v.preferred_node = first_node(*nodes);
174 if (policy->v.preferred_node >= MAX_NUMNODES)
175 policy->v.preferred_node = -1;
178 policy->v.zonelist = bind_zonelist(nodes);
179 if (policy->v.zonelist == NULL) {
180 kmem_cache_free(policy_cache, policy);
181 return ERR_PTR(-ENOMEM);
185 policy->policy = mode;
186 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
190 static void gather_stats(struct page *, void *);
191 static void migrate_page_add(struct page *page, struct list_head *pagelist,
192 unsigned long flags);
194 /* Scan through pages checking if pages follow certain conditions. */
195 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
196 unsigned long addr, unsigned long end,
197 const nodemask_t *nodes, unsigned long flags,
204 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
209 if (!pte_present(*pte))
211 page = vm_normal_page(vma, addr, *pte);
215 * The check for PageReserved here is important to avoid
216 * handling zero pages and other pages that may have been
217 * marked special by the system.
219 * If the PageReserved would not be checked here then f.e.
220 * the location of the zero page could have an influence
221 * on MPOL_MF_STRICT, zero pages would be counted for
222 * the per node stats, and there would be useless attempts
223 * to put zero pages on the migration list.
225 if (PageReserved(page))
227 nid = page_to_nid(page);
228 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
231 if (flags & MPOL_MF_STATS)
232 gather_stats(page, private);
233 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
234 migrate_page_add(page, private, flags);
237 } while (pte++, addr += PAGE_SIZE, addr != end);
238 pte_unmap_unlock(orig_pte, ptl);
242 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
243 unsigned long addr, unsigned long end,
244 const nodemask_t *nodes, unsigned long flags,
250 pmd = pmd_offset(pud, addr);
252 next = pmd_addr_end(addr, end);
253 if (pmd_none_or_clear_bad(pmd))
255 if (check_pte_range(vma, pmd, addr, next, nodes,
258 } while (pmd++, addr = next, addr != end);
262 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
263 unsigned long addr, unsigned long end,
264 const nodemask_t *nodes, unsigned long flags,
270 pud = pud_offset(pgd, addr);
272 next = pud_addr_end(addr, end);
273 if (pud_none_or_clear_bad(pud))
275 if (check_pmd_range(vma, pud, addr, next, nodes,
278 } while (pud++, addr = next, addr != end);
282 static inline int check_pgd_range(struct vm_area_struct *vma,
283 unsigned long addr, unsigned long end,
284 const nodemask_t *nodes, unsigned long flags,
290 pgd = pgd_offset(vma->vm_mm, addr);
292 next = pgd_addr_end(addr, end);
293 if (pgd_none_or_clear_bad(pgd))
295 if (check_pud_range(vma, pgd, addr, next, nodes,
298 } while (pgd++, addr = next, addr != end);
302 /* Check if a vma is migratable */
303 static inline int vma_migratable(struct vm_area_struct *vma)
305 if (vma->vm_flags & (
306 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED))
312 * Check if all pages in a range are on a set of nodes.
313 * If pagelist != NULL then isolate pages from the LRU and
314 * put them on the pagelist.
316 static struct vm_area_struct *
317 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
318 const nodemask_t *nodes, unsigned long flags, void *private)
321 struct vm_area_struct *first, *vma, *prev;
323 /* Clear the LRU lists so pages can be isolated */
324 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
327 first = find_vma(mm, start);
329 return ERR_PTR(-EFAULT);
331 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
332 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
333 if (!vma->vm_next && vma->vm_end < end)
334 return ERR_PTR(-EFAULT);
335 if (prev && prev->vm_end < vma->vm_start)
336 return ERR_PTR(-EFAULT);
338 if (!is_vm_hugetlb_page(vma) &&
339 ((flags & MPOL_MF_STRICT) ||
340 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
341 vma_migratable(vma)))) {
342 unsigned long endvma = vma->vm_end;
346 if (vma->vm_start > start)
347 start = vma->vm_start;
348 err = check_pgd_range(vma, start, endvma, nodes,
351 first = ERR_PTR(err);
360 /* Apply policy to a single VMA */
361 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
364 struct mempolicy *old = vma->vm_policy;
366 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
367 vma->vm_start, vma->vm_end, vma->vm_pgoff,
368 vma->vm_ops, vma->vm_file,
369 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
371 if (vma->vm_ops && vma->vm_ops->set_policy)
372 err = vma->vm_ops->set_policy(vma, new);
375 vma->vm_policy = new;
381 /* Step 2: apply policy to a range and do splits. */
382 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
383 unsigned long end, struct mempolicy *new)
385 struct vm_area_struct *next;
389 for (; vma && vma->vm_start < end; vma = next) {
391 if (vma->vm_start < start)
392 err = split_vma(vma->vm_mm, vma, start, 1);
393 if (!err && vma->vm_end > end)
394 err = split_vma(vma->vm_mm, vma, end, 0);
396 err = policy_vma(vma, new);
403 static int contextualize_policy(int mode, nodemask_t *nodes)
408 cpuset_update_task_memory_state();
409 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
411 return mpol_check_policy(mode, nodes);
414 /* Set the process memory policy */
415 long do_set_mempolicy(int mode, nodemask_t *nodes)
417 struct mempolicy *new;
419 if (contextualize_policy(mode, nodes))
421 new = mpol_new(mode, nodes);
424 mpol_free(current->mempolicy);
425 current->mempolicy = new;
426 if (new && new->policy == MPOL_INTERLEAVE)
427 current->il_next = first_node(new->v.nodes);
431 /* Fill a zone bitmap for a policy */
432 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
439 for (i = 0; p->v.zonelist->zones[i]; i++)
440 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
445 case MPOL_INTERLEAVE:
449 /* or use current node instead of online map? */
450 if (p->v.preferred_node < 0)
451 *nodes = node_online_map;
453 node_set(p->v.preferred_node, *nodes);
460 static int lookup_node(struct mm_struct *mm, unsigned long addr)
465 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
467 err = page_to_nid(p);
473 /* Retrieve NUMA policy */
474 long do_get_mempolicy(int *policy, nodemask_t *nmask,
475 unsigned long addr, unsigned long flags)
478 struct mm_struct *mm = current->mm;
479 struct vm_area_struct *vma = NULL;
480 struct mempolicy *pol = current->mempolicy;
482 cpuset_update_task_memory_state();
483 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
485 if (flags & MPOL_F_ADDR) {
486 down_read(&mm->mmap_sem);
487 vma = find_vma_intersection(mm, addr, addr+1);
489 up_read(&mm->mmap_sem);
492 if (vma->vm_ops && vma->vm_ops->get_policy)
493 pol = vma->vm_ops->get_policy(vma, addr);
495 pol = vma->vm_policy;
500 pol = &default_policy;
502 if (flags & MPOL_F_NODE) {
503 if (flags & MPOL_F_ADDR) {
504 err = lookup_node(mm, addr);
508 } else if (pol == current->mempolicy &&
509 pol->policy == MPOL_INTERLEAVE) {
510 *policy = current->il_next;
516 *policy = pol->policy;
519 up_read(¤t->mm->mmap_sem);
525 get_zonemask(pol, nmask);
529 up_read(¤t->mm->mmap_sem);
537 static void migrate_page_add(struct page *page, struct list_head *pagelist,
541 * Avoid migrating a page that is shared with others.
543 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
544 if (isolate_lru_page(page))
545 list_add(&page->lru, pagelist);
550 * Migrate the list 'pagelist' of pages to a certain destination.
552 * Specify destination with either non-NULL vma or dest_node >= 0
553 * Return the number of pages not migrated or error code
555 static int migrate_pages_to(struct list_head *pagelist,
556 struct vm_area_struct *vma, int dest)
568 list_for_each(p, pagelist) {
570 page = alloc_page_vma(GFP_HIGHUSER, vma, vma->vm_start);
572 page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
578 list_add(&page->lru, &newlist);
580 if (nr_pages > MIGRATE_CHUNK_SIZE);
583 err = migrate_pages(pagelist, &newlist, &moved, &failed);
585 putback_lru_pages(&moved); /* Call release pages instead ?? */
587 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
590 /* Return leftover allocated pages */
591 while (!list_empty(&newlist)) {
592 page = list_entry(newlist.next, struct page, lru);
593 list_del(&page->lru);
596 list_splice(&failed, pagelist);
600 /* Calculate number of leftover pages */
602 list_for_each(p, pagelist)
608 * Migrate pages from one node to a target node.
609 * Returns error or the number of pages not migrated.
611 int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags)
618 node_set(source, nmask);
620 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
621 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
623 if (!list_empty(&pagelist)) {
624 err = migrate_pages_to(&pagelist, NULL, dest);
625 if (!list_empty(&pagelist))
626 putback_lru_pages(&pagelist);
632 * Move pages between the two nodesets so as to preserve the physical
633 * layout as much as possible.
635 * Returns the number of page that could not be moved.
637 int do_migrate_pages(struct mm_struct *mm,
638 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
645 down_read(&mm->mmap_sem);
648 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
649 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
650 * bit in 'tmp', and return that <source, dest> pair for migration.
651 * The pair of nodemasks 'to' and 'from' define the map.
653 * If no pair of bits is found that way, fallback to picking some
654 * pair of 'source' and 'dest' bits that are not the same. If the
655 * 'source' and 'dest' bits are the same, this represents a node
656 * that will be migrating to itself, so no pages need move.
658 * If no bits are left in 'tmp', or if all remaining bits left
659 * in 'tmp' correspond to the same bit in 'to', return false
660 * (nothing left to migrate).
662 * This lets us pick a pair of nodes to migrate between, such that
663 * if possible the dest node is not already occupied by some other
664 * source node, minimizing the risk of overloading the memory on a
665 * node that would happen if we migrated incoming memory to a node
666 * before migrating outgoing memory source that same node.
668 * A single scan of tmp is sufficient. As we go, we remember the
669 * most recent <s, d> pair that moved (s != d). If we find a pair
670 * that not only moved, but what's better, moved to an empty slot
671 * (d is not set in tmp), then we break out then, with that pair.
672 * Otherwise when we finish scannng from_tmp, we at least have the
673 * most recent <s, d> pair that moved. If we get all the way through
674 * the scan of tmp without finding any node that moved, much less
675 * moved to an empty node, then there is nothing left worth migrating.
679 while (!nodes_empty(tmp)) {
684 for_each_node_mask(s, tmp) {
685 d = node_remap(s, *from_nodes, *to_nodes);
689 source = s; /* Node moved. Memorize */
692 /* dest not in remaining from nodes? */
693 if (!node_isset(dest, tmp))
699 node_clear(source, tmp);
700 err = migrate_to_node(mm, source, dest, flags);
707 up_read(&mm->mmap_sem);
713 long do_mbind(unsigned long start, unsigned long len,
714 unsigned long mode, nodemask_t *nmask, unsigned long flags)
716 struct vm_area_struct *vma;
717 struct mm_struct *mm = current->mm;
718 struct mempolicy *new;
723 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
724 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
727 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
730 if (start & ~PAGE_MASK)
733 if (mode == MPOL_DEFAULT)
734 flags &= ~MPOL_MF_STRICT;
736 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
744 if (mpol_check_policy(mode, nmask))
747 new = mpol_new(mode, nmask);
752 * If we are using the default policy then operation
753 * on discontinuous address spaces is okay after all
756 flags |= MPOL_MF_DISCONTIG_OK;
758 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
759 mode,nodes_addr(nodes)[0]);
761 down_write(&mm->mmap_sem);
762 vma = check_range(mm, start, end, nmask,
763 flags | MPOL_MF_INVERT, &pagelist);
769 err = mbind_range(vma, start, end, new);
771 if (!list_empty(&pagelist))
772 nr_failed = migrate_pages_to(&pagelist, vma, -1);
774 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
777 if (!list_empty(&pagelist))
778 putback_lru_pages(&pagelist);
780 up_write(&mm->mmap_sem);
786 * User space interface with variable sized bitmaps for nodelists.
789 /* Copy a node mask from user space. */
790 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
791 unsigned long maxnode)
794 unsigned long nlongs;
795 unsigned long endmask;
799 if (maxnode == 0 || !nmask)
802 nlongs = BITS_TO_LONGS(maxnode);
803 if ((maxnode % BITS_PER_LONG) == 0)
806 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
808 /* When the user specified more nodes than supported just check
809 if the non supported part is all zero. */
810 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
811 if (nlongs > PAGE_SIZE/sizeof(long))
813 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
815 if (get_user(t, nmask + k))
817 if (k == nlongs - 1) {
823 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
827 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
829 nodes_addr(*nodes)[nlongs-1] &= endmask;
833 /* Copy a kernel node mask to user space */
834 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
837 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
838 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
841 if (copy > PAGE_SIZE)
843 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
847 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
850 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
852 unsigned long __user *nmask, unsigned long maxnode,
858 err = get_nodes(&nodes, nmask, maxnode);
861 return do_mbind(start, len, mode, &nodes, flags);
864 /* Set the process memory policy */
865 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
866 unsigned long maxnode)
871 if (mode < 0 || mode > MPOL_MAX)
873 err = get_nodes(&nodes, nmask, maxnode);
876 return do_set_mempolicy(mode, &nodes);
879 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
880 const unsigned long __user *old_nodes,
881 const unsigned long __user *new_nodes)
883 struct mm_struct *mm;
884 struct task_struct *task;
887 nodemask_t task_nodes;
890 err = get_nodes(&old, old_nodes, maxnode);
894 err = get_nodes(&new, new_nodes, maxnode);
898 /* Find the mm_struct */
899 read_lock(&tasklist_lock);
900 task = pid ? find_task_by_pid(pid) : current;
902 read_unlock(&tasklist_lock);
905 mm = get_task_mm(task);
906 read_unlock(&tasklist_lock);
912 * Check if this process has the right to modify the specified
913 * process. The right exists if the process has administrative
914 * capabilities, superuser priviledges or the same
915 * userid as the target process.
917 if ((current->euid != task->suid) && (current->euid != task->uid) &&
918 (current->uid != task->suid) && (current->uid != task->uid) &&
919 !capable(CAP_SYS_ADMIN)) {
924 task_nodes = cpuset_mems_allowed(task);
925 /* Is the user allowed to access the target nodes? */
926 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
931 err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
938 /* Retrieve NUMA policy */
939 asmlinkage long sys_get_mempolicy(int __user *policy,
940 unsigned long __user *nmask,
941 unsigned long maxnode,
942 unsigned long addr, unsigned long flags)
947 if (nmask != NULL && maxnode < MAX_NUMNODES)
950 err = do_get_mempolicy(&pval, &nodes, addr, flags);
955 if (policy && put_user(pval, policy))
959 err = copy_nodes_to_user(nmask, maxnode, &nodes);
966 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
967 compat_ulong_t __user *nmask,
968 compat_ulong_t maxnode,
969 compat_ulong_t addr, compat_ulong_t flags)
972 unsigned long __user *nm = NULL;
973 unsigned long nr_bits, alloc_size;
974 DECLARE_BITMAP(bm, MAX_NUMNODES);
976 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
977 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
980 nm = compat_alloc_user_space(alloc_size);
982 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
985 err = copy_from_user(bm, nm, alloc_size);
986 /* ensure entire bitmap is zeroed */
987 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
988 err |= compat_put_bitmap(nmask, bm, nr_bits);
994 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
995 compat_ulong_t maxnode)
998 unsigned long __user *nm = NULL;
999 unsigned long nr_bits, alloc_size;
1000 DECLARE_BITMAP(bm, MAX_NUMNODES);
1002 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1003 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1006 err = compat_get_bitmap(bm, nmask, nr_bits);
1007 nm = compat_alloc_user_space(alloc_size);
1008 err |= copy_to_user(nm, bm, alloc_size);
1014 return sys_set_mempolicy(mode, nm, nr_bits+1);
1017 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1018 compat_ulong_t mode, compat_ulong_t __user *nmask,
1019 compat_ulong_t maxnode, compat_ulong_t flags)
1022 unsigned long __user *nm = NULL;
1023 unsigned long nr_bits, alloc_size;
1026 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1027 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1030 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1031 nm = compat_alloc_user_space(alloc_size);
1032 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1038 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1043 /* Return effective policy for a VMA */
1044 static struct mempolicy * get_vma_policy(struct task_struct *task,
1045 struct vm_area_struct *vma, unsigned long addr)
1047 struct mempolicy *pol = task->mempolicy;
1050 if (vma->vm_ops && vma->vm_ops->get_policy)
1051 pol = vma->vm_ops->get_policy(vma, addr);
1052 else if (vma->vm_policy &&
1053 vma->vm_policy->policy != MPOL_DEFAULT)
1054 pol = vma->vm_policy;
1057 pol = &default_policy;
1061 /* Return a zonelist representing a mempolicy */
1062 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1066 switch (policy->policy) {
1067 case MPOL_PREFERRED:
1068 nd = policy->v.preferred_node;
1070 nd = numa_node_id();
1073 /* Lower zones don't get a policy applied */
1074 /* Careful: current->mems_allowed might have moved */
1075 if (gfp_zone(gfp) >= policy_zone)
1076 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
1077 return policy->v.zonelist;
1079 case MPOL_INTERLEAVE: /* should not happen */
1081 nd = numa_node_id();
1087 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
1090 /* Do dynamic interleaving for a process */
1091 static unsigned interleave_nodes(struct mempolicy *policy)
1094 struct task_struct *me = current;
1097 next = next_node(nid, policy->v.nodes);
1098 if (next >= MAX_NUMNODES)
1099 next = first_node(policy->v.nodes);
1105 * Depending on the memory policy provide a node from which to allocate the
1108 unsigned slab_node(struct mempolicy *policy)
1110 switch (policy->policy) {
1111 case MPOL_INTERLEAVE:
1112 return interleave_nodes(policy);
1116 * Follow bind policy behavior and start allocation at the
1119 return policy->v.zonelist->zones[0]->zone_pgdat->node_id;
1121 case MPOL_PREFERRED:
1122 if (policy->v.preferred_node >= 0)
1123 return policy->v.preferred_node;
1127 return numa_node_id();
1131 /* Do static interleaving for a VMA with known offset. */
1132 static unsigned offset_il_node(struct mempolicy *pol,
1133 struct vm_area_struct *vma, unsigned long off)
1135 unsigned nnodes = nodes_weight(pol->v.nodes);
1136 unsigned target = (unsigned)off % nnodes;
1142 nid = next_node(nid, pol->v.nodes);
1144 } while (c <= target);
1148 /* Determine a node number for interleave */
1149 static inline unsigned interleave_nid(struct mempolicy *pol,
1150 struct vm_area_struct *vma, unsigned long addr, int shift)
1155 off = vma->vm_pgoff;
1156 off += (addr - vma->vm_start) >> shift;
1157 return offset_il_node(pol, vma, off);
1159 return interleave_nodes(pol);
1162 /* Return a zonelist suitable for a huge page allocation. */
1163 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1165 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1167 if (pol->policy == MPOL_INTERLEAVE) {
1170 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1171 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1173 return zonelist_policy(GFP_HIGHUSER, pol);
1176 /* Allocate a page in interleaved policy.
1177 Own path because it needs to do special accounting. */
1178 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1181 struct zonelist *zl;
1184 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1185 page = __alloc_pages(gfp, order, zl);
1186 if (page && page_zone(page) == zl->zones[0]) {
1187 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1194 * alloc_page_vma - Allocate a page for a VMA.
1197 * %GFP_USER user allocation.
1198 * %GFP_KERNEL kernel allocations,
1199 * %GFP_HIGHMEM highmem/user allocations,
1200 * %GFP_FS allocation should not call back into a file system.
1201 * %GFP_ATOMIC don't sleep.
1203 * @vma: Pointer to VMA or NULL if not available.
1204 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1206 * This function allocates a page from the kernel page pool and applies
1207 * a NUMA policy associated with the VMA or the current process.
1208 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1209 * mm_struct of the VMA to prevent it from going away. Should be used for
1210 * all allocations for pages that will be mapped into
1211 * user space. Returns NULL when no page can be allocated.
1213 * Should be called with the mm_sem of the vma hold.
1216 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1218 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1220 cpuset_update_task_memory_state();
1222 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1225 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1226 return alloc_page_interleave(gfp, 0, nid);
1228 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1232 * alloc_pages_current - Allocate pages.
1235 * %GFP_USER user allocation,
1236 * %GFP_KERNEL kernel allocation,
1237 * %GFP_HIGHMEM highmem allocation,
1238 * %GFP_FS don't call back into a file system.
1239 * %GFP_ATOMIC don't sleep.
1240 * @order: Power of two of allocation size in pages. 0 is a single page.
1242 * Allocate a page from the kernel page pool. When not in
1243 * interrupt context and apply the current process NUMA policy.
1244 * Returns NULL when no page can be allocated.
1246 * Don't call cpuset_update_task_memory_state() unless
1247 * 1) it's ok to take cpuset_sem (can WAIT), and
1248 * 2) allocating for current task (not interrupt).
1250 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1252 struct mempolicy *pol = current->mempolicy;
1254 if ((gfp & __GFP_WAIT) && !in_interrupt())
1255 cpuset_update_task_memory_state();
1256 if (!pol || in_interrupt())
1257 pol = &default_policy;
1258 if (pol->policy == MPOL_INTERLEAVE)
1259 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1260 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1262 EXPORT_SYMBOL(alloc_pages_current);
1265 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1266 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1267 * with the mems_allowed returned by cpuset_mems_allowed(). This
1268 * keeps mempolicies cpuset relative after its cpuset moves. See
1269 * further kernel/cpuset.c update_nodemask().
1271 void *cpuset_being_rebound;
1273 /* Slow path of a mempolicy copy */
1274 struct mempolicy *__mpol_copy(struct mempolicy *old)
1276 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1279 return ERR_PTR(-ENOMEM);
1280 if (current_cpuset_is_being_rebound()) {
1281 nodemask_t mems = cpuset_mems_allowed(current);
1282 mpol_rebind_policy(old, &mems);
1285 atomic_set(&new->refcnt, 1);
1286 if (new->policy == MPOL_BIND) {
1287 int sz = ksize(old->v.zonelist);
1288 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1289 if (!new->v.zonelist) {
1290 kmem_cache_free(policy_cache, new);
1291 return ERR_PTR(-ENOMEM);
1293 memcpy(new->v.zonelist, old->v.zonelist, sz);
1298 /* Slow path of a mempolicy comparison */
1299 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1303 if (a->policy != b->policy)
1305 switch (a->policy) {
1308 case MPOL_INTERLEAVE:
1309 return nodes_equal(a->v.nodes, b->v.nodes);
1310 case MPOL_PREFERRED:
1311 return a->v.preferred_node == b->v.preferred_node;
1314 for (i = 0; a->v.zonelist->zones[i]; i++)
1315 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1317 return b->v.zonelist->zones[i] == NULL;
1325 /* Slow path of a mpol destructor. */
1326 void __mpol_free(struct mempolicy *p)
1328 if (!atomic_dec_and_test(&p->refcnt))
1330 if (p->policy == MPOL_BIND)
1331 kfree(p->v.zonelist);
1332 p->policy = MPOL_DEFAULT;
1333 kmem_cache_free(policy_cache, p);
1337 * Shared memory backing store policy support.
1339 * Remember policies even when nobody has shared memory mapped.
1340 * The policies are kept in Red-Black tree linked from the inode.
1341 * They are protected by the sp->lock spinlock, which should be held
1342 * for any accesses to the tree.
1345 /* lookup first element intersecting start-end */
1346 /* Caller holds sp->lock */
1347 static struct sp_node *
1348 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1350 struct rb_node *n = sp->root.rb_node;
1353 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1355 if (start >= p->end)
1357 else if (end <= p->start)
1365 struct sp_node *w = NULL;
1366 struct rb_node *prev = rb_prev(n);
1369 w = rb_entry(prev, struct sp_node, nd);
1370 if (w->end <= start)
1374 return rb_entry(n, struct sp_node, nd);
1377 /* Insert a new shared policy into the list. */
1378 /* Caller holds sp->lock */
1379 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1381 struct rb_node **p = &sp->root.rb_node;
1382 struct rb_node *parent = NULL;
1387 nd = rb_entry(parent, struct sp_node, nd);
1388 if (new->start < nd->start)
1390 else if (new->end > nd->end)
1391 p = &(*p)->rb_right;
1395 rb_link_node(&new->nd, parent, p);
1396 rb_insert_color(&new->nd, &sp->root);
1397 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1398 new->policy ? new->policy->policy : 0);
1401 /* Find shared policy intersecting idx */
1403 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1405 struct mempolicy *pol = NULL;
1408 if (!sp->root.rb_node)
1410 spin_lock(&sp->lock);
1411 sn = sp_lookup(sp, idx, idx+1);
1413 mpol_get(sn->policy);
1416 spin_unlock(&sp->lock);
1420 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1422 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1423 rb_erase(&n->nd, &sp->root);
1424 mpol_free(n->policy);
1425 kmem_cache_free(sn_cache, n);
1429 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1431 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1442 /* Replace a policy range. */
1443 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1444 unsigned long end, struct sp_node *new)
1446 struct sp_node *n, *new2 = NULL;
1449 spin_lock(&sp->lock);
1450 n = sp_lookup(sp, start, end);
1451 /* Take care of old policies in the same range. */
1452 while (n && n->start < end) {
1453 struct rb_node *next = rb_next(&n->nd);
1454 if (n->start >= start) {
1460 /* Old policy spanning whole new range. */
1463 spin_unlock(&sp->lock);
1464 new2 = sp_alloc(end, n->end, n->policy);
1470 sp_insert(sp, new2);
1478 n = rb_entry(next, struct sp_node, nd);
1482 spin_unlock(&sp->lock);
1484 mpol_free(new2->policy);
1485 kmem_cache_free(sn_cache, new2);
1490 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1491 nodemask_t *policy_nodes)
1493 info->root = RB_ROOT;
1494 spin_lock_init(&info->lock);
1496 if (policy != MPOL_DEFAULT) {
1497 struct mempolicy *newpol;
1499 /* Falls back to MPOL_DEFAULT on any error */
1500 newpol = mpol_new(policy, policy_nodes);
1501 if (!IS_ERR(newpol)) {
1502 /* Create pseudo-vma that contains just the policy */
1503 struct vm_area_struct pvma;
1505 memset(&pvma, 0, sizeof(struct vm_area_struct));
1506 /* Policy covers entire file */
1507 pvma.vm_end = TASK_SIZE;
1508 mpol_set_shared_policy(info, &pvma, newpol);
1514 int mpol_set_shared_policy(struct shared_policy *info,
1515 struct vm_area_struct *vma, struct mempolicy *npol)
1518 struct sp_node *new = NULL;
1519 unsigned long sz = vma_pages(vma);
1521 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1523 sz, npol? npol->policy : -1,
1524 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1527 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1531 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1533 kmem_cache_free(sn_cache, new);
1537 /* Free a backing policy store on inode delete. */
1538 void mpol_free_shared_policy(struct shared_policy *p)
1541 struct rb_node *next;
1543 if (!p->root.rb_node)
1545 spin_lock(&p->lock);
1546 next = rb_first(&p->root);
1548 n = rb_entry(next, struct sp_node, nd);
1549 next = rb_next(&n->nd);
1550 rb_erase(&n->nd, &p->root);
1551 mpol_free(n->policy);
1552 kmem_cache_free(sn_cache, n);
1554 spin_unlock(&p->lock);
1557 /* assumes fs == KERNEL_DS */
1558 void __init numa_policy_init(void)
1560 policy_cache = kmem_cache_create("numa_policy",
1561 sizeof(struct mempolicy),
1562 0, SLAB_PANIC, NULL, NULL);
1564 sn_cache = kmem_cache_create("shared_policy_node",
1565 sizeof(struct sp_node),
1566 0, SLAB_PANIC, NULL, NULL);
1568 /* Set interleaving policy for system init. This way not all
1569 the data structures allocated at system boot end up in node zero. */
1571 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1572 printk("numa_policy_init: interleaving failed\n");
1575 /* Reset policy of current process to default */
1576 void numa_default_policy(void)
1578 do_set_mempolicy(MPOL_DEFAULT, NULL);
1581 /* Migrate a policy to a different set of nodes */
1582 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1584 nodemask_t *mpolmask;
1589 mpolmask = &pol->cpuset_mems_allowed;
1590 if (nodes_equal(*mpolmask, *newmask))
1593 switch (pol->policy) {
1596 case MPOL_INTERLEAVE:
1597 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1599 *mpolmask = *newmask;
1600 current->il_next = node_remap(current->il_next,
1601 *mpolmask, *newmask);
1603 case MPOL_PREFERRED:
1604 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1605 *mpolmask, *newmask);
1606 *mpolmask = *newmask;
1611 struct zonelist *zonelist;
1614 for (z = pol->v.zonelist->zones; *z; z++)
1615 node_set((*z)->zone_pgdat->node_id, nodes);
1616 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1619 zonelist = bind_zonelist(&nodes);
1621 /* If no mem, then zonelist is NULL and we keep old zonelist.
1622 * If that old zonelist has no remaining mems_allowed nodes,
1623 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1627 /* Good - got mem - substitute new zonelist */
1628 kfree(pol->v.zonelist);
1629 pol->v.zonelist = zonelist;
1631 *mpolmask = *newmask;
1641 * Wrapper for mpol_rebind_policy() that just requires task
1642 * pointer, and updates task mempolicy.
1645 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1647 mpol_rebind_policy(tsk->mempolicy, new);
1651 * Rebind each vma in mm to new nodemask.
1653 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1656 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1658 struct vm_area_struct *vma;
1660 down_write(&mm->mmap_sem);
1661 for (vma = mm->mmap; vma; vma = vma->vm_next)
1662 mpol_rebind_policy(vma->vm_policy, new);
1663 up_write(&mm->mmap_sem);
1667 * Display pages allocated per node and memory policy via /proc.
1670 static const char *policy_types[] = { "default", "prefer", "bind",
1674 * Convert a mempolicy into a string.
1675 * Returns the number of characters in buffer (if positive)
1676 * or an error (negative)
1678 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1683 int mode = pol ? pol->policy : MPOL_DEFAULT;
1690 case MPOL_PREFERRED:
1692 node_set(pol->v.preferred_node, nodes);
1696 get_zonemask(pol, &nodes);
1699 case MPOL_INTERLEAVE:
1700 nodes = pol->v.nodes;
1708 l = strlen(policy_types[mode]);
1709 if (buffer + maxlen < p + l + 1)
1712 strcpy(p, policy_types[mode]);
1715 if (!nodes_empty(nodes)) {
1716 if (buffer + maxlen < p + 2)
1719 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1725 unsigned long pages;
1727 unsigned long mapped;
1728 unsigned long mapcount_max;
1729 unsigned long node[MAX_NUMNODES];
1732 static void gather_stats(struct page *page, void *private)
1734 struct numa_maps *md = private;
1735 int count = page_mapcount(page);
1740 if (count > md->mapcount_max)
1741 md->mapcount_max = count;
1748 md->node[page_to_nid(page)]++;
1752 int show_numa_map(struct seq_file *m, void *v)
1754 struct task_struct *task = m->private;
1755 struct vm_area_struct *vma = v;
1756 struct numa_maps *md;
1763 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1767 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1768 &node_online_map, MPOL_MF_STATS, md);
1771 mpol_to_str(buffer, sizeof(buffer),
1772 get_vma_policy(task, vma, vma->vm_start));
1774 seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1775 vma->vm_start, buffer, md->pages,
1776 md->mapped, md->mapcount_max);
1779 seq_printf(m," anon=%lu",md->anon);
1781 for_each_online_node(n)
1783 seq_printf(m, " N%d=%lu", n, md->node[n]);
1789 if (m->count < m->size)
1790 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;