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/vs_cvirt.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);
592 * Migrate pages from one node to a target node.
593 * Returns error or the number of pages not migrated.
595 int migrate_to_node(struct mm_struct *mm, int source, int dest, int flags)
602 node_set(source, nmask);
604 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
605 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
607 if (!list_empty(&pagelist)) {
608 err = migrate_pages_to(&pagelist, NULL, dest);
609 if (!list_empty(&pagelist))
610 putback_lru_pages(&pagelist);
616 * Move pages between the two nodesets so as to preserve the physical
617 * layout as much as possible.
619 * Returns the number of page that could not be moved.
621 int do_migrate_pages(struct mm_struct *mm,
622 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
629 down_read(&mm->mmap_sem);
632 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
633 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
634 * bit in 'tmp', and return that <source, dest> pair for migration.
635 * The pair of nodemasks 'to' and 'from' define the map.
637 * If no pair of bits is found that way, fallback to picking some
638 * pair of 'source' and 'dest' bits that are not the same. If the
639 * 'source' and 'dest' bits are the same, this represents a node
640 * that will be migrating to itself, so no pages need move.
642 * If no bits are left in 'tmp', or if all remaining bits left
643 * in 'tmp' correspond to the same bit in 'to', return false
644 * (nothing left to migrate).
646 * This lets us pick a pair of nodes to migrate between, such that
647 * if possible the dest node is not already occupied by some other
648 * source node, minimizing the risk of overloading the memory on a
649 * node that would happen if we migrated incoming memory to a node
650 * before migrating outgoing memory source that same node.
652 * A single scan of tmp is sufficient. As we go, we remember the
653 * most recent <s, d> pair that moved (s != d). If we find a pair
654 * that not only moved, but what's better, moved to an empty slot
655 * (d is not set in tmp), then we break out then, with that pair.
656 * Otherwise when we finish scannng from_tmp, we at least have the
657 * most recent <s, d> pair that moved. If we get all the way through
658 * the scan of tmp without finding any node that moved, much less
659 * moved to an empty node, then there is nothing left worth migrating.
663 while (!nodes_empty(tmp)) {
668 for_each_node_mask(s, tmp) {
669 d = node_remap(s, *from_nodes, *to_nodes);
673 source = s; /* Node moved. Memorize */
676 /* dest not in remaining from nodes? */
677 if (!node_isset(dest, tmp))
683 node_clear(source, tmp);
684 err = migrate_to_node(mm, source, dest, flags);
691 up_read(&mm->mmap_sem);
700 static void migrate_page_add(struct page *page, struct list_head *pagelist,
705 int do_migrate_pages(struct mm_struct *mm,
706 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
712 long do_mbind(unsigned long start, unsigned long len,
713 unsigned long mode, nodemask_t *nmask, unsigned long flags)
715 struct vm_area_struct *vma;
716 struct mm_struct *mm = current->mm;
717 struct mempolicy *new;
722 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
723 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
726 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
729 if (start & ~PAGE_MASK)
732 if (mode == MPOL_DEFAULT)
733 flags &= ~MPOL_MF_STRICT;
735 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
743 if (mpol_check_policy(mode, nmask))
746 new = mpol_new(mode, nmask);
751 * If we are using the default policy then operation
752 * on discontinuous address spaces is okay after all
755 flags |= MPOL_MF_DISCONTIG_OK;
757 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
758 mode,nodes_addr(nodes)[0]);
760 down_write(&mm->mmap_sem);
761 vma = check_range(mm, start, end, nmask,
762 flags | MPOL_MF_INVERT, &pagelist);
768 err = mbind_range(vma, start, end, new);
770 if (!list_empty(&pagelist))
771 nr_failed = migrate_pages_to(&pagelist, vma, -1);
773 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)
801 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
804 nlongs = BITS_TO_LONGS(maxnode);
805 if ((maxnode % BITS_PER_LONG) == 0)
808 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
810 /* When the user specified more nodes than supported just check
811 if the non supported part is all zero. */
812 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
813 if (nlongs > PAGE_SIZE/sizeof(long))
815 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
817 if (get_user(t, nmask + k))
819 if (k == nlongs - 1) {
825 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
829 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
831 nodes_addr(*nodes)[nlongs-1] &= endmask;
835 /* Copy a kernel node mask to user space */
836 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
839 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
840 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
843 if (copy > PAGE_SIZE)
845 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
849 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
852 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
854 unsigned long __user *nmask, unsigned long maxnode,
860 err = get_nodes(&nodes, nmask, maxnode);
863 return do_mbind(start, len, mode, &nodes, flags);
866 /* Set the process memory policy */
867 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
868 unsigned long maxnode)
873 if (mode < 0 || mode > MPOL_MAX)
875 err = get_nodes(&nodes, nmask, maxnode);
878 return do_set_mempolicy(mode, &nodes);
881 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
882 const unsigned long __user *old_nodes,
883 const unsigned long __user *new_nodes)
885 struct mm_struct *mm;
886 struct task_struct *task;
889 nodemask_t task_nodes;
892 err = get_nodes(&old, old_nodes, maxnode);
896 err = get_nodes(&new, new_nodes, maxnode);
900 /* Find the mm_struct */
901 read_lock(&tasklist_lock);
902 task = pid ? find_task_by_pid(pid) : current;
904 read_unlock(&tasklist_lock);
907 mm = get_task_mm(task);
908 read_unlock(&tasklist_lock);
914 * Check if this process has the right to modify the specified
915 * process. The right exists if the process has administrative
916 * capabilities, superuser privileges or the same
917 * userid as the target process.
919 if ((current->euid != task->suid) && (current->euid != task->uid) &&
920 (current->uid != task->suid) && (current->uid != task->uid) &&
921 !capable(CAP_SYS_NICE)) {
926 task_nodes = cpuset_mems_allowed(task);
927 /* Is the user allowed to access the target nodes? */
928 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
933 err = do_migrate_pages(mm, &old, &new,
934 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
941 /* Retrieve NUMA policy */
942 asmlinkage long sys_get_mempolicy(int __user *policy,
943 unsigned long __user *nmask,
944 unsigned long maxnode,
945 unsigned long addr, unsigned long flags)
950 if (nmask != NULL && maxnode < MAX_NUMNODES)
953 err = do_get_mempolicy(&pval, &nodes, addr, flags);
958 if (policy && put_user(pval, policy))
962 err = copy_nodes_to_user(nmask, maxnode, &nodes);
969 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
970 compat_ulong_t __user *nmask,
971 compat_ulong_t maxnode,
972 compat_ulong_t addr, compat_ulong_t flags)
975 unsigned long __user *nm = NULL;
976 unsigned long nr_bits, alloc_size;
977 DECLARE_BITMAP(bm, MAX_NUMNODES);
979 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
980 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
983 nm = compat_alloc_user_space(alloc_size);
985 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
988 err = copy_from_user(bm, nm, alloc_size);
989 /* ensure entire bitmap is zeroed */
990 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
991 err |= compat_put_bitmap(nmask, bm, nr_bits);
997 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
998 compat_ulong_t maxnode)
1001 unsigned long __user *nm = NULL;
1002 unsigned long nr_bits, alloc_size;
1003 DECLARE_BITMAP(bm, MAX_NUMNODES);
1005 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1006 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1009 err = compat_get_bitmap(bm, nmask, nr_bits);
1010 nm = compat_alloc_user_space(alloc_size);
1011 err |= copy_to_user(nm, bm, alloc_size);
1017 return sys_set_mempolicy(mode, nm, nr_bits+1);
1020 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1021 compat_ulong_t mode, compat_ulong_t __user *nmask,
1022 compat_ulong_t maxnode, compat_ulong_t flags)
1025 unsigned long __user *nm = NULL;
1026 unsigned long nr_bits, alloc_size;
1029 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1030 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1033 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1034 nm = compat_alloc_user_space(alloc_size);
1035 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1041 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1046 /* Return effective policy for a VMA */
1047 static struct mempolicy * get_vma_policy(struct task_struct *task,
1048 struct vm_area_struct *vma, unsigned long addr)
1050 struct mempolicy *pol = task->mempolicy;
1053 if (vma->vm_ops && vma->vm_ops->get_policy)
1054 pol = vma->vm_ops->get_policy(vma, addr);
1055 else if (vma->vm_policy &&
1056 vma->vm_policy->policy != MPOL_DEFAULT)
1057 pol = vma->vm_policy;
1060 pol = &default_policy;
1064 /* Return a zonelist representing a mempolicy */
1065 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1069 switch (policy->policy) {
1070 case MPOL_PREFERRED:
1071 nd = policy->v.preferred_node;
1073 nd = numa_node_id();
1076 /* Lower zones don't get a policy applied */
1077 /* Careful: current->mems_allowed might have moved */
1078 if (gfp_zone(gfp) >= policy_zone)
1079 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
1080 return policy->v.zonelist;
1082 case MPOL_INTERLEAVE: /* should not happen */
1084 nd = numa_node_id();
1090 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
1093 /* Do dynamic interleaving for a process */
1094 static unsigned interleave_nodes(struct mempolicy *policy)
1097 struct task_struct *me = current;
1100 next = next_node(nid, policy->v.nodes);
1101 if (next >= MAX_NUMNODES)
1102 next = first_node(policy->v.nodes);
1108 * Depending on the memory policy provide a node from which to allocate the
1111 unsigned slab_node(struct mempolicy *policy)
1113 switch (policy->policy) {
1114 case MPOL_INTERLEAVE:
1115 return interleave_nodes(policy);
1119 * Follow bind policy behavior and start allocation at the
1122 return policy->v.zonelist->zones[0]->zone_pgdat->node_id;
1124 case MPOL_PREFERRED:
1125 if (policy->v.preferred_node >= 0)
1126 return policy->v.preferred_node;
1130 return numa_node_id();
1134 /* Do static interleaving for a VMA with known offset. */
1135 static unsigned offset_il_node(struct mempolicy *pol,
1136 struct vm_area_struct *vma, unsigned long off)
1138 unsigned nnodes = nodes_weight(pol->v.nodes);
1139 unsigned target = (unsigned)off % nnodes;
1145 nid = next_node(nid, pol->v.nodes);
1147 } while (c <= target);
1151 /* Determine a node number for interleave */
1152 static inline unsigned interleave_nid(struct mempolicy *pol,
1153 struct vm_area_struct *vma, unsigned long addr, int shift)
1158 off = vma->vm_pgoff;
1159 off += (addr - vma->vm_start) >> shift;
1160 return offset_il_node(pol, vma, off);
1162 return interleave_nodes(pol);
1165 #ifdef CONFIG_HUGETLBFS
1166 /* Return a zonelist suitable for a huge page allocation. */
1167 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1169 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1171 if (pol->policy == MPOL_INTERLEAVE) {
1174 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1175 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1177 return zonelist_policy(GFP_HIGHUSER, pol);
1181 /* Allocate a page in interleaved policy.
1182 Own path because it needs to do special accounting. */
1183 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1186 struct zonelist *zl;
1189 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1190 page = __alloc_pages(gfp, order, zl);
1191 if (page && page_zone(page) == zl->zones[0]) {
1192 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1199 * alloc_page_vma - Allocate a page for a VMA.
1202 * %GFP_USER user allocation.
1203 * %GFP_KERNEL kernel allocations,
1204 * %GFP_HIGHMEM highmem/user allocations,
1205 * %GFP_FS allocation should not call back into a file system.
1206 * %GFP_ATOMIC don't sleep.
1208 * @vma: Pointer to VMA or NULL if not available.
1209 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1211 * This function allocates a page from the kernel page pool and applies
1212 * a NUMA policy associated with the VMA or the current process.
1213 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1214 * mm_struct of the VMA to prevent it from going away. Should be used for
1215 * all allocations for pages that will be mapped into
1216 * user space. Returns NULL when no page can be allocated.
1218 * Should be called with the mm_sem of the vma hold.
1221 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1223 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1225 cpuset_update_task_memory_state();
1227 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1230 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1231 return alloc_page_interleave(gfp, 0, nid);
1233 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1237 * alloc_pages_current - Allocate pages.
1240 * %GFP_USER user allocation,
1241 * %GFP_KERNEL kernel allocation,
1242 * %GFP_HIGHMEM highmem allocation,
1243 * %GFP_FS don't call back into a file system.
1244 * %GFP_ATOMIC don't sleep.
1245 * @order: Power of two of allocation size in pages. 0 is a single page.
1247 * Allocate a page from the kernel page pool. When not in
1248 * interrupt context and apply the current process NUMA policy.
1249 * Returns NULL when no page can be allocated.
1251 * Don't call cpuset_update_task_memory_state() unless
1252 * 1) it's ok to take cpuset_sem (can WAIT), and
1253 * 2) allocating for current task (not interrupt).
1255 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1257 struct mempolicy *pol = current->mempolicy;
1259 if ((gfp & __GFP_WAIT) && !in_interrupt())
1260 cpuset_update_task_memory_state();
1261 if (!pol || in_interrupt())
1262 pol = &default_policy;
1263 if (pol->policy == MPOL_INTERLEAVE)
1264 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1265 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1267 EXPORT_SYMBOL(alloc_pages_current);
1270 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1271 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1272 * with the mems_allowed returned by cpuset_mems_allowed(). This
1273 * keeps mempolicies cpuset relative after its cpuset moves. See
1274 * further kernel/cpuset.c update_nodemask().
1276 void *cpuset_being_rebound;
1278 /* Slow path of a mempolicy copy */
1279 struct mempolicy *__mpol_copy(struct mempolicy *old)
1281 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1284 return ERR_PTR(-ENOMEM);
1285 if (current_cpuset_is_being_rebound()) {
1286 nodemask_t mems = cpuset_mems_allowed(current);
1287 mpol_rebind_policy(old, &mems);
1290 atomic_set(&new->refcnt, 1);
1291 if (new->policy == MPOL_BIND) {
1292 int sz = ksize(old->v.zonelist);
1293 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1294 if (!new->v.zonelist) {
1295 kmem_cache_free(policy_cache, new);
1296 return ERR_PTR(-ENOMEM);
1298 memcpy(new->v.zonelist, old->v.zonelist, sz);
1303 /* Slow path of a mempolicy comparison */
1304 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1308 if (a->policy != b->policy)
1310 switch (a->policy) {
1313 case MPOL_INTERLEAVE:
1314 return nodes_equal(a->v.nodes, b->v.nodes);
1315 case MPOL_PREFERRED:
1316 return a->v.preferred_node == b->v.preferred_node;
1319 for (i = 0; a->v.zonelist->zones[i]; i++)
1320 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1322 return b->v.zonelist->zones[i] == NULL;
1330 /* Slow path of a mpol destructor. */
1331 void __mpol_free(struct mempolicy *p)
1333 if (!atomic_dec_and_test(&p->refcnt))
1335 if (p->policy == MPOL_BIND)
1336 kfree(p->v.zonelist);
1337 p->policy = MPOL_DEFAULT;
1338 kmem_cache_free(policy_cache, p);
1342 * Shared memory backing store policy support.
1344 * Remember policies even when nobody has shared memory mapped.
1345 * The policies are kept in Red-Black tree linked from the inode.
1346 * They are protected by the sp->lock spinlock, which should be held
1347 * for any accesses to the tree.
1350 /* lookup first element intersecting start-end */
1351 /* Caller holds sp->lock */
1352 static struct sp_node *
1353 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1355 struct rb_node *n = sp->root.rb_node;
1358 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1360 if (start >= p->end)
1362 else if (end <= p->start)
1370 struct sp_node *w = NULL;
1371 struct rb_node *prev = rb_prev(n);
1374 w = rb_entry(prev, struct sp_node, nd);
1375 if (w->end <= start)
1379 return rb_entry(n, struct sp_node, nd);
1382 /* Insert a new shared policy into the list. */
1383 /* Caller holds sp->lock */
1384 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1386 struct rb_node **p = &sp->root.rb_node;
1387 struct rb_node *parent = NULL;
1392 nd = rb_entry(parent, struct sp_node, nd);
1393 if (new->start < nd->start)
1395 else if (new->end > nd->end)
1396 p = &(*p)->rb_right;
1400 rb_link_node(&new->nd, parent, p);
1401 rb_insert_color(&new->nd, &sp->root);
1402 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1403 new->policy ? new->policy->policy : 0);
1406 /* Find shared policy intersecting idx */
1408 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1410 struct mempolicy *pol = NULL;
1413 if (!sp->root.rb_node)
1415 spin_lock(&sp->lock);
1416 sn = sp_lookup(sp, idx, idx+1);
1418 mpol_get(sn->policy);
1421 spin_unlock(&sp->lock);
1425 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1427 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1428 rb_erase(&n->nd, &sp->root);
1429 mpol_free(n->policy);
1430 kmem_cache_free(sn_cache, n);
1434 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1436 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1447 /* Replace a policy range. */
1448 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1449 unsigned long end, struct sp_node *new)
1451 struct sp_node *n, *new2 = NULL;
1454 spin_lock(&sp->lock);
1455 n = sp_lookup(sp, start, end);
1456 /* Take care of old policies in the same range. */
1457 while (n && n->start < end) {
1458 struct rb_node *next = rb_next(&n->nd);
1459 if (n->start >= start) {
1465 /* Old policy spanning whole new range. */
1468 spin_unlock(&sp->lock);
1469 new2 = sp_alloc(end, n->end, n->policy);
1475 sp_insert(sp, new2);
1483 n = rb_entry(next, struct sp_node, nd);
1487 spin_unlock(&sp->lock);
1489 mpol_free(new2->policy);
1490 kmem_cache_free(sn_cache, new2);
1495 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1496 nodemask_t *policy_nodes)
1498 info->root = RB_ROOT;
1499 spin_lock_init(&info->lock);
1501 if (policy != MPOL_DEFAULT) {
1502 struct mempolicy *newpol;
1504 /* Falls back to MPOL_DEFAULT on any error */
1505 newpol = mpol_new(policy, policy_nodes);
1506 if (!IS_ERR(newpol)) {
1507 /* Create pseudo-vma that contains just the policy */
1508 struct vm_area_struct pvma;
1510 memset(&pvma, 0, sizeof(struct vm_area_struct));
1511 /* Policy covers entire file */
1512 pvma.vm_end = TASK_SIZE;
1513 mpol_set_shared_policy(info, &pvma, newpol);
1519 int mpol_set_shared_policy(struct shared_policy *info,
1520 struct vm_area_struct *vma, struct mempolicy *npol)
1523 struct sp_node *new = NULL;
1524 unsigned long sz = vma_pages(vma);
1526 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1528 sz, npol? npol->policy : -1,
1529 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1532 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1536 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1538 kmem_cache_free(sn_cache, new);
1542 /* Free a backing policy store on inode delete. */
1543 void mpol_free_shared_policy(struct shared_policy *p)
1546 struct rb_node *next;
1548 if (!p->root.rb_node)
1550 spin_lock(&p->lock);
1551 next = rb_first(&p->root);
1553 n = rb_entry(next, struct sp_node, nd);
1554 next = rb_next(&n->nd);
1555 rb_erase(&n->nd, &p->root);
1556 mpol_free(n->policy);
1557 kmem_cache_free(sn_cache, n);
1559 spin_unlock(&p->lock);
1562 /* assumes fs == KERNEL_DS */
1563 void __init numa_policy_init(void)
1565 policy_cache = kmem_cache_create("numa_policy",
1566 sizeof(struct mempolicy),
1567 0, SLAB_PANIC, NULL, NULL);
1569 sn_cache = kmem_cache_create("shared_policy_node",
1570 sizeof(struct sp_node),
1571 0, SLAB_PANIC, NULL, NULL);
1573 /* Set interleaving policy for system init. This way not all
1574 the data structures allocated at system boot end up in node zero. */
1576 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1577 printk("numa_policy_init: interleaving failed\n");
1580 /* Reset policy of current process to default */
1581 void numa_default_policy(void)
1583 do_set_mempolicy(MPOL_DEFAULT, NULL);
1586 /* Migrate a policy to a different set of nodes */
1587 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1589 nodemask_t *mpolmask;
1594 mpolmask = &pol->cpuset_mems_allowed;
1595 if (nodes_equal(*mpolmask, *newmask))
1598 switch (pol->policy) {
1601 case MPOL_INTERLEAVE:
1602 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1604 *mpolmask = *newmask;
1605 current->il_next = node_remap(current->il_next,
1606 *mpolmask, *newmask);
1608 case MPOL_PREFERRED:
1609 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1610 *mpolmask, *newmask);
1611 *mpolmask = *newmask;
1616 struct zonelist *zonelist;
1619 for (z = pol->v.zonelist->zones; *z; z++)
1620 node_set((*z)->zone_pgdat->node_id, nodes);
1621 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1624 zonelist = bind_zonelist(&nodes);
1626 /* If no mem, then zonelist is NULL and we keep old zonelist.
1627 * If that old zonelist has no remaining mems_allowed nodes,
1628 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1632 /* Good - got mem - substitute new zonelist */
1633 kfree(pol->v.zonelist);
1634 pol->v.zonelist = zonelist;
1636 *mpolmask = *newmask;
1646 * Wrapper for mpol_rebind_policy() that just requires task
1647 * pointer, and updates task mempolicy.
1650 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1652 mpol_rebind_policy(tsk->mempolicy, new);
1656 * Rebind each vma in mm to new nodemask.
1658 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1661 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1663 struct vm_area_struct *vma;
1665 down_write(&mm->mmap_sem);
1666 for (vma = mm->mmap; vma; vma = vma->vm_next)
1667 mpol_rebind_policy(vma->vm_policy, new);
1668 up_write(&mm->mmap_sem);
1672 * Display pages allocated per node and memory policy via /proc.
1675 static const char *policy_types[] = { "default", "prefer", "bind",
1679 * Convert a mempolicy into a string.
1680 * Returns the number of characters in buffer (if positive)
1681 * or an error (negative)
1683 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1688 int mode = pol ? pol->policy : MPOL_DEFAULT;
1695 case MPOL_PREFERRED:
1697 node_set(pol->v.preferred_node, nodes);
1701 get_zonemask(pol, &nodes);
1704 case MPOL_INTERLEAVE:
1705 nodes = pol->v.nodes;
1713 l = strlen(policy_types[mode]);
1714 if (buffer + maxlen < p + l + 1)
1717 strcpy(p, policy_types[mode]);
1720 if (!nodes_empty(nodes)) {
1721 if (buffer + maxlen < p + 2)
1724 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1730 unsigned long pages;
1732 unsigned long active;
1733 unsigned long writeback;
1734 unsigned long mapcount_max;
1735 unsigned long dirty;
1736 unsigned long swapcache;
1737 unsigned long node[MAX_NUMNODES];
1740 static void gather_stats(struct page *page, void *private, int pte_dirty)
1742 struct numa_maps *md = private;
1743 int count = page_mapcount(page);
1746 if (pte_dirty || PageDirty(page))
1749 if (PageSwapCache(page))
1752 if (PageActive(page))
1755 if (PageWriteback(page))
1761 if (count > md->mapcount_max)
1762 md->mapcount_max = count;
1764 md->node[page_to_nid(page)]++;
1767 #ifdef CONFIG_HUGETLB_PAGE
1768 static void check_huge_range(struct vm_area_struct *vma,
1769 unsigned long start, unsigned long end,
1770 struct numa_maps *md)
1775 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1776 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1786 page = pte_page(pte);
1790 gather_stats(page, md, pte_dirty(*ptep));
1794 static inline void check_huge_range(struct vm_area_struct *vma,
1795 unsigned long start, unsigned long end,
1796 struct numa_maps *md)
1801 int show_numa_map(struct seq_file *m, void *v)
1803 struct task_struct *task = m->private;
1804 struct vm_area_struct *vma = v;
1805 struct numa_maps *md;
1806 struct file *file = vma->vm_file;
1807 struct mm_struct *mm = vma->vm_mm;
1814 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1818 mpol_to_str(buffer, sizeof(buffer),
1819 get_vma_policy(task, vma, vma->vm_start));
1821 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1824 seq_printf(m, " file=");
1825 seq_path(m, file->f_vfsmnt, file->f_dentry, "\n\t= ");
1826 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1827 seq_printf(m, " heap");
1828 } else if (vma->vm_start <= mm->start_stack &&
1829 vma->vm_end >= mm->start_stack) {
1830 seq_printf(m, " stack");
1833 if (is_vm_hugetlb_page(vma)) {
1834 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
1835 seq_printf(m, " huge");
1837 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1838 &node_online_map, MPOL_MF_STATS, md);
1845 seq_printf(m," anon=%lu",md->anon);
1848 seq_printf(m," dirty=%lu",md->dirty);
1850 if (md->pages != md->anon && md->pages != md->dirty)
1851 seq_printf(m, " mapped=%lu", md->pages);
1853 if (md->mapcount_max > 1)
1854 seq_printf(m, " mapmax=%lu", md->mapcount_max);
1857 seq_printf(m," swapcache=%lu", md->swapcache);
1859 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
1860 seq_printf(m," active=%lu", md->active);
1863 seq_printf(m," writeback=%lu", md->writeback);
1865 for_each_online_node(n)
1867 seq_printf(m, " N%d=%lu", n, md->node[n]);
1872 if (m->count < m->size)
1873 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;