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 static kmem_cache_t *policy_cache;
99 static kmem_cache_t *sn_cache;
101 #define PDprintk(fmt...)
103 /* Highest zone. An specific allocation for a zone below that is not
105 int policy_zone = ZONE_DMA;
107 struct mempolicy default_policy = {
108 .refcnt = ATOMIC_INIT(1), /* never free it */
109 .policy = MPOL_DEFAULT,
112 /* Do sanity checking on a policy */
113 static int mpol_check_policy(int mode, nodemask_t *nodes)
115 int empty = nodes_empty(*nodes);
123 case MPOL_INTERLEAVE:
124 /* Preferred will only use the first bit, but allow
130 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
132 /* Generate a custom zonelist for the BIND policy. */
133 static struct zonelist *bind_zonelist(nodemask_t *nodes)
138 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
139 zl = kmalloc(sizeof(void *) * max, GFP_KERNEL);
143 for_each_node_mask(nd, *nodes)
144 zl->zones[num++] = &NODE_DATA(nd)->node_zones[policy_zone];
145 zl->zones[num] = NULL;
149 /* Create a new policy */
150 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
152 struct mempolicy *policy;
154 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
155 if (mode == MPOL_DEFAULT)
157 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
159 return ERR_PTR(-ENOMEM);
160 atomic_set(&policy->refcnt, 1);
162 case MPOL_INTERLEAVE:
163 policy->v.nodes = *nodes;
164 if (nodes_weight(*nodes) == 0) {
165 kmem_cache_free(policy_cache, policy);
166 return ERR_PTR(-EINVAL);
170 policy->v.preferred_node = first_node(*nodes);
171 if (policy->v.preferred_node >= MAX_NUMNODES)
172 policy->v.preferred_node = -1;
175 policy->v.zonelist = bind_zonelist(nodes);
176 if (policy->v.zonelist == NULL) {
177 kmem_cache_free(policy_cache, policy);
178 return ERR_PTR(-ENOMEM);
182 policy->policy = mode;
183 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
187 static void gather_stats(struct page *, void *);
188 static void migrate_page_add(struct page *page, struct list_head *pagelist,
189 unsigned long flags);
191 /* Scan through pages checking if pages follow certain conditions. */
192 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
193 unsigned long addr, unsigned long end,
194 const nodemask_t *nodes, unsigned long flags,
201 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
206 if (!pte_present(*pte))
208 page = vm_normal_page(vma, addr, *pte);
212 * The check for PageReserved here is important to avoid
213 * handling zero pages and other pages that may have been
214 * marked special by the system.
216 * If the PageReserved would not be checked here then f.e.
217 * the location of the zero page could have an influence
218 * on MPOL_MF_STRICT, zero pages would be counted for
219 * the per node stats, and there would be useless attempts
220 * to put zero pages on the migration list.
222 if (PageReserved(page))
224 nid = page_to_nid(page);
225 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
228 if (flags & MPOL_MF_STATS)
229 gather_stats(page, private);
230 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
231 migrate_page_add(page, private, flags);
234 } while (pte++, addr += PAGE_SIZE, addr != end);
235 pte_unmap_unlock(orig_pte, ptl);
239 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
240 unsigned long addr, unsigned long end,
241 const nodemask_t *nodes, unsigned long flags,
247 pmd = pmd_offset(pud, addr);
249 next = pmd_addr_end(addr, end);
250 if (pmd_none_or_clear_bad(pmd))
252 if (check_pte_range(vma, pmd, addr, next, nodes,
255 } while (pmd++, addr = next, addr != end);
259 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
260 unsigned long addr, unsigned long end,
261 const nodemask_t *nodes, unsigned long flags,
267 pud = pud_offset(pgd, addr);
269 next = pud_addr_end(addr, end);
270 if (pud_none_or_clear_bad(pud))
272 if (check_pmd_range(vma, pud, addr, next, nodes,
275 } while (pud++, addr = next, addr != end);
279 static inline int check_pgd_range(struct vm_area_struct *vma,
280 unsigned long addr, unsigned long end,
281 const nodemask_t *nodes, unsigned long flags,
287 pgd = pgd_offset(vma->vm_mm, addr);
289 next = pgd_addr_end(addr, end);
290 if (pgd_none_or_clear_bad(pgd))
292 if (check_pud_range(vma, pgd, addr, next, nodes,
295 } while (pgd++, addr = next, addr != end);
299 /* Check if a vma is migratable */
300 static inline int vma_migratable(struct vm_area_struct *vma)
302 if (vma->vm_flags & (
303 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP|VM_RESERVED))
309 * Check if all pages in a range are on a set of nodes.
310 * If pagelist != NULL then isolate pages from the LRU and
311 * put them on the pagelist.
313 static struct vm_area_struct *
314 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
315 const nodemask_t *nodes, unsigned long flags, void *private)
318 struct vm_area_struct *first, *vma, *prev;
320 /* Clear the LRU lists so pages can be isolated */
321 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
324 first = find_vma(mm, start);
326 return ERR_PTR(-EFAULT);
328 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
329 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
330 if (!vma->vm_next && vma->vm_end < end)
331 return ERR_PTR(-EFAULT);
332 if (prev && prev->vm_end < vma->vm_start)
333 return ERR_PTR(-EFAULT);
335 if (!is_vm_hugetlb_page(vma) &&
336 ((flags & MPOL_MF_STRICT) ||
337 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
338 vma_migratable(vma)))) {
339 unsigned long endvma = vma->vm_end;
343 if (vma->vm_start > start)
344 start = vma->vm_start;
345 err = check_pgd_range(vma, start, endvma, nodes,
348 first = ERR_PTR(err);
357 /* Apply policy to a single VMA */
358 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
361 struct mempolicy *old = vma->vm_policy;
363 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
364 vma->vm_start, vma->vm_end, vma->vm_pgoff,
365 vma->vm_ops, vma->vm_file,
366 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
368 if (vma->vm_ops && vma->vm_ops->set_policy)
369 err = vma->vm_ops->set_policy(vma, new);
372 vma->vm_policy = new;
378 /* Step 2: apply policy to a range and do splits. */
379 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
380 unsigned long end, struct mempolicy *new)
382 struct vm_area_struct *next;
386 for (; vma && vma->vm_start < end; vma = next) {
388 if (vma->vm_start < start)
389 err = split_vma(vma->vm_mm, vma, start, 1);
390 if (!err && vma->vm_end > end)
391 err = split_vma(vma->vm_mm, vma, end, 0);
393 err = policy_vma(vma, new);
400 static int contextualize_policy(int mode, nodemask_t *nodes)
405 cpuset_update_task_memory_state();
406 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
408 return mpol_check_policy(mode, nodes);
411 /* Set the process memory policy */
412 long do_set_mempolicy(int mode, nodemask_t *nodes)
414 struct mempolicy *new;
416 if (contextualize_policy(mode, nodes))
418 new = mpol_new(mode, nodes);
421 mpol_free(current->mempolicy);
422 current->mempolicy = new;
423 if (new && new->policy == MPOL_INTERLEAVE)
424 current->il_next = first_node(new->v.nodes);
428 /* Fill a zone bitmap for a policy */
429 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
436 for (i = 0; p->v.zonelist->zones[i]; i++)
437 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
442 case MPOL_INTERLEAVE:
446 /* or use current node instead of online map? */
447 if (p->v.preferred_node < 0)
448 *nodes = node_online_map;
450 node_set(p->v.preferred_node, *nodes);
457 static int lookup_node(struct mm_struct *mm, unsigned long addr)
462 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
464 err = page_to_nid(p);
470 /* Retrieve NUMA policy */
471 long do_get_mempolicy(int *policy, nodemask_t *nmask,
472 unsigned long addr, unsigned long flags)
475 struct mm_struct *mm = current->mm;
476 struct vm_area_struct *vma = NULL;
477 struct mempolicy *pol = current->mempolicy;
479 cpuset_update_task_memory_state();
480 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
482 if (flags & MPOL_F_ADDR) {
483 down_read(&mm->mmap_sem);
484 vma = find_vma_intersection(mm, addr, addr+1);
486 up_read(&mm->mmap_sem);
489 if (vma->vm_ops && vma->vm_ops->get_policy)
490 pol = vma->vm_ops->get_policy(vma, addr);
492 pol = vma->vm_policy;
497 pol = &default_policy;
499 if (flags & MPOL_F_NODE) {
500 if (flags & MPOL_F_ADDR) {
501 err = lookup_node(mm, addr);
505 } else if (pol == current->mempolicy &&
506 pol->policy == MPOL_INTERLEAVE) {
507 *policy = current->il_next;
513 *policy = pol->policy;
516 up_read(¤t->mm->mmap_sem);
522 get_zonemask(pol, nmask);
526 up_read(¤t->mm->mmap_sem);
534 static void migrate_page_add(struct page *page, struct list_head *pagelist,
538 * Avoid migrating a page that is shared with others.
540 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
541 if (isolate_lru_page(page))
542 list_add(&page->lru, pagelist);
546 static int swap_pages(struct list_head *pagelist)
552 n = migrate_pages(pagelist, NULL, &moved, &failed);
553 putback_lru_pages(&failed);
554 putback_lru_pages(&moved);
560 * For now migrate_pages simply swaps out the pages from nodes that are in
561 * the source set but not in the target set. In the future, we would
562 * want a function that moves pages between the two nodesets in such
563 * a way as to preserve the physical layout as much as possible.
565 * Returns the number of page that could not be moved.
567 int do_migrate_pages(struct mm_struct *mm,
568 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
574 nodes_andnot(nodes, *from_nodes, *to_nodes);
576 down_read(&mm->mmap_sem);
577 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nodes,
578 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
580 if (!list_empty(&pagelist)) {
581 count = swap_pages(&pagelist);
582 putback_lru_pages(&pagelist);
585 up_read(&mm->mmap_sem);
589 long do_mbind(unsigned long start, unsigned long len,
590 unsigned long mode, nodemask_t *nmask, unsigned long flags)
592 struct vm_area_struct *vma;
593 struct mm_struct *mm = current->mm;
594 struct mempolicy *new;
599 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
600 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
603 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
606 if (start & ~PAGE_MASK)
609 if (mode == MPOL_DEFAULT)
610 flags &= ~MPOL_MF_STRICT;
612 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
620 if (mpol_check_policy(mode, nmask))
623 new = mpol_new(mode, nmask);
628 * If we are using the default policy then operation
629 * on discontinuous address spaces is okay after all
632 flags |= MPOL_MF_DISCONTIG_OK;
634 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
635 mode,nodes_addr(nodes)[0]);
637 down_write(&mm->mmap_sem);
638 vma = check_range(mm, start, end, nmask,
639 flags | MPOL_MF_INVERT, &pagelist);
645 err = mbind_range(vma, start, end, new);
646 if (!list_empty(&pagelist))
647 nr_failed = swap_pages(&pagelist);
649 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
652 if (!list_empty(&pagelist))
653 putback_lru_pages(&pagelist);
655 up_write(&mm->mmap_sem);
661 * User space interface with variable sized bitmaps for nodelists.
664 /* Copy a node mask from user space. */
665 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
666 unsigned long maxnode)
669 unsigned long nlongs;
670 unsigned long endmask;
674 if (maxnode == 0 || !nmask)
677 nlongs = BITS_TO_LONGS(maxnode);
678 if ((maxnode % BITS_PER_LONG) == 0)
681 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
683 /* When the user specified more nodes than supported just check
684 if the non supported part is all zero. */
685 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
686 if (nlongs > PAGE_SIZE/sizeof(long))
688 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
690 if (get_user(t, nmask + k))
692 if (k == nlongs - 1) {
698 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
702 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
704 nodes_addr(*nodes)[nlongs-1] &= endmask;
708 /* Copy a kernel node mask to user space */
709 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
712 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
713 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
716 if (copy > PAGE_SIZE)
718 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
722 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
725 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
727 unsigned long __user *nmask, unsigned long maxnode,
733 err = get_nodes(&nodes, nmask, maxnode);
736 return do_mbind(start, len, mode, &nodes, flags);
739 /* Set the process memory policy */
740 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
741 unsigned long maxnode)
746 if (mode < 0 || mode > MPOL_MAX)
748 err = get_nodes(&nodes, nmask, maxnode);
751 return do_set_mempolicy(mode, &nodes);
754 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
755 const unsigned long __user *old_nodes,
756 const unsigned long __user *new_nodes)
758 struct mm_struct *mm;
759 struct task_struct *task;
762 nodemask_t task_nodes;
765 err = get_nodes(&old, old_nodes, maxnode);
769 err = get_nodes(&new, new_nodes, maxnode);
773 /* Find the mm_struct */
774 read_lock(&tasklist_lock);
775 task = pid ? find_task_by_pid(pid) : current;
777 read_unlock(&tasklist_lock);
780 mm = get_task_mm(task);
781 read_unlock(&tasklist_lock);
787 * Check if this process has the right to modify the specified
788 * process. The right exists if the process has administrative
789 * capabilities, superuser priviledges or the same
790 * userid as the target process.
792 if ((current->euid != task->suid) && (current->euid != task->uid) &&
793 (current->uid != task->suid) && (current->uid != task->uid) &&
794 !capable(CAP_SYS_ADMIN)) {
799 task_nodes = cpuset_mems_allowed(task);
800 /* Is the user allowed to access the target nodes? */
801 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
806 err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
813 /* Retrieve NUMA policy */
814 asmlinkage long sys_get_mempolicy(int __user *policy,
815 unsigned long __user *nmask,
816 unsigned long maxnode,
817 unsigned long addr, unsigned long flags)
822 if (nmask != NULL && maxnode < MAX_NUMNODES)
825 err = do_get_mempolicy(&pval, &nodes, addr, flags);
830 if (policy && put_user(pval, policy))
834 err = copy_nodes_to_user(nmask, maxnode, &nodes);
841 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
842 compat_ulong_t __user *nmask,
843 compat_ulong_t maxnode,
844 compat_ulong_t addr, compat_ulong_t flags)
847 unsigned long __user *nm = NULL;
848 unsigned long nr_bits, alloc_size;
849 DECLARE_BITMAP(bm, MAX_NUMNODES);
851 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
852 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
855 nm = compat_alloc_user_space(alloc_size);
857 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
860 err = copy_from_user(bm, nm, alloc_size);
861 /* ensure entire bitmap is zeroed */
862 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
863 err |= compat_put_bitmap(nmask, bm, nr_bits);
869 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
870 compat_ulong_t maxnode)
873 unsigned long __user *nm = NULL;
874 unsigned long nr_bits, alloc_size;
875 DECLARE_BITMAP(bm, MAX_NUMNODES);
877 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
878 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
881 err = compat_get_bitmap(bm, nmask, nr_bits);
882 nm = compat_alloc_user_space(alloc_size);
883 err |= copy_to_user(nm, bm, alloc_size);
889 return sys_set_mempolicy(mode, nm, nr_bits+1);
892 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
893 compat_ulong_t mode, compat_ulong_t __user *nmask,
894 compat_ulong_t maxnode, compat_ulong_t flags)
897 unsigned long __user *nm = NULL;
898 unsigned long nr_bits, alloc_size;
901 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
902 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
905 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
906 nm = compat_alloc_user_space(alloc_size);
907 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
913 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
918 /* Return effective policy for a VMA */
919 static struct mempolicy * get_vma_policy(struct task_struct *task,
920 struct vm_area_struct *vma, unsigned long addr)
922 struct mempolicy *pol = task->mempolicy;
925 if (vma->vm_ops && vma->vm_ops->get_policy)
926 pol = vma->vm_ops->get_policy(vma, addr);
927 else if (vma->vm_policy &&
928 vma->vm_policy->policy != MPOL_DEFAULT)
929 pol = vma->vm_policy;
932 pol = &default_policy;
936 /* Return a zonelist representing a mempolicy */
937 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
941 switch (policy->policy) {
943 nd = policy->v.preferred_node;
948 /* Lower zones don't get a policy applied */
949 /* Careful: current->mems_allowed might have moved */
950 if (gfp_zone(gfp) >= policy_zone)
951 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
952 return policy->v.zonelist;
954 case MPOL_INTERLEAVE: /* should not happen */
962 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
965 /* Do dynamic interleaving for a process */
966 static unsigned interleave_nodes(struct mempolicy *policy)
969 struct task_struct *me = current;
972 next = next_node(nid, policy->v.nodes);
973 if (next >= MAX_NUMNODES)
974 next = first_node(policy->v.nodes);
980 * Depending on the memory policy provide a node from which to allocate the
983 unsigned slab_node(struct mempolicy *policy)
986 return numa_node_id();
988 switch (policy->policy) {
989 case MPOL_INTERLEAVE:
990 return interleave_nodes(policy);
994 * Follow bind policy behavior and start allocation at the
997 return policy->v.zonelist->zones[0]->zone_pgdat->node_id;
1000 if (policy->v.preferred_node >= 0)
1001 return policy->v.preferred_node;
1005 return numa_node_id();
1009 /* Do static interleaving for a VMA with known offset. */
1010 static unsigned offset_il_node(struct mempolicy *pol,
1011 struct vm_area_struct *vma, unsigned long off)
1013 unsigned nnodes = nodes_weight(pol->v.nodes);
1014 unsigned target = (unsigned)off % nnodes;
1020 nid = next_node(nid, pol->v.nodes);
1022 } while (c <= target);
1026 /* Determine a node number for interleave */
1027 static inline unsigned interleave_nid(struct mempolicy *pol,
1028 struct vm_area_struct *vma, unsigned long addr, int shift)
1033 off = vma->vm_pgoff;
1034 off += (addr - vma->vm_start) >> shift;
1035 return offset_il_node(pol, vma, off);
1037 return interleave_nodes(pol);
1040 /* Return a zonelist suitable for a huge page allocation. */
1041 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1043 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1045 if (pol->policy == MPOL_INTERLEAVE) {
1048 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1049 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1051 return zonelist_policy(GFP_HIGHUSER, pol);
1054 /* Allocate a page in interleaved policy.
1055 Own path because it needs to do special accounting. */
1056 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1059 struct zonelist *zl;
1062 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1063 page = __alloc_pages(gfp, order, zl);
1064 if (page && page_zone(page) == zl->zones[0]) {
1065 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1072 * alloc_page_vma - Allocate a page for a VMA.
1075 * %GFP_USER user allocation.
1076 * %GFP_KERNEL kernel allocations,
1077 * %GFP_HIGHMEM highmem/user allocations,
1078 * %GFP_FS allocation should not call back into a file system.
1079 * %GFP_ATOMIC don't sleep.
1081 * @vma: Pointer to VMA or NULL if not available.
1082 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1084 * This function allocates a page from the kernel page pool and applies
1085 * a NUMA policy associated with the VMA or the current process.
1086 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1087 * mm_struct of the VMA to prevent it from going away. Should be used for
1088 * all allocations for pages that will be mapped into
1089 * user space. Returns NULL when no page can be allocated.
1091 * Should be called with the mm_sem of the vma hold.
1094 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1096 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1098 cpuset_update_task_memory_state();
1100 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1103 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1104 return alloc_page_interleave(gfp, 0, nid);
1106 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1110 * alloc_pages_current - Allocate pages.
1113 * %GFP_USER user allocation,
1114 * %GFP_KERNEL kernel allocation,
1115 * %GFP_HIGHMEM highmem allocation,
1116 * %GFP_FS don't call back into a file system.
1117 * %GFP_ATOMIC don't sleep.
1118 * @order: Power of two of allocation size in pages. 0 is a single page.
1120 * Allocate a page from the kernel page pool. When not in
1121 * interrupt context and apply the current process NUMA policy.
1122 * Returns NULL when no page can be allocated.
1124 * Don't call cpuset_update_task_memory_state() unless
1125 * 1) it's ok to take cpuset_sem (can WAIT), and
1126 * 2) allocating for current task (not interrupt).
1128 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1130 struct mempolicy *pol = current->mempolicy;
1132 if ((gfp & __GFP_WAIT) && !in_interrupt())
1133 cpuset_update_task_memory_state();
1134 if (!pol || in_interrupt())
1135 pol = &default_policy;
1136 if (pol->policy == MPOL_INTERLEAVE)
1137 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1138 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1140 EXPORT_SYMBOL(alloc_pages_current);
1143 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1144 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1145 * with the mems_allowed returned by cpuset_mems_allowed(). This
1146 * keeps mempolicies cpuset relative after its cpuset moves. See
1147 * further kernel/cpuset.c update_nodemask().
1149 void *cpuset_being_rebound;
1151 /* Slow path of a mempolicy copy */
1152 struct mempolicy *__mpol_copy(struct mempolicy *old)
1154 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1157 return ERR_PTR(-ENOMEM);
1158 if (current_cpuset_is_being_rebound()) {
1159 nodemask_t mems = cpuset_mems_allowed(current);
1160 mpol_rebind_policy(old, &mems);
1163 atomic_set(&new->refcnt, 1);
1164 if (new->policy == MPOL_BIND) {
1165 int sz = ksize(old->v.zonelist);
1166 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1167 if (!new->v.zonelist) {
1168 kmem_cache_free(policy_cache, new);
1169 return ERR_PTR(-ENOMEM);
1171 memcpy(new->v.zonelist, old->v.zonelist, sz);
1176 /* Slow path of a mempolicy comparison */
1177 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1181 if (a->policy != b->policy)
1183 switch (a->policy) {
1186 case MPOL_INTERLEAVE:
1187 return nodes_equal(a->v.nodes, b->v.nodes);
1188 case MPOL_PREFERRED:
1189 return a->v.preferred_node == b->v.preferred_node;
1192 for (i = 0; a->v.zonelist->zones[i]; i++)
1193 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1195 return b->v.zonelist->zones[i] == NULL;
1203 /* Slow path of a mpol destructor. */
1204 void __mpol_free(struct mempolicy *p)
1206 if (!atomic_dec_and_test(&p->refcnt))
1208 if (p->policy == MPOL_BIND)
1209 kfree(p->v.zonelist);
1210 p->policy = MPOL_DEFAULT;
1211 kmem_cache_free(policy_cache, p);
1215 * Shared memory backing store policy support.
1217 * Remember policies even when nobody has shared memory mapped.
1218 * The policies are kept in Red-Black tree linked from the inode.
1219 * They are protected by the sp->lock spinlock, which should be held
1220 * for any accesses to the tree.
1223 /* lookup first element intersecting start-end */
1224 /* Caller holds sp->lock */
1225 static struct sp_node *
1226 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1228 struct rb_node *n = sp->root.rb_node;
1231 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1233 if (start >= p->end)
1235 else if (end <= p->start)
1243 struct sp_node *w = NULL;
1244 struct rb_node *prev = rb_prev(n);
1247 w = rb_entry(prev, struct sp_node, nd);
1248 if (w->end <= start)
1252 return rb_entry(n, struct sp_node, nd);
1255 /* Insert a new shared policy into the list. */
1256 /* Caller holds sp->lock */
1257 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1259 struct rb_node **p = &sp->root.rb_node;
1260 struct rb_node *parent = NULL;
1265 nd = rb_entry(parent, struct sp_node, nd);
1266 if (new->start < nd->start)
1268 else if (new->end > nd->end)
1269 p = &(*p)->rb_right;
1273 rb_link_node(&new->nd, parent, p);
1274 rb_insert_color(&new->nd, &sp->root);
1275 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1276 new->policy ? new->policy->policy : 0);
1279 /* Find shared policy intersecting idx */
1281 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1283 struct mempolicy *pol = NULL;
1286 if (!sp->root.rb_node)
1288 spin_lock(&sp->lock);
1289 sn = sp_lookup(sp, idx, idx+1);
1291 mpol_get(sn->policy);
1294 spin_unlock(&sp->lock);
1298 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1300 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1301 rb_erase(&n->nd, &sp->root);
1302 mpol_free(n->policy);
1303 kmem_cache_free(sn_cache, n);
1307 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1309 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1320 /* Replace a policy range. */
1321 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1322 unsigned long end, struct sp_node *new)
1324 struct sp_node *n, *new2 = NULL;
1327 spin_lock(&sp->lock);
1328 n = sp_lookup(sp, start, end);
1329 /* Take care of old policies in the same range. */
1330 while (n && n->start < end) {
1331 struct rb_node *next = rb_next(&n->nd);
1332 if (n->start >= start) {
1338 /* Old policy spanning whole new range. */
1341 spin_unlock(&sp->lock);
1342 new2 = sp_alloc(end, n->end, n->policy);
1348 sp_insert(sp, new2);
1356 n = rb_entry(next, struct sp_node, nd);
1360 spin_unlock(&sp->lock);
1362 mpol_free(new2->policy);
1363 kmem_cache_free(sn_cache, new2);
1368 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1369 nodemask_t *policy_nodes)
1371 info->root = RB_ROOT;
1372 spin_lock_init(&info->lock);
1374 if (policy != MPOL_DEFAULT) {
1375 struct mempolicy *newpol;
1377 /* Falls back to MPOL_DEFAULT on any error */
1378 newpol = mpol_new(policy, policy_nodes);
1379 if (!IS_ERR(newpol)) {
1380 /* Create pseudo-vma that contains just the policy */
1381 struct vm_area_struct pvma;
1383 memset(&pvma, 0, sizeof(struct vm_area_struct));
1384 /* Policy covers entire file */
1385 pvma.vm_end = TASK_SIZE;
1386 mpol_set_shared_policy(info, &pvma, newpol);
1392 int mpol_set_shared_policy(struct shared_policy *info,
1393 struct vm_area_struct *vma, struct mempolicy *npol)
1396 struct sp_node *new = NULL;
1397 unsigned long sz = vma_pages(vma);
1399 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1401 sz, npol? npol->policy : -1,
1402 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1405 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1409 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1411 kmem_cache_free(sn_cache, new);
1415 /* Free a backing policy store on inode delete. */
1416 void mpol_free_shared_policy(struct shared_policy *p)
1419 struct rb_node *next;
1421 if (!p->root.rb_node)
1423 spin_lock(&p->lock);
1424 next = rb_first(&p->root);
1426 n = rb_entry(next, struct sp_node, nd);
1427 next = rb_next(&n->nd);
1428 rb_erase(&n->nd, &p->root);
1429 mpol_free(n->policy);
1430 kmem_cache_free(sn_cache, n);
1432 spin_unlock(&p->lock);
1435 /* assumes fs == KERNEL_DS */
1436 void __init numa_policy_init(void)
1438 policy_cache = kmem_cache_create("numa_policy",
1439 sizeof(struct mempolicy),
1440 0, SLAB_PANIC, NULL, NULL);
1442 sn_cache = kmem_cache_create("shared_policy_node",
1443 sizeof(struct sp_node),
1444 0, SLAB_PANIC, NULL, NULL);
1446 /* Set interleaving policy for system init. This way not all
1447 the data structures allocated at system boot end up in node zero. */
1449 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1450 printk("numa_policy_init: interleaving failed\n");
1453 /* Reset policy of current process to default */
1454 void numa_default_policy(void)
1456 do_set_mempolicy(MPOL_DEFAULT, NULL);
1459 /* Migrate a policy to a different set of nodes */
1460 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1462 nodemask_t *mpolmask;
1467 mpolmask = &pol->cpuset_mems_allowed;
1468 if (nodes_equal(*mpolmask, *newmask))
1471 switch (pol->policy) {
1474 case MPOL_INTERLEAVE:
1475 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1477 *mpolmask = *newmask;
1478 current->il_next = node_remap(current->il_next,
1479 *mpolmask, *newmask);
1481 case MPOL_PREFERRED:
1482 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1483 *mpolmask, *newmask);
1484 *mpolmask = *newmask;
1489 struct zonelist *zonelist;
1492 for (z = pol->v.zonelist->zones; *z; z++)
1493 node_set((*z)->zone_pgdat->node_id, nodes);
1494 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1497 zonelist = bind_zonelist(&nodes);
1499 /* If no mem, then zonelist is NULL and we keep old zonelist.
1500 * If that old zonelist has no remaining mems_allowed nodes,
1501 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1505 /* Good - got mem - substitute new zonelist */
1506 kfree(pol->v.zonelist);
1507 pol->v.zonelist = zonelist;
1509 *mpolmask = *newmask;
1519 * Wrapper for mpol_rebind_policy() that just requires task
1520 * pointer, and updates task mempolicy.
1523 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1525 mpol_rebind_policy(tsk->mempolicy, new);
1529 * Rebind each vma in mm to new nodemask.
1531 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1534 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1536 struct vm_area_struct *vma;
1538 down_write(&mm->mmap_sem);
1539 for (vma = mm->mmap; vma; vma = vma->vm_next)
1540 mpol_rebind_policy(vma->vm_policy, new);
1541 up_write(&mm->mmap_sem);
1545 * Display pages allocated per node and memory policy via /proc.
1548 static const char *policy_types[] = { "default", "prefer", "bind",
1552 * Convert a mempolicy into a string.
1553 * Returns the number of characters in buffer (if positive)
1554 * or an error (negative)
1556 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1561 int mode = pol ? pol->policy : MPOL_DEFAULT;
1568 case MPOL_PREFERRED:
1570 node_set(pol->v.preferred_node, nodes);
1574 get_zonemask(pol, &nodes);
1577 case MPOL_INTERLEAVE:
1578 nodes = pol->v.nodes;
1586 l = strlen(policy_types[mode]);
1587 if (buffer + maxlen < p + l + 1)
1590 strcpy(p, policy_types[mode]);
1593 if (!nodes_empty(nodes)) {
1594 if (buffer + maxlen < p + 2)
1597 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1603 unsigned long pages;
1605 unsigned long mapped;
1606 unsigned long mapcount_max;
1607 unsigned long node[MAX_NUMNODES];
1610 static void gather_stats(struct page *page, void *private)
1612 struct numa_maps *md = private;
1613 int count = page_mapcount(page);
1618 if (count > md->mapcount_max)
1619 md->mapcount_max = count;
1626 md->node[page_to_nid(page)]++;
1630 int show_numa_map(struct seq_file *m, void *v)
1632 struct task_struct *task = m->private;
1633 struct vm_area_struct *vma = v;
1634 struct numa_maps *md;
1641 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1645 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1646 &node_online_map, MPOL_MF_STATS, md);
1649 mpol_to_str(buffer, sizeof(buffer),
1650 get_vma_policy(task, vma, vma->vm_start));
1652 seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1653 vma->vm_start, buffer, md->pages,
1654 md->mapped, md->mapcount_max);
1657 seq_printf(m," anon=%lu",md->anon);
1659 for_each_online_node(n)
1661 seq_printf(m, " N%d=%lu", n, md->node[n]);
1667 if (m->count < m->size)
1668 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;