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 vm_area_struct *vma,
189 struct page *page, struct list_head *pagelist, 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(vma, 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 /* Check if we are the only process mapping the page in question */
535 static inline int single_mm_mapping(struct mm_struct *mm,
536 struct address_space *mapping)
538 struct vm_area_struct *vma;
539 struct prio_tree_iter iter;
542 spin_lock(&mapping->i_mmap_lock);
543 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
544 if (mm != vma->vm_mm) {
548 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
549 if (mm != vma->vm_mm) {
554 spin_unlock(&mapping->i_mmap_lock);
559 * Add a page to be migrated to the pagelist
561 static void migrate_page_add(struct vm_area_struct *vma,
562 struct page *page, struct list_head *pagelist, unsigned long flags)
565 * Avoid migrating a page that is shared by others and not writable.
567 if ((flags & MPOL_MF_MOVE_ALL) || !page->mapping || PageAnon(page) ||
568 mapping_writably_mapped(page->mapping) ||
569 single_mm_mapping(vma->vm_mm, page->mapping)) {
570 if (isolate_lru_page(page))
571 list_add(&page->lru, pagelist);
575 static int swap_pages(struct list_head *pagelist)
581 n = migrate_pages(pagelist, NULL, &moved, &failed);
582 putback_lru_pages(&failed);
583 putback_lru_pages(&moved);
589 * For now migrate_pages simply swaps out the pages from nodes that are in
590 * the source set but not in the target set. In the future, we would
591 * want a function that moves pages between the two nodesets in such
592 * a way as to preserve the physical layout as much as possible.
594 * Returns the number of page that could not be moved.
596 int do_migrate_pages(struct mm_struct *mm,
597 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
603 nodes_andnot(nodes, *from_nodes, *to_nodes);
605 down_read(&mm->mmap_sem);
606 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nodes,
607 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
609 if (!list_empty(&pagelist)) {
610 count = swap_pages(&pagelist);
611 putback_lru_pages(&pagelist);
614 up_read(&mm->mmap_sem);
618 long do_mbind(unsigned long start, unsigned long len,
619 unsigned long mode, nodemask_t *nmask, unsigned long flags)
621 struct vm_area_struct *vma;
622 struct mm_struct *mm = current->mm;
623 struct mempolicy *new;
628 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
629 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
632 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
635 if (start & ~PAGE_MASK)
638 if (mode == MPOL_DEFAULT)
639 flags &= ~MPOL_MF_STRICT;
641 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
649 if (mpol_check_policy(mode, nmask))
652 new = mpol_new(mode, nmask);
657 * If we are using the default policy then operation
658 * on discontinuous address spaces is okay after all
661 flags |= MPOL_MF_DISCONTIG_OK;
663 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
664 mode,nodes_addr(nodes)[0]);
666 down_write(&mm->mmap_sem);
667 vma = check_range(mm, start, end, nmask,
668 flags | MPOL_MF_INVERT, &pagelist);
674 err = mbind_range(vma, start, end, new);
675 if (!list_empty(&pagelist))
676 nr_failed = swap_pages(&pagelist);
678 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
681 if (!list_empty(&pagelist))
682 putback_lru_pages(&pagelist);
684 up_write(&mm->mmap_sem);
690 * User space interface with variable sized bitmaps for nodelists.
693 /* Copy a node mask from user space. */
694 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
695 unsigned long maxnode)
698 unsigned long nlongs;
699 unsigned long endmask;
703 if (maxnode == 0 || !nmask)
706 nlongs = BITS_TO_LONGS(maxnode);
707 if ((maxnode % BITS_PER_LONG) == 0)
710 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
712 /* When the user specified more nodes than supported just check
713 if the non supported part is all zero. */
714 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
715 if (nlongs > PAGE_SIZE/sizeof(long))
717 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
719 if (get_user(t, nmask + k))
721 if (k == nlongs - 1) {
727 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
731 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
733 nodes_addr(*nodes)[nlongs-1] &= endmask;
737 /* Copy a kernel node mask to user space */
738 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
741 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
742 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
745 if (copy > PAGE_SIZE)
747 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
751 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
754 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
756 unsigned long __user *nmask, unsigned long maxnode,
762 err = get_nodes(&nodes, nmask, maxnode);
765 return do_mbind(start, len, mode, &nodes, flags);
768 /* Set the process memory policy */
769 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
770 unsigned long maxnode)
775 if (mode < 0 || mode > MPOL_MAX)
777 err = get_nodes(&nodes, nmask, maxnode);
780 return do_set_mempolicy(mode, &nodes);
783 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
784 const unsigned long __user *old_nodes,
785 const unsigned long __user *new_nodes)
787 struct mm_struct *mm;
788 struct task_struct *task;
791 nodemask_t task_nodes;
794 err = get_nodes(&old, old_nodes, maxnode);
798 err = get_nodes(&new, new_nodes, maxnode);
802 /* Find the mm_struct */
803 read_lock(&tasklist_lock);
804 task = pid ? find_task_by_pid(pid) : current;
806 read_unlock(&tasklist_lock);
809 mm = get_task_mm(task);
810 read_unlock(&tasklist_lock);
816 * Check if this process has the right to modify the specified
817 * process. The right exists if the process has administrative
818 * capabilities, superuser priviledges or the same
819 * userid as the target process.
821 if ((current->euid != task->suid) && (current->euid != task->uid) &&
822 (current->uid != task->suid) && (current->uid != task->uid) &&
823 !capable(CAP_SYS_ADMIN)) {
828 task_nodes = cpuset_mems_allowed(task);
829 /* Is the user allowed to access the target nodes? */
830 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
835 err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
842 /* Retrieve NUMA policy */
843 asmlinkage long sys_get_mempolicy(int __user *policy,
844 unsigned long __user *nmask,
845 unsigned long maxnode,
846 unsigned long addr, unsigned long flags)
851 if (nmask != NULL && maxnode < MAX_NUMNODES)
854 err = do_get_mempolicy(&pval, &nodes, addr, flags);
859 if (policy && put_user(pval, policy))
863 err = copy_nodes_to_user(nmask, maxnode, &nodes);
870 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
871 compat_ulong_t __user *nmask,
872 compat_ulong_t maxnode,
873 compat_ulong_t addr, compat_ulong_t flags)
876 unsigned long __user *nm = NULL;
877 unsigned long nr_bits, alloc_size;
878 DECLARE_BITMAP(bm, MAX_NUMNODES);
880 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
881 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
884 nm = compat_alloc_user_space(alloc_size);
886 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
889 err = copy_from_user(bm, nm, alloc_size);
890 /* ensure entire bitmap is zeroed */
891 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
892 err |= compat_put_bitmap(nmask, bm, nr_bits);
898 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
899 compat_ulong_t maxnode)
902 unsigned long __user *nm = NULL;
903 unsigned long nr_bits, alloc_size;
904 DECLARE_BITMAP(bm, MAX_NUMNODES);
906 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
907 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
910 err = compat_get_bitmap(bm, nmask, nr_bits);
911 nm = compat_alloc_user_space(alloc_size);
912 err |= copy_to_user(nm, bm, alloc_size);
918 return sys_set_mempolicy(mode, nm, nr_bits+1);
921 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
922 compat_ulong_t mode, compat_ulong_t __user *nmask,
923 compat_ulong_t maxnode, compat_ulong_t flags)
926 unsigned long __user *nm = NULL;
927 unsigned long nr_bits, alloc_size;
930 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
931 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
934 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
935 nm = compat_alloc_user_space(alloc_size);
936 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
942 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
947 /* Return effective policy for a VMA */
948 static struct mempolicy * get_vma_policy(struct task_struct *task,
949 struct vm_area_struct *vma, unsigned long addr)
951 struct mempolicy *pol = task->mempolicy;
954 if (vma->vm_ops && vma->vm_ops->get_policy)
955 pol = vma->vm_ops->get_policy(vma, addr);
956 else if (vma->vm_policy &&
957 vma->vm_policy->policy != MPOL_DEFAULT)
958 pol = vma->vm_policy;
961 pol = &default_policy;
965 /* Return a zonelist representing a mempolicy */
966 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
970 switch (policy->policy) {
972 nd = policy->v.preferred_node;
977 /* Lower zones don't get a policy applied */
978 /* Careful: current->mems_allowed might have moved */
979 if (gfp_zone(gfp) >= policy_zone)
980 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
981 return policy->v.zonelist;
983 case MPOL_INTERLEAVE: /* should not happen */
991 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
994 /* Do dynamic interleaving for a process */
995 static unsigned interleave_nodes(struct mempolicy *policy)
998 struct task_struct *me = current;
1001 next = next_node(nid, policy->v.nodes);
1002 if (next >= MAX_NUMNODES)
1003 next = first_node(policy->v.nodes);
1008 /* Do static interleaving for a VMA with known offset. */
1009 static unsigned offset_il_node(struct mempolicy *pol,
1010 struct vm_area_struct *vma, unsigned long off)
1012 unsigned nnodes = nodes_weight(pol->v.nodes);
1013 unsigned target = (unsigned)off % nnodes;
1019 nid = next_node(nid, pol->v.nodes);
1021 } while (c <= target);
1025 /* Determine a node number for interleave */
1026 static inline unsigned interleave_nid(struct mempolicy *pol,
1027 struct vm_area_struct *vma, unsigned long addr, int shift)
1032 off = vma->vm_pgoff;
1033 off += (addr - vma->vm_start) >> shift;
1034 return offset_il_node(pol, vma, off);
1036 return interleave_nodes(pol);
1039 /* Return a zonelist suitable for a huge page allocation. */
1040 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1042 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1044 if (pol->policy == MPOL_INTERLEAVE) {
1047 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1048 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1050 return zonelist_policy(GFP_HIGHUSER, pol);
1053 /* Allocate a page in interleaved policy.
1054 Own path because it needs to do special accounting. */
1055 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1058 struct zonelist *zl;
1061 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1062 page = __alloc_pages(gfp, order, zl);
1063 if (page && page_zone(page) == zl->zones[0]) {
1064 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1071 * alloc_page_vma - Allocate a page for a VMA.
1074 * %GFP_USER user allocation.
1075 * %GFP_KERNEL kernel allocations,
1076 * %GFP_HIGHMEM highmem/user allocations,
1077 * %GFP_FS allocation should not call back into a file system.
1078 * %GFP_ATOMIC don't sleep.
1080 * @vma: Pointer to VMA or NULL if not available.
1081 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1083 * This function allocates a page from the kernel page pool and applies
1084 * a NUMA policy associated with the VMA or the current process.
1085 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1086 * mm_struct of the VMA to prevent it from going away. Should be used for
1087 * all allocations for pages that will be mapped into
1088 * user space. Returns NULL when no page can be allocated.
1090 * Should be called with the mm_sem of the vma hold.
1093 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1095 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1097 cpuset_update_task_memory_state();
1099 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1102 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1103 return alloc_page_interleave(gfp, 0, nid);
1105 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1109 * alloc_pages_current - Allocate pages.
1112 * %GFP_USER user allocation,
1113 * %GFP_KERNEL kernel allocation,
1114 * %GFP_HIGHMEM highmem allocation,
1115 * %GFP_FS don't call back into a file system.
1116 * %GFP_ATOMIC don't sleep.
1117 * @order: Power of two of allocation size in pages. 0 is a single page.
1119 * Allocate a page from the kernel page pool. When not in
1120 * interrupt context and apply the current process NUMA policy.
1121 * Returns NULL when no page can be allocated.
1123 * Don't call cpuset_update_task_memory_state() unless
1124 * 1) it's ok to take cpuset_sem (can WAIT), and
1125 * 2) allocating for current task (not interrupt).
1127 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1129 struct mempolicy *pol = current->mempolicy;
1131 if ((gfp & __GFP_WAIT) && !in_interrupt())
1132 cpuset_update_task_memory_state();
1133 if (!pol || in_interrupt())
1134 pol = &default_policy;
1135 if (pol->policy == MPOL_INTERLEAVE)
1136 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1137 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1139 EXPORT_SYMBOL(alloc_pages_current);
1142 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1143 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1144 * with the mems_allowed returned by cpuset_mems_allowed(). This
1145 * keeps mempolicies cpuset relative after its cpuset moves. See
1146 * further kernel/cpuset.c update_nodemask().
1148 void *cpuset_being_rebound;
1150 /* Slow path of a mempolicy copy */
1151 struct mempolicy *__mpol_copy(struct mempolicy *old)
1153 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1156 return ERR_PTR(-ENOMEM);
1157 if (current_cpuset_is_being_rebound()) {
1158 nodemask_t mems = cpuset_mems_allowed(current);
1159 mpol_rebind_policy(old, &mems);
1162 atomic_set(&new->refcnt, 1);
1163 if (new->policy == MPOL_BIND) {
1164 int sz = ksize(old->v.zonelist);
1165 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1166 if (!new->v.zonelist) {
1167 kmem_cache_free(policy_cache, new);
1168 return ERR_PTR(-ENOMEM);
1170 memcpy(new->v.zonelist, old->v.zonelist, sz);
1175 /* Slow path of a mempolicy comparison */
1176 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1180 if (a->policy != b->policy)
1182 switch (a->policy) {
1185 case MPOL_INTERLEAVE:
1186 return nodes_equal(a->v.nodes, b->v.nodes);
1187 case MPOL_PREFERRED:
1188 return a->v.preferred_node == b->v.preferred_node;
1191 for (i = 0; a->v.zonelist->zones[i]; i++)
1192 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1194 return b->v.zonelist->zones[i] == NULL;
1202 /* Slow path of a mpol destructor. */
1203 void __mpol_free(struct mempolicy *p)
1205 if (!atomic_dec_and_test(&p->refcnt))
1207 if (p->policy == MPOL_BIND)
1208 kfree(p->v.zonelist);
1209 p->policy = MPOL_DEFAULT;
1210 kmem_cache_free(policy_cache, p);
1214 * Shared memory backing store policy support.
1216 * Remember policies even when nobody has shared memory mapped.
1217 * The policies are kept in Red-Black tree linked from the inode.
1218 * They are protected by the sp->lock spinlock, which should be held
1219 * for any accesses to the tree.
1222 /* lookup first element intersecting start-end */
1223 /* Caller holds sp->lock */
1224 static struct sp_node *
1225 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1227 struct rb_node *n = sp->root.rb_node;
1230 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1232 if (start >= p->end)
1234 else if (end <= p->start)
1242 struct sp_node *w = NULL;
1243 struct rb_node *prev = rb_prev(n);
1246 w = rb_entry(prev, struct sp_node, nd);
1247 if (w->end <= start)
1251 return rb_entry(n, struct sp_node, nd);
1254 /* Insert a new shared policy into the list. */
1255 /* Caller holds sp->lock */
1256 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1258 struct rb_node **p = &sp->root.rb_node;
1259 struct rb_node *parent = NULL;
1264 nd = rb_entry(parent, struct sp_node, nd);
1265 if (new->start < nd->start)
1267 else if (new->end > nd->end)
1268 p = &(*p)->rb_right;
1272 rb_link_node(&new->nd, parent, p);
1273 rb_insert_color(&new->nd, &sp->root);
1274 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1275 new->policy ? new->policy->policy : 0);
1278 /* Find shared policy intersecting idx */
1280 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1282 struct mempolicy *pol = NULL;
1285 if (!sp->root.rb_node)
1287 spin_lock(&sp->lock);
1288 sn = sp_lookup(sp, idx, idx+1);
1290 mpol_get(sn->policy);
1293 spin_unlock(&sp->lock);
1297 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1299 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1300 rb_erase(&n->nd, &sp->root);
1301 mpol_free(n->policy);
1302 kmem_cache_free(sn_cache, n);
1306 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1308 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1319 /* Replace a policy range. */
1320 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1321 unsigned long end, struct sp_node *new)
1323 struct sp_node *n, *new2 = NULL;
1326 spin_lock(&sp->lock);
1327 n = sp_lookup(sp, start, end);
1328 /* Take care of old policies in the same range. */
1329 while (n && n->start < end) {
1330 struct rb_node *next = rb_next(&n->nd);
1331 if (n->start >= start) {
1337 /* Old policy spanning whole new range. */
1340 spin_unlock(&sp->lock);
1341 new2 = sp_alloc(end, n->end, n->policy);
1347 sp_insert(sp, new2);
1355 n = rb_entry(next, struct sp_node, nd);
1359 spin_unlock(&sp->lock);
1361 mpol_free(new2->policy);
1362 kmem_cache_free(sn_cache, new2);
1367 void mpol_shared_policy_init(struct shared_policy *info, int policy,
1368 nodemask_t *policy_nodes)
1370 info->root = RB_ROOT;
1371 spin_lock_init(&info->lock);
1373 if (policy != MPOL_DEFAULT) {
1374 struct mempolicy *newpol;
1376 /* Falls back to MPOL_DEFAULT on any error */
1377 newpol = mpol_new(policy, policy_nodes);
1378 if (!IS_ERR(newpol)) {
1379 /* Create pseudo-vma that contains just the policy */
1380 struct vm_area_struct pvma;
1382 memset(&pvma, 0, sizeof(struct vm_area_struct));
1383 /* Policy covers entire file */
1384 pvma.vm_end = TASK_SIZE;
1385 mpol_set_shared_policy(info, &pvma, newpol);
1391 int mpol_set_shared_policy(struct shared_policy *info,
1392 struct vm_area_struct *vma, struct mempolicy *npol)
1395 struct sp_node *new = NULL;
1396 unsigned long sz = vma_pages(vma);
1398 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1400 sz, npol? npol->policy : -1,
1401 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1404 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1408 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1410 kmem_cache_free(sn_cache, new);
1414 /* Free a backing policy store on inode delete. */
1415 void mpol_free_shared_policy(struct shared_policy *p)
1418 struct rb_node *next;
1420 if (!p->root.rb_node)
1422 spin_lock(&p->lock);
1423 next = rb_first(&p->root);
1425 n = rb_entry(next, struct sp_node, nd);
1426 next = rb_next(&n->nd);
1427 rb_erase(&n->nd, &p->root);
1428 mpol_free(n->policy);
1429 kmem_cache_free(sn_cache, n);
1431 spin_unlock(&p->lock);
1434 /* assumes fs == KERNEL_DS */
1435 void __init numa_policy_init(void)
1437 policy_cache = kmem_cache_create("numa_policy",
1438 sizeof(struct mempolicy),
1439 0, SLAB_PANIC, NULL, NULL);
1441 sn_cache = kmem_cache_create("shared_policy_node",
1442 sizeof(struct sp_node),
1443 0, SLAB_PANIC, NULL, NULL);
1445 /* Set interleaving policy for system init. This way not all
1446 the data structures allocated at system boot end up in node zero. */
1448 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1449 printk("numa_policy_init: interleaving failed\n");
1452 /* Reset policy of current process to default */
1453 void numa_default_policy(void)
1455 do_set_mempolicy(MPOL_DEFAULT, NULL);
1458 /* Migrate a policy to a different set of nodes */
1459 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1461 nodemask_t *mpolmask;
1466 mpolmask = &pol->cpuset_mems_allowed;
1467 if (nodes_equal(*mpolmask, *newmask))
1470 switch (pol->policy) {
1473 case MPOL_INTERLEAVE:
1474 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1476 *mpolmask = *newmask;
1477 current->il_next = node_remap(current->il_next,
1478 *mpolmask, *newmask);
1480 case MPOL_PREFERRED:
1481 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1482 *mpolmask, *newmask);
1483 *mpolmask = *newmask;
1488 struct zonelist *zonelist;
1491 for (z = pol->v.zonelist->zones; *z; z++)
1492 node_set((*z)->zone_pgdat->node_id, nodes);
1493 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1496 zonelist = bind_zonelist(&nodes);
1498 /* If no mem, then zonelist is NULL and we keep old zonelist.
1499 * If that old zonelist has no remaining mems_allowed nodes,
1500 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1504 /* Good - got mem - substitute new zonelist */
1505 kfree(pol->v.zonelist);
1506 pol->v.zonelist = zonelist;
1508 *mpolmask = *newmask;
1518 * Wrapper for mpol_rebind_policy() that just requires task
1519 * pointer, and updates task mempolicy.
1522 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1524 mpol_rebind_policy(tsk->mempolicy, new);
1528 * Rebind each vma in mm to new nodemask.
1530 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1533 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1535 struct vm_area_struct *vma;
1537 down_write(&mm->mmap_sem);
1538 for (vma = mm->mmap; vma; vma = vma->vm_next)
1539 mpol_rebind_policy(vma->vm_policy, new);
1540 up_write(&mm->mmap_sem);
1544 * Display pages allocated per node and memory policy via /proc.
1547 static const char *policy_types[] = { "default", "prefer", "bind",
1551 * Convert a mempolicy into a string.
1552 * Returns the number of characters in buffer (if positive)
1553 * or an error (negative)
1555 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1560 int mode = pol ? pol->policy : MPOL_DEFAULT;
1567 case MPOL_PREFERRED:
1569 node_set(pol->v.preferred_node, nodes);
1573 get_zonemask(pol, &nodes);
1576 case MPOL_INTERLEAVE:
1577 nodes = pol->v.nodes;
1585 l = strlen(policy_types[mode]);
1586 if (buffer + maxlen < p + l + 1)
1589 strcpy(p, policy_types[mode]);
1592 if (!nodes_empty(nodes)) {
1593 if (buffer + maxlen < p + 2)
1596 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1602 unsigned long pages;
1604 unsigned long mapped;
1605 unsigned long mapcount_max;
1606 unsigned long node[MAX_NUMNODES];
1609 static void gather_stats(struct page *page, void *private)
1611 struct numa_maps *md = private;
1612 int count = page_mapcount(page);
1617 if (count > md->mapcount_max)
1618 md->mapcount_max = count;
1625 md->node[page_to_nid(page)]++;
1629 int show_numa_map(struct seq_file *m, void *v)
1631 struct task_struct *task = m->private;
1632 struct vm_area_struct *vma = v;
1633 struct numa_maps *md;
1640 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1644 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1645 &node_online_map, MPOL_MF_STATS, md);
1648 mpol_to_str(buffer, sizeof(buffer),
1649 get_vma_policy(task, vma, vma->vm_start));
1651 seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1652 vma->vm_start, buffer, md->pages,
1653 md->mapped, md->mapcount_max);
1656 seq_printf(m," anon=%lu",md->anon);
1658 for_each_online_node(n)
1660 seq_printf(m, " N%d=%lu", n, md->node[n]);
1666 if (m->count < m->size)
1667 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;