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);
211 nid = page_to_nid(page);
212 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
215 if (flags & MPOL_MF_STATS)
216 gather_stats(page, private);
217 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
219 migrate_page_add(vma, page, private, flags);
224 } while (pte++, addr += PAGE_SIZE, addr != end);
225 pte_unmap_unlock(orig_pte, ptl);
229 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
230 unsigned long addr, unsigned long end,
231 const nodemask_t *nodes, unsigned long flags,
237 pmd = pmd_offset(pud, addr);
239 next = pmd_addr_end(addr, end);
240 if (pmd_none_or_clear_bad(pmd))
242 if (check_pte_range(vma, pmd, addr, next, nodes,
245 } while (pmd++, addr = next, addr != end);
249 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
250 unsigned long addr, unsigned long end,
251 const nodemask_t *nodes, unsigned long flags,
257 pud = pud_offset(pgd, addr);
259 next = pud_addr_end(addr, end);
260 if (pud_none_or_clear_bad(pud))
262 if (check_pmd_range(vma, pud, addr, next, nodes,
265 } while (pud++, addr = next, addr != end);
269 static inline int check_pgd_range(struct vm_area_struct *vma,
270 unsigned long addr, unsigned long end,
271 const nodemask_t *nodes, unsigned long flags,
277 pgd = pgd_offset(vma->vm_mm, addr);
279 next = pgd_addr_end(addr, end);
280 if (pgd_none_or_clear_bad(pgd))
282 if (check_pud_range(vma, pgd, addr, next, nodes,
285 } while (pgd++, addr = next, addr != end);
289 /* Check if a vma is migratable */
290 static inline int vma_migratable(struct vm_area_struct *vma)
292 if (vma->vm_flags & (
293 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP))
299 * Check if all pages in a range are on a set of nodes.
300 * If pagelist != NULL then isolate pages from the LRU and
301 * put them on the pagelist.
303 static struct vm_area_struct *
304 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
305 const nodemask_t *nodes, unsigned long flags, void *private)
308 struct vm_area_struct *first, *vma, *prev;
310 first = find_vma(mm, start);
312 return ERR_PTR(-EFAULT);
314 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
315 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
316 if (!vma->vm_next && vma->vm_end < end)
317 return ERR_PTR(-EFAULT);
318 if (prev && prev->vm_end < vma->vm_start)
319 return ERR_PTR(-EFAULT);
321 if (!is_vm_hugetlb_page(vma) &&
322 ((flags & MPOL_MF_STRICT) ||
323 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
324 vma_migratable(vma)))) {
325 unsigned long endvma = vma->vm_end;
329 if (vma->vm_start > start)
330 start = vma->vm_start;
331 err = check_pgd_range(vma, start, endvma, nodes,
334 first = ERR_PTR(err);
343 /* Apply policy to a single VMA */
344 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
347 struct mempolicy *old = vma->vm_policy;
349 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
350 vma->vm_start, vma->vm_end, vma->vm_pgoff,
351 vma->vm_ops, vma->vm_file,
352 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
354 if (vma->vm_ops && vma->vm_ops->set_policy)
355 err = vma->vm_ops->set_policy(vma, new);
358 vma->vm_policy = new;
364 /* Step 2: apply policy to a range and do splits. */
365 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
366 unsigned long end, struct mempolicy *new)
368 struct vm_area_struct *next;
372 for (; vma && vma->vm_start < end; vma = next) {
374 if (vma->vm_start < start)
375 err = split_vma(vma->vm_mm, vma, start, 1);
376 if (!err && vma->vm_end > end)
377 err = split_vma(vma->vm_mm, vma, end, 0);
379 err = policy_vma(vma, new);
386 static int contextualize_policy(int mode, nodemask_t *nodes)
391 cpuset_update_task_memory_state();
392 if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
394 return mpol_check_policy(mode, nodes);
397 /* Set the process memory policy */
398 long do_set_mempolicy(int mode, nodemask_t *nodes)
400 struct mempolicy *new;
402 if (contextualize_policy(mode, nodes))
404 new = mpol_new(mode, nodes);
407 mpol_free(current->mempolicy);
408 current->mempolicy = new;
409 if (new && new->policy == MPOL_INTERLEAVE)
410 current->il_next = first_node(new->v.nodes);
414 /* Fill a zone bitmap for a policy */
415 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
422 for (i = 0; p->v.zonelist->zones[i]; i++)
423 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
428 case MPOL_INTERLEAVE:
432 /* or use current node instead of online map? */
433 if (p->v.preferred_node < 0)
434 *nodes = node_online_map;
436 node_set(p->v.preferred_node, *nodes);
443 static int lookup_node(struct mm_struct *mm, unsigned long addr)
448 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
450 err = page_to_nid(p);
456 /* Retrieve NUMA policy */
457 long do_get_mempolicy(int *policy, nodemask_t *nmask,
458 unsigned long addr, unsigned long flags)
461 struct mm_struct *mm = current->mm;
462 struct vm_area_struct *vma = NULL;
463 struct mempolicy *pol = current->mempolicy;
465 cpuset_update_task_memory_state();
466 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
468 if (flags & MPOL_F_ADDR) {
469 down_read(&mm->mmap_sem);
470 vma = find_vma_intersection(mm, addr, addr+1);
472 up_read(&mm->mmap_sem);
475 if (vma->vm_ops && vma->vm_ops->get_policy)
476 pol = vma->vm_ops->get_policy(vma, addr);
478 pol = vma->vm_policy;
483 pol = &default_policy;
485 if (flags & MPOL_F_NODE) {
486 if (flags & MPOL_F_ADDR) {
487 err = lookup_node(mm, addr);
491 } else if (pol == current->mempolicy &&
492 pol->policy == MPOL_INTERLEAVE) {
493 *policy = current->il_next;
499 *policy = pol->policy;
502 up_read(¤t->mm->mmap_sem);
508 get_zonemask(pol, nmask);
512 up_read(¤t->mm->mmap_sem);
520 /* Check if we are the only process mapping the page in question */
521 static inline int single_mm_mapping(struct mm_struct *mm,
522 struct address_space *mapping)
524 struct vm_area_struct *vma;
525 struct prio_tree_iter iter;
528 spin_lock(&mapping->i_mmap_lock);
529 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
530 if (mm != vma->vm_mm) {
534 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
535 if (mm != vma->vm_mm) {
540 spin_unlock(&mapping->i_mmap_lock);
545 * Add a page to be migrated to the pagelist
547 static void migrate_page_add(struct vm_area_struct *vma,
548 struct page *page, struct list_head *pagelist, unsigned long flags)
551 * Avoid migrating a page that is shared by others and not writable.
553 if ((flags & MPOL_MF_MOVE_ALL) || !page->mapping || PageAnon(page) ||
554 mapping_writably_mapped(page->mapping) ||
555 single_mm_mapping(vma->vm_mm, page->mapping)) {
556 int rc = isolate_lru_page(page);
559 list_add(&page->lru, pagelist);
561 * If the isolate attempt was not successful then we just
562 * encountered an unswappable page. Something must be wrong.
568 static int swap_pages(struct list_head *pagelist)
574 n = migrate_pages(pagelist, NULL, &moved, &failed);
575 putback_lru_pages(&failed);
576 putback_lru_pages(&moved);
582 * For now migrate_pages simply swaps out the pages from nodes that are in
583 * the source set but not in the target set. In the future, we would
584 * want a function that moves pages between the two nodesets in such
585 * a way as to preserve the physical layout as much as possible.
587 * Returns the number of page that could not be moved.
589 int do_migrate_pages(struct mm_struct *mm,
590 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
596 nodes_andnot(nodes, *from_nodes, *to_nodes);
598 down_read(&mm->mmap_sem);
599 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nodes,
600 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
602 if (!list_empty(&pagelist)) {
603 count = swap_pages(&pagelist);
604 putback_lru_pages(&pagelist);
607 up_read(&mm->mmap_sem);
611 long do_mbind(unsigned long start, unsigned long len,
612 unsigned long mode, nodemask_t *nmask, unsigned long flags)
614 struct vm_area_struct *vma;
615 struct mm_struct *mm = current->mm;
616 struct mempolicy *new;
621 if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
622 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
625 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
628 if (start & ~PAGE_MASK)
631 if (mode == MPOL_DEFAULT)
632 flags &= ~MPOL_MF_STRICT;
634 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
642 if (mpol_check_policy(mode, nmask))
645 new = mpol_new(mode, nmask);
650 * If we are using the default policy then operation
651 * on discontinuous address spaces is okay after all
654 flags |= MPOL_MF_DISCONTIG_OK;
656 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
657 mode,nodes_addr(nodes)[0]);
659 down_write(&mm->mmap_sem);
660 vma = check_range(mm, start, end, nmask,
661 flags | MPOL_MF_INVERT, &pagelist);
667 err = mbind_range(vma, start, end, new);
668 if (!list_empty(&pagelist))
669 nr_failed = swap_pages(&pagelist);
671 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
674 if (!list_empty(&pagelist))
675 putback_lru_pages(&pagelist);
677 up_write(&mm->mmap_sem);
683 * User space interface with variable sized bitmaps for nodelists.
686 /* Copy a node mask from user space. */
687 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
688 unsigned long maxnode)
691 unsigned long nlongs;
692 unsigned long endmask;
696 if (maxnode == 0 || !nmask)
699 nlongs = BITS_TO_LONGS(maxnode);
700 if ((maxnode % BITS_PER_LONG) == 0)
703 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
705 /* When the user specified more nodes than supported just check
706 if the non supported part is all zero. */
707 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
708 if (nlongs > PAGE_SIZE/sizeof(long))
710 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
712 if (get_user(t, nmask + k))
714 if (k == nlongs - 1) {
720 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
724 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
726 nodes_addr(*nodes)[nlongs-1] &= endmask;
730 /* Copy a kernel node mask to user space */
731 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
734 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
735 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
738 if (copy > PAGE_SIZE)
740 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
744 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
747 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
749 unsigned long __user *nmask, unsigned long maxnode,
755 err = get_nodes(&nodes, nmask, maxnode);
758 return do_mbind(start, len, mode, &nodes, flags);
761 /* Set the process memory policy */
762 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
763 unsigned long maxnode)
768 if (mode < 0 || mode > MPOL_MAX)
770 err = get_nodes(&nodes, nmask, maxnode);
773 return do_set_mempolicy(mode, &nodes);
776 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
777 const unsigned long __user *old_nodes,
778 const unsigned long __user *new_nodes)
780 struct mm_struct *mm;
781 struct task_struct *task;
784 nodemask_t task_nodes;
787 err = get_nodes(&old, old_nodes, maxnode);
791 err = get_nodes(&new, new_nodes, maxnode);
795 /* Find the mm_struct */
796 read_lock(&tasklist_lock);
797 task = pid ? find_task_by_pid(pid) : current;
799 read_unlock(&tasklist_lock);
802 mm = get_task_mm(task);
803 read_unlock(&tasklist_lock);
809 * Check if this process has the right to modify the specified
810 * process. The right exists if the process has administrative
811 * capabilities, superuser priviledges or the same
812 * userid as the target process.
814 if ((current->euid != task->suid) && (current->euid != task->uid) &&
815 (current->uid != task->suid) && (current->uid != task->uid) &&
816 !capable(CAP_SYS_ADMIN)) {
821 task_nodes = cpuset_mems_allowed(task);
822 /* Is the user allowed to access the target nodes? */
823 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
828 err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
835 /* Retrieve NUMA policy */
836 asmlinkage long sys_get_mempolicy(int __user *policy,
837 unsigned long __user *nmask,
838 unsigned long maxnode,
839 unsigned long addr, unsigned long flags)
844 if (nmask != NULL && maxnode < MAX_NUMNODES)
847 err = do_get_mempolicy(&pval, &nodes, addr, flags);
852 if (policy && put_user(pval, policy))
856 err = copy_nodes_to_user(nmask, maxnode, &nodes);
863 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
864 compat_ulong_t __user *nmask,
865 compat_ulong_t maxnode,
866 compat_ulong_t addr, compat_ulong_t flags)
869 unsigned long __user *nm = NULL;
870 unsigned long nr_bits, alloc_size;
871 DECLARE_BITMAP(bm, MAX_NUMNODES);
873 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
874 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
877 nm = compat_alloc_user_space(alloc_size);
879 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
882 err = copy_from_user(bm, nm, alloc_size);
883 /* ensure entire bitmap is zeroed */
884 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
885 err |= compat_put_bitmap(nmask, bm, nr_bits);
891 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
892 compat_ulong_t maxnode)
895 unsigned long __user *nm = NULL;
896 unsigned long nr_bits, alloc_size;
897 DECLARE_BITMAP(bm, MAX_NUMNODES);
899 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
900 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
903 err = compat_get_bitmap(bm, nmask, nr_bits);
904 nm = compat_alloc_user_space(alloc_size);
905 err |= copy_to_user(nm, bm, alloc_size);
911 return sys_set_mempolicy(mode, nm, nr_bits+1);
914 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
915 compat_ulong_t mode, compat_ulong_t __user *nmask,
916 compat_ulong_t maxnode, compat_ulong_t flags)
919 unsigned long __user *nm = NULL;
920 unsigned long nr_bits, alloc_size;
923 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
924 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
927 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
928 nm = compat_alloc_user_space(alloc_size);
929 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
935 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
940 /* Return effective policy for a VMA */
941 static struct mempolicy * get_vma_policy(struct task_struct *task,
942 struct vm_area_struct *vma, unsigned long addr)
944 struct mempolicy *pol = task->mempolicy;
947 if (vma->vm_ops && vma->vm_ops->get_policy)
948 pol = vma->vm_ops->get_policy(vma, addr);
949 else if (vma->vm_policy &&
950 vma->vm_policy->policy != MPOL_DEFAULT)
951 pol = vma->vm_policy;
954 pol = &default_policy;
958 /* Return a zonelist representing a mempolicy */
959 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
963 switch (policy->policy) {
965 nd = policy->v.preferred_node;
970 /* Lower zones don't get a policy applied */
971 /* Careful: current->mems_allowed might have moved */
972 if (gfp_zone(gfp) >= policy_zone)
973 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
974 return policy->v.zonelist;
976 case MPOL_INTERLEAVE: /* should not happen */
984 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
987 /* Do dynamic interleaving for a process */
988 static unsigned interleave_nodes(struct mempolicy *policy)
991 struct task_struct *me = current;
994 next = next_node(nid, policy->v.nodes);
995 if (next >= MAX_NUMNODES)
996 next = first_node(policy->v.nodes);
1001 /* Do static interleaving for a VMA with known offset. */
1002 static unsigned offset_il_node(struct mempolicy *pol,
1003 struct vm_area_struct *vma, unsigned long off)
1005 unsigned nnodes = nodes_weight(pol->v.nodes);
1006 unsigned target = (unsigned)off % nnodes;
1012 nid = next_node(nid, pol->v.nodes);
1014 } while (c <= target);
1018 /* Determine a node number for interleave */
1019 static inline unsigned interleave_nid(struct mempolicy *pol,
1020 struct vm_area_struct *vma, unsigned long addr, int shift)
1025 off = vma->vm_pgoff;
1026 off += (addr - vma->vm_start) >> shift;
1027 return offset_il_node(pol, vma, off);
1029 return interleave_nodes(pol);
1032 /* Return a zonelist suitable for a huge page allocation. */
1033 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1035 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1037 if (pol->policy == MPOL_INTERLEAVE) {
1040 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1041 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1043 return zonelist_policy(GFP_HIGHUSER, pol);
1046 /* Allocate a page in interleaved policy.
1047 Own path because it needs to do special accounting. */
1048 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1051 struct zonelist *zl;
1054 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1055 page = __alloc_pages(gfp, order, zl);
1056 if (page && page_zone(page) == zl->zones[0]) {
1057 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1064 * alloc_page_vma - Allocate a page for a VMA.
1067 * %GFP_USER user allocation.
1068 * %GFP_KERNEL kernel allocations,
1069 * %GFP_HIGHMEM highmem/user allocations,
1070 * %GFP_FS allocation should not call back into a file system.
1071 * %GFP_ATOMIC don't sleep.
1073 * @vma: Pointer to VMA or NULL if not available.
1074 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1076 * This function allocates a page from the kernel page pool and applies
1077 * a NUMA policy associated with the VMA or the current process.
1078 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1079 * mm_struct of the VMA to prevent it from going away. Should be used for
1080 * all allocations for pages that will be mapped into
1081 * user space. Returns NULL when no page can be allocated.
1083 * Should be called with the mm_sem of the vma hold.
1086 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1088 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1090 cpuset_update_task_memory_state();
1092 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1095 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1096 return alloc_page_interleave(gfp, 0, nid);
1098 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1102 * alloc_pages_current - Allocate pages.
1105 * %GFP_USER user allocation,
1106 * %GFP_KERNEL kernel allocation,
1107 * %GFP_HIGHMEM highmem allocation,
1108 * %GFP_FS don't call back into a file system.
1109 * %GFP_ATOMIC don't sleep.
1110 * @order: Power of two of allocation size in pages. 0 is a single page.
1112 * Allocate a page from the kernel page pool. When not in
1113 * interrupt context and apply the current process NUMA policy.
1114 * Returns NULL when no page can be allocated.
1116 * Don't call cpuset_update_task_memory_state() unless
1117 * 1) it's ok to take cpuset_sem (can WAIT), and
1118 * 2) allocating for current task (not interrupt).
1120 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1122 struct mempolicy *pol = current->mempolicy;
1124 if ((gfp & __GFP_WAIT) && !in_interrupt())
1125 cpuset_update_task_memory_state();
1126 if (!pol || in_interrupt())
1127 pol = &default_policy;
1128 if (pol->policy == MPOL_INTERLEAVE)
1129 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1130 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1132 EXPORT_SYMBOL(alloc_pages_current);
1134 /* Slow path of a mempolicy copy */
1135 struct mempolicy *__mpol_copy(struct mempolicy *old)
1137 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1140 return ERR_PTR(-ENOMEM);
1142 atomic_set(&new->refcnt, 1);
1143 if (new->policy == MPOL_BIND) {
1144 int sz = ksize(old->v.zonelist);
1145 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1146 if (!new->v.zonelist) {
1147 kmem_cache_free(policy_cache, new);
1148 return ERR_PTR(-ENOMEM);
1150 memcpy(new->v.zonelist, old->v.zonelist, sz);
1155 /* Slow path of a mempolicy comparison */
1156 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1160 if (a->policy != b->policy)
1162 switch (a->policy) {
1165 case MPOL_INTERLEAVE:
1166 return nodes_equal(a->v.nodes, b->v.nodes);
1167 case MPOL_PREFERRED:
1168 return a->v.preferred_node == b->v.preferred_node;
1171 for (i = 0; a->v.zonelist->zones[i]; i++)
1172 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1174 return b->v.zonelist->zones[i] == NULL;
1182 /* Slow path of a mpol destructor. */
1183 void __mpol_free(struct mempolicy *p)
1185 if (!atomic_dec_and_test(&p->refcnt))
1187 if (p->policy == MPOL_BIND)
1188 kfree(p->v.zonelist);
1189 p->policy = MPOL_DEFAULT;
1190 kmem_cache_free(policy_cache, p);
1194 * Shared memory backing store policy support.
1196 * Remember policies even when nobody has shared memory mapped.
1197 * The policies are kept in Red-Black tree linked from the inode.
1198 * They are protected by the sp->lock spinlock, which should be held
1199 * for any accesses to the tree.
1202 /* lookup first element intersecting start-end */
1203 /* Caller holds sp->lock */
1204 static struct sp_node *
1205 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1207 struct rb_node *n = sp->root.rb_node;
1210 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1212 if (start >= p->end)
1214 else if (end <= p->start)
1222 struct sp_node *w = NULL;
1223 struct rb_node *prev = rb_prev(n);
1226 w = rb_entry(prev, struct sp_node, nd);
1227 if (w->end <= start)
1231 return rb_entry(n, struct sp_node, nd);
1234 /* Insert a new shared policy into the list. */
1235 /* Caller holds sp->lock */
1236 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1238 struct rb_node **p = &sp->root.rb_node;
1239 struct rb_node *parent = NULL;
1244 nd = rb_entry(parent, struct sp_node, nd);
1245 if (new->start < nd->start)
1247 else if (new->end > nd->end)
1248 p = &(*p)->rb_right;
1252 rb_link_node(&new->nd, parent, p);
1253 rb_insert_color(&new->nd, &sp->root);
1254 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1255 new->policy ? new->policy->policy : 0);
1258 /* Find shared policy intersecting idx */
1260 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1262 struct mempolicy *pol = NULL;
1265 if (!sp->root.rb_node)
1267 spin_lock(&sp->lock);
1268 sn = sp_lookup(sp, idx, idx+1);
1270 mpol_get(sn->policy);
1273 spin_unlock(&sp->lock);
1277 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1279 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1280 rb_erase(&n->nd, &sp->root);
1281 mpol_free(n->policy);
1282 kmem_cache_free(sn_cache, n);
1286 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1288 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1299 /* Replace a policy range. */
1300 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1301 unsigned long end, struct sp_node *new)
1303 struct sp_node *n, *new2 = NULL;
1306 spin_lock(&sp->lock);
1307 n = sp_lookup(sp, start, end);
1308 /* Take care of old policies in the same range. */
1309 while (n && n->start < end) {
1310 struct rb_node *next = rb_next(&n->nd);
1311 if (n->start >= start) {
1317 /* Old policy spanning whole new range. */
1320 spin_unlock(&sp->lock);
1321 new2 = sp_alloc(end, n->end, n->policy);
1327 sp_insert(sp, new2);
1335 n = rb_entry(next, struct sp_node, nd);
1339 spin_unlock(&sp->lock);
1341 mpol_free(new2->policy);
1342 kmem_cache_free(sn_cache, new2);
1347 int mpol_set_shared_policy(struct shared_policy *info,
1348 struct vm_area_struct *vma, struct mempolicy *npol)
1351 struct sp_node *new = NULL;
1352 unsigned long sz = vma_pages(vma);
1354 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1356 sz, npol? npol->policy : -1,
1357 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1360 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1364 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1366 kmem_cache_free(sn_cache, new);
1370 /* Free a backing policy store on inode delete. */
1371 void mpol_free_shared_policy(struct shared_policy *p)
1374 struct rb_node *next;
1376 if (!p->root.rb_node)
1378 spin_lock(&p->lock);
1379 next = rb_first(&p->root);
1381 n = rb_entry(next, struct sp_node, nd);
1382 next = rb_next(&n->nd);
1383 rb_erase(&n->nd, &p->root);
1384 mpol_free(n->policy);
1385 kmem_cache_free(sn_cache, n);
1387 spin_unlock(&p->lock);
1390 /* assumes fs == KERNEL_DS */
1391 void __init numa_policy_init(void)
1393 policy_cache = kmem_cache_create("numa_policy",
1394 sizeof(struct mempolicy),
1395 0, SLAB_PANIC, NULL, NULL);
1397 sn_cache = kmem_cache_create("shared_policy_node",
1398 sizeof(struct sp_node),
1399 0, SLAB_PANIC, NULL, NULL);
1401 /* Set interleaving policy for system init. This way not all
1402 the data structures allocated at system boot end up in node zero. */
1404 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1405 printk("numa_policy_init: interleaving failed\n");
1408 /* Reset policy of current process to default */
1409 void numa_default_policy(void)
1411 do_set_mempolicy(MPOL_DEFAULT, NULL);
1414 /* Migrate a policy to a different set of nodes */
1415 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1417 nodemask_t *mpolmask;
1422 mpolmask = &pol->cpuset_mems_allowed;
1423 if (nodes_equal(*mpolmask, *newmask))
1426 switch (pol->policy) {
1429 case MPOL_INTERLEAVE:
1430 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1432 *mpolmask = *newmask;
1433 current->il_next = node_remap(current->il_next,
1434 *mpolmask, *newmask);
1436 case MPOL_PREFERRED:
1437 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1438 *mpolmask, *newmask);
1439 *mpolmask = *newmask;
1444 struct zonelist *zonelist;
1447 for (z = pol->v.zonelist->zones; *z; z++)
1448 node_set((*z)->zone_pgdat->node_id, nodes);
1449 nodes_remap(tmp, nodes, *mpolmask, *newmask);
1452 zonelist = bind_zonelist(&nodes);
1454 /* If no mem, then zonelist is NULL and we keep old zonelist.
1455 * If that old zonelist has no remaining mems_allowed nodes,
1456 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1460 /* Good - got mem - substitute new zonelist */
1461 kfree(pol->v.zonelist);
1462 pol->v.zonelist = zonelist;
1464 *mpolmask = *newmask;
1474 * Wrapper for mpol_rebind_policy() that just requires task
1475 * pointer, and updates task mempolicy.
1478 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1480 mpol_rebind_policy(tsk->mempolicy, new);
1484 * Display pages allocated per node and memory policy via /proc.
1487 static const char *policy_types[] = { "default", "prefer", "bind",
1491 * Convert a mempolicy into a string.
1492 * Returns the number of characters in buffer (if positive)
1493 * or an error (negative)
1495 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1500 int mode = pol ? pol->policy : MPOL_DEFAULT;
1507 case MPOL_PREFERRED:
1509 node_set(pol->v.preferred_node, nodes);
1513 get_zonemask(pol, &nodes);
1516 case MPOL_INTERLEAVE:
1517 nodes = pol->v.nodes;
1525 l = strlen(policy_types[mode]);
1526 if (buffer + maxlen < p + l + 1)
1529 strcpy(p, policy_types[mode]);
1532 if (!nodes_empty(nodes)) {
1533 if (buffer + maxlen < p + 2)
1536 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1542 unsigned long pages;
1544 unsigned long mapped;
1545 unsigned long mapcount_max;
1546 unsigned long node[MAX_NUMNODES];
1549 static void gather_stats(struct page *page, void *private)
1551 struct numa_maps *md = private;
1552 int count = page_mapcount(page);
1557 if (count > md->mapcount_max)
1558 md->mapcount_max = count;
1565 md->node[page_to_nid(page)]++;
1569 int show_numa_map(struct seq_file *m, void *v)
1571 struct task_struct *task = m->private;
1572 struct vm_area_struct *vma = v;
1573 struct numa_maps *md;
1580 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1584 check_pgd_range(vma, vma->vm_start, vma->vm_end,
1585 &node_online_map, MPOL_MF_STATS, md);
1588 mpol_to_str(buffer, sizeof(buffer),
1589 get_vma_policy(task, vma, vma->vm_start));
1591 seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1592 vma->vm_start, buffer, md->pages,
1593 md->mapped, md->mapcount_max);
1596 seq_printf(m," anon=%lu",md->anon);
1598 for_each_online_node(n)
1600 seq_printf(m, " N%d=%lu", n, md->node[n]);
1606 if (m->count < m->size)
1607 m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;