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.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/module.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
94 #include <asm/tlbflush.h>
95 #include <asm/uaccess.h>
100 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
101 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
102 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
104 static struct kmem_cache *policy_cache;
105 static struct kmem_cache *sn_cache;
107 /* Highest zone. An specific allocation for a zone below that is not
109 enum zone_type policy_zone = 0;
112 * run-time system-wide default policy => local allocation
114 struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
120 static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
128 * If we have a lock to protect task->mempolicy in read-side, we do
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
138 } mpol_ops[MPOL_MAX];
140 /* Check that the nodemask contains at least one populated zone */
141 static int is_valid_nodemask(const nodemask_t *nodemask)
145 for_each_node_mask(nd, *nodemask) {
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
158 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
160 return pol->flags & MPOL_MODE_FLAGS;
163 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
171 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
173 if (nodes_empty(*nodes))
175 pol->v.nodes = *nodes;
179 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
186 pol->v.preferred_node = first_node(*nodes);
190 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
192 if (!is_valid_nodemask(nodes))
194 pol->v.nodes = *nodes;
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
207 static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
238 ret = mpol_ops[pol->mode].create(pol, NULL);
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
246 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
249 struct mempolicy *policy;
251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 return NULL; /* simply delete any existing policy */
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
279 policy->flags = flags;
284 /* Slow path of a mpol destructor. */
285 void __mpol_put(struct mempolicy *p)
287 if (!atomic_dec_and_test(&p->refcnt))
289 kmem_cache_free(policy_cache, p);
292 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
303 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
328 if (nodes_empty(tmp))
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
347 static void mpol_rebind_preferred(struct mempolicy *pol,
348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
356 if (node_isset(node, *nodes)) {
357 pol->v.preferred_node = node;
358 pol->flags &= ~MPOL_F_LOCAL;
360 pol->flags |= MPOL_F_LOCAL;
361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
368 pol->w.cpuset_mems_allowed = *nodes;
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
380 * If we have a lock to protect task->mempolicy in read-side, we do
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
388 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
393 if (!mpol_store_user_nodemask(pol) && step == 0 &&
394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
417 * Called with task's alloc_lock held.
420 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
423 mpol_rebind_policy(tsk->mempolicy, new, step);
427 * Rebind each vma in mm to new nodemask.
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
432 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
434 struct vm_area_struct *vma;
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
439 up_write(&mm->mmap_sem);
442 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
444 .rebind = mpol_rebind_default,
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
460 static void gather_stats(struct page *, void *, int pte_dirty);
461 static void migrate_page_add(struct page *page, struct list_head *pagelist,
462 unsigned long flags);
464 /* Scan through pages checking if pages follow certain conditions. */
465 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
466 unsigned long addr, unsigned long end,
467 const nodemask_t *nodes, unsigned long flags,
474 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
479 if (!pte_present(*pte))
481 page = vm_normal_page(vma, addr, *pte);
485 * vm_normal_page() filters out zero pages, but there might
486 * still be PageReserved pages to skip, perhaps in a VDSO.
487 * And we cannot move PageKsm pages sensibly or safely yet.
489 if (PageReserved(page) || PageKsm(page))
491 nid = page_to_nid(page);
492 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
495 if (flags & MPOL_MF_STATS)
496 gather_stats(page, private, pte_dirty(*pte));
497 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
498 migrate_page_add(page, private, flags);
501 } while (pte++, addr += PAGE_SIZE, addr != end);
502 pte_unmap_unlock(orig_pte, ptl);
506 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
507 unsigned long addr, unsigned long end,
508 const nodemask_t *nodes, unsigned long flags,
514 pmd = pmd_offset(pud, addr);
516 next = pmd_addr_end(addr, end);
517 if (pmd_none_or_clear_bad(pmd))
519 if (check_pte_range(vma, pmd, addr, next, nodes,
522 } while (pmd++, addr = next, addr != end);
526 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
527 unsigned long addr, unsigned long end,
528 const nodemask_t *nodes, unsigned long flags,
534 pud = pud_offset(pgd, addr);
536 next = pud_addr_end(addr, end);
537 if (pud_none_or_clear_bad(pud))
539 if (check_pmd_range(vma, pud, addr, next, nodes,
542 } while (pud++, addr = next, addr != end);
546 static inline int check_pgd_range(struct vm_area_struct *vma,
547 unsigned long addr, unsigned long end,
548 const nodemask_t *nodes, unsigned long flags,
554 pgd = pgd_offset(vma->vm_mm, addr);
556 next = pgd_addr_end(addr, end);
557 if (pgd_none_or_clear_bad(pgd))
559 if (check_pud_range(vma, pgd, addr, next, nodes,
562 } while (pgd++, addr = next, addr != end);
567 * Check if all pages in a range are on a set of nodes.
568 * If pagelist != NULL then isolate pages from the LRU and
569 * put them on the pagelist.
571 static struct vm_area_struct *
572 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
573 const nodemask_t *nodes, unsigned long flags, void *private)
576 struct vm_area_struct *first, *vma, *prev;
579 first = find_vma(mm, start);
581 return ERR_PTR(-EFAULT);
583 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
584 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
585 if (!vma->vm_next && vma->vm_end < end)
586 return ERR_PTR(-EFAULT);
587 if (prev && prev->vm_end < vma->vm_start)
588 return ERR_PTR(-EFAULT);
590 if (!is_vm_hugetlb_page(vma) &&
591 ((flags & MPOL_MF_STRICT) ||
592 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
593 vma_migratable(vma)))) {
594 unsigned long endvma = vma->vm_end;
598 if (vma->vm_start > start)
599 start = vma->vm_start;
600 err = check_pgd_range(vma, start, endvma, nodes,
603 first = ERR_PTR(err);
612 /* Apply policy to a single VMA */
613 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
616 struct mempolicy *old = vma->vm_policy;
618 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
619 vma->vm_start, vma->vm_end, vma->vm_pgoff,
620 vma->vm_ops, vma->vm_file,
621 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
623 if (vma->vm_ops && vma->vm_ops->set_policy)
624 err = vma->vm_ops->set_policy(vma, new);
627 vma->vm_policy = new;
633 /* Step 2: apply policy to a range and do splits. */
634 static int mbind_range(struct mm_struct *mm, unsigned long start,
635 unsigned long end, struct mempolicy *new_pol)
637 struct vm_area_struct *next;
638 struct vm_area_struct *prev;
639 struct vm_area_struct *vma;
642 unsigned long vmstart;
645 vma = find_vma_prev(mm, start, &prev);
646 if (!vma || vma->vm_start > start)
649 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
651 vmstart = max(start, vma->vm_start);
652 vmend = min(end, vma->vm_end);
654 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
655 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
656 vma->anon_vma, vma->vm_file, pgoff, new_pol);
662 if (vma->vm_start != vmstart) {
663 err = split_vma(vma->vm_mm, vma, vmstart, 1);
667 if (vma->vm_end != vmend) {
668 err = split_vma(vma->vm_mm, vma, vmend, 0);
672 err = policy_vma(vma, new_pol);
682 * Update task->flags PF_MEMPOLICY bit: set iff non-default
683 * mempolicy. Allows more rapid checking of this (combined perhaps
684 * with other PF_* flag bits) on memory allocation hot code paths.
686 * If called from outside this file, the task 'p' should -only- be
687 * a newly forked child not yet visible on the task list, because
688 * manipulating the task flags of a visible task is not safe.
690 * The above limitation is why this routine has the funny name
691 * mpol_fix_fork_child_flag().
693 * It is also safe to call this with a task pointer of current,
694 * which the static wrapper mpol_set_task_struct_flag() does,
695 * for use within this file.
698 void mpol_fix_fork_child_flag(struct task_struct *p)
701 p->flags |= PF_MEMPOLICY;
703 p->flags &= ~PF_MEMPOLICY;
706 static void mpol_set_task_struct_flag(void)
708 mpol_fix_fork_child_flag(current);
711 /* Set the process memory policy */
712 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
715 struct mempolicy *new, *old;
716 struct mm_struct *mm = current->mm;
717 NODEMASK_SCRATCH(scratch);
723 new = mpol_new(mode, flags, nodes);
729 * prevent changing our mempolicy while show_numa_maps()
731 * Note: do_set_mempolicy() can be called at init time
735 down_write(&mm->mmap_sem);
737 ret = mpol_set_nodemask(new, nodes, scratch);
739 task_unlock(current);
741 up_write(&mm->mmap_sem);
745 old = current->mempolicy;
746 current->mempolicy = new;
747 mpol_set_task_struct_flag();
748 if (new && new->mode == MPOL_INTERLEAVE &&
749 nodes_weight(new->v.nodes))
750 current->il_next = first_node(new->v.nodes);
751 task_unlock(current);
753 up_write(&mm->mmap_sem);
758 NODEMASK_SCRATCH_FREE(scratch);
763 * Return nodemask for policy for get_mempolicy() query
765 * Called with task's alloc_lock held
767 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
770 if (p == &default_policy)
776 case MPOL_INTERLEAVE:
780 if (!(p->flags & MPOL_F_LOCAL))
781 node_set(p->v.preferred_node, *nodes);
782 /* else return empty node mask for local allocation */
789 static int lookup_node(struct mm_struct *mm, unsigned long addr)
794 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
796 err = page_to_nid(p);
802 /* Retrieve NUMA policy */
803 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
804 unsigned long addr, unsigned long flags)
807 struct mm_struct *mm = current->mm;
808 struct vm_area_struct *vma = NULL;
809 struct mempolicy *pol = current->mempolicy;
812 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
815 if (flags & MPOL_F_MEMS_ALLOWED) {
816 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
818 *policy = 0; /* just so it's initialized */
820 *nmask = cpuset_current_mems_allowed;
821 task_unlock(current);
825 if (flags & MPOL_F_ADDR) {
827 * Do NOT fall back to task policy if the
828 * vma/shared policy at addr is NULL. We
829 * want to return MPOL_DEFAULT in this case.
831 down_read(&mm->mmap_sem);
832 vma = find_vma_intersection(mm, addr, addr+1);
834 up_read(&mm->mmap_sem);
837 if (vma->vm_ops && vma->vm_ops->get_policy)
838 pol = vma->vm_ops->get_policy(vma, addr);
840 pol = vma->vm_policy;
845 pol = &default_policy; /* indicates default behavior */
847 if (flags & MPOL_F_NODE) {
848 if (flags & MPOL_F_ADDR) {
849 err = lookup_node(mm, addr);
853 } else if (pol == current->mempolicy &&
854 pol->mode == MPOL_INTERLEAVE) {
855 *policy = current->il_next;
861 *policy = pol == &default_policy ? MPOL_DEFAULT :
864 * Internal mempolicy flags must be masked off before exposing
865 * the policy to userspace.
867 *policy |= (pol->flags & MPOL_MODE_FLAGS);
871 up_read(¤t->mm->mmap_sem);
877 if (mpol_store_user_nodemask(pol)) {
878 *nmask = pol->w.user_nodemask;
881 get_policy_nodemask(pol, nmask);
882 task_unlock(current);
889 up_read(¤t->mm->mmap_sem);
893 #ifdef CONFIG_MIGRATION
897 static void migrate_page_add(struct page *page, struct list_head *pagelist,
901 * Avoid migrating a page that is shared with others.
903 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
904 if (!isolate_lru_page(page)) {
905 list_add_tail(&page->lru, pagelist);
906 inc_zone_page_state(page, NR_ISOLATED_ANON +
907 page_is_file_cache(page));
912 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
914 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
918 * Migrate pages from one node to a target node.
919 * Returns error or the number of pages not migrated.
921 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
929 node_set(source, nmask);
931 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
932 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
934 if (!list_empty(&pagelist)) {
935 err = migrate_pages(&pagelist, new_node_page, dest, 0);
937 putback_lru_pages(&pagelist);
944 * Move pages between the two nodesets so as to preserve the physical
945 * layout as much as possible.
947 * Returns the number of page that could not be moved.
949 int do_migrate_pages(struct mm_struct *mm,
950 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
956 err = migrate_prep();
960 down_read(&mm->mmap_sem);
962 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
967 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
968 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
969 * bit in 'tmp', and return that <source, dest> pair for migration.
970 * The pair of nodemasks 'to' and 'from' define the map.
972 * If no pair of bits is found that way, fallback to picking some
973 * pair of 'source' and 'dest' bits that are not the same. If the
974 * 'source' and 'dest' bits are the same, this represents a node
975 * that will be migrating to itself, so no pages need move.
977 * If no bits are left in 'tmp', or if all remaining bits left
978 * in 'tmp' correspond to the same bit in 'to', return false
979 * (nothing left to migrate).
981 * This lets us pick a pair of nodes to migrate between, such that
982 * if possible the dest node is not already occupied by some other
983 * source node, minimizing the risk of overloading the memory on a
984 * node that would happen if we migrated incoming memory to a node
985 * before migrating outgoing memory source that same node.
987 * A single scan of tmp is sufficient. As we go, we remember the
988 * most recent <s, d> pair that moved (s != d). If we find a pair
989 * that not only moved, but what's better, moved to an empty slot
990 * (d is not set in tmp), then we break out then, with that pair.
991 * Otherwise when we finish scannng from_tmp, we at least have the
992 * most recent <s, d> pair that moved. If we get all the way through
993 * the scan of tmp without finding any node that moved, much less
994 * moved to an empty node, then there is nothing left worth migrating.
998 while (!nodes_empty(tmp)) {
1003 for_each_node_mask(s, tmp) {
1004 d = node_remap(s, *from_nodes, *to_nodes);
1008 source = s; /* Node moved. Memorize */
1011 /* dest not in remaining from nodes? */
1012 if (!node_isset(dest, tmp))
1018 node_clear(source, tmp);
1019 err = migrate_to_node(mm, source, dest, flags);
1026 up_read(&mm->mmap_sem);
1034 * Allocate a new page for page migration based on vma policy.
1035 * Start assuming that page is mapped by vma pointed to by @private.
1036 * Search forward from there, if not. N.B., this assumes that the
1037 * list of pages handed to migrate_pages()--which is how we get here--
1038 * is in virtual address order.
1040 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1042 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1043 unsigned long uninitialized_var(address);
1046 address = page_address_in_vma(page, vma);
1047 if (address != -EFAULT)
1053 * if !vma, alloc_page_vma() will use task or system default policy
1055 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1059 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1060 unsigned long flags)
1064 int do_migrate_pages(struct mm_struct *mm,
1065 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1070 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1076 static long do_mbind(unsigned long start, unsigned long len,
1077 unsigned short mode, unsigned short mode_flags,
1078 nodemask_t *nmask, unsigned long flags)
1080 struct vm_area_struct *vma;
1081 struct mm_struct *mm = current->mm;
1082 struct mempolicy *new;
1085 LIST_HEAD(pagelist);
1087 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1088 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1090 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1093 if (start & ~PAGE_MASK)
1096 if (mode == MPOL_DEFAULT)
1097 flags &= ~MPOL_MF_STRICT;
1099 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1107 new = mpol_new(mode, mode_flags, nmask);
1109 return PTR_ERR(new);
1112 * If we are using the default policy then operation
1113 * on discontinuous address spaces is okay after all
1116 flags |= MPOL_MF_DISCONTIG_OK;
1118 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1119 start, start + len, mode, mode_flags,
1120 nmask ? nodes_addr(*nmask)[0] : -1);
1122 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1124 err = migrate_prep();
1129 NODEMASK_SCRATCH(scratch);
1131 down_write(&mm->mmap_sem);
1133 err = mpol_set_nodemask(new, nmask, scratch);
1134 task_unlock(current);
1136 up_write(&mm->mmap_sem);
1139 NODEMASK_SCRATCH_FREE(scratch);
1144 vma = check_range(mm, start, end, nmask,
1145 flags | MPOL_MF_INVERT, &pagelist);
1151 err = mbind_range(mm, start, end, new);
1153 if (!list_empty(&pagelist)) {
1154 nr_failed = migrate_pages(&pagelist, new_vma_page,
1155 (unsigned long)vma, 0);
1157 putback_lru_pages(&pagelist);
1160 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1163 putback_lru_pages(&pagelist);
1165 up_write(&mm->mmap_sem);
1172 * User space interface with variable sized bitmaps for nodelists.
1175 /* Copy a node mask from user space. */
1176 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1177 unsigned long maxnode)
1180 unsigned long nlongs;
1181 unsigned long endmask;
1184 nodes_clear(*nodes);
1185 if (maxnode == 0 || !nmask)
1187 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1190 nlongs = BITS_TO_LONGS(maxnode);
1191 if ((maxnode % BITS_PER_LONG) == 0)
1194 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1196 /* When the user specified more nodes than supported just check
1197 if the non supported part is all zero. */
1198 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1199 if (nlongs > PAGE_SIZE/sizeof(long))
1201 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1203 if (get_user(t, nmask + k))
1205 if (k == nlongs - 1) {
1211 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1215 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1217 nodes_addr(*nodes)[nlongs-1] &= endmask;
1221 /* Copy a kernel node mask to user space */
1222 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1225 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1226 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1228 if (copy > nbytes) {
1229 if (copy > PAGE_SIZE)
1231 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1235 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1238 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1239 unsigned long, mode, unsigned long __user *, nmask,
1240 unsigned long, maxnode, unsigned, flags)
1244 unsigned short mode_flags;
1246 mode_flags = mode & MPOL_MODE_FLAGS;
1247 mode &= ~MPOL_MODE_FLAGS;
1248 if (mode >= MPOL_MAX)
1250 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1251 (mode_flags & MPOL_F_RELATIVE_NODES))
1253 err = get_nodes(&nodes, nmask, maxnode);
1256 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1259 /* Set the process memory policy */
1260 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1261 unsigned long, maxnode)
1265 unsigned short flags;
1267 flags = mode & MPOL_MODE_FLAGS;
1268 mode &= ~MPOL_MODE_FLAGS;
1269 if ((unsigned int)mode >= MPOL_MAX)
1271 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1273 err = get_nodes(&nodes, nmask, maxnode);
1276 return do_set_mempolicy(mode, flags, &nodes);
1279 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1280 const unsigned long __user *, old_nodes,
1281 const unsigned long __user *, new_nodes)
1283 const struct cred *cred = current_cred(), *tcred;
1284 struct mm_struct *mm = NULL;
1285 struct task_struct *task;
1286 nodemask_t task_nodes;
1290 NODEMASK_SCRATCH(scratch);
1295 old = &scratch->mask1;
1296 new = &scratch->mask2;
1298 err = get_nodes(old, old_nodes, maxnode);
1302 err = get_nodes(new, new_nodes, maxnode);
1306 /* Find the mm_struct */
1307 read_lock(&tasklist_lock);
1308 task = pid ? find_task_by_vpid(pid) : current;
1310 read_unlock(&tasklist_lock);
1314 mm = get_task_mm(task);
1315 read_unlock(&tasklist_lock);
1322 * Check if this process has the right to modify the specified
1323 * process. The right exists if the process has administrative
1324 * capabilities, superuser privileges or the same
1325 * userid as the target process.
1328 tcred = __task_cred(task);
1329 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1330 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1331 !capable(CAP_SYS_NICE)) {
1338 task_nodes = cpuset_mems_allowed(task);
1339 /* Is the user allowed to access the target nodes? */
1340 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1345 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1350 err = security_task_movememory(task);
1354 err = do_migrate_pages(mm, old, new,
1355 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1359 NODEMASK_SCRATCH_FREE(scratch);
1365 /* Retrieve NUMA policy */
1366 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1367 unsigned long __user *, nmask, unsigned long, maxnode,
1368 unsigned long, addr, unsigned long, flags)
1371 int uninitialized_var(pval);
1374 if (nmask != NULL && maxnode < MAX_NUMNODES)
1377 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1382 if (policy && put_user(pval, policy))
1386 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1391 #ifdef CONFIG_COMPAT
1393 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1394 compat_ulong_t __user *nmask,
1395 compat_ulong_t maxnode,
1396 compat_ulong_t addr, compat_ulong_t flags)
1399 unsigned long __user *nm = NULL;
1400 unsigned long nr_bits, alloc_size;
1401 DECLARE_BITMAP(bm, MAX_NUMNODES);
1403 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1404 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1407 nm = compat_alloc_user_space(alloc_size);
1409 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1411 if (!err && nmask) {
1412 err = copy_from_user(bm, nm, alloc_size);
1413 /* ensure entire bitmap is zeroed */
1414 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1415 err |= compat_put_bitmap(nmask, bm, nr_bits);
1421 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1422 compat_ulong_t maxnode)
1425 unsigned long __user *nm = NULL;
1426 unsigned long nr_bits, alloc_size;
1427 DECLARE_BITMAP(bm, MAX_NUMNODES);
1429 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1430 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1433 err = compat_get_bitmap(bm, nmask, nr_bits);
1434 nm = compat_alloc_user_space(alloc_size);
1435 err |= copy_to_user(nm, bm, alloc_size);
1441 return sys_set_mempolicy(mode, nm, nr_bits+1);
1444 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1445 compat_ulong_t mode, compat_ulong_t __user *nmask,
1446 compat_ulong_t maxnode, compat_ulong_t flags)
1449 unsigned long __user *nm = NULL;
1450 unsigned long nr_bits, alloc_size;
1453 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1454 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1457 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1458 nm = compat_alloc_user_space(alloc_size);
1459 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1465 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1471 * get_vma_policy(@task, @vma, @addr)
1472 * @task - task for fallback if vma policy == default
1473 * @vma - virtual memory area whose policy is sought
1474 * @addr - address in @vma for shared policy lookup
1476 * Returns effective policy for a VMA at specified address.
1477 * Falls back to @task or system default policy, as necessary.
1478 * Current or other task's task mempolicy and non-shared vma policies
1479 * are protected by the task's mmap_sem, which must be held for read by
1481 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1482 * count--added by the get_policy() vm_op, as appropriate--to protect against
1483 * freeing by another task. It is the caller's responsibility to free the
1484 * extra reference for shared policies.
1486 static struct mempolicy *get_vma_policy(struct task_struct *task,
1487 struct vm_area_struct *vma, unsigned long addr)
1489 struct mempolicy *pol = task->mempolicy;
1492 if (vma->vm_ops && vma->vm_ops->get_policy) {
1493 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1497 } else if (vma->vm_policy)
1498 pol = vma->vm_policy;
1501 pol = &default_policy;
1506 * Return a nodemask representing a mempolicy for filtering nodes for
1509 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1511 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1512 if (unlikely(policy->mode == MPOL_BIND) &&
1513 gfp_zone(gfp) >= policy_zone &&
1514 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1515 return &policy->v.nodes;
1520 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1521 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1523 int nd = numa_node_id();
1525 switch (policy->mode) {
1526 case MPOL_PREFERRED:
1527 if (!(policy->flags & MPOL_F_LOCAL))
1528 nd = policy->v.preferred_node;
1532 * Normally, MPOL_BIND allocations are node-local within the
1533 * allowed nodemask. However, if __GFP_THISNODE is set and the
1534 * current node isn't part of the mask, we use the zonelist for
1535 * the first node in the mask instead.
1537 if (unlikely(gfp & __GFP_THISNODE) &&
1538 unlikely(!node_isset(nd, policy->v.nodes)))
1539 nd = first_node(policy->v.nodes);
1544 return node_zonelist(nd, gfp);
1547 /* Do dynamic interleaving for a process */
1548 static unsigned interleave_nodes(struct mempolicy *policy)
1551 struct task_struct *me = current;
1554 next = next_node(nid, policy->v.nodes);
1555 if (next >= MAX_NUMNODES)
1556 next = first_node(policy->v.nodes);
1557 if (next < MAX_NUMNODES)
1563 * Depending on the memory policy provide a node from which to allocate the
1565 * @policy must be protected by freeing by the caller. If @policy is
1566 * the current task's mempolicy, this protection is implicit, as only the
1567 * task can change it's policy. The system default policy requires no
1570 unsigned slab_node(struct mempolicy *policy)
1572 if (!policy || policy->flags & MPOL_F_LOCAL)
1573 return numa_node_id();
1575 switch (policy->mode) {
1576 case MPOL_PREFERRED:
1578 * handled MPOL_F_LOCAL above
1580 return policy->v.preferred_node;
1582 case MPOL_INTERLEAVE:
1583 return interleave_nodes(policy);
1587 * Follow bind policy behavior and start allocation at the
1590 struct zonelist *zonelist;
1592 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1593 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1594 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1605 /* Do static interleaving for a VMA with known offset. */
1606 static unsigned offset_il_node(struct mempolicy *pol,
1607 struct vm_area_struct *vma, unsigned long off)
1609 unsigned nnodes = nodes_weight(pol->v.nodes);
1615 return numa_node_id();
1616 target = (unsigned int)off % nnodes;
1619 nid = next_node(nid, pol->v.nodes);
1621 } while (c <= target);
1625 /* Determine a node number for interleave */
1626 static inline unsigned interleave_nid(struct mempolicy *pol,
1627 struct vm_area_struct *vma, unsigned long addr, int shift)
1633 * for small pages, there is no difference between
1634 * shift and PAGE_SHIFT, so the bit-shift is safe.
1635 * for huge pages, since vm_pgoff is in units of small
1636 * pages, we need to shift off the always 0 bits to get
1639 BUG_ON(shift < PAGE_SHIFT);
1640 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1641 off += (addr - vma->vm_start) >> shift;
1642 return offset_il_node(pol, vma, off);
1644 return interleave_nodes(pol);
1647 #ifdef CONFIG_HUGETLBFS
1649 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1650 * @vma = virtual memory area whose policy is sought
1651 * @addr = address in @vma for shared policy lookup and interleave policy
1652 * @gfp_flags = for requested zone
1653 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1654 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1656 * Returns a zonelist suitable for a huge page allocation and a pointer
1657 * to the struct mempolicy for conditional unref after allocation.
1658 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1659 * @nodemask for filtering the zonelist.
1661 * Must be protected by get_mems_allowed()
1663 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1664 gfp_t gfp_flags, struct mempolicy **mpol,
1665 nodemask_t **nodemask)
1667 struct zonelist *zl;
1669 *mpol = get_vma_policy(current, vma, addr);
1670 *nodemask = NULL; /* assume !MPOL_BIND */
1672 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1673 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1674 huge_page_shift(hstate_vma(vma))), gfp_flags);
1676 zl = policy_zonelist(gfp_flags, *mpol);
1677 if ((*mpol)->mode == MPOL_BIND)
1678 *nodemask = &(*mpol)->v.nodes;
1684 * init_nodemask_of_mempolicy
1686 * If the current task's mempolicy is "default" [NULL], return 'false'
1687 * to indicate default policy. Otherwise, extract the policy nodemask
1688 * for 'bind' or 'interleave' policy into the argument nodemask, or
1689 * initialize the argument nodemask to contain the single node for
1690 * 'preferred' or 'local' policy and return 'true' to indicate presence
1691 * of non-default mempolicy.
1693 * We don't bother with reference counting the mempolicy [mpol_get/put]
1694 * because the current task is examining it's own mempolicy and a task's
1695 * mempolicy is only ever changed by the task itself.
1697 * N.B., it is the caller's responsibility to free a returned nodemask.
1699 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1701 struct mempolicy *mempolicy;
1704 if (!(mask && current->mempolicy))
1708 mempolicy = current->mempolicy;
1709 switch (mempolicy->mode) {
1710 case MPOL_PREFERRED:
1711 if (mempolicy->flags & MPOL_F_LOCAL)
1712 nid = numa_node_id();
1714 nid = mempolicy->v.preferred_node;
1715 init_nodemask_of_node(mask, nid);
1720 case MPOL_INTERLEAVE:
1721 *mask = mempolicy->v.nodes;
1727 task_unlock(current);
1734 * mempolicy_nodemask_intersects
1736 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1737 * policy. Otherwise, check for intersection between mask and the policy
1738 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1739 * policy, always return true since it may allocate elsewhere on fallback.
1741 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1743 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1744 const nodemask_t *mask)
1746 struct mempolicy *mempolicy;
1752 mempolicy = tsk->mempolicy;
1756 switch (mempolicy->mode) {
1757 case MPOL_PREFERRED:
1759 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1760 * allocate from, they may fallback to other nodes when oom.
1761 * Thus, it's possible for tsk to have allocated memory from
1766 case MPOL_INTERLEAVE:
1767 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1777 /* Allocate a page in interleaved policy.
1778 Own path because it needs to do special accounting. */
1779 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1782 struct zonelist *zl;
1785 zl = node_zonelist(nid, gfp);
1786 page = __alloc_pages(gfp, order, zl);
1787 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1788 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1793 * alloc_page_vma - Allocate a page for a VMA.
1796 * %GFP_USER user allocation.
1797 * %GFP_KERNEL kernel allocations,
1798 * %GFP_HIGHMEM highmem/user allocations,
1799 * %GFP_FS allocation should not call back into a file system.
1800 * %GFP_ATOMIC don't sleep.
1802 * @vma: Pointer to VMA or NULL if not available.
1803 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1805 * This function allocates a page from the kernel page pool and applies
1806 * a NUMA policy associated with the VMA or the current process.
1807 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1808 * mm_struct of the VMA to prevent it from going away. Should be used for
1809 * all allocations for pages that will be mapped into
1810 * user space. Returns NULL when no page can be allocated.
1812 * Should be called with the mm_sem of the vma hold.
1815 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1817 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1818 struct zonelist *zl;
1822 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1825 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1827 page = alloc_page_interleave(gfp, 0, nid);
1831 zl = policy_zonelist(gfp, pol);
1832 if (unlikely(mpol_needs_cond_ref(pol))) {
1834 * slow path: ref counted shared policy
1836 struct page *page = __alloc_pages_nodemask(gfp, 0,
1837 zl, policy_nodemask(gfp, pol));
1843 * fast path: default or task policy
1845 page = __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1851 * alloc_pages_current - Allocate pages.
1854 * %GFP_USER user allocation,
1855 * %GFP_KERNEL kernel allocation,
1856 * %GFP_HIGHMEM highmem allocation,
1857 * %GFP_FS don't call back into a file system.
1858 * %GFP_ATOMIC don't sleep.
1859 * @order: Power of two of allocation size in pages. 0 is a single page.
1861 * Allocate a page from the kernel page pool. When not in
1862 * interrupt context and apply the current process NUMA policy.
1863 * Returns NULL when no page can be allocated.
1865 * Don't call cpuset_update_task_memory_state() unless
1866 * 1) it's ok to take cpuset_sem (can WAIT), and
1867 * 2) allocating for current task (not interrupt).
1869 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1871 struct mempolicy *pol = current->mempolicy;
1874 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1875 pol = &default_policy;
1879 * No reference counting needed for current->mempolicy
1880 * nor system default_policy
1882 if (pol->mode == MPOL_INTERLEAVE)
1883 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1885 page = __alloc_pages_nodemask(gfp, order,
1886 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1890 EXPORT_SYMBOL(alloc_pages_current);
1893 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1894 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1895 * with the mems_allowed returned by cpuset_mems_allowed(). This
1896 * keeps mempolicies cpuset relative after its cpuset moves. See
1897 * further kernel/cpuset.c update_nodemask().
1899 * current's mempolicy may be rebinded by the other task(the task that changes
1900 * cpuset's mems), so we needn't do rebind work for current task.
1903 /* Slow path of a mempolicy duplicate */
1904 struct mempolicy *__mpol_dup(struct mempolicy *old)
1906 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1909 return ERR_PTR(-ENOMEM);
1911 /* task's mempolicy is protected by alloc_lock */
1912 if (old == current->mempolicy) {
1915 task_unlock(current);
1920 if (current_cpuset_is_being_rebound()) {
1921 nodemask_t mems = cpuset_mems_allowed(current);
1922 if (new->flags & MPOL_F_REBINDING)
1923 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1925 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1928 atomic_set(&new->refcnt, 1);
1933 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1934 * eliminate the * MPOL_F_* flags that require conditional ref and
1935 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1936 * after return. Use the returned value.
1938 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1939 * policy lookup, even if the policy needs/has extra ref on lookup.
1940 * shmem_readahead needs this.
1942 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1943 struct mempolicy *frompol)
1945 if (!mpol_needs_cond_ref(frompol))
1949 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1950 __mpol_put(frompol);
1954 /* Slow path of a mempolicy comparison */
1955 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1959 if (a->mode != b->mode)
1961 if (a->flags != b->flags)
1963 if (mpol_store_user_nodemask(a))
1964 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1970 case MPOL_INTERLEAVE:
1971 return nodes_equal(a->v.nodes, b->v.nodes);
1972 case MPOL_PREFERRED:
1973 return a->v.preferred_node == b->v.preferred_node &&
1974 a->flags == b->flags;
1982 * Shared memory backing store policy support.
1984 * Remember policies even when nobody has shared memory mapped.
1985 * The policies are kept in Red-Black tree linked from the inode.
1986 * They are protected by the sp->lock spinlock, which should be held
1987 * for any accesses to the tree.
1990 /* lookup first element intersecting start-end */
1991 /* Caller holds sp->lock */
1992 static struct sp_node *
1993 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1995 struct rb_node *n = sp->root.rb_node;
1998 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2000 if (start >= p->end)
2002 else if (end <= p->start)
2010 struct sp_node *w = NULL;
2011 struct rb_node *prev = rb_prev(n);
2014 w = rb_entry(prev, struct sp_node, nd);
2015 if (w->end <= start)
2019 return rb_entry(n, struct sp_node, nd);
2022 /* Insert a new shared policy into the list. */
2023 /* Caller holds sp->lock */
2024 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2026 struct rb_node **p = &sp->root.rb_node;
2027 struct rb_node *parent = NULL;
2032 nd = rb_entry(parent, struct sp_node, nd);
2033 if (new->start < nd->start)
2035 else if (new->end > nd->end)
2036 p = &(*p)->rb_right;
2040 rb_link_node(&new->nd, parent, p);
2041 rb_insert_color(&new->nd, &sp->root);
2042 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2043 new->policy ? new->policy->mode : 0);
2046 /* Find shared policy intersecting idx */
2048 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2050 struct mempolicy *pol = NULL;
2053 if (!sp->root.rb_node)
2055 spin_lock(&sp->lock);
2056 sn = sp_lookup(sp, idx, idx+1);
2058 mpol_get(sn->policy);
2061 spin_unlock(&sp->lock);
2065 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2067 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2068 rb_erase(&n->nd, &sp->root);
2069 mpol_put(n->policy);
2070 kmem_cache_free(sn_cache, n);
2073 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2074 struct mempolicy *pol)
2076 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2083 pol->flags |= MPOL_F_SHARED; /* for unref */
2088 /* Replace a policy range. */
2089 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2090 unsigned long end, struct sp_node *new)
2092 struct sp_node *n, *new2 = NULL;
2095 spin_lock(&sp->lock);
2096 n = sp_lookup(sp, start, end);
2097 /* Take care of old policies in the same range. */
2098 while (n && n->start < end) {
2099 struct rb_node *next = rb_next(&n->nd);
2100 if (n->start >= start) {
2106 /* Old policy spanning whole new range. */
2109 spin_unlock(&sp->lock);
2110 new2 = sp_alloc(end, n->end, n->policy);
2116 sp_insert(sp, new2);
2124 n = rb_entry(next, struct sp_node, nd);
2128 spin_unlock(&sp->lock);
2130 mpol_put(new2->policy);
2131 kmem_cache_free(sn_cache, new2);
2137 * mpol_shared_policy_init - initialize shared policy for inode
2138 * @sp: pointer to inode shared policy
2139 * @mpol: struct mempolicy to install
2141 * Install non-NULL @mpol in inode's shared policy rb-tree.
2142 * On entry, the current task has a reference on a non-NULL @mpol.
2143 * This must be released on exit.
2144 * This is called at get_inode() calls and we can use GFP_KERNEL.
2146 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2150 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2151 spin_lock_init(&sp->lock);
2154 struct vm_area_struct pvma;
2155 struct mempolicy *new;
2156 NODEMASK_SCRATCH(scratch);
2160 /* contextualize the tmpfs mount point mempolicy */
2161 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2163 goto free_scratch; /* no valid nodemask intersection */
2166 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2167 task_unlock(current);
2171 /* Create pseudo-vma that contains just the policy */
2172 memset(&pvma, 0, sizeof(struct vm_area_struct));
2173 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2174 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2177 mpol_put(new); /* drop initial ref */
2179 NODEMASK_SCRATCH_FREE(scratch);
2181 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2185 int mpol_set_shared_policy(struct shared_policy *info,
2186 struct vm_area_struct *vma, struct mempolicy *npol)
2189 struct sp_node *new = NULL;
2190 unsigned long sz = vma_pages(vma);
2192 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2194 sz, npol ? npol->mode : -1,
2195 npol ? npol->flags : -1,
2196 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2199 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2203 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2205 kmem_cache_free(sn_cache, new);
2209 /* Free a backing policy store on inode delete. */
2210 void mpol_free_shared_policy(struct shared_policy *p)
2213 struct rb_node *next;
2215 if (!p->root.rb_node)
2217 spin_lock(&p->lock);
2218 next = rb_first(&p->root);
2220 n = rb_entry(next, struct sp_node, nd);
2221 next = rb_next(&n->nd);
2222 rb_erase(&n->nd, &p->root);
2223 mpol_put(n->policy);
2224 kmem_cache_free(sn_cache, n);
2226 spin_unlock(&p->lock);
2229 /* assumes fs == KERNEL_DS */
2230 void __init numa_policy_init(void)
2232 nodemask_t interleave_nodes;
2233 unsigned long largest = 0;
2234 int nid, prefer = 0;
2236 policy_cache = kmem_cache_create("numa_policy",
2237 sizeof(struct mempolicy),
2238 0, SLAB_PANIC, NULL);
2240 sn_cache = kmem_cache_create("shared_policy_node",
2241 sizeof(struct sp_node),
2242 0, SLAB_PANIC, NULL);
2245 * Set interleaving policy for system init. Interleaving is only
2246 * enabled across suitably sized nodes (default is >= 16MB), or
2247 * fall back to the largest node if they're all smaller.
2249 nodes_clear(interleave_nodes);
2250 for_each_node_state(nid, N_HIGH_MEMORY) {
2251 unsigned long total_pages = node_present_pages(nid);
2253 /* Preserve the largest node */
2254 if (largest < total_pages) {
2255 largest = total_pages;
2259 /* Interleave this node? */
2260 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2261 node_set(nid, interleave_nodes);
2264 /* All too small, use the largest */
2265 if (unlikely(nodes_empty(interleave_nodes)))
2266 node_set(prefer, interleave_nodes);
2268 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2269 printk("numa_policy_init: interleaving failed\n");
2272 /* Reset policy of current process to default */
2273 void numa_default_policy(void)
2275 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2279 * Parse and format mempolicy from/to strings
2283 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2284 * Used only for mpol_parse_str() and mpol_to_str()
2286 #define MPOL_LOCAL MPOL_MAX
2287 static const char * const policy_modes[] =
2289 [MPOL_DEFAULT] = "default",
2290 [MPOL_PREFERRED] = "prefer",
2291 [MPOL_BIND] = "bind",
2292 [MPOL_INTERLEAVE] = "interleave",
2293 [MPOL_LOCAL] = "local"
2299 * mpol_parse_str - parse string to mempolicy
2300 * @str: string containing mempolicy to parse
2301 * @mpol: pointer to struct mempolicy pointer, returned on success.
2302 * @no_context: flag whether to "contextualize" the mempolicy
2305 * <mode>[=<flags>][:<nodelist>]
2307 * if @no_context is true, save the input nodemask in w.user_nodemask in
2308 * the returned mempolicy. This will be used to "clone" the mempolicy in
2309 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2310 * mount option. Note that if 'static' or 'relative' mode flags were
2311 * specified, the input nodemask will already have been saved. Saving
2312 * it again is redundant, but safe.
2314 * On success, returns 0, else 1
2316 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2318 struct mempolicy *new = NULL;
2319 unsigned short mode;
2320 unsigned short uninitialized_var(mode_flags);
2322 char *nodelist = strchr(str, ':');
2323 char *flags = strchr(str, '=');
2327 /* NUL-terminate mode or flags string */
2329 if (nodelist_parse(nodelist, nodes))
2331 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2337 *flags++ = '\0'; /* terminate mode string */
2339 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2340 if (!strcmp(str, policy_modes[mode])) {
2344 if (mode > MPOL_LOCAL)
2348 case MPOL_PREFERRED:
2350 * Insist on a nodelist of one node only
2353 char *rest = nodelist;
2354 while (isdigit(*rest))
2360 case MPOL_INTERLEAVE:
2362 * Default to online nodes with memory if no nodelist
2365 nodes = node_states[N_HIGH_MEMORY];
2369 * Don't allow a nodelist; mpol_new() checks flags
2373 mode = MPOL_PREFERRED;
2377 * Insist on a empty nodelist
2384 * Insist on a nodelist
2393 * Currently, we only support two mutually exclusive
2396 if (!strcmp(flags, "static"))
2397 mode_flags |= MPOL_F_STATIC_NODES;
2398 else if (!strcmp(flags, "relative"))
2399 mode_flags |= MPOL_F_RELATIVE_NODES;
2404 new = mpol_new(mode, mode_flags, &nodes);
2409 /* save for contextualization */
2410 new->w.user_nodemask = nodes;
2413 NODEMASK_SCRATCH(scratch);
2416 ret = mpol_set_nodemask(new, &nodes, scratch);
2417 task_unlock(current);
2420 NODEMASK_SCRATCH_FREE(scratch);
2429 /* Restore string for error message */
2438 #endif /* CONFIG_TMPFS */
2441 * mpol_to_str - format a mempolicy structure for printing
2442 * @buffer: to contain formatted mempolicy string
2443 * @maxlen: length of @buffer
2444 * @pol: pointer to mempolicy to be formatted
2445 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2447 * Convert a mempolicy into a string.
2448 * Returns the number of characters in buffer (if positive)
2449 * or an error (negative)
2451 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2456 unsigned short mode;
2457 unsigned short flags = pol ? pol->flags : 0;
2460 * Sanity check: room for longest mode, flag and some nodes
2462 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2464 if (!pol || pol == &default_policy)
2465 mode = MPOL_DEFAULT;
2474 case MPOL_PREFERRED:
2476 if (flags & MPOL_F_LOCAL)
2477 mode = MPOL_LOCAL; /* pseudo-policy */
2479 node_set(pol->v.preferred_node, nodes);
2484 case MPOL_INTERLEAVE:
2486 nodes = pol->w.user_nodemask;
2488 nodes = pol->v.nodes;
2495 l = strlen(policy_modes[mode]);
2496 if (buffer + maxlen < p + l + 1)
2499 strcpy(p, policy_modes[mode]);
2502 if (flags & MPOL_MODE_FLAGS) {
2503 if (buffer + maxlen < p + 2)
2508 * Currently, the only defined flags are mutually exclusive
2510 if (flags & MPOL_F_STATIC_NODES)
2511 p += snprintf(p, buffer + maxlen - p, "static");
2512 else if (flags & MPOL_F_RELATIVE_NODES)
2513 p += snprintf(p, buffer + maxlen - p, "relative");
2516 if (!nodes_empty(nodes)) {
2517 if (buffer + maxlen < p + 2)
2520 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2526 unsigned long pages;
2528 unsigned long active;
2529 unsigned long writeback;
2530 unsigned long mapcount_max;
2531 unsigned long dirty;
2532 unsigned long swapcache;
2533 unsigned long node[MAX_NUMNODES];
2536 static void gather_stats(struct page *page, void *private, int pte_dirty)
2538 struct numa_maps *md = private;
2539 int count = page_mapcount(page);
2542 if (pte_dirty || PageDirty(page))
2545 if (PageSwapCache(page))
2548 if (PageActive(page) || PageUnevictable(page))
2551 if (PageWriteback(page))
2557 if (count > md->mapcount_max)
2558 md->mapcount_max = count;
2560 md->node[page_to_nid(page)]++;
2563 #ifdef CONFIG_HUGETLB_PAGE
2564 static void check_huge_range(struct vm_area_struct *vma,
2565 unsigned long start, unsigned long end,
2566 struct numa_maps *md)
2570 struct hstate *h = hstate_vma(vma);
2571 unsigned long sz = huge_page_size(h);
2573 for (addr = start; addr < end; addr += sz) {
2574 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2575 addr & huge_page_mask(h));
2585 page = pte_page(pte);
2589 gather_stats(page, md, pte_dirty(*ptep));
2593 static inline void check_huge_range(struct vm_area_struct *vma,
2594 unsigned long start, unsigned long end,
2595 struct numa_maps *md)
2601 * Display pages allocated per node and memory policy via /proc.
2603 int show_numa_map(struct seq_file *m, void *v)
2605 struct proc_maps_private *priv = m->private;
2606 struct vm_area_struct *vma = v;
2607 struct numa_maps *md;
2608 struct file *file = vma->vm_file;
2609 struct mm_struct *mm = vma->vm_mm;
2610 struct mempolicy *pol;
2617 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2621 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2622 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2625 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2628 seq_printf(m, " file=");
2629 seq_path(m, &file->f_path, "\n\t= ");
2630 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2631 seq_printf(m, " heap");
2632 } else if (vma->vm_start <= mm->start_stack &&
2633 vma->vm_end >= mm->start_stack) {
2634 seq_printf(m, " stack");
2637 if (is_vm_hugetlb_page(vma)) {
2638 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2639 seq_printf(m, " huge");
2641 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2642 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2649 seq_printf(m," anon=%lu",md->anon);
2652 seq_printf(m," dirty=%lu",md->dirty);
2654 if (md->pages != md->anon && md->pages != md->dirty)
2655 seq_printf(m, " mapped=%lu", md->pages);
2657 if (md->mapcount_max > 1)
2658 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2661 seq_printf(m," swapcache=%lu", md->swapcache);
2663 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2664 seq_printf(m," active=%lu", md->active);
2667 seq_printf(m," writeback=%lu", md->writeback);
2669 for_each_node_state(n, N_HIGH_MEMORY)
2671 seq_printf(m, " N%d=%lu", n, md->node[n]);
2676 if (m->count < m->size)
2677 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;