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, 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);
941 * Move pages between the two nodesets so as to preserve the physical
942 * layout as much as possible.
944 * Returns the number of page that could not be moved.
946 int do_migrate_pages(struct mm_struct *mm,
947 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
953 err = migrate_prep();
957 down_read(&mm->mmap_sem);
959 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
964 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
965 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
966 * bit in 'tmp', and return that <source, dest> pair for migration.
967 * The pair of nodemasks 'to' and 'from' define the map.
969 * If no pair of bits is found that way, fallback to picking some
970 * pair of 'source' and 'dest' bits that are not the same. If the
971 * 'source' and 'dest' bits are the same, this represents a node
972 * that will be migrating to itself, so no pages need move.
974 * If no bits are left in 'tmp', or if all remaining bits left
975 * in 'tmp' correspond to the same bit in 'to', return false
976 * (nothing left to migrate).
978 * This lets us pick a pair of nodes to migrate between, such that
979 * if possible the dest node is not already occupied by some other
980 * source node, minimizing the risk of overloading the memory on a
981 * node that would happen if we migrated incoming memory to a node
982 * before migrating outgoing memory source that same node.
984 * A single scan of tmp is sufficient. As we go, we remember the
985 * most recent <s, d> pair that moved (s != d). If we find a pair
986 * that not only moved, but what's better, moved to an empty slot
987 * (d is not set in tmp), then we break out then, with that pair.
988 * Otherwise when we finish scannng from_tmp, we at least have the
989 * most recent <s, d> pair that moved. If we get all the way through
990 * the scan of tmp without finding any node that moved, much less
991 * moved to an empty node, then there is nothing left worth migrating.
995 while (!nodes_empty(tmp)) {
1000 for_each_node_mask(s, tmp) {
1001 d = node_remap(s, *from_nodes, *to_nodes);
1005 source = s; /* Node moved. Memorize */
1008 /* dest not in remaining from nodes? */
1009 if (!node_isset(dest, tmp))
1015 node_clear(source, tmp);
1016 err = migrate_to_node(mm, source, dest, flags);
1023 up_read(&mm->mmap_sem);
1031 * Allocate a new page for page migration based on vma policy.
1032 * Start assuming that page is mapped by vma pointed to by @private.
1033 * Search forward from there, if not. N.B., this assumes that the
1034 * list of pages handed to migrate_pages()--which is how we get here--
1035 * is in virtual address order.
1037 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1039 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1040 unsigned long uninitialized_var(address);
1043 address = page_address_in_vma(page, vma);
1044 if (address != -EFAULT)
1050 * if !vma, alloc_page_vma() will use task or system default policy
1052 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1056 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1057 unsigned long flags)
1061 int do_migrate_pages(struct mm_struct *mm,
1062 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1067 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1073 static long do_mbind(unsigned long start, unsigned long len,
1074 unsigned short mode, unsigned short mode_flags,
1075 nodemask_t *nmask, unsigned long flags)
1077 struct vm_area_struct *vma;
1078 struct mm_struct *mm = current->mm;
1079 struct mempolicy *new;
1082 LIST_HEAD(pagelist);
1084 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1085 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1087 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1090 if (start & ~PAGE_MASK)
1093 if (mode == MPOL_DEFAULT)
1094 flags &= ~MPOL_MF_STRICT;
1096 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1104 new = mpol_new(mode, mode_flags, nmask);
1106 return PTR_ERR(new);
1109 * If we are using the default policy then operation
1110 * on discontinuous address spaces is okay after all
1113 flags |= MPOL_MF_DISCONTIG_OK;
1115 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1116 start, start + len, mode, mode_flags,
1117 nmask ? nodes_addr(*nmask)[0] : -1);
1119 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1121 err = migrate_prep();
1126 NODEMASK_SCRATCH(scratch);
1128 down_write(&mm->mmap_sem);
1130 err = mpol_set_nodemask(new, nmask, scratch);
1131 task_unlock(current);
1133 up_write(&mm->mmap_sem);
1136 NODEMASK_SCRATCH_FREE(scratch);
1141 vma = check_range(mm, start, end, nmask,
1142 flags | MPOL_MF_INVERT, &pagelist);
1148 err = mbind_range(mm, start, end, new);
1150 if (!list_empty(&pagelist))
1151 nr_failed = migrate_pages(&pagelist, new_vma_page,
1152 (unsigned long)vma, 0);
1154 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1157 putback_lru_pages(&pagelist);
1159 up_write(&mm->mmap_sem);
1166 * User space interface with variable sized bitmaps for nodelists.
1169 /* Copy a node mask from user space. */
1170 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1171 unsigned long maxnode)
1174 unsigned long nlongs;
1175 unsigned long endmask;
1178 nodes_clear(*nodes);
1179 if (maxnode == 0 || !nmask)
1181 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1184 nlongs = BITS_TO_LONGS(maxnode);
1185 if ((maxnode % BITS_PER_LONG) == 0)
1188 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1190 /* When the user specified more nodes than supported just check
1191 if the non supported part is all zero. */
1192 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1193 if (nlongs > PAGE_SIZE/sizeof(long))
1195 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1197 if (get_user(t, nmask + k))
1199 if (k == nlongs - 1) {
1205 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1209 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1211 nodes_addr(*nodes)[nlongs-1] &= endmask;
1215 /* Copy a kernel node mask to user space */
1216 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1219 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1220 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1222 if (copy > nbytes) {
1223 if (copy > PAGE_SIZE)
1225 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1229 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1232 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1233 unsigned long, mode, unsigned long __user *, nmask,
1234 unsigned long, maxnode, unsigned, flags)
1238 unsigned short mode_flags;
1240 mode_flags = mode & MPOL_MODE_FLAGS;
1241 mode &= ~MPOL_MODE_FLAGS;
1242 if (mode >= MPOL_MAX)
1244 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1245 (mode_flags & MPOL_F_RELATIVE_NODES))
1247 err = get_nodes(&nodes, nmask, maxnode);
1250 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1253 /* Set the process memory policy */
1254 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1255 unsigned long, maxnode)
1259 unsigned short flags;
1261 flags = mode & MPOL_MODE_FLAGS;
1262 mode &= ~MPOL_MODE_FLAGS;
1263 if ((unsigned int)mode >= MPOL_MAX)
1265 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1267 err = get_nodes(&nodes, nmask, maxnode);
1270 return do_set_mempolicy(mode, flags, &nodes);
1273 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1274 const unsigned long __user *, old_nodes,
1275 const unsigned long __user *, new_nodes)
1277 const struct cred *cred = current_cred(), *tcred;
1278 struct mm_struct *mm;
1279 struct task_struct *task;
1282 nodemask_t task_nodes;
1285 err = get_nodes(&old, old_nodes, maxnode);
1289 err = get_nodes(&new, new_nodes, maxnode);
1293 /* Find the mm_struct */
1294 read_lock(&tasklist_lock);
1295 task = pid ? find_task_by_vpid(pid) : current;
1297 read_unlock(&tasklist_lock);
1300 mm = get_task_mm(task);
1301 read_unlock(&tasklist_lock);
1307 * Check if this process has the right to modify the specified
1308 * process. The right exists if the process has administrative
1309 * capabilities, superuser privileges or the same
1310 * userid as the target process.
1313 tcred = __task_cred(task);
1314 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1315 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1316 !capable(CAP_SYS_NICE)) {
1323 task_nodes = cpuset_mems_allowed(task);
1324 /* Is the user allowed to access the target nodes? */
1325 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1330 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1335 err = security_task_movememory(task);
1339 err = do_migrate_pages(mm, &old, &new,
1340 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1347 /* Retrieve NUMA policy */
1348 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1349 unsigned long __user *, nmask, unsigned long, maxnode,
1350 unsigned long, addr, unsigned long, flags)
1353 int uninitialized_var(pval);
1356 if (nmask != NULL && maxnode < MAX_NUMNODES)
1359 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1364 if (policy && put_user(pval, policy))
1368 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1373 #ifdef CONFIG_COMPAT
1375 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1376 compat_ulong_t __user *nmask,
1377 compat_ulong_t maxnode,
1378 compat_ulong_t addr, compat_ulong_t flags)
1381 unsigned long __user *nm = NULL;
1382 unsigned long nr_bits, alloc_size;
1383 DECLARE_BITMAP(bm, MAX_NUMNODES);
1385 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1386 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1389 nm = compat_alloc_user_space(alloc_size);
1391 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1393 if (!err && nmask) {
1394 err = copy_from_user(bm, nm, alloc_size);
1395 /* ensure entire bitmap is zeroed */
1396 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1397 err |= compat_put_bitmap(nmask, bm, nr_bits);
1403 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1404 compat_ulong_t maxnode)
1407 unsigned long __user *nm = NULL;
1408 unsigned long nr_bits, alloc_size;
1409 DECLARE_BITMAP(bm, MAX_NUMNODES);
1411 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1412 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1415 err = compat_get_bitmap(bm, nmask, nr_bits);
1416 nm = compat_alloc_user_space(alloc_size);
1417 err |= copy_to_user(nm, bm, alloc_size);
1423 return sys_set_mempolicy(mode, nm, nr_bits+1);
1426 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1427 compat_ulong_t mode, compat_ulong_t __user *nmask,
1428 compat_ulong_t maxnode, compat_ulong_t flags)
1431 unsigned long __user *nm = NULL;
1432 unsigned long nr_bits, alloc_size;
1435 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1436 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1439 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1440 nm = compat_alloc_user_space(alloc_size);
1441 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1447 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1453 * get_vma_policy(@task, @vma, @addr)
1454 * @task - task for fallback if vma policy == default
1455 * @vma - virtual memory area whose policy is sought
1456 * @addr - address in @vma for shared policy lookup
1458 * Returns effective policy for a VMA at specified address.
1459 * Falls back to @task or system default policy, as necessary.
1460 * Current or other task's task mempolicy and non-shared vma policies
1461 * are protected by the task's mmap_sem, which must be held for read by
1463 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1464 * count--added by the get_policy() vm_op, as appropriate--to protect against
1465 * freeing by another task. It is the caller's responsibility to free the
1466 * extra reference for shared policies.
1468 static struct mempolicy *get_vma_policy(struct task_struct *task,
1469 struct vm_area_struct *vma, unsigned long addr)
1471 struct mempolicy *pol = task->mempolicy;
1474 if (vma->vm_ops && vma->vm_ops->get_policy) {
1475 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1479 } else if (vma->vm_policy)
1480 pol = vma->vm_policy;
1483 pol = &default_policy;
1488 * Return a nodemask representing a mempolicy for filtering nodes for
1491 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1493 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1494 if (unlikely(policy->mode == MPOL_BIND) &&
1495 gfp_zone(gfp) >= policy_zone &&
1496 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1497 return &policy->v.nodes;
1502 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1503 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1505 int nd = numa_node_id();
1507 switch (policy->mode) {
1508 case MPOL_PREFERRED:
1509 if (!(policy->flags & MPOL_F_LOCAL))
1510 nd = policy->v.preferred_node;
1514 * Normally, MPOL_BIND allocations are node-local within the
1515 * allowed nodemask. However, if __GFP_THISNODE is set and the
1516 * current node isn't part of the mask, we use the zonelist for
1517 * the first node in the mask instead.
1519 if (unlikely(gfp & __GFP_THISNODE) &&
1520 unlikely(!node_isset(nd, policy->v.nodes)))
1521 nd = first_node(policy->v.nodes);
1526 return node_zonelist(nd, gfp);
1529 /* Do dynamic interleaving for a process */
1530 static unsigned interleave_nodes(struct mempolicy *policy)
1533 struct task_struct *me = current;
1536 next = next_node(nid, policy->v.nodes);
1537 if (next >= MAX_NUMNODES)
1538 next = first_node(policy->v.nodes);
1539 if (next < MAX_NUMNODES)
1545 * Depending on the memory policy provide a node from which to allocate the
1547 * @policy must be protected by freeing by the caller. If @policy is
1548 * the current task's mempolicy, this protection is implicit, as only the
1549 * task can change it's policy. The system default policy requires no
1552 unsigned slab_node(struct mempolicy *policy)
1554 if (!policy || policy->flags & MPOL_F_LOCAL)
1555 return numa_node_id();
1557 switch (policy->mode) {
1558 case MPOL_PREFERRED:
1560 * handled MPOL_F_LOCAL above
1562 return policy->v.preferred_node;
1564 case MPOL_INTERLEAVE:
1565 return interleave_nodes(policy);
1569 * Follow bind policy behavior and start allocation at the
1572 struct zonelist *zonelist;
1574 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1575 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1576 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1587 /* Do static interleaving for a VMA with known offset. */
1588 static unsigned offset_il_node(struct mempolicy *pol,
1589 struct vm_area_struct *vma, unsigned long off)
1591 unsigned nnodes = nodes_weight(pol->v.nodes);
1597 return numa_node_id();
1598 target = (unsigned int)off % nnodes;
1601 nid = next_node(nid, pol->v.nodes);
1603 } while (c <= target);
1607 /* Determine a node number for interleave */
1608 static inline unsigned interleave_nid(struct mempolicy *pol,
1609 struct vm_area_struct *vma, unsigned long addr, int shift)
1615 * for small pages, there is no difference between
1616 * shift and PAGE_SHIFT, so the bit-shift is safe.
1617 * for huge pages, since vm_pgoff is in units of small
1618 * pages, we need to shift off the always 0 bits to get
1621 BUG_ON(shift < PAGE_SHIFT);
1622 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1623 off += (addr - vma->vm_start) >> shift;
1624 return offset_il_node(pol, vma, off);
1626 return interleave_nodes(pol);
1629 #ifdef CONFIG_HUGETLBFS
1631 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1632 * @vma = virtual memory area whose policy is sought
1633 * @addr = address in @vma for shared policy lookup and interleave policy
1634 * @gfp_flags = for requested zone
1635 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1636 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1638 * Returns a zonelist suitable for a huge page allocation and a pointer
1639 * to the struct mempolicy for conditional unref after allocation.
1640 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1641 * @nodemask for filtering the zonelist.
1643 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1644 gfp_t gfp_flags, struct mempolicy **mpol,
1645 nodemask_t **nodemask)
1647 struct zonelist *zl;
1649 *mpol = get_vma_policy(current, vma, addr);
1650 *nodemask = NULL; /* assume !MPOL_BIND */
1652 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1653 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1654 huge_page_shift(hstate_vma(vma))), gfp_flags);
1656 zl = policy_zonelist(gfp_flags, *mpol);
1657 if ((*mpol)->mode == MPOL_BIND)
1658 *nodemask = &(*mpol)->v.nodes;
1664 * init_nodemask_of_mempolicy
1666 * If the current task's mempolicy is "default" [NULL], return 'false'
1667 * to indicate default policy. Otherwise, extract the policy nodemask
1668 * for 'bind' or 'interleave' policy into the argument nodemask, or
1669 * initialize the argument nodemask to contain the single node for
1670 * 'preferred' or 'local' policy and return 'true' to indicate presence
1671 * of non-default mempolicy.
1673 * We don't bother with reference counting the mempolicy [mpol_get/put]
1674 * because the current task is examining it's own mempolicy and a task's
1675 * mempolicy is only ever changed by the task itself.
1677 * N.B., it is the caller's responsibility to free a returned nodemask.
1679 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1681 struct mempolicy *mempolicy;
1684 if (!(mask && current->mempolicy))
1687 mempolicy = current->mempolicy;
1688 switch (mempolicy->mode) {
1689 case MPOL_PREFERRED:
1690 if (mempolicy->flags & MPOL_F_LOCAL)
1691 nid = numa_node_id();
1693 nid = mempolicy->v.preferred_node;
1694 init_nodemask_of_node(mask, nid);
1699 case MPOL_INTERLEAVE:
1700 *mask = mempolicy->v.nodes;
1711 /* Allocate a page in interleaved policy.
1712 Own path because it needs to do special accounting. */
1713 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1716 struct zonelist *zl;
1719 zl = node_zonelist(nid, gfp);
1720 page = __alloc_pages(gfp, order, zl);
1721 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1722 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1727 * alloc_page_vma - Allocate a page for a VMA.
1730 * %GFP_USER user allocation.
1731 * %GFP_KERNEL kernel allocations,
1732 * %GFP_HIGHMEM highmem/user allocations,
1733 * %GFP_FS allocation should not call back into a file system.
1734 * %GFP_ATOMIC don't sleep.
1736 * @vma: Pointer to VMA or NULL if not available.
1737 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1739 * This function allocates a page from the kernel page pool and applies
1740 * a NUMA policy associated with the VMA or the current process.
1741 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1742 * mm_struct of the VMA to prevent it from going away. Should be used for
1743 * all allocations for pages that will be mapped into
1744 * user space. Returns NULL when no page can be allocated.
1746 * Should be called with the mm_sem of the vma hold.
1749 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1751 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1752 struct zonelist *zl;
1754 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1757 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1759 return alloc_page_interleave(gfp, 0, nid);
1761 zl = policy_zonelist(gfp, pol);
1762 if (unlikely(mpol_needs_cond_ref(pol))) {
1764 * slow path: ref counted shared policy
1766 struct page *page = __alloc_pages_nodemask(gfp, 0,
1767 zl, policy_nodemask(gfp, pol));
1772 * fast path: default or task policy
1774 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1778 * alloc_pages_current - Allocate pages.
1781 * %GFP_USER user allocation,
1782 * %GFP_KERNEL kernel allocation,
1783 * %GFP_HIGHMEM highmem allocation,
1784 * %GFP_FS don't call back into a file system.
1785 * %GFP_ATOMIC don't sleep.
1786 * @order: Power of two of allocation size in pages. 0 is a single page.
1788 * Allocate a page from the kernel page pool. When not in
1789 * interrupt context and apply the current process NUMA policy.
1790 * Returns NULL when no page can be allocated.
1792 * Don't call cpuset_update_task_memory_state() unless
1793 * 1) it's ok to take cpuset_sem (can WAIT), and
1794 * 2) allocating for current task (not interrupt).
1796 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1798 struct mempolicy *pol = current->mempolicy;
1800 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1801 pol = &default_policy;
1804 * No reference counting needed for current->mempolicy
1805 * nor system default_policy
1807 if (pol->mode == MPOL_INTERLEAVE)
1808 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1809 return __alloc_pages_nodemask(gfp, order,
1810 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1812 EXPORT_SYMBOL(alloc_pages_current);
1815 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1816 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1817 * with the mems_allowed returned by cpuset_mems_allowed(). This
1818 * keeps mempolicies cpuset relative after its cpuset moves. See
1819 * further kernel/cpuset.c update_nodemask().
1821 * current's mempolicy may be rebinded by the other task(the task that changes
1822 * cpuset's mems), so we needn't do rebind work for current task.
1825 /* Slow path of a mempolicy duplicate */
1826 struct mempolicy *__mpol_dup(struct mempolicy *old)
1828 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1831 return ERR_PTR(-ENOMEM);
1833 /* task's mempolicy is protected by alloc_lock */
1834 if (old == current->mempolicy) {
1837 task_unlock(current);
1842 if (current_cpuset_is_being_rebound()) {
1843 nodemask_t mems = cpuset_mems_allowed(current);
1844 if (new->flags & MPOL_F_REBINDING)
1845 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1847 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1850 atomic_set(&new->refcnt, 1);
1855 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1856 * eliminate the * MPOL_F_* flags that require conditional ref and
1857 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1858 * after return. Use the returned value.
1860 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1861 * policy lookup, even if the policy needs/has extra ref on lookup.
1862 * shmem_readahead needs this.
1864 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1865 struct mempolicy *frompol)
1867 if (!mpol_needs_cond_ref(frompol))
1871 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1872 __mpol_put(frompol);
1876 /* Slow path of a mempolicy comparison */
1877 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1881 if (a->mode != b->mode)
1883 if (a->flags != b->flags)
1885 if (mpol_store_user_nodemask(a))
1886 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1892 case MPOL_INTERLEAVE:
1893 return nodes_equal(a->v.nodes, b->v.nodes);
1894 case MPOL_PREFERRED:
1895 return a->v.preferred_node == b->v.preferred_node &&
1896 a->flags == b->flags;
1904 * Shared memory backing store policy support.
1906 * Remember policies even when nobody has shared memory mapped.
1907 * The policies are kept in Red-Black tree linked from the inode.
1908 * They are protected by the sp->lock spinlock, which should be held
1909 * for any accesses to the tree.
1912 /* lookup first element intersecting start-end */
1913 /* Caller holds sp->lock */
1914 static struct sp_node *
1915 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1917 struct rb_node *n = sp->root.rb_node;
1920 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1922 if (start >= p->end)
1924 else if (end <= p->start)
1932 struct sp_node *w = NULL;
1933 struct rb_node *prev = rb_prev(n);
1936 w = rb_entry(prev, struct sp_node, nd);
1937 if (w->end <= start)
1941 return rb_entry(n, struct sp_node, nd);
1944 /* Insert a new shared policy into the list. */
1945 /* Caller holds sp->lock */
1946 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1948 struct rb_node **p = &sp->root.rb_node;
1949 struct rb_node *parent = NULL;
1954 nd = rb_entry(parent, struct sp_node, nd);
1955 if (new->start < nd->start)
1957 else if (new->end > nd->end)
1958 p = &(*p)->rb_right;
1962 rb_link_node(&new->nd, parent, p);
1963 rb_insert_color(&new->nd, &sp->root);
1964 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1965 new->policy ? new->policy->mode : 0);
1968 /* Find shared policy intersecting idx */
1970 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1972 struct mempolicy *pol = NULL;
1975 if (!sp->root.rb_node)
1977 spin_lock(&sp->lock);
1978 sn = sp_lookup(sp, idx, idx+1);
1980 mpol_get(sn->policy);
1983 spin_unlock(&sp->lock);
1987 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1989 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1990 rb_erase(&n->nd, &sp->root);
1991 mpol_put(n->policy);
1992 kmem_cache_free(sn_cache, n);
1995 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1996 struct mempolicy *pol)
1998 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2005 pol->flags |= MPOL_F_SHARED; /* for unref */
2010 /* Replace a policy range. */
2011 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2012 unsigned long end, struct sp_node *new)
2014 struct sp_node *n, *new2 = NULL;
2017 spin_lock(&sp->lock);
2018 n = sp_lookup(sp, start, end);
2019 /* Take care of old policies in the same range. */
2020 while (n && n->start < end) {
2021 struct rb_node *next = rb_next(&n->nd);
2022 if (n->start >= start) {
2028 /* Old policy spanning whole new range. */
2031 spin_unlock(&sp->lock);
2032 new2 = sp_alloc(end, n->end, n->policy);
2038 sp_insert(sp, new2);
2046 n = rb_entry(next, struct sp_node, nd);
2050 spin_unlock(&sp->lock);
2052 mpol_put(new2->policy);
2053 kmem_cache_free(sn_cache, new2);
2059 * mpol_shared_policy_init - initialize shared policy for inode
2060 * @sp: pointer to inode shared policy
2061 * @mpol: struct mempolicy to install
2063 * Install non-NULL @mpol in inode's shared policy rb-tree.
2064 * On entry, the current task has a reference on a non-NULL @mpol.
2065 * This must be released on exit.
2066 * This is called at get_inode() calls and we can use GFP_KERNEL.
2068 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2072 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2073 spin_lock_init(&sp->lock);
2076 struct vm_area_struct pvma;
2077 struct mempolicy *new;
2078 NODEMASK_SCRATCH(scratch);
2082 /* contextualize the tmpfs mount point mempolicy */
2083 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2085 goto put_free; /* no valid nodemask intersection */
2088 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2089 task_unlock(current);
2090 mpol_put(mpol); /* drop our ref on sb mpol */
2094 /* Create pseudo-vma that contains just the policy */
2095 memset(&pvma, 0, sizeof(struct vm_area_struct));
2096 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2097 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2100 mpol_put(new); /* drop initial ref */
2101 NODEMASK_SCRATCH_FREE(scratch);
2105 int mpol_set_shared_policy(struct shared_policy *info,
2106 struct vm_area_struct *vma, struct mempolicy *npol)
2109 struct sp_node *new = NULL;
2110 unsigned long sz = vma_pages(vma);
2112 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2114 sz, npol ? npol->mode : -1,
2115 npol ? npol->flags : -1,
2116 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2119 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2123 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2125 kmem_cache_free(sn_cache, new);
2129 /* Free a backing policy store on inode delete. */
2130 void mpol_free_shared_policy(struct shared_policy *p)
2133 struct rb_node *next;
2135 if (!p->root.rb_node)
2137 spin_lock(&p->lock);
2138 next = rb_first(&p->root);
2140 n = rb_entry(next, struct sp_node, nd);
2141 next = rb_next(&n->nd);
2142 rb_erase(&n->nd, &p->root);
2143 mpol_put(n->policy);
2144 kmem_cache_free(sn_cache, n);
2146 spin_unlock(&p->lock);
2149 /* assumes fs == KERNEL_DS */
2150 void __init numa_policy_init(void)
2152 nodemask_t interleave_nodes;
2153 unsigned long largest = 0;
2154 int nid, prefer = 0;
2156 policy_cache = kmem_cache_create("numa_policy",
2157 sizeof(struct mempolicy),
2158 0, SLAB_PANIC, NULL);
2160 sn_cache = kmem_cache_create("shared_policy_node",
2161 sizeof(struct sp_node),
2162 0, SLAB_PANIC, NULL);
2165 * Set interleaving policy for system init. Interleaving is only
2166 * enabled across suitably sized nodes (default is >= 16MB), or
2167 * fall back to the largest node if they're all smaller.
2169 nodes_clear(interleave_nodes);
2170 for_each_node_state(nid, N_HIGH_MEMORY) {
2171 unsigned long total_pages = node_present_pages(nid);
2173 /* Preserve the largest node */
2174 if (largest < total_pages) {
2175 largest = total_pages;
2179 /* Interleave this node? */
2180 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2181 node_set(nid, interleave_nodes);
2184 /* All too small, use the largest */
2185 if (unlikely(nodes_empty(interleave_nodes)))
2186 node_set(prefer, interleave_nodes);
2188 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2189 printk("numa_policy_init: interleaving failed\n");
2192 /* Reset policy of current process to default */
2193 void numa_default_policy(void)
2195 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2199 * Parse and format mempolicy from/to strings
2203 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2204 * Used only for mpol_parse_str() and mpol_to_str()
2206 #define MPOL_LOCAL MPOL_MAX
2207 static const char * const policy_modes[] =
2209 [MPOL_DEFAULT] = "default",
2210 [MPOL_PREFERRED] = "prefer",
2211 [MPOL_BIND] = "bind",
2212 [MPOL_INTERLEAVE] = "interleave",
2213 [MPOL_LOCAL] = "local"
2219 * mpol_parse_str - parse string to mempolicy
2220 * @str: string containing mempolicy to parse
2221 * @mpol: pointer to struct mempolicy pointer, returned on success.
2222 * @no_context: flag whether to "contextualize" the mempolicy
2225 * <mode>[=<flags>][:<nodelist>]
2227 * if @no_context is true, save the input nodemask in w.user_nodemask in
2228 * the returned mempolicy. This will be used to "clone" the mempolicy in
2229 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2230 * mount option. Note that if 'static' or 'relative' mode flags were
2231 * specified, the input nodemask will already have been saved. Saving
2232 * it again is redundant, but safe.
2234 * On success, returns 0, else 1
2236 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2238 struct mempolicy *new = NULL;
2239 unsigned short mode;
2240 unsigned short uninitialized_var(mode_flags);
2242 char *nodelist = strchr(str, ':');
2243 char *flags = strchr(str, '=');
2247 /* NUL-terminate mode or flags string */
2249 if (nodelist_parse(nodelist, nodes))
2251 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2257 *flags++ = '\0'; /* terminate mode string */
2259 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2260 if (!strcmp(str, policy_modes[mode])) {
2264 if (mode > MPOL_LOCAL)
2268 case MPOL_PREFERRED:
2270 * Insist on a nodelist of one node only
2273 char *rest = nodelist;
2274 while (isdigit(*rest))
2280 case MPOL_INTERLEAVE:
2282 * Default to online nodes with memory if no nodelist
2285 nodes = node_states[N_HIGH_MEMORY];
2289 * Don't allow a nodelist; mpol_new() checks flags
2293 mode = MPOL_PREFERRED;
2297 * Insist on a empty nodelist
2304 * Insist on a nodelist
2313 * Currently, we only support two mutually exclusive
2316 if (!strcmp(flags, "static"))
2317 mode_flags |= MPOL_F_STATIC_NODES;
2318 else if (!strcmp(flags, "relative"))
2319 mode_flags |= MPOL_F_RELATIVE_NODES;
2324 new = mpol_new(mode, mode_flags, &nodes);
2329 /* save for contextualization */
2330 new->w.user_nodemask = nodes;
2333 NODEMASK_SCRATCH(scratch);
2336 ret = mpol_set_nodemask(new, &nodes, scratch);
2337 task_unlock(current);
2340 NODEMASK_SCRATCH_FREE(scratch);
2349 /* Restore string for error message */
2358 #endif /* CONFIG_TMPFS */
2361 * mpol_to_str - format a mempolicy structure for printing
2362 * @buffer: to contain formatted mempolicy string
2363 * @maxlen: length of @buffer
2364 * @pol: pointer to mempolicy to be formatted
2365 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2367 * Convert a mempolicy into a string.
2368 * Returns the number of characters in buffer (if positive)
2369 * or an error (negative)
2371 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2376 unsigned short mode;
2377 unsigned short flags = pol ? pol->flags : 0;
2380 * Sanity check: room for longest mode, flag and some nodes
2382 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2384 if (!pol || pol == &default_policy)
2385 mode = MPOL_DEFAULT;
2394 case MPOL_PREFERRED:
2396 if (flags & MPOL_F_LOCAL)
2397 mode = MPOL_LOCAL; /* pseudo-policy */
2399 node_set(pol->v.preferred_node, nodes);
2404 case MPOL_INTERLEAVE:
2406 nodes = pol->w.user_nodemask;
2408 nodes = pol->v.nodes;
2415 l = strlen(policy_modes[mode]);
2416 if (buffer + maxlen < p + l + 1)
2419 strcpy(p, policy_modes[mode]);
2422 if (flags & MPOL_MODE_FLAGS) {
2423 if (buffer + maxlen < p + 2)
2428 * Currently, the only defined flags are mutually exclusive
2430 if (flags & MPOL_F_STATIC_NODES)
2431 p += snprintf(p, buffer + maxlen - p, "static");
2432 else if (flags & MPOL_F_RELATIVE_NODES)
2433 p += snprintf(p, buffer + maxlen - p, "relative");
2436 if (!nodes_empty(nodes)) {
2437 if (buffer + maxlen < p + 2)
2440 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2446 unsigned long pages;
2448 unsigned long active;
2449 unsigned long writeback;
2450 unsigned long mapcount_max;
2451 unsigned long dirty;
2452 unsigned long swapcache;
2453 unsigned long node[MAX_NUMNODES];
2456 static void gather_stats(struct page *page, void *private, int pte_dirty)
2458 struct numa_maps *md = private;
2459 int count = page_mapcount(page);
2462 if (pte_dirty || PageDirty(page))
2465 if (PageSwapCache(page))
2468 if (PageActive(page) || PageUnevictable(page))
2471 if (PageWriteback(page))
2477 if (count > md->mapcount_max)
2478 md->mapcount_max = count;
2480 md->node[page_to_nid(page)]++;
2483 #ifdef CONFIG_HUGETLB_PAGE
2484 static void check_huge_range(struct vm_area_struct *vma,
2485 unsigned long start, unsigned long end,
2486 struct numa_maps *md)
2490 struct hstate *h = hstate_vma(vma);
2491 unsigned long sz = huge_page_size(h);
2493 for (addr = start; addr < end; addr += sz) {
2494 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2495 addr & huge_page_mask(h));
2505 page = pte_page(pte);
2509 gather_stats(page, md, pte_dirty(*ptep));
2513 static inline void check_huge_range(struct vm_area_struct *vma,
2514 unsigned long start, unsigned long end,
2515 struct numa_maps *md)
2521 * Display pages allocated per node and memory policy via /proc.
2523 int show_numa_map(struct seq_file *m, void *v)
2525 struct proc_maps_private *priv = m->private;
2526 struct vm_area_struct *vma = v;
2527 struct numa_maps *md;
2528 struct file *file = vma->vm_file;
2529 struct mm_struct *mm = vma->vm_mm;
2530 struct mempolicy *pol;
2537 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2541 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2542 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2545 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2548 seq_printf(m, " file=");
2549 seq_path(m, &file->f_path, "\n\t= ");
2550 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2551 seq_printf(m, " heap");
2552 } else if (vma->vm_start <= mm->start_stack &&
2553 vma->vm_end >= mm->start_stack) {
2554 seq_printf(m, " stack");
2557 if (is_vm_hugetlb_page(vma)) {
2558 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2559 seq_printf(m, " huge");
2561 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2562 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2569 seq_printf(m," anon=%lu",md->anon);
2572 seq_printf(m," dirty=%lu",md->dirty);
2574 if (md->pages != md->anon && md->pages != md->dirty)
2575 seq_printf(m, " mapped=%lu", md->pages);
2577 if (md->mapcount_max > 1)
2578 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2581 seq_printf(m," swapcache=%lu", md->swapcache);
2583 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2584 seq_printf(m," active=%lu", md->active);
2587 seq_printf(m," writeback=%lu", md->writeback);
2589 for_each_node_state(n, N_HIGH_MEMORY)
2591 seq_printf(m, " N%d=%lu", n, md->node[n]);
2596 if (m->count < m->size)
2597 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;