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,
927 struct vm_area_struct *vma;
930 node_set(source, nmask);
932 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
933 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
937 if (!list_empty(&pagelist)) {
938 err = migrate_pages(&pagelist, new_node_page, dest,
941 putback_lru_pages(&pagelist);
948 * Move pages between the two nodesets so as to preserve the physical
949 * layout as much as possible.
951 * Returns the number of page that could not be moved.
953 int do_migrate_pages(struct mm_struct *mm,
954 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
960 err = migrate_prep();
964 down_read(&mm->mmap_sem);
966 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
971 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
972 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
973 * bit in 'tmp', and return that <source, dest> pair for migration.
974 * The pair of nodemasks 'to' and 'from' define the map.
976 * If no pair of bits is found that way, fallback to picking some
977 * pair of 'source' and 'dest' bits that are not the same. If the
978 * 'source' and 'dest' bits are the same, this represents a node
979 * that will be migrating to itself, so no pages need move.
981 * If no bits are left in 'tmp', or if all remaining bits left
982 * in 'tmp' correspond to the same bit in 'to', return false
983 * (nothing left to migrate).
985 * This lets us pick a pair of nodes to migrate between, such that
986 * if possible the dest node is not already occupied by some other
987 * source node, minimizing the risk of overloading the memory on a
988 * node that would happen if we migrated incoming memory to a node
989 * before migrating outgoing memory source that same node.
991 * A single scan of tmp is sufficient. As we go, we remember the
992 * most recent <s, d> pair that moved (s != d). If we find a pair
993 * that not only moved, but what's better, moved to an empty slot
994 * (d is not set in tmp), then we break out then, with that pair.
995 * Otherwise when we finish scannng from_tmp, we at least have the
996 * most recent <s, d> pair that moved. If we get all the way through
997 * the scan of tmp without finding any node that moved, much less
998 * moved to an empty node, then there is nothing left worth migrating.
1002 while (!nodes_empty(tmp)) {
1007 for_each_node_mask(s, tmp) {
1008 d = node_remap(s, *from_nodes, *to_nodes);
1012 source = s; /* Node moved. Memorize */
1015 /* dest not in remaining from nodes? */
1016 if (!node_isset(dest, tmp))
1022 node_clear(source, tmp);
1023 err = migrate_to_node(mm, source, dest, flags);
1030 up_read(&mm->mmap_sem);
1038 * Allocate a new page for page migration based on vma policy.
1039 * Start assuming that page is mapped by vma pointed to by @private.
1040 * Search forward from there, if not. N.B., this assumes that the
1041 * list of pages handed to migrate_pages()--which is how we get here--
1042 * is in virtual address order.
1044 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1046 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1047 unsigned long uninitialized_var(address);
1050 address = page_address_in_vma(page, vma);
1051 if (address != -EFAULT)
1057 * if !vma, alloc_page_vma() will use task or system default policy
1059 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1063 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1064 unsigned long flags)
1068 int do_migrate_pages(struct mm_struct *mm,
1069 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1074 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1080 static long do_mbind(unsigned long start, unsigned long len,
1081 unsigned short mode, unsigned short mode_flags,
1082 nodemask_t *nmask, unsigned long flags)
1084 struct vm_area_struct *vma;
1085 struct mm_struct *mm = current->mm;
1086 struct mempolicy *new;
1089 LIST_HEAD(pagelist);
1091 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1092 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1094 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1097 if (start & ~PAGE_MASK)
1100 if (mode == MPOL_DEFAULT)
1101 flags &= ~MPOL_MF_STRICT;
1103 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1111 new = mpol_new(mode, mode_flags, nmask);
1113 return PTR_ERR(new);
1116 * If we are using the default policy then operation
1117 * on discontinuous address spaces is okay after all
1120 flags |= MPOL_MF_DISCONTIG_OK;
1122 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1123 start, start + len, mode, mode_flags,
1124 nmask ? nodes_addr(*nmask)[0] : -1);
1126 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1128 err = migrate_prep();
1133 NODEMASK_SCRATCH(scratch);
1135 down_write(&mm->mmap_sem);
1137 err = mpol_set_nodemask(new, nmask, scratch);
1138 task_unlock(current);
1140 up_write(&mm->mmap_sem);
1143 NODEMASK_SCRATCH_FREE(scratch);
1148 vma = check_range(mm, start, end, nmask,
1149 flags | MPOL_MF_INVERT, &pagelist);
1155 err = mbind_range(mm, start, end, new);
1157 if (!list_empty(&pagelist)) {
1158 nr_failed = migrate_pages(&pagelist, new_vma_page,
1162 putback_lru_pages(&pagelist);
1165 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1168 putback_lru_pages(&pagelist);
1170 up_write(&mm->mmap_sem);
1177 * User space interface with variable sized bitmaps for nodelists.
1180 /* Copy a node mask from user space. */
1181 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1182 unsigned long maxnode)
1185 unsigned long nlongs;
1186 unsigned long endmask;
1189 nodes_clear(*nodes);
1190 if (maxnode == 0 || !nmask)
1192 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1195 nlongs = BITS_TO_LONGS(maxnode);
1196 if ((maxnode % BITS_PER_LONG) == 0)
1199 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1201 /* When the user specified more nodes than supported just check
1202 if the non supported part is all zero. */
1203 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1204 if (nlongs > PAGE_SIZE/sizeof(long))
1206 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1208 if (get_user(t, nmask + k))
1210 if (k == nlongs - 1) {
1216 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1220 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1222 nodes_addr(*nodes)[nlongs-1] &= endmask;
1226 /* Copy a kernel node mask to user space */
1227 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1230 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1231 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1233 if (copy > nbytes) {
1234 if (copy > PAGE_SIZE)
1236 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1240 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1243 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1244 unsigned long, mode, unsigned long __user *, nmask,
1245 unsigned long, maxnode, unsigned, flags)
1249 unsigned short mode_flags;
1251 mode_flags = mode & MPOL_MODE_FLAGS;
1252 mode &= ~MPOL_MODE_FLAGS;
1253 if (mode >= MPOL_MAX)
1255 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1256 (mode_flags & MPOL_F_RELATIVE_NODES))
1258 err = get_nodes(&nodes, nmask, maxnode);
1261 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1264 /* Set the process memory policy */
1265 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1266 unsigned long, maxnode)
1270 unsigned short flags;
1272 flags = mode & MPOL_MODE_FLAGS;
1273 mode &= ~MPOL_MODE_FLAGS;
1274 if ((unsigned int)mode >= MPOL_MAX)
1276 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1278 err = get_nodes(&nodes, nmask, maxnode);
1281 return do_set_mempolicy(mode, flags, &nodes);
1284 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1285 const unsigned long __user *, old_nodes,
1286 const unsigned long __user *, new_nodes)
1288 const struct cred *cred = current_cred(), *tcred;
1289 struct mm_struct *mm = NULL;
1290 struct task_struct *task;
1291 nodemask_t task_nodes;
1295 NODEMASK_SCRATCH(scratch);
1300 old = &scratch->mask1;
1301 new = &scratch->mask2;
1303 err = get_nodes(old, old_nodes, maxnode);
1307 err = get_nodes(new, new_nodes, maxnode);
1311 /* Find the mm_struct */
1313 read_lock(&tasklist_lock);
1314 task = pid ? find_task_by_vpid(pid) : current;
1316 read_unlock(&tasklist_lock);
1321 mm = get_task_mm(task);
1322 read_unlock(&tasklist_lock);
1330 * Check if this process has the right to modify the specified
1331 * process. The right exists if the process has administrative
1332 * capabilities, superuser privileges or the same
1333 * userid as the target process.
1336 tcred = __task_cred(task);
1337 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1338 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1339 !capable(CAP_SYS_NICE)) {
1346 task_nodes = cpuset_mems_allowed(task);
1347 /* Is the user allowed to access the target nodes? */
1348 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1353 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1358 err = security_task_movememory(task);
1362 err = do_migrate_pages(mm, old, new,
1363 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1367 NODEMASK_SCRATCH_FREE(scratch);
1373 /* Retrieve NUMA policy */
1374 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1375 unsigned long __user *, nmask, unsigned long, maxnode,
1376 unsigned long, addr, unsigned long, flags)
1379 int uninitialized_var(pval);
1382 if (nmask != NULL && maxnode < MAX_NUMNODES)
1385 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1390 if (policy && put_user(pval, policy))
1394 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1399 #ifdef CONFIG_COMPAT
1401 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1402 compat_ulong_t __user *nmask,
1403 compat_ulong_t maxnode,
1404 compat_ulong_t addr, compat_ulong_t flags)
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 nm = compat_alloc_user_space(alloc_size);
1417 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1419 if (!err && nmask) {
1420 err = copy_from_user(bm, nm, alloc_size);
1421 /* ensure entire bitmap is zeroed */
1422 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1423 err |= compat_put_bitmap(nmask, bm, nr_bits);
1429 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1430 compat_ulong_t maxnode)
1433 unsigned long __user *nm = NULL;
1434 unsigned long nr_bits, alloc_size;
1435 DECLARE_BITMAP(bm, MAX_NUMNODES);
1437 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1438 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1441 err = compat_get_bitmap(bm, nmask, nr_bits);
1442 nm = compat_alloc_user_space(alloc_size);
1443 err |= copy_to_user(nm, bm, alloc_size);
1449 return sys_set_mempolicy(mode, nm, nr_bits+1);
1452 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1453 compat_ulong_t mode, compat_ulong_t __user *nmask,
1454 compat_ulong_t maxnode, compat_ulong_t flags)
1457 unsigned long __user *nm = NULL;
1458 unsigned long nr_bits, alloc_size;
1461 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1462 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1465 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1466 nm = compat_alloc_user_space(alloc_size);
1467 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1473 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1479 * get_vma_policy(@task, @vma, @addr)
1480 * @task - task for fallback if vma policy == default
1481 * @vma - virtual memory area whose policy is sought
1482 * @addr - address in @vma for shared policy lookup
1484 * Returns effective policy for a VMA at specified address.
1485 * Falls back to @task or system default policy, as necessary.
1486 * Current or other task's task mempolicy and non-shared vma policies
1487 * are protected by the task's mmap_sem, which must be held for read by
1489 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1490 * count--added by the get_policy() vm_op, as appropriate--to protect against
1491 * freeing by another task. It is the caller's responsibility to free the
1492 * extra reference for shared policies.
1494 static struct mempolicy *get_vma_policy(struct task_struct *task,
1495 struct vm_area_struct *vma, unsigned long addr)
1497 struct mempolicy *pol = task->mempolicy;
1500 if (vma->vm_ops && vma->vm_ops->get_policy) {
1501 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1505 } else if (vma->vm_policy)
1506 pol = vma->vm_policy;
1509 pol = &default_policy;
1514 * Return a nodemask representing a mempolicy for filtering nodes for
1517 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1519 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1520 if (unlikely(policy->mode == MPOL_BIND) &&
1521 gfp_zone(gfp) >= policy_zone &&
1522 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1523 return &policy->v.nodes;
1528 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1529 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1531 int nd = numa_node_id();
1533 switch (policy->mode) {
1534 case MPOL_PREFERRED:
1535 if (!(policy->flags & MPOL_F_LOCAL))
1536 nd = policy->v.preferred_node;
1540 * Normally, MPOL_BIND allocations are node-local within the
1541 * allowed nodemask. However, if __GFP_THISNODE is set and the
1542 * current node isn't part of the mask, we use the zonelist for
1543 * the first node in the mask instead.
1545 if (unlikely(gfp & __GFP_THISNODE) &&
1546 unlikely(!node_isset(nd, policy->v.nodes)))
1547 nd = first_node(policy->v.nodes);
1552 return node_zonelist(nd, gfp);
1555 /* Do dynamic interleaving for a process */
1556 static unsigned interleave_nodes(struct mempolicy *policy)
1559 struct task_struct *me = current;
1562 next = next_node(nid, policy->v.nodes);
1563 if (next >= MAX_NUMNODES)
1564 next = first_node(policy->v.nodes);
1565 if (next < MAX_NUMNODES)
1571 * Depending on the memory policy provide a node from which to allocate the
1573 * @policy must be protected by freeing by the caller. If @policy is
1574 * the current task's mempolicy, this protection is implicit, as only the
1575 * task can change it's policy. The system default policy requires no
1578 unsigned slab_node(struct mempolicy *policy)
1580 if (!policy || policy->flags & MPOL_F_LOCAL)
1581 return numa_node_id();
1583 switch (policy->mode) {
1584 case MPOL_PREFERRED:
1586 * handled MPOL_F_LOCAL above
1588 return policy->v.preferred_node;
1590 case MPOL_INTERLEAVE:
1591 return interleave_nodes(policy);
1595 * Follow bind policy behavior and start allocation at the
1598 struct zonelist *zonelist;
1600 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1601 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1602 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1605 return zone ? zone->node : numa_node_id();
1613 /* Do static interleaving for a VMA with known offset. */
1614 static unsigned offset_il_node(struct mempolicy *pol,
1615 struct vm_area_struct *vma, unsigned long off)
1617 unsigned nnodes = nodes_weight(pol->v.nodes);
1623 return numa_node_id();
1624 target = (unsigned int)off % nnodes;
1627 nid = next_node(nid, pol->v.nodes);
1629 } while (c <= target);
1633 /* Determine a node number for interleave */
1634 static inline unsigned interleave_nid(struct mempolicy *pol,
1635 struct vm_area_struct *vma, unsigned long addr, int shift)
1641 * for small pages, there is no difference between
1642 * shift and PAGE_SHIFT, so the bit-shift is safe.
1643 * for huge pages, since vm_pgoff is in units of small
1644 * pages, we need to shift off the always 0 bits to get
1647 BUG_ON(shift < PAGE_SHIFT);
1648 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1649 off += (addr - vma->vm_start) >> shift;
1650 return offset_il_node(pol, vma, off);
1652 return interleave_nodes(pol);
1655 #ifdef CONFIG_HUGETLBFS
1657 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1658 * @vma = virtual memory area whose policy is sought
1659 * @addr = address in @vma for shared policy lookup and interleave policy
1660 * @gfp_flags = for requested zone
1661 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1662 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1664 * Returns a zonelist suitable for a huge page allocation and a pointer
1665 * to the struct mempolicy for conditional unref after allocation.
1666 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1667 * @nodemask for filtering the zonelist.
1669 * Must be protected by get_mems_allowed()
1671 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1672 gfp_t gfp_flags, struct mempolicy **mpol,
1673 nodemask_t **nodemask)
1675 struct zonelist *zl;
1677 *mpol = get_vma_policy(current, vma, addr);
1678 *nodemask = NULL; /* assume !MPOL_BIND */
1680 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1681 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1682 huge_page_shift(hstate_vma(vma))), gfp_flags);
1684 zl = policy_zonelist(gfp_flags, *mpol);
1685 if ((*mpol)->mode == MPOL_BIND)
1686 *nodemask = &(*mpol)->v.nodes;
1692 * init_nodemask_of_mempolicy
1694 * If the current task's mempolicy is "default" [NULL], return 'false'
1695 * to indicate default policy. Otherwise, extract the policy nodemask
1696 * for 'bind' or 'interleave' policy into the argument nodemask, or
1697 * initialize the argument nodemask to contain the single node for
1698 * 'preferred' or 'local' policy and return 'true' to indicate presence
1699 * of non-default mempolicy.
1701 * We don't bother with reference counting the mempolicy [mpol_get/put]
1702 * because the current task is examining it's own mempolicy and a task's
1703 * mempolicy is only ever changed by the task itself.
1705 * N.B., it is the caller's responsibility to free a returned nodemask.
1707 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1709 struct mempolicy *mempolicy;
1712 if (!(mask && current->mempolicy))
1716 mempolicy = current->mempolicy;
1717 switch (mempolicy->mode) {
1718 case MPOL_PREFERRED:
1719 if (mempolicy->flags & MPOL_F_LOCAL)
1720 nid = numa_node_id();
1722 nid = mempolicy->v.preferred_node;
1723 init_nodemask_of_node(mask, nid);
1728 case MPOL_INTERLEAVE:
1729 *mask = mempolicy->v.nodes;
1735 task_unlock(current);
1742 * mempolicy_nodemask_intersects
1744 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1745 * policy. Otherwise, check for intersection between mask and the policy
1746 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1747 * policy, always return true since it may allocate elsewhere on fallback.
1749 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1751 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1752 const nodemask_t *mask)
1754 struct mempolicy *mempolicy;
1760 mempolicy = tsk->mempolicy;
1764 switch (mempolicy->mode) {
1765 case MPOL_PREFERRED:
1767 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1768 * allocate from, they may fallback to other nodes when oom.
1769 * Thus, it's possible for tsk to have allocated memory from
1774 case MPOL_INTERLEAVE:
1775 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1785 /* Allocate a page in interleaved policy.
1786 Own path because it needs to do special accounting. */
1787 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1790 struct zonelist *zl;
1793 zl = node_zonelist(nid, gfp);
1794 page = __alloc_pages(gfp, order, zl);
1795 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1796 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1801 * alloc_page_vma - Allocate a page for a VMA.
1804 * %GFP_USER user allocation.
1805 * %GFP_KERNEL kernel allocations,
1806 * %GFP_HIGHMEM highmem/user allocations,
1807 * %GFP_FS allocation should not call back into a file system.
1808 * %GFP_ATOMIC don't sleep.
1810 * @vma: Pointer to VMA or NULL if not available.
1811 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1813 * This function allocates a page from the kernel page pool and applies
1814 * a NUMA policy associated with the VMA or the current process.
1815 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1816 * mm_struct of the VMA to prevent it from going away. Should be used for
1817 * all allocations for pages that will be mapped into
1818 * user space. Returns NULL when no page can be allocated.
1820 * Should be called with the mm_sem of the vma hold.
1823 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1825 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1826 struct zonelist *zl;
1830 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1833 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1835 page = alloc_page_interleave(gfp, 0, nid);
1839 zl = policy_zonelist(gfp, pol);
1840 if (unlikely(mpol_needs_cond_ref(pol))) {
1842 * slow path: ref counted shared policy
1844 struct page *page = __alloc_pages_nodemask(gfp, 0,
1845 zl, policy_nodemask(gfp, pol));
1851 * fast path: default or task policy
1853 page = __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1859 * alloc_pages_current - Allocate pages.
1862 * %GFP_USER user allocation,
1863 * %GFP_KERNEL kernel allocation,
1864 * %GFP_HIGHMEM highmem allocation,
1865 * %GFP_FS don't call back into a file system.
1866 * %GFP_ATOMIC don't sleep.
1867 * @order: Power of two of allocation size in pages. 0 is a single page.
1869 * Allocate a page from the kernel page pool. When not in
1870 * interrupt context and apply the current process NUMA policy.
1871 * Returns NULL when no page can be allocated.
1873 * Don't call cpuset_update_task_memory_state() unless
1874 * 1) it's ok to take cpuset_sem (can WAIT), and
1875 * 2) allocating for current task (not interrupt).
1877 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1879 struct mempolicy *pol = current->mempolicy;
1882 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1883 pol = &default_policy;
1887 * No reference counting needed for current->mempolicy
1888 * nor system default_policy
1890 if (pol->mode == MPOL_INTERLEAVE)
1891 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1893 page = __alloc_pages_nodemask(gfp, order,
1894 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1898 EXPORT_SYMBOL(alloc_pages_current);
1901 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1902 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1903 * with the mems_allowed returned by cpuset_mems_allowed(). This
1904 * keeps mempolicies cpuset relative after its cpuset moves. See
1905 * further kernel/cpuset.c update_nodemask().
1907 * current's mempolicy may be rebinded by the other task(the task that changes
1908 * cpuset's mems), so we needn't do rebind work for current task.
1911 /* Slow path of a mempolicy duplicate */
1912 struct mempolicy *__mpol_dup(struct mempolicy *old)
1914 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1917 return ERR_PTR(-ENOMEM);
1919 /* task's mempolicy is protected by alloc_lock */
1920 if (old == current->mempolicy) {
1923 task_unlock(current);
1928 if (current_cpuset_is_being_rebound()) {
1929 nodemask_t mems = cpuset_mems_allowed(current);
1930 if (new->flags & MPOL_F_REBINDING)
1931 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1933 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1936 atomic_set(&new->refcnt, 1);
1941 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1942 * eliminate the * MPOL_F_* flags that require conditional ref and
1943 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1944 * after return. Use the returned value.
1946 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1947 * policy lookup, even if the policy needs/has extra ref on lookup.
1948 * shmem_readahead needs this.
1950 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1951 struct mempolicy *frompol)
1953 if (!mpol_needs_cond_ref(frompol))
1957 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1958 __mpol_put(frompol);
1962 /* Slow path of a mempolicy comparison */
1963 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1967 if (a->mode != b->mode)
1969 if (a->flags != b->flags)
1971 if (mpol_store_user_nodemask(a))
1972 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1978 case MPOL_INTERLEAVE:
1979 return nodes_equal(a->v.nodes, b->v.nodes);
1980 case MPOL_PREFERRED:
1981 return a->v.preferred_node == b->v.preferred_node &&
1982 a->flags == b->flags;
1990 * Shared memory backing store policy support.
1992 * Remember policies even when nobody has shared memory mapped.
1993 * The policies are kept in Red-Black tree linked from the inode.
1994 * They are protected by the sp->lock spinlock, which should be held
1995 * for any accesses to the tree.
1998 /* lookup first element intersecting start-end */
1999 /* Caller holds sp->lock */
2000 static struct sp_node *
2001 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2003 struct rb_node *n = sp->root.rb_node;
2006 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2008 if (start >= p->end)
2010 else if (end <= p->start)
2018 struct sp_node *w = NULL;
2019 struct rb_node *prev = rb_prev(n);
2022 w = rb_entry(prev, struct sp_node, nd);
2023 if (w->end <= start)
2027 return rb_entry(n, struct sp_node, nd);
2030 /* Insert a new shared policy into the list. */
2031 /* Caller holds sp->lock */
2032 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2034 struct rb_node **p = &sp->root.rb_node;
2035 struct rb_node *parent = NULL;
2040 nd = rb_entry(parent, struct sp_node, nd);
2041 if (new->start < nd->start)
2043 else if (new->end > nd->end)
2044 p = &(*p)->rb_right;
2048 rb_link_node(&new->nd, parent, p);
2049 rb_insert_color(&new->nd, &sp->root);
2050 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2051 new->policy ? new->policy->mode : 0);
2054 /* Find shared policy intersecting idx */
2056 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2058 struct mempolicy *pol = NULL;
2061 if (!sp->root.rb_node)
2063 spin_lock(&sp->lock);
2064 sn = sp_lookup(sp, idx, idx+1);
2066 mpol_get(sn->policy);
2069 spin_unlock(&sp->lock);
2073 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2075 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2076 rb_erase(&n->nd, &sp->root);
2077 mpol_put(n->policy);
2078 kmem_cache_free(sn_cache, n);
2081 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2082 struct mempolicy *pol)
2084 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2091 pol->flags |= MPOL_F_SHARED; /* for unref */
2096 /* Replace a policy range. */
2097 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2098 unsigned long end, struct sp_node *new)
2100 struct sp_node *n, *new2 = NULL;
2103 spin_lock(&sp->lock);
2104 n = sp_lookup(sp, start, end);
2105 /* Take care of old policies in the same range. */
2106 while (n && n->start < end) {
2107 struct rb_node *next = rb_next(&n->nd);
2108 if (n->start >= start) {
2114 /* Old policy spanning whole new range. */
2117 spin_unlock(&sp->lock);
2118 new2 = sp_alloc(end, n->end, n->policy);
2124 sp_insert(sp, new2);
2132 n = rb_entry(next, struct sp_node, nd);
2136 spin_unlock(&sp->lock);
2138 mpol_put(new2->policy);
2139 kmem_cache_free(sn_cache, new2);
2145 * mpol_shared_policy_init - initialize shared policy for inode
2146 * @sp: pointer to inode shared policy
2147 * @mpol: struct mempolicy to install
2149 * Install non-NULL @mpol in inode's shared policy rb-tree.
2150 * On entry, the current task has a reference on a non-NULL @mpol.
2151 * This must be released on exit.
2152 * This is called at get_inode() calls and we can use GFP_KERNEL.
2154 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2158 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2159 spin_lock_init(&sp->lock);
2162 struct vm_area_struct pvma;
2163 struct mempolicy *new;
2164 NODEMASK_SCRATCH(scratch);
2168 /* contextualize the tmpfs mount point mempolicy */
2169 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2171 goto free_scratch; /* no valid nodemask intersection */
2174 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2175 task_unlock(current);
2179 /* Create pseudo-vma that contains just the policy */
2180 memset(&pvma, 0, sizeof(struct vm_area_struct));
2181 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2182 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2185 mpol_put(new); /* drop initial ref */
2187 NODEMASK_SCRATCH_FREE(scratch);
2189 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2193 int mpol_set_shared_policy(struct shared_policy *info,
2194 struct vm_area_struct *vma, struct mempolicy *npol)
2197 struct sp_node *new = NULL;
2198 unsigned long sz = vma_pages(vma);
2200 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2202 sz, npol ? npol->mode : -1,
2203 npol ? npol->flags : -1,
2204 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2207 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2211 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2213 kmem_cache_free(sn_cache, new);
2217 /* Free a backing policy store on inode delete. */
2218 void mpol_free_shared_policy(struct shared_policy *p)
2221 struct rb_node *next;
2223 if (!p->root.rb_node)
2225 spin_lock(&p->lock);
2226 next = rb_first(&p->root);
2228 n = rb_entry(next, struct sp_node, nd);
2229 next = rb_next(&n->nd);
2230 rb_erase(&n->nd, &p->root);
2231 mpol_put(n->policy);
2232 kmem_cache_free(sn_cache, n);
2234 spin_unlock(&p->lock);
2237 /* assumes fs == KERNEL_DS */
2238 void __init numa_policy_init(void)
2240 nodemask_t interleave_nodes;
2241 unsigned long largest = 0;
2242 int nid, prefer = 0;
2244 policy_cache = kmem_cache_create("numa_policy",
2245 sizeof(struct mempolicy),
2246 0, SLAB_PANIC, NULL);
2248 sn_cache = kmem_cache_create("shared_policy_node",
2249 sizeof(struct sp_node),
2250 0, SLAB_PANIC, NULL);
2253 * Set interleaving policy for system init. Interleaving is only
2254 * enabled across suitably sized nodes (default is >= 16MB), or
2255 * fall back to the largest node if they're all smaller.
2257 nodes_clear(interleave_nodes);
2258 for_each_node_state(nid, N_HIGH_MEMORY) {
2259 unsigned long total_pages = node_present_pages(nid);
2261 /* Preserve the largest node */
2262 if (largest < total_pages) {
2263 largest = total_pages;
2267 /* Interleave this node? */
2268 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2269 node_set(nid, interleave_nodes);
2272 /* All too small, use the largest */
2273 if (unlikely(nodes_empty(interleave_nodes)))
2274 node_set(prefer, interleave_nodes);
2276 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2277 printk("numa_policy_init: interleaving failed\n");
2280 /* Reset policy of current process to default */
2281 void numa_default_policy(void)
2283 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2287 * Parse and format mempolicy from/to strings
2291 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2292 * Used only for mpol_parse_str() and mpol_to_str()
2294 #define MPOL_LOCAL MPOL_MAX
2295 static const char * const policy_modes[] =
2297 [MPOL_DEFAULT] = "default",
2298 [MPOL_PREFERRED] = "prefer",
2299 [MPOL_BIND] = "bind",
2300 [MPOL_INTERLEAVE] = "interleave",
2301 [MPOL_LOCAL] = "local"
2307 * mpol_parse_str - parse string to mempolicy
2308 * @str: string containing mempolicy to parse
2309 * @mpol: pointer to struct mempolicy pointer, returned on success.
2310 * @no_context: flag whether to "contextualize" the mempolicy
2313 * <mode>[=<flags>][:<nodelist>]
2315 * if @no_context is true, save the input nodemask in w.user_nodemask in
2316 * the returned mempolicy. This will be used to "clone" the mempolicy in
2317 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2318 * mount option. Note that if 'static' or 'relative' mode flags were
2319 * specified, the input nodemask will already have been saved. Saving
2320 * it again is redundant, but safe.
2322 * On success, returns 0, else 1
2324 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2326 struct mempolicy *new = NULL;
2327 unsigned short mode;
2328 unsigned short uninitialized_var(mode_flags);
2330 char *nodelist = strchr(str, ':');
2331 char *flags = strchr(str, '=');
2335 /* NUL-terminate mode or flags string */
2337 if (nodelist_parse(nodelist, nodes))
2339 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2345 *flags++ = '\0'; /* terminate mode string */
2347 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2348 if (!strcmp(str, policy_modes[mode])) {
2352 if (mode > MPOL_LOCAL)
2356 case MPOL_PREFERRED:
2358 * Insist on a nodelist of one node only
2361 char *rest = nodelist;
2362 while (isdigit(*rest))
2368 case MPOL_INTERLEAVE:
2370 * Default to online nodes with memory if no nodelist
2373 nodes = node_states[N_HIGH_MEMORY];
2377 * Don't allow a nodelist; mpol_new() checks flags
2381 mode = MPOL_PREFERRED;
2385 * Insist on a empty nodelist
2392 * Insist on a nodelist
2401 * Currently, we only support two mutually exclusive
2404 if (!strcmp(flags, "static"))
2405 mode_flags |= MPOL_F_STATIC_NODES;
2406 else if (!strcmp(flags, "relative"))
2407 mode_flags |= MPOL_F_RELATIVE_NODES;
2412 new = mpol_new(mode, mode_flags, &nodes);
2417 /* save for contextualization */
2418 new->w.user_nodemask = nodes;
2421 NODEMASK_SCRATCH(scratch);
2424 ret = mpol_set_nodemask(new, &nodes, scratch);
2425 task_unlock(current);
2428 NODEMASK_SCRATCH_FREE(scratch);
2437 /* Restore string for error message */
2446 #endif /* CONFIG_TMPFS */
2449 * mpol_to_str - format a mempolicy structure for printing
2450 * @buffer: to contain formatted mempolicy string
2451 * @maxlen: length of @buffer
2452 * @pol: pointer to mempolicy to be formatted
2453 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2455 * Convert a mempolicy into a string.
2456 * Returns the number of characters in buffer (if positive)
2457 * or an error (negative)
2459 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2464 unsigned short mode;
2465 unsigned short flags = pol ? pol->flags : 0;
2468 * Sanity check: room for longest mode, flag and some nodes
2470 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2472 if (!pol || pol == &default_policy)
2473 mode = MPOL_DEFAULT;
2482 case MPOL_PREFERRED:
2484 if (flags & MPOL_F_LOCAL)
2485 mode = MPOL_LOCAL; /* pseudo-policy */
2487 node_set(pol->v.preferred_node, nodes);
2492 case MPOL_INTERLEAVE:
2494 nodes = pol->w.user_nodemask;
2496 nodes = pol->v.nodes;
2503 l = strlen(policy_modes[mode]);
2504 if (buffer + maxlen < p + l + 1)
2507 strcpy(p, policy_modes[mode]);
2510 if (flags & MPOL_MODE_FLAGS) {
2511 if (buffer + maxlen < p + 2)
2516 * Currently, the only defined flags are mutually exclusive
2518 if (flags & MPOL_F_STATIC_NODES)
2519 p += snprintf(p, buffer + maxlen - p, "static");
2520 else if (flags & MPOL_F_RELATIVE_NODES)
2521 p += snprintf(p, buffer + maxlen - p, "relative");
2524 if (!nodes_empty(nodes)) {
2525 if (buffer + maxlen < p + 2)
2528 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2534 unsigned long pages;
2536 unsigned long active;
2537 unsigned long writeback;
2538 unsigned long mapcount_max;
2539 unsigned long dirty;
2540 unsigned long swapcache;
2541 unsigned long node[MAX_NUMNODES];
2544 static void gather_stats(struct page *page, void *private, int pte_dirty)
2546 struct numa_maps *md = private;
2547 int count = page_mapcount(page);
2550 if (pte_dirty || PageDirty(page))
2553 if (PageSwapCache(page))
2556 if (PageActive(page) || PageUnevictable(page))
2559 if (PageWriteback(page))
2565 if (count > md->mapcount_max)
2566 md->mapcount_max = count;
2568 md->node[page_to_nid(page)]++;
2571 #ifdef CONFIG_HUGETLB_PAGE
2572 static void check_huge_range(struct vm_area_struct *vma,
2573 unsigned long start, unsigned long end,
2574 struct numa_maps *md)
2578 struct hstate *h = hstate_vma(vma);
2579 unsigned long sz = huge_page_size(h);
2581 for (addr = start; addr < end; addr += sz) {
2582 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2583 addr & huge_page_mask(h));
2593 page = pte_page(pte);
2597 gather_stats(page, md, pte_dirty(*ptep));
2601 static inline void check_huge_range(struct vm_area_struct *vma,
2602 unsigned long start, unsigned long end,
2603 struct numa_maps *md)
2609 * Display pages allocated per node and memory policy via /proc.
2611 int show_numa_map(struct seq_file *m, void *v)
2613 struct proc_maps_private *priv = m->private;
2614 struct vm_area_struct *vma = v;
2615 struct numa_maps *md;
2616 struct file *file = vma->vm_file;
2617 struct mm_struct *mm = vma->vm_mm;
2618 struct mempolicy *pol;
2625 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2629 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2630 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2633 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2636 seq_printf(m, " file=");
2637 seq_path(m, &file->f_path, "\n\t= ");
2638 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2639 seq_printf(m, " heap");
2640 } else if (vma->vm_start <= mm->start_stack &&
2641 vma->vm_end >= mm->start_stack) {
2642 seq_printf(m, " stack");
2645 if (is_vm_hugetlb_page(vma)) {
2646 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2647 seq_printf(m, " huge");
2649 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2650 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2657 seq_printf(m," anon=%lu",md->anon);
2660 seq_printf(m," dirty=%lu",md->dirty);
2662 if (md->pages != md->anon && md->pages != md->dirty)
2663 seq_printf(m, " mapped=%lu", md->pages);
2665 if (md->mapcount_max > 1)
2666 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2669 seq_printf(m," swapcache=%lu", md->swapcache);
2671 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2672 seq_printf(m," active=%lu", md->active);
2675 seq_printf(m," writeback=%lu", md->writeback);
2677 for_each_node_state(n, N_HIGH_MEMORY)
2679 seq_printf(m, " N%d=%lu", n, md->node[n]);
2684 if (m->count < m->size)
2685 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;