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/export.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>
96 #include <linux/random.h>
101 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
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 static 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 migrate_page_add(struct page *page, struct list_head *pagelist,
461 unsigned long flags);
463 /* Scan through pages checking if pages follow certain conditions. */
464 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
465 unsigned long addr, unsigned long end,
466 const nodemask_t *nodes, unsigned long flags,
473 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
478 if (!pte_present(*pte))
480 page = vm_normal_page(vma, addr, *pte);
484 * vm_normal_page() filters out zero pages, but there might
485 * still be PageReserved pages to skip, perhaps in a VDSO.
486 * And we cannot move PageKsm pages sensibly or safely yet.
488 if (PageReserved(page) || PageKsm(page))
490 nid = page_to_nid(page);
491 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
495 migrate_page_add(page, private, flags);
498 } while (pte++, addr += PAGE_SIZE, addr != end);
499 pte_unmap_unlock(orig_pte, ptl);
503 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
504 unsigned long addr, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags,
511 pmd = pmd_offset(pud, addr);
513 next = pmd_addr_end(addr, end);
514 split_huge_page_pmd(vma->vm_mm, pmd);
515 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
517 if (check_pte_range(vma, pmd, addr, next, nodes,
520 } while (pmd++, addr = next, addr != end);
524 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
525 unsigned long addr, unsigned long end,
526 const nodemask_t *nodes, unsigned long flags,
532 pud = pud_offset(pgd, addr);
534 next = pud_addr_end(addr, end);
535 if (pud_none_or_clear_bad(pud))
537 if (check_pmd_range(vma, pud, addr, next, nodes,
540 } while (pud++, addr = next, addr != end);
544 static inline int check_pgd_range(struct vm_area_struct *vma,
545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
552 pgd = pgd_offset(vma->vm_mm, addr);
554 next = pgd_addr_end(addr, end);
555 if (pgd_none_or_clear_bad(pgd))
557 if (check_pud_range(vma, pgd, addr, next, nodes,
560 } while (pgd++, addr = next, addr != end);
565 * Check if all pages in a range are on a set of nodes.
566 * If pagelist != NULL then isolate pages from the LRU and
567 * put them on the pagelist.
569 static struct vm_area_struct *
570 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
571 const nodemask_t *nodes, unsigned long flags, void *private)
574 struct vm_area_struct *first, *vma, *prev;
577 first = find_vma(mm, start);
579 return ERR_PTR(-EFAULT);
581 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
582 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
583 if (!vma->vm_next && vma->vm_end < end)
584 return ERR_PTR(-EFAULT);
585 if (prev && prev->vm_end < vma->vm_start)
586 return ERR_PTR(-EFAULT);
588 if (!is_vm_hugetlb_page(vma) &&
589 ((flags & MPOL_MF_STRICT) ||
590 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
591 vma_migratable(vma)))) {
592 unsigned long endvma = vma->vm_end;
596 if (vma->vm_start > start)
597 start = vma->vm_start;
598 err = check_pgd_range(vma, start, endvma, nodes,
601 first = ERR_PTR(err);
610 /* Apply policy to a single VMA */
611 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
614 struct mempolicy *old = vma->vm_policy;
616 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
617 vma->vm_start, vma->vm_end, vma->vm_pgoff,
618 vma->vm_ops, vma->vm_file,
619 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
621 if (vma->vm_ops && vma->vm_ops->set_policy)
622 err = vma->vm_ops->set_policy(vma, new);
625 vma->vm_policy = new;
631 /* Step 2: apply policy to a range and do splits. */
632 static int mbind_range(struct mm_struct *mm, unsigned long start,
633 unsigned long end, struct mempolicy *new_pol)
635 struct vm_area_struct *next;
636 struct vm_area_struct *prev;
637 struct vm_area_struct *vma;
640 unsigned long vmstart;
643 vma = find_vma(mm, start);
644 if (!vma || vma->vm_start > start)
648 if (start > vma->vm_start)
651 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
653 vmstart = max(start, vma->vm_start);
654 vmend = min(end, vma->vm_end);
656 if (mpol_equal(vma_policy(vma), new_pol))
659 pgoff = vma->vm_pgoff +
660 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
661 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
662 vma->anon_vma, vma->vm_file, pgoff,
669 if (vma->vm_start != vmstart) {
670 err = split_vma(vma->vm_mm, vma, vmstart, 1);
674 if (vma->vm_end != vmend) {
675 err = split_vma(vma->vm_mm, vma, vmend, 0);
679 err = policy_vma(vma, new_pol);
689 * Update task->flags PF_MEMPOLICY bit: set iff non-default
690 * mempolicy. Allows more rapid checking of this (combined perhaps
691 * with other PF_* flag bits) on memory allocation hot code paths.
693 * If called from outside this file, the task 'p' should -only- be
694 * a newly forked child not yet visible on the task list, because
695 * manipulating the task flags of a visible task is not safe.
697 * The above limitation is why this routine has the funny name
698 * mpol_fix_fork_child_flag().
700 * It is also safe to call this with a task pointer of current,
701 * which the static wrapper mpol_set_task_struct_flag() does,
702 * for use within this file.
705 void mpol_fix_fork_child_flag(struct task_struct *p)
708 p->flags |= PF_MEMPOLICY;
710 p->flags &= ~PF_MEMPOLICY;
713 static void mpol_set_task_struct_flag(void)
715 mpol_fix_fork_child_flag(current);
718 /* Set the process memory policy */
719 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
722 struct mempolicy *new, *old;
723 struct mm_struct *mm = current->mm;
724 NODEMASK_SCRATCH(scratch);
730 new = mpol_new(mode, flags, nodes);
736 * prevent changing our mempolicy while show_numa_maps()
738 * Note: do_set_mempolicy() can be called at init time
742 down_write(&mm->mmap_sem);
744 ret = mpol_set_nodemask(new, nodes, scratch);
746 task_unlock(current);
748 up_write(&mm->mmap_sem);
752 old = current->mempolicy;
753 current->mempolicy = new;
754 mpol_set_task_struct_flag();
755 if (new && new->mode == MPOL_INTERLEAVE &&
756 nodes_weight(new->v.nodes))
757 current->il_next = first_node(new->v.nodes);
758 task_unlock(current);
760 up_write(&mm->mmap_sem);
765 NODEMASK_SCRATCH_FREE(scratch);
770 * Return nodemask for policy for get_mempolicy() query
772 * Called with task's alloc_lock held
774 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
777 if (p == &default_policy)
783 case MPOL_INTERLEAVE:
787 if (!(p->flags & MPOL_F_LOCAL))
788 node_set(p->v.preferred_node, *nodes);
789 /* else return empty node mask for local allocation */
796 static int lookup_node(struct mm_struct *mm, unsigned long addr)
801 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
803 err = page_to_nid(p);
809 /* Retrieve NUMA policy */
810 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
811 unsigned long addr, unsigned long flags)
814 struct mm_struct *mm = current->mm;
815 struct vm_area_struct *vma = NULL;
816 struct mempolicy *pol = current->mempolicy;
819 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
822 if (flags & MPOL_F_MEMS_ALLOWED) {
823 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
825 *policy = 0; /* just so it's initialized */
827 *nmask = cpuset_current_mems_allowed;
828 task_unlock(current);
832 if (flags & MPOL_F_ADDR) {
834 * Do NOT fall back to task policy if the
835 * vma/shared policy at addr is NULL. We
836 * want to return MPOL_DEFAULT in this case.
838 down_read(&mm->mmap_sem);
839 vma = find_vma_intersection(mm, addr, addr+1);
841 up_read(&mm->mmap_sem);
844 if (vma->vm_ops && vma->vm_ops->get_policy)
845 pol = vma->vm_ops->get_policy(vma, addr);
847 pol = vma->vm_policy;
852 pol = &default_policy; /* indicates default behavior */
854 if (flags & MPOL_F_NODE) {
855 if (flags & MPOL_F_ADDR) {
856 err = lookup_node(mm, addr);
860 } else if (pol == current->mempolicy &&
861 pol->mode == MPOL_INTERLEAVE) {
862 *policy = current->il_next;
868 *policy = pol == &default_policy ? MPOL_DEFAULT :
871 * Internal mempolicy flags must be masked off before exposing
872 * the policy to userspace.
874 *policy |= (pol->flags & MPOL_MODE_FLAGS);
878 up_read(¤t->mm->mmap_sem);
884 if (mpol_store_user_nodemask(pol)) {
885 *nmask = pol->w.user_nodemask;
888 get_policy_nodemask(pol, nmask);
889 task_unlock(current);
896 up_read(¤t->mm->mmap_sem);
900 #ifdef CONFIG_MIGRATION
904 static void migrate_page_add(struct page *page, struct list_head *pagelist,
908 * Avoid migrating a page that is shared with others.
910 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
911 if (!isolate_lru_page(page)) {
912 list_add_tail(&page->lru, pagelist);
913 inc_zone_page_state(page, NR_ISOLATED_ANON +
914 page_is_file_cache(page));
919 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
921 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
925 * Migrate pages from one node to a target node.
926 * Returns error or the number of pages not migrated.
928 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
934 struct vm_area_struct *vma;
937 node_set(source, nmask);
939 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
940 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
944 if (!list_empty(&pagelist)) {
945 err = migrate_pages(&pagelist, new_node_page, dest,
946 false, MIGRATE_SYNC);
948 putback_lru_pages(&pagelist);
955 * Move pages between the two nodesets so as to preserve the physical
956 * layout as much as possible.
958 * Returns the number of page that could not be moved.
960 int do_migrate_pages(struct mm_struct *mm,
961 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
967 err = migrate_prep();
971 down_read(&mm->mmap_sem);
973 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
978 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
979 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
980 * bit in 'tmp', and return that <source, dest> pair for migration.
981 * The pair of nodemasks 'to' and 'from' define the map.
983 * If no pair of bits is found that way, fallback to picking some
984 * pair of 'source' and 'dest' bits that are not the same. If the
985 * 'source' and 'dest' bits are the same, this represents a node
986 * that will be migrating to itself, so no pages need move.
988 * If no bits are left in 'tmp', or if all remaining bits left
989 * in 'tmp' correspond to the same bit in 'to', return false
990 * (nothing left to migrate).
992 * This lets us pick a pair of nodes to migrate between, such that
993 * if possible the dest node is not already occupied by some other
994 * source node, minimizing the risk of overloading the memory on a
995 * node that would happen if we migrated incoming memory to a node
996 * before migrating outgoing memory source that same node.
998 * A single scan of tmp is sufficient. As we go, we remember the
999 * most recent <s, d> pair that moved (s != d). If we find a pair
1000 * that not only moved, but what's better, moved to an empty slot
1001 * (d is not set in tmp), then we break out then, with that pair.
1002 * Otherwise when we finish scanning from_tmp, we at least have the
1003 * most recent <s, d> pair that moved. If we get all the way through
1004 * the scan of tmp without finding any node that moved, much less
1005 * moved to an empty node, then there is nothing left worth migrating.
1009 while (!nodes_empty(tmp)) {
1014 for_each_node_mask(s, tmp) {
1015 d = node_remap(s, *from_nodes, *to_nodes);
1019 source = s; /* Node moved. Memorize */
1022 /* dest not in remaining from nodes? */
1023 if (!node_isset(dest, tmp))
1029 node_clear(source, tmp);
1030 err = migrate_to_node(mm, source, dest, flags);
1037 up_read(&mm->mmap_sem);
1045 * Allocate a new page for page migration based on vma policy.
1046 * Start assuming that page is mapped by vma pointed to by @private.
1047 * Search forward from there, if not. N.B., this assumes that the
1048 * list of pages handed to migrate_pages()--which is how we get here--
1049 * is in virtual address order.
1051 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1053 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1054 unsigned long uninitialized_var(address);
1057 address = page_address_in_vma(page, vma);
1058 if (address != -EFAULT)
1064 * if !vma, alloc_page_vma() will use task or system default policy
1066 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1070 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1071 unsigned long flags)
1075 int do_migrate_pages(struct mm_struct *mm,
1076 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1081 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1087 static long do_mbind(unsigned long start, unsigned long len,
1088 unsigned short mode, unsigned short mode_flags,
1089 nodemask_t *nmask, unsigned long flags)
1091 struct vm_area_struct *vma;
1092 struct mm_struct *mm = current->mm;
1093 struct mempolicy *new;
1096 LIST_HEAD(pagelist);
1098 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1099 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1101 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1104 if (start & ~PAGE_MASK)
1107 if (mode == MPOL_DEFAULT)
1108 flags &= ~MPOL_MF_STRICT;
1110 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1118 new = mpol_new(mode, mode_flags, nmask);
1120 return PTR_ERR(new);
1123 * If we are using the default policy then operation
1124 * on discontinuous address spaces is okay after all
1127 flags |= MPOL_MF_DISCONTIG_OK;
1129 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1130 start, start + len, mode, mode_flags,
1131 nmask ? nodes_addr(*nmask)[0] : -1);
1133 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1135 err = migrate_prep();
1140 NODEMASK_SCRATCH(scratch);
1142 down_write(&mm->mmap_sem);
1144 err = mpol_set_nodemask(new, nmask, scratch);
1145 task_unlock(current);
1147 up_write(&mm->mmap_sem);
1150 NODEMASK_SCRATCH_FREE(scratch);
1155 vma = check_range(mm, start, end, nmask,
1156 flags | MPOL_MF_INVERT, &pagelist);
1162 err = mbind_range(mm, start, end, new);
1164 if (!list_empty(&pagelist)) {
1165 nr_failed = migrate_pages(&pagelist, new_vma_page,
1169 putback_lru_pages(&pagelist);
1172 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1175 putback_lru_pages(&pagelist);
1177 up_write(&mm->mmap_sem);
1184 * User space interface with variable sized bitmaps for nodelists.
1187 /* Copy a node mask from user space. */
1188 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1189 unsigned long maxnode)
1192 unsigned long nlongs;
1193 unsigned long endmask;
1196 nodes_clear(*nodes);
1197 if (maxnode == 0 || !nmask)
1199 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1202 nlongs = BITS_TO_LONGS(maxnode);
1203 if ((maxnode % BITS_PER_LONG) == 0)
1206 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1208 /* When the user specified more nodes than supported just check
1209 if the non supported part is all zero. */
1210 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1211 if (nlongs > PAGE_SIZE/sizeof(long))
1213 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1215 if (get_user(t, nmask + k))
1217 if (k == nlongs - 1) {
1223 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1227 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1229 nodes_addr(*nodes)[nlongs-1] &= endmask;
1233 /* Copy a kernel node mask to user space */
1234 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1237 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1238 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1240 if (copy > nbytes) {
1241 if (copy > PAGE_SIZE)
1243 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1247 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1250 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1251 unsigned long, mode, unsigned long __user *, nmask,
1252 unsigned long, maxnode, unsigned, flags)
1256 unsigned short mode_flags;
1258 mode_flags = mode & MPOL_MODE_FLAGS;
1259 mode &= ~MPOL_MODE_FLAGS;
1260 if (mode >= MPOL_MAX)
1262 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1263 (mode_flags & MPOL_F_RELATIVE_NODES))
1265 err = get_nodes(&nodes, nmask, maxnode);
1268 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1271 /* Set the process memory policy */
1272 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1273 unsigned long, maxnode)
1277 unsigned short flags;
1279 flags = mode & MPOL_MODE_FLAGS;
1280 mode &= ~MPOL_MODE_FLAGS;
1281 if ((unsigned int)mode >= MPOL_MAX)
1283 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1285 err = get_nodes(&nodes, nmask, maxnode);
1288 return do_set_mempolicy(mode, flags, &nodes);
1291 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1292 const unsigned long __user *, old_nodes,
1293 const unsigned long __user *, new_nodes)
1295 const struct cred *cred = current_cred(), *tcred;
1296 struct mm_struct *mm = NULL;
1297 struct task_struct *task;
1298 nodemask_t task_nodes;
1302 NODEMASK_SCRATCH(scratch);
1307 old = &scratch->mask1;
1308 new = &scratch->mask2;
1310 err = get_nodes(old, old_nodes, maxnode);
1314 err = get_nodes(new, new_nodes, maxnode);
1318 /* Find the mm_struct */
1320 task = pid ? find_task_by_vpid(pid) : current;
1326 get_task_struct(task);
1331 * Check if this process has the right to modify the specified
1332 * process. The right exists if the process has administrative
1333 * capabilities, superuser privileges or the same
1334 * 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 mm = get_task_mm(task);
1363 put_task_struct(task);
1365 err = do_migrate_pages(mm, old, new,
1366 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1372 NODEMASK_SCRATCH_FREE(scratch);
1377 put_task_struct(task);
1383 /* Retrieve NUMA policy */
1384 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1385 unsigned long __user *, nmask, unsigned long, maxnode,
1386 unsigned long, addr, unsigned long, flags)
1389 int uninitialized_var(pval);
1392 if (nmask != NULL && maxnode < MAX_NUMNODES)
1395 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1400 if (policy && put_user(pval, policy))
1404 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1409 #ifdef CONFIG_COMPAT
1411 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1412 compat_ulong_t __user *nmask,
1413 compat_ulong_t maxnode,
1414 compat_ulong_t addr, compat_ulong_t flags)
1417 unsigned long __user *nm = NULL;
1418 unsigned long nr_bits, alloc_size;
1419 DECLARE_BITMAP(bm, MAX_NUMNODES);
1421 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1422 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1425 nm = compat_alloc_user_space(alloc_size);
1427 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1429 if (!err && nmask) {
1430 unsigned long copy_size;
1431 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1432 err = copy_from_user(bm, nm, copy_size);
1433 /* ensure entire bitmap is zeroed */
1434 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1435 err |= compat_put_bitmap(nmask, bm, nr_bits);
1441 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1442 compat_ulong_t maxnode)
1445 unsigned long __user *nm = NULL;
1446 unsigned long nr_bits, alloc_size;
1447 DECLARE_BITMAP(bm, MAX_NUMNODES);
1449 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1450 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1453 err = compat_get_bitmap(bm, nmask, nr_bits);
1454 nm = compat_alloc_user_space(alloc_size);
1455 err |= copy_to_user(nm, bm, alloc_size);
1461 return sys_set_mempolicy(mode, nm, nr_bits+1);
1464 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1465 compat_ulong_t mode, compat_ulong_t __user *nmask,
1466 compat_ulong_t maxnode, compat_ulong_t flags)
1469 unsigned long __user *nm = NULL;
1470 unsigned long nr_bits, alloc_size;
1473 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1474 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1477 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1478 nm = compat_alloc_user_space(alloc_size);
1479 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1485 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1491 * get_vma_policy(@task, @vma, @addr)
1492 * @task - task for fallback if vma policy == default
1493 * @vma - virtual memory area whose policy is sought
1494 * @addr - address in @vma for shared policy lookup
1496 * Returns effective policy for a VMA at specified address.
1497 * Falls back to @task or system default policy, as necessary.
1498 * Current or other task's task mempolicy and non-shared vma policies
1499 * are protected by the task's mmap_sem, which must be held for read by
1501 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1502 * count--added by the get_policy() vm_op, as appropriate--to protect against
1503 * freeing by another task. It is the caller's responsibility to free the
1504 * extra reference for shared policies.
1506 struct mempolicy *get_vma_policy(struct task_struct *task,
1507 struct vm_area_struct *vma, unsigned long addr)
1509 struct mempolicy *pol = task->mempolicy;
1512 if (vma->vm_ops && vma->vm_ops->get_policy) {
1513 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1517 } else if (vma->vm_policy)
1518 pol = vma->vm_policy;
1521 pol = &default_policy;
1526 * Return a nodemask representing a mempolicy for filtering nodes for
1529 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1531 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1532 if (unlikely(policy->mode == MPOL_BIND) &&
1533 gfp_zone(gfp) >= policy_zone &&
1534 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1535 return &policy->v.nodes;
1540 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1541 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1544 switch (policy->mode) {
1545 case MPOL_PREFERRED:
1546 if (!(policy->flags & MPOL_F_LOCAL))
1547 nd = policy->v.preferred_node;
1551 * Normally, MPOL_BIND allocations are node-local within the
1552 * allowed nodemask. However, if __GFP_THISNODE is set and the
1553 * current node isn't part of the mask, we use the zonelist for
1554 * the first node in the mask instead.
1556 if (unlikely(gfp & __GFP_THISNODE) &&
1557 unlikely(!node_isset(nd, policy->v.nodes)))
1558 nd = first_node(policy->v.nodes);
1563 return node_zonelist(nd, gfp);
1566 /* Do dynamic interleaving for a process */
1567 static unsigned interleave_nodes(struct mempolicy *policy)
1570 struct task_struct *me = current;
1573 next = next_node(nid, policy->v.nodes);
1574 if (next >= MAX_NUMNODES)
1575 next = first_node(policy->v.nodes);
1576 if (next < MAX_NUMNODES)
1582 * Depending on the memory policy provide a node from which to allocate the
1584 * @policy must be protected by freeing by the caller. If @policy is
1585 * the current task's mempolicy, this protection is implicit, as only the
1586 * task can change it's policy. The system default policy requires no
1589 unsigned slab_node(struct mempolicy *policy)
1591 if (!policy || policy->flags & MPOL_F_LOCAL)
1592 return numa_node_id();
1594 switch (policy->mode) {
1595 case MPOL_PREFERRED:
1597 * handled MPOL_F_LOCAL above
1599 return policy->v.preferred_node;
1601 case MPOL_INTERLEAVE:
1602 return interleave_nodes(policy);
1606 * Follow bind policy behavior and start allocation at the
1609 struct zonelist *zonelist;
1611 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1612 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1613 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1616 return zone ? zone->node : numa_node_id();
1624 /* Do static interleaving for a VMA with known offset. */
1625 static unsigned offset_il_node(struct mempolicy *pol,
1626 struct vm_area_struct *vma, unsigned long off)
1628 unsigned nnodes = nodes_weight(pol->v.nodes);
1634 return numa_node_id();
1635 target = (unsigned int)off % nnodes;
1638 nid = next_node(nid, pol->v.nodes);
1640 } while (c <= target);
1644 /* Determine a node number for interleave */
1645 static inline unsigned interleave_nid(struct mempolicy *pol,
1646 struct vm_area_struct *vma, unsigned long addr, int shift)
1652 * for small pages, there is no difference between
1653 * shift and PAGE_SHIFT, so the bit-shift is safe.
1654 * for huge pages, since vm_pgoff is in units of small
1655 * pages, we need to shift off the always 0 bits to get
1658 BUG_ON(shift < PAGE_SHIFT);
1659 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1660 off += (addr - vma->vm_start) >> shift;
1661 return offset_il_node(pol, vma, off);
1663 return interleave_nodes(pol);
1667 * Return the bit number of a random bit set in the nodemask.
1668 * (returns -1 if nodemask is empty)
1670 int node_random(const nodemask_t *maskp)
1674 w = nodes_weight(*maskp);
1676 bit = bitmap_ord_to_pos(maskp->bits,
1677 get_random_int() % w, MAX_NUMNODES);
1681 #ifdef CONFIG_HUGETLBFS
1683 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1684 * @vma = virtual memory area whose policy is sought
1685 * @addr = address in @vma for shared policy lookup and interleave policy
1686 * @gfp_flags = for requested zone
1687 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1688 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1690 * Returns a zonelist suitable for a huge page allocation and a pointer
1691 * to the struct mempolicy for conditional unref after allocation.
1692 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1693 * @nodemask for filtering the zonelist.
1695 * Must be protected by get_mems_allowed()
1697 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1698 gfp_t gfp_flags, struct mempolicy **mpol,
1699 nodemask_t **nodemask)
1701 struct zonelist *zl;
1703 *mpol = get_vma_policy(current, vma, addr);
1704 *nodemask = NULL; /* assume !MPOL_BIND */
1706 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1707 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1708 huge_page_shift(hstate_vma(vma))), gfp_flags);
1710 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1711 if ((*mpol)->mode == MPOL_BIND)
1712 *nodemask = &(*mpol)->v.nodes;
1718 * init_nodemask_of_mempolicy
1720 * If the current task's mempolicy is "default" [NULL], return 'false'
1721 * to indicate default policy. Otherwise, extract the policy nodemask
1722 * for 'bind' or 'interleave' policy into the argument nodemask, or
1723 * initialize the argument nodemask to contain the single node for
1724 * 'preferred' or 'local' policy and return 'true' to indicate presence
1725 * of non-default mempolicy.
1727 * We don't bother with reference counting the mempolicy [mpol_get/put]
1728 * because the current task is examining it's own mempolicy and a task's
1729 * mempolicy is only ever changed by the task itself.
1731 * N.B., it is the caller's responsibility to free a returned nodemask.
1733 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1735 struct mempolicy *mempolicy;
1738 if (!(mask && current->mempolicy))
1742 mempolicy = current->mempolicy;
1743 switch (mempolicy->mode) {
1744 case MPOL_PREFERRED:
1745 if (mempolicy->flags & MPOL_F_LOCAL)
1746 nid = numa_node_id();
1748 nid = mempolicy->v.preferred_node;
1749 init_nodemask_of_node(mask, nid);
1754 case MPOL_INTERLEAVE:
1755 *mask = mempolicy->v.nodes;
1761 task_unlock(current);
1768 * mempolicy_nodemask_intersects
1770 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1771 * policy. Otherwise, check for intersection between mask and the policy
1772 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1773 * policy, always return true since it may allocate elsewhere on fallback.
1775 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1777 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1778 const nodemask_t *mask)
1780 struct mempolicy *mempolicy;
1786 mempolicy = tsk->mempolicy;
1790 switch (mempolicy->mode) {
1791 case MPOL_PREFERRED:
1793 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1794 * allocate from, they may fallback to other nodes when oom.
1795 * Thus, it's possible for tsk to have allocated memory from
1800 case MPOL_INTERLEAVE:
1801 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1811 /* Allocate a page in interleaved policy.
1812 Own path because it needs to do special accounting. */
1813 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1816 struct zonelist *zl;
1819 zl = node_zonelist(nid, gfp);
1820 page = __alloc_pages(gfp, order, zl);
1821 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1822 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1827 * alloc_pages_vma - Allocate a page for a VMA.
1830 * %GFP_USER user allocation.
1831 * %GFP_KERNEL kernel allocations,
1832 * %GFP_HIGHMEM highmem/user allocations,
1833 * %GFP_FS allocation should not call back into a file system.
1834 * %GFP_ATOMIC don't sleep.
1836 * @order:Order of the GFP allocation.
1837 * @vma: Pointer to VMA or NULL if not available.
1838 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1840 * This function allocates a page from the kernel page pool and applies
1841 * a NUMA policy associated with the VMA or the current process.
1842 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1843 * mm_struct of the VMA to prevent it from going away. Should be used for
1844 * all allocations for pages that will be mapped into
1845 * user space. Returns NULL when no page can be allocated.
1847 * Should be called with the mm_sem of the vma hold.
1850 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1851 unsigned long addr, int node)
1853 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1854 struct zonelist *zl;
1858 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1861 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1863 page = alloc_page_interleave(gfp, order, nid);
1867 zl = policy_zonelist(gfp, pol, node);
1868 if (unlikely(mpol_needs_cond_ref(pol))) {
1870 * slow path: ref counted shared policy
1872 struct page *page = __alloc_pages_nodemask(gfp, order,
1873 zl, policy_nodemask(gfp, pol));
1879 * fast path: default or task policy
1881 page = __alloc_pages_nodemask(gfp, order, zl,
1882 policy_nodemask(gfp, pol));
1888 * alloc_pages_current - Allocate pages.
1891 * %GFP_USER user allocation,
1892 * %GFP_KERNEL kernel allocation,
1893 * %GFP_HIGHMEM highmem allocation,
1894 * %GFP_FS don't call back into a file system.
1895 * %GFP_ATOMIC don't sleep.
1896 * @order: Power of two of allocation size in pages. 0 is a single page.
1898 * Allocate a page from the kernel page pool. When not in
1899 * interrupt context and apply the current process NUMA policy.
1900 * Returns NULL when no page can be allocated.
1902 * Don't call cpuset_update_task_memory_state() unless
1903 * 1) it's ok to take cpuset_sem (can WAIT), and
1904 * 2) allocating for current task (not interrupt).
1906 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1908 struct mempolicy *pol = current->mempolicy;
1911 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1912 pol = &default_policy;
1916 * No reference counting needed for current->mempolicy
1917 * nor system default_policy
1919 if (pol->mode == MPOL_INTERLEAVE)
1920 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1922 page = __alloc_pages_nodemask(gfp, order,
1923 policy_zonelist(gfp, pol, numa_node_id()),
1924 policy_nodemask(gfp, pol));
1928 EXPORT_SYMBOL(alloc_pages_current);
1931 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1932 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1933 * with the mems_allowed returned by cpuset_mems_allowed(). This
1934 * keeps mempolicies cpuset relative after its cpuset moves. See
1935 * further kernel/cpuset.c update_nodemask().
1937 * current's mempolicy may be rebinded by the other task(the task that changes
1938 * cpuset's mems), so we needn't do rebind work for current task.
1941 /* Slow path of a mempolicy duplicate */
1942 struct mempolicy *__mpol_dup(struct mempolicy *old)
1944 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1947 return ERR_PTR(-ENOMEM);
1949 /* task's mempolicy is protected by alloc_lock */
1950 if (old == current->mempolicy) {
1953 task_unlock(current);
1958 if (current_cpuset_is_being_rebound()) {
1959 nodemask_t mems = cpuset_mems_allowed(current);
1960 if (new->flags & MPOL_F_REBINDING)
1961 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1963 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1966 atomic_set(&new->refcnt, 1);
1971 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1972 * eliminate the * MPOL_F_* flags that require conditional ref and
1973 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1974 * after return. Use the returned value.
1976 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1977 * policy lookup, even if the policy needs/has extra ref on lookup.
1978 * shmem_readahead needs this.
1980 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1981 struct mempolicy *frompol)
1983 if (!mpol_needs_cond_ref(frompol))
1987 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1988 __mpol_put(frompol);
1992 /* Slow path of a mempolicy comparison */
1993 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1997 if (a->mode != b->mode)
1999 if (a->flags != b->flags)
2001 if (mpol_store_user_nodemask(a))
2002 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2008 case MPOL_INTERLEAVE:
2009 return !!nodes_equal(a->v.nodes, b->v.nodes);
2010 case MPOL_PREFERRED:
2011 return a->v.preferred_node == b->v.preferred_node;
2019 * Shared memory backing store policy support.
2021 * Remember policies even when nobody has shared memory mapped.
2022 * The policies are kept in Red-Black tree linked from the inode.
2023 * They are protected by the sp->lock spinlock, which should be held
2024 * for any accesses to the tree.
2027 /* lookup first element intersecting start-end */
2028 /* Caller holds sp->lock */
2029 static struct sp_node *
2030 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2032 struct rb_node *n = sp->root.rb_node;
2035 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2037 if (start >= p->end)
2039 else if (end <= p->start)
2047 struct sp_node *w = NULL;
2048 struct rb_node *prev = rb_prev(n);
2051 w = rb_entry(prev, struct sp_node, nd);
2052 if (w->end <= start)
2056 return rb_entry(n, struct sp_node, nd);
2059 /* Insert a new shared policy into the list. */
2060 /* Caller holds sp->lock */
2061 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2063 struct rb_node **p = &sp->root.rb_node;
2064 struct rb_node *parent = NULL;
2069 nd = rb_entry(parent, struct sp_node, nd);
2070 if (new->start < nd->start)
2072 else if (new->end > nd->end)
2073 p = &(*p)->rb_right;
2077 rb_link_node(&new->nd, parent, p);
2078 rb_insert_color(&new->nd, &sp->root);
2079 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2080 new->policy ? new->policy->mode : 0);
2083 /* Find shared policy intersecting idx */
2085 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2087 struct mempolicy *pol = NULL;
2090 if (!sp->root.rb_node)
2092 spin_lock(&sp->lock);
2093 sn = sp_lookup(sp, idx, idx+1);
2095 mpol_get(sn->policy);
2098 spin_unlock(&sp->lock);
2102 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2104 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2105 rb_erase(&n->nd, &sp->root);
2106 mpol_put(n->policy);
2107 kmem_cache_free(sn_cache, n);
2110 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2111 struct mempolicy *pol)
2113 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2120 pol->flags |= MPOL_F_SHARED; /* for unref */
2125 /* Replace a policy range. */
2126 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2127 unsigned long end, struct sp_node *new)
2129 struct sp_node *n, *new2 = NULL;
2132 spin_lock(&sp->lock);
2133 n = sp_lookup(sp, start, end);
2134 /* Take care of old policies in the same range. */
2135 while (n && n->start < end) {
2136 struct rb_node *next = rb_next(&n->nd);
2137 if (n->start >= start) {
2143 /* Old policy spanning whole new range. */
2146 spin_unlock(&sp->lock);
2147 new2 = sp_alloc(end, n->end, n->policy);
2153 sp_insert(sp, new2);
2161 n = rb_entry(next, struct sp_node, nd);
2165 spin_unlock(&sp->lock);
2167 mpol_put(new2->policy);
2168 kmem_cache_free(sn_cache, new2);
2174 * mpol_shared_policy_init - initialize shared policy for inode
2175 * @sp: pointer to inode shared policy
2176 * @mpol: struct mempolicy to install
2178 * Install non-NULL @mpol in inode's shared policy rb-tree.
2179 * On entry, the current task has a reference on a non-NULL @mpol.
2180 * This must be released on exit.
2181 * This is called at get_inode() calls and we can use GFP_KERNEL.
2183 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2187 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2188 spin_lock_init(&sp->lock);
2191 struct vm_area_struct pvma;
2192 struct mempolicy *new;
2193 NODEMASK_SCRATCH(scratch);
2197 /* contextualize the tmpfs mount point mempolicy */
2198 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2200 goto free_scratch; /* no valid nodemask intersection */
2203 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2204 task_unlock(current);
2208 /* Create pseudo-vma that contains just the policy */
2209 memset(&pvma, 0, sizeof(struct vm_area_struct));
2210 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2211 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2214 mpol_put(new); /* drop initial ref */
2216 NODEMASK_SCRATCH_FREE(scratch);
2218 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2222 int mpol_set_shared_policy(struct shared_policy *info,
2223 struct vm_area_struct *vma, struct mempolicy *npol)
2226 struct sp_node *new = NULL;
2227 unsigned long sz = vma_pages(vma);
2229 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2231 sz, npol ? npol->mode : -1,
2232 npol ? npol->flags : -1,
2233 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2236 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2240 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2242 kmem_cache_free(sn_cache, new);
2246 /* Free a backing policy store on inode delete. */
2247 void mpol_free_shared_policy(struct shared_policy *p)
2250 struct rb_node *next;
2252 if (!p->root.rb_node)
2254 spin_lock(&p->lock);
2255 next = rb_first(&p->root);
2257 n = rb_entry(next, struct sp_node, nd);
2258 next = rb_next(&n->nd);
2259 rb_erase(&n->nd, &p->root);
2260 mpol_put(n->policy);
2261 kmem_cache_free(sn_cache, n);
2263 spin_unlock(&p->lock);
2266 /* assumes fs == KERNEL_DS */
2267 void __init numa_policy_init(void)
2269 nodemask_t interleave_nodes;
2270 unsigned long largest = 0;
2271 int nid, prefer = 0;
2273 policy_cache = kmem_cache_create("numa_policy",
2274 sizeof(struct mempolicy),
2275 0, SLAB_PANIC, NULL);
2277 sn_cache = kmem_cache_create("shared_policy_node",
2278 sizeof(struct sp_node),
2279 0, SLAB_PANIC, NULL);
2282 * Set interleaving policy for system init. Interleaving is only
2283 * enabled across suitably sized nodes (default is >= 16MB), or
2284 * fall back to the largest node if they're all smaller.
2286 nodes_clear(interleave_nodes);
2287 for_each_node_state(nid, N_HIGH_MEMORY) {
2288 unsigned long total_pages = node_present_pages(nid);
2290 /* Preserve the largest node */
2291 if (largest < total_pages) {
2292 largest = total_pages;
2296 /* Interleave this node? */
2297 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2298 node_set(nid, interleave_nodes);
2301 /* All too small, use the largest */
2302 if (unlikely(nodes_empty(interleave_nodes)))
2303 node_set(prefer, interleave_nodes);
2305 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2306 printk("numa_policy_init: interleaving failed\n");
2309 /* Reset policy of current process to default */
2310 void numa_default_policy(void)
2312 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2316 * Parse and format mempolicy from/to strings
2320 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2321 * Used only for mpol_parse_str() and mpol_to_str()
2323 #define MPOL_LOCAL MPOL_MAX
2324 static const char * const policy_modes[] =
2326 [MPOL_DEFAULT] = "default",
2327 [MPOL_PREFERRED] = "prefer",
2328 [MPOL_BIND] = "bind",
2329 [MPOL_INTERLEAVE] = "interleave",
2330 [MPOL_LOCAL] = "local"
2336 * mpol_parse_str - parse string to mempolicy
2337 * @str: string containing mempolicy to parse
2338 * @mpol: pointer to struct mempolicy pointer, returned on success.
2339 * @no_context: flag whether to "contextualize" the mempolicy
2342 * <mode>[=<flags>][:<nodelist>]
2344 * if @no_context is true, save the input nodemask in w.user_nodemask in
2345 * the returned mempolicy. This will be used to "clone" the mempolicy in
2346 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2347 * mount option. Note that if 'static' or 'relative' mode flags were
2348 * specified, the input nodemask will already have been saved. Saving
2349 * it again is redundant, but safe.
2351 * On success, returns 0, else 1
2353 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2355 struct mempolicy *new = NULL;
2356 unsigned short mode;
2357 unsigned short uninitialized_var(mode_flags);
2359 char *nodelist = strchr(str, ':');
2360 char *flags = strchr(str, '=');
2364 /* NUL-terminate mode or flags string */
2366 if (nodelist_parse(nodelist, nodes))
2368 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2374 *flags++ = '\0'; /* terminate mode string */
2376 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2377 if (!strcmp(str, policy_modes[mode])) {
2381 if (mode > MPOL_LOCAL)
2385 case MPOL_PREFERRED:
2387 * Insist on a nodelist of one node only
2390 char *rest = nodelist;
2391 while (isdigit(*rest))
2397 case MPOL_INTERLEAVE:
2399 * Default to online nodes with memory if no nodelist
2402 nodes = node_states[N_HIGH_MEMORY];
2406 * Don't allow a nodelist; mpol_new() checks flags
2410 mode = MPOL_PREFERRED;
2414 * Insist on a empty nodelist
2421 * Insist on a nodelist
2430 * Currently, we only support two mutually exclusive
2433 if (!strcmp(flags, "static"))
2434 mode_flags |= MPOL_F_STATIC_NODES;
2435 else if (!strcmp(flags, "relative"))
2436 mode_flags |= MPOL_F_RELATIVE_NODES;
2441 new = mpol_new(mode, mode_flags, &nodes);
2446 /* save for contextualization */
2447 new->w.user_nodemask = nodes;
2450 NODEMASK_SCRATCH(scratch);
2453 ret = mpol_set_nodemask(new, &nodes, scratch);
2454 task_unlock(current);
2457 NODEMASK_SCRATCH_FREE(scratch);
2466 /* Restore string for error message */
2475 #endif /* CONFIG_TMPFS */
2478 * mpol_to_str - format a mempolicy structure for printing
2479 * @buffer: to contain formatted mempolicy string
2480 * @maxlen: length of @buffer
2481 * @pol: pointer to mempolicy to be formatted
2482 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2484 * Convert a mempolicy into a string.
2485 * Returns the number of characters in buffer (if positive)
2486 * or an error (negative)
2488 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2493 unsigned short mode;
2494 unsigned short flags = pol ? pol->flags : 0;
2497 * Sanity check: room for longest mode, flag and some nodes
2499 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2501 if (!pol || pol == &default_policy)
2502 mode = MPOL_DEFAULT;
2511 case MPOL_PREFERRED:
2513 if (flags & MPOL_F_LOCAL)
2514 mode = MPOL_LOCAL; /* pseudo-policy */
2516 node_set(pol->v.preferred_node, nodes);
2521 case MPOL_INTERLEAVE:
2523 nodes = pol->w.user_nodemask;
2525 nodes = pol->v.nodes;
2532 l = strlen(policy_modes[mode]);
2533 if (buffer + maxlen < p + l + 1)
2536 strcpy(p, policy_modes[mode]);
2539 if (flags & MPOL_MODE_FLAGS) {
2540 if (buffer + maxlen < p + 2)
2545 * Currently, the only defined flags are mutually exclusive
2547 if (flags & MPOL_F_STATIC_NODES)
2548 p += snprintf(p, buffer + maxlen - p, "static");
2549 else if (flags & MPOL_F_RELATIVE_NODES)
2550 p += snprintf(p, buffer + maxlen - p, "relative");
2553 if (!nodes_empty(nodes)) {
2554 if (buffer + maxlen < p + 2)
2557 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);