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 NUMA_NO_NODE 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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/export.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/ksm.h>
90 #include <linux/rmap.h>
91 #include <linux/security.h>
92 #include <linux/syscalls.h>
93 #include <linux/ctype.h>
94 #include <linux/mm_inline.h>
95 #include <linux/mmu_notifier.h>
96 #include <linux/printk.h>
98 #include <asm/tlbflush.h>
99 #include <asm/uaccess.h>
100 #include <linux/random.h>
102 #include "internal.h"
105 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
106 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
108 static struct kmem_cache *policy_cache;
109 static struct kmem_cache *sn_cache;
111 /* Highest zone. An specific allocation for a zone below that is not
113 enum zone_type policy_zone = 0;
116 * run-time system-wide default policy => local allocation
118 static struct mempolicy default_policy = {
119 .refcnt = ATOMIC_INIT(1), /* never free it */
120 .mode = MPOL_PREFERRED,
121 .flags = MPOL_F_LOCAL,
124 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
126 struct mempolicy *get_task_policy(struct task_struct *p)
128 struct mempolicy *pol = p->mempolicy;
134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
142 return &default_policy;
145 static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
153 * If we have a lock to protect task->mempolicy in read-side, we do
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
163 } mpol_ops[MPOL_MAX];
165 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
167 return pol->flags & MPOL_MODE_FLAGS;
170 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
171 const nodemask_t *rel)
174 nodes_fold(tmp, *orig, nodes_weight(*rel));
175 nodes_onto(*ret, tmp, *rel);
178 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
180 if (nodes_empty(*nodes))
182 pol->v.nodes = *nodes;
186 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
189 pol->flags |= MPOL_F_LOCAL; /* local allocation */
190 else if (nodes_empty(*nodes))
191 return -EINVAL; /* no allowed nodes */
193 pol->v.preferred_node = first_node(*nodes);
197 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
199 if (nodes_empty(*nodes))
201 pol->v.nodes = *nodes;
206 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
207 * any, for the new policy. mpol_new() has already validated the nodes
208 * parameter with respect to the policy mode and flags. But, we need to
209 * handle an empty nodemask with MPOL_PREFERRED here.
211 * Must be called holding task's alloc_lock to protect task's mems_allowed
212 * and mempolicy. May also be called holding the mmap_semaphore for write.
214 static int mpol_set_nodemask(struct mempolicy *pol,
215 const nodemask_t *nodes, struct nodemask_scratch *nsc)
219 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
223 nodes_and(nsc->mask1,
224 cpuset_current_mems_allowed, node_states[N_MEMORY]);
227 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
228 nodes = NULL; /* explicit local allocation */
230 if (pol->flags & MPOL_F_RELATIVE_NODES)
231 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
233 nodes_and(nsc->mask2, *nodes, nsc->mask1);
235 if (mpol_store_user_nodemask(pol))
236 pol->w.user_nodemask = *nodes;
238 pol->w.cpuset_mems_allowed =
239 cpuset_current_mems_allowed;
243 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
245 ret = mpol_ops[pol->mode].create(pol, NULL);
250 * This function just creates a new policy, does some check and simple
251 * initialization. You must invoke mpol_set_nodemask() to set nodes.
253 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
256 struct mempolicy *policy;
258 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
259 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
261 if (mode == MPOL_DEFAULT) {
262 if (nodes && !nodes_empty(*nodes))
263 return ERR_PTR(-EINVAL);
269 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
270 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
271 * All other modes require a valid pointer to a non-empty nodemask.
273 if (mode == MPOL_PREFERRED) {
274 if (nodes_empty(*nodes)) {
275 if (((flags & MPOL_F_STATIC_NODES) ||
276 (flags & MPOL_F_RELATIVE_NODES)))
277 return ERR_PTR(-EINVAL);
279 } else if (mode == MPOL_LOCAL) {
280 if (!nodes_empty(*nodes))
281 return ERR_PTR(-EINVAL);
282 mode = MPOL_PREFERRED;
283 } else if (nodes_empty(*nodes))
284 return ERR_PTR(-EINVAL);
285 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
287 return ERR_PTR(-ENOMEM);
288 atomic_set(&policy->refcnt, 1);
290 policy->flags = flags;
295 /* Slow path of a mpol destructor. */
296 void __mpol_put(struct mempolicy *p)
298 if (!atomic_dec_and_test(&p->refcnt))
300 kmem_cache_free(policy_cache, p);
303 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
310 * MPOL_REBIND_ONCE - do rebind work at once
311 * MPOL_REBIND_STEP1 - set all the newly nodes
312 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
314 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
315 enum mpol_rebind_step step)
319 if (pol->flags & MPOL_F_STATIC_NODES)
320 nodes_and(tmp, pol->w.user_nodemask, *nodes);
321 else if (pol->flags & MPOL_F_RELATIVE_NODES)
322 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
325 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
328 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
329 nodes_remap(tmp, pol->v.nodes,
330 pol->w.cpuset_mems_allowed, *nodes);
331 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
332 } else if (step == MPOL_REBIND_STEP2) {
333 tmp = pol->w.cpuset_mems_allowed;
334 pol->w.cpuset_mems_allowed = *nodes;
339 if (nodes_empty(tmp))
342 if (step == MPOL_REBIND_STEP1)
343 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
344 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
349 if (!node_isset(current->il_next, tmp)) {
350 current->il_next = next_node(current->il_next, tmp);
351 if (current->il_next >= MAX_NUMNODES)
352 current->il_next = first_node(tmp);
353 if (current->il_next >= MAX_NUMNODES)
354 current->il_next = numa_node_id();
358 static void mpol_rebind_preferred(struct mempolicy *pol,
359 const nodemask_t *nodes,
360 enum mpol_rebind_step step)
364 if (pol->flags & MPOL_F_STATIC_NODES) {
365 int node = first_node(pol->w.user_nodemask);
367 if (node_isset(node, *nodes)) {
368 pol->v.preferred_node = node;
369 pol->flags &= ~MPOL_F_LOCAL;
371 pol->flags |= MPOL_F_LOCAL;
372 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
373 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
374 pol->v.preferred_node = first_node(tmp);
375 } else if (!(pol->flags & MPOL_F_LOCAL)) {
376 pol->v.preferred_node = node_remap(pol->v.preferred_node,
377 pol->w.cpuset_mems_allowed,
379 pol->w.cpuset_mems_allowed = *nodes;
384 * mpol_rebind_policy - Migrate a policy to a different set of nodes
386 * If read-side task has no lock to protect task->mempolicy, write-side
387 * task will rebind the task->mempolicy by two step. The first step is
388 * setting all the newly nodes, and the second step is cleaning all the
389 * disallowed nodes. In this way, we can avoid finding no node to alloc
391 * If we have a lock to protect task->mempolicy in read-side, we do
395 * MPOL_REBIND_ONCE - do rebind work at once
396 * MPOL_REBIND_STEP1 - set all the newly nodes
397 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
399 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
400 enum mpol_rebind_step step)
404 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
405 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
408 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
411 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
414 if (step == MPOL_REBIND_STEP1)
415 pol->flags |= MPOL_F_REBINDING;
416 else if (step == MPOL_REBIND_STEP2)
417 pol->flags &= ~MPOL_F_REBINDING;
418 else if (step >= MPOL_REBIND_NSTEP)
421 mpol_ops[pol->mode].rebind(pol, newmask, step);
425 * Wrapper for mpol_rebind_policy() that just requires task
426 * pointer, and updates task mempolicy.
428 * Called with task's alloc_lock held.
431 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
432 enum mpol_rebind_step step)
434 mpol_rebind_policy(tsk->mempolicy, new, step);
438 * Rebind each vma in mm to new nodemask.
440 * Call holding a reference to mm. Takes mm->mmap_sem during call.
443 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
445 struct vm_area_struct *vma;
447 down_write(&mm->mmap_sem);
448 for (vma = mm->mmap; vma; vma = vma->vm_next)
449 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
450 up_write(&mm->mmap_sem);
453 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
455 .rebind = mpol_rebind_default,
457 [MPOL_INTERLEAVE] = {
458 .create = mpol_new_interleave,
459 .rebind = mpol_rebind_nodemask,
462 .create = mpol_new_preferred,
463 .rebind = mpol_rebind_preferred,
466 .create = mpol_new_bind,
467 .rebind = mpol_rebind_nodemask,
471 static void migrate_page_add(struct page *page, struct list_head *pagelist,
472 unsigned long flags);
475 struct list_head *pagelist;
478 struct vm_area_struct *prev;
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
485 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
486 unsigned long end, struct mm_walk *walk)
488 struct vm_area_struct *vma = walk->vma;
490 struct queue_pages *qp = walk->private;
491 unsigned long flags = qp->flags;
496 split_huge_page_pmd(vma, addr, pmd);
497 if (pmd_trans_unstable(pmd))
500 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
501 for (; addr != end; pte++, addr += PAGE_SIZE) {
502 if (!pte_present(*pte))
504 page = vm_normal_page(vma, addr, *pte);
508 * vm_normal_page() filters out zero pages, but there might
509 * still be PageReserved pages to skip, perhaps in a VDSO.
511 if (PageReserved(page))
513 nid = page_to_nid(page);
514 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
517 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
518 if (!vma_migratable(vma))
520 migrate_page_add(page, qp->pagelist, flags);
524 pte_unmap_unlock(pte - 1, ptl);
526 return addr != end ? -EIO : 0;
529 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
530 unsigned long addr, unsigned long end,
531 struct mm_walk *walk)
533 #ifdef CONFIG_HUGETLB_PAGE
534 struct queue_pages *qp = walk->private;
535 unsigned long flags = qp->flags;
541 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
542 entry = huge_ptep_get(pte);
543 if (!pte_present(entry))
545 page = pte_page(entry);
546 nid = page_to_nid(page);
547 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
549 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
550 if (flags & (MPOL_MF_MOVE_ALL) ||
551 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
552 isolate_huge_page(page, qp->pagelist);
561 #ifdef CONFIG_NUMA_BALANCING
563 * This is used to mark a range of virtual addresses to be inaccessible.
564 * These are later cleared by a NUMA hinting fault. Depending on these
565 * faults, pages may be migrated for better NUMA placement.
567 * This is assuming that NUMA faults are handled using PROT_NONE. If
568 * an architecture makes a different choice, it will need further
569 * changes to the core.
571 unsigned long change_prot_numa(struct vm_area_struct *vma,
572 unsigned long addr, unsigned long end)
576 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
578 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
583 static unsigned long change_prot_numa(struct vm_area_struct *vma,
584 unsigned long addr, unsigned long end)
588 #endif /* CONFIG_NUMA_BALANCING */
590 static int queue_pages_test_walk(unsigned long start, unsigned long end,
591 struct mm_walk *walk)
593 struct vm_area_struct *vma = walk->vma;
594 struct queue_pages *qp = walk->private;
595 unsigned long endvma = vma->vm_end;
596 unsigned long flags = qp->flags;
598 if (vma->vm_flags & VM_PFNMAP)
603 if (vma->vm_start > start)
604 start = vma->vm_start;
606 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
607 if (!vma->vm_next && vma->vm_end < end)
609 if (qp->prev && qp->prev->vm_end < vma->vm_start)
615 if (flags & MPOL_MF_LAZY) {
616 /* Similar to task_numa_work, skip inaccessible VMAs */
617 if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
618 change_prot_numa(vma, start, endvma);
622 if ((flags & MPOL_MF_STRICT) ||
623 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
624 vma_migratable(vma)))
625 /* queue pages from current vma */
631 * Walk through page tables and collect pages to be migrated.
633 * If pages found in a given range are on a set of nodes (determined by
634 * @nodes and @flags,) it's isolated and queued to the pagelist which is
635 * passed via @private.)
638 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
639 nodemask_t *nodes, unsigned long flags,
640 struct list_head *pagelist)
642 struct queue_pages qp = {
643 .pagelist = pagelist,
648 struct mm_walk queue_pages_walk = {
649 .hugetlb_entry = queue_pages_hugetlb,
650 .pmd_entry = queue_pages_pte_range,
651 .test_walk = queue_pages_test_walk,
656 return walk_page_range(start, end, &queue_pages_walk);
660 * Apply policy to a single VMA
661 * This must be called with the mmap_sem held for writing.
663 static int vma_replace_policy(struct vm_area_struct *vma,
664 struct mempolicy *pol)
667 struct mempolicy *old;
668 struct mempolicy *new;
670 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
671 vma->vm_start, vma->vm_end, vma->vm_pgoff,
672 vma->vm_ops, vma->vm_file,
673 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
679 if (vma->vm_ops && vma->vm_ops->set_policy) {
680 err = vma->vm_ops->set_policy(vma, new);
685 old = vma->vm_policy;
686 vma->vm_policy = new; /* protected by mmap_sem */
695 /* Step 2: apply policy to a range and do splits. */
696 static int mbind_range(struct mm_struct *mm, unsigned long start,
697 unsigned long end, struct mempolicy *new_pol)
699 struct vm_area_struct *next;
700 struct vm_area_struct *prev;
701 struct vm_area_struct *vma;
704 unsigned long vmstart;
707 vma = find_vma(mm, start);
708 if (!vma || vma->vm_start > start)
712 if (start > vma->vm_start)
715 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
717 vmstart = max(start, vma->vm_start);
718 vmend = min(end, vma->vm_end);
720 if (mpol_equal(vma_policy(vma), new_pol))
723 pgoff = vma->vm_pgoff +
724 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
725 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
726 vma->anon_vma, vma->vm_file, pgoff,
727 new_pol, vma->vm_userfaultfd_ctx);
731 if (mpol_equal(vma_policy(vma), new_pol))
733 /* vma_merge() joined vma && vma->next, case 8 */
736 if (vma->vm_start != vmstart) {
737 err = split_vma(vma->vm_mm, vma, vmstart, 1);
741 if (vma->vm_end != vmend) {
742 err = split_vma(vma->vm_mm, vma, vmend, 0);
747 err = vma_replace_policy(vma, new_pol);
756 /* Set the process memory policy */
757 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
760 struct mempolicy *new, *old;
761 NODEMASK_SCRATCH(scratch);
767 new = mpol_new(mode, flags, nodes);
774 ret = mpol_set_nodemask(new, nodes, scratch);
776 task_unlock(current);
780 old = current->mempolicy;
781 current->mempolicy = new;
782 if (new && new->mode == MPOL_INTERLEAVE &&
783 nodes_weight(new->v.nodes))
784 current->il_next = first_node(new->v.nodes);
785 task_unlock(current);
789 NODEMASK_SCRATCH_FREE(scratch);
794 * Return nodemask for policy for get_mempolicy() query
796 * Called with task's alloc_lock held
798 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
801 if (p == &default_policy)
807 case MPOL_INTERLEAVE:
811 if (!(p->flags & MPOL_F_LOCAL))
812 node_set(p->v.preferred_node, *nodes);
813 /* else return empty node mask for local allocation */
820 static int lookup_node(struct mm_struct *mm, unsigned long addr)
825 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, &p, NULL);
827 err = page_to_nid(p);
833 /* Retrieve NUMA policy */
834 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
835 unsigned long addr, unsigned long flags)
838 struct mm_struct *mm = current->mm;
839 struct vm_area_struct *vma = NULL;
840 struct mempolicy *pol = current->mempolicy;
843 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
846 if (flags & MPOL_F_MEMS_ALLOWED) {
847 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
849 *policy = 0; /* just so it's initialized */
851 *nmask = cpuset_current_mems_allowed;
852 task_unlock(current);
856 if (flags & MPOL_F_ADDR) {
858 * Do NOT fall back to task policy if the
859 * vma/shared policy at addr is NULL. We
860 * want to return MPOL_DEFAULT in this case.
862 down_read(&mm->mmap_sem);
863 vma = find_vma_intersection(mm, addr, addr+1);
865 up_read(&mm->mmap_sem);
868 if (vma->vm_ops && vma->vm_ops->get_policy)
869 pol = vma->vm_ops->get_policy(vma, addr);
871 pol = vma->vm_policy;
876 pol = &default_policy; /* indicates default behavior */
878 if (flags & MPOL_F_NODE) {
879 if (flags & MPOL_F_ADDR) {
880 err = lookup_node(mm, addr);
884 } else if (pol == current->mempolicy &&
885 pol->mode == MPOL_INTERLEAVE) {
886 *policy = current->il_next;
892 *policy = pol == &default_policy ? MPOL_DEFAULT :
895 * Internal mempolicy flags must be masked off before exposing
896 * the policy to userspace.
898 *policy |= (pol->flags & MPOL_MODE_FLAGS);
903 if (mpol_store_user_nodemask(pol)) {
904 *nmask = pol->w.user_nodemask;
907 get_policy_nodemask(pol, nmask);
908 task_unlock(current);
915 up_read(¤t->mm->mmap_sem);
919 #ifdef CONFIG_MIGRATION
923 static void migrate_page_add(struct page *page, struct list_head *pagelist,
927 * Avoid migrating a page that is shared with others.
929 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
930 if (!isolate_lru_page(page)) {
931 list_add_tail(&page->lru, pagelist);
932 inc_zone_page_state(page, NR_ISOLATED_ANON +
933 page_is_file_cache(page));
938 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
941 return alloc_huge_page_node(page_hstate(compound_head(page)),
944 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
949 * Migrate pages from one node to a target node.
950 * Returns error or the number of pages not migrated.
952 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
960 node_set(source, nmask);
963 * This does not "check" the range but isolates all pages that
964 * need migration. Between passing in the full user address
965 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
967 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
968 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
969 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
971 if (!list_empty(&pagelist)) {
972 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
973 MIGRATE_SYNC, MR_SYSCALL);
975 putback_movable_pages(&pagelist);
982 * Move pages between the two nodesets so as to preserve the physical
983 * layout as much as possible.
985 * Returns the number of page that could not be moved.
987 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
988 const nodemask_t *to, int flags)
994 err = migrate_prep();
998 down_read(&mm->mmap_sem);
1001 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1002 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1003 * bit in 'tmp', and return that <source, dest> pair for migration.
1004 * The pair of nodemasks 'to' and 'from' define the map.
1006 * If no pair of bits is found that way, fallback to picking some
1007 * pair of 'source' and 'dest' bits that are not the same. If the
1008 * 'source' and 'dest' bits are the same, this represents a node
1009 * that will be migrating to itself, so no pages need move.
1011 * If no bits are left in 'tmp', or if all remaining bits left
1012 * in 'tmp' correspond to the same bit in 'to', return false
1013 * (nothing left to migrate).
1015 * This lets us pick a pair of nodes to migrate between, such that
1016 * if possible the dest node is not already occupied by some other
1017 * source node, minimizing the risk of overloading the memory on a
1018 * node that would happen if we migrated incoming memory to a node
1019 * before migrating outgoing memory source that same node.
1021 * A single scan of tmp is sufficient. As we go, we remember the
1022 * most recent <s, d> pair that moved (s != d). If we find a pair
1023 * that not only moved, but what's better, moved to an empty slot
1024 * (d is not set in tmp), then we break out then, with that pair.
1025 * Otherwise when we finish scanning from_tmp, we at least have the
1026 * most recent <s, d> pair that moved. If we get all the way through
1027 * the scan of tmp without finding any node that moved, much less
1028 * moved to an empty node, then there is nothing left worth migrating.
1032 while (!nodes_empty(tmp)) {
1034 int source = NUMA_NO_NODE;
1037 for_each_node_mask(s, tmp) {
1040 * do_migrate_pages() tries to maintain the relative
1041 * node relationship of the pages established between
1042 * threads and memory areas.
1044 * However if the number of source nodes is not equal to
1045 * the number of destination nodes we can not preserve
1046 * this node relative relationship. In that case, skip
1047 * copying memory from a node that is in the destination
1050 * Example: [2,3,4] -> [3,4,5] moves everything.
1051 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1054 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1055 (node_isset(s, *to)))
1058 d = node_remap(s, *from, *to);
1062 source = s; /* Node moved. Memorize */
1065 /* dest not in remaining from nodes? */
1066 if (!node_isset(dest, tmp))
1069 if (source == NUMA_NO_NODE)
1072 node_clear(source, tmp);
1073 err = migrate_to_node(mm, source, dest, flags);
1079 up_read(&mm->mmap_sem);
1087 * Allocate a new page for page migration based on vma policy.
1088 * Start by assuming the page is mapped by the same vma as contains @start.
1089 * Search forward from there, if not. N.B., this assumes that the
1090 * list of pages handed to migrate_pages()--which is how we get here--
1091 * is in virtual address order.
1093 static struct page *new_page(struct page *page, unsigned long start, int **x)
1095 struct vm_area_struct *vma;
1096 unsigned long uninitialized_var(address);
1098 vma = find_vma(current->mm, start);
1100 address = page_address_in_vma(page, vma);
1101 if (address != -EFAULT)
1106 if (PageHuge(page)) {
1108 return alloc_huge_page_noerr(vma, address, 1);
1111 * if !vma, alloc_page_vma() will use task or system default policy
1113 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1117 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1118 unsigned long flags)
1122 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1123 const nodemask_t *to, int flags)
1128 static struct page *new_page(struct page *page, unsigned long start, int **x)
1134 static long do_mbind(unsigned long start, unsigned long len,
1135 unsigned short mode, unsigned short mode_flags,
1136 nodemask_t *nmask, unsigned long flags)
1138 struct mm_struct *mm = current->mm;
1139 struct mempolicy *new;
1142 LIST_HEAD(pagelist);
1144 if (flags & ~(unsigned long)MPOL_MF_VALID)
1146 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1149 if (start & ~PAGE_MASK)
1152 if (mode == MPOL_DEFAULT)
1153 flags &= ~MPOL_MF_STRICT;
1155 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1163 new = mpol_new(mode, mode_flags, nmask);
1165 return PTR_ERR(new);
1167 if (flags & MPOL_MF_LAZY)
1168 new->flags |= MPOL_F_MOF;
1171 * If we are using the default policy then operation
1172 * on discontinuous address spaces is okay after all
1175 flags |= MPOL_MF_DISCONTIG_OK;
1177 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1178 start, start + len, mode, mode_flags,
1179 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1181 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1183 err = migrate_prep();
1188 NODEMASK_SCRATCH(scratch);
1190 down_write(&mm->mmap_sem);
1192 err = mpol_set_nodemask(new, nmask, scratch);
1193 task_unlock(current);
1195 up_write(&mm->mmap_sem);
1198 NODEMASK_SCRATCH_FREE(scratch);
1203 err = queue_pages_range(mm, start, end, nmask,
1204 flags | MPOL_MF_INVERT, &pagelist);
1206 err = mbind_range(mm, start, end, new);
1211 if (!list_empty(&pagelist)) {
1212 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1213 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1214 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1216 putback_movable_pages(&pagelist);
1219 if (nr_failed && (flags & MPOL_MF_STRICT))
1222 putback_movable_pages(&pagelist);
1224 up_write(&mm->mmap_sem);
1231 * User space interface with variable sized bitmaps for nodelists.
1234 /* Copy a node mask from user space. */
1235 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1236 unsigned long maxnode)
1240 unsigned long nlongs;
1241 unsigned long endmask;
1244 nodes_clear(*nodes);
1245 if (maxnode == 0 || !nmask)
1247 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1250 nlongs = BITS_TO_LONGS(maxnode);
1251 if ((maxnode % BITS_PER_LONG) == 0)
1254 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1257 * When the user specified more nodes than supported just check
1258 * if the non supported part is all zero.
1260 * If maxnode have more longs than MAX_NUMNODES, check
1261 * the bits in that area first. And then go through to
1262 * check the rest bits which equal or bigger than MAX_NUMNODES.
1263 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1265 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1266 if (nlongs > PAGE_SIZE/sizeof(long))
1268 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1269 if (get_user(t, nmask + k))
1271 if (k == nlongs - 1) {
1277 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1281 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1282 unsigned long valid_mask = endmask;
1284 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1285 if (get_user(t, nmask + nlongs - 1))
1291 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1293 nodes_addr(*nodes)[nlongs-1] &= endmask;
1297 /* Copy a kernel node mask to user space */
1298 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1301 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1302 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1304 if (copy > nbytes) {
1305 if (copy > PAGE_SIZE)
1307 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1311 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1314 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1315 unsigned long, mode, const unsigned long __user *, nmask,
1316 unsigned long, maxnode, unsigned, flags)
1320 unsigned short mode_flags;
1322 mode_flags = mode & MPOL_MODE_FLAGS;
1323 mode &= ~MPOL_MODE_FLAGS;
1324 if (mode >= MPOL_MAX)
1326 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1327 (mode_flags & MPOL_F_RELATIVE_NODES))
1329 err = get_nodes(&nodes, nmask, maxnode);
1332 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1335 /* Set the process memory policy */
1336 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1337 unsigned long, maxnode)
1341 unsigned short flags;
1343 flags = mode & MPOL_MODE_FLAGS;
1344 mode &= ~MPOL_MODE_FLAGS;
1345 if ((unsigned int)mode >= MPOL_MAX)
1347 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1349 err = get_nodes(&nodes, nmask, maxnode);
1352 return do_set_mempolicy(mode, flags, &nodes);
1355 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1356 const unsigned long __user *, old_nodes,
1357 const unsigned long __user *, new_nodes)
1359 const struct cred *cred = current_cred(), *tcred;
1360 struct mm_struct *mm = NULL;
1361 struct task_struct *task;
1362 nodemask_t task_nodes;
1366 NODEMASK_SCRATCH(scratch);
1371 old = &scratch->mask1;
1372 new = &scratch->mask2;
1374 err = get_nodes(old, old_nodes, maxnode);
1378 err = get_nodes(new, new_nodes, maxnode);
1382 /* Find the mm_struct */
1384 task = pid ? find_task_by_vpid(pid) : current;
1390 get_task_struct(task);
1395 * Check if this process has the right to modify the specified
1396 * process. The right exists if the process has administrative
1397 * capabilities, superuser privileges or the same
1398 * userid as the target process.
1400 tcred = __task_cred(task);
1401 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1402 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1403 !capable(CAP_SYS_NICE)) {
1410 task_nodes = cpuset_mems_allowed(task);
1411 /* Is the user allowed to access the target nodes? */
1412 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1417 task_nodes = cpuset_mems_allowed(current);
1418 nodes_and(*new, *new, task_nodes);
1419 if (nodes_empty(*new))
1422 nodes_and(*new, *new, node_states[N_MEMORY]);
1423 if (nodes_empty(*new))
1426 err = security_task_movememory(task);
1430 mm = get_task_mm(task);
1431 put_task_struct(task);
1438 err = do_migrate_pages(mm, old, new,
1439 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1443 NODEMASK_SCRATCH_FREE(scratch);
1448 put_task_struct(task);
1454 /* Retrieve NUMA policy */
1455 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1456 unsigned long __user *, nmask, unsigned long, maxnode,
1457 unsigned long, addr, unsigned long, flags)
1460 int uninitialized_var(pval);
1463 if (nmask != NULL && maxnode < nr_node_ids)
1466 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1471 if (policy && put_user(pval, policy))
1475 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1480 #ifdef CONFIG_COMPAT
1482 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1483 compat_ulong_t __user *, nmask,
1484 compat_ulong_t, maxnode,
1485 compat_ulong_t, addr, compat_ulong_t, flags)
1488 unsigned long __user *nm = NULL;
1489 unsigned long nr_bits, alloc_size;
1490 DECLARE_BITMAP(bm, MAX_NUMNODES);
1492 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1493 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1496 nm = compat_alloc_user_space(alloc_size);
1498 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1500 if (!err && nmask) {
1501 unsigned long copy_size;
1502 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1503 err = copy_from_user(bm, nm, copy_size);
1504 /* ensure entire bitmap is zeroed */
1505 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1506 err |= compat_put_bitmap(nmask, bm, nr_bits);
1512 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1513 compat_ulong_t, maxnode)
1515 unsigned long __user *nm = NULL;
1516 unsigned long nr_bits, alloc_size;
1517 DECLARE_BITMAP(bm, MAX_NUMNODES);
1519 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1520 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1523 if (compat_get_bitmap(bm, nmask, nr_bits))
1525 nm = compat_alloc_user_space(alloc_size);
1526 if (copy_to_user(nm, bm, alloc_size))
1530 return sys_set_mempolicy(mode, nm, nr_bits+1);
1533 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1534 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1535 compat_ulong_t, maxnode, compat_ulong_t, flags)
1537 unsigned long __user *nm = NULL;
1538 unsigned long nr_bits, alloc_size;
1541 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1542 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1545 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1547 nm = compat_alloc_user_space(alloc_size);
1548 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1552 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1557 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1560 struct mempolicy *pol = NULL;
1563 if (vma->vm_ops && vma->vm_ops->get_policy) {
1564 pol = vma->vm_ops->get_policy(vma, addr);
1565 } else if (vma->vm_policy) {
1566 pol = vma->vm_policy;
1569 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1570 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1571 * count on these policies which will be dropped by
1572 * mpol_cond_put() later
1574 if (mpol_needs_cond_ref(pol))
1583 * get_vma_policy(@vma, @addr)
1584 * @vma: virtual memory area whose policy is sought
1585 * @addr: address in @vma for shared policy lookup
1587 * Returns effective policy for a VMA at specified address.
1588 * Falls back to current->mempolicy or system default policy, as necessary.
1589 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1590 * count--added by the get_policy() vm_op, as appropriate--to protect against
1591 * freeing by another task. It is the caller's responsibility to free the
1592 * extra reference for shared policies.
1594 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1597 struct mempolicy *pol = __get_vma_policy(vma, addr);
1600 pol = get_task_policy(current);
1605 bool vma_policy_mof(struct vm_area_struct *vma)
1607 struct mempolicy *pol;
1609 if (vma->vm_ops && vma->vm_ops->get_policy) {
1612 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1613 if (pol && (pol->flags & MPOL_F_MOF))
1620 pol = vma->vm_policy;
1622 pol = get_task_policy(current);
1624 return pol->flags & MPOL_F_MOF;
1627 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1629 enum zone_type dynamic_policy_zone = policy_zone;
1631 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1634 * if policy->v.nodes has movable memory only,
1635 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1637 * policy->v.nodes is intersect with node_states[N_MEMORY].
1638 * so if the following test faile, it implies
1639 * policy->v.nodes has movable memory only.
1641 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1642 dynamic_policy_zone = ZONE_MOVABLE;
1644 return zone >= dynamic_policy_zone;
1648 * Return a nodemask representing a mempolicy for filtering nodes for
1651 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1653 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1654 if (unlikely(policy->mode == MPOL_BIND) &&
1655 apply_policy_zone(policy, gfp_zone(gfp)) &&
1656 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1657 return &policy->v.nodes;
1662 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1663 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1666 switch (policy->mode) {
1667 case MPOL_PREFERRED:
1668 if (!(policy->flags & MPOL_F_LOCAL))
1669 nd = policy->v.preferred_node;
1673 * Normally, MPOL_BIND allocations are node-local within the
1674 * allowed nodemask. However, if __GFP_THISNODE is set and the
1675 * current node isn't part of the mask, we use the zonelist for
1676 * the first node in the mask instead.
1678 if (unlikely(gfp & __GFP_THISNODE) &&
1679 unlikely(!node_isset(nd, policy->v.nodes)))
1680 nd = first_node(policy->v.nodes);
1685 return node_zonelist(nd, gfp);
1688 /* Do dynamic interleaving for a process */
1689 static unsigned interleave_nodes(struct mempolicy *policy)
1692 struct task_struct *me = current;
1695 next = next_node(nid, policy->v.nodes);
1696 if (next >= MAX_NUMNODES)
1697 next = first_node(policy->v.nodes);
1698 if (next < MAX_NUMNODES)
1704 * Depending on the memory policy provide a node from which to allocate the
1707 unsigned int mempolicy_slab_node(void)
1709 struct mempolicy *policy;
1710 int node = numa_mem_id();
1715 policy = current->mempolicy;
1716 if (!policy || policy->flags & MPOL_F_LOCAL)
1719 switch (policy->mode) {
1720 case MPOL_PREFERRED:
1722 * handled MPOL_F_LOCAL above
1724 return policy->v.preferred_node;
1726 case MPOL_INTERLEAVE:
1727 return interleave_nodes(policy);
1731 * Follow bind policy behavior and start allocation at the
1734 struct zonelist *zonelist;
1736 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1737 zonelist = &NODE_DATA(node)->node_zonelists[0];
1738 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1741 return zone ? zone->node : node;
1749 /* Do static interleaving for a VMA with known offset. */
1750 static unsigned offset_il_node(struct mempolicy *pol,
1751 struct vm_area_struct *vma, unsigned long off)
1753 unsigned nnodes = nodes_weight(pol->v.nodes);
1756 int nid = NUMA_NO_NODE;
1759 return numa_node_id();
1760 target = (unsigned int)off % nnodes;
1763 nid = next_node(nid, pol->v.nodes);
1765 } while (c <= target);
1769 /* Determine a node number for interleave */
1770 static inline unsigned interleave_nid(struct mempolicy *pol,
1771 struct vm_area_struct *vma, unsigned long addr, int shift)
1777 * for small pages, there is no difference between
1778 * shift and PAGE_SHIFT, so the bit-shift is safe.
1779 * for huge pages, since vm_pgoff is in units of small
1780 * pages, we need to shift off the always 0 bits to get
1783 BUG_ON(shift < PAGE_SHIFT);
1784 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1785 off += (addr - vma->vm_start) >> shift;
1786 return offset_il_node(pol, vma, off);
1788 return interleave_nodes(pol);
1792 * Return the bit number of a random bit set in the nodemask.
1793 * (returns NUMA_NO_NODE if nodemask is empty)
1795 int node_random(const nodemask_t *maskp)
1797 int w, bit = NUMA_NO_NODE;
1799 w = nodes_weight(*maskp);
1801 bit = bitmap_ord_to_pos(maskp->bits,
1802 get_random_int() % w, MAX_NUMNODES);
1806 #ifdef CONFIG_HUGETLBFS
1808 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1809 * @vma: virtual memory area whose policy is sought
1810 * @addr: address in @vma for shared policy lookup and interleave policy
1811 * @gfp_flags: for requested zone
1812 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1813 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1815 * Returns a zonelist suitable for a huge page allocation and a pointer
1816 * to the struct mempolicy for conditional unref after allocation.
1817 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1818 * @nodemask for filtering the zonelist.
1820 * Must be protected by read_mems_allowed_begin()
1822 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1823 gfp_t gfp_flags, struct mempolicy **mpol,
1824 nodemask_t **nodemask)
1826 struct zonelist *zl;
1828 *mpol = get_vma_policy(vma, addr);
1829 *nodemask = NULL; /* assume !MPOL_BIND */
1831 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1832 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1833 huge_page_shift(hstate_vma(vma))), gfp_flags);
1835 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1836 if ((*mpol)->mode == MPOL_BIND)
1837 *nodemask = &(*mpol)->v.nodes;
1843 * init_nodemask_of_mempolicy
1845 * If the current task's mempolicy is "default" [NULL], return 'false'
1846 * to indicate default policy. Otherwise, extract the policy nodemask
1847 * for 'bind' or 'interleave' policy into the argument nodemask, or
1848 * initialize the argument nodemask to contain the single node for
1849 * 'preferred' or 'local' policy and return 'true' to indicate presence
1850 * of non-default mempolicy.
1852 * We don't bother with reference counting the mempolicy [mpol_get/put]
1853 * because the current task is examining it's own mempolicy and a task's
1854 * mempolicy is only ever changed by the task itself.
1856 * N.B., it is the caller's responsibility to free a returned nodemask.
1858 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1860 struct mempolicy *mempolicy;
1863 if (!(mask && current->mempolicy))
1867 mempolicy = current->mempolicy;
1868 switch (mempolicy->mode) {
1869 case MPOL_PREFERRED:
1870 if (mempolicy->flags & MPOL_F_LOCAL)
1871 nid = numa_node_id();
1873 nid = mempolicy->v.preferred_node;
1874 init_nodemask_of_node(mask, nid);
1879 case MPOL_INTERLEAVE:
1880 *mask = mempolicy->v.nodes;
1886 task_unlock(current);
1893 * mempolicy_nodemask_intersects
1895 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1896 * policy. Otherwise, check for intersection between mask and the policy
1897 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1898 * policy, always return true since it may allocate elsewhere on fallback.
1900 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1902 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1903 const nodemask_t *mask)
1905 struct mempolicy *mempolicy;
1911 mempolicy = tsk->mempolicy;
1915 switch (mempolicy->mode) {
1916 case MPOL_PREFERRED:
1918 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1919 * allocate from, they may fallback to other nodes when oom.
1920 * Thus, it's possible for tsk to have allocated memory from
1925 case MPOL_INTERLEAVE:
1926 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1936 /* Allocate a page in interleaved policy.
1937 Own path because it needs to do special accounting. */
1938 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1941 struct zonelist *zl;
1944 zl = node_zonelist(nid, gfp);
1945 page = __alloc_pages(gfp, order, zl);
1946 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1947 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1952 * alloc_pages_vma - Allocate a page for a VMA.
1955 * %GFP_USER user allocation.
1956 * %GFP_KERNEL kernel allocations,
1957 * %GFP_HIGHMEM highmem/user allocations,
1958 * %GFP_FS allocation should not call back into a file system.
1959 * %GFP_ATOMIC don't sleep.
1961 * @order:Order of the GFP allocation.
1962 * @vma: Pointer to VMA or NULL if not available.
1963 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1964 * @node: Which node to prefer for allocation (modulo policy).
1965 * @hugepage: for hugepages try only the preferred node if possible
1967 * This function allocates a page from the kernel page pool and applies
1968 * a NUMA policy associated with the VMA or the current process.
1969 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1970 * mm_struct of the VMA to prevent it from going away. Should be used for
1971 * all allocations for pages that will be mapped into user space. Returns
1972 * NULL when no page can be allocated.
1975 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1976 unsigned long addr, int node, bool hugepage)
1978 struct mempolicy *pol;
1980 unsigned int cpuset_mems_cookie;
1981 struct zonelist *zl;
1985 pol = get_vma_policy(vma, addr);
1986 cpuset_mems_cookie = read_mems_allowed_begin();
1988 if (pol->mode == MPOL_INTERLEAVE) {
1991 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1993 page = alloc_page_interleave(gfp, order, nid);
1997 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1998 int hpage_node = node;
2001 * For hugepage allocation and non-interleave policy which
2002 * allows the current node (or other explicitly preferred
2003 * node) we only try to allocate from the current/preferred
2004 * node and don't fall back to other nodes, as the cost of
2005 * remote accesses would likely offset THP benefits.
2007 * If the policy is interleave, or does not allow the current
2008 * node in its nodemask, we allocate the standard way.
2010 if (pol->mode == MPOL_PREFERRED &&
2011 !(pol->flags & MPOL_F_LOCAL))
2012 hpage_node = pol->v.preferred_node;
2014 nmask = policy_nodemask(gfp, pol);
2015 if (!nmask || node_isset(hpage_node, *nmask)) {
2018 * We cannot invoke reclaim if __GFP_THISNODE
2019 * is set. Invoking reclaim with
2020 * __GFP_THISNODE set, would cause THP
2021 * allocations to trigger heavy swapping
2022 * despite there may be tons of free memory
2023 * (including potentially plenty of THP
2024 * already available in the buddy) on all the
2027 * At most we could invoke compaction when
2028 * __GFP_THISNODE is set (but we would need to
2029 * refrain from invoking reclaim even if
2030 * compaction returned COMPACT_SKIPPED because
2031 * there wasn't not enough memory to succeed
2032 * compaction). For now just avoid
2033 * __GFP_THISNODE instead of limiting the
2034 * allocation path to a strict and single
2035 * compaction invocation.
2037 * Supposedly if direct reclaim was enabled by
2038 * the caller, the app prefers THP regardless
2039 * of the node it comes from so this would be
2040 * more desiderable behavior than only
2041 * providing THP originated from the local
2042 * node in such case.
2044 if (!(gfp & __GFP_DIRECT_RECLAIM))
2045 gfp |= __GFP_THISNODE;
2046 page = __alloc_pages_node(hpage_node, gfp, order);
2051 nmask = policy_nodemask(gfp, pol);
2052 zl = policy_zonelist(gfp, pol, node);
2053 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2056 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2062 * alloc_pages_current - Allocate pages.
2065 * %GFP_USER user allocation,
2066 * %GFP_KERNEL kernel allocation,
2067 * %GFP_HIGHMEM highmem allocation,
2068 * %GFP_FS don't call back into a file system.
2069 * %GFP_ATOMIC don't sleep.
2070 * @order: Power of two of allocation size in pages. 0 is a single page.
2072 * Allocate a page from the kernel page pool. When not in
2073 * interrupt context and apply the current process NUMA policy.
2074 * Returns NULL when no page can be allocated.
2076 * Don't call cpuset_update_task_memory_state() unless
2077 * 1) it's ok to take cpuset_sem (can WAIT), and
2078 * 2) allocating for current task (not interrupt).
2080 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2082 struct mempolicy *pol = &default_policy;
2084 unsigned int cpuset_mems_cookie;
2086 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2087 pol = get_task_policy(current);
2090 cpuset_mems_cookie = read_mems_allowed_begin();
2093 * No reference counting needed for current->mempolicy
2094 * nor system default_policy
2096 if (pol->mode == MPOL_INTERLEAVE)
2097 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2099 page = __alloc_pages_nodemask(gfp, order,
2100 policy_zonelist(gfp, pol, numa_node_id()),
2101 policy_nodemask(gfp, pol));
2103 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2108 EXPORT_SYMBOL(alloc_pages_current);
2110 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2112 struct mempolicy *pol = mpol_dup(vma_policy(src));
2115 return PTR_ERR(pol);
2116 dst->vm_policy = pol;
2121 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2122 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2123 * with the mems_allowed returned by cpuset_mems_allowed(). This
2124 * keeps mempolicies cpuset relative after its cpuset moves. See
2125 * further kernel/cpuset.c update_nodemask().
2127 * current's mempolicy may be rebinded by the other task(the task that changes
2128 * cpuset's mems), so we needn't do rebind work for current task.
2131 /* Slow path of a mempolicy duplicate */
2132 struct mempolicy *__mpol_dup(struct mempolicy *old)
2134 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2137 return ERR_PTR(-ENOMEM);
2139 /* task's mempolicy is protected by alloc_lock */
2140 if (old == current->mempolicy) {
2143 task_unlock(current);
2147 if (current_cpuset_is_being_rebound()) {
2148 nodemask_t mems = cpuset_mems_allowed(current);
2149 if (new->flags & MPOL_F_REBINDING)
2150 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2152 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2154 atomic_set(&new->refcnt, 1);
2158 /* Slow path of a mempolicy comparison */
2159 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2163 if (a->mode != b->mode)
2165 if (a->flags != b->flags)
2167 if (mpol_store_user_nodemask(a))
2168 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2174 case MPOL_INTERLEAVE:
2175 return !!nodes_equal(a->v.nodes, b->v.nodes);
2176 case MPOL_PREFERRED:
2177 /* a's ->flags is the same as b's */
2178 if (a->flags & MPOL_F_LOCAL)
2180 return a->v.preferred_node == b->v.preferred_node;
2188 * Shared memory backing store policy support.
2190 * Remember policies even when nobody has shared memory mapped.
2191 * The policies are kept in Red-Black tree linked from the inode.
2192 * They are protected by the sp->lock spinlock, which should be held
2193 * for any accesses to the tree.
2196 /* lookup first element intersecting start-end */
2197 /* Caller holds sp->lock */
2198 static struct sp_node *
2199 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2201 struct rb_node *n = sp->root.rb_node;
2204 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2206 if (start >= p->end)
2208 else if (end <= p->start)
2216 struct sp_node *w = NULL;
2217 struct rb_node *prev = rb_prev(n);
2220 w = rb_entry(prev, struct sp_node, nd);
2221 if (w->end <= start)
2225 return rb_entry(n, struct sp_node, nd);
2228 /* Insert a new shared policy into the list. */
2229 /* Caller holds sp->lock */
2230 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2232 struct rb_node **p = &sp->root.rb_node;
2233 struct rb_node *parent = NULL;
2238 nd = rb_entry(parent, struct sp_node, nd);
2239 if (new->start < nd->start)
2241 else if (new->end > nd->end)
2242 p = &(*p)->rb_right;
2246 rb_link_node(&new->nd, parent, p);
2247 rb_insert_color(&new->nd, &sp->root);
2248 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2249 new->policy ? new->policy->mode : 0);
2252 /* Find shared policy intersecting idx */
2254 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2256 struct mempolicy *pol = NULL;
2259 if (!sp->root.rb_node)
2261 spin_lock(&sp->lock);
2262 sn = sp_lookup(sp, idx, idx+1);
2264 mpol_get(sn->policy);
2267 spin_unlock(&sp->lock);
2271 static void sp_free(struct sp_node *n)
2273 mpol_put(n->policy);
2274 kmem_cache_free(sn_cache, n);
2278 * mpol_misplaced - check whether current page node is valid in policy
2280 * @page: page to be checked
2281 * @vma: vm area where page mapped
2282 * @addr: virtual address where page mapped
2284 * Lookup current policy node id for vma,addr and "compare to" page's
2288 * -1 - not misplaced, page is in the right node
2289 * node - node id where the page should be
2291 * Policy determination "mimics" alloc_page_vma().
2292 * Called from fault path where we know the vma and faulting address.
2294 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2296 struct mempolicy *pol;
2298 int curnid = page_to_nid(page);
2299 unsigned long pgoff;
2300 int thiscpu = raw_smp_processor_id();
2301 int thisnid = cpu_to_node(thiscpu);
2307 pol = get_vma_policy(vma, addr);
2308 if (!(pol->flags & MPOL_F_MOF))
2311 switch (pol->mode) {
2312 case MPOL_INTERLEAVE:
2313 BUG_ON(addr >= vma->vm_end);
2314 BUG_ON(addr < vma->vm_start);
2316 pgoff = vma->vm_pgoff;
2317 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2318 polnid = offset_il_node(pol, vma, pgoff);
2321 case MPOL_PREFERRED:
2322 if (pol->flags & MPOL_F_LOCAL)
2323 polnid = numa_node_id();
2325 polnid = pol->v.preferred_node;
2330 * allows binding to multiple nodes.
2331 * use current page if in policy nodemask,
2332 * else select nearest allowed node, if any.
2333 * If no allowed nodes, use current [!misplaced].
2335 if (node_isset(curnid, pol->v.nodes))
2337 (void)first_zones_zonelist(
2338 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2339 gfp_zone(GFP_HIGHUSER),
2340 &pol->v.nodes, &zone);
2341 polnid = zone->node;
2348 /* Migrate the page towards the node whose CPU is referencing it */
2349 if (pol->flags & MPOL_F_MORON) {
2352 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2356 if (curnid != polnid)
2364 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2366 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2367 rb_erase(&n->nd, &sp->root);
2371 static void sp_node_init(struct sp_node *node, unsigned long start,
2372 unsigned long end, struct mempolicy *pol)
2374 node->start = start;
2379 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2380 struct mempolicy *pol)
2383 struct mempolicy *newpol;
2385 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2389 newpol = mpol_dup(pol);
2390 if (IS_ERR(newpol)) {
2391 kmem_cache_free(sn_cache, n);
2394 newpol->flags |= MPOL_F_SHARED;
2395 sp_node_init(n, start, end, newpol);
2400 /* Replace a policy range. */
2401 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2402 unsigned long end, struct sp_node *new)
2405 struct sp_node *n_new = NULL;
2406 struct mempolicy *mpol_new = NULL;
2410 spin_lock(&sp->lock);
2411 n = sp_lookup(sp, start, end);
2412 /* Take care of old policies in the same range. */
2413 while (n && n->start < end) {
2414 struct rb_node *next = rb_next(&n->nd);
2415 if (n->start >= start) {
2421 /* Old policy spanning whole new range. */
2426 *mpol_new = *n->policy;
2427 atomic_set(&mpol_new->refcnt, 1);
2428 sp_node_init(n_new, end, n->end, mpol_new);
2430 sp_insert(sp, n_new);
2439 n = rb_entry(next, struct sp_node, nd);
2443 spin_unlock(&sp->lock);
2450 kmem_cache_free(sn_cache, n_new);
2455 spin_unlock(&sp->lock);
2457 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2460 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2467 * mpol_shared_policy_init - initialize shared policy for inode
2468 * @sp: pointer to inode shared policy
2469 * @mpol: struct mempolicy to install
2471 * Install non-NULL @mpol in inode's shared policy rb-tree.
2472 * On entry, the current task has a reference on a non-NULL @mpol.
2473 * This must be released on exit.
2474 * This is called at get_inode() calls and we can use GFP_KERNEL.
2476 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2480 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2481 spin_lock_init(&sp->lock);
2484 struct vm_area_struct pvma;
2485 struct mempolicy *new;
2486 NODEMASK_SCRATCH(scratch);
2490 /* contextualize the tmpfs mount point mempolicy */
2491 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2493 goto free_scratch; /* no valid nodemask intersection */
2496 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2497 task_unlock(current);
2501 /* Create pseudo-vma that contains just the policy */
2502 memset(&pvma, 0, sizeof(struct vm_area_struct));
2503 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2504 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2507 mpol_put(new); /* drop initial ref */
2509 NODEMASK_SCRATCH_FREE(scratch);
2511 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2515 int mpol_set_shared_policy(struct shared_policy *info,
2516 struct vm_area_struct *vma, struct mempolicy *npol)
2519 struct sp_node *new = NULL;
2520 unsigned long sz = vma_pages(vma);
2522 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2524 sz, npol ? npol->mode : -1,
2525 npol ? npol->flags : -1,
2526 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2529 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2533 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2539 /* Free a backing policy store on inode delete. */
2540 void mpol_free_shared_policy(struct shared_policy *p)
2543 struct rb_node *next;
2545 if (!p->root.rb_node)
2547 spin_lock(&p->lock);
2548 next = rb_first(&p->root);
2550 n = rb_entry(next, struct sp_node, nd);
2551 next = rb_next(&n->nd);
2554 spin_unlock(&p->lock);
2557 #ifdef CONFIG_NUMA_BALANCING
2558 static int __initdata numabalancing_override;
2560 static void __init check_numabalancing_enable(void)
2562 bool numabalancing_default = false;
2564 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2565 numabalancing_default = true;
2567 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2568 if (numabalancing_override)
2569 set_numabalancing_state(numabalancing_override == 1);
2571 if (num_online_nodes() > 1 && !numabalancing_override) {
2572 pr_info("%s automatic NUMA balancing. "
2573 "Configure with numa_balancing= or the "
2574 "kernel.numa_balancing sysctl",
2575 numabalancing_default ? "Enabling" : "Disabling");
2576 set_numabalancing_state(numabalancing_default);
2580 static int __init setup_numabalancing(char *str)
2586 if (!strcmp(str, "enable")) {
2587 numabalancing_override = 1;
2589 } else if (!strcmp(str, "disable")) {
2590 numabalancing_override = -1;
2595 pr_warn("Unable to parse numa_balancing=\n");
2599 __setup("numa_balancing=", setup_numabalancing);
2601 static inline void __init check_numabalancing_enable(void)
2604 #endif /* CONFIG_NUMA_BALANCING */
2606 /* assumes fs == KERNEL_DS */
2607 void __init numa_policy_init(void)
2609 nodemask_t interleave_nodes;
2610 unsigned long largest = 0;
2611 int nid, prefer = 0;
2613 policy_cache = kmem_cache_create("numa_policy",
2614 sizeof(struct mempolicy),
2615 0, SLAB_PANIC, NULL);
2617 sn_cache = kmem_cache_create("shared_policy_node",
2618 sizeof(struct sp_node),
2619 0, SLAB_PANIC, NULL);
2621 for_each_node(nid) {
2622 preferred_node_policy[nid] = (struct mempolicy) {
2623 .refcnt = ATOMIC_INIT(1),
2624 .mode = MPOL_PREFERRED,
2625 .flags = MPOL_F_MOF | MPOL_F_MORON,
2626 .v = { .preferred_node = nid, },
2631 * Set interleaving policy for system init. Interleaving is only
2632 * enabled across suitably sized nodes (default is >= 16MB), or
2633 * fall back to the largest node if they're all smaller.
2635 nodes_clear(interleave_nodes);
2636 for_each_node_state(nid, N_MEMORY) {
2637 unsigned long total_pages = node_present_pages(nid);
2639 /* Preserve the largest node */
2640 if (largest < total_pages) {
2641 largest = total_pages;
2645 /* Interleave this node? */
2646 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2647 node_set(nid, interleave_nodes);
2650 /* All too small, use the largest */
2651 if (unlikely(nodes_empty(interleave_nodes)))
2652 node_set(prefer, interleave_nodes);
2654 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2655 pr_err("%s: interleaving failed\n", __func__);
2657 check_numabalancing_enable();
2660 /* Reset policy of current process to default */
2661 void numa_default_policy(void)
2663 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2667 * Parse and format mempolicy from/to strings
2671 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2673 static const char * const policy_modes[] =
2675 [MPOL_DEFAULT] = "default",
2676 [MPOL_PREFERRED] = "prefer",
2677 [MPOL_BIND] = "bind",
2678 [MPOL_INTERLEAVE] = "interleave",
2679 [MPOL_LOCAL] = "local",
2685 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2686 * @str: string containing mempolicy to parse
2687 * @mpol: pointer to struct mempolicy pointer, returned on success.
2690 * <mode>[=<flags>][:<nodelist>]
2692 * On success, returns 0, else 1
2694 int mpol_parse_str(char *str, struct mempolicy **mpol)
2696 struct mempolicy *new = NULL;
2697 unsigned short mode;
2698 unsigned short mode_flags;
2700 char *nodelist = strchr(str, ':');
2701 char *flags = strchr(str, '=');
2705 /* NUL-terminate mode or flags string */
2707 if (nodelist_parse(nodelist, nodes))
2709 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2715 *flags++ = '\0'; /* terminate mode string */
2717 for (mode = 0; mode < MPOL_MAX; mode++) {
2718 if (!strcmp(str, policy_modes[mode])) {
2722 if (mode >= MPOL_MAX)
2726 case MPOL_PREFERRED:
2728 * Insist on a nodelist of one node only
2731 char *rest = nodelist;
2732 while (isdigit(*rest))
2738 case MPOL_INTERLEAVE:
2740 * Default to online nodes with memory if no nodelist
2743 nodes = node_states[N_MEMORY];
2747 * Don't allow a nodelist; mpol_new() checks flags
2751 mode = MPOL_PREFERRED;
2755 * Insist on a empty nodelist
2762 * Insist on a nodelist
2771 * Currently, we only support two mutually exclusive
2774 if (!strcmp(flags, "static"))
2775 mode_flags |= MPOL_F_STATIC_NODES;
2776 else if (!strcmp(flags, "relative"))
2777 mode_flags |= MPOL_F_RELATIVE_NODES;
2782 new = mpol_new(mode, mode_flags, &nodes);
2787 * Save nodes for mpol_to_str() to show the tmpfs mount options
2788 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2790 if (mode != MPOL_PREFERRED)
2791 new->v.nodes = nodes;
2793 new->v.preferred_node = first_node(nodes);
2795 new->flags |= MPOL_F_LOCAL;
2798 * Save nodes for contextualization: this will be used to "clone"
2799 * the mempolicy in a specific context [cpuset] at a later time.
2801 new->w.user_nodemask = nodes;
2806 /* Restore string for error message */
2815 #endif /* CONFIG_TMPFS */
2818 * mpol_to_str - format a mempolicy structure for printing
2819 * @buffer: to contain formatted mempolicy string
2820 * @maxlen: length of @buffer
2821 * @pol: pointer to mempolicy to be formatted
2823 * Convert @pol into a string. If @buffer is too short, truncate the string.
2824 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2825 * longest flag, "relative", and to display at least a few node ids.
2827 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2830 nodemask_t nodes = NODE_MASK_NONE;
2831 unsigned short mode = MPOL_DEFAULT;
2832 unsigned short flags = 0;
2834 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2842 case MPOL_PREFERRED:
2843 if (flags & MPOL_F_LOCAL)
2846 node_set(pol->v.preferred_node, nodes);
2849 case MPOL_INTERLEAVE:
2850 nodes = pol->v.nodes;
2854 snprintf(p, maxlen, "unknown");
2858 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2860 if (flags & MPOL_MODE_FLAGS) {
2861 p += snprintf(p, buffer + maxlen - p, "=");
2864 * Currently, the only defined flags are mutually exclusive
2866 if (flags & MPOL_F_STATIC_NODES)
2867 p += snprintf(p, buffer + maxlen - p, "static");
2868 else if (flags & MPOL_F_RELATIVE_NODES)
2869 p += snprintf(p, buffer + maxlen - p, "relative");
2872 if (!nodes_empty(nodes))
2873 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2874 nodemask_pr_args(&nodes));