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 if (pmd_trans_huge(*pmd)) {
497 ptl = pmd_lock(walk->mm, pmd);
498 if (pmd_trans_huge(*pmd)) {
499 page = pmd_page(*pmd);
500 if (is_huge_zero_page(page)) {
502 split_huge_pmd(vma, pmd, addr);
507 ret = split_huge_page(page);
517 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
518 for (; addr != end; pte++, addr += PAGE_SIZE) {
519 if (!pte_present(*pte))
521 page = vm_normal_page(vma, addr, *pte);
525 * vm_normal_page() filters out zero pages, but there might
526 * still be PageReserved pages to skip, perhaps in a VDSO.
528 if (PageReserved(page))
530 nid = page_to_nid(page);
531 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
533 if (PageTail(page) && PageAnon(page)) {
535 pte_unmap_unlock(pte, ptl);
537 ret = split_huge_page(page);
540 /* Failed to split -- skip. */
542 pte = pte_offset_map_lock(walk->mm, pmd,
549 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
550 migrate_page_add(page, qp->pagelist, flags);
552 pte_unmap_unlock(pte - 1, ptl);
557 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
558 unsigned long addr, unsigned long end,
559 struct mm_walk *walk)
561 #ifdef CONFIG_HUGETLB_PAGE
562 struct queue_pages *qp = walk->private;
563 unsigned long flags = qp->flags;
569 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
570 entry = huge_ptep_get(pte);
571 if (!pte_present(entry))
573 page = pte_page(entry);
574 nid = page_to_nid(page);
575 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
577 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
578 if (flags & (MPOL_MF_MOVE_ALL) ||
579 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
580 isolate_huge_page(page, qp->pagelist);
589 #ifdef CONFIG_NUMA_BALANCING
591 * This is used to mark a range of virtual addresses to be inaccessible.
592 * These are later cleared by a NUMA hinting fault. Depending on these
593 * faults, pages may be migrated for better NUMA placement.
595 * This is assuming that NUMA faults are handled using PROT_NONE. If
596 * an architecture makes a different choice, it will need further
597 * changes to the core.
599 unsigned long change_prot_numa(struct vm_area_struct *vma,
600 unsigned long addr, unsigned long end)
604 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
606 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
611 static unsigned long change_prot_numa(struct vm_area_struct *vma,
612 unsigned long addr, unsigned long end)
616 #endif /* CONFIG_NUMA_BALANCING */
618 static int queue_pages_test_walk(unsigned long start, unsigned long end,
619 struct mm_walk *walk)
621 struct vm_area_struct *vma = walk->vma;
622 struct queue_pages *qp = walk->private;
623 unsigned long endvma = vma->vm_end;
624 unsigned long flags = qp->flags;
626 if (vma->vm_flags & VM_PFNMAP)
631 if (vma->vm_start > start)
632 start = vma->vm_start;
634 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
635 if (!vma->vm_next && vma->vm_end < end)
637 if (qp->prev && qp->prev->vm_end < vma->vm_start)
643 if (flags & MPOL_MF_LAZY) {
644 /* Similar to task_numa_work, skip inaccessible VMAs */
645 if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
646 change_prot_numa(vma, start, endvma);
650 if ((flags & MPOL_MF_STRICT) ||
651 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
652 vma_migratable(vma)))
653 /* queue pages from current vma */
659 * Walk through page tables and collect pages to be migrated.
661 * If pages found in a given range are on a set of nodes (determined by
662 * @nodes and @flags,) it's isolated and queued to the pagelist which is
663 * passed via @private.)
666 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
667 nodemask_t *nodes, unsigned long flags,
668 struct list_head *pagelist)
670 struct queue_pages qp = {
671 .pagelist = pagelist,
676 struct mm_walk queue_pages_walk = {
677 .hugetlb_entry = queue_pages_hugetlb,
678 .pmd_entry = queue_pages_pte_range,
679 .test_walk = queue_pages_test_walk,
684 return walk_page_range(start, end, &queue_pages_walk);
688 * Apply policy to a single VMA
689 * This must be called with the mmap_sem held for writing.
691 static int vma_replace_policy(struct vm_area_struct *vma,
692 struct mempolicy *pol)
695 struct mempolicy *old;
696 struct mempolicy *new;
698 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
699 vma->vm_start, vma->vm_end, vma->vm_pgoff,
700 vma->vm_ops, vma->vm_file,
701 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
707 if (vma->vm_ops && vma->vm_ops->set_policy) {
708 err = vma->vm_ops->set_policy(vma, new);
713 old = vma->vm_policy;
714 vma->vm_policy = new; /* protected by mmap_sem */
723 /* Step 2: apply policy to a range and do splits. */
724 static int mbind_range(struct mm_struct *mm, unsigned long start,
725 unsigned long end, struct mempolicy *new_pol)
727 struct vm_area_struct *next;
728 struct vm_area_struct *prev;
729 struct vm_area_struct *vma;
732 unsigned long vmstart;
735 vma = find_vma(mm, start);
736 if (!vma || vma->vm_start > start)
740 if (start > vma->vm_start)
743 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
745 vmstart = max(start, vma->vm_start);
746 vmend = min(end, vma->vm_end);
748 if (mpol_equal(vma_policy(vma), new_pol))
751 pgoff = vma->vm_pgoff +
752 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
753 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
754 vma->anon_vma, vma->vm_file, pgoff,
755 new_pol, vma->vm_userfaultfd_ctx);
759 if (mpol_equal(vma_policy(vma), new_pol))
761 /* vma_merge() joined vma && vma->next, case 8 */
764 if (vma->vm_start != vmstart) {
765 err = split_vma(vma->vm_mm, vma, vmstart, 1);
769 if (vma->vm_end != vmend) {
770 err = split_vma(vma->vm_mm, vma, vmend, 0);
775 err = vma_replace_policy(vma, new_pol);
784 /* Set the process memory policy */
785 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
788 struct mempolicy *new, *old;
789 NODEMASK_SCRATCH(scratch);
795 new = mpol_new(mode, flags, nodes);
802 ret = mpol_set_nodemask(new, nodes, scratch);
804 task_unlock(current);
808 old = current->mempolicy;
809 current->mempolicy = new;
810 if (new && new->mode == MPOL_INTERLEAVE &&
811 nodes_weight(new->v.nodes))
812 current->il_next = first_node(new->v.nodes);
813 task_unlock(current);
817 NODEMASK_SCRATCH_FREE(scratch);
822 * Return nodemask for policy for get_mempolicy() query
824 * Called with task's alloc_lock held
826 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
829 if (p == &default_policy)
835 case MPOL_INTERLEAVE:
839 if (!(p->flags & MPOL_F_LOCAL))
840 node_set(p->v.preferred_node, *nodes);
841 /* else return empty node mask for local allocation */
848 static int lookup_node(struct mm_struct *mm, unsigned long addr)
853 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
855 err = page_to_nid(p);
861 /* Retrieve NUMA policy */
862 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
863 unsigned long addr, unsigned long flags)
866 struct mm_struct *mm = current->mm;
867 struct vm_area_struct *vma = NULL;
868 struct mempolicy *pol = current->mempolicy;
871 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
874 if (flags & MPOL_F_MEMS_ALLOWED) {
875 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
877 *policy = 0; /* just so it's initialized */
879 *nmask = cpuset_current_mems_allowed;
880 task_unlock(current);
884 if (flags & MPOL_F_ADDR) {
886 * Do NOT fall back to task policy if the
887 * vma/shared policy at addr is NULL. We
888 * want to return MPOL_DEFAULT in this case.
890 down_read(&mm->mmap_sem);
891 vma = find_vma_intersection(mm, addr, addr+1);
893 up_read(&mm->mmap_sem);
896 if (vma->vm_ops && vma->vm_ops->get_policy)
897 pol = vma->vm_ops->get_policy(vma, addr);
899 pol = vma->vm_policy;
904 pol = &default_policy; /* indicates default behavior */
906 if (flags & MPOL_F_NODE) {
907 if (flags & MPOL_F_ADDR) {
908 err = lookup_node(mm, addr);
912 } else if (pol == current->mempolicy &&
913 pol->mode == MPOL_INTERLEAVE) {
914 *policy = current->il_next;
920 *policy = pol == &default_policy ? MPOL_DEFAULT :
923 * Internal mempolicy flags must be masked off before exposing
924 * the policy to userspace.
926 *policy |= (pol->flags & MPOL_MODE_FLAGS);
930 up_read(¤t->mm->mmap_sem);
936 if (mpol_store_user_nodemask(pol)) {
937 *nmask = pol->w.user_nodemask;
940 get_policy_nodemask(pol, nmask);
941 task_unlock(current);
948 up_read(¤t->mm->mmap_sem);
952 #ifdef CONFIG_MIGRATION
956 static void migrate_page_add(struct page *page, struct list_head *pagelist,
960 * Avoid migrating a page that is shared with others.
962 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
963 if (!isolate_lru_page(page)) {
964 list_add_tail(&page->lru, pagelist);
965 inc_zone_page_state(page, NR_ISOLATED_ANON +
966 page_is_file_cache(page));
971 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
974 return alloc_huge_page_node(page_hstate(compound_head(page)),
977 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
982 * Migrate pages from one node to a target node.
983 * Returns error or the number of pages not migrated.
985 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
993 node_set(source, nmask);
996 * This does not "check" the range but isolates all pages that
997 * need migration. Between passing in the full user address
998 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1000 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1001 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1002 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1004 if (!list_empty(&pagelist)) {
1005 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1006 MIGRATE_SYNC, MR_SYSCALL);
1008 putback_movable_pages(&pagelist);
1015 * Move pages between the two nodesets so as to preserve the physical
1016 * layout as much as possible.
1018 * Returns the number of page that could not be moved.
1020 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1021 const nodemask_t *to, int flags)
1027 err = migrate_prep();
1031 down_read(&mm->mmap_sem);
1034 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1035 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1036 * bit in 'tmp', and return that <source, dest> pair for migration.
1037 * The pair of nodemasks 'to' and 'from' define the map.
1039 * If no pair of bits is found that way, fallback to picking some
1040 * pair of 'source' and 'dest' bits that are not the same. If the
1041 * 'source' and 'dest' bits are the same, this represents a node
1042 * that will be migrating to itself, so no pages need move.
1044 * If no bits are left in 'tmp', or if all remaining bits left
1045 * in 'tmp' correspond to the same bit in 'to', return false
1046 * (nothing left to migrate).
1048 * This lets us pick a pair of nodes to migrate between, such that
1049 * if possible the dest node is not already occupied by some other
1050 * source node, minimizing the risk of overloading the memory on a
1051 * node that would happen if we migrated incoming memory to a node
1052 * before migrating outgoing memory source that same node.
1054 * A single scan of tmp is sufficient. As we go, we remember the
1055 * most recent <s, d> pair that moved (s != d). If we find a pair
1056 * that not only moved, but what's better, moved to an empty slot
1057 * (d is not set in tmp), then we break out then, with that pair.
1058 * Otherwise when we finish scanning from_tmp, we at least have the
1059 * most recent <s, d> pair that moved. If we get all the way through
1060 * the scan of tmp without finding any node that moved, much less
1061 * moved to an empty node, then there is nothing left worth migrating.
1065 while (!nodes_empty(tmp)) {
1067 int source = NUMA_NO_NODE;
1070 for_each_node_mask(s, tmp) {
1073 * do_migrate_pages() tries to maintain the relative
1074 * node relationship of the pages established between
1075 * threads and memory areas.
1077 * However if the number of source nodes is not equal to
1078 * the number of destination nodes we can not preserve
1079 * this node relative relationship. In that case, skip
1080 * copying memory from a node that is in the destination
1083 * Example: [2,3,4] -> [3,4,5] moves everything.
1084 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1087 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1088 (node_isset(s, *to)))
1091 d = node_remap(s, *from, *to);
1095 source = s; /* Node moved. Memorize */
1098 /* dest not in remaining from nodes? */
1099 if (!node_isset(dest, tmp))
1102 if (source == NUMA_NO_NODE)
1105 node_clear(source, tmp);
1106 err = migrate_to_node(mm, source, dest, flags);
1112 up_read(&mm->mmap_sem);
1120 * Allocate a new page for page migration based on vma policy.
1121 * Start by assuming the page is mapped by the same vma as contains @start.
1122 * Search forward from there, if not. N.B., this assumes that the
1123 * list of pages handed to migrate_pages()--which is how we get here--
1124 * is in virtual address order.
1126 static struct page *new_page(struct page *page, unsigned long start, int **x)
1128 struct vm_area_struct *vma;
1129 unsigned long uninitialized_var(address);
1131 vma = find_vma(current->mm, start);
1133 address = page_address_in_vma(page, vma);
1134 if (address != -EFAULT)
1139 if (PageHuge(page)) {
1141 return alloc_huge_page_noerr(vma, address, 1);
1144 * if !vma, alloc_page_vma() will use task or system default policy
1146 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1150 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1151 unsigned long flags)
1155 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1156 const nodemask_t *to, int flags)
1161 static struct page *new_page(struct page *page, unsigned long start, int **x)
1167 static long do_mbind(unsigned long start, unsigned long len,
1168 unsigned short mode, unsigned short mode_flags,
1169 nodemask_t *nmask, unsigned long flags)
1171 struct mm_struct *mm = current->mm;
1172 struct mempolicy *new;
1175 LIST_HEAD(pagelist);
1177 if (flags & ~(unsigned long)MPOL_MF_VALID)
1179 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1182 if (start & ~PAGE_MASK)
1185 if (mode == MPOL_DEFAULT)
1186 flags &= ~MPOL_MF_STRICT;
1188 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1196 new = mpol_new(mode, mode_flags, nmask);
1198 return PTR_ERR(new);
1200 if (flags & MPOL_MF_LAZY)
1201 new->flags |= MPOL_F_MOF;
1204 * If we are using the default policy then operation
1205 * on discontinuous address spaces is okay after all
1208 flags |= MPOL_MF_DISCONTIG_OK;
1210 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1211 start, start + len, mode, mode_flags,
1212 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1214 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1216 err = migrate_prep();
1221 NODEMASK_SCRATCH(scratch);
1223 down_write(&mm->mmap_sem);
1225 err = mpol_set_nodemask(new, nmask, scratch);
1226 task_unlock(current);
1228 up_write(&mm->mmap_sem);
1231 NODEMASK_SCRATCH_FREE(scratch);
1236 err = queue_pages_range(mm, start, end, nmask,
1237 flags | MPOL_MF_INVERT, &pagelist);
1239 err = mbind_range(mm, start, end, new);
1244 if (!list_empty(&pagelist)) {
1245 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1246 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1247 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1249 putback_movable_pages(&pagelist);
1252 if (nr_failed && (flags & MPOL_MF_STRICT))
1255 putback_movable_pages(&pagelist);
1257 up_write(&mm->mmap_sem);
1264 * User space interface with variable sized bitmaps for nodelists.
1267 /* Copy a node mask from user space. */
1268 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1269 unsigned long maxnode)
1272 unsigned long nlongs;
1273 unsigned long endmask;
1276 nodes_clear(*nodes);
1277 if (maxnode == 0 || !nmask)
1279 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1282 nlongs = BITS_TO_LONGS(maxnode);
1283 if ((maxnode % BITS_PER_LONG) == 0)
1286 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1288 /* When the user specified more nodes than supported just check
1289 if the non supported part is all zero. */
1290 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1291 if (nlongs > PAGE_SIZE/sizeof(long))
1293 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1295 if (get_user(t, nmask + k))
1297 if (k == nlongs - 1) {
1303 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1307 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1309 nodes_addr(*nodes)[nlongs-1] &= endmask;
1313 /* Copy a kernel node mask to user space */
1314 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1317 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1318 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1320 if (copy > nbytes) {
1321 if (copy > PAGE_SIZE)
1323 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1327 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1330 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1331 unsigned long, mode, const unsigned long __user *, nmask,
1332 unsigned long, maxnode, unsigned, flags)
1336 unsigned short mode_flags;
1338 mode_flags = mode & MPOL_MODE_FLAGS;
1339 mode &= ~MPOL_MODE_FLAGS;
1340 if (mode >= MPOL_MAX)
1342 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1343 (mode_flags & MPOL_F_RELATIVE_NODES))
1345 err = get_nodes(&nodes, nmask, maxnode);
1348 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1351 /* Set the process memory policy */
1352 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1353 unsigned long, maxnode)
1357 unsigned short flags;
1359 flags = mode & MPOL_MODE_FLAGS;
1360 mode &= ~MPOL_MODE_FLAGS;
1361 if ((unsigned int)mode >= MPOL_MAX)
1363 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1365 err = get_nodes(&nodes, nmask, maxnode);
1368 return do_set_mempolicy(mode, flags, &nodes);
1371 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1372 const unsigned long __user *, old_nodes,
1373 const unsigned long __user *, new_nodes)
1375 const struct cred *cred = current_cred(), *tcred;
1376 struct mm_struct *mm = NULL;
1377 struct task_struct *task;
1378 nodemask_t task_nodes;
1382 NODEMASK_SCRATCH(scratch);
1387 old = &scratch->mask1;
1388 new = &scratch->mask2;
1390 err = get_nodes(old, old_nodes, maxnode);
1394 err = get_nodes(new, new_nodes, maxnode);
1398 /* Find the mm_struct */
1400 task = pid ? find_task_by_vpid(pid) : current;
1406 get_task_struct(task);
1411 * Check if this process has the right to modify the specified
1412 * process. The right exists if the process has administrative
1413 * capabilities, superuser privileges or the same
1414 * userid as the target process.
1416 tcred = __task_cred(task);
1417 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1418 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1419 !capable(CAP_SYS_NICE)) {
1426 task_nodes = cpuset_mems_allowed(task);
1427 /* Is the user allowed to access the target nodes? */
1428 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1433 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1438 err = security_task_movememory(task);
1442 mm = get_task_mm(task);
1443 put_task_struct(task);
1450 err = do_migrate_pages(mm, old, new,
1451 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1455 NODEMASK_SCRATCH_FREE(scratch);
1460 put_task_struct(task);
1466 /* Retrieve NUMA policy */
1467 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1468 unsigned long __user *, nmask, unsigned long, maxnode,
1469 unsigned long, addr, unsigned long, flags)
1472 int uninitialized_var(pval);
1475 if (nmask != NULL && maxnode < MAX_NUMNODES)
1478 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1483 if (policy && put_user(pval, policy))
1487 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1492 #ifdef CONFIG_COMPAT
1494 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1495 compat_ulong_t __user *, nmask,
1496 compat_ulong_t, maxnode,
1497 compat_ulong_t, addr, compat_ulong_t, flags)
1500 unsigned long __user *nm = NULL;
1501 unsigned long nr_bits, alloc_size;
1502 DECLARE_BITMAP(bm, MAX_NUMNODES);
1504 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1505 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1508 nm = compat_alloc_user_space(alloc_size);
1510 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1512 if (!err && nmask) {
1513 unsigned long copy_size;
1514 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1515 err = copy_from_user(bm, nm, copy_size);
1516 /* ensure entire bitmap is zeroed */
1517 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1518 err |= compat_put_bitmap(nmask, bm, nr_bits);
1524 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1525 compat_ulong_t, maxnode)
1528 unsigned long __user *nm = NULL;
1529 unsigned long nr_bits, alloc_size;
1530 DECLARE_BITMAP(bm, MAX_NUMNODES);
1532 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1533 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1536 err = compat_get_bitmap(bm, nmask, nr_bits);
1537 nm = compat_alloc_user_space(alloc_size);
1538 err |= copy_to_user(nm, bm, alloc_size);
1544 return sys_set_mempolicy(mode, nm, nr_bits+1);
1547 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1548 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1549 compat_ulong_t, maxnode, compat_ulong_t, flags)
1552 unsigned long __user *nm = NULL;
1553 unsigned long nr_bits, alloc_size;
1556 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1557 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1560 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1561 nm = compat_alloc_user_space(alloc_size);
1562 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1568 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1573 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1576 struct mempolicy *pol = NULL;
1579 if (vma->vm_ops && vma->vm_ops->get_policy) {
1580 pol = vma->vm_ops->get_policy(vma, addr);
1581 } else if (vma->vm_policy) {
1582 pol = vma->vm_policy;
1585 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1586 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1587 * count on these policies which will be dropped by
1588 * mpol_cond_put() later
1590 if (mpol_needs_cond_ref(pol))
1599 * get_vma_policy(@vma, @addr)
1600 * @vma: virtual memory area whose policy is sought
1601 * @addr: address in @vma for shared policy lookup
1603 * Returns effective policy for a VMA at specified address.
1604 * Falls back to current->mempolicy or system default policy, as necessary.
1605 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1606 * count--added by the get_policy() vm_op, as appropriate--to protect against
1607 * freeing by another task. It is the caller's responsibility to free the
1608 * extra reference for shared policies.
1610 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1613 struct mempolicy *pol = __get_vma_policy(vma, addr);
1616 pol = get_task_policy(current);
1621 bool vma_policy_mof(struct vm_area_struct *vma)
1623 struct mempolicy *pol;
1625 if (vma->vm_ops && vma->vm_ops->get_policy) {
1628 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1629 if (pol && (pol->flags & MPOL_F_MOF))
1636 pol = vma->vm_policy;
1638 pol = get_task_policy(current);
1640 return pol->flags & MPOL_F_MOF;
1643 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1645 enum zone_type dynamic_policy_zone = policy_zone;
1647 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1650 * if policy->v.nodes has movable memory only,
1651 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1653 * policy->v.nodes is intersect with node_states[N_MEMORY].
1654 * so if the following test faile, it implies
1655 * policy->v.nodes has movable memory only.
1657 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1658 dynamic_policy_zone = ZONE_MOVABLE;
1660 return zone >= dynamic_policy_zone;
1664 * Return a nodemask representing a mempolicy for filtering nodes for
1667 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1669 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1670 if (unlikely(policy->mode == MPOL_BIND) &&
1671 apply_policy_zone(policy, gfp_zone(gfp)) &&
1672 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1673 return &policy->v.nodes;
1678 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1679 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1682 switch (policy->mode) {
1683 case MPOL_PREFERRED:
1684 if (!(policy->flags & MPOL_F_LOCAL))
1685 nd = policy->v.preferred_node;
1689 * Normally, MPOL_BIND allocations are node-local within the
1690 * allowed nodemask. However, if __GFP_THISNODE is set and the
1691 * current node isn't part of the mask, we use the zonelist for
1692 * the first node in the mask instead.
1694 if (unlikely(gfp & __GFP_THISNODE) &&
1695 unlikely(!node_isset(nd, policy->v.nodes)))
1696 nd = first_node(policy->v.nodes);
1701 return node_zonelist(nd, gfp);
1704 /* Do dynamic interleaving for a process */
1705 static unsigned interleave_nodes(struct mempolicy *policy)
1708 struct task_struct *me = current;
1711 next = next_node(nid, policy->v.nodes);
1712 if (next >= MAX_NUMNODES)
1713 next = first_node(policy->v.nodes);
1714 if (next < MAX_NUMNODES)
1720 * Depending on the memory policy provide a node from which to allocate the
1723 unsigned int mempolicy_slab_node(void)
1725 struct mempolicy *policy;
1726 int node = numa_mem_id();
1731 policy = current->mempolicy;
1732 if (!policy || policy->flags & MPOL_F_LOCAL)
1735 switch (policy->mode) {
1736 case MPOL_PREFERRED:
1738 * handled MPOL_F_LOCAL above
1740 return policy->v.preferred_node;
1742 case MPOL_INTERLEAVE:
1743 return interleave_nodes(policy);
1747 * Follow bind policy behavior and start allocation at the
1750 struct zonelist *zonelist;
1752 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1753 zonelist = &NODE_DATA(node)->node_zonelists[0];
1754 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1757 return zone ? zone->node : node;
1765 /* Do static interleaving for a VMA with known offset. */
1766 static unsigned offset_il_node(struct mempolicy *pol,
1767 struct vm_area_struct *vma, unsigned long off)
1769 unsigned nnodes = nodes_weight(pol->v.nodes);
1772 int nid = NUMA_NO_NODE;
1775 return numa_node_id();
1776 target = (unsigned int)off % nnodes;
1779 nid = next_node(nid, pol->v.nodes);
1781 } while (c <= target);
1785 /* Determine a node number for interleave */
1786 static inline unsigned interleave_nid(struct mempolicy *pol,
1787 struct vm_area_struct *vma, unsigned long addr, int shift)
1793 * for small pages, there is no difference between
1794 * shift and PAGE_SHIFT, so the bit-shift is safe.
1795 * for huge pages, since vm_pgoff is in units of small
1796 * pages, we need to shift off the always 0 bits to get
1799 BUG_ON(shift < PAGE_SHIFT);
1800 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1801 off += (addr - vma->vm_start) >> shift;
1802 return offset_il_node(pol, vma, off);
1804 return interleave_nodes(pol);
1808 * Return the bit number of a random bit set in the nodemask.
1809 * (returns NUMA_NO_NODE if nodemask is empty)
1811 int node_random(const nodemask_t *maskp)
1813 int w, bit = NUMA_NO_NODE;
1815 w = nodes_weight(*maskp);
1817 bit = bitmap_ord_to_pos(maskp->bits,
1818 get_random_int() % w, MAX_NUMNODES);
1822 #ifdef CONFIG_HUGETLBFS
1824 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1825 * @vma: virtual memory area whose policy is sought
1826 * @addr: address in @vma for shared policy lookup and interleave policy
1827 * @gfp_flags: for requested zone
1828 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1829 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1831 * Returns a zonelist suitable for a huge page allocation and a pointer
1832 * to the struct mempolicy for conditional unref after allocation.
1833 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1834 * @nodemask for filtering the zonelist.
1836 * Must be protected by read_mems_allowed_begin()
1838 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1839 gfp_t gfp_flags, struct mempolicy **mpol,
1840 nodemask_t **nodemask)
1842 struct zonelist *zl;
1844 *mpol = get_vma_policy(vma, addr);
1845 *nodemask = NULL; /* assume !MPOL_BIND */
1847 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1848 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1849 huge_page_shift(hstate_vma(vma))), gfp_flags);
1851 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1852 if ((*mpol)->mode == MPOL_BIND)
1853 *nodemask = &(*mpol)->v.nodes;
1859 * init_nodemask_of_mempolicy
1861 * If the current task's mempolicy is "default" [NULL], return 'false'
1862 * to indicate default policy. Otherwise, extract the policy nodemask
1863 * for 'bind' or 'interleave' policy into the argument nodemask, or
1864 * initialize the argument nodemask to contain the single node for
1865 * 'preferred' or 'local' policy and return 'true' to indicate presence
1866 * of non-default mempolicy.
1868 * We don't bother with reference counting the mempolicy [mpol_get/put]
1869 * because the current task is examining it's own mempolicy and a task's
1870 * mempolicy is only ever changed by the task itself.
1872 * N.B., it is the caller's responsibility to free a returned nodemask.
1874 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1876 struct mempolicy *mempolicy;
1879 if (!(mask && current->mempolicy))
1883 mempolicy = current->mempolicy;
1884 switch (mempolicy->mode) {
1885 case MPOL_PREFERRED:
1886 if (mempolicy->flags & MPOL_F_LOCAL)
1887 nid = numa_node_id();
1889 nid = mempolicy->v.preferred_node;
1890 init_nodemask_of_node(mask, nid);
1895 case MPOL_INTERLEAVE:
1896 *mask = mempolicy->v.nodes;
1902 task_unlock(current);
1909 * mempolicy_nodemask_intersects
1911 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1912 * policy. Otherwise, check for intersection between mask and the policy
1913 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1914 * policy, always return true since it may allocate elsewhere on fallback.
1916 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1918 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1919 const nodemask_t *mask)
1921 struct mempolicy *mempolicy;
1927 mempolicy = tsk->mempolicy;
1931 switch (mempolicy->mode) {
1932 case MPOL_PREFERRED:
1934 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1935 * allocate from, they may fallback to other nodes when oom.
1936 * Thus, it's possible for tsk to have allocated memory from
1941 case MPOL_INTERLEAVE:
1942 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1952 /* Allocate a page in interleaved policy.
1953 Own path because it needs to do special accounting. */
1954 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1957 struct zonelist *zl;
1960 zl = node_zonelist(nid, gfp);
1961 page = __alloc_pages(gfp, order, zl);
1962 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1963 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1968 * alloc_pages_vma - Allocate a page for a VMA.
1971 * %GFP_USER user allocation.
1972 * %GFP_KERNEL kernel allocations,
1973 * %GFP_HIGHMEM highmem/user allocations,
1974 * %GFP_FS allocation should not call back into a file system.
1975 * %GFP_ATOMIC don't sleep.
1977 * @order:Order of the GFP allocation.
1978 * @vma: Pointer to VMA or NULL if not available.
1979 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1980 * @node: Which node to prefer for allocation (modulo policy).
1981 * @hugepage: for hugepages try only the preferred node if possible
1983 * This function allocates a page from the kernel page pool and applies
1984 * a NUMA policy associated with the VMA or the current process.
1985 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1986 * mm_struct of the VMA to prevent it from going away. Should be used for
1987 * all allocations for pages that will be mapped into user space. Returns
1988 * NULL when no page can be allocated.
1991 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1992 unsigned long addr, int node, bool hugepage)
1994 struct mempolicy *pol;
1996 unsigned int cpuset_mems_cookie;
1997 struct zonelist *zl;
2001 pol = get_vma_policy(vma, addr);
2002 cpuset_mems_cookie = read_mems_allowed_begin();
2004 if (pol->mode == MPOL_INTERLEAVE) {
2007 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2009 page = alloc_page_interleave(gfp, order, nid);
2013 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2014 int hpage_node = node;
2017 * For hugepage allocation and non-interleave policy which
2018 * allows the current node (or other explicitly preferred
2019 * node) we only try to allocate from the current/preferred
2020 * node and don't fall back to other nodes, as the cost of
2021 * remote accesses would likely offset THP benefits.
2023 * If the policy is interleave, or does not allow the current
2024 * node in its nodemask, we allocate the standard way.
2026 if (pol->mode == MPOL_PREFERRED &&
2027 !(pol->flags & MPOL_F_LOCAL))
2028 hpage_node = pol->v.preferred_node;
2030 nmask = policy_nodemask(gfp, pol);
2031 if (!nmask || node_isset(hpage_node, *nmask)) {
2033 page = __alloc_pages_node(hpage_node,
2034 gfp | __GFP_THISNODE, order);
2039 nmask = policy_nodemask(gfp, pol);
2040 zl = policy_zonelist(gfp, pol, node);
2042 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2044 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2050 * alloc_pages_current - Allocate pages.
2053 * %GFP_USER user allocation,
2054 * %GFP_KERNEL kernel allocation,
2055 * %GFP_HIGHMEM highmem allocation,
2056 * %GFP_FS don't call back into a file system.
2057 * %GFP_ATOMIC don't sleep.
2058 * @order: Power of two of allocation size in pages. 0 is a single page.
2060 * Allocate a page from the kernel page pool. When not in
2061 * interrupt context and apply the current process NUMA policy.
2062 * Returns NULL when no page can be allocated.
2064 * Don't call cpuset_update_task_memory_state() unless
2065 * 1) it's ok to take cpuset_sem (can WAIT), and
2066 * 2) allocating for current task (not interrupt).
2068 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2070 struct mempolicy *pol = &default_policy;
2072 unsigned int cpuset_mems_cookie;
2074 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2075 pol = get_task_policy(current);
2078 cpuset_mems_cookie = read_mems_allowed_begin();
2081 * No reference counting needed for current->mempolicy
2082 * nor system default_policy
2084 if (pol->mode == MPOL_INTERLEAVE)
2085 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2087 page = __alloc_pages_nodemask(gfp, order,
2088 policy_zonelist(gfp, pol, numa_node_id()),
2089 policy_nodemask(gfp, pol));
2091 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2096 EXPORT_SYMBOL(alloc_pages_current);
2098 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2100 struct mempolicy *pol = mpol_dup(vma_policy(src));
2103 return PTR_ERR(pol);
2104 dst->vm_policy = pol;
2109 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2110 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2111 * with the mems_allowed returned by cpuset_mems_allowed(). This
2112 * keeps mempolicies cpuset relative after its cpuset moves. See
2113 * further kernel/cpuset.c update_nodemask().
2115 * current's mempolicy may be rebinded by the other task(the task that changes
2116 * cpuset's mems), so we needn't do rebind work for current task.
2119 /* Slow path of a mempolicy duplicate */
2120 struct mempolicy *__mpol_dup(struct mempolicy *old)
2122 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2125 return ERR_PTR(-ENOMEM);
2127 /* task's mempolicy is protected by alloc_lock */
2128 if (old == current->mempolicy) {
2131 task_unlock(current);
2135 if (current_cpuset_is_being_rebound()) {
2136 nodemask_t mems = cpuset_mems_allowed(current);
2137 if (new->flags & MPOL_F_REBINDING)
2138 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2140 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2142 atomic_set(&new->refcnt, 1);
2146 /* Slow path of a mempolicy comparison */
2147 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2151 if (a->mode != b->mode)
2153 if (a->flags != b->flags)
2155 if (mpol_store_user_nodemask(a))
2156 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2162 case MPOL_INTERLEAVE:
2163 return !!nodes_equal(a->v.nodes, b->v.nodes);
2164 case MPOL_PREFERRED:
2165 return a->v.preferred_node == b->v.preferred_node;
2173 * Shared memory backing store policy support.
2175 * Remember policies even when nobody has shared memory mapped.
2176 * The policies are kept in Red-Black tree linked from the inode.
2177 * They are protected by the sp->lock spinlock, which should be held
2178 * for any accesses to the tree.
2181 /* lookup first element intersecting start-end */
2182 /* Caller holds sp->lock */
2183 static struct sp_node *
2184 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2186 struct rb_node *n = sp->root.rb_node;
2189 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2191 if (start >= p->end)
2193 else if (end <= p->start)
2201 struct sp_node *w = NULL;
2202 struct rb_node *prev = rb_prev(n);
2205 w = rb_entry(prev, struct sp_node, nd);
2206 if (w->end <= start)
2210 return rb_entry(n, struct sp_node, nd);
2213 /* Insert a new shared policy into the list. */
2214 /* Caller holds sp->lock */
2215 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2217 struct rb_node **p = &sp->root.rb_node;
2218 struct rb_node *parent = NULL;
2223 nd = rb_entry(parent, struct sp_node, nd);
2224 if (new->start < nd->start)
2226 else if (new->end > nd->end)
2227 p = &(*p)->rb_right;
2231 rb_link_node(&new->nd, parent, p);
2232 rb_insert_color(&new->nd, &sp->root);
2233 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2234 new->policy ? new->policy->mode : 0);
2237 /* Find shared policy intersecting idx */
2239 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2241 struct mempolicy *pol = NULL;
2244 if (!sp->root.rb_node)
2246 spin_lock(&sp->lock);
2247 sn = sp_lookup(sp, idx, idx+1);
2249 mpol_get(sn->policy);
2252 spin_unlock(&sp->lock);
2256 static void sp_free(struct sp_node *n)
2258 mpol_put(n->policy);
2259 kmem_cache_free(sn_cache, n);
2263 * mpol_misplaced - check whether current page node is valid in policy
2265 * @page: page to be checked
2266 * @vma: vm area where page mapped
2267 * @addr: virtual address where page mapped
2269 * Lookup current policy node id for vma,addr and "compare to" page's
2273 * -1 - not misplaced, page is in the right node
2274 * node - node id where the page should be
2276 * Policy determination "mimics" alloc_page_vma().
2277 * Called from fault path where we know the vma and faulting address.
2279 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2281 struct mempolicy *pol;
2283 int curnid = page_to_nid(page);
2284 unsigned long pgoff;
2285 int thiscpu = raw_smp_processor_id();
2286 int thisnid = cpu_to_node(thiscpu);
2292 pol = get_vma_policy(vma, addr);
2293 if (!(pol->flags & MPOL_F_MOF))
2296 switch (pol->mode) {
2297 case MPOL_INTERLEAVE:
2298 BUG_ON(addr >= vma->vm_end);
2299 BUG_ON(addr < vma->vm_start);
2301 pgoff = vma->vm_pgoff;
2302 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2303 polnid = offset_il_node(pol, vma, pgoff);
2306 case MPOL_PREFERRED:
2307 if (pol->flags & MPOL_F_LOCAL)
2308 polnid = numa_node_id();
2310 polnid = pol->v.preferred_node;
2315 * allows binding to multiple nodes.
2316 * use current page if in policy nodemask,
2317 * else select nearest allowed node, if any.
2318 * If no allowed nodes, use current [!misplaced].
2320 if (node_isset(curnid, pol->v.nodes))
2322 (void)first_zones_zonelist(
2323 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2324 gfp_zone(GFP_HIGHUSER),
2325 &pol->v.nodes, &zone);
2326 polnid = zone->node;
2333 /* Migrate the page towards the node whose CPU is referencing it */
2334 if (pol->flags & MPOL_F_MORON) {
2337 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2341 if (curnid != polnid)
2349 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2351 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2352 rb_erase(&n->nd, &sp->root);
2356 static void sp_node_init(struct sp_node *node, unsigned long start,
2357 unsigned long end, struct mempolicy *pol)
2359 node->start = start;
2364 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2365 struct mempolicy *pol)
2368 struct mempolicy *newpol;
2370 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2374 newpol = mpol_dup(pol);
2375 if (IS_ERR(newpol)) {
2376 kmem_cache_free(sn_cache, n);
2379 newpol->flags |= MPOL_F_SHARED;
2380 sp_node_init(n, start, end, newpol);
2385 /* Replace a policy range. */
2386 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2387 unsigned long end, struct sp_node *new)
2390 struct sp_node *n_new = NULL;
2391 struct mempolicy *mpol_new = NULL;
2395 spin_lock(&sp->lock);
2396 n = sp_lookup(sp, start, end);
2397 /* Take care of old policies in the same range. */
2398 while (n && n->start < end) {
2399 struct rb_node *next = rb_next(&n->nd);
2400 if (n->start >= start) {
2406 /* Old policy spanning whole new range. */
2411 *mpol_new = *n->policy;
2412 atomic_set(&mpol_new->refcnt, 1);
2413 sp_node_init(n_new, end, n->end, mpol_new);
2415 sp_insert(sp, n_new);
2424 n = rb_entry(next, struct sp_node, nd);
2428 spin_unlock(&sp->lock);
2435 kmem_cache_free(sn_cache, n_new);
2440 spin_unlock(&sp->lock);
2442 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2445 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2452 * mpol_shared_policy_init - initialize shared policy for inode
2453 * @sp: pointer to inode shared policy
2454 * @mpol: struct mempolicy to install
2456 * Install non-NULL @mpol in inode's shared policy rb-tree.
2457 * On entry, the current task has a reference on a non-NULL @mpol.
2458 * This must be released on exit.
2459 * This is called at get_inode() calls and we can use GFP_KERNEL.
2461 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2465 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2466 spin_lock_init(&sp->lock);
2469 struct vm_area_struct pvma;
2470 struct mempolicy *new;
2471 NODEMASK_SCRATCH(scratch);
2475 /* contextualize the tmpfs mount point mempolicy */
2476 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2478 goto free_scratch; /* no valid nodemask intersection */
2481 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2482 task_unlock(current);
2486 /* Create pseudo-vma that contains just the policy */
2487 memset(&pvma, 0, sizeof(struct vm_area_struct));
2488 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2489 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2492 mpol_put(new); /* drop initial ref */
2494 NODEMASK_SCRATCH_FREE(scratch);
2496 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2500 int mpol_set_shared_policy(struct shared_policy *info,
2501 struct vm_area_struct *vma, struct mempolicy *npol)
2504 struct sp_node *new = NULL;
2505 unsigned long sz = vma_pages(vma);
2507 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2509 sz, npol ? npol->mode : -1,
2510 npol ? npol->flags : -1,
2511 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2514 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2518 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2524 /* Free a backing policy store on inode delete. */
2525 void mpol_free_shared_policy(struct shared_policy *p)
2528 struct rb_node *next;
2530 if (!p->root.rb_node)
2532 spin_lock(&p->lock);
2533 next = rb_first(&p->root);
2535 n = rb_entry(next, struct sp_node, nd);
2536 next = rb_next(&n->nd);
2539 spin_unlock(&p->lock);
2542 #ifdef CONFIG_NUMA_BALANCING
2543 static int __initdata numabalancing_override;
2545 static void __init check_numabalancing_enable(void)
2547 bool numabalancing_default = false;
2549 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2550 numabalancing_default = true;
2552 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2553 if (numabalancing_override)
2554 set_numabalancing_state(numabalancing_override == 1);
2556 if (num_online_nodes() > 1 && !numabalancing_override) {
2557 pr_info("%s automatic NUMA balancing. "
2558 "Configure with numa_balancing= or the "
2559 "kernel.numa_balancing sysctl",
2560 numabalancing_default ? "Enabling" : "Disabling");
2561 set_numabalancing_state(numabalancing_default);
2565 static int __init setup_numabalancing(char *str)
2571 if (!strcmp(str, "enable")) {
2572 numabalancing_override = 1;
2574 } else if (!strcmp(str, "disable")) {
2575 numabalancing_override = -1;
2580 pr_warn("Unable to parse numa_balancing=\n");
2584 __setup("numa_balancing=", setup_numabalancing);
2586 static inline void __init check_numabalancing_enable(void)
2589 #endif /* CONFIG_NUMA_BALANCING */
2591 /* assumes fs == KERNEL_DS */
2592 void __init numa_policy_init(void)
2594 nodemask_t interleave_nodes;
2595 unsigned long largest = 0;
2596 int nid, prefer = 0;
2598 policy_cache = kmem_cache_create("numa_policy",
2599 sizeof(struct mempolicy),
2600 0, SLAB_PANIC, NULL);
2602 sn_cache = kmem_cache_create("shared_policy_node",
2603 sizeof(struct sp_node),
2604 0, SLAB_PANIC, NULL);
2606 for_each_node(nid) {
2607 preferred_node_policy[nid] = (struct mempolicy) {
2608 .refcnt = ATOMIC_INIT(1),
2609 .mode = MPOL_PREFERRED,
2610 .flags = MPOL_F_MOF | MPOL_F_MORON,
2611 .v = { .preferred_node = nid, },
2616 * Set interleaving policy for system init. Interleaving is only
2617 * enabled across suitably sized nodes (default is >= 16MB), or
2618 * fall back to the largest node if they're all smaller.
2620 nodes_clear(interleave_nodes);
2621 for_each_node_state(nid, N_MEMORY) {
2622 unsigned long total_pages = node_present_pages(nid);
2624 /* Preserve the largest node */
2625 if (largest < total_pages) {
2626 largest = total_pages;
2630 /* Interleave this node? */
2631 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2632 node_set(nid, interleave_nodes);
2635 /* All too small, use the largest */
2636 if (unlikely(nodes_empty(interleave_nodes)))
2637 node_set(prefer, interleave_nodes);
2639 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2640 pr_err("%s: interleaving failed\n", __func__);
2642 check_numabalancing_enable();
2645 /* Reset policy of current process to default */
2646 void numa_default_policy(void)
2648 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2652 * Parse and format mempolicy from/to strings
2656 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2658 static const char * const policy_modes[] =
2660 [MPOL_DEFAULT] = "default",
2661 [MPOL_PREFERRED] = "prefer",
2662 [MPOL_BIND] = "bind",
2663 [MPOL_INTERLEAVE] = "interleave",
2664 [MPOL_LOCAL] = "local",
2670 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2671 * @str: string containing mempolicy to parse
2672 * @mpol: pointer to struct mempolicy pointer, returned on success.
2675 * <mode>[=<flags>][:<nodelist>]
2677 * On success, returns 0, else 1
2679 int mpol_parse_str(char *str, struct mempolicy **mpol)
2681 struct mempolicy *new = NULL;
2682 unsigned short mode;
2683 unsigned short mode_flags;
2685 char *nodelist = strchr(str, ':');
2686 char *flags = strchr(str, '=');
2690 /* NUL-terminate mode or flags string */
2692 if (nodelist_parse(nodelist, nodes))
2694 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2700 *flags++ = '\0'; /* terminate mode string */
2702 for (mode = 0; mode < MPOL_MAX; mode++) {
2703 if (!strcmp(str, policy_modes[mode])) {
2707 if (mode >= MPOL_MAX)
2711 case MPOL_PREFERRED:
2713 * Insist on a nodelist of one node only
2716 char *rest = nodelist;
2717 while (isdigit(*rest))
2723 case MPOL_INTERLEAVE:
2725 * Default to online nodes with memory if no nodelist
2728 nodes = node_states[N_MEMORY];
2732 * Don't allow a nodelist; mpol_new() checks flags
2736 mode = MPOL_PREFERRED;
2740 * Insist on a empty nodelist
2747 * Insist on a nodelist
2756 * Currently, we only support two mutually exclusive
2759 if (!strcmp(flags, "static"))
2760 mode_flags |= MPOL_F_STATIC_NODES;
2761 else if (!strcmp(flags, "relative"))
2762 mode_flags |= MPOL_F_RELATIVE_NODES;
2767 new = mpol_new(mode, mode_flags, &nodes);
2772 * Save nodes for mpol_to_str() to show the tmpfs mount options
2773 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2775 if (mode != MPOL_PREFERRED)
2776 new->v.nodes = nodes;
2778 new->v.preferred_node = first_node(nodes);
2780 new->flags |= MPOL_F_LOCAL;
2783 * Save nodes for contextualization: this will be used to "clone"
2784 * the mempolicy in a specific context [cpuset] at a later time.
2786 new->w.user_nodemask = nodes;
2791 /* Restore string for error message */
2800 #endif /* CONFIG_TMPFS */
2803 * mpol_to_str - format a mempolicy structure for printing
2804 * @buffer: to contain formatted mempolicy string
2805 * @maxlen: length of @buffer
2806 * @pol: pointer to mempolicy to be formatted
2808 * Convert @pol into a string. If @buffer is too short, truncate the string.
2809 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2810 * longest flag, "relative", and to display at least a few node ids.
2812 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2815 nodemask_t nodes = NODE_MASK_NONE;
2816 unsigned short mode = MPOL_DEFAULT;
2817 unsigned short flags = 0;
2819 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2827 case MPOL_PREFERRED:
2828 if (flags & MPOL_F_LOCAL)
2831 node_set(pol->v.preferred_node, nodes);
2834 case MPOL_INTERLEAVE:
2835 nodes = pol->v.nodes;
2839 snprintf(p, maxlen, "unknown");
2843 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2845 if (flags & MPOL_MODE_FLAGS) {
2846 p += snprintf(p, buffer + maxlen - p, "=");
2849 * Currently, the only defined flags are mutually exclusive
2851 if (flags & MPOL_F_STATIC_NODES)
2852 p += snprintf(p, buffer + maxlen - p, "static");
2853 else if (flags & MPOL_F_RELATIVE_NODES)
2854 p += snprintf(p, buffer + maxlen - p, "relative");
2857 if (!nodes_empty(nodes))
2858 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2859 nodemask_pr_args(&nodes));