1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple NUMA memory policy for the Linux kernel.
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
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 * preferred many Try a set of nodes first before normal fallback. This is
35 * similar to preferred without the special case.
37 * default Allocate on the local node first, or when on a VMA
38 * use the process policy. This is what Linux always did
39 * in a NUMA aware kernel and still does by, ahem, default.
41 * The process policy is applied for most non interrupt memory allocations
42 * in that process' context. Interrupts ignore the policies and always
43 * try to allocate on the local CPU. The VMA policy is only applied for memory
44 * allocations for a VMA in the VM.
46 * Currently there are a few corner cases in swapping where the policy
47 * is not applied, but the majority should be handled. When process policy
48 * is used it is not remembered over swap outs/swap ins.
50 * Only the highest zone in the zone hierarchy gets policied. Allocations
51 * requesting a lower zone just use default policy. This implies that
52 * on systems with highmem kernel lowmem allocation don't get policied.
53 * Same with GFP_DMA allocations.
55 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
56 * all users and remembered even when nobody has memory mapped.
60 fix mmap readahead to honour policy and enable policy for any page cache
62 statistics for bigpages
63 global policy for page cache? currently it uses process policy. Requires
65 handle mremap for shared memory (currently ignored for the policy)
67 make bind policy root only? It can trigger oom much faster and the
68 kernel is not always grateful with that.
71 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
73 #include <linux/mempolicy.h>
74 #include <linux/pagewalk.h>
75 #include <linux/highmem.h>
76 #include <linux/hugetlb.h>
77 #include <linux/kernel.h>
78 #include <linux/sched.h>
79 #include <linux/sched/mm.h>
80 #include <linux/sched/numa_balancing.h>
81 #include <linux/sched/task.h>
82 #include <linux/nodemask.h>
83 #include <linux/cpuset.h>
84 #include <linux/slab.h>
85 #include <linux/string.h>
86 #include <linux/export.h>
87 #include <linux/nsproxy.h>
88 #include <linux/interrupt.h>
89 #include <linux/init.h>
90 #include <linux/compat.h>
91 #include <linux/ptrace.h>
92 #include <linux/swap.h>
93 #include <linux/seq_file.h>
94 #include <linux/proc_fs.h>
95 #include <linux/migrate.h>
96 #include <linux/ksm.h>
97 #include <linux/rmap.h>
98 #include <linux/security.h>
99 #include <linux/syscalls.h>
100 #include <linux/ctype.h>
101 #include <linux/mm_inline.h>
102 #include <linux/mmu_notifier.h>
103 #include <linux/printk.h>
104 #include <linux/swapops.h>
106 #include <asm/tlbflush.h>
107 #include <linux/uaccess.h>
109 #include "internal.h"
112 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
113 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
115 static struct kmem_cache *policy_cache;
116 static struct kmem_cache *sn_cache;
118 /* Highest zone. An specific allocation for a zone below that is not
120 enum zone_type policy_zone = 0;
123 * run-time system-wide default policy => local allocation
125 static struct mempolicy default_policy = {
126 .refcnt = ATOMIC_INIT(1), /* never free it */
130 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
133 * numa_map_to_online_node - Find closest online node
134 * @node: Node id to start the search
136 * Lookup the next closest node by distance if @nid is not online.
138 int numa_map_to_online_node(int node)
140 int min_dist = INT_MAX, dist, n, min_node;
142 if (node == NUMA_NO_NODE || node_online(node))
146 for_each_online_node(n) {
147 dist = node_distance(node, n);
148 if (dist < min_dist) {
156 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
158 struct mempolicy *get_task_policy(struct task_struct *p)
160 struct mempolicy *pol = p->mempolicy;
166 node = numa_node_id();
167 if (node != NUMA_NO_NODE) {
168 pol = &preferred_node_policy[node];
169 /* preferred_node_policy is not initialised early in boot */
174 return &default_policy;
177 static const struct mempolicy_operations {
178 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
179 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
180 } mpol_ops[MPOL_MAX];
182 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
184 return pol->flags & MPOL_MODE_FLAGS;
187 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
188 const nodemask_t *rel)
191 nodes_fold(tmp, *orig, nodes_weight(*rel));
192 nodes_onto(*ret, tmp, *rel);
195 static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
197 if (nodes_empty(*nodes))
203 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
205 if (nodes_empty(*nodes))
208 nodes_clear(pol->nodes);
209 node_set(first_node(*nodes), pol->nodes);
214 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
215 * any, for the new policy. mpol_new() has already validated the nodes
216 * parameter with respect to the policy mode and flags.
218 * Must be called holding task's alloc_lock to protect task's mems_allowed
219 * and mempolicy. May also be called holding the mmap_lock for write.
221 static int mpol_set_nodemask(struct mempolicy *pol,
222 const nodemask_t *nodes, struct nodemask_scratch *nsc)
227 * Default (pol==NULL) resp. local memory policies are not a
228 * subject of any remapping. They also do not need any special
231 if (!pol || pol->mode == MPOL_LOCAL)
235 nodes_and(nsc->mask1,
236 cpuset_current_mems_allowed, node_states[N_MEMORY]);
240 if (pol->flags & MPOL_F_RELATIVE_NODES)
241 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
243 nodes_and(nsc->mask2, *nodes, nsc->mask1);
245 if (mpol_store_user_nodemask(pol))
246 pol->w.user_nodemask = *nodes;
248 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
250 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
286 } else if (mode == MPOL_LOCAL) {
287 if (!nodes_empty(*nodes) ||
288 (flags & MPOL_F_STATIC_NODES) ||
289 (flags & MPOL_F_RELATIVE_NODES))
290 return ERR_PTR(-EINVAL);
291 } else if (nodes_empty(*nodes))
292 return ERR_PTR(-EINVAL);
293 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
295 return ERR_PTR(-ENOMEM);
296 atomic_set(&policy->refcnt, 1);
298 policy->flags = flags;
303 /* Slow path of a mpol destructor. */
304 void __mpol_put(struct mempolicy *p)
306 if (!atomic_dec_and_test(&p->refcnt))
308 kmem_cache_free(policy_cache, p);
311 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
315 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
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);
324 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
326 pol->w.cpuset_mems_allowed = *nodes;
329 if (nodes_empty(tmp))
335 static void mpol_rebind_preferred(struct mempolicy *pol,
336 const nodemask_t *nodes)
338 pol->w.cpuset_mems_allowed = *nodes;
342 * mpol_rebind_policy - Migrate a policy to a different set of nodes
344 * Per-vma policies are protected by mmap_lock. Allocations using per-task
345 * policies are protected by task->mems_allowed_seq to prevent a premature
346 * OOM/allocation failure due to parallel nodemask modification.
348 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
350 if (!pol || pol->mode == MPOL_LOCAL)
352 if (!mpol_store_user_nodemask(pol) &&
353 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
356 mpol_ops[pol->mode].rebind(pol, newmask);
360 * Wrapper for mpol_rebind_policy() that just requires task
361 * pointer, and updates task mempolicy.
363 * Called with task's alloc_lock held.
366 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
368 mpol_rebind_policy(tsk->mempolicy, new);
372 * Rebind each vma in mm to new nodemask.
374 * Call holding a reference to mm. Takes mm->mmap_lock during call.
377 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
379 struct vm_area_struct *vma;
382 for (vma = mm->mmap; vma; vma = vma->vm_next)
383 mpol_rebind_policy(vma->vm_policy, new);
384 mmap_write_unlock(mm);
387 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
389 .rebind = mpol_rebind_default,
391 [MPOL_INTERLEAVE] = {
392 .create = mpol_new_nodemask,
393 .rebind = mpol_rebind_nodemask,
396 .create = mpol_new_preferred,
397 .rebind = mpol_rebind_preferred,
400 .create = mpol_new_nodemask,
401 .rebind = mpol_rebind_nodemask,
404 .rebind = mpol_rebind_default,
406 [MPOL_PREFERRED_MANY] = {
407 .create = mpol_new_nodemask,
408 .rebind = mpol_rebind_preferred,
412 static int migrate_page_add(struct page *page, struct list_head *pagelist,
413 unsigned long flags);
416 struct list_head *pagelist;
421 struct vm_area_struct *first;
425 * Check if the page's nid is in qp->nmask.
427 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
428 * in the invert of qp->nmask.
430 static inline bool queue_pages_required(struct page *page,
431 struct queue_pages *qp)
433 int nid = page_to_nid(page);
434 unsigned long flags = qp->flags;
436 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
440 * queue_pages_pmd() has four possible return values:
441 * 0 - pages are placed on the right node or queued successfully, or
442 * special page is met, i.e. huge zero page.
443 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
446 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
447 * existing page was already on a node that does not follow the
450 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
451 unsigned long end, struct mm_walk *walk)
456 struct queue_pages *qp = walk->private;
459 if (unlikely(is_pmd_migration_entry(*pmd))) {
463 page = pmd_page(*pmd);
464 if (is_huge_zero_page(page)) {
466 walk->action = ACTION_CONTINUE;
469 if (!queue_pages_required(page, qp))
473 /* go to thp migration */
474 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
475 if (!vma_migratable(walk->vma) ||
476 migrate_page_add(page, qp->pagelist, flags)) {
489 * Scan through pages checking if pages follow certain conditions,
490 * and move them to the pagelist if they do.
492 * queue_pages_pte_range() has three possible return values:
493 * 0 - pages are placed on the right node or queued successfully, or
494 * special page is met, i.e. zero page.
495 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
497 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
498 * on a node that does not follow the policy.
500 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
501 unsigned long end, struct mm_walk *walk)
503 struct vm_area_struct *vma = walk->vma;
505 struct queue_pages *qp = walk->private;
506 unsigned long flags = qp->flags;
508 bool has_unmovable = false;
509 pte_t *pte, *mapped_pte;
512 ptl = pmd_trans_huge_lock(pmd, vma);
514 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
518 /* THP was split, fall through to pte walk */
520 if (pmd_trans_unstable(pmd))
523 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
524 for (; addr != end; pte++, addr += PAGE_SIZE) {
525 if (!pte_present(*pte))
527 page = vm_normal_page(vma, addr, *pte);
531 * vm_normal_page() filters out zero pages, but there might
532 * still be PageReserved pages to skip, perhaps in a VDSO.
534 if (PageReserved(page))
536 if (!queue_pages_required(page, qp))
538 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
539 /* MPOL_MF_STRICT must be specified if we get here */
540 if (!vma_migratable(vma)) {
541 has_unmovable = true;
546 * Do not abort immediately since there may be
547 * temporary off LRU pages in the range. Still
548 * need migrate other LRU pages.
550 if (migrate_page_add(page, qp->pagelist, flags))
551 has_unmovable = true;
555 pte_unmap_unlock(mapped_pte, ptl);
561 return addr != end ? -EIO : 0;
564 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
565 unsigned long addr, unsigned long end,
566 struct mm_walk *walk)
569 #ifdef CONFIG_HUGETLB_PAGE
570 struct queue_pages *qp = walk->private;
571 unsigned long flags = (qp->flags & MPOL_MF_VALID);
576 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
577 entry = huge_ptep_get(pte);
578 if (!pte_present(entry))
580 page = pte_page(entry);
581 if (!queue_pages_required(page, qp))
584 if (flags == MPOL_MF_STRICT) {
586 * STRICT alone means only detecting misplaced page and no
587 * need to further check other vma.
593 if (!vma_migratable(walk->vma)) {
595 * Must be STRICT with MOVE*, otherwise .test_walk() have
596 * stopped walking current vma.
597 * Detecting misplaced page but allow migrating pages which
604 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
605 if (flags & (MPOL_MF_MOVE_ALL) ||
606 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1 &&
607 !hugetlb_pmd_shared(pte))) {
608 if (isolate_hugetlb(page, qp->pagelist) &&
609 (flags & MPOL_MF_STRICT))
611 * Failed to isolate page but allow migrating pages
612 * which have been queued.
624 #ifdef CONFIG_NUMA_BALANCING
626 * This is used to mark a range of virtual addresses to be inaccessible.
627 * These are later cleared by a NUMA hinting fault. Depending on these
628 * faults, pages may be migrated for better NUMA placement.
630 * This is assuming that NUMA faults are handled using PROT_NONE. If
631 * an architecture makes a different choice, it will need further
632 * changes to the core.
634 unsigned long change_prot_numa(struct vm_area_struct *vma,
635 unsigned long addr, unsigned long end)
639 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
641 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
646 static unsigned long change_prot_numa(struct vm_area_struct *vma,
647 unsigned long addr, unsigned long end)
651 #endif /* CONFIG_NUMA_BALANCING */
653 static int queue_pages_test_walk(unsigned long start, unsigned long end,
654 struct mm_walk *walk)
656 struct vm_area_struct *vma = walk->vma;
657 struct queue_pages *qp = walk->private;
658 unsigned long endvma = vma->vm_end;
659 unsigned long flags = qp->flags;
661 /* range check first */
662 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
666 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
667 (qp->start < vma->vm_start))
668 /* hole at head side of range */
671 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
672 ((vma->vm_end < qp->end) &&
673 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
674 /* hole at middle or tail of range */
678 * Need check MPOL_MF_STRICT to return -EIO if possible
679 * regardless of vma_migratable
681 if (!vma_migratable(vma) &&
682 !(flags & MPOL_MF_STRICT))
688 if (flags & MPOL_MF_LAZY) {
689 /* Similar to task_numa_work, skip inaccessible VMAs */
690 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
691 !(vma->vm_flags & VM_MIXEDMAP))
692 change_prot_numa(vma, start, endvma);
696 /* queue pages from current vma */
697 if (flags & MPOL_MF_VALID)
702 static const struct mm_walk_ops queue_pages_walk_ops = {
703 .hugetlb_entry = queue_pages_hugetlb,
704 .pmd_entry = queue_pages_pte_range,
705 .test_walk = queue_pages_test_walk,
709 * Walk through page tables and collect pages to be migrated.
711 * If pages found in a given range are on a set of nodes (determined by
712 * @nodes and @flags,) it's isolated and queued to the pagelist which is
713 * passed via @private.
715 * queue_pages_range() has three possible return values:
716 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
718 * 0 - queue pages successfully or no misplaced page.
719 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
720 * memory range specified by nodemask and maxnode points outside
721 * your accessible address space (-EFAULT)
724 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
725 nodemask_t *nodes, unsigned long flags,
726 struct list_head *pagelist)
729 struct queue_pages qp = {
730 .pagelist = pagelist,
738 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
741 /* whole range in hole */
748 * Apply policy to a single VMA
749 * This must be called with the mmap_lock held for writing.
751 static int vma_replace_policy(struct vm_area_struct *vma,
752 struct mempolicy *pol)
755 struct mempolicy *old;
756 struct mempolicy *new;
758 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
759 vma->vm_start, vma->vm_end, vma->vm_pgoff,
760 vma->vm_ops, vma->vm_file,
761 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
767 if (vma->vm_ops && vma->vm_ops->set_policy) {
768 err = vma->vm_ops->set_policy(vma, new);
773 old = vma->vm_policy;
774 vma->vm_policy = new; /* protected by mmap_lock */
783 /* Step 2: apply policy to a range and do splits. */
784 static int mbind_range(struct mm_struct *mm, unsigned long start,
785 unsigned long end, struct mempolicy *new_pol)
787 struct vm_area_struct *prev;
788 struct vm_area_struct *vma;
791 unsigned long vmstart;
794 vma = find_vma(mm, start);
798 if (start > vma->vm_start)
801 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
802 vmstart = max(start, vma->vm_start);
803 vmend = min(end, vma->vm_end);
805 if (mpol_equal(vma_policy(vma), new_pol))
808 pgoff = vma->vm_pgoff +
809 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
810 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
811 vma->anon_vma, vma->vm_file, pgoff,
812 new_pol, vma->vm_userfaultfd_ctx);
817 if (vma->vm_start != vmstart) {
818 err = split_vma(vma->vm_mm, vma, vmstart, 1);
822 if (vma->vm_end != vmend) {
823 err = split_vma(vma->vm_mm, vma, vmend, 0);
828 err = vma_replace_policy(vma, new_pol);
837 /* Set the process memory policy */
838 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
841 struct mempolicy *new, *old;
842 NODEMASK_SCRATCH(scratch);
848 new = mpol_new(mode, flags, nodes);
854 ret = mpol_set_nodemask(new, nodes, scratch);
860 old = current->mempolicy;
861 current->mempolicy = new;
862 if (new && new->mode == MPOL_INTERLEAVE)
863 current->il_prev = MAX_NUMNODES-1;
864 task_unlock(current);
868 NODEMASK_SCRATCH_FREE(scratch);
873 * Return nodemask for policy for get_mempolicy() query
875 * Called with task's alloc_lock held
877 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
880 if (p == &default_policy)
885 case MPOL_INTERLEAVE:
887 case MPOL_PREFERRED_MANY:
891 /* return empty node mask for local allocation */
898 static int lookup_node(struct mm_struct *mm, unsigned long addr)
900 struct page *p = NULL;
904 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
906 err = page_to_nid(p);
910 mmap_read_unlock(mm);
914 /* Retrieve NUMA policy */
915 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
916 unsigned long addr, unsigned long flags)
919 struct mm_struct *mm = current->mm;
920 struct vm_area_struct *vma = NULL;
921 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
924 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
927 if (flags & MPOL_F_MEMS_ALLOWED) {
928 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
930 *policy = 0; /* just so it's initialized */
932 *nmask = cpuset_current_mems_allowed;
933 task_unlock(current);
937 if (flags & MPOL_F_ADDR) {
939 * Do NOT fall back to task policy if the
940 * vma/shared policy at addr is NULL. We
941 * want to return MPOL_DEFAULT in this case.
944 vma = vma_lookup(mm, addr);
946 mmap_read_unlock(mm);
949 if (vma->vm_ops && vma->vm_ops->get_policy)
950 pol = vma->vm_ops->get_policy(vma, addr);
952 pol = vma->vm_policy;
957 pol = &default_policy; /* indicates default behavior */
959 if (flags & MPOL_F_NODE) {
960 if (flags & MPOL_F_ADDR) {
962 * Take a refcount on the mpol, lookup_node()
963 * will drop the mmap_lock, so after calling
964 * lookup_node() only "pol" remains valid, "vma"
970 err = lookup_node(mm, addr);
974 } else if (pol == current->mempolicy &&
975 pol->mode == MPOL_INTERLEAVE) {
976 *policy = next_node_in(current->il_prev, pol->nodes);
982 *policy = pol == &default_policy ? MPOL_DEFAULT :
985 * Internal mempolicy flags must be masked off before exposing
986 * the policy to userspace.
988 *policy |= (pol->flags & MPOL_MODE_FLAGS);
993 if (mpol_store_user_nodemask(pol)) {
994 *nmask = pol->w.user_nodemask;
997 get_policy_nodemask(pol, nmask);
998 task_unlock(current);
1005 mmap_read_unlock(mm);
1007 mpol_put(pol_refcount);
1011 #ifdef CONFIG_MIGRATION
1013 * page migration, thp tail pages can be passed.
1015 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1016 unsigned long flags)
1018 struct page *head = compound_head(page);
1020 * Avoid migrating a page that is shared with others.
1022 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1023 if (!isolate_lru_page(head)) {
1024 list_add_tail(&head->lru, pagelist);
1025 mod_node_page_state(page_pgdat(head),
1026 NR_ISOLATED_ANON + page_is_file_lru(head),
1027 thp_nr_pages(head));
1028 } else if (flags & MPOL_MF_STRICT) {
1030 * Non-movable page may reach here. And, there may be
1031 * temporary off LRU pages or non-LRU movable pages.
1032 * Treat them as unmovable pages since they can't be
1033 * isolated, so they can't be moved at the moment. It
1034 * should return -EIO for this case too.
1044 * Migrate pages from one node to a target node.
1045 * Returns error or the number of pages not migrated.
1047 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1051 LIST_HEAD(pagelist);
1053 struct migration_target_control mtc = {
1055 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1059 node_set(source, nmask);
1062 * This does not "check" the range but isolates all pages that
1063 * need migration. Between passing in the full user address
1064 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1066 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1067 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1068 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1070 if (!list_empty(&pagelist)) {
1071 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1072 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1074 putback_movable_pages(&pagelist);
1081 * Move pages between the two nodesets so as to preserve the physical
1082 * layout as much as possible.
1084 * Returns the number of page that could not be moved.
1086 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1087 const nodemask_t *to, int flags)
1093 lru_cache_disable();
1098 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1099 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1100 * bit in 'tmp', and return that <source, dest> pair for migration.
1101 * The pair of nodemasks 'to' and 'from' define the map.
1103 * If no pair of bits is found that way, fallback to picking some
1104 * pair of 'source' and 'dest' bits that are not the same. If the
1105 * 'source' and 'dest' bits are the same, this represents a node
1106 * that will be migrating to itself, so no pages need move.
1108 * If no bits are left in 'tmp', or if all remaining bits left
1109 * in 'tmp' correspond to the same bit in 'to', return false
1110 * (nothing left to migrate).
1112 * This lets us pick a pair of nodes to migrate between, such that
1113 * if possible the dest node is not already occupied by some other
1114 * source node, minimizing the risk of overloading the memory on a
1115 * node that would happen if we migrated incoming memory to a node
1116 * before migrating outgoing memory source that same node.
1118 * A single scan of tmp is sufficient. As we go, we remember the
1119 * most recent <s, d> pair that moved (s != d). If we find a pair
1120 * that not only moved, but what's better, moved to an empty slot
1121 * (d is not set in tmp), then we break out then, with that pair.
1122 * Otherwise when we finish scanning from_tmp, we at least have the
1123 * most recent <s, d> pair that moved. If we get all the way through
1124 * the scan of tmp without finding any node that moved, much less
1125 * moved to an empty node, then there is nothing left worth migrating.
1129 while (!nodes_empty(tmp)) {
1131 int source = NUMA_NO_NODE;
1134 for_each_node_mask(s, tmp) {
1137 * do_migrate_pages() tries to maintain the relative
1138 * node relationship of the pages established between
1139 * threads and memory areas.
1141 * However if the number of source nodes is not equal to
1142 * the number of destination nodes we can not preserve
1143 * this node relative relationship. In that case, skip
1144 * copying memory from a node that is in the destination
1147 * Example: [2,3,4] -> [3,4,5] moves everything.
1148 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1151 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1152 (node_isset(s, *to)))
1155 d = node_remap(s, *from, *to);
1159 source = s; /* Node moved. Memorize */
1162 /* dest not in remaining from nodes? */
1163 if (!node_isset(dest, tmp))
1166 if (source == NUMA_NO_NODE)
1169 node_clear(source, tmp);
1170 err = migrate_to_node(mm, source, dest, flags);
1176 mmap_read_unlock(mm);
1186 * Allocate a new page for page migration based on vma policy.
1187 * Start by assuming the page is mapped by the same vma as contains @start.
1188 * Search forward from there, if not. N.B., this assumes that the
1189 * list of pages handed to migrate_pages()--which is how we get here--
1190 * is in virtual address order.
1192 static struct page *new_page(struct page *page, unsigned long start)
1194 struct vm_area_struct *vma;
1195 unsigned long address;
1197 vma = find_vma(current->mm, start);
1199 address = page_address_in_vma(page, vma);
1200 if (address != -EFAULT)
1205 if (PageHuge(page)) {
1206 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1208 } else if (PageTransHuge(page)) {
1211 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1215 prep_transhuge_page(thp);
1219 * if !vma, alloc_page_vma() will use task or system default policy
1221 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1226 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1227 unsigned long flags)
1232 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1233 const nodemask_t *to, int flags)
1238 static struct page *new_page(struct page *page, unsigned long start)
1244 static long do_mbind(unsigned long start, unsigned long len,
1245 unsigned short mode, unsigned short mode_flags,
1246 nodemask_t *nmask, unsigned long flags)
1248 struct mm_struct *mm = current->mm;
1249 struct mempolicy *new;
1253 LIST_HEAD(pagelist);
1255 if (flags & ~(unsigned long)MPOL_MF_VALID)
1257 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1260 if (start & ~PAGE_MASK)
1263 if (mode == MPOL_DEFAULT)
1264 flags &= ~MPOL_MF_STRICT;
1266 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1274 new = mpol_new(mode, mode_flags, nmask);
1276 return PTR_ERR(new);
1278 if (flags & MPOL_MF_LAZY)
1279 new->flags |= MPOL_F_MOF;
1282 * If we are using the default policy then operation
1283 * on discontinuous address spaces is okay after all
1286 flags |= MPOL_MF_DISCONTIG_OK;
1288 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1289 start, start + len, mode, mode_flags,
1290 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1292 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1294 lru_cache_disable();
1297 NODEMASK_SCRATCH(scratch);
1299 mmap_write_lock(mm);
1300 err = mpol_set_nodemask(new, nmask, scratch);
1302 mmap_write_unlock(mm);
1305 NODEMASK_SCRATCH_FREE(scratch);
1310 ret = queue_pages_range(mm, start, end, nmask,
1311 flags | MPOL_MF_INVERT, &pagelist);
1318 err = mbind_range(mm, start, end, new);
1323 if (!list_empty(&pagelist)) {
1324 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1325 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1326 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1328 putback_movable_pages(&pagelist);
1331 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1335 if (!list_empty(&pagelist))
1336 putback_movable_pages(&pagelist);
1339 mmap_write_unlock(mm);
1342 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1348 * User space interface with variable sized bitmaps for nodelists.
1350 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1351 unsigned long maxnode)
1353 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1356 if (in_compat_syscall())
1357 ret = compat_get_bitmap(mask,
1358 (const compat_ulong_t __user *)nmask,
1361 ret = copy_from_user(mask, nmask,
1362 nlongs * sizeof(unsigned long));
1367 if (maxnode % BITS_PER_LONG)
1368 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1373 /* Copy a node mask from user space. */
1374 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1375 unsigned long maxnode)
1378 nodes_clear(*nodes);
1379 if (maxnode == 0 || !nmask)
1381 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1385 * When the user specified more nodes than supported just check
1386 * if the non supported part is all zero, one word at a time,
1387 * starting at the end.
1389 while (maxnode > MAX_NUMNODES) {
1390 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1393 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1396 if (maxnode - bits >= MAX_NUMNODES) {
1399 maxnode = MAX_NUMNODES;
1400 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1406 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1409 /* Copy a kernel node mask to user space */
1410 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1413 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1414 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1415 bool compat = in_compat_syscall();
1418 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1420 if (copy > nbytes) {
1421 if (copy > PAGE_SIZE)
1423 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1426 maxnode = nr_node_ids;
1430 return compat_put_bitmap((compat_ulong_t __user *)mask,
1431 nodes_addr(*nodes), maxnode);
1433 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1436 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1437 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1439 *flags = *mode & MPOL_MODE_FLAGS;
1440 *mode &= ~MPOL_MODE_FLAGS;
1442 if ((unsigned int)(*mode) >= MPOL_MAX)
1444 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1446 if (*flags & MPOL_F_NUMA_BALANCING) {
1447 if (*mode != MPOL_BIND)
1449 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1454 static long kernel_mbind(unsigned long start, unsigned long len,
1455 unsigned long mode, const unsigned long __user *nmask,
1456 unsigned long maxnode, unsigned int flags)
1458 unsigned short mode_flags;
1463 start = untagged_addr(start);
1464 err = sanitize_mpol_flags(&lmode, &mode_flags);
1468 err = get_nodes(&nodes, nmask, maxnode);
1472 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1475 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1476 unsigned long, mode, const unsigned long __user *, nmask,
1477 unsigned long, maxnode, unsigned int, flags)
1479 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1482 /* Set the process memory policy */
1483 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1484 unsigned long maxnode)
1486 unsigned short mode_flags;
1491 err = sanitize_mpol_flags(&lmode, &mode_flags);
1495 err = get_nodes(&nodes, nmask, maxnode);
1499 return do_set_mempolicy(lmode, mode_flags, &nodes);
1502 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1503 unsigned long, maxnode)
1505 return kernel_set_mempolicy(mode, nmask, maxnode);
1508 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1509 const unsigned long __user *old_nodes,
1510 const unsigned long __user *new_nodes)
1512 struct mm_struct *mm = NULL;
1513 struct task_struct *task;
1514 nodemask_t task_nodes;
1518 NODEMASK_SCRATCH(scratch);
1523 old = &scratch->mask1;
1524 new = &scratch->mask2;
1526 err = get_nodes(old, old_nodes, maxnode);
1530 err = get_nodes(new, new_nodes, maxnode);
1534 /* Find the mm_struct */
1536 task = pid ? find_task_by_vpid(pid) : current;
1542 get_task_struct(task);
1547 * Check if this process has the right to modify the specified process.
1548 * Use the regular "ptrace_may_access()" checks.
1550 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1557 task_nodes = cpuset_mems_allowed(task);
1558 /* Is the user allowed to access the target nodes? */
1559 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1564 task_nodes = cpuset_mems_allowed(current);
1565 nodes_and(*new, *new, task_nodes);
1566 if (nodes_empty(*new))
1569 err = security_task_movememory(task);
1573 mm = get_task_mm(task);
1574 put_task_struct(task);
1581 err = do_migrate_pages(mm, old, new,
1582 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1586 NODEMASK_SCRATCH_FREE(scratch);
1591 put_task_struct(task);
1596 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1597 const unsigned long __user *, old_nodes,
1598 const unsigned long __user *, new_nodes)
1600 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1604 /* Retrieve NUMA policy */
1605 static int kernel_get_mempolicy(int __user *policy,
1606 unsigned long __user *nmask,
1607 unsigned long maxnode,
1609 unsigned long flags)
1615 if (nmask != NULL && maxnode < nr_node_ids)
1618 addr = untagged_addr(addr);
1620 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1625 if (policy && put_user(pval, policy))
1629 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1634 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1635 unsigned long __user *, nmask, unsigned long, maxnode,
1636 unsigned long, addr, unsigned long, flags)
1638 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1641 bool vma_migratable(struct vm_area_struct *vma)
1643 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1647 * DAX device mappings require predictable access latency, so avoid
1648 * incurring periodic faults.
1650 if (vma_is_dax(vma))
1653 if (is_vm_hugetlb_page(vma) &&
1654 !hugepage_migration_supported(hstate_vma(vma)))
1658 * Migration allocates pages in the highest zone. If we cannot
1659 * do so then migration (at least from node to node) is not
1663 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1669 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1672 struct mempolicy *pol = NULL;
1675 if (vma->vm_ops && vma->vm_ops->get_policy) {
1676 pol = vma->vm_ops->get_policy(vma, addr);
1677 } else if (vma->vm_policy) {
1678 pol = vma->vm_policy;
1681 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1682 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1683 * count on these policies which will be dropped by
1684 * mpol_cond_put() later
1686 if (mpol_needs_cond_ref(pol))
1695 * get_vma_policy(@vma, @addr)
1696 * @vma: virtual memory area whose policy is sought
1697 * @addr: address in @vma for shared policy lookup
1699 * Returns effective policy for a VMA at specified address.
1700 * Falls back to current->mempolicy or system default policy, as necessary.
1701 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1702 * count--added by the get_policy() vm_op, as appropriate--to protect against
1703 * freeing by another task. It is the caller's responsibility to free the
1704 * extra reference for shared policies.
1706 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1709 struct mempolicy *pol = __get_vma_policy(vma, addr);
1712 pol = get_task_policy(current);
1717 bool vma_policy_mof(struct vm_area_struct *vma)
1719 struct mempolicy *pol;
1721 if (vma->vm_ops && vma->vm_ops->get_policy) {
1724 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1725 if (pol && (pol->flags & MPOL_F_MOF))
1732 pol = vma->vm_policy;
1734 pol = get_task_policy(current);
1736 return pol->flags & MPOL_F_MOF;
1739 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1741 enum zone_type dynamic_policy_zone = policy_zone;
1743 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1746 * if policy->nodes has movable memory only,
1747 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1749 * policy->nodes is intersect with node_states[N_MEMORY].
1750 * so if the following test fails, it implies
1751 * policy->nodes has movable memory only.
1753 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1754 dynamic_policy_zone = ZONE_MOVABLE;
1756 return zone >= dynamic_policy_zone;
1760 * Return a nodemask representing a mempolicy for filtering nodes for
1763 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1765 int mode = policy->mode;
1767 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1768 if (unlikely(mode == MPOL_BIND) &&
1769 apply_policy_zone(policy, gfp_zone(gfp)) &&
1770 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1771 return &policy->nodes;
1773 if (mode == MPOL_PREFERRED_MANY)
1774 return &policy->nodes;
1780 * Return the preferred node id for 'prefer' mempolicy, and return
1781 * the given id for all other policies.
1783 * policy_node() is always coupled with policy_nodemask(), which
1784 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1786 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1788 if (policy->mode == MPOL_PREFERRED) {
1789 nd = first_node(policy->nodes);
1792 * __GFP_THISNODE shouldn't even be used with the bind policy
1793 * because we might easily break the expectation to stay on the
1794 * requested node and not break the policy.
1796 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1802 /* Do dynamic interleaving for a process */
1803 static unsigned interleave_nodes(struct mempolicy *policy)
1806 struct task_struct *me = current;
1808 next = next_node_in(me->il_prev, policy->nodes);
1809 if (next < MAX_NUMNODES)
1815 * Depending on the memory policy provide a node from which to allocate the
1818 unsigned int mempolicy_slab_node(void)
1820 struct mempolicy *policy;
1821 int node = numa_mem_id();
1826 policy = current->mempolicy;
1830 switch (policy->mode) {
1831 case MPOL_PREFERRED:
1832 return first_node(policy->nodes);
1834 case MPOL_INTERLEAVE:
1835 return interleave_nodes(policy);
1838 case MPOL_PREFERRED_MANY:
1843 * Follow bind policy behavior and start allocation at the
1846 struct zonelist *zonelist;
1847 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1848 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1849 z = first_zones_zonelist(zonelist, highest_zoneidx,
1851 return z->zone ? zone_to_nid(z->zone) : node;
1862 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1863 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1864 * number of present nodes.
1866 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1868 nodemask_t nodemask = pol->nodes;
1869 unsigned int target, nnodes;
1873 * The barrier will stabilize the nodemask in a register or on
1874 * the stack so that it will stop changing under the code.
1876 * Between first_node() and next_node(), pol->nodes could be changed
1877 * by other threads. So we put pol->nodes in a local stack.
1881 nnodes = nodes_weight(nodemask);
1883 return numa_node_id();
1884 target = (unsigned int)n % nnodes;
1885 nid = first_node(nodemask);
1886 for (i = 0; i < target; i++)
1887 nid = next_node(nid, nodemask);
1891 /* Determine a node number for interleave */
1892 static inline unsigned interleave_nid(struct mempolicy *pol,
1893 struct vm_area_struct *vma, unsigned long addr, int shift)
1899 * for small pages, there is no difference between
1900 * shift and PAGE_SHIFT, so the bit-shift is safe.
1901 * for huge pages, since vm_pgoff is in units of small
1902 * pages, we need to shift off the always 0 bits to get
1905 BUG_ON(shift < PAGE_SHIFT);
1906 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1907 off += (addr - vma->vm_start) >> shift;
1908 return offset_il_node(pol, off);
1910 return interleave_nodes(pol);
1913 #ifdef CONFIG_HUGETLBFS
1915 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1916 * @vma: virtual memory area whose policy is sought
1917 * @addr: address in @vma for shared policy lookup and interleave policy
1918 * @gfp_flags: for requested zone
1919 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1920 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1922 * Returns a nid suitable for a huge page allocation and a pointer
1923 * to the struct mempolicy for conditional unref after allocation.
1924 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
1925 * to the mempolicy's @nodemask for filtering the zonelist.
1927 * Must be protected by read_mems_allowed_begin()
1929 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1930 struct mempolicy **mpol, nodemask_t **nodemask)
1935 *mpol = get_vma_policy(vma, addr);
1937 mode = (*mpol)->mode;
1939 if (unlikely(mode == MPOL_INTERLEAVE)) {
1940 nid = interleave_nid(*mpol, vma, addr,
1941 huge_page_shift(hstate_vma(vma)));
1943 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1944 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
1945 *nodemask = &(*mpol)->nodes;
1951 * init_nodemask_of_mempolicy
1953 * If the current task's mempolicy is "default" [NULL], return 'false'
1954 * to indicate default policy. Otherwise, extract the policy nodemask
1955 * for 'bind' or 'interleave' policy into the argument nodemask, or
1956 * initialize the argument nodemask to contain the single node for
1957 * 'preferred' or 'local' policy and return 'true' to indicate presence
1958 * of non-default mempolicy.
1960 * We don't bother with reference counting the mempolicy [mpol_get/put]
1961 * because the current task is examining it's own mempolicy and a task's
1962 * mempolicy is only ever changed by the task itself.
1964 * N.B., it is the caller's responsibility to free a returned nodemask.
1966 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1968 struct mempolicy *mempolicy;
1970 if (!(mask && current->mempolicy))
1974 mempolicy = current->mempolicy;
1975 switch (mempolicy->mode) {
1976 case MPOL_PREFERRED:
1977 case MPOL_PREFERRED_MANY:
1979 case MPOL_INTERLEAVE:
1980 *mask = mempolicy->nodes;
1984 init_nodemask_of_node(mask, numa_node_id());
1990 task_unlock(current);
1997 * mempolicy_in_oom_domain
1999 * If tsk's mempolicy is "bind", check for intersection between mask and
2000 * the policy nodemask. Otherwise, return true for all other policies
2001 * including "interleave", as a tsk with "interleave" policy may have
2002 * memory allocated from all nodes in system.
2004 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2006 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2007 const nodemask_t *mask)
2009 struct mempolicy *mempolicy;
2016 mempolicy = tsk->mempolicy;
2017 if (mempolicy && mempolicy->mode == MPOL_BIND)
2018 ret = nodes_intersects(mempolicy->nodes, *mask);
2024 /* Allocate a page in interleaved policy.
2025 Own path because it needs to do special accounting. */
2026 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2031 page = __alloc_pages(gfp, order, nid, NULL);
2032 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2033 if (!static_branch_likely(&vm_numa_stat_key))
2035 if (page && page_to_nid(page) == nid) {
2037 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2043 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2044 int nid, struct mempolicy *pol)
2047 gfp_t preferred_gfp;
2050 * This is a two pass approach. The first pass will only try the
2051 * preferred nodes but skip the direct reclaim and allow the
2052 * allocation to fail, while the second pass will try all the
2055 preferred_gfp = gfp | __GFP_NOWARN;
2056 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2057 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2059 page = __alloc_pages(gfp, order, numa_node_id(), NULL);
2065 * alloc_pages_vma - Allocate a page for a VMA.
2067 * @order: Order of the GFP allocation.
2068 * @vma: Pointer to VMA or NULL if not available.
2069 * @addr: Virtual address of the allocation. Must be inside @vma.
2070 * @node: Which node to prefer for allocation (modulo policy).
2071 * @hugepage: For hugepages try only the preferred node if possible.
2073 * Allocate a page for a specific address in @vma, using the appropriate
2074 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2075 * of the mm_struct of the VMA to prevent it from going away. Should be
2076 * used for all allocations for pages that will be mapped into user space.
2078 * Return: The page on success or NULL if allocation fails.
2080 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2081 unsigned long addr, int node, bool hugepage)
2083 struct mempolicy *pol;
2088 pol = get_vma_policy(vma, addr);
2090 if (pol->mode == MPOL_INTERLEAVE) {
2093 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2095 page = alloc_page_interleave(gfp, order, nid);
2099 if (pol->mode == MPOL_PREFERRED_MANY) {
2100 page = alloc_pages_preferred_many(gfp, order, node, pol);
2105 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2106 int hpage_node = node;
2109 * For hugepage allocation and non-interleave policy which
2110 * allows the current node (or other explicitly preferred
2111 * node) we only try to allocate from the current/preferred
2112 * node and don't fall back to other nodes, as the cost of
2113 * remote accesses would likely offset THP benefits.
2115 * If the policy is interleave or does not allow the current
2116 * node in its nodemask, we allocate the standard way.
2118 if (pol->mode == MPOL_PREFERRED)
2119 hpage_node = first_node(pol->nodes);
2121 nmask = policy_nodemask(gfp, pol);
2122 if (!nmask || node_isset(hpage_node, *nmask)) {
2125 * First, try to allocate THP only on local node, but
2126 * don't reclaim unnecessarily, just compact.
2128 page = __alloc_pages_node(hpage_node,
2129 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2132 * If hugepage allocations are configured to always
2133 * synchronous compact or the vma has been madvised
2134 * to prefer hugepage backing, retry allowing remote
2135 * memory with both reclaim and compact as well.
2137 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2138 page = __alloc_pages(gfp, order, hpage_node, nmask);
2144 nmask = policy_nodemask(gfp, pol);
2145 preferred_nid = policy_node(gfp, pol, node);
2146 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2151 EXPORT_SYMBOL(alloc_pages_vma);
2154 * alloc_pages - Allocate pages.
2156 * @order: Power of two of number of pages to allocate.
2158 * Allocate 1 << @order contiguous pages. The physical address of the
2159 * first page is naturally aligned (eg an order-3 allocation will be aligned
2160 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2161 * process is honoured when in process context.
2163 * Context: Can be called from any context, providing the appropriate GFP
2165 * Return: The page on success or NULL if allocation fails.
2167 struct page *alloc_pages(gfp_t gfp, unsigned order)
2169 struct mempolicy *pol = &default_policy;
2172 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2173 pol = get_task_policy(current);
2176 * No reference counting needed for current->mempolicy
2177 * nor system default_policy
2179 if (pol->mode == MPOL_INTERLEAVE)
2180 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2181 else if (pol->mode == MPOL_PREFERRED_MANY)
2182 page = alloc_pages_preferred_many(gfp, order,
2183 numa_node_id(), pol);
2185 page = __alloc_pages(gfp, order,
2186 policy_node(gfp, pol, numa_node_id()),
2187 policy_nodemask(gfp, pol));
2191 EXPORT_SYMBOL(alloc_pages);
2193 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2195 struct mempolicy *pol = mpol_dup(vma_policy(src));
2198 return PTR_ERR(pol);
2199 dst->vm_policy = pol;
2204 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2205 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2206 * with the mems_allowed returned by cpuset_mems_allowed(). This
2207 * keeps mempolicies cpuset relative after its cpuset moves. See
2208 * further kernel/cpuset.c update_nodemask().
2210 * current's mempolicy may be rebinded by the other task(the task that changes
2211 * cpuset's mems), so we needn't do rebind work for current task.
2214 /* Slow path of a mempolicy duplicate */
2215 struct mempolicy *__mpol_dup(struct mempolicy *old)
2217 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2220 return ERR_PTR(-ENOMEM);
2222 /* task's mempolicy is protected by alloc_lock */
2223 if (old == current->mempolicy) {
2226 task_unlock(current);
2230 if (current_cpuset_is_being_rebound()) {
2231 nodemask_t mems = cpuset_mems_allowed(current);
2232 mpol_rebind_policy(new, &mems);
2234 atomic_set(&new->refcnt, 1);
2238 /* Slow path of a mempolicy comparison */
2239 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2243 if (a->mode != b->mode)
2245 if (a->flags != b->flags)
2247 if (mpol_store_user_nodemask(a))
2248 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2253 case MPOL_INTERLEAVE:
2254 case MPOL_PREFERRED:
2255 case MPOL_PREFERRED_MANY:
2256 return !!nodes_equal(a->nodes, b->nodes);
2266 * Shared memory backing store policy support.
2268 * Remember policies even when nobody has shared memory mapped.
2269 * The policies are kept in Red-Black tree linked from the inode.
2270 * They are protected by the sp->lock rwlock, which should be held
2271 * for any accesses to the tree.
2275 * lookup first element intersecting start-end. Caller holds sp->lock for
2276 * reading or for writing
2278 static struct sp_node *
2279 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2281 struct rb_node *n = sp->root.rb_node;
2284 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2286 if (start >= p->end)
2288 else if (end <= p->start)
2296 struct sp_node *w = NULL;
2297 struct rb_node *prev = rb_prev(n);
2300 w = rb_entry(prev, struct sp_node, nd);
2301 if (w->end <= start)
2305 return rb_entry(n, struct sp_node, nd);
2309 * Insert a new shared policy into the list. Caller holds sp->lock for
2312 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2314 struct rb_node **p = &sp->root.rb_node;
2315 struct rb_node *parent = NULL;
2320 nd = rb_entry(parent, struct sp_node, nd);
2321 if (new->start < nd->start)
2323 else if (new->end > nd->end)
2324 p = &(*p)->rb_right;
2328 rb_link_node(&new->nd, parent, p);
2329 rb_insert_color(&new->nd, &sp->root);
2330 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2331 new->policy ? new->policy->mode : 0);
2334 /* Find shared policy intersecting idx */
2336 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2338 struct mempolicy *pol = NULL;
2341 if (!sp->root.rb_node)
2343 read_lock(&sp->lock);
2344 sn = sp_lookup(sp, idx, idx+1);
2346 mpol_get(sn->policy);
2349 read_unlock(&sp->lock);
2353 static void sp_free(struct sp_node *n)
2355 mpol_put(n->policy);
2356 kmem_cache_free(sn_cache, n);
2360 * mpol_misplaced - check whether current page node is valid in policy
2362 * @page: page to be checked
2363 * @vma: vm area where page mapped
2364 * @addr: virtual address where page mapped
2366 * Lookup current policy node id for vma,addr and "compare to" page's
2367 * node id. Policy determination "mimics" alloc_page_vma().
2368 * Called from fault path where we know the vma and faulting address.
2370 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2371 * policy, or a suitable node ID to allocate a replacement page from.
2373 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2375 struct mempolicy *pol;
2377 int curnid = page_to_nid(page);
2378 unsigned long pgoff;
2379 int thiscpu = raw_smp_processor_id();
2380 int thisnid = cpu_to_node(thiscpu);
2381 int polnid = NUMA_NO_NODE;
2382 int ret = NUMA_NO_NODE;
2384 pol = get_vma_policy(vma, addr);
2385 if (!(pol->flags & MPOL_F_MOF))
2388 switch (pol->mode) {
2389 case MPOL_INTERLEAVE:
2390 pgoff = vma->vm_pgoff;
2391 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2392 polnid = offset_il_node(pol, pgoff);
2395 case MPOL_PREFERRED:
2396 if (node_isset(curnid, pol->nodes))
2398 polnid = first_node(pol->nodes);
2402 polnid = numa_node_id();
2406 /* Optimize placement among multiple nodes via NUMA balancing */
2407 if (pol->flags & MPOL_F_MORON) {
2408 if (node_isset(thisnid, pol->nodes))
2414 case MPOL_PREFERRED_MANY:
2416 * use current page if in policy nodemask,
2417 * else select nearest allowed node, if any.
2418 * If no allowed nodes, use current [!misplaced].
2420 if (node_isset(curnid, pol->nodes))
2422 z = first_zones_zonelist(
2423 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2424 gfp_zone(GFP_HIGHUSER),
2426 polnid = zone_to_nid(z->zone);
2433 /* Migrate the page towards the node whose CPU is referencing it */
2434 if (pol->flags & MPOL_F_MORON) {
2437 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2441 if (curnid != polnid)
2450 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2451 * dropped after task->mempolicy is set to NULL so that any allocation done as
2452 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2455 void mpol_put_task_policy(struct task_struct *task)
2457 struct mempolicy *pol;
2460 pol = task->mempolicy;
2461 task->mempolicy = NULL;
2466 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2468 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2469 rb_erase(&n->nd, &sp->root);
2473 static void sp_node_init(struct sp_node *node, unsigned long start,
2474 unsigned long end, struct mempolicy *pol)
2476 node->start = start;
2481 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2482 struct mempolicy *pol)
2485 struct mempolicy *newpol;
2487 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2491 newpol = mpol_dup(pol);
2492 if (IS_ERR(newpol)) {
2493 kmem_cache_free(sn_cache, n);
2496 newpol->flags |= MPOL_F_SHARED;
2497 sp_node_init(n, start, end, newpol);
2502 /* Replace a policy range. */
2503 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2504 unsigned long end, struct sp_node *new)
2507 struct sp_node *n_new = NULL;
2508 struct mempolicy *mpol_new = NULL;
2512 write_lock(&sp->lock);
2513 n = sp_lookup(sp, start, end);
2514 /* Take care of old policies in the same range. */
2515 while (n && n->start < end) {
2516 struct rb_node *next = rb_next(&n->nd);
2517 if (n->start >= start) {
2523 /* Old policy spanning whole new range. */
2528 *mpol_new = *n->policy;
2529 atomic_set(&mpol_new->refcnt, 1);
2530 sp_node_init(n_new, end, n->end, mpol_new);
2532 sp_insert(sp, n_new);
2541 n = rb_entry(next, struct sp_node, nd);
2545 write_unlock(&sp->lock);
2552 kmem_cache_free(sn_cache, n_new);
2557 write_unlock(&sp->lock);
2559 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2562 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2565 atomic_set(&mpol_new->refcnt, 1);
2570 * mpol_shared_policy_init - initialize shared policy for inode
2571 * @sp: pointer to inode shared policy
2572 * @mpol: struct mempolicy to install
2574 * Install non-NULL @mpol in inode's shared policy rb-tree.
2575 * On entry, the current task has a reference on a non-NULL @mpol.
2576 * This must be released on exit.
2577 * This is called at get_inode() calls and we can use GFP_KERNEL.
2579 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2583 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2584 rwlock_init(&sp->lock);
2587 struct vm_area_struct pvma;
2588 struct mempolicy *new;
2589 NODEMASK_SCRATCH(scratch);
2593 /* contextualize the tmpfs mount point mempolicy */
2594 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2596 goto free_scratch; /* no valid nodemask intersection */
2599 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2600 task_unlock(current);
2604 /* Create pseudo-vma that contains just the policy */
2605 vma_init(&pvma, NULL);
2606 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2607 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2610 mpol_put(new); /* drop initial ref */
2612 NODEMASK_SCRATCH_FREE(scratch);
2614 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2618 int mpol_set_shared_policy(struct shared_policy *info,
2619 struct vm_area_struct *vma, struct mempolicy *npol)
2622 struct sp_node *new = NULL;
2623 unsigned long sz = vma_pages(vma);
2625 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2627 sz, npol ? npol->mode : -1,
2628 npol ? npol->flags : -1,
2629 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2632 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2636 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2642 /* Free a backing policy store on inode delete. */
2643 void mpol_free_shared_policy(struct shared_policy *p)
2646 struct rb_node *next;
2648 if (!p->root.rb_node)
2650 write_lock(&p->lock);
2651 next = rb_first(&p->root);
2653 n = rb_entry(next, struct sp_node, nd);
2654 next = rb_next(&n->nd);
2657 write_unlock(&p->lock);
2660 #ifdef CONFIG_NUMA_BALANCING
2661 static int __initdata numabalancing_override;
2663 static void __init check_numabalancing_enable(void)
2665 bool numabalancing_default = false;
2667 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2668 numabalancing_default = true;
2670 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2671 if (numabalancing_override)
2672 set_numabalancing_state(numabalancing_override == 1);
2674 if (num_online_nodes() > 1 && !numabalancing_override) {
2675 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2676 numabalancing_default ? "Enabling" : "Disabling");
2677 set_numabalancing_state(numabalancing_default);
2681 static int __init setup_numabalancing(char *str)
2687 if (!strcmp(str, "enable")) {
2688 numabalancing_override = 1;
2690 } else if (!strcmp(str, "disable")) {
2691 numabalancing_override = -1;
2696 pr_warn("Unable to parse numa_balancing=\n");
2700 __setup("numa_balancing=", setup_numabalancing);
2702 static inline void __init check_numabalancing_enable(void)
2705 #endif /* CONFIG_NUMA_BALANCING */
2707 /* assumes fs == KERNEL_DS */
2708 void __init numa_policy_init(void)
2710 nodemask_t interleave_nodes;
2711 unsigned long largest = 0;
2712 int nid, prefer = 0;
2714 policy_cache = kmem_cache_create("numa_policy",
2715 sizeof(struct mempolicy),
2716 0, SLAB_PANIC, NULL);
2718 sn_cache = kmem_cache_create("shared_policy_node",
2719 sizeof(struct sp_node),
2720 0, SLAB_PANIC, NULL);
2722 for_each_node(nid) {
2723 preferred_node_policy[nid] = (struct mempolicy) {
2724 .refcnt = ATOMIC_INIT(1),
2725 .mode = MPOL_PREFERRED,
2726 .flags = MPOL_F_MOF | MPOL_F_MORON,
2727 .nodes = nodemask_of_node(nid),
2732 * Set interleaving policy for system init. Interleaving is only
2733 * enabled across suitably sized nodes (default is >= 16MB), or
2734 * fall back to the largest node if they're all smaller.
2736 nodes_clear(interleave_nodes);
2737 for_each_node_state(nid, N_MEMORY) {
2738 unsigned long total_pages = node_present_pages(nid);
2740 /* Preserve the largest node */
2741 if (largest < total_pages) {
2742 largest = total_pages;
2746 /* Interleave this node? */
2747 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2748 node_set(nid, interleave_nodes);
2751 /* All too small, use the largest */
2752 if (unlikely(nodes_empty(interleave_nodes)))
2753 node_set(prefer, interleave_nodes);
2755 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2756 pr_err("%s: interleaving failed\n", __func__);
2758 check_numabalancing_enable();
2761 /* Reset policy of current process to default */
2762 void numa_default_policy(void)
2764 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2768 * Parse and format mempolicy from/to strings
2771 static const char * const policy_modes[] =
2773 [MPOL_DEFAULT] = "default",
2774 [MPOL_PREFERRED] = "prefer",
2775 [MPOL_BIND] = "bind",
2776 [MPOL_INTERLEAVE] = "interleave",
2777 [MPOL_LOCAL] = "local",
2778 [MPOL_PREFERRED_MANY] = "prefer (many)",
2784 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2785 * @str: string containing mempolicy to parse
2786 * @mpol: pointer to struct mempolicy pointer, returned on success.
2789 * <mode>[=<flags>][:<nodelist>]
2791 * On success, returns 0, else 1
2793 int mpol_parse_str(char *str, struct mempolicy **mpol)
2795 struct mempolicy *new = NULL;
2796 unsigned short mode_flags;
2798 char *nodelist = strchr(str, ':');
2799 char *flags = strchr(str, '=');
2803 *flags++ = '\0'; /* terminate mode string */
2806 /* NUL-terminate mode or flags string */
2808 if (nodelist_parse(nodelist, nodes))
2810 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2815 mode = match_string(policy_modes, MPOL_MAX, str);
2820 case MPOL_PREFERRED:
2822 * Insist on a nodelist of one node only, although later
2823 * we use first_node(nodes) to grab a single node, so here
2824 * nodelist (or nodes) cannot be empty.
2827 char *rest = nodelist;
2828 while (isdigit(*rest))
2832 if (nodes_empty(nodes))
2836 case MPOL_INTERLEAVE:
2838 * Default to online nodes with memory if no nodelist
2841 nodes = node_states[N_MEMORY];
2845 * Don't allow a nodelist; mpol_new() checks flags
2852 * Insist on a empty nodelist
2857 case MPOL_PREFERRED_MANY:
2860 * Insist on a nodelist
2869 * Currently, we only support two mutually exclusive
2872 if (!strcmp(flags, "static"))
2873 mode_flags |= MPOL_F_STATIC_NODES;
2874 else if (!strcmp(flags, "relative"))
2875 mode_flags |= MPOL_F_RELATIVE_NODES;
2880 new = mpol_new(mode, mode_flags, &nodes);
2885 * Save nodes for mpol_to_str() to show the tmpfs mount options
2886 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2888 if (mode != MPOL_PREFERRED) {
2890 } else if (nodelist) {
2891 nodes_clear(new->nodes);
2892 node_set(first_node(nodes), new->nodes);
2894 new->mode = MPOL_LOCAL;
2898 * Save nodes for contextualization: this will be used to "clone"
2899 * the mempolicy in a specific context [cpuset] at a later time.
2901 new->w.user_nodemask = nodes;
2906 /* Restore string for error message */
2915 #endif /* CONFIG_TMPFS */
2918 * mpol_to_str - format a mempolicy structure for printing
2919 * @buffer: to contain formatted mempolicy string
2920 * @maxlen: length of @buffer
2921 * @pol: pointer to mempolicy to be formatted
2923 * Convert @pol into a string. If @buffer is too short, truncate the string.
2924 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2925 * longest flag, "relative", and to display at least a few node ids.
2927 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2930 nodemask_t nodes = NODE_MASK_NONE;
2931 unsigned short mode = MPOL_DEFAULT;
2932 unsigned short flags = 0;
2934 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2943 case MPOL_PREFERRED:
2944 case MPOL_PREFERRED_MANY:
2946 case MPOL_INTERLEAVE:
2951 snprintf(p, maxlen, "unknown");
2955 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2957 if (flags & MPOL_MODE_FLAGS) {
2958 p += snprintf(p, buffer + maxlen - p, "=");
2961 * Currently, the only defined flags are mutually exclusive
2963 if (flags & MPOL_F_STATIC_NODES)
2964 p += snprintf(p, buffer + maxlen - p, "static");
2965 else if (flags & MPOL_F_RELATIVE_NODES)
2966 p += snprintf(p, buffer + maxlen - p, "relative");
2969 if (!nodes_empty(nodes))
2970 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2971 nodemask_pr_args(&nodes));
2974 bool numa_demotion_enabled = false;
2977 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2978 struct kobj_attribute *attr, char *buf)
2980 return sysfs_emit(buf, "%s\n",
2981 numa_demotion_enabled? "true" : "false");
2984 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2985 struct kobj_attribute *attr,
2986 const char *buf, size_t count)
2988 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
2989 numa_demotion_enabled = true;
2990 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
2991 numa_demotion_enabled = false;
2998 static struct kobj_attribute numa_demotion_enabled_attr =
2999 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3000 numa_demotion_enabled_store);
3002 static struct attribute *numa_attrs[] = {
3003 &numa_demotion_enabled_attr.attr,
3007 static const struct attribute_group numa_attr_group = {
3008 .attrs = numa_attrs,
3011 static int __init numa_init_sysfs(void)
3014 struct kobject *numa_kobj;
3016 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3018 pr_err("failed to create numa kobject\n");
3021 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3023 pr_err("failed to register numa group\n");
3029 kobject_put(numa_kobj);
3032 subsys_initcall(numa_init_sysfs);