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>
108 #include <linux/uaccess.h>
110 #include "internal.h"
113 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
114 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
115 #define MPOL_MF_WRLOCK (MPOL_MF_INTERNAL << 2) /* Write-lock walked vmas */
117 static struct kmem_cache *policy_cache;
118 static struct kmem_cache *sn_cache;
120 /* Highest zone. An specific allocation for a zone below that is not
122 enum zone_type policy_zone = 0;
125 * run-time system-wide default policy => local allocation
127 static struct mempolicy default_policy = {
128 .refcnt = ATOMIC_INIT(1), /* never free it */
132 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
135 * numa_map_to_online_node - Find closest online node
136 * @node: Node id to start the search
138 * Lookup the next closest node by distance if @nid is not online.
140 * Return: this @node if it is online, otherwise the closest node by distance
142 int numa_map_to_online_node(int node)
144 int min_dist = INT_MAX, dist, n, min_node;
146 if (node == NUMA_NO_NODE || node_online(node))
150 for_each_online_node(n) {
151 dist = node_distance(node, n);
152 if (dist < min_dist) {
160 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
162 struct mempolicy *get_task_policy(struct task_struct *p)
164 struct mempolicy *pol = p->mempolicy;
170 node = numa_node_id();
171 if (node != NUMA_NO_NODE) {
172 pol = &preferred_node_policy[node];
173 /* preferred_node_policy is not initialised early in boot */
178 return &default_policy;
181 static const struct mempolicy_operations {
182 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
183 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
184 } mpol_ops[MPOL_MAX];
186 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
188 return pol->flags & MPOL_MODE_FLAGS;
191 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
192 const nodemask_t *rel)
195 nodes_fold(tmp, *orig, nodes_weight(*rel));
196 nodes_onto(*ret, tmp, *rel);
199 static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
201 if (nodes_empty(*nodes))
207 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
209 if (nodes_empty(*nodes))
212 nodes_clear(pol->nodes);
213 node_set(first_node(*nodes), pol->nodes);
218 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
219 * any, for the new policy. mpol_new() has already validated the nodes
220 * parameter with respect to the policy mode and flags.
222 * Must be called holding task's alloc_lock to protect task's mems_allowed
223 * and mempolicy. May also be called holding the mmap_lock for write.
225 static int mpol_set_nodemask(struct mempolicy *pol,
226 const nodemask_t *nodes, struct nodemask_scratch *nsc)
231 * Default (pol==NULL) resp. local memory policies are not a
232 * subject of any remapping. They also do not need any special
235 if (!pol || pol->mode == MPOL_LOCAL)
239 nodes_and(nsc->mask1,
240 cpuset_current_mems_allowed, node_states[N_MEMORY]);
244 if (pol->flags & MPOL_F_RELATIVE_NODES)
245 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
247 nodes_and(nsc->mask2, *nodes, nsc->mask1);
249 if (mpol_store_user_nodemask(pol))
250 pol->w.user_nodemask = *nodes;
252 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
254 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
259 * This function just creates a new policy, does some check and simple
260 * initialization. You must invoke mpol_set_nodemask() to set nodes.
262 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
265 struct mempolicy *policy;
267 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
268 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
270 if (mode == MPOL_DEFAULT) {
271 if (nodes && !nodes_empty(*nodes))
272 return ERR_PTR(-EINVAL);
278 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
279 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
280 * All other modes require a valid pointer to a non-empty nodemask.
282 if (mode == MPOL_PREFERRED) {
283 if (nodes_empty(*nodes)) {
284 if (((flags & MPOL_F_STATIC_NODES) ||
285 (flags & MPOL_F_RELATIVE_NODES)))
286 return ERR_PTR(-EINVAL);
290 } else if (mode == MPOL_LOCAL) {
291 if (!nodes_empty(*nodes) ||
292 (flags & MPOL_F_STATIC_NODES) ||
293 (flags & MPOL_F_RELATIVE_NODES))
294 return ERR_PTR(-EINVAL);
295 } else if (nodes_empty(*nodes))
296 return ERR_PTR(-EINVAL);
297 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
299 return ERR_PTR(-ENOMEM);
300 atomic_set(&policy->refcnt, 1);
302 policy->flags = flags;
303 policy->home_node = NUMA_NO_NODE;
308 /* Slow path of a mpol destructor. */
309 void __mpol_put(struct mempolicy *p)
311 if (!atomic_dec_and_test(&p->refcnt))
313 kmem_cache_free(policy_cache, p);
316 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
320 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
329 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
331 pol->w.cpuset_mems_allowed = *nodes;
334 if (nodes_empty(tmp))
340 static void mpol_rebind_preferred(struct mempolicy *pol,
341 const nodemask_t *nodes)
343 pol->w.cpuset_mems_allowed = *nodes;
347 * mpol_rebind_policy - Migrate a policy to a different set of nodes
349 * Per-vma policies are protected by mmap_lock. Allocations using per-task
350 * policies are protected by task->mems_allowed_seq to prevent a premature
351 * OOM/allocation failure due to parallel nodemask modification.
353 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
355 if (!pol || pol->mode == MPOL_LOCAL)
357 if (!mpol_store_user_nodemask(pol) &&
358 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
361 mpol_ops[pol->mode].rebind(pol, newmask);
365 * Wrapper for mpol_rebind_policy() that just requires task
366 * pointer, and updates task mempolicy.
368 * Called with task's alloc_lock held.
371 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
373 mpol_rebind_policy(tsk->mempolicy, new);
377 * Rebind each vma in mm to new nodemask.
379 * Call holding a reference to mm. Takes mm->mmap_lock during call.
382 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
384 struct vm_area_struct *vma;
385 VMA_ITERATOR(vmi, mm, 0);
388 for_each_vma(vmi, vma) {
389 vma_start_write(vma);
390 mpol_rebind_policy(vma->vm_policy, new);
392 mmap_write_unlock(mm);
395 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
397 .rebind = mpol_rebind_default,
399 [MPOL_INTERLEAVE] = {
400 .create = mpol_new_nodemask,
401 .rebind = mpol_rebind_nodemask,
404 .create = mpol_new_preferred,
405 .rebind = mpol_rebind_preferred,
408 .create = mpol_new_nodemask,
409 .rebind = mpol_rebind_nodemask,
412 .rebind = mpol_rebind_default,
414 [MPOL_PREFERRED_MANY] = {
415 .create = mpol_new_nodemask,
416 .rebind = mpol_rebind_preferred,
420 static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
421 unsigned long flags);
423 static bool strictly_unmovable(unsigned long flags)
426 * STRICT without MOVE flags lets do_mbind() fail immediately with -EIO
427 * if any misplaced page is found.
429 return (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ==
434 struct list_head *pagelist;
439 struct vm_area_struct *first;
440 struct folio *large; /* note last large folio encountered */
441 long nr_failed; /* could not be isolated at this time */
445 * Check if the folio's nid is in qp->nmask.
447 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
448 * in the invert of qp->nmask.
450 static inline bool queue_folio_required(struct folio *folio,
451 struct queue_pages *qp)
453 int nid = folio_nid(folio);
454 unsigned long flags = qp->flags;
456 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
459 static void queue_folios_pmd(pmd_t *pmd, struct mm_walk *walk)
462 struct queue_pages *qp = walk->private;
464 if (unlikely(is_pmd_migration_entry(*pmd))) {
468 folio = pfn_folio(pmd_pfn(*pmd));
469 if (is_huge_zero_page(&folio->page)) {
470 walk->action = ACTION_CONTINUE;
473 if (!queue_folio_required(folio, qp))
475 if (!(qp->flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
476 !vma_migratable(walk->vma) ||
477 !migrate_folio_add(folio, qp->pagelist, qp->flags))
482 * Scan through folios, checking if they satisfy the required conditions,
483 * moving them from LRU to local pagelist for migration if they do (or not).
485 * queue_folios_pte_range() has two possible return values:
486 * 0 - continue walking to scan for more, even if an existing folio on the
487 * wrong node could not be isolated and queued for migration.
488 * -EIO - only MPOL_MF_STRICT was specified, without MPOL_MF_MOVE or ..._ALL,
489 * and an existing folio was on a node that does not follow the policy.
491 static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
492 unsigned long end, struct mm_walk *walk)
494 struct vm_area_struct *vma = walk->vma;
496 struct queue_pages *qp = walk->private;
497 unsigned long flags = qp->flags;
498 pte_t *pte, *mapped_pte;
502 ptl = pmd_trans_huge_lock(pmd, vma);
504 queue_folios_pmd(pmd, walk);
509 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
511 walk->action = ACTION_AGAIN;
514 for (; addr != end; pte++, addr += PAGE_SIZE) {
515 ptent = ptep_get(pte);
518 if (!pte_present(ptent)) {
519 if (is_migration_entry(pte_to_swp_entry(ptent)))
523 folio = vm_normal_folio(vma, addr, ptent);
524 if (!folio || folio_is_zone_device(folio))
527 * vm_normal_folio() filters out zero pages, but there might
528 * still be reserved folios to skip, perhaps in a VDSO.
530 if (folio_test_reserved(folio))
532 if (!queue_folio_required(folio, qp))
534 if (folio_test_large(folio)) {
536 * A large folio can only be isolated from LRU once,
537 * but may be mapped by many PTEs (and Copy-On-Write may
538 * intersperse PTEs of other, order 0, folios). This is
539 * a common case, so don't mistake it for failure (but
540 * there can be other cases of multi-mapped pages which
541 * this quick check does not help to filter out - and a
542 * search of the pagelist might grow to be prohibitive).
544 * migrate_pages(&pagelist) returns nr_failed folios, so
545 * check "large" now so that queue_pages_range() returns
546 * a comparable nr_failed folios. This does imply that
547 * if folio could not be isolated for some racy reason
548 * at its first PTE, later PTEs will not give it another
549 * chance of isolation; but keeps the accounting simple.
551 if (folio == qp->large)
555 if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
556 !vma_migratable(vma) ||
557 !migrate_folio_add(folio, qp->pagelist, flags)) {
559 if (strictly_unmovable(flags))
563 pte_unmap_unlock(mapped_pte, ptl);
566 if (qp->nr_failed && strictly_unmovable(flags))
571 static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
572 unsigned long addr, unsigned long end,
573 struct mm_walk *walk)
575 #ifdef CONFIG_HUGETLB_PAGE
576 struct queue_pages *qp = walk->private;
577 unsigned long flags = qp->flags;
582 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
583 entry = huge_ptep_get(pte);
584 if (!pte_present(entry)) {
585 if (unlikely(is_hugetlb_entry_migration(entry)))
589 folio = pfn_folio(pte_pfn(entry));
590 if (!queue_folio_required(folio, qp))
592 if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
593 !vma_migratable(walk->vma)) {
598 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
599 * Choosing not to migrate a shared folio is not counted as a failure.
601 * To check if the folio is shared, ideally we want to make sure
602 * every page is mapped to the same process. Doing that is very
603 * expensive, so check the estimated sharers of the folio instead.
605 if ((flags & MPOL_MF_MOVE_ALL) ||
606 (folio_estimated_sharers(folio) == 1 && !hugetlb_pmd_shared(pte)))
607 if (!isolate_hugetlb(folio, qp->pagelist))
611 if (qp->nr_failed && strictly_unmovable(flags))
617 #ifdef CONFIG_NUMA_BALANCING
619 * This is used to mark a range of virtual addresses to be inaccessible.
620 * These are later cleared by a NUMA hinting fault. Depending on these
621 * faults, pages may be migrated for better NUMA placement.
623 * This is assuming that NUMA faults are handled using PROT_NONE. If
624 * an architecture makes a different choice, it will need further
625 * changes to the core.
627 unsigned long change_prot_numa(struct vm_area_struct *vma,
628 unsigned long addr, unsigned long end)
630 struct mmu_gather tlb;
633 tlb_gather_mmu(&tlb, vma->vm_mm);
635 nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
637 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
639 tlb_finish_mmu(&tlb);
644 static unsigned long change_prot_numa(struct vm_area_struct *vma,
645 unsigned long addr, unsigned long end)
649 #endif /* CONFIG_NUMA_BALANCING */
651 static int queue_pages_test_walk(unsigned long start, unsigned long end,
652 struct mm_walk *walk)
654 struct vm_area_struct *next, *vma = walk->vma;
655 struct queue_pages *qp = walk->private;
656 unsigned long endvma = vma->vm_end;
657 unsigned long flags = qp->flags;
659 /* range check first */
660 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
664 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
665 (qp->start < vma->vm_start))
666 /* hole at head side of range */
669 next = find_vma(vma->vm_mm, vma->vm_end);
670 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
671 ((vma->vm_end < qp->end) &&
672 (!next || vma->vm_end < next->vm_start)))
673 /* hole at middle or tail of range */
677 * Need check MPOL_MF_STRICT to return -EIO if possible
678 * regardless of vma_migratable
680 if (!vma_migratable(vma) &&
681 !(flags & MPOL_MF_STRICT))
687 if (flags & MPOL_MF_LAZY) {
688 /* Similar to task_numa_work, skip inaccessible VMAs */
689 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
690 !(vma->vm_flags & VM_MIXEDMAP))
691 change_prot_numa(vma, start, endvma);
696 * Check page nodes, and queue pages to move, in the current vma.
697 * But if no moving, and no strict checking, the scan can be skipped.
699 if (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
704 static const struct mm_walk_ops queue_pages_walk_ops = {
705 .hugetlb_entry = queue_folios_hugetlb,
706 .pmd_entry = queue_folios_pte_range,
707 .test_walk = queue_pages_test_walk,
708 .walk_lock = PGWALK_RDLOCK,
711 static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
712 .hugetlb_entry = queue_folios_hugetlb,
713 .pmd_entry = queue_folios_pte_range,
714 .test_walk = queue_pages_test_walk,
715 .walk_lock = PGWALK_WRLOCK,
719 * Walk through page tables and collect pages to be migrated.
721 * If pages found in a given range are not on the required set of @nodes,
722 * and migration is allowed, they are isolated and queued to @pagelist.
724 * queue_pages_range() may return:
725 * 0 - all pages already on the right node, or successfully queued for moving
726 * (or neither strict checking nor moving requested: only range checking).
727 * >0 - this number of misplaced folios could not be queued for moving
728 * (a hugetlbfs page or a transparent huge page being counted as 1).
729 * -EIO - a misplaced page found, when MPOL_MF_STRICT specified without MOVEs.
730 * -EFAULT - a hole in the memory range, when MPOL_MF_DISCONTIG_OK unspecified.
733 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
734 nodemask_t *nodes, unsigned long flags,
735 struct list_head *pagelist)
738 struct queue_pages qp = {
739 .pagelist = pagelist,
746 const struct mm_walk_ops *ops = (flags & MPOL_MF_WRLOCK) ?
747 &queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
749 err = walk_page_range(mm, start, end, ops, &qp);
752 /* whole range in hole */
755 return err ? : qp.nr_failed;
759 * Apply policy to a single VMA
760 * This must be called with the mmap_lock held for writing.
762 static int vma_replace_policy(struct vm_area_struct *vma,
763 struct mempolicy *pol)
766 struct mempolicy *old;
767 struct mempolicy *new;
769 vma_assert_write_locked(vma);
771 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
772 vma->vm_start, vma->vm_end, vma->vm_pgoff,
773 vma->vm_ops, vma->vm_file,
774 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
780 if (vma->vm_ops && vma->vm_ops->set_policy) {
781 err = vma->vm_ops->set_policy(vma, new);
786 old = vma->vm_policy;
787 vma->vm_policy = new; /* protected by mmap_lock */
796 /* Split or merge the VMA (if required) and apply the new policy */
797 static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
798 struct vm_area_struct **prev, unsigned long start,
799 unsigned long end, struct mempolicy *new_pol)
801 unsigned long vmstart, vmend;
803 vmend = min(end, vma->vm_end);
804 if (start > vma->vm_start) {
808 vmstart = vma->vm_start;
811 if (mpol_equal(vma_policy(vma), new_pol)) {
816 vma = vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
821 return vma_replace_policy(vma, new_pol);
824 /* Set the process memory policy */
825 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
828 struct mempolicy *new, *old;
829 NODEMASK_SCRATCH(scratch);
835 new = mpol_new(mode, flags, nodes);
842 ret = mpol_set_nodemask(new, nodes, scratch);
844 task_unlock(current);
849 old = current->mempolicy;
850 current->mempolicy = new;
851 if (new && new->mode == MPOL_INTERLEAVE)
852 current->il_prev = MAX_NUMNODES-1;
853 task_unlock(current);
857 NODEMASK_SCRATCH_FREE(scratch);
862 * Return nodemask for policy for get_mempolicy() query
864 * Called with task's alloc_lock held
866 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
869 if (p == &default_policy)
874 case MPOL_INTERLEAVE:
876 case MPOL_PREFERRED_MANY:
880 /* return empty node mask for local allocation */
887 static int lookup_node(struct mm_struct *mm, unsigned long addr)
889 struct page *p = NULL;
892 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
894 ret = page_to_nid(p);
900 /* Retrieve NUMA policy */
901 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
902 unsigned long addr, unsigned long flags)
905 struct mm_struct *mm = current->mm;
906 struct vm_area_struct *vma = NULL;
907 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
910 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
913 if (flags & MPOL_F_MEMS_ALLOWED) {
914 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
916 *policy = 0; /* just so it's initialized */
918 *nmask = cpuset_current_mems_allowed;
919 task_unlock(current);
923 if (flags & MPOL_F_ADDR) {
925 * Do NOT fall back to task policy if the
926 * vma/shared policy at addr is NULL. We
927 * want to return MPOL_DEFAULT in this case.
930 vma = vma_lookup(mm, addr);
932 mmap_read_unlock(mm);
935 if (vma->vm_ops && vma->vm_ops->get_policy)
936 pol = vma->vm_ops->get_policy(vma, addr);
938 pol = vma->vm_policy;
943 pol = &default_policy; /* indicates default behavior */
945 if (flags & MPOL_F_NODE) {
946 if (flags & MPOL_F_ADDR) {
948 * Take a refcount on the mpol, because we are about to
949 * drop the mmap_lock, after which only "pol" remains
950 * valid, "vma" is stale.
955 mmap_read_unlock(mm);
956 err = lookup_node(mm, addr);
960 } else if (pol == current->mempolicy &&
961 pol->mode == MPOL_INTERLEAVE) {
962 *policy = next_node_in(current->il_prev, pol->nodes);
968 *policy = pol == &default_policy ? MPOL_DEFAULT :
971 * Internal mempolicy flags must be masked off before exposing
972 * the policy to userspace.
974 *policy |= (pol->flags & MPOL_MODE_FLAGS);
979 if (mpol_store_user_nodemask(pol)) {
980 *nmask = pol->w.user_nodemask;
983 get_policy_nodemask(pol, nmask);
984 task_unlock(current);
991 mmap_read_unlock(mm);
993 mpol_put(pol_refcount);
997 #ifdef CONFIG_MIGRATION
998 static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1002 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
1003 * Choosing not to migrate a shared folio is not counted as a failure.
1005 * To check if the folio is shared, ideally we want to make sure
1006 * every page is mapped to the same process. Doing that is very
1007 * expensive, so check the estimated sharers of the folio instead.
1009 if ((flags & MPOL_MF_MOVE_ALL) || folio_estimated_sharers(folio) == 1) {
1010 if (folio_isolate_lru(folio)) {
1011 list_add_tail(&folio->lru, foliolist);
1012 node_stat_mod_folio(folio,
1013 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1014 folio_nr_pages(folio));
1017 * Non-movable folio may reach here. And, there may be
1018 * temporary off LRU folios or non-LRU movable folios.
1019 * Treat them as unmovable folios since they can't be
1020 * isolated, so they can't be moved at the moment.
1029 * Migrate pages from one node to a target node.
1030 * Returns error or the number of pages not migrated.
1032 static long migrate_to_node(struct mm_struct *mm, int source, int dest,
1036 struct vm_area_struct *vma;
1037 LIST_HEAD(pagelist);
1040 struct migration_target_control mtc = {
1042 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1046 node_set(source, nmask);
1048 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1049 vma = find_vma(mm, 0);
1052 * This does not migrate the range, but isolates all pages that
1053 * need migration. Between passing in the full user address
1054 * space range and MPOL_MF_DISCONTIG_OK, this call cannot fail,
1055 * but passes back the count of pages which could not be isolated.
1057 nr_failed = queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1058 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1060 if (!list_empty(&pagelist)) {
1061 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1062 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1064 putback_movable_pages(&pagelist);
1073 * Move pages between the two nodesets so as to preserve the physical
1074 * layout as much as possible.
1076 * Returns the number of page that could not be moved.
1078 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1079 const nodemask_t *to, int flags)
1085 lru_cache_disable();
1090 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1091 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1092 * bit in 'tmp', and return that <source, dest> pair for migration.
1093 * The pair of nodemasks 'to' and 'from' define the map.
1095 * If no pair of bits is found that way, fallback to picking some
1096 * pair of 'source' and 'dest' bits that are not the same. If the
1097 * 'source' and 'dest' bits are the same, this represents a node
1098 * that will be migrating to itself, so no pages need move.
1100 * If no bits are left in 'tmp', or if all remaining bits left
1101 * in 'tmp' correspond to the same bit in 'to', return false
1102 * (nothing left to migrate).
1104 * This lets us pick a pair of nodes to migrate between, such that
1105 * if possible the dest node is not already occupied by some other
1106 * source node, minimizing the risk of overloading the memory on a
1107 * node that would happen if we migrated incoming memory to a node
1108 * before migrating outgoing memory source that same node.
1110 * A single scan of tmp is sufficient. As we go, we remember the
1111 * most recent <s, d> pair that moved (s != d). If we find a pair
1112 * that not only moved, but what's better, moved to an empty slot
1113 * (d is not set in tmp), then we break out then, with that pair.
1114 * Otherwise when we finish scanning from_tmp, we at least have the
1115 * most recent <s, d> pair that moved. If we get all the way through
1116 * the scan of tmp without finding any node that moved, much less
1117 * moved to an empty node, then there is nothing left worth migrating.
1121 while (!nodes_empty(tmp)) {
1123 int source = NUMA_NO_NODE;
1126 for_each_node_mask(s, tmp) {
1129 * do_migrate_pages() tries to maintain the relative
1130 * node relationship of the pages established between
1131 * threads and memory areas.
1133 * However if the number of source nodes is not equal to
1134 * the number of destination nodes we can not preserve
1135 * this node relative relationship. In that case, skip
1136 * copying memory from a node that is in the destination
1139 * Example: [2,3,4] -> [3,4,5] moves everything.
1140 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1143 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1144 (node_isset(s, *to)))
1147 d = node_remap(s, *from, *to);
1151 source = s; /* Node moved. Memorize */
1154 /* dest not in remaining from nodes? */
1155 if (!node_isset(dest, tmp))
1158 if (source == NUMA_NO_NODE)
1161 node_clear(source, tmp);
1162 err = migrate_to_node(mm, source, dest, flags);
1168 mmap_read_unlock(mm);
1173 return (nr_failed < INT_MAX) ? nr_failed : INT_MAX;
1177 * Allocate a new page for page migration based on vma policy.
1178 * Start by assuming the page is mapped by the same vma as contains @start.
1179 * Search forward from there, if not. N.B., this assumes that the
1180 * list of pages handed to migrate_pages()--which is how we get here--
1181 * is in virtual address order.
1183 static struct folio *new_folio(struct folio *src, unsigned long start)
1185 struct vm_area_struct *vma;
1186 unsigned long address;
1187 VMA_ITERATOR(vmi, current->mm, start);
1188 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1190 for_each_vma(vmi, vma) {
1191 address = page_address_in_vma(&src->page, vma);
1192 if (address != -EFAULT)
1196 if (folio_test_hugetlb(src)) {
1197 return alloc_hugetlb_folio_vma(folio_hstate(src),
1201 if (folio_test_large(src))
1202 gfp = GFP_TRANSHUGE;
1205 * if !vma, vma_alloc_folio() will use task or system default policy
1207 return vma_alloc_folio(gfp, folio_order(src), vma, address,
1208 folio_test_large(src));
1212 static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1213 unsigned long flags)
1218 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1219 const nodemask_t *to, int flags)
1224 static struct folio *new_folio(struct folio *src, unsigned long start)
1230 static long do_mbind(unsigned long start, unsigned long len,
1231 unsigned short mode, unsigned short mode_flags,
1232 nodemask_t *nmask, unsigned long flags)
1234 struct mm_struct *mm = current->mm;
1235 struct vm_area_struct *vma, *prev;
1236 struct vma_iterator vmi;
1237 struct mempolicy *new;
1241 LIST_HEAD(pagelist);
1243 if (flags & ~(unsigned long)MPOL_MF_VALID)
1245 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1248 if (start & ~PAGE_MASK)
1251 if (mode == MPOL_DEFAULT)
1252 flags &= ~MPOL_MF_STRICT;
1254 len = PAGE_ALIGN(len);
1262 new = mpol_new(mode, mode_flags, nmask);
1264 return PTR_ERR(new);
1266 if (flags & MPOL_MF_LAZY)
1267 new->flags |= MPOL_F_MOF;
1270 * If we are using the default policy then operation
1271 * on discontinuous address spaces is okay after all
1274 flags |= MPOL_MF_DISCONTIG_OK;
1276 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1277 start, start + len, mode, mode_flags,
1278 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1280 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1281 lru_cache_disable();
1283 NODEMASK_SCRATCH(scratch);
1285 mmap_write_lock(mm);
1286 err = mpol_set_nodemask(new, nmask, scratch);
1288 mmap_write_unlock(mm);
1291 NODEMASK_SCRATCH_FREE(scratch);
1297 * Lock the VMAs before scanning for pages to migrate,
1298 * to ensure we don't miss a concurrently inserted page.
1300 nr_failed = queue_pages_range(mm, start, end, nmask,
1301 flags | MPOL_MF_INVERT | MPOL_MF_WRLOCK, &pagelist);
1303 if (nr_failed < 0) {
1306 vma_iter_init(&vmi, mm, start);
1307 prev = vma_prev(&vmi);
1308 for_each_vma_range(vmi, vma, end) {
1309 err = mbind_range(&vmi, vma, &prev, start, end, new);
1316 if (!list_empty(&pagelist)) {
1317 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1318 nr_failed |= migrate_pages(&pagelist, new_folio, NULL,
1319 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1321 if (nr_failed && (flags & MPOL_MF_STRICT))
1325 if (!list_empty(&pagelist))
1326 putback_movable_pages(&pagelist);
1328 mmap_write_unlock(mm);
1331 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1337 * User space interface with variable sized bitmaps for nodelists.
1339 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1340 unsigned long maxnode)
1342 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1345 if (in_compat_syscall())
1346 ret = compat_get_bitmap(mask,
1347 (const compat_ulong_t __user *)nmask,
1350 ret = copy_from_user(mask, nmask,
1351 nlongs * sizeof(unsigned long));
1356 if (maxnode % BITS_PER_LONG)
1357 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1362 /* Copy a node mask from user space. */
1363 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1364 unsigned long maxnode)
1367 nodes_clear(*nodes);
1368 if (maxnode == 0 || !nmask)
1370 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1374 * When the user specified more nodes than supported just check
1375 * if the non supported part is all zero, one word at a time,
1376 * starting at the end.
1378 while (maxnode > MAX_NUMNODES) {
1379 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1382 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1385 if (maxnode - bits >= MAX_NUMNODES) {
1388 maxnode = MAX_NUMNODES;
1389 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1395 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1398 /* Copy a kernel node mask to user space */
1399 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1402 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1403 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1404 bool compat = in_compat_syscall();
1407 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1409 if (copy > nbytes) {
1410 if (copy > PAGE_SIZE)
1412 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1415 maxnode = nr_node_ids;
1419 return compat_put_bitmap((compat_ulong_t __user *)mask,
1420 nodes_addr(*nodes), maxnode);
1422 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1425 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1426 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1428 *flags = *mode & MPOL_MODE_FLAGS;
1429 *mode &= ~MPOL_MODE_FLAGS;
1431 if ((unsigned int)(*mode) >= MPOL_MAX)
1433 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1435 if (*flags & MPOL_F_NUMA_BALANCING) {
1436 if (*mode != MPOL_BIND)
1438 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1443 static long kernel_mbind(unsigned long start, unsigned long len,
1444 unsigned long mode, const unsigned long __user *nmask,
1445 unsigned long maxnode, unsigned int flags)
1447 unsigned short mode_flags;
1452 start = untagged_addr(start);
1453 err = sanitize_mpol_flags(&lmode, &mode_flags);
1457 err = get_nodes(&nodes, nmask, maxnode);
1461 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1464 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1465 unsigned long, home_node, unsigned long, flags)
1467 struct mm_struct *mm = current->mm;
1468 struct vm_area_struct *vma, *prev;
1469 struct mempolicy *new, *old;
1472 VMA_ITERATOR(vmi, mm, start);
1474 start = untagged_addr(start);
1475 if (start & ~PAGE_MASK)
1478 * flags is used for future extension if any.
1484 * Check home_node is online to avoid accessing uninitialized
1487 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1490 len = PAGE_ALIGN(len);
1497 mmap_write_lock(mm);
1498 prev = vma_prev(&vmi);
1499 for_each_vma_range(vmi, vma, end) {
1501 * If any vma in the range got policy other than MPOL_BIND
1502 * or MPOL_PREFERRED_MANY we return error. We don't reset
1503 * the home node for vmas we already updated before.
1505 old = vma_policy(vma);
1510 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1514 new = mpol_dup(old);
1520 vma_start_write(vma);
1521 new->home_node = home_node;
1522 err = mbind_range(&vmi, vma, &prev, start, end, new);
1527 mmap_write_unlock(mm);
1531 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1532 unsigned long, mode, const unsigned long __user *, nmask,
1533 unsigned long, maxnode, unsigned int, flags)
1535 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1538 /* Set the process memory policy */
1539 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1540 unsigned long maxnode)
1542 unsigned short mode_flags;
1547 err = sanitize_mpol_flags(&lmode, &mode_flags);
1551 err = get_nodes(&nodes, nmask, maxnode);
1555 return do_set_mempolicy(lmode, mode_flags, &nodes);
1558 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1559 unsigned long, maxnode)
1561 return kernel_set_mempolicy(mode, nmask, maxnode);
1564 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1565 const unsigned long __user *old_nodes,
1566 const unsigned long __user *new_nodes)
1568 struct mm_struct *mm = NULL;
1569 struct task_struct *task;
1570 nodemask_t task_nodes;
1574 NODEMASK_SCRATCH(scratch);
1579 old = &scratch->mask1;
1580 new = &scratch->mask2;
1582 err = get_nodes(old, old_nodes, maxnode);
1586 err = get_nodes(new, new_nodes, maxnode);
1590 /* Find the mm_struct */
1592 task = pid ? find_task_by_vpid(pid) : current;
1598 get_task_struct(task);
1603 * Check if this process has the right to modify the specified process.
1604 * Use the regular "ptrace_may_access()" checks.
1606 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1613 task_nodes = cpuset_mems_allowed(task);
1614 /* Is the user allowed to access the target nodes? */
1615 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1620 task_nodes = cpuset_mems_allowed(current);
1621 nodes_and(*new, *new, task_nodes);
1622 if (nodes_empty(*new))
1625 err = security_task_movememory(task);
1629 mm = get_task_mm(task);
1630 put_task_struct(task);
1637 err = do_migrate_pages(mm, old, new,
1638 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1642 NODEMASK_SCRATCH_FREE(scratch);
1647 put_task_struct(task);
1652 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1653 const unsigned long __user *, old_nodes,
1654 const unsigned long __user *, new_nodes)
1656 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1660 /* Retrieve NUMA policy */
1661 static int kernel_get_mempolicy(int __user *policy,
1662 unsigned long __user *nmask,
1663 unsigned long maxnode,
1665 unsigned long flags)
1671 if (nmask != NULL && maxnode < nr_node_ids)
1674 addr = untagged_addr(addr);
1676 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1681 if (policy && put_user(pval, policy))
1685 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1690 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1691 unsigned long __user *, nmask, unsigned long, maxnode,
1692 unsigned long, addr, unsigned long, flags)
1694 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1697 bool vma_migratable(struct vm_area_struct *vma)
1699 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1703 * DAX device mappings require predictable access latency, so avoid
1704 * incurring periodic faults.
1706 if (vma_is_dax(vma))
1709 if (is_vm_hugetlb_page(vma) &&
1710 !hugepage_migration_supported(hstate_vma(vma)))
1714 * Migration allocates pages in the highest zone. If we cannot
1715 * do so then migration (at least from node to node) is not
1719 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1725 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1728 struct mempolicy *pol = NULL;
1731 if (vma->vm_ops && vma->vm_ops->get_policy) {
1732 pol = vma->vm_ops->get_policy(vma, addr);
1733 } else if (vma->vm_policy) {
1734 pol = vma->vm_policy;
1737 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1738 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1739 * count on these policies which will be dropped by
1740 * mpol_cond_put() later
1742 if (mpol_needs_cond_ref(pol))
1751 * get_vma_policy(@vma, @addr)
1752 * @vma: virtual memory area whose policy is sought
1753 * @addr: address in @vma for shared policy lookup
1755 * Returns effective policy for a VMA at specified address.
1756 * Falls back to current->mempolicy or system default policy, as necessary.
1757 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1758 * count--added by the get_policy() vm_op, as appropriate--to protect against
1759 * freeing by another task. It is the caller's responsibility to free the
1760 * extra reference for shared policies.
1762 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1765 struct mempolicy *pol = __get_vma_policy(vma, addr);
1768 pol = get_task_policy(current);
1773 bool vma_policy_mof(struct vm_area_struct *vma)
1775 struct mempolicy *pol;
1777 if (vma->vm_ops && vma->vm_ops->get_policy) {
1780 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1781 if (pol && (pol->flags & MPOL_F_MOF))
1788 pol = vma->vm_policy;
1790 pol = get_task_policy(current);
1792 return pol->flags & MPOL_F_MOF;
1795 bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1797 enum zone_type dynamic_policy_zone = policy_zone;
1799 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1802 * if policy->nodes has movable memory only,
1803 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1805 * policy->nodes is intersect with node_states[N_MEMORY].
1806 * so if the following test fails, it implies
1807 * policy->nodes has movable memory only.
1809 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1810 dynamic_policy_zone = ZONE_MOVABLE;
1812 return zone >= dynamic_policy_zone;
1816 * Return a nodemask representing a mempolicy for filtering nodes for
1819 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1821 int mode = policy->mode;
1823 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1824 if (unlikely(mode == MPOL_BIND) &&
1825 apply_policy_zone(policy, gfp_zone(gfp)) &&
1826 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1827 return &policy->nodes;
1829 if (mode == MPOL_PREFERRED_MANY)
1830 return &policy->nodes;
1836 * Return the preferred node id for 'prefer' mempolicy, and return
1837 * the given id for all other policies.
1839 * policy_node() is always coupled with policy_nodemask(), which
1840 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1842 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1844 if (policy->mode == MPOL_PREFERRED) {
1845 nd = first_node(policy->nodes);
1848 * __GFP_THISNODE shouldn't even be used with the bind policy
1849 * because we might easily break the expectation to stay on the
1850 * requested node and not break the policy.
1852 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1855 if ((policy->mode == MPOL_BIND ||
1856 policy->mode == MPOL_PREFERRED_MANY) &&
1857 policy->home_node != NUMA_NO_NODE)
1858 return policy->home_node;
1863 /* Do dynamic interleaving for a process */
1864 static unsigned interleave_nodes(struct mempolicy *policy)
1867 struct task_struct *me = current;
1869 next = next_node_in(me->il_prev, policy->nodes);
1870 if (next < MAX_NUMNODES)
1876 * Depending on the memory policy provide a node from which to allocate the
1879 unsigned int mempolicy_slab_node(void)
1881 struct mempolicy *policy;
1882 int node = numa_mem_id();
1887 policy = current->mempolicy;
1891 switch (policy->mode) {
1892 case MPOL_PREFERRED:
1893 return first_node(policy->nodes);
1895 case MPOL_INTERLEAVE:
1896 return interleave_nodes(policy);
1899 case MPOL_PREFERRED_MANY:
1904 * Follow bind policy behavior and start allocation at the
1907 struct zonelist *zonelist;
1908 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1909 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1910 z = first_zones_zonelist(zonelist, highest_zoneidx,
1912 return z->zone ? zone_to_nid(z->zone) : node;
1923 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1924 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1925 * number of present nodes.
1927 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1929 nodemask_t nodemask = pol->nodes;
1930 unsigned int target, nnodes;
1934 * The barrier will stabilize the nodemask in a register or on
1935 * the stack so that it will stop changing under the code.
1937 * Between first_node() and next_node(), pol->nodes could be changed
1938 * by other threads. So we put pol->nodes in a local stack.
1942 nnodes = nodes_weight(nodemask);
1944 return numa_node_id();
1945 target = (unsigned int)n % nnodes;
1946 nid = first_node(nodemask);
1947 for (i = 0; i < target; i++)
1948 nid = next_node(nid, nodemask);
1952 /* Determine a node number for interleave */
1953 static inline unsigned interleave_nid(struct mempolicy *pol,
1954 struct vm_area_struct *vma, unsigned long addr, int shift)
1960 * for small pages, there is no difference between
1961 * shift and PAGE_SHIFT, so the bit-shift is safe.
1962 * for huge pages, since vm_pgoff is in units of small
1963 * pages, we need to shift off the always 0 bits to get
1966 BUG_ON(shift < PAGE_SHIFT);
1967 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1968 off += (addr - vma->vm_start) >> shift;
1969 return offset_il_node(pol, off);
1971 return interleave_nodes(pol);
1974 #ifdef CONFIG_HUGETLBFS
1976 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1977 * @vma: virtual memory area whose policy is sought
1978 * @addr: address in @vma for shared policy lookup and interleave policy
1979 * @gfp_flags: for requested zone
1980 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1981 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1983 * Returns a nid suitable for a huge page allocation and a pointer
1984 * to the struct mempolicy for conditional unref after allocation.
1985 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
1986 * to the mempolicy's @nodemask for filtering the zonelist.
1988 * Must be protected by read_mems_allowed_begin()
1990 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1991 struct mempolicy **mpol, nodemask_t **nodemask)
1996 *mpol = get_vma_policy(vma, addr);
1998 mode = (*mpol)->mode;
2000 if (unlikely(mode == MPOL_INTERLEAVE)) {
2001 nid = interleave_nid(*mpol, vma, addr,
2002 huge_page_shift(hstate_vma(vma)));
2004 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2005 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2006 *nodemask = &(*mpol)->nodes;
2012 * init_nodemask_of_mempolicy
2014 * If the current task's mempolicy is "default" [NULL], return 'false'
2015 * to indicate default policy. Otherwise, extract the policy nodemask
2016 * for 'bind' or 'interleave' policy into the argument nodemask, or
2017 * initialize the argument nodemask to contain the single node for
2018 * 'preferred' or 'local' policy and return 'true' to indicate presence
2019 * of non-default mempolicy.
2021 * We don't bother with reference counting the mempolicy [mpol_get/put]
2022 * because the current task is examining it's own mempolicy and a task's
2023 * mempolicy is only ever changed by the task itself.
2025 * N.B., it is the caller's responsibility to free a returned nodemask.
2027 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2029 struct mempolicy *mempolicy;
2031 if (!(mask && current->mempolicy))
2035 mempolicy = current->mempolicy;
2036 switch (mempolicy->mode) {
2037 case MPOL_PREFERRED:
2038 case MPOL_PREFERRED_MANY:
2040 case MPOL_INTERLEAVE:
2041 *mask = mempolicy->nodes;
2045 init_nodemask_of_node(mask, numa_node_id());
2051 task_unlock(current);
2058 * mempolicy_in_oom_domain
2060 * If tsk's mempolicy is "bind", check for intersection between mask and
2061 * the policy nodemask. Otherwise, return true for all other policies
2062 * including "interleave", as a tsk with "interleave" policy may have
2063 * memory allocated from all nodes in system.
2065 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2067 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2068 const nodemask_t *mask)
2070 struct mempolicy *mempolicy;
2077 mempolicy = tsk->mempolicy;
2078 if (mempolicy && mempolicy->mode == MPOL_BIND)
2079 ret = nodes_intersects(mempolicy->nodes, *mask);
2085 /* Allocate a page in interleaved policy.
2086 Own path because it needs to do special accounting. */
2087 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2092 page = __alloc_pages(gfp, order, nid, NULL);
2093 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2094 if (!static_branch_likely(&vm_numa_stat_key))
2096 if (page && page_to_nid(page) == nid) {
2098 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2104 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2105 int nid, struct mempolicy *pol)
2108 gfp_t preferred_gfp;
2111 * This is a two pass approach. The first pass will only try the
2112 * preferred nodes but skip the direct reclaim and allow the
2113 * allocation to fail, while the second pass will try all the
2116 preferred_gfp = gfp | __GFP_NOWARN;
2117 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2118 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2120 page = __alloc_pages(gfp, order, nid, NULL);
2126 * vma_alloc_folio - Allocate a folio for a VMA.
2128 * @order: Order of the folio.
2129 * @vma: Pointer to VMA or NULL if not available.
2130 * @addr: Virtual address of the allocation. Must be inside @vma.
2131 * @hugepage: For hugepages try only the preferred node if possible.
2133 * Allocate a folio for a specific address in @vma, using the appropriate
2134 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2135 * of the mm_struct of the VMA to prevent it from going away. Should be
2136 * used for all allocations for folios that will be mapped into user space.
2138 * Return: The folio on success or NULL if allocation fails.
2140 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2141 unsigned long addr, bool hugepage)
2143 struct mempolicy *pol;
2144 int node = numa_node_id();
2145 struct folio *folio;
2149 pol = get_vma_policy(vma, addr);
2151 if (pol->mode == MPOL_INTERLEAVE) {
2155 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2158 page = alloc_page_interleave(gfp, order, nid);
2159 folio = (struct folio *)page;
2160 if (folio && order > 1)
2161 folio_prep_large_rmappable(folio);
2165 if (pol->mode == MPOL_PREFERRED_MANY) {
2168 node = policy_node(gfp, pol, node);
2170 page = alloc_pages_preferred_many(gfp, order, node, pol);
2172 folio = (struct folio *)page;
2173 if (folio && order > 1)
2174 folio_prep_large_rmappable(folio);
2178 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2179 int hpage_node = node;
2182 * For hugepage allocation and non-interleave policy which
2183 * allows the current node (or other explicitly preferred
2184 * node) we only try to allocate from the current/preferred
2185 * node and don't fall back to other nodes, as the cost of
2186 * remote accesses would likely offset THP benefits.
2188 * If the policy is interleave or does not allow the current
2189 * node in its nodemask, we allocate the standard way.
2191 if (pol->mode == MPOL_PREFERRED)
2192 hpage_node = first_node(pol->nodes);
2194 nmask = policy_nodemask(gfp, pol);
2195 if (!nmask || node_isset(hpage_node, *nmask)) {
2198 * First, try to allocate THP only on local node, but
2199 * don't reclaim unnecessarily, just compact.
2201 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2202 __GFP_NORETRY, order, hpage_node);
2205 * If hugepage allocations are configured to always
2206 * synchronous compact or the vma has been madvised
2207 * to prefer hugepage backing, retry allowing remote
2208 * memory with both reclaim and compact as well.
2210 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2211 folio = __folio_alloc(gfp, order, hpage_node,
2218 nmask = policy_nodemask(gfp, pol);
2219 preferred_nid = policy_node(gfp, pol, node);
2220 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2225 EXPORT_SYMBOL(vma_alloc_folio);
2228 * alloc_pages - Allocate pages.
2230 * @order: Power of two of number of pages to allocate.
2232 * Allocate 1 << @order contiguous pages. The physical address of the
2233 * first page is naturally aligned (eg an order-3 allocation will be aligned
2234 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2235 * process is honoured when in process context.
2237 * Context: Can be called from any context, providing the appropriate GFP
2239 * Return: The page on success or NULL if allocation fails.
2241 struct page *alloc_pages(gfp_t gfp, unsigned order)
2243 struct mempolicy *pol = &default_policy;
2246 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2247 pol = get_task_policy(current);
2250 * No reference counting needed for current->mempolicy
2251 * nor system default_policy
2253 if (pol->mode == MPOL_INTERLEAVE)
2254 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2255 else if (pol->mode == MPOL_PREFERRED_MANY)
2256 page = alloc_pages_preferred_many(gfp, order,
2257 policy_node(gfp, pol, numa_node_id()), pol);
2259 page = __alloc_pages(gfp, order,
2260 policy_node(gfp, pol, numa_node_id()),
2261 policy_nodemask(gfp, pol));
2265 EXPORT_SYMBOL(alloc_pages);
2267 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2269 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2270 struct folio *folio = (struct folio *)page;
2272 if (folio && order > 1)
2273 folio_prep_large_rmappable(folio);
2276 EXPORT_SYMBOL(folio_alloc);
2278 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2279 struct mempolicy *pol, unsigned long nr_pages,
2280 struct page **page_array)
2283 unsigned long nr_pages_per_node;
2286 unsigned long nr_allocated;
2287 unsigned long total_allocated = 0;
2289 nodes = nodes_weight(pol->nodes);
2290 nr_pages_per_node = nr_pages / nodes;
2291 delta = nr_pages - nodes * nr_pages_per_node;
2293 for (i = 0; i < nodes; i++) {
2295 nr_allocated = __alloc_pages_bulk(gfp,
2296 interleave_nodes(pol), NULL,
2297 nr_pages_per_node + 1, NULL,
2301 nr_allocated = __alloc_pages_bulk(gfp,
2302 interleave_nodes(pol), NULL,
2303 nr_pages_per_node, NULL, page_array);
2306 page_array += nr_allocated;
2307 total_allocated += nr_allocated;
2310 return total_allocated;
2313 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2314 struct mempolicy *pol, unsigned long nr_pages,
2315 struct page **page_array)
2317 gfp_t preferred_gfp;
2318 unsigned long nr_allocated = 0;
2320 preferred_gfp = gfp | __GFP_NOWARN;
2321 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2323 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2324 nr_pages, NULL, page_array);
2326 if (nr_allocated < nr_pages)
2327 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2328 nr_pages - nr_allocated, NULL,
2329 page_array + nr_allocated);
2330 return nr_allocated;
2333 /* alloc pages bulk and mempolicy should be considered at the
2334 * same time in some situation such as vmalloc.
2336 * It can accelerate memory allocation especially interleaving
2339 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2340 unsigned long nr_pages, struct page **page_array)
2342 struct mempolicy *pol = &default_policy;
2344 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2345 pol = get_task_policy(current);
2347 if (pol->mode == MPOL_INTERLEAVE)
2348 return alloc_pages_bulk_array_interleave(gfp, pol,
2349 nr_pages, page_array);
2351 if (pol->mode == MPOL_PREFERRED_MANY)
2352 return alloc_pages_bulk_array_preferred_many(gfp,
2353 numa_node_id(), pol, nr_pages, page_array);
2355 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2356 policy_nodemask(gfp, pol), nr_pages, NULL,
2360 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2362 struct mempolicy *pol = mpol_dup(vma_policy(src));
2365 return PTR_ERR(pol);
2366 dst->vm_policy = pol;
2371 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2372 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2373 * with the mems_allowed returned by cpuset_mems_allowed(). This
2374 * keeps mempolicies cpuset relative after its cpuset moves. See
2375 * further kernel/cpuset.c update_nodemask().
2377 * current's mempolicy may be rebinded by the other task(the task that changes
2378 * cpuset's mems), so we needn't do rebind work for current task.
2381 /* Slow path of a mempolicy duplicate */
2382 struct mempolicy *__mpol_dup(struct mempolicy *old)
2384 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2387 return ERR_PTR(-ENOMEM);
2389 /* task's mempolicy is protected by alloc_lock */
2390 if (old == current->mempolicy) {
2393 task_unlock(current);
2397 if (current_cpuset_is_being_rebound()) {
2398 nodemask_t mems = cpuset_mems_allowed(current);
2399 mpol_rebind_policy(new, &mems);
2401 atomic_set(&new->refcnt, 1);
2405 /* Slow path of a mempolicy comparison */
2406 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2410 if (a->mode != b->mode)
2412 if (a->flags != b->flags)
2414 if (a->home_node != b->home_node)
2416 if (mpol_store_user_nodemask(a))
2417 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2422 case MPOL_INTERLEAVE:
2423 case MPOL_PREFERRED:
2424 case MPOL_PREFERRED_MANY:
2425 return !!nodes_equal(a->nodes, b->nodes);
2435 * Shared memory backing store policy support.
2437 * Remember policies even when nobody has shared memory mapped.
2438 * The policies are kept in Red-Black tree linked from the inode.
2439 * They are protected by the sp->lock rwlock, which should be held
2440 * for any accesses to the tree.
2444 * lookup first element intersecting start-end. Caller holds sp->lock for
2445 * reading or for writing
2447 static struct sp_node *
2448 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2450 struct rb_node *n = sp->root.rb_node;
2453 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2455 if (start >= p->end)
2457 else if (end <= p->start)
2465 struct sp_node *w = NULL;
2466 struct rb_node *prev = rb_prev(n);
2469 w = rb_entry(prev, struct sp_node, nd);
2470 if (w->end <= start)
2474 return rb_entry(n, struct sp_node, nd);
2478 * Insert a new shared policy into the list. Caller holds sp->lock for
2481 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2483 struct rb_node **p = &sp->root.rb_node;
2484 struct rb_node *parent = NULL;
2489 nd = rb_entry(parent, struct sp_node, nd);
2490 if (new->start < nd->start)
2492 else if (new->end > nd->end)
2493 p = &(*p)->rb_right;
2497 rb_link_node(&new->nd, parent, p);
2498 rb_insert_color(&new->nd, &sp->root);
2499 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2500 new->policy ? new->policy->mode : 0);
2503 /* Find shared policy intersecting idx */
2505 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2507 struct mempolicy *pol = NULL;
2510 if (!sp->root.rb_node)
2512 read_lock(&sp->lock);
2513 sn = sp_lookup(sp, idx, idx+1);
2515 mpol_get(sn->policy);
2518 read_unlock(&sp->lock);
2522 static void sp_free(struct sp_node *n)
2524 mpol_put(n->policy);
2525 kmem_cache_free(sn_cache, n);
2529 * mpol_misplaced - check whether current folio node is valid in policy
2531 * @folio: folio to be checked
2532 * @vma: vm area where folio mapped
2533 * @addr: virtual address in @vma for shared policy lookup and interleave policy
2535 * Lookup current policy node id for vma,addr and "compare to" folio's
2536 * node id. Policy determination "mimics" alloc_page_vma().
2537 * Called from fault path where we know the vma and faulting address.
2539 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2540 * policy, or a suitable node ID to allocate a replacement folio from.
2542 int mpol_misplaced(struct folio *folio, struct vm_area_struct *vma,
2545 struct mempolicy *pol;
2547 int curnid = folio_nid(folio);
2548 unsigned long pgoff;
2549 int thiscpu = raw_smp_processor_id();
2550 int thisnid = cpu_to_node(thiscpu);
2551 int polnid = NUMA_NO_NODE;
2552 int ret = NUMA_NO_NODE;
2554 pol = get_vma_policy(vma, addr);
2555 if (!(pol->flags & MPOL_F_MOF))
2558 switch (pol->mode) {
2559 case MPOL_INTERLEAVE:
2560 pgoff = vma->vm_pgoff;
2561 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2562 polnid = offset_il_node(pol, pgoff);
2565 case MPOL_PREFERRED:
2566 if (node_isset(curnid, pol->nodes))
2568 polnid = first_node(pol->nodes);
2572 polnid = numa_node_id();
2576 /* Optimize placement among multiple nodes via NUMA balancing */
2577 if (pol->flags & MPOL_F_MORON) {
2578 if (node_isset(thisnid, pol->nodes))
2584 case MPOL_PREFERRED_MANY:
2586 * use current page if in policy nodemask,
2587 * else select nearest allowed node, if any.
2588 * If no allowed nodes, use current [!misplaced].
2590 if (node_isset(curnid, pol->nodes))
2592 z = first_zones_zonelist(
2593 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2594 gfp_zone(GFP_HIGHUSER),
2596 polnid = zone_to_nid(z->zone);
2603 /* Migrate the folio towards the node whose CPU is referencing it */
2604 if (pol->flags & MPOL_F_MORON) {
2607 if (!should_numa_migrate_memory(current, folio, curnid,
2612 if (curnid != polnid)
2621 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2622 * dropped after task->mempolicy is set to NULL so that any allocation done as
2623 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2626 void mpol_put_task_policy(struct task_struct *task)
2628 struct mempolicy *pol;
2631 pol = task->mempolicy;
2632 task->mempolicy = NULL;
2637 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2639 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2640 rb_erase(&n->nd, &sp->root);
2644 static void sp_node_init(struct sp_node *node, unsigned long start,
2645 unsigned long end, struct mempolicy *pol)
2647 node->start = start;
2652 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2653 struct mempolicy *pol)
2656 struct mempolicy *newpol;
2658 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2662 newpol = mpol_dup(pol);
2663 if (IS_ERR(newpol)) {
2664 kmem_cache_free(sn_cache, n);
2667 newpol->flags |= MPOL_F_SHARED;
2668 sp_node_init(n, start, end, newpol);
2673 /* Replace a policy range. */
2674 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2675 unsigned long end, struct sp_node *new)
2678 struct sp_node *n_new = NULL;
2679 struct mempolicy *mpol_new = NULL;
2683 write_lock(&sp->lock);
2684 n = sp_lookup(sp, start, end);
2685 /* Take care of old policies in the same range. */
2686 while (n && n->start < end) {
2687 struct rb_node *next = rb_next(&n->nd);
2688 if (n->start >= start) {
2694 /* Old policy spanning whole new range. */
2699 *mpol_new = *n->policy;
2700 atomic_set(&mpol_new->refcnt, 1);
2701 sp_node_init(n_new, end, n->end, mpol_new);
2703 sp_insert(sp, n_new);
2712 n = rb_entry(next, struct sp_node, nd);
2716 write_unlock(&sp->lock);
2723 kmem_cache_free(sn_cache, n_new);
2728 write_unlock(&sp->lock);
2730 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2733 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2736 atomic_set(&mpol_new->refcnt, 1);
2741 * mpol_shared_policy_init - initialize shared policy for inode
2742 * @sp: pointer to inode shared policy
2743 * @mpol: struct mempolicy to install
2745 * Install non-NULL @mpol in inode's shared policy rb-tree.
2746 * On entry, the current task has a reference on a non-NULL @mpol.
2747 * This must be released on exit.
2748 * This is called at get_inode() calls and we can use GFP_KERNEL.
2750 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2754 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2755 rwlock_init(&sp->lock);
2758 struct vm_area_struct pvma;
2759 struct mempolicy *new;
2760 NODEMASK_SCRATCH(scratch);
2764 /* contextualize the tmpfs mount point mempolicy */
2765 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2767 goto free_scratch; /* no valid nodemask intersection */
2770 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2771 task_unlock(current);
2775 /* Create pseudo-vma that contains just the policy */
2776 vma_init(&pvma, NULL);
2777 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2778 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2781 mpol_put(new); /* drop initial ref */
2783 NODEMASK_SCRATCH_FREE(scratch);
2785 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2789 int mpol_set_shared_policy(struct shared_policy *info,
2790 struct vm_area_struct *vma, struct mempolicy *npol)
2793 struct sp_node *new = NULL;
2794 unsigned long sz = vma_pages(vma);
2796 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2798 sz, npol ? npol->mode : -1,
2799 npol ? npol->flags : -1,
2800 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2803 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2807 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2813 /* Free a backing policy store on inode delete. */
2814 void mpol_free_shared_policy(struct shared_policy *p)
2817 struct rb_node *next;
2819 if (!p->root.rb_node)
2821 write_lock(&p->lock);
2822 next = rb_first(&p->root);
2824 n = rb_entry(next, struct sp_node, nd);
2825 next = rb_next(&n->nd);
2828 write_unlock(&p->lock);
2831 #ifdef CONFIG_NUMA_BALANCING
2832 static int __initdata numabalancing_override;
2834 static void __init check_numabalancing_enable(void)
2836 bool numabalancing_default = false;
2838 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2839 numabalancing_default = true;
2841 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2842 if (numabalancing_override)
2843 set_numabalancing_state(numabalancing_override == 1);
2845 if (num_online_nodes() > 1 && !numabalancing_override) {
2846 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2847 numabalancing_default ? "Enabling" : "Disabling");
2848 set_numabalancing_state(numabalancing_default);
2852 static int __init setup_numabalancing(char *str)
2858 if (!strcmp(str, "enable")) {
2859 numabalancing_override = 1;
2861 } else if (!strcmp(str, "disable")) {
2862 numabalancing_override = -1;
2867 pr_warn("Unable to parse numa_balancing=\n");
2871 __setup("numa_balancing=", setup_numabalancing);
2873 static inline void __init check_numabalancing_enable(void)
2876 #endif /* CONFIG_NUMA_BALANCING */
2878 /* assumes fs == KERNEL_DS */
2879 void __init numa_policy_init(void)
2881 nodemask_t interleave_nodes;
2882 unsigned long largest = 0;
2883 int nid, prefer = 0;
2885 policy_cache = kmem_cache_create("numa_policy",
2886 sizeof(struct mempolicy),
2887 0, SLAB_PANIC, NULL);
2889 sn_cache = kmem_cache_create("shared_policy_node",
2890 sizeof(struct sp_node),
2891 0, SLAB_PANIC, NULL);
2893 for_each_node(nid) {
2894 preferred_node_policy[nid] = (struct mempolicy) {
2895 .refcnt = ATOMIC_INIT(1),
2896 .mode = MPOL_PREFERRED,
2897 .flags = MPOL_F_MOF | MPOL_F_MORON,
2898 .nodes = nodemask_of_node(nid),
2903 * Set interleaving policy for system init. Interleaving is only
2904 * enabled across suitably sized nodes (default is >= 16MB), or
2905 * fall back to the largest node if they're all smaller.
2907 nodes_clear(interleave_nodes);
2908 for_each_node_state(nid, N_MEMORY) {
2909 unsigned long total_pages = node_present_pages(nid);
2911 /* Preserve the largest node */
2912 if (largest < total_pages) {
2913 largest = total_pages;
2917 /* Interleave this node? */
2918 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2919 node_set(nid, interleave_nodes);
2922 /* All too small, use the largest */
2923 if (unlikely(nodes_empty(interleave_nodes)))
2924 node_set(prefer, interleave_nodes);
2926 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2927 pr_err("%s: interleaving failed\n", __func__);
2929 check_numabalancing_enable();
2932 /* Reset policy of current process to default */
2933 void numa_default_policy(void)
2935 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2939 * Parse and format mempolicy from/to strings
2942 static const char * const policy_modes[] =
2944 [MPOL_DEFAULT] = "default",
2945 [MPOL_PREFERRED] = "prefer",
2946 [MPOL_BIND] = "bind",
2947 [MPOL_INTERLEAVE] = "interleave",
2948 [MPOL_LOCAL] = "local",
2949 [MPOL_PREFERRED_MANY] = "prefer (many)",
2955 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2956 * @str: string containing mempolicy to parse
2957 * @mpol: pointer to struct mempolicy pointer, returned on success.
2960 * <mode>[=<flags>][:<nodelist>]
2962 * Return: %0 on success, else %1
2964 int mpol_parse_str(char *str, struct mempolicy **mpol)
2966 struct mempolicy *new = NULL;
2967 unsigned short mode_flags;
2969 char *nodelist = strchr(str, ':');
2970 char *flags = strchr(str, '=');
2974 *flags++ = '\0'; /* terminate mode string */
2977 /* NUL-terminate mode or flags string */
2979 if (nodelist_parse(nodelist, nodes))
2981 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2986 mode = match_string(policy_modes, MPOL_MAX, str);
2991 case MPOL_PREFERRED:
2993 * Insist on a nodelist of one node only, although later
2994 * we use first_node(nodes) to grab a single node, so here
2995 * nodelist (or nodes) cannot be empty.
2998 char *rest = nodelist;
2999 while (isdigit(*rest))
3003 if (nodes_empty(nodes))
3007 case MPOL_INTERLEAVE:
3009 * Default to online nodes with memory if no nodelist
3012 nodes = node_states[N_MEMORY];
3016 * Don't allow a nodelist; mpol_new() checks flags
3023 * Insist on a empty nodelist
3028 case MPOL_PREFERRED_MANY:
3031 * Insist on a nodelist
3040 * Currently, we only support two mutually exclusive
3043 if (!strcmp(flags, "static"))
3044 mode_flags |= MPOL_F_STATIC_NODES;
3045 else if (!strcmp(flags, "relative"))
3046 mode_flags |= MPOL_F_RELATIVE_NODES;
3051 new = mpol_new(mode, mode_flags, &nodes);
3056 * Save nodes for mpol_to_str() to show the tmpfs mount options
3057 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3059 if (mode != MPOL_PREFERRED) {
3061 } else if (nodelist) {
3062 nodes_clear(new->nodes);
3063 node_set(first_node(nodes), new->nodes);
3065 new->mode = MPOL_LOCAL;
3069 * Save nodes for contextualization: this will be used to "clone"
3070 * the mempolicy in a specific context [cpuset] at a later time.
3072 new->w.user_nodemask = nodes;
3077 /* Restore string for error message */
3086 #endif /* CONFIG_TMPFS */
3089 * mpol_to_str - format a mempolicy structure for printing
3090 * @buffer: to contain formatted mempolicy string
3091 * @maxlen: length of @buffer
3092 * @pol: pointer to mempolicy to be formatted
3094 * Convert @pol into a string. If @buffer is too short, truncate the string.
3095 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3096 * longest flag, "relative", and to display at least a few node ids.
3098 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3101 nodemask_t nodes = NODE_MASK_NONE;
3102 unsigned short mode = MPOL_DEFAULT;
3103 unsigned short flags = 0;
3105 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3114 case MPOL_PREFERRED:
3115 case MPOL_PREFERRED_MANY:
3117 case MPOL_INTERLEAVE:
3122 snprintf(p, maxlen, "unknown");
3126 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3128 if (flags & MPOL_MODE_FLAGS) {
3129 p += snprintf(p, buffer + maxlen - p, "=");
3132 * Currently, the only defined flags are mutually exclusive
3134 if (flags & MPOL_F_STATIC_NODES)
3135 p += snprintf(p, buffer + maxlen - p, "static");
3136 else if (flags & MPOL_F_RELATIVE_NODES)
3137 p += snprintf(p, buffer + maxlen - p, "relative");
3140 if (!nodes_empty(nodes))
3141 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3142 nodemask_pr_args(&nodes));