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 */
116 static struct kmem_cache *policy_cache;
117 static struct kmem_cache *sn_cache;
119 /* Highest zone. An specific allocation for a zone below that is not
121 enum zone_type policy_zone = 0;
124 * run-time system-wide default policy => local allocation
126 static struct mempolicy default_policy = {
127 .refcnt = ATOMIC_INIT(1), /* never free it */
131 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
134 * numa_map_to_online_node - Find closest online node
135 * @node: Node id to start the search
137 * Lookup the next closest node by distance if @nid is not online.
139 * Return: this @node if it is online, otherwise the closest node by distance
141 int numa_map_to_online_node(int node)
143 int min_dist = INT_MAX, dist, n, min_node;
145 if (node == NUMA_NO_NODE || node_online(node))
149 for_each_online_node(n) {
150 dist = node_distance(node, n);
151 if (dist < min_dist) {
159 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
161 struct mempolicy *get_task_policy(struct task_struct *p)
163 struct mempolicy *pol = p->mempolicy;
169 node = numa_node_id();
170 if (node != NUMA_NO_NODE) {
171 pol = &preferred_node_policy[node];
172 /* preferred_node_policy is not initialised early in boot */
177 return &default_policy;
180 static const struct mempolicy_operations {
181 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
182 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
183 } mpol_ops[MPOL_MAX];
185 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
187 return pol->flags & MPOL_MODE_FLAGS;
190 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
191 const nodemask_t *rel)
194 nodes_fold(tmp, *orig, nodes_weight(*rel));
195 nodes_onto(*ret, tmp, *rel);
198 static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
200 if (nodes_empty(*nodes))
206 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
208 if (nodes_empty(*nodes))
211 nodes_clear(pol->nodes);
212 node_set(first_node(*nodes), pol->nodes);
217 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
218 * any, for the new policy. mpol_new() has already validated the nodes
219 * parameter with respect to the policy mode and flags.
221 * Must be called holding task's alloc_lock to protect task's mems_allowed
222 * and mempolicy. May also be called holding the mmap_lock for write.
224 static int mpol_set_nodemask(struct mempolicy *pol,
225 const nodemask_t *nodes, struct nodemask_scratch *nsc)
230 * Default (pol==NULL) resp. local memory policies are not a
231 * subject of any remapping. They also do not need any special
234 if (!pol || pol->mode == MPOL_LOCAL)
238 nodes_and(nsc->mask1,
239 cpuset_current_mems_allowed, node_states[N_MEMORY]);
243 if (pol->flags & MPOL_F_RELATIVE_NODES)
244 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
246 nodes_and(nsc->mask2, *nodes, nsc->mask1);
248 if (mpol_store_user_nodemask(pol))
249 pol->w.user_nodemask = *nodes;
251 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
253 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
258 * This function just creates a new policy, does some check and simple
259 * initialization. You must invoke mpol_set_nodemask() to set nodes.
261 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
264 struct mempolicy *policy;
266 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
267 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
269 if (mode == MPOL_DEFAULT) {
270 if (nodes && !nodes_empty(*nodes))
271 return ERR_PTR(-EINVAL);
277 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
278 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
279 * All other modes require a valid pointer to a non-empty nodemask.
281 if (mode == MPOL_PREFERRED) {
282 if (nodes_empty(*nodes)) {
283 if (((flags & MPOL_F_STATIC_NODES) ||
284 (flags & MPOL_F_RELATIVE_NODES)))
285 return ERR_PTR(-EINVAL);
289 } else if (mode == MPOL_LOCAL) {
290 if (!nodes_empty(*nodes) ||
291 (flags & MPOL_F_STATIC_NODES) ||
292 (flags & MPOL_F_RELATIVE_NODES))
293 return ERR_PTR(-EINVAL);
294 } else if (nodes_empty(*nodes))
295 return ERR_PTR(-EINVAL);
296 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
298 return ERR_PTR(-ENOMEM);
299 atomic_set(&policy->refcnt, 1);
301 policy->flags = flags;
302 policy->home_node = NUMA_NO_NODE;
307 /* Slow path of a mpol destructor. */
308 void __mpol_put(struct mempolicy *p)
310 if (!atomic_dec_and_test(&p->refcnt))
312 kmem_cache_free(policy_cache, p);
315 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
319 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
323 if (pol->flags & MPOL_F_STATIC_NODES)
324 nodes_and(tmp, pol->w.user_nodemask, *nodes);
325 else if (pol->flags & MPOL_F_RELATIVE_NODES)
326 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
328 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
330 pol->w.cpuset_mems_allowed = *nodes;
333 if (nodes_empty(tmp))
339 static void mpol_rebind_preferred(struct mempolicy *pol,
340 const nodemask_t *nodes)
342 pol->w.cpuset_mems_allowed = *nodes;
346 * mpol_rebind_policy - Migrate a policy to a different set of nodes
348 * Per-vma policies are protected by mmap_lock. Allocations using per-task
349 * policies are protected by task->mems_allowed_seq to prevent a premature
350 * OOM/allocation failure due to parallel nodemask modification.
352 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
354 if (!pol || pol->mode == MPOL_LOCAL)
356 if (!mpol_store_user_nodemask(pol) &&
357 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
360 mpol_ops[pol->mode].rebind(pol, newmask);
364 * Wrapper for mpol_rebind_policy() that just requires task
365 * pointer, and updates task mempolicy.
367 * Called with task's alloc_lock held.
370 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
372 mpol_rebind_policy(tsk->mempolicy, new);
376 * Rebind each vma in mm to new nodemask.
378 * Call holding a reference to mm. Takes mm->mmap_lock during call.
381 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
383 struct vm_area_struct *vma;
384 VMA_ITERATOR(vmi, mm, 0);
387 for_each_vma(vmi, vma) {
388 vma_start_write(vma);
389 mpol_rebind_policy(vma->vm_policy, new);
391 mmap_write_unlock(mm);
394 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
396 .rebind = mpol_rebind_default,
398 [MPOL_INTERLEAVE] = {
399 .create = mpol_new_nodemask,
400 .rebind = mpol_rebind_nodemask,
403 .create = mpol_new_preferred,
404 .rebind = mpol_rebind_preferred,
407 .create = mpol_new_nodemask,
408 .rebind = mpol_rebind_nodemask,
411 .rebind = mpol_rebind_default,
413 [MPOL_PREFERRED_MANY] = {
414 .create = mpol_new_nodemask,
415 .rebind = mpol_rebind_preferred,
419 static int migrate_folio_add(struct folio *folio, struct list_head *foliolist,
420 unsigned long flags);
423 struct list_head *pagelist;
428 struct vm_area_struct *first;
433 * Check if the folio's nid is in qp->nmask.
435 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
436 * in the invert of qp->nmask.
438 static inline bool queue_folio_required(struct folio *folio,
439 struct queue_pages *qp)
441 int nid = folio_nid(folio);
442 unsigned long flags = qp->flags;
444 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
448 * queue_folios_pmd() has three possible return values:
449 * 0 - folios are placed on the right node or queued successfully, or
450 * special page is met, i.e. zero page, or unmovable page is found
451 * but continue walking (indicated by queue_pages.has_unmovable).
452 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
453 * existing folio was already on a node that does not follow the
456 static int queue_folios_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
457 unsigned long end, struct mm_walk *walk)
462 struct queue_pages *qp = walk->private;
465 if (unlikely(is_pmd_migration_entry(*pmd))) {
469 folio = pfn_folio(pmd_pfn(*pmd));
470 if (is_huge_zero_page(&folio->page)) {
471 walk->action = ACTION_CONTINUE;
474 if (!queue_folio_required(folio, qp))
478 /* go to folio migration */
479 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
480 if (!vma_migratable(walk->vma) ||
481 migrate_folio_add(folio, qp->pagelist, flags)) {
482 qp->has_unmovable = true;
493 * Scan through pages checking if pages follow certain conditions,
494 * and move them to the pagelist if they do.
496 * queue_folios_pte_range() has three possible return values:
497 * 0 - folios are placed on the right node or queued successfully, or
498 * special page is met, i.e. zero page, or unmovable page is found
499 * but continue walking (indicated by queue_pages.has_unmovable).
500 * -EIO - only MPOL_MF_STRICT was specified and an existing folio was already
501 * on a node that does not follow the policy.
503 static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
504 unsigned long end, struct mm_walk *walk)
506 struct vm_area_struct *vma = walk->vma;
508 struct queue_pages *qp = walk->private;
509 unsigned long flags = qp->flags;
510 pte_t *pte, *mapped_pte;
514 ptl = pmd_trans_huge_lock(pmd, vma);
516 return queue_folios_pmd(pmd, ptl, addr, end, walk);
518 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
520 walk->action = ACTION_AGAIN;
523 for (; addr != end; pte++, addr += PAGE_SIZE) {
524 ptent = ptep_get(pte);
525 if (!pte_present(ptent))
527 folio = vm_normal_folio(vma, addr, ptent);
528 if (!folio || folio_is_zone_device(folio))
531 * vm_normal_folio() filters out zero pages, but there might
532 * still be reserved folios to skip, perhaps in a VDSO.
534 if (folio_test_reserved(folio))
536 if (!queue_folio_required(folio, qp))
538 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
540 * MPOL_MF_STRICT must be specified if we get here.
541 * Continue walking vmas due to MPOL_MF_MOVE* flags.
543 if (!vma_migratable(vma))
544 qp->has_unmovable = true;
547 * Do not abort immediately since there may be
548 * temporary off LRU pages in the range. Still
549 * need migrate other LRU pages.
551 if (migrate_folio_add(folio, qp->pagelist, flags))
552 qp->has_unmovable = true;
556 pte_unmap_unlock(mapped_pte, ptl);
559 return addr != end ? -EIO : 0;
562 static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
563 unsigned long addr, unsigned long end,
564 struct mm_walk *walk)
567 #ifdef CONFIG_HUGETLB_PAGE
568 struct queue_pages *qp = walk->private;
569 unsigned long flags = (qp->flags & MPOL_MF_VALID);
574 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
575 entry = huge_ptep_get(pte);
576 if (!pte_present(entry))
578 folio = pfn_folio(pte_pfn(entry));
579 if (!queue_folio_required(folio, qp))
582 if (flags == MPOL_MF_STRICT) {
584 * STRICT alone means only detecting misplaced folio and no
585 * need to further check other vma.
591 if (!vma_migratable(walk->vma)) {
593 * Must be STRICT with MOVE*, otherwise .test_walk() have
594 * stopped walking current vma.
595 * Detecting misplaced folio but allow migrating folios which
598 qp->has_unmovable = true;
603 * With MPOL_MF_MOVE, we try to migrate only unshared folios. If it
604 * is shared it is likely not worth migrating.
606 * To check if the folio is shared, ideally we want to make sure
607 * every page is mapped to the same process. Doing that is very
608 * expensive, so check the estimated mapcount of the folio instead.
610 if (flags & (MPOL_MF_MOVE_ALL) ||
611 (flags & MPOL_MF_MOVE && folio_estimated_sharers(folio) == 1 &&
612 !hugetlb_pmd_shared(pte))) {
613 if (!isolate_hugetlb(folio, qp->pagelist) &&
614 (flags & MPOL_MF_STRICT))
616 * Failed to isolate folio but allow migrating pages
617 * which have been queued.
619 qp->has_unmovable = true;
629 #ifdef CONFIG_NUMA_BALANCING
631 * This is used to mark a range of virtual addresses to be inaccessible.
632 * These are later cleared by a NUMA hinting fault. Depending on these
633 * faults, pages may be migrated for better NUMA placement.
635 * This is assuming that NUMA faults are handled using PROT_NONE. If
636 * an architecture makes a different choice, it will need further
637 * changes to the core.
639 unsigned long change_prot_numa(struct vm_area_struct *vma,
640 unsigned long addr, unsigned long end)
642 struct mmu_gather tlb;
645 tlb_gather_mmu(&tlb, vma->vm_mm);
647 nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
649 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
651 tlb_finish_mmu(&tlb);
656 static unsigned long change_prot_numa(struct vm_area_struct *vma,
657 unsigned long addr, unsigned long end)
661 #endif /* CONFIG_NUMA_BALANCING */
663 static int queue_pages_test_walk(unsigned long start, unsigned long end,
664 struct mm_walk *walk)
666 struct vm_area_struct *next, *vma = walk->vma;
667 struct queue_pages *qp = walk->private;
668 unsigned long endvma = vma->vm_end;
669 unsigned long flags = qp->flags;
671 /* range check first */
672 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
676 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
677 (qp->start < vma->vm_start))
678 /* hole at head side of range */
681 next = find_vma(vma->vm_mm, vma->vm_end);
682 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
683 ((vma->vm_end < qp->end) &&
684 (!next || vma->vm_end < next->vm_start)))
685 /* hole at middle or tail of range */
689 * Need check MPOL_MF_STRICT to return -EIO if possible
690 * regardless of vma_migratable
692 if (!vma_migratable(vma) &&
693 !(flags & MPOL_MF_STRICT))
699 if (flags & MPOL_MF_LAZY) {
700 /* Similar to task_numa_work, skip inaccessible VMAs */
701 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
702 !(vma->vm_flags & VM_MIXEDMAP))
703 change_prot_numa(vma, start, endvma);
707 /* queue pages from current vma */
708 if (flags & MPOL_MF_VALID)
713 static const struct mm_walk_ops queue_pages_walk_ops = {
714 .hugetlb_entry = queue_folios_hugetlb,
715 .pmd_entry = queue_folios_pte_range,
716 .test_walk = queue_pages_test_walk,
717 .walk_lock = PGWALK_RDLOCK,
720 static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
721 .hugetlb_entry = queue_folios_hugetlb,
722 .pmd_entry = queue_folios_pte_range,
723 .test_walk = queue_pages_test_walk,
724 .walk_lock = PGWALK_WRLOCK,
728 * Walk through page tables and collect pages to be migrated.
730 * If pages found in a given range are on a set of nodes (determined by
731 * @nodes and @flags,) it's isolated and queued to the pagelist which is
732 * passed via @private.
734 * queue_pages_range() has three possible return values:
735 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
737 * 0 - queue pages successfully or no misplaced page.
738 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
739 * memory range specified by nodemask and maxnode points outside
740 * your accessible address space (-EFAULT)
743 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
744 nodemask_t *nodes, unsigned long flags,
745 struct list_head *pagelist, bool lock_vma)
748 struct queue_pages qp = {
749 .pagelist = pagelist,
755 .has_unmovable = false,
757 const struct mm_walk_ops *ops = lock_vma ?
758 &queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
760 err = walk_page_range(mm, start, end, ops, &qp);
762 if (qp.has_unmovable)
765 /* whole range in hole */
772 * Apply policy to a single VMA
773 * This must be called with the mmap_lock held for writing.
775 static int vma_replace_policy(struct vm_area_struct *vma,
776 struct mempolicy *pol)
779 struct mempolicy *old;
780 struct mempolicy *new;
782 vma_assert_write_locked(vma);
784 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
785 vma->vm_start, vma->vm_end, vma->vm_pgoff,
786 vma->vm_ops, vma->vm_file,
787 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
793 if (vma->vm_ops && vma->vm_ops->set_policy) {
794 err = vma->vm_ops->set_policy(vma, new);
799 old = vma->vm_policy;
800 vma->vm_policy = new; /* protected by mmap_lock */
809 /* Split or merge the VMA (if required) and apply the new policy */
810 static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
811 struct vm_area_struct **prev, unsigned long start,
812 unsigned long end, struct mempolicy *new_pol)
814 unsigned long vmstart, vmend;
816 vmend = min(end, vma->vm_end);
817 if (start > vma->vm_start) {
821 vmstart = vma->vm_start;
824 if (mpol_equal(vma_policy(vma), new_pol)) {
829 vma = vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
834 return 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);
855 ret = mpol_set_nodemask(new, nodes, scratch);
857 task_unlock(current);
862 old = current->mempolicy;
863 current->mempolicy = new;
864 if (new && new->mode == MPOL_INTERLEAVE)
865 current->il_prev = MAX_NUMNODES-1;
866 task_unlock(current);
870 NODEMASK_SCRATCH_FREE(scratch);
875 * Return nodemask for policy for get_mempolicy() query
877 * Called with task's alloc_lock held
879 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
882 if (p == &default_policy)
887 case MPOL_INTERLEAVE:
889 case MPOL_PREFERRED_MANY:
893 /* return empty node mask for local allocation */
900 static int lookup_node(struct mm_struct *mm, unsigned long addr)
902 struct page *p = NULL;
905 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
907 ret = page_to_nid(p);
913 /* Retrieve NUMA policy */
914 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
915 unsigned long addr, unsigned long flags)
918 struct mm_struct *mm = current->mm;
919 struct vm_area_struct *vma = NULL;
920 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
923 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
926 if (flags & MPOL_F_MEMS_ALLOWED) {
927 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
929 *policy = 0; /* just so it's initialized */
931 *nmask = cpuset_current_mems_allowed;
932 task_unlock(current);
936 if (flags & MPOL_F_ADDR) {
938 * Do NOT fall back to task policy if the
939 * vma/shared policy at addr is NULL. We
940 * want to return MPOL_DEFAULT in this case.
943 vma = vma_lookup(mm, addr);
945 mmap_read_unlock(mm);
948 if (vma->vm_ops && vma->vm_ops->get_policy)
949 pol = vma->vm_ops->get_policy(vma, addr);
951 pol = vma->vm_policy;
956 pol = &default_policy; /* indicates default behavior */
958 if (flags & MPOL_F_NODE) {
959 if (flags & MPOL_F_ADDR) {
961 * Take a refcount on the mpol, because we are about to
962 * drop the mmap_lock, after which only "pol" remains
963 * valid, "vma" is stale.
968 mmap_read_unlock(mm);
969 err = lookup_node(mm, addr);
973 } else if (pol == current->mempolicy &&
974 pol->mode == MPOL_INTERLEAVE) {
975 *policy = next_node_in(current->il_prev, pol->nodes);
981 *policy = pol == &default_policy ? MPOL_DEFAULT :
984 * Internal mempolicy flags must be masked off before exposing
985 * the policy to userspace.
987 *policy |= (pol->flags & MPOL_MODE_FLAGS);
992 if (mpol_store_user_nodemask(pol)) {
993 *nmask = pol->w.user_nodemask;
996 get_policy_nodemask(pol, nmask);
997 task_unlock(current);
1004 mmap_read_unlock(mm);
1006 mpol_put(pol_refcount);
1010 #ifdef CONFIG_MIGRATION
1011 static int migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1012 unsigned long flags)
1015 * We try to migrate only unshared folios. If it is shared it
1016 * is likely not worth migrating.
1018 * To check if the folio is shared, ideally we want to make sure
1019 * every page is mapped to the same process. Doing that is very
1020 * expensive, so check the estimated mapcount of the folio instead.
1022 if ((flags & MPOL_MF_MOVE_ALL) || folio_estimated_sharers(folio) == 1) {
1023 if (folio_isolate_lru(folio)) {
1024 list_add_tail(&folio->lru, foliolist);
1025 node_stat_mod_folio(folio,
1026 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1027 folio_nr_pages(folio));
1028 } else if (flags & MPOL_MF_STRICT) {
1030 * Non-movable folio may reach here. And, there may be
1031 * temporary off LRU folios or non-LRU movable folios.
1032 * Treat them as unmovable folios 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 struct vm_area_struct *vma;
1052 LIST_HEAD(pagelist);
1054 struct migration_target_control mtc = {
1056 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1060 node_set(source, nmask);
1063 * This does not "check" the range but isolates all pages that
1064 * need migration. Between passing in the full user address
1065 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1067 vma = find_vma(mm, 0);
1068 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1069 queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1070 flags | MPOL_MF_DISCONTIG_OK, &pagelist, false);
1072 if (!list_empty(&pagelist)) {
1073 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1074 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1076 putback_movable_pages(&pagelist);
1083 * Move pages between the two nodesets so as to preserve the physical
1084 * layout as much as possible.
1086 * Returns the number of page that could not be moved.
1088 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1089 const nodemask_t *to, int flags)
1095 lru_cache_disable();
1100 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1101 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1102 * bit in 'tmp', and return that <source, dest> pair for migration.
1103 * The pair of nodemasks 'to' and 'from' define the map.
1105 * If no pair of bits is found that way, fallback to picking some
1106 * pair of 'source' and 'dest' bits that are not the same. If the
1107 * 'source' and 'dest' bits are the same, this represents a node
1108 * that will be migrating to itself, so no pages need move.
1110 * If no bits are left in 'tmp', or if all remaining bits left
1111 * in 'tmp' correspond to the same bit in 'to', return false
1112 * (nothing left to migrate).
1114 * This lets us pick a pair of nodes to migrate between, such that
1115 * if possible the dest node is not already occupied by some other
1116 * source node, minimizing the risk of overloading the memory on a
1117 * node that would happen if we migrated incoming memory to a node
1118 * before migrating outgoing memory source that same node.
1120 * A single scan of tmp is sufficient. As we go, we remember the
1121 * most recent <s, d> pair that moved (s != d). If we find a pair
1122 * that not only moved, but what's better, moved to an empty slot
1123 * (d is not set in tmp), then we break out then, with that pair.
1124 * Otherwise when we finish scanning from_tmp, we at least have the
1125 * most recent <s, d> pair that moved. If we get all the way through
1126 * the scan of tmp without finding any node that moved, much less
1127 * moved to an empty node, then there is nothing left worth migrating.
1131 while (!nodes_empty(tmp)) {
1133 int source = NUMA_NO_NODE;
1136 for_each_node_mask(s, tmp) {
1139 * do_migrate_pages() tries to maintain the relative
1140 * node relationship of the pages established between
1141 * threads and memory areas.
1143 * However if the number of source nodes is not equal to
1144 * the number of destination nodes we can not preserve
1145 * this node relative relationship. In that case, skip
1146 * copying memory from a node that is in the destination
1149 * Example: [2,3,4] -> [3,4,5] moves everything.
1150 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1153 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1154 (node_isset(s, *to)))
1157 d = node_remap(s, *from, *to);
1161 source = s; /* Node moved. Memorize */
1164 /* dest not in remaining from nodes? */
1165 if (!node_isset(dest, tmp))
1168 if (source == NUMA_NO_NODE)
1171 node_clear(source, tmp);
1172 err = migrate_to_node(mm, source, dest, flags);
1178 mmap_read_unlock(mm);
1188 * Allocate a new page for page migration based on vma policy.
1189 * Start by assuming the page is mapped by the same vma as contains @start.
1190 * Search forward from there, if not. N.B., this assumes that the
1191 * list of pages handed to migrate_pages()--which is how we get here--
1192 * is in virtual address order.
1194 static struct folio *new_folio(struct folio *src, unsigned long start)
1196 struct vm_area_struct *vma;
1197 unsigned long address;
1198 VMA_ITERATOR(vmi, current->mm, start);
1199 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1201 for_each_vma(vmi, vma) {
1202 address = page_address_in_vma(&src->page, vma);
1203 if (address != -EFAULT)
1207 if (folio_test_hugetlb(src)) {
1208 return alloc_hugetlb_folio_vma(folio_hstate(src),
1212 if (folio_test_large(src))
1213 gfp = GFP_TRANSHUGE;
1216 * if !vma, vma_alloc_folio() will use task or system default policy
1218 return vma_alloc_folio(gfp, folio_order(src), vma, address,
1219 folio_test_large(src));
1223 static int migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1224 unsigned long flags)
1229 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1230 const nodemask_t *to, int flags)
1235 static struct folio *new_folio(struct folio *src, unsigned long start)
1241 static long do_mbind(unsigned long start, unsigned long len,
1242 unsigned short mode, unsigned short mode_flags,
1243 nodemask_t *nmask, unsigned long flags)
1245 struct mm_struct *mm = current->mm;
1246 struct vm_area_struct *vma, *prev;
1247 struct vma_iterator vmi;
1248 struct mempolicy *new;
1252 LIST_HEAD(pagelist);
1254 if (flags & ~(unsigned long)MPOL_MF_VALID)
1256 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1259 if (start & ~PAGE_MASK)
1262 if (mode == MPOL_DEFAULT)
1263 flags &= ~MPOL_MF_STRICT;
1265 len = PAGE_ALIGN(len);
1273 new = mpol_new(mode, mode_flags, nmask);
1275 return PTR_ERR(new);
1277 if (flags & MPOL_MF_LAZY)
1278 new->flags |= MPOL_F_MOF;
1281 * If we are using the default policy then operation
1282 * on discontinuous address spaces is okay after all
1285 flags |= MPOL_MF_DISCONTIG_OK;
1287 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1288 start, start + len, mode, mode_flags,
1289 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1291 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1293 lru_cache_disable();
1296 NODEMASK_SCRATCH(scratch);
1298 mmap_write_lock(mm);
1299 err = mpol_set_nodemask(new, nmask, scratch);
1301 mmap_write_unlock(mm);
1304 NODEMASK_SCRATCH_FREE(scratch);
1310 * Lock the VMAs before scanning for pages to migrate, to ensure we don't
1311 * miss a concurrently inserted page.
1313 ret = queue_pages_range(mm, start, end, nmask,
1314 flags | MPOL_MF_INVERT, &pagelist, true);
1321 vma_iter_init(&vmi, mm, start);
1322 prev = vma_prev(&vmi);
1323 for_each_vma_range(vmi, vma, end) {
1324 err = mbind_range(&vmi, vma, &prev, start, end, new);
1332 if (!list_empty(&pagelist)) {
1333 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1334 nr_failed = migrate_pages(&pagelist, new_folio, NULL,
1335 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1337 putback_movable_pages(&pagelist);
1340 if (((ret > 0) || nr_failed) && (flags & MPOL_MF_STRICT))
1344 if (!list_empty(&pagelist))
1345 putback_movable_pages(&pagelist);
1348 mmap_write_unlock(mm);
1351 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1357 * User space interface with variable sized bitmaps for nodelists.
1359 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1360 unsigned long maxnode)
1362 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1365 if (in_compat_syscall())
1366 ret = compat_get_bitmap(mask,
1367 (const compat_ulong_t __user *)nmask,
1370 ret = copy_from_user(mask, nmask,
1371 nlongs * sizeof(unsigned long));
1376 if (maxnode % BITS_PER_LONG)
1377 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1382 /* Copy a node mask from user space. */
1383 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1384 unsigned long maxnode)
1387 nodes_clear(*nodes);
1388 if (maxnode == 0 || !nmask)
1390 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1394 * When the user specified more nodes than supported just check
1395 * if the non supported part is all zero, one word at a time,
1396 * starting at the end.
1398 while (maxnode > MAX_NUMNODES) {
1399 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1402 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1405 if (maxnode - bits >= MAX_NUMNODES) {
1408 maxnode = MAX_NUMNODES;
1409 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1415 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1418 /* Copy a kernel node mask to user space */
1419 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1422 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1423 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1424 bool compat = in_compat_syscall();
1427 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1429 if (copy > nbytes) {
1430 if (copy > PAGE_SIZE)
1432 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1435 maxnode = nr_node_ids;
1439 return compat_put_bitmap((compat_ulong_t __user *)mask,
1440 nodes_addr(*nodes), maxnode);
1442 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1445 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1446 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1448 *flags = *mode & MPOL_MODE_FLAGS;
1449 *mode &= ~MPOL_MODE_FLAGS;
1451 if ((unsigned int)(*mode) >= MPOL_MAX)
1453 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1455 if (*flags & MPOL_F_NUMA_BALANCING) {
1456 if (*mode != MPOL_BIND)
1458 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1463 static long kernel_mbind(unsigned long start, unsigned long len,
1464 unsigned long mode, const unsigned long __user *nmask,
1465 unsigned long maxnode, unsigned int flags)
1467 unsigned short mode_flags;
1472 start = untagged_addr(start);
1473 err = sanitize_mpol_flags(&lmode, &mode_flags);
1477 err = get_nodes(&nodes, nmask, maxnode);
1481 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1484 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1485 unsigned long, home_node, unsigned long, flags)
1487 struct mm_struct *mm = current->mm;
1488 struct vm_area_struct *vma, *prev;
1489 struct mempolicy *new, *old;
1492 VMA_ITERATOR(vmi, mm, start);
1494 start = untagged_addr(start);
1495 if (start & ~PAGE_MASK)
1498 * flags is used for future extension if any.
1504 * Check home_node is online to avoid accessing uninitialized
1507 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1510 len = PAGE_ALIGN(len);
1517 mmap_write_lock(mm);
1518 prev = vma_prev(&vmi);
1519 for_each_vma_range(vmi, vma, end) {
1521 * If any vma in the range got policy other than MPOL_BIND
1522 * or MPOL_PREFERRED_MANY we return error. We don't reset
1523 * the home node for vmas we already updated before.
1525 old = vma_policy(vma);
1530 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1534 new = mpol_dup(old);
1540 vma_start_write(vma);
1541 new->home_node = home_node;
1542 err = mbind_range(&vmi, vma, &prev, start, end, new);
1547 mmap_write_unlock(mm);
1551 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1552 unsigned long, mode, const unsigned long __user *, nmask,
1553 unsigned long, maxnode, unsigned int, flags)
1555 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1558 /* Set the process memory policy */
1559 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1560 unsigned long maxnode)
1562 unsigned short mode_flags;
1567 err = sanitize_mpol_flags(&lmode, &mode_flags);
1571 err = get_nodes(&nodes, nmask, maxnode);
1575 return do_set_mempolicy(lmode, mode_flags, &nodes);
1578 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1579 unsigned long, maxnode)
1581 return kernel_set_mempolicy(mode, nmask, maxnode);
1584 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1585 const unsigned long __user *old_nodes,
1586 const unsigned long __user *new_nodes)
1588 struct mm_struct *mm = NULL;
1589 struct task_struct *task;
1590 nodemask_t task_nodes;
1594 NODEMASK_SCRATCH(scratch);
1599 old = &scratch->mask1;
1600 new = &scratch->mask2;
1602 err = get_nodes(old, old_nodes, maxnode);
1606 err = get_nodes(new, new_nodes, maxnode);
1610 /* Find the mm_struct */
1612 task = pid ? find_task_by_vpid(pid) : current;
1618 get_task_struct(task);
1623 * Check if this process has the right to modify the specified process.
1624 * Use the regular "ptrace_may_access()" checks.
1626 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1633 task_nodes = cpuset_mems_allowed(task);
1634 /* Is the user allowed to access the target nodes? */
1635 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1640 task_nodes = cpuset_mems_allowed(current);
1641 nodes_and(*new, *new, task_nodes);
1642 if (nodes_empty(*new))
1645 err = security_task_movememory(task);
1649 mm = get_task_mm(task);
1650 put_task_struct(task);
1657 err = do_migrate_pages(mm, old, new,
1658 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1662 NODEMASK_SCRATCH_FREE(scratch);
1667 put_task_struct(task);
1672 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1673 const unsigned long __user *, old_nodes,
1674 const unsigned long __user *, new_nodes)
1676 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1680 /* Retrieve NUMA policy */
1681 static int kernel_get_mempolicy(int __user *policy,
1682 unsigned long __user *nmask,
1683 unsigned long maxnode,
1685 unsigned long flags)
1691 if (nmask != NULL && maxnode < nr_node_ids)
1694 addr = untagged_addr(addr);
1696 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1701 if (policy && put_user(pval, policy))
1705 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1710 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1711 unsigned long __user *, nmask, unsigned long, maxnode,
1712 unsigned long, addr, unsigned long, flags)
1714 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1717 bool vma_migratable(struct vm_area_struct *vma)
1719 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1723 * DAX device mappings require predictable access latency, so avoid
1724 * incurring periodic faults.
1726 if (vma_is_dax(vma))
1729 if (is_vm_hugetlb_page(vma) &&
1730 !hugepage_migration_supported(hstate_vma(vma)))
1734 * Migration allocates pages in the highest zone. If we cannot
1735 * do so then migration (at least from node to node) is not
1739 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1745 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1748 struct mempolicy *pol = NULL;
1751 if (vma->vm_ops && vma->vm_ops->get_policy) {
1752 pol = vma->vm_ops->get_policy(vma, addr);
1753 } else if (vma->vm_policy) {
1754 pol = vma->vm_policy;
1757 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1758 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1759 * count on these policies which will be dropped by
1760 * mpol_cond_put() later
1762 if (mpol_needs_cond_ref(pol))
1771 * get_vma_policy(@vma, @addr)
1772 * @vma: virtual memory area whose policy is sought
1773 * @addr: address in @vma for shared policy lookup
1775 * Returns effective policy for a VMA at specified address.
1776 * Falls back to current->mempolicy or system default policy, as necessary.
1777 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1778 * count--added by the get_policy() vm_op, as appropriate--to protect against
1779 * freeing by another task. It is the caller's responsibility to free the
1780 * extra reference for shared policies.
1782 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1785 struct mempolicy *pol = __get_vma_policy(vma, addr);
1788 pol = get_task_policy(current);
1793 bool vma_policy_mof(struct vm_area_struct *vma)
1795 struct mempolicy *pol;
1797 if (vma->vm_ops && vma->vm_ops->get_policy) {
1800 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1801 if (pol && (pol->flags & MPOL_F_MOF))
1808 pol = vma->vm_policy;
1810 pol = get_task_policy(current);
1812 return pol->flags & MPOL_F_MOF;
1815 bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1817 enum zone_type dynamic_policy_zone = policy_zone;
1819 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1822 * if policy->nodes has movable memory only,
1823 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1825 * policy->nodes is intersect with node_states[N_MEMORY].
1826 * so if the following test fails, it implies
1827 * policy->nodes has movable memory only.
1829 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1830 dynamic_policy_zone = ZONE_MOVABLE;
1832 return zone >= dynamic_policy_zone;
1836 * Return a nodemask representing a mempolicy for filtering nodes for
1839 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1841 int mode = policy->mode;
1843 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1844 if (unlikely(mode == MPOL_BIND) &&
1845 apply_policy_zone(policy, gfp_zone(gfp)) &&
1846 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1847 return &policy->nodes;
1849 if (mode == MPOL_PREFERRED_MANY)
1850 return &policy->nodes;
1856 * Return the preferred node id for 'prefer' mempolicy, and return
1857 * the given id for all other policies.
1859 * policy_node() is always coupled with policy_nodemask(), which
1860 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1862 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1864 if (policy->mode == MPOL_PREFERRED) {
1865 nd = first_node(policy->nodes);
1868 * __GFP_THISNODE shouldn't even be used with the bind policy
1869 * because we might easily break the expectation to stay on the
1870 * requested node and not break the policy.
1872 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1875 if ((policy->mode == MPOL_BIND ||
1876 policy->mode == MPOL_PREFERRED_MANY) &&
1877 policy->home_node != NUMA_NO_NODE)
1878 return policy->home_node;
1883 /* Do dynamic interleaving for a process */
1884 static unsigned interleave_nodes(struct mempolicy *policy)
1887 struct task_struct *me = current;
1889 next = next_node_in(me->il_prev, policy->nodes);
1890 if (next < MAX_NUMNODES)
1896 * Depending on the memory policy provide a node from which to allocate the
1899 unsigned int mempolicy_slab_node(void)
1901 struct mempolicy *policy;
1902 int node = numa_mem_id();
1907 policy = current->mempolicy;
1911 switch (policy->mode) {
1912 case MPOL_PREFERRED:
1913 return first_node(policy->nodes);
1915 case MPOL_INTERLEAVE:
1916 return interleave_nodes(policy);
1919 case MPOL_PREFERRED_MANY:
1924 * Follow bind policy behavior and start allocation at the
1927 struct zonelist *zonelist;
1928 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1929 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1930 z = first_zones_zonelist(zonelist, highest_zoneidx,
1932 return z->zone ? zone_to_nid(z->zone) : node;
1943 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1944 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1945 * number of present nodes.
1947 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1949 nodemask_t nodemask = pol->nodes;
1950 unsigned int target, nnodes;
1954 * The barrier will stabilize the nodemask in a register or on
1955 * the stack so that it will stop changing under the code.
1957 * Between first_node() and next_node(), pol->nodes could be changed
1958 * by other threads. So we put pol->nodes in a local stack.
1962 nnodes = nodes_weight(nodemask);
1964 return numa_node_id();
1965 target = (unsigned int)n % nnodes;
1966 nid = first_node(nodemask);
1967 for (i = 0; i < target; i++)
1968 nid = next_node(nid, nodemask);
1972 /* Determine a node number for interleave */
1973 static inline unsigned interleave_nid(struct mempolicy *pol,
1974 struct vm_area_struct *vma, unsigned long addr, int shift)
1980 * for small pages, there is no difference between
1981 * shift and PAGE_SHIFT, so the bit-shift is safe.
1982 * for huge pages, since vm_pgoff is in units of small
1983 * pages, we need to shift off the always 0 bits to get
1986 BUG_ON(shift < PAGE_SHIFT);
1987 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1988 off += (addr - vma->vm_start) >> shift;
1989 return offset_il_node(pol, off);
1991 return interleave_nodes(pol);
1994 #ifdef CONFIG_HUGETLBFS
1996 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1997 * @vma: virtual memory area whose policy is sought
1998 * @addr: address in @vma for shared policy lookup and interleave policy
1999 * @gfp_flags: for requested zone
2000 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2001 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2003 * Returns a nid suitable for a huge page allocation and a pointer
2004 * to the struct mempolicy for conditional unref after allocation.
2005 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2006 * to the mempolicy's @nodemask for filtering the zonelist.
2008 * Must be protected by read_mems_allowed_begin()
2010 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2011 struct mempolicy **mpol, nodemask_t **nodemask)
2016 *mpol = get_vma_policy(vma, addr);
2018 mode = (*mpol)->mode;
2020 if (unlikely(mode == MPOL_INTERLEAVE)) {
2021 nid = interleave_nid(*mpol, vma, addr,
2022 huge_page_shift(hstate_vma(vma)));
2024 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2025 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2026 *nodemask = &(*mpol)->nodes;
2032 * init_nodemask_of_mempolicy
2034 * If the current task's mempolicy is "default" [NULL], return 'false'
2035 * to indicate default policy. Otherwise, extract the policy nodemask
2036 * for 'bind' or 'interleave' policy into the argument nodemask, or
2037 * initialize the argument nodemask to contain the single node for
2038 * 'preferred' or 'local' policy and return 'true' to indicate presence
2039 * of non-default mempolicy.
2041 * We don't bother with reference counting the mempolicy [mpol_get/put]
2042 * because the current task is examining it's own mempolicy and a task's
2043 * mempolicy is only ever changed by the task itself.
2045 * N.B., it is the caller's responsibility to free a returned nodemask.
2047 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2049 struct mempolicy *mempolicy;
2051 if (!(mask && current->mempolicy))
2055 mempolicy = current->mempolicy;
2056 switch (mempolicy->mode) {
2057 case MPOL_PREFERRED:
2058 case MPOL_PREFERRED_MANY:
2060 case MPOL_INTERLEAVE:
2061 *mask = mempolicy->nodes;
2065 init_nodemask_of_node(mask, numa_node_id());
2071 task_unlock(current);
2078 * mempolicy_in_oom_domain
2080 * If tsk's mempolicy is "bind", check for intersection between mask and
2081 * the policy nodemask. Otherwise, return true for all other policies
2082 * including "interleave", as a tsk with "interleave" policy may have
2083 * memory allocated from all nodes in system.
2085 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2087 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2088 const nodemask_t *mask)
2090 struct mempolicy *mempolicy;
2097 mempolicy = tsk->mempolicy;
2098 if (mempolicy && mempolicy->mode == MPOL_BIND)
2099 ret = nodes_intersects(mempolicy->nodes, *mask);
2105 /* Allocate a page in interleaved policy.
2106 Own path because it needs to do special accounting. */
2107 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2112 page = __alloc_pages(gfp, order, nid, NULL);
2113 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2114 if (!static_branch_likely(&vm_numa_stat_key))
2116 if (page && page_to_nid(page) == nid) {
2118 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2124 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2125 int nid, struct mempolicy *pol)
2128 gfp_t preferred_gfp;
2131 * This is a two pass approach. The first pass will only try the
2132 * preferred nodes but skip the direct reclaim and allow the
2133 * allocation to fail, while the second pass will try all the
2136 preferred_gfp = gfp | __GFP_NOWARN;
2137 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2138 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2140 page = __alloc_pages(gfp, order, nid, NULL);
2146 * vma_alloc_folio - Allocate a folio for a VMA.
2148 * @order: Order of the folio.
2149 * @vma: Pointer to VMA or NULL if not available.
2150 * @addr: Virtual address of the allocation. Must be inside @vma.
2151 * @hugepage: For hugepages try only the preferred node if possible.
2153 * Allocate a folio for a specific address in @vma, using the appropriate
2154 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2155 * of the mm_struct of the VMA to prevent it from going away. Should be
2156 * used for all allocations for folios that will be mapped into user space.
2158 * Return: The folio on success or NULL if allocation fails.
2160 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2161 unsigned long addr, bool hugepage)
2163 struct mempolicy *pol;
2164 int node = numa_node_id();
2165 struct folio *folio;
2169 pol = get_vma_policy(vma, addr);
2171 if (pol->mode == MPOL_INTERLEAVE) {
2175 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2178 page = alloc_page_interleave(gfp, order, nid);
2179 folio = (struct folio *)page;
2180 if (folio && order > 1)
2181 folio_prep_large_rmappable(folio);
2185 if (pol->mode == MPOL_PREFERRED_MANY) {
2188 node = policy_node(gfp, pol, node);
2190 page = alloc_pages_preferred_many(gfp, order, node, pol);
2192 folio = (struct folio *)page;
2193 if (folio && order > 1)
2194 folio_prep_large_rmappable(folio);
2198 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2199 int hpage_node = node;
2202 * For hugepage allocation and non-interleave policy which
2203 * allows the current node (or other explicitly preferred
2204 * node) we only try to allocate from the current/preferred
2205 * node and don't fall back to other nodes, as the cost of
2206 * remote accesses would likely offset THP benefits.
2208 * If the policy is interleave or does not allow the current
2209 * node in its nodemask, we allocate the standard way.
2211 if (pol->mode == MPOL_PREFERRED)
2212 hpage_node = first_node(pol->nodes);
2214 nmask = policy_nodemask(gfp, pol);
2215 if (!nmask || node_isset(hpage_node, *nmask)) {
2218 * First, try to allocate THP only on local node, but
2219 * don't reclaim unnecessarily, just compact.
2221 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2222 __GFP_NORETRY, order, hpage_node);
2225 * If hugepage allocations are configured to always
2226 * synchronous compact or the vma has been madvised
2227 * to prefer hugepage backing, retry allowing remote
2228 * memory with both reclaim and compact as well.
2230 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2231 folio = __folio_alloc(gfp, order, hpage_node,
2238 nmask = policy_nodemask(gfp, pol);
2239 preferred_nid = policy_node(gfp, pol, node);
2240 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2245 EXPORT_SYMBOL(vma_alloc_folio);
2248 * alloc_pages - Allocate pages.
2250 * @order: Power of two of number of pages to allocate.
2252 * Allocate 1 << @order contiguous pages. The physical address of the
2253 * first page is naturally aligned (eg an order-3 allocation will be aligned
2254 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2255 * process is honoured when in process context.
2257 * Context: Can be called from any context, providing the appropriate GFP
2259 * Return: The page on success or NULL if allocation fails.
2261 struct page *alloc_pages(gfp_t gfp, unsigned order)
2263 struct mempolicy *pol = &default_policy;
2266 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2267 pol = get_task_policy(current);
2270 * No reference counting needed for current->mempolicy
2271 * nor system default_policy
2273 if (pol->mode == MPOL_INTERLEAVE)
2274 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2275 else if (pol->mode == MPOL_PREFERRED_MANY)
2276 page = alloc_pages_preferred_many(gfp, order,
2277 policy_node(gfp, pol, numa_node_id()), pol);
2279 page = __alloc_pages(gfp, order,
2280 policy_node(gfp, pol, numa_node_id()),
2281 policy_nodemask(gfp, pol));
2285 EXPORT_SYMBOL(alloc_pages);
2287 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2289 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2290 struct folio *folio = (struct folio *)page;
2292 if (folio && order > 1)
2293 folio_prep_large_rmappable(folio);
2296 EXPORT_SYMBOL(folio_alloc);
2298 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2299 struct mempolicy *pol, unsigned long nr_pages,
2300 struct page **page_array)
2303 unsigned long nr_pages_per_node;
2306 unsigned long nr_allocated;
2307 unsigned long total_allocated = 0;
2309 nodes = nodes_weight(pol->nodes);
2310 nr_pages_per_node = nr_pages / nodes;
2311 delta = nr_pages - nodes * nr_pages_per_node;
2313 for (i = 0; i < nodes; i++) {
2315 nr_allocated = __alloc_pages_bulk(gfp,
2316 interleave_nodes(pol), NULL,
2317 nr_pages_per_node + 1, NULL,
2321 nr_allocated = __alloc_pages_bulk(gfp,
2322 interleave_nodes(pol), NULL,
2323 nr_pages_per_node, NULL, page_array);
2326 page_array += nr_allocated;
2327 total_allocated += nr_allocated;
2330 return total_allocated;
2333 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2334 struct mempolicy *pol, unsigned long nr_pages,
2335 struct page **page_array)
2337 gfp_t preferred_gfp;
2338 unsigned long nr_allocated = 0;
2340 preferred_gfp = gfp | __GFP_NOWARN;
2341 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2343 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2344 nr_pages, NULL, page_array);
2346 if (nr_allocated < nr_pages)
2347 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2348 nr_pages - nr_allocated, NULL,
2349 page_array + nr_allocated);
2350 return nr_allocated;
2353 /* alloc pages bulk and mempolicy should be considered at the
2354 * same time in some situation such as vmalloc.
2356 * It can accelerate memory allocation especially interleaving
2359 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2360 unsigned long nr_pages, struct page **page_array)
2362 struct mempolicy *pol = &default_policy;
2364 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2365 pol = get_task_policy(current);
2367 if (pol->mode == MPOL_INTERLEAVE)
2368 return alloc_pages_bulk_array_interleave(gfp, pol,
2369 nr_pages, page_array);
2371 if (pol->mode == MPOL_PREFERRED_MANY)
2372 return alloc_pages_bulk_array_preferred_many(gfp,
2373 numa_node_id(), pol, nr_pages, page_array);
2375 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2376 policy_nodemask(gfp, pol), nr_pages, NULL,
2380 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2382 struct mempolicy *pol = mpol_dup(vma_policy(src));
2385 return PTR_ERR(pol);
2386 dst->vm_policy = pol;
2391 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2392 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2393 * with the mems_allowed returned by cpuset_mems_allowed(). This
2394 * keeps mempolicies cpuset relative after its cpuset moves. See
2395 * further kernel/cpuset.c update_nodemask().
2397 * current's mempolicy may be rebinded by the other task(the task that changes
2398 * cpuset's mems), so we needn't do rebind work for current task.
2401 /* Slow path of a mempolicy duplicate */
2402 struct mempolicy *__mpol_dup(struct mempolicy *old)
2404 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2407 return ERR_PTR(-ENOMEM);
2409 /* task's mempolicy is protected by alloc_lock */
2410 if (old == current->mempolicy) {
2413 task_unlock(current);
2417 if (current_cpuset_is_being_rebound()) {
2418 nodemask_t mems = cpuset_mems_allowed(current);
2419 mpol_rebind_policy(new, &mems);
2421 atomic_set(&new->refcnt, 1);
2425 /* Slow path of a mempolicy comparison */
2426 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2430 if (a->mode != b->mode)
2432 if (a->flags != b->flags)
2434 if (a->home_node != b->home_node)
2436 if (mpol_store_user_nodemask(a))
2437 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2442 case MPOL_INTERLEAVE:
2443 case MPOL_PREFERRED:
2444 case MPOL_PREFERRED_MANY:
2445 return !!nodes_equal(a->nodes, b->nodes);
2455 * Shared memory backing store policy support.
2457 * Remember policies even when nobody has shared memory mapped.
2458 * The policies are kept in Red-Black tree linked from the inode.
2459 * They are protected by the sp->lock rwlock, which should be held
2460 * for any accesses to the tree.
2464 * lookup first element intersecting start-end. Caller holds sp->lock for
2465 * reading or for writing
2467 static struct sp_node *
2468 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2470 struct rb_node *n = sp->root.rb_node;
2473 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2475 if (start >= p->end)
2477 else if (end <= p->start)
2485 struct sp_node *w = NULL;
2486 struct rb_node *prev = rb_prev(n);
2489 w = rb_entry(prev, struct sp_node, nd);
2490 if (w->end <= start)
2494 return rb_entry(n, struct sp_node, nd);
2498 * Insert a new shared policy into the list. Caller holds sp->lock for
2501 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2503 struct rb_node **p = &sp->root.rb_node;
2504 struct rb_node *parent = NULL;
2509 nd = rb_entry(parent, struct sp_node, nd);
2510 if (new->start < nd->start)
2512 else if (new->end > nd->end)
2513 p = &(*p)->rb_right;
2517 rb_link_node(&new->nd, parent, p);
2518 rb_insert_color(&new->nd, &sp->root);
2519 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2520 new->policy ? new->policy->mode : 0);
2523 /* Find shared policy intersecting idx */
2525 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2527 struct mempolicy *pol = NULL;
2530 if (!sp->root.rb_node)
2532 read_lock(&sp->lock);
2533 sn = sp_lookup(sp, idx, idx+1);
2535 mpol_get(sn->policy);
2538 read_unlock(&sp->lock);
2542 static void sp_free(struct sp_node *n)
2544 mpol_put(n->policy);
2545 kmem_cache_free(sn_cache, n);
2549 * mpol_misplaced - check whether current folio node is valid in policy
2551 * @folio: folio to be checked
2552 * @vma: vm area where folio mapped
2553 * @addr: virtual address in @vma for shared policy lookup and interleave policy
2555 * Lookup current policy node id for vma,addr and "compare to" folio's
2556 * node id. Policy determination "mimics" alloc_page_vma().
2557 * Called from fault path where we know the vma and faulting address.
2559 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2560 * policy, or a suitable node ID to allocate a replacement folio from.
2562 int mpol_misplaced(struct folio *folio, struct vm_area_struct *vma,
2565 struct mempolicy *pol;
2567 int curnid = folio_nid(folio);
2568 unsigned long pgoff;
2569 int thiscpu = raw_smp_processor_id();
2570 int thisnid = cpu_to_node(thiscpu);
2571 int polnid = NUMA_NO_NODE;
2572 int ret = NUMA_NO_NODE;
2574 pol = get_vma_policy(vma, addr);
2575 if (!(pol->flags & MPOL_F_MOF))
2578 switch (pol->mode) {
2579 case MPOL_INTERLEAVE:
2580 pgoff = vma->vm_pgoff;
2581 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2582 polnid = offset_il_node(pol, pgoff);
2585 case MPOL_PREFERRED:
2586 if (node_isset(curnid, pol->nodes))
2588 polnid = first_node(pol->nodes);
2592 polnid = numa_node_id();
2596 /* Optimize placement among multiple nodes via NUMA balancing */
2597 if (pol->flags & MPOL_F_MORON) {
2598 if (node_isset(thisnid, pol->nodes))
2604 case MPOL_PREFERRED_MANY:
2606 * use current page if in policy nodemask,
2607 * else select nearest allowed node, if any.
2608 * If no allowed nodes, use current [!misplaced].
2610 if (node_isset(curnid, pol->nodes))
2612 z = first_zones_zonelist(
2613 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2614 gfp_zone(GFP_HIGHUSER),
2616 polnid = zone_to_nid(z->zone);
2623 /* Migrate the folio towards the node whose CPU is referencing it */
2624 if (pol->flags & MPOL_F_MORON) {
2627 if (!should_numa_migrate_memory(current, folio, curnid,
2632 if (curnid != polnid)
2641 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2642 * dropped after task->mempolicy is set to NULL so that any allocation done as
2643 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2646 void mpol_put_task_policy(struct task_struct *task)
2648 struct mempolicy *pol;
2651 pol = task->mempolicy;
2652 task->mempolicy = NULL;
2657 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2659 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2660 rb_erase(&n->nd, &sp->root);
2664 static void sp_node_init(struct sp_node *node, unsigned long start,
2665 unsigned long end, struct mempolicy *pol)
2667 node->start = start;
2672 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2673 struct mempolicy *pol)
2676 struct mempolicy *newpol;
2678 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2682 newpol = mpol_dup(pol);
2683 if (IS_ERR(newpol)) {
2684 kmem_cache_free(sn_cache, n);
2687 newpol->flags |= MPOL_F_SHARED;
2688 sp_node_init(n, start, end, newpol);
2693 /* Replace a policy range. */
2694 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2695 unsigned long end, struct sp_node *new)
2698 struct sp_node *n_new = NULL;
2699 struct mempolicy *mpol_new = NULL;
2703 write_lock(&sp->lock);
2704 n = sp_lookup(sp, start, end);
2705 /* Take care of old policies in the same range. */
2706 while (n && n->start < end) {
2707 struct rb_node *next = rb_next(&n->nd);
2708 if (n->start >= start) {
2714 /* Old policy spanning whole new range. */
2719 *mpol_new = *n->policy;
2720 atomic_set(&mpol_new->refcnt, 1);
2721 sp_node_init(n_new, end, n->end, mpol_new);
2723 sp_insert(sp, n_new);
2732 n = rb_entry(next, struct sp_node, nd);
2736 write_unlock(&sp->lock);
2743 kmem_cache_free(sn_cache, n_new);
2748 write_unlock(&sp->lock);
2750 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2753 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2756 atomic_set(&mpol_new->refcnt, 1);
2761 * mpol_shared_policy_init - initialize shared policy for inode
2762 * @sp: pointer to inode shared policy
2763 * @mpol: struct mempolicy to install
2765 * Install non-NULL @mpol in inode's shared policy rb-tree.
2766 * On entry, the current task has a reference on a non-NULL @mpol.
2767 * This must be released on exit.
2768 * This is called at get_inode() calls and we can use GFP_KERNEL.
2770 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2774 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2775 rwlock_init(&sp->lock);
2778 struct vm_area_struct pvma;
2779 struct mempolicy *new;
2780 NODEMASK_SCRATCH(scratch);
2784 /* contextualize the tmpfs mount point mempolicy */
2785 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2787 goto free_scratch; /* no valid nodemask intersection */
2790 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2791 task_unlock(current);
2795 /* Create pseudo-vma that contains just the policy */
2796 vma_init(&pvma, NULL);
2797 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2798 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2801 mpol_put(new); /* drop initial ref */
2803 NODEMASK_SCRATCH_FREE(scratch);
2805 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2809 int mpol_set_shared_policy(struct shared_policy *info,
2810 struct vm_area_struct *vma, struct mempolicy *npol)
2813 struct sp_node *new = NULL;
2814 unsigned long sz = vma_pages(vma);
2816 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2818 sz, npol ? npol->mode : -1,
2819 npol ? npol->flags : -1,
2820 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2823 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2827 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2833 /* Free a backing policy store on inode delete. */
2834 void mpol_free_shared_policy(struct shared_policy *p)
2837 struct rb_node *next;
2839 if (!p->root.rb_node)
2841 write_lock(&p->lock);
2842 next = rb_first(&p->root);
2844 n = rb_entry(next, struct sp_node, nd);
2845 next = rb_next(&n->nd);
2848 write_unlock(&p->lock);
2851 #ifdef CONFIG_NUMA_BALANCING
2852 static int __initdata numabalancing_override;
2854 static void __init check_numabalancing_enable(void)
2856 bool numabalancing_default = false;
2858 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2859 numabalancing_default = true;
2861 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2862 if (numabalancing_override)
2863 set_numabalancing_state(numabalancing_override == 1);
2865 if (num_online_nodes() > 1 && !numabalancing_override) {
2866 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2867 numabalancing_default ? "Enabling" : "Disabling");
2868 set_numabalancing_state(numabalancing_default);
2872 static int __init setup_numabalancing(char *str)
2878 if (!strcmp(str, "enable")) {
2879 numabalancing_override = 1;
2881 } else if (!strcmp(str, "disable")) {
2882 numabalancing_override = -1;
2887 pr_warn("Unable to parse numa_balancing=\n");
2891 __setup("numa_balancing=", setup_numabalancing);
2893 static inline void __init check_numabalancing_enable(void)
2896 #endif /* CONFIG_NUMA_BALANCING */
2898 /* assumes fs == KERNEL_DS */
2899 void __init numa_policy_init(void)
2901 nodemask_t interleave_nodes;
2902 unsigned long largest = 0;
2903 int nid, prefer = 0;
2905 policy_cache = kmem_cache_create("numa_policy",
2906 sizeof(struct mempolicy),
2907 0, SLAB_PANIC, NULL);
2909 sn_cache = kmem_cache_create("shared_policy_node",
2910 sizeof(struct sp_node),
2911 0, SLAB_PANIC, NULL);
2913 for_each_node(nid) {
2914 preferred_node_policy[nid] = (struct mempolicy) {
2915 .refcnt = ATOMIC_INIT(1),
2916 .mode = MPOL_PREFERRED,
2917 .flags = MPOL_F_MOF | MPOL_F_MORON,
2918 .nodes = nodemask_of_node(nid),
2923 * Set interleaving policy for system init. Interleaving is only
2924 * enabled across suitably sized nodes (default is >= 16MB), or
2925 * fall back to the largest node if they're all smaller.
2927 nodes_clear(interleave_nodes);
2928 for_each_node_state(nid, N_MEMORY) {
2929 unsigned long total_pages = node_present_pages(nid);
2931 /* Preserve the largest node */
2932 if (largest < total_pages) {
2933 largest = total_pages;
2937 /* Interleave this node? */
2938 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2939 node_set(nid, interleave_nodes);
2942 /* All too small, use the largest */
2943 if (unlikely(nodes_empty(interleave_nodes)))
2944 node_set(prefer, interleave_nodes);
2946 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2947 pr_err("%s: interleaving failed\n", __func__);
2949 check_numabalancing_enable();
2952 /* Reset policy of current process to default */
2953 void numa_default_policy(void)
2955 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2959 * Parse and format mempolicy from/to strings
2962 static const char * const policy_modes[] =
2964 [MPOL_DEFAULT] = "default",
2965 [MPOL_PREFERRED] = "prefer",
2966 [MPOL_BIND] = "bind",
2967 [MPOL_INTERLEAVE] = "interleave",
2968 [MPOL_LOCAL] = "local",
2969 [MPOL_PREFERRED_MANY] = "prefer (many)",
2975 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2976 * @str: string containing mempolicy to parse
2977 * @mpol: pointer to struct mempolicy pointer, returned on success.
2980 * <mode>[=<flags>][:<nodelist>]
2982 * Return: %0 on success, else %1
2984 int mpol_parse_str(char *str, struct mempolicy **mpol)
2986 struct mempolicy *new = NULL;
2987 unsigned short mode_flags;
2989 char *nodelist = strchr(str, ':');
2990 char *flags = strchr(str, '=');
2994 *flags++ = '\0'; /* terminate mode string */
2997 /* NUL-terminate mode or flags string */
2999 if (nodelist_parse(nodelist, nodes))
3001 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3006 mode = match_string(policy_modes, MPOL_MAX, str);
3011 case MPOL_PREFERRED:
3013 * Insist on a nodelist of one node only, although later
3014 * we use first_node(nodes) to grab a single node, so here
3015 * nodelist (or nodes) cannot be empty.
3018 char *rest = nodelist;
3019 while (isdigit(*rest))
3023 if (nodes_empty(nodes))
3027 case MPOL_INTERLEAVE:
3029 * Default to online nodes with memory if no nodelist
3032 nodes = node_states[N_MEMORY];
3036 * Don't allow a nodelist; mpol_new() checks flags
3043 * Insist on a empty nodelist
3048 case MPOL_PREFERRED_MANY:
3051 * Insist on a nodelist
3060 * Currently, we only support two mutually exclusive
3063 if (!strcmp(flags, "static"))
3064 mode_flags |= MPOL_F_STATIC_NODES;
3065 else if (!strcmp(flags, "relative"))
3066 mode_flags |= MPOL_F_RELATIVE_NODES;
3071 new = mpol_new(mode, mode_flags, &nodes);
3076 * Save nodes for mpol_to_str() to show the tmpfs mount options
3077 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3079 if (mode != MPOL_PREFERRED) {
3081 } else if (nodelist) {
3082 nodes_clear(new->nodes);
3083 node_set(first_node(nodes), new->nodes);
3085 new->mode = MPOL_LOCAL;
3089 * Save nodes for contextualization: this will be used to "clone"
3090 * the mempolicy in a specific context [cpuset] at a later time.
3092 new->w.user_nodemask = nodes;
3097 /* Restore string for error message */
3106 #endif /* CONFIG_TMPFS */
3109 * mpol_to_str - format a mempolicy structure for printing
3110 * @buffer: to contain formatted mempolicy string
3111 * @maxlen: length of @buffer
3112 * @pol: pointer to mempolicy to be formatted
3114 * Convert @pol into a string. If @buffer is too short, truncate the string.
3115 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3116 * longest flag, "relative", and to display at least a few node ids.
3118 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3121 nodemask_t nodes = NODE_MASK_NONE;
3122 unsigned short mode = MPOL_DEFAULT;
3123 unsigned short flags = 0;
3125 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3134 case MPOL_PREFERRED:
3135 case MPOL_PREFERRED_MANY:
3137 case MPOL_INTERLEAVE:
3142 snprintf(p, maxlen, "unknown");
3146 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3148 if (flags & MPOL_MODE_FLAGS) {
3149 p += snprintf(p, buffer + maxlen - p, "=");
3152 * Currently, the only defined flags are mutually exclusive
3154 if (flags & MPOL_F_STATIC_NODES)
3155 p += snprintf(p, buffer + maxlen - p, "static");
3156 else if (flags & MPOL_F_RELATIVE_NODES)
3157 p += snprintf(p, buffer + maxlen - p, "relative");
3160 if (!nodes_empty(nodes))
3161 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3162 nodemask_pr_args(&nodes));