1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple NUMA memory policy for the Linux kernel.
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * preferred many Try a set of nodes first before normal fallback. This is
35 * similar to preferred without the special case.
37 * default Allocate on the local node first, or when on a VMA
38 * use the process policy. This is what Linux always did
39 * in a NUMA aware kernel and still does by, ahem, default.
41 * The process policy is applied for most non interrupt memory allocations
42 * in that process' context. Interrupts ignore the policies and always
43 * try to allocate on the local CPU. The VMA policy is only applied for memory
44 * allocations for a VMA in the VM.
46 * Currently there are a few corner cases in swapping where the policy
47 * is not applied, but the majority should be handled. When process policy
48 * is used it is not remembered over swap outs/swap ins.
50 * Only the highest zone in the zone hierarchy gets policied. Allocations
51 * requesting a lower zone just use default policy. This implies that
52 * on systems with highmem kernel lowmem allocation don't get policied.
53 * Same with GFP_DMA allocations.
55 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
56 * all users and remembered even when nobody has memory mapped.
60 fix mmap readahead to honour policy and enable policy for any page cache
62 statistics for bigpages
63 global policy for page cache? currently it uses process policy. Requires
65 handle mremap for shared memory (currently ignored for the policy)
67 make bind policy root only? It can trigger oom much faster and the
68 kernel is not always grateful with that.
71 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
73 #include <linux/mempolicy.h>
74 #include <linux/pagewalk.h>
75 #include <linux/highmem.h>
76 #include <linux/hugetlb.h>
77 #include <linux/kernel.h>
78 #include <linux/sched.h>
79 #include <linux/sched/mm.h>
80 #include <linux/sched/numa_balancing.h>
81 #include <linux/sched/task.h>
82 #include <linux/nodemask.h>
83 #include <linux/cpuset.h>
84 #include <linux/slab.h>
85 #include <linux/string.h>
86 #include <linux/export.h>
87 #include <linux/nsproxy.h>
88 #include <linux/interrupt.h>
89 #include <linux/init.h>
90 #include <linux/compat.h>
91 #include <linux/ptrace.h>
92 #include <linux/swap.h>
93 #include <linux/seq_file.h>
94 #include <linux/proc_fs.h>
95 #include <linux/migrate.h>
96 #include <linux/ksm.h>
97 #include <linux/rmap.h>
98 #include <linux/security.h>
99 #include <linux/syscalls.h>
100 #include <linux/ctype.h>
101 #include <linux/mm_inline.h>
102 #include <linux/mmu_notifier.h>
103 #include <linux/printk.h>
104 #include <linux/swapops.h>
106 #include <asm/tlbflush.h>
107 #include <linux/uaccess.h>
109 #include "internal.h"
112 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
113 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
115 static struct kmem_cache *policy_cache;
116 static struct kmem_cache *sn_cache;
118 /* Highest zone. An specific allocation for a zone below that is not
120 enum zone_type policy_zone = 0;
123 * run-time system-wide default policy => local allocation
125 static struct mempolicy default_policy = {
126 .refcnt = ATOMIC_INIT(1), /* never free it */
130 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
133 * numa_map_to_online_node - Find closest online node
134 * @node: Node id to start the search
136 * Lookup the next closest node by distance if @nid is not online.
138 * Return: this @node if it is online, otherwise the closest node by distance
140 int numa_map_to_online_node(int node)
142 int min_dist = INT_MAX, dist, n, min_node;
144 if (node == NUMA_NO_NODE || node_online(node))
148 for_each_online_node(n) {
149 dist = node_distance(node, n);
150 if (dist < min_dist) {
158 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
160 struct mempolicy *get_task_policy(struct task_struct *p)
162 struct mempolicy *pol = p->mempolicy;
168 node = numa_node_id();
169 if (node != NUMA_NO_NODE) {
170 pol = &preferred_node_policy[node];
171 /* preferred_node_policy is not initialised early in boot */
176 return &default_policy;
179 static const struct mempolicy_operations {
180 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
181 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
182 } mpol_ops[MPOL_MAX];
184 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
186 return pol->flags & MPOL_MODE_FLAGS;
189 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
190 const nodemask_t *rel)
193 nodes_fold(tmp, *orig, nodes_weight(*rel));
194 nodes_onto(*ret, tmp, *rel);
197 static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
199 if (nodes_empty(*nodes))
205 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
207 if (nodes_empty(*nodes))
210 nodes_clear(pol->nodes);
211 node_set(first_node(*nodes), pol->nodes);
216 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
217 * any, for the new policy. mpol_new() has already validated the nodes
218 * parameter with respect to the policy mode and flags.
220 * Must be called holding task's alloc_lock to protect task's mems_allowed
221 * and mempolicy. May also be called holding the mmap_lock for write.
223 static int mpol_set_nodemask(struct mempolicy *pol,
224 const nodemask_t *nodes, struct nodemask_scratch *nsc)
229 * Default (pol==NULL) resp. local memory policies are not a
230 * subject of any remapping. They also do not need any special
233 if (!pol || pol->mode == MPOL_LOCAL)
237 nodes_and(nsc->mask1,
238 cpuset_current_mems_allowed, node_states[N_MEMORY]);
242 if (pol->flags & MPOL_F_RELATIVE_NODES)
243 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
245 nodes_and(nsc->mask2, *nodes, nsc->mask1);
247 if (mpol_store_user_nodemask(pol))
248 pol->w.user_nodemask = *nodes;
250 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
252 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
257 * This function just creates a new policy, does some check and simple
258 * initialization. You must invoke mpol_set_nodemask() to set nodes.
260 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
263 struct mempolicy *policy;
265 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
266 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
268 if (mode == MPOL_DEFAULT) {
269 if (nodes && !nodes_empty(*nodes))
270 return ERR_PTR(-EINVAL);
276 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
277 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
278 * All other modes require a valid pointer to a non-empty nodemask.
280 if (mode == MPOL_PREFERRED) {
281 if (nodes_empty(*nodes)) {
282 if (((flags & MPOL_F_STATIC_NODES) ||
283 (flags & MPOL_F_RELATIVE_NODES)))
284 return ERR_PTR(-EINVAL);
288 } else if (mode == MPOL_LOCAL) {
289 if (!nodes_empty(*nodes) ||
290 (flags & MPOL_F_STATIC_NODES) ||
291 (flags & MPOL_F_RELATIVE_NODES))
292 return ERR_PTR(-EINVAL);
293 } else if (nodes_empty(*nodes))
294 return ERR_PTR(-EINVAL);
295 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
297 return ERR_PTR(-ENOMEM);
298 atomic_set(&policy->refcnt, 1);
300 policy->flags = flags;
301 policy->home_node = NUMA_NO_NODE;
306 /* Slow path of a mpol destructor. */
307 void __mpol_put(struct mempolicy *p)
309 if (!atomic_dec_and_test(&p->refcnt))
311 kmem_cache_free(policy_cache, p);
314 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
318 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
322 if (pol->flags & MPOL_F_STATIC_NODES)
323 nodes_and(tmp, pol->w.user_nodemask, *nodes);
324 else if (pol->flags & MPOL_F_RELATIVE_NODES)
325 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
327 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
329 pol->w.cpuset_mems_allowed = *nodes;
332 if (nodes_empty(tmp))
338 static void mpol_rebind_preferred(struct mempolicy *pol,
339 const nodemask_t *nodes)
341 pol->w.cpuset_mems_allowed = *nodes;
345 * mpol_rebind_policy - Migrate a policy to a different set of nodes
347 * Per-vma policies are protected by mmap_lock. Allocations using per-task
348 * policies are protected by task->mems_allowed_seq to prevent a premature
349 * OOM/allocation failure due to parallel nodemask modification.
351 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
355 if (!mpol_store_user_nodemask(pol) &&
356 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
359 mpol_ops[pol->mode].rebind(pol, newmask);
363 * Wrapper for mpol_rebind_policy() that just requires task
364 * pointer, and updates task mempolicy.
366 * Called with task's alloc_lock held.
369 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
371 mpol_rebind_policy(tsk->mempolicy, new);
375 * Rebind each vma in mm to new nodemask.
377 * Call holding a reference to mm. Takes mm->mmap_lock during call.
380 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
382 struct vm_area_struct *vma;
385 for (vma = mm->mmap; vma; vma = vma->vm_next)
386 mpol_rebind_policy(vma->vm_policy, new);
387 mmap_write_unlock(mm);
390 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
392 .rebind = mpol_rebind_default,
394 [MPOL_INTERLEAVE] = {
395 .create = mpol_new_nodemask,
396 .rebind = mpol_rebind_nodemask,
399 .create = mpol_new_preferred,
400 .rebind = mpol_rebind_preferred,
403 .create = mpol_new_nodemask,
404 .rebind = mpol_rebind_nodemask,
407 .rebind = mpol_rebind_default,
409 [MPOL_PREFERRED_MANY] = {
410 .create = mpol_new_nodemask,
411 .rebind = mpol_rebind_preferred,
415 static int migrate_page_add(struct page *page, struct list_head *pagelist,
416 unsigned long flags);
419 struct list_head *pagelist;
424 struct vm_area_struct *first;
428 * Check if the page's nid is in qp->nmask.
430 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
431 * in the invert of qp->nmask.
433 static inline bool queue_pages_required(struct page *page,
434 struct queue_pages *qp)
436 int nid = page_to_nid(page);
437 unsigned long flags = qp->flags;
439 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
443 * queue_pages_pmd() has three possible return values:
444 * 0 - pages are placed on the right node or queued successfully, or
445 * special page is met, i.e. huge zero page.
446 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
448 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
449 * existing page was already on a node that does not follow the
452 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
453 unsigned long end, struct mm_walk *walk)
458 struct queue_pages *qp = walk->private;
461 if (unlikely(is_pmd_migration_entry(*pmd))) {
465 page = pmd_page(*pmd);
466 if (is_huge_zero_page(page)) {
468 walk->action = ACTION_CONTINUE;
471 if (!queue_pages_required(page, qp))
475 /* go to thp migration */
476 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
477 if (!vma_migratable(walk->vma) ||
478 migrate_page_add(page, qp->pagelist, flags)) {
491 * Scan through pages checking if pages follow certain conditions,
492 * and move them to the pagelist if they do.
494 * queue_pages_pte_range() has three possible return values:
495 * 0 - pages are placed on the right node or queued successfully, or
496 * special page is met, i.e. zero page.
497 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
499 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
500 * on a node that does not follow the policy.
502 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
503 unsigned long end, struct mm_walk *walk)
505 struct vm_area_struct *vma = walk->vma;
507 struct queue_pages *qp = walk->private;
508 unsigned long flags = qp->flags;
509 bool has_unmovable = false;
510 pte_t *pte, *mapped_pte;
513 ptl = pmd_trans_huge_lock(pmd, vma);
515 return queue_pages_pmd(pmd, ptl, addr, end, walk);
517 if (pmd_trans_unstable(pmd))
520 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
521 for (; addr != end; pte++, addr += PAGE_SIZE) {
522 if (!pte_present(*pte))
524 page = vm_normal_page(vma, addr, *pte);
528 * vm_normal_page() filters out zero pages, but there might
529 * still be PageReserved pages to skip, perhaps in a VDSO.
531 if (PageReserved(page))
533 if (!queue_pages_required(page, qp))
535 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
536 /* MPOL_MF_STRICT must be specified if we get here */
537 if (!vma_migratable(vma)) {
538 has_unmovable = true;
543 * Do not abort immediately since there may be
544 * temporary off LRU pages in the range. Still
545 * need migrate other LRU pages.
547 if (migrate_page_add(page, qp->pagelist, flags))
548 has_unmovable = true;
552 pte_unmap_unlock(mapped_pte, ptl);
558 return addr != end ? -EIO : 0;
561 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
562 unsigned long addr, unsigned long end,
563 struct mm_walk *walk)
566 #ifdef CONFIG_HUGETLB_PAGE
567 struct queue_pages *qp = walk->private;
568 unsigned long flags = (qp->flags & MPOL_MF_VALID);
573 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
574 entry = huge_ptep_get(pte);
575 if (!pte_present(entry))
577 page = pte_page(entry);
578 if (!queue_pages_required(page, qp))
581 if (flags == MPOL_MF_STRICT) {
583 * STRICT alone means only detecting misplaced page and no
584 * need to further check other vma.
590 if (!vma_migratable(walk->vma)) {
592 * Must be STRICT with MOVE*, otherwise .test_walk() have
593 * stopped walking current vma.
594 * Detecting misplaced page but allow migrating pages which
601 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
602 if (flags & (MPOL_MF_MOVE_ALL) ||
603 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
604 if (!isolate_huge_page(page, qp->pagelist) &&
605 (flags & MPOL_MF_STRICT))
607 * Failed to isolate page but allow migrating pages
608 * which have been queued.
620 #ifdef CONFIG_NUMA_BALANCING
622 * This is used to mark a range of virtual addresses to be inaccessible.
623 * These are later cleared by a NUMA hinting fault. Depending on these
624 * faults, pages may be migrated for better NUMA placement.
626 * This is assuming that NUMA faults are handled using PROT_NONE. If
627 * an architecture makes a different choice, it will need further
628 * changes to the core.
630 unsigned long change_prot_numa(struct vm_area_struct *vma,
631 unsigned long addr, unsigned long end)
635 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
637 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
642 static unsigned long change_prot_numa(struct vm_area_struct *vma,
643 unsigned long addr, unsigned long end)
647 #endif /* CONFIG_NUMA_BALANCING */
649 static int queue_pages_test_walk(unsigned long start, unsigned long end,
650 struct mm_walk *walk)
652 struct vm_area_struct *vma = walk->vma;
653 struct queue_pages *qp = walk->private;
654 unsigned long endvma = vma->vm_end;
655 unsigned long flags = qp->flags;
657 /* range check first */
658 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
662 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
663 (qp->start < vma->vm_start))
664 /* hole at head side of range */
667 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
668 ((vma->vm_end < qp->end) &&
669 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
670 /* hole at middle or tail of range */
674 * Need check MPOL_MF_STRICT to return -EIO if possible
675 * regardless of vma_migratable
677 if (!vma_migratable(vma) &&
678 !(flags & MPOL_MF_STRICT))
684 if (flags & MPOL_MF_LAZY) {
685 /* Similar to task_numa_work, skip inaccessible VMAs */
686 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
687 !(vma->vm_flags & VM_MIXEDMAP))
688 change_prot_numa(vma, start, endvma);
692 /* queue pages from current vma */
693 if (flags & MPOL_MF_VALID)
698 static const struct mm_walk_ops queue_pages_walk_ops = {
699 .hugetlb_entry = queue_pages_hugetlb,
700 .pmd_entry = queue_pages_pte_range,
701 .test_walk = queue_pages_test_walk,
705 * Walk through page tables and collect pages to be migrated.
707 * If pages found in a given range are on a set of nodes (determined by
708 * @nodes and @flags,) it's isolated and queued to the pagelist which is
709 * passed via @private.
711 * queue_pages_range() has three possible return values:
712 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
714 * 0 - queue pages successfully or no misplaced page.
715 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
716 * memory range specified by nodemask and maxnode points outside
717 * your accessible address space (-EFAULT)
720 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
721 nodemask_t *nodes, unsigned long flags,
722 struct list_head *pagelist)
725 struct queue_pages qp = {
726 .pagelist = pagelist,
734 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
737 /* whole range in hole */
744 * Apply policy to a single VMA
745 * This must be called with the mmap_lock held for writing.
747 static int vma_replace_policy(struct vm_area_struct *vma,
748 struct mempolicy *pol)
751 struct mempolicy *old;
752 struct mempolicy *new;
754 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
755 vma->vm_start, vma->vm_end, vma->vm_pgoff,
756 vma->vm_ops, vma->vm_file,
757 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
763 if (vma->vm_ops && vma->vm_ops->set_policy) {
764 err = vma->vm_ops->set_policy(vma, new);
769 old = vma->vm_policy;
770 vma->vm_policy = new; /* protected by mmap_lock */
779 /* Step 2: apply policy to a range and do splits. */
780 static int mbind_range(struct mm_struct *mm, unsigned long start,
781 unsigned long end, struct mempolicy *new_pol)
783 struct vm_area_struct *prev;
784 struct vm_area_struct *vma;
787 unsigned long vmstart;
790 vma = find_vma(mm, start);
794 if (start > vma->vm_start)
797 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
798 vmstart = max(start, vma->vm_start);
799 vmend = min(end, vma->vm_end);
801 if (mpol_equal(vma_policy(vma), new_pol))
804 pgoff = vma->vm_pgoff +
805 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
806 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
807 vma->anon_vma, vma->vm_file, pgoff,
808 new_pol, vma->vm_userfaultfd_ctx,
814 if (vma->vm_start != vmstart) {
815 err = split_vma(vma->vm_mm, vma, vmstart, 1);
819 if (vma->vm_end != vmend) {
820 err = split_vma(vma->vm_mm, vma, vmend, 0);
825 err = vma_replace_policy(vma, new_pol);
834 /* Set the process memory policy */
835 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
838 struct mempolicy *new, *old;
839 NODEMASK_SCRATCH(scratch);
845 new = mpol_new(mode, flags, nodes);
851 ret = mpol_set_nodemask(new, nodes, scratch);
857 old = current->mempolicy;
858 current->mempolicy = new;
859 if (new && new->mode == MPOL_INTERLEAVE)
860 current->il_prev = MAX_NUMNODES-1;
861 task_unlock(current);
865 NODEMASK_SCRATCH_FREE(scratch);
870 * Return nodemask for policy for get_mempolicy() query
872 * Called with task's alloc_lock held
874 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
877 if (p == &default_policy)
882 case MPOL_INTERLEAVE:
884 case MPOL_PREFERRED_MANY:
888 /* return empty node mask for local allocation */
895 static int lookup_node(struct mm_struct *mm, unsigned long addr)
897 struct page *p = NULL;
900 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
902 ret = page_to_nid(p);
908 /* Retrieve NUMA policy */
909 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
910 unsigned long addr, unsigned long flags)
913 struct mm_struct *mm = current->mm;
914 struct vm_area_struct *vma = NULL;
915 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
918 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
921 if (flags & MPOL_F_MEMS_ALLOWED) {
922 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
924 *policy = 0; /* just so it's initialized */
926 *nmask = cpuset_current_mems_allowed;
927 task_unlock(current);
931 if (flags & MPOL_F_ADDR) {
933 * Do NOT fall back to task policy if the
934 * vma/shared policy at addr is NULL. We
935 * want to return MPOL_DEFAULT in this case.
938 vma = vma_lookup(mm, addr);
940 mmap_read_unlock(mm);
943 if (vma->vm_ops && vma->vm_ops->get_policy)
944 pol = vma->vm_ops->get_policy(vma, addr);
946 pol = vma->vm_policy;
951 pol = &default_policy; /* indicates default behavior */
953 if (flags & MPOL_F_NODE) {
954 if (flags & MPOL_F_ADDR) {
956 * Take a refcount on the mpol, because we are about to
957 * drop the mmap_lock, after which only "pol" remains
958 * valid, "vma" is stale.
963 mmap_read_unlock(mm);
964 err = lookup_node(mm, addr);
968 } else if (pol == current->mempolicy &&
969 pol->mode == MPOL_INTERLEAVE) {
970 *policy = next_node_in(current->il_prev, pol->nodes);
976 *policy = pol == &default_policy ? MPOL_DEFAULT :
979 * Internal mempolicy flags must be masked off before exposing
980 * the policy to userspace.
982 *policy |= (pol->flags & MPOL_MODE_FLAGS);
987 if (mpol_store_user_nodemask(pol)) {
988 *nmask = pol->w.user_nodemask;
991 get_policy_nodemask(pol, nmask);
992 task_unlock(current);
999 mmap_read_unlock(mm);
1001 mpol_put(pol_refcount);
1005 #ifdef CONFIG_MIGRATION
1007 * page migration, thp tail pages can be passed.
1009 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1010 unsigned long flags)
1012 struct page *head = compound_head(page);
1014 * Avoid migrating a page that is shared with others.
1016 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1017 if (!isolate_lru_page(head)) {
1018 list_add_tail(&head->lru, pagelist);
1019 mod_node_page_state(page_pgdat(head),
1020 NR_ISOLATED_ANON + page_is_file_lru(head),
1021 thp_nr_pages(head));
1022 } else if (flags & MPOL_MF_STRICT) {
1024 * Non-movable page may reach here. And, there may be
1025 * temporary off LRU pages or non-LRU movable pages.
1026 * Treat them as unmovable pages since they can't be
1027 * isolated, so they can't be moved at the moment. It
1028 * should return -EIO for this case too.
1038 * Migrate pages from one node to a target node.
1039 * Returns error or the number of pages not migrated.
1041 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1045 LIST_HEAD(pagelist);
1047 struct migration_target_control mtc = {
1049 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1053 node_set(source, nmask);
1056 * This does not "check" the range but isolates all pages that
1057 * need migration. Between passing in the full user address
1058 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1060 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1061 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1062 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1064 if (!list_empty(&pagelist)) {
1065 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1066 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1068 putback_movable_pages(&pagelist);
1075 * Move pages between the two nodesets so as to preserve the physical
1076 * layout as much as possible.
1078 * Returns the number of page that could not be moved.
1080 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1081 const nodemask_t *to, int flags)
1087 lru_cache_disable();
1092 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1093 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1094 * bit in 'tmp', and return that <source, dest> pair for migration.
1095 * The pair of nodemasks 'to' and 'from' define the map.
1097 * If no pair of bits is found that way, fallback to picking some
1098 * pair of 'source' and 'dest' bits that are not the same. If the
1099 * 'source' and 'dest' bits are the same, this represents a node
1100 * that will be migrating to itself, so no pages need move.
1102 * If no bits are left in 'tmp', or if all remaining bits left
1103 * in 'tmp' correspond to the same bit in 'to', return false
1104 * (nothing left to migrate).
1106 * This lets us pick a pair of nodes to migrate between, such that
1107 * if possible the dest node is not already occupied by some other
1108 * source node, minimizing the risk of overloading the memory on a
1109 * node that would happen if we migrated incoming memory to a node
1110 * before migrating outgoing memory source that same node.
1112 * A single scan of tmp is sufficient. As we go, we remember the
1113 * most recent <s, d> pair that moved (s != d). If we find a pair
1114 * that not only moved, but what's better, moved to an empty slot
1115 * (d is not set in tmp), then we break out then, with that pair.
1116 * Otherwise when we finish scanning from_tmp, we at least have the
1117 * most recent <s, d> pair that moved. If we get all the way through
1118 * the scan of tmp without finding any node that moved, much less
1119 * moved to an empty node, then there is nothing left worth migrating.
1123 while (!nodes_empty(tmp)) {
1125 int source = NUMA_NO_NODE;
1128 for_each_node_mask(s, tmp) {
1131 * do_migrate_pages() tries to maintain the relative
1132 * node relationship of the pages established between
1133 * threads and memory areas.
1135 * However if the number of source nodes is not equal to
1136 * the number of destination nodes we can not preserve
1137 * this node relative relationship. In that case, skip
1138 * copying memory from a node that is in the destination
1141 * Example: [2,3,4] -> [3,4,5] moves everything.
1142 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1145 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1146 (node_isset(s, *to)))
1149 d = node_remap(s, *from, *to);
1153 source = s; /* Node moved. Memorize */
1156 /* dest not in remaining from nodes? */
1157 if (!node_isset(dest, tmp))
1160 if (source == NUMA_NO_NODE)
1163 node_clear(source, tmp);
1164 err = migrate_to_node(mm, source, dest, flags);
1170 mmap_read_unlock(mm);
1180 * Allocate a new page for page migration based on vma policy.
1181 * Start by assuming the page is mapped by the same vma as contains @start.
1182 * Search forward from there, if not. N.B., this assumes that the
1183 * list of pages handed to migrate_pages()--which is how we get here--
1184 * is in virtual address order.
1186 static struct page *new_page(struct page *page, unsigned long start)
1188 struct folio *dst, *src = page_folio(page);
1189 struct vm_area_struct *vma;
1190 unsigned long address;
1191 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1193 vma = find_vma(current->mm, start);
1195 address = page_address_in_vma(page, vma);
1196 if (address != -EFAULT)
1201 if (folio_test_hugetlb(src))
1202 return alloc_huge_page_vma(page_hstate(&src->page),
1205 if (folio_test_large(src))
1206 gfp = GFP_TRANSHUGE;
1209 * if !vma, vma_alloc_folio() will use task or system default policy
1211 dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1212 folio_test_large(src));
1217 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1218 unsigned long flags)
1223 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1224 const nodemask_t *to, int flags)
1229 static struct page *new_page(struct page *page, unsigned long start)
1235 static long do_mbind(unsigned long start, unsigned long len,
1236 unsigned short mode, unsigned short mode_flags,
1237 nodemask_t *nmask, unsigned long flags)
1239 struct mm_struct *mm = current->mm;
1240 struct mempolicy *new;
1244 LIST_HEAD(pagelist);
1246 if (flags & ~(unsigned long)MPOL_MF_VALID)
1248 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1251 if (start & ~PAGE_MASK)
1254 if (mode == MPOL_DEFAULT)
1255 flags &= ~MPOL_MF_STRICT;
1257 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1265 new = mpol_new(mode, mode_flags, nmask);
1267 return PTR_ERR(new);
1269 if (flags & MPOL_MF_LAZY)
1270 new->flags |= MPOL_F_MOF;
1273 * If we are using the default policy then operation
1274 * on discontinuous address spaces is okay after all
1277 flags |= MPOL_MF_DISCONTIG_OK;
1279 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1280 start, start + len, mode, mode_flags,
1281 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1283 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1285 lru_cache_disable();
1288 NODEMASK_SCRATCH(scratch);
1290 mmap_write_lock(mm);
1291 err = mpol_set_nodemask(new, nmask, scratch);
1293 mmap_write_unlock(mm);
1296 NODEMASK_SCRATCH_FREE(scratch);
1301 ret = queue_pages_range(mm, start, end, nmask,
1302 flags | MPOL_MF_INVERT, &pagelist);
1309 err = mbind_range(mm, start, end, new);
1314 if (!list_empty(&pagelist)) {
1315 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1316 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1317 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1319 putback_movable_pages(&pagelist);
1322 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1326 if (!list_empty(&pagelist))
1327 putback_movable_pages(&pagelist);
1330 mmap_write_unlock(mm);
1333 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1339 * User space interface with variable sized bitmaps for nodelists.
1341 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1342 unsigned long maxnode)
1344 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1347 if (in_compat_syscall())
1348 ret = compat_get_bitmap(mask,
1349 (const compat_ulong_t __user *)nmask,
1352 ret = copy_from_user(mask, nmask,
1353 nlongs * sizeof(unsigned long));
1358 if (maxnode % BITS_PER_LONG)
1359 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1364 /* Copy a node mask from user space. */
1365 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1366 unsigned long maxnode)
1369 nodes_clear(*nodes);
1370 if (maxnode == 0 || !nmask)
1372 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1376 * When the user specified more nodes than supported just check
1377 * if the non supported part is all zero, one word at a time,
1378 * starting at the end.
1380 while (maxnode > MAX_NUMNODES) {
1381 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1384 if (get_bitmap(&t, &nmask[maxnode / BITS_PER_LONG], bits))
1387 if (maxnode - bits >= MAX_NUMNODES) {
1390 maxnode = MAX_NUMNODES;
1391 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1397 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1400 /* Copy a kernel node mask to user space */
1401 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1404 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1405 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1406 bool compat = in_compat_syscall();
1409 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1411 if (copy > nbytes) {
1412 if (copy > PAGE_SIZE)
1414 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1417 maxnode = nr_node_ids;
1421 return compat_put_bitmap((compat_ulong_t __user *)mask,
1422 nodes_addr(*nodes), maxnode);
1424 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1427 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1428 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1430 *flags = *mode & MPOL_MODE_FLAGS;
1431 *mode &= ~MPOL_MODE_FLAGS;
1433 if ((unsigned int)(*mode) >= MPOL_MAX)
1435 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1437 if (*flags & MPOL_F_NUMA_BALANCING) {
1438 if (*mode != MPOL_BIND)
1440 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1445 static long kernel_mbind(unsigned long start, unsigned long len,
1446 unsigned long mode, const unsigned long __user *nmask,
1447 unsigned long maxnode, unsigned int flags)
1449 unsigned short mode_flags;
1454 start = untagged_addr(start);
1455 err = sanitize_mpol_flags(&lmode, &mode_flags);
1459 err = get_nodes(&nodes, nmask, maxnode);
1463 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1466 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1467 unsigned long, home_node, unsigned long, flags)
1469 struct mm_struct *mm = current->mm;
1470 struct vm_area_struct *vma;
1471 struct mempolicy *new;
1472 unsigned long vmstart;
1473 unsigned long vmend;
1477 start = untagged_addr(start);
1478 if (start & ~PAGE_MASK)
1481 * flags is used for future extension if any.
1487 * Check home_node is online to avoid accessing uninitialized
1490 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1493 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1500 mmap_write_lock(mm);
1501 vma = find_vma(mm, start);
1502 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
1504 vmstart = max(start, vma->vm_start);
1505 vmend = min(end, vma->vm_end);
1506 new = mpol_dup(vma_policy(vma));
1512 * Only update home node if there is an existing vma policy
1518 * If any vma in the range got policy other than MPOL_BIND
1519 * or MPOL_PREFERRED_MANY we return error. We don't reset
1520 * the home node for vmas we already updated before.
1522 if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1527 new->home_node = home_node;
1528 err = mbind_range(mm, vmstart, vmend, new);
1533 mmap_write_unlock(mm);
1537 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1538 unsigned long, mode, const unsigned long __user *, nmask,
1539 unsigned long, maxnode, unsigned int, flags)
1541 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1544 /* Set the process memory policy */
1545 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1546 unsigned long maxnode)
1548 unsigned short mode_flags;
1553 err = sanitize_mpol_flags(&lmode, &mode_flags);
1557 err = get_nodes(&nodes, nmask, maxnode);
1561 return do_set_mempolicy(lmode, mode_flags, &nodes);
1564 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1565 unsigned long, maxnode)
1567 return kernel_set_mempolicy(mode, nmask, maxnode);
1570 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1571 const unsigned long __user *old_nodes,
1572 const unsigned long __user *new_nodes)
1574 struct mm_struct *mm = NULL;
1575 struct task_struct *task;
1576 nodemask_t task_nodes;
1580 NODEMASK_SCRATCH(scratch);
1585 old = &scratch->mask1;
1586 new = &scratch->mask2;
1588 err = get_nodes(old, old_nodes, maxnode);
1592 err = get_nodes(new, new_nodes, maxnode);
1596 /* Find the mm_struct */
1598 task = pid ? find_task_by_vpid(pid) : current;
1604 get_task_struct(task);
1609 * Check if this process has the right to modify the specified process.
1610 * Use the regular "ptrace_may_access()" checks.
1612 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1619 task_nodes = cpuset_mems_allowed(task);
1620 /* Is the user allowed to access the target nodes? */
1621 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1626 task_nodes = cpuset_mems_allowed(current);
1627 nodes_and(*new, *new, task_nodes);
1628 if (nodes_empty(*new))
1631 err = security_task_movememory(task);
1635 mm = get_task_mm(task);
1636 put_task_struct(task);
1643 err = do_migrate_pages(mm, old, new,
1644 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1648 NODEMASK_SCRATCH_FREE(scratch);
1653 put_task_struct(task);
1658 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1659 const unsigned long __user *, old_nodes,
1660 const unsigned long __user *, new_nodes)
1662 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1666 /* Retrieve NUMA policy */
1667 static int kernel_get_mempolicy(int __user *policy,
1668 unsigned long __user *nmask,
1669 unsigned long maxnode,
1671 unsigned long flags)
1677 if (nmask != NULL && maxnode < nr_node_ids)
1680 addr = untagged_addr(addr);
1682 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1687 if (policy && put_user(pval, policy))
1691 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1696 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1697 unsigned long __user *, nmask, unsigned long, maxnode,
1698 unsigned long, addr, unsigned long, flags)
1700 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1703 bool vma_migratable(struct vm_area_struct *vma)
1705 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1709 * DAX device mappings require predictable access latency, so avoid
1710 * incurring periodic faults.
1712 if (vma_is_dax(vma))
1715 if (is_vm_hugetlb_page(vma) &&
1716 !hugepage_migration_supported(hstate_vma(vma)))
1720 * Migration allocates pages in the highest zone. If we cannot
1721 * do so then migration (at least from node to node) is not
1725 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1731 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1734 struct mempolicy *pol = NULL;
1737 if (vma->vm_ops && vma->vm_ops->get_policy) {
1738 pol = vma->vm_ops->get_policy(vma, addr);
1739 } else if (vma->vm_policy) {
1740 pol = vma->vm_policy;
1743 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1744 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1745 * count on these policies which will be dropped by
1746 * mpol_cond_put() later
1748 if (mpol_needs_cond_ref(pol))
1757 * get_vma_policy(@vma, @addr)
1758 * @vma: virtual memory area whose policy is sought
1759 * @addr: address in @vma for shared policy lookup
1761 * Returns effective policy for a VMA at specified address.
1762 * Falls back to current->mempolicy or system default policy, as necessary.
1763 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1764 * count--added by the get_policy() vm_op, as appropriate--to protect against
1765 * freeing by another task. It is the caller's responsibility to free the
1766 * extra reference for shared policies.
1768 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1771 struct mempolicy *pol = __get_vma_policy(vma, addr);
1774 pol = get_task_policy(current);
1779 bool vma_policy_mof(struct vm_area_struct *vma)
1781 struct mempolicy *pol;
1783 if (vma->vm_ops && vma->vm_ops->get_policy) {
1786 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1787 if (pol && (pol->flags & MPOL_F_MOF))
1794 pol = vma->vm_policy;
1796 pol = get_task_policy(current);
1798 return pol->flags & MPOL_F_MOF;
1801 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1803 enum zone_type dynamic_policy_zone = policy_zone;
1805 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1808 * if policy->nodes has movable memory only,
1809 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1811 * policy->nodes is intersect with node_states[N_MEMORY].
1812 * so if the following test fails, it implies
1813 * policy->nodes has movable memory only.
1815 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1816 dynamic_policy_zone = ZONE_MOVABLE;
1818 return zone >= dynamic_policy_zone;
1822 * Return a nodemask representing a mempolicy for filtering nodes for
1825 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1827 int mode = policy->mode;
1829 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1830 if (unlikely(mode == MPOL_BIND) &&
1831 apply_policy_zone(policy, gfp_zone(gfp)) &&
1832 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1833 return &policy->nodes;
1835 if (mode == MPOL_PREFERRED_MANY)
1836 return &policy->nodes;
1842 * Return the preferred node id for 'prefer' mempolicy, and return
1843 * the given id for all other policies.
1845 * policy_node() is always coupled with policy_nodemask(), which
1846 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1848 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1850 if (policy->mode == MPOL_PREFERRED) {
1851 nd = first_node(policy->nodes);
1854 * __GFP_THISNODE shouldn't even be used with the bind policy
1855 * because we might easily break the expectation to stay on the
1856 * requested node and not break the policy.
1858 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1861 if ((policy->mode == MPOL_BIND ||
1862 policy->mode == MPOL_PREFERRED_MANY) &&
1863 policy->home_node != NUMA_NO_NODE)
1864 return policy->home_node;
1869 /* Do dynamic interleaving for a process */
1870 static unsigned interleave_nodes(struct mempolicy *policy)
1873 struct task_struct *me = current;
1875 next = next_node_in(me->il_prev, policy->nodes);
1876 if (next < MAX_NUMNODES)
1882 * Depending on the memory policy provide a node from which to allocate the
1885 unsigned int mempolicy_slab_node(void)
1887 struct mempolicy *policy;
1888 int node = numa_mem_id();
1893 policy = current->mempolicy;
1897 switch (policy->mode) {
1898 case MPOL_PREFERRED:
1899 return first_node(policy->nodes);
1901 case MPOL_INTERLEAVE:
1902 return interleave_nodes(policy);
1905 case MPOL_PREFERRED_MANY:
1910 * Follow bind policy behavior and start allocation at the
1913 struct zonelist *zonelist;
1914 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1915 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1916 z = first_zones_zonelist(zonelist, highest_zoneidx,
1918 return z->zone ? zone_to_nid(z->zone) : node;
1929 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1930 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1931 * number of present nodes.
1933 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1935 nodemask_t nodemask = pol->nodes;
1936 unsigned int target, nnodes;
1940 * The barrier will stabilize the nodemask in a register or on
1941 * the stack so that it will stop changing under the code.
1943 * Between first_node() and next_node(), pol->nodes could be changed
1944 * by other threads. So we put pol->nodes in a local stack.
1948 nnodes = nodes_weight(nodemask);
1950 return numa_node_id();
1951 target = (unsigned int)n % nnodes;
1952 nid = first_node(nodemask);
1953 for (i = 0; i < target; i++)
1954 nid = next_node(nid, nodemask);
1958 /* Determine a node number for interleave */
1959 static inline unsigned interleave_nid(struct mempolicy *pol,
1960 struct vm_area_struct *vma, unsigned long addr, int shift)
1966 * for small pages, there is no difference between
1967 * shift and PAGE_SHIFT, so the bit-shift is safe.
1968 * for huge pages, since vm_pgoff is in units of small
1969 * pages, we need to shift off the always 0 bits to get
1972 BUG_ON(shift < PAGE_SHIFT);
1973 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1974 off += (addr - vma->vm_start) >> shift;
1975 return offset_il_node(pol, off);
1977 return interleave_nodes(pol);
1980 #ifdef CONFIG_HUGETLBFS
1982 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1983 * @vma: virtual memory area whose policy is sought
1984 * @addr: address in @vma for shared policy lookup and interleave policy
1985 * @gfp_flags: for requested zone
1986 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1987 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1989 * Returns a nid suitable for a huge page allocation and a pointer
1990 * to the struct mempolicy for conditional unref after allocation.
1991 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
1992 * to the mempolicy's @nodemask for filtering the zonelist.
1994 * Must be protected by read_mems_allowed_begin()
1996 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1997 struct mempolicy **mpol, nodemask_t **nodemask)
2002 *mpol = get_vma_policy(vma, addr);
2004 mode = (*mpol)->mode;
2006 if (unlikely(mode == MPOL_INTERLEAVE)) {
2007 nid = interleave_nid(*mpol, vma, addr,
2008 huge_page_shift(hstate_vma(vma)));
2010 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2011 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2012 *nodemask = &(*mpol)->nodes;
2018 * init_nodemask_of_mempolicy
2020 * If the current task's mempolicy is "default" [NULL], return 'false'
2021 * to indicate default policy. Otherwise, extract the policy nodemask
2022 * for 'bind' or 'interleave' policy into the argument nodemask, or
2023 * initialize the argument nodemask to contain the single node for
2024 * 'preferred' or 'local' policy and return 'true' to indicate presence
2025 * of non-default mempolicy.
2027 * We don't bother with reference counting the mempolicy [mpol_get/put]
2028 * because the current task is examining it's own mempolicy and a task's
2029 * mempolicy is only ever changed by the task itself.
2031 * N.B., it is the caller's responsibility to free a returned nodemask.
2033 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2035 struct mempolicy *mempolicy;
2037 if (!(mask && current->mempolicy))
2041 mempolicy = current->mempolicy;
2042 switch (mempolicy->mode) {
2043 case MPOL_PREFERRED:
2044 case MPOL_PREFERRED_MANY:
2046 case MPOL_INTERLEAVE:
2047 *mask = mempolicy->nodes;
2051 init_nodemask_of_node(mask, numa_node_id());
2057 task_unlock(current);
2064 * mempolicy_in_oom_domain
2066 * If tsk's mempolicy is "bind", check for intersection between mask and
2067 * the policy nodemask. Otherwise, return true for all other policies
2068 * including "interleave", as a tsk with "interleave" policy may have
2069 * memory allocated from all nodes in system.
2071 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2073 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2074 const nodemask_t *mask)
2076 struct mempolicy *mempolicy;
2083 mempolicy = tsk->mempolicy;
2084 if (mempolicy && mempolicy->mode == MPOL_BIND)
2085 ret = nodes_intersects(mempolicy->nodes, *mask);
2091 /* Allocate a page in interleaved policy.
2092 Own path because it needs to do special accounting. */
2093 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2098 page = __alloc_pages(gfp, order, nid, NULL);
2099 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2100 if (!static_branch_likely(&vm_numa_stat_key))
2102 if (page && page_to_nid(page) == nid) {
2104 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2110 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2111 int nid, struct mempolicy *pol)
2114 gfp_t preferred_gfp;
2117 * This is a two pass approach. The first pass will only try the
2118 * preferred nodes but skip the direct reclaim and allow the
2119 * allocation to fail, while the second pass will try all the
2122 preferred_gfp = gfp | __GFP_NOWARN;
2123 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2124 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2126 page = __alloc_pages(gfp, order, nid, NULL);
2132 * alloc_pages_vma - Allocate a page for a VMA.
2134 * @order: Order of the GFP allocation.
2135 * @vma: Pointer to VMA or NULL if not available.
2136 * @addr: Virtual address of the allocation. Must be inside @vma.
2137 * @hugepage: For hugepages try only the preferred node if possible.
2139 * Allocate a page for a specific address in @vma, using the appropriate
2140 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2141 * of the mm_struct of the VMA to prevent it from going away. Should be
2142 * used for all allocations for pages that will be mapped into user space.
2144 * Return: The page on success or NULL if allocation fails.
2146 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2147 unsigned long addr, bool hugepage)
2149 struct mempolicy *pol;
2150 int node = numa_node_id();
2155 pol = get_vma_policy(vma, addr);
2157 if (pol->mode == MPOL_INTERLEAVE) {
2160 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2162 page = alloc_page_interleave(gfp, order, nid);
2166 if (pol->mode == MPOL_PREFERRED_MANY) {
2167 node = policy_node(gfp, pol, node);
2168 page = alloc_pages_preferred_many(gfp, order, node, pol);
2173 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2174 int hpage_node = node;
2177 * For hugepage allocation and non-interleave policy which
2178 * allows the current node (or other explicitly preferred
2179 * node) we only try to allocate from the current/preferred
2180 * node and don't fall back to other nodes, as the cost of
2181 * remote accesses would likely offset THP benefits.
2183 * If the policy is interleave or does not allow the current
2184 * node in its nodemask, we allocate the standard way.
2186 if (pol->mode == MPOL_PREFERRED)
2187 hpage_node = first_node(pol->nodes);
2189 nmask = policy_nodemask(gfp, pol);
2190 if (!nmask || node_isset(hpage_node, *nmask)) {
2193 * First, try to allocate THP only on local node, but
2194 * don't reclaim unnecessarily, just compact.
2196 page = __alloc_pages_node(hpage_node,
2197 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2200 * If hugepage allocations are configured to always
2201 * synchronous compact or the vma has been madvised
2202 * to prefer hugepage backing, retry allowing remote
2203 * memory with both reclaim and compact as well.
2205 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2206 page = __alloc_pages(gfp, order, hpage_node, nmask);
2212 nmask = policy_nodemask(gfp, pol);
2213 preferred_nid = policy_node(gfp, pol, node);
2214 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2219 EXPORT_SYMBOL(alloc_pages_vma);
2221 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2222 unsigned long addr, bool hugepage)
2224 struct folio *folio;
2226 folio = (struct folio *)alloc_pages_vma(gfp, order, vma, addr,
2228 if (folio && order > 1)
2229 prep_transhuge_page(&folio->page);
2235 * alloc_pages - Allocate pages.
2237 * @order: Power of two of number of pages to allocate.
2239 * Allocate 1 << @order contiguous pages. The physical address of the
2240 * first page is naturally aligned (eg an order-3 allocation will be aligned
2241 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2242 * process is honoured when in process context.
2244 * Context: Can be called from any context, providing the appropriate GFP
2246 * Return: The page on success or NULL if allocation fails.
2248 struct page *alloc_pages(gfp_t gfp, unsigned order)
2250 struct mempolicy *pol = &default_policy;
2253 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2254 pol = get_task_policy(current);
2257 * No reference counting needed for current->mempolicy
2258 * nor system default_policy
2260 if (pol->mode == MPOL_INTERLEAVE)
2261 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2262 else if (pol->mode == MPOL_PREFERRED_MANY)
2263 page = alloc_pages_preferred_many(gfp, order,
2264 policy_node(gfp, pol, numa_node_id()), pol);
2266 page = __alloc_pages(gfp, order,
2267 policy_node(gfp, pol, numa_node_id()),
2268 policy_nodemask(gfp, pol));
2272 EXPORT_SYMBOL(alloc_pages);
2274 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2276 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2278 if (page && order > 1)
2279 prep_transhuge_page(page);
2280 return (struct folio *)page;
2282 EXPORT_SYMBOL(folio_alloc);
2284 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2285 struct mempolicy *pol, unsigned long nr_pages,
2286 struct page **page_array)
2289 unsigned long nr_pages_per_node;
2292 unsigned long nr_allocated;
2293 unsigned long total_allocated = 0;
2295 nodes = nodes_weight(pol->nodes);
2296 nr_pages_per_node = nr_pages / nodes;
2297 delta = nr_pages - nodes * nr_pages_per_node;
2299 for (i = 0; i < nodes; i++) {
2301 nr_allocated = __alloc_pages_bulk(gfp,
2302 interleave_nodes(pol), NULL,
2303 nr_pages_per_node + 1, NULL,
2307 nr_allocated = __alloc_pages_bulk(gfp,
2308 interleave_nodes(pol), NULL,
2309 nr_pages_per_node, NULL, page_array);
2312 page_array += nr_allocated;
2313 total_allocated += nr_allocated;
2316 return total_allocated;
2319 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2320 struct mempolicy *pol, unsigned long nr_pages,
2321 struct page **page_array)
2323 gfp_t preferred_gfp;
2324 unsigned long nr_allocated = 0;
2326 preferred_gfp = gfp | __GFP_NOWARN;
2327 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2329 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2330 nr_pages, NULL, page_array);
2332 if (nr_allocated < nr_pages)
2333 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2334 nr_pages - nr_allocated, NULL,
2335 page_array + nr_allocated);
2336 return nr_allocated;
2339 /* alloc pages bulk and mempolicy should be considered at the
2340 * same time in some situation such as vmalloc.
2342 * It can accelerate memory allocation especially interleaving
2345 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2346 unsigned long nr_pages, struct page **page_array)
2348 struct mempolicy *pol = &default_policy;
2350 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2351 pol = get_task_policy(current);
2353 if (pol->mode == MPOL_INTERLEAVE)
2354 return alloc_pages_bulk_array_interleave(gfp, pol,
2355 nr_pages, page_array);
2357 if (pol->mode == MPOL_PREFERRED_MANY)
2358 return alloc_pages_bulk_array_preferred_many(gfp,
2359 numa_node_id(), pol, nr_pages, page_array);
2361 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2362 policy_nodemask(gfp, pol), nr_pages, NULL,
2366 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2368 struct mempolicy *pol = mpol_dup(vma_policy(src));
2371 return PTR_ERR(pol);
2372 dst->vm_policy = pol;
2377 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2378 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2379 * with the mems_allowed returned by cpuset_mems_allowed(). This
2380 * keeps mempolicies cpuset relative after its cpuset moves. See
2381 * further kernel/cpuset.c update_nodemask().
2383 * current's mempolicy may be rebinded by the other task(the task that changes
2384 * cpuset's mems), so we needn't do rebind work for current task.
2387 /* Slow path of a mempolicy duplicate */
2388 struct mempolicy *__mpol_dup(struct mempolicy *old)
2390 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2393 return ERR_PTR(-ENOMEM);
2395 /* task's mempolicy is protected by alloc_lock */
2396 if (old == current->mempolicy) {
2399 task_unlock(current);
2403 if (current_cpuset_is_being_rebound()) {
2404 nodemask_t mems = cpuset_mems_allowed(current);
2405 mpol_rebind_policy(new, &mems);
2407 atomic_set(&new->refcnt, 1);
2411 /* Slow path of a mempolicy comparison */
2412 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2416 if (a->mode != b->mode)
2418 if (a->flags != b->flags)
2420 if (a->home_node != b->home_node)
2422 if (mpol_store_user_nodemask(a))
2423 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2428 case MPOL_INTERLEAVE:
2429 case MPOL_PREFERRED:
2430 case MPOL_PREFERRED_MANY:
2431 return !!nodes_equal(a->nodes, b->nodes);
2441 * Shared memory backing store policy support.
2443 * Remember policies even when nobody has shared memory mapped.
2444 * The policies are kept in Red-Black tree linked from the inode.
2445 * They are protected by the sp->lock rwlock, which should be held
2446 * for any accesses to the tree.
2450 * lookup first element intersecting start-end. Caller holds sp->lock for
2451 * reading or for writing
2453 static struct sp_node *
2454 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2456 struct rb_node *n = sp->root.rb_node;
2459 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2461 if (start >= p->end)
2463 else if (end <= p->start)
2471 struct sp_node *w = NULL;
2472 struct rb_node *prev = rb_prev(n);
2475 w = rb_entry(prev, struct sp_node, nd);
2476 if (w->end <= start)
2480 return rb_entry(n, struct sp_node, nd);
2484 * Insert a new shared policy into the list. Caller holds sp->lock for
2487 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2489 struct rb_node **p = &sp->root.rb_node;
2490 struct rb_node *parent = NULL;
2495 nd = rb_entry(parent, struct sp_node, nd);
2496 if (new->start < nd->start)
2498 else if (new->end > nd->end)
2499 p = &(*p)->rb_right;
2503 rb_link_node(&new->nd, parent, p);
2504 rb_insert_color(&new->nd, &sp->root);
2505 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2506 new->policy ? new->policy->mode : 0);
2509 /* Find shared policy intersecting idx */
2511 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2513 struct mempolicy *pol = NULL;
2516 if (!sp->root.rb_node)
2518 read_lock(&sp->lock);
2519 sn = sp_lookup(sp, idx, idx+1);
2521 mpol_get(sn->policy);
2524 read_unlock(&sp->lock);
2528 static void sp_free(struct sp_node *n)
2530 mpol_put(n->policy);
2531 kmem_cache_free(sn_cache, n);
2535 * mpol_misplaced - check whether current page node is valid in policy
2537 * @page: page to be checked
2538 * @vma: vm area where page mapped
2539 * @addr: virtual address where page mapped
2541 * Lookup current policy node id for vma,addr and "compare to" page's
2542 * node id. Policy determination "mimics" alloc_page_vma().
2543 * Called from fault path where we know the vma and faulting address.
2545 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2546 * policy, or a suitable node ID to allocate a replacement page from.
2548 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2550 struct mempolicy *pol;
2552 int curnid = page_to_nid(page);
2553 unsigned long pgoff;
2554 int thiscpu = raw_smp_processor_id();
2555 int thisnid = cpu_to_node(thiscpu);
2556 int polnid = NUMA_NO_NODE;
2557 int ret = NUMA_NO_NODE;
2559 pol = get_vma_policy(vma, addr);
2560 if (!(pol->flags & MPOL_F_MOF))
2563 switch (pol->mode) {
2564 case MPOL_INTERLEAVE:
2565 pgoff = vma->vm_pgoff;
2566 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2567 polnid = offset_il_node(pol, pgoff);
2570 case MPOL_PREFERRED:
2571 if (node_isset(curnid, pol->nodes))
2573 polnid = first_node(pol->nodes);
2577 polnid = numa_node_id();
2581 /* Optimize placement among multiple nodes via NUMA balancing */
2582 if (pol->flags & MPOL_F_MORON) {
2583 if (node_isset(thisnid, pol->nodes))
2589 case MPOL_PREFERRED_MANY:
2591 * use current page if in policy nodemask,
2592 * else select nearest allowed node, if any.
2593 * If no allowed nodes, use current [!misplaced].
2595 if (node_isset(curnid, pol->nodes))
2597 z = first_zones_zonelist(
2598 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2599 gfp_zone(GFP_HIGHUSER),
2601 polnid = zone_to_nid(z->zone);
2608 /* Migrate the page towards the node whose CPU is referencing it */
2609 if (pol->flags & MPOL_F_MORON) {
2612 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2616 if (curnid != polnid)
2625 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2626 * dropped after task->mempolicy is set to NULL so that any allocation done as
2627 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2630 void mpol_put_task_policy(struct task_struct *task)
2632 struct mempolicy *pol;
2635 pol = task->mempolicy;
2636 task->mempolicy = NULL;
2641 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2643 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2644 rb_erase(&n->nd, &sp->root);
2648 static void sp_node_init(struct sp_node *node, unsigned long start,
2649 unsigned long end, struct mempolicy *pol)
2651 node->start = start;
2656 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2657 struct mempolicy *pol)
2660 struct mempolicy *newpol;
2662 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2666 newpol = mpol_dup(pol);
2667 if (IS_ERR(newpol)) {
2668 kmem_cache_free(sn_cache, n);
2671 newpol->flags |= MPOL_F_SHARED;
2672 sp_node_init(n, start, end, newpol);
2677 /* Replace a policy range. */
2678 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2679 unsigned long end, struct sp_node *new)
2682 struct sp_node *n_new = NULL;
2683 struct mempolicy *mpol_new = NULL;
2687 write_lock(&sp->lock);
2688 n = sp_lookup(sp, start, end);
2689 /* Take care of old policies in the same range. */
2690 while (n && n->start < end) {
2691 struct rb_node *next = rb_next(&n->nd);
2692 if (n->start >= start) {
2698 /* Old policy spanning whole new range. */
2703 *mpol_new = *n->policy;
2704 atomic_set(&mpol_new->refcnt, 1);
2705 sp_node_init(n_new, end, n->end, mpol_new);
2707 sp_insert(sp, n_new);
2716 n = rb_entry(next, struct sp_node, nd);
2720 write_unlock(&sp->lock);
2727 kmem_cache_free(sn_cache, n_new);
2732 write_unlock(&sp->lock);
2734 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2737 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2740 atomic_set(&mpol_new->refcnt, 1);
2745 * mpol_shared_policy_init - initialize shared policy for inode
2746 * @sp: pointer to inode shared policy
2747 * @mpol: struct mempolicy to install
2749 * Install non-NULL @mpol in inode's shared policy rb-tree.
2750 * On entry, the current task has a reference on a non-NULL @mpol.
2751 * This must be released on exit.
2752 * This is called at get_inode() calls and we can use GFP_KERNEL.
2754 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2758 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2759 rwlock_init(&sp->lock);
2762 struct vm_area_struct pvma;
2763 struct mempolicy *new;
2764 NODEMASK_SCRATCH(scratch);
2768 /* contextualize the tmpfs mount point mempolicy */
2769 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2771 goto free_scratch; /* no valid nodemask intersection */
2774 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2775 task_unlock(current);
2779 /* Create pseudo-vma that contains just the policy */
2780 vma_init(&pvma, NULL);
2781 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2782 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2785 mpol_put(new); /* drop initial ref */
2787 NODEMASK_SCRATCH_FREE(scratch);
2789 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2793 int mpol_set_shared_policy(struct shared_policy *info,
2794 struct vm_area_struct *vma, struct mempolicy *npol)
2797 struct sp_node *new = NULL;
2798 unsigned long sz = vma_pages(vma);
2800 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2802 sz, npol ? npol->mode : -1,
2803 npol ? npol->flags : -1,
2804 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2807 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2811 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2817 /* Free a backing policy store on inode delete. */
2818 void mpol_free_shared_policy(struct shared_policy *p)
2821 struct rb_node *next;
2823 if (!p->root.rb_node)
2825 write_lock(&p->lock);
2826 next = rb_first(&p->root);
2828 n = rb_entry(next, struct sp_node, nd);
2829 next = rb_next(&n->nd);
2832 write_unlock(&p->lock);
2835 #ifdef CONFIG_NUMA_BALANCING
2836 static int __initdata numabalancing_override;
2838 static void __init check_numabalancing_enable(void)
2840 bool numabalancing_default = false;
2842 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2843 numabalancing_default = true;
2845 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2846 if (numabalancing_override)
2847 set_numabalancing_state(numabalancing_override == 1);
2849 if (num_online_nodes() > 1 && !numabalancing_override) {
2850 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2851 numabalancing_default ? "Enabling" : "Disabling");
2852 set_numabalancing_state(numabalancing_default);
2856 static int __init setup_numabalancing(char *str)
2862 if (!strcmp(str, "enable")) {
2863 numabalancing_override = 1;
2865 } else if (!strcmp(str, "disable")) {
2866 numabalancing_override = -1;
2871 pr_warn("Unable to parse numa_balancing=\n");
2875 __setup("numa_balancing=", setup_numabalancing);
2877 static inline void __init check_numabalancing_enable(void)
2880 #endif /* CONFIG_NUMA_BALANCING */
2882 /* assumes fs == KERNEL_DS */
2883 void __init numa_policy_init(void)
2885 nodemask_t interleave_nodes;
2886 unsigned long largest = 0;
2887 int nid, prefer = 0;
2889 policy_cache = kmem_cache_create("numa_policy",
2890 sizeof(struct mempolicy),
2891 0, SLAB_PANIC, NULL);
2893 sn_cache = kmem_cache_create("shared_policy_node",
2894 sizeof(struct sp_node),
2895 0, SLAB_PANIC, NULL);
2897 for_each_node(nid) {
2898 preferred_node_policy[nid] = (struct mempolicy) {
2899 .refcnt = ATOMIC_INIT(1),
2900 .mode = MPOL_PREFERRED,
2901 .flags = MPOL_F_MOF | MPOL_F_MORON,
2902 .nodes = nodemask_of_node(nid),
2907 * Set interleaving policy for system init. Interleaving is only
2908 * enabled across suitably sized nodes (default is >= 16MB), or
2909 * fall back to the largest node if they're all smaller.
2911 nodes_clear(interleave_nodes);
2912 for_each_node_state(nid, N_MEMORY) {
2913 unsigned long total_pages = node_present_pages(nid);
2915 /* Preserve the largest node */
2916 if (largest < total_pages) {
2917 largest = total_pages;
2921 /* Interleave this node? */
2922 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2923 node_set(nid, interleave_nodes);
2926 /* All too small, use the largest */
2927 if (unlikely(nodes_empty(interleave_nodes)))
2928 node_set(prefer, interleave_nodes);
2930 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2931 pr_err("%s: interleaving failed\n", __func__);
2933 check_numabalancing_enable();
2936 /* Reset policy of current process to default */
2937 void numa_default_policy(void)
2939 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2943 * Parse and format mempolicy from/to strings
2946 static const char * const policy_modes[] =
2948 [MPOL_DEFAULT] = "default",
2949 [MPOL_PREFERRED] = "prefer",
2950 [MPOL_BIND] = "bind",
2951 [MPOL_INTERLEAVE] = "interleave",
2952 [MPOL_LOCAL] = "local",
2953 [MPOL_PREFERRED_MANY] = "prefer (many)",
2959 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2960 * @str: string containing mempolicy to parse
2961 * @mpol: pointer to struct mempolicy pointer, returned on success.
2964 * <mode>[=<flags>][:<nodelist>]
2966 * Return: %0 on success, else %1
2968 int mpol_parse_str(char *str, struct mempolicy **mpol)
2970 struct mempolicy *new = NULL;
2971 unsigned short mode_flags;
2973 char *nodelist = strchr(str, ':');
2974 char *flags = strchr(str, '=');
2978 *flags++ = '\0'; /* terminate mode string */
2981 /* NUL-terminate mode or flags string */
2983 if (nodelist_parse(nodelist, nodes))
2985 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2990 mode = match_string(policy_modes, MPOL_MAX, str);
2995 case MPOL_PREFERRED:
2997 * Insist on a nodelist of one node only, although later
2998 * we use first_node(nodes) to grab a single node, so here
2999 * nodelist (or nodes) cannot be empty.
3002 char *rest = nodelist;
3003 while (isdigit(*rest))
3007 if (nodes_empty(nodes))
3011 case MPOL_INTERLEAVE:
3013 * Default to online nodes with memory if no nodelist
3016 nodes = node_states[N_MEMORY];
3020 * Don't allow a nodelist; mpol_new() checks flags
3027 * Insist on a empty nodelist
3032 case MPOL_PREFERRED_MANY:
3035 * Insist on a nodelist
3044 * Currently, we only support two mutually exclusive
3047 if (!strcmp(flags, "static"))
3048 mode_flags |= MPOL_F_STATIC_NODES;
3049 else if (!strcmp(flags, "relative"))
3050 mode_flags |= MPOL_F_RELATIVE_NODES;
3055 new = mpol_new(mode, mode_flags, &nodes);
3060 * Save nodes for mpol_to_str() to show the tmpfs mount options
3061 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3063 if (mode != MPOL_PREFERRED) {
3065 } else if (nodelist) {
3066 nodes_clear(new->nodes);
3067 node_set(first_node(nodes), new->nodes);
3069 new->mode = MPOL_LOCAL;
3073 * Save nodes for contextualization: this will be used to "clone"
3074 * the mempolicy in a specific context [cpuset] at a later time.
3076 new->w.user_nodemask = nodes;
3081 /* Restore string for error message */
3090 #endif /* CONFIG_TMPFS */
3093 * mpol_to_str - format a mempolicy structure for printing
3094 * @buffer: to contain formatted mempolicy string
3095 * @maxlen: length of @buffer
3096 * @pol: pointer to mempolicy to be formatted
3098 * Convert @pol into a string. If @buffer is too short, truncate the string.
3099 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3100 * longest flag, "relative", and to display at least a few node ids.
3102 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3105 nodemask_t nodes = NODE_MASK_NONE;
3106 unsigned short mode = MPOL_DEFAULT;
3107 unsigned short flags = 0;
3109 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3118 case MPOL_PREFERRED:
3119 case MPOL_PREFERRED_MANY:
3121 case MPOL_INTERLEAVE:
3126 snprintf(p, maxlen, "unknown");
3130 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3132 if (flags & MPOL_MODE_FLAGS) {
3133 p += snprintf(p, buffer + maxlen - p, "=");
3136 * Currently, the only defined flags are mutually exclusive
3138 if (flags & MPOL_F_STATIC_NODES)
3139 p += snprintf(p, buffer + maxlen - p, "static");
3140 else if (flags & MPOL_F_RELATIVE_NODES)
3141 p += snprintf(p, buffer + maxlen - p, "relative");
3144 if (!nodes_empty(nodes))
3145 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3146 nodemask_pr_args(&nodes));