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 mpol_rebind_policy(vma->vm_policy, new);
389 mmap_write_unlock(mm);
392 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
394 .rebind = mpol_rebind_default,
396 [MPOL_INTERLEAVE] = {
397 .create = mpol_new_nodemask,
398 .rebind = mpol_rebind_nodemask,
401 .create = mpol_new_preferred,
402 .rebind = mpol_rebind_preferred,
405 .create = mpol_new_nodemask,
406 .rebind = mpol_rebind_nodemask,
409 .rebind = mpol_rebind_default,
411 [MPOL_PREFERRED_MANY] = {
412 .create = mpol_new_nodemask,
413 .rebind = mpol_rebind_preferred,
417 static int migrate_page_add(struct page *page, struct list_head *pagelist,
418 unsigned long flags);
421 struct list_head *pagelist;
426 struct vm_area_struct *first;
430 * Check if the page's nid is in qp->nmask.
432 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
433 * in the invert of qp->nmask.
435 static inline bool queue_pages_required(struct page *page,
436 struct queue_pages *qp)
438 int nid = page_to_nid(page);
439 unsigned long flags = qp->flags;
441 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
445 * queue_pages_pmd() has three possible return values:
446 * 0 - pages are placed on the right node or queued successfully, or
447 * special page is met, i.e. huge zero page.
448 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
450 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
451 * existing page was already on a node that does not follow the
454 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
455 unsigned long end, struct mm_walk *walk)
460 struct queue_pages *qp = walk->private;
463 if (unlikely(is_pmd_migration_entry(*pmd))) {
467 page = pmd_page(*pmd);
468 if (is_huge_zero_page(page)) {
469 walk->action = ACTION_CONTINUE;
472 if (!queue_pages_required(page, qp))
476 /* go to thp migration */
477 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
478 if (!vma_migratable(walk->vma) ||
479 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);
525 if (!page || is_zone_device_page(page))
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 !hugetlb_pmd_shared(pte))) {
605 if (isolate_hugetlb(page, qp->pagelist) &&
606 (flags & MPOL_MF_STRICT))
608 * Failed to isolate page but allow migrating pages
609 * which have been queued.
621 #ifdef CONFIG_NUMA_BALANCING
623 * This is used to mark a range of virtual addresses to be inaccessible.
624 * These are later cleared by a NUMA hinting fault. Depending on these
625 * faults, pages may be migrated for better NUMA placement.
627 * This is assuming that NUMA faults are handled using PROT_NONE. If
628 * an architecture makes a different choice, it will need further
629 * changes to the core.
631 unsigned long change_prot_numa(struct vm_area_struct *vma,
632 unsigned long addr, unsigned long end)
634 struct mmu_gather tlb;
637 tlb_gather_mmu(&tlb, vma->vm_mm);
639 nr_updated = change_protection(&tlb, vma, addr, end, PAGE_NONE,
642 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
644 tlb_finish_mmu(&tlb);
649 static unsigned long change_prot_numa(struct vm_area_struct *vma,
650 unsigned long addr, unsigned long end)
654 #endif /* CONFIG_NUMA_BALANCING */
656 static int queue_pages_test_walk(unsigned long start, unsigned long end,
657 struct mm_walk *walk)
659 struct vm_area_struct *next, *vma = walk->vma;
660 struct queue_pages *qp = walk->private;
661 unsigned long endvma = vma->vm_end;
662 unsigned long flags = qp->flags;
664 /* range check first */
665 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
669 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
670 (qp->start < vma->vm_start))
671 /* hole at head side of range */
674 next = find_vma(vma->vm_mm, vma->vm_end);
675 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
676 ((vma->vm_end < qp->end) &&
677 (!next || vma->vm_end < next->vm_start)))
678 /* hole at middle or tail of range */
682 * Need check MPOL_MF_STRICT to return -EIO if possible
683 * regardless of vma_migratable
685 if (!vma_migratable(vma) &&
686 !(flags & MPOL_MF_STRICT))
692 if (flags & MPOL_MF_LAZY) {
693 /* Similar to task_numa_work, skip inaccessible VMAs */
694 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
695 !(vma->vm_flags & VM_MIXEDMAP))
696 change_prot_numa(vma, start, endvma);
700 /* queue pages from current vma */
701 if (flags & MPOL_MF_VALID)
706 static const struct mm_walk_ops queue_pages_walk_ops = {
707 .hugetlb_entry = queue_pages_hugetlb,
708 .pmd_entry = queue_pages_pte_range,
709 .test_walk = queue_pages_test_walk,
713 * Walk through page tables and collect pages to be migrated.
715 * If pages found in a given range are on a set of nodes (determined by
716 * @nodes and @flags,) it's isolated and queued to the pagelist which is
717 * passed via @private.
719 * queue_pages_range() has three possible return values:
720 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
722 * 0 - queue pages successfully or no misplaced page.
723 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
724 * memory range specified by nodemask and maxnode points outside
725 * your accessible address space (-EFAULT)
728 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
729 nodemask_t *nodes, unsigned long flags,
730 struct list_head *pagelist)
733 struct queue_pages qp = {
734 .pagelist = pagelist,
742 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
745 /* whole range in hole */
752 * Apply policy to a single VMA
753 * This must be called with the mmap_lock held for writing.
755 static int vma_replace_policy(struct vm_area_struct *vma,
756 struct mempolicy *pol)
759 struct mempolicy *old;
760 struct mempolicy *new;
762 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
763 vma->vm_start, vma->vm_end, vma->vm_pgoff,
764 vma->vm_ops, vma->vm_file,
765 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
771 if (vma->vm_ops && vma->vm_ops->set_policy) {
772 err = vma->vm_ops->set_policy(vma, new);
777 old = vma->vm_policy;
778 vma->vm_policy = new; /* protected by mmap_lock */
787 /* Step 2: apply policy to a range and do splits. */
788 static int mbind_range(struct mm_struct *mm, unsigned long start,
789 unsigned long end, struct mempolicy *new_pol)
791 MA_STATE(mas, &mm->mm_mt, start, start);
792 struct vm_area_struct *prev;
793 struct vm_area_struct *vma;
797 prev = mas_prev(&mas, 0);
799 mas_set(&mas, start);
801 vma = mas_find(&mas, end - 1);
805 if (start > vma->vm_start)
808 for (; vma; vma = mas_next(&mas, end - 1)) {
809 unsigned long vmstart = max(start, vma->vm_start);
810 unsigned long vmend = min(end, vma->vm_end);
812 if (mpol_equal(vma_policy(vma), new_pol))
815 pgoff = vma->vm_pgoff +
816 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
817 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
818 vma->anon_vma, vma->vm_file, pgoff,
819 new_pol, vma->vm_userfaultfd_ctx,
822 /* vma_merge() invalidated the mas */
827 if (vma->vm_start != vmstart) {
828 err = split_vma(vma->vm_mm, vma, vmstart, 1);
831 /* split_vma() invalidated the mas */
834 if (vma->vm_end != vmend) {
835 err = split_vma(vma->vm_mm, vma, vmend, 0);
838 /* split_vma() invalidated the mas */
842 err = vma_replace_policy(vma, new_pol);
853 /* Set the process memory policy */
854 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
857 struct mempolicy *new, *old;
858 NODEMASK_SCRATCH(scratch);
864 new = mpol_new(mode, flags, nodes);
871 ret = mpol_set_nodemask(new, nodes, scratch);
873 task_unlock(current);
878 old = current->mempolicy;
879 current->mempolicy = new;
880 if (new && new->mode == MPOL_INTERLEAVE)
881 current->il_prev = MAX_NUMNODES-1;
882 task_unlock(current);
886 NODEMASK_SCRATCH_FREE(scratch);
891 * Return nodemask for policy for get_mempolicy() query
893 * Called with task's alloc_lock held
895 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
898 if (p == &default_policy)
903 case MPOL_INTERLEAVE:
905 case MPOL_PREFERRED_MANY:
909 /* return empty node mask for local allocation */
916 static int lookup_node(struct mm_struct *mm, unsigned long addr)
918 struct page *p = NULL;
921 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
923 ret = page_to_nid(p);
929 /* Retrieve NUMA policy */
930 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
931 unsigned long addr, unsigned long flags)
934 struct mm_struct *mm = current->mm;
935 struct vm_area_struct *vma = NULL;
936 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
939 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
942 if (flags & MPOL_F_MEMS_ALLOWED) {
943 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
945 *policy = 0; /* just so it's initialized */
947 *nmask = cpuset_current_mems_allowed;
948 task_unlock(current);
952 if (flags & MPOL_F_ADDR) {
954 * Do NOT fall back to task policy if the
955 * vma/shared policy at addr is NULL. We
956 * want to return MPOL_DEFAULT in this case.
959 vma = vma_lookup(mm, addr);
961 mmap_read_unlock(mm);
964 if (vma->vm_ops && vma->vm_ops->get_policy)
965 pol = vma->vm_ops->get_policy(vma, addr);
967 pol = vma->vm_policy;
972 pol = &default_policy; /* indicates default behavior */
974 if (flags & MPOL_F_NODE) {
975 if (flags & MPOL_F_ADDR) {
977 * Take a refcount on the mpol, because we are about to
978 * drop the mmap_lock, after which only "pol" remains
979 * valid, "vma" is stale.
984 mmap_read_unlock(mm);
985 err = lookup_node(mm, addr);
989 } else if (pol == current->mempolicy &&
990 pol->mode == MPOL_INTERLEAVE) {
991 *policy = next_node_in(current->il_prev, pol->nodes);
997 *policy = pol == &default_policy ? MPOL_DEFAULT :
1000 * Internal mempolicy flags must be masked off before exposing
1001 * the policy to userspace.
1003 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1008 if (mpol_store_user_nodemask(pol)) {
1009 *nmask = pol->w.user_nodemask;
1012 get_policy_nodemask(pol, nmask);
1013 task_unlock(current);
1020 mmap_read_unlock(mm);
1022 mpol_put(pol_refcount);
1026 #ifdef CONFIG_MIGRATION
1028 * page migration, thp tail pages can be passed.
1030 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1031 unsigned long flags)
1033 struct page *head = compound_head(page);
1035 * Avoid migrating a page that is shared with others.
1037 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1038 if (!isolate_lru_page(head)) {
1039 list_add_tail(&head->lru, pagelist);
1040 mod_node_page_state(page_pgdat(head),
1041 NR_ISOLATED_ANON + page_is_file_lru(head),
1042 thp_nr_pages(head));
1043 } else if (flags & MPOL_MF_STRICT) {
1045 * Non-movable page may reach here. And, there may be
1046 * temporary off LRU pages or non-LRU movable pages.
1047 * Treat them as unmovable pages since they can't be
1048 * isolated, so they can't be moved at the moment. It
1049 * should return -EIO for this case too.
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1062 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1066 struct vm_area_struct *vma;
1067 LIST_HEAD(pagelist);
1069 struct migration_target_control mtc = {
1071 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1075 node_set(source, nmask);
1078 * This does not "check" the range but isolates all pages that
1079 * need migration. Between passing in the full user address
1080 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1082 vma = find_vma(mm, 0);
1083 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1084 queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1085 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1087 if (!list_empty(&pagelist)) {
1088 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1089 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1091 putback_movable_pages(&pagelist);
1098 * Move pages between the two nodesets so as to preserve the physical
1099 * layout as much as possible.
1101 * Returns the number of page that could not be moved.
1103 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1104 const nodemask_t *to, int flags)
1110 lru_cache_disable();
1115 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1116 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1117 * bit in 'tmp', and return that <source, dest> pair for migration.
1118 * The pair of nodemasks 'to' and 'from' define the map.
1120 * If no pair of bits is found that way, fallback to picking some
1121 * pair of 'source' and 'dest' bits that are not the same. If the
1122 * 'source' and 'dest' bits are the same, this represents a node
1123 * that will be migrating to itself, so no pages need move.
1125 * If no bits are left in 'tmp', or if all remaining bits left
1126 * in 'tmp' correspond to the same bit in 'to', return false
1127 * (nothing left to migrate).
1129 * This lets us pick a pair of nodes to migrate between, such that
1130 * if possible the dest node is not already occupied by some other
1131 * source node, minimizing the risk of overloading the memory on a
1132 * node that would happen if we migrated incoming memory to a node
1133 * before migrating outgoing memory source that same node.
1135 * A single scan of tmp is sufficient. As we go, we remember the
1136 * most recent <s, d> pair that moved (s != d). If we find a pair
1137 * that not only moved, but what's better, moved to an empty slot
1138 * (d is not set in tmp), then we break out then, with that pair.
1139 * Otherwise when we finish scanning from_tmp, we at least have the
1140 * most recent <s, d> pair that moved. If we get all the way through
1141 * the scan of tmp without finding any node that moved, much less
1142 * moved to an empty node, then there is nothing left worth migrating.
1146 while (!nodes_empty(tmp)) {
1148 int source = NUMA_NO_NODE;
1151 for_each_node_mask(s, tmp) {
1154 * do_migrate_pages() tries to maintain the relative
1155 * node relationship of the pages established between
1156 * threads and memory areas.
1158 * However if the number of source nodes is not equal to
1159 * the number of destination nodes we can not preserve
1160 * this node relative relationship. In that case, skip
1161 * copying memory from a node that is in the destination
1164 * Example: [2,3,4] -> [3,4,5] moves everything.
1165 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1168 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1169 (node_isset(s, *to)))
1172 d = node_remap(s, *from, *to);
1176 source = s; /* Node moved. Memorize */
1179 /* dest not in remaining from nodes? */
1180 if (!node_isset(dest, tmp))
1183 if (source == NUMA_NO_NODE)
1186 node_clear(source, tmp);
1187 err = migrate_to_node(mm, source, dest, flags);
1193 mmap_read_unlock(mm);
1203 * Allocate a new page for page migration based on vma policy.
1204 * Start by assuming the page is mapped by the same vma as contains @start.
1205 * Search forward from there, if not. N.B., this assumes that the
1206 * list of pages handed to migrate_pages()--which is how we get here--
1207 * is in virtual address order.
1209 static struct page *new_page(struct page *page, unsigned long start)
1211 struct folio *dst, *src = page_folio(page);
1212 struct vm_area_struct *vma;
1213 unsigned long address;
1214 VMA_ITERATOR(vmi, current->mm, start);
1215 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1217 for_each_vma(vmi, vma) {
1218 address = page_address_in_vma(page, vma);
1219 if (address != -EFAULT)
1223 if (folio_test_hugetlb(src))
1224 return alloc_huge_page_vma(page_hstate(&src->page),
1227 if (folio_test_large(src))
1228 gfp = GFP_TRANSHUGE;
1231 * if !vma, vma_alloc_folio() will use task or system default policy
1233 dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1234 folio_test_large(src));
1239 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1240 unsigned long flags)
1245 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1246 const nodemask_t *to, int flags)
1251 static struct page *new_page(struct page *page, unsigned long start)
1257 static long do_mbind(unsigned long start, unsigned long len,
1258 unsigned short mode, unsigned short mode_flags,
1259 nodemask_t *nmask, unsigned long flags)
1261 struct mm_struct *mm = current->mm;
1262 struct mempolicy *new;
1266 LIST_HEAD(pagelist);
1268 if (flags & ~(unsigned long)MPOL_MF_VALID)
1270 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1273 if (start & ~PAGE_MASK)
1276 if (mode == MPOL_DEFAULT)
1277 flags &= ~MPOL_MF_STRICT;
1279 len = PAGE_ALIGN(len);
1287 new = mpol_new(mode, mode_flags, nmask);
1289 return PTR_ERR(new);
1291 if (flags & MPOL_MF_LAZY)
1292 new->flags |= MPOL_F_MOF;
1295 * If we are using the default policy then operation
1296 * on discontinuous address spaces is okay after all
1299 flags |= MPOL_MF_DISCONTIG_OK;
1301 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1302 start, start + len, mode, mode_flags,
1303 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1305 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1307 lru_cache_disable();
1310 NODEMASK_SCRATCH(scratch);
1312 mmap_write_lock(mm);
1313 err = mpol_set_nodemask(new, nmask, scratch);
1315 mmap_write_unlock(mm);
1318 NODEMASK_SCRATCH_FREE(scratch);
1323 ret = queue_pages_range(mm, start, end, nmask,
1324 flags | MPOL_MF_INVERT, &pagelist);
1331 err = mbind_range(mm, start, end, new);
1336 if (!list_empty(&pagelist)) {
1337 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1338 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1339 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1341 putback_movable_pages(&pagelist);
1344 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1348 if (!list_empty(&pagelist))
1349 putback_movable_pages(&pagelist);
1352 mmap_write_unlock(mm);
1355 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1361 * User space interface with variable sized bitmaps for nodelists.
1363 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1364 unsigned long maxnode)
1366 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1369 if (in_compat_syscall())
1370 ret = compat_get_bitmap(mask,
1371 (const compat_ulong_t __user *)nmask,
1374 ret = copy_from_user(mask, nmask,
1375 nlongs * sizeof(unsigned long));
1380 if (maxnode % BITS_PER_LONG)
1381 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1386 /* Copy a node mask from user space. */
1387 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1388 unsigned long maxnode)
1391 nodes_clear(*nodes);
1392 if (maxnode == 0 || !nmask)
1394 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1398 * When the user specified more nodes than supported just check
1399 * if the non supported part is all zero, one word at a time,
1400 * starting at the end.
1402 while (maxnode > MAX_NUMNODES) {
1403 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1406 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1409 if (maxnode - bits >= MAX_NUMNODES) {
1412 maxnode = MAX_NUMNODES;
1413 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1419 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1422 /* Copy a kernel node mask to user space */
1423 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1426 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1427 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1428 bool compat = in_compat_syscall();
1431 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1433 if (copy > nbytes) {
1434 if (copy > PAGE_SIZE)
1436 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1439 maxnode = nr_node_ids;
1443 return compat_put_bitmap((compat_ulong_t __user *)mask,
1444 nodes_addr(*nodes), maxnode);
1446 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1449 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1450 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1452 *flags = *mode & MPOL_MODE_FLAGS;
1453 *mode &= ~MPOL_MODE_FLAGS;
1455 if ((unsigned int)(*mode) >= MPOL_MAX)
1457 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1459 if (*flags & MPOL_F_NUMA_BALANCING) {
1460 if (*mode != MPOL_BIND)
1462 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1467 static long kernel_mbind(unsigned long start, unsigned long len,
1468 unsigned long mode, const unsigned long __user *nmask,
1469 unsigned long maxnode, unsigned int flags)
1471 unsigned short mode_flags;
1476 start = untagged_addr(start);
1477 err = sanitize_mpol_flags(&lmode, &mode_flags);
1481 err = get_nodes(&nodes, nmask, maxnode);
1485 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1488 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1489 unsigned long, home_node, unsigned long, flags)
1491 struct mm_struct *mm = current->mm;
1492 struct vm_area_struct *vma;
1493 struct mempolicy *new;
1494 unsigned long vmstart;
1495 unsigned long vmend;
1498 VMA_ITERATOR(vmi, mm, start);
1500 start = untagged_addr(start);
1501 if (start & ~PAGE_MASK)
1504 * flags is used for future extension if any.
1510 * Check home_node is online to avoid accessing uninitialized
1513 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1516 len = PAGE_ALIGN(len);
1523 mmap_write_lock(mm);
1524 for_each_vma_range(vmi, vma, end) {
1525 vmstart = max(start, vma->vm_start);
1526 vmend = min(end, vma->vm_end);
1527 new = mpol_dup(vma_policy(vma));
1533 * Only update home node if there is an existing vma policy
1539 * If any vma in the range got policy other than MPOL_BIND
1540 * or MPOL_PREFERRED_MANY we return error. We don't reset
1541 * the home node for vmas we already updated before.
1543 if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1549 new->home_node = home_node;
1550 err = mbind_range(mm, vmstart, vmend, new);
1555 mmap_write_unlock(mm);
1559 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1560 unsigned long, mode, const unsigned long __user *, nmask,
1561 unsigned long, maxnode, unsigned int, flags)
1563 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1566 /* Set the process memory policy */
1567 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1568 unsigned long maxnode)
1570 unsigned short mode_flags;
1575 err = sanitize_mpol_flags(&lmode, &mode_flags);
1579 err = get_nodes(&nodes, nmask, maxnode);
1583 return do_set_mempolicy(lmode, mode_flags, &nodes);
1586 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1587 unsigned long, maxnode)
1589 return kernel_set_mempolicy(mode, nmask, maxnode);
1592 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1593 const unsigned long __user *old_nodes,
1594 const unsigned long __user *new_nodes)
1596 struct mm_struct *mm = NULL;
1597 struct task_struct *task;
1598 nodemask_t task_nodes;
1602 NODEMASK_SCRATCH(scratch);
1607 old = &scratch->mask1;
1608 new = &scratch->mask2;
1610 err = get_nodes(old, old_nodes, maxnode);
1614 err = get_nodes(new, new_nodes, maxnode);
1618 /* Find the mm_struct */
1620 task = pid ? find_task_by_vpid(pid) : current;
1626 get_task_struct(task);
1631 * Check if this process has the right to modify the specified process.
1632 * Use the regular "ptrace_may_access()" checks.
1634 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1641 task_nodes = cpuset_mems_allowed(task);
1642 /* Is the user allowed to access the target nodes? */
1643 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1648 task_nodes = cpuset_mems_allowed(current);
1649 nodes_and(*new, *new, task_nodes);
1650 if (nodes_empty(*new))
1653 err = security_task_movememory(task);
1657 mm = get_task_mm(task);
1658 put_task_struct(task);
1665 err = do_migrate_pages(mm, old, new,
1666 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1670 NODEMASK_SCRATCH_FREE(scratch);
1675 put_task_struct(task);
1680 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1681 const unsigned long __user *, old_nodes,
1682 const unsigned long __user *, new_nodes)
1684 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1688 /* Retrieve NUMA policy */
1689 static int kernel_get_mempolicy(int __user *policy,
1690 unsigned long __user *nmask,
1691 unsigned long maxnode,
1693 unsigned long flags)
1699 if (nmask != NULL && maxnode < nr_node_ids)
1702 addr = untagged_addr(addr);
1704 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1709 if (policy && put_user(pval, policy))
1713 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1718 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1719 unsigned long __user *, nmask, unsigned long, maxnode,
1720 unsigned long, addr, unsigned long, flags)
1722 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1725 bool vma_migratable(struct vm_area_struct *vma)
1727 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1731 * DAX device mappings require predictable access latency, so avoid
1732 * incurring periodic faults.
1734 if (vma_is_dax(vma))
1737 if (is_vm_hugetlb_page(vma) &&
1738 !hugepage_migration_supported(hstate_vma(vma)))
1742 * Migration allocates pages in the highest zone. If we cannot
1743 * do so then migration (at least from node to node) is not
1747 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1753 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1756 struct mempolicy *pol = NULL;
1759 if (vma->vm_ops && vma->vm_ops->get_policy) {
1760 pol = vma->vm_ops->get_policy(vma, addr);
1761 } else if (vma->vm_policy) {
1762 pol = vma->vm_policy;
1765 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1766 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1767 * count on these policies which will be dropped by
1768 * mpol_cond_put() later
1770 if (mpol_needs_cond_ref(pol))
1779 * get_vma_policy(@vma, @addr)
1780 * @vma: virtual memory area whose policy is sought
1781 * @addr: address in @vma for shared policy lookup
1783 * Returns effective policy for a VMA at specified address.
1784 * Falls back to current->mempolicy or system default policy, as necessary.
1785 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1786 * count--added by the get_policy() vm_op, as appropriate--to protect against
1787 * freeing by another task. It is the caller's responsibility to free the
1788 * extra reference for shared policies.
1790 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1793 struct mempolicy *pol = __get_vma_policy(vma, addr);
1796 pol = get_task_policy(current);
1801 bool vma_policy_mof(struct vm_area_struct *vma)
1803 struct mempolicy *pol;
1805 if (vma->vm_ops && vma->vm_ops->get_policy) {
1808 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1809 if (pol && (pol->flags & MPOL_F_MOF))
1816 pol = vma->vm_policy;
1818 pol = get_task_policy(current);
1820 return pol->flags & MPOL_F_MOF;
1823 bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1825 enum zone_type dynamic_policy_zone = policy_zone;
1827 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1830 * if policy->nodes has movable memory only,
1831 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1833 * policy->nodes is intersect with node_states[N_MEMORY].
1834 * so if the following test fails, it implies
1835 * policy->nodes has movable memory only.
1837 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1838 dynamic_policy_zone = ZONE_MOVABLE;
1840 return zone >= dynamic_policy_zone;
1844 * Return a nodemask representing a mempolicy for filtering nodes for
1847 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1849 int mode = policy->mode;
1851 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1852 if (unlikely(mode == MPOL_BIND) &&
1853 apply_policy_zone(policy, gfp_zone(gfp)) &&
1854 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1855 return &policy->nodes;
1857 if (mode == MPOL_PREFERRED_MANY)
1858 return &policy->nodes;
1864 * Return the preferred node id for 'prefer' mempolicy, and return
1865 * the given id for all other policies.
1867 * policy_node() is always coupled with policy_nodemask(), which
1868 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1870 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1872 if (policy->mode == MPOL_PREFERRED) {
1873 nd = first_node(policy->nodes);
1876 * __GFP_THISNODE shouldn't even be used with the bind policy
1877 * because we might easily break the expectation to stay on the
1878 * requested node and not break the policy.
1880 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1883 if ((policy->mode == MPOL_BIND ||
1884 policy->mode == MPOL_PREFERRED_MANY) &&
1885 policy->home_node != NUMA_NO_NODE)
1886 return policy->home_node;
1891 /* Do dynamic interleaving for a process */
1892 static unsigned interleave_nodes(struct mempolicy *policy)
1895 struct task_struct *me = current;
1897 next = next_node_in(me->il_prev, policy->nodes);
1898 if (next < MAX_NUMNODES)
1904 * Depending on the memory policy provide a node from which to allocate the
1907 unsigned int mempolicy_slab_node(void)
1909 struct mempolicy *policy;
1910 int node = numa_mem_id();
1915 policy = current->mempolicy;
1919 switch (policy->mode) {
1920 case MPOL_PREFERRED:
1921 return first_node(policy->nodes);
1923 case MPOL_INTERLEAVE:
1924 return interleave_nodes(policy);
1927 case MPOL_PREFERRED_MANY:
1932 * Follow bind policy behavior and start allocation at the
1935 struct zonelist *zonelist;
1936 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1937 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1938 z = first_zones_zonelist(zonelist, highest_zoneidx,
1940 return z->zone ? zone_to_nid(z->zone) : node;
1951 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1952 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1953 * number of present nodes.
1955 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1957 nodemask_t nodemask = pol->nodes;
1958 unsigned int target, nnodes;
1962 * The barrier will stabilize the nodemask in a register or on
1963 * the stack so that it will stop changing under the code.
1965 * Between first_node() and next_node(), pol->nodes could be changed
1966 * by other threads. So we put pol->nodes in a local stack.
1970 nnodes = nodes_weight(nodemask);
1972 return numa_node_id();
1973 target = (unsigned int)n % nnodes;
1974 nid = first_node(nodemask);
1975 for (i = 0; i < target; i++)
1976 nid = next_node(nid, nodemask);
1980 /* Determine a node number for interleave */
1981 static inline unsigned interleave_nid(struct mempolicy *pol,
1982 struct vm_area_struct *vma, unsigned long addr, int shift)
1988 * for small pages, there is no difference between
1989 * shift and PAGE_SHIFT, so the bit-shift is safe.
1990 * for huge pages, since vm_pgoff is in units of small
1991 * pages, we need to shift off the always 0 bits to get
1994 BUG_ON(shift < PAGE_SHIFT);
1995 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1996 off += (addr - vma->vm_start) >> shift;
1997 return offset_il_node(pol, off);
1999 return interleave_nodes(pol);
2002 #ifdef CONFIG_HUGETLBFS
2004 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2005 * @vma: virtual memory area whose policy is sought
2006 * @addr: address in @vma for shared policy lookup and interleave policy
2007 * @gfp_flags: for requested zone
2008 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2009 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2011 * Returns a nid suitable for a huge page allocation and a pointer
2012 * to the struct mempolicy for conditional unref after allocation.
2013 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2014 * to the mempolicy's @nodemask for filtering the zonelist.
2016 * Must be protected by read_mems_allowed_begin()
2018 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2019 struct mempolicy **mpol, nodemask_t **nodemask)
2024 *mpol = get_vma_policy(vma, addr);
2026 mode = (*mpol)->mode;
2028 if (unlikely(mode == MPOL_INTERLEAVE)) {
2029 nid = interleave_nid(*mpol, vma, addr,
2030 huge_page_shift(hstate_vma(vma)));
2032 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2033 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2034 *nodemask = &(*mpol)->nodes;
2040 * init_nodemask_of_mempolicy
2042 * If the current task's mempolicy is "default" [NULL], return 'false'
2043 * to indicate default policy. Otherwise, extract the policy nodemask
2044 * for 'bind' or 'interleave' policy into the argument nodemask, or
2045 * initialize the argument nodemask to contain the single node for
2046 * 'preferred' or 'local' policy and return 'true' to indicate presence
2047 * of non-default mempolicy.
2049 * We don't bother with reference counting the mempolicy [mpol_get/put]
2050 * because the current task is examining it's own mempolicy and a task's
2051 * mempolicy is only ever changed by the task itself.
2053 * N.B., it is the caller's responsibility to free a returned nodemask.
2055 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2057 struct mempolicy *mempolicy;
2059 if (!(mask && current->mempolicy))
2063 mempolicy = current->mempolicy;
2064 switch (mempolicy->mode) {
2065 case MPOL_PREFERRED:
2066 case MPOL_PREFERRED_MANY:
2068 case MPOL_INTERLEAVE:
2069 *mask = mempolicy->nodes;
2073 init_nodemask_of_node(mask, numa_node_id());
2079 task_unlock(current);
2086 * mempolicy_in_oom_domain
2088 * If tsk's mempolicy is "bind", check for intersection between mask and
2089 * the policy nodemask. Otherwise, return true for all other policies
2090 * including "interleave", as a tsk with "interleave" policy may have
2091 * memory allocated from all nodes in system.
2093 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2095 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2096 const nodemask_t *mask)
2098 struct mempolicy *mempolicy;
2105 mempolicy = tsk->mempolicy;
2106 if (mempolicy && mempolicy->mode == MPOL_BIND)
2107 ret = nodes_intersects(mempolicy->nodes, *mask);
2113 /* Allocate a page in interleaved policy.
2114 Own path because it needs to do special accounting. */
2115 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2120 page = __alloc_pages(gfp, order, nid, NULL);
2121 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2122 if (!static_branch_likely(&vm_numa_stat_key))
2124 if (page && page_to_nid(page) == nid) {
2126 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2132 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2133 int nid, struct mempolicy *pol)
2136 gfp_t preferred_gfp;
2139 * This is a two pass approach. The first pass will only try the
2140 * preferred nodes but skip the direct reclaim and allow the
2141 * allocation to fail, while the second pass will try all the
2144 preferred_gfp = gfp | __GFP_NOWARN;
2145 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2146 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2148 page = __alloc_pages(gfp, order, nid, NULL);
2154 * vma_alloc_folio - Allocate a folio for a VMA.
2156 * @order: Order of the folio.
2157 * @vma: Pointer to VMA or NULL if not available.
2158 * @addr: Virtual address of the allocation. Must be inside @vma.
2159 * @hugepage: For hugepages try only the preferred node if possible.
2161 * Allocate a folio for a specific address in @vma, using the appropriate
2162 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2163 * of the mm_struct of the VMA to prevent it from going away. Should be
2164 * used for all allocations for folios that will be mapped into user space.
2166 * Return: The folio on success or NULL if allocation fails.
2168 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2169 unsigned long addr, bool hugepage)
2171 struct mempolicy *pol;
2172 int node = numa_node_id();
2173 struct folio *folio;
2177 pol = get_vma_policy(vma, addr);
2179 if (pol->mode == MPOL_INTERLEAVE) {
2183 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2186 page = alloc_page_interleave(gfp, order, nid);
2187 if (page && order > 1)
2188 prep_transhuge_page(page);
2189 folio = (struct folio *)page;
2193 if (pol->mode == MPOL_PREFERRED_MANY) {
2196 node = policy_node(gfp, pol, node);
2198 page = alloc_pages_preferred_many(gfp, order, node, pol);
2200 if (page && order > 1)
2201 prep_transhuge_page(page);
2202 folio = (struct folio *)page;
2206 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2207 int hpage_node = node;
2210 * For hugepage allocation and non-interleave policy which
2211 * allows the current node (or other explicitly preferred
2212 * node) we only try to allocate from the current/preferred
2213 * node and don't fall back to other nodes, as the cost of
2214 * remote accesses would likely offset THP benefits.
2216 * If the policy is interleave or does not allow the current
2217 * node in its nodemask, we allocate the standard way.
2219 if (pol->mode == MPOL_PREFERRED)
2220 hpage_node = first_node(pol->nodes);
2222 nmask = policy_nodemask(gfp, pol);
2223 if (!nmask || node_isset(hpage_node, *nmask)) {
2226 * First, try to allocate THP only on local node, but
2227 * don't reclaim unnecessarily, just compact.
2229 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2230 __GFP_NORETRY, order, hpage_node);
2233 * If hugepage allocations are configured to always
2234 * synchronous compact or the vma has been madvised
2235 * to prefer hugepage backing, retry allowing remote
2236 * memory with both reclaim and compact as well.
2238 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2239 folio = __folio_alloc(gfp, order, hpage_node,
2246 nmask = policy_nodemask(gfp, pol);
2247 preferred_nid = policy_node(gfp, pol, node);
2248 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2253 EXPORT_SYMBOL(vma_alloc_folio);
2256 * alloc_pages - Allocate pages.
2258 * @order: Power of two of number of pages to allocate.
2260 * Allocate 1 << @order contiguous pages. The physical address of the
2261 * first page is naturally aligned (eg an order-3 allocation will be aligned
2262 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2263 * process is honoured when in process context.
2265 * Context: Can be called from any context, providing the appropriate GFP
2267 * Return: The page on success or NULL if allocation fails.
2269 struct page *alloc_pages(gfp_t gfp, unsigned order)
2271 struct mempolicy *pol = &default_policy;
2274 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2275 pol = get_task_policy(current);
2278 * No reference counting needed for current->mempolicy
2279 * nor system default_policy
2281 if (pol->mode == MPOL_INTERLEAVE)
2282 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2283 else if (pol->mode == MPOL_PREFERRED_MANY)
2284 page = alloc_pages_preferred_many(gfp, order,
2285 policy_node(gfp, pol, numa_node_id()), pol);
2287 page = __alloc_pages(gfp, order,
2288 policy_node(gfp, pol, numa_node_id()),
2289 policy_nodemask(gfp, pol));
2293 EXPORT_SYMBOL(alloc_pages);
2295 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2297 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2299 if (page && order > 1)
2300 prep_transhuge_page(page);
2301 return (struct folio *)page;
2303 EXPORT_SYMBOL(folio_alloc);
2305 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2306 struct mempolicy *pol, unsigned long nr_pages,
2307 struct page **page_array)
2310 unsigned long nr_pages_per_node;
2313 unsigned long nr_allocated;
2314 unsigned long total_allocated = 0;
2316 nodes = nodes_weight(pol->nodes);
2317 nr_pages_per_node = nr_pages / nodes;
2318 delta = nr_pages - nodes * nr_pages_per_node;
2320 for (i = 0; i < nodes; i++) {
2322 nr_allocated = __alloc_pages_bulk(gfp,
2323 interleave_nodes(pol), NULL,
2324 nr_pages_per_node + 1, NULL,
2328 nr_allocated = __alloc_pages_bulk(gfp,
2329 interleave_nodes(pol), NULL,
2330 nr_pages_per_node, NULL, page_array);
2333 page_array += nr_allocated;
2334 total_allocated += nr_allocated;
2337 return total_allocated;
2340 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2341 struct mempolicy *pol, unsigned long nr_pages,
2342 struct page **page_array)
2344 gfp_t preferred_gfp;
2345 unsigned long nr_allocated = 0;
2347 preferred_gfp = gfp | __GFP_NOWARN;
2348 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2350 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2351 nr_pages, NULL, page_array);
2353 if (nr_allocated < nr_pages)
2354 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2355 nr_pages - nr_allocated, NULL,
2356 page_array + nr_allocated);
2357 return nr_allocated;
2360 /* alloc pages bulk and mempolicy should be considered at the
2361 * same time in some situation such as vmalloc.
2363 * It can accelerate memory allocation especially interleaving
2366 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2367 unsigned long nr_pages, struct page **page_array)
2369 struct mempolicy *pol = &default_policy;
2371 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2372 pol = get_task_policy(current);
2374 if (pol->mode == MPOL_INTERLEAVE)
2375 return alloc_pages_bulk_array_interleave(gfp, pol,
2376 nr_pages, page_array);
2378 if (pol->mode == MPOL_PREFERRED_MANY)
2379 return alloc_pages_bulk_array_preferred_many(gfp,
2380 numa_node_id(), pol, nr_pages, page_array);
2382 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2383 policy_nodemask(gfp, pol), nr_pages, NULL,
2387 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2389 struct mempolicy *pol = mpol_dup(vma_policy(src));
2392 return PTR_ERR(pol);
2393 dst->vm_policy = pol;
2398 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2399 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2400 * with the mems_allowed returned by cpuset_mems_allowed(). This
2401 * keeps mempolicies cpuset relative after its cpuset moves. See
2402 * further kernel/cpuset.c update_nodemask().
2404 * current's mempolicy may be rebinded by the other task(the task that changes
2405 * cpuset's mems), so we needn't do rebind work for current task.
2408 /* Slow path of a mempolicy duplicate */
2409 struct mempolicy *__mpol_dup(struct mempolicy *old)
2411 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2414 return ERR_PTR(-ENOMEM);
2416 /* task's mempolicy is protected by alloc_lock */
2417 if (old == current->mempolicy) {
2420 task_unlock(current);
2424 if (current_cpuset_is_being_rebound()) {
2425 nodemask_t mems = cpuset_mems_allowed(current);
2426 mpol_rebind_policy(new, &mems);
2428 atomic_set(&new->refcnt, 1);
2432 /* Slow path of a mempolicy comparison */
2433 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2437 if (a->mode != b->mode)
2439 if (a->flags != b->flags)
2441 if (a->home_node != b->home_node)
2443 if (mpol_store_user_nodemask(a))
2444 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2449 case MPOL_INTERLEAVE:
2450 case MPOL_PREFERRED:
2451 case MPOL_PREFERRED_MANY:
2452 return !!nodes_equal(a->nodes, b->nodes);
2462 * Shared memory backing store policy support.
2464 * Remember policies even when nobody has shared memory mapped.
2465 * The policies are kept in Red-Black tree linked from the inode.
2466 * They are protected by the sp->lock rwlock, which should be held
2467 * for any accesses to the tree.
2471 * lookup first element intersecting start-end. Caller holds sp->lock for
2472 * reading or for writing
2474 static struct sp_node *
2475 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2477 struct rb_node *n = sp->root.rb_node;
2480 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2482 if (start >= p->end)
2484 else if (end <= p->start)
2492 struct sp_node *w = NULL;
2493 struct rb_node *prev = rb_prev(n);
2496 w = rb_entry(prev, struct sp_node, nd);
2497 if (w->end <= start)
2501 return rb_entry(n, struct sp_node, nd);
2505 * Insert a new shared policy into the list. Caller holds sp->lock for
2508 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2510 struct rb_node **p = &sp->root.rb_node;
2511 struct rb_node *parent = NULL;
2516 nd = rb_entry(parent, struct sp_node, nd);
2517 if (new->start < nd->start)
2519 else if (new->end > nd->end)
2520 p = &(*p)->rb_right;
2524 rb_link_node(&new->nd, parent, p);
2525 rb_insert_color(&new->nd, &sp->root);
2526 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2527 new->policy ? new->policy->mode : 0);
2530 /* Find shared policy intersecting idx */
2532 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2534 struct mempolicy *pol = NULL;
2537 if (!sp->root.rb_node)
2539 read_lock(&sp->lock);
2540 sn = sp_lookup(sp, idx, idx+1);
2542 mpol_get(sn->policy);
2545 read_unlock(&sp->lock);
2549 static void sp_free(struct sp_node *n)
2551 mpol_put(n->policy);
2552 kmem_cache_free(sn_cache, n);
2556 * mpol_misplaced - check whether current page node is valid in policy
2558 * @page: page to be checked
2559 * @vma: vm area where page mapped
2560 * @addr: virtual address where page mapped
2562 * Lookup current policy node id for vma,addr and "compare to" page's
2563 * node id. Policy determination "mimics" alloc_page_vma().
2564 * Called from fault path where we know the vma and faulting address.
2566 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2567 * policy, or a suitable node ID to allocate a replacement page from.
2569 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2571 struct mempolicy *pol;
2573 int curnid = page_to_nid(page);
2574 unsigned long pgoff;
2575 int thiscpu = raw_smp_processor_id();
2576 int thisnid = cpu_to_node(thiscpu);
2577 int polnid = NUMA_NO_NODE;
2578 int ret = NUMA_NO_NODE;
2580 pol = get_vma_policy(vma, addr);
2581 if (!(pol->flags & MPOL_F_MOF))
2584 switch (pol->mode) {
2585 case MPOL_INTERLEAVE:
2586 pgoff = vma->vm_pgoff;
2587 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2588 polnid = offset_il_node(pol, pgoff);
2591 case MPOL_PREFERRED:
2592 if (node_isset(curnid, pol->nodes))
2594 polnid = first_node(pol->nodes);
2598 polnid = numa_node_id();
2602 /* Optimize placement among multiple nodes via NUMA balancing */
2603 if (pol->flags & MPOL_F_MORON) {
2604 if (node_isset(thisnid, pol->nodes))
2610 case MPOL_PREFERRED_MANY:
2612 * use current page if in policy nodemask,
2613 * else select nearest allowed node, if any.
2614 * If no allowed nodes, use current [!misplaced].
2616 if (node_isset(curnid, pol->nodes))
2618 z = first_zones_zonelist(
2619 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2620 gfp_zone(GFP_HIGHUSER),
2622 polnid = zone_to_nid(z->zone);
2629 /* Migrate the page towards the node whose CPU is referencing it */
2630 if (pol->flags & MPOL_F_MORON) {
2633 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2637 if (curnid != polnid)
2646 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2647 * dropped after task->mempolicy is set to NULL so that any allocation done as
2648 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2651 void mpol_put_task_policy(struct task_struct *task)
2653 struct mempolicy *pol;
2656 pol = task->mempolicy;
2657 task->mempolicy = NULL;
2662 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2664 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2665 rb_erase(&n->nd, &sp->root);
2669 static void sp_node_init(struct sp_node *node, unsigned long start,
2670 unsigned long end, struct mempolicy *pol)
2672 node->start = start;
2677 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2678 struct mempolicy *pol)
2681 struct mempolicy *newpol;
2683 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2687 newpol = mpol_dup(pol);
2688 if (IS_ERR(newpol)) {
2689 kmem_cache_free(sn_cache, n);
2692 newpol->flags |= MPOL_F_SHARED;
2693 sp_node_init(n, start, end, newpol);
2698 /* Replace a policy range. */
2699 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2700 unsigned long end, struct sp_node *new)
2703 struct sp_node *n_new = NULL;
2704 struct mempolicy *mpol_new = NULL;
2708 write_lock(&sp->lock);
2709 n = sp_lookup(sp, start, end);
2710 /* Take care of old policies in the same range. */
2711 while (n && n->start < end) {
2712 struct rb_node *next = rb_next(&n->nd);
2713 if (n->start >= start) {
2719 /* Old policy spanning whole new range. */
2724 *mpol_new = *n->policy;
2725 atomic_set(&mpol_new->refcnt, 1);
2726 sp_node_init(n_new, end, n->end, mpol_new);
2728 sp_insert(sp, n_new);
2737 n = rb_entry(next, struct sp_node, nd);
2741 write_unlock(&sp->lock);
2748 kmem_cache_free(sn_cache, n_new);
2753 write_unlock(&sp->lock);
2755 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2758 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2761 atomic_set(&mpol_new->refcnt, 1);
2766 * mpol_shared_policy_init - initialize shared policy for inode
2767 * @sp: pointer to inode shared policy
2768 * @mpol: struct mempolicy to install
2770 * Install non-NULL @mpol in inode's shared policy rb-tree.
2771 * On entry, the current task has a reference on a non-NULL @mpol.
2772 * This must be released on exit.
2773 * This is called at get_inode() calls and we can use GFP_KERNEL.
2775 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2779 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2780 rwlock_init(&sp->lock);
2783 struct vm_area_struct pvma;
2784 struct mempolicy *new;
2785 NODEMASK_SCRATCH(scratch);
2789 /* contextualize the tmpfs mount point mempolicy */
2790 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2792 goto free_scratch; /* no valid nodemask intersection */
2795 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2796 task_unlock(current);
2800 /* Create pseudo-vma that contains just the policy */
2801 vma_init(&pvma, NULL);
2802 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2803 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2806 mpol_put(new); /* drop initial ref */
2808 NODEMASK_SCRATCH_FREE(scratch);
2810 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2814 int mpol_set_shared_policy(struct shared_policy *info,
2815 struct vm_area_struct *vma, struct mempolicy *npol)
2818 struct sp_node *new = NULL;
2819 unsigned long sz = vma_pages(vma);
2821 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2823 sz, npol ? npol->mode : -1,
2824 npol ? npol->flags : -1,
2825 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2828 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2832 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2838 /* Free a backing policy store on inode delete. */
2839 void mpol_free_shared_policy(struct shared_policy *p)
2842 struct rb_node *next;
2844 if (!p->root.rb_node)
2846 write_lock(&p->lock);
2847 next = rb_first(&p->root);
2849 n = rb_entry(next, struct sp_node, nd);
2850 next = rb_next(&n->nd);
2853 write_unlock(&p->lock);
2856 #ifdef CONFIG_NUMA_BALANCING
2857 static int __initdata numabalancing_override;
2859 static void __init check_numabalancing_enable(void)
2861 bool numabalancing_default = false;
2863 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2864 numabalancing_default = true;
2866 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2867 if (numabalancing_override)
2868 set_numabalancing_state(numabalancing_override == 1);
2870 if (num_online_nodes() > 1 && !numabalancing_override) {
2871 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2872 numabalancing_default ? "Enabling" : "Disabling");
2873 set_numabalancing_state(numabalancing_default);
2877 static int __init setup_numabalancing(char *str)
2883 if (!strcmp(str, "enable")) {
2884 numabalancing_override = 1;
2886 } else if (!strcmp(str, "disable")) {
2887 numabalancing_override = -1;
2892 pr_warn("Unable to parse numa_balancing=\n");
2896 __setup("numa_balancing=", setup_numabalancing);
2898 static inline void __init check_numabalancing_enable(void)
2901 #endif /* CONFIG_NUMA_BALANCING */
2903 /* assumes fs == KERNEL_DS */
2904 void __init numa_policy_init(void)
2906 nodemask_t interleave_nodes;
2907 unsigned long largest = 0;
2908 int nid, prefer = 0;
2910 policy_cache = kmem_cache_create("numa_policy",
2911 sizeof(struct mempolicy),
2912 0, SLAB_PANIC, NULL);
2914 sn_cache = kmem_cache_create("shared_policy_node",
2915 sizeof(struct sp_node),
2916 0, SLAB_PANIC, NULL);
2918 for_each_node(nid) {
2919 preferred_node_policy[nid] = (struct mempolicy) {
2920 .refcnt = ATOMIC_INIT(1),
2921 .mode = MPOL_PREFERRED,
2922 .flags = MPOL_F_MOF | MPOL_F_MORON,
2923 .nodes = nodemask_of_node(nid),
2928 * Set interleaving policy for system init. Interleaving is only
2929 * enabled across suitably sized nodes (default is >= 16MB), or
2930 * fall back to the largest node if they're all smaller.
2932 nodes_clear(interleave_nodes);
2933 for_each_node_state(nid, N_MEMORY) {
2934 unsigned long total_pages = node_present_pages(nid);
2936 /* Preserve the largest node */
2937 if (largest < total_pages) {
2938 largest = total_pages;
2942 /* Interleave this node? */
2943 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2944 node_set(nid, interleave_nodes);
2947 /* All too small, use the largest */
2948 if (unlikely(nodes_empty(interleave_nodes)))
2949 node_set(prefer, interleave_nodes);
2951 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2952 pr_err("%s: interleaving failed\n", __func__);
2954 check_numabalancing_enable();
2957 /* Reset policy of current process to default */
2958 void numa_default_policy(void)
2960 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2964 * Parse and format mempolicy from/to strings
2967 static const char * const policy_modes[] =
2969 [MPOL_DEFAULT] = "default",
2970 [MPOL_PREFERRED] = "prefer",
2971 [MPOL_BIND] = "bind",
2972 [MPOL_INTERLEAVE] = "interleave",
2973 [MPOL_LOCAL] = "local",
2974 [MPOL_PREFERRED_MANY] = "prefer (many)",
2980 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2981 * @str: string containing mempolicy to parse
2982 * @mpol: pointer to struct mempolicy pointer, returned on success.
2985 * <mode>[=<flags>][:<nodelist>]
2987 * Return: %0 on success, else %1
2989 int mpol_parse_str(char *str, struct mempolicy **mpol)
2991 struct mempolicy *new = NULL;
2992 unsigned short mode_flags;
2994 char *nodelist = strchr(str, ':');
2995 char *flags = strchr(str, '=');
2999 *flags++ = '\0'; /* terminate mode string */
3002 /* NUL-terminate mode or flags string */
3004 if (nodelist_parse(nodelist, nodes))
3006 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3011 mode = match_string(policy_modes, MPOL_MAX, str);
3016 case MPOL_PREFERRED:
3018 * Insist on a nodelist of one node only, although later
3019 * we use first_node(nodes) to grab a single node, so here
3020 * nodelist (or nodes) cannot be empty.
3023 char *rest = nodelist;
3024 while (isdigit(*rest))
3028 if (nodes_empty(nodes))
3032 case MPOL_INTERLEAVE:
3034 * Default to online nodes with memory if no nodelist
3037 nodes = node_states[N_MEMORY];
3041 * Don't allow a nodelist; mpol_new() checks flags
3048 * Insist on a empty nodelist
3053 case MPOL_PREFERRED_MANY:
3056 * Insist on a nodelist
3065 * Currently, we only support two mutually exclusive
3068 if (!strcmp(flags, "static"))
3069 mode_flags |= MPOL_F_STATIC_NODES;
3070 else if (!strcmp(flags, "relative"))
3071 mode_flags |= MPOL_F_RELATIVE_NODES;
3076 new = mpol_new(mode, mode_flags, &nodes);
3081 * Save nodes for mpol_to_str() to show the tmpfs mount options
3082 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3084 if (mode != MPOL_PREFERRED) {
3086 } else if (nodelist) {
3087 nodes_clear(new->nodes);
3088 node_set(first_node(nodes), new->nodes);
3090 new->mode = MPOL_LOCAL;
3094 * Save nodes for contextualization: this will be used to "clone"
3095 * the mempolicy in a specific context [cpuset] at a later time.
3097 new->w.user_nodemask = nodes;
3102 /* Restore string for error message */
3111 #endif /* CONFIG_TMPFS */
3114 * mpol_to_str - format a mempolicy structure for printing
3115 * @buffer: to contain formatted mempolicy string
3116 * @maxlen: length of @buffer
3117 * @pol: pointer to mempolicy to be formatted
3119 * Convert @pol into a string. If @buffer is too short, truncate the string.
3120 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3121 * longest flag, "relative", and to display at least a few node ids.
3123 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3126 nodemask_t nodes = NODE_MASK_NONE;
3127 unsigned short mode = MPOL_DEFAULT;
3128 unsigned short flags = 0;
3130 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3139 case MPOL_PREFERRED:
3140 case MPOL_PREFERRED_MANY:
3142 case MPOL_INTERLEAVE:
3147 snprintf(p, maxlen, "unknown");
3151 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3153 if (flags & MPOL_MODE_FLAGS) {
3154 p += snprintf(p, buffer + maxlen - p, "=");
3157 * Currently, the only defined flags are mutually exclusive
3159 if (flags & MPOL_F_STATIC_NODES)
3160 p += snprintf(p, buffer + maxlen - p, "static");
3161 else if (flags & MPOL_F_RELATIVE_NODES)
3162 p += snprintf(p, buffer + maxlen - p, "relative");
3165 if (!nodes_empty(nodes))
3166 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3167 nodemask_pr_args(&nodes));