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 if (isolate_hugetlb(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)
633 struct mmu_gather tlb;
636 tlb_gather_mmu(&tlb, vma->vm_mm);
638 nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
640 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
642 tlb_finish_mmu(&tlb);
647 static unsigned long change_prot_numa(struct vm_area_struct *vma,
648 unsigned long addr, unsigned long end)
652 #endif /* CONFIG_NUMA_BALANCING */
654 static int queue_pages_test_walk(unsigned long start, unsigned long end,
655 struct mm_walk *walk)
657 struct vm_area_struct *next, *vma = walk->vma;
658 struct queue_pages *qp = walk->private;
659 unsigned long endvma = vma->vm_end;
660 unsigned long flags = qp->flags;
662 /* range check first */
663 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
667 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
668 (qp->start < vma->vm_start))
669 /* hole at head side of range */
672 next = find_vma(vma->vm_mm, vma->vm_end);
673 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
674 ((vma->vm_end < qp->end) &&
675 (!next || vma->vm_end < next->vm_start)))
676 /* hole at middle or tail of range */
680 * Need check MPOL_MF_STRICT to return -EIO if possible
681 * regardless of vma_migratable
683 if (!vma_migratable(vma) &&
684 !(flags & MPOL_MF_STRICT))
690 if (flags & MPOL_MF_LAZY) {
691 /* Similar to task_numa_work, skip inaccessible VMAs */
692 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
693 !(vma->vm_flags & VM_MIXEDMAP))
694 change_prot_numa(vma, start, endvma);
698 /* queue pages from current vma */
699 if (flags & MPOL_MF_VALID)
704 static const struct mm_walk_ops queue_pages_walk_ops = {
705 .hugetlb_entry = queue_pages_hugetlb,
706 .pmd_entry = queue_pages_pte_range,
707 .test_walk = queue_pages_test_walk,
711 * Walk through page tables and collect pages to be migrated.
713 * If pages found in a given range are on a set of nodes (determined by
714 * @nodes and @flags,) it's isolated and queued to the pagelist which is
715 * passed via @private.
717 * queue_pages_range() has three possible return values:
718 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
720 * 0 - queue pages successfully or no misplaced page.
721 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
722 * memory range specified by nodemask and maxnode points outside
723 * your accessible address space (-EFAULT)
726 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
727 nodemask_t *nodes, unsigned long flags,
728 struct list_head *pagelist)
731 struct queue_pages qp = {
732 .pagelist = pagelist,
740 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
743 /* whole range in hole */
750 * Apply policy to a single VMA
751 * This must be called with the mmap_lock held for writing.
753 static int vma_replace_policy(struct vm_area_struct *vma,
754 struct mempolicy *pol)
757 struct mempolicy *old;
758 struct mempolicy *new;
760 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
761 vma->vm_start, vma->vm_end, vma->vm_pgoff,
762 vma->vm_ops, vma->vm_file,
763 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
769 if (vma->vm_ops && vma->vm_ops->set_policy) {
770 err = vma->vm_ops->set_policy(vma, new);
775 old = vma->vm_policy;
776 vma->vm_policy = new; /* protected by mmap_lock */
785 /* Step 2: apply policy to a range and do splits. */
786 static int mbind_range(struct mm_struct *mm, unsigned long start,
787 unsigned long end, struct mempolicy *new_pol)
789 MA_STATE(mas, &mm->mm_mt, start, start);
790 struct vm_area_struct *prev;
791 struct vm_area_struct *vma;
795 prev = mas_prev(&mas, 0);
797 mas_set(&mas, start);
799 vma = mas_find(&mas, end - 1);
803 if (start > vma->vm_start)
806 for (; vma; vma = mas_next(&mas, end - 1)) {
807 unsigned long vmstart = max(start, vma->vm_start);
808 unsigned long vmend = min(end, vma->vm_end);
810 if (mpol_equal(vma_policy(vma), new_pol))
813 pgoff = vma->vm_pgoff +
814 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
815 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
816 vma->anon_vma, vma->vm_file, pgoff,
817 new_pol, vma->vm_userfaultfd_ctx,
820 /* vma_merge() invalidated the mas */
825 if (vma->vm_start != vmstart) {
826 err = split_vma(vma->vm_mm, vma, vmstart, 1);
829 /* split_vma() invalidated the mas */
832 if (vma->vm_end != vmend) {
833 err = split_vma(vma->vm_mm, vma, vmend, 0);
836 /* split_vma() invalidated the mas */
840 err = vma_replace_policy(vma, new_pol);
851 /* Set the process memory policy */
852 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
855 struct mempolicy *new, *old;
856 NODEMASK_SCRATCH(scratch);
862 new = mpol_new(mode, flags, nodes);
869 ret = mpol_set_nodemask(new, nodes, scratch);
871 task_unlock(current);
876 old = current->mempolicy;
877 current->mempolicy = new;
878 if (new && new->mode == MPOL_INTERLEAVE)
879 current->il_prev = MAX_NUMNODES-1;
880 task_unlock(current);
884 NODEMASK_SCRATCH_FREE(scratch);
889 * Return nodemask for policy for get_mempolicy() query
891 * Called with task's alloc_lock held
893 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
896 if (p == &default_policy)
901 case MPOL_INTERLEAVE:
903 case MPOL_PREFERRED_MANY:
907 /* return empty node mask for local allocation */
914 static int lookup_node(struct mm_struct *mm, unsigned long addr)
916 struct page *p = NULL;
919 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
921 ret = page_to_nid(p);
927 /* Retrieve NUMA policy */
928 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
929 unsigned long addr, unsigned long flags)
932 struct mm_struct *mm = current->mm;
933 struct vm_area_struct *vma = NULL;
934 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
937 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
940 if (flags & MPOL_F_MEMS_ALLOWED) {
941 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
943 *policy = 0; /* just so it's initialized */
945 *nmask = cpuset_current_mems_allowed;
946 task_unlock(current);
950 if (flags & MPOL_F_ADDR) {
952 * Do NOT fall back to task policy if the
953 * vma/shared policy at addr is NULL. We
954 * want to return MPOL_DEFAULT in this case.
957 vma = vma_lookup(mm, addr);
959 mmap_read_unlock(mm);
962 if (vma->vm_ops && vma->vm_ops->get_policy)
963 pol = vma->vm_ops->get_policy(vma, addr);
965 pol = vma->vm_policy;
970 pol = &default_policy; /* indicates default behavior */
972 if (flags & MPOL_F_NODE) {
973 if (flags & MPOL_F_ADDR) {
975 * Take a refcount on the mpol, because we are about to
976 * drop the mmap_lock, after which only "pol" remains
977 * valid, "vma" is stale.
982 mmap_read_unlock(mm);
983 err = lookup_node(mm, addr);
987 } else if (pol == current->mempolicy &&
988 pol->mode == MPOL_INTERLEAVE) {
989 *policy = next_node_in(current->il_prev, pol->nodes);
995 *policy = pol == &default_policy ? MPOL_DEFAULT :
998 * Internal mempolicy flags must be masked off before exposing
999 * the policy to userspace.
1001 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1006 if (mpol_store_user_nodemask(pol)) {
1007 *nmask = pol->w.user_nodemask;
1010 get_policy_nodemask(pol, nmask);
1011 task_unlock(current);
1018 mmap_read_unlock(mm);
1020 mpol_put(pol_refcount);
1024 #ifdef CONFIG_MIGRATION
1026 * page migration, thp tail pages can be passed.
1028 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1029 unsigned long flags)
1031 struct page *head = compound_head(page);
1033 * Avoid migrating a page that is shared with others.
1035 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1036 if (!isolate_lru_page(head)) {
1037 list_add_tail(&head->lru, pagelist);
1038 mod_node_page_state(page_pgdat(head),
1039 NR_ISOLATED_ANON + page_is_file_lru(head),
1040 thp_nr_pages(head));
1041 } else if (flags & MPOL_MF_STRICT) {
1043 * Non-movable page may reach here. And, there may be
1044 * temporary off LRU pages or non-LRU movable pages.
1045 * Treat them as unmovable pages since they can't be
1046 * isolated, so they can't be moved at the moment. It
1047 * should return -EIO for this case too.
1057 * Migrate pages from one node to a target node.
1058 * Returns error or the number of pages not migrated.
1060 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1064 struct vm_area_struct *vma;
1065 LIST_HEAD(pagelist);
1067 struct migration_target_control mtc = {
1069 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1073 node_set(source, nmask);
1076 * This does not "check" the range but isolates all pages that
1077 * need migration. Between passing in the full user address
1078 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1080 vma = find_vma(mm, 0);
1081 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1082 queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1083 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1085 if (!list_empty(&pagelist)) {
1086 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1087 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1089 putback_movable_pages(&pagelist);
1096 * Move pages between the two nodesets so as to preserve the physical
1097 * layout as much as possible.
1099 * Returns the number of page that could not be moved.
1101 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1102 const nodemask_t *to, int flags)
1108 lru_cache_disable();
1113 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1114 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1115 * bit in 'tmp', and return that <source, dest> pair for migration.
1116 * The pair of nodemasks 'to' and 'from' define the map.
1118 * If no pair of bits is found that way, fallback to picking some
1119 * pair of 'source' and 'dest' bits that are not the same. If the
1120 * 'source' and 'dest' bits are the same, this represents a node
1121 * that will be migrating to itself, so no pages need move.
1123 * If no bits are left in 'tmp', or if all remaining bits left
1124 * in 'tmp' correspond to the same bit in 'to', return false
1125 * (nothing left to migrate).
1127 * This lets us pick a pair of nodes to migrate between, such that
1128 * if possible the dest node is not already occupied by some other
1129 * source node, minimizing the risk of overloading the memory on a
1130 * node that would happen if we migrated incoming memory to a node
1131 * before migrating outgoing memory source that same node.
1133 * A single scan of tmp is sufficient. As we go, we remember the
1134 * most recent <s, d> pair that moved (s != d). If we find a pair
1135 * that not only moved, but what's better, moved to an empty slot
1136 * (d is not set in tmp), then we break out then, with that pair.
1137 * Otherwise when we finish scanning from_tmp, we at least have the
1138 * most recent <s, d> pair that moved. If we get all the way through
1139 * the scan of tmp without finding any node that moved, much less
1140 * moved to an empty node, then there is nothing left worth migrating.
1144 while (!nodes_empty(tmp)) {
1146 int source = NUMA_NO_NODE;
1149 for_each_node_mask(s, tmp) {
1152 * do_migrate_pages() tries to maintain the relative
1153 * node relationship of the pages established between
1154 * threads and memory areas.
1156 * However if the number of source nodes is not equal to
1157 * the number of destination nodes we can not preserve
1158 * this node relative relationship. In that case, skip
1159 * copying memory from a node that is in the destination
1162 * Example: [2,3,4] -> [3,4,5] moves everything.
1163 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1166 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1167 (node_isset(s, *to)))
1170 d = node_remap(s, *from, *to);
1174 source = s; /* Node moved. Memorize */
1177 /* dest not in remaining from nodes? */
1178 if (!node_isset(dest, tmp))
1181 if (source == NUMA_NO_NODE)
1184 node_clear(source, tmp);
1185 err = migrate_to_node(mm, source, dest, flags);
1191 mmap_read_unlock(mm);
1201 * Allocate a new page for page migration based on vma policy.
1202 * Start by assuming the page is mapped by the same vma as contains @start.
1203 * Search forward from there, if not. N.B., this assumes that the
1204 * list of pages handed to migrate_pages()--which is how we get here--
1205 * is in virtual address order.
1207 static struct page *new_page(struct page *page, unsigned long start)
1209 struct folio *dst, *src = page_folio(page);
1210 struct vm_area_struct *vma;
1211 unsigned long address;
1212 VMA_ITERATOR(vmi, current->mm, start);
1213 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1215 for_each_vma(vmi, vma) {
1216 address = page_address_in_vma(page, vma);
1217 if (address != -EFAULT)
1221 if (folio_test_hugetlb(src))
1222 return alloc_huge_page_vma(page_hstate(&src->page),
1225 if (folio_test_large(src))
1226 gfp = GFP_TRANSHUGE;
1229 * if !vma, vma_alloc_folio() will use task or system default policy
1231 dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1232 folio_test_large(src));
1237 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1238 unsigned long flags)
1243 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1244 const nodemask_t *to, int flags)
1249 static struct page *new_page(struct page *page, unsigned long start)
1255 static long do_mbind(unsigned long start, unsigned long len,
1256 unsigned short mode, unsigned short mode_flags,
1257 nodemask_t *nmask, unsigned long flags)
1259 struct mm_struct *mm = current->mm;
1260 struct mempolicy *new;
1264 LIST_HEAD(pagelist);
1266 if (flags & ~(unsigned long)MPOL_MF_VALID)
1268 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1271 if (start & ~PAGE_MASK)
1274 if (mode == MPOL_DEFAULT)
1275 flags &= ~MPOL_MF_STRICT;
1277 len = PAGE_ALIGN(len);
1285 new = mpol_new(mode, mode_flags, nmask);
1287 return PTR_ERR(new);
1289 if (flags & MPOL_MF_LAZY)
1290 new->flags |= MPOL_F_MOF;
1293 * If we are using the default policy then operation
1294 * on discontinuous address spaces is okay after all
1297 flags |= MPOL_MF_DISCONTIG_OK;
1299 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1300 start, start + len, mode, mode_flags,
1301 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1303 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1305 lru_cache_disable();
1308 NODEMASK_SCRATCH(scratch);
1310 mmap_write_lock(mm);
1311 err = mpol_set_nodemask(new, nmask, scratch);
1313 mmap_write_unlock(mm);
1316 NODEMASK_SCRATCH_FREE(scratch);
1321 ret = queue_pages_range(mm, start, end, nmask,
1322 flags | MPOL_MF_INVERT, &pagelist);
1329 err = mbind_range(mm, start, end, new);
1334 if (!list_empty(&pagelist)) {
1335 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1336 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1337 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1339 putback_movable_pages(&pagelist);
1342 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1346 if (!list_empty(&pagelist))
1347 putback_movable_pages(&pagelist);
1350 mmap_write_unlock(mm);
1353 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1359 * User space interface with variable sized bitmaps for nodelists.
1361 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1362 unsigned long maxnode)
1364 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1367 if (in_compat_syscall())
1368 ret = compat_get_bitmap(mask,
1369 (const compat_ulong_t __user *)nmask,
1372 ret = copy_from_user(mask, nmask,
1373 nlongs * sizeof(unsigned long));
1378 if (maxnode % BITS_PER_LONG)
1379 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1384 /* Copy a node mask from user space. */
1385 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1386 unsigned long maxnode)
1389 nodes_clear(*nodes);
1390 if (maxnode == 0 || !nmask)
1392 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1396 * When the user specified more nodes than supported just check
1397 * if the non supported part is all zero, one word at a time,
1398 * starting at the end.
1400 while (maxnode > MAX_NUMNODES) {
1401 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1404 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1407 if (maxnode - bits >= MAX_NUMNODES) {
1410 maxnode = MAX_NUMNODES;
1411 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1417 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1420 /* Copy a kernel node mask to user space */
1421 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1424 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1425 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1426 bool compat = in_compat_syscall();
1429 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1431 if (copy > nbytes) {
1432 if (copy > PAGE_SIZE)
1434 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1437 maxnode = nr_node_ids;
1441 return compat_put_bitmap((compat_ulong_t __user *)mask,
1442 nodes_addr(*nodes), maxnode);
1444 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1447 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1448 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1450 *flags = *mode & MPOL_MODE_FLAGS;
1451 *mode &= ~MPOL_MODE_FLAGS;
1453 if ((unsigned int)(*mode) >= MPOL_MAX)
1455 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1457 if (*flags & MPOL_F_NUMA_BALANCING) {
1458 if (*mode != MPOL_BIND)
1460 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1465 static long kernel_mbind(unsigned long start, unsigned long len,
1466 unsigned long mode, const unsigned long __user *nmask,
1467 unsigned long maxnode, unsigned int flags)
1469 unsigned short mode_flags;
1474 start = untagged_addr(start);
1475 err = sanitize_mpol_flags(&lmode, &mode_flags);
1479 err = get_nodes(&nodes, nmask, maxnode);
1483 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1486 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1487 unsigned long, home_node, unsigned long, flags)
1489 struct mm_struct *mm = current->mm;
1490 struct vm_area_struct *vma;
1491 struct mempolicy *new, *old;
1492 unsigned long vmstart;
1493 unsigned long vmend;
1496 VMA_ITERATOR(vmi, mm, start);
1498 start = untagged_addr(start);
1499 if (start & ~PAGE_MASK)
1502 * flags is used for future extension if any.
1508 * Check home_node is online to avoid accessing uninitialized
1511 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1514 len = PAGE_ALIGN(len);
1521 mmap_write_lock(mm);
1522 for_each_vma_range(vmi, vma, end) {
1524 * If any vma in the range got policy other than MPOL_BIND
1525 * or MPOL_PREFERRED_MANY we return error. We don't reset
1526 * the home node for vmas we already updated before.
1528 old = vma_policy(vma);
1531 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1535 new = mpol_dup(old);
1541 new->home_node = home_node;
1542 vmstart = max(start, vma->vm_start);
1543 vmend = min(end, vma->vm_end);
1544 err = mbind_range(mm, vmstart, vmend, new);
1549 mmap_write_unlock(mm);
1553 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1554 unsigned long, mode, const unsigned long __user *, nmask,
1555 unsigned long, maxnode, unsigned int, flags)
1557 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1560 /* Set the process memory policy */
1561 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1562 unsigned long maxnode)
1564 unsigned short mode_flags;
1569 err = sanitize_mpol_flags(&lmode, &mode_flags);
1573 err = get_nodes(&nodes, nmask, maxnode);
1577 return do_set_mempolicy(lmode, mode_flags, &nodes);
1580 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1581 unsigned long, maxnode)
1583 return kernel_set_mempolicy(mode, nmask, maxnode);
1586 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1587 const unsigned long __user *old_nodes,
1588 const unsigned long __user *new_nodes)
1590 struct mm_struct *mm = NULL;
1591 struct task_struct *task;
1592 nodemask_t task_nodes;
1596 NODEMASK_SCRATCH(scratch);
1601 old = &scratch->mask1;
1602 new = &scratch->mask2;
1604 err = get_nodes(old, old_nodes, maxnode);
1608 err = get_nodes(new, new_nodes, maxnode);
1612 /* Find the mm_struct */
1614 task = pid ? find_task_by_vpid(pid) : current;
1620 get_task_struct(task);
1625 * Check if this process has the right to modify the specified process.
1626 * Use the regular "ptrace_may_access()" checks.
1628 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1635 task_nodes = cpuset_mems_allowed(task);
1636 /* Is the user allowed to access the target nodes? */
1637 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1642 task_nodes = cpuset_mems_allowed(current);
1643 nodes_and(*new, *new, task_nodes);
1644 if (nodes_empty(*new))
1647 err = security_task_movememory(task);
1651 mm = get_task_mm(task);
1652 put_task_struct(task);
1659 err = do_migrate_pages(mm, old, new,
1660 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1664 NODEMASK_SCRATCH_FREE(scratch);
1669 put_task_struct(task);
1674 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1675 const unsigned long __user *, old_nodes,
1676 const unsigned long __user *, new_nodes)
1678 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1682 /* Retrieve NUMA policy */
1683 static int kernel_get_mempolicy(int __user *policy,
1684 unsigned long __user *nmask,
1685 unsigned long maxnode,
1687 unsigned long flags)
1693 if (nmask != NULL && maxnode < nr_node_ids)
1696 addr = untagged_addr(addr);
1698 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1703 if (policy && put_user(pval, policy))
1707 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1712 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1713 unsigned long __user *, nmask, unsigned long, maxnode,
1714 unsigned long, addr, unsigned long, flags)
1716 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1719 bool vma_migratable(struct vm_area_struct *vma)
1721 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1725 * DAX device mappings require predictable access latency, so avoid
1726 * incurring periodic faults.
1728 if (vma_is_dax(vma))
1731 if (is_vm_hugetlb_page(vma) &&
1732 !hugepage_migration_supported(hstate_vma(vma)))
1736 * Migration allocates pages in the highest zone. If we cannot
1737 * do so then migration (at least from node to node) is not
1741 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1747 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1750 struct mempolicy *pol = NULL;
1753 if (vma->vm_ops && vma->vm_ops->get_policy) {
1754 pol = vma->vm_ops->get_policy(vma, addr);
1755 } else if (vma->vm_policy) {
1756 pol = vma->vm_policy;
1759 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1760 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1761 * count on these policies which will be dropped by
1762 * mpol_cond_put() later
1764 if (mpol_needs_cond_ref(pol))
1773 * get_vma_policy(@vma, @addr)
1774 * @vma: virtual memory area whose policy is sought
1775 * @addr: address in @vma for shared policy lookup
1777 * Returns effective policy for a VMA at specified address.
1778 * Falls back to current->mempolicy or system default policy, as necessary.
1779 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1780 * count--added by the get_policy() vm_op, as appropriate--to protect against
1781 * freeing by another task. It is the caller's responsibility to free the
1782 * extra reference for shared policies.
1784 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1787 struct mempolicy *pol = __get_vma_policy(vma, addr);
1790 pol = get_task_policy(current);
1795 bool vma_policy_mof(struct vm_area_struct *vma)
1797 struct mempolicy *pol;
1799 if (vma->vm_ops && vma->vm_ops->get_policy) {
1802 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1803 if (pol && (pol->flags & MPOL_F_MOF))
1810 pol = vma->vm_policy;
1812 pol = get_task_policy(current);
1814 return pol->flags & MPOL_F_MOF;
1817 bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1819 enum zone_type dynamic_policy_zone = policy_zone;
1821 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1824 * if policy->nodes has movable memory only,
1825 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1827 * policy->nodes is intersect with node_states[N_MEMORY].
1828 * so if the following test fails, it implies
1829 * policy->nodes has movable memory only.
1831 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1832 dynamic_policy_zone = ZONE_MOVABLE;
1834 return zone >= dynamic_policy_zone;
1838 * Return a nodemask representing a mempolicy for filtering nodes for
1841 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1843 int mode = policy->mode;
1845 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1846 if (unlikely(mode == MPOL_BIND) &&
1847 apply_policy_zone(policy, gfp_zone(gfp)) &&
1848 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1849 return &policy->nodes;
1851 if (mode == MPOL_PREFERRED_MANY)
1852 return &policy->nodes;
1858 * Return the preferred node id for 'prefer' mempolicy, and return
1859 * the given id for all other policies.
1861 * policy_node() is always coupled with policy_nodemask(), which
1862 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1864 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1866 if (policy->mode == MPOL_PREFERRED) {
1867 nd = first_node(policy->nodes);
1870 * __GFP_THISNODE shouldn't even be used with the bind policy
1871 * because we might easily break the expectation to stay on the
1872 * requested node and not break the policy.
1874 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1877 if ((policy->mode == MPOL_BIND ||
1878 policy->mode == MPOL_PREFERRED_MANY) &&
1879 policy->home_node != NUMA_NO_NODE)
1880 return policy->home_node;
1885 /* Do dynamic interleaving for a process */
1886 static unsigned interleave_nodes(struct mempolicy *policy)
1889 struct task_struct *me = current;
1891 next = next_node_in(me->il_prev, policy->nodes);
1892 if (next < MAX_NUMNODES)
1898 * Depending on the memory policy provide a node from which to allocate the
1901 unsigned int mempolicy_slab_node(void)
1903 struct mempolicy *policy;
1904 int node = numa_mem_id();
1909 policy = current->mempolicy;
1913 switch (policy->mode) {
1914 case MPOL_PREFERRED:
1915 return first_node(policy->nodes);
1917 case MPOL_INTERLEAVE:
1918 return interleave_nodes(policy);
1921 case MPOL_PREFERRED_MANY:
1926 * Follow bind policy behavior and start allocation at the
1929 struct zonelist *zonelist;
1930 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1931 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1932 z = first_zones_zonelist(zonelist, highest_zoneidx,
1934 return z->zone ? zone_to_nid(z->zone) : node;
1945 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1946 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1947 * number of present nodes.
1949 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1951 nodemask_t nodemask = pol->nodes;
1952 unsigned int target, nnodes;
1956 * The barrier will stabilize the nodemask in a register or on
1957 * the stack so that it will stop changing under the code.
1959 * Between first_node() and next_node(), pol->nodes could be changed
1960 * by other threads. So we put pol->nodes in a local stack.
1964 nnodes = nodes_weight(nodemask);
1966 return numa_node_id();
1967 target = (unsigned int)n % nnodes;
1968 nid = first_node(nodemask);
1969 for (i = 0; i < target; i++)
1970 nid = next_node(nid, nodemask);
1974 /* Determine a node number for interleave */
1975 static inline unsigned interleave_nid(struct mempolicy *pol,
1976 struct vm_area_struct *vma, unsigned long addr, int shift)
1982 * for small pages, there is no difference between
1983 * shift and PAGE_SHIFT, so the bit-shift is safe.
1984 * for huge pages, since vm_pgoff is in units of small
1985 * pages, we need to shift off the always 0 bits to get
1988 BUG_ON(shift < PAGE_SHIFT);
1989 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1990 off += (addr - vma->vm_start) >> shift;
1991 return offset_il_node(pol, off);
1993 return interleave_nodes(pol);
1996 #ifdef CONFIG_HUGETLBFS
1998 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1999 * @vma: virtual memory area whose policy is sought
2000 * @addr: address in @vma for shared policy lookup and interleave policy
2001 * @gfp_flags: for requested zone
2002 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2003 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2005 * Returns a nid suitable for a huge page allocation and a pointer
2006 * to the struct mempolicy for conditional unref after allocation.
2007 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2008 * to the mempolicy's @nodemask for filtering the zonelist.
2010 * Must be protected by read_mems_allowed_begin()
2012 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2013 struct mempolicy **mpol, nodemask_t **nodemask)
2018 *mpol = get_vma_policy(vma, addr);
2020 mode = (*mpol)->mode;
2022 if (unlikely(mode == MPOL_INTERLEAVE)) {
2023 nid = interleave_nid(*mpol, vma, addr,
2024 huge_page_shift(hstate_vma(vma)));
2026 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2027 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2028 *nodemask = &(*mpol)->nodes;
2034 * init_nodemask_of_mempolicy
2036 * If the current task's mempolicy is "default" [NULL], return 'false'
2037 * to indicate default policy. Otherwise, extract the policy nodemask
2038 * for 'bind' or 'interleave' policy into the argument nodemask, or
2039 * initialize the argument nodemask to contain the single node for
2040 * 'preferred' or 'local' policy and return 'true' to indicate presence
2041 * of non-default mempolicy.
2043 * We don't bother with reference counting the mempolicy [mpol_get/put]
2044 * because the current task is examining it's own mempolicy and a task's
2045 * mempolicy is only ever changed by the task itself.
2047 * N.B., it is the caller's responsibility to free a returned nodemask.
2049 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2051 struct mempolicy *mempolicy;
2053 if (!(mask && current->mempolicy))
2057 mempolicy = current->mempolicy;
2058 switch (mempolicy->mode) {
2059 case MPOL_PREFERRED:
2060 case MPOL_PREFERRED_MANY:
2062 case MPOL_INTERLEAVE:
2063 *mask = mempolicy->nodes;
2067 init_nodemask_of_node(mask, numa_node_id());
2073 task_unlock(current);
2080 * mempolicy_in_oom_domain
2082 * If tsk's mempolicy is "bind", check for intersection between mask and
2083 * the policy nodemask. Otherwise, return true for all other policies
2084 * including "interleave", as a tsk with "interleave" policy may have
2085 * memory allocated from all nodes in system.
2087 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2089 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2090 const nodemask_t *mask)
2092 struct mempolicy *mempolicy;
2099 mempolicy = tsk->mempolicy;
2100 if (mempolicy && mempolicy->mode == MPOL_BIND)
2101 ret = nodes_intersects(mempolicy->nodes, *mask);
2107 /* Allocate a page in interleaved policy.
2108 Own path because it needs to do special accounting. */
2109 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2114 page = __alloc_pages(gfp, order, nid, NULL);
2115 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2116 if (!static_branch_likely(&vm_numa_stat_key))
2118 if (page && page_to_nid(page) == nid) {
2120 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2126 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2127 int nid, struct mempolicy *pol)
2130 gfp_t preferred_gfp;
2133 * This is a two pass approach. The first pass will only try the
2134 * preferred nodes but skip the direct reclaim and allow the
2135 * allocation to fail, while the second pass will try all the
2138 preferred_gfp = gfp | __GFP_NOWARN;
2139 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2140 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2142 page = __alloc_pages(gfp, order, nid, NULL);
2148 * vma_alloc_folio - Allocate a folio for a VMA.
2150 * @order: Order of the folio.
2151 * @vma: Pointer to VMA or NULL if not available.
2152 * @addr: Virtual address of the allocation. Must be inside @vma.
2153 * @hugepage: For hugepages try only the preferred node if possible.
2155 * Allocate a folio for a specific address in @vma, using the appropriate
2156 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2157 * of the mm_struct of the VMA to prevent it from going away. Should be
2158 * used for all allocations for folios that will be mapped into user space.
2160 * Return: The folio on success or NULL if allocation fails.
2162 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2163 unsigned long addr, bool hugepage)
2165 struct mempolicy *pol;
2166 int node = numa_node_id();
2167 struct folio *folio;
2171 pol = get_vma_policy(vma, addr);
2173 if (pol->mode == MPOL_INTERLEAVE) {
2177 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2180 page = alloc_page_interleave(gfp, order, nid);
2181 if (page && order > 1)
2182 prep_transhuge_page(page);
2183 folio = (struct folio *)page;
2187 if (pol->mode == MPOL_PREFERRED_MANY) {
2190 node = policy_node(gfp, pol, node);
2192 page = alloc_pages_preferred_many(gfp, order, node, pol);
2194 if (page && order > 1)
2195 prep_transhuge_page(page);
2196 folio = (struct folio *)page;
2200 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2201 int hpage_node = node;
2204 * For hugepage allocation and non-interleave policy which
2205 * allows the current node (or other explicitly preferred
2206 * node) we only try to allocate from the current/preferred
2207 * node and don't fall back to other nodes, as the cost of
2208 * remote accesses would likely offset THP benefits.
2210 * If the policy is interleave or does not allow the current
2211 * node in its nodemask, we allocate the standard way.
2213 if (pol->mode == MPOL_PREFERRED)
2214 hpage_node = first_node(pol->nodes);
2216 nmask = policy_nodemask(gfp, pol);
2217 if (!nmask || node_isset(hpage_node, *nmask)) {
2220 * First, try to allocate THP only on local node, but
2221 * don't reclaim unnecessarily, just compact.
2223 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2224 __GFP_NORETRY, order, hpage_node);
2227 * If hugepage allocations are configured to always
2228 * synchronous compact or the vma has been madvised
2229 * to prefer hugepage backing, retry allowing remote
2230 * memory with both reclaim and compact as well.
2232 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2233 folio = __folio_alloc(gfp, order, hpage_node,
2240 nmask = policy_nodemask(gfp, pol);
2241 preferred_nid = policy_node(gfp, pol, node);
2242 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2247 EXPORT_SYMBOL(vma_alloc_folio);
2250 * alloc_pages - Allocate pages.
2252 * @order: Power of two of number of pages to allocate.
2254 * Allocate 1 << @order contiguous pages. The physical address of the
2255 * first page is naturally aligned (eg an order-3 allocation will be aligned
2256 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2257 * process is honoured when in process context.
2259 * Context: Can be called from any context, providing the appropriate GFP
2261 * Return: The page on success or NULL if allocation fails.
2263 struct page *alloc_pages(gfp_t gfp, unsigned order)
2265 struct mempolicy *pol = &default_policy;
2268 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2269 pol = get_task_policy(current);
2272 * No reference counting needed for current->mempolicy
2273 * nor system default_policy
2275 if (pol->mode == MPOL_INTERLEAVE)
2276 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2277 else if (pol->mode == MPOL_PREFERRED_MANY)
2278 page = alloc_pages_preferred_many(gfp, order,
2279 policy_node(gfp, pol, numa_node_id()), pol);
2281 page = __alloc_pages(gfp, order,
2282 policy_node(gfp, pol, numa_node_id()),
2283 policy_nodemask(gfp, pol));
2287 EXPORT_SYMBOL(alloc_pages);
2289 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2291 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2293 if (page && order > 1)
2294 prep_transhuge_page(page);
2295 return (struct folio *)page;
2297 EXPORT_SYMBOL(folio_alloc);
2299 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2300 struct mempolicy *pol, unsigned long nr_pages,
2301 struct page **page_array)
2304 unsigned long nr_pages_per_node;
2307 unsigned long nr_allocated;
2308 unsigned long total_allocated = 0;
2310 nodes = nodes_weight(pol->nodes);
2311 nr_pages_per_node = nr_pages / nodes;
2312 delta = nr_pages - nodes * nr_pages_per_node;
2314 for (i = 0; i < nodes; i++) {
2316 nr_allocated = __alloc_pages_bulk(gfp,
2317 interleave_nodes(pol), NULL,
2318 nr_pages_per_node + 1, NULL,
2322 nr_allocated = __alloc_pages_bulk(gfp,
2323 interleave_nodes(pol), NULL,
2324 nr_pages_per_node, NULL, page_array);
2327 page_array += nr_allocated;
2328 total_allocated += nr_allocated;
2331 return total_allocated;
2334 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2335 struct mempolicy *pol, unsigned long nr_pages,
2336 struct page **page_array)
2338 gfp_t preferred_gfp;
2339 unsigned long nr_allocated = 0;
2341 preferred_gfp = gfp | __GFP_NOWARN;
2342 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2344 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2345 nr_pages, NULL, page_array);
2347 if (nr_allocated < nr_pages)
2348 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2349 nr_pages - nr_allocated, NULL,
2350 page_array + nr_allocated);
2351 return nr_allocated;
2354 /* alloc pages bulk and mempolicy should be considered at the
2355 * same time in some situation such as vmalloc.
2357 * It can accelerate memory allocation especially interleaving
2360 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2361 unsigned long nr_pages, struct page **page_array)
2363 struct mempolicy *pol = &default_policy;
2365 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2366 pol = get_task_policy(current);
2368 if (pol->mode == MPOL_INTERLEAVE)
2369 return alloc_pages_bulk_array_interleave(gfp, pol,
2370 nr_pages, page_array);
2372 if (pol->mode == MPOL_PREFERRED_MANY)
2373 return alloc_pages_bulk_array_preferred_many(gfp,
2374 numa_node_id(), pol, nr_pages, page_array);
2376 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2377 policy_nodemask(gfp, pol), nr_pages, NULL,
2381 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2383 struct mempolicy *pol = mpol_dup(vma_policy(src));
2386 return PTR_ERR(pol);
2387 dst->vm_policy = pol;
2392 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2393 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2394 * with the mems_allowed returned by cpuset_mems_allowed(). This
2395 * keeps mempolicies cpuset relative after its cpuset moves. See
2396 * further kernel/cpuset.c update_nodemask().
2398 * current's mempolicy may be rebinded by the other task(the task that changes
2399 * cpuset's mems), so we needn't do rebind work for current task.
2402 /* Slow path of a mempolicy duplicate */
2403 struct mempolicy *__mpol_dup(struct mempolicy *old)
2405 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2408 return ERR_PTR(-ENOMEM);
2410 /* task's mempolicy is protected by alloc_lock */
2411 if (old == current->mempolicy) {
2414 task_unlock(current);
2418 if (current_cpuset_is_being_rebound()) {
2419 nodemask_t mems = cpuset_mems_allowed(current);
2420 mpol_rebind_policy(new, &mems);
2422 atomic_set(&new->refcnt, 1);
2426 /* Slow path of a mempolicy comparison */
2427 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2431 if (a->mode != b->mode)
2433 if (a->flags != b->flags)
2435 if (a->home_node != b->home_node)
2437 if (mpol_store_user_nodemask(a))
2438 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2443 case MPOL_INTERLEAVE:
2444 case MPOL_PREFERRED:
2445 case MPOL_PREFERRED_MANY:
2446 return !!nodes_equal(a->nodes, b->nodes);
2456 * Shared memory backing store policy support.
2458 * Remember policies even when nobody has shared memory mapped.
2459 * The policies are kept in Red-Black tree linked from the inode.
2460 * They are protected by the sp->lock rwlock, which should be held
2461 * for any accesses to the tree.
2465 * lookup first element intersecting start-end. Caller holds sp->lock for
2466 * reading or for writing
2468 static struct sp_node *
2469 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2471 struct rb_node *n = sp->root.rb_node;
2474 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2476 if (start >= p->end)
2478 else if (end <= p->start)
2486 struct sp_node *w = NULL;
2487 struct rb_node *prev = rb_prev(n);
2490 w = rb_entry(prev, struct sp_node, nd);
2491 if (w->end <= start)
2495 return rb_entry(n, struct sp_node, nd);
2499 * Insert a new shared policy into the list. Caller holds sp->lock for
2502 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2504 struct rb_node **p = &sp->root.rb_node;
2505 struct rb_node *parent = NULL;
2510 nd = rb_entry(parent, struct sp_node, nd);
2511 if (new->start < nd->start)
2513 else if (new->end > nd->end)
2514 p = &(*p)->rb_right;
2518 rb_link_node(&new->nd, parent, p);
2519 rb_insert_color(&new->nd, &sp->root);
2520 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2521 new->policy ? new->policy->mode : 0);
2524 /* Find shared policy intersecting idx */
2526 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2528 struct mempolicy *pol = NULL;
2531 if (!sp->root.rb_node)
2533 read_lock(&sp->lock);
2534 sn = sp_lookup(sp, idx, idx+1);
2536 mpol_get(sn->policy);
2539 read_unlock(&sp->lock);
2543 static void sp_free(struct sp_node *n)
2545 mpol_put(n->policy);
2546 kmem_cache_free(sn_cache, n);
2550 * mpol_misplaced - check whether current page node is valid in policy
2552 * @page: page to be checked
2553 * @vma: vm area where page mapped
2554 * @addr: virtual address where page mapped
2556 * Lookup current policy node id for vma,addr and "compare to" page's
2557 * node id. Policy determination "mimics" alloc_page_vma().
2558 * Called from fault path where we know the vma and faulting address.
2560 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2561 * policy, or a suitable node ID to allocate a replacement page from.
2563 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2565 struct mempolicy *pol;
2567 int curnid = page_to_nid(page);
2568 unsigned long pgoff;
2569 int thiscpu = raw_smp_processor_id();
2570 int thisnid = cpu_to_node(thiscpu);
2571 int polnid = NUMA_NO_NODE;
2572 int ret = NUMA_NO_NODE;
2574 pol = get_vma_policy(vma, addr);
2575 if (!(pol->flags & MPOL_F_MOF))
2578 switch (pol->mode) {
2579 case MPOL_INTERLEAVE:
2580 pgoff = vma->vm_pgoff;
2581 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2582 polnid = offset_il_node(pol, pgoff);
2585 case MPOL_PREFERRED:
2586 if (node_isset(curnid, pol->nodes))
2588 polnid = first_node(pol->nodes);
2592 polnid = numa_node_id();
2596 /* Optimize placement among multiple nodes via NUMA balancing */
2597 if (pol->flags & MPOL_F_MORON) {
2598 if (node_isset(thisnid, pol->nodes))
2604 case MPOL_PREFERRED_MANY:
2606 * use current page if in policy nodemask,
2607 * else select nearest allowed node, if any.
2608 * If no allowed nodes, use current [!misplaced].
2610 if (node_isset(curnid, pol->nodes))
2612 z = first_zones_zonelist(
2613 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2614 gfp_zone(GFP_HIGHUSER),
2616 polnid = zone_to_nid(z->zone);
2623 /* Migrate the page towards the node whose CPU is referencing it */
2624 if (pol->flags & MPOL_F_MORON) {
2627 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2631 if (curnid != polnid)
2640 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2641 * dropped after task->mempolicy is set to NULL so that any allocation done as
2642 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2645 void mpol_put_task_policy(struct task_struct *task)
2647 struct mempolicy *pol;
2650 pol = task->mempolicy;
2651 task->mempolicy = NULL;
2656 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2658 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2659 rb_erase(&n->nd, &sp->root);
2663 static void sp_node_init(struct sp_node *node, unsigned long start,
2664 unsigned long end, struct mempolicy *pol)
2666 node->start = start;
2671 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2672 struct mempolicy *pol)
2675 struct mempolicy *newpol;
2677 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2681 newpol = mpol_dup(pol);
2682 if (IS_ERR(newpol)) {
2683 kmem_cache_free(sn_cache, n);
2686 newpol->flags |= MPOL_F_SHARED;
2687 sp_node_init(n, start, end, newpol);
2692 /* Replace a policy range. */
2693 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2694 unsigned long end, struct sp_node *new)
2697 struct sp_node *n_new = NULL;
2698 struct mempolicy *mpol_new = NULL;
2702 write_lock(&sp->lock);
2703 n = sp_lookup(sp, start, end);
2704 /* Take care of old policies in the same range. */
2705 while (n && n->start < end) {
2706 struct rb_node *next = rb_next(&n->nd);
2707 if (n->start >= start) {
2713 /* Old policy spanning whole new range. */
2718 *mpol_new = *n->policy;
2719 atomic_set(&mpol_new->refcnt, 1);
2720 sp_node_init(n_new, end, n->end, mpol_new);
2722 sp_insert(sp, n_new);
2731 n = rb_entry(next, struct sp_node, nd);
2735 write_unlock(&sp->lock);
2742 kmem_cache_free(sn_cache, n_new);
2747 write_unlock(&sp->lock);
2749 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2752 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2755 atomic_set(&mpol_new->refcnt, 1);
2760 * mpol_shared_policy_init - initialize shared policy for inode
2761 * @sp: pointer to inode shared policy
2762 * @mpol: struct mempolicy to install
2764 * Install non-NULL @mpol in inode's shared policy rb-tree.
2765 * On entry, the current task has a reference on a non-NULL @mpol.
2766 * This must be released on exit.
2767 * This is called at get_inode() calls and we can use GFP_KERNEL.
2769 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2773 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2774 rwlock_init(&sp->lock);
2777 struct vm_area_struct pvma;
2778 struct mempolicy *new;
2779 NODEMASK_SCRATCH(scratch);
2783 /* contextualize the tmpfs mount point mempolicy */
2784 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2786 goto free_scratch; /* no valid nodemask intersection */
2789 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2790 task_unlock(current);
2794 /* Create pseudo-vma that contains just the policy */
2795 vma_init(&pvma, NULL);
2796 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2797 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2800 mpol_put(new); /* drop initial ref */
2802 NODEMASK_SCRATCH_FREE(scratch);
2804 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2808 int mpol_set_shared_policy(struct shared_policy *info,
2809 struct vm_area_struct *vma, struct mempolicy *npol)
2812 struct sp_node *new = NULL;
2813 unsigned long sz = vma_pages(vma);
2815 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2817 sz, npol ? npol->mode : -1,
2818 npol ? npol->flags : -1,
2819 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2822 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2826 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2832 /* Free a backing policy store on inode delete. */
2833 void mpol_free_shared_policy(struct shared_policy *p)
2836 struct rb_node *next;
2838 if (!p->root.rb_node)
2840 write_lock(&p->lock);
2841 next = rb_first(&p->root);
2843 n = rb_entry(next, struct sp_node, nd);
2844 next = rb_next(&n->nd);
2847 write_unlock(&p->lock);
2850 #ifdef CONFIG_NUMA_BALANCING
2851 static int __initdata numabalancing_override;
2853 static void __init check_numabalancing_enable(void)
2855 bool numabalancing_default = false;
2857 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2858 numabalancing_default = true;
2860 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2861 if (numabalancing_override)
2862 set_numabalancing_state(numabalancing_override == 1);
2864 if (num_online_nodes() > 1 && !numabalancing_override) {
2865 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2866 numabalancing_default ? "Enabling" : "Disabling");
2867 set_numabalancing_state(numabalancing_default);
2871 static int __init setup_numabalancing(char *str)
2877 if (!strcmp(str, "enable")) {
2878 numabalancing_override = 1;
2880 } else if (!strcmp(str, "disable")) {
2881 numabalancing_override = -1;
2886 pr_warn("Unable to parse numa_balancing=\n");
2890 __setup("numa_balancing=", setup_numabalancing);
2892 static inline void __init check_numabalancing_enable(void)
2895 #endif /* CONFIG_NUMA_BALANCING */
2897 /* assumes fs == KERNEL_DS */
2898 void __init numa_policy_init(void)
2900 nodemask_t interleave_nodes;
2901 unsigned long largest = 0;
2902 int nid, prefer = 0;
2904 policy_cache = kmem_cache_create("numa_policy",
2905 sizeof(struct mempolicy),
2906 0, SLAB_PANIC, NULL);
2908 sn_cache = kmem_cache_create("shared_policy_node",
2909 sizeof(struct sp_node),
2910 0, SLAB_PANIC, NULL);
2912 for_each_node(nid) {
2913 preferred_node_policy[nid] = (struct mempolicy) {
2914 .refcnt = ATOMIC_INIT(1),
2915 .mode = MPOL_PREFERRED,
2916 .flags = MPOL_F_MOF | MPOL_F_MORON,
2917 .nodes = nodemask_of_node(nid),
2922 * Set interleaving policy for system init. Interleaving is only
2923 * enabled across suitably sized nodes (default is >= 16MB), or
2924 * fall back to the largest node if they're all smaller.
2926 nodes_clear(interleave_nodes);
2927 for_each_node_state(nid, N_MEMORY) {
2928 unsigned long total_pages = node_present_pages(nid);
2930 /* Preserve the largest node */
2931 if (largest < total_pages) {
2932 largest = total_pages;
2936 /* Interleave this node? */
2937 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2938 node_set(nid, interleave_nodes);
2941 /* All too small, use the largest */
2942 if (unlikely(nodes_empty(interleave_nodes)))
2943 node_set(prefer, interleave_nodes);
2945 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2946 pr_err("%s: interleaving failed\n", __func__);
2948 check_numabalancing_enable();
2951 /* Reset policy of current process to default */
2952 void numa_default_policy(void)
2954 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2958 * Parse and format mempolicy from/to strings
2961 static const char * const policy_modes[] =
2963 [MPOL_DEFAULT] = "default",
2964 [MPOL_PREFERRED] = "prefer",
2965 [MPOL_BIND] = "bind",
2966 [MPOL_INTERLEAVE] = "interleave",
2967 [MPOL_LOCAL] = "local",
2968 [MPOL_PREFERRED_MANY] = "prefer (many)",
2974 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2975 * @str: string containing mempolicy to parse
2976 * @mpol: pointer to struct mempolicy pointer, returned on success.
2979 * <mode>[=<flags>][:<nodelist>]
2981 * Return: %0 on success, else %1
2983 int mpol_parse_str(char *str, struct mempolicy **mpol)
2985 struct mempolicy *new = NULL;
2986 unsigned short mode_flags;
2988 char *nodelist = strchr(str, ':');
2989 char *flags = strchr(str, '=');
2993 *flags++ = '\0'; /* terminate mode string */
2996 /* NUL-terminate mode or flags string */
2998 if (nodelist_parse(nodelist, nodes))
3000 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3005 mode = match_string(policy_modes, MPOL_MAX, str);
3010 case MPOL_PREFERRED:
3012 * Insist on a nodelist of one node only, although later
3013 * we use first_node(nodes) to grab a single node, so here
3014 * nodelist (or nodes) cannot be empty.
3017 char *rest = nodelist;
3018 while (isdigit(*rest))
3022 if (nodes_empty(nodes))
3026 case MPOL_INTERLEAVE:
3028 * Default to online nodes with memory if no nodelist
3031 nodes = node_states[N_MEMORY];
3035 * Don't allow a nodelist; mpol_new() checks flags
3042 * Insist on a empty nodelist
3047 case MPOL_PREFERRED_MANY:
3050 * Insist on a nodelist
3059 * Currently, we only support two mutually exclusive
3062 if (!strcmp(flags, "static"))
3063 mode_flags |= MPOL_F_STATIC_NODES;
3064 else if (!strcmp(flags, "relative"))
3065 mode_flags |= MPOL_F_RELATIVE_NODES;
3070 new = mpol_new(mode, mode_flags, &nodes);
3075 * Save nodes for mpol_to_str() to show the tmpfs mount options
3076 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3078 if (mode != MPOL_PREFERRED) {
3080 } else if (nodelist) {
3081 nodes_clear(new->nodes);
3082 node_set(first_node(nodes), new->nodes);
3084 new->mode = MPOL_LOCAL;
3088 * Save nodes for contextualization: this will be used to "clone"
3089 * the mempolicy in a specific context [cpuset] at a later time.
3091 new->w.user_nodemask = nodes;
3096 /* Restore string for error message */
3105 #endif /* CONFIG_TMPFS */
3108 * mpol_to_str - format a mempolicy structure for printing
3109 * @buffer: to contain formatted mempolicy string
3110 * @maxlen: length of @buffer
3111 * @pol: pointer to mempolicy to be formatted
3113 * Convert @pol into a string. If @buffer is too short, truncate the string.
3114 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3115 * longest flag, "relative", and to display at least a few node ids.
3117 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3120 nodemask_t nodes = NODE_MASK_NONE;
3121 unsigned short mode = MPOL_DEFAULT;
3122 unsigned short flags = 0;
3124 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3133 case MPOL_PREFERRED:
3134 case MPOL_PREFERRED_MANY:
3136 case MPOL_INTERLEAVE:
3141 snprintf(p, maxlen, "unknown");
3145 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3147 if (flags & MPOL_MODE_FLAGS) {
3148 p += snprintf(p, buffer + maxlen - p, "=");
3151 * Currently, the only defined flags are mutually exclusive
3153 if (flags & MPOL_F_STATIC_NODES)
3154 p += snprintf(p, buffer + maxlen - p, "static");
3155 else if (flags & MPOL_F_RELATIVE_NODES)
3156 p += snprintf(p, buffer + maxlen - p, "relative");
3159 if (!nodes_empty(nodes))
3160 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3161 nodemask_pr_args(&nodes));