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_folio(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, MM_CP_PROT_NUMA);
641 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
643 tlb_finish_mmu(&tlb);
648 static unsigned long change_prot_numa(struct vm_area_struct *vma,
649 unsigned long addr, unsigned long end)
653 #endif /* CONFIG_NUMA_BALANCING */
655 static int queue_pages_test_walk(unsigned long start, unsigned long end,
656 struct mm_walk *walk)
658 struct vm_area_struct *next, *vma = walk->vma;
659 struct queue_pages *qp = walk->private;
660 unsigned long endvma = vma->vm_end;
661 unsigned long flags = qp->flags;
663 /* range check first */
664 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
668 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
669 (qp->start < vma->vm_start))
670 /* hole at head side of range */
673 next = find_vma(vma->vm_mm, vma->vm_end);
674 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
675 ((vma->vm_end < qp->end) &&
676 (!next || vma->vm_end < next->vm_start)))
677 /* hole at middle or tail of range */
681 * Need check MPOL_MF_STRICT to return -EIO if possible
682 * regardless of vma_migratable
684 if (!vma_migratable(vma) &&
685 !(flags & MPOL_MF_STRICT))
691 if (flags & MPOL_MF_LAZY) {
692 /* Similar to task_numa_work, skip inaccessible VMAs */
693 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
694 !(vma->vm_flags & VM_MIXEDMAP))
695 change_prot_numa(vma, start, endvma);
699 /* queue pages from current vma */
700 if (flags & MPOL_MF_VALID)
705 static const struct mm_walk_ops queue_pages_walk_ops = {
706 .hugetlb_entry = queue_pages_hugetlb,
707 .pmd_entry = queue_pages_pte_range,
708 .test_walk = queue_pages_test_walk,
712 * Walk through page tables and collect pages to be migrated.
714 * If pages found in a given range are on a set of nodes (determined by
715 * @nodes and @flags,) it's isolated and queued to the pagelist which is
716 * passed via @private.
718 * queue_pages_range() has three possible return values:
719 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
721 * 0 - queue pages successfully or no misplaced page.
722 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
723 * memory range specified by nodemask and maxnode points outside
724 * your accessible address space (-EFAULT)
727 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
728 nodemask_t *nodes, unsigned long flags,
729 struct list_head *pagelist)
732 struct queue_pages qp = {
733 .pagelist = pagelist,
741 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
744 /* whole range in hole */
751 * Apply policy to a single VMA
752 * This must be called with the mmap_lock held for writing.
754 static int vma_replace_policy(struct vm_area_struct *vma,
755 struct mempolicy *pol)
758 struct mempolicy *old;
759 struct mempolicy *new;
761 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
762 vma->vm_start, vma->vm_end, vma->vm_pgoff,
763 vma->vm_ops, vma->vm_file,
764 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
770 if (vma->vm_ops && vma->vm_ops->set_policy) {
771 err = vma->vm_ops->set_policy(vma, new);
776 old = vma->vm_policy;
777 vma->vm_policy = new; /* protected by mmap_lock */
786 /* Step 2: apply policy to a range and do splits. */
787 static int mbind_range(struct mm_struct *mm, unsigned long start,
788 unsigned long end, struct mempolicy *new_pol)
790 VMA_ITERATOR(vmi, mm, start);
791 struct vm_area_struct *prev;
792 struct vm_area_struct *vma;
796 prev = vma_prev(&vmi);
797 vma = vma_find(&vmi, end);
801 if (start > vma->vm_start)
805 unsigned long vmstart = max(start, vma->vm_start);
806 unsigned long vmend = min(end, vma->vm_end);
808 if (mpol_equal(vma_policy(vma), new_pol))
811 pgoff = vma->vm_pgoff +
812 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
813 prev = vma_merge(&vmi, mm, prev, vmstart, vmend, vma->vm_flags,
814 vma->anon_vma, vma->vm_file, pgoff,
815 new_pol, vma->vm_userfaultfd_ctx,
821 if (vma->vm_start != vmstart) {
822 err = split_vma(&vmi, vma, vmstart, 1);
826 if (vma->vm_end != vmend) {
827 err = split_vma(&vmi, vma, vmend, 0);
832 err = vma_replace_policy(vma, new_pol);
837 } for_each_vma_range(vmi, vma, end);
843 /* Set the process memory policy */
844 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
847 struct mempolicy *new, *old;
848 NODEMASK_SCRATCH(scratch);
854 new = mpol_new(mode, flags, nodes);
861 ret = mpol_set_nodemask(new, nodes, scratch);
863 task_unlock(current);
868 old = current->mempolicy;
869 current->mempolicy = new;
870 if (new && new->mode == MPOL_INTERLEAVE)
871 current->il_prev = MAX_NUMNODES-1;
872 task_unlock(current);
876 NODEMASK_SCRATCH_FREE(scratch);
881 * Return nodemask for policy for get_mempolicy() query
883 * Called with task's alloc_lock held
885 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
888 if (p == &default_policy)
893 case MPOL_INTERLEAVE:
895 case MPOL_PREFERRED_MANY:
899 /* return empty node mask for local allocation */
906 static int lookup_node(struct mm_struct *mm, unsigned long addr)
908 struct page *p = NULL;
911 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
913 ret = page_to_nid(p);
919 /* Retrieve NUMA policy */
920 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
921 unsigned long addr, unsigned long flags)
924 struct mm_struct *mm = current->mm;
925 struct vm_area_struct *vma = NULL;
926 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
929 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
932 if (flags & MPOL_F_MEMS_ALLOWED) {
933 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
935 *policy = 0; /* just so it's initialized */
937 *nmask = cpuset_current_mems_allowed;
938 task_unlock(current);
942 if (flags & MPOL_F_ADDR) {
944 * Do NOT fall back to task policy if the
945 * vma/shared policy at addr is NULL. We
946 * want to return MPOL_DEFAULT in this case.
949 vma = vma_lookup(mm, addr);
951 mmap_read_unlock(mm);
954 if (vma->vm_ops && vma->vm_ops->get_policy)
955 pol = vma->vm_ops->get_policy(vma, addr);
957 pol = vma->vm_policy;
962 pol = &default_policy; /* indicates default behavior */
964 if (flags & MPOL_F_NODE) {
965 if (flags & MPOL_F_ADDR) {
967 * Take a refcount on the mpol, because we are about to
968 * drop the mmap_lock, after which only "pol" remains
969 * valid, "vma" is stale.
974 mmap_read_unlock(mm);
975 err = lookup_node(mm, addr);
979 } else if (pol == current->mempolicy &&
980 pol->mode == MPOL_INTERLEAVE) {
981 *policy = next_node_in(current->il_prev, pol->nodes);
987 *policy = pol == &default_policy ? MPOL_DEFAULT :
990 * Internal mempolicy flags must be masked off before exposing
991 * the policy to userspace.
993 *policy |= (pol->flags & MPOL_MODE_FLAGS);
998 if (mpol_store_user_nodemask(pol)) {
999 *nmask = pol->w.user_nodemask;
1002 get_policy_nodemask(pol, nmask);
1003 task_unlock(current);
1010 mmap_read_unlock(mm);
1012 mpol_put(pol_refcount);
1016 #ifdef CONFIG_MIGRATION
1018 * page migration, thp tail pages can be passed.
1020 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1021 unsigned long flags)
1023 struct page *head = compound_head(page);
1025 * Avoid migrating a page that is shared with others.
1027 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1028 if (!isolate_lru_page(head)) {
1029 list_add_tail(&head->lru, pagelist);
1030 mod_node_page_state(page_pgdat(head),
1031 NR_ISOLATED_ANON + page_is_file_lru(head),
1032 thp_nr_pages(head));
1033 } else if (flags & MPOL_MF_STRICT) {
1035 * Non-movable page may reach here. And, there may be
1036 * temporary off LRU pages or non-LRU movable pages.
1037 * Treat them as unmovable pages since they can't be
1038 * isolated, so they can't be moved at the moment. It
1039 * should return -EIO for this case too.
1049 * Migrate pages from one node to a target node.
1050 * Returns error or the number of pages not migrated.
1052 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1056 struct vm_area_struct *vma;
1057 LIST_HEAD(pagelist);
1059 struct migration_target_control mtc = {
1061 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1065 node_set(source, nmask);
1068 * This does not "check" the range but isolates all pages that
1069 * need migration. Between passing in the full user address
1070 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1072 vma = find_vma(mm, 0);
1073 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1074 queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1075 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1077 if (!list_empty(&pagelist)) {
1078 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1079 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1081 putback_movable_pages(&pagelist);
1088 * Move pages between the two nodesets so as to preserve the physical
1089 * layout as much as possible.
1091 * Returns the number of page that could not be moved.
1093 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1094 const nodemask_t *to, int flags)
1100 lru_cache_disable();
1105 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1106 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1107 * bit in 'tmp', and return that <source, dest> pair for migration.
1108 * The pair of nodemasks 'to' and 'from' define the map.
1110 * If no pair of bits is found that way, fallback to picking some
1111 * pair of 'source' and 'dest' bits that are not the same. If the
1112 * 'source' and 'dest' bits are the same, this represents a node
1113 * that will be migrating to itself, so no pages need move.
1115 * If no bits are left in 'tmp', or if all remaining bits left
1116 * in 'tmp' correspond to the same bit in 'to', return false
1117 * (nothing left to migrate).
1119 * This lets us pick a pair of nodes to migrate between, such that
1120 * if possible the dest node is not already occupied by some other
1121 * source node, minimizing the risk of overloading the memory on a
1122 * node that would happen if we migrated incoming memory to a node
1123 * before migrating outgoing memory source that same node.
1125 * A single scan of tmp is sufficient. As we go, we remember the
1126 * most recent <s, d> pair that moved (s != d). If we find a pair
1127 * that not only moved, but what's better, moved to an empty slot
1128 * (d is not set in tmp), then we break out then, with that pair.
1129 * Otherwise when we finish scanning from_tmp, we at least have the
1130 * most recent <s, d> pair that moved. If we get all the way through
1131 * the scan of tmp without finding any node that moved, much less
1132 * moved to an empty node, then there is nothing left worth migrating.
1136 while (!nodes_empty(tmp)) {
1138 int source = NUMA_NO_NODE;
1141 for_each_node_mask(s, tmp) {
1144 * do_migrate_pages() tries to maintain the relative
1145 * node relationship of the pages established between
1146 * threads and memory areas.
1148 * However if the number of source nodes is not equal to
1149 * the number of destination nodes we can not preserve
1150 * this node relative relationship. In that case, skip
1151 * copying memory from a node that is in the destination
1154 * Example: [2,3,4] -> [3,4,5] moves everything.
1155 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1158 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1159 (node_isset(s, *to)))
1162 d = node_remap(s, *from, *to);
1166 source = s; /* Node moved. Memorize */
1169 /* dest not in remaining from nodes? */
1170 if (!node_isset(dest, tmp))
1173 if (source == NUMA_NO_NODE)
1176 node_clear(source, tmp);
1177 err = migrate_to_node(mm, source, dest, flags);
1183 mmap_read_unlock(mm);
1193 * Allocate a new page for page migration based on vma policy.
1194 * Start by assuming the page is mapped by the same vma as contains @start.
1195 * Search forward from there, if not. N.B., this assumes that the
1196 * list of pages handed to migrate_pages()--which is how we get here--
1197 * is in virtual address order.
1199 static struct page *new_page(struct page *page, unsigned long start)
1201 struct folio *dst, *src = page_folio(page);
1202 struct vm_area_struct *vma;
1203 unsigned long address;
1204 VMA_ITERATOR(vmi, current->mm, start);
1205 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1207 for_each_vma(vmi, vma) {
1208 address = page_address_in_vma(page, vma);
1209 if (address != -EFAULT)
1213 if (folio_test_hugetlb(src))
1214 return alloc_huge_page_vma(page_hstate(&src->page),
1217 if (folio_test_large(src))
1218 gfp = GFP_TRANSHUGE;
1221 * if !vma, vma_alloc_folio() will use task or system default policy
1223 dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1224 folio_test_large(src));
1229 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1230 unsigned long flags)
1235 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1236 const nodemask_t *to, int flags)
1241 static struct page *new_page(struct page *page, unsigned long start)
1247 static long do_mbind(unsigned long start, unsigned long len,
1248 unsigned short mode, unsigned short mode_flags,
1249 nodemask_t *nmask, unsigned long flags)
1251 struct mm_struct *mm = current->mm;
1252 struct mempolicy *new;
1256 LIST_HEAD(pagelist);
1258 if (flags & ~(unsigned long)MPOL_MF_VALID)
1260 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1263 if (start & ~PAGE_MASK)
1266 if (mode == MPOL_DEFAULT)
1267 flags &= ~MPOL_MF_STRICT;
1269 len = PAGE_ALIGN(len);
1277 new = mpol_new(mode, mode_flags, nmask);
1279 return PTR_ERR(new);
1281 if (flags & MPOL_MF_LAZY)
1282 new->flags |= MPOL_F_MOF;
1285 * If we are using the default policy then operation
1286 * on discontinuous address spaces is okay after all
1289 flags |= MPOL_MF_DISCONTIG_OK;
1291 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1292 start, start + len, mode, mode_flags,
1293 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1295 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1297 lru_cache_disable();
1300 NODEMASK_SCRATCH(scratch);
1302 mmap_write_lock(mm);
1303 err = mpol_set_nodemask(new, nmask, scratch);
1305 mmap_write_unlock(mm);
1308 NODEMASK_SCRATCH_FREE(scratch);
1313 ret = queue_pages_range(mm, start, end, nmask,
1314 flags | MPOL_MF_INVERT, &pagelist);
1321 err = mbind_range(mm, start, end, new);
1326 if (!list_empty(&pagelist)) {
1327 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1328 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1329 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1331 putback_movable_pages(&pagelist);
1334 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1338 if (!list_empty(&pagelist))
1339 putback_movable_pages(&pagelist);
1342 mmap_write_unlock(mm);
1345 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1351 * User space interface with variable sized bitmaps for nodelists.
1353 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1354 unsigned long maxnode)
1356 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1359 if (in_compat_syscall())
1360 ret = compat_get_bitmap(mask,
1361 (const compat_ulong_t __user *)nmask,
1364 ret = copy_from_user(mask, nmask,
1365 nlongs * sizeof(unsigned long));
1370 if (maxnode % BITS_PER_LONG)
1371 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1376 /* Copy a node mask from user space. */
1377 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1378 unsigned long maxnode)
1381 nodes_clear(*nodes);
1382 if (maxnode == 0 || !nmask)
1384 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1388 * When the user specified more nodes than supported just check
1389 * if the non supported part is all zero, one word at a time,
1390 * starting at the end.
1392 while (maxnode > MAX_NUMNODES) {
1393 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1396 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1399 if (maxnode - bits >= MAX_NUMNODES) {
1402 maxnode = MAX_NUMNODES;
1403 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1409 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1412 /* Copy a kernel node mask to user space */
1413 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1416 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1417 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1418 bool compat = in_compat_syscall();
1421 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1423 if (copy > nbytes) {
1424 if (copy > PAGE_SIZE)
1426 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1429 maxnode = nr_node_ids;
1433 return compat_put_bitmap((compat_ulong_t __user *)mask,
1434 nodes_addr(*nodes), maxnode);
1436 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1439 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1440 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1442 *flags = *mode & MPOL_MODE_FLAGS;
1443 *mode &= ~MPOL_MODE_FLAGS;
1445 if ((unsigned int)(*mode) >= MPOL_MAX)
1447 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1449 if (*flags & MPOL_F_NUMA_BALANCING) {
1450 if (*mode != MPOL_BIND)
1452 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1457 static long kernel_mbind(unsigned long start, unsigned long len,
1458 unsigned long mode, const unsigned long __user *nmask,
1459 unsigned long maxnode, unsigned int flags)
1461 unsigned short mode_flags;
1466 start = untagged_addr(start);
1467 err = sanitize_mpol_flags(&lmode, &mode_flags);
1471 err = get_nodes(&nodes, nmask, maxnode);
1475 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1478 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1479 unsigned long, home_node, unsigned long, flags)
1481 struct mm_struct *mm = current->mm;
1482 struct vm_area_struct *vma;
1483 struct mempolicy *new, *old;
1484 unsigned long vmstart;
1485 unsigned long vmend;
1488 VMA_ITERATOR(vmi, mm, start);
1490 start = untagged_addr(start);
1491 if (start & ~PAGE_MASK)
1494 * flags is used for future extension if any.
1500 * Check home_node is online to avoid accessing uninitialized
1503 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1506 len = PAGE_ALIGN(len);
1513 mmap_write_lock(mm);
1514 for_each_vma_range(vmi, vma, end) {
1516 * If any vma in the range got policy other than MPOL_BIND
1517 * or MPOL_PREFERRED_MANY we return error. We don't reset
1518 * the home node for vmas we already updated before.
1520 old = vma_policy(vma);
1523 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1527 new = mpol_dup(old);
1533 new->home_node = home_node;
1534 vmstart = max(start, vma->vm_start);
1535 vmend = min(end, vma->vm_end);
1536 err = mbind_range(mm, vmstart, vmend, new);
1541 mmap_write_unlock(mm);
1545 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1546 unsigned long, mode, const unsigned long __user *, nmask,
1547 unsigned long, maxnode, unsigned int, flags)
1549 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1552 /* Set the process memory policy */
1553 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1554 unsigned long maxnode)
1556 unsigned short mode_flags;
1561 err = sanitize_mpol_flags(&lmode, &mode_flags);
1565 err = get_nodes(&nodes, nmask, maxnode);
1569 return do_set_mempolicy(lmode, mode_flags, &nodes);
1572 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1573 unsigned long, maxnode)
1575 return kernel_set_mempolicy(mode, nmask, maxnode);
1578 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1579 const unsigned long __user *old_nodes,
1580 const unsigned long __user *new_nodes)
1582 struct mm_struct *mm = NULL;
1583 struct task_struct *task;
1584 nodemask_t task_nodes;
1588 NODEMASK_SCRATCH(scratch);
1593 old = &scratch->mask1;
1594 new = &scratch->mask2;
1596 err = get_nodes(old, old_nodes, maxnode);
1600 err = get_nodes(new, new_nodes, maxnode);
1604 /* Find the mm_struct */
1606 task = pid ? find_task_by_vpid(pid) : current;
1612 get_task_struct(task);
1617 * Check if this process has the right to modify the specified process.
1618 * Use the regular "ptrace_may_access()" checks.
1620 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1627 task_nodes = cpuset_mems_allowed(task);
1628 /* Is the user allowed to access the target nodes? */
1629 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1634 task_nodes = cpuset_mems_allowed(current);
1635 nodes_and(*new, *new, task_nodes);
1636 if (nodes_empty(*new))
1639 err = security_task_movememory(task);
1643 mm = get_task_mm(task);
1644 put_task_struct(task);
1651 err = do_migrate_pages(mm, old, new,
1652 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1656 NODEMASK_SCRATCH_FREE(scratch);
1661 put_task_struct(task);
1666 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1667 const unsigned long __user *, old_nodes,
1668 const unsigned long __user *, new_nodes)
1670 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1674 /* Retrieve NUMA policy */
1675 static int kernel_get_mempolicy(int __user *policy,
1676 unsigned long __user *nmask,
1677 unsigned long maxnode,
1679 unsigned long flags)
1685 if (nmask != NULL && maxnode < nr_node_ids)
1688 addr = untagged_addr(addr);
1690 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1695 if (policy && put_user(pval, policy))
1699 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1704 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1705 unsigned long __user *, nmask, unsigned long, maxnode,
1706 unsigned long, addr, unsigned long, flags)
1708 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1711 bool vma_migratable(struct vm_area_struct *vma)
1713 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1717 * DAX device mappings require predictable access latency, so avoid
1718 * incurring periodic faults.
1720 if (vma_is_dax(vma))
1723 if (is_vm_hugetlb_page(vma) &&
1724 !hugepage_migration_supported(hstate_vma(vma)))
1728 * Migration allocates pages in the highest zone. If we cannot
1729 * do so then migration (at least from node to node) is not
1733 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1739 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1742 struct mempolicy *pol = NULL;
1745 if (vma->vm_ops && vma->vm_ops->get_policy) {
1746 pol = vma->vm_ops->get_policy(vma, addr);
1747 } else if (vma->vm_policy) {
1748 pol = vma->vm_policy;
1751 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1752 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1753 * count on these policies which will be dropped by
1754 * mpol_cond_put() later
1756 if (mpol_needs_cond_ref(pol))
1765 * get_vma_policy(@vma, @addr)
1766 * @vma: virtual memory area whose policy is sought
1767 * @addr: address in @vma for shared policy lookup
1769 * Returns effective policy for a VMA at specified address.
1770 * Falls back to current->mempolicy or system default policy, as necessary.
1771 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1772 * count--added by the get_policy() vm_op, as appropriate--to protect against
1773 * freeing by another task. It is the caller's responsibility to free the
1774 * extra reference for shared policies.
1776 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1779 struct mempolicy *pol = __get_vma_policy(vma, addr);
1782 pol = get_task_policy(current);
1787 bool vma_policy_mof(struct vm_area_struct *vma)
1789 struct mempolicy *pol;
1791 if (vma->vm_ops && vma->vm_ops->get_policy) {
1794 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1795 if (pol && (pol->flags & MPOL_F_MOF))
1802 pol = vma->vm_policy;
1804 pol = get_task_policy(current);
1806 return pol->flags & MPOL_F_MOF;
1809 bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1811 enum zone_type dynamic_policy_zone = policy_zone;
1813 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1816 * if policy->nodes has movable memory only,
1817 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1819 * policy->nodes is intersect with node_states[N_MEMORY].
1820 * so if the following test fails, it implies
1821 * policy->nodes has movable memory only.
1823 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1824 dynamic_policy_zone = ZONE_MOVABLE;
1826 return zone >= dynamic_policy_zone;
1830 * Return a nodemask representing a mempolicy for filtering nodes for
1833 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1835 int mode = policy->mode;
1837 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1838 if (unlikely(mode == MPOL_BIND) &&
1839 apply_policy_zone(policy, gfp_zone(gfp)) &&
1840 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1841 return &policy->nodes;
1843 if (mode == MPOL_PREFERRED_MANY)
1844 return &policy->nodes;
1850 * Return the preferred node id for 'prefer' mempolicy, and return
1851 * the given id for all other policies.
1853 * policy_node() is always coupled with policy_nodemask(), which
1854 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1856 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1858 if (policy->mode == MPOL_PREFERRED) {
1859 nd = first_node(policy->nodes);
1862 * __GFP_THISNODE shouldn't even be used with the bind policy
1863 * because we might easily break the expectation to stay on the
1864 * requested node and not break the policy.
1866 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1869 if ((policy->mode == MPOL_BIND ||
1870 policy->mode == MPOL_PREFERRED_MANY) &&
1871 policy->home_node != NUMA_NO_NODE)
1872 return policy->home_node;
1877 /* Do dynamic interleaving for a process */
1878 static unsigned interleave_nodes(struct mempolicy *policy)
1881 struct task_struct *me = current;
1883 next = next_node_in(me->il_prev, policy->nodes);
1884 if (next < MAX_NUMNODES)
1890 * Depending on the memory policy provide a node from which to allocate the
1893 unsigned int mempolicy_slab_node(void)
1895 struct mempolicy *policy;
1896 int node = numa_mem_id();
1901 policy = current->mempolicy;
1905 switch (policy->mode) {
1906 case MPOL_PREFERRED:
1907 return first_node(policy->nodes);
1909 case MPOL_INTERLEAVE:
1910 return interleave_nodes(policy);
1913 case MPOL_PREFERRED_MANY:
1918 * Follow bind policy behavior and start allocation at the
1921 struct zonelist *zonelist;
1922 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1923 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1924 z = first_zones_zonelist(zonelist, highest_zoneidx,
1926 return z->zone ? zone_to_nid(z->zone) : node;
1937 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1938 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1939 * number of present nodes.
1941 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1943 nodemask_t nodemask = pol->nodes;
1944 unsigned int target, nnodes;
1948 * The barrier will stabilize the nodemask in a register or on
1949 * the stack so that it will stop changing under the code.
1951 * Between first_node() and next_node(), pol->nodes could be changed
1952 * by other threads. So we put pol->nodes in a local stack.
1956 nnodes = nodes_weight(nodemask);
1958 return numa_node_id();
1959 target = (unsigned int)n % nnodes;
1960 nid = first_node(nodemask);
1961 for (i = 0; i < target; i++)
1962 nid = next_node(nid, nodemask);
1966 /* Determine a node number for interleave */
1967 static inline unsigned interleave_nid(struct mempolicy *pol,
1968 struct vm_area_struct *vma, unsigned long addr, int shift)
1974 * for small pages, there is no difference between
1975 * shift and PAGE_SHIFT, so the bit-shift is safe.
1976 * for huge pages, since vm_pgoff is in units of small
1977 * pages, we need to shift off the always 0 bits to get
1980 BUG_ON(shift < PAGE_SHIFT);
1981 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1982 off += (addr - vma->vm_start) >> shift;
1983 return offset_il_node(pol, off);
1985 return interleave_nodes(pol);
1988 #ifdef CONFIG_HUGETLBFS
1990 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1991 * @vma: virtual memory area whose policy is sought
1992 * @addr: address in @vma for shared policy lookup and interleave policy
1993 * @gfp_flags: for requested zone
1994 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1995 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1997 * Returns a nid suitable for a huge page allocation and a pointer
1998 * to the struct mempolicy for conditional unref after allocation.
1999 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2000 * to the mempolicy's @nodemask for filtering the zonelist.
2002 * Must be protected by read_mems_allowed_begin()
2004 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2005 struct mempolicy **mpol, nodemask_t **nodemask)
2010 *mpol = get_vma_policy(vma, addr);
2012 mode = (*mpol)->mode;
2014 if (unlikely(mode == MPOL_INTERLEAVE)) {
2015 nid = interleave_nid(*mpol, vma, addr,
2016 huge_page_shift(hstate_vma(vma)));
2018 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2019 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2020 *nodemask = &(*mpol)->nodes;
2026 * init_nodemask_of_mempolicy
2028 * If the current task's mempolicy is "default" [NULL], return 'false'
2029 * to indicate default policy. Otherwise, extract the policy nodemask
2030 * for 'bind' or 'interleave' policy into the argument nodemask, or
2031 * initialize the argument nodemask to contain the single node for
2032 * 'preferred' or 'local' policy and return 'true' to indicate presence
2033 * of non-default mempolicy.
2035 * We don't bother with reference counting the mempolicy [mpol_get/put]
2036 * because the current task is examining it's own mempolicy and a task's
2037 * mempolicy is only ever changed by the task itself.
2039 * N.B., it is the caller's responsibility to free a returned nodemask.
2041 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2043 struct mempolicy *mempolicy;
2045 if (!(mask && current->mempolicy))
2049 mempolicy = current->mempolicy;
2050 switch (mempolicy->mode) {
2051 case MPOL_PREFERRED:
2052 case MPOL_PREFERRED_MANY:
2054 case MPOL_INTERLEAVE:
2055 *mask = mempolicy->nodes;
2059 init_nodemask_of_node(mask, numa_node_id());
2065 task_unlock(current);
2072 * mempolicy_in_oom_domain
2074 * If tsk's mempolicy is "bind", check for intersection between mask and
2075 * the policy nodemask. Otherwise, return true for all other policies
2076 * including "interleave", as a tsk with "interleave" policy may have
2077 * memory allocated from all nodes in system.
2079 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2081 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2082 const nodemask_t *mask)
2084 struct mempolicy *mempolicy;
2091 mempolicy = tsk->mempolicy;
2092 if (mempolicy && mempolicy->mode == MPOL_BIND)
2093 ret = nodes_intersects(mempolicy->nodes, *mask);
2099 /* Allocate a page in interleaved policy.
2100 Own path because it needs to do special accounting. */
2101 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2106 page = __alloc_pages(gfp, order, nid, NULL);
2107 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2108 if (!static_branch_likely(&vm_numa_stat_key))
2110 if (page && page_to_nid(page) == nid) {
2112 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2118 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2119 int nid, struct mempolicy *pol)
2122 gfp_t preferred_gfp;
2125 * This is a two pass approach. The first pass will only try the
2126 * preferred nodes but skip the direct reclaim and allow the
2127 * allocation to fail, while the second pass will try all the
2130 preferred_gfp = gfp | __GFP_NOWARN;
2131 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2132 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2134 page = __alloc_pages(gfp, order, nid, NULL);
2140 * vma_alloc_folio - Allocate a folio for a VMA.
2142 * @order: Order of the folio.
2143 * @vma: Pointer to VMA or NULL if not available.
2144 * @addr: Virtual address of the allocation. Must be inside @vma.
2145 * @hugepage: For hugepages try only the preferred node if possible.
2147 * Allocate a folio for a specific address in @vma, using the appropriate
2148 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2149 * of the mm_struct of the VMA to prevent it from going away. Should be
2150 * used for all allocations for folios that will be mapped into user space.
2152 * Return: The folio on success or NULL if allocation fails.
2154 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2155 unsigned long addr, bool hugepage)
2157 struct mempolicy *pol;
2158 int node = numa_node_id();
2159 struct folio *folio;
2163 pol = get_vma_policy(vma, addr);
2165 if (pol->mode == MPOL_INTERLEAVE) {
2169 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2172 page = alloc_page_interleave(gfp, order, nid);
2173 if (page && order > 1)
2174 prep_transhuge_page(page);
2175 folio = (struct folio *)page;
2179 if (pol->mode == MPOL_PREFERRED_MANY) {
2182 node = policy_node(gfp, pol, node);
2184 page = alloc_pages_preferred_many(gfp, order, node, pol);
2186 if (page && order > 1)
2187 prep_transhuge_page(page);
2188 folio = (struct folio *)page;
2192 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2193 int hpage_node = node;
2196 * For hugepage allocation and non-interleave policy which
2197 * allows the current node (or other explicitly preferred
2198 * node) we only try to allocate from the current/preferred
2199 * node and don't fall back to other nodes, as the cost of
2200 * remote accesses would likely offset THP benefits.
2202 * If the policy is interleave or does not allow the current
2203 * node in its nodemask, we allocate the standard way.
2205 if (pol->mode == MPOL_PREFERRED)
2206 hpage_node = first_node(pol->nodes);
2208 nmask = policy_nodemask(gfp, pol);
2209 if (!nmask || node_isset(hpage_node, *nmask)) {
2212 * First, try to allocate THP only on local node, but
2213 * don't reclaim unnecessarily, just compact.
2215 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2216 __GFP_NORETRY, order, hpage_node);
2219 * If hugepage allocations are configured to always
2220 * synchronous compact or the vma has been madvised
2221 * to prefer hugepage backing, retry allowing remote
2222 * memory with both reclaim and compact as well.
2224 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2225 folio = __folio_alloc(gfp, order, hpage_node,
2232 nmask = policy_nodemask(gfp, pol);
2233 preferred_nid = policy_node(gfp, pol, node);
2234 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2239 EXPORT_SYMBOL(vma_alloc_folio);
2242 * alloc_pages - Allocate pages.
2244 * @order: Power of two of number of pages to allocate.
2246 * Allocate 1 << @order contiguous pages. The physical address of the
2247 * first page is naturally aligned (eg an order-3 allocation will be aligned
2248 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2249 * process is honoured when in process context.
2251 * Context: Can be called from any context, providing the appropriate GFP
2253 * Return: The page on success or NULL if allocation fails.
2255 struct page *alloc_pages(gfp_t gfp, unsigned order)
2257 struct mempolicy *pol = &default_policy;
2260 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2261 pol = get_task_policy(current);
2264 * No reference counting needed for current->mempolicy
2265 * nor system default_policy
2267 if (pol->mode == MPOL_INTERLEAVE)
2268 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2269 else if (pol->mode == MPOL_PREFERRED_MANY)
2270 page = alloc_pages_preferred_many(gfp, order,
2271 policy_node(gfp, pol, numa_node_id()), pol);
2273 page = __alloc_pages(gfp, order,
2274 policy_node(gfp, pol, numa_node_id()),
2275 policy_nodemask(gfp, pol));
2279 EXPORT_SYMBOL(alloc_pages);
2281 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2283 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2285 if (page && order > 1)
2286 prep_transhuge_page(page);
2287 return (struct folio *)page;
2289 EXPORT_SYMBOL(folio_alloc);
2291 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2292 struct mempolicy *pol, unsigned long nr_pages,
2293 struct page **page_array)
2296 unsigned long nr_pages_per_node;
2299 unsigned long nr_allocated;
2300 unsigned long total_allocated = 0;
2302 nodes = nodes_weight(pol->nodes);
2303 nr_pages_per_node = nr_pages / nodes;
2304 delta = nr_pages - nodes * nr_pages_per_node;
2306 for (i = 0; i < nodes; i++) {
2308 nr_allocated = __alloc_pages_bulk(gfp,
2309 interleave_nodes(pol), NULL,
2310 nr_pages_per_node + 1, NULL,
2314 nr_allocated = __alloc_pages_bulk(gfp,
2315 interleave_nodes(pol), NULL,
2316 nr_pages_per_node, NULL, page_array);
2319 page_array += nr_allocated;
2320 total_allocated += nr_allocated;
2323 return total_allocated;
2326 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2327 struct mempolicy *pol, unsigned long nr_pages,
2328 struct page **page_array)
2330 gfp_t preferred_gfp;
2331 unsigned long nr_allocated = 0;
2333 preferred_gfp = gfp | __GFP_NOWARN;
2334 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2336 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2337 nr_pages, NULL, page_array);
2339 if (nr_allocated < nr_pages)
2340 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2341 nr_pages - nr_allocated, NULL,
2342 page_array + nr_allocated);
2343 return nr_allocated;
2346 /* alloc pages bulk and mempolicy should be considered at the
2347 * same time in some situation such as vmalloc.
2349 * It can accelerate memory allocation especially interleaving
2352 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2353 unsigned long nr_pages, struct page **page_array)
2355 struct mempolicy *pol = &default_policy;
2357 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2358 pol = get_task_policy(current);
2360 if (pol->mode == MPOL_INTERLEAVE)
2361 return alloc_pages_bulk_array_interleave(gfp, pol,
2362 nr_pages, page_array);
2364 if (pol->mode == MPOL_PREFERRED_MANY)
2365 return alloc_pages_bulk_array_preferred_many(gfp,
2366 numa_node_id(), pol, nr_pages, page_array);
2368 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2369 policy_nodemask(gfp, pol), nr_pages, NULL,
2373 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2375 struct mempolicy *pol = mpol_dup(vma_policy(src));
2378 return PTR_ERR(pol);
2379 dst->vm_policy = pol;
2384 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2385 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2386 * with the mems_allowed returned by cpuset_mems_allowed(). This
2387 * keeps mempolicies cpuset relative after its cpuset moves. See
2388 * further kernel/cpuset.c update_nodemask().
2390 * current's mempolicy may be rebinded by the other task(the task that changes
2391 * cpuset's mems), so we needn't do rebind work for current task.
2394 /* Slow path of a mempolicy duplicate */
2395 struct mempolicy *__mpol_dup(struct mempolicy *old)
2397 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2400 return ERR_PTR(-ENOMEM);
2402 /* task's mempolicy is protected by alloc_lock */
2403 if (old == current->mempolicy) {
2406 task_unlock(current);
2410 if (current_cpuset_is_being_rebound()) {
2411 nodemask_t mems = cpuset_mems_allowed(current);
2412 mpol_rebind_policy(new, &mems);
2414 atomic_set(&new->refcnt, 1);
2418 /* Slow path of a mempolicy comparison */
2419 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2423 if (a->mode != b->mode)
2425 if (a->flags != b->flags)
2427 if (a->home_node != b->home_node)
2429 if (mpol_store_user_nodemask(a))
2430 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2435 case MPOL_INTERLEAVE:
2436 case MPOL_PREFERRED:
2437 case MPOL_PREFERRED_MANY:
2438 return !!nodes_equal(a->nodes, b->nodes);
2448 * Shared memory backing store policy support.
2450 * Remember policies even when nobody has shared memory mapped.
2451 * The policies are kept in Red-Black tree linked from the inode.
2452 * They are protected by the sp->lock rwlock, which should be held
2453 * for any accesses to the tree.
2457 * lookup first element intersecting start-end. Caller holds sp->lock for
2458 * reading or for writing
2460 static struct sp_node *
2461 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2463 struct rb_node *n = sp->root.rb_node;
2466 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2468 if (start >= p->end)
2470 else if (end <= p->start)
2478 struct sp_node *w = NULL;
2479 struct rb_node *prev = rb_prev(n);
2482 w = rb_entry(prev, struct sp_node, nd);
2483 if (w->end <= start)
2487 return rb_entry(n, struct sp_node, nd);
2491 * Insert a new shared policy into the list. Caller holds sp->lock for
2494 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2496 struct rb_node **p = &sp->root.rb_node;
2497 struct rb_node *parent = NULL;
2502 nd = rb_entry(parent, struct sp_node, nd);
2503 if (new->start < nd->start)
2505 else if (new->end > nd->end)
2506 p = &(*p)->rb_right;
2510 rb_link_node(&new->nd, parent, p);
2511 rb_insert_color(&new->nd, &sp->root);
2512 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2513 new->policy ? new->policy->mode : 0);
2516 /* Find shared policy intersecting idx */
2518 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2520 struct mempolicy *pol = NULL;
2523 if (!sp->root.rb_node)
2525 read_lock(&sp->lock);
2526 sn = sp_lookup(sp, idx, idx+1);
2528 mpol_get(sn->policy);
2531 read_unlock(&sp->lock);
2535 static void sp_free(struct sp_node *n)
2537 mpol_put(n->policy);
2538 kmem_cache_free(sn_cache, n);
2542 * mpol_misplaced - check whether current page node is valid in policy
2544 * @page: page to be checked
2545 * @vma: vm area where page mapped
2546 * @addr: virtual address where page mapped
2548 * Lookup current policy node id for vma,addr and "compare to" page's
2549 * node id. Policy determination "mimics" alloc_page_vma().
2550 * Called from fault path where we know the vma and faulting address.
2552 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2553 * policy, or a suitable node ID to allocate a replacement page from.
2555 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2557 struct mempolicy *pol;
2559 int curnid = page_to_nid(page);
2560 unsigned long pgoff;
2561 int thiscpu = raw_smp_processor_id();
2562 int thisnid = cpu_to_node(thiscpu);
2563 int polnid = NUMA_NO_NODE;
2564 int ret = NUMA_NO_NODE;
2566 pol = get_vma_policy(vma, addr);
2567 if (!(pol->flags & MPOL_F_MOF))
2570 switch (pol->mode) {
2571 case MPOL_INTERLEAVE:
2572 pgoff = vma->vm_pgoff;
2573 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2574 polnid = offset_il_node(pol, pgoff);
2577 case MPOL_PREFERRED:
2578 if (node_isset(curnid, pol->nodes))
2580 polnid = first_node(pol->nodes);
2584 polnid = numa_node_id();
2588 /* Optimize placement among multiple nodes via NUMA balancing */
2589 if (pol->flags & MPOL_F_MORON) {
2590 if (node_isset(thisnid, pol->nodes))
2596 case MPOL_PREFERRED_MANY:
2598 * use current page if in policy nodemask,
2599 * else select nearest allowed node, if any.
2600 * If no allowed nodes, use current [!misplaced].
2602 if (node_isset(curnid, pol->nodes))
2604 z = first_zones_zonelist(
2605 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2606 gfp_zone(GFP_HIGHUSER),
2608 polnid = zone_to_nid(z->zone);
2615 /* Migrate the page towards the node whose CPU is referencing it */
2616 if (pol->flags & MPOL_F_MORON) {
2619 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2623 if (curnid != polnid)
2632 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2633 * dropped after task->mempolicy is set to NULL so that any allocation done as
2634 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2637 void mpol_put_task_policy(struct task_struct *task)
2639 struct mempolicy *pol;
2642 pol = task->mempolicy;
2643 task->mempolicy = NULL;
2648 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2650 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2651 rb_erase(&n->nd, &sp->root);
2655 static void sp_node_init(struct sp_node *node, unsigned long start,
2656 unsigned long end, struct mempolicy *pol)
2658 node->start = start;
2663 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2664 struct mempolicy *pol)
2667 struct mempolicy *newpol;
2669 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2673 newpol = mpol_dup(pol);
2674 if (IS_ERR(newpol)) {
2675 kmem_cache_free(sn_cache, n);
2678 newpol->flags |= MPOL_F_SHARED;
2679 sp_node_init(n, start, end, newpol);
2684 /* Replace a policy range. */
2685 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2686 unsigned long end, struct sp_node *new)
2689 struct sp_node *n_new = NULL;
2690 struct mempolicy *mpol_new = NULL;
2694 write_lock(&sp->lock);
2695 n = sp_lookup(sp, start, end);
2696 /* Take care of old policies in the same range. */
2697 while (n && n->start < end) {
2698 struct rb_node *next = rb_next(&n->nd);
2699 if (n->start >= start) {
2705 /* Old policy spanning whole new range. */
2710 *mpol_new = *n->policy;
2711 atomic_set(&mpol_new->refcnt, 1);
2712 sp_node_init(n_new, end, n->end, mpol_new);
2714 sp_insert(sp, n_new);
2723 n = rb_entry(next, struct sp_node, nd);
2727 write_unlock(&sp->lock);
2734 kmem_cache_free(sn_cache, n_new);
2739 write_unlock(&sp->lock);
2741 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2744 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2747 atomic_set(&mpol_new->refcnt, 1);
2752 * mpol_shared_policy_init - initialize shared policy for inode
2753 * @sp: pointer to inode shared policy
2754 * @mpol: struct mempolicy to install
2756 * Install non-NULL @mpol in inode's shared policy rb-tree.
2757 * On entry, the current task has a reference on a non-NULL @mpol.
2758 * This must be released on exit.
2759 * This is called at get_inode() calls and we can use GFP_KERNEL.
2761 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2765 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2766 rwlock_init(&sp->lock);
2769 struct vm_area_struct pvma;
2770 struct mempolicy *new;
2771 NODEMASK_SCRATCH(scratch);
2775 /* contextualize the tmpfs mount point mempolicy */
2776 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2778 goto free_scratch; /* no valid nodemask intersection */
2781 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2782 task_unlock(current);
2786 /* Create pseudo-vma that contains just the policy */
2787 vma_init(&pvma, NULL);
2788 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2789 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2792 mpol_put(new); /* drop initial ref */
2794 NODEMASK_SCRATCH_FREE(scratch);
2796 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2800 int mpol_set_shared_policy(struct shared_policy *info,
2801 struct vm_area_struct *vma, struct mempolicy *npol)
2804 struct sp_node *new = NULL;
2805 unsigned long sz = vma_pages(vma);
2807 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2809 sz, npol ? npol->mode : -1,
2810 npol ? npol->flags : -1,
2811 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2814 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2818 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2824 /* Free a backing policy store on inode delete. */
2825 void mpol_free_shared_policy(struct shared_policy *p)
2828 struct rb_node *next;
2830 if (!p->root.rb_node)
2832 write_lock(&p->lock);
2833 next = rb_first(&p->root);
2835 n = rb_entry(next, struct sp_node, nd);
2836 next = rb_next(&n->nd);
2839 write_unlock(&p->lock);
2842 #ifdef CONFIG_NUMA_BALANCING
2843 static int __initdata numabalancing_override;
2845 static void __init check_numabalancing_enable(void)
2847 bool numabalancing_default = false;
2849 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2850 numabalancing_default = true;
2852 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2853 if (numabalancing_override)
2854 set_numabalancing_state(numabalancing_override == 1);
2856 if (num_online_nodes() > 1 && !numabalancing_override) {
2857 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2858 numabalancing_default ? "Enabling" : "Disabling");
2859 set_numabalancing_state(numabalancing_default);
2863 static int __init setup_numabalancing(char *str)
2869 if (!strcmp(str, "enable")) {
2870 numabalancing_override = 1;
2872 } else if (!strcmp(str, "disable")) {
2873 numabalancing_override = -1;
2878 pr_warn("Unable to parse numa_balancing=\n");
2882 __setup("numa_balancing=", setup_numabalancing);
2884 static inline void __init check_numabalancing_enable(void)
2887 #endif /* CONFIG_NUMA_BALANCING */
2889 /* assumes fs == KERNEL_DS */
2890 void __init numa_policy_init(void)
2892 nodemask_t interleave_nodes;
2893 unsigned long largest = 0;
2894 int nid, prefer = 0;
2896 policy_cache = kmem_cache_create("numa_policy",
2897 sizeof(struct mempolicy),
2898 0, SLAB_PANIC, NULL);
2900 sn_cache = kmem_cache_create("shared_policy_node",
2901 sizeof(struct sp_node),
2902 0, SLAB_PANIC, NULL);
2904 for_each_node(nid) {
2905 preferred_node_policy[nid] = (struct mempolicy) {
2906 .refcnt = ATOMIC_INIT(1),
2907 .mode = MPOL_PREFERRED,
2908 .flags = MPOL_F_MOF | MPOL_F_MORON,
2909 .nodes = nodemask_of_node(nid),
2914 * Set interleaving policy for system init. Interleaving is only
2915 * enabled across suitably sized nodes (default is >= 16MB), or
2916 * fall back to the largest node if they're all smaller.
2918 nodes_clear(interleave_nodes);
2919 for_each_node_state(nid, N_MEMORY) {
2920 unsigned long total_pages = node_present_pages(nid);
2922 /* Preserve the largest node */
2923 if (largest < total_pages) {
2924 largest = total_pages;
2928 /* Interleave this node? */
2929 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2930 node_set(nid, interleave_nodes);
2933 /* All too small, use the largest */
2934 if (unlikely(nodes_empty(interleave_nodes)))
2935 node_set(prefer, interleave_nodes);
2937 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2938 pr_err("%s: interleaving failed\n", __func__);
2940 check_numabalancing_enable();
2943 /* Reset policy of current process to default */
2944 void numa_default_policy(void)
2946 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2950 * Parse and format mempolicy from/to strings
2953 static const char * const policy_modes[] =
2955 [MPOL_DEFAULT] = "default",
2956 [MPOL_PREFERRED] = "prefer",
2957 [MPOL_BIND] = "bind",
2958 [MPOL_INTERLEAVE] = "interleave",
2959 [MPOL_LOCAL] = "local",
2960 [MPOL_PREFERRED_MANY] = "prefer (many)",
2966 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2967 * @str: string containing mempolicy to parse
2968 * @mpol: pointer to struct mempolicy pointer, returned on success.
2971 * <mode>[=<flags>][:<nodelist>]
2973 * Return: %0 on success, else %1
2975 int mpol_parse_str(char *str, struct mempolicy **mpol)
2977 struct mempolicy *new = NULL;
2978 unsigned short mode_flags;
2980 char *nodelist = strchr(str, ':');
2981 char *flags = strchr(str, '=');
2985 *flags++ = '\0'; /* terminate mode string */
2988 /* NUL-terminate mode or flags string */
2990 if (nodelist_parse(nodelist, nodes))
2992 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2997 mode = match_string(policy_modes, MPOL_MAX, str);
3002 case MPOL_PREFERRED:
3004 * Insist on a nodelist of one node only, although later
3005 * we use first_node(nodes) to grab a single node, so here
3006 * nodelist (or nodes) cannot be empty.
3009 char *rest = nodelist;
3010 while (isdigit(*rest))
3014 if (nodes_empty(nodes))
3018 case MPOL_INTERLEAVE:
3020 * Default to online nodes with memory if no nodelist
3023 nodes = node_states[N_MEMORY];
3027 * Don't allow a nodelist; mpol_new() checks flags
3034 * Insist on a empty nodelist
3039 case MPOL_PREFERRED_MANY:
3042 * Insist on a nodelist
3051 * Currently, we only support two mutually exclusive
3054 if (!strcmp(flags, "static"))
3055 mode_flags |= MPOL_F_STATIC_NODES;
3056 else if (!strcmp(flags, "relative"))
3057 mode_flags |= MPOL_F_RELATIVE_NODES;
3062 new = mpol_new(mode, mode_flags, &nodes);
3067 * Save nodes for mpol_to_str() to show the tmpfs mount options
3068 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3070 if (mode != MPOL_PREFERRED) {
3072 } else if (nodelist) {
3073 nodes_clear(new->nodes);
3074 node_set(first_node(nodes), new->nodes);
3076 new->mode = MPOL_LOCAL;
3080 * Save nodes for contextualization: this will be used to "clone"
3081 * the mempolicy in a specific context [cpuset] at a later time.
3083 new->w.user_nodemask = nodes;
3088 /* Restore string for error message */
3097 #endif /* CONFIG_TMPFS */
3100 * mpol_to_str - format a mempolicy structure for printing
3101 * @buffer: to contain formatted mempolicy string
3102 * @maxlen: length of @buffer
3103 * @pol: pointer to mempolicy to be formatted
3105 * Convert @pol into a string. If @buffer is too short, truncate the string.
3106 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3107 * longest flag, "relative", and to display at least a few node ids.
3109 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3112 nodemask_t nodes = NODE_MASK_NONE;
3113 unsigned short mode = MPOL_DEFAULT;
3114 unsigned short flags = 0;
3116 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3125 case MPOL_PREFERRED:
3126 case MPOL_PREFERRED_MANY:
3128 case MPOL_INTERLEAVE:
3133 snprintf(p, maxlen, "unknown");
3137 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3139 if (flags & MPOL_MODE_FLAGS) {
3140 p += snprintf(p, buffer + maxlen - p, "=");
3143 * Currently, the only defined flags are mutually exclusive
3145 if (flags & MPOL_F_STATIC_NODES)
3146 p += snprintf(p, buffer + maxlen - p, "static");
3147 else if (flags & MPOL_F_RELATIVE_NODES)
3148 p += snprintf(p, buffer + maxlen - p, "relative");
3151 if (!nodes_empty(nodes))
3152 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3153 nodemask_pr_args(&nodes));