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, PAGE_NONE,
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 MA_STATE(mas, &mm->mm_mt, start, start);
791 struct vm_area_struct *prev;
792 struct vm_area_struct *vma;
796 prev = mas_prev(&mas, 0);
798 mas_set(&mas, start);
800 vma = mas_find(&mas, end - 1);
804 if (start > vma->vm_start)
807 for (; vma; vma = mas_next(&mas, end - 1)) {
808 unsigned long vmstart = max(start, vma->vm_start);
809 unsigned long vmend = min(end, vma->vm_end);
811 if (mpol_equal(vma_policy(vma), new_pol))
814 pgoff = vma->vm_pgoff +
815 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
816 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
817 vma->anon_vma, vma->vm_file, pgoff,
818 new_pol, vma->vm_userfaultfd_ctx,
821 /* vma_merge() invalidated the mas */
826 if (vma->vm_start != vmstart) {
827 err = split_vma(vma->vm_mm, vma, vmstart, 1);
830 /* split_vma() invalidated the mas */
833 if (vma->vm_end != vmend) {
834 err = split_vma(vma->vm_mm, vma, vmend, 0);
837 /* split_vma() invalidated the mas */
841 err = vma_replace_policy(vma, new_pol);
852 /* Set the process memory policy */
853 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
856 struct mempolicy *new, *old;
857 NODEMASK_SCRATCH(scratch);
863 new = mpol_new(mode, flags, nodes);
870 ret = mpol_set_nodemask(new, nodes, scratch);
872 task_unlock(current);
877 old = current->mempolicy;
878 current->mempolicy = new;
879 if (new && new->mode == MPOL_INTERLEAVE)
880 current->il_prev = MAX_NUMNODES-1;
881 task_unlock(current);
885 NODEMASK_SCRATCH_FREE(scratch);
890 * Return nodemask for policy for get_mempolicy() query
892 * Called with task's alloc_lock held
894 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
897 if (p == &default_policy)
902 case MPOL_INTERLEAVE:
904 case MPOL_PREFERRED_MANY:
908 /* return empty node mask for local allocation */
915 static int lookup_node(struct mm_struct *mm, unsigned long addr)
917 struct page *p = NULL;
920 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
922 ret = page_to_nid(p);
928 /* Retrieve NUMA policy */
929 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
930 unsigned long addr, unsigned long flags)
933 struct mm_struct *mm = current->mm;
934 struct vm_area_struct *vma = NULL;
935 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
938 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
941 if (flags & MPOL_F_MEMS_ALLOWED) {
942 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
944 *policy = 0; /* just so it's initialized */
946 *nmask = cpuset_current_mems_allowed;
947 task_unlock(current);
951 if (flags & MPOL_F_ADDR) {
953 * Do NOT fall back to task policy if the
954 * vma/shared policy at addr is NULL. We
955 * want to return MPOL_DEFAULT in this case.
958 vma = vma_lookup(mm, addr);
960 mmap_read_unlock(mm);
963 if (vma->vm_ops && vma->vm_ops->get_policy)
964 pol = vma->vm_ops->get_policy(vma, addr);
966 pol = vma->vm_policy;
971 pol = &default_policy; /* indicates default behavior */
973 if (flags & MPOL_F_NODE) {
974 if (flags & MPOL_F_ADDR) {
976 * Take a refcount on the mpol, because we are about to
977 * drop the mmap_lock, after which only "pol" remains
978 * valid, "vma" is stale.
983 mmap_read_unlock(mm);
984 err = lookup_node(mm, addr);
988 } else if (pol == current->mempolicy &&
989 pol->mode == MPOL_INTERLEAVE) {
990 *policy = next_node_in(current->il_prev, pol->nodes);
996 *policy = pol == &default_policy ? MPOL_DEFAULT :
999 * Internal mempolicy flags must be masked off before exposing
1000 * the policy to userspace.
1002 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1007 if (mpol_store_user_nodemask(pol)) {
1008 *nmask = pol->w.user_nodemask;
1011 get_policy_nodemask(pol, nmask);
1012 task_unlock(current);
1019 mmap_read_unlock(mm);
1021 mpol_put(pol_refcount);
1025 #ifdef CONFIG_MIGRATION
1027 * page migration, thp tail pages can be passed.
1029 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1030 unsigned long flags)
1032 struct page *head = compound_head(page);
1034 * Avoid migrating a page that is shared with others.
1036 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1037 if (!isolate_lru_page(head)) {
1038 list_add_tail(&head->lru, pagelist);
1039 mod_node_page_state(page_pgdat(head),
1040 NR_ISOLATED_ANON + page_is_file_lru(head),
1041 thp_nr_pages(head));
1042 } else if (flags & MPOL_MF_STRICT) {
1044 * Non-movable page may reach here. And, there may be
1045 * temporary off LRU pages or non-LRU movable pages.
1046 * Treat them as unmovable pages since they can't be
1047 * isolated, so they can't be moved at the moment. It
1048 * should return -EIO for this case too.
1058 * Migrate pages from one node to a target node.
1059 * Returns error or the number of pages not migrated.
1061 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1065 struct vm_area_struct *vma;
1066 LIST_HEAD(pagelist);
1068 struct migration_target_control mtc = {
1070 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1074 node_set(source, nmask);
1077 * This does not "check" the range but isolates all pages that
1078 * need migration. Between passing in the full user address
1079 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1081 vma = find_vma(mm, 0);
1082 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1083 queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1084 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1086 if (!list_empty(&pagelist)) {
1087 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1088 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1090 putback_movable_pages(&pagelist);
1097 * Move pages between the two nodesets so as to preserve the physical
1098 * layout as much as possible.
1100 * Returns the number of page that could not be moved.
1102 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1103 const nodemask_t *to, int flags)
1109 lru_cache_disable();
1114 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1115 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1116 * bit in 'tmp', and return that <source, dest> pair for migration.
1117 * The pair of nodemasks 'to' and 'from' define the map.
1119 * If no pair of bits is found that way, fallback to picking some
1120 * pair of 'source' and 'dest' bits that are not the same. If the
1121 * 'source' and 'dest' bits are the same, this represents a node
1122 * that will be migrating to itself, so no pages need move.
1124 * If no bits are left in 'tmp', or if all remaining bits left
1125 * in 'tmp' correspond to the same bit in 'to', return false
1126 * (nothing left to migrate).
1128 * This lets us pick a pair of nodes to migrate between, such that
1129 * if possible the dest node is not already occupied by some other
1130 * source node, minimizing the risk of overloading the memory on a
1131 * node that would happen if we migrated incoming memory to a node
1132 * before migrating outgoing memory source that same node.
1134 * A single scan of tmp is sufficient. As we go, we remember the
1135 * most recent <s, d> pair that moved (s != d). If we find a pair
1136 * that not only moved, but what's better, moved to an empty slot
1137 * (d is not set in tmp), then we break out then, with that pair.
1138 * Otherwise when we finish scanning from_tmp, we at least have the
1139 * most recent <s, d> pair that moved. If we get all the way through
1140 * the scan of tmp without finding any node that moved, much less
1141 * moved to an empty node, then there is nothing left worth migrating.
1145 while (!nodes_empty(tmp)) {
1147 int source = NUMA_NO_NODE;
1150 for_each_node_mask(s, tmp) {
1153 * do_migrate_pages() tries to maintain the relative
1154 * node relationship of the pages established between
1155 * threads and memory areas.
1157 * However if the number of source nodes is not equal to
1158 * the number of destination nodes we can not preserve
1159 * this node relative relationship. In that case, skip
1160 * copying memory from a node that is in the destination
1163 * Example: [2,3,4] -> [3,4,5] moves everything.
1164 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1167 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1168 (node_isset(s, *to)))
1171 d = node_remap(s, *from, *to);
1175 source = s; /* Node moved. Memorize */
1178 /* dest not in remaining from nodes? */
1179 if (!node_isset(dest, tmp))
1182 if (source == NUMA_NO_NODE)
1185 node_clear(source, tmp);
1186 err = migrate_to_node(mm, source, dest, flags);
1192 mmap_read_unlock(mm);
1202 * Allocate a new page for page migration based on vma policy.
1203 * Start by assuming the page is mapped by the same vma as contains @start.
1204 * Search forward from there, if not. N.B., this assumes that the
1205 * list of pages handed to migrate_pages()--which is how we get here--
1206 * is in virtual address order.
1208 static struct page *new_page(struct page *page, unsigned long start)
1210 struct folio *dst, *src = page_folio(page);
1211 struct vm_area_struct *vma;
1212 unsigned long address;
1213 VMA_ITERATOR(vmi, current->mm, start);
1214 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1216 for_each_vma(vmi, vma) {
1217 address = page_address_in_vma(page, vma);
1218 if (address != -EFAULT)
1222 if (folio_test_hugetlb(src))
1223 return alloc_huge_page_vma(page_hstate(&src->page),
1226 if (folio_test_large(src))
1227 gfp = GFP_TRANSHUGE;
1230 * if !vma, vma_alloc_folio() will use task or system default policy
1232 dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1233 folio_test_large(src));
1238 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1239 unsigned long flags)
1244 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1245 const nodemask_t *to, int flags)
1250 static struct page *new_page(struct page *page, unsigned long start)
1256 static long do_mbind(unsigned long start, unsigned long len,
1257 unsigned short mode, unsigned short mode_flags,
1258 nodemask_t *nmask, unsigned long flags)
1260 struct mm_struct *mm = current->mm;
1261 struct mempolicy *new;
1265 LIST_HEAD(pagelist);
1267 if (flags & ~(unsigned long)MPOL_MF_VALID)
1269 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1272 if (start & ~PAGE_MASK)
1275 if (mode == MPOL_DEFAULT)
1276 flags &= ~MPOL_MF_STRICT;
1278 len = PAGE_ALIGN(len);
1286 new = mpol_new(mode, mode_flags, nmask);
1288 return PTR_ERR(new);
1290 if (flags & MPOL_MF_LAZY)
1291 new->flags |= MPOL_F_MOF;
1294 * If we are using the default policy then operation
1295 * on discontinuous address spaces is okay after all
1298 flags |= MPOL_MF_DISCONTIG_OK;
1300 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1301 start, start + len, mode, mode_flags,
1302 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1304 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1306 lru_cache_disable();
1309 NODEMASK_SCRATCH(scratch);
1311 mmap_write_lock(mm);
1312 err = mpol_set_nodemask(new, nmask, scratch);
1314 mmap_write_unlock(mm);
1317 NODEMASK_SCRATCH_FREE(scratch);
1322 ret = queue_pages_range(mm, start, end, nmask,
1323 flags | MPOL_MF_INVERT, &pagelist);
1330 err = mbind_range(mm, start, end, new);
1335 if (!list_empty(&pagelist)) {
1336 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1337 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1338 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1340 putback_movable_pages(&pagelist);
1343 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1347 if (!list_empty(&pagelist))
1348 putback_movable_pages(&pagelist);
1351 mmap_write_unlock(mm);
1354 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1360 * User space interface with variable sized bitmaps for nodelists.
1362 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1363 unsigned long maxnode)
1365 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1368 if (in_compat_syscall())
1369 ret = compat_get_bitmap(mask,
1370 (const compat_ulong_t __user *)nmask,
1373 ret = copy_from_user(mask, nmask,
1374 nlongs * sizeof(unsigned long));
1379 if (maxnode % BITS_PER_LONG)
1380 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1385 /* Copy a node mask from user space. */
1386 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1387 unsigned long maxnode)
1390 nodes_clear(*nodes);
1391 if (maxnode == 0 || !nmask)
1393 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1397 * When the user specified more nodes than supported just check
1398 * if the non supported part is all zero, one word at a time,
1399 * starting at the end.
1401 while (maxnode > MAX_NUMNODES) {
1402 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1405 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1408 if (maxnode - bits >= MAX_NUMNODES) {
1411 maxnode = MAX_NUMNODES;
1412 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1418 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1421 /* Copy a kernel node mask to user space */
1422 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1425 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1426 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1427 bool compat = in_compat_syscall();
1430 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1432 if (copy > nbytes) {
1433 if (copy > PAGE_SIZE)
1435 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1438 maxnode = nr_node_ids;
1442 return compat_put_bitmap((compat_ulong_t __user *)mask,
1443 nodes_addr(*nodes), maxnode);
1445 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1448 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1449 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1451 *flags = *mode & MPOL_MODE_FLAGS;
1452 *mode &= ~MPOL_MODE_FLAGS;
1454 if ((unsigned int)(*mode) >= MPOL_MAX)
1456 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1458 if (*flags & MPOL_F_NUMA_BALANCING) {
1459 if (*mode != MPOL_BIND)
1461 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1466 static long kernel_mbind(unsigned long start, unsigned long len,
1467 unsigned long mode, const unsigned long __user *nmask,
1468 unsigned long maxnode, unsigned int flags)
1470 unsigned short mode_flags;
1475 start = untagged_addr(start);
1476 err = sanitize_mpol_flags(&lmode, &mode_flags);
1480 err = get_nodes(&nodes, nmask, maxnode);
1484 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1487 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1488 unsigned long, home_node, unsigned long, flags)
1490 struct mm_struct *mm = current->mm;
1491 struct vm_area_struct *vma;
1492 struct mempolicy *new, *old;
1493 unsigned long vmstart;
1494 unsigned long vmend;
1497 VMA_ITERATOR(vmi, mm, start);
1499 start = untagged_addr(start);
1500 if (start & ~PAGE_MASK)
1503 * flags is used for future extension if any.
1509 * Check home_node is online to avoid accessing uninitialized
1512 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1515 len = PAGE_ALIGN(len);
1522 mmap_write_lock(mm);
1523 for_each_vma_range(vmi, vma, end) {
1525 * If any vma in the range got policy other than MPOL_BIND
1526 * or MPOL_PREFERRED_MANY we return error. We don't reset
1527 * the home node for vmas we already updated before.
1529 old = vma_policy(vma);
1532 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1536 new = mpol_dup(old);
1542 new->home_node = home_node;
1543 vmstart = max(start, vma->vm_start);
1544 vmend = min(end, vma->vm_end);
1545 err = mbind_range(mm, vmstart, vmend, new);
1550 mmap_write_unlock(mm);
1554 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1555 unsigned long, mode, const unsigned long __user *, nmask,
1556 unsigned long, maxnode, unsigned int, flags)
1558 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1561 /* Set the process memory policy */
1562 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1563 unsigned long maxnode)
1565 unsigned short mode_flags;
1570 err = sanitize_mpol_flags(&lmode, &mode_flags);
1574 err = get_nodes(&nodes, nmask, maxnode);
1578 return do_set_mempolicy(lmode, mode_flags, &nodes);
1581 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1582 unsigned long, maxnode)
1584 return kernel_set_mempolicy(mode, nmask, maxnode);
1587 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1588 const unsigned long __user *old_nodes,
1589 const unsigned long __user *new_nodes)
1591 struct mm_struct *mm = NULL;
1592 struct task_struct *task;
1593 nodemask_t task_nodes;
1597 NODEMASK_SCRATCH(scratch);
1602 old = &scratch->mask1;
1603 new = &scratch->mask2;
1605 err = get_nodes(old, old_nodes, maxnode);
1609 err = get_nodes(new, new_nodes, maxnode);
1613 /* Find the mm_struct */
1615 task = pid ? find_task_by_vpid(pid) : current;
1621 get_task_struct(task);
1626 * Check if this process has the right to modify the specified process.
1627 * Use the regular "ptrace_may_access()" checks.
1629 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1636 task_nodes = cpuset_mems_allowed(task);
1637 /* Is the user allowed to access the target nodes? */
1638 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1643 task_nodes = cpuset_mems_allowed(current);
1644 nodes_and(*new, *new, task_nodes);
1645 if (nodes_empty(*new))
1648 err = security_task_movememory(task);
1652 mm = get_task_mm(task);
1653 put_task_struct(task);
1660 err = do_migrate_pages(mm, old, new,
1661 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1665 NODEMASK_SCRATCH_FREE(scratch);
1670 put_task_struct(task);
1675 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1676 const unsigned long __user *, old_nodes,
1677 const unsigned long __user *, new_nodes)
1679 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1683 /* Retrieve NUMA policy */
1684 static int kernel_get_mempolicy(int __user *policy,
1685 unsigned long __user *nmask,
1686 unsigned long maxnode,
1688 unsigned long flags)
1694 if (nmask != NULL && maxnode < nr_node_ids)
1697 addr = untagged_addr(addr);
1699 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1704 if (policy && put_user(pval, policy))
1708 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1713 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1714 unsigned long __user *, nmask, unsigned long, maxnode,
1715 unsigned long, addr, unsigned long, flags)
1717 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1720 bool vma_migratable(struct vm_area_struct *vma)
1722 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1726 * DAX device mappings require predictable access latency, so avoid
1727 * incurring periodic faults.
1729 if (vma_is_dax(vma))
1732 if (is_vm_hugetlb_page(vma) &&
1733 !hugepage_migration_supported(hstate_vma(vma)))
1737 * Migration allocates pages in the highest zone. If we cannot
1738 * do so then migration (at least from node to node) is not
1742 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1748 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1751 struct mempolicy *pol = NULL;
1754 if (vma->vm_ops && vma->vm_ops->get_policy) {
1755 pol = vma->vm_ops->get_policy(vma, addr);
1756 } else if (vma->vm_policy) {
1757 pol = vma->vm_policy;
1760 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1761 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1762 * count on these policies which will be dropped by
1763 * mpol_cond_put() later
1765 if (mpol_needs_cond_ref(pol))
1774 * get_vma_policy(@vma, @addr)
1775 * @vma: virtual memory area whose policy is sought
1776 * @addr: address in @vma for shared policy lookup
1778 * Returns effective policy for a VMA at specified address.
1779 * Falls back to current->mempolicy or system default policy, as necessary.
1780 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1781 * count--added by the get_policy() vm_op, as appropriate--to protect against
1782 * freeing by another task. It is the caller's responsibility to free the
1783 * extra reference for shared policies.
1785 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1788 struct mempolicy *pol = __get_vma_policy(vma, addr);
1791 pol = get_task_policy(current);
1796 bool vma_policy_mof(struct vm_area_struct *vma)
1798 struct mempolicy *pol;
1800 if (vma->vm_ops && vma->vm_ops->get_policy) {
1803 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1804 if (pol && (pol->flags & MPOL_F_MOF))
1811 pol = vma->vm_policy;
1813 pol = get_task_policy(current);
1815 return pol->flags & MPOL_F_MOF;
1818 bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1820 enum zone_type dynamic_policy_zone = policy_zone;
1822 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1825 * if policy->nodes has movable memory only,
1826 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1828 * policy->nodes is intersect with node_states[N_MEMORY].
1829 * so if the following test fails, it implies
1830 * policy->nodes has movable memory only.
1832 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1833 dynamic_policy_zone = ZONE_MOVABLE;
1835 return zone >= dynamic_policy_zone;
1839 * Return a nodemask representing a mempolicy for filtering nodes for
1842 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1844 int mode = policy->mode;
1846 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1847 if (unlikely(mode == MPOL_BIND) &&
1848 apply_policy_zone(policy, gfp_zone(gfp)) &&
1849 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1850 return &policy->nodes;
1852 if (mode == MPOL_PREFERRED_MANY)
1853 return &policy->nodes;
1859 * Return the preferred node id for 'prefer' mempolicy, and return
1860 * the given id for all other policies.
1862 * policy_node() is always coupled with policy_nodemask(), which
1863 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1865 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1867 if (policy->mode == MPOL_PREFERRED) {
1868 nd = first_node(policy->nodes);
1871 * __GFP_THISNODE shouldn't even be used with the bind policy
1872 * because we might easily break the expectation to stay on the
1873 * requested node and not break the policy.
1875 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1878 if ((policy->mode == MPOL_BIND ||
1879 policy->mode == MPOL_PREFERRED_MANY) &&
1880 policy->home_node != NUMA_NO_NODE)
1881 return policy->home_node;
1886 /* Do dynamic interleaving for a process */
1887 static unsigned interleave_nodes(struct mempolicy *policy)
1890 struct task_struct *me = current;
1892 next = next_node_in(me->il_prev, policy->nodes);
1893 if (next < MAX_NUMNODES)
1899 * Depending on the memory policy provide a node from which to allocate the
1902 unsigned int mempolicy_slab_node(void)
1904 struct mempolicy *policy;
1905 int node = numa_mem_id();
1910 policy = current->mempolicy;
1914 switch (policy->mode) {
1915 case MPOL_PREFERRED:
1916 return first_node(policy->nodes);
1918 case MPOL_INTERLEAVE:
1919 return interleave_nodes(policy);
1922 case MPOL_PREFERRED_MANY:
1927 * Follow bind policy behavior and start allocation at the
1930 struct zonelist *zonelist;
1931 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1932 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1933 z = first_zones_zonelist(zonelist, highest_zoneidx,
1935 return z->zone ? zone_to_nid(z->zone) : node;
1946 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1947 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1948 * number of present nodes.
1950 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1952 nodemask_t nodemask = pol->nodes;
1953 unsigned int target, nnodes;
1957 * The barrier will stabilize the nodemask in a register or on
1958 * the stack so that it will stop changing under the code.
1960 * Between first_node() and next_node(), pol->nodes could be changed
1961 * by other threads. So we put pol->nodes in a local stack.
1965 nnodes = nodes_weight(nodemask);
1967 return numa_node_id();
1968 target = (unsigned int)n % nnodes;
1969 nid = first_node(nodemask);
1970 for (i = 0; i < target; i++)
1971 nid = next_node(nid, nodemask);
1975 /* Determine a node number for interleave */
1976 static inline unsigned interleave_nid(struct mempolicy *pol,
1977 struct vm_area_struct *vma, unsigned long addr, int shift)
1983 * for small pages, there is no difference between
1984 * shift and PAGE_SHIFT, so the bit-shift is safe.
1985 * for huge pages, since vm_pgoff is in units of small
1986 * pages, we need to shift off the always 0 bits to get
1989 BUG_ON(shift < PAGE_SHIFT);
1990 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1991 off += (addr - vma->vm_start) >> shift;
1992 return offset_il_node(pol, off);
1994 return interleave_nodes(pol);
1997 #ifdef CONFIG_HUGETLBFS
1999 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2000 * @vma: virtual memory area whose policy is sought
2001 * @addr: address in @vma for shared policy lookup and interleave policy
2002 * @gfp_flags: for requested zone
2003 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2004 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2006 * Returns a nid suitable for a huge page allocation and a pointer
2007 * to the struct mempolicy for conditional unref after allocation.
2008 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2009 * to the mempolicy's @nodemask for filtering the zonelist.
2011 * Must be protected by read_mems_allowed_begin()
2013 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2014 struct mempolicy **mpol, nodemask_t **nodemask)
2019 *mpol = get_vma_policy(vma, addr);
2021 mode = (*mpol)->mode;
2023 if (unlikely(mode == MPOL_INTERLEAVE)) {
2024 nid = interleave_nid(*mpol, vma, addr,
2025 huge_page_shift(hstate_vma(vma)));
2027 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2028 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2029 *nodemask = &(*mpol)->nodes;
2035 * init_nodemask_of_mempolicy
2037 * If the current task's mempolicy is "default" [NULL], return 'false'
2038 * to indicate default policy. Otherwise, extract the policy nodemask
2039 * for 'bind' or 'interleave' policy into the argument nodemask, or
2040 * initialize the argument nodemask to contain the single node for
2041 * 'preferred' or 'local' policy and return 'true' to indicate presence
2042 * of non-default mempolicy.
2044 * We don't bother with reference counting the mempolicy [mpol_get/put]
2045 * because the current task is examining it's own mempolicy and a task's
2046 * mempolicy is only ever changed by the task itself.
2048 * N.B., it is the caller's responsibility to free a returned nodemask.
2050 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2052 struct mempolicy *mempolicy;
2054 if (!(mask && current->mempolicy))
2058 mempolicy = current->mempolicy;
2059 switch (mempolicy->mode) {
2060 case MPOL_PREFERRED:
2061 case MPOL_PREFERRED_MANY:
2063 case MPOL_INTERLEAVE:
2064 *mask = mempolicy->nodes;
2068 init_nodemask_of_node(mask, numa_node_id());
2074 task_unlock(current);
2081 * mempolicy_in_oom_domain
2083 * If tsk's mempolicy is "bind", check for intersection between mask and
2084 * the policy nodemask. Otherwise, return true for all other policies
2085 * including "interleave", as a tsk with "interleave" policy may have
2086 * memory allocated from all nodes in system.
2088 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2090 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2091 const nodemask_t *mask)
2093 struct mempolicy *mempolicy;
2100 mempolicy = tsk->mempolicy;
2101 if (mempolicy && mempolicy->mode == MPOL_BIND)
2102 ret = nodes_intersects(mempolicy->nodes, *mask);
2108 /* Allocate a page in interleaved policy.
2109 Own path because it needs to do special accounting. */
2110 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2115 page = __alloc_pages(gfp, order, nid, NULL);
2116 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2117 if (!static_branch_likely(&vm_numa_stat_key))
2119 if (page && page_to_nid(page) == nid) {
2121 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2127 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2128 int nid, struct mempolicy *pol)
2131 gfp_t preferred_gfp;
2134 * This is a two pass approach. The first pass will only try the
2135 * preferred nodes but skip the direct reclaim and allow the
2136 * allocation to fail, while the second pass will try all the
2139 preferred_gfp = gfp | __GFP_NOWARN;
2140 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2141 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2143 page = __alloc_pages(gfp, order, nid, NULL);
2149 * vma_alloc_folio - Allocate a folio for a VMA.
2151 * @order: Order of the folio.
2152 * @vma: Pointer to VMA or NULL if not available.
2153 * @addr: Virtual address of the allocation. Must be inside @vma.
2154 * @hugepage: For hugepages try only the preferred node if possible.
2156 * Allocate a folio for a specific address in @vma, using the appropriate
2157 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2158 * of the mm_struct of the VMA to prevent it from going away. Should be
2159 * used for all allocations for folios that will be mapped into user space.
2161 * Return: The folio on success or NULL if allocation fails.
2163 struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2164 unsigned long addr, bool hugepage)
2166 struct mempolicy *pol;
2167 int node = numa_node_id();
2168 struct folio *folio;
2172 pol = get_vma_policy(vma, addr);
2174 if (pol->mode == MPOL_INTERLEAVE) {
2178 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2181 page = alloc_page_interleave(gfp, order, nid);
2182 if (page && order > 1)
2183 prep_transhuge_page(page);
2184 folio = (struct folio *)page;
2188 if (pol->mode == MPOL_PREFERRED_MANY) {
2191 node = policy_node(gfp, pol, node);
2193 page = alloc_pages_preferred_many(gfp, order, node, pol);
2195 if (page && order > 1)
2196 prep_transhuge_page(page);
2197 folio = (struct folio *)page;
2201 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2202 int hpage_node = node;
2205 * For hugepage allocation and non-interleave policy which
2206 * allows the current node (or other explicitly preferred
2207 * node) we only try to allocate from the current/preferred
2208 * node and don't fall back to other nodes, as the cost of
2209 * remote accesses would likely offset THP benefits.
2211 * If the policy is interleave or does not allow the current
2212 * node in its nodemask, we allocate the standard way.
2214 if (pol->mode == MPOL_PREFERRED)
2215 hpage_node = first_node(pol->nodes);
2217 nmask = policy_nodemask(gfp, pol);
2218 if (!nmask || node_isset(hpage_node, *nmask)) {
2221 * First, try to allocate THP only on local node, but
2222 * don't reclaim unnecessarily, just compact.
2224 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2225 __GFP_NORETRY, order, hpage_node);
2228 * If hugepage allocations are configured to always
2229 * synchronous compact or the vma has been madvised
2230 * to prefer hugepage backing, retry allowing remote
2231 * memory with both reclaim and compact as well.
2233 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2234 folio = __folio_alloc(gfp, order, hpage_node,
2241 nmask = policy_nodemask(gfp, pol);
2242 preferred_nid = policy_node(gfp, pol, node);
2243 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2248 EXPORT_SYMBOL(vma_alloc_folio);
2251 * alloc_pages - Allocate pages.
2253 * @order: Power of two of number of pages to allocate.
2255 * Allocate 1 << @order contiguous pages. The physical address of the
2256 * first page is naturally aligned (eg an order-3 allocation will be aligned
2257 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2258 * process is honoured when in process context.
2260 * Context: Can be called from any context, providing the appropriate GFP
2262 * Return: The page on success or NULL if allocation fails.
2264 struct page *alloc_pages(gfp_t gfp, unsigned order)
2266 struct mempolicy *pol = &default_policy;
2269 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2270 pol = get_task_policy(current);
2273 * No reference counting needed for current->mempolicy
2274 * nor system default_policy
2276 if (pol->mode == MPOL_INTERLEAVE)
2277 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2278 else if (pol->mode == MPOL_PREFERRED_MANY)
2279 page = alloc_pages_preferred_many(gfp, order,
2280 policy_node(gfp, pol, numa_node_id()), pol);
2282 page = __alloc_pages(gfp, order,
2283 policy_node(gfp, pol, numa_node_id()),
2284 policy_nodemask(gfp, pol));
2288 EXPORT_SYMBOL(alloc_pages);
2290 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2292 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2294 if (page && order > 1)
2295 prep_transhuge_page(page);
2296 return (struct folio *)page;
2298 EXPORT_SYMBOL(folio_alloc);
2300 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2301 struct mempolicy *pol, unsigned long nr_pages,
2302 struct page **page_array)
2305 unsigned long nr_pages_per_node;
2308 unsigned long nr_allocated;
2309 unsigned long total_allocated = 0;
2311 nodes = nodes_weight(pol->nodes);
2312 nr_pages_per_node = nr_pages / nodes;
2313 delta = nr_pages - nodes * nr_pages_per_node;
2315 for (i = 0; i < nodes; i++) {
2317 nr_allocated = __alloc_pages_bulk(gfp,
2318 interleave_nodes(pol), NULL,
2319 nr_pages_per_node + 1, NULL,
2323 nr_allocated = __alloc_pages_bulk(gfp,
2324 interleave_nodes(pol), NULL,
2325 nr_pages_per_node, NULL, page_array);
2328 page_array += nr_allocated;
2329 total_allocated += nr_allocated;
2332 return total_allocated;
2335 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2336 struct mempolicy *pol, unsigned long nr_pages,
2337 struct page **page_array)
2339 gfp_t preferred_gfp;
2340 unsigned long nr_allocated = 0;
2342 preferred_gfp = gfp | __GFP_NOWARN;
2343 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2345 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2346 nr_pages, NULL, page_array);
2348 if (nr_allocated < nr_pages)
2349 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2350 nr_pages - nr_allocated, NULL,
2351 page_array + nr_allocated);
2352 return nr_allocated;
2355 /* alloc pages bulk and mempolicy should be considered at the
2356 * same time in some situation such as vmalloc.
2358 * It can accelerate memory allocation especially interleaving
2361 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2362 unsigned long nr_pages, struct page **page_array)
2364 struct mempolicy *pol = &default_policy;
2366 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2367 pol = get_task_policy(current);
2369 if (pol->mode == MPOL_INTERLEAVE)
2370 return alloc_pages_bulk_array_interleave(gfp, pol,
2371 nr_pages, page_array);
2373 if (pol->mode == MPOL_PREFERRED_MANY)
2374 return alloc_pages_bulk_array_preferred_many(gfp,
2375 numa_node_id(), pol, nr_pages, page_array);
2377 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2378 policy_nodemask(gfp, pol), nr_pages, NULL,
2382 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2384 struct mempolicy *pol = mpol_dup(vma_policy(src));
2387 return PTR_ERR(pol);
2388 dst->vm_policy = pol;
2393 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2394 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2395 * with the mems_allowed returned by cpuset_mems_allowed(). This
2396 * keeps mempolicies cpuset relative after its cpuset moves. See
2397 * further kernel/cpuset.c update_nodemask().
2399 * current's mempolicy may be rebinded by the other task(the task that changes
2400 * cpuset's mems), so we needn't do rebind work for current task.
2403 /* Slow path of a mempolicy duplicate */
2404 struct mempolicy *__mpol_dup(struct mempolicy *old)
2406 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2409 return ERR_PTR(-ENOMEM);
2411 /* task's mempolicy is protected by alloc_lock */
2412 if (old == current->mempolicy) {
2415 task_unlock(current);
2419 if (current_cpuset_is_being_rebound()) {
2420 nodemask_t mems = cpuset_mems_allowed(current);
2421 mpol_rebind_policy(new, &mems);
2423 atomic_set(&new->refcnt, 1);
2427 /* Slow path of a mempolicy comparison */
2428 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2432 if (a->mode != b->mode)
2434 if (a->flags != b->flags)
2436 if (a->home_node != b->home_node)
2438 if (mpol_store_user_nodemask(a))
2439 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2444 case MPOL_INTERLEAVE:
2445 case MPOL_PREFERRED:
2446 case MPOL_PREFERRED_MANY:
2447 return !!nodes_equal(a->nodes, b->nodes);
2457 * Shared memory backing store policy support.
2459 * Remember policies even when nobody has shared memory mapped.
2460 * The policies are kept in Red-Black tree linked from the inode.
2461 * They are protected by the sp->lock rwlock, which should be held
2462 * for any accesses to the tree.
2466 * lookup first element intersecting start-end. Caller holds sp->lock for
2467 * reading or for writing
2469 static struct sp_node *
2470 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2472 struct rb_node *n = sp->root.rb_node;
2475 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2477 if (start >= p->end)
2479 else if (end <= p->start)
2487 struct sp_node *w = NULL;
2488 struct rb_node *prev = rb_prev(n);
2491 w = rb_entry(prev, struct sp_node, nd);
2492 if (w->end <= start)
2496 return rb_entry(n, struct sp_node, nd);
2500 * Insert a new shared policy into the list. Caller holds sp->lock for
2503 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2505 struct rb_node **p = &sp->root.rb_node;
2506 struct rb_node *parent = NULL;
2511 nd = rb_entry(parent, struct sp_node, nd);
2512 if (new->start < nd->start)
2514 else if (new->end > nd->end)
2515 p = &(*p)->rb_right;
2519 rb_link_node(&new->nd, parent, p);
2520 rb_insert_color(&new->nd, &sp->root);
2521 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2522 new->policy ? new->policy->mode : 0);
2525 /* Find shared policy intersecting idx */
2527 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2529 struct mempolicy *pol = NULL;
2532 if (!sp->root.rb_node)
2534 read_lock(&sp->lock);
2535 sn = sp_lookup(sp, idx, idx+1);
2537 mpol_get(sn->policy);
2540 read_unlock(&sp->lock);
2544 static void sp_free(struct sp_node *n)
2546 mpol_put(n->policy);
2547 kmem_cache_free(sn_cache, n);
2551 * mpol_misplaced - check whether current page node is valid in policy
2553 * @page: page to be checked
2554 * @vma: vm area where page mapped
2555 * @addr: virtual address where page mapped
2557 * Lookup current policy node id for vma,addr and "compare to" page's
2558 * node id. Policy determination "mimics" alloc_page_vma().
2559 * Called from fault path where we know the vma and faulting address.
2561 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2562 * policy, or a suitable node ID to allocate a replacement page from.
2564 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2566 struct mempolicy *pol;
2568 int curnid = page_to_nid(page);
2569 unsigned long pgoff;
2570 int thiscpu = raw_smp_processor_id();
2571 int thisnid = cpu_to_node(thiscpu);
2572 int polnid = NUMA_NO_NODE;
2573 int ret = NUMA_NO_NODE;
2575 pol = get_vma_policy(vma, addr);
2576 if (!(pol->flags & MPOL_F_MOF))
2579 switch (pol->mode) {
2580 case MPOL_INTERLEAVE:
2581 pgoff = vma->vm_pgoff;
2582 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2583 polnid = offset_il_node(pol, pgoff);
2586 case MPOL_PREFERRED:
2587 if (node_isset(curnid, pol->nodes))
2589 polnid = first_node(pol->nodes);
2593 polnid = numa_node_id();
2597 /* Optimize placement among multiple nodes via NUMA balancing */
2598 if (pol->flags & MPOL_F_MORON) {
2599 if (node_isset(thisnid, pol->nodes))
2605 case MPOL_PREFERRED_MANY:
2607 * use current page if in policy nodemask,
2608 * else select nearest allowed node, if any.
2609 * If no allowed nodes, use current [!misplaced].
2611 if (node_isset(curnid, pol->nodes))
2613 z = first_zones_zonelist(
2614 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2615 gfp_zone(GFP_HIGHUSER),
2617 polnid = zone_to_nid(z->zone);
2624 /* Migrate the page towards the node whose CPU is referencing it */
2625 if (pol->flags & MPOL_F_MORON) {
2628 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2632 if (curnid != polnid)
2641 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2642 * dropped after task->mempolicy is set to NULL so that any allocation done as
2643 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2646 void mpol_put_task_policy(struct task_struct *task)
2648 struct mempolicy *pol;
2651 pol = task->mempolicy;
2652 task->mempolicy = NULL;
2657 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2659 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2660 rb_erase(&n->nd, &sp->root);
2664 static void sp_node_init(struct sp_node *node, unsigned long start,
2665 unsigned long end, struct mempolicy *pol)
2667 node->start = start;
2672 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2673 struct mempolicy *pol)
2676 struct mempolicy *newpol;
2678 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2682 newpol = mpol_dup(pol);
2683 if (IS_ERR(newpol)) {
2684 kmem_cache_free(sn_cache, n);
2687 newpol->flags |= MPOL_F_SHARED;
2688 sp_node_init(n, start, end, newpol);
2693 /* Replace a policy range. */
2694 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2695 unsigned long end, struct sp_node *new)
2698 struct sp_node *n_new = NULL;
2699 struct mempolicy *mpol_new = NULL;
2703 write_lock(&sp->lock);
2704 n = sp_lookup(sp, start, end);
2705 /* Take care of old policies in the same range. */
2706 while (n && n->start < end) {
2707 struct rb_node *next = rb_next(&n->nd);
2708 if (n->start >= start) {
2714 /* Old policy spanning whole new range. */
2719 *mpol_new = *n->policy;
2720 atomic_set(&mpol_new->refcnt, 1);
2721 sp_node_init(n_new, end, n->end, mpol_new);
2723 sp_insert(sp, n_new);
2732 n = rb_entry(next, struct sp_node, nd);
2736 write_unlock(&sp->lock);
2743 kmem_cache_free(sn_cache, n_new);
2748 write_unlock(&sp->lock);
2750 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2753 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2756 atomic_set(&mpol_new->refcnt, 1);
2761 * mpol_shared_policy_init - initialize shared policy for inode
2762 * @sp: pointer to inode shared policy
2763 * @mpol: struct mempolicy to install
2765 * Install non-NULL @mpol in inode's shared policy rb-tree.
2766 * On entry, the current task has a reference on a non-NULL @mpol.
2767 * This must be released on exit.
2768 * This is called at get_inode() calls and we can use GFP_KERNEL.
2770 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2774 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2775 rwlock_init(&sp->lock);
2778 struct vm_area_struct pvma;
2779 struct mempolicy *new;
2780 NODEMASK_SCRATCH(scratch);
2784 /* contextualize the tmpfs mount point mempolicy */
2785 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2787 goto free_scratch; /* no valid nodemask intersection */
2790 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2791 task_unlock(current);
2795 /* Create pseudo-vma that contains just the policy */
2796 vma_init(&pvma, NULL);
2797 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2798 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2801 mpol_put(new); /* drop initial ref */
2803 NODEMASK_SCRATCH_FREE(scratch);
2805 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2809 int mpol_set_shared_policy(struct shared_policy *info,
2810 struct vm_area_struct *vma, struct mempolicy *npol)
2813 struct sp_node *new = NULL;
2814 unsigned long sz = vma_pages(vma);
2816 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2818 sz, npol ? npol->mode : -1,
2819 npol ? npol->flags : -1,
2820 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2823 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2827 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2833 /* Free a backing policy store on inode delete. */
2834 void mpol_free_shared_policy(struct shared_policy *p)
2837 struct rb_node *next;
2839 if (!p->root.rb_node)
2841 write_lock(&p->lock);
2842 next = rb_first(&p->root);
2844 n = rb_entry(next, struct sp_node, nd);
2845 next = rb_next(&n->nd);
2848 write_unlock(&p->lock);
2851 #ifdef CONFIG_NUMA_BALANCING
2852 static int __initdata numabalancing_override;
2854 static void __init check_numabalancing_enable(void)
2856 bool numabalancing_default = false;
2858 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2859 numabalancing_default = true;
2861 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2862 if (numabalancing_override)
2863 set_numabalancing_state(numabalancing_override == 1);
2865 if (num_online_nodes() > 1 && !numabalancing_override) {
2866 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2867 numabalancing_default ? "Enabling" : "Disabling");
2868 set_numabalancing_state(numabalancing_default);
2872 static int __init setup_numabalancing(char *str)
2878 if (!strcmp(str, "enable")) {
2879 numabalancing_override = 1;
2881 } else if (!strcmp(str, "disable")) {
2882 numabalancing_override = -1;
2887 pr_warn("Unable to parse numa_balancing=\n");
2891 __setup("numa_balancing=", setup_numabalancing);
2893 static inline void __init check_numabalancing_enable(void)
2896 #endif /* CONFIG_NUMA_BALANCING */
2898 /* assumes fs == KERNEL_DS */
2899 void __init numa_policy_init(void)
2901 nodemask_t interleave_nodes;
2902 unsigned long largest = 0;
2903 int nid, prefer = 0;
2905 policy_cache = kmem_cache_create("numa_policy",
2906 sizeof(struct mempolicy),
2907 0, SLAB_PANIC, NULL);
2909 sn_cache = kmem_cache_create("shared_policy_node",
2910 sizeof(struct sp_node),
2911 0, SLAB_PANIC, NULL);
2913 for_each_node(nid) {
2914 preferred_node_policy[nid] = (struct mempolicy) {
2915 .refcnt = ATOMIC_INIT(1),
2916 .mode = MPOL_PREFERRED,
2917 .flags = MPOL_F_MOF | MPOL_F_MORON,
2918 .nodes = nodemask_of_node(nid),
2923 * Set interleaving policy for system init. Interleaving is only
2924 * enabled across suitably sized nodes (default is >= 16MB), or
2925 * fall back to the largest node if they're all smaller.
2927 nodes_clear(interleave_nodes);
2928 for_each_node_state(nid, N_MEMORY) {
2929 unsigned long total_pages = node_present_pages(nid);
2931 /* Preserve the largest node */
2932 if (largest < total_pages) {
2933 largest = total_pages;
2937 /* Interleave this node? */
2938 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2939 node_set(nid, interleave_nodes);
2942 /* All too small, use the largest */
2943 if (unlikely(nodes_empty(interleave_nodes)))
2944 node_set(prefer, interleave_nodes);
2946 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2947 pr_err("%s: interleaving failed\n", __func__);
2949 check_numabalancing_enable();
2952 /* Reset policy of current process to default */
2953 void numa_default_policy(void)
2955 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2959 * Parse and format mempolicy from/to strings
2962 static const char * const policy_modes[] =
2964 [MPOL_DEFAULT] = "default",
2965 [MPOL_PREFERRED] = "prefer",
2966 [MPOL_BIND] = "bind",
2967 [MPOL_INTERLEAVE] = "interleave",
2968 [MPOL_LOCAL] = "local",
2969 [MPOL_PREFERRED_MANY] = "prefer (many)",
2975 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2976 * @str: string containing mempolicy to parse
2977 * @mpol: pointer to struct mempolicy pointer, returned on success.
2980 * <mode>[=<flags>][:<nodelist>]
2982 * Return: %0 on success, else %1
2984 int mpol_parse_str(char *str, struct mempolicy **mpol)
2986 struct mempolicy *new = NULL;
2987 unsigned short mode_flags;
2989 char *nodelist = strchr(str, ':');
2990 char *flags = strchr(str, '=');
2994 *flags++ = '\0'; /* terminate mode string */
2997 /* NUL-terminate mode or flags string */
2999 if (nodelist_parse(nodelist, nodes))
3001 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3006 mode = match_string(policy_modes, MPOL_MAX, str);
3011 case MPOL_PREFERRED:
3013 * Insist on a nodelist of one node only, although later
3014 * we use first_node(nodes) to grab a single node, so here
3015 * nodelist (or nodes) cannot be empty.
3018 char *rest = nodelist;
3019 while (isdigit(*rest))
3023 if (nodes_empty(nodes))
3027 case MPOL_INTERLEAVE:
3029 * Default to online nodes with memory if no nodelist
3032 nodes = node_states[N_MEMORY];
3036 * Don't allow a nodelist; mpol_new() checks flags
3043 * Insist on a empty nodelist
3048 case MPOL_PREFERRED_MANY:
3051 * Insist on a nodelist
3060 * Currently, we only support two mutually exclusive
3063 if (!strcmp(flags, "static"))
3064 mode_flags |= MPOL_F_STATIC_NODES;
3065 else if (!strcmp(flags, "relative"))
3066 mode_flags |= MPOL_F_RELATIVE_NODES;
3071 new = mpol_new(mode, mode_flags, &nodes);
3076 * Save nodes for mpol_to_str() to show the tmpfs mount options
3077 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3079 if (mode != MPOL_PREFERRED) {
3081 } else if (nodelist) {
3082 nodes_clear(new->nodes);
3083 node_set(first_node(nodes), new->nodes);
3085 new->mode = MPOL_LOCAL;
3089 * Save nodes for contextualization: this will be used to "clone"
3090 * the mempolicy in a specific context [cpuset] at a later time.
3092 new->w.user_nodemask = nodes;
3097 /* Restore string for error message */
3106 #endif /* CONFIG_TMPFS */
3109 * mpol_to_str - format a mempolicy structure for printing
3110 * @buffer: to contain formatted mempolicy string
3111 * @maxlen: length of @buffer
3112 * @pol: pointer to mempolicy to be formatted
3114 * Convert @pol into a string. If @buffer is too short, truncate the string.
3115 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3116 * longest flag, "relative", and to display at least a few node ids.
3118 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3121 nodemask_t nodes = NODE_MASK_NONE;
3122 unsigned short mode = MPOL_DEFAULT;
3123 unsigned short flags = 0;
3125 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3134 case MPOL_PREFERRED:
3135 case MPOL_PREFERRED_MANY:
3137 case MPOL_INTERLEAVE:
3142 snprintf(p, maxlen, "unknown");
3146 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3148 if (flags & MPOL_MODE_FLAGS) {
3149 p += snprintf(p, buffer + maxlen - p, "=");
3152 * Currently, the only defined flags are mutually exclusive
3154 if (flags & MPOL_F_STATIC_NODES)
3155 p += snprintf(p, buffer + maxlen - p, "static");
3156 else if (flags & MPOL_F_RELATIVE_NODES)
3157 p += snprintf(p, buffer + maxlen - p, "relative");
3160 if (!nodes_empty(nodes))
3161 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3162 nodemask_pr_args(&nodes));