1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple NUMA memory policy for the Linux kernel.
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * preferred many Try a set of nodes first before normal fallback. This is
35 * similar to preferred without the special case.
37 * default Allocate on the local node first, or when on a VMA
38 * use the process policy. This is what Linux always did
39 * in a NUMA aware kernel and still does by, ahem, default.
41 * The process policy is applied for most non interrupt memory allocations
42 * in that process' context. Interrupts ignore the policies and always
43 * try to allocate on the local CPU. The VMA policy is only applied for memory
44 * allocations for a VMA in the VM.
46 * Currently there are a few corner cases in swapping where the policy
47 * is not applied, but the majority should be handled. When process policy
48 * is used it is not remembered over swap outs/swap ins.
50 * Only the highest zone in the zone hierarchy gets policied. Allocations
51 * requesting a lower zone just use default policy. This implies that
52 * on systems with highmem kernel lowmem allocation don't get policied.
53 * Same with GFP_DMA allocations.
55 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
56 * all users and remembered even when nobody has memory mapped.
60 fix mmap readahead to honour policy and enable policy for any page cache
62 statistics for bigpages
63 global policy for page cache? currently it uses process policy. Requires
65 handle mremap for shared memory (currently ignored for the policy)
67 make bind policy root only? It can trigger oom much faster and the
68 kernel is not always grateful with that.
71 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
73 #include <linux/mempolicy.h>
74 #include <linux/pagewalk.h>
75 #include <linux/highmem.h>
76 #include <linux/hugetlb.h>
77 #include <linux/kernel.h>
78 #include <linux/sched.h>
79 #include <linux/sched/mm.h>
80 #include <linux/sched/numa_balancing.h>
81 #include <linux/sched/task.h>
82 #include <linux/nodemask.h>
83 #include <linux/cpuset.h>
84 #include <linux/slab.h>
85 #include <linux/string.h>
86 #include <linux/export.h>
87 #include <linux/nsproxy.h>
88 #include <linux/interrupt.h>
89 #include <linux/init.h>
90 #include <linux/compat.h>
91 #include <linux/ptrace.h>
92 #include <linux/swap.h>
93 #include <linux/seq_file.h>
94 #include <linux/proc_fs.h>
95 #include <linux/migrate.h>
96 #include <linux/ksm.h>
97 #include <linux/rmap.h>
98 #include <linux/security.h>
99 #include <linux/syscalls.h>
100 #include <linux/ctype.h>
101 #include <linux/mm_inline.h>
102 #include <linux/mmu_notifier.h>
103 #include <linux/printk.h>
104 #include <linux/swapops.h>
106 #include <asm/tlbflush.h>
107 #include <linux/uaccess.h>
109 #include "internal.h"
112 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
113 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
115 static struct kmem_cache *policy_cache;
116 static struct kmem_cache *sn_cache;
118 /* Highest zone. An specific allocation for a zone below that is not
120 enum zone_type policy_zone = 0;
123 * run-time system-wide default policy => local allocation
125 static struct mempolicy default_policy = {
126 .refcnt = ATOMIC_INIT(1), /* never free it */
130 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
133 * numa_map_to_online_node - Find closest online node
134 * @node: Node id to start the search
136 * Lookup the next closest node by distance if @nid is not online.
138 * Return: this @node if it is online, otherwise the closest node by distance
140 int numa_map_to_online_node(int node)
142 int min_dist = INT_MAX, dist, n, min_node;
144 if (node == NUMA_NO_NODE || node_online(node))
148 for_each_online_node(n) {
149 dist = node_distance(node, n);
150 if (dist < min_dist) {
158 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
160 struct mempolicy *get_task_policy(struct task_struct *p)
162 struct mempolicy *pol = p->mempolicy;
168 node = numa_node_id();
169 if (node != NUMA_NO_NODE) {
170 pol = &preferred_node_policy[node];
171 /* preferred_node_policy is not initialised early in boot */
176 return &default_policy;
179 static const struct mempolicy_operations {
180 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
181 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
182 } mpol_ops[MPOL_MAX];
184 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
186 return pol->flags & MPOL_MODE_FLAGS;
189 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
190 const nodemask_t *rel)
193 nodes_fold(tmp, *orig, nodes_weight(*rel));
194 nodes_onto(*ret, tmp, *rel);
197 static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
199 if (nodes_empty(*nodes))
205 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
207 if (nodes_empty(*nodes))
210 nodes_clear(pol->nodes);
211 node_set(first_node(*nodes), pol->nodes);
216 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
217 * any, for the new policy. mpol_new() has already validated the nodes
218 * parameter with respect to the policy mode and flags.
220 * Must be called holding task's alloc_lock to protect task's mems_allowed
221 * and mempolicy. May also be called holding the mmap_lock for write.
223 static int mpol_set_nodemask(struct mempolicy *pol,
224 const nodemask_t *nodes, struct nodemask_scratch *nsc)
229 * Default (pol==NULL) resp. local memory policies are not a
230 * subject of any remapping. They also do not need any special
233 if (!pol || pol->mode == MPOL_LOCAL)
237 nodes_and(nsc->mask1,
238 cpuset_current_mems_allowed, node_states[N_MEMORY]);
242 if (pol->flags & MPOL_F_RELATIVE_NODES)
243 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
245 nodes_and(nsc->mask2, *nodes, nsc->mask1);
247 if (mpol_store_user_nodemask(pol))
248 pol->w.user_nodemask = *nodes;
250 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
252 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
257 * This function just creates a new policy, does some check and simple
258 * initialization. You must invoke mpol_set_nodemask() to set nodes.
260 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
263 struct mempolicy *policy;
265 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
266 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
268 if (mode == MPOL_DEFAULT) {
269 if (nodes && !nodes_empty(*nodes))
270 return ERR_PTR(-EINVAL);
276 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
277 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
278 * All other modes require a valid pointer to a non-empty nodemask.
280 if (mode == MPOL_PREFERRED) {
281 if (nodes_empty(*nodes)) {
282 if (((flags & MPOL_F_STATIC_NODES) ||
283 (flags & MPOL_F_RELATIVE_NODES)))
284 return ERR_PTR(-EINVAL);
288 } else if (mode == MPOL_LOCAL) {
289 if (!nodes_empty(*nodes) ||
290 (flags & MPOL_F_STATIC_NODES) ||
291 (flags & MPOL_F_RELATIVE_NODES))
292 return ERR_PTR(-EINVAL);
293 } else if (nodes_empty(*nodes))
294 return ERR_PTR(-EINVAL);
295 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
297 return ERR_PTR(-ENOMEM);
298 atomic_set(&policy->refcnt, 1);
300 policy->flags = flags;
301 policy->home_node = NUMA_NO_NODE;
306 /* Slow path of a mpol destructor. */
307 void __mpol_put(struct mempolicy *p)
309 if (!atomic_dec_and_test(&p->refcnt))
311 kmem_cache_free(policy_cache, p);
314 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
318 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
322 if (pol->flags & MPOL_F_STATIC_NODES)
323 nodes_and(tmp, pol->w.user_nodemask, *nodes);
324 else if (pol->flags & MPOL_F_RELATIVE_NODES)
325 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
327 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
329 pol->w.cpuset_mems_allowed = *nodes;
332 if (nodes_empty(tmp))
338 static void mpol_rebind_preferred(struct mempolicy *pol,
339 const nodemask_t *nodes)
341 pol->w.cpuset_mems_allowed = *nodes;
345 * mpol_rebind_policy - Migrate a policy to a different set of nodes
347 * Per-vma policies are protected by mmap_lock. Allocations using per-task
348 * policies are protected by task->mems_allowed_seq to prevent a premature
349 * OOM/allocation failure due to parallel nodemask modification.
351 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
355 if (!mpol_store_user_nodemask(pol) &&
356 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
359 mpol_ops[pol->mode].rebind(pol, newmask);
363 * Wrapper for mpol_rebind_policy() that just requires task
364 * pointer, and updates task mempolicy.
366 * Called with task's alloc_lock held.
369 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
371 mpol_rebind_policy(tsk->mempolicy, new);
375 * Rebind each vma in mm to new nodemask.
377 * Call holding a reference to mm. Takes mm->mmap_lock during call.
380 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
382 struct vm_area_struct *vma;
385 for (vma = mm->mmap; vma; vma = vma->vm_next)
386 mpol_rebind_policy(vma->vm_policy, new);
387 mmap_write_unlock(mm);
390 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
392 .rebind = mpol_rebind_default,
394 [MPOL_INTERLEAVE] = {
395 .create = mpol_new_nodemask,
396 .rebind = mpol_rebind_nodemask,
399 .create = mpol_new_preferred,
400 .rebind = mpol_rebind_preferred,
403 .create = mpol_new_nodemask,
404 .rebind = mpol_rebind_nodemask,
407 .rebind = mpol_rebind_default,
409 [MPOL_PREFERRED_MANY] = {
410 .create = mpol_new_nodemask,
411 .rebind = mpol_rebind_preferred,
415 static int migrate_page_add(struct page *page, struct list_head *pagelist,
416 unsigned long flags);
419 struct list_head *pagelist;
424 struct vm_area_struct *first;
428 * Check if the page's nid is in qp->nmask.
430 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
431 * in the invert of qp->nmask.
433 static inline bool queue_pages_required(struct page *page,
434 struct queue_pages *qp)
436 int nid = page_to_nid(page);
437 unsigned long flags = qp->flags;
439 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
443 * queue_pages_pmd() has four possible return values:
444 * 0 - pages are placed on the right node or queued successfully, or
445 * special page is met, i.e. huge zero page.
446 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
449 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
450 * existing page was already on a node that does not follow the
453 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
454 unsigned long end, struct mm_walk *walk)
459 struct queue_pages *qp = walk->private;
462 if (unlikely(is_pmd_migration_entry(*pmd))) {
466 page = pmd_page(*pmd);
467 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)) {
492 * Scan through pages checking if pages follow certain conditions,
493 * and move them to the pagelist if they do.
495 * queue_pages_pte_range() has three possible return values:
496 * 0 - pages are placed on the right node or queued successfully, or
497 * special page is met, i.e. zero page.
498 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
500 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
501 * on a node that does not follow the policy.
503 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
504 unsigned long end, struct mm_walk *walk)
506 struct vm_area_struct *vma = walk->vma;
508 struct queue_pages *qp = walk->private;
509 unsigned long flags = qp->flags;
511 bool has_unmovable = false;
512 pte_t *pte, *mapped_pte;
515 ptl = pmd_trans_huge_lock(pmd, vma);
517 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
521 /* THP was split, fall through to pte walk */
523 if (pmd_trans_unstable(pmd))
526 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
527 for (; addr != end; pte++, addr += PAGE_SIZE) {
528 if (!pte_present(*pte))
530 page = vm_normal_page(vma, addr, *pte);
534 * vm_normal_page() filters out zero pages, but there might
535 * still be PageReserved pages to skip, perhaps in a VDSO.
537 if (PageReserved(page))
539 if (!queue_pages_required(page, qp))
541 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
542 /* MPOL_MF_STRICT must be specified if we get here */
543 if (!vma_migratable(vma)) {
544 has_unmovable = true;
549 * Do not abort immediately since there may be
550 * temporary off LRU pages in the range. Still
551 * need migrate other LRU pages.
553 if (migrate_page_add(page, qp->pagelist, flags))
554 has_unmovable = true;
558 pte_unmap_unlock(mapped_pte, ptl);
564 return addr != end ? -EIO : 0;
567 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
568 unsigned long addr, unsigned long end,
569 struct mm_walk *walk)
572 #ifdef CONFIG_HUGETLB_PAGE
573 struct queue_pages *qp = walk->private;
574 unsigned long flags = (qp->flags & MPOL_MF_VALID);
579 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
580 entry = huge_ptep_get(pte);
581 if (!pte_present(entry))
583 page = pte_page(entry);
584 if (!queue_pages_required(page, qp))
587 if (flags == MPOL_MF_STRICT) {
589 * STRICT alone means only detecting misplaced page and no
590 * need to further check other vma.
596 if (!vma_migratable(walk->vma)) {
598 * Must be STRICT with MOVE*, otherwise .test_walk() have
599 * stopped walking current vma.
600 * Detecting misplaced page but allow migrating pages which
607 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
608 if (flags & (MPOL_MF_MOVE_ALL) ||
609 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
610 if (!isolate_huge_page(page, qp->pagelist) &&
611 (flags & MPOL_MF_STRICT))
613 * Failed to isolate page but allow migrating pages
614 * which have been queued.
626 #ifdef CONFIG_NUMA_BALANCING
628 * This is used to mark a range of virtual addresses to be inaccessible.
629 * These are later cleared by a NUMA hinting fault. Depending on these
630 * faults, pages may be migrated for better NUMA placement.
632 * This is assuming that NUMA faults are handled using PROT_NONE. If
633 * an architecture makes a different choice, it will need further
634 * changes to the core.
636 unsigned long change_prot_numa(struct vm_area_struct *vma,
637 unsigned long addr, unsigned long end)
641 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
643 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
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 *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 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
674 ((vma->vm_end < qp->end) &&
675 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
676 /* hole at middle or tail of range */
680 * Need check MPOL_MF_STRICT to return -EIO if possible
681 * regardless of vma_migratable
683 if (!vma_migratable(vma) &&
684 !(flags & MPOL_MF_STRICT))
690 if (flags & MPOL_MF_LAZY) {
691 /* Similar to task_numa_work, skip inaccessible VMAs */
692 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
693 !(vma->vm_flags & VM_MIXEDMAP))
694 change_prot_numa(vma, start, endvma);
698 /* queue pages from current vma */
699 if (flags & MPOL_MF_VALID)
704 static const struct mm_walk_ops queue_pages_walk_ops = {
705 .hugetlb_entry = queue_pages_hugetlb,
706 .pmd_entry = queue_pages_pte_range,
707 .test_walk = queue_pages_test_walk,
711 * Walk through page tables and collect pages to be migrated.
713 * If pages found in a given range are on a set of nodes (determined by
714 * @nodes and @flags,) it's isolated and queued to the pagelist which is
715 * passed via @private.
717 * queue_pages_range() has three possible return values:
718 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
720 * 0 - queue pages successfully or no misplaced page.
721 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
722 * memory range specified by nodemask and maxnode points outside
723 * your accessible address space (-EFAULT)
726 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
727 nodemask_t *nodes, unsigned long flags,
728 struct list_head *pagelist)
731 struct queue_pages qp = {
732 .pagelist = pagelist,
740 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
743 /* whole range in hole */
750 * Apply policy to a single VMA
751 * This must be called with the mmap_lock held for writing.
753 static int vma_replace_policy(struct vm_area_struct *vma,
754 struct mempolicy *pol)
757 struct mempolicy *old;
758 struct mempolicy *new;
760 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
761 vma->vm_start, vma->vm_end, vma->vm_pgoff,
762 vma->vm_ops, vma->vm_file,
763 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
769 if (vma->vm_ops && vma->vm_ops->set_policy) {
770 err = vma->vm_ops->set_policy(vma, new);
775 old = vma->vm_policy;
776 vma->vm_policy = new; /* protected by mmap_lock */
785 /* Step 2: apply policy to a range and do splits. */
786 static int mbind_range(struct mm_struct *mm, unsigned long start,
787 unsigned long end, struct mempolicy *new_pol)
789 struct vm_area_struct *prev;
790 struct vm_area_struct *vma;
793 unsigned long vmstart;
796 vma = find_vma(mm, start);
800 if (start > vma->vm_start)
803 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
804 vmstart = max(start, vma->vm_start);
805 vmend = min(end, vma->vm_end);
807 if (mpol_equal(vma_policy(vma), new_pol))
810 pgoff = vma->vm_pgoff +
811 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
812 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
813 vma->anon_vma, vma->vm_file, pgoff,
814 new_pol, vma->vm_userfaultfd_ctx,
820 if (vma->vm_start != vmstart) {
821 err = split_vma(vma->vm_mm, vma, vmstart, 1);
825 if (vma->vm_end != vmend) {
826 err = split_vma(vma->vm_mm, vma, vmend, 0);
831 err = vma_replace_policy(vma, new_pol);
840 /* Set the process memory policy */
841 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
844 struct mempolicy *new, *old;
845 NODEMASK_SCRATCH(scratch);
851 new = mpol_new(mode, flags, nodes);
857 ret = mpol_set_nodemask(new, nodes, scratch);
863 old = current->mempolicy;
864 current->mempolicy = new;
865 if (new && new->mode == MPOL_INTERLEAVE)
866 current->il_prev = MAX_NUMNODES-1;
867 task_unlock(current);
871 NODEMASK_SCRATCH_FREE(scratch);
876 * Return nodemask for policy for get_mempolicy() query
878 * Called with task's alloc_lock held
880 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
883 if (p == &default_policy)
888 case MPOL_INTERLEAVE:
890 case MPOL_PREFERRED_MANY:
894 /* return empty node mask for local allocation */
901 static int lookup_node(struct mm_struct *mm, unsigned long addr)
903 struct page *p = NULL;
907 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
909 err = page_to_nid(p);
913 mmap_read_unlock(mm);
917 /* Retrieve NUMA policy */
918 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
919 unsigned long addr, unsigned long flags)
922 struct mm_struct *mm = current->mm;
923 struct vm_area_struct *vma = NULL;
924 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
927 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
930 if (flags & MPOL_F_MEMS_ALLOWED) {
931 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
933 *policy = 0; /* just so it's initialized */
935 *nmask = cpuset_current_mems_allowed;
936 task_unlock(current);
940 if (flags & MPOL_F_ADDR) {
942 * Do NOT fall back to task policy if the
943 * vma/shared policy at addr is NULL. We
944 * want to return MPOL_DEFAULT in this case.
947 vma = vma_lookup(mm, addr);
949 mmap_read_unlock(mm);
952 if (vma->vm_ops && vma->vm_ops->get_policy)
953 pol = vma->vm_ops->get_policy(vma, addr);
955 pol = vma->vm_policy;
960 pol = &default_policy; /* indicates default behavior */
962 if (flags & MPOL_F_NODE) {
963 if (flags & MPOL_F_ADDR) {
965 * Take a refcount on the mpol, lookup_node()
966 * will drop the mmap_lock, so after calling
967 * lookup_node() only "pol" remains valid, "vma"
973 err = lookup_node(mm, addr);
977 } else if (pol == current->mempolicy &&
978 pol->mode == MPOL_INTERLEAVE) {
979 *policy = next_node_in(current->il_prev, pol->nodes);
985 *policy = pol == &default_policy ? MPOL_DEFAULT :
988 * Internal mempolicy flags must be masked off before exposing
989 * the policy to userspace.
991 *policy |= (pol->flags & MPOL_MODE_FLAGS);
996 if (mpol_store_user_nodemask(pol)) {
997 *nmask = pol->w.user_nodemask;
1000 get_policy_nodemask(pol, nmask);
1001 task_unlock(current);
1008 mmap_read_unlock(mm);
1010 mpol_put(pol_refcount);
1014 #ifdef CONFIG_MIGRATION
1016 * page migration, thp tail pages can be passed.
1018 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1019 unsigned long flags)
1021 struct page *head = compound_head(page);
1023 * Avoid migrating a page that is shared with others.
1025 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1026 if (!isolate_lru_page(head)) {
1027 list_add_tail(&head->lru, pagelist);
1028 mod_node_page_state(page_pgdat(head),
1029 NR_ISOLATED_ANON + page_is_file_lru(head),
1030 thp_nr_pages(head));
1031 } else if (flags & MPOL_MF_STRICT) {
1033 * Non-movable page may reach here. And, there may be
1034 * temporary off LRU pages or non-LRU movable pages.
1035 * Treat them as unmovable pages since they can't be
1036 * isolated, so they can't be moved at the moment. It
1037 * should return -EIO for this case too.
1047 * Migrate pages from one node to a target node.
1048 * Returns error or the number of pages not migrated.
1050 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1054 LIST_HEAD(pagelist);
1056 struct migration_target_control mtc = {
1058 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1062 node_set(source, nmask);
1065 * This does not "check" the range but isolates all pages that
1066 * need migration. Between passing in the full user address
1067 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1069 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1070 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1071 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1073 if (!list_empty(&pagelist)) {
1074 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1075 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1077 putback_movable_pages(&pagelist);
1084 * Move pages between the two nodesets so as to preserve the physical
1085 * layout as much as possible.
1087 * Returns the number of page that could not be moved.
1089 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1090 const nodemask_t *to, int flags)
1096 lru_cache_disable();
1101 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1102 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1103 * bit in 'tmp', and return that <source, dest> pair for migration.
1104 * The pair of nodemasks 'to' and 'from' define the map.
1106 * If no pair of bits is found that way, fallback to picking some
1107 * pair of 'source' and 'dest' bits that are not the same. If the
1108 * 'source' and 'dest' bits are the same, this represents a node
1109 * that will be migrating to itself, so no pages need move.
1111 * If no bits are left in 'tmp', or if all remaining bits left
1112 * in 'tmp' correspond to the same bit in 'to', return false
1113 * (nothing left to migrate).
1115 * This lets us pick a pair of nodes to migrate between, such that
1116 * if possible the dest node is not already occupied by some other
1117 * source node, minimizing the risk of overloading the memory on a
1118 * node that would happen if we migrated incoming memory to a node
1119 * before migrating outgoing memory source that same node.
1121 * A single scan of tmp is sufficient. As we go, we remember the
1122 * most recent <s, d> pair that moved (s != d). If we find a pair
1123 * that not only moved, but what's better, moved to an empty slot
1124 * (d is not set in tmp), then we break out then, with that pair.
1125 * Otherwise when we finish scanning from_tmp, we at least have the
1126 * most recent <s, d> pair that moved. If we get all the way through
1127 * the scan of tmp without finding any node that moved, much less
1128 * moved to an empty node, then there is nothing left worth migrating.
1132 while (!nodes_empty(tmp)) {
1134 int source = NUMA_NO_NODE;
1137 for_each_node_mask(s, tmp) {
1140 * do_migrate_pages() tries to maintain the relative
1141 * node relationship of the pages established between
1142 * threads and memory areas.
1144 * However if the number of source nodes is not equal to
1145 * the number of destination nodes we can not preserve
1146 * this node relative relationship. In that case, skip
1147 * copying memory from a node that is in the destination
1150 * Example: [2,3,4] -> [3,4,5] moves everything.
1151 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1154 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1155 (node_isset(s, *to)))
1158 d = node_remap(s, *from, *to);
1162 source = s; /* Node moved. Memorize */
1165 /* dest not in remaining from nodes? */
1166 if (!node_isset(dest, tmp))
1169 if (source == NUMA_NO_NODE)
1172 node_clear(source, tmp);
1173 err = migrate_to_node(mm, source, dest, flags);
1179 mmap_read_unlock(mm);
1189 * Allocate a new page for page migration based on vma policy.
1190 * Start by assuming the page is mapped by the same vma as contains @start.
1191 * Search forward from there, if not. N.B., this assumes that the
1192 * list of pages handed to migrate_pages()--which is how we get here--
1193 * is in virtual address order.
1195 static struct page *new_page(struct page *page, unsigned long start)
1197 struct vm_area_struct *vma;
1198 unsigned long address;
1200 vma = find_vma(current->mm, start);
1202 address = page_address_in_vma(page, vma);
1203 if (address != -EFAULT)
1208 if (PageHuge(page)) {
1209 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1211 } else if (PageTransHuge(page)) {
1214 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1218 prep_transhuge_page(thp);
1222 * if !vma, alloc_page_vma() will use task or system default policy
1224 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
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 = (len + PAGE_SIZE - 1) & PAGE_MASK;
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 / 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;
1484 unsigned long vmstart;
1485 unsigned long vmend;
1489 start = untagged_addr(start);
1490 if (start & ~PAGE_MASK)
1493 * flags is used for future extension if any.
1499 * Check home_node is online to avoid accessing uninitialized
1502 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1505 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1512 mmap_write_lock(mm);
1513 vma = find_vma(mm, start);
1514 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
1516 vmstart = max(start, vma->vm_start);
1517 vmend = min(end, vma->vm_end);
1518 new = mpol_dup(vma_policy(vma));
1524 * Only update home node if there is an existing vma policy
1530 * If any vma in the range got policy other than MPOL_BIND
1531 * or MPOL_PREFERRED_MANY we return error. We don't reset
1532 * the home node for vmas we already updated before.
1534 if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1539 new->home_node = home_node;
1540 err = mbind_range(mm, vmstart, vmend, new);
1545 mmap_write_unlock(mm);
1549 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1550 unsigned long, mode, const unsigned long __user *, nmask,
1551 unsigned long, maxnode, unsigned int, flags)
1553 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1556 /* Set the process memory policy */
1557 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1558 unsigned long maxnode)
1560 unsigned short mode_flags;
1565 err = sanitize_mpol_flags(&lmode, &mode_flags);
1569 err = get_nodes(&nodes, nmask, maxnode);
1573 return do_set_mempolicy(lmode, mode_flags, &nodes);
1576 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1577 unsigned long, maxnode)
1579 return kernel_set_mempolicy(mode, nmask, maxnode);
1582 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1583 const unsigned long __user *old_nodes,
1584 const unsigned long __user *new_nodes)
1586 struct mm_struct *mm = NULL;
1587 struct task_struct *task;
1588 nodemask_t task_nodes;
1592 NODEMASK_SCRATCH(scratch);
1597 old = &scratch->mask1;
1598 new = &scratch->mask2;
1600 err = get_nodes(old, old_nodes, maxnode);
1604 err = get_nodes(new, new_nodes, maxnode);
1608 /* Find the mm_struct */
1610 task = pid ? find_task_by_vpid(pid) : current;
1616 get_task_struct(task);
1621 * Check if this process has the right to modify the specified process.
1622 * Use the regular "ptrace_may_access()" checks.
1624 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1631 task_nodes = cpuset_mems_allowed(task);
1632 /* Is the user allowed to access the target nodes? */
1633 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1638 task_nodes = cpuset_mems_allowed(current);
1639 nodes_and(*new, *new, task_nodes);
1640 if (nodes_empty(*new))
1643 err = security_task_movememory(task);
1647 mm = get_task_mm(task);
1648 put_task_struct(task);
1655 err = do_migrate_pages(mm, old, new,
1656 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1660 NODEMASK_SCRATCH_FREE(scratch);
1665 put_task_struct(task);
1670 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1671 const unsigned long __user *, old_nodes,
1672 const unsigned long __user *, new_nodes)
1674 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1678 /* Retrieve NUMA policy */
1679 static int kernel_get_mempolicy(int __user *policy,
1680 unsigned long __user *nmask,
1681 unsigned long maxnode,
1683 unsigned long flags)
1689 if (nmask != NULL && maxnode < nr_node_ids)
1692 addr = untagged_addr(addr);
1694 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1699 if (policy && put_user(pval, policy))
1703 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1708 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1709 unsigned long __user *, nmask, unsigned long, maxnode,
1710 unsigned long, addr, unsigned long, flags)
1712 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1715 bool vma_migratable(struct vm_area_struct *vma)
1717 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1721 * DAX device mappings require predictable access latency, so avoid
1722 * incurring periodic faults.
1724 if (vma_is_dax(vma))
1727 if (is_vm_hugetlb_page(vma) &&
1728 !hugepage_migration_supported(hstate_vma(vma)))
1732 * Migration allocates pages in the highest zone. If we cannot
1733 * do so then migration (at least from node to node) is not
1737 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1743 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1746 struct mempolicy *pol = NULL;
1749 if (vma->vm_ops && vma->vm_ops->get_policy) {
1750 pol = vma->vm_ops->get_policy(vma, addr);
1751 } else if (vma->vm_policy) {
1752 pol = vma->vm_policy;
1755 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1756 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1757 * count on these policies which will be dropped by
1758 * mpol_cond_put() later
1760 if (mpol_needs_cond_ref(pol))
1769 * get_vma_policy(@vma, @addr)
1770 * @vma: virtual memory area whose policy is sought
1771 * @addr: address in @vma for shared policy lookup
1773 * Returns effective policy for a VMA at specified address.
1774 * Falls back to current->mempolicy or system default policy, as necessary.
1775 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1776 * count--added by the get_policy() vm_op, as appropriate--to protect against
1777 * freeing by another task. It is the caller's responsibility to free the
1778 * extra reference for shared policies.
1780 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1783 struct mempolicy *pol = __get_vma_policy(vma, addr);
1786 pol = get_task_policy(current);
1791 bool vma_policy_mof(struct vm_area_struct *vma)
1793 struct mempolicy *pol;
1795 if (vma->vm_ops && vma->vm_ops->get_policy) {
1798 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1799 if (pol && (pol->flags & MPOL_F_MOF))
1806 pol = vma->vm_policy;
1808 pol = get_task_policy(current);
1810 return pol->flags & MPOL_F_MOF;
1813 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1815 enum zone_type dynamic_policy_zone = policy_zone;
1817 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1820 * if policy->nodes has movable memory only,
1821 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1823 * policy->nodes is intersect with node_states[N_MEMORY].
1824 * so if the following test fails, it implies
1825 * policy->nodes has movable memory only.
1827 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1828 dynamic_policy_zone = ZONE_MOVABLE;
1830 return zone >= dynamic_policy_zone;
1834 * Return a nodemask representing a mempolicy for filtering nodes for
1837 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1839 int mode = policy->mode;
1841 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1842 if (unlikely(mode == MPOL_BIND) &&
1843 apply_policy_zone(policy, gfp_zone(gfp)) &&
1844 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1845 return &policy->nodes;
1847 if (mode == MPOL_PREFERRED_MANY)
1848 return &policy->nodes;
1854 * Return the preferred node id for 'prefer' mempolicy, and return
1855 * the given id for all other policies.
1857 * policy_node() is always coupled with policy_nodemask(), which
1858 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1860 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1862 if (policy->mode == MPOL_PREFERRED) {
1863 nd = first_node(policy->nodes);
1866 * __GFP_THISNODE shouldn't even be used with the bind policy
1867 * because we might easily break the expectation to stay on the
1868 * requested node and not break the policy.
1870 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1873 if ((policy->mode == MPOL_BIND ||
1874 policy->mode == MPOL_PREFERRED_MANY) &&
1875 policy->home_node != NUMA_NO_NODE)
1876 return policy->home_node;
1881 /* Do dynamic interleaving for a process */
1882 static unsigned interleave_nodes(struct mempolicy *policy)
1885 struct task_struct *me = current;
1887 next = next_node_in(me->il_prev, policy->nodes);
1888 if (next < MAX_NUMNODES)
1894 * Depending on the memory policy provide a node from which to allocate the
1897 unsigned int mempolicy_slab_node(void)
1899 struct mempolicy *policy;
1900 int node = numa_mem_id();
1905 policy = current->mempolicy;
1909 switch (policy->mode) {
1910 case MPOL_PREFERRED:
1911 return first_node(policy->nodes);
1913 case MPOL_INTERLEAVE:
1914 return interleave_nodes(policy);
1917 case MPOL_PREFERRED_MANY:
1922 * Follow bind policy behavior and start allocation at the
1925 struct zonelist *zonelist;
1926 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1927 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1928 z = first_zones_zonelist(zonelist, highest_zoneidx,
1930 return z->zone ? zone_to_nid(z->zone) : node;
1941 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1942 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1943 * number of present nodes.
1945 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1947 nodemask_t nodemask = pol->nodes;
1948 unsigned int target, nnodes;
1952 * The barrier will stabilize the nodemask in a register or on
1953 * the stack so that it will stop changing under the code.
1955 * Between first_node() and next_node(), pol->nodes could be changed
1956 * by other threads. So we put pol->nodes in a local stack.
1960 nnodes = nodes_weight(nodemask);
1962 return numa_node_id();
1963 target = (unsigned int)n % nnodes;
1964 nid = first_node(nodemask);
1965 for (i = 0; i < target; i++)
1966 nid = next_node(nid, nodemask);
1970 /* Determine a node number for interleave */
1971 static inline unsigned interleave_nid(struct mempolicy *pol,
1972 struct vm_area_struct *vma, unsigned long addr, int shift)
1978 * for small pages, there is no difference between
1979 * shift and PAGE_SHIFT, so the bit-shift is safe.
1980 * for huge pages, since vm_pgoff is in units of small
1981 * pages, we need to shift off the always 0 bits to get
1984 BUG_ON(shift < PAGE_SHIFT);
1985 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1986 off += (addr - vma->vm_start) >> shift;
1987 return offset_il_node(pol, off);
1989 return interleave_nodes(pol);
1992 #ifdef CONFIG_HUGETLBFS
1994 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1995 * @vma: virtual memory area whose policy is sought
1996 * @addr: address in @vma for shared policy lookup and interleave policy
1997 * @gfp_flags: for requested zone
1998 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1999 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2001 * Returns a nid suitable for a huge page allocation and a pointer
2002 * to the struct mempolicy for conditional unref after allocation.
2003 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2004 * to the mempolicy's @nodemask for filtering the zonelist.
2006 * Must be protected by read_mems_allowed_begin()
2008 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2009 struct mempolicy **mpol, nodemask_t **nodemask)
2014 *mpol = get_vma_policy(vma, addr);
2016 mode = (*mpol)->mode;
2018 if (unlikely(mode == MPOL_INTERLEAVE)) {
2019 nid = interleave_nid(*mpol, vma, addr,
2020 huge_page_shift(hstate_vma(vma)));
2022 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2023 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2024 *nodemask = &(*mpol)->nodes;
2030 * init_nodemask_of_mempolicy
2032 * If the current task's mempolicy is "default" [NULL], return 'false'
2033 * to indicate default policy. Otherwise, extract the policy nodemask
2034 * for 'bind' or 'interleave' policy into the argument nodemask, or
2035 * initialize the argument nodemask to contain the single node for
2036 * 'preferred' or 'local' policy and return 'true' to indicate presence
2037 * of non-default mempolicy.
2039 * We don't bother with reference counting the mempolicy [mpol_get/put]
2040 * because the current task is examining it's own mempolicy and a task's
2041 * mempolicy is only ever changed by the task itself.
2043 * N.B., it is the caller's responsibility to free a returned nodemask.
2045 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2047 struct mempolicy *mempolicy;
2049 if (!(mask && current->mempolicy))
2053 mempolicy = current->mempolicy;
2054 switch (mempolicy->mode) {
2055 case MPOL_PREFERRED:
2056 case MPOL_PREFERRED_MANY:
2058 case MPOL_INTERLEAVE:
2059 *mask = mempolicy->nodes;
2063 init_nodemask_of_node(mask, numa_node_id());
2069 task_unlock(current);
2076 * mempolicy_in_oom_domain
2078 * If tsk's mempolicy is "bind", check for intersection between mask and
2079 * the policy nodemask. Otherwise, return true for all other policies
2080 * including "interleave", as a tsk with "interleave" policy may have
2081 * memory allocated from all nodes in system.
2083 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2085 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2086 const nodemask_t *mask)
2088 struct mempolicy *mempolicy;
2095 mempolicy = tsk->mempolicy;
2096 if (mempolicy && mempolicy->mode == MPOL_BIND)
2097 ret = nodes_intersects(mempolicy->nodes, *mask);
2103 /* Allocate a page in interleaved policy.
2104 Own path because it needs to do special accounting. */
2105 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2110 page = __alloc_pages(gfp, order, nid, NULL);
2111 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2112 if (!static_branch_likely(&vm_numa_stat_key))
2114 if (page && page_to_nid(page) == nid) {
2116 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2122 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2123 int nid, struct mempolicy *pol)
2126 gfp_t preferred_gfp;
2129 * This is a two pass approach. The first pass will only try the
2130 * preferred nodes but skip the direct reclaim and allow the
2131 * allocation to fail, while the second pass will try all the
2134 preferred_gfp = gfp | __GFP_NOWARN;
2135 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2136 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2138 page = __alloc_pages(gfp, order, nid, NULL);
2144 * alloc_pages_vma - Allocate a page for a VMA.
2146 * @order: Order of the GFP allocation.
2147 * @vma: Pointer to VMA or NULL if not available.
2148 * @addr: Virtual address of the allocation. Must be inside @vma.
2149 * @hugepage: For hugepages try only the preferred node if possible.
2151 * Allocate a page for a specific address in @vma, using the appropriate
2152 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2153 * of the mm_struct of the VMA to prevent it from going away. Should be
2154 * used for all allocations for pages that will be mapped into user space.
2156 * Return: The page on success or NULL if allocation fails.
2158 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2159 unsigned long addr, bool hugepage)
2161 struct mempolicy *pol;
2162 int node = numa_node_id();
2167 pol = get_vma_policy(vma, addr);
2169 if (pol->mode == MPOL_INTERLEAVE) {
2172 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2174 page = alloc_page_interleave(gfp, order, nid);
2178 if (pol->mode == MPOL_PREFERRED_MANY) {
2179 node = policy_node(gfp, pol, node);
2180 page = alloc_pages_preferred_many(gfp, order, node, pol);
2185 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2186 int hpage_node = node;
2189 * For hugepage allocation and non-interleave policy which
2190 * allows the current node (or other explicitly preferred
2191 * node) we only try to allocate from the current/preferred
2192 * node and don't fall back to other nodes, as the cost of
2193 * remote accesses would likely offset THP benefits.
2195 * If the policy is interleave or does not allow the current
2196 * node in its nodemask, we allocate the standard way.
2198 if (pol->mode == MPOL_PREFERRED)
2199 hpage_node = first_node(pol->nodes);
2201 nmask = policy_nodemask(gfp, pol);
2202 if (!nmask || node_isset(hpage_node, *nmask)) {
2205 * First, try to allocate THP only on local node, but
2206 * don't reclaim unnecessarily, just compact.
2208 page = __alloc_pages_node(hpage_node,
2209 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2212 * If hugepage allocations are configured to always
2213 * synchronous compact or the vma has been madvised
2214 * to prefer hugepage backing, retry allowing remote
2215 * memory with both reclaim and compact as well.
2217 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2218 page = __alloc_pages(gfp, order, hpage_node, nmask);
2224 nmask = policy_nodemask(gfp, pol);
2225 preferred_nid = policy_node(gfp, pol, node);
2226 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2231 EXPORT_SYMBOL(alloc_pages_vma);
2234 * alloc_pages - Allocate pages.
2236 * @order: Power of two of number of pages to allocate.
2238 * Allocate 1 << @order contiguous pages. The physical address of the
2239 * first page is naturally aligned (eg an order-3 allocation will be aligned
2240 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2241 * process is honoured when in process context.
2243 * Context: Can be called from any context, providing the appropriate GFP
2245 * Return: The page on success or NULL if allocation fails.
2247 struct page *alloc_pages(gfp_t gfp, unsigned order)
2249 struct mempolicy *pol = &default_policy;
2252 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2253 pol = get_task_policy(current);
2256 * No reference counting needed for current->mempolicy
2257 * nor system default_policy
2259 if (pol->mode == MPOL_INTERLEAVE)
2260 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2261 else if (pol->mode == MPOL_PREFERRED_MANY)
2262 page = alloc_pages_preferred_many(gfp, order,
2263 policy_node(gfp, pol, numa_node_id()), pol);
2265 page = __alloc_pages(gfp, order,
2266 policy_node(gfp, pol, numa_node_id()),
2267 policy_nodemask(gfp, pol));
2271 EXPORT_SYMBOL(alloc_pages);
2273 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2275 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2277 if (page && order > 1)
2278 prep_transhuge_page(page);
2279 return (struct folio *)page;
2281 EXPORT_SYMBOL(folio_alloc);
2283 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2284 struct mempolicy *pol, unsigned long nr_pages,
2285 struct page **page_array)
2288 unsigned long nr_pages_per_node;
2291 unsigned long nr_allocated;
2292 unsigned long total_allocated = 0;
2294 nodes = nodes_weight(pol->nodes);
2295 nr_pages_per_node = nr_pages / nodes;
2296 delta = nr_pages - nodes * nr_pages_per_node;
2298 for (i = 0; i < nodes; i++) {
2300 nr_allocated = __alloc_pages_bulk(gfp,
2301 interleave_nodes(pol), NULL,
2302 nr_pages_per_node + 1, NULL,
2306 nr_allocated = __alloc_pages_bulk(gfp,
2307 interleave_nodes(pol), NULL,
2308 nr_pages_per_node, NULL, page_array);
2311 page_array += nr_allocated;
2312 total_allocated += nr_allocated;
2315 return total_allocated;
2318 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2319 struct mempolicy *pol, unsigned long nr_pages,
2320 struct page **page_array)
2322 gfp_t preferred_gfp;
2323 unsigned long nr_allocated = 0;
2325 preferred_gfp = gfp | __GFP_NOWARN;
2326 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2328 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2329 nr_pages, NULL, page_array);
2331 if (nr_allocated < nr_pages)
2332 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2333 nr_pages - nr_allocated, NULL,
2334 page_array + nr_allocated);
2335 return nr_allocated;
2338 /* alloc pages bulk and mempolicy should be considered at the
2339 * same time in some situation such as vmalloc.
2341 * It can accelerate memory allocation especially interleaving
2344 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2345 unsigned long nr_pages, struct page **page_array)
2347 struct mempolicy *pol = &default_policy;
2349 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2350 pol = get_task_policy(current);
2352 if (pol->mode == MPOL_INTERLEAVE)
2353 return alloc_pages_bulk_array_interleave(gfp, pol,
2354 nr_pages, page_array);
2356 if (pol->mode == MPOL_PREFERRED_MANY)
2357 return alloc_pages_bulk_array_preferred_many(gfp,
2358 numa_node_id(), pol, nr_pages, page_array);
2360 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2361 policy_nodemask(gfp, pol), nr_pages, NULL,
2365 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2367 struct mempolicy *pol = mpol_dup(vma_policy(src));
2370 return PTR_ERR(pol);
2371 dst->vm_policy = pol;
2376 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2377 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2378 * with the mems_allowed returned by cpuset_mems_allowed(). This
2379 * keeps mempolicies cpuset relative after its cpuset moves. See
2380 * further kernel/cpuset.c update_nodemask().
2382 * current's mempolicy may be rebinded by the other task(the task that changes
2383 * cpuset's mems), so we needn't do rebind work for current task.
2386 /* Slow path of a mempolicy duplicate */
2387 struct mempolicy *__mpol_dup(struct mempolicy *old)
2389 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2392 return ERR_PTR(-ENOMEM);
2394 /* task's mempolicy is protected by alloc_lock */
2395 if (old == current->mempolicy) {
2398 task_unlock(current);
2402 if (current_cpuset_is_being_rebound()) {
2403 nodemask_t mems = cpuset_mems_allowed(current);
2404 mpol_rebind_policy(new, &mems);
2406 atomic_set(&new->refcnt, 1);
2410 /* Slow path of a mempolicy comparison */
2411 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2415 if (a->mode != b->mode)
2417 if (a->flags != b->flags)
2419 if (a->home_node != b->home_node)
2421 if (mpol_store_user_nodemask(a))
2422 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2427 case MPOL_INTERLEAVE:
2428 case MPOL_PREFERRED:
2429 case MPOL_PREFERRED_MANY:
2430 return !!nodes_equal(a->nodes, b->nodes);
2440 * Shared memory backing store policy support.
2442 * Remember policies even when nobody has shared memory mapped.
2443 * The policies are kept in Red-Black tree linked from the inode.
2444 * They are protected by the sp->lock rwlock, which should be held
2445 * for any accesses to the tree.
2449 * lookup first element intersecting start-end. Caller holds sp->lock for
2450 * reading or for writing
2452 static struct sp_node *
2453 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2455 struct rb_node *n = sp->root.rb_node;
2458 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2460 if (start >= p->end)
2462 else if (end <= p->start)
2470 struct sp_node *w = NULL;
2471 struct rb_node *prev = rb_prev(n);
2474 w = rb_entry(prev, struct sp_node, nd);
2475 if (w->end <= start)
2479 return rb_entry(n, struct sp_node, nd);
2483 * Insert a new shared policy into the list. Caller holds sp->lock for
2486 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2488 struct rb_node **p = &sp->root.rb_node;
2489 struct rb_node *parent = NULL;
2494 nd = rb_entry(parent, struct sp_node, nd);
2495 if (new->start < nd->start)
2497 else if (new->end > nd->end)
2498 p = &(*p)->rb_right;
2502 rb_link_node(&new->nd, parent, p);
2503 rb_insert_color(&new->nd, &sp->root);
2504 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2505 new->policy ? new->policy->mode : 0);
2508 /* Find shared policy intersecting idx */
2510 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2512 struct mempolicy *pol = NULL;
2515 if (!sp->root.rb_node)
2517 read_lock(&sp->lock);
2518 sn = sp_lookup(sp, idx, idx+1);
2520 mpol_get(sn->policy);
2523 read_unlock(&sp->lock);
2527 static void sp_free(struct sp_node *n)
2529 mpol_put(n->policy);
2530 kmem_cache_free(sn_cache, n);
2534 * mpol_misplaced - check whether current page node is valid in policy
2536 * @page: page to be checked
2537 * @vma: vm area where page mapped
2538 * @addr: virtual address where page mapped
2540 * Lookup current policy node id for vma,addr and "compare to" page's
2541 * node id. Policy determination "mimics" alloc_page_vma().
2542 * Called from fault path where we know the vma and faulting address.
2544 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2545 * policy, or a suitable node ID to allocate a replacement page from.
2547 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2549 struct mempolicy *pol;
2551 int curnid = page_to_nid(page);
2552 unsigned long pgoff;
2553 int thiscpu = raw_smp_processor_id();
2554 int thisnid = cpu_to_node(thiscpu);
2555 int polnid = NUMA_NO_NODE;
2556 int ret = NUMA_NO_NODE;
2558 pol = get_vma_policy(vma, addr);
2559 if (!(pol->flags & MPOL_F_MOF))
2562 switch (pol->mode) {
2563 case MPOL_INTERLEAVE:
2564 pgoff = vma->vm_pgoff;
2565 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2566 polnid = offset_il_node(pol, pgoff);
2569 case MPOL_PREFERRED:
2570 if (node_isset(curnid, pol->nodes))
2572 polnid = first_node(pol->nodes);
2576 polnid = numa_node_id();
2580 /* Optimize placement among multiple nodes via NUMA balancing */
2581 if (pol->flags & MPOL_F_MORON) {
2582 if (node_isset(thisnid, pol->nodes))
2588 case MPOL_PREFERRED_MANY:
2590 * use current page if in policy nodemask,
2591 * else select nearest allowed node, if any.
2592 * If no allowed nodes, use current [!misplaced].
2594 if (node_isset(curnid, pol->nodes))
2596 z = first_zones_zonelist(
2597 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2598 gfp_zone(GFP_HIGHUSER),
2600 polnid = zone_to_nid(z->zone);
2607 /* Migrate the page towards the node whose CPU is referencing it */
2608 if (pol->flags & MPOL_F_MORON) {
2611 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2615 if (curnid != polnid)
2624 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2625 * dropped after task->mempolicy is set to NULL so that any allocation done as
2626 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2629 void mpol_put_task_policy(struct task_struct *task)
2631 struct mempolicy *pol;
2634 pol = task->mempolicy;
2635 task->mempolicy = NULL;
2640 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2642 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2643 rb_erase(&n->nd, &sp->root);
2647 static void sp_node_init(struct sp_node *node, unsigned long start,
2648 unsigned long end, struct mempolicy *pol)
2650 node->start = start;
2655 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2656 struct mempolicy *pol)
2659 struct mempolicy *newpol;
2661 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2665 newpol = mpol_dup(pol);
2666 if (IS_ERR(newpol)) {
2667 kmem_cache_free(sn_cache, n);
2670 newpol->flags |= MPOL_F_SHARED;
2671 sp_node_init(n, start, end, newpol);
2676 /* Replace a policy range. */
2677 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2678 unsigned long end, struct sp_node *new)
2681 struct sp_node *n_new = NULL;
2682 struct mempolicy *mpol_new = NULL;
2686 write_lock(&sp->lock);
2687 n = sp_lookup(sp, start, end);
2688 /* Take care of old policies in the same range. */
2689 while (n && n->start < end) {
2690 struct rb_node *next = rb_next(&n->nd);
2691 if (n->start >= start) {
2697 /* Old policy spanning whole new range. */
2702 *mpol_new = *n->policy;
2703 atomic_set(&mpol_new->refcnt, 1);
2704 sp_node_init(n_new, end, n->end, mpol_new);
2706 sp_insert(sp, n_new);
2715 n = rb_entry(next, struct sp_node, nd);
2719 write_unlock(&sp->lock);
2726 kmem_cache_free(sn_cache, n_new);
2731 write_unlock(&sp->lock);
2733 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2736 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2739 atomic_set(&mpol_new->refcnt, 1);
2744 * mpol_shared_policy_init - initialize shared policy for inode
2745 * @sp: pointer to inode shared policy
2746 * @mpol: struct mempolicy to install
2748 * Install non-NULL @mpol in inode's shared policy rb-tree.
2749 * On entry, the current task has a reference on a non-NULL @mpol.
2750 * This must be released on exit.
2751 * This is called at get_inode() calls and we can use GFP_KERNEL.
2753 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2757 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2758 rwlock_init(&sp->lock);
2761 struct vm_area_struct pvma;
2762 struct mempolicy *new;
2763 NODEMASK_SCRATCH(scratch);
2767 /* contextualize the tmpfs mount point mempolicy */
2768 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2770 goto free_scratch; /* no valid nodemask intersection */
2773 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2774 task_unlock(current);
2778 /* Create pseudo-vma that contains just the policy */
2779 vma_init(&pvma, NULL);
2780 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2781 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2784 mpol_put(new); /* drop initial ref */
2786 NODEMASK_SCRATCH_FREE(scratch);
2788 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2792 int mpol_set_shared_policy(struct shared_policy *info,
2793 struct vm_area_struct *vma, struct mempolicy *npol)
2796 struct sp_node *new = NULL;
2797 unsigned long sz = vma_pages(vma);
2799 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2801 sz, npol ? npol->mode : -1,
2802 npol ? npol->flags : -1,
2803 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2806 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2810 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2816 /* Free a backing policy store on inode delete. */
2817 void mpol_free_shared_policy(struct shared_policy *p)
2820 struct rb_node *next;
2822 if (!p->root.rb_node)
2824 write_lock(&p->lock);
2825 next = rb_first(&p->root);
2827 n = rb_entry(next, struct sp_node, nd);
2828 next = rb_next(&n->nd);
2831 write_unlock(&p->lock);
2834 #ifdef CONFIG_NUMA_BALANCING
2835 static int __initdata numabalancing_override;
2837 static void __init check_numabalancing_enable(void)
2839 bool numabalancing_default = false;
2841 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2842 numabalancing_default = true;
2844 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2845 if (numabalancing_override)
2846 set_numabalancing_state(numabalancing_override == 1);
2848 if (num_online_nodes() > 1 && !numabalancing_override) {
2849 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2850 numabalancing_default ? "Enabling" : "Disabling");
2851 set_numabalancing_state(numabalancing_default);
2855 static int __init setup_numabalancing(char *str)
2861 if (!strcmp(str, "enable")) {
2862 numabalancing_override = 1;
2864 } else if (!strcmp(str, "disable")) {
2865 numabalancing_override = -1;
2870 pr_warn("Unable to parse numa_balancing=\n");
2874 __setup("numa_balancing=", setup_numabalancing);
2876 static inline void __init check_numabalancing_enable(void)
2879 #endif /* CONFIG_NUMA_BALANCING */
2881 /* assumes fs == KERNEL_DS */
2882 void __init numa_policy_init(void)
2884 nodemask_t interleave_nodes;
2885 unsigned long largest = 0;
2886 int nid, prefer = 0;
2888 policy_cache = kmem_cache_create("numa_policy",
2889 sizeof(struct mempolicy),
2890 0, SLAB_PANIC, NULL);
2892 sn_cache = kmem_cache_create("shared_policy_node",
2893 sizeof(struct sp_node),
2894 0, SLAB_PANIC, NULL);
2896 for_each_node(nid) {
2897 preferred_node_policy[nid] = (struct mempolicy) {
2898 .refcnt = ATOMIC_INIT(1),
2899 .mode = MPOL_PREFERRED,
2900 .flags = MPOL_F_MOF | MPOL_F_MORON,
2901 .nodes = nodemask_of_node(nid),
2906 * Set interleaving policy for system init. Interleaving is only
2907 * enabled across suitably sized nodes (default is >= 16MB), or
2908 * fall back to the largest node if they're all smaller.
2910 nodes_clear(interleave_nodes);
2911 for_each_node_state(nid, N_MEMORY) {
2912 unsigned long total_pages = node_present_pages(nid);
2914 /* Preserve the largest node */
2915 if (largest < total_pages) {
2916 largest = total_pages;
2920 /* Interleave this node? */
2921 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2922 node_set(nid, interleave_nodes);
2925 /* All too small, use the largest */
2926 if (unlikely(nodes_empty(interleave_nodes)))
2927 node_set(prefer, interleave_nodes);
2929 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2930 pr_err("%s: interleaving failed\n", __func__);
2932 check_numabalancing_enable();
2935 /* Reset policy of current process to default */
2936 void numa_default_policy(void)
2938 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2942 * Parse and format mempolicy from/to strings
2945 static const char * const policy_modes[] =
2947 [MPOL_DEFAULT] = "default",
2948 [MPOL_PREFERRED] = "prefer",
2949 [MPOL_BIND] = "bind",
2950 [MPOL_INTERLEAVE] = "interleave",
2951 [MPOL_LOCAL] = "local",
2952 [MPOL_PREFERRED_MANY] = "prefer (many)",
2958 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2959 * @str: string containing mempolicy to parse
2960 * @mpol: pointer to struct mempolicy pointer, returned on success.
2963 * <mode>[=<flags>][:<nodelist>]
2965 * Return: %0 on success, else %1
2967 int mpol_parse_str(char *str, struct mempolicy **mpol)
2969 struct mempolicy *new = NULL;
2970 unsigned short mode_flags;
2972 char *nodelist = strchr(str, ':');
2973 char *flags = strchr(str, '=');
2977 *flags++ = '\0'; /* terminate mode string */
2980 /* NUL-terminate mode or flags string */
2982 if (nodelist_parse(nodelist, nodes))
2984 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2989 mode = match_string(policy_modes, MPOL_MAX, str);
2994 case MPOL_PREFERRED:
2996 * Insist on a nodelist of one node only, although later
2997 * we use first_node(nodes) to grab a single node, so here
2998 * nodelist (or nodes) cannot be empty.
3001 char *rest = nodelist;
3002 while (isdigit(*rest))
3006 if (nodes_empty(nodes))
3010 case MPOL_INTERLEAVE:
3012 * Default to online nodes with memory if no nodelist
3015 nodes = node_states[N_MEMORY];
3019 * Don't allow a nodelist; mpol_new() checks flags
3026 * Insist on a empty nodelist
3031 case MPOL_PREFERRED_MANY:
3034 * Insist on a nodelist
3043 * Currently, we only support two mutually exclusive
3046 if (!strcmp(flags, "static"))
3047 mode_flags |= MPOL_F_STATIC_NODES;
3048 else if (!strcmp(flags, "relative"))
3049 mode_flags |= MPOL_F_RELATIVE_NODES;
3054 new = mpol_new(mode, mode_flags, &nodes);
3059 * Save nodes for mpol_to_str() to show the tmpfs mount options
3060 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3062 if (mode != MPOL_PREFERRED) {
3064 } else if (nodelist) {
3065 nodes_clear(new->nodes);
3066 node_set(first_node(nodes), new->nodes);
3068 new->mode = MPOL_LOCAL;
3072 * Save nodes for contextualization: this will be used to "clone"
3073 * the mempolicy in a specific context [cpuset] at a later time.
3075 new->w.user_nodemask = nodes;
3080 /* Restore string for error message */
3089 #endif /* CONFIG_TMPFS */
3092 * mpol_to_str - format a mempolicy structure for printing
3093 * @buffer: to contain formatted mempolicy string
3094 * @maxlen: length of @buffer
3095 * @pol: pointer to mempolicy to be formatted
3097 * Convert @pol into a string. If @buffer is too short, truncate the string.
3098 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3099 * longest flag, "relative", and to display at least a few node ids.
3101 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3104 nodemask_t nodes = NODE_MASK_NONE;
3105 unsigned short mode = MPOL_DEFAULT;
3106 unsigned short flags = 0;
3108 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3117 case MPOL_PREFERRED:
3118 case MPOL_PREFERRED_MANY:
3120 case MPOL_INTERLEAVE:
3125 snprintf(p, maxlen, "unknown");
3129 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3131 if (flags & MPOL_MODE_FLAGS) {
3132 p += snprintf(p, buffer + maxlen - p, "=");
3135 * Currently, the only defined flags are mutually exclusive
3137 if (flags & MPOL_F_STATIC_NODES)
3138 p += snprintf(p, buffer + maxlen - p, "static");
3139 else if (flags & MPOL_F_RELATIVE_NODES)
3140 p += snprintf(p, buffer + maxlen - p, "relative");
3143 if (!nodes_empty(nodes))
3144 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3145 nodemask_pr_args(&nodes));