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;
906 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
908 ret = page_to_nid(p);
914 /* Retrieve NUMA policy */
915 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
916 unsigned long addr, unsigned long flags)
919 struct mm_struct *mm = current->mm;
920 struct vm_area_struct *vma = NULL;
921 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
924 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
927 if (flags & MPOL_F_MEMS_ALLOWED) {
928 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
930 *policy = 0; /* just so it's initialized */
932 *nmask = cpuset_current_mems_allowed;
933 task_unlock(current);
937 if (flags & MPOL_F_ADDR) {
939 * Do NOT fall back to task policy if the
940 * vma/shared policy at addr is NULL. We
941 * want to return MPOL_DEFAULT in this case.
944 vma = vma_lookup(mm, addr);
946 mmap_read_unlock(mm);
949 if (vma->vm_ops && vma->vm_ops->get_policy)
950 pol = vma->vm_ops->get_policy(vma, addr);
952 pol = vma->vm_policy;
957 pol = &default_policy; /* indicates default behavior */
959 if (flags & MPOL_F_NODE) {
960 if (flags & MPOL_F_ADDR) {
962 * Take a refcount on the mpol, because we are about to
963 * drop the mmap_lock, after which only "pol" remains
964 * valid, "vma" is stale.
969 mmap_read_unlock(mm);
970 err = lookup_node(mm, addr);
974 } else if (pol == current->mempolicy &&
975 pol->mode == MPOL_INTERLEAVE) {
976 *policy = next_node_in(current->il_prev, pol->nodes);
982 *policy = pol == &default_policy ? MPOL_DEFAULT :
985 * Internal mempolicy flags must be masked off before exposing
986 * the policy to userspace.
988 *policy |= (pol->flags & MPOL_MODE_FLAGS);
993 if (mpol_store_user_nodemask(pol)) {
994 *nmask = pol->w.user_nodemask;
997 get_policy_nodemask(pol, nmask);
998 task_unlock(current);
1005 mmap_read_unlock(mm);
1007 mpol_put(pol_refcount);
1011 #ifdef CONFIG_MIGRATION
1013 * page migration, thp tail pages can be passed.
1015 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1016 unsigned long flags)
1018 struct page *head = compound_head(page);
1020 * Avoid migrating a page that is shared with others.
1022 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1023 if (!isolate_lru_page(head)) {
1024 list_add_tail(&head->lru, pagelist);
1025 mod_node_page_state(page_pgdat(head),
1026 NR_ISOLATED_ANON + page_is_file_lru(head),
1027 thp_nr_pages(head));
1028 } else if (flags & MPOL_MF_STRICT) {
1030 * Non-movable page may reach here. And, there may be
1031 * temporary off LRU pages or non-LRU movable pages.
1032 * Treat them as unmovable pages since they can't be
1033 * isolated, so they can't be moved at the moment. It
1034 * should return -EIO for this case too.
1044 * Migrate pages from one node to a target node.
1045 * Returns error or the number of pages not migrated.
1047 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1051 LIST_HEAD(pagelist);
1053 struct migration_target_control mtc = {
1055 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1059 node_set(source, nmask);
1062 * This does not "check" the range but isolates all pages that
1063 * need migration. Between passing in the full user address
1064 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1066 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1067 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1068 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1070 if (!list_empty(&pagelist)) {
1071 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1072 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1074 putback_movable_pages(&pagelist);
1081 * Move pages between the two nodesets so as to preserve the physical
1082 * layout as much as possible.
1084 * Returns the number of page that could not be moved.
1086 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1087 const nodemask_t *to, int flags)
1093 lru_cache_disable();
1098 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1099 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1100 * bit in 'tmp', and return that <source, dest> pair for migration.
1101 * The pair of nodemasks 'to' and 'from' define the map.
1103 * If no pair of bits is found that way, fallback to picking some
1104 * pair of 'source' and 'dest' bits that are not the same. If the
1105 * 'source' and 'dest' bits are the same, this represents a node
1106 * that will be migrating to itself, so no pages need move.
1108 * If no bits are left in 'tmp', or if all remaining bits left
1109 * in 'tmp' correspond to the same bit in 'to', return false
1110 * (nothing left to migrate).
1112 * This lets us pick a pair of nodes to migrate between, such that
1113 * if possible the dest node is not already occupied by some other
1114 * source node, minimizing the risk of overloading the memory on a
1115 * node that would happen if we migrated incoming memory to a node
1116 * before migrating outgoing memory source that same node.
1118 * A single scan of tmp is sufficient. As we go, we remember the
1119 * most recent <s, d> pair that moved (s != d). If we find a pair
1120 * that not only moved, but what's better, moved to an empty slot
1121 * (d is not set in tmp), then we break out then, with that pair.
1122 * Otherwise when we finish scanning from_tmp, we at least have the
1123 * most recent <s, d> pair that moved. If we get all the way through
1124 * the scan of tmp without finding any node that moved, much less
1125 * moved to an empty node, then there is nothing left worth migrating.
1129 while (!nodes_empty(tmp)) {
1131 int source = NUMA_NO_NODE;
1134 for_each_node_mask(s, tmp) {
1137 * do_migrate_pages() tries to maintain the relative
1138 * node relationship of the pages established between
1139 * threads and memory areas.
1141 * However if the number of source nodes is not equal to
1142 * the number of destination nodes we can not preserve
1143 * this node relative relationship. In that case, skip
1144 * copying memory from a node that is in the destination
1147 * Example: [2,3,4] -> [3,4,5] moves everything.
1148 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1151 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1152 (node_isset(s, *to)))
1155 d = node_remap(s, *from, *to);
1159 source = s; /* Node moved. Memorize */
1162 /* dest not in remaining from nodes? */
1163 if (!node_isset(dest, tmp))
1166 if (source == NUMA_NO_NODE)
1169 node_clear(source, tmp);
1170 err = migrate_to_node(mm, source, dest, flags);
1176 mmap_read_unlock(mm);
1186 * Allocate a new page for page migration based on vma policy.
1187 * Start by assuming the page is mapped by the same vma as contains @start.
1188 * Search forward from there, if not. N.B., this assumes that the
1189 * list of pages handed to migrate_pages()--which is how we get here--
1190 * is in virtual address order.
1192 static struct page *new_page(struct page *page, unsigned long start)
1194 struct vm_area_struct *vma;
1195 unsigned long address;
1197 vma = find_vma(current->mm, start);
1199 address = page_address_in_vma(page, vma);
1200 if (address != -EFAULT)
1205 if (PageHuge(page)) {
1206 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1208 } else if (PageTransHuge(page)) {
1211 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1215 prep_transhuge_page(thp);
1219 * if !vma, alloc_page_vma() will use task or system default policy
1221 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1226 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1227 unsigned long flags)
1232 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1233 const nodemask_t *to, int flags)
1238 static struct page *new_page(struct page *page, unsigned long start)
1244 static long do_mbind(unsigned long start, unsigned long len,
1245 unsigned short mode, unsigned short mode_flags,
1246 nodemask_t *nmask, unsigned long flags)
1248 struct mm_struct *mm = current->mm;
1249 struct mempolicy *new;
1253 LIST_HEAD(pagelist);
1255 if (flags & ~(unsigned long)MPOL_MF_VALID)
1257 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1260 if (start & ~PAGE_MASK)
1263 if (mode == MPOL_DEFAULT)
1264 flags &= ~MPOL_MF_STRICT;
1266 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1274 new = mpol_new(mode, mode_flags, nmask);
1276 return PTR_ERR(new);
1278 if (flags & MPOL_MF_LAZY)
1279 new->flags |= MPOL_F_MOF;
1282 * If we are using the default policy then operation
1283 * on discontinuous address spaces is okay after all
1286 flags |= MPOL_MF_DISCONTIG_OK;
1288 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1289 start, start + len, mode, mode_flags,
1290 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1292 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1294 lru_cache_disable();
1297 NODEMASK_SCRATCH(scratch);
1299 mmap_write_lock(mm);
1300 err = mpol_set_nodemask(new, nmask, scratch);
1302 mmap_write_unlock(mm);
1305 NODEMASK_SCRATCH_FREE(scratch);
1310 ret = queue_pages_range(mm, start, end, nmask,
1311 flags | MPOL_MF_INVERT, &pagelist);
1318 err = mbind_range(mm, start, end, new);
1323 if (!list_empty(&pagelist)) {
1324 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1325 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1326 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1328 putback_movable_pages(&pagelist);
1331 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1335 if (!list_empty(&pagelist))
1336 putback_movable_pages(&pagelist);
1339 mmap_write_unlock(mm);
1342 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1348 * User space interface with variable sized bitmaps for nodelists.
1350 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1351 unsigned long maxnode)
1353 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1356 if (in_compat_syscall())
1357 ret = compat_get_bitmap(mask,
1358 (const compat_ulong_t __user *)nmask,
1361 ret = copy_from_user(mask, nmask,
1362 nlongs * sizeof(unsigned long));
1367 if (maxnode % BITS_PER_LONG)
1368 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1373 /* Copy a node mask from user space. */
1374 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1375 unsigned long maxnode)
1378 nodes_clear(*nodes);
1379 if (maxnode == 0 || !nmask)
1381 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1385 * When the user specified more nodes than supported just check
1386 * if the non supported part is all zero, one word at a time,
1387 * starting at the end.
1389 while (maxnode > MAX_NUMNODES) {
1390 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1393 if (get_bitmap(&t, &nmask[maxnode / BITS_PER_LONG], bits))
1396 if (maxnode - bits >= MAX_NUMNODES) {
1399 maxnode = MAX_NUMNODES;
1400 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1406 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1409 /* Copy a kernel node mask to user space */
1410 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1413 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1414 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1415 bool compat = in_compat_syscall();
1418 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1420 if (copy > nbytes) {
1421 if (copy > PAGE_SIZE)
1423 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1426 maxnode = nr_node_ids;
1430 return compat_put_bitmap((compat_ulong_t __user *)mask,
1431 nodes_addr(*nodes), maxnode);
1433 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1436 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1437 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1439 *flags = *mode & MPOL_MODE_FLAGS;
1440 *mode &= ~MPOL_MODE_FLAGS;
1442 if ((unsigned int)(*mode) >= MPOL_MAX)
1444 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1446 if (*flags & MPOL_F_NUMA_BALANCING) {
1447 if (*mode != MPOL_BIND)
1449 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1454 static long kernel_mbind(unsigned long start, unsigned long len,
1455 unsigned long mode, const unsigned long __user *nmask,
1456 unsigned long maxnode, unsigned int flags)
1458 unsigned short mode_flags;
1463 start = untagged_addr(start);
1464 err = sanitize_mpol_flags(&lmode, &mode_flags);
1468 err = get_nodes(&nodes, nmask, maxnode);
1472 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1475 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1476 unsigned long, home_node, unsigned long, flags)
1478 struct mm_struct *mm = current->mm;
1479 struct vm_area_struct *vma;
1480 struct mempolicy *new;
1481 unsigned long vmstart;
1482 unsigned long vmend;
1486 start = untagged_addr(start);
1487 if (start & ~PAGE_MASK)
1490 * flags is used for future extension if any.
1496 * Check home_node is online to avoid accessing uninitialized
1499 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1502 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1509 mmap_write_lock(mm);
1510 vma = find_vma(mm, start);
1511 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
1513 vmstart = max(start, vma->vm_start);
1514 vmend = min(end, vma->vm_end);
1515 new = mpol_dup(vma_policy(vma));
1521 * Only update home node if there is an existing vma policy
1527 * If any vma in the range got policy other than MPOL_BIND
1528 * or MPOL_PREFERRED_MANY we return error. We don't reset
1529 * the home node for vmas we already updated before.
1531 if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1536 new->home_node = home_node;
1537 err = mbind_range(mm, vmstart, vmend, new);
1542 mmap_write_unlock(mm);
1546 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1547 unsigned long, mode, const unsigned long __user *, nmask,
1548 unsigned long, maxnode, unsigned int, flags)
1550 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1553 /* Set the process memory policy */
1554 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1555 unsigned long maxnode)
1557 unsigned short mode_flags;
1562 err = sanitize_mpol_flags(&lmode, &mode_flags);
1566 err = get_nodes(&nodes, nmask, maxnode);
1570 return do_set_mempolicy(lmode, mode_flags, &nodes);
1573 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1574 unsigned long, maxnode)
1576 return kernel_set_mempolicy(mode, nmask, maxnode);
1579 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1580 const unsigned long __user *old_nodes,
1581 const unsigned long __user *new_nodes)
1583 struct mm_struct *mm = NULL;
1584 struct task_struct *task;
1585 nodemask_t task_nodes;
1589 NODEMASK_SCRATCH(scratch);
1594 old = &scratch->mask1;
1595 new = &scratch->mask2;
1597 err = get_nodes(old, old_nodes, maxnode);
1601 err = get_nodes(new, new_nodes, maxnode);
1605 /* Find the mm_struct */
1607 task = pid ? find_task_by_vpid(pid) : current;
1613 get_task_struct(task);
1618 * Check if this process has the right to modify the specified process.
1619 * Use the regular "ptrace_may_access()" checks.
1621 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1628 task_nodes = cpuset_mems_allowed(task);
1629 /* Is the user allowed to access the target nodes? */
1630 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1635 task_nodes = cpuset_mems_allowed(current);
1636 nodes_and(*new, *new, task_nodes);
1637 if (nodes_empty(*new))
1640 err = security_task_movememory(task);
1644 mm = get_task_mm(task);
1645 put_task_struct(task);
1652 err = do_migrate_pages(mm, old, new,
1653 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1657 NODEMASK_SCRATCH_FREE(scratch);
1662 put_task_struct(task);
1667 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1668 const unsigned long __user *, old_nodes,
1669 const unsigned long __user *, new_nodes)
1671 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1675 /* Retrieve NUMA policy */
1676 static int kernel_get_mempolicy(int __user *policy,
1677 unsigned long __user *nmask,
1678 unsigned long maxnode,
1680 unsigned long flags)
1686 if (nmask != NULL && maxnode < nr_node_ids)
1689 addr = untagged_addr(addr);
1691 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1696 if (policy && put_user(pval, policy))
1700 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1705 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1706 unsigned long __user *, nmask, unsigned long, maxnode,
1707 unsigned long, addr, unsigned long, flags)
1709 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1712 bool vma_migratable(struct vm_area_struct *vma)
1714 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1718 * DAX device mappings require predictable access latency, so avoid
1719 * incurring periodic faults.
1721 if (vma_is_dax(vma))
1724 if (is_vm_hugetlb_page(vma) &&
1725 !hugepage_migration_supported(hstate_vma(vma)))
1729 * Migration allocates pages in the highest zone. If we cannot
1730 * do so then migration (at least from node to node) is not
1734 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1740 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1743 struct mempolicy *pol = NULL;
1746 if (vma->vm_ops && vma->vm_ops->get_policy) {
1747 pol = vma->vm_ops->get_policy(vma, addr);
1748 } else if (vma->vm_policy) {
1749 pol = vma->vm_policy;
1752 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1753 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1754 * count on these policies which will be dropped by
1755 * mpol_cond_put() later
1757 if (mpol_needs_cond_ref(pol))
1766 * get_vma_policy(@vma, @addr)
1767 * @vma: virtual memory area whose policy is sought
1768 * @addr: address in @vma for shared policy lookup
1770 * Returns effective policy for a VMA at specified address.
1771 * Falls back to current->mempolicy or system default policy, as necessary.
1772 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1773 * count--added by the get_policy() vm_op, as appropriate--to protect against
1774 * freeing by another task. It is the caller's responsibility to free the
1775 * extra reference for shared policies.
1777 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1780 struct mempolicy *pol = __get_vma_policy(vma, addr);
1783 pol = get_task_policy(current);
1788 bool vma_policy_mof(struct vm_area_struct *vma)
1790 struct mempolicy *pol;
1792 if (vma->vm_ops && vma->vm_ops->get_policy) {
1795 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1796 if (pol && (pol->flags & MPOL_F_MOF))
1803 pol = vma->vm_policy;
1805 pol = get_task_policy(current);
1807 return pol->flags & MPOL_F_MOF;
1810 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1812 enum zone_type dynamic_policy_zone = policy_zone;
1814 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1817 * if policy->nodes has movable memory only,
1818 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1820 * policy->nodes is intersect with node_states[N_MEMORY].
1821 * so if the following test fails, it implies
1822 * policy->nodes has movable memory only.
1824 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1825 dynamic_policy_zone = ZONE_MOVABLE;
1827 return zone >= dynamic_policy_zone;
1831 * Return a nodemask representing a mempolicy for filtering nodes for
1834 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1836 int mode = policy->mode;
1838 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1839 if (unlikely(mode == MPOL_BIND) &&
1840 apply_policy_zone(policy, gfp_zone(gfp)) &&
1841 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1842 return &policy->nodes;
1844 if (mode == MPOL_PREFERRED_MANY)
1845 return &policy->nodes;
1851 * Return the preferred node id for 'prefer' mempolicy, and return
1852 * the given id for all other policies.
1854 * policy_node() is always coupled with policy_nodemask(), which
1855 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1857 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1859 if (policy->mode == MPOL_PREFERRED) {
1860 nd = first_node(policy->nodes);
1863 * __GFP_THISNODE shouldn't even be used with the bind policy
1864 * because we might easily break the expectation to stay on the
1865 * requested node and not break the policy.
1867 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1870 if ((policy->mode == MPOL_BIND ||
1871 policy->mode == MPOL_PREFERRED_MANY) &&
1872 policy->home_node != NUMA_NO_NODE)
1873 return policy->home_node;
1878 /* Do dynamic interleaving for a process */
1879 static unsigned interleave_nodes(struct mempolicy *policy)
1882 struct task_struct *me = current;
1884 next = next_node_in(me->il_prev, policy->nodes);
1885 if (next < MAX_NUMNODES)
1891 * Depending on the memory policy provide a node from which to allocate the
1894 unsigned int mempolicy_slab_node(void)
1896 struct mempolicy *policy;
1897 int node = numa_mem_id();
1902 policy = current->mempolicy;
1906 switch (policy->mode) {
1907 case MPOL_PREFERRED:
1908 return first_node(policy->nodes);
1910 case MPOL_INTERLEAVE:
1911 return interleave_nodes(policy);
1914 case MPOL_PREFERRED_MANY:
1919 * Follow bind policy behavior and start allocation at the
1922 struct zonelist *zonelist;
1923 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1924 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1925 z = first_zones_zonelist(zonelist, highest_zoneidx,
1927 return z->zone ? zone_to_nid(z->zone) : node;
1938 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1939 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1940 * number of present nodes.
1942 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1944 nodemask_t nodemask = pol->nodes;
1945 unsigned int target, nnodes;
1949 * The barrier will stabilize the nodemask in a register or on
1950 * the stack so that it will stop changing under the code.
1952 * Between first_node() and next_node(), pol->nodes could be changed
1953 * by other threads. So we put pol->nodes in a local stack.
1957 nnodes = nodes_weight(nodemask);
1959 return numa_node_id();
1960 target = (unsigned int)n % nnodes;
1961 nid = first_node(nodemask);
1962 for (i = 0; i < target; i++)
1963 nid = next_node(nid, nodemask);
1967 /* Determine a node number for interleave */
1968 static inline unsigned interleave_nid(struct mempolicy *pol,
1969 struct vm_area_struct *vma, unsigned long addr, int shift)
1975 * for small pages, there is no difference between
1976 * shift and PAGE_SHIFT, so the bit-shift is safe.
1977 * for huge pages, since vm_pgoff is in units of small
1978 * pages, we need to shift off the always 0 bits to get
1981 BUG_ON(shift < PAGE_SHIFT);
1982 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1983 off += (addr - vma->vm_start) >> shift;
1984 return offset_il_node(pol, off);
1986 return interleave_nodes(pol);
1989 #ifdef CONFIG_HUGETLBFS
1991 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1992 * @vma: virtual memory area whose policy is sought
1993 * @addr: address in @vma for shared policy lookup and interleave policy
1994 * @gfp_flags: for requested zone
1995 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1996 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1998 * Returns a nid suitable for a huge page allocation and a pointer
1999 * to the struct mempolicy for conditional unref after allocation.
2000 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2001 * to the mempolicy's @nodemask for filtering the zonelist.
2003 * Must be protected by read_mems_allowed_begin()
2005 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2006 struct mempolicy **mpol, nodemask_t **nodemask)
2011 *mpol = get_vma_policy(vma, addr);
2013 mode = (*mpol)->mode;
2015 if (unlikely(mode == MPOL_INTERLEAVE)) {
2016 nid = interleave_nid(*mpol, vma, addr,
2017 huge_page_shift(hstate_vma(vma)));
2019 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2020 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2021 *nodemask = &(*mpol)->nodes;
2027 * init_nodemask_of_mempolicy
2029 * If the current task's mempolicy is "default" [NULL], return 'false'
2030 * to indicate default policy. Otherwise, extract the policy nodemask
2031 * for 'bind' or 'interleave' policy into the argument nodemask, or
2032 * initialize the argument nodemask to contain the single node for
2033 * 'preferred' or 'local' policy and return 'true' to indicate presence
2034 * of non-default mempolicy.
2036 * We don't bother with reference counting the mempolicy [mpol_get/put]
2037 * because the current task is examining it's own mempolicy and a task's
2038 * mempolicy is only ever changed by the task itself.
2040 * N.B., it is the caller's responsibility to free a returned nodemask.
2042 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2044 struct mempolicy *mempolicy;
2046 if (!(mask && current->mempolicy))
2050 mempolicy = current->mempolicy;
2051 switch (mempolicy->mode) {
2052 case MPOL_PREFERRED:
2053 case MPOL_PREFERRED_MANY:
2055 case MPOL_INTERLEAVE:
2056 *mask = mempolicy->nodes;
2060 init_nodemask_of_node(mask, numa_node_id());
2066 task_unlock(current);
2073 * mempolicy_in_oom_domain
2075 * If tsk's mempolicy is "bind", check for intersection between mask and
2076 * the policy nodemask. Otherwise, return true for all other policies
2077 * including "interleave", as a tsk with "interleave" policy may have
2078 * memory allocated from all nodes in system.
2080 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2082 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2083 const nodemask_t *mask)
2085 struct mempolicy *mempolicy;
2092 mempolicy = tsk->mempolicy;
2093 if (mempolicy && mempolicy->mode == MPOL_BIND)
2094 ret = nodes_intersects(mempolicy->nodes, *mask);
2100 /* Allocate a page in interleaved policy.
2101 Own path because it needs to do special accounting. */
2102 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2107 page = __alloc_pages(gfp, order, nid, NULL);
2108 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2109 if (!static_branch_likely(&vm_numa_stat_key))
2111 if (page && page_to_nid(page) == nid) {
2113 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2119 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2120 int nid, struct mempolicy *pol)
2123 gfp_t preferred_gfp;
2126 * This is a two pass approach. The first pass will only try the
2127 * preferred nodes but skip the direct reclaim and allow the
2128 * allocation to fail, while the second pass will try all the
2131 preferred_gfp = gfp | __GFP_NOWARN;
2132 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2133 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2135 page = __alloc_pages(gfp, order, nid, NULL);
2141 * alloc_pages_vma - Allocate a page for a VMA.
2143 * @order: Order of the GFP allocation.
2144 * @vma: Pointer to VMA or NULL if not available.
2145 * @addr: Virtual address of the allocation. Must be inside @vma.
2146 * @hugepage: For hugepages try only the preferred node if possible.
2148 * Allocate a page for a specific address in @vma, using the appropriate
2149 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2150 * of the mm_struct of the VMA to prevent it from going away. Should be
2151 * used for all allocations for pages that will be mapped into user space.
2153 * Return: The page on success or NULL if allocation fails.
2155 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2156 unsigned long addr, bool hugepage)
2158 struct mempolicy *pol;
2159 int node = numa_node_id();
2164 pol = get_vma_policy(vma, addr);
2166 if (pol->mode == MPOL_INTERLEAVE) {
2169 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2171 page = alloc_page_interleave(gfp, order, nid);
2175 if (pol->mode == MPOL_PREFERRED_MANY) {
2176 node = policy_node(gfp, pol, node);
2177 page = alloc_pages_preferred_many(gfp, order, node, pol);
2182 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2183 int hpage_node = node;
2186 * For hugepage allocation and non-interleave policy which
2187 * allows the current node (or other explicitly preferred
2188 * node) we only try to allocate from the current/preferred
2189 * node and don't fall back to other nodes, as the cost of
2190 * remote accesses would likely offset THP benefits.
2192 * If the policy is interleave or does not allow the current
2193 * node in its nodemask, we allocate the standard way.
2195 if (pol->mode == MPOL_PREFERRED)
2196 hpage_node = first_node(pol->nodes);
2198 nmask = policy_nodemask(gfp, pol);
2199 if (!nmask || node_isset(hpage_node, *nmask)) {
2202 * First, try to allocate THP only on local node, but
2203 * don't reclaim unnecessarily, just compact.
2205 page = __alloc_pages_node(hpage_node,
2206 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2209 * If hugepage allocations are configured to always
2210 * synchronous compact or the vma has been madvised
2211 * to prefer hugepage backing, retry allowing remote
2212 * memory with both reclaim and compact as well.
2214 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2215 page = __alloc_pages(gfp, order, hpage_node, nmask);
2221 nmask = policy_nodemask(gfp, pol);
2222 preferred_nid = policy_node(gfp, pol, node);
2223 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2228 EXPORT_SYMBOL(alloc_pages_vma);
2231 * alloc_pages - Allocate pages.
2233 * @order: Power of two of number of pages to allocate.
2235 * Allocate 1 << @order contiguous pages. The physical address of the
2236 * first page is naturally aligned (eg an order-3 allocation will be aligned
2237 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2238 * process is honoured when in process context.
2240 * Context: Can be called from any context, providing the appropriate GFP
2242 * Return: The page on success or NULL if allocation fails.
2244 struct page *alloc_pages(gfp_t gfp, unsigned order)
2246 struct mempolicy *pol = &default_policy;
2249 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2250 pol = get_task_policy(current);
2253 * No reference counting needed for current->mempolicy
2254 * nor system default_policy
2256 if (pol->mode == MPOL_INTERLEAVE)
2257 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2258 else if (pol->mode == MPOL_PREFERRED_MANY)
2259 page = alloc_pages_preferred_many(gfp, order,
2260 policy_node(gfp, pol, numa_node_id()), pol);
2262 page = __alloc_pages(gfp, order,
2263 policy_node(gfp, pol, numa_node_id()),
2264 policy_nodemask(gfp, pol));
2268 EXPORT_SYMBOL(alloc_pages);
2270 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2272 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2274 if (page && order > 1)
2275 prep_transhuge_page(page);
2276 return (struct folio *)page;
2278 EXPORT_SYMBOL(folio_alloc);
2280 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2281 struct mempolicy *pol, unsigned long nr_pages,
2282 struct page **page_array)
2285 unsigned long nr_pages_per_node;
2288 unsigned long nr_allocated;
2289 unsigned long total_allocated = 0;
2291 nodes = nodes_weight(pol->nodes);
2292 nr_pages_per_node = nr_pages / nodes;
2293 delta = nr_pages - nodes * nr_pages_per_node;
2295 for (i = 0; i < nodes; i++) {
2297 nr_allocated = __alloc_pages_bulk(gfp,
2298 interleave_nodes(pol), NULL,
2299 nr_pages_per_node + 1, NULL,
2303 nr_allocated = __alloc_pages_bulk(gfp,
2304 interleave_nodes(pol), NULL,
2305 nr_pages_per_node, NULL, page_array);
2308 page_array += nr_allocated;
2309 total_allocated += nr_allocated;
2312 return total_allocated;
2315 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2316 struct mempolicy *pol, unsigned long nr_pages,
2317 struct page **page_array)
2319 gfp_t preferred_gfp;
2320 unsigned long nr_allocated = 0;
2322 preferred_gfp = gfp | __GFP_NOWARN;
2323 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2325 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2326 nr_pages, NULL, page_array);
2328 if (nr_allocated < nr_pages)
2329 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2330 nr_pages - nr_allocated, NULL,
2331 page_array + nr_allocated);
2332 return nr_allocated;
2335 /* alloc pages bulk and mempolicy should be considered at the
2336 * same time in some situation such as vmalloc.
2338 * It can accelerate memory allocation especially interleaving
2341 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2342 unsigned long nr_pages, struct page **page_array)
2344 struct mempolicy *pol = &default_policy;
2346 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2347 pol = get_task_policy(current);
2349 if (pol->mode == MPOL_INTERLEAVE)
2350 return alloc_pages_bulk_array_interleave(gfp, pol,
2351 nr_pages, page_array);
2353 if (pol->mode == MPOL_PREFERRED_MANY)
2354 return alloc_pages_bulk_array_preferred_many(gfp,
2355 numa_node_id(), pol, nr_pages, page_array);
2357 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2358 policy_nodemask(gfp, pol), nr_pages, NULL,
2362 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2364 struct mempolicy *pol = mpol_dup(vma_policy(src));
2367 return PTR_ERR(pol);
2368 dst->vm_policy = pol;
2373 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2374 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2375 * with the mems_allowed returned by cpuset_mems_allowed(). This
2376 * keeps mempolicies cpuset relative after its cpuset moves. See
2377 * further kernel/cpuset.c update_nodemask().
2379 * current's mempolicy may be rebinded by the other task(the task that changes
2380 * cpuset's mems), so we needn't do rebind work for current task.
2383 /* Slow path of a mempolicy duplicate */
2384 struct mempolicy *__mpol_dup(struct mempolicy *old)
2386 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2389 return ERR_PTR(-ENOMEM);
2391 /* task's mempolicy is protected by alloc_lock */
2392 if (old == current->mempolicy) {
2395 task_unlock(current);
2399 if (current_cpuset_is_being_rebound()) {
2400 nodemask_t mems = cpuset_mems_allowed(current);
2401 mpol_rebind_policy(new, &mems);
2403 atomic_set(&new->refcnt, 1);
2407 /* Slow path of a mempolicy comparison */
2408 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2412 if (a->mode != b->mode)
2414 if (a->flags != b->flags)
2416 if (a->home_node != b->home_node)
2418 if (mpol_store_user_nodemask(a))
2419 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2424 case MPOL_INTERLEAVE:
2425 case MPOL_PREFERRED:
2426 case MPOL_PREFERRED_MANY:
2427 return !!nodes_equal(a->nodes, b->nodes);
2437 * Shared memory backing store policy support.
2439 * Remember policies even when nobody has shared memory mapped.
2440 * The policies are kept in Red-Black tree linked from the inode.
2441 * They are protected by the sp->lock rwlock, which should be held
2442 * for any accesses to the tree.
2446 * lookup first element intersecting start-end. Caller holds sp->lock for
2447 * reading or for writing
2449 static struct sp_node *
2450 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2452 struct rb_node *n = sp->root.rb_node;
2455 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2457 if (start >= p->end)
2459 else if (end <= p->start)
2467 struct sp_node *w = NULL;
2468 struct rb_node *prev = rb_prev(n);
2471 w = rb_entry(prev, struct sp_node, nd);
2472 if (w->end <= start)
2476 return rb_entry(n, struct sp_node, nd);
2480 * Insert a new shared policy into the list. Caller holds sp->lock for
2483 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2485 struct rb_node **p = &sp->root.rb_node;
2486 struct rb_node *parent = NULL;
2491 nd = rb_entry(parent, struct sp_node, nd);
2492 if (new->start < nd->start)
2494 else if (new->end > nd->end)
2495 p = &(*p)->rb_right;
2499 rb_link_node(&new->nd, parent, p);
2500 rb_insert_color(&new->nd, &sp->root);
2501 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2502 new->policy ? new->policy->mode : 0);
2505 /* Find shared policy intersecting idx */
2507 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2509 struct mempolicy *pol = NULL;
2512 if (!sp->root.rb_node)
2514 read_lock(&sp->lock);
2515 sn = sp_lookup(sp, idx, idx+1);
2517 mpol_get(sn->policy);
2520 read_unlock(&sp->lock);
2524 static void sp_free(struct sp_node *n)
2526 mpol_put(n->policy);
2527 kmem_cache_free(sn_cache, n);
2531 * mpol_misplaced - check whether current page node is valid in policy
2533 * @page: page to be checked
2534 * @vma: vm area where page mapped
2535 * @addr: virtual address where page mapped
2537 * Lookup current policy node id for vma,addr and "compare to" page's
2538 * node id. Policy determination "mimics" alloc_page_vma().
2539 * Called from fault path where we know the vma and faulting address.
2541 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2542 * policy, or a suitable node ID to allocate a replacement page from.
2544 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2546 struct mempolicy *pol;
2548 int curnid = page_to_nid(page);
2549 unsigned long pgoff;
2550 int thiscpu = raw_smp_processor_id();
2551 int thisnid = cpu_to_node(thiscpu);
2552 int polnid = NUMA_NO_NODE;
2553 int ret = NUMA_NO_NODE;
2555 pol = get_vma_policy(vma, addr);
2556 if (!(pol->flags & MPOL_F_MOF))
2559 switch (pol->mode) {
2560 case MPOL_INTERLEAVE:
2561 pgoff = vma->vm_pgoff;
2562 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2563 polnid = offset_il_node(pol, pgoff);
2566 case MPOL_PREFERRED:
2567 if (node_isset(curnid, pol->nodes))
2569 polnid = first_node(pol->nodes);
2573 polnid = numa_node_id();
2577 /* Optimize placement among multiple nodes via NUMA balancing */
2578 if (pol->flags & MPOL_F_MORON) {
2579 if (node_isset(thisnid, pol->nodes))
2585 case MPOL_PREFERRED_MANY:
2587 * use current page if in policy nodemask,
2588 * else select nearest allowed node, if any.
2589 * If no allowed nodes, use current [!misplaced].
2591 if (node_isset(curnid, pol->nodes))
2593 z = first_zones_zonelist(
2594 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2595 gfp_zone(GFP_HIGHUSER),
2597 polnid = zone_to_nid(z->zone);
2604 /* Migrate the page towards the node whose CPU is referencing it */
2605 if (pol->flags & MPOL_F_MORON) {
2608 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2612 if (curnid != polnid)
2621 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2622 * dropped after task->mempolicy is set to NULL so that any allocation done as
2623 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2626 void mpol_put_task_policy(struct task_struct *task)
2628 struct mempolicy *pol;
2631 pol = task->mempolicy;
2632 task->mempolicy = NULL;
2637 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2639 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2640 rb_erase(&n->nd, &sp->root);
2644 static void sp_node_init(struct sp_node *node, unsigned long start,
2645 unsigned long end, struct mempolicy *pol)
2647 node->start = start;
2652 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2653 struct mempolicy *pol)
2656 struct mempolicy *newpol;
2658 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2662 newpol = mpol_dup(pol);
2663 if (IS_ERR(newpol)) {
2664 kmem_cache_free(sn_cache, n);
2667 newpol->flags |= MPOL_F_SHARED;
2668 sp_node_init(n, start, end, newpol);
2673 /* Replace a policy range. */
2674 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2675 unsigned long end, struct sp_node *new)
2678 struct sp_node *n_new = NULL;
2679 struct mempolicy *mpol_new = NULL;
2683 write_lock(&sp->lock);
2684 n = sp_lookup(sp, start, end);
2685 /* Take care of old policies in the same range. */
2686 while (n && n->start < end) {
2687 struct rb_node *next = rb_next(&n->nd);
2688 if (n->start >= start) {
2694 /* Old policy spanning whole new range. */
2699 *mpol_new = *n->policy;
2700 atomic_set(&mpol_new->refcnt, 1);
2701 sp_node_init(n_new, end, n->end, mpol_new);
2703 sp_insert(sp, n_new);
2712 n = rb_entry(next, struct sp_node, nd);
2716 write_unlock(&sp->lock);
2723 kmem_cache_free(sn_cache, n_new);
2728 write_unlock(&sp->lock);
2730 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2733 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2740 * mpol_shared_policy_init - initialize shared policy for inode
2741 * @sp: pointer to inode shared policy
2742 * @mpol: struct mempolicy to install
2744 * Install non-NULL @mpol in inode's shared policy rb-tree.
2745 * On entry, the current task has a reference on a non-NULL @mpol.
2746 * This must be released on exit.
2747 * This is called at get_inode() calls and we can use GFP_KERNEL.
2749 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2753 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2754 rwlock_init(&sp->lock);
2757 struct vm_area_struct pvma;
2758 struct mempolicy *new;
2759 NODEMASK_SCRATCH(scratch);
2763 /* contextualize the tmpfs mount point mempolicy */
2764 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2766 goto free_scratch; /* no valid nodemask intersection */
2769 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2770 task_unlock(current);
2774 /* Create pseudo-vma that contains just the policy */
2775 vma_init(&pvma, NULL);
2776 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2777 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2780 mpol_put(new); /* drop initial ref */
2782 NODEMASK_SCRATCH_FREE(scratch);
2784 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2788 int mpol_set_shared_policy(struct shared_policy *info,
2789 struct vm_area_struct *vma, struct mempolicy *npol)
2792 struct sp_node *new = NULL;
2793 unsigned long sz = vma_pages(vma);
2795 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2797 sz, npol ? npol->mode : -1,
2798 npol ? npol->flags : -1,
2799 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2802 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2806 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2812 /* Free a backing policy store on inode delete. */
2813 void mpol_free_shared_policy(struct shared_policy *p)
2816 struct rb_node *next;
2818 if (!p->root.rb_node)
2820 write_lock(&p->lock);
2821 next = rb_first(&p->root);
2823 n = rb_entry(next, struct sp_node, nd);
2824 next = rb_next(&n->nd);
2827 write_unlock(&p->lock);
2830 #ifdef CONFIG_NUMA_BALANCING
2831 static int __initdata numabalancing_override;
2833 static void __init check_numabalancing_enable(void)
2835 bool numabalancing_default = false;
2837 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2838 numabalancing_default = true;
2840 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2841 if (numabalancing_override)
2842 set_numabalancing_state(numabalancing_override == 1);
2844 if (num_online_nodes() > 1 && !numabalancing_override) {
2845 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2846 numabalancing_default ? "Enabling" : "Disabling");
2847 set_numabalancing_state(numabalancing_default);
2851 static int __init setup_numabalancing(char *str)
2857 if (!strcmp(str, "enable")) {
2858 numabalancing_override = 1;
2860 } else if (!strcmp(str, "disable")) {
2861 numabalancing_override = -1;
2866 pr_warn("Unable to parse numa_balancing=\n");
2870 __setup("numa_balancing=", setup_numabalancing);
2872 static inline void __init check_numabalancing_enable(void)
2875 #endif /* CONFIG_NUMA_BALANCING */
2877 /* assumes fs == KERNEL_DS */
2878 void __init numa_policy_init(void)
2880 nodemask_t interleave_nodes;
2881 unsigned long largest = 0;
2882 int nid, prefer = 0;
2884 policy_cache = kmem_cache_create("numa_policy",
2885 sizeof(struct mempolicy),
2886 0, SLAB_PANIC, NULL);
2888 sn_cache = kmem_cache_create("shared_policy_node",
2889 sizeof(struct sp_node),
2890 0, SLAB_PANIC, NULL);
2892 for_each_node(nid) {
2893 preferred_node_policy[nid] = (struct mempolicy) {
2894 .refcnt = ATOMIC_INIT(1),
2895 .mode = MPOL_PREFERRED,
2896 .flags = MPOL_F_MOF | MPOL_F_MORON,
2897 .nodes = nodemask_of_node(nid),
2902 * Set interleaving policy for system init. Interleaving is only
2903 * enabled across suitably sized nodes (default is >= 16MB), or
2904 * fall back to the largest node if they're all smaller.
2906 nodes_clear(interleave_nodes);
2907 for_each_node_state(nid, N_MEMORY) {
2908 unsigned long total_pages = node_present_pages(nid);
2910 /* Preserve the largest node */
2911 if (largest < total_pages) {
2912 largest = total_pages;
2916 /* Interleave this node? */
2917 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2918 node_set(nid, interleave_nodes);
2921 /* All too small, use the largest */
2922 if (unlikely(nodes_empty(interleave_nodes)))
2923 node_set(prefer, interleave_nodes);
2925 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2926 pr_err("%s: interleaving failed\n", __func__);
2928 check_numabalancing_enable();
2931 /* Reset policy of current process to default */
2932 void numa_default_policy(void)
2934 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2938 * Parse and format mempolicy from/to strings
2941 static const char * const policy_modes[] =
2943 [MPOL_DEFAULT] = "default",
2944 [MPOL_PREFERRED] = "prefer",
2945 [MPOL_BIND] = "bind",
2946 [MPOL_INTERLEAVE] = "interleave",
2947 [MPOL_LOCAL] = "local",
2948 [MPOL_PREFERRED_MANY] = "prefer (many)",
2954 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2955 * @str: string containing mempolicy to parse
2956 * @mpol: pointer to struct mempolicy pointer, returned on success.
2959 * <mode>[=<flags>][:<nodelist>]
2961 * Return: %0 on success, else %1
2963 int mpol_parse_str(char *str, struct mempolicy **mpol)
2965 struct mempolicy *new = NULL;
2966 unsigned short mode_flags;
2968 char *nodelist = strchr(str, ':');
2969 char *flags = strchr(str, '=');
2973 *flags++ = '\0'; /* terminate mode string */
2976 /* NUL-terminate mode or flags string */
2978 if (nodelist_parse(nodelist, nodes))
2980 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2985 mode = match_string(policy_modes, MPOL_MAX, str);
2990 case MPOL_PREFERRED:
2992 * Insist on a nodelist of one node only, although later
2993 * we use first_node(nodes) to grab a single node, so here
2994 * nodelist (or nodes) cannot be empty.
2997 char *rest = nodelist;
2998 while (isdigit(*rest))
3002 if (nodes_empty(nodes))
3006 case MPOL_INTERLEAVE:
3008 * Default to online nodes with memory if no nodelist
3011 nodes = node_states[N_MEMORY];
3015 * Don't allow a nodelist; mpol_new() checks flags
3022 * Insist on a empty nodelist
3027 case MPOL_PREFERRED_MANY:
3030 * Insist on a nodelist
3039 * Currently, we only support two mutually exclusive
3042 if (!strcmp(flags, "static"))
3043 mode_flags |= MPOL_F_STATIC_NODES;
3044 else if (!strcmp(flags, "relative"))
3045 mode_flags |= MPOL_F_RELATIVE_NODES;
3050 new = mpol_new(mode, mode_flags, &nodes);
3055 * Save nodes for mpol_to_str() to show the tmpfs mount options
3056 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3058 if (mode != MPOL_PREFERRED) {
3060 } else if (nodelist) {
3061 nodes_clear(new->nodes);
3062 node_set(first_node(nodes), new->nodes);
3064 new->mode = MPOL_LOCAL;
3068 * Save nodes for contextualization: this will be used to "clone"
3069 * the mempolicy in a specific context [cpuset] at a later time.
3071 new->w.user_nodemask = nodes;
3076 /* Restore string for error message */
3085 #endif /* CONFIG_TMPFS */
3088 * mpol_to_str - format a mempolicy structure for printing
3089 * @buffer: to contain formatted mempolicy string
3090 * @maxlen: length of @buffer
3091 * @pol: pointer to mempolicy to be formatted
3093 * Convert @pol into a string. If @buffer is too short, truncate the string.
3094 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3095 * longest flag, "relative", and to display at least a few node ids.
3097 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3100 nodemask_t nodes = NODE_MASK_NONE;
3101 unsigned short mode = MPOL_DEFAULT;
3102 unsigned short flags = 0;
3104 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3113 case MPOL_PREFERRED:
3114 case MPOL_PREFERRED_MANY:
3116 case MPOL_INTERLEAVE:
3121 snprintf(p, maxlen, "unknown");
3125 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3127 if (flags & MPOL_MODE_FLAGS) {
3128 p += snprintf(p, buffer + maxlen - p, "=");
3131 * Currently, the only defined flags are mutually exclusive
3133 if (flags & MPOL_F_STATIC_NODES)
3134 p += snprintf(p, buffer + maxlen - p, "static");
3135 else if (flags & MPOL_F_RELATIVE_NODES)
3136 p += snprintf(p, buffer + maxlen - p, "relative");
3139 if (!nodes_empty(nodes))
3140 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3141 nodemask_pr_args(&nodes));