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 *next;
790 struct vm_area_struct *prev;
791 struct vm_area_struct *vma;
794 unsigned long vmstart;
797 vma = find_vma(mm, start);
801 if (start > vma->vm_start)
804 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
806 vmstart = max(start, vma->vm_start);
807 vmend = min(end, vma->vm_end);
809 if (mpol_equal(vma_policy(vma), new_pol))
812 pgoff = vma->vm_pgoff +
813 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
814 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
815 vma->anon_vma, vma->vm_file, pgoff,
816 new_pol, vma->vm_userfaultfd_ctx,
821 if (mpol_equal(vma_policy(vma), new_pol))
823 /* vma_merge() joined vma && vma->next, case 8 */
826 if (vma->vm_start != vmstart) {
827 err = split_vma(vma->vm_mm, vma, vmstart, 1);
831 if (vma->vm_end != vmend) {
832 err = split_vma(vma->vm_mm, vma, vmend, 0);
837 err = vma_replace_policy(vma, new_pol);
846 /* Set the process memory policy */
847 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
850 struct mempolicy *new, *old;
851 NODEMASK_SCRATCH(scratch);
857 new = mpol_new(mode, flags, nodes);
863 ret = mpol_set_nodemask(new, nodes, scratch);
869 old = current->mempolicy;
870 current->mempolicy = new;
871 if (new && new->mode == MPOL_INTERLEAVE)
872 current->il_prev = MAX_NUMNODES-1;
873 task_unlock(current);
877 NODEMASK_SCRATCH_FREE(scratch);
882 * Return nodemask for policy for get_mempolicy() query
884 * Called with task's alloc_lock held
886 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
889 if (p == &default_policy)
894 case MPOL_INTERLEAVE:
896 case MPOL_PREFERRED_MANY:
900 /* return empty node mask for local allocation */
907 static int lookup_node(struct mm_struct *mm, unsigned long addr)
909 struct page *p = NULL;
913 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
915 err = page_to_nid(p);
919 mmap_read_unlock(mm);
923 /* Retrieve NUMA policy */
924 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
925 unsigned long addr, unsigned long flags)
928 struct mm_struct *mm = current->mm;
929 struct vm_area_struct *vma = NULL;
930 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
933 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
936 if (flags & MPOL_F_MEMS_ALLOWED) {
937 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
939 *policy = 0; /* just so it's initialized */
941 *nmask = cpuset_current_mems_allowed;
942 task_unlock(current);
946 if (flags & MPOL_F_ADDR) {
948 * Do NOT fall back to task policy if the
949 * vma/shared policy at addr is NULL. We
950 * want to return MPOL_DEFAULT in this case.
953 vma = vma_lookup(mm, addr);
955 mmap_read_unlock(mm);
958 if (vma->vm_ops && vma->vm_ops->get_policy)
959 pol = vma->vm_ops->get_policy(vma, addr);
961 pol = vma->vm_policy;
966 pol = &default_policy; /* indicates default behavior */
968 if (flags & MPOL_F_NODE) {
969 if (flags & MPOL_F_ADDR) {
971 * Take a refcount on the mpol, lookup_node()
972 * will drop the mmap_lock, so after calling
973 * lookup_node() only "pol" remains valid, "vma"
979 err = lookup_node(mm, addr);
983 } else if (pol == current->mempolicy &&
984 pol->mode == MPOL_INTERLEAVE) {
985 *policy = next_node_in(current->il_prev, pol->nodes);
991 *policy = pol == &default_policy ? MPOL_DEFAULT :
994 * Internal mempolicy flags must be masked off before exposing
995 * the policy to userspace.
997 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1002 if (mpol_store_user_nodemask(pol)) {
1003 *nmask = pol->w.user_nodemask;
1006 get_policy_nodemask(pol, nmask);
1007 task_unlock(current);
1014 mmap_read_unlock(mm);
1016 mpol_put(pol_refcount);
1020 #ifdef CONFIG_MIGRATION
1022 * page migration, thp tail pages can be passed.
1024 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1025 unsigned long flags)
1027 struct page *head = compound_head(page);
1029 * Avoid migrating a page that is shared with others.
1031 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1032 if (!isolate_lru_page(head)) {
1033 list_add_tail(&head->lru, pagelist);
1034 mod_node_page_state(page_pgdat(head),
1035 NR_ISOLATED_ANON + page_is_file_lru(head),
1036 thp_nr_pages(head));
1037 } else if (flags & MPOL_MF_STRICT) {
1039 * Non-movable page may reach here. And, there may be
1040 * temporary off LRU pages or non-LRU movable pages.
1041 * Treat them as unmovable pages since they can't be
1042 * isolated, so they can't be moved at the moment. It
1043 * should return -EIO for this case too.
1053 * Migrate pages from one node to a target node.
1054 * Returns error or the number of pages not migrated.
1056 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1060 LIST_HEAD(pagelist);
1062 struct migration_target_control mtc = {
1064 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1068 node_set(source, nmask);
1071 * This does not "check" the range but isolates all pages that
1072 * need migration. Between passing in the full user address
1073 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1075 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1076 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1077 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1079 if (!list_empty(&pagelist)) {
1080 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1081 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1083 putback_movable_pages(&pagelist);
1090 * Move pages between the two nodesets so as to preserve the physical
1091 * layout as much as possible.
1093 * Returns the number of page that could not be moved.
1095 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1096 const nodemask_t *to, int flags)
1102 lru_cache_disable();
1107 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1108 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1109 * bit in 'tmp', and return that <source, dest> pair for migration.
1110 * The pair of nodemasks 'to' and 'from' define the map.
1112 * If no pair of bits is found that way, fallback to picking some
1113 * pair of 'source' and 'dest' bits that are not the same. If the
1114 * 'source' and 'dest' bits are the same, this represents a node
1115 * that will be migrating to itself, so no pages need move.
1117 * If no bits are left in 'tmp', or if all remaining bits left
1118 * in 'tmp' correspond to the same bit in 'to', return false
1119 * (nothing left to migrate).
1121 * This lets us pick a pair of nodes to migrate between, such that
1122 * if possible the dest node is not already occupied by some other
1123 * source node, minimizing the risk of overloading the memory on a
1124 * node that would happen if we migrated incoming memory to a node
1125 * before migrating outgoing memory source that same node.
1127 * A single scan of tmp is sufficient. As we go, we remember the
1128 * most recent <s, d> pair that moved (s != d). If we find a pair
1129 * that not only moved, but what's better, moved to an empty slot
1130 * (d is not set in tmp), then we break out then, with that pair.
1131 * Otherwise when we finish scanning from_tmp, we at least have the
1132 * most recent <s, d> pair that moved. If we get all the way through
1133 * the scan of tmp without finding any node that moved, much less
1134 * moved to an empty node, then there is nothing left worth migrating.
1138 while (!nodes_empty(tmp)) {
1140 int source = NUMA_NO_NODE;
1143 for_each_node_mask(s, tmp) {
1146 * do_migrate_pages() tries to maintain the relative
1147 * node relationship of the pages established between
1148 * threads and memory areas.
1150 * However if the number of source nodes is not equal to
1151 * the number of destination nodes we can not preserve
1152 * this node relative relationship. In that case, skip
1153 * copying memory from a node that is in the destination
1156 * Example: [2,3,4] -> [3,4,5] moves everything.
1157 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1160 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1161 (node_isset(s, *to)))
1164 d = node_remap(s, *from, *to);
1168 source = s; /* Node moved. Memorize */
1171 /* dest not in remaining from nodes? */
1172 if (!node_isset(dest, tmp))
1175 if (source == NUMA_NO_NODE)
1178 node_clear(source, tmp);
1179 err = migrate_to_node(mm, source, dest, flags);
1185 mmap_read_unlock(mm);
1195 * Allocate a new page for page migration based on vma policy.
1196 * Start by assuming the page is mapped by the same vma as contains @start.
1197 * Search forward from there, if not. N.B., this assumes that the
1198 * list of pages handed to migrate_pages()--which is how we get here--
1199 * is in virtual address order.
1201 static struct page *new_page(struct page *page, unsigned long start)
1203 struct vm_area_struct *vma;
1204 unsigned long address;
1206 vma = find_vma(current->mm, start);
1208 address = page_address_in_vma(page, vma);
1209 if (address != -EFAULT)
1214 if (PageHuge(page)) {
1215 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1217 } else if (PageTransHuge(page)) {
1220 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1224 prep_transhuge_page(thp);
1228 * if !vma, alloc_page_vma() will use task or system default policy
1230 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1235 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1236 unsigned long flags)
1241 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1242 const nodemask_t *to, int flags)
1247 static struct page *new_page(struct page *page, unsigned long start)
1253 static long do_mbind(unsigned long start, unsigned long len,
1254 unsigned short mode, unsigned short mode_flags,
1255 nodemask_t *nmask, unsigned long flags)
1257 struct mm_struct *mm = current->mm;
1258 struct mempolicy *new;
1262 LIST_HEAD(pagelist);
1264 if (flags & ~(unsigned long)MPOL_MF_VALID)
1266 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1269 if (start & ~PAGE_MASK)
1272 if (mode == MPOL_DEFAULT)
1273 flags &= ~MPOL_MF_STRICT;
1275 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1283 new = mpol_new(mode, mode_flags, nmask);
1285 return PTR_ERR(new);
1287 if (flags & MPOL_MF_LAZY)
1288 new->flags |= MPOL_F_MOF;
1291 * If we are using the default policy then operation
1292 * on discontinuous address spaces is okay after all
1295 flags |= MPOL_MF_DISCONTIG_OK;
1297 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1298 start, start + len, mode, mode_flags,
1299 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1301 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1303 lru_cache_disable();
1306 NODEMASK_SCRATCH(scratch);
1308 mmap_write_lock(mm);
1309 err = mpol_set_nodemask(new, nmask, scratch);
1311 mmap_write_unlock(mm);
1314 NODEMASK_SCRATCH_FREE(scratch);
1319 ret = queue_pages_range(mm, start, end, nmask,
1320 flags | MPOL_MF_INVERT, &pagelist);
1327 err = mbind_range(mm, start, end, new);
1332 if (!list_empty(&pagelist)) {
1333 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1334 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1335 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1337 putback_movable_pages(&pagelist);
1340 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1344 if (!list_empty(&pagelist))
1345 putback_movable_pages(&pagelist);
1348 mmap_write_unlock(mm);
1351 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1357 * User space interface with variable sized bitmaps for nodelists.
1359 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1360 unsigned long maxnode)
1362 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1365 if (in_compat_syscall())
1366 ret = compat_get_bitmap(mask,
1367 (const compat_ulong_t __user *)nmask,
1370 ret = copy_from_user(mask, nmask,
1371 nlongs * sizeof(unsigned long));
1376 if (maxnode % BITS_PER_LONG)
1377 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1382 /* Copy a node mask from user space. */
1383 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1384 unsigned long maxnode)
1387 nodes_clear(*nodes);
1388 if (maxnode == 0 || !nmask)
1390 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1394 * When the user specified more nodes than supported just check
1395 * if the non supported part is all zero, one word at a time,
1396 * starting at the end.
1398 while (maxnode > MAX_NUMNODES) {
1399 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1402 if (get_bitmap(&t, &nmask[maxnode / BITS_PER_LONG], bits))
1405 if (maxnode - bits >= MAX_NUMNODES) {
1408 maxnode = MAX_NUMNODES;
1409 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1415 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1418 /* Copy a kernel node mask to user space */
1419 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1422 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1423 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1424 bool compat = in_compat_syscall();
1427 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1429 if (copy > nbytes) {
1430 if (copy > PAGE_SIZE)
1432 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1435 maxnode = nr_node_ids;
1439 return compat_put_bitmap((compat_ulong_t __user *)mask,
1440 nodes_addr(*nodes), maxnode);
1442 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1445 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1446 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1448 *flags = *mode & MPOL_MODE_FLAGS;
1449 *mode &= ~MPOL_MODE_FLAGS;
1451 if ((unsigned int)(*mode) >= MPOL_MAX)
1453 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1455 if (*flags & MPOL_F_NUMA_BALANCING) {
1456 if (*mode != MPOL_BIND)
1458 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1463 static long kernel_mbind(unsigned long start, unsigned long len,
1464 unsigned long mode, const unsigned long __user *nmask,
1465 unsigned long maxnode, unsigned int flags)
1467 unsigned short mode_flags;
1472 start = untagged_addr(start);
1473 err = sanitize_mpol_flags(&lmode, &mode_flags);
1477 err = get_nodes(&nodes, nmask, maxnode);
1481 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1484 SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1485 unsigned long, home_node, unsigned long, flags)
1487 struct mm_struct *mm = current->mm;
1488 struct vm_area_struct *vma;
1489 struct mempolicy *new;
1490 unsigned long vmstart;
1491 unsigned long vmend;
1495 start = untagged_addr(start);
1496 if (start & ~PAGE_MASK)
1499 * flags is used for future extension if any.
1505 * Check home_node is online to avoid accessing uninitialized
1508 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1511 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1518 mmap_write_lock(mm);
1519 vma = find_vma(mm, start);
1520 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
1522 vmstart = max(start, vma->vm_start);
1523 vmend = min(end, vma->vm_end);
1524 new = mpol_dup(vma_policy(vma));
1530 * Only update home node if there is an existing vma policy
1536 * If any vma in the range got policy other than MPOL_BIND
1537 * or MPOL_PREFERRED_MANY we return error. We don't reset
1538 * the home node for vmas we already updated before.
1540 if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1545 new->home_node = home_node;
1546 err = mbind_range(mm, vmstart, vmend, new);
1551 mmap_write_unlock(mm);
1555 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1556 unsigned long, mode, const unsigned long __user *, nmask,
1557 unsigned long, maxnode, unsigned int, flags)
1559 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1562 /* Set the process memory policy */
1563 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1564 unsigned long maxnode)
1566 unsigned short mode_flags;
1571 err = sanitize_mpol_flags(&lmode, &mode_flags);
1575 err = get_nodes(&nodes, nmask, maxnode);
1579 return do_set_mempolicy(lmode, mode_flags, &nodes);
1582 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1583 unsigned long, maxnode)
1585 return kernel_set_mempolicy(mode, nmask, maxnode);
1588 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1589 const unsigned long __user *old_nodes,
1590 const unsigned long __user *new_nodes)
1592 struct mm_struct *mm = NULL;
1593 struct task_struct *task;
1594 nodemask_t task_nodes;
1598 NODEMASK_SCRATCH(scratch);
1603 old = &scratch->mask1;
1604 new = &scratch->mask2;
1606 err = get_nodes(old, old_nodes, maxnode);
1610 err = get_nodes(new, new_nodes, maxnode);
1614 /* Find the mm_struct */
1616 task = pid ? find_task_by_vpid(pid) : current;
1622 get_task_struct(task);
1627 * Check if this process has the right to modify the specified process.
1628 * Use the regular "ptrace_may_access()" checks.
1630 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1637 task_nodes = cpuset_mems_allowed(task);
1638 /* Is the user allowed to access the target nodes? */
1639 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1644 task_nodes = cpuset_mems_allowed(current);
1645 nodes_and(*new, *new, task_nodes);
1646 if (nodes_empty(*new))
1649 err = security_task_movememory(task);
1653 mm = get_task_mm(task);
1654 put_task_struct(task);
1661 err = do_migrate_pages(mm, old, new,
1662 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1666 NODEMASK_SCRATCH_FREE(scratch);
1671 put_task_struct(task);
1676 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1677 const unsigned long __user *, old_nodes,
1678 const unsigned long __user *, new_nodes)
1680 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1684 /* Retrieve NUMA policy */
1685 static int kernel_get_mempolicy(int __user *policy,
1686 unsigned long __user *nmask,
1687 unsigned long maxnode,
1689 unsigned long flags)
1695 if (nmask != NULL && maxnode < nr_node_ids)
1698 addr = untagged_addr(addr);
1700 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1705 if (policy && put_user(pval, policy))
1709 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1714 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1715 unsigned long __user *, nmask, unsigned long, maxnode,
1716 unsigned long, addr, unsigned long, flags)
1718 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1721 bool vma_migratable(struct vm_area_struct *vma)
1723 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1727 * DAX device mappings require predictable access latency, so avoid
1728 * incurring periodic faults.
1730 if (vma_is_dax(vma))
1733 if (is_vm_hugetlb_page(vma) &&
1734 !hugepage_migration_supported(hstate_vma(vma)))
1738 * Migration allocates pages in the highest zone. If we cannot
1739 * do so then migration (at least from node to node) is not
1743 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1749 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1752 struct mempolicy *pol = NULL;
1755 if (vma->vm_ops && vma->vm_ops->get_policy) {
1756 pol = vma->vm_ops->get_policy(vma, addr);
1757 } else if (vma->vm_policy) {
1758 pol = vma->vm_policy;
1761 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1762 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1763 * count on these policies which will be dropped by
1764 * mpol_cond_put() later
1766 if (mpol_needs_cond_ref(pol))
1775 * get_vma_policy(@vma, @addr)
1776 * @vma: virtual memory area whose policy is sought
1777 * @addr: address in @vma for shared policy lookup
1779 * Returns effective policy for a VMA at specified address.
1780 * Falls back to current->mempolicy or system default policy, as necessary.
1781 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1782 * count--added by the get_policy() vm_op, as appropriate--to protect against
1783 * freeing by another task. It is the caller's responsibility to free the
1784 * extra reference for shared policies.
1786 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1789 struct mempolicy *pol = __get_vma_policy(vma, addr);
1792 pol = get_task_policy(current);
1797 bool vma_policy_mof(struct vm_area_struct *vma)
1799 struct mempolicy *pol;
1801 if (vma->vm_ops && vma->vm_ops->get_policy) {
1804 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1805 if (pol && (pol->flags & MPOL_F_MOF))
1812 pol = vma->vm_policy;
1814 pol = get_task_policy(current);
1816 return pol->flags & MPOL_F_MOF;
1819 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1821 enum zone_type dynamic_policy_zone = policy_zone;
1823 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1826 * if policy->nodes has movable memory only,
1827 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1829 * policy->nodes is intersect with node_states[N_MEMORY].
1830 * so if the following test fails, it implies
1831 * policy->nodes has movable memory only.
1833 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1834 dynamic_policy_zone = ZONE_MOVABLE;
1836 return zone >= dynamic_policy_zone;
1840 * Return a nodemask representing a mempolicy for filtering nodes for
1843 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1845 int mode = policy->mode;
1847 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1848 if (unlikely(mode == MPOL_BIND) &&
1849 apply_policy_zone(policy, gfp_zone(gfp)) &&
1850 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1851 return &policy->nodes;
1853 if (mode == MPOL_PREFERRED_MANY)
1854 return &policy->nodes;
1860 * Return the preferred node id for 'prefer' mempolicy, and return
1861 * the given id for all other policies.
1863 * policy_node() is always coupled with policy_nodemask(), which
1864 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1866 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1868 if (policy->mode == MPOL_PREFERRED) {
1869 nd = first_node(policy->nodes);
1872 * __GFP_THISNODE shouldn't even be used with the bind policy
1873 * because we might easily break the expectation to stay on the
1874 * requested node and not break the policy.
1876 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1879 if ((policy->mode == MPOL_BIND ||
1880 policy->mode == MPOL_PREFERRED_MANY) &&
1881 policy->home_node != NUMA_NO_NODE)
1882 return policy->home_node;
1887 /* Do dynamic interleaving for a process */
1888 static unsigned interleave_nodes(struct mempolicy *policy)
1891 struct task_struct *me = current;
1893 next = next_node_in(me->il_prev, policy->nodes);
1894 if (next < MAX_NUMNODES)
1900 * Depending on the memory policy provide a node from which to allocate the
1903 unsigned int mempolicy_slab_node(void)
1905 struct mempolicy *policy;
1906 int node = numa_mem_id();
1911 policy = current->mempolicy;
1915 switch (policy->mode) {
1916 case MPOL_PREFERRED:
1917 return first_node(policy->nodes);
1919 case MPOL_INTERLEAVE:
1920 return interleave_nodes(policy);
1923 case MPOL_PREFERRED_MANY:
1928 * Follow bind policy behavior and start allocation at the
1931 struct zonelist *zonelist;
1932 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1933 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1934 z = first_zones_zonelist(zonelist, highest_zoneidx,
1936 return z->zone ? zone_to_nid(z->zone) : node;
1947 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1948 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1949 * number of present nodes.
1951 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1953 nodemask_t nodemask = pol->nodes;
1954 unsigned int target, nnodes;
1958 * The barrier will stabilize the nodemask in a register or on
1959 * the stack so that it will stop changing under the code.
1961 * Between first_node() and next_node(), pol->nodes could be changed
1962 * by other threads. So we put pol->nodes in a local stack.
1966 nnodes = nodes_weight(nodemask);
1968 return numa_node_id();
1969 target = (unsigned int)n % nnodes;
1970 nid = first_node(nodemask);
1971 for (i = 0; i < target; i++)
1972 nid = next_node(nid, nodemask);
1976 /* Determine a node number for interleave */
1977 static inline unsigned interleave_nid(struct mempolicy *pol,
1978 struct vm_area_struct *vma, unsigned long addr, int shift)
1984 * for small pages, there is no difference between
1985 * shift and PAGE_SHIFT, so the bit-shift is safe.
1986 * for huge pages, since vm_pgoff is in units of small
1987 * pages, we need to shift off the always 0 bits to get
1990 BUG_ON(shift < PAGE_SHIFT);
1991 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1992 off += (addr - vma->vm_start) >> shift;
1993 return offset_il_node(pol, off);
1995 return interleave_nodes(pol);
1998 #ifdef CONFIG_HUGETLBFS
2000 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2001 * @vma: virtual memory area whose policy is sought
2002 * @addr: address in @vma for shared policy lookup and interleave policy
2003 * @gfp_flags: for requested zone
2004 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2005 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2007 * Returns a nid suitable for a huge page allocation and a pointer
2008 * to the struct mempolicy for conditional unref after allocation.
2009 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2010 * to the mempolicy's @nodemask for filtering the zonelist.
2012 * Must be protected by read_mems_allowed_begin()
2014 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2015 struct mempolicy **mpol, nodemask_t **nodemask)
2020 *mpol = get_vma_policy(vma, addr);
2022 mode = (*mpol)->mode;
2024 if (unlikely(mode == MPOL_INTERLEAVE)) {
2025 nid = interleave_nid(*mpol, vma, addr,
2026 huge_page_shift(hstate_vma(vma)));
2028 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2029 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2030 *nodemask = &(*mpol)->nodes;
2036 * init_nodemask_of_mempolicy
2038 * If the current task's mempolicy is "default" [NULL], return 'false'
2039 * to indicate default policy. Otherwise, extract the policy nodemask
2040 * for 'bind' or 'interleave' policy into the argument nodemask, or
2041 * initialize the argument nodemask to contain the single node for
2042 * 'preferred' or 'local' policy and return 'true' to indicate presence
2043 * of non-default mempolicy.
2045 * We don't bother with reference counting the mempolicy [mpol_get/put]
2046 * because the current task is examining it's own mempolicy and a task's
2047 * mempolicy is only ever changed by the task itself.
2049 * N.B., it is the caller's responsibility to free a returned nodemask.
2051 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2053 struct mempolicy *mempolicy;
2055 if (!(mask && current->mempolicy))
2059 mempolicy = current->mempolicy;
2060 switch (mempolicy->mode) {
2061 case MPOL_PREFERRED:
2062 case MPOL_PREFERRED_MANY:
2064 case MPOL_INTERLEAVE:
2065 *mask = mempolicy->nodes;
2069 init_nodemask_of_node(mask, numa_node_id());
2075 task_unlock(current);
2082 * mempolicy_in_oom_domain
2084 * If tsk's mempolicy is "bind", check for intersection between mask and
2085 * the policy nodemask. Otherwise, return true for all other policies
2086 * including "interleave", as a tsk with "interleave" policy may have
2087 * memory allocated from all nodes in system.
2089 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2091 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2092 const nodemask_t *mask)
2094 struct mempolicy *mempolicy;
2101 mempolicy = tsk->mempolicy;
2102 if (mempolicy && mempolicy->mode == MPOL_BIND)
2103 ret = nodes_intersects(mempolicy->nodes, *mask);
2109 /* Allocate a page in interleaved policy.
2110 Own path because it needs to do special accounting. */
2111 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2116 page = __alloc_pages(gfp, order, nid, NULL);
2117 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2118 if (!static_branch_likely(&vm_numa_stat_key))
2120 if (page && page_to_nid(page) == nid) {
2122 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2128 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2129 int nid, struct mempolicy *pol)
2132 gfp_t preferred_gfp;
2135 * This is a two pass approach. The first pass will only try the
2136 * preferred nodes but skip the direct reclaim and allow the
2137 * allocation to fail, while the second pass will try all the
2140 preferred_gfp = gfp | __GFP_NOWARN;
2141 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2142 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2144 page = __alloc_pages(gfp, order, nid, NULL);
2150 * alloc_pages_vma - Allocate a page for a VMA.
2152 * @order: Order of the GFP allocation.
2153 * @vma: Pointer to VMA or NULL if not available.
2154 * @addr: Virtual address of the allocation. Must be inside @vma.
2155 * @hugepage: For hugepages try only the preferred node if possible.
2157 * Allocate a page for a specific address in @vma, using the appropriate
2158 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2159 * of the mm_struct of the VMA to prevent it from going away. Should be
2160 * used for all allocations for pages that will be mapped into user space.
2162 * Return: The page on success or NULL if allocation fails.
2164 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2165 unsigned long addr, bool hugepage)
2167 struct mempolicy *pol;
2168 int node = numa_node_id();
2173 pol = get_vma_policy(vma, addr);
2175 if (pol->mode == MPOL_INTERLEAVE) {
2178 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2180 page = alloc_page_interleave(gfp, order, nid);
2184 if (pol->mode == MPOL_PREFERRED_MANY) {
2185 node = policy_node(gfp, pol, node);
2186 page = alloc_pages_preferred_many(gfp, order, node, pol);
2191 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2192 int hpage_node = node;
2195 * For hugepage allocation and non-interleave policy which
2196 * allows the current node (or other explicitly preferred
2197 * node) we only try to allocate from the current/preferred
2198 * node and don't fall back to other nodes, as the cost of
2199 * remote accesses would likely offset THP benefits.
2201 * If the policy is interleave or does not allow the current
2202 * node in its nodemask, we allocate the standard way.
2204 if (pol->mode == MPOL_PREFERRED)
2205 hpage_node = first_node(pol->nodes);
2207 nmask = policy_nodemask(gfp, pol);
2208 if (!nmask || node_isset(hpage_node, *nmask)) {
2211 * First, try to allocate THP only on local node, but
2212 * don't reclaim unnecessarily, just compact.
2214 page = __alloc_pages_node(hpage_node,
2215 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2218 * If hugepage allocations are configured to always
2219 * synchronous compact or the vma has been madvised
2220 * to prefer hugepage backing, retry allowing remote
2221 * memory with both reclaim and compact as well.
2223 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2224 page = __alloc_pages(gfp, order, hpage_node, nmask);
2230 nmask = policy_nodemask(gfp, pol);
2231 preferred_nid = policy_node(gfp, pol, node);
2232 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2237 EXPORT_SYMBOL(alloc_pages_vma);
2240 * alloc_pages - Allocate pages.
2242 * @order: Power of two of number of pages to allocate.
2244 * Allocate 1 << @order contiguous pages. The physical address of the
2245 * first page is naturally aligned (eg an order-3 allocation will be aligned
2246 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2247 * process is honoured when in process context.
2249 * Context: Can be called from any context, providing the appropriate GFP
2251 * Return: The page on success or NULL if allocation fails.
2253 struct page *alloc_pages(gfp_t gfp, unsigned order)
2255 struct mempolicy *pol = &default_policy;
2258 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2259 pol = get_task_policy(current);
2262 * No reference counting needed for current->mempolicy
2263 * nor system default_policy
2265 if (pol->mode == MPOL_INTERLEAVE)
2266 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2267 else if (pol->mode == MPOL_PREFERRED_MANY)
2268 page = alloc_pages_preferred_many(gfp, order,
2269 policy_node(gfp, pol, numa_node_id()), pol);
2271 page = __alloc_pages(gfp, order,
2272 policy_node(gfp, pol, numa_node_id()),
2273 policy_nodemask(gfp, pol));
2277 EXPORT_SYMBOL(alloc_pages);
2279 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2281 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2283 if (page && order > 1)
2284 prep_transhuge_page(page);
2285 return (struct folio *)page;
2287 EXPORT_SYMBOL(folio_alloc);
2289 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2290 struct mempolicy *pol, unsigned long nr_pages,
2291 struct page **page_array)
2294 unsigned long nr_pages_per_node;
2297 unsigned long nr_allocated;
2298 unsigned long total_allocated = 0;
2300 nodes = nodes_weight(pol->nodes);
2301 nr_pages_per_node = nr_pages / nodes;
2302 delta = nr_pages - nodes * nr_pages_per_node;
2304 for (i = 0; i < nodes; i++) {
2306 nr_allocated = __alloc_pages_bulk(gfp,
2307 interleave_nodes(pol), NULL,
2308 nr_pages_per_node + 1, NULL,
2312 nr_allocated = __alloc_pages_bulk(gfp,
2313 interleave_nodes(pol), NULL,
2314 nr_pages_per_node, NULL, page_array);
2317 page_array += nr_allocated;
2318 total_allocated += nr_allocated;
2321 return total_allocated;
2324 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2325 struct mempolicy *pol, unsigned long nr_pages,
2326 struct page **page_array)
2328 gfp_t preferred_gfp;
2329 unsigned long nr_allocated = 0;
2331 preferred_gfp = gfp | __GFP_NOWARN;
2332 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2334 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2335 nr_pages, NULL, page_array);
2337 if (nr_allocated < nr_pages)
2338 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2339 nr_pages - nr_allocated, NULL,
2340 page_array + nr_allocated);
2341 return nr_allocated;
2344 /* alloc pages bulk and mempolicy should be considered at the
2345 * same time in some situation such as vmalloc.
2347 * It can accelerate memory allocation especially interleaving
2350 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2351 unsigned long nr_pages, struct page **page_array)
2353 struct mempolicy *pol = &default_policy;
2355 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2356 pol = get_task_policy(current);
2358 if (pol->mode == MPOL_INTERLEAVE)
2359 return alloc_pages_bulk_array_interleave(gfp, pol,
2360 nr_pages, page_array);
2362 if (pol->mode == MPOL_PREFERRED_MANY)
2363 return alloc_pages_bulk_array_preferred_many(gfp,
2364 numa_node_id(), pol, nr_pages, page_array);
2366 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2367 policy_nodemask(gfp, pol), nr_pages, NULL,
2371 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2373 struct mempolicy *pol = mpol_dup(vma_policy(src));
2376 return PTR_ERR(pol);
2377 dst->vm_policy = pol;
2382 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2383 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2384 * with the mems_allowed returned by cpuset_mems_allowed(). This
2385 * keeps mempolicies cpuset relative after its cpuset moves. See
2386 * further kernel/cpuset.c update_nodemask().
2388 * current's mempolicy may be rebinded by the other task(the task that changes
2389 * cpuset's mems), so we needn't do rebind work for current task.
2392 /* Slow path of a mempolicy duplicate */
2393 struct mempolicy *__mpol_dup(struct mempolicy *old)
2395 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2398 return ERR_PTR(-ENOMEM);
2400 /* task's mempolicy is protected by alloc_lock */
2401 if (old == current->mempolicy) {
2404 task_unlock(current);
2408 if (current_cpuset_is_being_rebound()) {
2409 nodemask_t mems = cpuset_mems_allowed(current);
2410 mpol_rebind_policy(new, &mems);
2412 atomic_set(&new->refcnt, 1);
2416 /* Slow path of a mempolicy comparison */
2417 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2421 if (a->mode != b->mode)
2423 if (a->flags != b->flags)
2425 if (a->home_node != b->home_node)
2427 if (mpol_store_user_nodemask(a))
2428 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2433 case MPOL_INTERLEAVE:
2434 case MPOL_PREFERRED:
2435 case MPOL_PREFERRED_MANY:
2436 return !!nodes_equal(a->nodes, b->nodes);
2446 * Shared memory backing store policy support.
2448 * Remember policies even when nobody has shared memory mapped.
2449 * The policies are kept in Red-Black tree linked from the inode.
2450 * They are protected by the sp->lock rwlock, which should be held
2451 * for any accesses to the tree.
2455 * lookup first element intersecting start-end. Caller holds sp->lock for
2456 * reading or for writing
2458 static struct sp_node *
2459 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2461 struct rb_node *n = sp->root.rb_node;
2464 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2466 if (start >= p->end)
2468 else if (end <= p->start)
2476 struct sp_node *w = NULL;
2477 struct rb_node *prev = rb_prev(n);
2480 w = rb_entry(prev, struct sp_node, nd);
2481 if (w->end <= start)
2485 return rb_entry(n, struct sp_node, nd);
2489 * Insert a new shared policy into the list. Caller holds sp->lock for
2492 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2494 struct rb_node **p = &sp->root.rb_node;
2495 struct rb_node *parent = NULL;
2500 nd = rb_entry(parent, struct sp_node, nd);
2501 if (new->start < nd->start)
2503 else if (new->end > nd->end)
2504 p = &(*p)->rb_right;
2508 rb_link_node(&new->nd, parent, p);
2509 rb_insert_color(&new->nd, &sp->root);
2510 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2511 new->policy ? new->policy->mode : 0);
2514 /* Find shared policy intersecting idx */
2516 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2518 struct mempolicy *pol = NULL;
2521 if (!sp->root.rb_node)
2523 read_lock(&sp->lock);
2524 sn = sp_lookup(sp, idx, idx+1);
2526 mpol_get(sn->policy);
2529 read_unlock(&sp->lock);
2533 static void sp_free(struct sp_node *n)
2535 mpol_put(n->policy);
2536 kmem_cache_free(sn_cache, n);
2540 * mpol_misplaced - check whether current page node is valid in policy
2542 * @page: page to be checked
2543 * @vma: vm area where page mapped
2544 * @addr: virtual address where page mapped
2546 * Lookup current policy node id for vma,addr and "compare to" page's
2547 * node id. Policy determination "mimics" alloc_page_vma().
2548 * Called from fault path where we know the vma and faulting address.
2550 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2551 * policy, or a suitable node ID to allocate a replacement page from.
2553 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2555 struct mempolicy *pol;
2557 int curnid = page_to_nid(page);
2558 unsigned long pgoff;
2559 int thiscpu = raw_smp_processor_id();
2560 int thisnid = cpu_to_node(thiscpu);
2561 int polnid = NUMA_NO_NODE;
2562 int ret = NUMA_NO_NODE;
2564 pol = get_vma_policy(vma, addr);
2565 if (!(pol->flags & MPOL_F_MOF))
2568 switch (pol->mode) {
2569 case MPOL_INTERLEAVE:
2570 pgoff = vma->vm_pgoff;
2571 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2572 polnid = offset_il_node(pol, pgoff);
2575 case MPOL_PREFERRED:
2576 if (node_isset(curnid, pol->nodes))
2578 polnid = first_node(pol->nodes);
2582 polnid = numa_node_id();
2586 /* Optimize placement among multiple nodes via NUMA balancing */
2587 if (pol->flags & MPOL_F_MORON) {
2588 if (node_isset(thisnid, pol->nodes))
2594 case MPOL_PREFERRED_MANY:
2596 * use current page if in policy nodemask,
2597 * else select nearest allowed node, if any.
2598 * If no allowed nodes, use current [!misplaced].
2600 if (node_isset(curnid, pol->nodes))
2602 z = first_zones_zonelist(
2603 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2604 gfp_zone(GFP_HIGHUSER),
2606 polnid = zone_to_nid(z->zone);
2613 /* Migrate the page towards the node whose CPU is referencing it */
2614 if (pol->flags & MPOL_F_MORON) {
2617 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2621 if (curnid != polnid)
2630 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2631 * dropped after task->mempolicy is set to NULL so that any allocation done as
2632 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2635 void mpol_put_task_policy(struct task_struct *task)
2637 struct mempolicy *pol;
2640 pol = task->mempolicy;
2641 task->mempolicy = NULL;
2646 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2648 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2649 rb_erase(&n->nd, &sp->root);
2653 static void sp_node_init(struct sp_node *node, unsigned long start,
2654 unsigned long end, struct mempolicy *pol)
2656 node->start = start;
2661 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2662 struct mempolicy *pol)
2665 struct mempolicy *newpol;
2667 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2671 newpol = mpol_dup(pol);
2672 if (IS_ERR(newpol)) {
2673 kmem_cache_free(sn_cache, n);
2676 newpol->flags |= MPOL_F_SHARED;
2677 sp_node_init(n, start, end, newpol);
2682 /* Replace a policy range. */
2683 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2684 unsigned long end, struct sp_node *new)
2687 struct sp_node *n_new = NULL;
2688 struct mempolicy *mpol_new = NULL;
2692 write_lock(&sp->lock);
2693 n = sp_lookup(sp, start, end);
2694 /* Take care of old policies in the same range. */
2695 while (n && n->start < end) {
2696 struct rb_node *next = rb_next(&n->nd);
2697 if (n->start >= start) {
2703 /* Old policy spanning whole new range. */
2708 *mpol_new = *n->policy;
2709 atomic_set(&mpol_new->refcnt, 1);
2710 sp_node_init(n_new, end, n->end, mpol_new);
2712 sp_insert(sp, n_new);
2721 n = rb_entry(next, struct sp_node, nd);
2725 write_unlock(&sp->lock);
2732 kmem_cache_free(sn_cache, n_new);
2737 write_unlock(&sp->lock);
2739 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2742 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2749 * mpol_shared_policy_init - initialize shared policy for inode
2750 * @sp: pointer to inode shared policy
2751 * @mpol: struct mempolicy to install
2753 * Install non-NULL @mpol in inode's shared policy rb-tree.
2754 * On entry, the current task has a reference on a non-NULL @mpol.
2755 * This must be released on exit.
2756 * This is called at get_inode() calls and we can use GFP_KERNEL.
2758 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2762 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2763 rwlock_init(&sp->lock);
2766 struct vm_area_struct pvma;
2767 struct mempolicy *new;
2768 NODEMASK_SCRATCH(scratch);
2772 /* contextualize the tmpfs mount point mempolicy */
2773 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2775 goto free_scratch; /* no valid nodemask intersection */
2778 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2779 task_unlock(current);
2783 /* Create pseudo-vma that contains just the policy */
2784 vma_init(&pvma, NULL);
2785 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2786 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2789 mpol_put(new); /* drop initial ref */
2791 NODEMASK_SCRATCH_FREE(scratch);
2793 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2797 int mpol_set_shared_policy(struct shared_policy *info,
2798 struct vm_area_struct *vma, struct mempolicy *npol)
2801 struct sp_node *new = NULL;
2802 unsigned long sz = vma_pages(vma);
2804 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2806 sz, npol ? npol->mode : -1,
2807 npol ? npol->flags : -1,
2808 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2811 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2815 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2821 /* Free a backing policy store on inode delete. */
2822 void mpol_free_shared_policy(struct shared_policy *p)
2825 struct rb_node *next;
2827 if (!p->root.rb_node)
2829 write_lock(&p->lock);
2830 next = rb_first(&p->root);
2832 n = rb_entry(next, struct sp_node, nd);
2833 next = rb_next(&n->nd);
2836 write_unlock(&p->lock);
2839 #ifdef CONFIG_NUMA_BALANCING
2840 static int __initdata numabalancing_override;
2842 static void __init check_numabalancing_enable(void)
2844 bool numabalancing_default = false;
2846 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2847 numabalancing_default = true;
2849 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2850 if (numabalancing_override)
2851 set_numabalancing_state(numabalancing_override == 1);
2853 if (num_online_nodes() > 1 && !numabalancing_override) {
2854 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2855 numabalancing_default ? "Enabling" : "Disabling");
2856 set_numabalancing_state(numabalancing_default);
2860 static int __init setup_numabalancing(char *str)
2866 if (!strcmp(str, "enable")) {
2867 numabalancing_override = 1;
2869 } else if (!strcmp(str, "disable")) {
2870 numabalancing_override = -1;
2875 pr_warn("Unable to parse numa_balancing=\n");
2879 __setup("numa_balancing=", setup_numabalancing);
2881 static inline void __init check_numabalancing_enable(void)
2884 #endif /* CONFIG_NUMA_BALANCING */
2886 /* assumes fs == KERNEL_DS */
2887 void __init numa_policy_init(void)
2889 nodemask_t interleave_nodes;
2890 unsigned long largest = 0;
2891 int nid, prefer = 0;
2893 policy_cache = kmem_cache_create("numa_policy",
2894 sizeof(struct mempolicy),
2895 0, SLAB_PANIC, NULL);
2897 sn_cache = kmem_cache_create("shared_policy_node",
2898 sizeof(struct sp_node),
2899 0, SLAB_PANIC, NULL);
2901 for_each_node(nid) {
2902 preferred_node_policy[nid] = (struct mempolicy) {
2903 .refcnt = ATOMIC_INIT(1),
2904 .mode = MPOL_PREFERRED,
2905 .flags = MPOL_F_MOF | MPOL_F_MORON,
2906 .nodes = nodemask_of_node(nid),
2911 * Set interleaving policy for system init. Interleaving is only
2912 * enabled across suitably sized nodes (default is >= 16MB), or
2913 * fall back to the largest node if they're all smaller.
2915 nodes_clear(interleave_nodes);
2916 for_each_node_state(nid, N_MEMORY) {
2917 unsigned long total_pages = node_present_pages(nid);
2919 /* Preserve the largest node */
2920 if (largest < total_pages) {
2921 largest = total_pages;
2925 /* Interleave this node? */
2926 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2927 node_set(nid, interleave_nodes);
2930 /* All too small, use the largest */
2931 if (unlikely(nodes_empty(interleave_nodes)))
2932 node_set(prefer, interleave_nodes);
2934 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2935 pr_err("%s: interleaving failed\n", __func__);
2937 check_numabalancing_enable();
2940 /* Reset policy of current process to default */
2941 void numa_default_policy(void)
2943 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2947 * Parse and format mempolicy from/to strings
2950 static const char * const policy_modes[] =
2952 [MPOL_DEFAULT] = "default",
2953 [MPOL_PREFERRED] = "prefer",
2954 [MPOL_BIND] = "bind",
2955 [MPOL_INTERLEAVE] = "interleave",
2956 [MPOL_LOCAL] = "local",
2957 [MPOL_PREFERRED_MANY] = "prefer (many)",
2963 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2964 * @str: string containing mempolicy to parse
2965 * @mpol: pointer to struct mempolicy pointer, returned on success.
2968 * <mode>[=<flags>][:<nodelist>]
2970 * Return: %0 on success, else %1
2972 int mpol_parse_str(char *str, struct mempolicy **mpol)
2974 struct mempolicy *new = NULL;
2975 unsigned short mode_flags;
2977 char *nodelist = strchr(str, ':');
2978 char *flags = strchr(str, '=');
2982 *flags++ = '\0'; /* terminate mode string */
2985 /* NUL-terminate mode or flags string */
2987 if (nodelist_parse(nodelist, nodes))
2989 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2994 mode = match_string(policy_modes, MPOL_MAX, str);
2999 case MPOL_PREFERRED:
3001 * Insist on a nodelist of one node only, although later
3002 * we use first_node(nodes) to grab a single node, so here
3003 * nodelist (or nodes) cannot be empty.
3006 char *rest = nodelist;
3007 while (isdigit(*rest))
3011 if (nodes_empty(nodes))
3015 case MPOL_INTERLEAVE:
3017 * Default to online nodes with memory if no nodelist
3020 nodes = node_states[N_MEMORY];
3024 * Don't allow a nodelist; mpol_new() checks flags
3031 * Insist on a empty nodelist
3036 case MPOL_PREFERRED_MANY:
3039 * Insist on a nodelist
3048 * Currently, we only support two mutually exclusive
3051 if (!strcmp(flags, "static"))
3052 mode_flags |= MPOL_F_STATIC_NODES;
3053 else if (!strcmp(flags, "relative"))
3054 mode_flags |= MPOL_F_RELATIVE_NODES;
3059 new = mpol_new(mode, mode_flags, &nodes);
3064 * Save nodes for mpol_to_str() to show the tmpfs mount options
3065 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3067 if (mode != MPOL_PREFERRED) {
3069 } else if (nodelist) {
3070 nodes_clear(new->nodes);
3071 node_set(first_node(nodes), new->nodes);
3073 new->mode = MPOL_LOCAL;
3077 * Save nodes for contextualization: this will be used to "clone"
3078 * the mempolicy in a specific context [cpuset] at a later time.
3080 new->w.user_nodemask = nodes;
3085 /* Restore string for error message */
3094 #endif /* CONFIG_TMPFS */
3097 * mpol_to_str - format a mempolicy structure for printing
3098 * @buffer: to contain formatted mempolicy string
3099 * @maxlen: length of @buffer
3100 * @pol: pointer to mempolicy to be formatted
3102 * Convert @pol into a string. If @buffer is too short, truncate the string.
3103 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3104 * longest flag, "relative", and to display at least a few node ids.
3106 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3109 nodemask_t nodes = NODE_MASK_NONE;
3110 unsigned short mode = MPOL_DEFAULT;
3111 unsigned short flags = 0;
3113 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3122 case MPOL_PREFERRED:
3123 case MPOL_PREFERRED_MANY:
3125 case MPOL_INTERLEAVE:
3130 snprintf(p, maxlen, "unknown");
3134 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3136 if (flags & MPOL_MODE_FLAGS) {
3137 p += snprintf(p, buffer + maxlen - p, "=");
3140 * Currently, the only defined flags are mutually exclusive
3142 if (flags & MPOL_F_STATIC_NODES)
3143 p += snprintf(p, buffer + maxlen - p, "static");
3144 else if (flags & MPOL_F_RELATIVE_NODES)
3145 p += snprintf(p, buffer + maxlen - p, "relative");
3148 if (!nodes_empty(nodes))
3149 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3150 nodemask_pr_args(&nodes));