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 int numa_map_to_online_node(int node)
140 int min_dist = INT_MAX, dist, n, min_node;
142 if (node == NUMA_NO_NODE || node_online(node))
146 for_each_online_node(n) {
147 dist = node_distance(node, n);
148 if (dist < min_dist) {
156 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
158 struct mempolicy *get_task_policy(struct task_struct *p)
160 struct mempolicy *pol = p->mempolicy;
166 node = numa_node_id();
167 if (node != NUMA_NO_NODE) {
168 pol = &preferred_node_policy[node];
169 /* preferred_node_policy is not initialised early in boot */
174 return &default_policy;
177 static const struct mempolicy_operations {
178 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
179 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
180 } mpol_ops[MPOL_MAX];
182 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
184 return pol->flags & MPOL_MODE_FLAGS;
187 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
188 const nodemask_t *rel)
191 nodes_fold(tmp, *orig, nodes_weight(*rel));
192 nodes_onto(*ret, tmp, *rel);
195 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
197 if (nodes_empty(*nodes))
203 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
205 if (nodes_empty(*nodes))
208 nodes_clear(pol->nodes);
209 node_set(first_node(*nodes), pol->nodes);
213 static int mpol_new_preferred_many(struct mempolicy *pol, const nodemask_t *nodes)
215 if (nodes_empty(*nodes))
221 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
223 if (nodes_empty(*nodes))
230 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
231 * any, for the new policy. mpol_new() has already validated the nodes
232 * parameter with respect to the policy mode and flags.
234 * Must be called holding task's alloc_lock to protect task's mems_allowed
235 * and mempolicy. May also be called holding the mmap_lock for write.
237 static int mpol_set_nodemask(struct mempolicy *pol,
238 const nodemask_t *nodes, struct nodemask_scratch *nsc)
243 * Default (pol==NULL) resp. local memory policies are not a
244 * subject of any remapping. They also do not need any special
247 if (!pol || pol->mode == MPOL_LOCAL)
251 nodes_and(nsc->mask1,
252 cpuset_current_mems_allowed, node_states[N_MEMORY]);
256 if (pol->flags & MPOL_F_RELATIVE_NODES)
257 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
259 nodes_and(nsc->mask2, *nodes, nsc->mask1);
261 if (mpol_store_user_nodemask(pol))
262 pol->w.user_nodemask = *nodes;
264 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
266 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
271 * This function just creates a new policy, does some check and simple
272 * initialization. You must invoke mpol_set_nodemask() to set nodes.
274 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
277 struct mempolicy *policy;
279 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
280 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
282 if (mode == MPOL_DEFAULT) {
283 if (nodes && !nodes_empty(*nodes))
284 return ERR_PTR(-EINVAL);
290 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
291 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
292 * All other modes require a valid pointer to a non-empty nodemask.
294 if (mode == MPOL_PREFERRED) {
295 if (nodes_empty(*nodes)) {
296 if (((flags & MPOL_F_STATIC_NODES) ||
297 (flags & MPOL_F_RELATIVE_NODES)))
298 return ERR_PTR(-EINVAL);
302 } else if (mode == MPOL_LOCAL) {
303 if (!nodes_empty(*nodes) ||
304 (flags & MPOL_F_STATIC_NODES) ||
305 (flags & MPOL_F_RELATIVE_NODES))
306 return ERR_PTR(-EINVAL);
307 } else if (nodes_empty(*nodes))
308 return ERR_PTR(-EINVAL);
309 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
311 return ERR_PTR(-ENOMEM);
312 atomic_set(&policy->refcnt, 1);
314 policy->flags = flags;
319 /* Slow path of a mpol destructor. */
320 void __mpol_put(struct mempolicy *p)
322 if (!atomic_dec_and_test(&p->refcnt))
324 kmem_cache_free(policy_cache, p);
327 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
331 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
335 if (pol->flags & MPOL_F_STATIC_NODES)
336 nodes_and(tmp, pol->w.user_nodemask, *nodes);
337 else if (pol->flags & MPOL_F_RELATIVE_NODES)
338 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
340 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
342 pol->w.cpuset_mems_allowed = *nodes;
345 if (nodes_empty(tmp))
351 static void mpol_rebind_preferred(struct mempolicy *pol,
352 const nodemask_t *nodes)
354 pol->w.cpuset_mems_allowed = *nodes;
358 * mpol_rebind_policy - Migrate a policy to a different set of nodes
360 * Per-vma policies are protected by mmap_lock. Allocations using per-task
361 * policies are protected by task->mems_allowed_seq to prevent a premature
362 * OOM/allocation failure due to parallel nodemask modification.
364 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
368 if (!mpol_store_user_nodemask(pol) &&
369 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
372 mpol_ops[pol->mode].rebind(pol, newmask);
376 * Wrapper for mpol_rebind_policy() that just requires task
377 * pointer, and updates task mempolicy.
379 * Called with task's alloc_lock held.
382 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
384 mpol_rebind_policy(tsk->mempolicy, new);
388 * Rebind each vma in mm to new nodemask.
390 * Call holding a reference to mm. Takes mm->mmap_lock during call.
393 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
395 struct vm_area_struct *vma;
398 for (vma = mm->mmap; vma; vma = vma->vm_next)
399 mpol_rebind_policy(vma->vm_policy, new);
400 mmap_write_unlock(mm);
403 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
405 .rebind = mpol_rebind_default,
407 [MPOL_INTERLEAVE] = {
408 .create = mpol_new_interleave,
409 .rebind = mpol_rebind_nodemask,
412 .create = mpol_new_preferred,
413 .rebind = mpol_rebind_preferred,
416 .create = mpol_new_bind,
417 .rebind = mpol_rebind_nodemask,
420 .rebind = mpol_rebind_default,
422 [MPOL_PREFERRED_MANY] = {
423 .create = mpol_new_preferred_many,
424 .rebind = mpol_rebind_preferred,
428 static int migrate_page_add(struct page *page, struct list_head *pagelist,
429 unsigned long flags);
432 struct list_head *pagelist;
437 struct vm_area_struct *first;
441 * Check if the page's nid is in qp->nmask.
443 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
444 * in the invert of qp->nmask.
446 static inline bool queue_pages_required(struct page *page,
447 struct queue_pages *qp)
449 int nid = page_to_nid(page);
450 unsigned long flags = qp->flags;
452 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
456 * queue_pages_pmd() has four possible return values:
457 * 0 - pages are placed on the right node or queued successfully, or
458 * special page is met, i.e. huge zero page.
459 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
462 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
463 * existing page was already on a node that does not follow the
466 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
467 unsigned long end, struct mm_walk *walk)
472 struct queue_pages *qp = walk->private;
475 if (unlikely(is_pmd_migration_entry(*pmd))) {
479 page = pmd_page(*pmd);
480 if (is_huge_zero_page(page)) {
482 walk->action = ACTION_CONTINUE;
485 if (!queue_pages_required(page, qp))
489 /* go to thp migration */
490 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
491 if (!vma_migratable(walk->vma) ||
492 migrate_page_add(page, qp->pagelist, flags)) {
505 * Scan through pages checking if pages follow certain conditions,
506 * and move them to the pagelist if they do.
508 * queue_pages_pte_range() has three possible return values:
509 * 0 - pages are placed on the right node or queued successfully, or
510 * special page is met, i.e. zero page.
511 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
513 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
514 * on a node that does not follow the policy.
516 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
517 unsigned long end, struct mm_walk *walk)
519 struct vm_area_struct *vma = walk->vma;
521 struct queue_pages *qp = walk->private;
522 unsigned long flags = qp->flags;
524 bool has_unmovable = false;
525 pte_t *pte, *mapped_pte;
528 ptl = pmd_trans_huge_lock(pmd, vma);
530 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
534 /* THP was split, fall through to pte walk */
536 if (pmd_trans_unstable(pmd))
539 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
540 for (; addr != end; pte++, addr += PAGE_SIZE) {
541 if (!pte_present(*pte))
543 page = vm_normal_page(vma, addr, *pte);
547 * vm_normal_page() filters out zero pages, but there might
548 * still be PageReserved pages to skip, perhaps in a VDSO.
550 if (PageReserved(page))
552 if (!queue_pages_required(page, qp))
554 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
555 /* MPOL_MF_STRICT must be specified if we get here */
556 if (!vma_migratable(vma)) {
557 has_unmovable = true;
562 * Do not abort immediately since there may be
563 * temporary off LRU pages in the range. Still
564 * need migrate other LRU pages.
566 if (migrate_page_add(page, qp->pagelist, flags))
567 has_unmovable = true;
571 pte_unmap_unlock(mapped_pte, ptl);
577 return addr != end ? -EIO : 0;
580 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
581 unsigned long addr, unsigned long end,
582 struct mm_walk *walk)
585 #ifdef CONFIG_HUGETLB_PAGE
586 struct queue_pages *qp = walk->private;
587 unsigned long flags = (qp->flags & MPOL_MF_VALID);
592 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
593 entry = huge_ptep_get(pte);
594 if (!pte_present(entry))
596 page = pte_page(entry);
597 if (!queue_pages_required(page, qp))
600 if (flags == MPOL_MF_STRICT) {
602 * STRICT alone means only detecting misplaced page and no
603 * need to further check other vma.
609 if (!vma_migratable(walk->vma)) {
611 * Must be STRICT with MOVE*, otherwise .test_walk() have
612 * stopped walking current vma.
613 * Detecting misplaced page but allow migrating pages which
620 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
621 if (flags & (MPOL_MF_MOVE_ALL) ||
622 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
623 if (!isolate_huge_page(page, qp->pagelist) &&
624 (flags & MPOL_MF_STRICT))
626 * Failed to isolate page but allow migrating pages
627 * which have been queued.
639 #ifdef CONFIG_NUMA_BALANCING
641 * This is used to mark a range of virtual addresses to be inaccessible.
642 * These are later cleared by a NUMA hinting fault. Depending on these
643 * faults, pages may be migrated for better NUMA placement.
645 * This is assuming that NUMA faults are handled using PROT_NONE. If
646 * an architecture makes a different choice, it will need further
647 * changes to the core.
649 unsigned long change_prot_numa(struct vm_area_struct *vma,
650 unsigned long addr, unsigned long end)
654 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
656 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
661 static unsigned long change_prot_numa(struct vm_area_struct *vma,
662 unsigned long addr, unsigned long end)
666 #endif /* CONFIG_NUMA_BALANCING */
668 static int queue_pages_test_walk(unsigned long start, unsigned long end,
669 struct mm_walk *walk)
671 struct vm_area_struct *vma = walk->vma;
672 struct queue_pages *qp = walk->private;
673 unsigned long endvma = vma->vm_end;
674 unsigned long flags = qp->flags;
676 /* range check first */
677 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
681 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
682 (qp->start < vma->vm_start))
683 /* hole at head side of range */
686 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
687 ((vma->vm_end < qp->end) &&
688 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
689 /* hole at middle or tail of range */
693 * Need check MPOL_MF_STRICT to return -EIO if possible
694 * regardless of vma_migratable
696 if (!vma_migratable(vma) &&
697 !(flags & MPOL_MF_STRICT))
703 if (flags & MPOL_MF_LAZY) {
704 /* Similar to task_numa_work, skip inaccessible VMAs */
705 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
706 !(vma->vm_flags & VM_MIXEDMAP))
707 change_prot_numa(vma, start, endvma);
711 /* queue pages from current vma */
712 if (flags & MPOL_MF_VALID)
717 static const struct mm_walk_ops queue_pages_walk_ops = {
718 .hugetlb_entry = queue_pages_hugetlb,
719 .pmd_entry = queue_pages_pte_range,
720 .test_walk = queue_pages_test_walk,
724 * Walk through page tables and collect pages to be migrated.
726 * If pages found in a given range are on a set of nodes (determined by
727 * @nodes and @flags,) it's isolated and queued to the pagelist which is
728 * passed via @private.
730 * queue_pages_range() has three possible return values:
731 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
733 * 0 - queue pages successfully or no misplaced page.
734 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
735 * memory range specified by nodemask and maxnode points outside
736 * your accessible address space (-EFAULT)
739 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
740 nodemask_t *nodes, unsigned long flags,
741 struct list_head *pagelist)
744 struct queue_pages qp = {
745 .pagelist = pagelist,
753 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
756 /* whole range in hole */
763 * Apply policy to a single VMA
764 * This must be called with the mmap_lock held for writing.
766 static int vma_replace_policy(struct vm_area_struct *vma,
767 struct mempolicy *pol)
770 struct mempolicy *old;
771 struct mempolicy *new;
773 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
774 vma->vm_start, vma->vm_end, vma->vm_pgoff,
775 vma->vm_ops, vma->vm_file,
776 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
782 if (vma->vm_ops && vma->vm_ops->set_policy) {
783 err = vma->vm_ops->set_policy(vma, new);
788 old = vma->vm_policy;
789 vma->vm_policy = new; /* protected by mmap_lock */
798 /* Step 2: apply policy to a range and do splits. */
799 static int mbind_range(struct mm_struct *mm, unsigned long start,
800 unsigned long end, struct mempolicy *new_pol)
802 struct vm_area_struct *next;
803 struct vm_area_struct *prev;
804 struct vm_area_struct *vma;
807 unsigned long vmstart;
810 vma = find_vma(mm, start);
814 if (start > vma->vm_start)
817 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
819 vmstart = max(start, vma->vm_start);
820 vmend = min(end, vma->vm_end);
822 if (mpol_equal(vma_policy(vma), new_pol))
825 pgoff = vma->vm_pgoff +
826 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
827 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
828 vma->anon_vma, vma->vm_file, pgoff,
829 new_pol, vma->vm_userfaultfd_ctx);
833 if (mpol_equal(vma_policy(vma), new_pol))
835 /* vma_merge() joined vma && vma->next, case 8 */
838 if (vma->vm_start != vmstart) {
839 err = split_vma(vma->vm_mm, vma, vmstart, 1);
843 if (vma->vm_end != vmend) {
844 err = split_vma(vma->vm_mm, vma, vmend, 0);
849 err = vma_replace_policy(vma, new_pol);
858 /* Set the process memory policy */
859 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
862 struct mempolicy *new, *old;
863 NODEMASK_SCRATCH(scratch);
869 new = mpol_new(mode, flags, nodes);
875 if (flags & MPOL_F_NUMA_BALANCING) {
876 if (new && new->mode == MPOL_BIND) {
877 new->flags |= (MPOL_F_MOF | MPOL_F_MORON);
885 ret = mpol_set_nodemask(new, nodes, scratch);
891 old = current->mempolicy;
892 current->mempolicy = new;
893 if (new && new->mode == MPOL_INTERLEAVE)
894 current->il_prev = MAX_NUMNODES-1;
895 task_unlock(current);
899 NODEMASK_SCRATCH_FREE(scratch);
904 * Return nodemask for policy for get_mempolicy() query
906 * Called with task's alloc_lock held
908 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
911 if (p == &default_policy)
916 case MPOL_INTERLEAVE:
918 case MPOL_PREFERRED_MANY:
922 /* return empty node mask for local allocation */
929 static int lookup_node(struct mm_struct *mm, unsigned long addr)
931 struct page *p = NULL;
935 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
937 err = page_to_nid(p);
941 mmap_read_unlock(mm);
945 /* Retrieve NUMA policy */
946 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
947 unsigned long addr, unsigned long flags)
950 struct mm_struct *mm = current->mm;
951 struct vm_area_struct *vma = NULL;
952 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
955 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
958 if (flags & MPOL_F_MEMS_ALLOWED) {
959 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
961 *policy = 0; /* just so it's initialized */
963 *nmask = cpuset_current_mems_allowed;
964 task_unlock(current);
968 if (flags & MPOL_F_ADDR) {
970 * Do NOT fall back to task policy if the
971 * vma/shared policy at addr is NULL. We
972 * want to return MPOL_DEFAULT in this case.
975 vma = vma_lookup(mm, addr);
977 mmap_read_unlock(mm);
980 if (vma->vm_ops && vma->vm_ops->get_policy)
981 pol = vma->vm_ops->get_policy(vma, addr);
983 pol = vma->vm_policy;
988 pol = &default_policy; /* indicates default behavior */
990 if (flags & MPOL_F_NODE) {
991 if (flags & MPOL_F_ADDR) {
993 * Take a refcount on the mpol, lookup_node()
994 * will drop the mmap_lock, so after calling
995 * lookup_node() only "pol" remains valid, "vma"
1001 err = lookup_node(mm, addr);
1005 } else if (pol == current->mempolicy &&
1006 pol->mode == MPOL_INTERLEAVE) {
1007 *policy = next_node_in(current->il_prev, pol->nodes);
1013 *policy = pol == &default_policy ? MPOL_DEFAULT :
1016 * Internal mempolicy flags must be masked off before exposing
1017 * the policy to userspace.
1019 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1024 if (mpol_store_user_nodemask(pol)) {
1025 *nmask = pol->w.user_nodemask;
1028 get_policy_nodemask(pol, nmask);
1029 task_unlock(current);
1036 mmap_read_unlock(mm);
1038 mpol_put(pol_refcount);
1042 #ifdef CONFIG_MIGRATION
1044 * page migration, thp tail pages can be passed.
1046 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1047 unsigned long flags)
1049 struct page *head = compound_head(page);
1051 * Avoid migrating a page that is shared with others.
1053 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1054 if (!isolate_lru_page(head)) {
1055 list_add_tail(&head->lru, pagelist);
1056 mod_node_page_state(page_pgdat(head),
1057 NR_ISOLATED_ANON + page_is_file_lru(head),
1058 thp_nr_pages(head));
1059 } else if (flags & MPOL_MF_STRICT) {
1061 * Non-movable page may reach here. And, there may be
1062 * temporary off LRU pages or non-LRU movable pages.
1063 * Treat them as unmovable pages since they can't be
1064 * isolated, so they can't be moved at the moment. It
1065 * should return -EIO for this case too.
1075 * Migrate pages from one node to a target node.
1076 * Returns error or the number of pages not migrated.
1078 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1082 LIST_HEAD(pagelist);
1084 struct migration_target_control mtc = {
1086 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1090 node_set(source, nmask);
1093 * This does not "check" the range but isolates all pages that
1094 * need migration. Between passing in the full user address
1095 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1097 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1098 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1099 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1101 if (!list_empty(&pagelist)) {
1102 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1103 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1105 putback_movable_pages(&pagelist);
1112 * Move pages between the two nodesets so as to preserve the physical
1113 * layout as much as possible.
1115 * Returns the number of page that could not be moved.
1117 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1118 const nodemask_t *to, int flags)
1124 lru_cache_disable();
1129 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1130 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1131 * bit in 'tmp', and return that <source, dest> pair for migration.
1132 * The pair of nodemasks 'to' and 'from' define the map.
1134 * If no pair of bits is found that way, fallback to picking some
1135 * pair of 'source' and 'dest' bits that are not the same. If the
1136 * 'source' and 'dest' bits are the same, this represents a node
1137 * that will be migrating to itself, so no pages need move.
1139 * If no bits are left in 'tmp', or if all remaining bits left
1140 * in 'tmp' correspond to the same bit in 'to', return false
1141 * (nothing left to migrate).
1143 * This lets us pick a pair of nodes to migrate between, such that
1144 * if possible the dest node is not already occupied by some other
1145 * source node, minimizing the risk of overloading the memory on a
1146 * node that would happen if we migrated incoming memory to a node
1147 * before migrating outgoing memory source that same node.
1149 * A single scan of tmp is sufficient. As we go, we remember the
1150 * most recent <s, d> pair that moved (s != d). If we find a pair
1151 * that not only moved, but what's better, moved to an empty slot
1152 * (d is not set in tmp), then we break out then, with that pair.
1153 * Otherwise when we finish scanning from_tmp, we at least have the
1154 * most recent <s, d> pair that moved. If we get all the way through
1155 * the scan of tmp without finding any node that moved, much less
1156 * moved to an empty node, then there is nothing left worth migrating.
1160 while (!nodes_empty(tmp)) {
1162 int source = NUMA_NO_NODE;
1165 for_each_node_mask(s, tmp) {
1168 * do_migrate_pages() tries to maintain the relative
1169 * node relationship of the pages established between
1170 * threads and memory areas.
1172 * However if the number of source nodes is not equal to
1173 * the number of destination nodes we can not preserve
1174 * this node relative relationship. In that case, skip
1175 * copying memory from a node that is in the destination
1178 * Example: [2,3,4] -> [3,4,5] moves everything.
1179 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1182 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1183 (node_isset(s, *to)))
1186 d = node_remap(s, *from, *to);
1190 source = s; /* Node moved. Memorize */
1193 /* dest not in remaining from nodes? */
1194 if (!node_isset(dest, tmp))
1197 if (source == NUMA_NO_NODE)
1200 node_clear(source, tmp);
1201 err = migrate_to_node(mm, source, dest, flags);
1207 mmap_read_unlock(mm);
1217 * Allocate a new page for page migration based on vma policy.
1218 * Start by assuming the page is mapped by the same vma as contains @start.
1219 * Search forward from there, if not. N.B., this assumes that the
1220 * list of pages handed to migrate_pages()--which is how we get here--
1221 * is in virtual address order.
1223 static struct page *new_page(struct page *page, unsigned long start)
1225 struct vm_area_struct *vma;
1226 unsigned long address;
1228 vma = find_vma(current->mm, start);
1230 address = page_address_in_vma(page, vma);
1231 if (address != -EFAULT)
1236 if (PageHuge(page)) {
1237 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1239 } else if (PageTransHuge(page)) {
1242 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1246 prep_transhuge_page(thp);
1250 * if !vma, alloc_page_vma() will use task or system default policy
1252 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1257 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1258 unsigned long flags)
1263 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1264 const nodemask_t *to, int flags)
1269 static struct page *new_page(struct page *page, unsigned long start)
1275 static long do_mbind(unsigned long start, unsigned long len,
1276 unsigned short mode, unsigned short mode_flags,
1277 nodemask_t *nmask, unsigned long flags)
1279 struct mm_struct *mm = current->mm;
1280 struct mempolicy *new;
1284 LIST_HEAD(pagelist);
1286 if (flags & ~(unsigned long)MPOL_MF_VALID)
1288 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1291 if (start & ~PAGE_MASK)
1294 if (mode == MPOL_DEFAULT)
1295 flags &= ~MPOL_MF_STRICT;
1297 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1305 new = mpol_new(mode, mode_flags, nmask);
1307 return PTR_ERR(new);
1309 if (flags & MPOL_MF_LAZY)
1310 new->flags |= MPOL_F_MOF;
1313 * If we are using the default policy then operation
1314 * on discontinuous address spaces is okay after all
1317 flags |= MPOL_MF_DISCONTIG_OK;
1319 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1320 start, start + len, mode, mode_flags,
1321 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1323 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1325 lru_cache_disable();
1328 NODEMASK_SCRATCH(scratch);
1330 mmap_write_lock(mm);
1331 err = mpol_set_nodemask(new, nmask, scratch);
1333 mmap_write_unlock(mm);
1336 NODEMASK_SCRATCH_FREE(scratch);
1341 ret = queue_pages_range(mm, start, end, nmask,
1342 flags | MPOL_MF_INVERT, &pagelist);
1349 err = mbind_range(mm, start, end, new);
1354 if (!list_empty(&pagelist)) {
1355 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1356 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1357 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1359 putback_movable_pages(&pagelist);
1362 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1366 if (!list_empty(&pagelist))
1367 putback_movable_pages(&pagelist);
1370 mmap_write_unlock(mm);
1373 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1379 * User space interface with variable sized bitmaps for nodelists.
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)
1388 unsigned long nlongs;
1389 unsigned long endmask;
1392 nodes_clear(*nodes);
1393 if (maxnode == 0 || !nmask)
1395 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1398 nlongs = BITS_TO_LONGS(maxnode);
1399 if ((maxnode % BITS_PER_LONG) == 0)
1402 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1405 * When the user specified more nodes than supported just check
1406 * if the non supported part is all zero.
1408 * If maxnode have more longs than MAX_NUMNODES, check
1409 * the bits in that area first. And then go through to
1410 * check the rest bits which equal or bigger than MAX_NUMNODES.
1411 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1413 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1414 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1415 if (get_user(t, nmask + k))
1417 if (k == nlongs - 1) {
1423 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1427 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1428 unsigned long valid_mask = endmask;
1430 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1431 if (get_user(t, nmask + nlongs - 1))
1437 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1439 nodes_addr(*nodes)[nlongs-1] &= endmask;
1443 /* Copy a kernel node mask to user space */
1444 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1447 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1448 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1450 if (copy > nbytes) {
1451 if (copy > PAGE_SIZE)
1453 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1457 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1460 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1461 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1463 *flags = *mode & MPOL_MODE_FLAGS;
1464 *mode &= ~MPOL_MODE_FLAGS;
1467 * The check should be 'mode >= MPOL_MAX', but as 'prefer_many'
1468 * is not fully implemented, don't permit it to be used for now,
1469 * and the logic will be restored in following patch
1471 if ((unsigned int)(*mode) >= MPOL_PREFERRED_MANY)
1473 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1479 static long kernel_mbind(unsigned long start, unsigned long len,
1480 unsigned long mode, const unsigned long __user *nmask,
1481 unsigned long maxnode, unsigned int flags)
1483 unsigned short mode_flags;
1488 start = untagged_addr(start);
1489 err = sanitize_mpol_flags(&lmode, &mode_flags);
1493 err = get_nodes(&nodes, nmask, maxnode);
1497 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1500 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1501 unsigned long, mode, const unsigned long __user *, nmask,
1502 unsigned long, maxnode, unsigned int, flags)
1504 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1507 /* Set the process memory policy */
1508 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1509 unsigned long maxnode)
1511 unsigned short mode_flags;
1516 err = sanitize_mpol_flags(&lmode, &mode_flags);
1520 err = get_nodes(&nodes, nmask, maxnode);
1524 return do_set_mempolicy(lmode, mode_flags, &nodes);
1527 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1528 unsigned long, maxnode)
1530 return kernel_set_mempolicy(mode, nmask, maxnode);
1533 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1534 const unsigned long __user *old_nodes,
1535 const unsigned long __user *new_nodes)
1537 struct mm_struct *mm = NULL;
1538 struct task_struct *task;
1539 nodemask_t task_nodes;
1543 NODEMASK_SCRATCH(scratch);
1548 old = &scratch->mask1;
1549 new = &scratch->mask2;
1551 err = get_nodes(old, old_nodes, maxnode);
1555 err = get_nodes(new, new_nodes, maxnode);
1559 /* Find the mm_struct */
1561 task = pid ? find_task_by_vpid(pid) : current;
1567 get_task_struct(task);
1572 * Check if this process has the right to modify the specified process.
1573 * Use the regular "ptrace_may_access()" checks.
1575 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1582 task_nodes = cpuset_mems_allowed(task);
1583 /* Is the user allowed to access the target nodes? */
1584 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1589 task_nodes = cpuset_mems_allowed(current);
1590 nodes_and(*new, *new, task_nodes);
1591 if (nodes_empty(*new))
1594 err = security_task_movememory(task);
1598 mm = get_task_mm(task);
1599 put_task_struct(task);
1606 err = do_migrate_pages(mm, old, new,
1607 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1611 NODEMASK_SCRATCH_FREE(scratch);
1616 put_task_struct(task);
1621 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1622 const unsigned long __user *, old_nodes,
1623 const unsigned long __user *, new_nodes)
1625 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1629 /* Retrieve NUMA policy */
1630 static int kernel_get_mempolicy(int __user *policy,
1631 unsigned long __user *nmask,
1632 unsigned long maxnode,
1634 unsigned long flags)
1640 if (nmask != NULL && maxnode < nr_node_ids)
1643 addr = untagged_addr(addr);
1645 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1650 if (policy && put_user(pval, policy))
1654 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1659 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1660 unsigned long __user *, nmask, unsigned long, maxnode,
1661 unsigned long, addr, unsigned long, flags)
1663 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1666 #ifdef CONFIG_COMPAT
1668 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1669 compat_ulong_t __user *, nmask,
1670 compat_ulong_t, maxnode,
1671 compat_ulong_t, addr, compat_ulong_t, flags)
1674 unsigned long __user *nm = NULL;
1675 unsigned long nr_bits, alloc_size;
1676 DECLARE_BITMAP(bm, MAX_NUMNODES);
1678 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1679 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1682 nm = compat_alloc_user_space(alloc_size);
1684 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1686 if (!err && nmask) {
1687 unsigned long copy_size;
1688 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1689 err = copy_from_user(bm, nm, copy_size);
1690 /* ensure entire bitmap is zeroed */
1691 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1692 err |= compat_put_bitmap(nmask, bm, nr_bits);
1698 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1699 compat_ulong_t, maxnode)
1701 unsigned long __user *nm = NULL;
1702 unsigned long nr_bits, alloc_size;
1703 DECLARE_BITMAP(bm, MAX_NUMNODES);
1705 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1706 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1709 if (compat_get_bitmap(bm, nmask, nr_bits))
1711 nm = compat_alloc_user_space(alloc_size);
1712 if (copy_to_user(nm, bm, alloc_size))
1716 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1719 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1720 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1721 compat_ulong_t, maxnode, compat_ulong_t, flags)
1723 unsigned long __user *nm = NULL;
1724 unsigned long nr_bits, alloc_size;
1727 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1728 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1731 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1733 nm = compat_alloc_user_space(alloc_size);
1734 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1738 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1741 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1742 compat_ulong_t, maxnode,
1743 const compat_ulong_t __user *, old_nodes,
1744 const compat_ulong_t __user *, new_nodes)
1746 unsigned long __user *old = NULL;
1747 unsigned long __user *new = NULL;
1748 nodemask_t tmp_mask;
1749 unsigned long nr_bits;
1752 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1753 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1755 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1757 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1759 new = old + size / sizeof(unsigned long);
1760 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1764 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1767 new = compat_alloc_user_space(size);
1768 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1771 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1774 #endif /* CONFIG_COMPAT */
1776 bool vma_migratable(struct vm_area_struct *vma)
1778 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1782 * DAX device mappings require predictable access latency, so avoid
1783 * incurring periodic faults.
1785 if (vma_is_dax(vma))
1788 if (is_vm_hugetlb_page(vma) &&
1789 !hugepage_migration_supported(hstate_vma(vma)))
1793 * Migration allocates pages in the highest zone. If we cannot
1794 * do so then migration (at least from node to node) is not
1798 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1804 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1807 struct mempolicy *pol = NULL;
1810 if (vma->vm_ops && vma->vm_ops->get_policy) {
1811 pol = vma->vm_ops->get_policy(vma, addr);
1812 } else if (vma->vm_policy) {
1813 pol = vma->vm_policy;
1816 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1817 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1818 * count on these policies which will be dropped by
1819 * mpol_cond_put() later
1821 if (mpol_needs_cond_ref(pol))
1830 * get_vma_policy(@vma, @addr)
1831 * @vma: virtual memory area whose policy is sought
1832 * @addr: address in @vma for shared policy lookup
1834 * Returns effective policy for a VMA at specified address.
1835 * Falls back to current->mempolicy or system default policy, as necessary.
1836 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1837 * count--added by the get_policy() vm_op, as appropriate--to protect against
1838 * freeing by another task. It is the caller's responsibility to free the
1839 * extra reference for shared policies.
1841 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1844 struct mempolicy *pol = __get_vma_policy(vma, addr);
1847 pol = get_task_policy(current);
1852 bool vma_policy_mof(struct vm_area_struct *vma)
1854 struct mempolicy *pol;
1856 if (vma->vm_ops && vma->vm_ops->get_policy) {
1859 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1860 if (pol && (pol->flags & MPOL_F_MOF))
1867 pol = vma->vm_policy;
1869 pol = get_task_policy(current);
1871 return pol->flags & MPOL_F_MOF;
1874 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1876 enum zone_type dynamic_policy_zone = policy_zone;
1878 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1881 * if policy->nodes has movable memory only,
1882 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1884 * policy->nodes is intersect with node_states[N_MEMORY].
1885 * so if the following test fails, it implies
1886 * policy->nodes has movable memory only.
1888 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1889 dynamic_policy_zone = ZONE_MOVABLE;
1891 return zone >= dynamic_policy_zone;
1895 * Return a nodemask representing a mempolicy for filtering nodes for
1898 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1900 int mode = policy->mode;
1902 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1903 if (unlikely(mode == MPOL_BIND) &&
1904 apply_policy_zone(policy, gfp_zone(gfp)) &&
1905 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1906 return &policy->nodes;
1908 if (mode == MPOL_PREFERRED_MANY)
1909 return &policy->nodes;
1915 * Return the preferred node id for 'prefer' mempolicy, and return
1916 * the given id for all other policies.
1918 * policy_node() is always coupled with policy_nodemask(), which
1919 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1921 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1923 if (policy->mode == MPOL_PREFERRED) {
1924 nd = first_node(policy->nodes);
1927 * __GFP_THISNODE shouldn't even be used with the bind policy
1928 * because we might easily break the expectation to stay on the
1929 * requested node and not break the policy.
1931 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1937 /* Do dynamic interleaving for a process */
1938 static unsigned interleave_nodes(struct mempolicy *policy)
1941 struct task_struct *me = current;
1943 next = next_node_in(me->il_prev, policy->nodes);
1944 if (next < MAX_NUMNODES)
1950 * Depending on the memory policy provide a node from which to allocate the
1953 unsigned int mempolicy_slab_node(void)
1955 struct mempolicy *policy;
1956 int node = numa_mem_id();
1961 policy = current->mempolicy;
1965 switch (policy->mode) {
1966 case MPOL_PREFERRED:
1967 return first_node(policy->nodes);
1969 case MPOL_INTERLEAVE:
1970 return interleave_nodes(policy);
1973 case MPOL_PREFERRED_MANY:
1978 * Follow bind policy behavior and start allocation at the
1981 struct zonelist *zonelist;
1982 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1983 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1984 z = first_zones_zonelist(zonelist, highest_zoneidx,
1986 return z->zone ? zone_to_nid(z->zone) : node;
1997 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1998 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1999 * number of present nodes.
2001 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
2003 unsigned nnodes = nodes_weight(pol->nodes);
2009 return numa_node_id();
2010 target = (unsigned int)n % nnodes;
2011 nid = first_node(pol->nodes);
2012 for (i = 0; i < target; i++)
2013 nid = next_node(nid, pol->nodes);
2017 /* Determine a node number for interleave */
2018 static inline unsigned interleave_nid(struct mempolicy *pol,
2019 struct vm_area_struct *vma, unsigned long addr, int shift)
2025 * for small pages, there is no difference between
2026 * shift and PAGE_SHIFT, so the bit-shift is safe.
2027 * for huge pages, since vm_pgoff is in units of small
2028 * pages, we need to shift off the always 0 bits to get
2031 BUG_ON(shift < PAGE_SHIFT);
2032 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2033 off += (addr - vma->vm_start) >> shift;
2034 return offset_il_node(pol, off);
2036 return interleave_nodes(pol);
2039 #ifdef CONFIG_HUGETLBFS
2041 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2042 * @vma: virtual memory area whose policy is sought
2043 * @addr: address in @vma for shared policy lookup and interleave policy
2044 * @gfp_flags: for requested zone
2045 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2046 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2048 * Returns a nid suitable for a huge page allocation and a pointer
2049 * to the struct mempolicy for conditional unref after allocation.
2050 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2051 * to the mempolicy's @nodemask for filtering the zonelist.
2053 * Must be protected by read_mems_allowed_begin()
2055 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2056 struct mempolicy **mpol, nodemask_t **nodemask)
2061 *mpol = get_vma_policy(vma, addr);
2063 mode = (*mpol)->mode;
2065 if (unlikely(mode == MPOL_INTERLEAVE)) {
2066 nid = interleave_nid(*mpol, vma, addr,
2067 huge_page_shift(hstate_vma(vma)));
2069 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2070 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2071 *nodemask = &(*mpol)->nodes;
2077 * init_nodemask_of_mempolicy
2079 * If the current task's mempolicy is "default" [NULL], return 'false'
2080 * to indicate default policy. Otherwise, extract the policy nodemask
2081 * for 'bind' or 'interleave' policy into the argument nodemask, or
2082 * initialize the argument nodemask to contain the single node for
2083 * 'preferred' or 'local' policy and return 'true' to indicate presence
2084 * of non-default mempolicy.
2086 * We don't bother with reference counting the mempolicy [mpol_get/put]
2087 * because the current task is examining it's own mempolicy and a task's
2088 * mempolicy is only ever changed by the task itself.
2090 * N.B., it is the caller's responsibility to free a returned nodemask.
2092 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2094 struct mempolicy *mempolicy;
2096 if (!(mask && current->mempolicy))
2100 mempolicy = current->mempolicy;
2101 switch (mempolicy->mode) {
2102 case MPOL_PREFERRED:
2103 case MPOL_PREFERRED_MANY:
2105 case MPOL_INTERLEAVE:
2106 *mask = mempolicy->nodes;
2110 init_nodemask_of_node(mask, numa_node_id());
2116 task_unlock(current);
2123 * mempolicy_in_oom_domain
2125 * If tsk's mempolicy is "bind", check for intersection between mask and
2126 * the policy nodemask. Otherwise, return true for all other policies
2127 * including "interleave", as a tsk with "interleave" policy may have
2128 * memory allocated from all nodes in system.
2130 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2132 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2133 const nodemask_t *mask)
2135 struct mempolicy *mempolicy;
2142 mempolicy = tsk->mempolicy;
2143 if (mempolicy && mempolicy->mode == MPOL_BIND)
2144 ret = nodes_intersects(mempolicy->nodes, *mask);
2150 /* Allocate a page in interleaved policy.
2151 Own path because it needs to do special accounting. */
2152 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2157 page = __alloc_pages(gfp, order, nid, NULL);
2158 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2159 if (!static_branch_likely(&vm_numa_stat_key))
2161 if (page && page_to_nid(page) == nid) {
2163 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2169 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2170 int nid, struct mempolicy *pol)
2173 gfp_t preferred_gfp;
2176 * This is a two pass approach. The first pass will only try the
2177 * preferred nodes but skip the direct reclaim and allow the
2178 * allocation to fail, while the second pass will try all the
2181 preferred_gfp = gfp | __GFP_NOWARN;
2182 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2183 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2185 page = __alloc_pages(gfp, order, numa_node_id(), NULL);
2191 * alloc_pages_vma - Allocate a page for a VMA.
2193 * @order: Order of the GFP allocation.
2194 * @vma: Pointer to VMA or NULL if not available.
2195 * @addr: Virtual address of the allocation. Must be inside @vma.
2196 * @node: Which node to prefer for allocation (modulo policy).
2197 * @hugepage: For hugepages try only the preferred node if possible.
2199 * Allocate a page for a specific address in @vma, using the appropriate
2200 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2201 * of the mm_struct of the VMA to prevent it from going away. Should be
2202 * used for all allocations for pages that will be mapped into user space.
2204 * Return: The page on success or NULL if allocation fails.
2206 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2207 unsigned long addr, int node, bool hugepage)
2209 struct mempolicy *pol;
2214 pol = get_vma_policy(vma, addr);
2216 if (pol->mode == MPOL_INTERLEAVE) {
2219 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2221 page = alloc_page_interleave(gfp, order, nid);
2225 if (pol->mode == MPOL_PREFERRED_MANY) {
2226 page = alloc_pages_preferred_many(gfp, order, node, pol);
2231 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2232 int hpage_node = node;
2235 * For hugepage allocation and non-interleave policy which
2236 * allows the current node (or other explicitly preferred
2237 * node) we only try to allocate from the current/preferred
2238 * node and don't fall back to other nodes, as the cost of
2239 * remote accesses would likely offset THP benefits.
2241 * If the policy is interleave or does not allow the current
2242 * node in its nodemask, we allocate the standard way.
2244 if (pol->mode == MPOL_PREFERRED)
2245 hpage_node = first_node(pol->nodes);
2247 nmask = policy_nodemask(gfp, pol);
2248 if (!nmask || node_isset(hpage_node, *nmask)) {
2251 * First, try to allocate THP only on local node, but
2252 * don't reclaim unnecessarily, just compact.
2254 page = __alloc_pages_node(hpage_node,
2255 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2258 * If hugepage allocations are configured to always
2259 * synchronous compact or the vma has been madvised
2260 * to prefer hugepage backing, retry allowing remote
2261 * memory with both reclaim and compact as well.
2263 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2264 page = __alloc_pages_node(hpage_node,
2271 nmask = policy_nodemask(gfp, pol);
2272 preferred_nid = policy_node(gfp, pol, node);
2273 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2278 EXPORT_SYMBOL(alloc_pages_vma);
2281 * alloc_pages - Allocate pages.
2283 * @order: Power of two of number of pages to allocate.
2285 * Allocate 1 << @order contiguous pages. The physical address of the
2286 * first page is naturally aligned (eg an order-3 allocation will be aligned
2287 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2288 * process is honoured when in process context.
2290 * Context: Can be called from any context, providing the appropriate GFP
2292 * Return: The page on success or NULL if allocation fails.
2294 struct page *alloc_pages(gfp_t gfp, unsigned order)
2296 struct mempolicy *pol = &default_policy;
2299 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2300 pol = get_task_policy(current);
2303 * No reference counting needed for current->mempolicy
2304 * nor system default_policy
2306 if (pol->mode == MPOL_INTERLEAVE)
2307 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2308 else if (pol->mode == MPOL_PREFERRED_MANY)
2309 page = alloc_pages_preferred_many(gfp, order,
2310 numa_node_id(), pol);
2312 page = __alloc_pages(gfp, order,
2313 policy_node(gfp, pol, numa_node_id()),
2314 policy_nodemask(gfp, pol));
2318 EXPORT_SYMBOL(alloc_pages);
2320 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2322 struct mempolicy *pol = mpol_dup(vma_policy(src));
2325 return PTR_ERR(pol);
2326 dst->vm_policy = pol;
2331 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2332 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2333 * with the mems_allowed returned by cpuset_mems_allowed(). This
2334 * keeps mempolicies cpuset relative after its cpuset moves. See
2335 * further kernel/cpuset.c update_nodemask().
2337 * current's mempolicy may be rebinded by the other task(the task that changes
2338 * cpuset's mems), so we needn't do rebind work for current task.
2341 /* Slow path of a mempolicy duplicate */
2342 struct mempolicy *__mpol_dup(struct mempolicy *old)
2344 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2347 return ERR_PTR(-ENOMEM);
2349 /* task's mempolicy is protected by alloc_lock */
2350 if (old == current->mempolicy) {
2353 task_unlock(current);
2357 if (current_cpuset_is_being_rebound()) {
2358 nodemask_t mems = cpuset_mems_allowed(current);
2359 mpol_rebind_policy(new, &mems);
2361 atomic_set(&new->refcnt, 1);
2365 /* Slow path of a mempolicy comparison */
2366 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2370 if (a->mode != b->mode)
2372 if (a->flags != b->flags)
2374 if (mpol_store_user_nodemask(a))
2375 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2380 case MPOL_INTERLEAVE:
2381 case MPOL_PREFERRED:
2382 case MPOL_PREFERRED_MANY:
2383 return !!nodes_equal(a->nodes, b->nodes);
2393 * Shared memory backing store policy support.
2395 * Remember policies even when nobody has shared memory mapped.
2396 * The policies are kept in Red-Black tree linked from the inode.
2397 * They are protected by the sp->lock rwlock, which should be held
2398 * for any accesses to the tree.
2402 * lookup first element intersecting start-end. Caller holds sp->lock for
2403 * reading or for writing
2405 static struct sp_node *
2406 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2408 struct rb_node *n = sp->root.rb_node;
2411 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2413 if (start >= p->end)
2415 else if (end <= p->start)
2423 struct sp_node *w = NULL;
2424 struct rb_node *prev = rb_prev(n);
2427 w = rb_entry(prev, struct sp_node, nd);
2428 if (w->end <= start)
2432 return rb_entry(n, struct sp_node, nd);
2436 * Insert a new shared policy into the list. Caller holds sp->lock for
2439 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2441 struct rb_node **p = &sp->root.rb_node;
2442 struct rb_node *parent = NULL;
2447 nd = rb_entry(parent, struct sp_node, nd);
2448 if (new->start < nd->start)
2450 else if (new->end > nd->end)
2451 p = &(*p)->rb_right;
2455 rb_link_node(&new->nd, parent, p);
2456 rb_insert_color(&new->nd, &sp->root);
2457 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2458 new->policy ? new->policy->mode : 0);
2461 /* Find shared policy intersecting idx */
2463 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2465 struct mempolicy *pol = NULL;
2468 if (!sp->root.rb_node)
2470 read_lock(&sp->lock);
2471 sn = sp_lookup(sp, idx, idx+1);
2473 mpol_get(sn->policy);
2476 read_unlock(&sp->lock);
2480 static void sp_free(struct sp_node *n)
2482 mpol_put(n->policy);
2483 kmem_cache_free(sn_cache, n);
2487 * mpol_misplaced - check whether current page node is valid in policy
2489 * @page: page to be checked
2490 * @vma: vm area where page mapped
2491 * @addr: virtual address where page mapped
2493 * Lookup current policy node id for vma,addr and "compare to" page's
2494 * node id. Policy determination "mimics" alloc_page_vma().
2495 * Called from fault path where we know the vma and faulting address.
2497 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2498 * policy, or a suitable node ID to allocate a replacement page from.
2500 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2502 struct mempolicy *pol;
2504 int curnid = page_to_nid(page);
2505 unsigned long pgoff;
2506 int thiscpu = raw_smp_processor_id();
2507 int thisnid = cpu_to_node(thiscpu);
2508 int polnid = NUMA_NO_NODE;
2509 int ret = NUMA_NO_NODE;
2511 pol = get_vma_policy(vma, addr);
2512 if (!(pol->flags & MPOL_F_MOF))
2515 switch (pol->mode) {
2516 case MPOL_INTERLEAVE:
2517 pgoff = vma->vm_pgoff;
2518 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2519 polnid = offset_il_node(pol, pgoff);
2522 case MPOL_PREFERRED:
2523 if (node_isset(curnid, pol->nodes))
2525 polnid = first_node(pol->nodes);
2529 polnid = numa_node_id();
2533 /* Optimize placement among multiple nodes via NUMA balancing */
2534 if (pol->flags & MPOL_F_MORON) {
2535 if (node_isset(thisnid, pol->nodes))
2541 case MPOL_PREFERRED_MANY:
2543 * use current page if in policy nodemask,
2544 * else select nearest allowed node, if any.
2545 * If no allowed nodes, use current [!misplaced].
2547 if (node_isset(curnid, pol->nodes))
2549 z = first_zones_zonelist(
2550 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2551 gfp_zone(GFP_HIGHUSER),
2553 polnid = zone_to_nid(z->zone);
2560 /* Migrate the page towards the node whose CPU is referencing it */
2561 if (pol->flags & MPOL_F_MORON) {
2564 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2568 if (curnid != polnid)
2577 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2578 * dropped after task->mempolicy is set to NULL so that any allocation done as
2579 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2582 void mpol_put_task_policy(struct task_struct *task)
2584 struct mempolicy *pol;
2587 pol = task->mempolicy;
2588 task->mempolicy = NULL;
2593 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2595 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2596 rb_erase(&n->nd, &sp->root);
2600 static void sp_node_init(struct sp_node *node, unsigned long start,
2601 unsigned long end, struct mempolicy *pol)
2603 node->start = start;
2608 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2609 struct mempolicy *pol)
2612 struct mempolicy *newpol;
2614 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2618 newpol = mpol_dup(pol);
2619 if (IS_ERR(newpol)) {
2620 kmem_cache_free(sn_cache, n);
2623 newpol->flags |= MPOL_F_SHARED;
2624 sp_node_init(n, start, end, newpol);
2629 /* Replace a policy range. */
2630 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2631 unsigned long end, struct sp_node *new)
2634 struct sp_node *n_new = NULL;
2635 struct mempolicy *mpol_new = NULL;
2639 write_lock(&sp->lock);
2640 n = sp_lookup(sp, start, end);
2641 /* Take care of old policies in the same range. */
2642 while (n && n->start < end) {
2643 struct rb_node *next = rb_next(&n->nd);
2644 if (n->start >= start) {
2650 /* Old policy spanning whole new range. */
2655 *mpol_new = *n->policy;
2656 atomic_set(&mpol_new->refcnt, 1);
2657 sp_node_init(n_new, end, n->end, mpol_new);
2659 sp_insert(sp, n_new);
2668 n = rb_entry(next, struct sp_node, nd);
2672 write_unlock(&sp->lock);
2679 kmem_cache_free(sn_cache, n_new);
2684 write_unlock(&sp->lock);
2686 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2689 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2696 * mpol_shared_policy_init - initialize shared policy for inode
2697 * @sp: pointer to inode shared policy
2698 * @mpol: struct mempolicy to install
2700 * Install non-NULL @mpol in inode's shared policy rb-tree.
2701 * On entry, the current task has a reference on a non-NULL @mpol.
2702 * This must be released on exit.
2703 * This is called at get_inode() calls and we can use GFP_KERNEL.
2705 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2709 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2710 rwlock_init(&sp->lock);
2713 struct vm_area_struct pvma;
2714 struct mempolicy *new;
2715 NODEMASK_SCRATCH(scratch);
2719 /* contextualize the tmpfs mount point mempolicy */
2720 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2722 goto free_scratch; /* no valid nodemask intersection */
2725 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2726 task_unlock(current);
2730 /* Create pseudo-vma that contains just the policy */
2731 vma_init(&pvma, NULL);
2732 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2733 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2736 mpol_put(new); /* drop initial ref */
2738 NODEMASK_SCRATCH_FREE(scratch);
2740 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2744 int mpol_set_shared_policy(struct shared_policy *info,
2745 struct vm_area_struct *vma, struct mempolicy *npol)
2748 struct sp_node *new = NULL;
2749 unsigned long sz = vma_pages(vma);
2751 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2753 sz, npol ? npol->mode : -1,
2754 npol ? npol->flags : -1,
2755 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2758 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2762 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2768 /* Free a backing policy store on inode delete. */
2769 void mpol_free_shared_policy(struct shared_policy *p)
2772 struct rb_node *next;
2774 if (!p->root.rb_node)
2776 write_lock(&p->lock);
2777 next = rb_first(&p->root);
2779 n = rb_entry(next, struct sp_node, nd);
2780 next = rb_next(&n->nd);
2783 write_unlock(&p->lock);
2786 #ifdef CONFIG_NUMA_BALANCING
2787 static int __initdata numabalancing_override;
2789 static void __init check_numabalancing_enable(void)
2791 bool numabalancing_default = false;
2793 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2794 numabalancing_default = true;
2796 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2797 if (numabalancing_override)
2798 set_numabalancing_state(numabalancing_override == 1);
2800 if (num_online_nodes() > 1 && !numabalancing_override) {
2801 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2802 numabalancing_default ? "Enabling" : "Disabling");
2803 set_numabalancing_state(numabalancing_default);
2807 static int __init setup_numabalancing(char *str)
2813 if (!strcmp(str, "enable")) {
2814 numabalancing_override = 1;
2816 } else if (!strcmp(str, "disable")) {
2817 numabalancing_override = -1;
2822 pr_warn("Unable to parse numa_balancing=\n");
2826 __setup("numa_balancing=", setup_numabalancing);
2828 static inline void __init check_numabalancing_enable(void)
2831 #endif /* CONFIG_NUMA_BALANCING */
2833 /* assumes fs == KERNEL_DS */
2834 void __init numa_policy_init(void)
2836 nodemask_t interleave_nodes;
2837 unsigned long largest = 0;
2838 int nid, prefer = 0;
2840 policy_cache = kmem_cache_create("numa_policy",
2841 sizeof(struct mempolicy),
2842 0, SLAB_PANIC, NULL);
2844 sn_cache = kmem_cache_create("shared_policy_node",
2845 sizeof(struct sp_node),
2846 0, SLAB_PANIC, NULL);
2848 for_each_node(nid) {
2849 preferred_node_policy[nid] = (struct mempolicy) {
2850 .refcnt = ATOMIC_INIT(1),
2851 .mode = MPOL_PREFERRED,
2852 .flags = MPOL_F_MOF | MPOL_F_MORON,
2853 .nodes = nodemask_of_node(nid),
2858 * Set interleaving policy for system init. Interleaving is only
2859 * enabled across suitably sized nodes (default is >= 16MB), or
2860 * fall back to the largest node if they're all smaller.
2862 nodes_clear(interleave_nodes);
2863 for_each_node_state(nid, N_MEMORY) {
2864 unsigned long total_pages = node_present_pages(nid);
2866 /* Preserve the largest node */
2867 if (largest < total_pages) {
2868 largest = total_pages;
2872 /* Interleave this node? */
2873 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2874 node_set(nid, interleave_nodes);
2877 /* All too small, use the largest */
2878 if (unlikely(nodes_empty(interleave_nodes)))
2879 node_set(prefer, interleave_nodes);
2881 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2882 pr_err("%s: interleaving failed\n", __func__);
2884 check_numabalancing_enable();
2887 /* Reset policy of current process to default */
2888 void numa_default_policy(void)
2890 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2894 * Parse and format mempolicy from/to strings
2897 static const char * const policy_modes[] =
2899 [MPOL_DEFAULT] = "default",
2900 [MPOL_PREFERRED] = "prefer",
2901 [MPOL_BIND] = "bind",
2902 [MPOL_INTERLEAVE] = "interleave",
2903 [MPOL_LOCAL] = "local",
2904 [MPOL_PREFERRED_MANY] = "prefer (many)",
2910 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2911 * @str: string containing mempolicy to parse
2912 * @mpol: pointer to struct mempolicy pointer, returned on success.
2915 * <mode>[=<flags>][:<nodelist>]
2917 * On success, returns 0, else 1
2919 int mpol_parse_str(char *str, struct mempolicy **mpol)
2921 struct mempolicy *new = NULL;
2922 unsigned short mode_flags;
2924 char *nodelist = strchr(str, ':');
2925 char *flags = strchr(str, '=');
2929 *flags++ = '\0'; /* terminate mode string */
2932 /* NUL-terminate mode or flags string */
2934 if (nodelist_parse(nodelist, nodes))
2936 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2941 mode = match_string(policy_modes, MPOL_MAX, str);
2946 case MPOL_PREFERRED:
2948 * Insist on a nodelist of one node only, although later
2949 * we use first_node(nodes) to grab a single node, so here
2950 * nodelist (or nodes) cannot be empty.
2953 char *rest = nodelist;
2954 while (isdigit(*rest))
2958 if (nodes_empty(nodes))
2962 case MPOL_INTERLEAVE:
2964 * Default to online nodes with memory if no nodelist
2967 nodes = node_states[N_MEMORY];
2971 * Don't allow a nodelist; mpol_new() checks flags
2978 * Insist on a empty nodelist
2983 case MPOL_PREFERRED_MANY:
2986 * Insist on a nodelist
2995 * Currently, we only support two mutually exclusive
2998 if (!strcmp(flags, "static"))
2999 mode_flags |= MPOL_F_STATIC_NODES;
3000 else if (!strcmp(flags, "relative"))
3001 mode_flags |= MPOL_F_RELATIVE_NODES;
3006 new = mpol_new(mode, mode_flags, &nodes);
3011 * Save nodes for mpol_to_str() to show the tmpfs mount options
3012 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3014 if (mode != MPOL_PREFERRED) {
3016 } else if (nodelist) {
3017 nodes_clear(new->nodes);
3018 node_set(first_node(nodes), new->nodes);
3020 new->mode = MPOL_LOCAL;
3024 * Save nodes for contextualization: this will be used to "clone"
3025 * the mempolicy in a specific context [cpuset] at a later time.
3027 new->w.user_nodemask = nodes;
3032 /* Restore string for error message */
3041 #endif /* CONFIG_TMPFS */
3044 * mpol_to_str - format a mempolicy structure for printing
3045 * @buffer: to contain formatted mempolicy string
3046 * @maxlen: length of @buffer
3047 * @pol: pointer to mempolicy to be formatted
3049 * Convert @pol into a string. If @buffer is too short, truncate the string.
3050 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3051 * longest flag, "relative", and to display at least a few node ids.
3053 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3056 nodemask_t nodes = NODE_MASK_NONE;
3057 unsigned short mode = MPOL_DEFAULT;
3058 unsigned short flags = 0;
3060 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3069 case MPOL_PREFERRED:
3070 case MPOL_PREFERRED_MANY:
3072 case MPOL_INTERLEAVE:
3077 snprintf(p, maxlen, "unknown");
3081 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3083 if (flags & MPOL_MODE_FLAGS) {
3084 p += snprintf(p, buffer + maxlen - p, "=");
3087 * Currently, the only defined flags are mutually exclusive
3089 if (flags & MPOL_F_STATIC_NODES)
3090 p += snprintf(p, buffer + maxlen - p, "static");
3091 else if (flags & MPOL_F_RELATIVE_NODES)
3092 p += snprintf(p, buffer + maxlen - p, "relative");
3095 if (!nodes_empty(nodes))
3096 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3097 nodemask_pr_args(&nodes));
3100 bool numa_demotion_enabled = false;
3103 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
3104 struct kobj_attribute *attr, char *buf)
3106 return sysfs_emit(buf, "%s\n",
3107 numa_demotion_enabled? "true" : "false");
3110 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
3111 struct kobj_attribute *attr,
3112 const char *buf, size_t count)
3114 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
3115 numa_demotion_enabled = true;
3116 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
3117 numa_demotion_enabled = false;
3124 static struct kobj_attribute numa_demotion_enabled_attr =
3125 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3126 numa_demotion_enabled_store);
3128 static struct attribute *numa_attrs[] = {
3129 &numa_demotion_enabled_attr.attr,
3133 static const struct attribute_group numa_attr_group = {
3134 .attrs = numa_attrs,
3137 static int __init numa_init_sysfs(void)
3140 struct kobject *numa_kobj;
3142 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3144 pr_err("failed to create numa kobject\n");
3147 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3149 pr_err("failed to register numa group\n");
3155 kobject_put(numa_kobj);
3158 subsys_initcall(numa_init_sysfs);