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_nodemask(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);
214 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
215 * any, for the new policy. mpol_new() has already validated the nodes
216 * parameter with respect to the policy mode and flags.
218 * Must be called holding task's alloc_lock to protect task's mems_allowed
219 * and mempolicy. May also be called holding the mmap_lock for write.
221 static int mpol_set_nodemask(struct mempolicy *pol,
222 const nodemask_t *nodes, struct nodemask_scratch *nsc)
227 * Default (pol==NULL) resp. local memory policies are not a
228 * subject of any remapping. They also do not need any special
231 if (!pol || pol->mode == MPOL_LOCAL)
235 nodes_and(nsc->mask1,
236 cpuset_current_mems_allowed, node_states[N_MEMORY]);
240 if (pol->flags & MPOL_F_RELATIVE_NODES)
241 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
243 nodes_and(nsc->mask2, *nodes, nsc->mask1);
245 if (mpol_store_user_nodemask(pol))
246 pol->w.user_nodemask = *nodes;
248 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
250 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
286 } else if (mode == MPOL_LOCAL) {
287 if (!nodes_empty(*nodes) ||
288 (flags & MPOL_F_STATIC_NODES) ||
289 (flags & MPOL_F_RELATIVE_NODES))
290 return ERR_PTR(-EINVAL);
291 } else if (nodes_empty(*nodes))
292 return ERR_PTR(-EINVAL);
293 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
295 return ERR_PTR(-ENOMEM);
296 atomic_set(&policy->refcnt, 1);
298 policy->flags = flags;
303 /* Slow path of a mpol destructor. */
304 void __mpol_put(struct mempolicy *p)
306 if (!atomic_dec_and_test(&p->refcnt))
308 kmem_cache_free(policy_cache, p);
311 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
315 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
319 if (pol->flags & MPOL_F_STATIC_NODES)
320 nodes_and(tmp, pol->w.user_nodemask, *nodes);
321 else if (pol->flags & MPOL_F_RELATIVE_NODES)
322 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
324 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
326 pol->w.cpuset_mems_allowed = *nodes;
329 if (nodes_empty(tmp))
335 static void mpol_rebind_preferred(struct mempolicy *pol,
336 const nodemask_t *nodes)
338 pol->w.cpuset_mems_allowed = *nodes;
342 * mpol_rebind_policy - Migrate a policy to a different set of nodes
344 * Per-vma policies are protected by mmap_lock. Allocations using per-task
345 * policies are protected by task->mems_allowed_seq to prevent a premature
346 * OOM/allocation failure due to parallel nodemask modification.
348 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
352 if (!mpol_store_user_nodemask(pol) &&
353 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
356 mpol_ops[pol->mode].rebind(pol, newmask);
360 * Wrapper for mpol_rebind_policy() that just requires task
361 * pointer, and updates task mempolicy.
363 * Called with task's alloc_lock held.
366 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
368 mpol_rebind_policy(tsk->mempolicy, new);
372 * Rebind each vma in mm to new nodemask.
374 * Call holding a reference to mm. Takes mm->mmap_lock during call.
377 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
379 struct vm_area_struct *vma;
382 for (vma = mm->mmap; vma; vma = vma->vm_next)
383 mpol_rebind_policy(vma->vm_policy, new);
384 mmap_write_unlock(mm);
387 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
389 .rebind = mpol_rebind_default,
391 [MPOL_INTERLEAVE] = {
392 .create = mpol_new_nodemask,
393 .rebind = mpol_rebind_nodemask,
396 .create = mpol_new_preferred,
397 .rebind = mpol_rebind_preferred,
400 .create = mpol_new_nodemask,
401 .rebind = mpol_rebind_nodemask,
404 .rebind = mpol_rebind_default,
406 [MPOL_PREFERRED_MANY] = {
407 .create = mpol_new_nodemask,
408 .rebind = mpol_rebind_preferred,
412 static int migrate_page_add(struct page *page, struct list_head *pagelist,
413 unsigned long flags);
416 struct list_head *pagelist;
421 struct vm_area_struct *first;
425 * Check if the page's nid is in qp->nmask.
427 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
428 * in the invert of qp->nmask.
430 static inline bool queue_pages_required(struct page *page,
431 struct queue_pages *qp)
433 int nid = page_to_nid(page);
434 unsigned long flags = qp->flags;
436 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
440 * queue_pages_pmd() has four possible return values:
441 * 0 - pages are placed on the right node or queued successfully, or
442 * special page is met, i.e. huge zero page.
443 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
446 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
447 * existing page was already on a node that does not follow the
450 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
451 unsigned long end, struct mm_walk *walk)
456 struct queue_pages *qp = walk->private;
459 if (unlikely(is_pmd_migration_entry(*pmd))) {
463 page = pmd_page(*pmd);
464 if (is_huge_zero_page(page)) {
466 walk->action = ACTION_CONTINUE;
469 if (!queue_pages_required(page, qp))
473 /* go to thp migration */
474 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
475 if (!vma_migratable(walk->vma) ||
476 migrate_page_add(page, qp->pagelist, flags)) {
489 * Scan through pages checking if pages follow certain conditions,
490 * and move them to the pagelist if they do.
492 * queue_pages_pte_range() has three possible return values:
493 * 0 - pages are placed on the right node or queued successfully, or
494 * special page is met, i.e. zero page.
495 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
497 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
498 * on a node that does not follow the policy.
500 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
501 unsigned long end, struct mm_walk *walk)
503 struct vm_area_struct *vma = walk->vma;
505 struct queue_pages *qp = walk->private;
506 unsigned long flags = qp->flags;
508 bool has_unmovable = false;
509 pte_t *pte, *mapped_pte;
512 ptl = pmd_trans_huge_lock(pmd, vma);
514 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
518 /* THP was split, fall through to pte walk */
520 if (pmd_trans_unstable(pmd))
523 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
524 for (; addr != end; pte++, addr += PAGE_SIZE) {
525 if (!pte_present(*pte))
527 page = vm_normal_page(vma, addr, *pte);
531 * vm_normal_page() filters out zero pages, but there might
532 * still be PageReserved pages to skip, perhaps in a VDSO.
534 if (PageReserved(page))
536 if (!queue_pages_required(page, qp))
538 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
539 /* MPOL_MF_STRICT must be specified if we get here */
540 if (!vma_migratable(vma)) {
541 has_unmovable = true;
546 * Do not abort immediately since there may be
547 * temporary off LRU pages in the range. Still
548 * need migrate other LRU pages.
550 if (migrate_page_add(page, qp->pagelist, flags))
551 has_unmovable = true;
555 pte_unmap_unlock(mapped_pte, ptl);
561 return addr != end ? -EIO : 0;
564 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
565 unsigned long addr, unsigned long end,
566 struct mm_walk *walk)
569 #ifdef CONFIG_HUGETLB_PAGE
570 struct queue_pages *qp = walk->private;
571 unsigned long flags = (qp->flags & MPOL_MF_VALID);
576 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
577 entry = huge_ptep_get(pte);
578 if (!pte_present(entry))
580 page = pte_page(entry);
581 if (!queue_pages_required(page, qp))
584 if (flags == MPOL_MF_STRICT) {
586 * STRICT alone means only detecting misplaced page and no
587 * need to further check other vma.
593 if (!vma_migratable(walk->vma)) {
595 * Must be STRICT with MOVE*, otherwise .test_walk() have
596 * stopped walking current vma.
597 * Detecting misplaced page but allow migrating pages which
604 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
605 if (flags & (MPOL_MF_MOVE_ALL) ||
606 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
607 if (!isolate_huge_page(page, qp->pagelist) &&
608 (flags & MPOL_MF_STRICT))
610 * Failed to isolate page but allow migrating pages
611 * which have been queued.
623 #ifdef CONFIG_NUMA_BALANCING
625 * This is used to mark a range of virtual addresses to be inaccessible.
626 * These are later cleared by a NUMA hinting fault. Depending on these
627 * faults, pages may be migrated for better NUMA placement.
629 * This is assuming that NUMA faults are handled using PROT_NONE. If
630 * an architecture makes a different choice, it will need further
631 * changes to the core.
633 unsigned long change_prot_numa(struct vm_area_struct *vma,
634 unsigned long addr, unsigned long end)
638 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
640 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
645 static unsigned long change_prot_numa(struct vm_area_struct *vma,
646 unsigned long addr, unsigned long end)
650 #endif /* CONFIG_NUMA_BALANCING */
652 static int queue_pages_test_walk(unsigned long start, unsigned long end,
653 struct mm_walk *walk)
655 struct vm_area_struct *vma = walk->vma;
656 struct queue_pages *qp = walk->private;
657 unsigned long endvma = vma->vm_end;
658 unsigned long flags = qp->flags;
660 /* range check first */
661 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
665 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
666 (qp->start < vma->vm_start))
667 /* hole at head side of range */
670 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
671 ((vma->vm_end < qp->end) &&
672 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
673 /* hole at middle or tail of range */
677 * Need check MPOL_MF_STRICT to return -EIO if possible
678 * regardless of vma_migratable
680 if (!vma_migratable(vma) &&
681 !(flags & MPOL_MF_STRICT))
687 if (flags & MPOL_MF_LAZY) {
688 /* Similar to task_numa_work, skip inaccessible VMAs */
689 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
690 !(vma->vm_flags & VM_MIXEDMAP))
691 change_prot_numa(vma, start, endvma);
695 /* queue pages from current vma */
696 if (flags & MPOL_MF_VALID)
701 static const struct mm_walk_ops queue_pages_walk_ops = {
702 .hugetlb_entry = queue_pages_hugetlb,
703 .pmd_entry = queue_pages_pte_range,
704 .test_walk = queue_pages_test_walk,
708 * Walk through page tables and collect pages to be migrated.
710 * If pages found in a given range are on a set of nodes (determined by
711 * @nodes and @flags,) it's isolated and queued to the pagelist which is
712 * passed via @private.
714 * queue_pages_range() has three possible return values:
715 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
717 * 0 - queue pages successfully or no misplaced page.
718 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
719 * memory range specified by nodemask and maxnode points outside
720 * your accessible address space (-EFAULT)
723 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
724 nodemask_t *nodes, unsigned long flags,
725 struct list_head *pagelist)
728 struct queue_pages qp = {
729 .pagelist = pagelist,
737 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
740 /* whole range in hole */
747 * Apply policy to a single VMA
748 * This must be called with the mmap_lock held for writing.
750 static int vma_replace_policy(struct vm_area_struct *vma,
751 struct mempolicy *pol)
754 struct mempolicy *old;
755 struct mempolicy *new;
757 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
758 vma->vm_start, vma->vm_end, vma->vm_pgoff,
759 vma->vm_ops, vma->vm_file,
760 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
766 if (vma->vm_ops && vma->vm_ops->set_policy) {
767 err = vma->vm_ops->set_policy(vma, new);
772 old = vma->vm_policy;
773 vma->vm_policy = new; /* protected by mmap_lock */
782 /* Step 2: apply policy to a range and do splits. */
783 static int mbind_range(struct mm_struct *mm, unsigned long start,
784 unsigned long end, struct mempolicy *new_pol)
786 struct vm_area_struct *next;
787 struct vm_area_struct *prev;
788 struct vm_area_struct *vma;
791 unsigned long vmstart;
794 vma = find_vma(mm, start);
798 if (start > vma->vm_start)
801 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
803 vmstart = max(start, vma->vm_start);
804 vmend = min(end, vma->vm_end);
806 if (mpol_equal(vma_policy(vma), new_pol))
809 pgoff = vma->vm_pgoff +
810 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
811 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
812 vma->anon_vma, vma->vm_file, pgoff,
813 new_pol, vma->vm_userfaultfd_ctx);
817 if (mpol_equal(vma_policy(vma), new_pol))
819 /* vma_merge() joined vma && vma->next, case 8 */
822 if (vma->vm_start != vmstart) {
823 err = split_vma(vma->vm_mm, vma, vmstart, 1);
827 if (vma->vm_end != vmend) {
828 err = split_vma(vma->vm_mm, vma, vmend, 0);
833 err = vma_replace_policy(vma, new_pol);
842 /* Set the process memory policy */
843 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
846 struct mempolicy *new, *old;
847 NODEMASK_SCRATCH(scratch);
853 new = mpol_new(mode, flags, nodes);
859 if (flags & MPOL_F_NUMA_BALANCING) {
860 if (new && new->mode == MPOL_BIND) {
861 new->flags |= (MPOL_F_MOF | MPOL_F_MORON);
869 ret = mpol_set_nodemask(new, nodes, scratch);
875 old = current->mempolicy;
876 current->mempolicy = new;
877 if (new && new->mode == MPOL_INTERLEAVE)
878 current->il_prev = MAX_NUMNODES-1;
879 task_unlock(current);
883 NODEMASK_SCRATCH_FREE(scratch);
888 * Return nodemask for policy for get_mempolicy() query
890 * Called with task's alloc_lock held
892 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
895 if (p == &default_policy)
900 case MPOL_INTERLEAVE:
902 case MPOL_PREFERRED_MANY:
906 /* return empty node mask for local allocation */
913 static int lookup_node(struct mm_struct *mm, unsigned long addr)
915 struct page *p = NULL;
919 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
921 err = page_to_nid(p);
925 mmap_read_unlock(mm);
929 /* Retrieve NUMA policy */
930 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
931 unsigned long addr, unsigned long flags)
934 struct mm_struct *mm = current->mm;
935 struct vm_area_struct *vma = NULL;
936 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
939 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
942 if (flags & MPOL_F_MEMS_ALLOWED) {
943 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
945 *policy = 0; /* just so it's initialized */
947 *nmask = cpuset_current_mems_allowed;
948 task_unlock(current);
952 if (flags & MPOL_F_ADDR) {
954 * Do NOT fall back to task policy if the
955 * vma/shared policy at addr is NULL. We
956 * want to return MPOL_DEFAULT in this case.
959 vma = vma_lookup(mm, addr);
961 mmap_read_unlock(mm);
964 if (vma->vm_ops && vma->vm_ops->get_policy)
965 pol = vma->vm_ops->get_policy(vma, addr);
967 pol = vma->vm_policy;
972 pol = &default_policy; /* indicates default behavior */
974 if (flags & MPOL_F_NODE) {
975 if (flags & MPOL_F_ADDR) {
977 * Take a refcount on the mpol, lookup_node()
978 * will drop the mmap_lock, so after calling
979 * lookup_node() only "pol" remains valid, "vma"
985 err = lookup_node(mm, addr);
989 } else if (pol == current->mempolicy &&
990 pol->mode == MPOL_INTERLEAVE) {
991 *policy = next_node_in(current->il_prev, pol->nodes);
997 *policy = pol == &default_policy ? MPOL_DEFAULT :
1000 * Internal mempolicy flags must be masked off before exposing
1001 * the policy to userspace.
1003 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1008 if (mpol_store_user_nodemask(pol)) {
1009 *nmask = pol->w.user_nodemask;
1012 get_policy_nodemask(pol, nmask);
1013 task_unlock(current);
1020 mmap_read_unlock(mm);
1022 mpol_put(pol_refcount);
1026 #ifdef CONFIG_MIGRATION
1028 * page migration, thp tail pages can be passed.
1030 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1031 unsigned long flags)
1033 struct page *head = compound_head(page);
1035 * Avoid migrating a page that is shared with others.
1037 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1038 if (!isolate_lru_page(head)) {
1039 list_add_tail(&head->lru, pagelist);
1040 mod_node_page_state(page_pgdat(head),
1041 NR_ISOLATED_ANON + page_is_file_lru(head),
1042 thp_nr_pages(head));
1043 } else if (flags & MPOL_MF_STRICT) {
1045 * Non-movable page may reach here. And, there may be
1046 * temporary off LRU pages or non-LRU movable pages.
1047 * Treat them as unmovable pages since they can't be
1048 * isolated, so they can't be moved at the moment. It
1049 * should return -EIO for this case too.
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1062 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1066 LIST_HEAD(pagelist);
1068 struct migration_target_control mtc = {
1070 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1074 node_set(source, nmask);
1077 * This does not "check" the range but isolates all pages that
1078 * need migration. Between passing in the full user address
1079 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1081 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1082 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1083 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1085 if (!list_empty(&pagelist)) {
1086 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1087 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1089 putback_movable_pages(&pagelist);
1096 * Move pages between the two nodesets so as to preserve the physical
1097 * layout as much as possible.
1099 * Returns the number of page that could not be moved.
1101 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1102 const nodemask_t *to, int flags)
1108 lru_cache_disable();
1113 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1114 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1115 * bit in 'tmp', and return that <source, dest> pair for migration.
1116 * The pair of nodemasks 'to' and 'from' define the map.
1118 * If no pair of bits is found that way, fallback to picking some
1119 * pair of 'source' and 'dest' bits that are not the same. If the
1120 * 'source' and 'dest' bits are the same, this represents a node
1121 * that will be migrating to itself, so no pages need move.
1123 * If no bits are left in 'tmp', or if all remaining bits left
1124 * in 'tmp' correspond to the same bit in 'to', return false
1125 * (nothing left to migrate).
1127 * This lets us pick a pair of nodes to migrate between, such that
1128 * if possible the dest node is not already occupied by some other
1129 * source node, minimizing the risk of overloading the memory on a
1130 * node that would happen if we migrated incoming memory to a node
1131 * before migrating outgoing memory source that same node.
1133 * A single scan of tmp is sufficient. As we go, we remember the
1134 * most recent <s, d> pair that moved (s != d). If we find a pair
1135 * that not only moved, but what's better, moved to an empty slot
1136 * (d is not set in tmp), then we break out then, with that pair.
1137 * Otherwise when we finish scanning from_tmp, we at least have the
1138 * most recent <s, d> pair that moved. If we get all the way through
1139 * the scan of tmp without finding any node that moved, much less
1140 * moved to an empty node, then there is nothing left worth migrating.
1144 while (!nodes_empty(tmp)) {
1146 int source = NUMA_NO_NODE;
1149 for_each_node_mask(s, tmp) {
1152 * do_migrate_pages() tries to maintain the relative
1153 * node relationship of the pages established between
1154 * threads and memory areas.
1156 * However if the number of source nodes is not equal to
1157 * the number of destination nodes we can not preserve
1158 * this node relative relationship. In that case, skip
1159 * copying memory from a node that is in the destination
1162 * Example: [2,3,4] -> [3,4,5] moves everything.
1163 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1166 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1167 (node_isset(s, *to)))
1170 d = node_remap(s, *from, *to);
1174 source = s; /* Node moved. Memorize */
1177 /* dest not in remaining from nodes? */
1178 if (!node_isset(dest, tmp))
1181 if (source == NUMA_NO_NODE)
1184 node_clear(source, tmp);
1185 err = migrate_to_node(mm, source, dest, flags);
1191 mmap_read_unlock(mm);
1201 * Allocate a new page for page migration based on vma policy.
1202 * Start by assuming the page is mapped by the same vma as contains @start.
1203 * Search forward from there, if not. N.B., this assumes that the
1204 * list of pages handed to migrate_pages()--which is how we get here--
1205 * is in virtual address order.
1207 static struct page *new_page(struct page *page, unsigned long start)
1209 struct vm_area_struct *vma;
1210 unsigned long address;
1212 vma = find_vma(current->mm, start);
1214 address = page_address_in_vma(page, vma);
1215 if (address != -EFAULT)
1220 if (PageHuge(page)) {
1221 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1223 } else if (PageTransHuge(page)) {
1226 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1230 prep_transhuge_page(thp);
1234 * if !vma, alloc_page_vma() will use task or system default policy
1236 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1241 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1242 unsigned long flags)
1247 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1248 const nodemask_t *to, int flags)
1253 static struct page *new_page(struct page *page, unsigned long start)
1259 static long do_mbind(unsigned long start, unsigned long len,
1260 unsigned short mode, unsigned short mode_flags,
1261 nodemask_t *nmask, unsigned long flags)
1263 struct mm_struct *mm = current->mm;
1264 struct mempolicy *new;
1268 LIST_HEAD(pagelist);
1270 if (flags & ~(unsigned long)MPOL_MF_VALID)
1272 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1275 if (start & ~PAGE_MASK)
1278 if (mode == MPOL_DEFAULT)
1279 flags &= ~MPOL_MF_STRICT;
1281 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1289 new = mpol_new(mode, mode_flags, nmask);
1291 return PTR_ERR(new);
1293 if (flags & MPOL_MF_LAZY)
1294 new->flags |= MPOL_F_MOF;
1297 * If we are using the default policy then operation
1298 * on discontinuous address spaces is okay after all
1301 flags |= MPOL_MF_DISCONTIG_OK;
1303 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1304 start, start + len, mode, mode_flags,
1305 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1307 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1309 lru_cache_disable();
1312 NODEMASK_SCRATCH(scratch);
1314 mmap_write_lock(mm);
1315 err = mpol_set_nodemask(new, nmask, scratch);
1317 mmap_write_unlock(mm);
1320 NODEMASK_SCRATCH_FREE(scratch);
1325 ret = queue_pages_range(mm, start, end, nmask,
1326 flags | MPOL_MF_INVERT, &pagelist);
1333 err = mbind_range(mm, start, end, new);
1338 if (!list_empty(&pagelist)) {
1339 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1340 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1341 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1343 putback_movable_pages(&pagelist);
1346 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1350 if (!list_empty(&pagelist))
1351 putback_movable_pages(&pagelist);
1354 mmap_write_unlock(mm);
1357 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1363 * User space interface with variable sized bitmaps for nodelists.
1366 /* Copy a node mask from user space. */
1367 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1368 unsigned long maxnode)
1372 unsigned long nlongs;
1373 unsigned long endmask;
1376 nodes_clear(*nodes);
1377 if (maxnode == 0 || !nmask)
1379 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1382 nlongs = BITS_TO_LONGS(maxnode);
1383 if ((maxnode % BITS_PER_LONG) == 0)
1386 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1389 * When the user specified more nodes than supported just check
1390 * if the non supported part is all zero.
1392 * If maxnode have more longs than MAX_NUMNODES, check
1393 * the bits in that area first. And then go through to
1394 * check the rest bits which equal or bigger than MAX_NUMNODES.
1395 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1397 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1398 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1399 if (get_user(t, nmask + k))
1401 if (k == nlongs - 1) {
1407 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1411 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1412 unsigned long valid_mask = endmask;
1414 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1415 if (get_user(t, nmask + nlongs - 1))
1421 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1423 nodes_addr(*nodes)[nlongs-1] &= endmask;
1427 /* Copy a kernel node mask to user space */
1428 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1431 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1432 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1434 if (copy > nbytes) {
1435 if (copy > PAGE_SIZE)
1437 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1441 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1444 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1445 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1447 *flags = *mode & MPOL_MODE_FLAGS;
1448 *mode &= ~MPOL_MODE_FLAGS;
1450 if ((unsigned int)(*mode) >= MPOL_MAX)
1452 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1458 static long kernel_mbind(unsigned long start, unsigned long len,
1459 unsigned long mode, const unsigned long __user *nmask,
1460 unsigned long maxnode, unsigned int flags)
1462 unsigned short mode_flags;
1467 start = untagged_addr(start);
1468 err = sanitize_mpol_flags(&lmode, &mode_flags);
1472 err = get_nodes(&nodes, nmask, maxnode);
1476 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1479 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1480 unsigned long, mode, const unsigned long __user *, nmask,
1481 unsigned long, maxnode, unsigned int, flags)
1483 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1486 /* Set the process memory policy */
1487 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1488 unsigned long maxnode)
1490 unsigned short mode_flags;
1495 err = sanitize_mpol_flags(&lmode, &mode_flags);
1499 err = get_nodes(&nodes, nmask, maxnode);
1503 return do_set_mempolicy(lmode, mode_flags, &nodes);
1506 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1507 unsigned long, maxnode)
1509 return kernel_set_mempolicy(mode, nmask, maxnode);
1512 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1513 const unsigned long __user *old_nodes,
1514 const unsigned long __user *new_nodes)
1516 struct mm_struct *mm = NULL;
1517 struct task_struct *task;
1518 nodemask_t task_nodes;
1522 NODEMASK_SCRATCH(scratch);
1527 old = &scratch->mask1;
1528 new = &scratch->mask2;
1530 err = get_nodes(old, old_nodes, maxnode);
1534 err = get_nodes(new, new_nodes, maxnode);
1538 /* Find the mm_struct */
1540 task = pid ? find_task_by_vpid(pid) : current;
1546 get_task_struct(task);
1551 * Check if this process has the right to modify the specified process.
1552 * Use the regular "ptrace_may_access()" checks.
1554 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1561 task_nodes = cpuset_mems_allowed(task);
1562 /* Is the user allowed to access the target nodes? */
1563 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1568 task_nodes = cpuset_mems_allowed(current);
1569 nodes_and(*new, *new, task_nodes);
1570 if (nodes_empty(*new))
1573 err = security_task_movememory(task);
1577 mm = get_task_mm(task);
1578 put_task_struct(task);
1585 err = do_migrate_pages(mm, old, new,
1586 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1590 NODEMASK_SCRATCH_FREE(scratch);
1595 put_task_struct(task);
1600 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1601 const unsigned long __user *, old_nodes,
1602 const unsigned long __user *, new_nodes)
1604 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1608 /* Retrieve NUMA policy */
1609 static int kernel_get_mempolicy(int __user *policy,
1610 unsigned long __user *nmask,
1611 unsigned long maxnode,
1613 unsigned long flags)
1619 if (nmask != NULL && maxnode < nr_node_ids)
1622 addr = untagged_addr(addr);
1624 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1629 if (policy && put_user(pval, policy))
1633 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1638 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1639 unsigned long __user *, nmask, unsigned long, maxnode,
1640 unsigned long, addr, unsigned long, flags)
1642 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1645 #ifdef CONFIG_COMPAT
1647 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1648 compat_ulong_t __user *, nmask,
1649 compat_ulong_t, maxnode,
1650 compat_ulong_t, addr, compat_ulong_t, flags)
1653 unsigned long __user *nm = NULL;
1654 unsigned long nr_bits, alloc_size;
1655 DECLARE_BITMAP(bm, MAX_NUMNODES);
1657 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1658 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1661 nm = compat_alloc_user_space(alloc_size);
1663 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1665 if (!err && nmask) {
1666 unsigned long copy_size;
1667 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1668 err = copy_from_user(bm, nm, copy_size);
1669 /* ensure entire bitmap is zeroed */
1670 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1671 err |= compat_put_bitmap(nmask, bm, nr_bits);
1677 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1678 compat_ulong_t, maxnode)
1680 unsigned long __user *nm = NULL;
1681 unsigned long nr_bits, alloc_size;
1682 DECLARE_BITMAP(bm, MAX_NUMNODES);
1684 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1685 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1688 if (compat_get_bitmap(bm, nmask, nr_bits))
1690 nm = compat_alloc_user_space(alloc_size);
1691 if (copy_to_user(nm, bm, alloc_size))
1695 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1698 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1699 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1700 compat_ulong_t, maxnode, compat_ulong_t, flags)
1702 unsigned long __user *nm = NULL;
1703 unsigned long nr_bits, alloc_size;
1706 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1707 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1710 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1712 nm = compat_alloc_user_space(alloc_size);
1713 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1717 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1720 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1721 compat_ulong_t, maxnode,
1722 const compat_ulong_t __user *, old_nodes,
1723 const compat_ulong_t __user *, new_nodes)
1725 unsigned long __user *old = NULL;
1726 unsigned long __user *new = NULL;
1727 nodemask_t tmp_mask;
1728 unsigned long nr_bits;
1731 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1732 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1734 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1736 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1738 new = old + size / sizeof(unsigned long);
1739 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1743 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1746 new = compat_alloc_user_space(size);
1747 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1750 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1753 #endif /* CONFIG_COMPAT */
1755 bool vma_migratable(struct vm_area_struct *vma)
1757 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1761 * DAX device mappings require predictable access latency, so avoid
1762 * incurring periodic faults.
1764 if (vma_is_dax(vma))
1767 if (is_vm_hugetlb_page(vma) &&
1768 !hugepage_migration_supported(hstate_vma(vma)))
1772 * Migration allocates pages in the highest zone. If we cannot
1773 * do so then migration (at least from node to node) is not
1777 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1783 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1786 struct mempolicy *pol = NULL;
1789 if (vma->vm_ops && vma->vm_ops->get_policy) {
1790 pol = vma->vm_ops->get_policy(vma, addr);
1791 } else if (vma->vm_policy) {
1792 pol = vma->vm_policy;
1795 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1796 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1797 * count on these policies which will be dropped by
1798 * mpol_cond_put() later
1800 if (mpol_needs_cond_ref(pol))
1809 * get_vma_policy(@vma, @addr)
1810 * @vma: virtual memory area whose policy is sought
1811 * @addr: address in @vma for shared policy lookup
1813 * Returns effective policy for a VMA at specified address.
1814 * Falls back to current->mempolicy or system default policy, as necessary.
1815 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1816 * count--added by the get_policy() vm_op, as appropriate--to protect against
1817 * freeing by another task. It is the caller's responsibility to free the
1818 * extra reference for shared policies.
1820 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1823 struct mempolicy *pol = __get_vma_policy(vma, addr);
1826 pol = get_task_policy(current);
1831 bool vma_policy_mof(struct vm_area_struct *vma)
1833 struct mempolicy *pol;
1835 if (vma->vm_ops && vma->vm_ops->get_policy) {
1838 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1839 if (pol && (pol->flags & MPOL_F_MOF))
1846 pol = vma->vm_policy;
1848 pol = get_task_policy(current);
1850 return pol->flags & MPOL_F_MOF;
1853 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1855 enum zone_type dynamic_policy_zone = policy_zone;
1857 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1860 * if policy->nodes has movable memory only,
1861 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1863 * policy->nodes is intersect with node_states[N_MEMORY].
1864 * so if the following test fails, it implies
1865 * policy->nodes has movable memory only.
1867 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1868 dynamic_policy_zone = ZONE_MOVABLE;
1870 return zone >= dynamic_policy_zone;
1874 * Return a nodemask representing a mempolicy for filtering nodes for
1877 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1879 int mode = policy->mode;
1881 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1882 if (unlikely(mode == MPOL_BIND) &&
1883 apply_policy_zone(policy, gfp_zone(gfp)) &&
1884 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1885 return &policy->nodes;
1887 if (mode == MPOL_PREFERRED_MANY)
1888 return &policy->nodes;
1894 * Return the preferred node id for 'prefer' mempolicy, and return
1895 * the given id for all other policies.
1897 * policy_node() is always coupled with policy_nodemask(), which
1898 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1900 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1902 if (policy->mode == MPOL_PREFERRED) {
1903 nd = first_node(policy->nodes);
1906 * __GFP_THISNODE shouldn't even be used with the bind policy
1907 * because we might easily break the expectation to stay on the
1908 * requested node and not break the policy.
1910 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1916 /* Do dynamic interleaving for a process */
1917 static unsigned interleave_nodes(struct mempolicy *policy)
1920 struct task_struct *me = current;
1922 next = next_node_in(me->il_prev, policy->nodes);
1923 if (next < MAX_NUMNODES)
1929 * Depending on the memory policy provide a node from which to allocate the
1932 unsigned int mempolicy_slab_node(void)
1934 struct mempolicy *policy;
1935 int node = numa_mem_id();
1940 policy = current->mempolicy;
1944 switch (policy->mode) {
1945 case MPOL_PREFERRED:
1946 return first_node(policy->nodes);
1948 case MPOL_INTERLEAVE:
1949 return interleave_nodes(policy);
1952 case MPOL_PREFERRED_MANY:
1957 * Follow bind policy behavior and start allocation at the
1960 struct zonelist *zonelist;
1961 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1962 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1963 z = first_zones_zonelist(zonelist, highest_zoneidx,
1965 return z->zone ? zone_to_nid(z->zone) : node;
1976 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1977 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1978 * number of present nodes.
1980 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1982 unsigned nnodes = nodes_weight(pol->nodes);
1988 return numa_node_id();
1989 target = (unsigned int)n % nnodes;
1990 nid = first_node(pol->nodes);
1991 for (i = 0; i < target; i++)
1992 nid = next_node(nid, pol->nodes);
1996 /* Determine a node number for interleave */
1997 static inline unsigned interleave_nid(struct mempolicy *pol,
1998 struct vm_area_struct *vma, unsigned long addr, int shift)
2004 * for small pages, there is no difference between
2005 * shift and PAGE_SHIFT, so the bit-shift is safe.
2006 * for huge pages, since vm_pgoff is in units of small
2007 * pages, we need to shift off the always 0 bits to get
2010 BUG_ON(shift < PAGE_SHIFT);
2011 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2012 off += (addr - vma->vm_start) >> shift;
2013 return offset_il_node(pol, off);
2015 return interleave_nodes(pol);
2018 #ifdef CONFIG_HUGETLBFS
2020 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2021 * @vma: virtual memory area whose policy is sought
2022 * @addr: address in @vma for shared policy lookup and interleave policy
2023 * @gfp_flags: for requested zone
2024 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2025 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2027 * Returns a nid suitable for a huge page allocation and a pointer
2028 * to the struct mempolicy for conditional unref after allocation.
2029 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2030 * to the mempolicy's @nodemask for filtering the zonelist.
2032 * Must be protected by read_mems_allowed_begin()
2034 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2035 struct mempolicy **mpol, nodemask_t **nodemask)
2040 *mpol = get_vma_policy(vma, addr);
2042 mode = (*mpol)->mode;
2044 if (unlikely(mode == MPOL_INTERLEAVE)) {
2045 nid = interleave_nid(*mpol, vma, addr,
2046 huge_page_shift(hstate_vma(vma)));
2048 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2049 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2050 *nodemask = &(*mpol)->nodes;
2056 * init_nodemask_of_mempolicy
2058 * If the current task's mempolicy is "default" [NULL], return 'false'
2059 * to indicate default policy. Otherwise, extract the policy nodemask
2060 * for 'bind' or 'interleave' policy into the argument nodemask, or
2061 * initialize the argument nodemask to contain the single node for
2062 * 'preferred' or 'local' policy and return 'true' to indicate presence
2063 * of non-default mempolicy.
2065 * We don't bother with reference counting the mempolicy [mpol_get/put]
2066 * because the current task is examining it's own mempolicy and a task's
2067 * mempolicy is only ever changed by the task itself.
2069 * N.B., it is the caller's responsibility to free a returned nodemask.
2071 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2073 struct mempolicy *mempolicy;
2075 if (!(mask && current->mempolicy))
2079 mempolicy = current->mempolicy;
2080 switch (mempolicy->mode) {
2081 case MPOL_PREFERRED:
2082 case MPOL_PREFERRED_MANY:
2084 case MPOL_INTERLEAVE:
2085 *mask = mempolicy->nodes;
2089 init_nodemask_of_node(mask, numa_node_id());
2095 task_unlock(current);
2102 * mempolicy_in_oom_domain
2104 * If tsk's mempolicy is "bind", check for intersection between mask and
2105 * the policy nodemask. Otherwise, return true for all other policies
2106 * including "interleave", as a tsk with "interleave" policy may have
2107 * memory allocated from all nodes in system.
2109 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2111 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2112 const nodemask_t *mask)
2114 struct mempolicy *mempolicy;
2121 mempolicy = tsk->mempolicy;
2122 if (mempolicy && mempolicy->mode == MPOL_BIND)
2123 ret = nodes_intersects(mempolicy->nodes, *mask);
2129 /* Allocate a page in interleaved policy.
2130 Own path because it needs to do special accounting. */
2131 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2136 page = __alloc_pages(gfp, order, nid, NULL);
2137 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2138 if (!static_branch_likely(&vm_numa_stat_key))
2140 if (page && page_to_nid(page) == nid) {
2142 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2148 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2149 int nid, struct mempolicy *pol)
2152 gfp_t preferred_gfp;
2155 * This is a two pass approach. The first pass will only try the
2156 * preferred nodes but skip the direct reclaim and allow the
2157 * allocation to fail, while the second pass will try all the
2160 preferred_gfp = gfp | __GFP_NOWARN;
2161 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2162 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2164 page = __alloc_pages(gfp, order, numa_node_id(), NULL);
2170 * alloc_pages_vma - Allocate a page for a VMA.
2172 * @order: Order of the GFP allocation.
2173 * @vma: Pointer to VMA or NULL if not available.
2174 * @addr: Virtual address of the allocation. Must be inside @vma.
2175 * @node: Which node to prefer for allocation (modulo policy).
2176 * @hugepage: For hugepages try only the preferred node if possible.
2178 * Allocate a page for a specific address in @vma, using the appropriate
2179 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2180 * of the mm_struct of the VMA to prevent it from going away. Should be
2181 * used for all allocations for pages that will be mapped into user space.
2183 * Return: The page on success or NULL if allocation fails.
2185 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2186 unsigned long addr, int node, bool hugepage)
2188 struct mempolicy *pol;
2193 pol = get_vma_policy(vma, addr);
2195 if (pol->mode == MPOL_INTERLEAVE) {
2198 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2200 page = alloc_page_interleave(gfp, order, nid);
2204 if (pol->mode == MPOL_PREFERRED_MANY) {
2205 page = alloc_pages_preferred_many(gfp, order, node, pol);
2210 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2211 int hpage_node = node;
2214 * For hugepage allocation and non-interleave policy which
2215 * allows the current node (or other explicitly preferred
2216 * node) we only try to allocate from the current/preferred
2217 * node and don't fall back to other nodes, as the cost of
2218 * remote accesses would likely offset THP benefits.
2220 * If the policy is interleave or does not allow the current
2221 * node in its nodemask, we allocate the standard way.
2223 if (pol->mode == MPOL_PREFERRED)
2224 hpage_node = first_node(pol->nodes);
2226 nmask = policy_nodemask(gfp, pol);
2227 if (!nmask || node_isset(hpage_node, *nmask)) {
2230 * First, try to allocate THP only on local node, but
2231 * don't reclaim unnecessarily, just compact.
2233 page = __alloc_pages_node(hpage_node,
2234 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2237 * If hugepage allocations are configured to always
2238 * synchronous compact or the vma has been madvised
2239 * to prefer hugepage backing, retry allowing remote
2240 * memory with both reclaim and compact as well.
2242 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2243 page = __alloc_pages_node(hpage_node,
2250 nmask = policy_nodemask(gfp, pol);
2251 preferred_nid = policy_node(gfp, pol, node);
2252 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2257 EXPORT_SYMBOL(alloc_pages_vma);
2260 * alloc_pages - Allocate pages.
2262 * @order: Power of two of number of pages to allocate.
2264 * Allocate 1 << @order contiguous pages. The physical address of the
2265 * first page is naturally aligned (eg an order-3 allocation will be aligned
2266 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2267 * process is honoured when in process context.
2269 * Context: Can be called from any context, providing the appropriate GFP
2271 * Return: The page on success or NULL if allocation fails.
2273 struct page *alloc_pages(gfp_t gfp, unsigned order)
2275 struct mempolicy *pol = &default_policy;
2278 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2279 pol = get_task_policy(current);
2282 * No reference counting needed for current->mempolicy
2283 * nor system default_policy
2285 if (pol->mode == MPOL_INTERLEAVE)
2286 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2287 else if (pol->mode == MPOL_PREFERRED_MANY)
2288 page = alloc_pages_preferred_many(gfp, order,
2289 numa_node_id(), pol);
2291 page = __alloc_pages(gfp, order,
2292 policy_node(gfp, pol, numa_node_id()),
2293 policy_nodemask(gfp, pol));
2297 EXPORT_SYMBOL(alloc_pages);
2299 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2301 struct mempolicy *pol = mpol_dup(vma_policy(src));
2304 return PTR_ERR(pol);
2305 dst->vm_policy = pol;
2310 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2311 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2312 * with the mems_allowed returned by cpuset_mems_allowed(). This
2313 * keeps mempolicies cpuset relative after its cpuset moves. See
2314 * further kernel/cpuset.c update_nodemask().
2316 * current's mempolicy may be rebinded by the other task(the task that changes
2317 * cpuset's mems), so we needn't do rebind work for current task.
2320 /* Slow path of a mempolicy duplicate */
2321 struct mempolicy *__mpol_dup(struct mempolicy *old)
2323 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2326 return ERR_PTR(-ENOMEM);
2328 /* task's mempolicy is protected by alloc_lock */
2329 if (old == current->mempolicy) {
2332 task_unlock(current);
2336 if (current_cpuset_is_being_rebound()) {
2337 nodemask_t mems = cpuset_mems_allowed(current);
2338 mpol_rebind_policy(new, &mems);
2340 atomic_set(&new->refcnt, 1);
2344 /* Slow path of a mempolicy comparison */
2345 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2349 if (a->mode != b->mode)
2351 if (a->flags != b->flags)
2353 if (mpol_store_user_nodemask(a))
2354 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2359 case MPOL_INTERLEAVE:
2360 case MPOL_PREFERRED:
2361 case MPOL_PREFERRED_MANY:
2362 return !!nodes_equal(a->nodes, b->nodes);
2372 * Shared memory backing store policy support.
2374 * Remember policies even when nobody has shared memory mapped.
2375 * The policies are kept in Red-Black tree linked from the inode.
2376 * They are protected by the sp->lock rwlock, which should be held
2377 * for any accesses to the tree.
2381 * lookup first element intersecting start-end. Caller holds sp->lock for
2382 * reading or for writing
2384 static struct sp_node *
2385 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2387 struct rb_node *n = sp->root.rb_node;
2390 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2392 if (start >= p->end)
2394 else if (end <= p->start)
2402 struct sp_node *w = NULL;
2403 struct rb_node *prev = rb_prev(n);
2406 w = rb_entry(prev, struct sp_node, nd);
2407 if (w->end <= start)
2411 return rb_entry(n, struct sp_node, nd);
2415 * Insert a new shared policy into the list. Caller holds sp->lock for
2418 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2420 struct rb_node **p = &sp->root.rb_node;
2421 struct rb_node *parent = NULL;
2426 nd = rb_entry(parent, struct sp_node, nd);
2427 if (new->start < nd->start)
2429 else if (new->end > nd->end)
2430 p = &(*p)->rb_right;
2434 rb_link_node(&new->nd, parent, p);
2435 rb_insert_color(&new->nd, &sp->root);
2436 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2437 new->policy ? new->policy->mode : 0);
2440 /* Find shared policy intersecting idx */
2442 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2444 struct mempolicy *pol = NULL;
2447 if (!sp->root.rb_node)
2449 read_lock(&sp->lock);
2450 sn = sp_lookup(sp, idx, idx+1);
2452 mpol_get(sn->policy);
2455 read_unlock(&sp->lock);
2459 static void sp_free(struct sp_node *n)
2461 mpol_put(n->policy);
2462 kmem_cache_free(sn_cache, n);
2466 * mpol_misplaced - check whether current page node is valid in policy
2468 * @page: page to be checked
2469 * @vma: vm area where page mapped
2470 * @addr: virtual address where page mapped
2472 * Lookup current policy node id for vma,addr and "compare to" page's
2473 * node id. Policy determination "mimics" alloc_page_vma().
2474 * Called from fault path where we know the vma and faulting address.
2476 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2477 * policy, or a suitable node ID to allocate a replacement page from.
2479 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2481 struct mempolicy *pol;
2483 int curnid = page_to_nid(page);
2484 unsigned long pgoff;
2485 int thiscpu = raw_smp_processor_id();
2486 int thisnid = cpu_to_node(thiscpu);
2487 int polnid = NUMA_NO_NODE;
2488 int ret = NUMA_NO_NODE;
2490 pol = get_vma_policy(vma, addr);
2491 if (!(pol->flags & MPOL_F_MOF))
2494 switch (pol->mode) {
2495 case MPOL_INTERLEAVE:
2496 pgoff = vma->vm_pgoff;
2497 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2498 polnid = offset_il_node(pol, pgoff);
2501 case MPOL_PREFERRED:
2502 if (node_isset(curnid, pol->nodes))
2504 polnid = first_node(pol->nodes);
2508 polnid = numa_node_id();
2512 /* Optimize placement among multiple nodes via NUMA balancing */
2513 if (pol->flags & MPOL_F_MORON) {
2514 if (node_isset(thisnid, pol->nodes))
2520 case MPOL_PREFERRED_MANY:
2522 * use current page if in policy nodemask,
2523 * else select nearest allowed node, if any.
2524 * If no allowed nodes, use current [!misplaced].
2526 if (node_isset(curnid, pol->nodes))
2528 z = first_zones_zonelist(
2529 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2530 gfp_zone(GFP_HIGHUSER),
2532 polnid = zone_to_nid(z->zone);
2539 /* Migrate the page towards the node whose CPU is referencing it */
2540 if (pol->flags & MPOL_F_MORON) {
2543 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2547 if (curnid != polnid)
2556 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2557 * dropped after task->mempolicy is set to NULL so that any allocation done as
2558 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2561 void mpol_put_task_policy(struct task_struct *task)
2563 struct mempolicy *pol;
2566 pol = task->mempolicy;
2567 task->mempolicy = NULL;
2572 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2574 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2575 rb_erase(&n->nd, &sp->root);
2579 static void sp_node_init(struct sp_node *node, unsigned long start,
2580 unsigned long end, struct mempolicy *pol)
2582 node->start = start;
2587 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2588 struct mempolicy *pol)
2591 struct mempolicy *newpol;
2593 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2597 newpol = mpol_dup(pol);
2598 if (IS_ERR(newpol)) {
2599 kmem_cache_free(sn_cache, n);
2602 newpol->flags |= MPOL_F_SHARED;
2603 sp_node_init(n, start, end, newpol);
2608 /* Replace a policy range. */
2609 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2610 unsigned long end, struct sp_node *new)
2613 struct sp_node *n_new = NULL;
2614 struct mempolicy *mpol_new = NULL;
2618 write_lock(&sp->lock);
2619 n = sp_lookup(sp, start, end);
2620 /* Take care of old policies in the same range. */
2621 while (n && n->start < end) {
2622 struct rb_node *next = rb_next(&n->nd);
2623 if (n->start >= start) {
2629 /* Old policy spanning whole new range. */
2634 *mpol_new = *n->policy;
2635 atomic_set(&mpol_new->refcnt, 1);
2636 sp_node_init(n_new, end, n->end, mpol_new);
2638 sp_insert(sp, n_new);
2647 n = rb_entry(next, struct sp_node, nd);
2651 write_unlock(&sp->lock);
2658 kmem_cache_free(sn_cache, n_new);
2663 write_unlock(&sp->lock);
2665 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2668 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2675 * mpol_shared_policy_init - initialize shared policy for inode
2676 * @sp: pointer to inode shared policy
2677 * @mpol: struct mempolicy to install
2679 * Install non-NULL @mpol in inode's shared policy rb-tree.
2680 * On entry, the current task has a reference on a non-NULL @mpol.
2681 * This must be released on exit.
2682 * This is called at get_inode() calls and we can use GFP_KERNEL.
2684 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2688 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2689 rwlock_init(&sp->lock);
2692 struct vm_area_struct pvma;
2693 struct mempolicy *new;
2694 NODEMASK_SCRATCH(scratch);
2698 /* contextualize the tmpfs mount point mempolicy */
2699 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2701 goto free_scratch; /* no valid nodemask intersection */
2704 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2705 task_unlock(current);
2709 /* Create pseudo-vma that contains just the policy */
2710 vma_init(&pvma, NULL);
2711 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2712 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2715 mpol_put(new); /* drop initial ref */
2717 NODEMASK_SCRATCH_FREE(scratch);
2719 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2723 int mpol_set_shared_policy(struct shared_policy *info,
2724 struct vm_area_struct *vma, struct mempolicy *npol)
2727 struct sp_node *new = NULL;
2728 unsigned long sz = vma_pages(vma);
2730 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2732 sz, npol ? npol->mode : -1,
2733 npol ? npol->flags : -1,
2734 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2737 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2741 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2747 /* Free a backing policy store on inode delete. */
2748 void mpol_free_shared_policy(struct shared_policy *p)
2751 struct rb_node *next;
2753 if (!p->root.rb_node)
2755 write_lock(&p->lock);
2756 next = rb_first(&p->root);
2758 n = rb_entry(next, struct sp_node, nd);
2759 next = rb_next(&n->nd);
2762 write_unlock(&p->lock);
2765 #ifdef CONFIG_NUMA_BALANCING
2766 static int __initdata numabalancing_override;
2768 static void __init check_numabalancing_enable(void)
2770 bool numabalancing_default = false;
2772 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2773 numabalancing_default = true;
2775 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2776 if (numabalancing_override)
2777 set_numabalancing_state(numabalancing_override == 1);
2779 if (num_online_nodes() > 1 && !numabalancing_override) {
2780 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2781 numabalancing_default ? "Enabling" : "Disabling");
2782 set_numabalancing_state(numabalancing_default);
2786 static int __init setup_numabalancing(char *str)
2792 if (!strcmp(str, "enable")) {
2793 numabalancing_override = 1;
2795 } else if (!strcmp(str, "disable")) {
2796 numabalancing_override = -1;
2801 pr_warn("Unable to parse numa_balancing=\n");
2805 __setup("numa_balancing=", setup_numabalancing);
2807 static inline void __init check_numabalancing_enable(void)
2810 #endif /* CONFIG_NUMA_BALANCING */
2812 /* assumes fs == KERNEL_DS */
2813 void __init numa_policy_init(void)
2815 nodemask_t interleave_nodes;
2816 unsigned long largest = 0;
2817 int nid, prefer = 0;
2819 policy_cache = kmem_cache_create("numa_policy",
2820 sizeof(struct mempolicy),
2821 0, SLAB_PANIC, NULL);
2823 sn_cache = kmem_cache_create("shared_policy_node",
2824 sizeof(struct sp_node),
2825 0, SLAB_PANIC, NULL);
2827 for_each_node(nid) {
2828 preferred_node_policy[nid] = (struct mempolicy) {
2829 .refcnt = ATOMIC_INIT(1),
2830 .mode = MPOL_PREFERRED,
2831 .flags = MPOL_F_MOF | MPOL_F_MORON,
2832 .nodes = nodemask_of_node(nid),
2837 * Set interleaving policy for system init. Interleaving is only
2838 * enabled across suitably sized nodes (default is >= 16MB), or
2839 * fall back to the largest node if they're all smaller.
2841 nodes_clear(interleave_nodes);
2842 for_each_node_state(nid, N_MEMORY) {
2843 unsigned long total_pages = node_present_pages(nid);
2845 /* Preserve the largest node */
2846 if (largest < total_pages) {
2847 largest = total_pages;
2851 /* Interleave this node? */
2852 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2853 node_set(nid, interleave_nodes);
2856 /* All too small, use the largest */
2857 if (unlikely(nodes_empty(interleave_nodes)))
2858 node_set(prefer, interleave_nodes);
2860 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2861 pr_err("%s: interleaving failed\n", __func__);
2863 check_numabalancing_enable();
2866 /* Reset policy of current process to default */
2867 void numa_default_policy(void)
2869 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2873 * Parse and format mempolicy from/to strings
2876 static const char * const policy_modes[] =
2878 [MPOL_DEFAULT] = "default",
2879 [MPOL_PREFERRED] = "prefer",
2880 [MPOL_BIND] = "bind",
2881 [MPOL_INTERLEAVE] = "interleave",
2882 [MPOL_LOCAL] = "local",
2883 [MPOL_PREFERRED_MANY] = "prefer (many)",
2889 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2890 * @str: string containing mempolicy to parse
2891 * @mpol: pointer to struct mempolicy pointer, returned on success.
2894 * <mode>[=<flags>][:<nodelist>]
2896 * On success, returns 0, else 1
2898 int mpol_parse_str(char *str, struct mempolicy **mpol)
2900 struct mempolicy *new = NULL;
2901 unsigned short mode_flags;
2903 char *nodelist = strchr(str, ':');
2904 char *flags = strchr(str, '=');
2908 *flags++ = '\0'; /* terminate mode string */
2911 /* NUL-terminate mode or flags string */
2913 if (nodelist_parse(nodelist, nodes))
2915 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2920 mode = match_string(policy_modes, MPOL_MAX, str);
2925 case MPOL_PREFERRED:
2927 * Insist on a nodelist of one node only, although later
2928 * we use first_node(nodes) to grab a single node, so here
2929 * nodelist (or nodes) cannot be empty.
2932 char *rest = nodelist;
2933 while (isdigit(*rest))
2937 if (nodes_empty(nodes))
2941 case MPOL_INTERLEAVE:
2943 * Default to online nodes with memory if no nodelist
2946 nodes = node_states[N_MEMORY];
2950 * Don't allow a nodelist; mpol_new() checks flags
2957 * Insist on a empty nodelist
2962 case MPOL_PREFERRED_MANY:
2965 * Insist on a nodelist
2974 * Currently, we only support two mutually exclusive
2977 if (!strcmp(flags, "static"))
2978 mode_flags |= MPOL_F_STATIC_NODES;
2979 else if (!strcmp(flags, "relative"))
2980 mode_flags |= MPOL_F_RELATIVE_NODES;
2985 new = mpol_new(mode, mode_flags, &nodes);
2990 * Save nodes for mpol_to_str() to show the tmpfs mount options
2991 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2993 if (mode != MPOL_PREFERRED) {
2995 } else if (nodelist) {
2996 nodes_clear(new->nodes);
2997 node_set(first_node(nodes), new->nodes);
2999 new->mode = MPOL_LOCAL;
3003 * Save nodes for contextualization: this will be used to "clone"
3004 * the mempolicy in a specific context [cpuset] at a later time.
3006 new->w.user_nodemask = nodes;
3011 /* Restore string for error message */
3020 #endif /* CONFIG_TMPFS */
3023 * mpol_to_str - format a mempolicy structure for printing
3024 * @buffer: to contain formatted mempolicy string
3025 * @maxlen: length of @buffer
3026 * @pol: pointer to mempolicy to be formatted
3028 * Convert @pol into a string. If @buffer is too short, truncate the string.
3029 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3030 * longest flag, "relative", and to display at least a few node ids.
3032 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3035 nodemask_t nodes = NODE_MASK_NONE;
3036 unsigned short mode = MPOL_DEFAULT;
3037 unsigned short flags = 0;
3039 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3048 case MPOL_PREFERRED:
3049 case MPOL_PREFERRED_MANY:
3051 case MPOL_INTERLEAVE:
3056 snprintf(p, maxlen, "unknown");
3060 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3062 if (flags & MPOL_MODE_FLAGS) {
3063 p += snprintf(p, buffer + maxlen - p, "=");
3066 * Currently, the only defined flags are mutually exclusive
3068 if (flags & MPOL_F_STATIC_NODES)
3069 p += snprintf(p, buffer + maxlen - p, "static");
3070 else if (flags & MPOL_F_RELATIVE_NODES)
3071 p += snprintf(p, buffer + maxlen - p, "relative");
3074 if (!nodes_empty(nodes))
3075 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3076 nodemask_pr_args(&nodes));
3079 bool numa_demotion_enabled = false;
3082 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
3083 struct kobj_attribute *attr, char *buf)
3085 return sysfs_emit(buf, "%s\n",
3086 numa_demotion_enabled? "true" : "false");
3089 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
3090 struct kobj_attribute *attr,
3091 const char *buf, size_t count)
3093 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
3094 numa_demotion_enabled = true;
3095 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
3096 numa_demotion_enabled = false;
3103 static struct kobj_attribute numa_demotion_enabled_attr =
3104 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3105 numa_demotion_enabled_store);
3107 static struct attribute *numa_attrs[] = {
3108 &numa_demotion_enabled_attr.attr,
3112 static const struct attribute_group numa_attr_group = {
3113 .attrs = numa_attrs,
3116 static int __init numa_init_sysfs(void)
3119 struct kobject *numa_kobj;
3121 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3123 pr_err("failed to create numa kobject\n");
3126 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3128 pr_err("failed to register numa group\n");
3134 kobject_put(numa_kobj);
3137 subsys_initcall(numa_init_sysfs);