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,
818 if (mpol_equal(vma_policy(vma), new_pol))
820 /* vma_merge() joined vma && vma->next, case 8 */
823 if (vma->vm_start != vmstart) {
824 err = split_vma(vma->vm_mm, vma, vmstart, 1);
828 if (vma->vm_end != vmend) {
829 err = split_vma(vma->vm_mm, vma, vmend, 0);
834 err = vma_replace_policy(vma, new_pol);
843 /* Set the process memory policy */
844 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
847 struct mempolicy *new, *old;
848 NODEMASK_SCRATCH(scratch);
854 new = mpol_new(mode, flags, nodes);
860 ret = mpol_set_nodemask(new, nodes, scratch);
866 old = current->mempolicy;
867 current->mempolicy = new;
868 if (new && new->mode == MPOL_INTERLEAVE)
869 current->il_prev = MAX_NUMNODES-1;
870 task_unlock(current);
874 NODEMASK_SCRATCH_FREE(scratch);
879 * Return nodemask for policy for get_mempolicy() query
881 * Called with task's alloc_lock held
883 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
886 if (p == &default_policy)
891 case MPOL_INTERLEAVE:
893 case MPOL_PREFERRED_MANY:
897 /* return empty node mask for local allocation */
904 static int lookup_node(struct mm_struct *mm, unsigned long addr)
906 struct page *p = NULL;
910 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
912 err = page_to_nid(p);
916 mmap_read_unlock(mm);
920 /* Retrieve NUMA policy */
921 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
922 unsigned long addr, unsigned long flags)
925 struct mm_struct *mm = current->mm;
926 struct vm_area_struct *vma = NULL;
927 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
930 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
933 if (flags & MPOL_F_MEMS_ALLOWED) {
934 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
936 *policy = 0; /* just so it's initialized */
938 *nmask = cpuset_current_mems_allowed;
939 task_unlock(current);
943 if (flags & MPOL_F_ADDR) {
945 * Do NOT fall back to task policy if the
946 * vma/shared policy at addr is NULL. We
947 * want to return MPOL_DEFAULT in this case.
950 vma = vma_lookup(mm, addr);
952 mmap_read_unlock(mm);
955 if (vma->vm_ops && vma->vm_ops->get_policy)
956 pol = vma->vm_ops->get_policy(vma, addr);
958 pol = vma->vm_policy;
963 pol = &default_policy; /* indicates default behavior */
965 if (flags & MPOL_F_NODE) {
966 if (flags & MPOL_F_ADDR) {
968 * Take a refcount on the mpol, lookup_node()
969 * will drop the mmap_lock, so after calling
970 * lookup_node() only "pol" remains valid, "vma"
976 err = lookup_node(mm, addr);
980 } else if (pol == current->mempolicy &&
981 pol->mode == MPOL_INTERLEAVE) {
982 *policy = next_node_in(current->il_prev, pol->nodes);
988 *policy = pol == &default_policy ? MPOL_DEFAULT :
991 * Internal mempolicy flags must be masked off before exposing
992 * the policy to userspace.
994 *policy |= (pol->flags & MPOL_MODE_FLAGS);
999 if (mpol_store_user_nodemask(pol)) {
1000 *nmask = pol->w.user_nodemask;
1003 get_policy_nodemask(pol, nmask);
1004 task_unlock(current);
1011 mmap_read_unlock(mm);
1013 mpol_put(pol_refcount);
1017 #ifdef CONFIG_MIGRATION
1019 * page migration, thp tail pages can be passed.
1021 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1022 unsigned long flags)
1024 struct page *head = compound_head(page);
1026 * Avoid migrating a page that is shared with others.
1028 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1029 if (!isolate_lru_page(head)) {
1030 list_add_tail(&head->lru, pagelist);
1031 mod_node_page_state(page_pgdat(head),
1032 NR_ISOLATED_ANON + page_is_file_lru(head),
1033 thp_nr_pages(head));
1034 } else if (flags & MPOL_MF_STRICT) {
1036 * Non-movable page may reach here. And, there may be
1037 * temporary off LRU pages or non-LRU movable pages.
1038 * Treat them as unmovable pages since they can't be
1039 * isolated, so they can't be moved at the moment. It
1040 * should return -EIO for this case too.
1050 * Migrate pages from one node to a target node.
1051 * Returns error or the number of pages not migrated.
1053 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1057 LIST_HEAD(pagelist);
1059 struct migration_target_control mtc = {
1061 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1065 node_set(source, nmask);
1068 * This does not "check" the range but isolates all pages that
1069 * need migration. Between passing in the full user address
1070 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1072 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1073 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1074 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1076 if (!list_empty(&pagelist)) {
1077 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1078 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1080 putback_movable_pages(&pagelist);
1087 * Move pages between the two nodesets so as to preserve the physical
1088 * layout as much as possible.
1090 * Returns the number of page that could not be moved.
1092 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1093 const nodemask_t *to, int flags)
1099 lru_cache_disable();
1104 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1105 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1106 * bit in 'tmp', and return that <source, dest> pair for migration.
1107 * The pair of nodemasks 'to' and 'from' define the map.
1109 * If no pair of bits is found that way, fallback to picking some
1110 * pair of 'source' and 'dest' bits that are not the same. If the
1111 * 'source' and 'dest' bits are the same, this represents a node
1112 * that will be migrating to itself, so no pages need move.
1114 * If no bits are left in 'tmp', or if all remaining bits left
1115 * in 'tmp' correspond to the same bit in 'to', return false
1116 * (nothing left to migrate).
1118 * This lets us pick a pair of nodes to migrate between, such that
1119 * if possible the dest node is not already occupied by some other
1120 * source node, minimizing the risk of overloading the memory on a
1121 * node that would happen if we migrated incoming memory to a node
1122 * before migrating outgoing memory source that same node.
1124 * A single scan of tmp is sufficient. As we go, we remember the
1125 * most recent <s, d> pair that moved (s != d). If we find a pair
1126 * that not only moved, but what's better, moved to an empty slot
1127 * (d is not set in tmp), then we break out then, with that pair.
1128 * Otherwise when we finish scanning from_tmp, we at least have the
1129 * most recent <s, d> pair that moved. If we get all the way through
1130 * the scan of tmp without finding any node that moved, much less
1131 * moved to an empty node, then there is nothing left worth migrating.
1135 while (!nodes_empty(tmp)) {
1137 int source = NUMA_NO_NODE;
1140 for_each_node_mask(s, tmp) {
1143 * do_migrate_pages() tries to maintain the relative
1144 * node relationship of the pages established between
1145 * threads and memory areas.
1147 * However if the number of source nodes is not equal to
1148 * the number of destination nodes we can not preserve
1149 * this node relative relationship. In that case, skip
1150 * copying memory from a node that is in the destination
1153 * Example: [2,3,4] -> [3,4,5] moves everything.
1154 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1157 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1158 (node_isset(s, *to)))
1161 d = node_remap(s, *from, *to);
1165 source = s; /* Node moved. Memorize */
1168 /* dest not in remaining from nodes? */
1169 if (!node_isset(dest, tmp))
1172 if (source == NUMA_NO_NODE)
1175 node_clear(source, tmp);
1176 err = migrate_to_node(mm, source, dest, flags);
1182 mmap_read_unlock(mm);
1192 * Allocate a new page for page migration based on vma policy.
1193 * Start by assuming the page is mapped by the same vma as contains @start.
1194 * Search forward from there, if not. N.B., this assumes that the
1195 * list of pages handed to migrate_pages()--which is how we get here--
1196 * is in virtual address order.
1198 static struct page *new_page(struct page *page, unsigned long start)
1200 struct vm_area_struct *vma;
1201 unsigned long address;
1203 vma = find_vma(current->mm, start);
1205 address = page_address_in_vma(page, vma);
1206 if (address != -EFAULT)
1211 if (PageHuge(page)) {
1212 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1214 } else if (PageTransHuge(page)) {
1217 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1221 prep_transhuge_page(thp);
1225 * if !vma, alloc_page_vma() will use task or system default policy
1227 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1232 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1233 unsigned long flags)
1238 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1239 const nodemask_t *to, int flags)
1244 static struct page *new_page(struct page *page, unsigned long start)
1250 static long do_mbind(unsigned long start, unsigned long len,
1251 unsigned short mode, unsigned short mode_flags,
1252 nodemask_t *nmask, unsigned long flags)
1254 struct mm_struct *mm = current->mm;
1255 struct mempolicy *new;
1259 LIST_HEAD(pagelist);
1261 if (flags & ~(unsigned long)MPOL_MF_VALID)
1263 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1266 if (start & ~PAGE_MASK)
1269 if (mode == MPOL_DEFAULT)
1270 flags &= ~MPOL_MF_STRICT;
1272 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1280 new = mpol_new(mode, mode_flags, nmask);
1282 return PTR_ERR(new);
1284 if (flags & MPOL_MF_LAZY)
1285 new->flags |= MPOL_F_MOF;
1288 * If we are using the default policy then operation
1289 * on discontinuous address spaces is okay after all
1292 flags |= MPOL_MF_DISCONTIG_OK;
1294 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1295 start, start + len, mode, mode_flags,
1296 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1298 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1300 lru_cache_disable();
1303 NODEMASK_SCRATCH(scratch);
1305 mmap_write_lock(mm);
1306 err = mpol_set_nodemask(new, nmask, scratch);
1308 mmap_write_unlock(mm);
1311 NODEMASK_SCRATCH_FREE(scratch);
1316 ret = queue_pages_range(mm, start, end, nmask,
1317 flags | MPOL_MF_INVERT, &pagelist);
1324 err = mbind_range(mm, start, end, new);
1329 if (!list_empty(&pagelist)) {
1330 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1331 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1332 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1334 putback_movable_pages(&pagelist);
1337 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1341 if (!list_empty(&pagelist))
1342 putback_movable_pages(&pagelist);
1345 mmap_write_unlock(mm);
1348 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1354 * User space interface with variable sized bitmaps for nodelists.
1356 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1357 unsigned long maxnode)
1359 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1362 if (in_compat_syscall())
1363 ret = compat_get_bitmap(mask,
1364 (const compat_ulong_t __user *)nmask,
1367 ret = copy_from_user(mask, nmask,
1368 nlongs * sizeof(unsigned long));
1373 if (maxnode % BITS_PER_LONG)
1374 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1379 /* Copy a node mask from user space. */
1380 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1381 unsigned long maxnode)
1384 nodes_clear(*nodes);
1385 if (maxnode == 0 || !nmask)
1387 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1391 * When the user specified more nodes than supported just check
1392 * if the non supported part is all zero, one word at a time,
1393 * starting at the end.
1395 while (maxnode > MAX_NUMNODES) {
1396 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1399 if (get_bitmap(&t, &nmask[maxnode / BITS_PER_LONG], bits))
1402 if (maxnode - bits >= MAX_NUMNODES) {
1405 maxnode = MAX_NUMNODES;
1406 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1412 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1415 /* Copy a kernel node mask to user space */
1416 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1419 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1420 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1421 bool compat = in_compat_syscall();
1424 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1426 if (copy > nbytes) {
1427 if (copy > PAGE_SIZE)
1429 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1432 maxnode = nr_node_ids;
1436 return compat_put_bitmap((compat_ulong_t __user *)mask,
1437 nodes_addr(*nodes), maxnode);
1439 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1442 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1443 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1445 *flags = *mode & MPOL_MODE_FLAGS;
1446 *mode &= ~MPOL_MODE_FLAGS;
1448 if ((unsigned int)(*mode) >= MPOL_MAX)
1450 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1452 if (*flags & MPOL_F_NUMA_BALANCING) {
1453 if (*mode != MPOL_BIND)
1455 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1460 static long kernel_mbind(unsigned long start, unsigned long len,
1461 unsigned long mode, const unsigned long __user *nmask,
1462 unsigned long maxnode, unsigned int flags)
1464 unsigned short mode_flags;
1469 start = untagged_addr(start);
1470 err = sanitize_mpol_flags(&lmode, &mode_flags);
1474 err = get_nodes(&nodes, nmask, maxnode);
1478 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1481 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1482 unsigned long, mode, const unsigned long __user *, nmask,
1483 unsigned long, maxnode, unsigned int, flags)
1485 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1488 /* Set the process memory policy */
1489 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1490 unsigned long maxnode)
1492 unsigned short mode_flags;
1497 err = sanitize_mpol_flags(&lmode, &mode_flags);
1501 err = get_nodes(&nodes, nmask, maxnode);
1505 return do_set_mempolicy(lmode, mode_flags, &nodes);
1508 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1509 unsigned long, maxnode)
1511 return kernel_set_mempolicy(mode, nmask, maxnode);
1514 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1515 const unsigned long __user *old_nodes,
1516 const unsigned long __user *new_nodes)
1518 struct mm_struct *mm = NULL;
1519 struct task_struct *task;
1520 nodemask_t task_nodes;
1524 NODEMASK_SCRATCH(scratch);
1529 old = &scratch->mask1;
1530 new = &scratch->mask2;
1532 err = get_nodes(old, old_nodes, maxnode);
1536 err = get_nodes(new, new_nodes, maxnode);
1540 /* Find the mm_struct */
1542 task = pid ? find_task_by_vpid(pid) : current;
1548 get_task_struct(task);
1553 * Check if this process has the right to modify the specified process.
1554 * Use the regular "ptrace_may_access()" checks.
1556 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1563 task_nodes = cpuset_mems_allowed(task);
1564 /* Is the user allowed to access the target nodes? */
1565 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1570 task_nodes = cpuset_mems_allowed(current);
1571 nodes_and(*new, *new, task_nodes);
1572 if (nodes_empty(*new))
1575 err = security_task_movememory(task);
1579 mm = get_task_mm(task);
1580 put_task_struct(task);
1587 err = do_migrate_pages(mm, old, new,
1588 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1592 NODEMASK_SCRATCH_FREE(scratch);
1597 put_task_struct(task);
1602 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1603 const unsigned long __user *, old_nodes,
1604 const unsigned long __user *, new_nodes)
1606 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1610 /* Retrieve NUMA policy */
1611 static int kernel_get_mempolicy(int __user *policy,
1612 unsigned long __user *nmask,
1613 unsigned long maxnode,
1615 unsigned long flags)
1621 if (nmask != NULL && maxnode < nr_node_ids)
1624 addr = untagged_addr(addr);
1626 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1631 if (policy && put_user(pval, policy))
1635 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1640 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1641 unsigned long __user *, nmask, unsigned long, maxnode,
1642 unsigned long, addr, unsigned long, flags)
1644 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1647 bool vma_migratable(struct vm_area_struct *vma)
1649 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1653 * DAX device mappings require predictable access latency, so avoid
1654 * incurring periodic faults.
1656 if (vma_is_dax(vma))
1659 if (is_vm_hugetlb_page(vma) &&
1660 !hugepage_migration_supported(hstate_vma(vma)))
1664 * Migration allocates pages in the highest zone. If we cannot
1665 * do so then migration (at least from node to node) is not
1669 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1675 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1678 struct mempolicy *pol = NULL;
1681 if (vma->vm_ops && vma->vm_ops->get_policy) {
1682 pol = vma->vm_ops->get_policy(vma, addr);
1683 } else if (vma->vm_policy) {
1684 pol = vma->vm_policy;
1687 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1688 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1689 * count on these policies which will be dropped by
1690 * mpol_cond_put() later
1692 if (mpol_needs_cond_ref(pol))
1701 * get_vma_policy(@vma, @addr)
1702 * @vma: virtual memory area whose policy is sought
1703 * @addr: address in @vma for shared policy lookup
1705 * Returns effective policy for a VMA at specified address.
1706 * Falls back to current->mempolicy or system default policy, as necessary.
1707 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1708 * count--added by the get_policy() vm_op, as appropriate--to protect against
1709 * freeing by another task. It is the caller's responsibility to free the
1710 * extra reference for shared policies.
1712 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1715 struct mempolicy *pol = __get_vma_policy(vma, addr);
1718 pol = get_task_policy(current);
1723 bool vma_policy_mof(struct vm_area_struct *vma)
1725 struct mempolicy *pol;
1727 if (vma->vm_ops && vma->vm_ops->get_policy) {
1730 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1731 if (pol && (pol->flags & MPOL_F_MOF))
1738 pol = vma->vm_policy;
1740 pol = get_task_policy(current);
1742 return pol->flags & MPOL_F_MOF;
1745 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1747 enum zone_type dynamic_policy_zone = policy_zone;
1749 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1752 * if policy->nodes has movable memory only,
1753 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1755 * policy->nodes is intersect with node_states[N_MEMORY].
1756 * so if the following test fails, it implies
1757 * policy->nodes has movable memory only.
1759 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1760 dynamic_policy_zone = ZONE_MOVABLE;
1762 return zone >= dynamic_policy_zone;
1766 * Return a nodemask representing a mempolicy for filtering nodes for
1769 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1771 int mode = policy->mode;
1773 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1774 if (unlikely(mode == MPOL_BIND) &&
1775 apply_policy_zone(policy, gfp_zone(gfp)) &&
1776 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1777 return &policy->nodes;
1779 if (mode == MPOL_PREFERRED_MANY)
1780 return &policy->nodes;
1786 * Return the preferred node id for 'prefer' mempolicy, and return
1787 * the given id for all other policies.
1789 * policy_node() is always coupled with policy_nodemask(), which
1790 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1792 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1794 if (policy->mode == MPOL_PREFERRED) {
1795 nd = first_node(policy->nodes);
1798 * __GFP_THISNODE shouldn't even be used with the bind policy
1799 * because we might easily break the expectation to stay on the
1800 * requested node and not break the policy.
1802 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1808 /* Do dynamic interleaving for a process */
1809 static unsigned interleave_nodes(struct mempolicy *policy)
1812 struct task_struct *me = current;
1814 next = next_node_in(me->il_prev, policy->nodes);
1815 if (next < MAX_NUMNODES)
1821 * Depending on the memory policy provide a node from which to allocate the
1824 unsigned int mempolicy_slab_node(void)
1826 struct mempolicy *policy;
1827 int node = numa_mem_id();
1832 policy = current->mempolicy;
1836 switch (policy->mode) {
1837 case MPOL_PREFERRED:
1838 return first_node(policy->nodes);
1840 case MPOL_INTERLEAVE:
1841 return interleave_nodes(policy);
1844 case MPOL_PREFERRED_MANY:
1849 * Follow bind policy behavior and start allocation at the
1852 struct zonelist *zonelist;
1853 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1854 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1855 z = first_zones_zonelist(zonelist, highest_zoneidx,
1857 return z->zone ? zone_to_nid(z->zone) : node;
1868 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1869 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1870 * number of present nodes.
1872 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1874 nodemask_t nodemask = pol->nodes;
1875 unsigned int target, nnodes;
1879 * The barrier will stabilize the nodemask in a register or on
1880 * the stack so that it will stop changing under the code.
1882 * Between first_node() and next_node(), pol->nodes could be changed
1883 * by other threads. So we put pol->nodes in a local stack.
1887 nnodes = nodes_weight(nodemask);
1889 return numa_node_id();
1890 target = (unsigned int)n % nnodes;
1891 nid = first_node(nodemask);
1892 for (i = 0; i < target; i++)
1893 nid = next_node(nid, nodemask);
1897 /* Determine a node number for interleave */
1898 static inline unsigned interleave_nid(struct mempolicy *pol,
1899 struct vm_area_struct *vma, unsigned long addr, int shift)
1905 * for small pages, there is no difference between
1906 * shift and PAGE_SHIFT, so the bit-shift is safe.
1907 * for huge pages, since vm_pgoff is in units of small
1908 * pages, we need to shift off the always 0 bits to get
1911 BUG_ON(shift < PAGE_SHIFT);
1912 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1913 off += (addr - vma->vm_start) >> shift;
1914 return offset_il_node(pol, off);
1916 return interleave_nodes(pol);
1919 #ifdef CONFIG_HUGETLBFS
1921 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1922 * @vma: virtual memory area whose policy is sought
1923 * @addr: address in @vma for shared policy lookup and interleave policy
1924 * @gfp_flags: for requested zone
1925 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1926 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1928 * Returns a nid suitable for a huge page allocation and a pointer
1929 * to the struct mempolicy for conditional unref after allocation.
1930 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
1931 * to the mempolicy's @nodemask for filtering the zonelist.
1933 * Must be protected by read_mems_allowed_begin()
1935 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1936 struct mempolicy **mpol, nodemask_t **nodemask)
1941 *mpol = get_vma_policy(vma, addr);
1943 mode = (*mpol)->mode;
1945 if (unlikely(mode == MPOL_INTERLEAVE)) {
1946 nid = interleave_nid(*mpol, vma, addr,
1947 huge_page_shift(hstate_vma(vma)));
1949 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1950 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
1951 *nodemask = &(*mpol)->nodes;
1957 * init_nodemask_of_mempolicy
1959 * If the current task's mempolicy is "default" [NULL], return 'false'
1960 * to indicate default policy. Otherwise, extract the policy nodemask
1961 * for 'bind' or 'interleave' policy into the argument nodemask, or
1962 * initialize the argument nodemask to contain the single node for
1963 * 'preferred' or 'local' policy and return 'true' to indicate presence
1964 * of non-default mempolicy.
1966 * We don't bother with reference counting the mempolicy [mpol_get/put]
1967 * because the current task is examining it's own mempolicy and a task's
1968 * mempolicy is only ever changed by the task itself.
1970 * N.B., it is the caller's responsibility to free a returned nodemask.
1972 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1974 struct mempolicy *mempolicy;
1976 if (!(mask && current->mempolicy))
1980 mempolicy = current->mempolicy;
1981 switch (mempolicy->mode) {
1982 case MPOL_PREFERRED:
1983 case MPOL_PREFERRED_MANY:
1985 case MPOL_INTERLEAVE:
1986 *mask = mempolicy->nodes;
1990 init_nodemask_of_node(mask, numa_node_id());
1996 task_unlock(current);
2003 * mempolicy_in_oom_domain
2005 * If tsk's mempolicy is "bind", check for intersection between mask and
2006 * the policy nodemask. Otherwise, return true for all other policies
2007 * including "interleave", as a tsk with "interleave" policy may have
2008 * memory allocated from all nodes in system.
2010 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2012 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2013 const nodemask_t *mask)
2015 struct mempolicy *mempolicy;
2022 mempolicy = tsk->mempolicy;
2023 if (mempolicy && mempolicy->mode == MPOL_BIND)
2024 ret = nodes_intersects(mempolicy->nodes, *mask);
2030 /* Allocate a page in interleaved policy.
2031 Own path because it needs to do special accounting. */
2032 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2037 page = __alloc_pages(gfp, order, nid, NULL);
2038 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2039 if (!static_branch_likely(&vm_numa_stat_key))
2041 if (page && page_to_nid(page) == nid) {
2043 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2049 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2050 int nid, struct mempolicy *pol)
2053 gfp_t preferred_gfp;
2056 * This is a two pass approach. The first pass will only try the
2057 * preferred nodes but skip the direct reclaim and allow the
2058 * allocation to fail, while the second pass will try all the
2061 preferred_gfp = gfp | __GFP_NOWARN;
2062 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2063 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2065 page = __alloc_pages(gfp, order, numa_node_id(), NULL);
2071 * alloc_pages_vma - Allocate a page for a VMA.
2073 * @order: Order of the GFP allocation.
2074 * @vma: Pointer to VMA or NULL if not available.
2075 * @addr: Virtual address of the allocation. Must be inside @vma.
2076 * @node: Which node to prefer for allocation (modulo policy).
2077 * @hugepage: For hugepages try only the preferred node if possible.
2079 * Allocate a page for a specific address in @vma, using the appropriate
2080 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2081 * of the mm_struct of the VMA to prevent it from going away. Should be
2082 * used for all allocations for pages that will be mapped into user space.
2084 * Return: The page on success or NULL if allocation fails.
2086 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2087 unsigned long addr, bool hugepage)
2089 struct mempolicy *pol;
2090 int node = numa_node_id();
2095 pol = get_vma_policy(vma, addr);
2097 if (pol->mode == MPOL_INTERLEAVE) {
2100 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2102 page = alloc_page_interleave(gfp, order, nid);
2106 if (pol->mode == MPOL_PREFERRED_MANY) {
2107 page = alloc_pages_preferred_many(gfp, order, node, pol);
2112 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2113 int hpage_node = node;
2116 * For hugepage allocation and non-interleave policy which
2117 * allows the current node (or other explicitly preferred
2118 * node) we only try to allocate from the current/preferred
2119 * node and don't fall back to other nodes, as the cost of
2120 * remote accesses would likely offset THP benefits.
2122 * If the policy is interleave or does not allow the current
2123 * node in its nodemask, we allocate the standard way.
2125 if (pol->mode == MPOL_PREFERRED)
2126 hpage_node = first_node(pol->nodes);
2128 nmask = policy_nodemask(gfp, pol);
2129 if (!nmask || node_isset(hpage_node, *nmask)) {
2132 * First, try to allocate THP only on local node, but
2133 * don't reclaim unnecessarily, just compact.
2135 page = __alloc_pages_node(hpage_node,
2136 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2139 * If hugepage allocations are configured to always
2140 * synchronous compact or the vma has been madvised
2141 * to prefer hugepage backing, retry allowing remote
2142 * memory with both reclaim and compact as well.
2144 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2145 page = __alloc_pages(gfp, order, hpage_node, nmask);
2151 nmask = policy_nodemask(gfp, pol);
2152 preferred_nid = policy_node(gfp, pol, node);
2153 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2158 EXPORT_SYMBOL(alloc_pages_vma);
2161 * alloc_pages - Allocate pages.
2163 * @order: Power of two of number of pages to allocate.
2165 * Allocate 1 << @order contiguous pages. The physical address of the
2166 * first page is naturally aligned (eg an order-3 allocation will be aligned
2167 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2168 * process is honoured when in process context.
2170 * Context: Can be called from any context, providing the appropriate GFP
2172 * Return: The page on success or NULL if allocation fails.
2174 struct page *alloc_pages(gfp_t gfp, unsigned order)
2176 struct mempolicy *pol = &default_policy;
2179 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2180 pol = get_task_policy(current);
2183 * No reference counting needed for current->mempolicy
2184 * nor system default_policy
2186 if (pol->mode == MPOL_INTERLEAVE)
2187 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2188 else if (pol->mode == MPOL_PREFERRED_MANY)
2189 page = alloc_pages_preferred_many(gfp, order,
2190 numa_node_id(), pol);
2192 page = __alloc_pages(gfp, order,
2193 policy_node(gfp, pol, numa_node_id()),
2194 policy_nodemask(gfp, pol));
2198 EXPORT_SYMBOL(alloc_pages);
2200 struct folio *folio_alloc(gfp_t gfp, unsigned order)
2202 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2204 if (page && order > 1)
2205 prep_transhuge_page(page);
2206 return (struct folio *)page;
2208 EXPORT_SYMBOL(folio_alloc);
2210 static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2211 struct mempolicy *pol, unsigned long nr_pages,
2212 struct page **page_array)
2215 unsigned long nr_pages_per_node;
2218 unsigned long nr_allocated;
2219 unsigned long total_allocated = 0;
2221 nodes = nodes_weight(pol->nodes);
2222 nr_pages_per_node = nr_pages / nodes;
2223 delta = nr_pages - nodes * nr_pages_per_node;
2225 for (i = 0; i < nodes; i++) {
2227 nr_allocated = __alloc_pages_bulk(gfp,
2228 interleave_nodes(pol), NULL,
2229 nr_pages_per_node + 1, NULL,
2233 nr_allocated = __alloc_pages_bulk(gfp,
2234 interleave_nodes(pol), NULL,
2235 nr_pages_per_node, NULL, page_array);
2238 page_array += nr_allocated;
2239 total_allocated += nr_allocated;
2242 return total_allocated;
2245 static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2246 struct mempolicy *pol, unsigned long nr_pages,
2247 struct page **page_array)
2249 gfp_t preferred_gfp;
2250 unsigned long nr_allocated = 0;
2252 preferred_gfp = gfp | __GFP_NOWARN;
2253 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2255 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2256 nr_pages, NULL, page_array);
2258 if (nr_allocated < nr_pages)
2259 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2260 nr_pages - nr_allocated, NULL,
2261 page_array + nr_allocated);
2262 return nr_allocated;
2265 /* alloc pages bulk and mempolicy should be considered at the
2266 * same time in some situation such as vmalloc.
2268 * It can accelerate memory allocation especially interleaving
2271 unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2272 unsigned long nr_pages, struct page **page_array)
2274 struct mempolicy *pol = &default_policy;
2276 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2277 pol = get_task_policy(current);
2279 if (pol->mode == MPOL_INTERLEAVE)
2280 return alloc_pages_bulk_array_interleave(gfp, pol,
2281 nr_pages, page_array);
2283 if (pol->mode == MPOL_PREFERRED_MANY)
2284 return alloc_pages_bulk_array_preferred_many(gfp,
2285 numa_node_id(), pol, nr_pages, page_array);
2287 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2288 policy_nodemask(gfp, pol), nr_pages, NULL,
2292 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2294 struct mempolicy *pol = mpol_dup(vma_policy(src));
2297 return PTR_ERR(pol);
2298 dst->vm_policy = pol;
2303 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2304 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2305 * with the mems_allowed returned by cpuset_mems_allowed(). This
2306 * keeps mempolicies cpuset relative after its cpuset moves. See
2307 * further kernel/cpuset.c update_nodemask().
2309 * current's mempolicy may be rebinded by the other task(the task that changes
2310 * cpuset's mems), so we needn't do rebind work for current task.
2313 /* Slow path of a mempolicy duplicate */
2314 struct mempolicy *__mpol_dup(struct mempolicy *old)
2316 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2319 return ERR_PTR(-ENOMEM);
2321 /* task's mempolicy is protected by alloc_lock */
2322 if (old == current->mempolicy) {
2325 task_unlock(current);
2329 if (current_cpuset_is_being_rebound()) {
2330 nodemask_t mems = cpuset_mems_allowed(current);
2331 mpol_rebind_policy(new, &mems);
2333 atomic_set(&new->refcnt, 1);
2337 /* Slow path of a mempolicy comparison */
2338 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2342 if (a->mode != b->mode)
2344 if (a->flags != b->flags)
2346 if (mpol_store_user_nodemask(a))
2347 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2352 case MPOL_INTERLEAVE:
2353 case MPOL_PREFERRED:
2354 case MPOL_PREFERRED_MANY:
2355 return !!nodes_equal(a->nodes, b->nodes);
2365 * Shared memory backing store policy support.
2367 * Remember policies even when nobody has shared memory mapped.
2368 * The policies are kept in Red-Black tree linked from the inode.
2369 * They are protected by the sp->lock rwlock, which should be held
2370 * for any accesses to the tree.
2374 * lookup first element intersecting start-end. Caller holds sp->lock for
2375 * reading or for writing
2377 static struct sp_node *
2378 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2380 struct rb_node *n = sp->root.rb_node;
2383 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2385 if (start >= p->end)
2387 else if (end <= p->start)
2395 struct sp_node *w = NULL;
2396 struct rb_node *prev = rb_prev(n);
2399 w = rb_entry(prev, struct sp_node, nd);
2400 if (w->end <= start)
2404 return rb_entry(n, struct sp_node, nd);
2408 * Insert a new shared policy into the list. Caller holds sp->lock for
2411 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2413 struct rb_node **p = &sp->root.rb_node;
2414 struct rb_node *parent = NULL;
2419 nd = rb_entry(parent, struct sp_node, nd);
2420 if (new->start < nd->start)
2422 else if (new->end > nd->end)
2423 p = &(*p)->rb_right;
2427 rb_link_node(&new->nd, parent, p);
2428 rb_insert_color(&new->nd, &sp->root);
2429 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2430 new->policy ? new->policy->mode : 0);
2433 /* Find shared policy intersecting idx */
2435 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2437 struct mempolicy *pol = NULL;
2440 if (!sp->root.rb_node)
2442 read_lock(&sp->lock);
2443 sn = sp_lookup(sp, idx, idx+1);
2445 mpol_get(sn->policy);
2448 read_unlock(&sp->lock);
2452 static void sp_free(struct sp_node *n)
2454 mpol_put(n->policy);
2455 kmem_cache_free(sn_cache, n);
2459 * mpol_misplaced - check whether current page node is valid in policy
2461 * @page: page to be checked
2462 * @vma: vm area where page mapped
2463 * @addr: virtual address where page mapped
2465 * Lookup current policy node id for vma,addr and "compare to" page's
2466 * node id. Policy determination "mimics" alloc_page_vma().
2467 * Called from fault path where we know the vma and faulting address.
2469 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2470 * policy, or a suitable node ID to allocate a replacement page from.
2472 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2474 struct mempolicy *pol;
2476 int curnid = page_to_nid(page);
2477 unsigned long pgoff;
2478 int thiscpu = raw_smp_processor_id();
2479 int thisnid = cpu_to_node(thiscpu);
2480 int polnid = NUMA_NO_NODE;
2481 int ret = NUMA_NO_NODE;
2483 pol = get_vma_policy(vma, addr);
2484 if (!(pol->flags & MPOL_F_MOF))
2487 switch (pol->mode) {
2488 case MPOL_INTERLEAVE:
2489 pgoff = vma->vm_pgoff;
2490 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2491 polnid = offset_il_node(pol, pgoff);
2494 case MPOL_PREFERRED:
2495 if (node_isset(curnid, pol->nodes))
2497 polnid = first_node(pol->nodes);
2501 polnid = numa_node_id();
2505 /* Optimize placement among multiple nodes via NUMA balancing */
2506 if (pol->flags & MPOL_F_MORON) {
2507 if (node_isset(thisnid, pol->nodes))
2513 case MPOL_PREFERRED_MANY:
2515 * use current page if in policy nodemask,
2516 * else select nearest allowed node, if any.
2517 * If no allowed nodes, use current [!misplaced].
2519 if (node_isset(curnid, pol->nodes))
2521 z = first_zones_zonelist(
2522 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2523 gfp_zone(GFP_HIGHUSER),
2525 polnid = zone_to_nid(z->zone);
2532 /* Migrate the page towards the node whose CPU is referencing it */
2533 if (pol->flags & MPOL_F_MORON) {
2536 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2540 if (curnid != polnid)
2549 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2550 * dropped after task->mempolicy is set to NULL so that any allocation done as
2551 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2554 void mpol_put_task_policy(struct task_struct *task)
2556 struct mempolicy *pol;
2559 pol = task->mempolicy;
2560 task->mempolicy = NULL;
2565 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2567 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2568 rb_erase(&n->nd, &sp->root);
2572 static void sp_node_init(struct sp_node *node, unsigned long start,
2573 unsigned long end, struct mempolicy *pol)
2575 node->start = start;
2580 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2581 struct mempolicy *pol)
2584 struct mempolicy *newpol;
2586 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2590 newpol = mpol_dup(pol);
2591 if (IS_ERR(newpol)) {
2592 kmem_cache_free(sn_cache, n);
2595 newpol->flags |= MPOL_F_SHARED;
2596 sp_node_init(n, start, end, newpol);
2601 /* Replace a policy range. */
2602 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2603 unsigned long end, struct sp_node *new)
2606 struct sp_node *n_new = NULL;
2607 struct mempolicy *mpol_new = NULL;
2611 write_lock(&sp->lock);
2612 n = sp_lookup(sp, start, end);
2613 /* Take care of old policies in the same range. */
2614 while (n && n->start < end) {
2615 struct rb_node *next = rb_next(&n->nd);
2616 if (n->start >= start) {
2622 /* Old policy spanning whole new range. */
2627 *mpol_new = *n->policy;
2628 atomic_set(&mpol_new->refcnt, 1);
2629 sp_node_init(n_new, end, n->end, mpol_new);
2631 sp_insert(sp, n_new);
2640 n = rb_entry(next, struct sp_node, nd);
2644 write_unlock(&sp->lock);
2651 kmem_cache_free(sn_cache, n_new);
2656 write_unlock(&sp->lock);
2658 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2661 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2668 * mpol_shared_policy_init - initialize shared policy for inode
2669 * @sp: pointer to inode shared policy
2670 * @mpol: struct mempolicy to install
2672 * Install non-NULL @mpol in inode's shared policy rb-tree.
2673 * On entry, the current task has a reference on a non-NULL @mpol.
2674 * This must be released on exit.
2675 * This is called at get_inode() calls and we can use GFP_KERNEL.
2677 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2681 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2682 rwlock_init(&sp->lock);
2685 struct vm_area_struct pvma;
2686 struct mempolicy *new;
2687 NODEMASK_SCRATCH(scratch);
2691 /* contextualize the tmpfs mount point mempolicy */
2692 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2694 goto free_scratch; /* no valid nodemask intersection */
2697 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2698 task_unlock(current);
2702 /* Create pseudo-vma that contains just the policy */
2703 vma_init(&pvma, NULL);
2704 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2705 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2708 mpol_put(new); /* drop initial ref */
2710 NODEMASK_SCRATCH_FREE(scratch);
2712 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2716 int mpol_set_shared_policy(struct shared_policy *info,
2717 struct vm_area_struct *vma, struct mempolicy *npol)
2720 struct sp_node *new = NULL;
2721 unsigned long sz = vma_pages(vma);
2723 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2725 sz, npol ? npol->mode : -1,
2726 npol ? npol->flags : -1,
2727 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2730 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2734 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2740 /* Free a backing policy store on inode delete. */
2741 void mpol_free_shared_policy(struct shared_policy *p)
2744 struct rb_node *next;
2746 if (!p->root.rb_node)
2748 write_lock(&p->lock);
2749 next = rb_first(&p->root);
2751 n = rb_entry(next, struct sp_node, nd);
2752 next = rb_next(&n->nd);
2755 write_unlock(&p->lock);
2758 #ifdef CONFIG_NUMA_BALANCING
2759 static int __initdata numabalancing_override;
2761 static void __init check_numabalancing_enable(void)
2763 bool numabalancing_default = false;
2765 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2766 numabalancing_default = true;
2768 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2769 if (numabalancing_override)
2770 set_numabalancing_state(numabalancing_override == 1);
2772 if (num_online_nodes() > 1 && !numabalancing_override) {
2773 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2774 numabalancing_default ? "Enabling" : "Disabling");
2775 set_numabalancing_state(numabalancing_default);
2779 static int __init setup_numabalancing(char *str)
2785 if (!strcmp(str, "enable")) {
2786 numabalancing_override = 1;
2788 } else if (!strcmp(str, "disable")) {
2789 numabalancing_override = -1;
2794 pr_warn("Unable to parse numa_balancing=\n");
2798 __setup("numa_balancing=", setup_numabalancing);
2800 static inline void __init check_numabalancing_enable(void)
2803 #endif /* CONFIG_NUMA_BALANCING */
2805 /* assumes fs == KERNEL_DS */
2806 void __init numa_policy_init(void)
2808 nodemask_t interleave_nodes;
2809 unsigned long largest = 0;
2810 int nid, prefer = 0;
2812 policy_cache = kmem_cache_create("numa_policy",
2813 sizeof(struct mempolicy),
2814 0, SLAB_PANIC, NULL);
2816 sn_cache = kmem_cache_create("shared_policy_node",
2817 sizeof(struct sp_node),
2818 0, SLAB_PANIC, NULL);
2820 for_each_node(nid) {
2821 preferred_node_policy[nid] = (struct mempolicy) {
2822 .refcnt = ATOMIC_INIT(1),
2823 .mode = MPOL_PREFERRED,
2824 .flags = MPOL_F_MOF | MPOL_F_MORON,
2825 .nodes = nodemask_of_node(nid),
2830 * Set interleaving policy for system init. Interleaving is only
2831 * enabled across suitably sized nodes (default is >= 16MB), or
2832 * fall back to the largest node if they're all smaller.
2834 nodes_clear(interleave_nodes);
2835 for_each_node_state(nid, N_MEMORY) {
2836 unsigned long total_pages = node_present_pages(nid);
2838 /* Preserve the largest node */
2839 if (largest < total_pages) {
2840 largest = total_pages;
2844 /* Interleave this node? */
2845 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2846 node_set(nid, interleave_nodes);
2849 /* All too small, use the largest */
2850 if (unlikely(nodes_empty(interleave_nodes)))
2851 node_set(prefer, interleave_nodes);
2853 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2854 pr_err("%s: interleaving failed\n", __func__);
2856 check_numabalancing_enable();
2859 /* Reset policy of current process to default */
2860 void numa_default_policy(void)
2862 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2866 * Parse and format mempolicy from/to strings
2869 static const char * const policy_modes[] =
2871 [MPOL_DEFAULT] = "default",
2872 [MPOL_PREFERRED] = "prefer",
2873 [MPOL_BIND] = "bind",
2874 [MPOL_INTERLEAVE] = "interleave",
2875 [MPOL_LOCAL] = "local",
2876 [MPOL_PREFERRED_MANY] = "prefer (many)",
2882 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2883 * @str: string containing mempolicy to parse
2884 * @mpol: pointer to struct mempolicy pointer, returned on success.
2887 * <mode>[=<flags>][:<nodelist>]
2889 * On success, returns 0, else 1
2891 int mpol_parse_str(char *str, struct mempolicy **mpol)
2893 struct mempolicy *new = NULL;
2894 unsigned short mode_flags;
2896 char *nodelist = strchr(str, ':');
2897 char *flags = strchr(str, '=');
2901 *flags++ = '\0'; /* terminate mode string */
2904 /* NUL-terminate mode or flags string */
2906 if (nodelist_parse(nodelist, nodes))
2908 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2913 mode = match_string(policy_modes, MPOL_MAX, str);
2918 case MPOL_PREFERRED:
2920 * Insist on a nodelist of one node only, although later
2921 * we use first_node(nodes) to grab a single node, so here
2922 * nodelist (or nodes) cannot be empty.
2925 char *rest = nodelist;
2926 while (isdigit(*rest))
2930 if (nodes_empty(nodes))
2934 case MPOL_INTERLEAVE:
2936 * Default to online nodes with memory if no nodelist
2939 nodes = node_states[N_MEMORY];
2943 * Don't allow a nodelist; mpol_new() checks flags
2950 * Insist on a empty nodelist
2955 case MPOL_PREFERRED_MANY:
2958 * Insist on a nodelist
2967 * Currently, we only support two mutually exclusive
2970 if (!strcmp(flags, "static"))
2971 mode_flags |= MPOL_F_STATIC_NODES;
2972 else if (!strcmp(flags, "relative"))
2973 mode_flags |= MPOL_F_RELATIVE_NODES;
2978 new = mpol_new(mode, mode_flags, &nodes);
2983 * Save nodes for mpol_to_str() to show the tmpfs mount options
2984 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2986 if (mode != MPOL_PREFERRED) {
2988 } else if (nodelist) {
2989 nodes_clear(new->nodes);
2990 node_set(first_node(nodes), new->nodes);
2992 new->mode = MPOL_LOCAL;
2996 * Save nodes for contextualization: this will be used to "clone"
2997 * the mempolicy in a specific context [cpuset] at a later time.
2999 new->w.user_nodemask = nodes;
3004 /* Restore string for error message */
3013 #endif /* CONFIG_TMPFS */
3016 * mpol_to_str - format a mempolicy structure for printing
3017 * @buffer: to contain formatted mempolicy string
3018 * @maxlen: length of @buffer
3019 * @pol: pointer to mempolicy to be formatted
3021 * Convert @pol into a string. If @buffer is too short, truncate the string.
3022 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3023 * longest flag, "relative", and to display at least a few node ids.
3025 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3028 nodemask_t nodes = NODE_MASK_NONE;
3029 unsigned short mode = MPOL_DEFAULT;
3030 unsigned short flags = 0;
3032 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3041 case MPOL_PREFERRED:
3042 case MPOL_PREFERRED_MANY:
3044 case MPOL_INTERLEAVE:
3049 snprintf(p, maxlen, "unknown");
3053 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3055 if (flags & MPOL_MODE_FLAGS) {
3056 p += snprintf(p, buffer + maxlen - p, "=");
3059 * Currently, the only defined flags are mutually exclusive
3061 if (flags & MPOL_F_STATIC_NODES)
3062 p += snprintf(p, buffer + maxlen - p, "static");
3063 else if (flags & MPOL_F_RELATIVE_NODES)
3064 p += snprintf(p, buffer + maxlen - p, "relative");
3067 if (!nodes_empty(nodes))
3068 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3069 nodemask_pr_args(&nodes));