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 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
71 #include <linux/pagewalk.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/sched/mm.h>
77 #include <linux/sched/numa_balancing.h>
78 #include <linux/sched/task.h>
79 #include <linux/nodemask.h>
80 #include <linux/cpuset.h>
81 #include <linux/slab.h>
82 #include <linux/string.h>
83 #include <linux/export.h>
84 #include <linux/nsproxy.h>
85 #include <linux/interrupt.h>
86 #include <linux/init.h>
87 #include <linux/compat.h>
88 #include <linux/ptrace.h>
89 #include <linux/swap.h>
90 #include <linux/seq_file.h>
91 #include <linux/proc_fs.h>
92 #include <linux/migrate.h>
93 #include <linux/ksm.h>
94 #include <linux/rmap.h>
95 #include <linux/security.h>
96 #include <linux/syscalls.h>
97 #include <linux/ctype.h>
98 #include <linux/mm_inline.h>
99 #include <linux/mmu_notifier.h>
100 #include <linux/printk.h>
101 #include <linux/swapops.h>
103 #include <asm/tlbflush.h>
104 #include <linux/uaccess.h>
106 #include "internal.h"
109 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
110 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
112 static struct kmem_cache *policy_cache;
113 static struct kmem_cache *sn_cache;
115 /* Highest zone. An specific allocation for a zone below that is not
117 enum zone_type policy_zone = 0;
120 * run-time system-wide default policy => local allocation
122 static struct mempolicy default_policy = {
123 .refcnt = ATOMIC_INIT(1), /* never free it */
124 .mode = MPOL_PREFERRED,
125 .flags = MPOL_F_LOCAL,
128 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
131 * numa_map_to_online_node - Find closest online node
132 * @node: Node id to start the search
134 * Lookup the next closest node by distance if @nid is not online.
136 int numa_map_to_online_node(int node)
138 int min_dist = INT_MAX, dist, n, min_node;
140 if (node == NUMA_NO_NODE || node_online(node))
144 for_each_online_node(n) {
145 dist = node_distance(node, n);
146 if (dist < min_dist) {
154 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
156 struct mempolicy *get_task_policy(struct task_struct *p)
158 struct mempolicy *pol = p->mempolicy;
164 node = numa_node_id();
165 if (node != NUMA_NO_NODE) {
166 pol = &preferred_node_policy[node];
167 /* preferred_node_policy is not initialised early in boot */
172 return &default_policy;
175 static const struct mempolicy_operations {
176 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
177 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
178 } mpol_ops[MPOL_MAX];
180 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
182 return pol->flags & MPOL_MODE_FLAGS;
185 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
186 const nodemask_t *rel)
189 nodes_fold(tmp, *orig, nodes_weight(*rel));
190 nodes_onto(*ret, tmp, *rel);
193 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
195 if (nodes_empty(*nodes))
197 pol->v.nodes = *nodes;
201 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
204 pol->flags |= MPOL_F_LOCAL; /* local allocation */
205 else if (nodes_empty(*nodes))
206 return -EINVAL; /* no allowed nodes */
208 pol->v.preferred_node = first_node(*nodes);
212 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
214 if (nodes_empty(*nodes))
216 pol->v.nodes = *nodes;
221 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
222 * any, for the new policy. mpol_new() has already validated the nodes
223 * parameter with respect to the policy mode and flags. But, we need to
224 * handle an empty nodemask with MPOL_PREFERRED here.
226 * Must be called holding task's alloc_lock to protect task's mems_allowed
227 * and mempolicy. May also be called holding the mmap_lock for write.
229 static int mpol_set_nodemask(struct mempolicy *pol,
230 const nodemask_t *nodes, struct nodemask_scratch *nsc)
234 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
238 nodes_and(nsc->mask1,
239 cpuset_current_mems_allowed, node_states[N_MEMORY]);
242 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
243 nodes = NULL; /* explicit local allocation */
245 if (pol->flags & MPOL_F_RELATIVE_NODES)
246 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
248 nodes_and(nsc->mask2, *nodes, nsc->mask1);
250 if (mpol_store_user_nodemask(pol))
251 pol->w.user_nodemask = *nodes;
253 pol->w.cpuset_mems_allowed =
254 cpuset_current_mems_allowed;
258 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
260 ret = mpol_ops[pol->mode].create(pol, NULL);
265 * This function just creates a new policy, does some check and simple
266 * initialization. You must invoke mpol_set_nodemask() to set nodes.
268 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
271 struct mempolicy *policy;
273 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
274 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
276 if (mode == MPOL_DEFAULT) {
277 if (nodes && !nodes_empty(*nodes))
278 return ERR_PTR(-EINVAL);
284 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
285 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
286 * All other modes require a valid pointer to a non-empty nodemask.
288 if (mode == MPOL_PREFERRED) {
289 if (nodes_empty(*nodes)) {
290 if (((flags & MPOL_F_STATIC_NODES) ||
291 (flags & MPOL_F_RELATIVE_NODES)))
292 return ERR_PTR(-EINVAL);
294 } else if (mode == MPOL_LOCAL) {
295 if (!nodes_empty(*nodes) ||
296 (flags & MPOL_F_STATIC_NODES) ||
297 (flags & MPOL_F_RELATIVE_NODES))
298 return ERR_PTR(-EINVAL);
299 mode = MPOL_PREFERRED;
300 } else if (nodes_empty(*nodes))
301 return ERR_PTR(-EINVAL);
302 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
304 return ERR_PTR(-ENOMEM);
305 atomic_set(&policy->refcnt, 1);
307 policy->flags = flags;
312 /* Slow path of a mpol destructor. */
313 void __mpol_put(struct mempolicy *p)
315 if (!atomic_dec_and_test(&p->refcnt))
317 kmem_cache_free(policy_cache, p);
320 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
324 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
328 if (pol->flags & MPOL_F_STATIC_NODES)
329 nodes_and(tmp, pol->w.user_nodemask, *nodes);
330 else if (pol->flags & MPOL_F_RELATIVE_NODES)
331 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
333 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
335 pol->w.cpuset_mems_allowed = *nodes;
338 if (nodes_empty(tmp))
344 static void mpol_rebind_preferred(struct mempolicy *pol,
345 const nodemask_t *nodes)
349 if (pol->flags & MPOL_F_STATIC_NODES) {
350 int node = first_node(pol->w.user_nodemask);
352 if (node_isset(node, *nodes)) {
353 pol->v.preferred_node = node;
354 pol->flags &= ~MPOL_F_LOCAL;
356 pol->flags |= MPOL_F_LOCAL;
357 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
358 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
359 pol->v.preferred_node = first_node(tmp);
360 } else if (!(pol->flags & MPOL_F_LOCAL)) {
361 pol->v.preferred_node = node_remap(pol->v.preferred_node,
362 pol->w.cpuset_mems_allowed,
364 pol->w.cpuset_mems_allowed = *nodes;
369 * mpol_rebind_policy - Migrate a policy to a different set of nodes
371 * Per-vma policies are protected by mmap_lock. Allocations using per-task
372 * policies are protected by task->mems_allowed_seq to prevent a premature
373 * OOM/allocation failure due to parallel nodemask modification.
375 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
379 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
380 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
383 mpol_ops[pol->mode].rebind(pol, newmask);
387 * Wrapper for mpol_rebind_policy() that just requires task
388 * pointer, and updates task mempolicy.
390 * Called with task's alloc_lock held.
393 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
395 mpol_rebind_policy(tsk->mempolicy, new);
399 * Rebind each vma in mm to new nodemask.
401 * Call holding a reference to mm. Takes mm->mmap_lock during call.
404 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
406 struct vm_area_struct *vma;
409 for (vma = mm->mmap; vma; vma = vma->vm_next)
410 mpol_rebind_policy(vma->vm_policy, new);
411 mmap_write_unlock(mm);
414 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
416 .rebind = mpol_rebind_default,
418 [MPOL_INTERLEAVE] = {
419 .create = mpol_new_interleave,
420 .rebind = mpol_rebind_nodemask,
423 .create = mpol_new_preferred,
424 .rebind = mpol_rebind_preferred,
427 .create = mpol_new_bind,
428 .rebind = mpol_rebind_nodemask,
432 static int migrate_page_add(struct page *page, struct list_head *pagelist,
433 unsigned long flags);
436 struct list_head *pagelist;
441 struct vm_area_struct *first;
445 * Check if the page's nid is in qp->nmask.
447 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
448 * in the invert of qp->nmask.
450 static inline bool queue_pages_required(struct page *page,
451 struct queue_pages *qp)
453 int nid = page_to_nid(page);
454 unsigned long flags = qp->flags;
456 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
460 * queue_pages_pmd() has four possible return values:
461 * 0 - pages are placed on the right node or queued successfully.
462 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
465 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
466 * existing page was already on a node that does not follow the
469 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
470 unsigned long end, struct mm_walk *walk)
475 struct queue_pages *qp = walk->private;
478 if (unlikely(is_pmd_migration_entry(*pmd))) {
482 page = pmd_page(*pmd);
483 if (is_huge_zero_page(page)) {
485 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
489 if (!queue_pages_required(page, qp))
493 /* go to thp migration */
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
495 if (!vma_migratable(walk->vma) ||
496 migrate_page_add(page, qp->pagelist, flags)) {
509 * Scan through pages checking if pages follow certain conditions,
510 * and move them to the pagelist if they do.
512 * queue_pages_pte_range() has three possible return values:
513 * 0 - pages are placed on the right node or queued successfully.
514 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
516 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
517 * on a node that does not follow the policy.
519 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
520 unsigned long end, struct mm_walk *walk)
522 struct vm_area_struct *vma = walk->vma;
524 struct queue_pages *qp = walk->private;
525 unsigned long flags = qp->flags;
527 bool has_unmovable = false;
528 pte_t *pte, *mapped_pte;
531 ptl = pmd_trans_huge_lock(pmd, vma);
533 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
537 /* THP was split, fall through to pte walk */
539 if (pmd_trans_unstable(pmd))
542 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
543 for (; addr != end; pte++, addr += PAGE_SIZE) {
544 if (!pte_present(*pte))
546 page = vm_normal_page(vma, addr, *pte);
550 * vm_normal_page() filters out zero pages, but there might
551 * still be PageReserved pages to skip, perhaps in a VDSO.
553 if (PageReserved(page))
555 if (!queue_pages_required(page, qp))
557 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
558 /* MPOL_MF_STRICT must be specified if we get here */
559 if (!vma_migratable(vma)) {
560 has_unmovable = true;
565 * Do not abort immediately since there may be
566 * temporary off LRU pages in the range. Still
567 * need migrate other LRU pages.
569 if (migrate_page_add(page, qp->pagelist, flags))
570 has_unmovable = true;
574 pte_unmap_unlock(mapped_pte, ptl);
580 return addr != end ? -EIO : 0;
583 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
584 unsigned long addr, unsigned long end,
585 struct mm_walk *walk)
588 #ifdef CONFIG_HUGETLB_PAGE
589 struct queue_pages *qp = walk->private;
590 unsigned long flags = (qp->flags & MPOL_MF_VALID);
595 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
596 entry = huge_ptep_get(pte);
597 if (!pte_present(entry))
599 page = pte_page(entry);
600 if (!queue_pages_required(page, qp))
603 if (flags == MPOL_MF_STRICT) {
605 * STRICT alone means only detecting misplaced page and no
606 * need to further check other vma.
612 if (!vma_migratable(walk->vma)) {
614 * Must be STRICT with MOVE*, otherwise .test_walk() have
615 * stopped walking current vma.
616 * Detecting misplaced page but allow migrating pages which
623 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
624 if (flags & (MPOL_MF_MOVE_ALL) ||
625 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
626 if (!isolate_huge_page(page, qp->pagelist) &&
627 (flags & MPOL_MF_STRICT))
629 * Failed to isolate page but allow migrating pages
630 * which have been queued.
642 #ifdef CONFIG_NUMA_BALANCING
644 * This is used to mark a range of virtual addresses to be inaccessible.
645 * These are later cleared by a NUMA hinting fault. Depending on these
646 * faults, pages may be migrated for better NUMA placement.
648 * This is assuming that NUMA faults are handled using PROT_NONE. If
649 * an architecture makes a different choice, it will need further
650 * changes to the core.
652 unsigned long change_prot_numa(struct vm_area_struct *vma,
653 unsigned long addr, unsigned long end)
657 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
659 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
664 static unsigned long change_prot_numa(struct vm_area_struct *vma,
665 unsigned long addr, unsigned long end)
669 #endif /* CONFIG_NUMA_BALANCING */
671 static int queue_pages_test_walk(unsigned long start, unsigned long end,
672 struct mm_walk *walk)
674 struct vm_area_struct *vma = walk->vma;
675 struct queue_pages *qp = walk->private;
676 unsigned long endvma = vma->vm_end;
677 unsigned long flags = qp->flags;
679 /* range check first */
680 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
684 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
685 (qp->start < vma->vm_start))
686 /* hole at head side of range */
689 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
690 ((vma->vm_end < qp->end) &&
691 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
692 /* hole at middle or tail of range */
696 * Need check MPOL_MF_STRICT to return -EIO if possible
697 * regardless of vma_migratable
699 if (!vma_migratable(vma) &&
700 !(flags & MPOL_MF_STRICT))
706 if (flags & MPOL_MF_LAZY) {
707 /* Similar to task_numa_work, skip inaccessible VMAs */
708 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
709 !(vma->vm_flags & VM_MIXEDMAP))
710 change_prot_numa(vma, start, endvma);
714 /* queue pages from current vma */
715 if (flags & MPOL_MF_VALID)
720 static const struct mm_walk_ops queue_pages_walk_ops = {
721 .hugetlb_entry = queue_pages_hugetlb,
722 .pmd_entry = queue_pages_pte_range,
723 .test_walk = queue_pages_test_walk,
727 * Walk through page tables and collect pages to be migrated.
729 * If pages found in a given range are on a set of nodes (determined by
730 * @nodes and @flags,) it's isolated and queued to the pagelist which is
731 * passed via @private.
733 * queue_pages_range() has three possible return values:
734 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
736 * 0 - queue pages successfully or no misplaced page.
737 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
738 * memory range specified by nodemask and maxnode points outside
739 * your accessible address space (-EFAULT)
742 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
743 nodemask_t *nodes, unsigned long flags,
744 struct list_head *pagelist)
747 struct queue_pages qp = {
748 .pagelist = pagelist,
756 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
759 /* whole range in hole */
766 * Apply policy to a single VMA
767 * This must be called with the mmap_lock held for writing.
769 static int vma_replace_policy(struct vm_area_struct *vma,
770 struct mempolicy *pol)
773 struct mempolicy *old;
774 struct mempolicy *new;
776 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
777 vma->vm_start, vma->vm_end, vma->vm_pgoff,
778 vma->vm_ops, vma->vm_file,
779 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
785 if (vma->vm_ops && vma->vm_ops->set_policy) {
786 err = vma->vm_ops->set_policy(vma, new);
791 old = vma->vm_policy;
792 vma->vm_policy = new; /* protected by mmap_lock */
801 /* Step 2: apply policy to a range and do splits. */
802 static int mbind_range(struct mm_struct *mm, unsigned long start,
803 unsigned long end, struct mempolicy *new_pol)
805 struct vm_area_struct *next;
806 struct vm_area_struct *prev;
807 struct vm_area_struct *vma;
810 unsigned long vmstart;
813 vma = find_vma(mm, start);
817 if (start > vma->vm_start)
820 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
822 vmstart = max(start, vma->vm_start);
823 vmend = min(end, vma->vm_end);
825 if (mpol_equal(vma_policy(vma), new_pol))
828 pgoff = vma->vm_pgoff +
829 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
830 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
831 vma->anon_vma, vma->vm_file, pgoff,
832 new_pol, vma->vm_userfaultfd_ctx);
836 if (mpol_equal(vma_policy(vma), new_pol))
838 /* vma_merge() joined vma && vma->next, case 8 */
841 if (vma->vm_start != vmstart) {
842 err = split_vma(vma->vm_mm, vma, vmstart, 1);
846 if (vma->vm_end != vmend) {
847 err = split_vma(vma->vm_mm, vma, vmend, 0);
852 err = vma_replace_policy(vma, new_pol);
861 /* Set the process memory policy */
862 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
865 struct mempolicy *new, *old;
866 NODEMASK_SCRATCH(scratch);
872 new = mpol_new(mode, flags, nodes);
878 if (flags & MPOL_F_NUMA_BALANCING) {
879 if (new && new->mode == MPOL_BIND) {
880 new->flags |= (MPOL_F_MOF | MPOL_F_MORON);
888 ret = mpol_set_nodemask(new, nodes, scratch);
894 old = current->mempolicy;
895 current->mempolicy = new;
896 if (new && new->mode == MPOL_INTERLEAVE)
897 current->il_prev = MAX_NUMNODES-1;
898 task_unlock(current);
902 NODEMASK_SCRATCH_FREE(scratch);
907 * Return nodemask for policy for get_mempolicy() query
909 * Called with task's alloc_lock held
911 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
914 if (p == &default_policy)
919 case MPOL_INTERLEAVE:
923 if (!(p->flags & MPOL_F_LOCAL))
924 node_set(p->v.preferred_node, *nodes);
925 /* else return empty node mask for local allocation */
932 static int lookup_node(struct mm_struct *mm, unsigned long addr)
934 struct page *p = NULL;
938 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
940 err = page_to_nid(p);
944 mmap_read_unlock(mm);
948 /* Retrieve NUMA policy */
949 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
950 unsigned long addr, unsigned long flags)
953 struct mm_struct *mm = current->mm;
954 struct vm_area_struct *vma = NULL;
955 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
958 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
961 if (flags & MPOL_F_MEMS_ALLOWED) {
962 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
964 *policy = 0; /* just so it's initialized */
966 *nmask = cpuset_current_mems_allowed;
967 task_unlock(current);
971 if (flags & MPOL_F_ADDR) {
973 * Do NOT fall back to task policy if the
974 * vma/shared policy at addr is NULL. We
975 * want to return MPOL_DEFAULT in this case.
978 vma = find_vma_intersection(mm, addr, addr+1);
980 mmap_read_unlock(mm);
983 if (vma->vm_ops && vma->vm_ops->get_policy)
984 pol = vma->vm_ops->get_policy(vma, addr);
986 pol = vma->vm_policy;
991 pol = &default_policy; /* indicates default behavior */
993 if (flags & MPOL_F_NODE) {
994 if (flags & MPOL_F_ADDR) {
996 * Take a refcount on the mpol, lookup_node()
997 * wil drop the mmap_lock, so after calling
998 * lookup_node() only "pol" remains valid, "vma"
1004 err = lookup_node(mm, addr);
1008 } else if (pol == current->mempolicy &&
1009 pol->mode == MPOL_INTERLEAVE) {
1010 *policy = next_node_in(current->il_prev, pol->v.nodes);
1016 *policy = pol == &default_policy ? MPOL_DEFAULT :
1019 * Internal mempolicy flags must be masked off before exposing
1020 * the policy to userspace.
1022 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1027 if (mpol_store_user_nodemask(pol)) {
1028 *nmask = pol->w.user_nodemask;
1031 get_policy_nodemask(pol, nmask);
1032 task_unlock(current);
1039 mmap_read_unlock(mm);
1041 mpol_put(pol_refcount);
1045 #ifdef CONFIG_MIGRATION
1047 * page migration, thp tail pages can be passed.
1049 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1050 unsigned long flags)
1052 struct page *head = compound_head(page);
1054 * Avoid migrating a page that is shared with others.
1056 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1057 if (!isolate_lru_page(head)) {
1058 list_add_tail(&head->lru, pagelist);
1059 mod_node_page_state(page_pgdat(head),
1060 NR_ISOLATED_ANON + page_is_file_lru(head),
1061 thp_nr_pages(head));
1062 } else if (flags & MPOL_MF_STRICT) {
1064 * Non-movable page may reach here. And, there may be
1065 * temporary off LRU pages or non-LRU movable pages.
1066 * Treat them as unmovable pages since they can't be
1067 * isolated, so they can't be moved at the moment. It
1068 * should return -EIO for this case too.
1078 * Migrate pages from one node to a target node.
1079 * Returns error or the number of pages not migrated.
1081 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1085 LIST_HEAD(pagelist);
1087 struct migration_target_control mtc = {
1089 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1093 node_set(source, nmask);
1096 * This does not "check" the range but isolates all pages that
1097 * need migration. Between passing in the full user address
1098 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1100 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1101 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1102 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1104 if (!list_empty(&pagelist)) {
1105 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1106 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1108 putback_movable_pages(&pagelist);
1115 * Move pages between the two nodesets so as to preserve the physical
1116 * layout as much as possible.
1118 * Returns the number of page that could not be moved.
1120 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1121 const nodemask_t *to, int flags)
1132 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1133 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1134 * bit in 'tmp', and return that <source, dest> pair for migration.
1135 * The pair of nodemasks 'to' and 'from' define the map.
1137 * If no pair of bits is found that way, fallback to picking some
1138 * pair of 'source' and 'dest' bits that are not the same. If the
1139 * 'source' and 'dest' bits are the same, this represents a node
1140 * that will be migrating to itself, so no pages need move.
1142 * If no bits are left in 'tmp', or if all remaining bits left
1143 * in 'tmp' correspond to the same bit in 'to', return false
1144 * (nothing left to migrate).
1146 * This lets us pick a pair of nodes to migrate between, such that
1147 * if possible the dest node is not already occupied by some other
1148 * source node, minimizing the risk of overloading the memory on a
1149 * node that would happen if we migrated incoming memory to a node
1150 * before migrating outgoing memory source that same node.
1152 * A single scan of tmp is sufficient. As we go, we remember the
1153 * most recent <s, d> pair that moved (s != d). If we find a pair
1154 * that not only moved, but what's better, moved to an empty slot
1155 * (d is not set in tmp), then we break out then, with that pair.
1156 * Otherwise when we finish scanning from_tmp, we at least have the
1157 * most recent <s, d> pair that moved. If we get all the way through
1158 * the scan of tmp without finding any node that moved, much less
1159 * moved to an empty node, then there is nothing left worth migrating.
1163 while (!nodes_empty(tmp)) {
1165 int source = NUMA_NO_NODE;
1168 for_each_node_mask(s, tmp) {
1171 * do_migrate_pages() tries to maintain the relative
1172 * node relationship of the pages established between
1173 * threads and memory areas.
1175 * However if the number of source nodes is not equal to
1176 * the number of destination nodes we can not preserve
1177 * this node relative relationship. In that case, skip
1178 * copying memory from a node that is in the destination
1181 * Example: [2,3,4] -> [3,4,5] moves everything.
1182 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1185 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1186 (node_isset(s, *to)))
1189 d = node_remap(s, *from, *to);
1193 source = s; /* Node moved. Memorize */
1196 /* dest not in remaining from nodes? */
1197 if (!node_isset(dest, tmp))
1200 if (source == NUMA_NO_NODE)
1203 node_clear(source, tmp);
1204 err = migrate_to_node(mm, source, dest, flags);
1210 mmap_read_unlock(mm);
1218 * Allocate a new page for page migration based on vma policy.
1219 * Start by assuming the page is mapped by the same vma as contains @start.
1220 * Search forward from there, if not. N.B., this assumes that the
1221 * list of pages handed to migrate_pages()--which is how we get here--
1222 * is in virtual address order.
1224 static struct page *new_page(struct page *page, unsigned long start)
1226 struct vm_area_struct *vma;
1227 unsigned long address;
1229 vma = find_vma(current->mm, start);
1231 address = page_address_in_vma(page, vma);
1232 if (address != -EFAULT)
1237 if (PageHuge(page)) {
1238 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1240 } else if (PageTransHuge(page)) {
1243 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1247 prep_transhuge_page(thp);
1251 * if !vma, alloc_page_vma() will use task or system default policy
1253 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1258 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1259 unsigned long flags)
1264 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1265 const nodemask_t *to, int flags)
1270 static struct page *new_page(struct page *page, unsigned long start)
1276 static long do_mbind(unsigned long start, unsigned long len,
1277 unsigned short mode, unsigned short mode_flags,
1278 nodemask_t *nmask, unsigned long flags)
1280 struct mm_struct *mm = current->mm;
1281 struct mempolicy *new;
1285 LIST_HEAD(pagelist);
1287 if (flags & ~(unsigned long)MPOL_MF_VALID)
1289 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1292 if (start & ~PAGE_MASK)
1295 if (mode == MPOL_DEFAULT)
1296 flags &= ~MPOL_MF_STRICT;
1298 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1306 new = mpol_new(mode, mode_flags, nmask);
1308 return PTR_ERR(new);
1310 if (flags & MPOL_MF_LAZY)
1311 new->flags |= MPOL_F_MOF;
1314 * If we are using the default policy then operation
1315 * on discontinuous address spaces is okay after all
1318 flags |= MPOL_MF_DISCONTIG_OK;
1320 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1321 start, start + len, mode, mode_flags,
1322 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1324 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1329 NODEMASK_SCRATCH(scratch);
1331 mmap_write_lock(mm);
1332 err = mpol_set_nodemask(new, nmask, scratch);
1334 mmap_write_unlock(mm);
1337 NODEMASK_SCRATCH_FREE(scratch);
1342 ret = queue_pages_range(mm, start, end, nmask,
1343 flags | MPOL_MF_INVERT, &pagelist);
1350 err = mbind_range(mm, start, end, new);
1355 if (!list_empty(&pagelist)) {
1356 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1357 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1358 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1360 putback_movable_pages(&pagelist);
1363 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1367 if (!list_empty(&pagelist))
1368 putback_movable_pages(&pagelist);
1371 mmap_write_unlock(mm);
1378 * User space interface with variable sized bitmaps for nodelists.
1381 /* Copy a node mask from user space. */
1382 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1383 unsigned long maxnode)
1387 unsigned long nlongs;
1388 unsigned long endmask;
1391 nodes_clear(*nodes);
1392 if (maxnode == 0 || !nmask)
1394 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1397 nlongs = BITS_TO_LONGS(maxnode);
1398 if ((maxnode % BITS_PER_LONG) == 0)
1401 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1404 * When the user specified more nodes than supported just check
1405 * if the non supported part is all zero.
1407 * If maxnode have more longs than MAX_NUMNODES, check
1408 * the bits in that area first. And then go through to
1409 * check the rest bits which equal or bigger than MAX_NUMNODES.
1410 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1412 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1413 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1414 if (get_user(t, nmask + k))
1416 if (k == nlongs - 1) {
1422 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1426 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1427 unsigned long valid_mask = endmask;
1429 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1430 if (get_user(t, nmask + nlongs - 1))
1436 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1438 nodes_addr(*nodes)[nlongs-1] &= endmask;
1442 /* Copy a kernel node mask to user space */
1443 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1446 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1447 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1449 if (copy > nbytes) {
1450 if (copy > PAGE_SIZE)
1452 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1456 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1459 static long kernel_mbind(unsigned long start, unsigned long len,
1460 unsigned long mode, const unsigned long __user *nmask,
1461 unsigned long maxnode, unsigned int flags)
1465 unsigned short mode_flags;
1467 start = untagged_addr(start);
1468 mode_flags = mode & MPOL_MODE_FLAGS;
1469 mode &= ~MPOL_MODE_FLAGS;
1470 if (mode >= MPOL_MAX)
1472 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1473 (mode_flags & MPOL_F_RELATIVE_NODES))
1475 err = get_nodes(&nodes, nmask, maxnode);
1478 return do_mbind(start, len, mode, 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)
1494 unsigned short flags;
1496 flags = mode & MPOL_MODE_FLAGS;
1497 mode &= ~MPOL_MODE_FLAGS;
1498 if ((unsigned int)mode >= MPOL_MAX)
1500 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1502 err = get_nodes(&nodes, nmask, maxnode);
1505 return do_set_mempolicy(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 #ifdef CONFIG_COMPAT
1649 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1650 compat_ulong_t __user *, nmask,
1651 compat_ulong_t, maxnode,
1652 compat_ulong_t, addr, compat_ulong_t, flags)
1655 unsigned long __user *nm = NULL;
1656 unsigned long nr_bits, alloc_size;
1657 DECLARE_BITMAP(bm, MAX_NUMNODES);
1659 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1660 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1663 nm = compat_alloc_user_space(alloc_size);
1665 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1667 if (!err && nmask) {
1668 unsigned long copy_size;
1669 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1670 err = copy_from_user(bm, nm, copy_size);
1671 /* ensure entire bitmap is zeroed */
1672 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1673 err |= compat_put_bitmap(nmask, bm, nr_bits);
1679 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1680 compat_ulong_t, maxnode)
1682 unsigned long __user *nm = NULL;
1683 unsigned long nr_bits, alloc_size;
1684 DECLARE_BITMAP(bm, MAX_NUMNODES);
1686 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1687 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1690 if (compat_get_bitmap(bm, nmask, nr_bits))
1692 nm = compat_alloc_user_space(alloc_size);
1693 if (copy_to_user(nm, bm, alloc_size))
1697 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1700 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1701 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1702 compat_ulong_t, maxnode, compat_ulong_t, flags)
1704 unsigned long __user *nm = NULL;
1705 unsigned long nr_bits, alloc_size;
1708 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1709 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1712 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1714 nm = compat_alloc_user_space(alloc_size);
1715 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1719 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1722 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1723 compat_ulong_t, maxnode,
1724 const compat_ulong_t __user *, old_nodes,
1725 const compat_ulong_t __user *, new_nodes)
1727 unsigned long __user *old = NULL;
1728 unsigned long __user *new = NULL;
1729 nodemask_t tmp_mask;
1730 unsigned long nr_bits;
1733 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1734 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1736 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1738 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1740 new = old + size / sizeof(unsigned long);
1741 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1745 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1748 new = compat_alloc_user_space(size);
1749 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1752 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1755 #endif /* CONFIG_COMPAT */
1757 bool vma_migratable(struct vm_area_struct *vma)
1759 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1763 * DAX device mappings require predictable access latency, so avoid
1764 * incurring periodic faults.
1766 if (vma_is_dax(vma))
1769 if (is_vm_hugetlb_page(vma) &&
1770 !hugepage_migration_supported(hstate_vma(vma)))
1774 * Migration allocates pages in the highest zone. If we cannot
1775 * do so then migration (at least from node to node) is not
1779 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1785 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1788 struct mempolicy *pol = NULL;
1791 if (vma->vm_ops && vma->vm_ops->get_policy) {
1792 pol = vma->vm_ops->get_policy(vma, addr);
1793 } else if (vma->vm_policy) {
1794 pol = vma->vm_policy;
1797 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1798 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1799 * count on these policies which will be dropped by
1800 * mpol_cond_put() later
1802 if (mpol_needs_cond_ref(pol))
1811 * get_vma_policy(@vma, @addr)
1812 * @vma: virtual memory area whose policy is sought
1813 * @addr: address in @vma for shared policy lookup
1815 * Returns effective policy for a VMA at specified address.
1816 * Falls back to current->mempolicy or system default policy, as necessary.
1817 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1818 * count--added by the get_policy() vm_op, as appropriate--to protect against
1819 * freeing by another task. It is the caller's responsibility to free the
1820 * extra reference for shared policies.
1822 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1825 struct mempolicy *pol = __get_vma_policy(vma, addr);
1828 pol = get_task_policy(current);
1833 bool vma_policy_mof(struct vm_area_struct *vma)
1835 struct mempolicy *pol;
1837 if (vma->vm_ops && vma->vm_ops->get_policy) {
1840 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1841 if (pol && (pol->flags & MPOL_F_MOF))
1848 pol = vma->vm_policy;
1850 pol = get_task_policy(current);
1852 return pol->flags & MPOL_F_MOF;
1855 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1857 enum zone_type dynamic_policy_zone = policy_zone;
1859 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1862 * if policy->v.nodes has movable memory only,
1863 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1865 * policy->v.nodes is intersect with node_states[N_MEMORY].
1866 * so if the following test faile, it implies
1867 * policy->v.nodes has movable memory only.
1869 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1870 dynamic_policy_zone = ZONE_MOVABLE;
1872 return zone >= dynamic_policy_zone;
1876 * Return a nodemask representing a mempolicy for filtering nodes for
1879 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1881 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1882 if (unlikely(policy->mode == MPOL_BIND) &&
1883 apply_policy_zone(policy, gfp_zone(gfp)) &&
1884 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1885 return &policy->v.nodes;
1890 /* Return the node id preferred by the given mempolicy, or the given id */
1891 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1893 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1894 nd = policy->v.preferred_node;
1897 * __GFP_THISNODE shouldn't even be used with the bind policy
1898 * because we might easily break the expectation to stay on the
1899 * requested node and not break the policy.
1901 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1907 /* Do dynamic interleaving for a process */
1908 static unsigned interleave_nodes(struct mempolicy *policy)
1911 struct task_struct *me = current;
1913 next = next_node_in(me->il_prev, policy->v.nodes);
1914 if (next < MAX_NUMNODES)
1920 * Depending on the memory policy provide a node from which to allocate the
1923 unsigned int mempolicy_slab_node(void)
1925 struct mempolicy *policy;
1926 int node = numa_mem_id();
1931 policy = current->mempolicy;
1932 if (!policy || policy->flags & MPOL_F_LOCAL)
1935 switch (policy->mode) {
1936 case MPOL_PREFERRED:
1938 * handled MPOL_F_LOCAL above
1940 return policy->v.preferred_node;
1942 case MPOL_INTERLEAVE:
1943 return interleave_nodes(policy);
1949 * Follow bind policy behavior and start allocation at the
1952 struct zonelist *zonelist;
1953 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1954 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1955 z = first_zones_zonelist(zonelist, highest_zoneidx,
1957 return z->zone ? zone_to_nid(z->zone) : node;
1966 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1967 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1968 * number of present nodes.
1970 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1972 unsigned nnodes = nodes_weight(pol->v.nodes);
1978 return numa_node_id();
1979 target = (unsigned int)n % nnodes;
1980 nid = first_node(pol->v.nodes);
1981 for (i = 0; i < target; i++)
1982 nid = next_node(nid, pol->v.nodes);
1986 /* Determine a node number for interleave */
1987 static inline unsigned interleave_nid(struct mempolicy *pol,
1988 struct vm_area_struct *vma, unsigned long addr, int shift)
1994 * for small pages, there is no difference between
1995 * shift and PAGE_SHIFT, so the bit-shift is safe.
1996 * for huge pages, since vm_pgoff is in units of small
1997 * pages, we need to shift off the always 0 bits to get
2000 BUG_ON(shift < PAGE_SHIFT);
2001 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2002 off += (addr - vma->vm_start) >> shift;
2003 return offset_il_node(pol, off);
2005 return interleave_nodes(pol);
2008 #ifdef CONFIG_HUGETLBFS
2010 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2011 * @vma: virtual memory area whose policy is sought
2012 * @addr: address in @vma for shared policy lookup and interleave policy
2013 * @gfp_flags: for requested zone
2014 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2015 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2017 * Returns a nid suitable for a huge page allocation and a pointer
2018 * to the struct mempolicy for conditional unref after allocation.
2019 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2020 * @nodemask for filtering the zonelist.
2022 * Must be protected by read_mems_allowed_begin()
2024 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2025 struct mempolicy **mpol, nodemask_t **nodemask)
2029 *mpol = get_vma_policy(vma, addr);
2030 *nodemask = NULL; /* assume !MPOL_BIND */
2032 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
2033 nid = interleave_nid(*mpol, vma, addr,
2034 huge_page_shift(hstate_vma(vma)));
2036 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2037 if ((*mpol)->mode == MPOL_BIND)
2038 *nodemask = &(*mpol)->v.nodes;
2044 * init_nodemask_of_mempolicy
2046 * If the current task's mempolicy is "default" [NULL], return 'false'
2047 * to indicate default policy. Otherwise, extract the policy nodemask
2048 * for 'bind' or 'interleave' policy into the argument nodemask, or
2049 * initialize the argument nodemask to contain the single node for
2050 * 'preferred' or 'local' policy and return 'true' to indicate presence
2051 * of non-default mempolicy.
2053 * We don't bother with reference counting the mempolicy [mpol_get/put]
2054 * because the current task is examining it's own mempolicy and a task's
2055 * mempolicy is only ever changed by the task itself.
2057 * N.B., it is the caller's responsibility to free a returned nodemask.
2059 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2061 struct mempolicy *mempolicy;
2064 if (!(mask && current->mempolicy))
2068 mempolicy = current->mempolicy;
2069 switch (mempolicy->mode) {
2070 case MPOL_PREFERRED:
2071 if (mempolicy->flags & MPOL_F_LOCAL)
2072 nid = numa_node_id();
2074 nid = mempolicy->v.preferred_node;
2075 init_nodemask_of_node(mask, nid);
2079 case MPOL_INTERLEAVE:
2080 *mask = mempolicy->v.nodes;
2086 task_unlock(current);
2093 * mempolicy_nodemask_intersects
2095 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2096 * policy. Otherwise, check for intersection between mask and the policy
2097 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2098 * policy, always return true since it may allocate elsewhere on fallback.
2100 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2102 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2103 const nodemask_t *mask)
2105 struct mempolicy *mempolicy;
2111 mempolicy = tsk->mempolicy;
2115 switch (mempolicy->mode) {
2116 case MPOL_PREFERRED:
2118 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2119 * allocate from, they may fallback to other nodes when oom.
2120 * Thus, it's possible for tsk to have allocated memory from
2125 case MPOL_INTERLEAVE:
2126 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2136 /* Allocate a page in interleaved policy.
2137 Own path because it needs to do special accounting. */
2138 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2143 page = __alloc_pages(gfp, order, nid);
2144 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2145 if (!static_branch_likely(&vm_numa_stat_key))
2147 if (page && page_to_nid(page) == nid) {
2149 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2156 * alloc_pages_vma - Allocate a page for a VMA.
2159 * %GFP_USER user allocation.
2160 * %GFP_KERNEL kernel allocations,
2161 * %GFP_HIGHMEM highmem/user allocations,
2162 * %GFP_FS allocation should not call back into a file system.
2163 * %GFP_ATOMIC don't sleep.
2165 * @order:Order of the GFP allocation.
2166 * @vma: Pointer to VMA or NULL if not available.
2167 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2168 * @node: Which node to prefer for allocation (modulo policy).
2169 * @hugepage: for hugepages try only the preferred node if possible
2171 * This function allocates a page from the kernel page pool and applies
2172 * a NUMA policy associated with the VMA or the current process.
2173 * When VMA is not NULL caller must read-lock the mmap_lock of the
2174 * mm_struct of the VMA to prevent it from going away. Should be used for
2175 * all allocations for pages that will be mapped into user space. Returns
2176 * NULL when no page can be allocated.
2179 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2180 unsigned long addr, int node, bool hugepage)
2182 struct mempolicy *pol;
2187 pol = get_vma_policy(vma, addr);
2189 if (pol->mode == MPOL_INTERLEAVE) {
2192 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2194 page = alloc_page_interleave(gfp, order, nid);
2198 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2199 int hpage_node = node;
2202 * For hugepage allocation and non-interleave policy which
2203 * allows the current node (or other explicitly preferred
2204 * node) we only try to allocate from the current/preferred
2205 * node and don't fall back to other nodes, as the cost of
2206 * remote accesses would likely offset THP benefits.
2208 * If the policy is interleave, or does not allow the current
2209 * node in its nodemask, we allocate the standard way.
2211 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2212 hpage_node = pol->v.preferred_node;
2214 nmask = policy_nodemask(gfp, pol);
2215 if (!nmask || node_isset(hpage_node, *nmask)) {
2218 * First, try to allocate THP only on local node, but
2219 * don't reclaim unnecessarily, just compact.
2221 page = __alloc_pages_node(hpage_node,
2222 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2225 * If hugepage allocations are configured to always
2226 * synchronous compact or the vma has been madvised
2227 * to prefer hugepage backing, retry allowing remote
2228 * memory with both reclaim and compact as well.
2230 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2231 page = __alloc_pages_node(hpage_node,
2238 nmask = policy_nodemask(gfp, pol);
2239 preferred_nid = policy_node(gfp, pol, node);
2240 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2245 EXPORT_SYMBOL(alloc_pages_vma);
2248 * alloc_pages_current - Allocate pages.
2251 * %GFP_USER user allocation,
2252 * %GFP_KERNEL kernel allocation,
2253 * %GFP_HIGHMEM highmem allocation,
2254 * %GFP_FS don't call back into a file system.
2255 * %GFP_ATOMIC don't sleep.
2256 * @order: Power of two of allocation size in pages. 0 is a single page.
2258 * Allocate a page from the kernel page pool. When not in
2259 * interrupt context and apply the current process NUMA policy.
2260 * Returns NULL when no page can be allocated.
2262 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2264 struct mempolicy *pol = &default_policy;
2267 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2268 pol = get_task_policy(current);
2271 * No reference counting needed for current->mempolicy
2272 * nor system default_policy
2274 if (pol->mode == MPOL_INTERLEAVE)
2275 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2277 page = __alloc_pages_nodemask(gfp, order,
2278 policy_node(gfp, pol, numa_node_id()),
2279 policy_nodemask(gfp, pol));
2283 EXPORT_SYMBOL(alloc_pages_current);
2285 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2287 struct mempolicy *pol = mpol_dup(vma_policy(src));
2290 return PTR_ERR(pol);
2291 dst->vm_policy = pol;
2296 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2297 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2298 * with the mems_allowed returned by cpuset_mems_allowed(). This
2299 * keeps mempolicies cpuset relative after its cpuset moves. See
2300 * further kernel/cpuset.c update_nodemask().
2302 * current's mempolicy may be rebinded by the other task(the task that changes
2303 * cpuset's mems), so we needn't do rebind work for current task.
2306 /* Slow path of a mempolicy duplicate */
2307 struct mempolicy *__mpol_dup(struct mempolicy *old)
2309 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2312 return ERR_PTR(-ENOMEM);
2314 /* task's mempolicy is protected by alloc_lock */
2315 if (old == current->mempolicy) {
2318 task_unlock(current);
2322 if (current_cpuset_is_being_rebound()) {
2323 nodemask_t mems = cpuset_mems_allowed(current);
2324 mpol_rebind_policy(new, &mems);
2326 atomic_set(&new->refcnt, 1);
2330 /* Slow path of a mempolicy comparison */
2331 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2335 if (a->mode != b->mode)
2337 if (a->flags != b->flags)
2339 if (mpol_store_user_nodemask(a))
2340 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2345 case MPOL_INTERLEAVE:
2346 return !!nodes_equal(a->v.nodes, b->v.nodes);
2347 case MPOL_PREFERRED:
2348 /* a's ->flags is the same as b's */
2349 if (a->flags & MPOL_F_LOCAL)
2351 return a->v.preferred_node == b->v.preferred_node;
2359 * Shared memory backing store policy support.
2361 * Remember policies even when nobody has shared memory mapped.
2362 * The policies are kept in Red-Black tree linked from the inode.
2363 * They are protected by the sp->lock rwlock, which should be held
2364 * for any accesses to the tree.
2368 * lookup first element intersecting start-end. Caller holds sp->lock for
2369 * reading or for writing
2371 static struct sp_node *
2372 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2374 struct rb_node *n = sp->root.rb_node;
2377 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2379 if (start >= p->end)
2381 else if (end <= p->start)
2389 struct sp_node *w = NULL;
2390 struct rb_node *prev = rb_prev(n);
2393 w = rb_entry(prev, struct sp_node, nd);
2394 if (w->end <= start)
2398 return rb_entry(n, struct sp_node, nd);
2402 * Insert a new shared policy into the list. Caller holds sp->lock for
2405 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2407 struct rb_node **p = &sp->root.rb_node;
2408 struct rb_node *parent = NULL;
2413 nd = rb_entry(parent, struct sp_node, nd);
2414 if (new->start < nd->start)
2416 else if (new->end > nd->end)
2417 p = &(*p)->rb_right;
2421 rb_link_node(&new->nd, parent, p);
2422 rb_insert_color(&new->nd, &sp->root);
2423 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2424 new->policy ? new->policy->mode : 0);
2427 /* Find shared policy intersecting idx */
2429 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2431 struct mempolicy *pol = NULL;
2434 if (!sp->root.rb_node)
2436 read_lock(&sp->lock);
2437 sn = sp_lookup(sp, idx, idx+1);
2439 mpol_get(sn->policy);
2442 read_unlock(&sp->lock);
2446 static void sp_free(struct sp_node *n)
2448 mpol_put(n->policy);
2449 kmem_cache_free(sn_cache, n);
2453 * mpol_misplaced - check whether current page node is valid in policy
2455 * @page: page to be checked
2456 * @vma: vm area where page mapped
2457 * @addr: virtual address where page mapped
2459 * Lookup current policy node id for vma,addr and "compare to" page's
2463 * -1 - not misplaced, page is in the right node
2464 * node - node id where the page should be
2466 * Policy determination "mimics" alloc_page_vma().
2467 * Called from fault path where we know the vma and faulting address.
2469 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2471 struct mempolicy *pol;
2473 int curnid = page_to_nid(page);
2474 unsigned long pgoff;
2475 int thiscpu = raw_smp_processor_id();
2476 int thisnid = cpu_to_node(thiscpu);
2477 int polnid = NUMA_NO_NODE;
2480 pol = get_vma_policy(vma, addr);
2481 if (!(pol->flags & MPOL_F_MOF))
2484 switch (pol->mode) {
2485 case MPOL_INTERLEAVE:
2486 pgoff = vma->vm_pgoff;
2487 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2488 polnid = offset_il_node(pol, pgoff);
2491 case MPOL_PREFERRED:
2492 if (pol->flags & MPOL_F_LOCAL)
2493 polnid = numa_node_id();
2495 polnid = pol->v.preferred_node;
2499 /* Optimize placement among multiple nodes via NUMA balancing */
2500 if (pol->flags & MPOL_F_MORON) {
2501 if (node_isset(thisnid, pol->v.nodes))
2507 * allows binding to multiple nodes.
2508 * use current page if in policy nodemask,
2509 * else select nearest allowed node, if any.
2510 * If no allowed nodes, use current [!misplaced].
2512 if (node_isset(curnid, pol->v.nodes))
2514 z = first_zones_zonelist(
2515 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2516 gfp_zone(GFP_HIGHUSER),
2518 polnid = zone_to_nid(z->zone);
2525 /* Migrate the page towards the node whose CPU is referencing it */
2526 if (pol->flags & MPOL_F_MORON) {
2529 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2533 if (curnid != polnid)
2542 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2543 * dropped after task->mempolicy is set to NULL so that any allocation done as
2544 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2547 void mpol_put_task_policy(struct task_struct *task)
2549 struct mempolicy *pol;
2552 pol = task->mempolicy;
2553 task->mempolicy = NULL;
2558 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2560 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2561 rb_erase(&n->nd, &sp->root);
2565 static void sp_node_init(struct sp_node *node, unsigned long start,
2566 unsigned long end, struct mempolicy *pol)
2568 node->start = start;
2573 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2574 struct mempolicy *pol)
2577 struct mempolicy *newpol;
2579 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2583 newpol = mpol_dup(pol);
2584 if (IS_ERR(newpol)) {
2585 kmem_cache_free(sn_cache, n);
2588 newpol->flags |= MPOL_F_SHARED;
2589 sp_node_init(n, start, end, newpol);
2594 /* Replace a policy range. */
2595 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2596 unsigned long end, struct sp_node *new)
2599 struct sp_node *n_new = NULL;
2600 struct mempolicy *mpol_new = NULL;
2604 write_lock(&sp->lock);
2605 n = sp_lookup(sp, start, end);
2606 /* Take care of old policies in the same range. */
2607 while (n && n->start < end) {
2608 struct rb_node *next = rb_next(&n->nd);
2609 if (n->start >= start) {
2615 /* Old policy spanning whole new range. */
2620 *mpol_new = *n->policy;
2621 atomic_set(&mpol_new->refcnt, 1);
2622 sp_node_init(n_new, end, n->end, mpol_new);
2624 sp_insert(sp, n_new);
2633 n = rb_entry(next, struct sp_node, nd);
2637 write_unlock(&sp->lock);
2644 kmem_cache_free(sn_cache, n_new);
2649 write_unlock(&sp->lock);
2651 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2654 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2661 * mpol_shared_policy_init - initialize shared policy for inode
2662 * @sp: pointer to inode shared policy
2663 * @mpol: struct mempolicy to install
2665 * Install non-NULL @mpol in inode's shared policy rb-tree.
2666 * On entry, the current task has a reference on a non-NULL @mpol.
2667 * This must be released on exit.
2668 * This is called at get_inode() calls and we can use GFP_KERNEL.
2670 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2674 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2675 rwlock_init(&sp->lock);
2678 struct vm_area_struct pvma;
2679 struct mempolicy *new;
2680 NODEMASK_SCRATCH(scratch);
2684 /* contextualize the tmpfs mount point mempolicy */
2685 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2687 goto free_scratch; /* no valid nodemask intersection */
2690 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2691 task_unlock(current);
2695 /* Create pseudo-vma that contains just the policy */
2696 vma_init(&pvma, NULL);
2697 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2698 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2701 mpol_put(new); /* drop initial ref */
2703 NODEMASK_SCRATCH_FREE(scratch);
2705 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2709 int mpol_set_shared_policy(struct shared_policy *info,
2710 struct vm_area_struct *vma, struct mempolicy *npol)
2713 struct sp_node *new = NULL;
2714 unsigned long sz = vma_pages(vma);
2716 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2718 sz, npol ? npol->mode : -1,
2719 npol ? npol->flags : -1,
2720 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2723 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2727 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2733 /* Free a backing policy store on inode delete. */
2734 void mpol_free_shared_policy(struct shared_policy *p)
2737 struct rb_node *next;
2739 if (!p->root.rb_node)
2741 write_lock(&p->lock);
2742 next = rb_first(&p->root);
2744 n = rb_entry(next, struct sp_node, nd);
2745 next = rb_next(&n->nd);
2748 write_unlock(&p->lock);
2751 #ifdef CONFIG_NUMA_BALANCING
2752 static int __initdata numabalancing_override;
2754 static void __init check_numabalancing_enable(void)
2756 bool numabalancing_default = false;
2758 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2759 numabalancing_default = true;
2761 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2762 if (numabalancing_override)
2763 set_numabalancing_state(numabalancing_override == 1);
2765 if (num_online_nodes() > 1 && !numabalancing_override) {
2766 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2767 numabalancing_default ? "Enabling" : "Disabling");
2768 set_numabalancing_state(numabalancing_default);
2772 static int __init setup_numabalancing(char *str)
2778 if (!strcmp(str, "enable")) {
2779 numabalancing_override = 1;
2781 } else if (!strcmp(str, "disable")) {
2782 numabalancing_override = -1;
2787 pr_warn("Unable to parse numa_balancing=\n");
2791 __setup("numa_balancing=", setup_numabalancing);
2793 static inline void __init check_numabalancing_enable(void)
2796 #endif /* CONFIG_NUMA_BALANCING */
2798 /* assumes fs == KERNEL_DS */
2799 void __init numa_policy_init(void)
2801 nodemask_t interleave_nodes;
2802 unsigned long largest = 0;
2803 int nid, prefer = 0;
2805 policy_cache = kmem_cache_create("numa_policy",
2806 sizeof(struct mempolicy),
2807 0, SLAB_PANIC, NULL);
2809 sn_cache = kmem_cache_create("shared_policy_node",
2810 sizeof(struct sp_node),
2811 0, SLAB_PANIC, NULL);
2813 for_each_node(nid) {
2814 preferred_node_policy[nid] = (struct mempolicy) {
2815 .refcnt = ATOMIC_INIT(1),
2816 .mode = MPOL_PREFERRED,
2817 .flags = MPOL_F_MOF | MPOL_F_MORON,
2818 .v = { .preferred_node = nid, },
2823 * Set interleaving policy for system init. Interleaving is only
2824 * enabled across suitably sized nodes (default is >= 16MB), or
2825 * fall back to the largest node if they're all smaller.
2827 nodes_clear(interleave_nodes);
2828 for_each_node_state(nid, N_MEMORY) {
2829 unsigned long total_pages = node_present_pages(nid);
2831 /* Preserve the largest node */
2832 if (largest < total_pages) {
2833 largest = total_pages;
2837 /* Interleave this node? */
2838 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2839 node_set(nid, interleave_nodes);
2842 /* All too small, use the largest */
2843 if (unlikely(nodes_empty(interleave_nodes)))
2844 node_set(prefer, interleave_nodes);
2846 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2847 pr_err("%s: interleaving failed\n", __func__);
2849 check_numabalancing_enable();
2852 /* Reset policy of current process to default */
2853 void numa_default_policy(void)
2855 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2859 * Parse and format mempolicy from/to strings
2863 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2865 static const char * const policy_modes[] =
2867 [MPOL_DEFAULT] = "default",
2868 [MPOL_PREFERRED] = "prefer",
2869 [MPOL_BIND] = "bind",
2870 [MPOL_INTERLEAVE] = "interleave",
2871 [MPOL_LOCAL] = "local",
2877 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2878 * @str: string containing mempolicy to parse
2879 * @mpol: pointer to struct mempolicy pointer, returned on success.
2882 * <mode>[=<flags>][:<nodelist>]
2884 * On success, returns 0, else 1
2886 int mpol_parse_str(char *str, struct mempolicy **mpol)
2888 struct mempolicy *new = NULL;
2889 unsigned short mode_flags;
2891 char *nodelist = strchr(str, ':');
2892 char *flags = strchr(str, '=');
2896 *flags++ = '\0'; /* terminate mode string */
2899 /* NUL-terminate mode or flags string */
2901 if (nodelist_parse(nodelist, nodes))
2903 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2908 mode = match_string(policy_modes, MPOL_MAX, str);
2913 case MPOL_PREFERRED:
2915 * Insist on a nodelist of one node only, although later
2916 * we use first_node(nodes) to grab a single node, so here
2917 * nodelist (or nodes) cannot be empty.
2920 char *rest = nodelist;
2921 while (isdigit(*rest))
2925 if (nodes_empty(nodes))
2929 case MPOL_INTERLEAVE:
2931 * Default to online nodes with memory if no nodelist
2934 nodes = node_states[N_MEMORY];
2938 * Don't allow a nodelist; mpol_new() checks flags
2942 mode = MPOL_PREFERRED;
2946 * Insist on a empty nodelist
2953 * Insist on a nodelist
2962 * Currently, we only support two mutually exclusive
2965 if (!strcmp(flags, "static"))
2966 mode_flags |= MPOL_F_STATIC_NODES;
2967 else if (!strcmp(flags, "relative"))
2968 mode_flags |= MPOL_F_RELATIVE_NODES;
2973 new = mpol_new(mode, mode_flags, &nodes);
2978 * Save nodes for mpol_to_str() to show the tmpfs mount options
2979 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2981 if (mode != MPOL_PREFERRED)
2982 new->v.nodes = nodes;
2984 new->v.preferred_node = first_node(nodes);
2986 new->flags |= MPOL_F_LOCAL;
2989 * Save nodes for contextualization: this will be used to "clone"
2990 * the mempolicy in a specific context [cpuset] at a later time.
2992 new->w.user_nodemask = nodes;
2997 /* Restore string for error message */
3006 #endif /* CONFIG_TMPFS */
3009 * mpol_to_str - format a mempolicy structure for printing
3010 * @buffer: to contain formatted mempolicy string
3011 * @maxlen: length of @buffer
3012 * @pol: pointer to mempolicy to be formatted
3014 * Convert @pol into a string. If @buffer is too short, truncate the string.
3015 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3016 * longest flag, "relative", and to display at least a few node ids.
3018 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3021 nodemask_t nodes = NODE_MASK_NONE;
3022 unsigned short mode = MPOL_DEFAULT;
3023 unsigned short flags = 0;
3025 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3033 case MPOL_PREFERRED:
3034 if (flags & MPOL_F_LOCAL)
3037 node_set(pol->v.preferred_node, nodes);
3040 case MPOL_INTERLEAVE:
3041 nodes = pol->v.nodes;
3045 snprintf(p, maxlen, "unknown");
3049 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3051 if (flags & MPOL_MODE_FLAGS) {
3052 p += snprintf(p, buffer + maxlen - p, "=");
3055 * Currently, the only defined flags are mutually exclusive
3057 if (flags & MPOL_F_STATIC_NODES)
3058 p += snprintf(p, buffer + maxlen - p, "static");
3059 else if (flags & MPOL_F_RELATIVE_NODES)
3060 p += snprintf(p, buffer + maxlen - p, "relative");
3063 if (!nodes_empty(nodes))
3064 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3065 nodemask_pr_args(&nodes));