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;
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 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(pte - 1, 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((vma->vm_start > start) || (vma->vm_end < 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);
879 ret = mpol_set_nodemask(new, nodes, scratch);
881 task_unlock(current);
885 old = current->mempolicy;
886 current->mempolicy = new;
887 if (new && new->mode == MPOL_INTERLEAVE)
888 current->il_prev = MAX_NUMNODES-1;
889 task_unlock(current);
893 NODEMASK_SCRATCH_FREE(scratch);
898 * Return nodemask for policy for get_mempolicy() query
900 * Called with task's alloc_lock held
902 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
905 if (p == &default_policy)
910 case MPOL_INTERLEAVE:
914 if (!(p->flags & MPOL_F_LOCAL))
915 node_set(p->v.preferred_node, *nodes);
916 /* else return empty node mask for local allocation */
923 static int lookup_node(struct mm_struct *mm, unsigned long addr)
925 struct page *p = NULL;
929 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
931 err = page_to_nid(p);
935 mmap_read_unlock(mm);
939 /* Retrieve NUMA policy */
940 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
941 unsigned long addr, unsigned long flags)
944 struct mm_struct *mm = current->mm;
945 struct vm_area_struct *vma = NULL;
946 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
949 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
952 if (flags & MPOL_F_MEMS_ALLOWED) {
953 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
955 *policy = 0; /* just so it's initialized */
957 *nmask = cpuset_current_mems_allowed;
958 task_unlock(current);
962 if (flags & MPOL_F_ADDR) {
964 * Do NOT fall back to task policy if the
965 * vma/shared policy at addr is NULL. We
966 * want to return MPOL_DEFAULT in this case.
969 vma = find_vma_intersection(mm, addr, addr+1);
971 mmap_read_unlock(mm);
974 if (vma->vm_ops && vma->vm_ops->get_policy)
975 pol = vma->vm_ops->get_policy(vma, addr);
977 pol = vma->vm_policy;
982 pol = &default_policy; /* indicates default behavior */
984 if (flags & MPOL_F_NODE) {
985 if (flags & MPOL_F_ADDR) {
987 * Take a refcount on the mpol, lookup_node()
988 * wil drop the mmap_lock, so after calling
989 * lookup_node() only "pol" remains valid, "vma"
995 err = lookup_node(mm, addr);
999 } else if (pol == current->mempolicy &&
1000 pol->mode == MPOL_INTERLEAVE) {
1001 *policy = next_node_in(current->il_prev, pol->v.nodes);
1007 *policy = pol == &default_policy ? MPOL_DEFAULT :
1010 * Internal mempolicy flags must be masked off before exposing
1011 * the policy to userspace.
1013 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1018 if (mpol_store_user_nodemask(pol)) {
1019 *nmask = pol->w.user_nodemask;
1022 get_policy_nodemask(pol, nmask);
1023 task_unlock(current);
1030 mmap_read_unlock(mm);
1032 mpol_put(pol_refcount);
1036 #ifdef CONFIG_MIGRATION
1038 * page migration, thp tail pages can be passed.
1040 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1041 unsigned long flags)
1043 struct page *head = compound_head(page);
1045 * Avoid migrating a page that is shared with others.
1047 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1048 if (!isolate_lru_page(head)) {
1049 list_add_tail(&head->lru, pagelist);
1050 mod_node_page_state(page_pgdat(head),
1051 NR_ISOLATED_ANON + page_is_file_lru(head),
1052 thp_nr_pages(head));
1053 } else if (flags & MPOL_MF_STRICT) {
1055 * Non-movable page may reach here. And, there may be
1056 * temporary off LRU pages or non-LRU movable pages.
1057 * Treat them as unmovable pages since they can't be
1058 * isolated, so they can't be moved at the moment. It
1059 * should return -EIO for this case too.
1069 * Migrate pages from one node to a target node.
1070 * Returns error or the number of pages not migrated.
1072 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1076 LIST_HEAD(pagelist);
1078 struct migration_target_control mtc = {
1080 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1084 node_set(source, nmask);
1087 * This does not "check" the range but isolates all pages that
1088 * need migration. Between passing in the full user address
1089 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1091 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1092 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1093 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1095 if (!list_empty(&pagelist)) {
1096 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1097 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1099 putback_movable_pages(&pagelist);
1106 * Move pages between the two nodesets so as to preserve the physical
1107 * layout as much as possible.
1109 * Returns the number of page that could not be moved.
1111 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1112 const nodemask_t *to, int flags)
1118 err = migrate_prep();
1125 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1126 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1127 * bit in 'tmp', and return that <source, dest> pair for migration.
1128 * The pair of nodemasks 'to' and 'from' define the map.
1130 * If no pair of bits is found that way, fallback to picking some
1131 * pair of 'source' and 'dest' bits that are not the same. If the
1132 * 'source' and 'dest' bits are the same, this represents a node
1133 * that will be migrating to itself, so no pages need move.
1135 * If no bits are left in 'tmp', or if all remaining bits left
1136 * in 'tmp' correspond to the same bit in 'to', return false
1137 * (nothing left to migrate).
1139 * This lets us pick a pair of nodes to migrate between, such that
1140 * if possible the dest node is not already occupied by some other
1141 * source node, minimizing the risk of overloading the memory on a
1142 * node that would happen if we migrated incoming memory to a node
1143 * before migrating outgoing memory source that same node.
1145 * A single scan of tmp is sufficient. As we go, we remember the
1146 * most recent <s, d> pair that moved (s != d). If we find a pair
1147 * that not only moved, but what's better, moved to an empty slot
1148 * (d is not set in tmp), then we break out then, with that pair.
1149 * Otherwise when we finish scanning from_tmp, we at least have the
1150 * most recent <s, d> pair that moved. If we get all the way through
1151 * the scan of tmp without finding any node that moved, much less
1152 * moved to an empty node, then there is nothing left worth migrating.
1156 while (!nodes_empty(tmp)) {
1158 int source = NUMA_NO_NODE;
1161 for_each_node_mask(s, tmp) {
1164 * do_migrate_pages() tries to maintain the relative
1165 * node relationship of the pages established between
1166 * threads and memory areas.
1168 * However if the number of source nodes is not equal to
1169 * the number of destination nodes we can not preserve
1170 * this node relative relationship. In that case, skip
1171 * copying memory from a node that is in the destination
1174 * Example: [2,3,4] -> [3,4,5] moves everything.
1175 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1178 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1179 (node_isset(s, *to)))
1182 d = node_remap(s, *from, *to);
1186 source = s; /* Node moved. Memorize */
1189 /* dest not in remaining from nodes? */
1190 if (!node_isset(dest, tmp))
1193 if (source == NUMA_NO_NODE)
1196 node_clear(source, tmp);
1197 err = migrate_to_node(mm, source, dest, flags);
1203 mmap_read_unlock(mm);
1211 * Allocate a new page for page migration based on vma policy.
1212 * Start by assuming the page is mapped by the same vma as contains @start.
1213 * Search forward from there, if not. N.B., this assumes that the
1214 * list of pages handed to migrate_pages()--which is how we get here--
1215 * is in virtual address order.
1217 static struct page *new_page(struct page *page, unsigned long start)
1219 struct vm_area_struct *vma;
1220 unsigned long address;
1222 vma = find_vma(current->mm, start);
1224 address = page_address_in_vma(page, vma);
1225 if (address != -EFAULT)
1230 if (PageHuge(page)) {
1231 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1233 } else if (PageTransHuge(page)) {
1236 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1240 prep_transhuge_page(thp);
1244 * if !vma, alloc_page_vma() will use task or system default policy
1246 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1251 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1252 unsigned long flags)
1257 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1258 const nodemask_t *to, int flags)
1263 static struct page *new_page(struct page *page, unsigned long start)
1269 static long do_mbind(unsigned long start, unsigned long len,
1270 unsigned short mode, unsigned short mode_flags,
1271 nodemask_t *nmask, unsigned long flags)
1273 struct mm_struct *mm = current->mm;
1274 struct mempolicy *new;
1278 LIST_HEAD(pagelist);
1280 if (flags & ~(unsigned long)MPOL_MF_VALID)
1282 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1285 if (start & ~PAGE_MASK)
1288 if (mode == MPOL_DEFAULT)
1289 flags &= ~MPOL_MF_STRICT;
1291 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1299 new = mpol_new(mode, mode_flags, nmask);
1301 return PTR_ERR(new);
1303 if (flags & MPOL_MF_LAZY)
1304 new->flags |= MPOL_F_MOF;
1307 * If we are using the default policy then operation
1308 * on discontinuous address spaces is okay after all
1311 flags |= MPOL_MF_DISCONTIG_OK;
1313 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1314 start, start + len, mode, mode_flags,
1315 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1317 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1319 err = migrate_prep();
1324 NODEMASK_SCRATCH(scratch);
1326 mmap_write_lock(mm);
1328 err = mpol_set_nodemask(new, nmask, scratch);
1329 task_unlock(current);
1331 mmap_write_unlock(mm);
1334 NODEMASK_SCRATCH_FREE(scratch);
1339 ret = queue_pages_range(mm, start, end, nmask,
1340 flags | MPOL_MF_INVERT, &pagelist);
1347 err = mbind_range(mm, start, end, new);
1352 if (!list_empty(&pagelist)) {
1353 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1354 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1355 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1357 putback_movable_pages(&pagelist);
1360 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1364 if (!list_empty(&pagelist))
1365 putback_movable_pages(&pagelist);
1368 mmap_write_unlock(mm);
1375 * User space interface with variable sized bitmaps for nodelists.
1378 /* Copy a node mask from user space. */
1379 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1380 unsigned long maxnode)
1384 unsigned long nlongs;
1385 unsigned long endmask;
1388 nodes_clear(*nodes);
1389 if (maxnode == 0 || !nmask)
1391 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1394 nlongs = BITS_TO_LONGS(maxnode);
1395 if ((maxnode % BITS_PER_LONG) == 0)
1398 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1401 * When the user specified more nodes than supported just check
1402 * if the non supported part is all zero.
1404 * If maxnode have more longs than MAX_NUMNODES, check
1405 * the bits in that area first. And then go through to
1406 * check the rest bits which equal or bigger than MAX_NUMNODES.
1407 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1409 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1410 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1411 if (get_user(t, nmask + k))
1413 if (k == nlongs - 1) {
1419 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1423 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1424 unsigned long valid_mask = endmask;
1426 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1427 if (get_user(t, nmask + nlongs - 1))
1433 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1435 nodes_addr(*nodes)[nlongs-1] &= endmask;
1439 /* Copy a kernel node mask to user space */
1440 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1443 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1444 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1446 if (copy > nbytes) {
1447 if (copy > PAGE_SIZE)
1449 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1453 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1456 static long kernel_mbind(unsigned long start, unsigned long len,
1457 unsigned long mode, const unsigned long __user *nmask,
1458 unsigned long maxnode, unsigned int flags)
1462 unsigned short mode_flags;
1464 start = untagged_addr(start);
1465 mode_flags = mode & MPOL_MODE_FLAGS;
1466 mode &= ~MPOL_MODE_FLAGS;
1467 if (mode >= MPOL_MAX)
1469 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1470 (mode_flags & MPOL_F_RELATIVE_NODES))
1472 err = get_nodes(&nodes, nmask, maxnode);
1475 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1478 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1479 unsigned long, mode, const unsigned long __user *, nmask,
1480 unsigned long, maxnode, unsigned int, flags)
1482 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1485 /* Set the process memory policy */
1486 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1487 unsigned long maxnode)
1491 unsigned short flags;
1493 flags = mode & MPOL_MODE_FLAGS;
1494 mode &= ~MPOL_MODE_FLAGS;
1495 if ((unsigned int)mode >= MPOL_MAX)
1497 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1499 err = get_nodes(&nodes, nmask, maxnode);
1502 return do_set_mempolicy(mode, flags, &nodes);
1505 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1506 unsigned long, maxnode)
1508 return kernel_set_mempolicy(mode, nmask, maxnode);
1511 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1512 const unsigned long __user *old_nodes,
1513 const unsigned long __user *new_nodes)
1515 struct mm_struct *mm = NULL;
1516 struct task_struct *task;
1517 nodemask_t task_nodes;
1521 NODEMASK_SCRATCH(scratch);
1526 old = &scratch->mask1;
1527 new = &scratch->mask2;
1529 err = get_nodes(old, old_nodes, maxnode);
1533 err = get_nodes(new, new_nodes, maxnode);
1537 /* Find the mm_struct */
1539 task = pid ? find_task_by_vpid(pid) : current;
1545 get_task_struct(task);
1550 * Check if this process has the right to modify the specified process.
1551 * Use the regular "ptrace_may_access()" checks.
1553 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1560 task_nodes = cpuset_mems_allowed(task);
1561 /* Is the user allowed to access the target nodes? */
1562 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1567 task_nodes = cpuset_mems_allowed(current);
1568 nodes_and(*new, *new, task_nodes);
1569 if (nodes_empty(*new))
1572 err = security_task_movememory(task);
1576 mm = get_task_mm(task);
1577 put_task_struct(task);
1584 err = do_migrate_pages(mm, old, new,
1585 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1589 NODEMASK_SCRATCH_FREE(scratch);
1594 put_task_struct(task);
1599 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1600 const unsigned long __user *, old_nodes,
1601 const unsigned long __user *, new_nodes)
1603 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1607 /* Retrieve NUMA policy */
1608 static int kernel_get_mempolicy(int __user *policy,
1609 unsigned long __user *nmask,
1610 unsigned long maxnode,
1612 unsigned long flags)
1618 if (nmask != NULL && maxnode < nr_node_ids)
1621 addr = untagged_addr(addr);
1623 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1628 if (policy && put_user(pval, policy))
1632 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1637 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1638 unsigned long __user *, nmask, unsigned long, maxnode,
1639 unsigned long, addr, unsigned long, flags)
1641 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1644 #ifdef CONFIG_COMPAT
1646 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1647 compat_ulong_t __user *, nmask,
1648 compat_ulong_t, maxnode,
1649 compat_ulong_t, addr, compat_ulong_t, flags)
1652 unsigned long __user *nm = NULL;
1653 unsigned long nr_bits, alloc_size;
1654 DECLARE_BITMAP(bm, MAX_NUMNODES);
1656 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1657 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1660 nm = compat_alloc_user_space(alloc_size);
1662 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1664 if (!err && nmask) {
1665 unsigned long copy_size;
1666 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1667 err = copy_from_user(bm, nm, copy_size);
1668 /* ensure entire bitmap is zeroed */
1669 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1670 err |= compat_put_bitmap(nmask, bm, nr_bits);
1676 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1677 compat_ulong_t, maxnode)
1679 unsigned long __user *nm = NULL;
1680 unsigned long nr_bits, alloc_size;
1681 DECLARE_BITMAP(bm, MAX_NUMNODES);
1683 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1684 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1687 if (compat_get_bitmap(bm, nmask, nr_bits))
1689 nm = compat_alloc_user_space(alloc_size);
1690 if (copy_to_user(nm, bm, alloc_size))
1694 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1697 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1698 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1699 compat_ulong_t, maxnode, compat_ulong_t, flags)
1701 unsigned long __user *nm = NULL;
1702 unsigned long nr_bits, alloc_size;
1705 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1706 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1709 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1711 nm = compat_alloc_user_space(alloc_size);
1712 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1716 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1719 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1720 compat_ulong_t, maxnode,
1721 const compat_ulong_t __user *, old_nodes,
1722 const compat_ulong_t __user *, new_nodes)
1724 unsigned long __user *old = NULL;
1725 unsigned long __user *new = NULL;
1726 nodemask_t tmp_mask;
1727 unsigned long nr_bits;
1730 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1731 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1733 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1735 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1737 new = old + size / sizeof(unsigned long);
1738 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1742 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1745 new = compat_alloc_user_space(size);
1746 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1749 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1752 #endif /* CONFIG_COMPAT */
1754 bool vma_migratable(struct vm_area_struct *vma)
1756 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1760 * DAX device mappings require predictable access latency, so avoid
1761 * incurring periodic faults.
1763 if (vma_is_dax(vma))
1766 if (is_vm_hugetlb_page(vma) &&
1767 !hugepage_migration_supported(hstate_vma(vma)))
1771 * Migration allocates pages in the highest zone. If we cannot
1772 * do so then migration (at least from node to node) is not
1776 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1782 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1785 struct mempolicy *pol = NULL;
1788 if (vma->vm_ops && vma->vm_ops->get_policy) {
1789 pol = vma->vm_ops->get_policy(vma, addr);
1790 } else if (vma->vm_policy) {
1791 pol = vma->vm_policy;
1794 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1795 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1796 * count on these policies which will be dropped by
1797 * mpol_cond_put() later
1799 if (mpol_needs_cond_ref(pol))
1808 * get_vma_policy(@vma, @addr)
1809 * @vma: virtual memory area whose policy is sought
1810 * @addr: address in @vma for shared policy lookup
1812 * Returns effective policy for a VMA at specified address.
1813 * Falls back to current->mempolicy or system default policy, as necessary.
1814 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1815 * count--added by the get_policy() vm_op, as appropriate--to protect against
1816 * freeing by another task. It is the caller's responsibility to free the
1817 * extra reference for shared policies.
1819 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1822 struct mempolicy *pol = __get_vma_policy(vma, addr);
1825 pol = get_task_policy(current);
1830 bool vma_policy_mof(struct vm_area_struct *vma)
1832 struct mempolicy *pol;
1834 if (vma->vm_ops && vma->vm_ops->get_policy) {
1837 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1838 if (pol && (pol->flags & MPOL_F_MOF))
1845 pol = vma->vm_policy;
1847 pol = get_task_policy(current);
1849 return pol->flags & MPOL_F_MOF;
1852 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1854 enum zone_type dynamic_policy_zone = policy_zone;
1856 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1859 * if policy->v.nodes has movable memory only,
1860 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1862 * policy->v.nodes is intersect with node_states[N_MEMORY].
1863 * so if the following test faile, it implies
1864 * policy->v.nodes has movable memory only.
1866 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1867 dynamic_policy_zone = ZONE_MOVABLE;
1869 return zone >= dynamic_policy_zone;
1873 * Return a nodemask representing a mempolicy for filtering nodes for
1876 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1878 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1879 if (unlikely(policy->mode == MPOL_BIND) &&
1880 apply_policy_zone(policy, gfp_zone(gfp)) &&
1881 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1882 return &policy->v.nodes;
1887 /* Return the node id preferred by the given mempolicy, or the given id */
1888 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1891 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1892 nd = policy->v.preferred_node;
1895 * __GFP_THISNODE shouldn't even be used with the bind policy
1896 * because we might easily break the expectation to stay on the
1897 * requested node and not break the policy.
1899 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1905 /* Do dynamic interleaving for a process */
1906 static unsigned interleave_nodes(struct mempolicy *policy)
1909 struct task_struct *me = current;
1911 next = next_node_in(me->il_prev, policy->v.nodes);
1912 if (next < MAX_NUMNODES)
1918 * Depending on the memory policy provide a node from which to allocate the
1921 unsigned int mempolicy_slab_node(void)
1923 struct mempolicy *policy;
1924 int node = numa_mem_id();
1929 policy = current->mempolicy;
1930 if (!policy || policy->flags & MPOL_F_LOCAL)
1933 switch (policy->mode) {
1934 case MPOL_PREFERRED:
1936 * handled MPOL_F_LOCAL above
1938 return policy->v.preferred_node;
1940 case MPOL_INTERLEAVE:
1941 return interleave_nodes(policy);
1947 * Follow bind policy behavior and start allocation at the
1950 struct zonelist *zonelist;
1951 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1952 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1953 z = first_zones_zonelist(zonelist, highest_zoneidx,
1955 return z->zone ? zone_to_nid(z->zone) : node;
1964 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1965 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1966 * number of present nodes.
1968 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1970 unsigned nnodes = nodes_weight(pol->v.nodes);
1976 return numa_node_id();
1977 target = (unsigned int)n % nnodes;
1978 nid = first_node(pol->v.nodes);
1979 for (i = 0; i < target; i++)
1980 nid = next_node(nid, pol->v.nodes);
1984 /* Determine a node number for interleave */
1985 static inline unsigned interleave_nid(struct mempolicy *pol,
1986 struct vm_area_struct *vma, unsigned long addr, int shift)
1992 * for small pages, there is no difference between
1993 * shift and PAGE_SHIFT, so the bit-shift is safe.
1994 * for huge pages, since vm_pgoff is in units of small
1995 * pages, we need to shift off the always 0 bits to get
1998 BUG_ON(shift < PAGE_SHIFT);
1999 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2000 off += (addr - vma->vm_start) >> shift;
2001 return offset_il_node(pol, off);
2003 return interleave_nodes(pol);
2006 #ifdef CONFIG_HUGETLBFS
2008 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2009 * @vma: virtual memory area whose policy is sought
2010 * @addr: address in @vma for shared policy lookup and interleave policy
2011 * @gfp_flags: for requested zone
2012 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2013 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2015 * Returns a nid suitable for a huge page allocation and a pointer
2016 * to the struct mempolicy for conditional unref after allocation.
2017 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2018 * @nodemask for filtering the zonelist.
2020 * Must be protected by read_mems_allowed_begin()
2022 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2023 struct mempolicy **mpol, nodemask_t **nodemask)
2027 *mpol = get_vma_policy(vma, addr);
2028 *nodemask = NULL; /* assume !MPOL_BIND */
2030 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
2031 nid = interleave_nid(*mpol, vma, addr,
2032 huge_page_shift(hstate_vma(vma)));
2034 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2035 if ((*mpol)->mode == MPOL_BIND)
2036 *nodemask = &(*mpol)->v.nodes;
2042 * init_nodemask_of_mempolicy
2044 * If the current task's mempolicy is "default" [NULL], return 'false'
2045 * to indicate default policy. Otherwise, extract the policy nodemask
2046 * for 'bind' or 'interleave' policy into the argument nodemask, or
2047 * initialize the argument nodemask to contain the single node for
2048 * 'preferred' or 'local' policy and return 'true' to indicate presence
2049 * of non-default mempolicy.
2051 * We don't bother with reference counting the mempolicy [mpol_get/put]
2052 * because the current task is examining it's own mempolicy and a task's
2053 * mempolicy is only ever changed by the task itself.
2055 * N.B., it is the caller's responsibility to free a returned nodemask.
2057 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2059 struct mempolicy *mempolicy;
2062 if (!(mask && current->mempolicy))
2066 mempolicy = current->mempolicy;
2067 switch (mempolicy->mode) {
2068 case MPOL_PREFERRED:
2069 if (mempolicy->flags & MPOL_F_LOCAL)
2070 nid = numa_node_id();
2072 nid = mempolicy->v.preferred_node;
2073 init_nodemask_of_node(mask, nid);
2077 case MPOL_INTERLEAVE:
2078 *mask = mempolicy->v.nodes;
2084 task_unlock(current);
2091 * mempolicy_nodemask_intersects
2093 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2094 * policy. Otherwise, check for intersection between mask and the policy
2095 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2096 * policy, always return true since it may allocate elsewhere on fallback.
2098 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2100 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2101 const nodemask_t *mask)
2103 struct mempolicy *mempolicy;
2109 mempolicy = tsk->mempolicy;
2113 switch (mempolicy->mode) {
2114 case MPOL_PREFERRED:
2116 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2117 * allocate from, they may fallback to other nodes when oom.
2118 * Thus, it's possible for tsk to have allocated memory from
2123 case MPOL_INTERLEAVE:
2124 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2134 /* Allocate a page in interleaved policy.
2135 Own path because it needs to do special accounting. */
2136 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2141 page = __alloc_pages(gfp, order, nid);
2142 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2143 if (!static_branch_likely(&vm_numa_stat_key))
2145 if (page && page_to_nid(page) == nid) {
2147 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2154 * alloc_pages_vma - Allocate a page for a VMA.
2157 * %GFP_USER user allocation.
2158 * %GFP_KERNEL kernel allocations,
2159 * %GFP_HIGHMEM highmem/user allocations,
2160 * %GFP_FS allocation should not call back into a file system.
2161 * %GFP_ATOMIC don't sleep.
2163 * @order:Order of the GFP allocation.
2164 * @vma: Pointer to VMA or NULL if not available.
2165 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2166 * @node: Which node to prefer for allocation (modulo policy).
2167 * @hugepage: for hugepages try only the preferred node if possible
2169 * This function allocates a page from the kernel page pool and applies
2170 * a NUMA policy associated with the VMA or the current process.
2171 * When VMA is not NULL caller must read-lock the mmap_lock of the
2172 * mm_struct of the VMA to prevent it from going away. Should be used for
2173 * all allocations for pages that will be mapped into user space. Returns
2174 * NULL when no page can be allocated.
2177 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2178 unsigned long addr, int node, bool hugepage)
2180 struct mempolicy *pol;
2185 pol = get_vma_policy(vma, addr);
2187 if (pol->mode == MPOL_INTERLEAVE) {
2190 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2192 page = alloc_page_interleave(gfp, order, nid);
2196 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2197 int hpage_node = node;
2200 * For hugepage allocation and non-interleave policy which
2201 * allows the current node (or other explicitly preferred
2202 * node) we only try to allocate from the current/preferred
2203 * node and don't fall back to other nodes, as the cost of
2204 * remote accesses would likely offset THP benefits.
2206 * If the policy is interleave, or does not allow the current
2207 * node in its nodemask, we allocate the standard way.
2209 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2210 hpage_node = pol->v.preferred_node;
2212 nmask = policy_nodemask(gfp, pol);
2213 if (!nmask || node_isset(hpage_node, *nmask)) {
2216 * First, try to allocate THP only on local node, but
2217 * don't reclaim unnecessarily, just compact.
2219 page = __alloc_pages_node(hpage_node,
2220 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2223 * If hugepage allocations are configured to always
2224 * synchronous compact or the vma has been madvised
2225 * to prefer hugepage backing, retry allowing remote
2226 * memory with both reclaim and compact as well.
2228 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2229 page = __alloc_pages_node(hpage_node,
2236 nmask = policy_nodemask(gfp, pol);
2237 preferred_nid = policy_node(gfp, pol, node);
2238 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2243 EXPORT_SYMBOL(alloc_pages_vma);
2246 * alloc_pages_current - Allocate pages.
2249 * %GFP_USER user allocation,
2250 * %GFP_KERNEL kernel allocation,
2251 * %GFP_HIGHMEM highmem allocation,
2252 * %GFP_FS don't call back into a file system.
2253 * %GFP_ATOMIC don't sleep.
2254 * @order: Power of two of allocation size in pages. 0 is a single page.
2256 * Allocate a page from the kernel page pool. When not in
2257 * interrupt context and apply the current process NUMA policy.
2258 * Returns NULL when no page can be allocated.
2260 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2262 struct mempolicy *pol = &default_policy;
2265 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2266 pol = get_task_policy(current);
2269 * No reference counting needed for current->mempolicy
2270 * nor system default_policy
2272 if (pol->mode == MPOL_INTERLEAVE)
2273 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2275 page = __alloc_pages_nodemask(gfp, order,
2276 policy_node(gfp, pol, numa_node_id()),
2277 policy_nodemask(gfp, pol));
2281 EXPORT_SYMBOL(alloc_pages_current);
2283 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2285 struct mempolicy *pol = mpol_dup(vma_policy(src));
2288 return PTR_ERR(pol);
2289 dst->vm_policy = pol;
2294 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2295 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2296 * with the mems_allowed returned by cpuset_mems_allowed(). This
2297 * keeps mempolicies cpuset relative after its cpuset moves. See
2298 * further kernel/cpuset.c update_nodemask().
2300 * current's mempolicy may be rebinded by the other task(the task that changes
2301 * cpuset's mems), so we needn't do rebind work for current task.
2304 /* Slow path of a mempolicy duplicate */
2305 struct mempolicy *__mpol_dup(struct mempolicy *old)
2307 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2310 return ERR_PTR(-ENOMEM);
2312 /* task's mempolicy is protected by alloc_lock */
2313 if (old == current->mempolicy) {
2316 task_unlock(current);
2320 if (current_cpuset_is_being_rebound()) {
2321 nodemask_t mems = cpuset_mems_allowed(current);
2322 mpol_rebind_policy(new, &mems);
2324 atomic_set(&new->refcnt, 1);
2328 /* Slow path of a mempolicy comparison */
2329 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2333 if (a->mode != b->mode)
2335 if (a->flags != b->flags)
2337 if (mpol_store_user_nodemask(a))
2338 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2343 case MPOL_INTERLEAVE:
2344 return !!nodes_equal(a->v.nodes, b->v.nodes);
2345 case MPOL_PREFERRED:
2346 /* a's ->flags is the same as b's */
2347 if (a->flags & MPOL_F_LOCAL)
2349 return a->v.preferred_node == b->v.preferred_node;
2357 * Shared memory backing store policy support.
2359 * Remember policies even when nobody has shared memory mapped.
2360 * The policies are kept in Red-Black tree linked from the inode.
2361 * They are protected by the sp->lock rwlock, which should be held
2362 * for any accesses to the tree.
2366 * lookup first element intersecting start-end. Caller holds sp->lock for
2367 * reading or for writing
2369 static struct sp_node *
2370 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2372 struct rb_node *n = sp->root.rb_node;
2375 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2377 if (start >= p->end)
2379 else if (end <= p->start)
2387 struct sp_node *w = NULL;
2388 struct rb_node *prev = rb_prev(n);
2391 w = rb_entry(prev, struct sp_node, nd);
2392 if (w->end <= start)
2396 return rb_entry(n, struct sp_node, nd);
2400 * Insert a new shared policy into the list. Caller holds sp->lock for
2403 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2405 struct rb_node **p = &sp->root.rb_node;
2406 struct rb_node *parent = NULL;
2411 nd = rb_entry(parent, struct sp_node, nd);
2412 if (new->start < nd->start)
2414 else if (new->end > nd->end)
2415 p = &(*p)->rb_right;
2419 rb_link_node(&new->nd, parent, p);
2420 rb_insert_color(&new->nd, &sp->root);
2421 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2422 new->policy ? new->policy->mode : 0);
2425 /* Find shared policy intersecting idx */
2427 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2429 struct mempolicy *pol = NULL;
2432 if (!sp->root.rb_node)
2434 read_lock(&sp->lock);
2435 sn = sp_lookup(sp, idx, idx+1);
2437 mpol_get(sn->policy);
2440 read_unlock(&sp->lock);
2444 static void sp_free(struct sp_node *n)
2446 mpol_put(n->policy);
2447 kmem_cache_free(sn_cache, n);
2451 * mpol_misplaced - check whether current page node is valid in policy
2453 * @page: page to be checked
2454 * @vma: vm area where page mapped
2455 * @addr: virtual address where page mapped
2457 * Lookup current policy node id for vma,addr and "compare to" page's
2461 * -1 - not misplaced, page is in the right node
2462 * node - node id where the page should be
2464 * Policy determination "mimics" alloc_page_vma().
2465 * Called from fault path where we know the vma and faulting address.
2467 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2469 struct mempolicy *pol;
2471 int curnid = page_to_nid(page);
2472 unsigned long pgoff;
2473 int thiscpu = raw_smp_processor_id();
2474 int thisnid = cpu_to_node(thiscpu);
2475 int polnid = NUMA_NO_NODE;
2478 pol = get_vma_policy(vma, addr);
2479 if (!(pol->flags & MPOL_F_MOF))
2482 switch (pol->mode) {
2483 case MPOL_INTERLEAVE:
2484 pgoff = vma->vm_pgoff;
2485 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2486 polnid = offset_il_node(pol, pgoff);
2489 case MPOL_PREFERRED:
2490 if (pol->flags & MPOL_F_LOCAL)
2491 polnid = numa_node_id();
2493 polnid = pol->v.preferred_node;
2499 * allows binding to multiple nodes.
2500 * use current page if in policy nodemask,
2501 * else select nearest allowed node, if any.
2502 * If no allowed nodes, use current [!misplaced].
2504 if (node_isset(curnid, pol->v.nodes))
2506 z = first_zones_zonelist(
2507 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2508 gfp_zone(GFP_HIGHUSER),
2510 polnid = zone_to_nid(z->zone);
2517 /* Migrate the page towards the node whose CPU is referencing it */
2518 if (pol->flags & MPOL_F_MORON) {
2521 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2525 if (curnid != polnid)
2534 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2535 * dropped after task->mempolicy is set to NULL so that any allocation done as
2536 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2539 void mpol_put_task_policy(struct task_struct *task)
2541 struct mempolicy *pol;
2544 pol = task->mempolicy;
2545 task->mempolicy = NULL;
2550 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2552 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2553 rb_erase(&n->nd, &sp->root);
2557 static void sp_node_init(struct sp_node *node, unsigned long start,
2558 unsigned long end, struct mempolicy *pol)
2560 node->start = start;
2565 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2566 struct mempolicy *pol)
2569 struct mempolicy *newpol;
2571 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2575 newpol = mpol_dup(pol);
2576 if (IS_ERR(newpol)) {
2577 kmem_cache_free(sn_cache, n);
2580 newpol->flags |= MPOL_F_SHARED;
2581 sp_node_init(n, start, end, newpol);
2586 /* Replace a policy range. */
2587 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2588 unsigned long end, struct sp_node *new)
2591 struct sp_node *n_new = NULL;
2592 struct mempolicy *mpol_new = NULL;
2596 write_lock(&sp->lock);
2597 n = sp_lookup(sp, start, end);
2598 /* Take care of old policies in the same range. */
2599 while (n && n->start < end) {
2600 struct rb_node *next = rb_next(&n->nd);
2601 if (n->start >= start) {
2607 /* Old policy spanning whole new range. */
2612 *mpol_new = *n->policy;
2613 atomic_set(&mpol_new->refcnt, 1);
2614 sp_node_init(n_new, end, n->end, mpol_new);
2616 sp_insert(sp, n_new);
2625 n = rb_entry(next, struct sp_node, nd);
2629 write_unlock(&sp->lock);
2636 kmem_cache_free(sn_cache, n_new);
2641 write_unlock(&sp->lock);
2643 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2646 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2653 * mpol_shared_policy_init - initialize shared policy for inode
2654 * @sp: pointer to inode shared policy
2655 * @mpol: struct mempolicy to install
2657 * Install non-NULL @mpol in inode's shared policy rb-tree.
2658 * On entry, the current task has a reference on a non-NULL @mpol.
2659 * This must be released on exit.
2660 * This is called at get_inode() calls and we can use GFP_KERNEL.
2662 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2666 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2667 rwlock_init(&sp->lock);
2670 struct vm_area_struct pvma;
2671 struct mempolicy *new;
2672 NODEMASK_SCRATCH(scratch);
2676 /* contextualize the tmpfs mount point mempolicy */
2677 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2679 goto free_scratch; /* no valid nodemask intersection */
2682 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2683 task_unlock(current);
2687 /* Create pseudo-vma that contains just the policy */
2688 vma_init(&pvma, NULL);
2689 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2690 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2693 mpol_put(new); /* drop initial ref */
2695 NODEMASK_SCRATCH_FREE(scratch);
2697 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2701 int mpol_set_shared_policy(struct shared_policy *info,
2702 struct vm_area_struct *vma, struct mempolicy *npol)
2705 struct sp_node *new = NULL;
2706 unsigned long sz = vma_pages(vma);
2708 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2710 sz, npol ? npol->mode : -1,
2711 npol ? npol->flags : -1,
2712 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2715 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2719 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2725 /* Free a backing policy store on inode delete. */
2726 void mpol_free_shared_policy(struct shared_policy *p)
2729 struct rb_node *next;
2731 if (!p->root.rb_node)
2733 write_lock(&p->lock);
2734 next = rb_first(&p->root);
2736 n = rb_entry(next, struct sp_node, nd);
2737 next = rb_next(&n->nd);
2740 write_unlock(&p->lock);
2743 #ifdef CONFIG_NUMA_BALANCING
2744 static int __initdata numabalancing_override;
2746 static void __init check_numabalancing_enable(void)
2748 bool numabalancing_default = false;
2750 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2751 numabalancing_default = true;
2753 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2754 if (numabalancing_override)
2755 set_numabalancing_state(numabalancing_override == 1);
2757 if (num_online_nodes() > 1 && !numabalancing_override) {
2758 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2759 numabalancing_default ? "Enabling" : "Disabling");
2760 set_numabalancing_state(numabalancing_default);
2764 static int __init setup_numabalancing(char *str)
2770 if (!strcmp(str, "enable")) {
2771 numabalancing_override = 1;
2773 } else if (!strcmp(str, "disable")) {
2774 numabalancing_override = -1;
2779 pr_warn("Unable to parse numa_balancing=\n");
2783 __setup("numa_balancing=", setup_numabalancing);
2785 static inline void __init check_numabalancing_enable(void)
2788 #endif /* CONFIG_NUMA_BALANCING */
2790 /* assumes fs == KERNEL_DS */
2791 void __init numa_policy_init(void)
2793 nodemask_t interleave_nodes;
2794 unsigned long largest = 0;
2795 int nid, prefer = 0;
2797 policy_cache = kmem_cache_create("numa_policy",
2798 sizeof(struct mempolicy),
2799 0, SLAB_PANIC, NULL);
2801 sn_cache = kmem_cache_create("shared_policy_node",
2802 sizeof(struct sp_node),
2803 0, SLAB_PANIC, NULL);
2805 for_each_node(nid) {
2806 preferred_node_policy[nid] = (struct mempolicy) {
2807 .refcnt = ATOMIC_INIT(1),
2808 .mode = MPOL_PREFERRED,
2809 .flags = MPOL_F_MOF | MPOL_F_MORON,
2810 .v = { .preferred_node = nid, },
2815 * Set interleaving policy for system init. Interleaving is only
2816 * enabled across suitably sized nodes (default is >= 16MB), or
2817 * fall back to the largest node if they're all smaller.
2819 nodes_clear(interleave_nodes);
2820 for_each_node_state(nid, N_MEMORY) {
2821 unsigned long total_pages = node_present_pages(nid);
2823 /* Preserve the largest node */
2824 if (largest < total_pages) {
2825 largest = total_pages;
2829 /* Interleave this node? */
2830 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2831 node_set(nid, interleave_nodes);
2834 /* All too small, use the largest */
2835 if (unlikely(nodes_empty(interleave_nodes)))
2836 node_set(prefer, interleave_nodes);
2838 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2839 pr_err("%s: interleaving failed\n", __func__);
2841 check_numabalancing_enable();
2844 /* Reset policy of current process to default */
2845 void numa_default_policy(void)
2847 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2851 * Parse and format mempolicy from/to strings
2855 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2857 static const char * const policy_modes[] =
2859 [MPOL_DEFAULT] = "default",
2860 [MPOL_PREFERRED] = "prefer",
2861 [MPOL_BIND] = "bind",
2862 [MPOL_INTERLEAVE] = "interleave",
2863 [MPOL_LOCAL] = "local",
2869 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2870 * @str: string containing mempolicy to parse
2871 * @mpol: pointer to struct mempolicy pointer, returned on success.
2874 * <mode>[=<flags>][:<nodelist>]
2876 * On success, returns 0, else 1
2878 int mpol_parse_str(char *str, struct mempolicy **mpol)
2880 struct mempolicy *new = NULL;
2881 unsigned short mode_flags;
2883 char *nodelist = strchr(str, ':');
2884 char *flags = strchr(str, '=');
2888 *flags++ = '\0'; /* terminate mode string */
2891 /* NUL-terminate mode or flags string */
2893 if (nodelist_parse(nodelist, nodes))
2895 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2900 mode = match_string(policy_modes, MPOL_MAX, str);
2905 case MPOL_PREFERRED:
2907 * Insist on a nodelist of one node only, although later
2908 * we use first_node(nodes) to grab a single node, so here
2909 * nodelist (or nodes) cannot be empty.
2912 char *rest = nodelist;
2913 while (isdigit(*rest))
2917 if (nodes_empty(nodes))
2921 case MPOL_INTERLEAVE:
2923 * Default to online nodes with memory if no nodelist
2926 nodes = node_states[N_MEMORY];
2930 * Don't allow a nodelist; mpol_new() checks flags
2934 mode = MPOL_PREFERRED;
2938 * Insist on a empty nodelist
2945 * Insist on a nodelist
2954 * Currently, we only support two mutually exclusive
2957 if (!strcmp(flags, "static"))
2958 mode_flags |= MPOL_F_STATIC_NODES;
2959 else if (!strcmp(flags, "relative"))
2960 mode_flags |= MPOL_F_RELATIVE_NODES;
2965 new = mpol_new(mode, mode_flags, &nodes);
2970 * Save nodes for mpol_to_str() to show the tmpfs mount options
2971 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2973 if (mode != MPOL_PREFERRED)
2974 new->v.nodes = nodes;
2976 new->v.preferred_node = first_node(nodes);
2978 new->flags |= MPOL_F_LOCAL;
2981 * Save nodes for contextualization: this will be used to "clone"
2982 * the mempolicy in a specific context [cpuset] at a later time.
2984 new->w.user_nodemask = nodes;
2989 /* Restore string for error message */
2998 #endif /* CONFIG_TMPFS */
3001 * mpol_to_str - format a mempolicy structure for printing
3002 * @buffer: to contain formatted mempolicy string
3003 * @maxlen: length of @buffer
3004 * @pol: pointer to mempolicy to be formatted
3006 * Convert @pol into a string. If @buffer is too short, truncate the string.
3007 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3008 * longest flag, "relative", and to display at least a few node ids.
3010 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3013 nodemask_t nodes = NODE_MASK_NONE;
3014 unsigned short mode = MPOL_DEFAULT;
3015 unsigned short flags = 0;
3017 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3025 case MPOL_PREFERRED:
3026 if (flags & MPOL_F_LOCAL)
3029 node_set(pol->v.preferred_node, nodes);
3032 case MPOL_INTERLEAVE:
3033 nodes = pol->v.nodes;
3037 snprintf(p, maxlen, "unknown");
3041 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3043 if (flags & MPOL_MODE_FLAGS) {
3044 p += snprintf(p, buffer + maxlen - p, "=");
3047 * Currently, the only defined flags are mutually exclusive
3049 if (flags & MPOL_F_STATIC_NODES)
3050 p += snprintf(p, buffer + maxlen - p, "static");
3051 else if (flags & MPOL_F_RELATIVE_NODES)
3052 p += snprintf(p, buffer + maxlen - p, "relative");
3055 if (!nodes_empty(nodes))
3056 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3057 nodemask_pr_args(&nodes));