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];
130 struct mempolicy *get_task_policy(struct task_struct *p)
132 struct mempolicy *pol = p->mempolicy;
138 node = numa_node_id();
139 if (node != NUMA_NO_NODE) {
140 pol = &preferred_node_policy[node];
141 /* preferred_node_policy is not initialised early in boot */
146 return &default_policy;
149 static const struct mempolicy_operations {
150 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
151 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
152 } mpol_ops[MPOL_MAX];
154 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
156 return pol->flags & MPOL_MODE_FLAGS;
159 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
160 const nodemask_t *rel)
163 nodes_fold(tmp, *orig, nodes_weight(*rel));
164 nodes_onto(*ret, tmp, *rel);
167 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
169 if (nodes_empty(*nodes))
171 pol->v.nodes = *nodes;
175 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
178 pol->flags |= MPOL_F_LOCAL; /* local allocation */
179 else if (nodes_empty(*nodes))
180 return -EINVAL; /* no allowed nodes */
182 pol->v.preferred_node = first_node(*nodes);
186 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
188 if (nodes_empty(*nodes))
190 pol->v.nodes = *nodes;
195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
196 * any, for the new policy. mpol_new() has already validated the nodes
197 * parameter with respect to the policy mode and flags. But, we need to
198 * handle an empty nodemask with MPOL_PREFERRED here.
200 * Must be called holding task's alloc_lock to protect task's mems_allowed
201 * and mempolicy. May also be called holding the mmap_semaphore for write.
203 static int mpol_set_nodemask(struct mempolicy *pol,
204 const nodemask_t *nodes, struct nodemask_scratch *nsc)
208 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
212 nodes_and(nsc->mask1,
213 cpuset_current_mems_allowed, node_states[N_MEMORY]);
216 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
217 nodes = NULL; /* explicit local allocation */
219 if (pol->flags & MPOL_F_RELATIVE_NODES)
220 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
222 nodes_and(nsc->mask2, *nodes, nsc->mask1);
224 if (mpol_store_user_nodemask(pol))
225 pol->w.user_nodemask = *nodes;
227 pol->w.cpuset_mems_allowed =
228 cpuset_current_mems_allowed;
232 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
234 ret = mpol_ops[pol->mode].create(pol, NULL);
239 * This function just creates a new policy, does some check and simple
240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
242 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
245 struct mempolicy *policy;
247 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
248 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
250 if (mode == MPOL_DEFAULT) {
251 if (nodes && !nodes_empty(*nodes))
252 return ERR_PTR(-EINVAL);
258 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
259 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
260 * All other modes require a valid pointer to a non-empty nodemask.
262 if (mode == MPOL_PREFERRED) {
263 if (nodes_empty(*nodes)) {
264 if (((flags & MPOL_F_STATIC_NODES) ||
265 (flags & MPOL_F_RELATIVE_NODES)))
266 return ERR_PTR(-EINVAL);
268 } else if (mode == MPOL_LOCAL) {
269 if (!nodes_empty(*nodes) ||
270 (flags & MPOL_F_STATIC_NODES) ||
271 (flags & MPOL_F_RELATIVE_NODES))
272 return ERR_PTR(-EINVAL);
273 mode = MPOL_PREFERRED;
274 } else if (nodes_empty(*nodes))
275 return ERR_PTR(-EINVAL);
276 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
278 return ERR_PTR(-ENOMEM);
279 atomic_set(&policy->refcnt, 1);
281 policy->flags = flags;
286 /* Slow path of a mpol destructor. */
287 void __mpol_put(struct mempolicy *p)
289 if (!atomic_dec_and_test(&p->refcnt))
291 kmem_cache_free(policy_cache, p);
294 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
298 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
302 if (pol->flags & MPOL_F_STATIC_NODES)
303 nodes_and(tmp, pol->w.user_nodemask, *nodes);
304 else if (pol->flags & MPOL_F_RELATIVE_NODES)
305 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
307 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
309 pol->w.cpuset_mems_allowed = *nodes;
312 if (nodes_empty(tmp))
318 static void mpol_rebind_preferred(struct mempolicy *pol,
319 const nodemask_t *nodes)
323 if (pol->flags & MPOL_F_STATIC_NODES) {
324 int node = first_node(pol->w.user_nodemask);
326 if (node_isset(node, *nodes)) {
327 pol->v.preferred_node = node;
328 pol->flags &= ~MPOL_F_LOCAL;
330 pol->flags |= MPOL_F_LOCAL;
331 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
332 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
333 pol->v.preferred_node = first_node(tmp);
334 } else if (!(pol->flags & MPOL_F_LOCAL)) {
335 pol->v.preferred_node = node_remap(pol->v.preferred_node,
336 pol->w.cpuset_mems_allowed,
338 pol->w.cpuset_mems_allowed = *nodes;
343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
346 * policies are protected by task->mems_allowed_seq to prevent a premature
347 * OOM/allocation failure due to parallel nodemask modification.
349 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
353 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
354 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
357 mpol_ops[pol->mode].rebind(pol, newmask);
361 * Wrapper for mpol_rebind_policy() that just requires task
362 * pointer, and updates task mempolicy.
364 * Called with task's alloc_lock held.
367 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
369 mpol_rebind_policy(tsk->mempolicy, new);
373 * Rebind each vma in mm to new nodemask.
375 * Call holding a reference to mm. Takes mm->mmap_sem during call.
378 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
380 struct vm_area_struct *vma;
382 down_write(&mm->mmap_sem);
383 for (vma = mm->mmap; vma; vma = vma->vm_next)
384 mpol_rebind_policy(vma->vm_policy, new);
385 up_write(&mm->mmap_sem);
388 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
390 .rebind = mpol_rebind_default,
392 [MPOL_INTERLEAVE] = {
393 .create = mpol_new_interleave,
394 .rebind = mpol_rebind_nodemask,
397 .create = mpol_new_preferred,
398 .rebind = mpol_rebind_preferred,
401 .create = mpol_new_bind,
402 .rebind = mpol_rebind_nodemask,
406 static int migrate_page_add(struct page *page, struct list_head *pagelist,
407 unsigned long flags);
410 struct list_head *pagelist;
415 struct vm_area_struct *first;
419 * Check if the page's nid is in qp->nmask.
421 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
422 * in the invert of qp->nmask.
424 static inline bool queue_pages_required(struct page *page,
425 struct queue_pages *qp)
427 int nid = page_to_nid(page);
428 unsigned long flags = qp->flags;
430 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
434 * queue_pages_pmd() has four possible return values:
435 * 0 - pages are placed on the right node or queued successfully.
436 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
439 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
440 * existing page was already on a node that does not follow the
443 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
444 unsigned long end, struct mm_walk *walk)
449 struct queue_pages *qp = walk->private;
452 if (unlikely(is_pmd_migration_entry(*pmd))) {
456 page = pmd_page(*pmd);
457 if (is_huge_zero_page(page)) {
459 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
463 if (!queue_pages_required(page, qp))
467 /* go to thp migration */
468 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
469 if (!vma_migratable(walk->vma) ||
470 migrate_page_add(page, qp->pagelist, flags)) {
483 * Scan through pages checking if pages follow certain conditions,
484 * and move them to the pagelist if they do.
486 * queue_pages_pte_range() has three possible return values:
487 * 0 - pages are placed on the right node or queued successfully.
488 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
490 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
491 * on a node that does not follow the policy.
493 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
494 unsigned long end, struct mm_walk *walk)
496 struct vm_area_struct *vma = walk->vma;
498 struct queue_pages *qp = walk->private;
499 unsigned long flags = qp->flags;
501 bool has_unmovable = false;
505 ptl = pmd_trans_huge_lock(pmd, vma);
507 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
511 /* THP was split, fall through to pte walk */
513 if (pmd_trans_unstable(pmd))
516 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
517 for (; addr != end; pte++, addr += PAGE_SIZE) {
518 if (!pte_present(*pte))
520 page = vm_normal_page(vma, addr, *pte);
524 * vm_normal_page() filters out zero pages, but there might
525 * still be PageReserved pages to skip, perhaps in a VDSO.
527 if (PageReserved(page))
529 if (!queue_pages_required(page, qp))
531 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
532 /* MPOL_MF_STRICT must be specified if we get here */
533 if (!vma_migratable(vma)) {
534 has_unmovable = true;
539 * Do not abort immediately since there may be
540 * temporary off LRU pages in the range. Still
541 * need migrate other LRU pages.
543 if (migrate_page_add(page, qp->pagelist, flags))
544 has_unmovable = true;
548 pte_unmap_unlock(pte - 1, ptl);
554 return addr != end ? -EIO : 0;
557 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
558 unsigned long addr, unsigned long end,
559 struct mm_walk *walk)
562 #ifdef CONFIG_HUGETLB_PAGE
563 struct queue_pages *qp = walk->private;
564 unsigned long flags = (qp->flags & MPOL_MF_VALID);
569 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
570 entry = huge_ptep_get(pte);
571 if (!pte_present(entry))
573 page = pte_page(entry);
574 if (!queue_pages_required(page, qp))
577 if (flags == MPOL_MF_STRICT) {
579 * STRICT alone means only detecting misplaced page and no
580 * need to further check other vma.
586 if (!vma_migratable(walk->vma)) {
588 * Must be STRICT with MOVE*, otherwise .test_walk() have
589 * stopped walking current vma.
590 * Detecting misplaced page but allow migrating pages which
597 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
598 if (flags & (MPOL_MF_MOVE_ALL) ||
599 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
600 if (!isolate_huge_page(page, qp->pagelist) &&
601 (flags & MPOL_MF_STRICT))
603 * Failed to isolate page but allow migrating pages
604 * which have been queued.
616 #ifdef CONFIG_NUMA_BALANCING
618 * This is used to mark a range of virtual addresses to be inaccessible.
619 * These are later cleared by a NUMA hinting fault. Depending on these
620 * faults, pages may be migrated for better NUMA placement.
622 * This is assuming that NUMA faults are handled using PROT_NONE. If
623 * an architecture makes a different choice, it will need further
624 * changes to the core.
626 unsigned long change_prot_numa(struct vm_area_struct *vma,
627 unsigned long addr, unsigned long end)
631 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
633 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
638 static unsigned long change_prot_numa(struct vm_area_struct *vma,
639 unsigned long addr, unsigned long end)
643 #endif /* CONFIG_NUMA_BALANCING */
645 static int queue_pages_test_walk(unsigned long start, unsigned long end,
646 struct mm_walk *walk)
648 struct vm_area_struct *vma = walk->vma;
649 struct queue_pages *qp = walk->private;
650 unsigned long endvma = vma->vm_end;
651 unsigned long flags = qp->flags;
653 /* range check first */
654 VM_BUG_ON_VMA((vma->vm_start > start) || (vma->vm_end < end), vma);
658 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
659 (qp->start < vma->vm_start))
660 /* hole at head side of range */
663 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
664 ((vma->vm_end < qp->end) &&
665 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
666 /* hole at middle or tail of range */
670 * Need check MPOL_MF_STRICT to return -EIO if possible
671 * regardless of vma_migratable
673 if (!vma_migratable(vma) &&
674 !(flags & MPOL_MF_STRICT))
680 if (flags & MPOL_MF_LAZY) {
681 /* Similar to task_numa_work, skip inaccessible VMAs */
682 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
683 !(vma->vm_flags & VM_MIXEDMAP))
684 change_prot_numa(vma, start, endvma);
688 /* queue pages from current vma */
689 if (flags & MPOL_MF_VALID)
694 static const struct mm_walk_ops queue_pages_walk_ops = {
695 .hugetlb_entry = queue_pages_hugetlb,
696 .pmd_entry = queue_pages_pte_range,
697 .test_walk = queue_pages_test_walk,
701 * Walk through page tables and collect pages to be migrated.
703 * If pages found in a given range are on a set of nodes (determined by
704 * @nodes and @flags,) it's isolated and queued to the pagelist which is
705 * passed via @private.
707 * queue_pages_range() has three possible return values:
708 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
710 * 0 - queue pages successfully or no misplaced page.
711 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
712 * memory range specified by nodemask and maxnode points outside
713 * your accessible address space (-EFAULT)
716 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
717 nodemask_t *nodes, unsigned long flags,
718 struct list_head *pagelist)
721 struct queue_pages qp = {
722 .pagelist = pagelist,
730 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
733 /* whole range in hole */
740 * Apply policy to a single VMA
741 * This must be called with the mmap_sem held for writing.
743 static int vma_replace_policy(struct vm_area_struct *vma,
744 struct mempolicy *pol)
747 struct mempolicy *old;
748 struct mempolicy *new;
750 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
751 vma->vm_start, vma->vm_end, vma->vm_pgoff,
752 vma->vm_ops, vma->vm_file,
753 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
759 if (vma->vm_ops && vma->vm_ops->set_policy) {
760 err = vma->vm_ops->set_policy(vma, new);
765 old = vma->vm_policy;
766 vma->vm_policy = new; /* protected by mmap_sem */
775 /* Step 2: apply policy to a range and do splits. */
776 static int mbind_range(struct mm_struct *mm, unsigned long start,
777 unsigned long end, struct mempolicy *new_pol)
779 struct vm_area_struct *next;
780 struct vm_area_struct *prev;
781 struct vm_area_struct *vma;
784 unsigned long vmstart;
787 vma = find_vma(mm, start);
791 if (start > vma->vm_start)
794 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
796 vmstart = max(start, vma->vm_start);
797 vmend = min(end, vma->vm_end);
799 if (mpol_equal(vma_policy(vma), new_pol))
802 pgoff = vma->vm_pgoff +
803 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
804 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
805 vma->anon_vma, vma->vm_file, pgoff,
806 new_pol, vma->vm_userfaultfd_ctx);
810 if (mpol_equal(vma_policy(vma), new_pol))
812 /* vma_merge() joined vma && vma->next, case 8 */
815 if (vma->vm_start != vmstart) {
816 err = split_vma(vma->vm_mm, vma, vmstart, 1);
820 if (vma->vm_end != vmend) {
821 err = split_vma(vma->vm_mm, vma, vmend, 0);
826 err = vma_replace_policy(vma, new_pol);
835 /* Set the process memory policy */
836 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
839 struct mempolicy *new, *old;
840 NODEMASK_SCRATCH(scratch);
846 new = mpol_new(mode, flags, nodes);
853 ret = mpol_set_nodemask(new, nodes, scratch);
855 task_unlock(current);
859 old = current->mempolicy;
860 current->mempolicy = new;
861 if (new && new->mode == MPOL_INTERLEAVE)
862 current->il_prev = MAX_NUMNODES-1;
863 task_unlock(current);
867 NODEMASK_SCRATCH_FREE(scratch);
872 * Return nodemask for policy for get_mempolicy() query
874 * Called with task's alloc_lock held
876 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
879 if (p == &default_policy)
884 case MPOL_INTERLEAVE:
888 if (!(p->flags & MPOL_F_LOCAL))
889 node_set(p->v.preferred_node, *nodes);
890 /* else return empty node mask for local allocation */
897 static int lookup_node(struct mm_struct *mm, unsigned long addr)
899 struct page *p = NULL;
903 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
905 /* E.g. GUP interrupted by fatal signal */
907 } else if (err > 0) {
908 err = page_to_nid(p);
912 up_read(&mm->mmap_sem);
916 /* Retrieve NUMA policy */
917 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
918 unsigned long addr, unsigned long flags)
921 struct mm_struct *mm = current->mm;
922 struct vm_area_struct *vma = NULL;
923 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
926 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
929 if (flags & MPOL_F_MEMS_ALLOWED) {
930 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
932 *policy = 0; /* just so it's initialized */
934 *nmask = cpuset_current_mems_allowed;
935 task_unlock(current);
939 if (flags & MPOL_F_ADDR) {
941 * Do NOT fall back to task policy if the
942 * vma/shared policy at addr is NULL. We
943 * want to return MPOL_DEFAULT in this case.
945 down_read(&mm->mmap_sem);
946 vma = find_vma_intersection(mm, addr, addr+1);
948 up_read(&mm->mmap_sem);
951 if (vma->vm_ops && vma->vm_ops->get_policy)
952 pol = vma->vm_ops->get_policy(vma, addr);
954 pol = vma->vm_policy;
959 pol = &default_policy; /* indicates default behavior */
961 if (flags & MPOL_F_NODE) {
962 if (flags & MPOL_F_ADDR) {
964 * Take a refcount on the mpol, lookup_node()
965 * wil drop the mmap_sem, so after calling
966 * lookup_node() only "pol" remains valid, "vma"
972 err = lookup_node(mm, addr);
976 } else if (pol == current->mempolicy &&
977 pol->mode == MPOL_INTERLEAVE) {
978 *policy = next_node_in(current->il_prev, pol->v.nodes);
984 *policy = pol == &default_policy ? MPOL_DEFAULT :
987 * Internal mempolicy flags must be masked off before exposing
988 * the policy to userspace.
990 *policy |= (pol->flags & MPOL_MODE_FLAGS);
995 if (mpol_store_user_nodemask(pol)) {
996 *nmask = pol->w.user_nodemask;
999 get_policy_nodemask(pol, nmask);
1000 task_unlock(current);
1007 up_read(&mm->mmap_sem);
1009 mpol_put(pol_refcount);
1013 #ifdef CONFIG_MIGRATION
1015 * page migration, thp tail pages can be passed.
1017 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1018 unsigned long flags)
1020 struct page *head = compound_head(page);
1022 * Avoid migrating a page that is shared with others.
1024 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1025 if (!isolate_lru_page(head)) {
1026 list_add_tail(&head->lru, pagelist);
1027 mod_node_page_state(page_pgdat(head),
1028 NR_ISOLATED_ANON + page_is_file_lru(head),
1029 hpage_nr_pages(head));
1030 } else if (flags & MPOL_MF_STRICT) {
1032 * Non-movable page may reach here. And, there may be
1033 * temporary off LRU pages or non-LRU movable pages.
1034 * Treat them as unmovable pages since they can't be
1035 * isolated, so they can't be moved at the moment. It
1036 * should return -EIO for this case too.
1045 /* page allocation callback for NUMA node migration */
1046 struct page *alloc_new_node_page(struct page *page, unsigned long node)
1049 return alloc_huge_page_node(page_hstate(compound_head(page)),
1051 else if (PageTransHuge(page)) {
1054 thp = alloc_pages_node(node,
1055 (GFP_TRANSHUGE | __GFP_THISNODE),
1059 prep_transhuge_page(thp);
1062 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1067 * Migrate pages from one node to a target node.
1068 * Returns error or the number of pages not migrated.
1070 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1074 LIST_HEAD(pagelist);
1078 node_set(source, nmask);
1081 * This does not "check" the range but isolates all pages that
1082 * need migration. Between passing in the full user address
1083 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1085 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1086 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1087 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1089 if (!list_empty(&pagelist)) {
1090 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1091 MIGRATE_SYNC, MR_SYSCALL);
1093 putback_movable_pages(&pagelist);
1100 * Move pages between the two nodesets so as to preserve the physical
1101 * layout as much as possible.
1103 * Returns the number of page that could not be moved.
1105 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1106 const nodemask_t *to, int flags)
1112 err = migrate_prep();
1116 down_read(&mm->mmap_sem);
1119 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1120 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1121 * bit in 'tmp', and return that <source, dest> pair for migration.
1122 * The pair of nodemasks 'to' and 'from' define the map.
1124 * If no pair of bits is found that way, fallback to picking some
1125 * pair of 'source' and 'dest' bits that are not the same. If the
1126 * 'source' and 'dest' bits are the same, this represents a node
1127 * that will be migrating to itself, so no pages need move.
1129 * If no bits are left in 'tmp', or if all remaining bits left
1130 * in 'tmp' correspond to the same bit in 'to', return false
1131 * (nothing left to migrate).
1133 * This lets us pick a pair of nodes to migrate between, such that
1134 * if possible the dest node is not already occupied by some other
1135 * source node, minimizing the risk of overloading the memory on a
1136 * node that would happen if we migrated incoming memory to a node
1137 * before migrating outgoing memory source that same node.
1139 * A single scan of tmp is sufficient. As we go, we remember the
1140 * most recent <s, d> pair that moved (s != d). If we find a pair
1141 * that not only moved, but what's better, moved to an empty slot
1142 * (d is not set in tmp), then we break out then, with that pair.
1143 * Otherwise when we finish scanning from_tmp, we at least have the
1144 * most recent <s, d> pair that moved. If we get all the way through
1145 * the scan of tmp without finding any node that moved, much less
1146 * moved to an empty node, then there is nothing left worth migrating.
1150 while (!nodes_empty(tmp)) {
1152 int source = NUMA_NO_NODE;
1155 for_each_node_mask(s, tmp) {
1158 * do_migrate_pages() tries to maintain the relative
1159 * node relationship of the pages established between
1160 * threads and memory areas.
1162 * However if the number of source nodes is not equal to
1163 * the number of destination nodes we can not preserve
1164 * this node relative relationship. In that case, skip
1165 * copying memory from a node that is in the destination
1168 * Example: [2,3,4] -> [3,4,5] moves everything.
1169 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1172 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1173 (node_isset(s, *to)))
1176 d = node_remap(s, *from, *to);
1180 source = s; /* Node moved. Memorize */
1183 /* dest not in remaining from nodes? */
1184 if (!node_isset(dest, tmp))
1187 if (source == NUMA_NO_NODE)
1190 node_clear(source, tmp);
1191 err = migrate_to_node(mm, source, dest, flags);
1197 up_read(&mm->mmap_sem);
1205 * Allocate a new page for page migration based on vma policy.
1206 * Start by assuming the page is mapped by the same vma as contains @start.
1207 * Search forward from there, if not. N.B., this assumes that the
1208 * list of pages handed to migrate_pages()--which is how we get here--
1209 * is in virtual address order.
1211 static struct page *new_page(struct page *page, unsigned long start)
1213 struct vm_area_struct *vma;
1214 unsigned long uninitialized_var(address);
1216 vma = find_vma(current->mm, start);
1218 address = page_address_in_vma(page, vma);
1219 if (address != -EFAULT)
1224 if (PageHuge(page)) {
1225 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1227 } else if (PageTransHuge(page)) {
1230 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1234 prep_transhuge_page(thp);
1238 * if !vma, alloc_page_vma() will use task or system default policy
1240 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1245 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1246 unsigned long flags)
1251 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1252 const nodemask_t *to, int flags)
1257 static struct page *new_page(struct page *page, unsigned long start)
1263 static long do_mbind(unsigned long start, unsigned long len,
1264 unsigned short mode, unsigned short mode_flags,
1265 nodemask_t *nmask, unsigned long flags)
1267 struct mm_struct *mm = current->mm;
1268 struct mempolicy *new;
1272 LIST_HEAD(pagelist);
1274 if (flags & ~(unsigned long)MPOL_MF_VALID)
1276 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1279 if (start & ~PAGE_MASK)
1282 if (mode == MPOL_DEFAULT)
1283 flags &= ~MPOL_MF_STRICT;
1285 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1293 new = mpol_new(mode, mode_flags, nmask);
1295 return PTR_ERR(new);
1297 if (flags & MPOL_MF_LAZY)
1298 new->flags |= MPOL_F_MOF;
1301 * If we are using the default policy then operation
1302 * on discontinuous address spaces is okay after all
1305 flags |= MPOL_MF_DISCONTIG_OK;
1307 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1308 start, start + len, mode, mode_flags,
1309 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1311 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1313 err = migrate_prep();
1318 NODEMASK_SCRATCH(scratch);
1320 down_write(&mm->mmap_sem);
1322 err = mpol_set_nodemask(new, nmask, scratch);
1323 task_unlock(current);
1325 up_write(&mm->mmap_sem);
1328 NODEMASK_SCRATCH_FREE(scratch);
1333 ret = queue_pages_range(mm, start, end, nmask,
1334 flags | MPOL_MF_INVERT, &pagelist);
1341 err = mbind_range(mm, start, end, new);
1346 if (!list_empty(&pagelist)) {
1347 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1348 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1349 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1351 putback_movable_pages(&pagelist);
1354 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1358 if (!list_empty(&pagelist))
1359 putback_movable_pages(&pagelist);
1362 up_write(&mm->mmap_sem);
1369 * User space interface with variable sized bitmaps for nodelists.
1372 /* Copy a node mask from user space. */
1373 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1374 unsigned long maxnode)
1378 unsigned long nlongs;
1379 unsigned long endmask;
1382 nodes_clear(*nodes);
1383 if (maxnode == 0 || !nmask)
1385 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1388 nlongs = BITS_TO_LONGS(maxnode);
1389 if ((maxnode % BITS_PER_LONG) == 0)
1392 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1395 * When the user specified more nodes than supported just check
1396 * if the non supported part is all zero.
1398 * If maxnode have more longs than MAX_NUMNODES, check
1399 * the bits in that area first. And then go through to
1400 * check the rest bits which equal or bigger than MAX_NUMNODES.
1401 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1403 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1404 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1405 if (get_user(t, nmask + k))
1407 if (k == nlongs - 1) {
1413 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1417 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1418 unsigned long valid_mask = endmask;
1420 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1421 if (get_user(t, nmask + nlongs - 1))
1427 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1429 nodes_addr(*nodes)[nlongs-1] &= endmask;
1433 /* Copy a kernel node mask to user space */
1434 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1437 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1438 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1440 if (copy > nbytes) {
1441 if (copy > PAGE_SIZE)
1443 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1447 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1450 static long kernel_mbind(unsigned long start, unsigned long len,
1451 unsigned long mode, const unsigned long __user *nmask,
1452 unsigned long maxnode, unsigned int flags)
1456 unsigned short mode_flags;
1458 start = untagged_addr(start);
1459 mode_flags = mode & MPOL_MODE_FLAGS;
1460 mode &= ~MPOL_MODE_FLAGS;
1461 if (mode >= MPOL_MAX)
1463 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1464 (mode_flags & MPOL_F_RELATIVE_NODES))
1466 err = get_nodes(&nodes, nmask, maxnode);
1469 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1472 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1473 unsigned long, mode, const unsigned long __user *, nmask,
1474 unsigned long, maxnode, unsigned int, flags)
1476 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1479 /* Set the process memory policy */
1480 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1481 unsigned long maxnode)
1485 unsigned short flags;
1487 flags = mode & MPOL_MODE_FLAGS;
1488 mode &= ~MPOL_MODE_FLAGS;
1489 if ((unsigned int)mode >= MPOL_MAX)
1491 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1493 err = get_nodes(&nodes, nmask, maxnode);
1496 return do_set_mempolicy(mode, flags, &nodes);
1499 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1500 unsigned long, maxnode)
1502 return kernel_set_mempolicy(mode, nmask, maxnode);
1505 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1506 const unsigned long __user *old_nodes,
1507 const unsigned long __user *new_nodes)
1509 struct mm_struct *mm = NULL;
1510 struct task_struct *task;
1511 nodemask_t task_nodes;
1515 NODEMASK_SCRATCH(scratch);
1520 old = &scratch->mask1;
1521 new = &scratch->mask2;
1523 err = get_nodes(old, old_nodes, maxnode);
1527 err = get_nodes(new, new_nodes, maxnode);
1531 /* Find the mm_struct */
1533 task = pid ? find_task_by_vpid(pid) : current;
1539 get_task_struct(task);
1544 * Check if this process has the right to modify the specified process.
1545 * Use the regular "ptrace_may_access()" checks.
1547 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1554 task_nodes = cpuset_mems_allowed(task);
1555 /* Is the user allowed to access the target nodes? */
1556 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1561 task_nodes = cpuset_mems_allowed(current);
1562 nodes_and(*new, *new, task_nodes);
1563 if (nodes_empty(*new))
1566 err = security_task_movememory(task);
1570 mm = get_task_mm(task);
1571 put_task_struct(task);
1578 err = do_migrate_pages(mm, old, new,
1579 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1583 NODEMASK_SCRATCH_FREE(scratch);
1588 put_task_struct(task);
1593 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1594 const unsigned long __user *, old_nodes,
1595 const unsigned long __user *, new_nodes)
1597 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1601 /* Retrieve NUMA policy */
1602 static int kernel_get_mempolicy(int __user *policy,
1603 unsigned long __user *nmask,
1604 unsigned long maxnode,
1606 unsigned long flags)
1609 int uninitialized_var(pval);
1612 addr = untagged_addr(addr);
1614 if (nmask != NULL && maxnode < nr_node_ids)
1617 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1622 if (policy && put_user(pval, policy))
1626 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1631 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1632 unsigned long __user *, nmask, unsigned long, maxnode,
1633 unsigned long, addr, unsigned long, flags)
1635 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1638 #ifdef CONFIG_COMPAT
1640 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1641 compat_ulong_t __user *, nmask,
1642 compat_ulong_t, maxnode,
1643 compat_ulong_t, addr, compat_ulong_t, flags)
1646 unsigned long __user *nm = NULL;
1647 unsigned long nr_bits, alloc_size;
1648 DECLARE_BITMAP(bm, MAX_NUMNODES);
1650 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1651 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1654 nm = compat_alloc_user_space(alloc_size);
1656 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1658 if (!err && nmask) {
1659 unsigned long copy_size;
1660 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1661 err = copy_from_user(bm, nm, copy_size);
1662 /* ensure entire bitmap is zeroed */
1663 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1664 err |= compat_put_bitmap(nmask, bm, nr_bits);
1670 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1671 compat_ulong_t, maxnode)
1673 unsigned long __user *nm = NULL;
1674 unsigned long nr_bits, alloc_size;
1675 DECLARE_BITMAP(bm, MAX_NUMNODES);
1677 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1678 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1681 if (compat_get_bitmap(bm, nmask, nr_bits))
1683 nm = compat_alloc_user_space(alloc_size);
1684 if (copy_to_user(nm, bm, alloc_size))
1688 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1691 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1692 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1693 compat_ulong_t, maxnode, compat_ulong_t, flags)
1695 unsigned long __user *nm = NULL;
1696 unsigned long nr_bits, alloc_size;
1699 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1700 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1703 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1705 nm = compat_alloc_user_space(alloc_size);
1706 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1710 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1713 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1714 compat_ulong_t, maxnode,
1715 const compat_ulong_t __user *, old_nodes,
1716 const compat_ulong_t __user *, new_nodes)
1718 unsigned long __user *old = NULL;
1719 unsigned long __user *new = NULL;
1720 nodemask_t tmp_mask;
1721 unsigned long nr_bits;
1724 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1725 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1727 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1729 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1731 new = old + size / sizeof(unsigned long);
1732 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1736 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1739 new = compat_alloc_user_space(size);
1740 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1743 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1746 #endif /* CONFIG_COMPAT */
1748 bool vma_migratable(struct vm_area_struct *vma)
1750 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1754 * DAX device mappings require predictable access latency, so avoid
1755 * incurring periodic faults.
1757 if (vma_is_dax(vma))
1760 if (is_vm_hugetlb_page(vma) &&
1761 !hugepage_migration_supported(hstate_vma(vma)))
1765 * Migration allocates pages in the highest zone. If we cannot
1766 * do so then migration (at least from node to node) is not
1770 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1776 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1779 struct mempolicy *pol = NULL;
1782 if (vma->vm_ops && vma->vm_ops->get_policy) {
1783 pol = vma->vm_ops->get_policy(vma, addr);
1784 } else if (vma->vm_policy) {
1785 pol = vma->vm_policy;
1788 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1789 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1790 * count on these policies which will be dropped by
1791 * mpol_cond_put() later
1793 if (mpol_needs_cond_ref(pol))
1802 * get_vma_policy(@vma, @addr)
1803 * @vma: virtual memory area whose policy is sought
1804 * @addr: address in @vma for shared policy lookup
1806 * Returns effective policy for a VMA at specified address.
1807 * Falls back to current->mempolicy or system default policy, as necessary.
1808 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1809 * count--added by the get_policy() vm_op, as appropriate--to protect against
1810 * freeing by another task. It is the caller's responsibility to free the
1811 * extra reference for shared policies.
1813 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1816 struct mempolicy *pol = __get_vma_policy(vma, addr);
1819 pol = get_task_policy(current);
1824 bool vma_policy_mof(struct vm_area_struct *vma)
1826 struct mempolicy *pol;
1828 if (vma->vm_ops && vma->vm_ops->get_policy) {
1831 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1832 if (pol && (pol->flags & MPOL_F_MOF))
1839 pol = vma->vm_policy;
1841 pol = get_task_policy(current);
1843 return pol->flags & MPOL_F_MOF;
1846 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1848 enum zone_type dynamic_policy_zone = policy_zone;
1850 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1853 * if policy->v.nodes has movable memory only,
1854 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1856 * policy->v.nodes is intersect with node_states[N_MEMORY].
1857 * so if the following test faile, it implies
1858 * policy->v.nodes has movable memory only.
1860 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1861 dynamic_policy_zone = ZONE_MOVABLE;
1863 return zone >= dynamic_policy_zone;
1867 * Return a nodemask representing a mempolicy for filtering nodes for
1870 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1872 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1873 if (unlikely(policy->mode == MPOL_BIND) &&
1874 apply_policy_zone(policy, gfp_zone(gfp)) &&
1875 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1876 return &policy->v.nodes;
1881 /* Return the node id preferred by the given mempolicy, or the given id */
1882 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1885 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1886 nd = policy->v.preferred_node;
1889 * __GFP_THISNODE shouldn't even be used with the bind policy
1890 * because we might easily break the expectation to stay on the
1891 * requested node and not break the policy.
1893 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1899 /* Do dynamic interleaving for a process */
1900 static unsigned interleave_nodes(struct mempolicy *policy)
1903 struct task_struct *me = current;
1905 next = next_node_in(me->il_prev, policy->v.nodes);
1906 if (next < MAX_NUMNODES)
1912 * Depending on the memory policy provide a node from which to allocate the
1915 unsigned int mempolicy_slab_node(void)
1917 struct mempolicy *policy;
1918 int node = numa_mem_id();
1923 policy = current->mempolicy;
1924 if (!policy || policy->flags & MPOL_F_LOCAL)
1927 switch (policy->mode) {
1928 case MPOL_PREFERRED:
1930 * handled MPOL_F_LOCAL above
1932 return policy->v.preferred_node;
1934 case MPOL_INTERLEAVE:
1935 return interleave_nodes(policy);
1941 * Follow bind policy behavior and start allocation at the
1944 struct zonelist *zonelist;
1945 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1946 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1947 z = first_zones_zonelist(zonelist, highest_zoneidx,
1949 return z->zone ? zone_to_nid(z->zone) : node;
1958 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1959 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1960 * number of present nodes.
1962 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1964 unsigned nnodes = nodes_weight(pol->v.nodes);
1970 return numa_node_id();
1971 target = (unsigned int)n % nnodes;
1972 nid = first_node(pol->v.nodes);
1973 for (i = 0; i < target; i++)
1974 nid = next_node(nid, pol->v.nodes);
1978 /* Determine a node number for interleave */
1979 static inline unsigned interleave_nid(struct mempolicy *pol,
1980 struct vm_area_struct *vma, unsigned long addr, int shift)
1986 * for small pages, there is no difference between
1987 * shift and PAGE_SHIFT, so the bit-shift is safe.
1988 * for huge pages, since vm_pgoff is in units of small
1989 * pages, we need to shift off the always 0 bits to get
1992 BUG_ON(shift < PAGE_SHIFT);
1993 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1994 off += (addr - vma->vm_start) >> shift;
1995 return offset_il_node(pol, off);
1997 return interleave_nodes(pol);
2000 #ifdef CONFIG_HUGETLBFS
2002 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2003 * @vma: virtual memory area whose policy is sought
2004 * @addr: address in @vma for shared policy lookup and interleave policy
2005 * @gfp_flags: for requested zone
2006 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2007 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2009 * Returns a nid suitable for a huge page allocation and a pointer
2010 * to the struct mempolicy for conditional unref after allocation.
2011 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2012 * @nodemask for filtering the zonelist.
2014 * Must be protected by read_mems_allowed_begin()
2016 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2017 struct mempolicy **mpol, nodemask_t **nodemask)
2021 *mpol = get_vma_policy(vma, addr);
2022 *nodemask = NULL; /* assume !MPOL_BIND */
2024 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
2025 nid = interleave_nid(*mpol, vma, addr,
2026 huge_page_shift(hstate_vma(vma)));
2028 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2029 if ((*mpol)->mode == MPOL_BIND)
2030 *nodemask = &(*mpol)->v.nodes;
2036 * init_nodemask_of_mempolicy
2038 * If the current task's mempolicy is "default" [NULL], return 'false'
2039 * to indicate default policy. Otherwise, extract the policy nodemask
2040 * for 'bind' or 'interleave' policy into the argument nodemask, or
2041 * initialize the argument nodemask to contain the single node for
2042 * 'preferred' or 'local' policy and return 'true' to indicate presence
2043 * of non-default mempolicy.
2045 * We don't bother with reference counting the mempolicy [mpol_get/put]
2046 * because the current task is examining it's own mempolicy and a task's
2047 * mempolicy is only ever changed by the task itself.
2049 * N.B., it is the caller's responsibility to free a returned nodemask.
2051 bool init_nodemask_of_mempolicy(nodemask_t *mask)
2053 struct mempolicy *mempolicy;
2056 if (!(mask && current->mempolicy))
2060 mempolicy = current->mempolicy;
2061 switch (mempolicy->mode) {
2062 case MPOL_PREFERRED:
2063 if (mempolicy->flags & MPOL_F_LOCAL)
2064 nid = numa_node_id();
2066 nid = mempolicy->v.preferred_node;
2067 init_nodemask_of_node(mask, nid);
2071 case MPOL_INTERLEAVE:
2072 *mask = mempolicy->v.nodes;
2078 task_unlock(current);
2085 * mempolicy_nodemask_intersects
2087 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2088 * policy. Otherwise, check for intersection between mask and the policy
2089 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2090 * policy, always return true since it may allocate elsewhere on fallback.
2092 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2094 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2095 const nodemask_t *mask)
2097 struct mempolicy *mempolicy;
2103 mempolicy = tsk->mempolicy;
2107 switch (mempolicy->mode) {
2108 case MPOL_PREFERRED:
2110 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2111 * allocate from, they may fallback to other nodes when oom.
2112 * Thus, it's possible for tsk to have allocated memory from
2117 case MPOL_INTERLEAVE:
2118 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2128 /* Allocate a page in interleaved policy.
2129 Own path because it needs to do special accounting. */
2130 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2135 page = __alloc_pages(gfp, order, nid);
2136 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2137 if (!static_branch_likely(&vm_numa_stat_key))
2139 if (page && page_to_nid(page) == nid) {
2141 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2148 * alloc_pages_vma - Allocate a page for a VMA.
2151 * %GFP_USER user allocation.
2152 * %GFP_KERNEL kernel allocations,
2153 * %GFP_HIGHMEM highmem/user allocations,
2154 * %GFP_FS allocation should not call back into a file system.
2155 * %GFP_ATOMIC don't sleep.
2157 * @order:Order of the GFP allocation.
2158 * @vma: Pointer to VMA or NULL if not available.
2159 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2160 * @node: Which node to prefer for allocation (modulo policy).
2161 * @hugepage: for hugepages try only the preferred node if possible
2163 * This function allocates a page from the kernel page pool and applies
2164 * a NUMA policy associated with the VMA or the current process.
2165 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2166 * mm_struct of the VMA to prevent it from going away. Should be used for
2167 * all allocations for pages that will be mapped into user space. Returns
2168 * NULL when no page can be allocated.
2171 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2172 unsigned long addr, int node, bool hugepage)
2174 struct mempolicy *pol;
2179 pol = get_vma_policy(vma, addr);
2181 if (pol->mode == MPOL_INTERLEAVE) {
2184 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2186 page = alloc_page_interleave(gfp, order, nid);
2190 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2191 int hpage_node = node;
2194 * For hugepage allocation and non-interleave policy which
2195 * allows the current node (or other explicitly preferred
2196 * node) we only try to allocate from the current/preferred
2197 * node and don't fall back to other nodes, as the cost of
2198 * remote accesses would likely offset THP benefits.
2200 * If the policy is interleave, or does not allow the current
2201 * node in its nodemask, we allocate the standard way.
2203 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2204 hpage_node = pol->v.preferred_node;
2206 nmask = policy_nodemask(gfp, pol);
2207 if (!nmask || node_isset(hpage_node, *nmask)) {
2210 * First, try to allocate THP only on local node, but
2211 * don't reclaim unnecessarily, just compact.
2213 page = __alloc_pages_node(hpage_node,
2214 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2217 * If hugepage allocations are configured to always
2218 * synchronous compact or the vma has been madvised
2219 * to prefer hugepage backing, retry allowing remote
2220 * memory with both reclaim and compact as well.
2222 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2223 page = __alloc_pages_node(hpage_node,
2230 nmask = policy_nodemask(gfp, pol);
2231 preferred_nid = policy_node(gfp, pol, node);
2232 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2237 EXPORT_SYMBOL(alloc_pages_vma);
2240 * alloc_pages_current - Allocate pages.
2243 * %GFP_USER user allocation,
2244 * %GFP_KERNEL kernel allocation,
2245 * %GFP_HIGHMEM highmem allocation,
2246 * %GFP_FS don't call back into a file system.
2247 * %GFP_ATOMIC don't sleep.
2248 * @order: Power of two of allocation size in pages. 0 is a single page.
2250 * Allocate a page from the kernel page pool. When not in
2251 * interrupt context and apply the current process NUMA policy.
2252 * Returns NULL when no page can be allocated.
2254 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2256 struct mempolicy *pol = &default_policy;
2259 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2260 pol = get_task_policy(current);
2263 * No reference counting needed for current->mempolicy
2264 * nor system default_policy
2266 if (pol->mode == MPOL_INTERLEAVE)
2267 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2269 page = __alloc_pages_nodemask(gfp, order,
2270 policy_node(gfp, pol, numa_node_id()),
2271 policy_nodemask(gfp, pol));
2275 EXPORT_SYMBOL(alloc_pages_current);
2277 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2279 struct mempolicy *pol = mpol_dup(vma_policy(src));
2282 return PTR_ERR(pol);
2283 dst->vm_policy = pol;
2288 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2289 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2290 * with the mems_allowed returned by cpuset_mems_allowed(). This
2291 * keeps mempolicies cpuset relative after its cpuset moves. See
2292 * further kernel/cpuset.c update_nodemask().
2294 * current's mempolicy may be rebinded by the other task(the task that changes
2295 * cpuset's mems), so we needn't do rebind work for current task.
2298 /* Slow path of a mempolicy duplicate */
2299 struct mempolicy *__mpol_dup(struct mempolicy *old)
2301 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2304 return ERR_PTR(-ENOMEM);
2306 /* task's mempolicy is protected by alloc_lock */
2307 if (old == current->mempolicy) {
2310 task_unlock(current);
2314 if (current_cpuset_is_being_rebound()) {
2315 nodemask_t mems = cpuset_mems_allowed(current);
2316 mpol_rebind_policy(new, &mems);
2318 atomic_set(&new->refcnt, 1);
2322 /* Slow path of a mempolicy comparison */
2323 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2327 if (a->mode != b->mode)
2329 if (a->flags != b->flags)
2331 if (mpol_store_user_nodemask(a))
2332 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2337 case MPOL_INTERLEAVE:
2338 return !!nodes_equal(a->v.nodes, b->v.nodes);
2339 case MPOL_PREFERRED:
2340 /* a's ->flags is the same as b's */
2341 if (a->flags & MPOL_F_LOCAL)
2343 return a->v.preferred_node == b->v.preferred_node;
2351 * Shared memory backing store policy support.
2353 * Remember policies even when nobody has shared memory mapped.
2354 * The policies are kept in Red-Black tree linked from the inode.
2355 * They are protected by the sp->lock rwlock, which should be held
2356 * for any accesses to the tree.
2360 * lookup first element intersecting start-end. Caller holds sp->lock for
2361 * reading or for writing
2363 static struct sp_node *
2364 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2366 struct rb_node *n = sp->root.rb_node;
2369 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2371 if (start >= p->end)
2373 else if (end <= p->start)
2381 struct sp_node *w = NULL;
2382 struct rb_node *prev = rb_prev(n);
2385 w = rb_entry(prev, struct sp_node, nd);
2386 if (w->end <= start)
2390 return rb_entry(n, struct sp_node, nd);
2394 * Insert a new shared policy into the list. Caller holds sp->lock for
2397 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2399 struct rb_node **p = &sp->root.rb_node;
2400 struct rb_node *parent = NULL;
2405 nd = rb_entry(parent, struct sp_node, nd);
2406 if (new->start < nd->start)
2408 else if (new->end > nd->end)
2409 p = &(*p)->rb_right;
2413 rb_link_node(&new->nd, parent, p);
2414 rb_insert_color(&new->nd, &sp->root);
2415 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2416 new->policy ? new->policy->mode : 0);
2419 /* Find shared policy intersecting idx */
2421 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2423 struct mempolicy *pol = NULL;
2426 if (!sp->root.rb_node)
2428 read_lock(&sp->lock);
2429 sn = sp_lookup(sp, idx, idx+1);
2431 mpol_get(sn->policy);
2434 read_unlock(&sp->lock);
2438 static void sp_free(struct sp_node *n)
2440 mpol_put(n->policy);
2441 kmem_cache_free(sn_cache, n);
2445 * mpol_misplaced - check whether current page node is valid in policy
2447 * @page: page to be checked
2448 * @vma: vm area where page mapped
2449 * @addr: virtual address where page mapped
2451 * Lookup current policy node id for vma,addr and "compare to" page's
2455 * -1 - not misplaced, page is in the right node
2456 * node - node id where the page should be
2458 * Policy determination "mimics" alloc_page_vma().
2459 * Called from fault path where we know the vma and faulting address.
2461 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2463 struct mempolicy *pol;
2465 int curnid = page_to_nid(page);
2466 unsigned long pgoff;
2467 int thiscpu = raw_smp_processor_id();
2468 int thisnid = cpu_to_node(thiscpu);
2469 int polnid = NUMA_NO_NODE;
2472 pol = get_vma_policy(vma, addr);
2473 if (!(pol->flags & MPOL_F_MOF))
2476 switch (pol->mode) {
2477 case MPOL_INTERLEAVE:
2478 pgoff = vma->vm_pgoff;
2479 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2480 polnid = offset_il_node(pol, pgoff);
2483 case MPOL_PREFERRED:
2484 if (pol->flags & MPOL_F_LOCAL)
2485 polnid = numa_node_id();
2487 polnid = pol->v.preferred_node;
2493 * allows binding to multiple nodes.
2494 * use current page if in policy nodemask,
2495 * else select nearest allowed node, if any.
2496 * If no allowed nodes, use current [!misplaced].
2498 if (node_isset(curnid, pol->v.nodes))
2500 z = first_zones_zonelist(
2501 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2502 gfp_zone(GFP_HIGHUSER),
2504 polnid = zone_to_nid(z->zone);
2511 /* Migrate the page towards the node whose CPU is referencing it */
2512 if (pol->flags & MPOL_F_MORON) {
2515 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2519 if (curnid != polnid)
2528 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2529 * dropped after task->mempolicy is set to NULL so that any allocation done as
2530 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2533 void mpol_put_task_policy(struct task_struct *task)
2535 struct mempolicy *pol;
2538 pol = task->mempolicy;
2539 task->mempolicy = NULL;
2544 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2546 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2547 rb_erase(&n->nd, &sp->root);
2551 static void sp_node_init(struct sp_node *node, unsigned long start,
2552 unsigned long end, struct mempolicy *pol)
2554 node->start = start;
2559 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2560 struct mempolicy *pol)
2563 struct mempolicy *newpol;
2565 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2569 newpol = mpol_dup(pol);
2570 if (IS_ERR(newpol)) {
2571 kmem_cache_free(sn_cache, n);
2574 newpol->flags |= MPOL_F_SHARED;
2575 sp_node_init(n, start, end, newpol);
2580 /* Replace a policy range. */
2581 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2582 unsigned long end, struct sp_node *new)
2585 struct sp_node *n_new = NULL;
2586 struct mempolicy *mpol_new = NULL;
2590 write_lock(&sp->lock);
2591 n = sp_lookup(sp, start, end);
2592 /* Take care of old policies in the same range. */
2593 while (n && n->start < end) {
2594 struct rb_node *next = rb_next(&n->nd);
2595 if (n->start >= start) {
2601 /* Old policy spanning whole new range. */
2606 *mpol_new = *n->policy;
2607 atomic_set(&mpol_new->refcnt, 1);
2608 sp_node_init(n_new, end, n->end, mpol_new);
2610 sp_insert(sp, n_new);
2619 n = rb_entry(next, struct sp_node, nd);
2623 write_unlock(&sp->lock);
2630 kmem_cache_free(sn_cache, n_new);
2635 write_unlock(&sp->lock);
2637 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2640 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2647 * mpol_shared_policy_init - initialize shared policy for inode
2648 * @sp: pointer to inode shared policy
2649 * @mpol: struct mempolicy to install
2651 * Install non-NULL @mpol in inode's shared policy rb-tree.
2652 * On entry, the current task has a reference on a non-NULL @mpol.
2653 * This must be released on exit.
2654 * This is called at get_inode() calls and we can use GFP_KERNEL.
2656 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2660 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2661 rwlock_init(&sp->lock);
2664 struct vm_area_struct pvma;
2665 struct mempolicy *new;
2666 NODEMASK_SCRATCH(scratch);
2670 /* contextualize the tmpfs mount point mempolicy */
2671 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2673 goto free_scratch; /* no valid nodemask intersection */
2676 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2677 task_unlock(current);
2681 /* Create pseudo-vma that contains just the policy */
2682 vma_init(&pvma, NULL);
2683 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2684 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2687 mpol_put(new); /* drop initial ref */
2689 NODEMASK_SCRATCH_FREE(scratch);
2691 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2695 int mpol_set_shared_policy(struct shared_policy *info,
2696 struct vm_area_struct *vma, struct mempolicy *npol)
2699 struct sp_node *new = NULL;
2700 unsigned long sz = vma_pages(vma);
2702 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2704 sz, npol ? npol->mode : -1,
2705 npol ? npol->flags : -1,
2706 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2709 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2713 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2719 /* Free a backing policy store on inode delete. */
2720 void mpol_free_shared_policy(struct shared_policy *p)
2723 struct rb_node *next;
2725 if (!p->root.rb_node)
2727 write_lock(&p->lock);
2728 next = rb_first(&p->root);
2730 n = rb_entry(next, struct sp_node, nd);
2731 next = rb_next(&n->nd);
2734 write_unlock(&p->lock);
2737 #ifdef CONFIG_NUMA_BALANCING
2738 static int __initdata numabalancing_override;
2740 static void __init check_numabalancing_enable(void)
2742 bool numabalancing_default = false;
2744 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2745 numabalancing_default = true;
2747 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2748 if (numabalancing_override)
2749 set_numabalancing_state(numabalancing_override == 1);
2751 if (num_online_nodes() > 1 && !numabalancing_override) {
2752 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2753 numabalancing_default ? "Enabling" : "Disabling");
2754 set_numabalancing_state(numabalancing_default);
2758 static int __init setup_numabalancing(char *str)
2764 if (!strcmp(str, "enable")) {
2765 numabalancing_override = 1;
2767 } else if (!strcmp(str, "disable")) {
2768 numabalancing_override = -1;
2773 pr_warn("Unable to parse numa_balancing=\n");
2777 __setup("numa_balancing=", setup_numabalancing);
2779 static inline void __init check_numabalancing_enable(void)
2782 #endif /* CONFIG_NUMA_BALANCING */
2784 /* assumes fs == KERNEL_DS */
2785 void __init numa_policy_init(void)
2787 nodemask_t interleave_nodes;
2788 unsigned long largest = 0;
2789 int nid, prefer = 0;
2791 policy_cache = kmem_cache_create("numa_policy",
2792 sizeof(struct mempolicy),
2793 0, SLAB_PANIC, NULL);
2795 sn_cache = kmem_cache_create("shared_policy_node",
2796 sizeof(struct sp_node),
2797 0, SLAB_PANIC, NULL);
2799 for_each_node(nid) {
2800 preferred_node_policy[nid] = (struct mempolicy) {
2801 .refcnt = ATOMIC_INIT(1),
2802 .mode = MPOL_PREFERRED,
2803 .flags = MPOL_F_MOF | MPOL_F_MORON,
2804 .v = { .preferred_node = nid, },
2809 * Set interleaving policy for system init. Interleaving is only
2810 * enabled across suitably sized nodes (default is >= 16MB), or
2811 * fall back to the largest node if they're all smaller.
2813 nodes_clear(interleave_nodes);
2814 for_each_node_state(nid, N_MEMORY) {
2815 unsigned long total_pages = node_present_pages(nid);
2817 /* Preserve the largest node */
2818 if (largest < total_pages) {
2819 largest = total_pages;
2823 /* Interleave this node? */
2824 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2825 node_set(nid, interleave_nodes);
2828 /* All too small, use the largest */
2829 if (unlikely(nodes_empty(interleave_nodes)))
2830 node_set(prefer, interleave_nodes);
2832 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2833 pr_err("%s: interleaving failed\n", __func__);
2835 check_numabalancing_enable();
2838 /* Reset policy of current process to default */
2839 void numa_default_policy(void)
2841 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2845 * Parse and format mempolicy from/to strings
2849 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2851 static const char * const policy_modes[] =
2853 [MPOL_DEFAULT] = "default",
2854 [MPOL_PREFERRED] = "prefer",
2855 [MPOL_BIND] = "bind",
2856 [MPOL_INTERLEAVE] = "interleave",
2857 [MPOL_LOCAL] = "local",
2863 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2864 * @str: string containing mempolicy to parse
2865 * @mpol: pointer to struct mempolicy pointer, returned on success.
2868 * <mode>[=<flags>][:<nodelist>]
2870 * On success, returns 0, else 1
2872 int mpol_parse_str(char *str, struct mempolicy **mpol)
2874 struct mempolicy *new = NULL;
2875 unsigned short mode_flags;
2877 char *nodelist = strchr(str, ':');
2878 char *flags = strchr(str, '=');
2882 *flags++ = '\0'; /* terminate mode string */
2885 /* NUL-terminate mode or flags string */
2887 if (nodelist_parse(nodelist, nodes))
2889 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2894 mode = match_string(policy_modes, MPOL_MAX, str);
2899 case MPOL_PREFERRED:
2901 * Insist on a nodelist of one node only, although later
2902 * we use first_node(nodes) to grab a single node, so here
2903 * nodelist (or nodes) cannot be empty.
2906 char *rest = nodelist;
2907 while (isdigit(*rest))
2911 if (nodes_empty(nodes))
2915 case MPOL_INTERLEAVE:
2917 * Default to online nodes with memory if no nodelist
2920 nodes = node_states[N_MEMORY];
2924 * Don't allow a nodelist; mpol_new() checks flags
2928 mode = MPOL_PREFERRED;
2932 * Insist on a empty nodelist
2939 * Insist on a nodelist
2948 * Currently, we only support two mutually exclusive
2951 if (!strcmp(flags, "static"))
2952 mode_flags |= MPOL_F_STATIC_NODES;
2953 else if (!strcmp(flags, "relative"))
2954 mode_flags |= MPOL_F_RELATIVE_NODES;
2959 new = mpol_new(mode, mode_flags, &nodes);
2964 * Save nodes for mpol_to_str() to show the tmpfs mount options
2965 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2967 if (mode != MPOL_PREFERRED)
2968 new->v.nodes = nodes;
2970 new->v.preferred_node = first_node(nodes);
2972 new->flags |= MPOL_F_LOCAL;
2975 * Save nodes for contextualization: this will be used to "clone"
2976 * the mempolicy in a specific context [cpuset] at a later time.
2978 new->w.user_nodemask = nodes;
2983 /* Restore string for error message */
2992 #endif /* CONFIG_TMPFS */
2995 * mpol_to_str - format a mempolicy structure for printing
2996 * @buffer: to contain formatted mempolicy string
2997 * @maxlen: length of @buffer
2998 * @pol: pointer to mempolicy to be formatted
3000 * Convert @pol into a string. If @buffer is too short, truncate the string.
3001 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3002 * longest flag, "relative", and to display at least a few node ids.
3004 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3007 nodemask_t nodes = NODE_MASK_NONE;
3008 unsigned short mode = MPOL_DEFAULT;
3009 unsigned short flags = 0;
3011 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3019 case MPOL_PREFERRED:
3020 if (flags & MPOL_F_LOCAL)
3023 node_set(pol->v.preferred_node, nodes);
3026 case MPOL_INTERLEAVE:
3027 nodes = pol->v.nodes;
3031 snprintf(p, maxlen, "unknown");
3035 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3037 if (flags & MPOL_MODE_FLAGS) {
3038 p += snprintf(p, buffer + maxlen - p, "=");
3041 * Currently, the only defined flags are mutually exclusive
3043 if (flags & MPOL_F_STATIC_NODES)
3044 p += snprintf(p, buffer + maxlen - p, "static");
3045 else if (flags & MPOL_F_RELATIVE_NODES)
3046 p += snprintf(p, buffer + maxlen - p, "relative");
3049 if (!nodes_empty(nodes))
3050 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3051 nodemask_pr_args(&nodes));