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)
351 if (!pol || pol->mode == MPOL_LOCAL)
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;
413 struct vm_area_struct *prev;
417 * Check if the page's nid is in qp->nmask.
419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
420 * in the invert of qp->nmask.
422 static inline bool queue_pages_required(struct page *page,
423 struct queue_pages *qp)
425 int nid = page_to_nid(page);
426 unsigned long flags = qp->flags;
428 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
432 * queue_pages_pmd() has four possible return values:
433 * 0 - pages are placed on the right node or queued successfully.
434 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
437 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
438 * existing page was already on a node that does not follow the
441 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
442 unsigned long end, struct mm_walk *walk)
446 struct queue_pages *qp = walk->private;
449 if (unlikely(is_pmd_migration_entry(*pmd))) {
453 page = pmd_page(*pmd);
454 if (is_huge_zero_page(page)) {
456 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
460 if (!queue_pages_required(page, qp))
464 /* go to thp migration */
465 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
466 if (!vma_migratable(walk->vma) ||
467 migrate_page_add(page, qp->pagelist, flags)) {
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
483 * queue_pages_pte_range() has three possible return values:
484 * 0 - pages are placed on the right node or queued successfully.
485 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
487 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
488 * on a node that does not follow the policy.
490 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
491 unsigned long end, struct mm_walk *walk)
493 struct vm_area_struct *vma = walk->vma;
495 struct queue_pages *qp = walk->private;
496 unsigned long flags = qp->flags;
498 bool has_unmovable = false;
499 pte_t *pte, *mapped_pte;
502 ptl = pmd_trans_huge_lock(pmd, vma);
504 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
508 /* THP was split, fall through to pte walk */
510 if (pmd_trans_unstable(pmd))
513 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
514 for (; addr != end; pte++, addr += PAGE_SIZE) {
515 if (!pte_present(*pte))
517 page = vm_normal_page(vma, addr, *pte);
521 * vm_normal_page() filters out zero pages, but there might
522 * still be PageReserved pages to skip, perhaps in a VDSO.
524 if (PageReserved(page))
526 if (!queue_pages_required(page, qp))
528 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
529 /* MPOL_MF_STRICT must be specified if we get here */
530 if (!vma_migratable(vma)) {
531 has_unmovable = true;
536 * Do not abort immediately since there may be
537 * temporary off LRU pages in the range. Still
538 * need migrate other LRU pages.
540 if (migrate_page_add(page, qp->pagelist, flags))
541 has_unmovable = true;
545 pte_unmap_unlock(mapped_pte, ptl);
551 return addr != end ? -EIO : 0;
554 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
555 unsigned long addr, unsigned long end,
556 struct mm_walk *walk)
558 #ifdef CONFIG_HUGETLB_PAGE
559 struct queue_pages *qp = walk->private;
560 unsigned long flags = qp->flags;
565 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
566 entry = huge_ptep_get(pte);
567 if (!pte_present(entry))
569 page = pte_page(entry);
570 if (!queue_pages_required(page, qp))
572 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
573 if (flags & (MPOL_MF_MOVE_ALL) ||
574 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1 &&
575 !hugetlb_pmd_shared(pte)))
576 isolate_huge_page(page, qp->pagelist);
585 #ifdef CONFIG_NUMA_BALANCING
587 * This is used to mark a range of virtual addresses to be inaccessible.
588 * These are later cleared by a NUMA hinting fault. Depending on these
589 * faults, pages may be migrated for better NUMA placement.
591 * This is assuming that NUMA faults are handled using PROT_NONE. If
592 * an architecture makes a different choice, it will need further
593 * changes to the core.
595 unsigned long change_prot_numa(struct vm_area_struct *vma,
596 unsigned long addr, unsigned long end)
600 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
602 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
607 static unsigned long change_prot_numa(struct vm_area_struct *vma,
608 unsigned long addr, unsigned long end)
612 #endif /* CONFIG_NUMA_BALANCING */
614 static int queue_pages_test_walk(unsigned long start, unsigned long end,
615 struct mm_walk *walk)
617 struct vm_area_struct *vma = walk->vma;
618 struct queue_pages *qp = walk->private;
619 unsigned long endvma = vma->vm_end;
620 unsigned long flags = qp->flags;
623 * Need check MPOL_MF_STRICT to return -EIO if possible
624 * regardless of vma_migratable
626 if (!vma_migratable(vma) &&
627 !(flags & MPOL_MF_STRICT))
632 if (vma->vm_start > start)
633 start = vma->vm_start;
635 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
636 if (!vma->vm_next && vma->vm_end < end)
638 if (qp->prev && qp->prev->vm_end < vma->vm_start)
644 if (flags & MPOL_MF_LAZY) {
645 /* Similar to task_numa_work, skip inaccessible VMAs */
646 if (!is_vm_hugetlb_page(vma) &&
647 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
648 !(vma->vm_flags & VM_MIXEDMAP))
649 change_prot_numa(vma, start, endvma);
653 /* queue pages from current vma */
654 if (flags & MPOL_MF_VALID)
659 static const struct mm_walk_ops queue_pages_walk_ops = {
660 .hugetlb_entry = queue_pages_hugetlb,
661 .pmd_entry = queue_pages_pte_range,
662 .test_walk = queue_pages_test_walk,
666 * Walk through page tables and collect pages to be migrated.
668 * If pages found in a given range are on a set of nodes (determined by
669 * @nodes and @flags,) it's isolated and queued to the pagelist which is
670 * passed via @private.
672 * queue_pages_range() has three possible return values:
673 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
675 * 0 - queue pages successfully or no misplaced page.
676 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
677 * memory range specified by nodemask and maxnode points outside
678 * your accessible address space (-EFAULT)
681 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
682 nodemask_t *nodes, unsigned long flags,
683 struct list_head *pagelist)
685 struct queue_pages qp = {
686 .pagelist = pagelist,
692 return walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
696 * Apply policy to a single VMA
697 * This must be called with the mmap_sem held for writing.
699 static int vma_replace_policy(struct vm_area_struct *vma,
700 struct mempolicy *pol)
703 struct mempolicy *old;
704 struct mempolicy *new;
706 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
707 vma->vm_start, vma->vm_end, vma->vm_pgoff,
708 vma->vm_ops, vma->vm_file,
709 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
715 if (vma->vm_ops && vma->vm_ops->set_policy) {
716 err = vma->vm_ops->set_policy(vma, new);
721 old = vma->vm_policy;
722 vma->vm_policy = new; /* protected by mmap_sem */
731 /* Step 2: apply policy to a range and do splits. */
732 static int mbind_range(struct mm_struct *mm, unsigned long start,
733 unsigned long end, struct mempolicy *new_pol)
735 struct vm_area_struct *prev;
736 struct vm_area_struct *vma;
739 unsigned long vmstart;
742 vma = find_vma(mm, start);
743 if (!vma || vma->vm_start > start)
747 if (start > vma->vm_start)
750 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
751 vmstart = max(start, vma->vm_start);
752 vmend = min(end, vma->vm_end);
754 if (mpol_equal(vma_policy(vma), new_pol))
757 pgoff = vma->vm_pgoff +
758 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
759 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
760 vma->anon_vma, vma->vm_file, pgoff,
761 new_pol, vma->vm_userfaultfd_ctx);
766 if (vma->vm_start != vmstart) {
767 err = split_vma(vma->vm_mm, vma, vmstart, 1);
771 if (vma->vm_end != vmend) {
772 err = split_vma(vma->vm_mm, vma, vmend, 0);
777 err = vma_replace_policy(vma, new_pol);
786 /* Set the process memory policy */
787 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
790 struct mempolicy *new, *old;
791 NODEMASK_SCRATCH(scratch);
797 new = mpol_new(mode, flags, nodes);
804 ret = mpol_set_nodemask(new, nodes, scratch);
806 task_unlock(current);
810 old = current->mempolicy;
811 current->mempolicy = new;
812 if (new && new->mode == MPOL_INTERLEAVE)
813 current->il_prev = MAX_NUMNODES-1;
814 task_unlock(current);
818 NODEMASK_SCRATCH_FREE(scratch);
823 * Return nodemask for policy for get_mempolicy() query
825 * Called with task's alloc_lock held
827 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
830 if (p == &default_policy)
836 case MPOL_INTERLEAVE:
840 if (!(p->flags & MPOL_F_LOCAL))
841 node_set(p->v.preferred_node, *nodes);
842 /* else return empty node mask for local allocation */
849 static int lookup_node(struct mm_struct *mm, unsigned long addr)
855 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
857 err = page_to_nid(p);
861 up_read(&mm->mmap_sem);
865 /* Retrieve NUMA policy */
866 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
867 unsigned long addr, unsigned long flags)
870 struct mm_struct *mm = current->mm;
871 struct vm_area_struct *vma = NULL;
872 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
875 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
878 if (flags & MPOL_F_MEMS_ALLOWED) {
879 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
881 *policy = 0; /* just so it's initialized */
883 *nmask = cpuset_current_mems_allowed;
884 task_unlock(current);
888 if (flags & MPOL_F_ADDR) {
890 * Do NOT fall back to task policy if the
891 * vma/shared policy at addr is NULL. We
892 * want to return MPOL_DEFAULT in this case.
894 down_read(&mm->mmap_sem);
895 vma = find_vma_intersection(mm, addr, addr+1);
897 up_read(&mm->mmap_sem);
900 if (vma->vm_ops && vma->vm_ops->get_policy)
901 pol = vma->vm_ops->get_policy(vma, addr);
903 pol = vma->vm_policy;
908 pol = &default_policy; /* indicates default behavior */
910 if (flags & MPOL_F_NODE) {
911 if (flags & MPOL_F_ADDR) {
913 * Take a refcount on the mpol, lookup_node()
914 * wil drop the mmap_sem, so after calling
915 * lookup_node() only "pol" remains valid, "vma"
921 err = lookup_node(mm, addr);
925 } else if (pol == current->mempolicy &&
926 pol->mode == MPOL_INTERLEAVE) {
927 *policy = next_node_in(current->il_prev, pol->v.nodes);
933 *policy = pol == &default_policy ? MPOL_DEFAULT :
936 * Internal mempolicy flags must be masked off before exposing
937 * the policy to userspace.
939 *policy |= (pol->flags & MPOL_MODE_FLAGS);
944 if (mpol_store_user_nodemask(pol)) {
945 *nmask = pol->w.user_nodemask;
948 get_policy_nodemask(pol, nmask);
949 task_unlock(current);
956 up_read(&mm->mmap_sem);
958 mpol_put(pol_refcount);
962 #ifdef CONFIG_MIGRATION
964 * page migration, thp tail pages can be passed.
966 static int migrate_page_add(struct page *page, struct list_head *pagelist,
969 struct page *head = compound_head(page);
971 * Avoid migrating a page that is shared with others.
973 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
974 if (!isolate_lru_page(head)) {
975 list_add_tail(&head->lru, pagelist);
976 mod_node_page_state(page_pgdat(head),
977 NR_ISOLATED_ANON + page_is_file_cache(head),
978 hpage_nr_pages(head));
979 } else if (flags & MPOL_MF_STRICT) {
981 * Non-movable page may reach here. And, there may be
982 * temporary off LRU pages or non-LRU movable pages.
983 * Treat them as unmovable pages since they can't be
984 * isolated, so they can't be moved at the moment. It
985 * should return -EIO for this case too.
994 /* page allocation callback for NUMA node migration */
995 struct page *alloc_new_node_page(struct page *page, unsigned long node)
998 return alloc_huge_page_node(page_hstate(compound_head(page)),
1000 else if (PageTransHuge(page)) {
1003 thp = alloc_pages_node(node,
1004 (GFP_TRANSHUGE | __GFP_THISNODE),
1008 prep_transhuge_page(thp);
1011 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1016 * Migrate pages from one node to a target node.
1017 * Returns error or the number of pages not migrated.
1019 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1023 LIST_HEAD(pagelist);
1027 node_set(source, nmask);
1030 * This does not "check" the range but isolates all pages that
1031 * need migration. Between passing in the full user address
1032 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1034 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1035 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1036 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1038 if (!list_empty(&pagelist)) {
1039 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1040 MIGRATE_SYNC, MR_SYSCALL);
1042 putback_movable_pages(&pagelist);
1049 * Move pages between the two nodesets so as to preserve the physical
1050 * layout as much as possible.
1052 * Returns the number of page that could not be moved.
1054 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1055 const nodemask_t *to, int flags)
1061 err = migrate_prep();
1065 down_read(&mm->mmap_sem);
1068 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1069 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1070 * bit in 'tmp', and return that <source, dest> pair for migration.
1071 * The pair of nodemasks 'to' and 'from' define the map.
1073 * If no pair of bits is found that way, fallback to picking some
1074 * pair of 'source' and 'dest' bits that are not the same. If the
1075 * 'source' and 'dest' bits are the same, this represents a node
1076 * that will be migrating to itself, so no pages need move.
1078 * If no bits are left in 'tmp', or if all remaining bits left
1079 * in 'tmp' correspond to the same bit in 'to', return false
1080 * (nothing left to migrate).
1082 * This lets us pick a pair of nodes to migrate between, such that
1083 * if possible the dest node is not already occupied by some other
1084 * source node, minimizing the risk of overloading the memory on a
1085 * node that would happen if we migrated incoming memory to a node
1086 * before migrating outgoing memory source that same node.
1088 * A single scan of tmp is sufficient. As we go, we remember the
1089 * most recent <s, d> pair that moved (s != d). If we find a pair
1090 * that not only moved, but what's better, moved to an empty slot
1091 * (d is not set in tmp), then we break out then, with that pair.
1092 * Otherwise when we finish scanning from_tmp, we at least have the
1093 * most recent <s, d> pair that moved. If we get all the way through
1094 * the scan of tmp without finding any node that moved, much less
1095 * moved to an empty node, then there is nothing left worth migrating.
1099 while (!nodes_empty(tmp)) {
1101 int source = NUMA_NO_NODE;
1104 for_each_node_mask(s, tmp) {
1107 * do_migrate_pages() tries to maintain the relative
1108 * node relationship of the pages established between
1109 * threads and memory areas.
1111 * However if the number of source nodes is not equal to
1112 * the number of destination nodes we can not preserve
1113 * this node relative relationship. In that case, skip
1114 * copying memory from a node that is in the destination
1117 * Example: [2,3,4] -> [3,4,5] moves everything.
1118 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1121 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1122 (node_isset(s, *to)))
1125 d = node_remap(s, *from, *to);
1129 source = s; /* Node moved. Memorize */
1132 /* dest not in remaining from nodes? */
1133 if (!node_isset(dest, tmp))
1136 if (source == NUMA_NO_NODE)
1139 node_clear(source, tmp);
1140 err = migrate_to_node(mm, source, dest, flags);
1146 up_read(&mm->mmap_sem);
1154 * Allocate a new page for page migration based on vma policy.
1155 * Start by assuming the page is mapped by the same vma as contains @start.
1156 * Search forward from there, if not. N.B., this assumes that the
1157 * list of pages handed to migrate_pages()--which is how we get here--
1158 * is in virtual address order.
1160 static struct page *new_page(struct page *page, unsigned long start)
1162 struct vm_area_struct *vma;
1163 unsigned long address;
1165 vma = find_vma(current->mm, start);
1167 address = page_address_in_vma(page, vma);
1168 if (address != -EFAULT)
1173 if (PageHuge(page)) {
1174 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1176 } else if (PageTransHuge(page)) {
1179 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1183 prep_transhuge_page(thp);
1187 * if !vma, alloc_page_vma() will use task or system default policy
1189 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1194 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1195 unsigned long flags)
1200 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1201 const nodemask_t *to, int flags)
1206 static struct page *new_page(struct page *page, unsigned long start)
1212 static long do_mbind(unsigned long start, unsigned long len,
1213 unsigned short mode, unsigned short mode_flags,
1214 nodemask_t *nmask, unsigned long flags)
1216 struct mm_struct *mm = current->mm;
1217 struct mempolicy *new;
1221 LIST_HEAD(pagelist);
1223 if (flags & ~(unsigned long)MPOL_MF_VALID)
1225 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1228 if (start & ~PAGE_MASK)
1231 if (mode == MPOL_DEFAULT)
1232 flags &= ~MPOL_MF_STRICT;
1234 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1242 new = mpol_new(mode, mode_flags, nmask);
1244 return PTR_ERR(new);
1246 if (flags & MPOL_MF_LAZY)
1247 new->flags |= MPOL_F_MOF;
1250 * If we are using the default policy then operation
1251 * on discontinuous address spaces is okay after all
1254 flags |= MPOL_MF_DISCONTIG_OK;
1256 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1257 start, start + len, mode, mode_flags,
1258 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1260 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1262 err = migrate_prep();
1267 NODEMASK_SCRATCH(scratch);
1269 down_write(&mm->mmap_sem);
1271 err = mpol_set_nodemask(new, nmask, scratch);
1272 task_unlock(current);
1274 up_write(&mm->mmap_sem);
1277 NODEMASK_SCRATCH_FREE(scratch);
1282 ret = queue_pages_range(mm, start, end, nmask,
1283 flags | MPOL_MF_INVERT, &pagelist);
1290 err = mbind_range(mm, start, end, new);
1295 if (!list_empty(&pagelist)) {
1296 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1297 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1298 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1300 putback_movable_pages(&pagelist);
1303 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1307 if (!list_empty(&pagelist))
1308 putback_movable_pages(&pagelist);
1311 up_write(&mm->mmap_sem);
1318 * User space interface with variable sized bitmaps for nodelists.
1321 /* Copy a node mask from user space. */
1322 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1323 unsigned long maxnode)
1327 unsigned long nlongs;
1328 unsigned long endmask;
1331 nodes_clear(*nodes);
1332 if (maxnode == 0 || !nmask)
1334 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1337 nlongs = BITS_TO_LONGS(maxnode);
1338 if ((maxnode % BITS_PER_LONG) == 0)
1341 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1344 * When the user specified more nodes than supported just check
1345 * if the non supported part is all zero.
1347 * If maxnode have more longs than MAX_NUMNODES, check
1348 * the bits in that area first. And then go through to
1349 * check the rest bits which equal or bigger than MAX_NUMNODES.
1350 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1352 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1353 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1354 if (get_user(t, nmask + k))
1356 if (k == nlongs - 1) {
1362 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1366 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1367 unsigned long valid_mask = endmask;
1369 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1370 if (get_user(t, nmask + nlongs - 1))
1376 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1378 nodes_addr(*nodes)[nlongs-1] &= endmask;
1382 /* Copy a kernel node mask to user space */
1383 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1386 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1387 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1389 if (copy > nbytes) {
1390 if (copy > PAGE_SIZE)
1392 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1396 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1399 static long kernel_mbind(unsigned long start, unsigned long len,
1400 unsigned long mode, const unsigned long __user *nmask,
1401 unsigned long maxnode, unsigned int flags)
1405 unsigned short mode_flags;
1407 start = untagged_addr(start);
1408 mode_flags = mode & MPOL_MODE_FLAGS;
1409 mode &= ~MPOL_MODE_FLAGS;
1410 if (mode >= MPOL_MAX)
1412 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1413 (mode_flags & MPOL_F_RELATIVE_NODES))
1415 err = get_nodes(&nodes, nmask, maxnode);
1418 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1421 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1422 unsigned long, mode, const unsigned long __user *, nmask,
1423 unsigned long, maxnode, unsigned int, flags)
1425 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1428 /* Set the process memory policy */
1429 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1430 unsigned long maxnode)
1434 unsigned short flags;
1436 flags = mode & MPOL_MODE_FLAGS;
1437 mode &= ~MPOL_MODE_FLAGS;
1438 if ((unsigned int)mode >= MPOL_MAX)
1440 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1442 err = get_nodes(&nodes, nmask, maxnode);
1445 return do_set_mempolicy(mode, flags, &nodes);
1448 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1449 unsigned long, maxnode)
1451 return kernel_set_mempolicy(mode, nmask, maxnode);
1454 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1455 const unsigned long __user *old_nodes,
1456 const unsigned long __user *new_nodes)
1458 struct mm_struct *mm = NULL;
1459 struct task_struct *task;
1460 nodemask_t task_nodes;
1464 NODEMASK_SCRATCH(scratch);
1469 old = &scratch->mask1;
1470 new = &scratch->mask2;
1472 err = get_nodes(old, old_nodes, maxnode);
1476 err = get_nodes(new, new_nodes, maxnode);
1480 /* Find the mm_struct */
1482 task = pid ? find_task_by_vpid(pid) : current;
1488 get_task_struct(task);
1493 * Check if this process has the right to modify the specified process.
1494 * Use the regular "ptrace_may_access()" checks.
1496 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1503 task_nodes = cpuset_mems_allowed(task);
1504 /* Is the user allowed to access the target nodes? */
1505 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1510 task_nodes = cpuset_mems_allowed(current);
1511 nodes_and(*new, *new, task_nodes);
1512 if (nodes_empty(*new))
1515 err = security_task_movememory(task);
1519 mm = get_task_mm(task);
1520 put_task_struct(task);
1527 err = do_migrate_pages(mm, old, new,
1528 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1532 NODEMASK_SCRATCH_FREE(scratch);
1537 put_task_struct(task);
1542 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1543 const unsigned long __user *, old_nodes,
1544 const unsigned long __user *, new_nodes)
1546 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1550 /* Retrieve NUMA policy */
1551 static int kernel_get_mempolicy(int __user *policy,
1552 unsigned long __user *nmask,
1553 unsigned long maxnode,
1555 unsigned long flags)
1561 addr = untagged_addr(addr);
1563 if (nmask != NULL && maxnode < nr_node_ids)
1566 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1571 if (policy && put_user(pval, policy))
1575 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1580 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1581 unsigned long __user *, nmask, unsigned long, maxnode,
1582 unsigned long, addr, unsigned long, flags)
1584 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1587 #ifdef CONFIG_COMPAT
1589 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1590 compat_ulong_t __user *, nmask,
1591 compat_ulong_t, maxnode,
1592 compat_ulong_t, addr, compat_ulong_t, flags)
1595 unsigned long __user *nm = NULL;
1596 unsigned long nr_bits, alloc_size;
1597 DECLARE_BITMAP(bm, MAX_NUMNODES);
1599 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1600 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1603 nm = compat_alloc_user_space(alloc_size);
1605 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1607 if (!err && nmask) {
1608 unsigned long copy_size;
1609 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1610 err = copy_from_user(bm, nm, copy_size);
1611 /* ensure entire bitmap is zeroed */
1612 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1613 err |= compat_put_bitmap(nmask, bm, nr_bits);
1619 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1620 compat_ulong_t, maxnode)
1622 unsigned long __user *nm = NULL;
1623 unsigned long nr_bits, alloc_size;
1624 DECLARE_BITMAP(bm, MAX_NUMNODES);
1626 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1627 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1630 if (compat_get_bitmap(bm, nmask, nr_bits))
1632 nm = compat_alloc_user_space(alloc_size);
1633 if (copy_to_user(nm, bm, alloc_size))
1637 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1640 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1641 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1642 compat_ulong_t, maxnode, compat_ulong_t, flags)
1644 unsigned long __user *nm = NULL;
1645 unsigned long nr_bits, alloc_size;
1648 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1649 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1652 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1654 nm = compat_alloc_user_space(alloc_size);
1655 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1659 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1662 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1663 compat_ulong_t, maxnode,
1664 const compat_ulong_t __user *, old_nodes,
1665 const compat_ulong_t __user *, new_nodes)
1667 unsigned long __user *old = NULL;
1668 unsigned long __user *new = NULL;
1669 nodemask_t tmp_mask;
1670 unsigned long nr_bits;
1673 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1674 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1676 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1678 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1680 new = old + size / sizeof(unsigned long);
1681 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1685 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1688 new = compat_alloc_user_space(size);
1689 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1692 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1695 #endif /* CONFIG_COMPAT */
1697 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1700 struct mempolicy *pol = NULL;
1703 if (vma->vm_ops && vma->vm_ops->get_policy) {
1704 pol = vma->vm_ops->get_policy(vma, addr);
1705 } else if (vma->vm_policy) {
1706 pol = vma->vm_policy;
1709 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1710 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1711 * count on these policies which will be dropped by
1712 * mpol_cond_put() later
1714 if (mpol_needs_cond_ref(pol))
1723 * get_vma_policy(@vma, @addr)
1724 * @vma: virtual memory area whose policy is sought
1725 * @addr: address in @vma for shared policy lookup
1727 * Returns effective policy for a VMA at specified address.
1728 * Falls back to current->mempolicy or system default policy, as necessary.
1729 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1730 * count--added by the get_policy() vm_op, as appropriate--to protect against
1731 * freeing by another task. It is the caller's responsibility to free the
1732 * extra reference for shared policies.
1734 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1737 struct mempolicy *pol = __get_vma_policy(vma, addr);
1740 pol = get_task_policy(current);
1745 bool vma_policy_mof(struct vm_area_struct *vma)
1747 struct mempolicy *pol;
1749 if (vma->vm_ops && vma->vm_ops->get_policy) {
1752 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1753 if (pol && (pol->flags & MPOL_F_MOF))
1760 pol = vma->vm_policy;
1762 pol = get_task_policy(current);
1764 return pol->flags & MPOL_F_MOF;
1767 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1769 enum zone_type dynamic_policy_zone = policy_zone;
1771 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1774 * if policy->v.nodes has movable memory only,
1775 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1777 * policy->v.nodes is intersect with node_states[N_MEMORY].
1778 * so if the following test faile, it implies
1779 * policy->v.nodes has movable memory only.
1781 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1782 dynamic_policy_zone = ZONE_MOVABLE;
1784 return zone >= dynamic_policy_zone;
1788 * Return a nodemask representing a mempolicy for filtering nodes for
1791 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1793 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1794 if (unlikely(policy->mode == MPOL_BIND) &&
1795 apply_policy_zone(policy, gfp_zone(gfp)) &&
1796 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1797 return &policy->v.nodes;
1802 /* Return the node id preferred by the given mempolicy, or the given id */
1803 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1806 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1807 nd = policy->v.preferred_node;
1810 * __GFP_THISNODE shouldn't even be used with the bind policy
1811 * because we might easily break the expectation to stay on the
1812 * requested node and not break the policy.
1814 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1820 /* Do dynamic interleaving for a process */
1821 static unsigned interleave_nodes(struct mempolicy *policy)
1824 struct task_struct *me = current;
1826 next = next_node_in(me->il_prev, policy->v.nodes);
1827 if (next < MAX_NUMNODES)
1833 * Depending on the memory policy provide a node from which to allocate the
1836 unsigned int mempolicy_slab_node(void)
1838 struct mempolicy *policy;
1839 int node = numa_mem_id();
1844 policy = current->mempolicy;
1845 if (!policy || policy->flags & MPOL_F_LOCAL)
1848 switch (policy->mode) {
1849 case MPOL_PREFERRED:
1851 * handled MPOL_F_LOCAL above
1853 return policy->v.preferred_node;
1855 case MPOL_INTERLEAVE:
1856 return interleave_nodes(policy);
1862 * Follow bind policy behavior and start allocation at the
1865 struct zonelist *zonelist;
1866 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1867 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1868 z = first_zones_zonelist(zonelist, highest_zoneidx,
1870 return z->zone ? zone_to_nid(z->zone) : node;
1879 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1880 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1881 * number of present nodes.
1883 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1885 unsigned nnodes = nodes_weight(pol->v.nodes);
1891 return numa_node_id();
1892 target = (unsigned int)n % nnodes;
1893 nid = first_node(pol->v.nodes);
1894 for (i = 0; i < target; i++)
1895 nid = next_node(nid, pol->v.nodes);
1899 /* Determine a node number for interleave */
1900 static inline unsigned interleave_nid(struct mempolicy *pol,
1901 struct vm_area_struct *vma, unsigned long addr, int shift)
1907 * for small pages, there is no difference between
1908 * shift and PAGE_SHIFT, so the bit-shift is safe.
1909 * for huge pages, since vm_pgoff is in units of small
1910 * pages, we need to shift off the always 0 bits to get
1913 BUG_ON(shift < PAGE_SHIFT);
1914 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1915 off += (addr - vma->vm_start) >> shift;
1916 return offset_il_node(pol, off);
1918 return interleave_nodes(pol);
1921 #ifdef CONFIG_HUGETLBFS
1923 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1924 * @vma: virtual memory area whose policy is sought
1925 * @addr: address in @vma for shared policy lookup and interleave policy
1926 * @gfp_flags: for requested zone
1927 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1928 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1930 * Returns a nid suitable for a huge page allocation and a pointer
1931 * to the struct mempolicy for conditional unref after allocation.
1932 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1933 * @nodemask for filtering the zonelist.
1935 * Must be protected by read_mems_allowed_begin()
1937 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1938 struct mempolicy **mpol, nodemask_t **nodemask)
1942 *mpol = get_vma_policy(vma, addr);
1943 *nodemask = NULL; /* assume !MPOL_BIND */
1945 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1946 nid = interleave_nid(*mpol, vma, addr,
1947 huge_page_shift(hstate_vma(vma)));
1949 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1950 if ((*mpol)->mode == MPOL_BIND)
1951 *nodemask = &(*mpol)->v.nodes;
1957 * init_nodemask_of_mempolicy
1959 * If the current task's mempolicy is "default" [NULL], return 'false'
1960 * to indicate default policy. Otherwise, extract the policy nodemask
1961 * for 'bind' or 'interleave' policy into the argument nodemask, or
1962 * initialize the argument nodemask to contain the single node for
1963 * 'preferred' or 'local' policy and return 'true' to indicate presence
1964 * of non-default mempolicy.
1966 * We don't bother with reference counting the mempolicy [mpol_get/put]
1967 * because the current task is examining it's own mempolicy and a task's
1968 * mempolicy is only ever changed by the task itself.
1970 * N.B., it is the caller's responsibility to free a returned nodemask.
1972 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1974 struct mempolicy *mempolicy;
1977 if (!(mask && current->mempolicy))
1981 mempolicy = current->mempolicy;
1982 switch (mempolicy->mode) {
1983 case MPOL_PREFERRED:
1984 if (mempolicy->flags & MPOL_F_LOCAL)
1985 nid = numa_node_id();
1987 nid = mempolicy->v.preferred_node;
1988 init_nodemask_of_node(mask, nid);
1993 case MPOL_INTERLEAVE:
1994 *mask = mempolicy->v.nodes;
2000 task_unlock(current);
2007 * mempolicy_nodemask_intersects
2009 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2010 * policy. Otherwise, check for intersection between mask and the policy
2011 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2012 * policy, always return true since it may allocate elsewhere on fallback.
2014 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2016 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2017 const nodemask_t *mask)
2019 struct mempolicy *mempolicy;
2025 mempolicy = tsk->mempolicy;
2029 switch (mempolicy->mode) {
2030 case MPOL_PREFERRED:
2032 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2033 * allocate from, they may fallback to other nodes when oom.
2034 * Thus, it's possible for tsk to have allocated memory from
2039 case MPOL_INTERLEAVE:
2040 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2050 /* Allocate a page in interleaved policy.
2051 Own path because it needs to do special accounting. */
2052 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2057 page = __alloc_pages(gfp, order, nid);
2058 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2059 if (!static_branch_likely(&vm_numa_stat_key))
2061 if (page && page_to_nid(page) == nid) {
2063 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2070 * alloc_pages_vma - Allocate a page for a VMA.
2073 * %GFP_USER user allocation.
2074 * %GFP_KERNEL kernel allocations,
2075 * %GFP_HIGHMEM highmem/user allocations,
2076 * %GFP_FS allocation should not call back into a file system.
2077 * %GFP_ATOMIC don't sleep.
2079 * @order:Order of the GFP allocation.
2080 * @vma: Pointer to VMA or NULL if not available.
2081 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2082 * @node: Which node to prefer for allocation (modulo policy).
2083 * @hugepage: for hugepages try only the preferred node if possible
2085 * This function allocates a page from the kernel page pool and applies
2086 * a NUMA policy associated with the VMA or the current process.
2087 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2088 * mm_struct of the VMA to prevent it from going away. Should be used for
2089 * all allocations for pages that will be mapped into user space. Returns
2090 * NULL when no page can be allocated.
2093 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2094 unsigned long addr, int node, bool hugepage)
2096 struct mempolicy *pol;
2101 pol = get_vma_policy(vma, addr);
2103 if (pol->mode == MPOL_INTERLEAVE) {
2106 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2108 page = alloc_page_interleave(gfp, order, nid);
2112 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2113 int hpage_node = node;
2116 * For hugepage allocation and non-interleave policy which
2117 * allows the current node (or other explicitly preferred
2118 * node) we only try to allocate from the current/preferred
2119 * node and don't fall back to other nodes, as the cost of
2120 * remote accesses would likely offset THP benefits.
2122 * If the policy is interleave, or does not allow the current
2123 * node in its nodemask, we allocate the standard way.
2125 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2126 hpage_node = pol->v.preferred_node;
2128 nmask = policy_nodemask(gfp, pol);
2129 if (!nmask || node_isset(hpage_node, *nmask)) {
2131 page = __alloc_pages_node(hpage_node,
2132 gfp | __GFP_THISNODE, order);
2135 * If hugepage allocations are configured to always
2136 * synchronous compact or the vma has been madvised
2137 * to prefer hugepage backing, retry allowing remote
2140 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2141 page = __alloc_pages_nodemask(gfp | __GFP_NORETRY,
2149 nmask = policy_nodemask(gfp, pol);
2150 preferred_nid = policy_node(gfp, pol, node);
2151 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2156 EXPORT_SYMBOL(alloc_pages_vma);
2159 * alloc_pages_current - Allocate pages.
2162 * %GFP_USER user allocation,
2163 * %GFP_KERNEL kernel allocation,
2164 * %GFP_HIGHMEM highmem allocation,
2165 * %GFP_FS don't call back into a file system.
2166 * %GFP_ATOMIC don't sleep.
2167 * @order: Power of two of allocation size in pages. 0 is a single page.
2169 * Allocate a page from the kernel page pool. When not in
2170 * interrupt context and apply the current process NUMA policy.
2171 * Returns NULL when no page can be allocated.
2173 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2175 struct mempolicy *pol = &default_policy;
2178 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2179 pol = get_task_policy(current);
2182 * No reference counting needed for current->mempolicy
2183 * nor system default_policy
2185 if (pol->mode == MPOL_INTERLEAVE)
2186 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2188 page = __alloc_pages_nodemask(gfp, order,
2189 policy_node(gfp, pol, numa_node_id()),
2190 policy_nodemask(gfp, pol));
2194 EXPORT_SYMBOL(alloc_pages_current);
2196 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2198 struct mempolicy *pol = mpol_dup(vma_policy(src));
2201 return PTR_ERR(pol);
2202 dst->vm_policy = pol;
2207 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2208 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2209 * with the mems_allowed returned by cpuset_mems_allowed(). This
2210 * keeps mempolicies cpuset relative after its cpuset moves. See
2211 * further kernel/cpuset.c update_nodemask().
2213 * current's mempolicy may be rebinded by the other task(the task that changes
2214 * cpuset's mems), so we needn't do rebind work for current task.
2217 /* Slow path of a mempolicy duplicate */
2218 struct mempolicy *__mpol_dup(struct mempolicy *old)
2220 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2223 return ERR_PTR(-ENOMEM);
2225 /* task's mempolicy is protected by alloc_lock */
2226 if (old == current->mempolicy) {
2229 task_unlock(current);
2233 if (current_cpuset_is_being_rebound()) {
2234 nodemask_t mems = cpuset_mems_allowed(current);
2235 mpol_rebind_policy(new, &mems);
2237 atomic_set(&new->refcnt, 1);
2241 /* Slow path of a mempolicy comparison */
2242 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2246 if (a->mode != b->mode)
2248 if (a->flags != b->flags)
2250 if (mpol_store_user_nodemask(a))
2251 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2257 case MPOL_INTERLEAVE:
2258 return !!nodes_equal(a->v.nodes, b->v.nodes);
2259 case MPOL_PREFERRED:
2260 /* a's ->flags is the same as b's */
2261 if (a->flags & MPOL_F_LOCAL)
2263 return a->v.preferred_node == b->v.preferred_node;
2271 * Shared memory backing store policy support.
2273 * Remember policies even when nobody has shared memory mapped.
2274 * The policies are kept in Red-Black tree linked from the inode.
2275 * They are protected by the sp->lock rwlock, which should be held
2276 * for any accesses to the tree.
2280 * lookup first element intersecting start-end. Caller holds sp->lock for
2281 * reading or for writing
2283 static struct sp_node *
2284 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2286 struct rb_node *n = sp->root.rb_node;
2289 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2291 if (start >= p->end)
2293 else if (end <= p->start)
2301 struct sp_node *w = NULL;
2302 struct rb_node *prev = rb_prev(n);
2305 w = rb_entry(prev, struct sp_node, nd);
2306 if (w->end <= start)
2310 return rb_entry(n, struct sp_node, nd);
2314 * Insert a new shared policy into the list. Caller holds sp->lock for
2317 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2319 struct rb_node **p = &sp->root.rb_node;
2320 struct rb_node *parent = NULL;
2325 nd = rb_entry(parent, struct sp_node, nd);
2326 if (new->start < nd->start)
2328 else if (new->end > nd->end)
2329 p = &(*p)->rb_right;
2333 rb_link_node(&new->nd, parent, p);
2334 rb_insert_color(&new->nd, &sp->root);
2335 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2336 new->policy ? new->policy->mode : 0);
2339 /* Find shared policy intersecting idx */
2341 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2343 struct mempolicy *pol = NULL;
2346 if (!sp->root.rb_node)
2348 read_lock(&sp->lock);
2349 sn = sp_lookup(sp, idx, idx+1);
2351 mpol_get(sn->policy);
2354 read_unlock(&sp->lock);
2358 static void sp_free(struct sp_node *n)
2360 mpol_put(n->policy);
2361 kmem_cache_free(sn_cache, n);
2365 * mpol_misplaced - check whether current page node is valid in policy
2367 * @page: page to be checked
2368 * @vma: vm area where page mapped
2369 * @addr: virtual address where page mapped
2371 * Lookup current policy node id for vma,addr and "compare to" page's
2375 * -1 - not misplaced, page is in the right node
2376 * node - node id where the page should be
2378 * Policy determination "mimics" alloc_page_vma().
2379 * Called from fault path where we know the vma and faulting address.
2381 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2383 struct mempolicy *pol;
2385 int curnid = page_to_nid(page);
2386 unsigned long pgoff;
2387 int thiscpu = raw_smp_processor_id();
2388 int thisnid = cpu_to_node(thiscpu);
2389 int polnid = NUMA_NO_NODE;
2392 pol = get_vma_policy(vma, addr);
2393 if (!(pol->flags & MPOL_F_MOF))
2396 switch (pol->mode) {
2397 case MPOL_INTERLEAVE:
2398 pgoff = vma->vm_pgoff;
2399 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2400 polnid = offset_il_node(pol, pgoff);
2403 case MPOL_PREFERRED:
2404 if (pol->flags & MPOL_F_LOCAL)
2405 polnid = numa_node_id();
2407 polnid = pol->v.preferred_node;
2413 * allows binding to multiple nodes.
2414 * use current page if in policy nodemask,
2415 * else select nearest allowed node, if any.
2416 * If no allowed nodes, use current [!misplaced].
2418 if (node_isset(curnid, pol->v.nodes))
2420 z = first_zones_zonelist(
2421 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2422 gfp_zone(GFP_HIGHUSER),
2424 polnid = zone_to_nid(z->zone);
2431 /* Migrate the page towards the node whose CPU is referencing it */
2432 if (pol->flags & MPOL_F_MORON) {
2435 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2439 if (curnid != polnid)
2448 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2449 * dropped after task->mempolicy is set to NULL so that any allocation done as
2450 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2453 void mpol_put_task_policy(struct task_struct *task)
2455 struct mempolicy *pol;
2458 pol = task->mempolicy;
2459 task->mempolicy = NULL;
2464 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2466 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2467 rb_erase(&n->nd, &sp->root);
2471 static void sp_node_init(struct sp_node *node, unsigned long start,
2472 unsigned long end, struct mempolicy *pol)
2474 node->start = start;
2479 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2480 struct mempolicy *pol)
2483 struct mempolicy *newpol;
2485 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2489 newpol = mpol_dup(pol);
2490 if (IS_ERR(newpol)) {
2491 kmem_cache_free(sn_cache, n);
2494 newpol->flags |= MPOL_F_SHARED;
2495 sp_node_init(n, start, end, newpol);
2500 /* Replace a policy range. */
2501 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2502 unsigned long end, struct sp_node *new)
2505 struct sp_node *n_new = NULL;
2506 struct mempolicy *mpol_new = NULL;
2510 write_lock(&sp->lock);
2511 n = sp_lookup(sp, start, end);
2512 /* Take care of old policies in the same range. */
2513 while (n && n->start < end) {
2514 struct rb_node *next = rb_next(&n->nd);
2515 if (n->start >= start) {
2521 /* Old policy spanning whole new range. */
2526 *mpol_new = *n->policy;
2527 atomic_set(&mpol_new->refcnt, 1);
2528 sp_node_init(n_new, end, n->end, mpol_new);
2530 sp_insert(sp, n_new);
2539 n = rb_entry(next, struct sp_node, nd);
2543 write_unlock(&sp->lock);
2550 kmem_cache_free(sn_cache, n_new);
2555 write_unlock(&sp->lock);
2557 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2560 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2563 atomic_set(&mpol_new->refcnt, 1);
2568 * mpol_shared_policy_init - initialize shared policy for inode
2569 * @sp: pointer to inode shared policy
2570 * @mpol: struct mempolicy to install
2572 * Install non-NULL @mpol in inode's shared policy rb-tree.
2573 * On entry, the current task has a reference on a non-NULL @mpol.
2574 * This must be released on exit.
2575 * This is called at get_inode() calls and we can use GFP_KERNEL.
2577 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2581 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2582 rwlock_init(&sp->lock);
2585 struct vm_area_struct pvma;
2586 struct mempolicy *new;
2587 NODEMASK_SCRATCH(scratch);
2591 /* contextualize the tmpfs mount point mempolicy */
2592 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2594 goto free_scratch; /* no valid nodemask intersection */
2597 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2598 task_unlock(current);
2602 /* Create pseudo-vma that contains just the policy */
2603 vma_init(&pvma, NULL);
2604 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2605 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2608 mpol_put(new); /* drop initial ref */
2610 NODEMASK_SCRATCH_FREE(scratch);
2612 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2616 int mpol_set_shared_policy(struct shared_policy *info,
2617 struct vm_area_struct *vma, struct mempolicy *npol)
2620 struct sp_node *new = NULL;
2621 unsigned long sz = vma_pages(vma);
2623 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2625 sz, npol ? npol->mode : -1,
2626 npol ? npol->flags : -1,
2627 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2630 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2634 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2640 /* Free a backing policy store on inode delete. */
2641 void mpol_free_shared_policy(struct shared_policy *p)
2644 struct rb_node *next;
2646 if (!p->root.rb_node)
2648 write_lock(&p->lock);
2649 next = rb_first(&p->root);
2651 n = rb_entry(next, struct sp_node, nd);
2652 next = rb_next(&n->nd);
2655 write_unlock(&p->lock);
2658 #ifdef CONFIG_NUMA_BALANCING
2659 static int __initdata numabalancing_override;
2661 static void __init check_numabalancing_enable(void)
2663 bool numabalancing_default = false;
2665 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2666 numabalancing_default = true;
2668 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2669 if (numabalancing_override)
2670 set_numabalancing_state(numabalancing_override == 1);
2672 if (num_online_nodes() > 1 && !numabalancing_override) {
2673 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2674 numabalancing_default ? "Enabling" : "Disabling");
2675 set_numabalancing_state(numabalancing_default);
2679 static int __init setup_numabalancing(char *str)
2685 if (!strcmp(str, "enable")) {
2686 numabalancing_override = 1;
2688 } else if (!strcmp(str, "disable")) {
2689 numabalancing_override = -1;
2694 pr_warn("Unable to parse numa_balancing=\n");
2698 __setup("numa_balancing=", setup_numabalancing);
2700 static inline void __init check_numabalancing_enable(void)
2703 #endif /* CONFIG_NUMA_BALANCING */
2705 /* assumes fs == KERNEL_DS */
2706 void __init numa_policy_init(void)
2708 nodemask_t interleave_nodes;
2709 unsigned long largest = 0;
2710 int nid, prefer = 0;
2712 policy_cache = kmem_cache_create("numa_policy",
2713 sizeof(struct mempolicy),
2714 0, SLAB_PANIC, NULL);
2716 sn_cache = kmem_cache_create("shared_policy_node",
2717 sizeof(struct sp_node),
2718 0, SLAB_PANIC, NULL);
2720 for_each_node(nid) {
2721 preferred_node_policy[nid] = (struct mempolicy) {
2722 .refcnt = ATOMIC_INIT(1),
2723 .mode = MPOL_PREFERRED,
2724 .flags = MPOL_F_MOF | MPOL_F_MORON,
2725 .v = { .preferred_node = nid, },
2730 * Set interleaving policy for system init. Interleaving is only
2731 * enabled across suitably sized nodes (default is >= 16MB), or
2732 * fall back to the largest node if they're all smaller.
2734 nodes_clear(interleave_nodes);
2735 for_each_node_state(nid, N_MEMORY) {
2736 unsigned long total_pages = node_present_pages(nid);
2738 /* Preserve the largest node */
2739 if (largest < total_pages) {
2740 largest = total_pages;
2744 /* Interleave this node? */
2745 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2746 node_set(nid, interleave_nodes);
2749 /* All too small, use the largest */
2750 if (unlikely(nodes_empty(interleave_nodes)))
2751 node_set(prefer, interleave_nodes);
2753 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2754 pr_err("%s: interleaving failed\n", __func__);
2756 check_numabalancing_enable();
2759 /* Reset policy of current process to default */
2760 void numa_default_policy(void)
2762 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2766 * Parse and format mempolicy from/to strings
2770 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2772 static const char * const policy_modes[] =
2774 [MPOL_DEFAULT] = "default",
2775 [MPOL_PREFERRED] = "prefer",
2776 [MPOL_BIND] = "bind",
2777 [MPOL_INTERLEAVE] = "interleave",
2778 [MPOL_LOCAL] = "local",
2784 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2785 * @str: string containing mempolicy to parse
2786 * @mpol: pointer to struct mempolicy pointer, returned on success.
2789 * <mode>[=<flags>][:<nodelist>]
2791 * On success, returns 0, else 1
2793 int mpol_parse_str(char *str, struct mempolicy **mpol)
2795 struct mempolicy *new = NULL;
2796 unsigned short mode_flags;
2798 char *nodelist = strchr(str, ':');
2799 char *flags = strchr(str, '=');
2803 *flags++ = '\0'; /* terminate mode string */
2806 /* NUL-terminate mode or flags string */
2808 if (nodelist_parse(nodelist, nodes))
2810 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2815 mode = match_string(policy_modes, MPOL_MAX, str);
2820 case MPOL_PREFERRED:
2822 * Insist on a nodelist of one node only, although later
2823 * we use first_node(nodes) to grab a single node, so here
2824 * nodelist (or nodes) cannot be empty.
2827 char *rest = nodelist;
2828 while (isdigit(*rest))
2832 if (nodes_empty(nodes))
2836 case MPOL_INTERLEAVE:
2838 * Default to online nodes with memory if no nodelist
2841 nodes = node_states[N_MEMORY];
2845 * Don't allow a nodelist; mpol_new() checks flags
2849 mode = MPOL_PREFERRED;
2853 * Insist on a empty nodelist
2860 * Insist on a nodelist
2869 * Currently, we only support two mutually exclusive
2872 if (!strcmp(flags, "static"))
2873 mode_flags |= MPOL_F_STATIC_NODES;
2874 else if (!strcmp(flags, "relative"))
2875 mode_flags |= MPOL_F_RELATIVE_NODES;
2880 new = mpol_new(mode, mode_flags, &nodes);
2885 * Save nodes for mpol_to_str() to show the tmpfs mount options
2886 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2888 if (mode != MPOL_PREFERRED)
2889 new->v.nodes = nodes;
2891 new->v.preferred_node = first_node(nodes);
2893 new->flags |= MPOL_F_LOCAL;
2896 * Save nodes for contextualization: this will be used to "clone"
2897 * the mempolicy in a specific context [cpuset] at a later time.
2899 new->w.user_nodemask = nodes;
2904 /* Restore string for error message */
2913 #endif /* CONFIG_TMPFS */
2916 * mpol_to_str - format a mempolicy structure for printing
2917 * @buffer: to contain formatted mempolicy string
2918 * @maxlen: length of @buffer
2919 * @pol: pointer to mempolicy to be formatted
2921 * Convert @pol into a string. If @buffer is too short, truncate the string.
2922 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2923 * longest flag, "relative", and to display at least a few node ids.
2925 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2928 nodemask_t nodes = NODE_MASK_NONE;
2929 unsigned short mode = MPOL_DEFAULT;
2930 unsigned short flags = 0;
2932 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2940 case MPOL_PREFERRED:
2941 if (flags & MPOL_F_LOCAL)
2944 node_set(pol->v.preferred_node, nodes);
2947 case MPOL_INTERLEAVE:
2948 nodes = pol->v.nodes;
2952 snprintf(p, maxlen, "unknown");
2956 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2958 if (flags & MPOL_MODE_FLAGS) {
2959 p += snprintf(p, buffer + maxlen - p, "=");
2962 * Currently, the only defined flags are mutually exclusive
2964 if (flags & MPOL_F_STATIC_NODES)
2965 p += snprintf(p, buffer + maxlen - p, "static");
2966 else if (flags & MPOL_F_RELATIVE_NODES)
2967 p += snprintf(p, buffer + maxlen - p, "relative");
2970 if (!nodes_empty(nodes))
2971 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2972 nodemask_pr_args(&nodes));