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;
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 isolate_huge_page(page, qp->pagelist);
584 #ifdef CONFIG_NUMA_BALANCING
586 * This is used to mark a range of virtual addresses to be inaccessible.
587 * These are later cleared by a NUMA hinting fault. Depending on these
588 * faults, pages may be migrated for better NUMA placement.
590 * This is assuming that NUMA faults are handled using PROT_NONE. If
591 * an architecture makes a different choice, it will need further
592 * changes to the core.
594 unsigned long change_prot_numa(struct vm_area_struct *vma,
595 unsigned long addr, unsigned long end)
599 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
601 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
606 static unsigned long change_prot_numa(struct vm_area_struct *vma,
607 unsigned long addr, unsigned long end)
611 #endif /* CONFIG_NUMA_BALANCING */
613 static int queue_pages_test_walk(unsigned long start, unsigned long end,
614 struct mm_walk *walk)
616 struct vm_area_struct *vma = walk->vma;
617 struct queue_pages *qp = walk->private;
618 unsigned long endvma = vma->vm_end;
619 unsigned long flags = qp->flags;
622 * Need check MPOL_MF_STRICT to return -EIO if possible
623 * regardless of vma_migratable
625 if (!vma_migratable(vma) &&
626 !(flags & MPOL_MF_STRICT))
631 if (vma->vm_start > start)
632 start = vma->vm_start;
634 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
635 if (!vma->vm_next && vma->vm_end < end)
637 if (qp->prev && qp->prev->vm_end < vma->vm_start)
643 if (flags & MPOL_MF_LAZY) {
644 /* Similar to task_numa_work, skip inaccessible VMAs */
645 if (!is_vm_hugetlb_page(vma) &&
646 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
647 !(vma->vm_flags & VM_MIXEDMAP))
648 change_prot_numa(vma, start, endvma);
652 /* queue pages from current vma */
653 if (flags & MPOL_MF_VALID)
658 static const struct mm_walk_ops queue_pages_walk_ops = {
659 .hugetlb_entry = queue_pages_hugetlb,
660 .pmd_entry = queue_pages_pte_range,
661 .test_walk = queue_pages_test_walk,
665 * Walk through page tables and collect pages to be migrated.
667 * If pages found in a given range are on a set of nodes (determined by
668 * @nodes and @flags,) it's isolated and queued to the pagelist which is
669 * passed via @private.
671 * queue_pages_range() has three possible return values:
672 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
674 * 0 - queue pages successfully or no misplaced page.
675 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
676 * memory range specified by nodemask and maxnode points outside
677 * your accessible address space (-EFAULT)
680 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
681 nodemask_t *nodes, unsigned long flags,
682 struct list_head *pagelist)
684 struct queue_pages qp = {
685 .pagelist = pagelist,
691 return walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
695 * Apply policy to a single VMA
696 * This must be called with the mmap_sem held for writing.
698 static int vma_replace_policy(struct vm_area_struct *vma,
699 struct mempolicy *pol)
702 struct mempolicy *old;
703 struct mempolicy *new;
705 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
706 vma->vm_start, vma->vm_end, vma->vm_pgoff,
707 vma->vm_ops, vma->vm_file,
708 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
714 if (vma->vm_ops && vma->vm_ops->set_policy) {
715 err = vma->vm_ops->set_policy(vma, new);
720 old = vma->vm_policy;
721 vma->vm_policy = new; /* protected by mmap_sem */
730 /* Step 2: apply policy to a range and do splits. */
731 static int mbind_range(struct mm_struct *mm, unsigned long start,
732 unsigned long end, struct mempolicy *new_pol)
734 struct vm_area_struct *prev;
735 struct vm_area_struct *vma;
738 unsigned long vmstart;
741 vma = find_vma(mm, start);
742 if (!vma || vma->vm_start > start)
746 if (start > vma->vm_start)
749 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
750 vmstart = max(start, vma->vm_start);
751 vmend = min(end, vma->vm_end);
753 if (mpol_equal(vma_policy(vma), new_pol))
756 pgoff = vma->vm_pgoff +
757 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
758 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
759 vma->anon_vma, vma->vm_file, pgoff,
760 new_pol, vma->vm_userfaultfd_ctx);
765 if (vma->vm_start != vmstart) {
766 err = split_vma(vma->vm_mm, vma, vmstart, 1);
770 if (vma->vm_end != vmend) {
771 err = split_vma(vma->vm_mm, vma, vmend, 0);
776 err = vma_replace_policy(vma, new_pol);
785 /* Set the process memory policy */
786 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
789 struct mempolicy *new, *old;
790 NODEMASK_SCRATCH(scratch);
796 new = mpol_new(mode, flags, nodes);
803 ret = mpol_set_nodemask(new, nodes, scratch);
805 task_unlock(current);
809 old = current->mempolicy;
810 current->mempolicy = new;
811 if (new && new->mode == MPOL_INTERLEAVE)
812 current->il_prev = MAX_NUMNODES-1;
813 task_unlock(current);
817 NODEMASK_SCRATCH_FREE(scratch);
822 * Return nodemask for policy for get_mempolicy() query
824 * Called with task's alloc_lock held
826 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
829 if (p == &default_policy)
835 case MPOL_INTERLEAVE:
839 if (!(p->flags & MPOL_F_LOCAL))
840 node_set(p->v.preferred_node, *nodes);
841 /* else return empty node mask for local allocation */
848 static int lookup_node(struct mm_struct *mm, unsigned long addr)
854 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
856 err = page_to_nid(p);
860 up_read(&mm->mmap_sem);
864 /* Retrieve NUMA policy */
865 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
866 unsigned long addr, unsigned long flags)
869 struct mm_struct *mm = current->mm;
870 struct vm_area_struct *vma = NULL;
871 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
874 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
877 if (flags & MPOL_F_MEMS_ALLOWED) {
878 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
880 *policy = 0; /* just so it's initialized */
882 *nmask = cpuset_current_mems_allowed;
883 task_unlock(current);
887 if (flags & MPOL_F_ADDR) {
889 * Do NOT fall back to task policy if the
890 * vma/shared policy at addr is NULL. We
891 * want to return MPOL_DEFAULT in this case.
893 down_read(&mm->mmap_sem);
894 vma = find_vma_intersection(mm, addr, addr+1);
896 up_read(&mm->mmap_sem);
899 if (vma->vm_ops && vma->vm_ops->get_policy)
900 pol = vma->vm_ops->get_policy(vma, addr);
902 pol = vma->vm_policy;
907 pol = &default_policy; /* indicates default behavior */
909 if (flags & MPOL_F_NODE) {
910 if (flags & MPOL_F_ADDR) {
912 * Take a refcount on the mpol, lookup_node()
913 * wil drop the mmap_sem, so after calling
914 * lookup_node() only "pol" remains valid, "vma"
920 err = lookup_node(mm, addr);
924 } else if (pol == current->mempolicy &&
925 pol->mode == MPOL_INTERLEAVE) {
926 *policy = next_node_in(current->il_prev, pol->v.nodes);
932 *policy = pol == &default_policy ? MPOL_DEFAULT :
935 * Internal mempolicy flags must be masked off before exposing
936 * the policy to userspace.
938 *policy |= (pol->flags & MPOL_MODE_FLAGS);
943 if (mpol_store_user_nodemask(pol)) {
944 *nmask = pol->w.user_nodemask;
947 get_policy_nodemask(pol, nmask);
948 task_unlock(current);
955 up_read(&mm->mmap_sem);
957 mpol_put(pol_refcount);
961 #ifdef CONFIG_MIGRATION
963 * page migration, thp tail pages can be passed.
965 static int migrate_page_add(struct page *page, struct list_head *pagelist,
968 struct page *head = compound_head(page);
970 * Avoid migrating a page that is shared with others.
972 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
973 if (!isolate_lru_page(head)) {
974 list_add_tail(&head->lru, pagelist);
975 mod_node_page_state(page_pgdat(head),
976 NR_ISOLATED_ANON + page_is_file_cache(head),
977 hpage_nr_pages(head));
978 } else if (flags & MPOL_MF_STRICT) {
980 * Non-movable page may reach here. And, there may be
981 * temporary off LRU pages or non-LRU movable pages.
982 * Treat them as unmovable pages since they can't be
983 * isolated, so they can't be moved at the moment. It
984 * should return -EIO for this case too.
993 /* page allocation callback for NUMA node migration */
994 struct page *alloc_new_node_page(struct page *page, unsigned long node)
997 return alloc_huge_page_node(page_hstate(compound_head(page)),
999 else if (PageTransHuge(page)) {
1002 thp = alloc_pages_node(node,
1003 (GFP_TRANSHUGE | __GFP_THISNODE),
1007 prep_transhuge_page(thp);
1010 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1015 * Migrate pages from one node to a target node.
1016 * Returns error or the number of pages not migrated.
1018 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1022 LIST_HEAD(pagelist);
1026 node_set(source, nmask);
1029 * This does not "check" the range but isolates all pages that
1030 * need migration. Between passing in the full user address
1031 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1033 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1034 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1035 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1037 if (!list_empty(&pagelist)) {
1038 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1039 MIGRATE_SYNC, MR_SYSCALL);
1041 putback_movable_pages(&pagelist);
1048 * Move pages between the two nodesets so as to preserve the physical
1049 * layout as much as possible.
1051 * Returns the number of page that could not be moved.
1053 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1054 const nodemask_t *to, int flags)
1060 err = migrate_prep();
1064 down_read(&mm->mmap_sem);
1067 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1068 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1069 * bit in 'tmp', and return that <source, dest> pair for migration.
1070 * The pair of nodemasks 'to' and 'from' define the map.
1072 * If no pair of bits is found that way, fallback to picking some
1073 * pair of 'source' and 'dest' bits that are not the same. If the
1074 * 'source' and 'dest' bits are the same, this represents a node
1075 * that will be migrating to itself, so no pages need move.
1077 * If no bits are left in 'tmp', or if all remaining bits left
1078 * in 'tmp' correspond to the same bit in 'to', return false
1079 * (nothing left to migrate).
1081 * This lets us pick a pair of nodes to migrate between, such that
1082 * if possible the dest node is not already occupied by some other
1083 * source node, minimizing the risk of overloading the memory on a
1084 * node that would happen if we migrated incoming memory to a node
1085 * before migrating outgoing memory source that same node.
1087 * A single scan of tmp is sufficient. As we go, we remember the
1088 * most recent <s, d> pair that moved (s != d). If we find a pair
1089 * that not only moved, but what's better, moved to an empty slot
1090 * (d is not set in tmp), then we break out then, with that pair.
1091 * Otherwise when we finish scanning from_tmp, we at least have the
1092 * most recent <s, d> pair that moved. If we get all the way through
1093 * the scan of tmp without finding any node that moved, much less
1094 * moved to an empty node, then there is nothing left worth migrating.
1098 while (!nodes_empty(tmp)) {
1100 int source = NUMA_NO_NODE;
1103 for_each_node_mask(s, tmp) {
1106 * do_migrate_pages() tries to maintain the relative
1107 * node relationship of the pages established between
1108 * threads and memory areas.
1110 * However if the number of source nodes is not equal to
1111 * the number of destination nodes we can not preserve
1112 * this node relative relationship. In that case, skip
1113 * copying memory from a node that is in the destination
1116 * Example: [2,3,4] -> [3,4,5] moves everything.
1117 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1120 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1121 (node_isset(s, *to)))
1124 d = node_remap(s, *from, *to);
1128 source = s; /* Node moved. Memorize */
1131 /* dest not in remaining from nodes? */
1132 if (!node_isset(dest, tmp))
1135 if (source == NUMA_NO_NODE)
1138 node_clear(source, tmp);
1139 err = migrate_to_node(mm, source, dest, flags);
1145 up_read(&mm->mmap_sem);
1153 * Allocate a new page for page migration based on vma policy.
1154 * Start by assuming the page is mapped by the same vma as contains @start.
1155 * Search forward from there, if not. N.B., this assumes that the
1156 * list of pages handed to migrate_pages()--which is how we get here--
1157 * is in virtual address order.
1159 static struct page *new_page(struct page *page, unsigned long start)
1161 struct vm_area_struct *vma;
1162 unsigned long uninitialized_var(address);
1164 vma = find_vma(current->mm, start);
1166 address = page_address_in_vma(page, vma);
1167 if (address != -EFAULT)
1172 if (PageHuge(page)) {
1173 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1175 } else if (PageTransHuge(page)) {
1178 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1182 prep_transhuge_page(thp);
1186 * if !vma, alloc_page_vma() will use task or system default policy
1188 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1193 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1194 unsigned long flags)
1199 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1200 const nodemask_t *to, int flags)
1205 static struct page *new_page(struct page *page, unsigned long start)
1211 static long do_mbind(unsigned long start, unsigned long len,
1212 unsigned short mode, unsigned short mode_flags,
1213 nodemask_t *nmask, unsigned long flags)
1215 struct mm_struct *mm = current->mm;
1216 struct mempolicy *new;
1220 LIST_HEAD(pagelist);
1222 if (flags & ~(unsigned long)MPOL_MF_VALID)
1224 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1227 if (start & ~PAGE_MASK)
1230 if (mode == MPOL_DEFAULT)
1231 flags &= ~MPOL_MF_STRICT;
1233 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1241 new = mpol_new(mode, mode_flags, nmask);
1243 return PTR_ERR(new);
1245 if (flags & MPOL_MF_LAZY)
1246 new->flags |= MPOL_F_MOF;
1249 * If we are using the default policy then operation
1250 * on discontinuous address spaces is okay after all
1253 flags |= MPOL_MF_DISCONTIG_OK;
1255 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1256 start, start + len, mode, mode_flags,
1257 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1259 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1261 err = migrate_prep();
1266 NODEMASK_SCRATCH(scratch);
1268 down_write(&mm->mmap_sem);
1270 err = mpol_set_nodemask(new, nmask, scratch);
1271 task_unlock(current);
1273 up_write(&mm->mmap_sem);
1276 NODEMASK_SCRATCH_FREE(scratch);
1281 ret = queue_pages_range(mm, start, end, nmask,
1282 flags | MPOL_MF_INVERT, &pagelist);
1289 err = mbind_range(mm, start, end, new);
1294 if (!list_empty(&pagelist)) {
1295 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1296 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1297 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1299 putback_movable_pages(&pagelist);
1302 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1306 if (!list_empty(&pagelist))
1307 putback_movable_pages(&pagelist);
1310 up_write(&mm->mmap_sem);
1317 * User space interface with variable sized bitmaps for nodelists.
1320 /* Copy a node mask from user space. */
1321 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1322 unsigned long maxnode)
1326 unsigned long nlongs;
1327 unsigned long endmask;
1330 nodes_clear(*nodes);
1331 if (maxnode == 0 || !nmask)
1333 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1336 nlongs = BITS_TO_LONGS(maxnode);
1337 if ((maxnode % BITS_PER_LONG) == 0)
1340 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1343 * When the user specified more nodes than supported just check
1344 * if the non supported part is all zero.
1346 * If maxnode have more longs than MAX_NUMNODES, check
1347 * the bits in that area first. And then go through to
1348 * check the rest bits which equal or bigger than MAX_NUMNODES.
1349 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1351 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1352 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1353 if (get_user(t, nmask + k))
1355 if (k == nlongs - 1) {
1361 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1365 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1366 unsigned long valid_mask = endmask;
1368 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1369 if (get_user(t, nmask + nlongs - 1))
1375 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1377 nodes_addr(*nodes)[nlongs-1] &= endmask;
1381 /* Copy a kernel node mask to user space */
1382 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1385 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1386 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1388 if (copy > nbytes) {
1389 if (copy > PAGE_SIZE)
1391 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1395 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1398 static long kernel_mbind(unsigned long start, unsigned long len,
1399 unsigned long mode, const unsigned long __user *nmask,
1400 unsigned long maxnode, unsigned int flags)
1404 unsigned short mode_flags;
1406 start = untagged_addr(start);
1407 mode_flags = mode & MPOL_MODE_FLAGS;
1408 mode &= ~MPOL_MODE_FLAGS;
1409 if (mode >= MPOL_MAX)
1411 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1412 (mode_flags & MPOL_F_RELATIVE_NODES))
1414 err = get_nodes(&nodes, nmask, maxnode);
1417 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1420 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1421 unsigned long, mode, const unsigned long __user *, nmask,
1422 unsigned long, maxnode, unsigned int, flags)
1424 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1427 /* Set the process memory policy */
1428 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1429 unsigned long maxnode)
1433 unsigned short flags;
1435 flags = mode & MPOL_MODE_FLAGS;
1436 mode &= ~MPOL_MODE_FLAGS;
1437 if ((unsigned int)mode >= MPOL_MAX)
1439 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1441 err = get_nodes(&nodes, nmask, maxnode);
1444 return do_set_mempolicy(mode, flags, &nodes);
1447 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1448 unsigned long, maxnode)
1450 return kernel_set_mempolicy(mode, nmask, maxnode);
1453 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1454 const unsigned long __user *old_nodes,
1455 const unsigned long __user *new_nodes)
1457 struct mm_struct *mm = NULL;
1458 struct task_struct *task;
1459 nodemask_t task_nodes;
1463 NODEMASK_SCRATCH(scratch);
1468 old = &scratch->mask1;
1469 new = &scratch->mask2;
1471 err = get_nodes(old, old_nodes, maxnode);
1475 err = get_nodes(new, new_nodes, maxnode);
1479 /* Find the mm_struct */
1481 task = pid ? find_task_by_vpid(pid) : current;
1487 get_task_struct(task);
1492 * Check if this process has the right to modify the specified process.
1493 * Use the regular "ptrace_may_access()" checks.
1495 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1502 task_nodes = cpuset_mems_allowed(task);
1503 /* Is the user allowed to access the target nodes? */
1504 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1509 task_nodes = cpuset_mems_allowed(current);
1510 nodes_and(*new, *new, task_nodes);
1511 if (nodes_empty(*new))
1514 err = security_task_movememory(task);
1518 mm = get_task_mm(task);
1519 put_task_struct(task);
1526 err = do_migrate_pages(mm, old, new,
1527 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1531 NODEMASK_SCRATCH_FREE(scratch);
1536 put_task_struct(task);
1541 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1542 const unsigned long __user *, old_nodes,
1543 const unsigned long __user *, new_nodes)
1545 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1549 /* Retrieve NUMA policy */
1550 static int kernel_get_mempolicy(int __user *policy,
1551 unsigned long __user *nmask,
1552 unsigned long maxnode,
1554 unsigned long flags)
1557 int uninitialized_var(pval);
1560 addr = untagged_addr(addr);
1562 if (nmask != NULL && maxnode < nr_node_ids)
1565 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1570 if (policy && put_user(pval, policy))
1574 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1579 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1580 unsigned long __user *, nmask, unsigned long, maxnode,
1581 unsigned long, addr, unsigned long, flags)
1583 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1586 #ifdef CONFIG_COMPAT
1588 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1589 compat_ulong_t __user *, nmask,
1590 compat_ulong_t, maxnode,
1591 compat_ulong_t, addr, compat_ulong_t, flags)
1594 unsigned long __user *nm = NULL;
1595 unsigned long nr_bits, alloc_size;
1596 DECLARE_BITMAP(bm, MAX_NUMNODES);
1598 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1599 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1602 nm = compat_alloc_user_space(alloc_size);
1604 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1606 if (!err && nmask) {
1607 unsigned long copy_size;
1608 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1609 err = copy_from_user(bm, nm, copy_size);
1610 /* ensure entire bitmap is zeroed */
1611 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1612 err |= compat_put_bitmap(nmask, bm, nr_bits);
1618 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1619 compat_ulong_t, maxnode)
1621 unsigned long __user *nm = NULL;
1622 unsigned long nr_bits, alloc_size;
1623 DECLARE_BITMAP(bm, MAX_NUMNODES);
1625 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1626 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1629 if (compat_get_bitmap(bm, nmask, nr_bits))
1631 nm = compat_alloc_user_space(alloc_size);
1632 if (copy_to_user(nm, bm, alloc_size))
1636 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1639 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1640 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1641 compat_ulong_t, maxnode, compat_ulong_t, flags)
1643 unsigned long __user *nm = NULL;
1644 unsigned long nr_bits, alloc_size;
1647 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1648 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1651 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1653 nm = compat_alloc_user_space(alloc_size);
1654 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1658 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1661 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1662 compat_ulong_t, maxnode,
1663 const compat_ulong_t __user *, old_nodes,
1664 const compat_ulong_t __user *, new_nodes)
1666 unsigned long __user *old = NULL;
1667 unsigned long __user *new = NULL;
1668 nodemask_t tmp_mask;
1669 unsigned long nr_bits;
1672 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1673 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1675 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1677 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1679 new = old + size / sizeof(unsigned long);
1680 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1684 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1687 new = compat_alloc_user_space(size);
1688 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1691 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1694 #endif /* CONFIG_COMPAT */
1696 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1699 struct mempolicy *pol = NULL;
1702 if (vma->vm_ops && vma->vm_ops->get_policy) {
1703 pol = vma->vm_ops->get_policy(vma, addr);
1704 } else if (vma->vm_policy) {
1705 pol = vma->vm_policy;
1708 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1709 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1710 * count on these policies which will be dropped by
1711 * mpol_cond_put() later
1713 if (mpol_needs_cond_ref(pol))
1722 * get_vma_policy(@vma, @addr)
1723 * @vma: virtual memory area whose policy is sought
1724 * @addr: address in @vma for shared policy lookup
1726 * Returns effective policy for a VMA at specified address.
1727 * Falls back to current->mempolicy or system default policy, as necessary.
1728 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1729 * count--added by the get_policy() vm_op, as appropriate--to protect against
1730 * freeing by another task. It is the caller's responsibility to free the
1731 * extra reference for shared policies.
1733 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1736 struct mempolicy *pol = __get_vma_policy(vma, addr);
1739 pol = get_task_policy(current);
1744 bool vma_policy_mof(struct vm_area_struct *vma)
1746 struct mempolicy *pol;
1748 if (vma->vm_ops && vma->vm_ops->get_policy) {
1751 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1752 if (pol && (pol->flags & MPOL_F_MOF))
1759 pol = vma->vm_policy;
1761 pol = get_task_policy(current);
1763 return pol->flags & MPOL_F_MOF;
1766 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1768 enum zone_type dynamic_policy_zone = policy_zone;
1770 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1773 * if policy->v.nodes has movable memory only,
1774 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1776 * policy->v.nodes is intersect with node_states[N_MEMORY].
1777 * so if the following test faile, it implies
1778 * policy->v.nodes has movable memory only.
1780 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1781 dynamic_policy_zone = ZONE_MOVABLE;
1783 return zone >= dynamic_policy_zone;
1787 * Return a nodemask representing a mempolicy for filtering nodes for
1790 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1792 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1793 if (unlikely(policy->mode == MPOL_BIND) &&
1794 apply_policy_zone(policy, gfp_zone(gfp)) &&
1795 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1796 return &policy->v.nodes;
1801 /* Return the node id preferred by the given mempolicy, or the given id */
1802 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1805 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1806 nd = policy->v.preferred_node;
1809 * __GFP_THISNODE shouldn't even be used with the bind policy
1810 * because we might easily break the expectation to stay on the
1811 * requested node and not break the policy.
1813 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1819 /* Do dynamic interleaving for a process */
1820 static unsigned interleave_nodes(struct mempolicy *policy)
1823 struct task_struct *me = current;
1825 next = next_node_in(me->il_prev, policy->v.nodes);
1826 if (next < MAX_NUMNODES)
1832 * Depending on the memory policy provide a node from which to allocate the
1835 unsigned int mempolicy_slab_node(void)
1837 struct mempolicy *policy;
1838 int node = numa_mem_id();
1843 policy = current->mempolicy;
1844 if (!policy || policy->flags & MPOL_F_LOCAL)
1847 switch (policy->mode) {
1848 case MPOL_PREFERRED:
1850 * handled MPOL_F_LOCAL above
1852 return policy->v.preferred_node;
1854 case MPOL_INTERLEAVE:
1855 return interleave_nodes(policy);
1861 * Follow bind policy behavior and start allocation at the
1864 struct zonelist *zonelist;
1865 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1866 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1867 z = first_zones_zonelist(zonelist, highest_zoneidx,
1869 return z->zone ? zone_to_nid(z->zone) : node;
1878 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1879 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1880 * number of present nodes.
1882 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1884 unsigned nnodes = nodes_weight(pol->v.nodes);
1890 return numa_node_id();
1891 target = (unsigned int)n % nnodes;
1892 nid = first_node(pol->v.nodes);
1893 for (i = 0; i < target; i++)
1894 nid = next_node(nid, pol->v.nodes);
1898 /* Determine a node number for interleave */
1899 static inline unsigned interleave_nid(struct mempolicy *pol,
1900 struct vm_area_struct *vma, unsigned long addr, int shift)
1906 * for small pages, there is no difference between
1907 * shift and PAGE_SHIFT, so the bit-shift is safe.
1908 * for huge pages, since vm_pgoff is in units of small
1909 * pages, we need to shift off the always 0 bits to get
1912 BUG_ON(shift < PAGE_SHIFT);
1913 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1914 off += (addr - vma->vm_start) >> shift;
1915 return offset_il_node(pol, off);
1917 return interleave_nodes(pol);
1920 #ifdef CONFIG_HUGETLBFS
1922 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1923 * @vma: virtual memory area whose policy is sought
1924 * @addr: address in @vma for shared policy lookup and interleave policy
1925 * @gfp_flags: for requested zone
1926 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1927 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1929 * Returns a nid suitable for a huge page allocation and a pointer
1930 * to the struct mempolicy for conditional unref after allocation.
1931 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1932 * @nodemask for filtering the zonelist.
1934 * Must be protected by read_mems_allowed_begin()
1936 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1937 struct mempolicy **mpol, nodemask_t **nodemask)
1941 *mpol = get_vma_policy(vma, addr);
1942 *nodemask = NULL; /* assume !MPOL_BIND */
1944 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1945 nid = interleave_nid(*mpol, vma, addr,
1946 huge_page_shift(hstate_vma(vma)));
1948 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1949 if ((*mpol)->mode == MPOL_BIND)
1950 *nodemask = &(*mpol)->v.nodes;
1956 * init_nodemask_of_mempolicy
1958 * If the current task's mempolicy is "default" [NULL], return 'false'
1959 * to indicate default policy. Otherwise, extract the policy nodemask
1960 * for 'bind' or 'interleave' policy into the argument nodemask, or
1961 * initialize the argument nodemask to contain the single node for
1962 * 'preferred' or 'local' policy and return 'true' to indicate presence
1963 * of non-default mempolicy.
1965 * We don't bother with reference counting the mempolicy [mpol_get/put]
1966 * because the current task is examining it's own mempolicy and a task's
1967 * mempolicy is only ever changed by the task itself.
1969 * N.B., it is the caller's responsibility to free a returned nodemask.
1971 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1973 struct mempolicy *mempolicy;
1976 if (!(mask && current->mempolicy))
1980 mempolicy = current->mempolicy;
1981 switch (mempolicy->mode) {
1982 case MPOL_PREFERRED:
1983 if (mempolicy->flags & MPOL_F_LOCAL)
1984 nid = numa_node_id();
1986 nid = mempolicy->v.preferred_node;
1987 init_nodemask_of_node(mask, nid);
1992 case MPOL_INTERLEAVE:
1993 *mask = mempolicy->v.nodes;
1999 task_unlock(current);
2006 * mempolicy_nodemask_intersects
2008 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2009 * policy. Otherwise, check for intersection between mask and the policy
2010 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2011 * policy, always return true since it may allocate elsewhere on fallback.
2013 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2015 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2016 const nodemask_t *mask)
2018 struct mempolicy *mempolicy;
2024 mempolicy = tsk->mempolicy;
2028 switch (mempolicy->mode) {
2029 case MPOL_PREFERRED:
2031 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2032 * allocate from, they may fallback to other nodes when oom.
2033 * Thus, it's possible for tsk to have allocated memory from
2038 case MPOL_INTERLEAVE:
2039 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2049 /* Allocate a page in interleaved policy.
2050 Own path because it needs to do special accounting. */
2051 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2056 page = __alloc_pages(gfp, order, nid);
2057 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2058 if (!static_branch_likely(&vm_numa_stat_key))
2060 if (page && page_to_nid(page) == nid) {
2062 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2069 * alloc_pages_vma - Allocate a page for a VMA.
2072 * %GFP_USER user allocation.
2073 * %GFP_KERNEL kernel allocations,
2074 * %GFP_HIGHMEM highmem/user allocations,
2075 * %GFP_FS allocation should not call back into a file system.
2076 * %GFP_ATOMIC don't sleep.
2078 * @order:Order of the GFP allocation.
2079 * @vma: Pointer to VMA or NULL if not available.
2080 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2081 * @node: Which node to prefer for allocation (modulo policy).
2082 * @hugepage: for hugepages try only the preferred node if possible
2084 * This function allocates a page from the kernel page pool and applies
2085 * a NUMA policy associated with the VMA or the current process.
2086 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2087 * mm_struct of the VMA to prevent it from going away. Should be used for
2088 * all allocations for pages that will be mapped into user space. Returns
2089 * NULL when no page can be allocated.
2092 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2093 unsigned long addr, int node, bool hugepage)
2095 struct mempolicy *pol;
2100 pol = get_vma_policy(vma, addr);
2102 if (pol->mode == MPOL_INTERLEAVE) {
2105 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2107 page = alloc_page_interleave(gfp, order, nid);
2111 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2112 int hpage_node = node;
2115 * For hugepage allocation and non-interleave policy which
2116 * allows the current node (or other explicitly preferred
2117 * node) we only try to allocate from the current/preferred
2118 * node and don't fall back to other nodes, as the cost of
2119 * remote accesses would likely offset THP benefits.
2121 * If the policy is interleave, or does not allow the current
2122 * node in its nodemask, we allocate the standard way.
2124 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2125 hpage_node = pol->v.preferred_node;
2127 nmask = policy_nodemask(gfp, pol);
2128 if (!nmask || node_isset(hpage_node, *nmask)) {
2130 page = __alloc_pages_node(hpage_node,
2131 gfp | __GFP_THISNODE, order);
2134 * If hugepage allocations are configured to always
2135 * synchronous compact or the vma has been madvised
2136 * to prefer hugepage backing, retry allowing remote
2139 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2140 page = __alloc_pages_nodemask(gfp | __GFP_NORETRY,
2148 nmask = policy_nodemask(gfp, pol);
2149 preferred_nid = policy_node(gfp, pol, node);
2150 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2155 EXPORT_SYMBOL(alloc_pages_vma);
2158 * alloc_pages_current - Allocate pages.
2161 * %GFP_USER user allocation,
2162 * %GFP_KERNEL kernel allocation,
2163 * %GFP_HIGHMEM highmem allocation,
2164 * %GFP_FS don't call back into a file system.
2165 * %GFP_ATOMIC don't sleep.
2166 * @order: Power of two of allocation size in pages. 0 is a single page.
2168 * Allocate a page from the kernel page pool. When not in
2169 * interrupt context and apply the current process NUMA policy.
2170 * Returns NULL when no page can be allocated.
2172 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2174 struct mempolicy *pol = &default_policy;
2177 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2178 pol = get_task_policy(current);
2181 * No reference counting needed for current->mempolicy
2182 * nor system default_policy
2184 if (pol->mode == MPOL_INTERLEAVE)
2185 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2187 page = __alloc_pages_nodemask(gfp, order,
2188 policy_node(gfp, pol, numa_node_id()),
2189 policy_nodemask(gfp, pol));
2193 EXPORT_SYMBOL(alloc_pages_current);
2195 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2197 struct mempolicy *pol = mpol_dup(vma_policy(src));
2200 return PTR_ERR(pol);
2201 dst->vm_policy = pol;
2206 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2207 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2208 * with the mems_allowed returned by cpuset_mems_allowed(). This
2209 * keeps mempolicies cpuset relative after its cpuset moves. See
2210 * further kernel/cpuset.c update_nodemask().
2212 * current's mempolicy may be rebinded by the other task(the task that changes
2213 * cpuset's mems), so we needn't do rebind work for current task.
2216 /* Slow path of a mempolicy duplicate */
2217 struct mempolicy *__mpol_dup(struct mempolicy *old)
2219 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2222 return ERR_PTR(-ENOMEM);
2224 /* task's mempolicy is protected by alloc_lock */
2225 if (old == current->mempolicy) {
2228 task_unlock(current);
2232 if (current_cpuset_is_being_rebound()) {
2233 nodemask_t mems = cpuset_mems_allowed(current);
2234 mpol_rebind_policy(new, &mems);
2236 atomic_set(&new->refcnt, 1);
2240 /* Slow path of a mempolicy comparison */
2241 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2245 if (a->mode != b->mode)
2247 if (a->flags != b->flags)
2249 if (mpol_store_user_nodemask(a))
2250 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2256 case MPOL_INTERLEAVE:
2257 return !!nodes_equal(a->v.nodes, b->v.nodes);
2258 case MPOL_PREFERRED:
2259 /* a's ->flags is the same as b's */
2260 if (a->flags & MPOL_F_LOCAL)
2262 return a->v.preferred_node == b->v.preferred_node;
2270 * Shared memory backing store policy support.
2272 * Remember policies even when nobody has shared memory mapped.
2273 * The policies are kept in Red-Black tree linked from the inode.
2274 * They are protected by the sp->lock rwlock, which should be held
2275 * for any accesses to the tree.
2279 * lookup first element intersecting start-end. Caller holds sp->lock for
2280 * reading or for writing
2282 static struct sp_node *
2283 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2285 struct rb_node *n = sp->root.rb_node;
2288 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2290 if (start >= p->end)
2292 else if (end <= p->start)
2300 struct sp_node *w = NULL;
2301 struct rb_node *prev = rb_prev(n);
2304 w = rb_entry(prev, struct sp_node, nd);
2305 if (w->end <= start)
2309 return rb_entry(n, struct sp_node, nd);
2313 * Insert a new shared policy into the list. Caller holds sp->lock for
2316 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2318 struct rb_node **p = &sp->root.rb_node;
2319 struct rb_node *parent = NULL;
2324 nd = rb_entry(parent, struct sp_node, nd);
2325 if (new->start < nd->start)
2327 else if (new->end > nd->end)
2328 p = &(*p)->rb_right;
2332 rb_link_node(&new->nd, parent, p);
2333 rb_insert_color(&new->nd, &sp->root);
2334 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2335 new->policy ? new->policy->mode : 0);
2338 /* Find shared policy intersecting idx */
2340 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2342 struct mempolicy *pol = NULL;
2345 if (!sp->root.rb_node)
2347 read_lock(&sp->lock);
2348 sn = sp_lookup(sp, idx, idx+1);
2350 mpol_get(sn->policy);
2353 read_unlock(&sp->lock);
2357 static void sp_free(struct sp_node *n)
2359 mpol_put(n->policy);
2360 kmem_cache_free(sn_cache, n);
2364 * mpol_misplaced - check whether current page node is valid in policy
2366 * @page: page to be checked
2367 * @vma: vm area where page mapped
2368 * @addr: virtual address where page mapped
2370 * Lookup current policy node id for vma,addr and "compare to" page's
2374 * -1 - not misplaced, page is in the right node
2375 * node - node id where the page should be
2377 * Policy determination "mimics" alloc_page_vma().
2378 * Called from fault path where we know the vma and faulting address.
2380 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2382 struct mempolicy *pol;
2384 int curnid = page_to_nid(page);
2385 unsigned long pgoff;
2386 int thiscpu = raw_smp_processor_id();
2387 int thisnid = cpu_to_node(thiscpu);
2388 int polnid = NUMA_NO_NODE;
2391 pol = get_vma_policy(vma, addr);
2392 if (!(pol->flags & MPOL_F_MOF))
2395 switch (pol->mode) {
2396 case MPOL_INTERLEAVE:
2397 pgoff = vma->vm_pgoff;
2398 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2399 polnid = offset_il_node(pol, pgoff);
2402 case MPOL_PREFERRED:
2403 if (pol->flags & MPOL_F_LOCAL)
2404 polnid = numa_node_id();
2406 polnid = pol->v.preferred_node;
2412 * allows binding to multiple nodes.
2413 * use current page if in policy nodemask,
2414 * else select nearest allowed node, if any.
2415 * If no allowed nodes, use current [!misplaced].
2417 if (node_isset(curnid, pol->v.nodes))
2419 z = first_zones_zonelist(
2420 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2421 gfp_zone(GFP_HIGHUSER),
2423 polnid = zone_to_nid(z->zone);
2430 /* Migrate the page towards the node whose CPU is referencing it */
2431 if (pol->flags & MPOL_F_MORON) {
2434 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2438 if (curnid != polnid)
2447 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2448 * dropped after task->mempolicy is set to NULL so that any allocation done as
2449 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2452 void mpol_put_task_policy(struct task_struct *task)
2454 struct mempolicy *pol;
2457 pol = task->mempolicy;
2458 task->mempolicy = NULL;
2463 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2465 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2466 rb_erase(&n->nd, &sp->root);
2470 static void sp_node_init(struct sp_node *node, unsigned long start,
2471 unsigned long end, struct mempolicy *pol)
2473 node->start = start;
2478 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2479 struct mempolicy *pol)
2482 struct mempolicy *newpol;
2484 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2488 newpol = mpol_dup(pol);
2489 if (IS_ERR(newpol)) {
2490 kmem_cache_free(sn_cache, n);
2493 newpol->flags |= MPOL_F_SHARED;
2494 sp_node_init(n, start, end, newpol);
2499 /* Replace a policy range. */
2500 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2501 unsigned long end, struct sp_node *new)
2504 struct sp_node *n_new = NULL;
2505 struct mempolicy *mpol_new = NULL;
2509 write_lock(&sp->lock);
2510 n = sp_lookup(sp, start, end);
2511 /* Take care of old policies in the same range. */
2512 while (n && n->start < end) {
2513 struct rb_node *next = rb_next(&n->nd);
2514 if (n->start >= start) {
2520 /* Old policy spanning whole new range. */
2525 *mpol_new = *n->policy;
2526 atomic_set(&mpol_new->refcnt, 1);
2527 sp_node_init(n_new, end, n->end, mpol_new);
2529 sp_insert(sp, n_new);
2538 n = rb_entry(next, struct sp_node, nd);
2542 write_unlock(&sp->lock);
2549 kmem_cache_free(sn_cache, n_new);
2554 write_unlock(&sp->lock);
2556 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2559 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2562 atomic_set(&mpol_new->refcnt, 1);
2567 * mpol_shared_policy_init - initialize shared policy for inode
2568 * @sp: pointer to inode shared policy
2569 * @mpol: struct mempolicy to install
2571 * Install non-NULL @mpol in inode's shared policy rb-tree.
2572 * On entry, the current task has a reference on a non-NULL @mpol.
2573 * This must be released on exit.
2574 * This is called at get_inode() calls and we can use GFP_KERNEL.
2576 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2580 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2581 rwlock_init(&sp->lock);
2584 struct vm_area_struct pvma;
2585 struct mempolicy *new;
2586 NODEMASK_SCRATCH(scratch);
2590 /* contextualize the tmpfs mount point mempolicy */
2591 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2593 goto free_scratch; /* no valid nodemask intersection */
2596 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2597 task_unlock(current);
2601 /* Create pseudo-vma that contains just the policy */
2602 vma_init(&pvma, NULL);
2603 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2604 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2607 mpol_put(new); /* drop initial ref */
2609 NODEMASK_SCRATCH_FREE(scratch);
2611 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2615 int mpol_set_shared_policy(struct shared_policy *info,
2616 struct vm_area_struct *vma, struct mempolicy *npol)
2619 struct sp_node *new = NULL;
2620 unsigned long sz = vma_pages(vma);
2622 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2624 sz, npol ? npol->mode : -1,
2625 npol ? npol->flags : -1,
2626 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2629 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2633 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2639 /* Free a backing policy store on inode delete. */
2640 void mpol_free_shared_policy(struct shared_policy *p)
2643 struct rb_node *next;
2645 if (!p->root.rb_node)
2647 write_lock(&p->lock);
2648 next = rb_first(&p->root);
2650 n = rb_entry(next, struct sp_node, nd);
2651 next = rb_next(&n->nd);
2654 write_unlock(&p->lock);
2657 #ifdef CONFIG_NUMA_BALANCING
2658 static int __initdata numabalancing_override;
2660 static void __init check_numabalancing_enable(void)
2662 bool numabalancing_default = false;
2664 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2665 numabalancing_default = true;
2667 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2668 if (numabalancing_override)
2669 set_numabalancing_state(numabalancing_override == 1);
2671 if (num_online_nodes() > 1 && !numabalancing_override) {
2672 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2673 numabalancing_default ? "Enabling" : "Disabling");
2674 set_numabalancing_state(numabalancing_default);
2678 static int __init setup_numabalancing(char *str)
2684 if (!strcmp(str, "enable")) {
2685 numabalancing_override = 1;
2687 } else if (!strcmp(str, "disable")) {
2688 numabalancing_override = -1;
2693 pr_warn("Unable to parse numa_balancing=\n");
2697 __setup("numa_balancing=", setup_numabalancing);
2699 static inline void __init check_numabalancing_enable(void)
2702 #endif /* CONFIG_NUMA_BALANCING */
2704 /* assumes fs == KERNEL_DS */
2705 void __init numa_policy_init(void)
2707 nodemask_t interleave_nodes;
2708 unsigned long largest = 0;
2709 int nid, prefer = 0;
2711 policy_cache = kmem_cache_create("numa_policy",
2712 sizeof(struct mempolicy),
2713 0, SLAB_PANIC, NULL);
2715 sn_cache = kmem_cache_create("shared_policy_node",
2716 sizeof(struct sp_node),
2717 0, SLAB_PANIC, NULL);
2719 for_each_node(nid) {
2720 preferred_node_policy[nid] = (struct mempolicy) {
2721 .refcnt = ATOMIC_INIT(1),
2722 .mode = MPOL_PREFERRED,
2723 .flags = MPOL_F_MOF | MPOL_F_MORON,
2724 .v = { .preferred_node = nid, },
2729 * Set interleaving policy for system init. Interleaving is only
2730 * enabled across suitably sized nodes (default is >= 16MB), or
2731 * fall back to the largest node if they're all smaller.
2733 nodes_clear(interleave_nodes);
2734 for_each_node_state(nid, N_MEMORY) {
2735 unsigned long total_pages = node_present_pages(nid);
2737 /* Preserve the largest node */
2738 if (largest < total_pages) {
2739 largest = total_pages;
2743 /* Interleave this node? */
2744 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2745 node_set(nid, interleave_nodes);
2748 /* All too small, use the largest */
2749 if (unlikely(nodes_empty(interleave_nodes)))
2750 node_set(prefer, interleave_nodes);
2752 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2753 pr_err("%s: interleaving failed\n", __func__);
2755 check_numabalancing_enable();
2758 /* Reset policy of current process to default */
2759 void numa_default_policy(void)
2761 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2765 * Parse and format mempolicy from/to strings
2769 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2771 static const char * const policy_modes[] =
2773 [MPOL_DEFAULT] = "default",
2774 [MPOL_PREFERRED] = "prefer",
2775 [MPOL_BIND] = "bind",
2776 [MPOL_INTERLEAVE] = "interleave",
2777 [MPOL_LOCAL] = "local",
2783 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2784 * @str: string containing mempolicy to parse
2785 * @mpol: pointer to struct mempolicy pointer, returned on success.
2788 * <mode>[=<flags>][:<nodelist>]
2790 * On success, returns 0, else 1
2792 int mpol_parse_str(char *str, struct mempolicy **mpol)
2794 struct mempolicy *new = NULL;
2795 unsigned short mode_flags;
2797 char *nodelist = strchr(str, ':');
2798 char *flags = strchr(str, '=');
2802 *flags++ = '\0'; /* terminate mode string */
2805 /* NUL-terminate mode or flags string */
2807 if (nodelist_parse(nodelist, nodes))
2809 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2814 mode = match_string(policy_modes, MPOL_MAX, str);
2819 case MPOL_PREFERRED:
2821 * Insist on a nodelist of one node only, although later
2822 * we use first_node(nodes) to grab a single node, so here
2823 * nodelist (or nodes) cannot be empty.
2826 char *rest = nodelist;
2827 while (isdigit(*rest))
2831 if (nodes_empty(nodes))
2835 case MPOL_INTERLEAVE:
2837 * Default to online nodes with memory if no nodelist
2840 nodes = node_states[N_MEMORY];
2844 * Don't allow a nodelist; mpol_new() checks flags
2848 mode = MPOL_PREFERRED;
2852 * Insist on a empty nodelist
2859 * Insist on a nodelist
2868 * Currently, we only support two mutually exclusive
2871 if (!strcmp(flags, "static"))
2872 mode_flags |= MPOL_F_STATIC_NODES;
2873 else if (!strcmp(flags, "relative"))
2874 mode_flags |= MPOL_F_RELATIVE_NODES;
2879 new = mpol_new(mode, mode_flags, &nodes);
2884 * Save nodes for mpol_to_str() to show the tmpfs mount options
2885 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2887 if (mode != MPOL_PREFERRED)
2888 new->v.nodes = nodes;
2890 new->v.preferred_node = first_node(nodes);
2892 new->flags |= MPOL_F_LOCAL;
2895 * Save nodes for contextualization: this will be used to "clone"
2896 * the mempolicy in a specific context [cpuset] at a later time.
2898 new->w.user_nodemask = nodes;
2903 /* Restore string for error message */
2912 #endif /* CONFIG_TMPFS */
2915 * mpol_to_str - format a mempolicy structure for printing
2916 * @buffer: to contain formatted mempolicy string
2917 * @maxlen: length of @buffer
2918 * @pol: pointer to mempolicy to be formatted
2920 * Convert @pol into a string. If @buffer is too short, truncate the string.
2921 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2922 * longest flag, "relative", and to display at least a few node ids.
2924 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2927 nodemask_t nodes = NODE_MASK_NONE;
2928 unsigned short mode = MPOL_DEFAULT;
2929 unsigned short flags = 0;
2931 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2939 case MPOL_PREFERRED:
2940 if (flags & MPOL_F_LOCAL)
2943 node_set(pol->v.preferred_node, nodes);
2946 case MPOL_INTERLEAVE:
2947 nodes = pol->v.nodes;
2951 snprintf(p, maxlen, "unknown");
2955 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2957 if (flags & MPOL_MODE_FLAGS) {
2958 p += snprintf(p, buffer + maxlen - p, "=");
2961 * Currently, the only defined flags are mutually exclusive
2963 if (flags & MPOL_F_STATIC_NODES)
2964 p += snprintf(p, buffer + maxlen - p, "static");
2965 else if (flags & MPOL_F_RELATIVE_NODES)
2966 p += snprintf(p, buffer + maxlen - p, "relative");
2969 if (!nodes_empty(nodes))
2970 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2971 nodemask_pr_args(&nodes));