2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
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>
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)
660 * Walk through page tables and collect pages to be migrated.
662 * If pages found in a given range are on a set of nodes (determined by
663 * @nodes and @flags,) it's isolated and queued to the pagelist which is
664 * passed via @private.
666 * queue_pages_range() has three possible return values:
667 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
669 * 0 - queue pages successfully or no misplaced page.
670 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
671 * memory range specified by nodemask and maxnode points outside
672 * your accessible address space (-EFAULT)
675 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
676 nodemask_t *nodes, unsigned long flags,
677 struct list_head *pagelist)
679 struct queue_pages qp = {
680 .pagelist = pagelist,
685 struct mm_walk queue_pages_walk = {
686 .hugetlb_entry = queue_pages_hugetlb,
687 .pmd_entry = queue_pages_pte_range,
688 .test_walk = queue_pages_test_walk,
693 return walk_page_range(start, end, &queue_pages_walk);
697 * Apply policy to a single VMA
698 * This must be called with the mmap_sem held for writing.
700 static int vma_replace_policy(struct vm_area_struct *vma,
701 struct mempolicy *pol)
704 struct mempolicy *old;
705 struct mempolicy *new;
707 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
708 vma->vm_start, vma->vm_end, vma->vm_pgoff,
709 vma->vm_ops, vma->vm_file,
710 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
716 if (vma->vm_ops && vma->vm_ops->set_policy) {
717 err = vma->vm_ops->set_policy(vma, new);
722 old = vma->vm_policy;
723 vma->vm_policy = new; /* protected by mmap_sem */
732 /* Step 2: apply policy to a range and do splits. */
733 static int mbind_range(struct mm_struct *mm, unsigned long start,
734 unsigned long end, struct mempolicy *new_pol)
736 struct vm_area_struct *prev;
737 struct vm_area_struct *vma;
740 unsigned long vmstart;
743 vma = find_vma(mm, start);
744 if (!vma || vma->vm_start > start)
748 if (start > vma->vm_start)
751 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
752 vmstart = max(start, vma->vm_start);
753 vmend = min(end, vma->vm_end);
755 if (mpol_equal(vma_policy(vma), new_pol))
758 pgoff = vma->vm_pgoff +
759 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
760 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
761 vma->anon_vma, vma->vm_file, pgoff,
762 new_pol, vma->vm_userfaultfd_ctx);
767 if (vma->vm_start != vmstart) {
768 err = split_vma(vma->vm_mm, vma, vmstart, 1);
772 if (vma->vm_end != vmend) {
773 err = split_vma(vma->vm_mm, vma, vmend, 0);
778 err = vma_replace_policy(vma, new_pol);
787 /* Set the process memory policy */
788 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
791 struct mempolicy *new, *old;
792 NODEMASK_SCRATCH(scratch);
798 new = mpol_new(mode, flags, nodes);
805 ret = mpol_set_nodemask(new, nodes, scratch);
807 task_unlock(current);
811 old = current->mempolicy;
812 current->mempolicy = new;
813 if (new && new->mode == MPOL_INTERLEAVE)
814 current->il_prev = MAX_NUMNODES-1;
815 task_unlock(current);
819 NODEMASK_SCRATCH_FREE(scratch);
824 * Return nodemask for policy for get_mempolicy() query
826 * Called with task's alloc_lock held
828 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
831 if (p == &default_policy)
837 case MPOL_INTERLEAVE:
841 if (!(p->flags & MPOL_F_LOCAL))
842 node_set(p->v.preferred_node, *nodes);
843 /* else return empty node mask for local allocation */
850 static int lookup_node(unsigned long addr)
855 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
857 err = page_to_nid(p);
863 /* Retrieve NUMA policy */
864 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
865 unsigned long addr, unsigned long flags)
868 struct mm_struct *mm = current->mm;
869 struct vm_area_struct *vma = NULL;
870 struct mempolicy *pol = current->mempolicy;
873 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
876 if (flags & MPOL_F_MEMS_ALLOWED) {
877 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
879 *policy = 0; /* just so it's initialized */
881 *nmask = cpuset_current_mems_allowed;
882 task_unlock(current);
886 if (flags & MPOL_F_ADDR) {
888 * Do NOT fall back to task policy if the
889 * vma/shared policy at addr is NULL. We
890 * want to return MPOL_DEFAULT in this case.
892 down_read(&mm->mmap_sem);
893 vma = find_vma_intersection(mm, addr, addr+1);
895 up_read(&mm->mmap_sem);
898 if (vma->vm_ops && vma->vm_ops->get_policy)
899 pol = vma->vm_ops->get_policy(vma, addr);
901 pol = vma->vm_policy;
906 pol = &default_policy; /* indicates default behavior */
908 if (flags & MPOL_F_NODE) {
909 if (flags & MPOL_F_ADDR) {
910 err = lookup_node(addr);
914 } else if (pol == current->mempolicy &&
915 pol->mode == MPOL_INTERLEAVE) {
916 *policy = next_node_in(current->il_prev, pol->v.nodes);
922 *policy = pol == &default_policy ? MPOL_DEFAULT :
925 * Internal mempolicy flags must be masked off before exposing
926 * the policy to userspace.
928 *policy |= (pol->flags & MPOL_MODE_FLAGS);
933 if (mpol_store_user_nodemask(pol)) {
934 *nmask = pol->w.user_nodemask;
937 get_policy_nodemask(pol, nmask);
938 task_unlock(current);
945 up_read(¤t->mm->mmap_sem);
949 #ifdef CONFIG_MIGRATION
951 * page migration, thp tail pages can be passed.
953 static int migrate_page_add(struct page *page, struct list_head *pagelist,
956 struct page *head = compound_head(page);
958 * Avoid migrating a page that is shared with others.
960 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
961 if (!isolate_lru_page(head)) {
962 list_add_tail(&head->lru, pagelist);
963 mod_node_page_state(page_pgdat(head),
964 NR_ISOLATED_ANON + page_is_file_cache(head),
965 hpage_nr_pages(head));
966 } else if (flags & MPOL_MF_STRICT) {
968 * Non-movable page may reach here. And, there may be
969 * temporary off LRU pages or non-LRU movable pages.
970 * Treat them as unmovable pages since they can't be
971 * isolated, so they can't be moved at the moment. It
972 * should return -EIO for this case too.
981 /* page allocation callback for NUMA node migration */
982 struct page *alloc_new_node_page(struct page *page, unsigned long node)
985 return alloc_huge_page_node(page_hstate(compound_head(page)),
987 else if (PageTransHuge(page)) {
990 thp = alloc_pages_node(node,
991 (GFP_TRANSHUGE | __GFP_THISNODE),
995 prep_transhuge_page(thp);
998 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1003 * Migrate pages from one node to a target node.
1004 * Returns error or the number of pages not migrated.
1006 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1010 LIST_HEAD(pagelist);
1014 node_set(source, nmask);
1017 * This does not "check" the range but isolates all pages that
1018 * need migration. Between passing in the full user address
1019 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1021 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1022 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1023 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1025 if (!list_empty(&pagelist)) {
1026 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1027 MIGRATE_SYNC, MR_SYSCALL);
1029 putback_movable_pages(&pagelist);
1036 * Move pages between the two nodesets so as to preserve the physical
1037 * layout as much as possible.
1039 * Returns the number of page that could not be moved.
1041 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1042 const nodemask_t *to, int flags)
1048 err = migrate_prep();
1052 down_read(&mm->mmap_sem);
1055 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1056 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1057 * bit in 'tmp', and return that <source, dest> pair for migration.
1058 * The pair of nodemasks 'to' and 'from' define the map.
1060 * If no pair of bits is found that way, fallback to picking some
1061 * pair of 'source' and 'dest' bits that are not the same. If the
1062 * 'source' and 'dest' bits are the same, this represents a node
1063 * that will be migrating to itself, so no pages need move.
1065 * If no bits are left in 'tmp', or if all remaining bits left
1066 * in 'tmp' correspond to the same bit in 'to', return false
1067 * (nothing left to migrate).
1069 * This lets us pick a pair of nodes to migrate between, such that
1070 * if possible the dest node is not already occupied by some other
1071 * source node, minimizing the risk of overloading the memory on a
1072 * node that would happen if we migrated incoming memory to a node
1073 * before migrating outgoing memory source that same node.
1075 * A single scan of tmp is sufficient. As we go, we remember the
1076 * most recent <s, d> pair that moved (s != d). If we find a pair
1077 * that not only moved, but what's better, moved to an empty slot
1078 * (d is not set in tmp), then we break out then, with that pair.
1079 * Otherwise when we finish scanning from_tmp, we at least have the
1080 * most recent <s, d> pair that moved. If we get all the way through
1081 * the scan of tmp without finding any node that moved, much less
1082 * moved to an empty node, then there is nothing left worth migrating.
1086 while (!nodes_empty(tmp)) {
1088 int source = NUMA_NO_NODE;
1091 for_each_node_mask(s, tmp) {
1094 * do_migrate_pages() tries to maintain the relative
1095 * node relationship of the pages established between
1096 * threads and memory areas.
1098 * However if the number of source nodes is not equal to
1099 * the number of destination nodes we can not preserve
1100 * this node relative relationship. In that case, skip
1101 * copying memory from a node that is in the destination
1104 * Example: [2,3,4] -> [3,4,5] moves everything.
1105 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1108 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1109 (node_isset(s, *to)))
1112 d = node_remap(s, *from, *to);
1116 source = s; /* Node moved. Memorize */
1119 /* dest not in remaining from nodes? */
1120 if (!node_isset(dest, tmp))
1123 if (source == NUMA_NO_NODE)
1126 node_clear(source, tmp);
1127 err = migrate_to_node(mm, source, dest, flags);
1133 up_read(&mm->mmap_sem);
1141 * Allocate a new page for page migration based on vma policy.
1142 * Start by assuming the page is mapped by the same vma as contains @start.
1143 * Search forward from there, if not. N.B., this assumes that the
1144 * list of pages handed to migrate_pages()--which is how we get here--
1145 * is in virtual address order.
1147 static struct page *new_page(struct page *page, unsigned long start)
1149 struct vm_area_struct *vma;
1150 unsigned long address;
1152 vma = find_vma(current->mm, start);
1154 address = page_address_in_vma(page, vma);
1155 if (address != -EFAULT)
1160 if (PageHuge(page)) {
1161 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1163 } else if (PageTransHuge(page)) {
1166 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1170 prep_transhuge_page(thp);
1174 * if !vma, alloc_page_vma() will use task or system default policy
1176 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1181 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1182 unsigned long flags)
1187 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1188 const nodemask_t *to, int flags)
1193 static struct page *new_page(struct page *page, unsigned long start)
1199 static long do_mbind(unsigned long start, unsigned long len,
1200 unsigned short mode, unsigned short mode_flags,
1201 nodemask_t *nmask, unsigned long flags)
1203 struct mm_struct *mm = current->mm;
1204 struct mempolicy *new;
1208 LIST_HEAD(pagelist);
1210 if (flags & ~(unsigned long)MPOL_MF_VALID)
1212 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1215 if (start & ~PAGE_MASK)
1218 if (mode == MPOL_DEFAULT)
1219 flags &= ~MPOL_MF_STRICT;
1221 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1229 new = mpol_new(mode, mode_flags, nmask);
1231 return PTR_ERR(new);
1233 if (flags & MPOL_MF_LAZY)
1234 new->flags |= MPOL_F_MOF;
1237 * If we are using the default policy then operation
1238 * on discontinuous address spaces is okay after all
1241 flags |= MPOL_MF_DISCONTIG_OK;
1243 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1244 start, start + len, mode, mode_flags,
1245 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1247 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1249 err = migrate_prep();
1254 NODEMASK_SCRATCH(scratch);
1256 down_write(&mm->mmap_sem);
1258 err = mpol_set_nodemask(new, nmask, scratch);
1259 task_unlock(current);
1261 up_write(&mm->mmap_sem);
1264 NODEMASK_SCRATCH_FREE(scratch);
1269 ret = queue_pages_range(mm, start, end, nmask,
1270 flags | MPOL_MF_INVERT, &pagelist);
1277 err = mbind_range(mm, start, end, new);
1282 if (!list_empty(&pagelist)) {
1283 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1284 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1285 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1287 putback_movable_pages(&pagelist);
1290 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1294 if (!list_empty(&pagelist))
1295 putback_movable_pages(&pagelist);
1298 up_write(&mm->mmap_sem);
1305 * User space interface with variable sized bitmaps for nodelists.
1308 /* Copy a node mask from user space. */
1309 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1310 unsigned long maxnode)
1314 unsigned long nlongs;
1315 unsigned long endmask;
1318 nodes_clear(*nodes);
1319 if (maxnode == 0 || !nmask)
1321 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1324 nlongs = BITS_TO_LONGS(maxnode);
1325 if ((maxnode % BITS_PER_LONG) == 0)
1328 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1331 * When the user specified more nodes than supported just check
1332 * if the non supported part is all zero.
1334 * If maxnode have more longs than MAX_NUMNODES, check
1335 * the bits in that area first. And then go through to
1336 * check the rest bits which equal or bigger than MAX_NUMNODES.
1337 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1339 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1340 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1341 if (get_user(t, nmask + k))
1343 if (k == nlongs - 1) {
1349 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1353 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1354 unsigned long valid_mask = endmask;
1356 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1357 if (get_user(t, nmask + nlongs - 1))
1363 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1365 nodes_addr(*nodes)[nlongs-1] &= endmask;
1369 /* Copy a kernel node mask to user space */
1370 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1373 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1374 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1376 if (copy > nbytes) {
1377 if (copy > PAGE_SIZE)
1379 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1383 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1386 static long kernel_mbind(unsigned long start, unsigned long len,
1387 unsigned long mode, const unsigned long __user *nmask,
1388 unsigned long maxnode, unsigned int flags)
1392 unsigned short mode_flags;
1394 mode_flags = mode & MPOL_MODE_FLAGS;
1395 mode &= ~MPOL_MODE_FLAGS;
1396 if (mode >= MPOL_MAX)
1398 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1399 (mode_flags & MPOL_F_RELATIVE_NODES))
1401 err = get_nodes(&nodes, nmask, maxnode);
1404 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1407 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1408 unsigned long, mode, const unsigned long __user *, nmask,
1409 unsigned long, maxnode, unsigned int, flags)
1411 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1414 /* Set the process memory policy */
1415 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1416 unsigned long maxnode)
1420 unsigned short flags;
1422 flags = mode & MPOL_MODE_FLAGS;
1423 mode &= ~MPOL_MODE_FLAGS;
1424 if ((unsigned int)mode >= MPOL_MAX)
1426 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1428 err = get_nodes(&nodes, nmask, maxnode);
1431 return do_set_mempolicy(mode, flags, &nodes);
1434 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1435 unsigned long, maxnode)
1437 return kernel_set_mempolicy(mode, nmask, maxnode);
1440 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1441 const unsigned long __user *old_nodes,
1442 const unsigned long __user *new_nodes)
1444 struct mm_struct *mm = NULL;
1445 struct task_struct *task;
1446 nodemask_t task_nodes;
1450 NODEMASK_SCRATCH(scratch);
1455 old = &scratch->mask1;
1456 new = &scratch->mask2;
1458 err = get_nodes(old, old_nodes, maxnode);
1462 err = get_nodes(new, new_nodes, maxnode);
1466 /* Find the mm_struct */
1468 task = pid ? find_task_by_vpid(pid) : current;
1474 get_task_struct(task);
1479 * Check if this process has the right to modify the specified process.
1480 * Use the regular "ptrace_may_access()" checks.
1482 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1489 task_nodes = cpuset_mems_allowed(task);
1490 /* Is the user allowed to access the target nodes? */
1491 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1496 task_nodes = cpuset_mems_allowed(current);
1497 nodes_and(*new, *new, task_nodes);
1498 if (nodes_empty(*new))
1501 nodes_and(*new, *new, node_states[N_MEMORY]);
1502 if (nodes_empty(*new))
1505 err = security_task_movememory(task);
1509 mm = get_task_mm(task);
1510 put_task_struct(task);
1517 err = do_migrate_pages(mm, old, new,
1518 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1522 NODEMASK_SCRATCH_FREE(scratch);
1527 put_task_struct(task);
1532 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1533 const unsigned long __user *, old_nodes,
1534 const unsigned long __user *, new_nodes)
1536 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1540 /* Retrieve NUMA policy */
1541 static int kernel_get_mempolicy(int __user *policy,
1542 unsigned long __user *nmask,
1543 unsigned long maxnode,
1545 unsigned long flags)
1551 if (nmask != NULL && maxnode < nr_node_ids)
1554 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1559 if (policy && put_user(pval, policy))
1563 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1568 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1569 unsigned long __user *, nmask, unsigned long, maxnode,
1570 unsigned long, addr, unsigned long, flags)
1572 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1575 #ifdef CONFIG_COMPAT
1577 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1578 compat_ulong_t __user *, nmask,
1579 compat_ulong_t, maxnode,
1580 compat_ulong_t, addr, compat_ulong_t, flags)
1583 unsigned long __user *nm = NULL;
1584 unsigned long nr_bits, alloc_size;
1585 DECLARE_BITMAP(bm, MAX_NUMNODES);
1587 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1588 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1591 nm = compat_alloc_user_space(alloc_size);
1593 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1595 if (!err && nmask) {
1596 unsigned long copy_size;
1597 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1598 err = copy_from_user(bm, nm, copy_size);
1599 /* ensure entire bitmap is zeroed */
1600 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1601 err |= compat_put_bitmap(nmask, bm, nr_bits);
1607 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1608 compat_ulong_t, maxnode)
1610 unsigned long __user *nm = NULL;
1611 unsigned long nr_bits, alloc_size;
1612 DECLARE_BITMAP(bm, MAX_NUMNODES);
1614 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1615 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1618 if (compat_get_bitmap(bm, nmask, nr_bits))
1620 nm = compat_alloc_user_space(alloc_size);
1621 if (copy_to_user(nm, bm, alloc_size))
1625 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1628 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1629 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1630 compat_ulong_t, maxnode, compat_ulong_t, flags)
1632 unsigned long __user *nm = NULL;
1633 unsigned long nr_bits, alloc_size;
1636 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1637 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1640 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1642 nm = compat_alloc_user_space(alloc_size);
1643 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1647 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1650 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1651 compat_ulong_t, maxnode,
1652 const compat_ulong_t __user *, old_nodes,
1653 const compat_ulong_t __user *, new_nodes)
1655 unsigned long __user *old = NULL;
1656 unsigned long __user *new = NULL;
1657 nodemask_t tmp_mask;
1658 unsigned long nr_bits;
1661 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1662 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1664 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1666 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1668 new = old + size / sizeof(unsigned long);
1669 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1673 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1676 new = compat_alloc_user_space(size);
1677 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1680 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1683 #endif /* CONFIG_COMPAT */
1685 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1688 struct mempolicy *pol = NULL;
1691 if (vma->vm_ops && vma->vm_ops->get_policy) {
1692 pol = vma->vm_ops->get_policy(vma, addr);
1693 } else if (vma->vm_policy) {
1694 pol = vma->vm_policy;
1697 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1698 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1699 * count on these policies which will be dropped by
1700 * mpol_cond_put() later
1702 if (mpol_needs_cond_ref(pol))
1711 * get_vma_policy(@vma, @addr)
1712 * @vma: virtual memory area whose policy is sought
1713 * @addr: address in @vma for shared policy lookup
1715 * Returns effective policy for a VMA at specified address.
1716 * Falls back to current->mempolicy or system default policy, as necessary.
1717 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1718 * count--added by the get_policy() vm_op, as appropriate--to protect against
1719 * freeing by another task. It is the caller's responsibility to free the
1720 * extra reference for shared policies.
1722 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1725 struct mempolicy *pol = __get_vma_policy(vma, addr);
1728 pol = get_task_policy(current);
1733 bool vma_policy_mof(struct vm_area_struct *vma)
1735 struct mempolicy *pol;
1737 if (vma->vm_ops && vma->vm_ops->get_policy) {
1740 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1741 if (pol && (pol->flags & MPOL_F_MOF))
1748 pol = vma->vm_policy;
1750 pol = get_task_policy(current);
1752 return pol->flags & MPOL_F_MOF;
1755 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1757 enum zone_type dynamic_policy_zone = policy_zone;
1759 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1762 * if policy->v.nodes has movable memory only,
1763 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1765 * policy->v.nodes is intersect with node_states[N_MEMORY].
1766 * so if the following test faile, it implies
1767 * policy->v.nodes has movable memory only.
1769 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1770 dynamic_policy_zone = ZONE_MOVABLE;
1772 return zone >= dynamic_policy_zone;
1776 * Return a nodemask representing a mempolicy for filtering nodes for
1779 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1781 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1782 if (unlikely(policy->mode == MPOL_BIND) &&
1783 apply_policy_zone(policy, gfp_zone(gfp)) &&
1784 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1785 return &policy->v.nodes;
1790 /* Return the node id preferred by the given mempolicy, or the given id */
1791 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1794 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1795 nd = policy->v.preferred_node;
1798 * __GFP_THISNODE shouldn't even be used with the bind policy
1799 * because we might easily break the expectation to stay on the
1800 * requested node and not break the policy.
1802 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1808 /* Do dynamic interleaving for a process */
1809 static unsigned interleave_nodes(struct mempolicy *policy)
1812 struct task_struct *me = current;
1814 next = next_node_in(me->il_prev, policy->v.nodes);
1815 if (next < MAX_NUMNODES)
1821 * Depending on the memory policy provide a node from which to allocate the
1824 unsigned int mempolicy_slab_node(void)
1826 struct mempolicy *policy;
1827 int node = numa_mem_id();
1832 policy = current->mempolicy;
1833 if (!policy || policy->flags & MPOL_F_LOCAL)
1836 switch (policy->mode) {
1837 case MPOL_PREFERRED:
1839 * handled MPOL_F_LOCAL above
1841 return policy->v.preferred_node;
1843 case MPOL_INTERLEAVE:
1844 return interleave_nodes(policy);
1850 * Follow bind policy behavior and start allocation at the
1853 struct zonelist *zonelist;
1854 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1855 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1856 z = first_zones_zonelist(zonelist, highest_zoneidx,
1858 return z->zone ? zone_to_nid(z->zone) : node;
1867 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1868 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1869 * number of present nodes.
1871 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1873 unsigned nnodes = nodes_weight(pol->v.nodes);
1879 return numa_node_id();
1880 target = (unsigned int)n % nnodes;
1881 nid = first_node(pol->v.nodes);
1882 for (i = 0; i < target; i++)
1883 nid = next_node(nid, pol->v.nodes);
1887 /* Determine a node number for interleave */
1888 static inline unsigned interleave_nid(struct mempolicy *pol,
1889 struct vm_area_struct *vma, unsigned long addr, int shift)
1895 * for small pages, there is no difference between
1896 * shift and PAGE_SHIFT, so the bit-shift is safe.
1897 * for huge pages, since vm_pgoff is in units of small
1898 * pages, we need to shift off the always 0 bits to get
1901 BUG_ON(shift < PAGE_SHIFT);
1902 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1903 off += (addr - vma->vm_start) >> shift;
1904 return offset_il_node(pol, off);
1906 return interleave_nodes(pol);
1909 #ifdef CONFIG_HUGETLBFS
1911 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1912 * @vma: virtual memory area whose policy is sought
1913 * @addr: address in @vma for shared policy lookup and interleave policy
1914 * @gfp_flags: for requested zone
1915 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1916 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1918 * Returns a nid suitable for a huge page allocation and a pointer
1919 * to the struct mempolicy for conditional unref after allocation.
1920 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1921 * @nodemask for filtering the zonelist.
1923 * Must be protected by read_mems_allowed_begin()
1925 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1926 struct mempolicy **mpol, nodemask_t **nodemask)
1930 *mpol = get_vma_policy(vma, addr);
1931 *nodemask = NULL; /* assume !MPOL_BIND */
1933 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1934 nid = interleave_nid(*mpol, vma, addr,
1935 huge_page_shift(hstate_vma(vma)));
1937 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1938 if ((*mpol)->mode == MPOL_BIND)
1939 *nodemask = &(*mpol)->v.nodes;
1945 * init_nodemask_of_mempolicy
1947 * If the current task's mempolicy is "default" [NULL], return 'false'
1948 * to indicate default policy. Otherwise, extract the policy nodemask
1949 * for 'bind' or 'interleave' policy into the argument nodemask, or
1950 * initialize the argument nodemask to contain the single node for
1951 * 'preferred' or 'local' policy and return 'true' to indicate presence
1952 * of non-default mempolicy.
1954 * We don't bother with reference counting the mempolicy [mpol_get/put]
1955 * because the current task is examining it's own mempolicy and a task's
1956 * mempolicy is only ever changed by the task itself.
1958 * N.B., it is the caller's responsibility to free a returned nodemask.
1960 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1962 struct mempolicy *mempolicy;
1965 if (!(mask && current->mempolicy))
1969 mempolicy = current->mempolicy;
1970 switch (mempolicy->mode) {
1971 case MPOL_PREFERRED:
1972 if (mempolicy->flags & MPOL_F_LOCAL)
1973 nid = numa_node_id();
1975 nid = mempolicy->v.preferred_node;
1976 init_nodemask_of_node(mask, nid);
1981 case MPOL_INTERLEAVE:
1982 *mask = mempolicy->v.nodes;
1988 task_unlock(current);
1995 * mempolicy_nodemask_intersects
1997 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1998 * policy. Otherwise, check for intersection between mask and the policy
1999 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2000 * policy, always return true since it may allocate elsewhere on fallback.
2002 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2004 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2005 const nodemask_t *mask)
2007 struct mempolicy *mempolicy;
2013 mempolicy = tsk->mempolicy;
2017 switch (mempolicy->mode) {
2018 case MPOL_PREFERRED:
2020 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2021 * allocate from, they may fallback to other nodes when oom.
2022 * Thus, it's possible for tsk to have allocated memory from
2027 case MPOL_INTERLEAVE:
2028 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2038 /* Allocate a page in interleaved policy.
2039 Own path because it needs to do special accounting. */
2040 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2045 page = __alloc_pages(gfp, order, nid);
2046 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2047 if (!static_branch_likely(&vm_numa_stat_key))
2049 if (page && page_to_nid(page) == nid) {
2051 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2058 * alloc_pages_vma - Allocate a page for a VMA.
2061 * %GFP_USER user allocation.
2062 * %GFP_KERNEL kernel allocations,
2063 * %GFP_HIGHMEM highmem/user allocations,
2064 * %GFP_FS allocation should not call back into a file system.
2065 * %GFP_ATOMIC don't sleep.
2067 * @order:Order of the GFP allocation.
2068 * @vma: Pointer to VMA or NULL if not available.
2069 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2070 * @node: Which node to prefer for allocation (modulo policy).
2071 * @hugepage: for hugepages try only the preferred node if possible
2073 * This function allocates a page from the kernel page pool and applies
2074 * a NUMA policy associated with the VMA or the current process.
2075 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2076 * mm_struct of the VMA to prevent it from going away. Should be used for
2077 * all allocations for pages that will be mapped into user space. Returns
2078 * NULL when no page can be allocated.
2081 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2082 unsigned long addr, int node, bool hugepage)
2084 struct mempolicy *pol;
2089 pol = get_vma_policy(vma, addr);
2091 if (pol->mode == MPOL_INTERLEAVE) {
2094 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2096 page = alloc_page_interleave(gfp, order, nid);
2100 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2101 int hpage_node = node;
2104 * For hugepage allocation and non-interleave policy which
2105 * allows the current node (or other explicitly preferred
2106 * node) we only try to allocate from the current/preferred
2107 * node and don't fall back to other nodes, as the cost of
2108 * remote accesses would likely offset THP benefits.
2110 * If the policy is interleave, or does not allow the current
2111 * node in its nodemask, we allocate the standard way.
2113 if (pol->mode == MPOL_PREFERRED &&
2114 !(pol->flags & MPOL_F_LOCAL))
2115 hpage_node = pol->v.preferred_node;
2117 nmask = policy_nodemask(gfp, pol);
2118 if (!nmask || node_isset(hpage_node, *nmask)) {
2121 * We cannot invoke reclaim if __GFP_THISNODE
2122 * is set. Invoking reclaim with
2123 * __GFP_THISNODE set, would cause THP
2124 * allocations to trigger heavy swapping
2125 * despite there may be tons of free memory
2126 * (including potentially plenty of THP
2127 * already available in the buddy) on all the
2130 * At most we could invoke compaction when
2131 * __GFP_THISNODE is set (but we would need to
2132 * refrain from invoking reclaim even if
2133 * compaction returned COMPACT_SKIPPED because
2134 * there wasn't not enough memory to succeed
2135 * compaction). For now just avoid
2136 * __GFP_THISNODE instead of limiting the
2137 * allocation path to a strict and single
2138 * compaction invocation.
2140 * Supposedly if direct reclaim was enabled by
2141 * the caller, the app prefers THP regardless
2142 * of the node it comes from so this would be
2143 * more desiderable behavior than only
2144 * providing THP originated from the local
2145 * node in such case.
2147 if (!(gfp & __GFP_DIRECT_RECLAIM))
2148 gfp |= __GFP_THISNODE;
2149 page = __alloc_pages_node(hpage_node, gfp, order);
2154 nmask = policy_nodemask(gfp, pol);
2155 preferred_nid = policy_node(gfp, pol, node);
2156 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2163 * alloc_pages_current - Allocate pages.
2166 * %GFP_USER user allocation,
2167 * %GFP_KERNEL kernel allocation,
2168 * %GFP_HIGHMEM highmem allocation,
2169 * %GFP_FS don't call back into a file system.
2170 * %GFP_ATOMIC don't sleep.
2171 * @order: Power of two of allocation size in pages. 0 is a single page.
2173 * Allocate a page from the kernel page pool. When not in
2174 * interrupt context and apply the current process NUMA policy.
2175 * Returns NULL when no page can be allocated.
2177 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2179 struct mempolicy *pol = &default_policy;
2182 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2183 pol = get_task_policy(current);
2186 * No reference counting needed for current->mempolicy
2187 * nor system default_policy
2189 if (pol->mode == MPOL_INTERLEAVE)
2190 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2192 page = __alloc_pages_nodemask(gfp, order,
2193 policy_node(gfp, pol, numa_node_id()),
2194 policy_nodemask(gfp, pol));
2198 EXPORT_SYMBOL(alloc_pages_current);
2200 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2202 struct mempolicy *pol = mpol_dup(vma_policy(src));
2205 return PTR_ERR(pol);
2206 dst->vm_policy = pol;
2211 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2212 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2213 * with the mems_allowed returned by cpuset_mems_allowed(). This
2214 * keeps mempolicies cpuset relative after its cpuset moves. See
2215 * further kernel/cpuset.c update_nodemask().
2217 * current's mempolicy may be rebinded by the other task(the task that changes
2218 * cpuset's mems), so we needn't do rebind work for current task.
2221 /* Slow path of a mempolicy duplicate */
2222 struct mempolicy *__mpol_dup(struct mempolicy *old)
2224 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2227 return ERR_PTR(-ENOMEM);
2229 /* task's mempolicy is protected by alloc_lock */
2230 if (old == current->mempolicy) {
2233 task_unlock(current);
2237 if (current_cpuset_is_being_rebound()) {
2238 nodemask_t mems = cpuset_mems_allowed(current);
2239 mpol_rebind_policy(new, &mems);
2241 atomic_set(&new->refcnt, 1);
2245 /* Slow path of a mempolicy comparison */
2246 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2250 if (a->mode != b->mode)
2252 if (a->flags != b->flags)
2254 if (mpol_store_user_nodemask(a))
2255 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2261 case MPOL_INTERLEAVE:
2262 return !!nodes_equal(a->v.nodes, b->v.nodes);
2263 case MPOL_PREFERRED:
2264 /* a's ->flags is the same as b's */
2265 if (a->flags & MPOL_F_LOCAL)
2267 return a->v.preferred_node == b->v.preferred_node;
2275 * Shared memory backing store policy support.
2277 * Remember policies even when nobody has shared memory mapped.
2278 * The policies are kept in Red-Black tree linked from the inode.
2279 * They are protected by the sp->lock rwlock, which should be held
2280 * for any accesses to the tree.
2284 * lookup first element intersecting start-end. Caller holds sp->lock for
2285 * reading or for writing
2287 static struct sp_node *
2288 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2290 struct rb_node *n = sp->root.rb_node;
2293 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2295 if (start >= p->end)
2297 else if (end <= p->start)
2305 struct sp_node *w = NULL;
2306 struct rb_node *prev = rb_prev(n);
2309 w = rb_entry(prev, struct sp_node, nd);
2310 if (w->end <= start)
2314 return rb_entry(n, struct sp_node, nd);
2318 * Insert a new shared policy into the list. Caller holds sp->lock for
2321 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2323 struct rb_node **p = &sp->root.rb_node;
2324 struct rb_node *parent = NULL;
2329 nd = rb_entry(parent, struct sp_node, nd);
2330 if (new->start < nd->start)
2332 else if (new->end > nd->end)
2333 p = &(*p)->rb_right;
2337 rb_link_node(&new->nd, parent, p);
2338 rb_insert_color(&new->nd, &sp->root);
2339 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2340 new->policy ? new->policy->mode : 0);
2343 /* Find shared policy intersecting idx */
2345 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2347 struct mempolicy *pol = NULL;
2350 if (!sp->root.rb_node)
2352 read_lock(&sp->lock);
2353 sn = sp_lookup(sp, idx, idx+1);
2355 mpol_get(sn->policy);
2358 read_unlock(&sp->lock);
2362 static void sp_free(struct sp_node *n)
2364 mpol_put(n->policy);
2365 kmem_cache_free(sn_cache, n);
2369 * mpol_misplaced - check whether current page node is valid in policy
2371 * @page: page to be checked
2372 * @vma: vm area where page mapped
2373 * @addr: virtual address where page mapped
2375 * Lookup current policy node id for vma,addr and "compare to" page's
2379 * -1 - not misplaced, page is in the right node
2380 * node - node id where the page should be
2382 * Policy determination "mimics" alloc_page_vma().
2383 * Called from fault path where we know the vma and faulting address.
2385 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2387 struct mempolicy *pol;
2389 int curnid = page_to_nid(page);
2390 unsigned long pgoff;
2391 int thiscpu = raw_smp_processor_id();
2392 int thisnid = cpu_to_node(thiscpu);
2396 pol = get_vma_policy(vma, addr);
2397 if (!(pol->flags & MPOL_F_MOF))
2400 switch (pol->mode) {
2401 case MPOL_INTERLEAVE:
2402 pgoff = vma->vm_pgoff;
2403 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2404 polnid = offset_il_node(pol, pgoff);
2407 case MPOL_PREFERRED:
2408 if (pol->flags & MPOL_F_LOCAL)
2409 polnid = numa_node_id();
2411 polnid = pol->v.preferred_node;
2417 * allows binding to multiple nodes.
2418 * use current page if in policy nodemask,
2419 * else select nearest allowed node, if any.
2420 * If no allowed nodes, use current [!misplaced].
2422 if (node_isset(curnid, pol->v.nodes))
2424 z = first_zones_zonelist(
2425 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2426 gfp_zone(GFP_HIGHUSER),
2428 polnid = zone_to_nid(z->zone);
2435 /* Migrate the page towards the node whose CPU is referencing it */
2436 if (pol->flags & MPOL_F_MORON) {
2439 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2443 if (curnid != polnid)
2452 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2453 * dropped after task->mempolicy is set to NULL so that any allocation done as
2454 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2457 void mpol_put_task_policy(struct task_struct *task)
2459 struct mempolicy *pol;
2462 pol = task->mempolicy;
2463 task->mempolicy = NULL;
2468 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2470 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2471 rb_erase(&n->nd, &sp->root);
2475 static void sp_node_init(struct sp_node *node, unsigned long start,
2476 unsigned long end, struct mempolicy *pol)
2478 node->start = start;
2483 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2484 struct mempolicy *pol)
2487 struct mempolicy *newpol;
2489 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2493 newpol = mpol_dup(pol);
2494 if (IS_ERR(newpol)) {
2495 kmem_cache_free(sn_cache, n);
2498 newpol->flags |= MPOL_F_SHARED;
2499 sp_node_init(n, start, end, newpol);
2504 /* Replace a policy range. */
2505 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2506 unsigned long end, struct sp_node *new)
2509 struct sp_node *n_new = NULL;
2510 struct mempolicy *mpol_new = NULL;
2514 write_lock(&sp->lock);
2515 n = sp_lookup(sp, start, end);
2516 /* Take care of old policies in the same range. */
2517 while (n && n->start < end) {
2518 struct rb_node *next = rb_next(&n->nd);
2519 if (n->start >= start) {
2525 /* Old policy spanning whole new range. */
2530 *mpol_new = *n->policy;
2531 atomic_set(&mpol_new->refcnt, 1);
2532 sp_node_init(n_new, end, n->end, mpol_new);
2534 sp_insert(sp, n_new);
2543 n = rb_entry(next, struct sp_node, nd);
2547 write_unlock(&sp->lock);
2554 kmem_cache_free(sn_cache, n_new);
2559 write_unlock(&sp->lock);
2561 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2564 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2567 atomic_set(&mpol_new->refcnt, 1);
2572 * mpol_shared_policy_init - initialize shared policy for inode
2573 * @sp: pointer to inode shared policy
2574 * @mpol: struct mempolicy to install
2576 * Install non-NULL @mpol in inode's shared policy rb-tree.
2577 * On entry, the current task has a reference on a non-NULL @mpol.
2578 * This must be released on exit.
2579 * This is called at get_inode() calls and we can use GFP_KERNEL.
2581 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2585 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2586 rwlock_init(&sp->lock);
2589 struct vm_area_struct pvma;
2590 struct mempolicy *new;
2591 NODEMASK_SCRATCH(scratch);
2595 /* contextualize the tmpfs mount point mempolicy */
2596 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2598 goto free_scratch; /* no valid nodemask intersection */
2601 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2602 task_unlock(current);
2606 /* Create pseudo-vma that contains just the policy */
2607 vma_init(&pvma, NULL);
2608 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2609 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2612 mpol_put(new); /* drop initial ref */
2614 NODEMASK_SCRATCH_FREE(scratch);
2616 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2620 int mpol_set_shared_policy(struct shared_policy *info,
2621 struct vm_area_struct *vma, struct mempolicy *npol)
2624 struct sp_node *new = NULL;
2625 unsigned long sz = vma_pages(vma);
2627 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2629 sz, npol ? npol->mode : -1,
2630 npol ? npol->flags : -1,
2631 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2634 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2638 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2644 /* Free a backing policy store on inode delete. */
2645 void mpol_free_shared_policy(struct shared_policy *p)
2648 struct rb_node *next;
2650 if (!p->root.rb_node)
2652 write_lock(&p->lock);
2653 next = rb_first(&p->root);
2655 n = rb_entry(next, struct sp_node, nd);
2656 next = rb_next(&n->nd);
2659 write_unlock(&p->lock);
2662 #ifdef CONFIG_NUMA_BALANCING
2663 static int __initdata numabalancing_override;
2665 static void __init check_numabalancing_enable(void)
2667 bool numabalancing_default = false;
2669 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2670 numabalancing_default = true;
2672 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2673 if (numabalancing_override)
2674 set_numabalancing_state(numabalancing_override == 1);
2676 if (num_online_nodes() > 1 && !numabalancing_override) {
2677 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2678 numabalancing_default ? "Enabling" : "Disabling");
2679 set_numabalancing_state(numabalancing_default);
2683 static int __init setup_numabalancing(char *str)
2689 if (!strcmp(str, "enable")) {
2690 numabalancing_override = 1;
2692 } else if (!strcmp(str, "disable")) {
2693 numabalancing_override = -1;
2698 pr_warn("Unable to parse numa_balancing=\n");
2702 __setup("numa_balancing=", setup_numabalancing);
2704 static inline void __init check_numabalancing_enable(void)
2707 #endif /* CONFIG_NUMA_BALANCING */
2709 /* assumes fs == KERNEL_DS */
2710 void __init numa_policy_init(void)
2712 nodemask_t interleave_nodes;
2713 unsigned long largest = 0;
2714 int nid, prefer = 0;
2716 policy_cache = kmem_cache_create("numa_policy",
2717 sizeof(struct mempolicy),
2718 0, SLAB_PANIC, NULL);
2720 sn_cache = kmem_cache_create("shared_policy_node",
2721 sizeof(struct sp_node),
2722 0, SLAB_PANIC, NULL);
2724 for_each_node(nid) {
2725 preferred_node_policy[nid] = (struct mempolicy) {
2726 .refcnt = ATOMIC_INIT(1),
2727 .mode = MPOL_PREFERRED,
2728 .flags = MPOL_F_MOF | MPOL_F_MORON,
2729 .v = { .preferred_node = nid, },
2734 * Set interleaving policy for system init. Interleaving is only
2735 * enabled across suitably sized nodes (default is >= 16MB), or
2736 * fall back to the largest node if they're all smaller.
2738 nodes_clear(interleave_nodes);
2739 for_each_node_state(nid, N_MEMORY) {
2740 unsigned long total_pages = node_present_pages(nid);
2742 /* Preserve the largest node */
2743 if (largest < total_pages) {
2744 largest = total_pages;
2748 /* Interleave this node? */
2749 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2750 node_set(nid, interleave_nodes);
2753 /* All too small, use the largest */
2754 if (unlikely(nodes_empty(interleave_nodes)))
2755 node_set(prefer, interleave_nodes);
2757 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2758 pr_err("%s: interleaving failed\n", __func__);
2760 check_numabalancing_enable();
2763 /* Reset policy of current process to default */
2764 void numa_default_policy(void)
2766 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2770 * Parse and format mempolicy from/to strings
2774 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2776 static const char * const policy_modes[] =
2778 [MPOL_DEFAULT] = "default",
2779 [MPOL_PREFERRED] = "prefer",
2780 [MPOL_BIND] = "bind",
2781 [MPOL_INTERLEAVE] = "interleave",
2782 [MPOL_LOCAL] = "local",
2788 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2789 * @str: string containing mempolicy to parse
2790 * @mpol: pointer to struct mempolicy pointer, returned on success.
2793 * <mode>[=<flags>][:<nodelist>]
2795 * On success, returns 0, else 1
2797 int mpol_parse_str(char *str, struct mempolicy **mpol)
2799 struct mempolicy *new = NULL;
2800 unsigned short mode;
2801 unsigned short mode_flags;
2803 char *nodelist = strchr(str, ':');
2804 char *flags = strchr(str, '=');
2808 *flags++ = '\0'; /* terminate mode string */
2811 /* NUL-terminate mode or flags string */
2813 if (nodelist_parse(nodelist, nodes))
2815 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2820 for (mode = 0; mode < MPOL_MAX; mode++) {
2821 if (!strcmp(str, policy_modes[mode])) {
2825 if (mode >= MPOL_MAX)
2829 case MPOL_PREFERRED:
2831 * Insist on a nodelist of one node only, although later
2832 * we use first_node(nodes) to grab a single node, so here
2833 * nodelist (or nodes) cannot be empty.
2836 char *rest = nodelist;
2837 while (isdigit(*rest))
2841 if (nodes_empty(nodes))
2845 case MPOL_INTERLEAVE:
2847 * Default to online nodes with memory if no nodelist
2850 nodes = node_states[N_MEMORY];
2854 * Don't allow a nodelist; mpol_new() checks flags
2858 mode = MPOL_PREFERRED;
2862 * Insist on a empty nodelist
2869 * Insist on a nodelist
2878 * Currently, we only support two mutually exclusive
2881 if (!strcmp(flags, "static"))
2882 mode_flags |= MPOL_F_STATIC_NODES;
2883 else if (!strcmp(flags, "relative"))
2884 mode_flags |= MPOL_F_RELATIVE_NODES;
2889 new = mpol_new(mode, mode_flags, &nodes);
2894 * Save nodes for mpol_to_str() to show the tmpfs mount options
2895 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2897 if (mode != MPOL_PREFERRED)
2898 new->v.nodes = nodes;
2900 new->v.preferred_node = first_node(nodes);
2902 new->flags |= MPOL_F_LOCAL;
2905 * Save nodes for contextualization: this will be used to "clone"
2906 * the mempolicy in a specific context [cpuset] at a later time.
2908 new->w.user_nodemask = nodes;
2913 /* Restore string for error message */
2922 #endif /* CONFIG_TMPFS */
2925 * mpol_to_str - format a mempolicy structure for printing
2926 * @buffer: to contain formatted mempolicy string
2927 * @maxlen: length of @buffer
2928 * @pol: pointer to mempolicy to be formatted
2930 * Convert @pol into a string. If @buffer is too short, truncate the string.
2931 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2932 * longest flag, "relative", and to display at least a few node ids.
2934 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2937 nodemask_t nodes = NODE_MASK_NONE;
2938 unsigned short mode = MPOL_DEFAULT;
2939 unsigned short flags = 0;
2941 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2949 case MPOL_PREFERRED:
2950 if (flags & MPOL_F_LOCAL)
2953 node_set(pol->v.preferred_node, nodes);
2956 case MPOL_INTERLEAVE:
2957 nodes = pol->v.nodes;
2961 snprintf(p, maxlen, "unknown");
2965 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2967 if (flags & MPOL_MODE_FLAGS) {
2968 p += snprintf(p, buffer + maxlen - p, "=");
2971 * Currently, the only defined flags are mutually exclusive
2973 if (flags & MPOL_F_STATIC_NODES)
2974 p += snprintf(p, buffer + maxlen - p, "static");
2975 else if (flags & MPOL_F_RELATIVE_NODES)
2976 p += snprintf(p, buffer + maxlen - p, "relative");
2979 if (!nodes_empty(nodes))
2980 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2981 nodemask_pr_args(&nodes));