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)
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)
659 * Walk through page tables and collect pages to be migrated.
661 * If pages found in a given range are on a set of nodes (determined by
662 * @nodes and @flags,) it's isolated and queued to the pagelist which is
663 * passed via @private.
665 * queue_pages_range() has three possible return values:
666 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
668 * 0 - queue pages successfully or no misplaced page.
669 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
670 * memory range specified by nodemask and maxnode points outside
671 * your accessible address space (-EFAULT)
674 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
675 nodemask_t *nodes, unsigned long flags,
676 struct list_head *pagelist)
678 struct queue_pages qp = {
679 .pagelist = pagelist,
684 struct mm_walk queue_pages_walk = {
685 .hugetlb_entry = queue_pages_hugetlb,
686 .pmd_entry = queue_pages_pte_range,
687 .test_walk = queue_pages_test_walk,
692 return walk_page_range(start, end, &queue_pages_walk);
696 * Apply policy to a single VMA
697 * This must be called with the mmap_sem held for writing.
699 static int vma_replace_policy(struct vm_area_struct *vma,
700 struct mempolicy *pol)
703 struct mempolicy *old;
704 struct mempolicy *new;
706 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
707 vma->vm_start, vma->vm_end, vma->vm_pgoff,
708 vma->vm_ops, vma->vm_file,
709 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
715 if (vma->vm_ops && vma->vm_ops->set_policy) {
716 err = vma->vm_ops->set_policy(vma, new);
721 old = vma->vm_policy;
722 vma->vm_policy = new; /* protected by mmap_sem */
731 /* Step 2: apply policy to a range and do splits. */
732 static int mbind_range(struct mm_struct *mm, unsigned long start,
733 unsigned long end, struct mempolicy *new_pol)
735 struct vm_area_struct *prev;
736 struct vm_area_struct *vma;
739 unsigned long vmstart;
742 vma = find_vma(mm, start);
743 if (!vma || vma->vm_start > start)
747 if (start > vma->vm_start)
750 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
751 vmstart = max(start, vma->vm_start);
752 vmend = min(end, vma->vm_end);
754 if (mpol_equal(vma_policy(vma), new_pol))
757 pgoff = vma->vm_pgoff +
758 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
759 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
760 vma->anon_vma, vma->vm_file, pgoff,
761 new_pol, vma->vm_userfaultfd_ctx);
766 if (vma->vm_start != vmstart) {
767 err = split_vma(vma->vm_mm, vma, vmstart, 1);
771 if (vma->vm_end != vmend) {
772 err = split_vma(vma->vm_mm, vma, vmend, 0);
777 err = vma_replace_policy(vma, new_pol);
786 /* Set the process memory policy */
787 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
790 struct mempolicy *new, *old;
791 NODEMASK_SCRATCH(scratch);
797 new = mpol_new(mode, flags, nodes);
804 ret = mpol_set_nodemask(new, nodes, scratch);
806 task_unlock(current);
810 old = current->mempolicy;
811 current->mempolicy = new;
812 if (new && new->mode == MPOL_INTERLEAVE)
813 current->il_prev = MAX_NUMNODES-1;
814 task_unlock(current);
818 NODEMASK_SCRATCH_FREE(scratch);
823 * Return nodemask for policy for get_mempolicy() query
825 * Called with task's alloc_lock held
827 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
830 if (p == &default_policy)
836 case MPOL_INTERLEAVE:
840 if (!(p->flags & MPOL_F_LOCAL))
841 node_set(p->v.preferred_node, *nodes);
842 /* else return empty node mask for local allocation */
849 static int lookup_node(unsigned long addr)
854 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
856 err = page_to_nid(p);
862 /* Retrieve NUMA policy */
863 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
864 unsigned long addr, unsigned long flags)
867 struct mm_struct *mm = current->mm;
868 struct vm_area_struct *vma = NULL;
869 struct mempolicy *pol = current->mempolicy;
872 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
875 if (flags & MPOL_F_MEMS_ALLOWED) {
876 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
878 *policy = 0; /* just so it's initialized */
880 *nmask = cpuset_current_mems_allowed;
881 task_unlock(current);
885 if (flags & MPOL_F_ADDR) {
887 * Do NOT fall back to task policy if the
888 * vma/shared policy at addr is NULL. We
889 * want to return MPOL_DEFAULT in this case.
891 down_read(&mm->mmap_sem);
892 vma = find_vma_intersection(mm, addr, addr+1);
894 up_read(&mm->mmap_sem);
897 if (vma->vm_ops && vma->vm_ops->get_policy)
898 pol = vma->vm_ops->get_policy(vma, addr);
900 pol = vma->vm_policy;
905 pol = &default_policy; /* indicates default behavior */
907 if (flags & MPOL_F_NODE) {
908 if (flags & MPOL_F_ADDR) {
909 err = lookup_node(addr);
913 } else if (pol == current->mempolicy &&
914 pol->mode == MPOL_INTERLEAVE) {
915 *policy = next_node_in(current->il_prev, pol->v.nodes);
921 *policy = pol == &default_policy ? MPOL_DEFAULT :
924 * Internal mempolicy flags must be masked off before exposing
925 * the policy to userspace.
927 *policy |= (pol->flags & MPOL_MODE_FLAGS);
932 if (mpol_store_user_nodemask(pol)) {
933 *nmask = pol->w.user_nodemask;
936 get_policy_nodemask(pol, nmask);
937 task_unlock(current);
944 up_read(¤t->mm->mmap_sem);
948 #ifdef CONFIG_MIGRATION
950 * page migration, thp tail pages can be passed.
952 static int migrate_page_add(struct page *page, struct list_head *pagelist,
955 struct page *head = compound_head(page);
957 * Avoid migrating a page that is shared with others.
959 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
960 if (!isolate_lru_page(head)) {
961 list_add_tail(&head->lru, pagelist);
962 mod_node_page_state(page_pgdat(head),
963 NR_ISOLATED_ANON + page_is_file_cache(head),
964 hpage_nr_pages(head));
965 } else if (flags & MPOL_MF_STRICT) {
967 * Non-movable page may reach here. And, there may be
968 * temporary off LRU pages or non-LRU movable pages.
969 * Treat them as unmovable pages since they can't be
970 * isolated, so they can't be moved at the moment. It
971 * should return -EIO for this case too.
980 /* page allocation callback for NUMA node migration */
981 struct page *alloc_new_node_page(struct page *page, unsigned long node)
984 return alloc_huge_page_node(page_hstate(compound_head(page)),
986 else if (PageTransHuge(page)) {
989 thp = alloc_pages_node(node,
990 (GFP_TRANSHUGE | __GFP_THISNODE),
994 prep_transhuge_page(thp);
997 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1002 * Migrate pages from one node to a target node.
1003 * Returns error or the number of pages not migrated.
1005 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1009 LIST_HEAD(pagelist);
1013 node_set(source, nmask);
1016 * This does not "check" the range but isolates all pages that
1017 * need migration. Between passing in the full user address
1018 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1020 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1021 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1022 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1024 if (!list_empty(&pagelist)) {
1025 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1026 MIGRATE_SYNC, MR_SYSCALL);
1028 putback_movable_pages(&pagelist);
1035 * Move pages between the two nodesets so as to preserve the physical
1036 * layout as much as possible.
1038 * Returns the number of page that could not be moved.
1040 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1041 const nodemask_t *to, int flags)
1047 err = migrate_prep();
1051 down_read(&mm->mmap_sem);
1054 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1055 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1056 * bit in 'tmp', and return that <source, dest> pair for migration.
1057 * The pair of nodemasks 'to' and 'from' define the map.
1059 * If no pair of bits is found that way, fallback to picking some
1060 * pair of 'source' and 'dest' bits that are not the same. If the
1061 * 'source' and 'dest' bits are the same, this represents a node
1062 * that will be migrating to itself, so no pages need move.
1064 * If no bits are left in 'tmp', or if all remaining bits left
1065 * in 'tmp' correspond to the same bit in 'to', return false
1066 * (nothing left to migrate).
1068 * This lets us pick a pair of nodes to migrate between, such that
1069 * if possible the dest node is not already occupied by some other
1070 * source node, minimizing the risk of overloading the memory on a
1071 * node that would happen if we migrated incoming memory to a node
1072 * before migrating outgoing memory source that same node.
1074 * A single scan of tmp is sufficient. As we go, we remember the
1075 * most recent <s, d> pair that moved (s != d). If we find a pair
1076 * that not only moved, but what's better, moved to an empty slot
1077 * (d is not set in tmp), then we break out then, with that pair.
1078 * Otherwise when we finish scanning from_tmp, we at least have the
1079 * most recent <s, d> pair that moved. If we get all the way through
1080 * the scan of tmp without finding any node that moved, much less
1081 * moved to an empty node, then there is nothing left worth migrating.
1085 while (!nodes_empty(tmp)) {
1087 int source = NUMA_NO_NODE;
1090 for_each_node_mask(s, tmp) {
1093 * do_migrate_pages() tries to maintain the relative
1094 * node relationship of the pages established between
1095 * threads and memory areas.
1097 * However if the number of source nodes is not equal to
1098 * the number of destination nodes we can not preserve
1099 * this node relative relationship. In that case, skip
1100 * copying memory from a node that is in the destination
1103 * Example: [2,3,4] -> [3,4,5] moves everything.
1104 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1107 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1108 (node_isset(s, *to)))
1111 d = node_remap(s, *from, *to);
1115 source = s; /* Node moved. Memorize */
1118 /* dest not in remaining from nodes? */
1119 if (!node_isset(dest, tmp))
1122 if (source == NUMA_NO_NODE)
1125 node_clear(source, tmp);
1126 err = migrate_to_node(mm, source, dest, flags);
1132 up_read(&mm->mmap_sem);
1140 * Allocate a new page for page migration based on vma policy.
1141 * Start by assuming the page is mapped by the same vma as contains @start.
1142 * Search forward from there, if not. N.B., this assumes that the
1143 * list of pages handed to migrate_pages()--which is how we get here--
1144 * is in virtual address order.
1146 static struct page *new_page(struct page *page, unsigned long start)
1148 struct vm_area_struct *vma;
1149 unsigned long uninitialized_var(address);
1151 vma = find_vma(current->mm, start);
1153 address = page_address_in_vma(page, vma);
1154 if (address != -EFAULT)
1159 if (PageHuge(page)) {
1160 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1162 } else if (PageTransHuge(page)) {
1165 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1169 prep_transhuge_page(thp);
1173 * if !vma, alloc_page_vma() will use task or system default policy
1175 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1180 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1181 unsigned long flags)
1186 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1187 const nodemask_t *to, int flags)
1192 static struct page *new_page(struct page *page, unsigned long start)
1198 static long do_mbind(unsigned long start, unsigned long len,
1199 unsigned short mode, unsigned short mode_flags,
1200 nodemask_t *nmask, unsigned long flags)
1202 struct mm_struct *mm = current->mm;
1203 struct mempolicy *new;
1207 LIST_HEAD(pagelist);
1209 if (flags & ~(unsigned long)MPOL_MF_VALID)
1211 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1214 if (start & ~PAGE_MASK)
1217 if (mode == MPOL_DEFAULT)
1218 flags &= ~MPOL_MF_STRICT;
1220 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1228 new = mpol_new(mode, mode_flags, nmask);
1230 return PTR_ERR(new);
1232 if (flags & MPOL_MF_LAZY)
1233 new->flags |= MPOL_F_MOF;
1236 * If we are using the default policy then operation
1237 * on discontinuous address spaces is okay after all
1240 flags |= MPOL_MF_DISCONTIG_OK;
1242 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1243 start, start + len, mode, mode_flags,
1244 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1246 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1248 err = migrate_prep();
1253 NODEMASK_SCRATCH(scratch);
1255 down_write(&mm->mmap_sem);
1257 err = mpol_set_nodemask(new, nmask, scratch);
1258 task_unlock(current);
1260 up_write(&mm->mmap_sem);
1263 NODEMASK_SCRATCH_FREE(scratch);
1268 ret = queue_pages_range(mm, start, end, nmask,
1269 flags | MPOL_MF_INVERT, &pagelist);
1276 err = mbind_range(mm, start, end, new);
1281 if (!list_empty(&pagelist)) {
1282 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1283 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1284 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1286 putback_movable_pages(&pagelist);
1289 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1293 if (!list_empty(&pagelist))
1294 putback_movable_pages(&pagelist);
1297 up_write(&mm->mmap_sem);
1304 * User space interface with variable sized bitmaps for nodelists.
1307 /* Copy a node mask from user space. */
1308 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1309 unsigned long maxnode)
1313 unsigned long nlongs;
1314 unsigned long endmask;
1317 nodes_clear(*nodes);
1318 if (maxnode == 0 || !nmask)
1320 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1323 nlongs = BITS_TO_LONGS(maxnode);
1324 if ((maxnode % BITS_PER_LONG) == 0)
1327 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1330 * When the user specified more nodes than supported just check
1331 * if the non supported part is all zero.
1333 * If maxnode have more longs than MAX_NUMNODES, check
1334 * the bits in that area first. And then go through to
1335 * check the rest bits which equal or bigger than MAX_NUMNODES.
1336 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1338 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1339 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1340 if (get_user(t, nmask + k))
1342 if (k == nlongs - 1) {
1348 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1352 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1353 unsigned long valid_mask = endmask;
1355 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1356 if (get_user(t, nmask + nlongs - 1))
1362 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1364 nodes_addr(*nodes)[nlongs-1] &= endmask;
1368 /* Copy a kernel node mask to user space */
1369 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1372 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1373 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1375 if (copy > nbytes) {
1376 if (copy > PAGE_SIZE)
1378 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1382 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1385 static long kernel_mbind(unsigned long start, unsigned long len,
1386 unsigned long mode, const unsigned long __user *nmask,
1387 unsigned long maxnode, unsigned int flags)
1391 unsigned short mode_flags;
1393 mode_flags = mode & MPOL_MODE_FLAGS;
1394 mode &= ~MPOL_MODE_FLAGS;
1395 if (mode >= MPOL_MAX)
1397 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1398 (mode_flags & MPOL_F_RELATIVE_NODES))
1400 err = get_nodes(&nodes, nmask, maxnode);
1403 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1406 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1407 unsigned long, mode, const unsigned long __user *, nmask,
1408 unsigned long, maxnode, unsigned int, flags)
1410 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1413 /* Set the process memory policy */
1414 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1415 unsigned long maxnode)
1419 unsigned short flags;
1421 flags = mode & MPOL_MODE_FLAGS;
1422 mode &= ~MPOL_MODE_FLAGS;
1423 if ((unsigned int)mode >= MPOL_MAX)
1425 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1427 err = get_nodes(&nodes, nmask, maxnode);
1430 return do_set_mempolicy(mode, flags, &nodes);
1433 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1434 unsigned long, maxnode)
1436 return kernel_set_mempolicy(mode, nmask, maxnode);
1439 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1440 const unsigned long __user *old_nodes,
1441 const unsigned long __user *new_nodes)
1443 struct mm_struct *mm = NULL;
1444 struct task_struct *task;
1445 nodemask_t task_nodes;
1449 NODEMASK_SCRATCH(scratch);
1454 old = &scratch->mask1;
1455 new = &scratch->mask2;
1457 err = get_nodes(old, old_nodes, maxnode);
1461 err = get_nodes(new, new_nodes, maxnode);
1465 /* Find the mm_struct */
1467 task = pid ? find_task_by_vpid(pid) : current;
1473 get_task_struct(task);
1478 * Check if this process has the right to modify the specified process.
1479 * Use the regular "ptrace_may_access()" checks.
1481 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1488 task_nodes = cpuset_mems_allowed(task);
1489 /* Is the user allowed to access the target nodes? */
1490 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1495 task_nodes = cpuset_mems_allowed(current);
1496 nodes_and(*new, *new, task_nodes);
1497 if (nodes_empty(*new))
1500 nodes_and(*new, *new, node_states[N_MEMORY]);
1501 if (nodes_empty(*new))
1504 err = security_task_movememory(task);
1508 mm = get_task_mm(task);
1509 put_task_struct(task);
1516 err = do_migrate_pages(mm, old, new,
1517 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1521 NODEMASK_SCRATCH_FREE(scratch);
1526 put_task_struct(task);
1531 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1532 const unsigned long __user *, old_nodes,
1533 const unsigned long __user *, new_nodes)
1535 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1539 /* Retrieve NUMA policy */
1540 static int kernel_get_mempolicy(int __user *policy,
1541 unsigned long __user *nmask,
1542 unsigned long maxnode,
1544 unsigned long flags)
1547 int uninitialized_var(pval);
1550 if (nmask != NULL && maxnode < nr_node_ids)
1553 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1558 if (policy && put_user(pval, policy))
1562 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1567 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1568 unsigned long __user *, nmask, unsigned long, maxnode,
1569 unsigned long, addr, unsigned long, flags)
1571 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1574 #ifdef CONFIG_COMPAT
1576 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1577 compat_ulong_t __user *, nmask,
1578 compat_ulong_t, maxnode,
1579 compat_ulong_t, addr, compat_ulong_t, flags)
1582 unsigned long __user *nm = NULL;
1583 unsigned long nr_bits, alloc_size;
1584 DECLARE_BITMAP(bm, MAX_NUMNODES);
1586 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1587 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1590 nm = compat_alloc_user_space(alloc_size);
1592 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1594 if (!err && nmask) {
1595 unsigned long copy_size;
1596 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1597 err = copy_from_user(bm, nm, copy_size);
1598 /* ensure entire bitmap is zeroed */
1599 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1600 err |= compat_put_bitmap(nmask, bm, nr_bits);
1606 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1607 compat_ulong_t, maxnode)
1609 unsigned long __user *nm = NULL;
1610 unsigned long nr_bits, alloc_size;
1611 DECLARE_BITMAP(bm, MAX_NUMNODES);
1613 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1614 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1617 if (compat_get_bitmap(bm, nmask, nr_bits))
1619 nm = compat_alloc_user_space(alloc_size);
1620 if (copy_to_user(nm, bm, alloc_size))
1624 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1627 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1628 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1629 compat_ulong_t, maxnode, compat_ulong_t, flags)
1631 unsigned long __user *nm = NULL;
1632 unsigned long nr_bits, alloc_size;
1635 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1636 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1639 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1641 nm = compat_alloc_user_space(alloc_size);
1642 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1646 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1649 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1650 compat_ulong_t, maxnode,
1651 const compat_ulong_t __user *, old_nodes,
1652 const compat_ulong_t __user *, new_nodes)
1654 unsigned long __user *old = NULL;
1655 unsigned long __user *new = NULL;
1656 nodemask_t tmp_mask;
1657 unsigned long nr_bits;
1660 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1661 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1663 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1665 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1667 new = old + size / sizeof(unsigned long);
1668 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1672 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1675 new = compat_alloc_user_space(size);
1676 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1679 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1682 #endif /* CONFIG_COMPAT */
1684 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1687 struct mempolicy *pol = NULL;
1690 if (vma->vm_ops && vma->vm_ops->get_policy) {
1691 pol = vma->vm_ops->get_policy(vma, addr);
1692 } else if (vma->vm_policy) {
1693 pol = vma->vm_policy;
1696 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1697 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1698 * count on these policies which will be dropped by
1699 * mpol_cond_put() later
1701 if (mpol_needs_cond_ref(pol))
1710 * get_vma_policy(@vma, @addr)
1711 * @vma: virtual memory area whose policy is sought
1712 * @addr: address in @vma for shared policy lookup
1714 * Returns effective policy for a VMA at specified address.
1715 * Falls back to current->mempolicy or system default policy, as necessary.
1716 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1717 * count--added by the get_policy() vm_op, as appropriate--to protect against
1718 * freeing by another task. It is the caller's responsibility to free the
1719 * extra reference for shared policies.
1721 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1724 struct mempolicy *pol = __get_vma_policy(vma, addr);
1727 pol = get_task_policy(current);
1732 bool vma_policy_mof(struct vm_area_struct *vma)
1734 struct mempolicy *pol;
1736 if (vma->vm_ops && vma->vm_ops->get_policy) {
1739 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1740 if (pol && (pol->flags & MPOL_F_MOF))
1747 pol = vma->vm_policy;
1749 pol = get_task_policy(current);
1751 return pol->flags & MPOL_F_MOF;
1754 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1756 enum zone_type dynamic_policy_zone = policy_zone;
1758 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1761 * if policy->v.nodes has movable memory only,
1762 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1764 * policy->v.nodes is intersect with node_states[N_MEMORY].
1765 * so if the following test faile, it implies
1766 * policy->v.nodes has movable memory only.
1768 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1769 dynamic_policy_zone = ZONE_MOVABLE;
1771 return zone >= dynamic_policy_zone;
1775 * Return a nodemask representing a mempolicy for filtering nodes for
1778 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1780 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1781 if (unlikely(policy->mode == MPOL_BIND) &&
1782 apply_policy_zone(policy, gfp_zone(gfp)) &&
1783 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1784 return &policy->v.nodes;
1789 /* Return the node id preferred by the given mempolicy, or the given id */
1790 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1793 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1794 nd = policy->v.preferred_node;
1797 * __GFP_THISNODE shouldn't even be used with the bind policy
1798 * because we might easily break the expectation to stay on the
1799 * requested node and not break the policy.
1801 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1807 /* Do dynamic interleaving for a process */
1808 static unsigned interleave_nodes(struct mempolicy *policy)
1811 struct task_struct *me = current;
1813 next = next_node_in(me->il_prev, policy->v.nodes);
1814 if (next < MAX_NUMNODES)
1820 * Depending on the memory policy provide a node from which to allocate the
1823 unsigned int mempolicy_slab_node(void)
1825 struct mempolicy *policy;
1826 int node = numa_mem_id();
1831 policy = current->mempolicy;
1832 if (!policy || policy->flags & MPOL_F_LOCAL)
1835 switch (policy->mode) {
1836 case MPOL_PREFERRED:
1838 * handled MPOL_F_LOCAL above
1840 return policy->v.preferred_node;
1842 case MPOL_INTERLEAVE:
1843 return interleave_nodes(policy);
1849 * Follow bind policy behavior and start allocation at the
1852 struct zonelist *zonelist;
1853 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1854 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1855 z = first_zones_zonelist(zonelist, highest_zoneidx,
1857 return z->zone ? zone_to_nid(z->zone) : node;
1866 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1867 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1868 * number of present nodes.
1870 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1872 unsigned nnodes = nodes_weight(pol->v.nodes);
1878 return numa_node_id();
1879 target = (unsigned int)n % nnodes;
1880 nid = first_node(pol->v.nodes);
1881 for (i = 0; i < target; i++)
1882 nid = next_node(nid, pol->v.nodes);
1886 /* Determine a node number for interleave */
1887 static inline unsigned interleave_nid(struct mempolicy *pol,
1888 struct vm_area_struct *vma, unsigned long addr, int shift)
1894 * for small pages, there is no difference between
1895 * shift and PAGE_SHIFT, so the bit-shift is safe.
1896 * for huge pages, since vm_pgoff is in units of small
1897 * pages, we need to shift off the always 0 bits to get
1900 BUG_ON(shift < PAGE_SHIFT);
1901 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1902 off += (addr - vma->vm_start) >> shift;
1903 return offset_il_node(pol, off);
1905 return interleave_nodes(pol);
1908 #ifdef CONFIG_HUGETLBFS
1910 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1911 * @vma: virtual memory area whose policy is sought
1912 * @addr: address in @vma for shared policy lookup and interleave policy
1913 * @gfp_flags: for requested zone
1914 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1915 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1917 * Returns a nid suitable for a huge page allocation and a pointer
1918 * to the struct mempolicy for conditional unref after allocation.
1919 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1920 * @nodemask for filtering the zonelist.
1922 * Must be protected by read_mems_allowed_begin()
1924 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1925 struct mempolicy **mpol, nodemask_t **nodemask)
1929 *mpol = get_vma_policy(vma, addr);
1930 *nodemask = NULL; /* assume !MPOL_BIND */
1932 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1933 nid = interleave_nid(*mpol, vma, addr,
1934 huge_page_shift(hstate_vma(vma)));
1936 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1937 if ((*mpol)->mode == MPOL_BIND)
1938 *nodemask = &(*mpol)->v.nodes;
1944 * init_nodemask_of_mempolicy
1946 * If the current task's mempolicy is "default" [NULL], return 'false'
1947 * to indicate default policy. Otherwise, extract the policy nodemask
1948 * for 'bind' or 'interleave' policy into the argument nodemask, or
1949 * initialize the argument nodemask to contain the single node for
1950 * 'preferred' or 'local' policy and return 'true' to indicate presence
1951 * of non-default mempolicy.
1953 * We don't bother with reference counting the mempolicy [mpol_get/put]
1954 * because the current task is examining it's own mempolicy and a task's
1955 * mempolicy is only ever changed by the task itself.
1957 * N.B., it is the caller's responsibility to free a returned nodemask.
1959 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1961 struct mempolicy *mempolicy;
1964 if (!(mask && current->mempolicy))
1968 mempolicy = current->mempolicy;
1969 switch (mempolicy->mode) {
1970 case MPOL_PREFERRED:
1971 if (mempolicy->flags & MPOL_F_LOCAL)
1972 nid = numa_node_id();
1974 nid = mempolicy->v.preferred_node;
1975 init_nodemask_of_node(mask, nid);
1980 case MPOL_INTERLEAVE:
1981 *mask = mempolicy->v.nodes;
1987 task_unlock(current);
1994 * mempolicy_nodemask_intersects
1996 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1997 * policy. Otherwise, check for intersection between mask and the policy
1998 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1999 * policy, always return true since it may allocate elsewhere on fallback.
2001 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2003 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2004 const nodemask_t *mask)
2006 struct mempolicy *mempolicy;
2012 mempolicy = tsk->mempolicy;
2016 switch (mempolicy->mode) {
2017 case MPOL_PREFERRED:
2019 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2020 * allocate from, they may fallback to other nodes when oom.
2021 * Thus, it's possible for tsk to have allocated memory from
2026 case MPOL_INTERLEAVE:
2027 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2037 /* Allocate a page in interleaved policy.
2038 Own path because it needs to do special accounting. */
2039 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2044 page = __alloc_pages(gfp, order, nid);
2045 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2046 if (!static_branch_likely(&vm_numa_stat_key))
2048 if (page && page_to_nid(page) == nid) {
2050 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2057 * alloc_pages_vma - Allocate a page for a VMA.
2060 * %GFP_USER user allocation.
2061 * %GFP_KERNEL kernel allocations,
2062 * %GFP_HIGHMEM highmem/user allocations,
2063 * %GFP_FS allocation should not call back into a file system.
2064 * %GFP_ATOMIC don't sleep.
2066 * @order:Order of the GFP allocation.
2067 * @vma: Pointer to VMA or NULL if not available.
2068 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2069 * @node: Which node to prefer for allocation (modulo policy).
2070 * @hugepage: for hugepages try only the preferred node if possible
2072 * This function allocates a page from the kernel page pool and applies
2073 * a NUMA policy associated with the VMA or the current process.
2074 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2075 * mm_struct of the VMA to prevent it from going away. Should be used for
2076 * all allocations for pages that will be mapped into user space. Returns
2077 * NULL when no page can be allocated.
2080 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2081 unsigned long addr, int node, bool hugepage)
2083 struct mempolicy *pol;
2088 pol = get_vma_policy(vma, addr);
2090 if (pol->mode == MPOL_INTERLEAVE) {
2093 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2095 page = alloc_page_interleave(gfp, order, nid);
2099 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2100 int hpage_node = node;
2103 * For hugepage allocation and non-interleave policy which
2104 * allows the current node (or other explicitly preferred
2105 * node) we only try to allocate from the current/preferred
2106 * node and don't fall back to other nodes, as the cost of
2107 * remote accesses would likely offset THP benefits.
2109 * If the policy is interleave, or does not allow the current
2110 * node in its nodemask, we allocate the standard way.
2112 if (pol->mode == MPOL_PREFERRED &&
2113 !(pol->flags & MPOL_F_LOCAL))
2114 hpage_node = pol->v.preferred_node;
2116 nmask = policy_nodemask(gfp, pol);
2117 if (!nmask || node_isset(hpage_node, *nmask)) {
2120 * We cannot invoke reclaim if __GFP_THISNODE
2121 * is set. Invoking reclaim with
2122 * __GFP_THISNODE set, would cause THP
2123 * allocations to trigger heavy swapping
2124 * despite there may be tons of free memory
2125 * (including potentially plenty of THP
2126 * already available in the buddy) on all the
2129 * At most we could invoke compaction when
2130 * __GFP_THISNODE is set (but we would need to
2131 * refrain from invoking reclaim even if
2132 * compaction returned COMPACT_SKIPPED because
2133 * there wasn't not enough memory to succeed
2134 * compaction). For now just avoid
2135 * __GFP_THISNODE instead of limiting the
2136 * allocation path to a strict and single
2137 * compaction invocation.
2139 * Supposedly if direct reclaim was enabled by
2140 * the caller, the app prefers THP regardless
2141 * of the node it comes from so this would be
2142 * more desiderable behavior than only
2143 * providing THP originated from the local
2144 * node in such case.
2146 if (!(gfp & __GFP_DIRECT_RECLAIM))
2147 gfp |= __GFP_THISNODE;
2148 page = __alloc_pages_node(hpage_node, gfp, order);
2153 nmask = policy_nodemask(gfp, pol);
2154 preferred_nid = policy_node(gfp, pol, node);
2155 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2162 * alloc_pages_current - Allocate pages.
2165 * %GFP_USER user allocation,
2166 * %GFP_KERNEL kernel allocation,
2167 * %GFP_HIGHMEM highmem allocation,
2168 * %GFP_FS don't call back into a file system.
2169 * %GFP_ATOMIC don't sleep.
2170 * @order: Power of two of allocation size in pages. 0 is a single page.
2172 * Allocate a page from the kernel page pool. When not in
2173 * interrupt context and apply the current process NUMA policy.
2174 * Returns NULL when no page can be allocated.
2176 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2178 struct mempolicy *pol = &default_policy;
2181 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2182 pol = get_task_policy(current);
2185 * No reference counting needed for current->mempolicy
2186 * nor system default_policy
2188 if (pol->mode == MPOL_INTERLEAVE)
2189 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2191 page = __alloc_pages_nodemask(gfp, order,
2192 policy_node(gfp, pol, numa_node_id()),
2193 policy_nodemask(gfp, pol));
2197 EXPORT_SYMBOL(alloc_pages_current);
2199 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2201 struct mempolicy *pol = mpol_dup(vma_policy(src));
2204 return PTR_ERR(pol);
2205 dst->vm_policy = pol;
2210 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2211 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2212 * with the mems_allowed returned by cpuset_mems_allowed(). This
2213 * keeps mempolicies cpuset relative after its cpuset moves. See
2214 * further kernel/cpuset.c update_nodemask().
2216 * current's mempolicy may be rebinded by the other task(the task that changes
2217 * cpuset's mems), so we needn't do rebind work for current task.
2220 /* Slow path of a mempolicy duplicate */
2221 struct mempolicy *__mpol_dup(struct mempolicy *old)
2223 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2226 return ERR_PTR(-ENOMEM);
2228 /* task's mempolicy is protected by alloc_lock */
2229 if (old == current->mempolicy) {
2232 task_unlock(current);
2236 if (current_cpuset_is_being_rebound()) {
2237 nodemask_t mems = cpuset_mems_allowed(current);
2238 mpol_rebind_policy(new, &mems);
2240 atomic_set(&new->refcnt, 1);
2244 /* Slow path of a mempolicy comparison */
2245 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2249 if (a->mode != b->mode)
2251 if (a->flags != b->flags)
2253 if (mpol_store_user_nodemask(a))
2254 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2260 case MPOL_INTERLEAVE:
2261 return !!nodes_equal(a->v.nodes, b->v.nodes);
2262 case MPOL_PREFERRED:
2263 /* a's ->flags is the same as b's */
2264 if (a->flags & MPOL_F_LOCAL)
2266 return a->v.preferred_node == b->v.preferred_node;
2274 * Shared memory backing store policy support.
2276 * Remember policies even when nobody has shared memory mapped.
2277 * The policies are kept in Red-Black tree linked from the inode.
2278 * They are protected by the sp->lock rwlock, which should be held
2279 * for any accesses to the tree.
2283 * lookup first element intersecting start-end. Caller holds sp->lock for
2284 * reading or for writing
2286 static struct sp_node *
2287 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2289 struct rb_node *n = sp->root.rb_node;
2292 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2294 if (start >= p->end)
2296 else if (end <= p->start)
2304 struct sp_node *w = NULL;
2305 struct rb_node *prev = rb_prev(n);
2308 w = rb_entry(prev, struct sp_node, nd);
2309 if (w->end <= start)
2313 return rb_entry(n, struct sp_node, nd);
2317 * Insert a new shared policy into the list. Caller holds sp->lock for
2320 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2322 struct rb_node **p = &sp->root.rb_node;
2323 struct rb_node *parent = NULL;
2328 nd = rb_entry(parent, struct sp_node, nd);
2329 if (new->start < nd->start)
2331 else if (new->end > nd->end)
2332 p = &(*p)->rb_right;
2336 rb_link_node(&new->nd, parent, p);
2337 rb_insert_color(&new->nd, &sp->root);
2338 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2339 new->policy ? new->policy->mode : 0);
2342 /* Find shared policy intersecting idx */
2344 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2346 struct mempolicy *pol = NULL;
2349 if (!sp->root.rb_node)
2351 read_lock(&sp->lock);
2352 sn = sp_lookup(sp, idx, idx+1);
2354 mpol_get(sn->policy);
2357 read_unlock(&sp->lock);
2361 static void sp_free(struct sp_node *n)
2363 mpol_put(n->policy);
2364 kmem_cache_free(sn_cache, n);
2368 * mpol_misplaced - check whether current page node is valid in policy
2370 * @page: page to be checked
2371 * @vma: vm area where page mapped
2372 * @addr: virtual address where page mapped
2374 * Lookup current policy node id for vma,addr and "compare to" page's
2378 * -1 - not misplaced, page is in the right node
2379 * node - node id where the page should be
2381 * Policy determination "mimics" alloc_page_vma().
2382 * Called from fault path where we know the vma and faulting address.
2384 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2386 struct mempolicy *pol;
2388 int curnid = page_to_nid(page);
2389 unsigned long pgoff;
2390 int thiscpu = raw_smp_processor_id();
2391 int thisnid = cpu_to_node(thiscpu);
2395 pol = get_vma_policy(vma, addr);
2396 if (!(pol->flags & MPOL_F_MOF))
2399 switch (pol->mode) {
2400 case MPOL_INTERLEAVE:
2401 pgoff = vma->vm_pgoff;
2402 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2403 polnid = offset_il_node(pol, pgoff);
2406 case MPOL_PREFERRED:
2407 if (pol->flags & MPOL_F_LOCAL)
2408 polnid = numa_node_id();
2410 polnid = pol->v.preferred_node;
2416 * allows binding to multiple nodes.
2417 * use current page if in policy nodemask,
2418 * else select nearest allowed node, if any.
2419 * If no allowed nodes, use current [!misplaced].
2421 if (node_isset(curnid, pol->v.nodes))
2423 z = first_zones_zonelist(
2424 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2425 gfp_zone(GFP_HIGHUSER),
2427 polnid = zone_to_nid(z->zone);
2434 /* Migrate the page towards the node whose CPU is referencing it */
2435 if (pol->flags & MPOL_F_MORON) {
2438 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2442 if (curnid != polnid)
2451 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2452 * dropped after task->mempolicy is set to NULL so that any allocation done as
2453 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2456 void mpol_put_task_policy(struct task_struct *task)
2458 struct mempolicy *pol;
2461 pol = task->mempolicy;
2462 task->mempolicy = NULL;
2467 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2469 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2470 rb_erase(&n->nd, &sp->root);
2474 static void sp_node_init(struct sp_node *node, unsigned long start,
2475 unsigned long end, struct mempolicy *pol)
2477 node->start = start;
2482 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2483 struct mempolicy *pol)
2486 struct mempolicy *newpol;
2488 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2492 newpol = mpol_dup(pol);
2493 if (IS_ERR(newpol)) {
2494 kmem_cache_free(sn_cache, n);
2497 newpol->flags |= MPOL_F_SHARED;
2498 sp_node_init(n, start, end, newpol);
2503 /* Replace a policy range. */
2504 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2505 unsigned long end, struct sp_node *new)
2508 struct sp_node *n_new = NULL;
2509 struct mempolicy *mpol_new = NULL;
2513 write_lock(&sp->lock);
2514 n = sp_lookup(sp, start, end);
2515 /* Take care of old policies in the same range. */
2516 while (n && n->start < end) {
2517 struct rb_node *next = rb_next(&n->nd);
2518 if (n->start >= start) {
2524 /* Old policy spanning whole new range. */
2529 *mpol_new = *n->policy;
2530 atomic_set(&mpol_new->refcnt, 1);
2531 sp_node_init(n_new, end, n->end, mpol_new);
2533 sp_insert(sp, n_new);
2542 n = rb_entry(next, struct sp_node, nd);
2546 write_unlock(&sp->lock);
2553 kmem_cache_free(sn_cache, n_new);
2558 write_unlock(&sp->lock);
2560 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2563 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2566 atomic_set(&mpol_new->refcnt, 1);
2571 * mpol_shared_policy_init - initialize shared policy for inode
2572 * @sp: pointer to inode shared policy
2573 * @mpol: struct mempolicy to install
2575 * Install non-NULL @mpol in inode's shared policy rb-tree.
2576 * On entry, the current task has a reference on a non-NULL @mpol.
2577 * This must be released on exit.
2578 * This is called at get_inode() calls and we can use GFP_KERNEL.
2580 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2584 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2585 rwlock_init(&sp->lock);
2588 struct vm_area_struct pvma;
2589 struct mempolicy *new;
2590 NODEMASK_SCRATCH(scratch);
2594 /* contextualize the tmpfs mount point mempolicy */
2595 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2597 goto free_scratch; /* no valid nodemask intersection */
2600 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2601 task_unlock(current);
2605 /* Create pseudo-vma that contains just the policy */
2606 vma_init(&pvma, NULL);
2607 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2608 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2611 mpol_put(new); /* drop initial ref */
2613 NODEMASK_SCRATCH_FREE(scratch);
2615 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2619 int mpol_set_shared_policy(struct shared_policy *info,
2620 struct vm_area_struct *vma, struct mempolicy *npol)
2623 struct sp_node *new = NULL;
2624 unsigned long sz = vma_pages(vma);
2626 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2628 sz, npol ? npol->mode : -1,
2629 npol ? npol->flags : -1,
2630 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2633 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2637 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2643 /* Free a backing policy store on inode delete. */
2644 void mpol_free_shared_policy(struct shared_policy *p)
2647 struct rb_node *next;
2649 if (!p->root.rb_node)
2651 write_lock(&p->lock);
2652 next = rb_first(&p->root);
2654 n = rb_entry(next, struct sp_node, nd);
2655 next = rb_next(&n->nd);
2658 write_unlock(&p->lock);
2661 #ifdef CONFIG_NUMA_BALANCING
2662 static int __initdata numabalancing_override;
2664 static void __init check_numabalancing_enable(void)
2666 bool numabalancing_default = false;
2668 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2669 numabalancing_default = true;
2671 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2672 if (numabalancing_override)
2673 set_numabalancing_state(numabalancing_override == 1);
2675 if (num_online_nodes() > 1 && !numabalancing_override) {
2676 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2677 numabalancing_default ? "Enabling" : "Disabling");
2678 set_numabalancing_state(numabalancing_default);
2682 static int __init setup_numabalancing(char *str)
2688 if (!strcmp(str, "enable")) {
2689 numabalancing_override = 1;
2691 } else if (!strcmp(str, "disable")) {
2692 numabalancing_override = -1;
2697 pr_warn("Unable to parse numa_balancing=\n");
2701 __setup("numa_balancing=", setup_numabalancing);
2703 static inline void __init check_numabalancing_enable(void)
2706 #endif /* CONFIG_NUMA_BALANCING */
2708 /* assumes fs == KERNEL_DS */
2709 void __init numa_policy_init(void)
2711 nodemask_t interleave_nodes;
2712 unsigned long largest = 0;
2713 int nid, prefer = 0;
2715 policy_cache = kmem_cache_create("numa_policy",
2716 sizeof(struct mempolicy),
2717 0, SLAB_PANIC, NULL);
2719 sn_cache = kmem_cache_create("shared_policy_node",
2720 sizeof(struct sp_node),
2721 0, SLAB_PANIC, NULL);
2723 for_each_node(nid) {
2724 preferred_node_policy[nid] = (struct mempolicy) {
2725 .refcnt = ATOMIC_INIT(1),
2726 .mode = MPOL_PREFERRED,
2727 .flags = MPOL_F_MOF | MPOL_F_MORON,
2728 .v = { .preferred_node = nid, },
2733 * Set interleaving policy for system init. Interleaving is only
2734 * enabled across suitably sized nodes (default is >= 16MB), or
2735 * fall back to the largest node if they're all smaller.
2737 nodes_clear(interleave_nodes);
2738 for_each_node_state(nid, N_MEMORY) {
2739 unsigned long total_pages = node_present_pages(nid);
2741 /* Preserve the largest node */
2742 if (largest < total_pages) {
2743 largest = total_pages;
2747 /* Interleave this node? */
2748 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2749 node_set(nid, interleave_nodes);
2752 /* All too small, use the largest */
2753 if (unlikely(nodes_empty(interleave_nodes)))
2754 node_set(prefer, interleave_nodes);
2756 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2757 pr_err("%s: interleaving failed\n", __func__);
2759 check_numabalancing_enable();
2762 /* Reset policy of current process to default */
2763 void numa_default_policy(void)
2765 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2769 * Parse and format mempolicy from/to strings
2773 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2775 static const char * const policy_modes[] =
2777 [MPOL_DEFAULT] = "default",
2778 [MPOL_PREFERRED] = "prefer",
2779 [MPOL_BIND] = "bind",
2780 [MPOL_INTERLEAVE] = "interleave",
2781 [MPOL_LOCAL] = "local",
2787 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2788 * @str: string containing mempolicy to parse
2789 * @mpol: pointer to struct mempolicy pointer, returned on success.
2792 * <mode>[=<flags>][:<nodelist>]
2794 * On success, returns 0, else 1
2796 int mpol_parse_str(char *str, struct mempolicy **mpol)
2798 struct mempolicy *new = NULL;
2799 unsigned short mode;
2800 unsigned short mode_flags;
2802 char *nodelist = strchr(str, ':');
2803 char *flags = strchr(str, '=');
2807 *flags++ = '\0'; /* terminate mode string */
2810 /* NUL-terminate mode or flags string */
2812 if (nodelist_parse(nodelist, nodes))
2814 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2819 for (mode = 0; mode < MPOL_MAX; mode++) {
2820 if (!strcmp(str, policy_modes[mode])) {
2824 if (mode >= MPOL_MAX)
2828 case MPOL_PREFERRED:
2830 * Insist on a nodelist of one node only, although later
2831 * we use first_node(nodes) to grab a single node, so here
2832 * nodelist (or nodes) cannot be empty.
2835 char *rest = nodelist;
2836 while (isdigit(*rest))
2840 if (nodes_empty(nodes))
2844 case MPOL_INTERLEAVE:
2846 * Default to online nodes with memory if no nodelist
2849 nodes = node_states[N_MEMORY];
2853 * Don't allow a nodelist; mpol_new() checks flags
2857 mode = MPOL_PREFERRED;
2861 * Insist on a empty nodelist
2868 * Insist on a nodelist
2877 * Currently, we only support two mutually exclusive
2880 if (!strcmp(flags, "static"))
2881 mode_flags |= MPOL_F_STATIC_NODES;
2882 else if (!strcmp(flags, "relative"))
2883 mode_flags |= MPOL_F_RELATIVE_NODES;
2888 new = mpol_new(mode, mode_flags, &nodes);
2893 * Save nodes for mpol_to_str() to show the tmpfs mount options
2894 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2896 if (mode != MPOL_PREFERRED)
2897 new->v.nodes = nodes;
2899 new->v.preferred_node = first_node(nodes);
2901 new->flags |= MPOL_F_LOCAL;
2904 * Save nodes for contextualization: this will be used to "clone"
2905 * the mempolicy in a specific context [cpuset] at a later time.
2907 new->w.user_nodemask = nodes;
2912 /* Restore string for error message */
2921 #endif /* CONFIG_TMPFS */
2924 * mpol_to_str - format a mempolicy structure for printing
2925 * @buffer: to contain formatted mempolicy string
2926 * @maxlen: length of @buffer
2927 * @pol: pointer to mempolicy to be formatted
2929 * Convert @pol into a string. If @buffer is too short, truncate the string.
2930 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2931 * longest flag, "relative", and to display at least a few node ids.
2933 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2936 nodemask_t nodes = NODE_MASK_NONE;
2937 unsigned short mode = MPOL_DEFAULT;
2938 unsigned short flags = 0;
2940 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2948 case MPOL_PREFERRED:
2949 if (flags & MPOL_F_LOCAL)
2952 node_set(pol->v.preferred_node, nodes);
2955 case MPOL_INTERLEAVE:
2956 nodes = pol->v.nodes;
2960 snprintf(p, maxlen, "unknown");
2964 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2966 if (flags & MPOL_MODE_FLAGS) {
2967 p += snprintf(p, buffer + maxlen - p, "=");
2970 * Currently, the only defined flags are mutually exclusive
2972 if (flags & MPOL_F_STATIC_NODES)
2973 p += snprintf(p, buffer + maxlen - p, "static");
2974 else if (flags & MPOL_F_RELATIVE_NODES)
2975 p += snprintf(p, buffer + maxlen - p, "relative");
2978 if (!nodes_empty(nodes))
2979 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2980 nodemask_pr_args(&nodes));