1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple NUMA memory policy for the Linux kernel.
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * preferred many Try a set of nodes first before normal fallback. This is
35 * similar to preferred without the special case.
37 * default Allocate on the local node first, or when on a VMA
38 * use the process policy. This is what Linux always did
39 * in a NUMA aware kernel and still does by, ahem, default.
41 * The process policy is applied for most non interrupt memory allocations
42 * in that process' context. Interrupts ignore the policies and always
43 * try to allocate on the local CPU. The VMA policy is only applied for memory
44 * allocations for a VMA in the VM.
46 * Currently there are a few corner cases in swapping where the policy
47 * is not applied, but the majority should be handled. When process policy
48 * is used it is not remembered over swap outs/swap ins.
50 * Only the highest zone in the zone hierarchy gets policied. Allocations
51 * requesting a lower zone just use default policy. This implies that
52 * on systems with highmem kernel lowmem allocation don't get policied.
53 * Same with GFP_DMA allocations.
55 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
56 * all users and remembered even when nobody has memory mapped.
60 fix mmap readahead to honour policy and enable policy for any page cache
62 statistics for bigpages
63 global policy for page cache? currently it uses process policy. Requires
65 handle mremap for shared memory (currently ignored for the policy)
67 make bind policy root only? It can trigger oom much faster and the
68 kernel is not always grateful with that.
71 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
73 #include <linux/mempolicy.h>
74 #include <linux/pagewalk.h>
75 #include <linux/highmem.h>
76 #include <linux/hugetlb.h>
77 #include <linux/kernel.h>
78 #include <linux/sched.h>
79 #include <linux/sched/mm.h>
80 #include <linux/sched/numa_balancing.h>
81 #include <linux/sched/task.h>
82 #include <linux/nodemask.h>
83 #include <linux/cpuset.h>
84 #include <linux/slab.h>
85 #include <linux/string.h>
86 #include <linux/export.h>
87 #include <linux/nsproxy.h>
88 #include <linux/interrupt.h>
89 #include <linux/init.h>
90 #include <linux/compat.h>
91 #include <linux/ptrace.h>
92 #include <linux/swap.h>
93 #include <linux/seq_file.h>
94 #include <linux/proc_fs.h>
95 #include <linux/migrate.h>
96 #include <linux/ksm.h>
97 #include <linux/rmap.h>
98 #include <linux/security.h>
99 #include <linux/syscalls.h>
100 #include <linux/ctype.h>
101 #include <linux/mm_inline.h>
102 #include <linux/mmu_notifier.h>
103 #include <linux/printk.h>
104 #include <linux/swapops.h>
106 #include <asm/tlbflush.h>
107 #include <linux/uaccess.h>
109 #include "internal.h"
112 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
113 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
115 static struct kmem_cache *policy_cache;
116 static struct kmem_cache *sn_cache;
118 /* Highest zone. An specific allocation for a zone below that is not
120 enum zone_type policy_zone = 0;
123 * run-time system-wide default policy => local allocation
125 static struct mempolicy default_policy = {
126 .refcnt = ATOMIC_INIT(1), /* never free it */
130 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
133 * numa_map_to_online_node - Find closest online node
134 * @node: Node id to start the search
136 * Lookup the next closest node by distance if @nid is not online.
138 int numa_map_to_online_node(int node)
140 int min_dist = INT_MAX, dist, n, min_node;
142 if (node == NUMA_NO_NODE || node_online(node))
146 for_each_online_node(n) {
147 dist = node_distance(node, n);
148 if (dist < min_dist) {
156 EXPORT_SYMBOL_GPL(numa_map_to_online_node);
158 struct mempolicy *get_task_policy(struct task_struct *p)
160 struct mempolicy *pol = p->mempolicy;
166 node = numa_node_id();
167 if (node != NUMA_NO_NODE) {
168 pol = &preferred_node_policy[node];
169 /* preferred_node_policy is not initialised early in boot */
174 return &default_policy;
177 static const struct mempolicy_operations {
178 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
179 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
180 } mpol_ops[MPOL_MAX];
182 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
184 return pol->flags & MPOL_MODE_FLAGS;
187 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
188 const nodemask_t *rel)
191 nodes_fold(tmp, *orig, nodes_weight(*rel));
192 nodes_onto(*ret, tmp, *rel);
195 static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
197 if (nodes_empty(*nodes))
203 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
205 if (nodes_empty(*nodes))
208 nodes_clear(pol->nodes);
209 node_set(first_node(*nodes), pol->nodes);
214 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
215 * any, for the new policy. mpol_new() has already validated the nodes
216 * parameter with respect to the policy mode and flags.
218 * Must be called holding task's alloc_lock to protect task's mems_allowed
219 * and mempolicy. May also be called holding the mmap_lock for write.
221 static int mpol_set_nodemask(struct mempolicy *pol,
222 const nodemask_t *nodes, struct nodemask_scratch *nsc)
227 * Default (pol==NULL) resp. local memory policies are not a
228 * subject of any remapping. They also do not need any special
231 if (!pol || pol->mode == MPOL_LOCAL)
235 nodes_and(nsc->mask1,
236 cpuset_current_mems_allowed, node_states[N_MEMORY]);
240 if (pol->flags & MPOL_F_RELATIVE_NODES)
241 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
243 nodes_and(nsc->mask2, *nodes, nsc->mask1);
245 if (mpol_store_user_nodemask(pol))
246 pol->w.user_nodemask = *nodes;
248 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
250 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
286 } else if (mode == MPOL_LOCAL) {
287 if (!nodes_empty(*nodes) ||
288 (flags & MPOL_F_STATIC_NODES) ||
289 (flags & MPOL_F_RELATIVE_NODES))
290 return ERR_PTR(-EINVAL);
291 } else if (nodes_empty(*nodes))
292 return ERR_PTR(-EINVAL);
293 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
295 return ERR_PTR(-ENOMEM);
296 atomic_set(&policy->refcnt, 1);
298 policy->flags = flags;
303 /* Slow path of a mpol destructor. */
304 void __mpol_put(struct mempolicy *p)
306 if (!atomic_dec_and_test(&p->refcnt))
308 kmem_cache_free(policy_cache, p);
311 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
315 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
319 if (pol->flags & MPOL_F_STATIC_NODES)
320 nodes_and(tmp, pol->w.user_nodemask, *nodes);
321 else if (pol->flags & MPOL_F_RELATIVE_NODES)
322 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
324 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
326 pol->w.cpuset_mems_allowed = *nodes;
329 if (nodes_empty(tmp))
335 static void mpol_rebind_preferred(struct mempolicy *pol,
336 const nodemask_t *nodes)
338 pol->w.cpuset_mems_allowed = *nodes;
342 * mpol_rebind_policy - Migrate a policy to a different set of nodes
344 * Per-vma policies are protected by mmap_lock. Allocations using per-task
345 * policies are protected by task->mems_allowed_seq to prevent a premature
346 * OOM/allocation failure due to parallel nodemask modification.
348 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
352 if (!mpol_store_user_nodemask(pol) &&
353 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
356 mpol_ops[pol->mode].rebind(pol, newmask);
360 * Wrapper for mpol_rebind_policy() that just requires task
361 * pointer, and updates task mempolicy.
363 * Called with task's alloc_lock held.
366 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
368 mpol_rebind_policy(tsk->mempolicy, new);
372 * Rebind each vma in mm to new nodemask.
374 * Call holding a reference to mm. Takes mm->mmap_lock during call.
377 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
379 struct vm_area_struct *vma;
382 for (vma = mm->mmap; vma; vma = vma->vm_next)
383 mpol_rebind_policy(vma->vm_policy, new);
384 mmap_write_unlock(mm);
387 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
389 .rebind = mpol_rebind_default,
391 [MPOL_INTERLEAVE] = {
392 .create = mpol_new_nodemask,
393 .rebind = mpol_rebind_nodemask,
396 .create = mpol_new_preferred,
397 .rebind = mpol_rebind_preferred,
400 .create = mpol_new_nodemask,
401 .rebind = mpol_rebind_nodemask,
404 .rebind = mpol_rebind_default,
406 [MPOL_PREFERRED_MANY] = {
407 .create = mpol_new_nodemask,
408 .rebind = mpol_rebind_preferred,
412 static int migrate_page_add(struct page *page, struct list_head *pagelist,
413 unsigned long flags);
416 struct list_head *pagelist;
421 struct vm_area_struct *first;
425 * Check if the page's nid is in qp->nmask.
427 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
428 * in the invert of qp->nmask.
430 static inline bool queue_pages_required(struct page *page,
431 struct queue_pages *qp)
433 int nid = page_to_nid(page);
434 unsigned long flags = qp->flags;
436 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
440 * queue_pages_pmd() has four possible return values:
441 * 0 - pages are placed on the right node or queued successfully, or
442 * special page is met, i.e. huge zero page.
443 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
446 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
447 * existing page was already on a node that does not follow the
450 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
451 unsigned long end, struct mm_walk *walk)
456 struct queue_pages *qp = walk->private;
459 if (unlikely(is_pmd_migration_entry(*pmd))) {
463 page = pmd_page(*pmd);
464 if (is_huge_zero_page(page)) {
466 walk->action = ACTION_CONTINUE;
469 if (!queue_pages_required(page, qp))
473 /* go to thp migration */
474 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
475 if (!vma_migratable(walk->vma) ||
476 migrate_page_add(page, qp->pagelist, flags)) {
489 * Scan through pages checking if pages follow certain conditions,
490 * and move them to the pagelist if they do.
492 * queue_pages_pte_range() has three possible return values:
493 * 0 - pages are placed on the right node or queued successfully, or
494 * special page is met, i.e. zero page.
495 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
497 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
498 * on a node that does not follow the policy.
500 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
501 unsigned long end, struct mm_walk *walk)
503 struct vm_area_struct *vma = walk->vma;
505 struct queue_pages *qp = walk->private;
506 unsigned long flags = qp->flags;
508 bool has_unmovable = false;
509 pte_t *pte, *mapped_pte;
512 ptl = pmd_trans_huge_lock(pmd, vma);
514 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
518 /* THP was split, fall through to pte walk */
520 if (pmd_trans_unstable(pmd))
523 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
524 for (; addr != end; pte++, addr += PAGE_SIZE) {
525 if (!pte_present(*pte))
527 page = vm_normal_page(vma, addr, *pte);
531 * vm_normal_page() filters out zero pages, but there might
532 * still be PageReserved pages to skip, perhaps in a VDSO.
534 if (PageReserved(page))
536 if (!queue_pages_required(page, qp))
538 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
539 /* MPOL_MF_STRICT must be specified if we get here */
540 if (!vma_migratable(vma)) {
541 has_unmovable = true;
546 * Do not abort immediately since there may be
547 * temporary off LRU pages in the range. Still
548 * need migrate other LRU pages.
550 if (migrate_page_add(page, qp->pagelist, flags))
551 has_unmovable = true;
555 pte_unmap_unlock(mapped_pte, ptl);
561 return addr != end ? -EIO : 0;
564 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
565 unsigned long addr, unsigned long end,
566 struct mm_walk *walk)
569 #ifdef CONFIG_HUGETLB_PAGE
570 struct queue_pages *qp = walk->private;
571 unsigned long flags = (qp->flags & MPOL_MF_VALID);
576 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
577 entry = huge_ptep_get(pte);
578 if (!pte_present(entry))
580 page = pte_page(entry);
581 if (!queue_pages_required(page, qp))
584 if (flags == MPOL_MF_STRICT) {
586 * STRICT alone means only detecting misplaced page and no
587 * need to further check other vma.
593 if (!vma_migratable(walk->vma)) {
595 * Must be STRICT with MOVE*, otherwise .test_walk() have
596 * stopped walking current vma.
597 * Detecting misplaced page but allow migrating pages which
604 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
605 if (flags & (MPOL_MF_MOVE_ALL) ||
606 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
607 if (!isolate_huge_page(page, qp->pagelist) &&
608 (flags & MPOL_MF_STRICT))
610 * Failed to isolate page but allow migrating pages
611 * which have been queued.
623 #ifdef CONFIG_NUMA_BALANCING
625 * This is used to mark a range of virtual addresses to be inaccessible.
626 * These are later cleared by a NUMA hinting fault. Depending on these
627 * faults, pages may be migrated for better NUMA placement.
629 * This is assuming that NUMA faults are handled using PROT_NONE. If
630 * an architecture makes a different choice, it will need further
631 * changes to the core.
633 unsigned long change_prot_numa(struct vm_area_struct *vma,
634 unsigned long addr, unsigned long end)
638 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
640 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
645 static unsigned long change_prot_numa(struct vm_area_struct *vma,
646 unsigned long addr, unsigned long end)
650 #endif /* CONFIG_NUMA_BALANCING */
652 static int queue_pages_test_walk(unsigned long start, unsigned long end,
653 struct mm_walk *walk)
655 struct vm_area_struct *vma = walk->vma;
656 struct queue_pages *qp = walk->private;
657 unsigned long endvma = vma->vm_end;
658 unsigned long flags = qp->flags;
660 /* range check first */
661 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
665 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
666 (qp->start < vma->vm_start))
667 /* hole at head side of range */
670 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
671 ((vma->vm_end < qp->end) &&
672 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
673 /* hole at middle or tail of range */
677 * Need check MPOL_MF_STRICT to return -EIO if possible
678 * regardless of vma_migratable
680 if (!vma_migratable(vma) &&
681 !(flags & MPOL_MF_STRICT))
687 if (flags & MPOL_MF_LAZY) {
688 /* Similar to task_numa_work, skip inaccessible VMAs */
689 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
690 !(vma->vm_flags & VM_MIXEDMAP))
691 change_prot_numa(vma, start, endvma);
695 /* queue pages from current vma */
696 if (flags & MPOL_MF_VALID)
701 static const struct mm_walk_ops queue_pages_walk_ops = {
702 .hugetlb_entry = queue_pages_hugetlb,
703 .pmd_entry = queue_pages_pte_range,
704 .test_walk = queue_pages_test_walk,
708 * Walk through page tables and collect pages to be migrated.
710 * If pages found in a given range are on a set of nodes (determined by
711 * @nodes and @flags,) it's isolated and queued to the pagelist which is
712 * passed via @private.
714 * queue_pages_range() has three possible return values:
715 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
717 * 0 - queue pages successfully or no misplaced page.
718 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
719 * memory range specified by nodemask and maxnode points outside
720 * your accessible address space (-EFAULT)
723 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
724 nodemask_t *nodes, unsigned long flags,
725 struct list_head *pagelist)
728 struct queue_pages qp = {
729 .pagelist = pagelist,
737 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
740 /* whole range in hole */
747 * Apply policy to a single VMA
748 * This must be called with the mmap_lock held for writing.
750 static int vma_replace_policy(struct vm_area_struct *vma,
751 struct mempolicy *pol)
754 struct mempolicy *old;
755 struct mempolicy *new;
757 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
758 vma->vm_start, vma->vm_end, vma->vm_pgoff,
759 vma->vm_ops, vma->vm_file,
760 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
766 if (vma->vm_ops && vma->vm_ops->set_policy) {
767 err = vma->vm_ops->set_policy(vma, new);
772 old = vma->vm_policy;
773 vma->vm_policy = new; /* protected by mmap_lock */
782 /* Step 2: apply policy to a range and do splits. */
783 static int mbind_range(struct mm_struct *mm, unsigned long start,
784 unsigned long end, struct mempolicy *new_pol)
786 struct vm_area_struct *next;
787 struct vm_area_struct *prev;
788 struct vm_area_struct *vma;
791 unsigned long vmstart;
794 vma = find_vma(mm, start);
798 if (start > vma->vm_start)
801 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
803 vmstart = max(start, vma->vm_start);
804 vmend = min(end, vma->vm_end);
806 if (mpol_equal(vma_policy(vma), new_pol))
809 pgoff = vma->vm_pgoff +
810 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
811 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
812 vma->anon_vma, vma->vm_file, pgoff,
813 new_pol, vma->vm_userfaultfd_ctx);
817 if (mpol_equal(vma_policy(vma), new_pol))
819 /* vma_merge() joined vma && vma->next, case 8 */
822 if (vma->vm_start != vmstart) {
823 err = split_vma(vma->vm_mm, vma, vmstart, 1);
827 if (vma->vm_end != vmend) {
828 err = split_vma(vma->vm_mm, vma, vmend, 0);
833 err = vma_replace_policy(vma, new_pol);
842 /* Set the process memory policy */
843 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
846 struct mempolicy *new, *old;
847 NODEMASK_SCRATCH(scratch);
853 new = mpol_new(mode, flags, nodes);
859 if (flags & MPOL_F_NUMA_BALANCING) {
860 if (new && new->mode == MPOL_BIND) {
861 new->flags |= (MPOL_F_MOF | MPOL_F_MORON);
869 ret = mpol_set_nodemask(new, nodes, scratch);
875 old = current->mempolicy;
876 current->mempolicy = new;
877 if (new && new->mode == MPOL_INTERLEAVE)
878 current->il_prev = MAX_NUMNODES-1;
879 task_unlock(current);
883 NODEMASK_SCRATCH_FREE(scratch);
888 * Return nodemask for policy for get_mempolicy() query
890 * Called with task's alloc_lock held
892 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
895 if (p == &default_policy)
900 case MPOL_INTERLEAVE:
902 case MPOL_PREFERRED_MANY:
906 /* return empty node mask for local allocation */
913 static int lookup_node(struct mm_struct *mm, unsigned long addr)
915 struct page *p = NULL;
919 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
921 err = page_to_nid(p);
925 mmap_read_unlock(mm);
929 /* Retrieve NUMA policy */
930 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
931 unsigned long addr, unsigned long flags)
934 struct mm_struct *mm = current->mm;
935 struct vm_area_struct *vma = NULL;
936 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
939 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
942 if (flags & MPOL_F_MEMS_ALLOWED) {
943 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
945 *policy = 0; /* just so it's initialized */
947 *nmask = cpuset_current_mems_allowed;
948 task_unlock(current);
952 if (flags & MPOL_F_ADDR) {
954 * Do NOT fall back to task policy if the
955 * vma/shared policy at addr is NULL. We
956 * want to return MPOL_DEFAULT in this case.
959 vma = vma_lookup(mm, addr);
961 mmap_read_unlock(mm);
964 if (vma->vm_ops && vma->vm_ops->get_policy)
965 pol = vma->vm_ops->get_policy(vma, addr);
967 pol = vma->vm_policy;
972 pol = &default_policy; /* indicates default behavior */
974 if (flags & MPOL_F_NODE) {
975 if (flags & MPOL_F_ADDR) {
977 * Take a refcount on the mpol, lookup_node()
978 * will drop the mmap_lock, so after calling
979 * lookup_node() only "pol" remains valid, "vma"
985 err = lookup_node(mm, addr);
989 } else if (pol == current->mempolicy &&
990 pol->mode == MPOL_INTERLEAVE) {
991 *policy = next_node_in(current->il_prev, pol->nodes);
997 *policy = pol == &default_policy ? MPOL_DEFAULT :
1000 * Internal mempolicy flags must be masked off before exposing
1001 * the policy to userspace.
1003 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1008 if (mpol_store_user_nodemask(pol)) {
1009 *nmask = pol->w.user_nodemask;
1012 get_policy_nodemask(pol, nmask);
1013 task_unlock(current);
1020 mmap_read_unlock(mm);
1022 mpol_put(pol_refcount);
1026 #ifdef CONFIG_MIGRATION
1028 * page migration, thp tail pages can be passed.
1030 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1031 unsigned long flags)
1033 struct page *head = compound_head(page);
1035 * Avoid migrating a page that is shared with others.
1037 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1038 if (!isolate_lru_page(head)) {
1039 list_add_tail(&head->lru, pagelist);
1040 mod_node_page_state(page_pgdat(head),
1041 NR_ISOLATED_ANON + page_is_file_lru(head),
1042 thp_nr_pages(head));
1043 } else if (flags & MPOL_MF_STRICT) {
1045 * Non-movable page may reach here. And, there may be
1046 * temporary off LRU pages or non-LRU movable pages.
1047 * Treat them as unmovable pages since they can't be
1048 * isolated, so they can't be moved at the moment. It
1049 * should return -EIO for this case too.
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1062 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1066 LIST_HEAD(pagelist);
1068 struct migration_target_control mtc = {
1070 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1074 node_set(source, nmask);
1077 * This does not "check" the range but isolates all pages that
1078 * need migration. Between passing in the full user address
1079 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1081 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1082 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1083 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1085 if (!list_empty(&pagelist)) {
1086 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1087 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1089 putback_movable_pages(&pagelist);
1096 * Move pages between the two nodesets so as to preserve the physical
1097 * layout as much as possible.
1099 * Returns the number of page that could not be moved.
1101 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1102 const nodemask_t *to, int flags)
1108 lru_cache_disable();
1113 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1114 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1115 * bit in 'tmp', and return that <source, dest> pair for migration.
1116 * The pair of nodemasks 'to' and 'from' define the map.
1118 * If no pair of bits is found that way, fallback to picking some
1119 * pair of 'source' and 'dest' bits that are not the same. If the
1120 * 'source' and 'dest' bits are the same, this represents a node
1121 * that will be migrating to itself, so no pages need move.
1123 * If no bits are left in 'tmp', or if all remaining bits left
1124 * in 'tmp' correspond to the same bit in 'to', return false
1125 * (nothing left to migrate).
1127 * This lets us pick a pair of nodes to migrate between, such that
1128 * if possible the dest node is not already occupied by some other
1129 * source node, minimizing the risk of overloading the memory on a
1130 * node that would happen if we migrated incoming memory to a node
1131 * before migrating outgoing memory source that same node.
1133 * A single scan of tmp is sufficient. As we go, we remember the
1134 * most recent <s, d> pair that moved (s != d). If we find a pair
1135 * that not only moved, but what's better, moved to an empty slot
1136 * (d is not set in tmp), then we break out then, with that pair.
1137 * Otherwise when we finish scanning from_tmp, we at least have the
1138 * most recent <s, d> pair that moved. If we get all the way through
1139 * the scan of tmp without finding any node that moved, much less
1140 * moved to an empty node, then there is nothing left worth migrating.
1144 while (!nodes_empty(tmp)) {
1146 int source = NUMA_NO_NODE;
1149 for_each_node_mask(s, tmp) {
1152 * do_migrate_pages() tries to maintain the relative
1153 * node relationship of the pages established between
1154 * threads and memory areas.
1156 * However if the number of source nodes is not equal to
1157 * the number of destination nodes we can not preserve
1158 * this node relative relationship. In that case, skip
1159 * copying memory from a node that is in the destination
1162 * Example: [2,3,4] -> [3,4,5] moves everything.
1163 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1166 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1167 (node_isset(s, *to)))
1170 d = node_remap(s, *from, *to);
1174 source = s; /* Node moved. Memorize */
1177 /* dest not in remaining from nodes? */
1178 if (!node_isset(dest, tmp))
1181 if (source == NUMA_NO_NODE)
1184 node_clear(source, tmp);
1185 err = migrate_to_node(mm, source, dest, flags);
1191 mmap_read_unlock(mm);
1201 * Allocate a new page for page migration based on vma policy.
1202 * Start by assuming the page is mapped by the same vma as contains @start.
1203 * Search forward from there, if not. N.B., this assumes that the
1204 * list of pages handed to migrate_pages()--which is how we get here--
1205 * is in virtual address order.
1207 static struct page *new_page(struct page *page, unsigned long start)
1209 struct vm_area_struct *vma;
1210 unsigned long address;
1212 vma = find_vma(current->mm, start);
1214 address = page_address_in_vma(page, vma);
1215 if (address != -EFAULT)
1220 if (PageHuge(page)) {
1221 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1223 } else if (PageTransHuge(page)) {
1226 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1230 prep_transhuge_page(thp);
1234 * if !vma, alloc_page_vma() will use task or system default policy
1236 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1241 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1242 unsigned long flags)
1247 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1248 const nodemask_t *to, int flags)
1253 static struct page *new_page(struct page *page, unsigned long start)
1259 static long do_mbind(unsigned long start, unsigned long len,
1260 unsigned short mode, unsigned short mode_flags,
1261 nodemask_t *nmask, unsigned long flags)
1263 struct mm_struct *mm = current->mm;
1264 struct mempolicy *new;
1268 LIST_HEAD(pagelist);
1270 if (flags & ~(unsigned long)MPOL_MF_VALID)
1272 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1275 if (start & ~PAGE_MASK)
1278 if (mode == MPOL_DEFAULT)
1279 flags &= ~MPOL_MF_STRICT;
1281 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1289 new = mpol_new(mode, mode_flags, nmask);
1291 return PTR_ERR(new);
1293 if (flags & MPOL_MF_LAZY)
1294 new->flags |= MPOL_F_MOF;
1297 * If we are using the default policy then operation
1298 * on discontinuous address spaces is okay after all
1301 flags |= MPOL_MF_DISCONTIG_OK;
1303 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1304 start, start + len, mode, mode_flags,
1305 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1307 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1309 lru_cache_disable();
1312 NODEMASK_SCRATCH(scratch);
1314 mmap_write_lock(mm);
1315 err = mpol_set_nodemask(new, nmask, scratch);
1317 mmap_write_unlock(mm);
1320 NODEMASK_SCRATCH_FREE(scratch);
1325 ret = queue_pages_range(mm, start, end, nmask,
1326 flags | MPOL_MF_INVERT, &pagelist);
1333 err = mbind_range(mm, start, end, new);
1338 if (!list_empty(&pagelist)) {
1339 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1340 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1341 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1343 putback_movable_pages(&pagelist);
1346 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1350 if (!list_empty(&pagelist))
1351 putback_movable_pages(&pagelist);
1354 mmap_write_unlock(mm);
1357 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1363 * User space interface with variable sized bitmaps for nodelists.
1365 static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1366 unsigned long maxnode)
1368 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1371 if (in_compat_syscall())
1372 ret = compat_get_bitmap(mask,
1373 (const compat_ulong_t __user *)nmask,
1376 ret = copy_from_user(mask, nmask,
1377 nlongs * sizeof(unsigned long));
1382 if (maxnode % BITS_PER_LONG)
1383 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1388 /* Copy a node mask from user space. */
1389 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1390 unsigned long maxnode)
1393 nodes_clear(*nodes);
1394 if (maxnode == 0 || !nmask)
1396 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1400 * When the user specified more nodes than supported just check
1401 * if the non supported part is all zero, one word at a time,
1402 * starting at the end.
1404 while (maxnode > MAX_NUMNODES) {
1405 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1408 if (get_bitmap(&t, &nmask[maxnode / BITS_PER_LONG], bits))
1411 if (maxnode - bits >= MAX_NUMNODES) {
1414 maxnode = MAX_NUMNODES;
1415 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1421 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1424 /* Copy a kernel node mask to user space */
1425 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1428 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1429 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1430 bool compat = in_compat_syscall();
1433 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1435 if (copy > nbytes) {
1436 if (copy > PAGE_SIZE)
1438 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1441 maxnode = nr_node_ids;
1445 return compat_put_bitmap((compat_ulong_t __user *)mask,
1446 nodes_addr(*nodes), maxnode);
1448 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1451 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1452 static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1454 *flags = *mode & MPOL_MODE_FLAGS;
1455 *mode &= ~MPOL_MODE_FLAGS;
1457 if ((unsigned int)(*mode) >= MPOL_MAX)
1459 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1465 static long kernel_mbind(unsigned long start, unsigned long len,
1466 unsigned long mode, const unsigned long __user *nmask,
1467 unsigned long maxnode, unsigned int flags)
1469 unsigned short mode_flags;
1474 start = untagged_addr(start);
1475 err = sanitize_mpol_flags(&lmode, &mode_flags);
1479 err = get_nodes(&nodes, nmask, maxnode);
1483 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1486 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1487 unsigned long, mode, const unsigned long __user *, nmask,
1488 unsigned long, maxnode, unsigned int, flags)
1490 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1493 /* Set the process memory policy */
1494 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1495 unsigned long maxnode)
1497 unsigned short mode_flags;
1502 err = sanitize_mpol_flags(&lmode, &mode_flags);
1506 err = get_nodes(&nodes, nmask, maxnode);
1510 return do_set_mempolicy(lmode, mode_flags, &nodes);
1513 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1514 unsigned long, maxnode)
1516 return kernel_set_mempolicy(mode, nmask, maxnode);
1519 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1520 const unsigned long __user *old_nodes,
1521 const unsigned long __user *new_nodes)
1523 struct mm_struct *mm = NULL;
1524 struct task_struct *task;
1525 nodemask_t task_nodes;
1529 NODEMASK_SCRATCH(scratch);
1534 old = &scratch->mask1;
1535 new = &scratch->mask2;
1537 err = get_nodes(old, old_nodes, maxnode);
1541 err = get_nodes(new, new_nodes, maxnode);
1545 /* Find the mm_struct */
1547 task = pid ? find_task_by_vpid(pid) : current;
1553 get_task_struct(task);
1558 * Check if this process has the right to modify the specified process.
1559 * Use the regular "ptrace_may_access()" checks.
1561 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1568 task_nodes = cpuset_mems_allowed(task);
1569 /* Is the user allowed to access the target nodes? */
1570 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1575 task_nodes = cpuset_mems_allowed(current);
1576 nodes_and(*new, *new, task_nodes);
1577 if (nodes_empty(*new))
1580 err = security_task_movememory(task);
1584 mm = get_task_mm(task);
1585 put_task_struct(task);
1592 err = do_migrate_pages(mm, old, new,
1593 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1597 NODEMASK_SCRATCH_FREE(scratch);
1602 put_task_struct(task);
1607 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1608 const unsigned long __user *, old_nodes,
1609 const unsigned long __user *, new_nodes)
1611 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1615 /* Retrieve NUMA policy */
1616 static int kernel_get_mempolicy(int __user *policy,
1617 unsigned long __user *nmask,
1618 unsigned long maxnode,
1620 unsigned long flags)
1626 if (nmask != NULL && maxnode < nr_node_ids)
1629 addr = untagged_addr(addr);
1631 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1636 if (policy && put_user(pval, policy))
1640 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1645 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1646 unsigned long __user *, nmask, unsigned long, maxnode,
1647 unsigned long, addr, unsigned long, flags)
1649 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1652 bool vma_migratable(struct vm_area_struct *vma)
1654 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1658 * DAX device mappings require predictable access latency, so avoid
1659 * incurring periodic faults.
1661 if (vma_is_dax(vma))
1664 if (is_vm_hugetlb_page(vma) &&
1665 !hugepage_migration_supported(hstate_vma(vma)))
1669 * Migration allocates pages in the highest zone. If we cannot
1670 * do so then migration (at least from node to node) is not
1674 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1680 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1683 struct mempolicy *pol = NULL;
1686 if (vma->vm_ops && vma->vm_ops->get_policy) {
1687 pol = vma->vm_ops->get_policy(vma, addr);
1688 } else if (vma->vm_policy) {
1689 pol = vma->vm_policy;
1692 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1693 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1694 * count on these policies which will be dropped by
1695 * mpol_cond_put() later
1697 if (mpol_needs_cond_ref(pol))
1706 * get_vma_policy(@vma, @addr)
1707 * @vma: virtual memory area whose policy is sought
1708 * @addr: address in @vma for shared policy lookup
1710 * Returns effective policy for a VMA at specified address.
1711 * Falls back to current->mempolicy or system default policy, as necessary.
1712 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1713 * count--added by the get_policy() vm_op, as appropriate--to protect against
1714 * freeing by another task. It is the caller's responsibility to free the
1715 * extra reference for shared policies.
1717 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1720 struct mempolicy *pol = __get_vma_policy(vma, addr);
1723 pol = get_task_policy(current);
1728 bool vma_policy_mof(struct vm_area_struct *vma)
1730 struct mempolicy *pol;
1732 if (vma->vm_ops && vma->vm_ops->get_policy) {
1735 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1736 if (pol && (pol->flags & MPOL_F_MOF))
1743 pol = vma->vm_policy;
1745 pol = get_task_policy(current);
1747 return pol->flags & MPOL_F_MOF;
1750 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1752 enum zone_type dynamic_policy_zone = policy_zone;
1754 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1757 * if policy->nodes has movable memory only,
1758 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1760 * policy->nodes is intersect with node_states[N_MEMORY].
1761 * so if the following test fails, it implies
1762 * policy->nodes has movable memory only.
1764 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1765 dynamic_policy_zone = ZONE_MOVABLE;
1767 return zone >= dynamic_policy_zone;
1771 * Return a nodemask representing a mempolicy for filtering nodes for
1774 nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1776 int mode = policy->mode;
1778 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1779 if (unlikely(mode == MPOL_BIND) &&
1780 apply_policy_zone(policy, gfp_zone(gfp)) &&
1781 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1782 return &policy->nodes;
1784 if (mode == MPOL_PREFERRED_MANY)
1785 return &policy->nodes;
1791 * Return the preferred node id for 'prefer' mempolicy, and return
1792 * the given id for all other policies.
1794 * policy_node() is always coupled with policy_nodemask(), which
1795 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1797 static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1799 if (policy->mode == MPOL_PREFERRED) {
1800 nd = first_node(policy->nodes);
1803 * __GFP_THISNODE shouldn't even be used with the bind policy
1804 * because we might easily break the expectation to stay on the
1805 * requested node and not break the policy.
1807 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1813 /* Do dynamic interleaving for a process */
1814 static unsigned interleave_nodes(struct mempolicy *policy)
1817 struct task_struct *me = current;
1819 next = next_node_in(me->il_prev, policy->nodes);
1820 if (next < MAX_NUMNODES)
1826 * Depending on the memory policy provide a node from which to allocate the
1829 unsigned int mempolicy_slab_node(void)
1831 struct mempolicy *policy;
1832 int node = numa_mem_id();
1837 policy = current->mempolicy;
1841 switch (policy->mode) {
1842 case MPOL_PREFERRED:
1843 return first_node(policy->nodes);
1845 case MPOL_INTERLEAVE:
1846 return interleave_nodes(policy);
1849 case MPOL_PREFERRED_MANY:
1854 * Follow bind policy behavior and start allocation at the
1857 struct zonelist *zonelist;
1858 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1859 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1860 z = first_zones_zonelist(zonelist, highest_zoneidx,
1862 return z->zone ? zone_to_nid(z->zone) : node;
1873 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1874 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1875 * number of present nodes.
1877 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1879 unsigned nnodes = nodes_weight(pol->nodes);
1885 return numa_node_id();
1886 target = (unsigned int)n % nnodes;
1887 nid = first_node(pol->nodes);
1888 for (i = 0; i < target; i++)
1889 nid = next_node(nid, pol->nodes);
1893 /* Determine a node number for interleave */
1894 static inline unsigned interleave_nid(struct mempolicy *pol,
1895 struct vm_area_struct *vma, unsigned long addr, int shift)
1901 * for small pages, there is no difference between
1902 * shift and PAGE_SHIFT, so the bit-shift is safe.
1903 * for huge pages, since vm_pgoff is in units of small
1904 * pages, we need to shift off the always 0 bits to get
1907 BUG_ON(shift < PAGE_SHIFT);
1908 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1909 off += (addr - vma->vm_start) >> shift;
1910 return offset_il_node(pol, off);
1912 return interleave_nodes(pol);
1915 #ifdef CONFIG_HUGETLBFS
1917 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1918 * @vma: virtual memory area whose policy is sought
1919 * @addr: address in @vma for shared policy lookup and interleave policy
1920 * @gfp_flags: for requested zone
1921 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1922 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1924 * Returns a nid suitable for a huge page allocation and a pointer
1925 * to the struct mempolicy for conditional unref after allocation.
1926 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
1927 * to the mempolicy's @nodemask for filtering the zonelist.
1929 * Must be protected by read_mems_allowed_begin()
1931 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1932 struct mempolicy **mpol, nodemask_t **nodemask)
1937 *mpol = get_vma_policy(vma, addr);
1939 mode = (*mpol)->mode;
1941 if (unlikely(mode == MPOL_INTERLEAVE)) {
1942 nid = interleave_nid(*mpol, vma, addr,
1943 huge_page_shift(hstate_vma(vma)));
1945 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1946 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
1947 *nodemask = &(*mpol)->nodes;
1953 * init_nodemask_of_mempolicy
1955 * If the current task's mempolicy is "default" [NULL], return 'false'
1956 * to indicate default policy. Otherwise, extract the policy nodemask
1957 * for 'bind' or 'interleave' policy into the argument nodemask, or
1958 * initialize the argument nodemask to contain the single node for
1959 * 'preferred' or 'local' policy and return 'true' to indicate presence
1960 * of non-default mempolicy.
1962 * We don't bother with reference counting the mempolicy [mpol_get/put]
1963 * because the current task is examining it's own mempolicy and a task's
1964 * mempolicy is only ever changed by the task itself.
1966 * N.B., it is the caller's responsibility to free a returned nodemask.
1968 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1970 struct mempolicy *mempolicy;
1972 if (!(mask && current->mempolicy))
1976 mempolicy = current->mempolicy;
1977 switch (mempolicy->mode) {
1978 case MPOL_PREFERRED:
1979 case MPOL_PREFERRED_MANY:
1981 case MPOL_INTERLEAVE:
1982 *mask = mempolicy->nodes;
1986 init_nodemask_of_node(mask, numa_node_id());
1992 task_unlock(current);
1999 * mempolicy_in_oom_domain
2001 * If tsk's mempolicy is "bind", check for intersection between mask and
2002 * the policy nodemask. Otherwise, return true for all other policies
2003 * including "interleave", as a tsk with "interleave" policy may have
2004 * memory allocated from all nodes in system.
2006 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2008 bool mempolicy_in_oom_domain(struct task_struct *tsk,
2009 const nodemask_t *mask)
2011 struct mempolicy *mempolicy;
2018 mempolicy = tsk->mempolicy;
2019 if (mempolicy && mempolicy->mode == MPOL_BIND)
2020 ret = nodes_intersects(mempolicy->nodes, *mask);
2026 /* Allocate a page in interleaved policy.
2027 Own path because it needs to do special accounting. */
2028 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2033 page = __alloc_pages(gfp, order, nid, NULL);
2034 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2035 if (!static_branch_likely(&vm_numa_stat_key))
2037 if (page && page_to_nid(page) == nid) {
2039 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2045 static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2046 int nid, struct mempolicy *pol)
2049 gfp_t preferred_gfp;
2052 * This is a two pass approach. The first pass will only try the
2053 * preferred nodes but skip the direct reclaim and allow the
2054 * allocation to fail, while the second pass will try all the
2057 preferred_gfp = gfp | __GFP_NOWARN;
2058 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2059 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2061 page = __alloc_pages(gfp, order, numa_node_id(), NULL);
2067 * alloc_pages_vma - Allocate a page for a VMA.
2069 * @order: Order of the GFP allocation.
2070 * @vma: Pointer to VMA or NULL if not available.
2071 * @addr: Virtual address of the allocation. Must be inside @vma.
2072 * @node: Which node to prefer for allocation (modulo policy).
2073 * @hugepage: For hugepages try only the preferred node if possible.
2075 * Allocate a page for a specific address in @vma, using the appropriate
2076 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2077 * of the mm_struct of the VMA to prevent it from going away. Should be
2078 * used for all allocations for pages that will be mapped into user space.
2080 * Return: The page on success or NULL if allocation fails.
2082 struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2083 unsigned long addr, int node, bool hugepage)
2085 struct mempolicy *pol;
2090 pol = get_vma_policy(vma, addr);
2092 if (pol->mode == MPOL_INTERLEAVE) {
2095 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2097 page = alloc_page_interleave(gfp, order, nid);
2101 if (pol->mode == MPOL_PREFERRED_MANY) {
2102 page = alloc_pages_preferred_many(gfp, order, node, pol);
2107 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2108 int hpage_node = node;
2111 * For hugepage allocation and non-interleave policy which
2112 * allows the current node (or other explicitly preferred
2113 * node) we only try to allocate from the current/preferred
2114 * node and don't fall back to other nodes, as the cost of
2115 * remote accesses would likely offset THP benefits.
2117 * If the policy is interleave or does not allow the current
2118 * node in its nodemask, we allocate the standard way.
2120 if (pol->mode == MPOL_PREFERRED)
2121 hpage_node = first_node(pol->nodes);
2123 nmask = policy_nodemask(gfp, pol);
2124 if (!nmask || node_isset(hpage_node, *nmask)) {
2127 * First, try to allocate THP only on local node, but
2128 * don't reclaim unnecessarily, just compact.
2130 page = __alloc_pages_node(hpage_node,
2131 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2134 * If hugepage allocations are configured to always
2135 * synchronous compact or the vma has been madvised
2136 * to prefer hugepage backing, retry allowing remote
2137 * memory with both reclaim and compact as well.
2139 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2140 page = __alloc_pages_node(hpage_node,
2147 nmask = policy_nodemask(gfp, pol);
2148 preferred_nid = policy_node(gfp, pol, node);
2149 page = __alloc_pages(gfp, order, preferred_nid, nmask);
2154 EXPORT_SYMBOL(alloc_pages_vma);
2157 * alloc_pages - Allocate pages.
2159 * @order: Power of two of number of pages to allocate.
2161 * Allocate 1 << @order contiguous pages. The physical address of the
2162 * first page is naturally aligned (eg an order-3 allocation will be aligned
2163 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2164 * process is honoured when in process context.
2166 * Context: Can be called from any context, providing the appropriate GFP
2168 * Return: The page on success or NULL if allocation fails.
2170 struct page *alloc_pages(gfp_t gfp, unsigned order)
2172 struct mempolicy *pol = &default_policy;
2175 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2176 pol = get_task_policy(current);
2179 * No reference counting needed for current->mempolicy
2180 * nor system default_policy
2182 if (pol->mode == MPOL_INTERLEAVE)
2183 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2184 else if (pol->mode == MPOL_PREFERRED_MANY)
2185 page = alloc_pages_preferred_many(gfp, order,
2186 numa_node_id(), pol);
2188 page = __alloc_pages(gfp, order,
2189 policy_node(gfp, pol, numa_node_id()),
2190 policy_nodemask(gfp, pol));
2194 EXPORT_SYMBOL(alloc_pages);
2196 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2198 struct mempolicy *pol = mpol_dup(vma_policy(src));
2201 return PTR_ERR(pol);
2202 dst->vm_policy = pol;
2207 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2208 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2209 * with the mems_allowed returned by cpuset_mems_allowed(). This
2210 * keeps mempolicies cpuset relative after its cpuset moves. See
2211 * further kernel/cpuset.c update_nodemask().
2213 * current's mempolicy may be rebinded by the other task(the task that changes
2214 * cpuset's mems), so we needn't do rebind work for current task.
2217 /* Slow path of a mempolicy duplicate */
2218 struct mempolicy *__mpol_dup(struct mempolicy *old)
2220 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2223 return ERR_PTR(-ENOMEM);
2225 /* task's mempolicy is protected by alloc_lock */
2226 if (old == current->mempolicy) {
2229 task_unlock(current);
2233 if (current_cpuset_is_being_rebound()) {
2234 nodemask_t mems = cpuset_mems_allowed(current);
2235 mpol_rebind_policy(new, &mems);
2237 atomic_set(&new->refcnt, 1);
2241 /* Slow path of a mempolicy comparison */
2242 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2246 if (a->mode != b->mode)
2248 if (a->flags != b->flags)
2250 if (mpol_store_user_nodemask(a))
2251 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2256 case MPOL_INTERLEAVE:
2257 case MPOL_PREFERRED:
2258 case MPOL_PREFERRED_MANY:
2259 return !!nodes_equal(a->nodes, b->nodes);
2269 * Shared memory backing store policy support.
2271 * Remember policies even when nobody has shared memory mapped.
2272 * The policies are kept in Red-Black tree linked from the inode.
2273 * They are protected by the sp->lock rwlock, which should be held
2274 * for any accesses to the tree.
2278 * lookup first element intersecting start-end. Caller holds sp->lock for
2279 * reading or for writing
2281 static struct sp_node *
2282 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2284 struct rb_node *n = sp->root.rb_node;
2287 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2289 if (start >= p->end)
2291 else if (end <= p->start)
2299 struct sp_node *w = NULL;
2300 struct rb_node *prev = rb_prev(n);
2303 w = rb_entry(prev, struct sp_node, nd);
2304 if (w->end <= start)
2308 return rb_entry(n, struct sp_node, nd);
2312 * Insert a new shared policy into the list. Caller holds sp->lock for
2315 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2317 struct rb_node **p = &sp->root.rb_node;
2318 struct rb_node *parent = NULL;
2323 nd = rb_entry(parent, struct sp_node, nd);
2324 if (new->start < nd->start)
2326 else if (new->end > nd->end)
2327 p = &(*p)->rb_right;
2331 rb_link_node(&new->nd, parent, p);
2332 rb_insert_color(&new->nd, &sp->root);
2333 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2334 new->policy ? new->policy->mode : 0);
2337 /* Find shared policy intersecting idx */
2339 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2341 struct mempolicy *pol = NULL;
2344 if (!sp->root.rb_node)
2346 read_lock(&sp->lock);
2347 sn = sp_lookup(sp, idx, idx+1);
2349 mpol_get(sn->policy);
2352 read_unlock(&sp->lock);
2356 static void sp_free(struct sp_node *n)
2358 mpol_put(n->policy);
2359 kmem_cache_free(sn_cache, n);
2363 * mpol_misplaced - check whether current page node is valid in policy
2365 * @page: page to be checked
2366 * @vma: vm area where page mapped
2367 * @addr: virtual address where page mapped
2369 * Lookup current policy node id for vma,addr and "compare to" page's
2370 * node id. Policy determination "mimics" alloc_page_vma().
2371 * Called from fault path where we know the vma and faulting address.
2373 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2374 * policy, or a suitable node ID to allocate a replacement page from.
2376 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2378 struct mempolicy *pol;
2380 int curnid = page_to_nid(page);
2381 unsigned long pgoff;
2382 int thiscpu = raw_smp_processor_id();
2383 int thisnid = cpu_to_node(thiscpu);
2384 int polnid = NUMA_NO_NODE;
2385 int ret = NUMA_NO_NODE;
2387 pol = get_vma_policy(vma, addr);
2388 if (!(pol->flags & MPOL_F_MOF))
2391 switch (pol->mode) {
2392 case MPOL_INTERLEAVE:
2393 pgoff = vma->vm_pgoff;
2394 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2395 polnid = offset_il_node(pol, pgoff);
2398 case MPOL_PREFERRED:
2399 if (node_isset(curnid, pol->nodes))
2401 polnid = first_node(pol->nodes);
2405 polnid = numa_node_id();
2409 /* Optimize placement among multiple nodes via NUMA balancing */
2410 if (pol->flags & MPOL_F_MORON) {
2411 if (node_isset(thisnid, pol->nodes))
2417 case MPOL_PREFERRED_MANY:
2419 * use current page if in policy nodemask,
2420 * else select nearest allowed node, if any.
2421 * If no allowed nodes, use current [!misplaced].
2423 if (node_isset(curnid, pol->nodes))
2425 z = first_zones_zonelist(
2426 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2427 gfp_zone(GFP_HIGHUSER),
2429 polnid = zone_to_nid(z->zone);
2436 /* Migrate the page towards the node whose CPU is referencing it */
2437 if (pol->flags & MPOL_F_MORON) {
2440 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2444 if (curnid != polnid)
2453 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2454 * dropped after task->mempolicy is set to NULL so that any allocation done as
2455 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2458 void mpol_put_task_policy(struct task_struct *task)
2460 struct mempolicy *pol;
2463 pol = task->mempolicy;
2464 task->mempolicy = NULL;
2469 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2471 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2472 rb_erase(&n->nd, &sp->root);
2476 static void sp_node_init(struct sp_node *node, unsigned long start,
2477 unsigned long end, struct mempolicy *pol)
2479 node->start = start;
2484 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2485 struct mempolicy *pol)
2488 struct mempolicy *newpol;
2490 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2494 newpol = mpol_dup(pol);
2495 if (IS_ERR(newpol)) {
2496 kmem_cache_free(sn_cache, n);
2499 newpol->flags |= MPOL_F_SHARED;
2500 sp_node_init(n, start, end, newpol);
2505 /* Replace a policy range. */
2506 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2507 unsigned long end, struct sp_node *new)
2510 struct sp_node *n_new = NULL;
2511 struct mempolicy *mpol_new = NULL;
2515 write_lock(&sp->lock);
2516 n = sp_lookup(sp, start, end);
2517 /* Take care of old policies in the same range. */
2518 while (n && n->start < end) {
2519 struct rb_node *next = rb_next(&n->nd);
2520 if (n->start >= start) {
2526 /* Old policy spanning whole new range. */
2531 *mpol_new = *n->policy;
2532 atomic_set(&mpol_new->refcnt, 1);
2533 sp_node_init(n_new, end, n->end, mpol_new);
2535 sp_insert(sp, n_new);
2544 n = rb_entry(next, struct sp_node, nd);
2548 write_unlock(&sp->lock);
2555 kmem_cache_free(sn_cache, n_new);
2560 write_unlock(&sp->lock);
2562 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2565 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
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->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 .nodes = nodemask_of_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
2773 static const char * const policy_modes[] =
2775 [MPOL_DEFAULT] = "default",
2776 [MPOL_PREFERRED] = "prefer",
2777 [MPOL_BIND] = "bind",
2778 [MPOL_INTERLEAVE] = "interleave",
2779 [MPOL_LOCAL] = "local",
2780 [MPOL_PREFERRED_MANY] = "prefer (many)",
2786 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2787 * @str: string containing mempolicy to parse
2788 * @mpol: pointer to struct mempolicy pointer, returned on success.
2791 * <mode>[=<flags>][:<nodelist>]
2793 * On success, returns 0, else 1
2795 int mpol_parse_str(char *str, struct mempolicy **mpol)
2797 struct mempolicy *new = NULL;
2798 unsigned short mode_flags;
2800 char *nodelist = strchr(str, ':');
2801 char *flags = strchr(str, '=');
2805 *flags++ = '\0'; /* terminate mode string */
2808 /* NUL-terminate mode or flags string */
2810 if (nodelist_parse(nodelist, nodes))
2812 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2817 mode = match_string(policy_modes, MPOL_MAX, str);
2822 case MPOL_PREFERRED:
2824 * Insist on a nodelist of one node only, although later
2825 * we use first_node(nodes) to grab a single node, so here
2826 * nodelist (or nodes) cannot be empty.
2829 char *rest = nodelist;
2830 while (isdigit(*rest))
2834 if (nodes_empty(nodes))
2838 case MPOL_INTERLEAVE:
2840 * Default to online nodes with memory if no nodelist
2843 nodes = node_states[N_MEMORY];
2847 * Don't allow a nodelist; mpol_new() checks flags
2854 * Insist on a empty nodelist
2859 case MPOL_PREFERRED_MANY:
2862 * Insist on a nodelist
2871 * Currently, we only support two mutually exclusive
2874 if (!strcmp(flags, "static"))
2875 mode_flags |= MPOL_F_STATIC_NODES;
2876 else if (!strcmp(flags, "relative"))
2877 mode_flags |= MPOL_F_RELATIVE_NODES;
2882 new = mpol_new(mode, mode_flags, &nodes);
2887 * Save nodes for mpol_to_str() to show the tmpfs mount options
2888 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2890 if (mode != MPOL_PREFERRED) {
2892 } else if (nodelist) {
2893 nodes_clear(new->nodes);
2894 node_set(first_node(nodes), new->nodes);
2896 new->mode = MPOL_LOCAL;
2900 * Save nodes for contextualization: this will be used to "clone"
2901 * the mempolicy in a specific context [cpuset] at a later time.
2903 new->w.user_nodemask = nodes;
2908 /* Restore string for error message */
2917 #endif /* CONFIG_TMPFS */
2920 * mpol_to_str - format a mempolicy structure for printing
2921 * @buffer: to contain formatted mempolicy string
2922 * @maxlen: length of @buffer
2923 * @pol: pointer to mempolicy to be formatted
2925 * Convert @pol into a string. If @buffer is too short, truncate the string.
2926 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2927 * longest flag, "relative", and to display at least a few node ids.
2929 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2932 nodemask_t nodes = NODE_MASK_NONE;
2933 unsigned short mode = MPOL_DEFAULT;
2934 unsigned short flags = 0;
2936 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2945 case MPOL_PREFERRED:
2946 case MPOL_PREFERRED_MANY:
2948 case MPOL_INTERLEAVE:
2953 snprintf(p, maxlen, "unknown");
2957 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2959 if (flags & MPOL_MODE_FLAGS) {
2960 p += snprintf(p, buffer + maxlen - p, "=");
2963 * Currently, the only defined flags are mutually exclusive
2965 if (flags & MPOL_F_STATIC_NODES)
2966 p += snprintf(p, buffer + maxlen - p, "static");
2967 else if (flags & MPOL_F_RELATIVE_NODES)
2968 p += snprintf(p, buffer + maxlen - p, "relative");
2971 if (!nodes_empty(nodes))
2972 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2973 nodemask_pr_args(&nodes));
2976 bool numa_demotion_enabled = false;
2979 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2980 struct kobj_attribute *attr, char *buf)
2982 return sysfs_emit(buf, "%s\n",
2983 numa_demotion_enabled? "true" : "false");
2986 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2987 struct kobj_attribute *attr,
2988 const char *buf, size_t count)
2990 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
2991 numa_demotion_enabled = true;
2992 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
2993 numa_demotion_enabled = false;
3000 static struct kobj_attribute numa_demotion_enabled_attr =
3001 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3002 numa_demotion_enabled_store);
3004 static struct attribute *numa_attrs[] = {
3005 &numa_demotion_enabled_attr.attr,
3009 static const struct attribute_group numa_attr_group = {
3010 .attrs = numa_attrs,
3013 static int __init numa_init_sysfs(void)
3016 struct kobject *numa_kobj;
3018 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3020 pr_err("failed to create numa kobject\n");
3023 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3025 pr_err("failed to register numa group\n");
3031 kobject_put(numa_kobj);
3034 subsys_initcall(numa_init_sysfs);