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 = tmp;
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) &&
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 void 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);
431 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
432 unsigned long end, struct mm_walk *walk)
436 struct queue_pages *qp = walk->private;
439 if (unlikely(is_pmd_migration_entry(*pmd))) {
443 page = pmd_page(*pmd);
444 if (is_huge_zero_page(page)) {
446 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
449 if (!thp_migration_supported()) {
453 ret = split_huge_page(page);
458 if (!queue_pages_required(page, qp)) {
465 /* go to thp migration */
466 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
467 migrate_page_add(page, qp->pagelist, flags);
475 * Scan through pages checking if pages follow certain conditions,
476 * and move them to the pagelist if they do.
478 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
479 unsigned long end, struct mm_walk *walk)
481 struct vm_area_struct *vma = walk->vma;
483 struct queue_pages *qp = walk->private;
484 unsigned long flags = qp->flags;
489 ptl = pmd_trans_huge_lock(pmd, vma);
491 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
496 if (pmd_trans_unstable(pmd))
499 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
500 for (; addr != end; pte++, addr += PAGE_SIZE) {
501 if (!pte_present(*pte))
503 page = vm_normal_page(vma, addr, *pte);
507 * vm_normal_page() filters out zero pages, but there might
508 * still be PageReserved pages to skip, perhaps in a VDSO.
510 if (PageReserved(page))
512 if (!queue_pages_required(page, qp))
514 if (PageTransCompound(page) && !thp_migration_supported()) {
516 pte_unmap_unlock(pte, ptl);
518 ret = split_huge_page(page);
521 /* Failed to split -- skip. */
523 pte = pte_offset_map_lock(walk->mm, pmd,
530 migrate_page_add(page, qp->pagelist, flags);
532 pte_unmap_unlock(pte - 1, ptl);
537 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
538 unsigned long addr, unsigned long end,
539 struct mm_walk *walk)
541 #ifdef CONFIG_HUGETLB_PAGE
542 struct queue_pages *qp = walk->private;
543 unsigned long flags = qp->flags;
548 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
549 entry = huge_ptep_get(pte);
550 if (!pte_present(entry))
552 page = pte_page(entry);
553 if (!queue_pages_required(page, qp))
555 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
556 if (flags & (MPOL_MF_MOVE_ALL) ||
557 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
558 isolate_huge_page(page, qp->pagelist);
567 #ifdef CONFIG_NUMA_BALANCING
569 * This is used to mark a range of virtual addresses to be inaccessible.
570 * These are later cleared by a NUMA hinting fault. Depending on these
571 * faults, pages may be migrated for better NUMA placement.
573 * This is assuming that NUMA faults are handled using PROT_NONE. If
574 * an architecture makes a different choice, it will need further
575 * changes to the core.
577 unsigned long change_prot_numa(struct vm_area_struct *vma,
578 unsigned long addr, unsigned long end)
582 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
584 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
589 static unsigned long change_prot_numa(struct vm_area_struct *vma,
590 unsigned long addr, unsigned long end)
594 #endif /* CONFIG_NUMA_BALANCING */
596 static int queue_pages_test_walk(unsigned long start, unsigned long end,
597 struct mm_walk *walk)
599 struct vm_area_struct *vma = walk->vma;
600 struct queue_pages *qp = walk->private;
601 unsigned long endvma = vma->vm_end;
602 unsigned long flags = qp->flags;
604 if (!vma_migratable(vma))
609 if (vma->vm_start > start)
610 start = vma->vm_start;
612 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
613 if (!vma->vm_next && vma->vm_end < end)
615 if (qp->prev && qp->prev->vm_end < vma->vm_start)
621 if (flags & MPOL_MF_LAZY) {
622 /* Similar to task_numa_work, skip inaccessible VMAs */
623 if (!is_vm_hugetlb_page(vma) &&
624 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
625 !(vma->vm_flags & VM_MIXEDMAP))
626 change_prot_numa(vma, start, endvma);
630 /* queue pages from current vma */
631 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
637 * Walk through page tables and collect pages to be migrated.
639 * If pages found in a given range are on a set of nodes (determined by
640 * @nodes and @flags,) it's isolated and queued to the pagelist which is
641 * passed via @private.)
644 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
645 nodemask_t *nodes, unsigned long flags,
646 struct list_head *pagelist)
648 struct queue_pages qp = {
649 .pagelist = pagelist,
654 struct mm_walk queue_pages_walk = {
655 .hugetlb_entry = queue_pages_hugetlb,
656 .pmd_entry = queue_pages_pte_range,
657 .test_walk = queue_pages_test_walk,
662 return walk_page_range(start, end, &queue_pages_walk);
666 * Apply policy to a single VMA
667 * This must be called with the mmap_sem held for writing.
669 static int vma_replace_policy(struct vm_area_struct *vma,
670 struct mempolicy *pol)
673 struct mempolicy *old;
674 struct mempolicy *new;
676 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
677 vma->vm_start, vma->vm_end, vma->vm_pgoff,
678 vma->vm_ops, vma->vm_file,
679 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
685 if (vma->vm_ops && vma->vm_ops->set_policy) {
686 err = vma->vm_ops->set_policy(vma, new);
691 old = vma->vm_policy;
692 vma->vm_policy = new; /* protected by mmap_sem */
701 /* Step 2: apply policy to a range and do splits. */
702 static int mbind_range(struct mm_struct *mm, unsigned long start,
703 unsigned long end, struct mempolicy *new_pol)
705 struct vm_area_struct *next;
706 struct vm_area_struct *prev;
707 struct vm_area_struct *vma;
710 unsigned long vmstart;
713 vma = find_vma(mm, start);
714 if (!vma || vma->vm_start > start)
718 if (start > vma->vm_start)
721 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
723 vmstart = max(start, vma->vm_start);
724 vmend = min(end, vma->vm_end);
726 if (mpol_equal(vma_policy(vma), new_pol))
729 pgoff = vma->vm_pgoff +
730 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
731 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
732 vma->anon_vma, vma->vm_file, pgoff,
733 new_pol, vma->vm_userfaultfd_ctx);
737 if (mpol_equal(vma_policy(vma), new_pol))
739 /* vma_merge() joined vma && vma->next, case 8 */
742 if (vma->vm_start != vmstart) {
743 err = split_vma(vma->vm_mm, vma, vmstart, 1);
747 if (vma->vm_end != vmend) {
748 err = split_vma(vma->vm_mm, vma, vmend, 0);
753 err = vma_replace_policy(vma, new_pol);
762 /* Set the process memory policy */
763 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
766 struct mempolicy *new, *old;
767 NODEMASK_SCRATCH(scratch);
773 new = mpol_new(mode, flags, nodes);
780 ret = mpol_set_nodemask(new, nodes, scratch);
782 task_unlock(current);
786 old = current->mempolicy;
787 current->mempolicy = new;
788 if (new && new->mode == MPOL_INTERLEAVE)
789 current->il_prev = MAX_NUMNODES-1;
790 task_unlock(current);
794 NODEMASK_SCRATCH_FREE(scratch);
799 * Return nodemask for policy for get_mempolicy() query
801 * Called with task's alloc_lock held
803 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
806 if (p == &default_policy)
812 case MPOL_INTERLEAVE:
816 if (!(p->flags & MPOL_F_LOCAL))
817 node_set(p->v.preferred_node, *nodes);
818 /* else return empty node mask for local allocation */
825 static int lookup_node(unsigned long addr)
830 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
832 err = page_to_nid(p);
838 /* Retrieve NUMA policy */
839 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
840 unsigned long addr, unsigned long flags)
843 struct mm_struct *mm = current->mm;
844 struct vm_area_struct *vma = NULL;
845 struct mempolicy *pol = current->mempolicy;
848 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
851 if (flags & MPOL_F_MEMS_ALLOWED) {
852 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
854 *policy = 0; /* just so it's initialized */
856 *nmask = cpuset_current_mems_allowed;
857 task_unlock(current);
861 if (flags & MPOL_F_ADDR) {
863 * Do NOT fall back to task policy if the
864 * vma/shared policy at addr is NULL. We
865 * want to return MPOL_DEFAULT in this case.
867 down_read(&mm->mmap_sem);
868 vma = find_vma_intersection(mm, addr, addr+1);
870 up_read(&mm->mmap_sem);
873 if (vma->vm_ops && vma->vm_ops->get_policy)
874 pol = vma->vm_ops->get_policy(vma, addr);
876 pol = vma->vm_policy;
881 pol = &default_policy; /* indicates default behavior */
883 if (flags & MPOL_F_NODE) {
884 if (flags & MPOL_F_ADDR) {
885 err = lookup_node(addr);
889 } else if (pol == current->mempolicy &&
890 pol->mode == MPOL_INTERLEAVE) {
891 *policy = next_node_in(current->il_prev, pol->v.nodes);
897 *policy = pol == &default_policy ? MPOL_DEFAULT :
900 * Internal mempolicy flags must be masked off before exposing
901 * the policy to userspace.
903 *policy |= (pol->flags & MPOL_MODE_FLAGS);
908 if (mpol_store_user_nodemask(pol)) {
909 *nmask = pol->w.user_nodemask;
912 get_policy_nodemask(pol, nmask);
913 task_unlock(current);
920 up_read(¤t->mm->mmap_sem);
924 #ifdef CONFIG_MIGRATION
926 * page migration, thp tail pages can be passed.
928 static void migrate_page_add(struct page *page, struct list_head *pagelist,
931 struct page *head = compound_head(page);
933 * Avoid migrating a page that is shared with others.
935 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
936 if (!isolate_lru_page(head)) {
937 list_add_tail(&head->lru, pagelist);
938 mod_node_page_state(page_pgdat(head),
939 NR_ISOLATED_ANON + page_is_file_cache(head),
940 hpage_nr_pages(head));
945 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
948 return alloc_huge_page_node(page_hstate(compound_head(page)),
950 else if (thp_migration_supported() && PageTransHuge(page)) {
953 thp = alloc_pages_node(node,
954 (GFP_TRANSHUGE | __GFP_THISNODE),
958 prep_transhuge_page(thp);
961 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
966 * Migrate pages from one node to a target node.
967 * Returns error or the number of pages not migrated.
969 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
977 node_set(source, nmask);
980 * This does not "check" the range but isolates all pages that
981 * need migration. Between passing in the full user address
982 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
984 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
985 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
986 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
988 if (!list_empty(&pagelist)) {
989 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
990 MIGRATE_SYNC, MR_SYSCALL);
992 putback_movable_pages(&pagelist);
999 * Move pages between the two nodesets so as to preserve the physical
1000 * layout as much as possible.
1002 * Returns the number of page that could not be moved.
1004 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1005 const nodemask_t *to, int flags)
1011 err = migrate_prep();
1015 down_read(&mm->mmap_sem);
1018 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1019 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1020 * bit in 'tmp', and return that <source, dest> pair for migration.
1021 * The pair of nodemasks 'to' and 'from' define the map.
1023 * If no pair of bits is found that way, fallback to picking some
1024 * pair of 'source' and 'dest' bits that are not the same. If the
1025 * 'source' and 'dest' bits are the same, this represents a node
1026 * that will be migrating to itself, so no pages need move.
1028 * If no bits are left in 'tmp', or if all remaining bits left
1029 * in 'tmp' correspond to the same bit in 'to', return false
1030 * (nothing left to migrate).
1032 * This lets us pick a pair of nodes to migrate between, such that
1033 * if possible the dest node is not already occupied by some other
1034 * source node, minimizing the risk of overloading the memory on a
1035 * node that would happen if we migrated incoming memory to a node
1036 * before migrating outgoing memory source that same node.
1038 * A single scan of tmp is sufficient. As we go, we remember the
1039 * most recent <s, d> pair that moved (s != d). If we find a pair
1040 * that not only moved, but what's better, moved to an empty slot
1041 * (d is not set in tmp), then we break out then, with that pair.
1042 * Otherwise when we finish scanning from_tmp, we at least have the
1043 * most recent <s, d> pair that moved. If we get all the way through
1044 * the scan of tmp without finding any node that moved, much less
1045 * moved to an empty node, then there is nothing left worth migrating.
1049 while (!nodes_empty(tmp)) {
1051 int source = NUMA_NO_NODE;
1054 for_each_node_mask(s, tmp) {
1057 * do_migrate_pages() tries to maintain the relative
1058 * node relationship of the pages established between
1059 * threads and memory areas.
1061 * However if the number of source nodes is not equal to
1062 * the number of destination nodes we can not preserve
1063 * this node relative relationship. In that case, skip
1064 * copying memory from a node that is in the destination
1067 * Example: [2,3,4] -> [3,4,5] moves everything.
1068 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1071 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1072 (node_isset(s, *to)))
1075 d = node_remap(s, *from, *to);
1079 source = s; /* Node moved. Memorize */
1082 /* dest not in remaining from nodes? */
1083 if (!node_isset(dest, tmp))
1086 if (source == NUMA_NO_NODE)
1089 node_clear(source, tmp);
1090 err = migrate_to_node(mm, source, dest, flags);
1096 up_read(&mm->mmap_sem);
1104 * Allocate a new page for page migration based on vma policy.
1105 * Start by assuming the page is mapped by the same vma as contains @start.
1106 * Search forward from there, if not. N.B., this assumes that the
1107 * list of pages handed to migrate_pages()--which is how we get here--
1108 * is in virtual address order.
1110 static struct page *new_page(struct page *page, unsigned long start, int **x)
1112 struct vm_area_struct *vma;
1113 unsigned long uninitialized_var(address);
1115 vma = find_vma(current->mm, start);
1117 address = page_address_in_vma(page, vma);
1118 if (address != -EFAULT)
1123 if (PageHuge(page)) {
1124 return alloc_huge_page_vma(vma, address);
1125 } else if (thp_migration_supported() && PageTransHuge(page)) {
1128 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1132 prep_transhuge_page(thp);
1136 * if !vma, alloc_page_vma() will use task or system default policy
1138 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1143 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1144 unsigned long flags)
1148 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1149 const nodemask_t *to, int flags)
1154 static struct page *new_page(struct page *page, unsigned long start, int **x)
1160 static long do_mbind(unsigned long start, unsigned long len,
1161 unsigned short mode, unsigned short mode_flags,
1162 nodemask_t *nmask, unsigned long flags)
1164 struct mm_struct *mm = current->mm;
1165 struct mempolicy *new;
1168 LIST_HEAD(pagelist);
1170 if (flags & ~(unsigned long)MPOL_MF_VALID)
1172 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1175 if (start & ~PAGE_MASK)
1178 if (mode == MPOL_DEFAULT)
1179 flags &= ~MPOL_MF_STRICT;
1181 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1189 new = mpol_new(mode, mode_flags, nmask);
1191 return PTR_ERR(new);
1193 if (flags & MPOL_MF_LAZY)
1194 new->flags |= MPOL_F_MOF;
1197 * If we are using the default policy then operation
1198 * on discontinuous address spaces is okay after all
1201 flags |= MPOL_MF_DISCONTIG_OK;
1203 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1204 start, start + len, mode, mode_flags,
1205 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1207 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1209 err = migrate_prep();
1214 NODEMASK_SCRATCH(scratch);
1216 down_write(&mm->mmap_sem);
1218 err = mpol_set_nodemask(new, nmask, scratch);
1219 task_unlock(current);
1221 up_write(&mm->mmap_sem);
1224 NODEMASK_SCRATCH_FREE(scratch);
1229 err = queue_pages_range(mm, start, end, nmask,
1230 flags | MPOL_MF_INVERT, &pagelist);
1232 err = mbind_range(mm, start, end, new);
1237 if (!list_empty(&pagelist)) {
1238 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1239 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1240 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1242 putback_movable_pages(&pagelist);
1245 if (nr_failed && (flags & MPOL_MF_STRICT))
1248 putback_movable_pages(&pagelist);
1250 up_write(&mm->mmap_sem);
1257 * User space interface with variable sized bitmaps for nodelists.
1260 /* Copy a node mask from user space. */
1261 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1262 unsigned long maxnode)
1266 unsigned long nlongs;
1267 unsigned long endmask;
1270 nodes_clear(*nodes);
1271 if (maxnode == 0 || !nmask)
1273 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1276 nlongs = BITS_TO_LONGS(maxnode);
1277 if ((maxnode % BITS_PER_LONG) == 0)
1280 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1283 * When the user specified more nodes than supported just check
1284 * if the non supported part is all zero.
1286 * If maxnode have more longs than MAX_NUMNODES, check
1287 * the bits in that area first. And then go through to
1288 * check the rest bits which equal or bigger than MAX_NUMNODES.
1289 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1291 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1292 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1293 if (get_user(t, nmask + k))
1295 if (k == nlongs - 1) {
1301 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1305 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1306 unsigned long valid_mask = endmask;
1308 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1309 if (get_user(t, nmask + nlongs - 1))
1315 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1317 nodes_addr(*nodes)[nlongs-1] &= endmask;
1321 /* Copy a kernel node mask to user space */
1322 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1325 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1326 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1328 if (copy > nbytes) {
1329 if (copy > PAGE_SIZE)
1331 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1335 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1338 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1339 unsigned long, mode, const unsigned long __user *, nmask,
1340 unsigned long, maxnode, unsigned, flags)
1344 unsigned short mode_flags;
1346 mode_flags = mode & MPOL_MODE_FLAGS;
1347 mode &= ~MPOL_MODE_FLAGS;
1348 if (mode >= MPOL_MAX)
1350 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1351 (mode_flags & MPOL_F_RELATIVE_NODES))
1353 err = get_nodes(&nodes, nmask, maxnode);
1356 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1359 /* Set the process memory policy */
1360 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1361 unsigned long, maxnode)
1365 unsigned short flags;
1367 flags = mode & MPOL_MODE_FLAGS;
1368 mode &= ~MPOL_MODE_FLAGS;
1369 if ((unsigned int)mode >= MPOL_MAX)
1371 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1373 err = get_nodes(&nodes, nmask, maxnode);
1376 return do_set_mempolicy(mode, flags, &nodes);
1379 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1380 const unsigned long __user *, old_nodes,
1381 const unsigned long __user *, new_nodes)
1383 struct mm_struct *mm = NULL;
1384 struct task_struct *task;
1385 nodemask_t task_nodes;
1389 NODEMASK_SCRATCH(scratch);
1394 old = &scratch->mask1;
1395 new = &scratch->mask2;
1397 err = get_nodes(old, old_nodes, maxnode);
1401 err = get_nodes(new, new_nodes, maxnode);
1405 /* Find the mm_struct */
1407 task = pid ? find_task_by_vpid(pid) : current;
1413 get_task_struct(task);
1418 * Check if this process has the right to modify the specified process.
1419 * Use the regular "ptrace_may_access()" checks.
1421 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1428 task_nodes = cpuset_mems_allowed(task);
1429 /* Is the user allowed to access the target nodes? */
1430 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1435 task_nodes = cpuset_mems_allowed(current);
1436 nodes_and(*new, *new, task_nodes);
1437 if (nodes_empty(*new))
1440 nodes_and(*new, *new, node_states[N_MEMORY]);
1441 if (nodes_empty(*new))
1444 err = security_task_movememory(task);
1448 mm = get_task_mm(task);
1449 put_task_struct(task);
1456 err = do_migrate_pages(mm, old, new,
1457 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1461 NODEMASK_SCRATCH_FREE(scratch);
1466 put_task_struct(task);
1472 /* Retrieve NUMA policy */
1473 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1474 unsigned long __user *, nmask, unsigned long, maxnode,
1475 unsigned long, addr, unsigned long, flags)
1478 int uninitialized_var(pval);
1481 if (nmask != NULL && maxnode < MAX_NUMNODES)
1484 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1489 if (policy && put_user(pval, policy))
1493 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1498 #ifdef CONFIG_COMPAT
1500 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1501 compat_ulong_t __user *, nmask,
1502 compat_ulong_t, maxnode,
1503 compat_ulong_t, addr, compat_ulong_t, flags)
1506 unsigned long __user *nm = NULL;
1507 unsigned long nr_bits, alloc_size;
1508 DECLARE_BITMAP(bm, MAX_NUMNODES);
1510 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1511 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1514 nm = compat_alloc_user_space(alloc_size);
1516 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1518 if (!err && nmask) {
1519 unsigned long copy_size;
1520 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1521 err = copy_from_user(bm, nm, copy_size);
1522 /* ensure entire bitmap is zeroed */
1523 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1524 err |= compat_put_bitmap(nmask, bm, nr_bits);
1530 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1531 compat_ulong_t, maxnode)
1533 unsigned long __user *nm = NULL;
1534 unsigned long nr_bits, alloc_size;
1535 DECLARE_BITMAP(bm, MAX_NUMNODES);
1537 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1538 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1541 if (compat_get_bitmap(bm, nmask, nr_bits))
1543 nm = compat_alloc_user_space(alloc_size);
1544 if (copy_to_user(nm, bm, alloc_size))
1548 return sys_set_mempolicy(mode, nm, nr_bits+1);
1551 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1552 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1553 compat_ulong_t, maxnode, compat_ulong_t, flags)
1555 unsigned long __user *nm = NULL;
1556 unsigned long nr_bits, alloc_size;
1559 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1560 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1563 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1565 nm = compat_alloc_user_space(alloc_size);
1566 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1570 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1575 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1578 struct mempolicy *pol = NULL;
1581 if (vma->vm_ops && vma->vm_ops->get_policy) {
1582 pol = vma->vm_ops->get_policy(vma, addr);
1583 } else if (vma->vm_policy) {
1584 pol = vma->vm_policy;
1587 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1588 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1589 * count on these policies which will be dropped by
1590 * mpol_cond_put() later
1592 if (mpol_needs_cond_ref(pol))
1601 * get_vma_policy(@vma, @addr)
1602 * @vma: virtual memory area whose policy is sought
1603 * @addr: address in @vma for shared policy lookup
1605 * Returns effective policy for a VMA at specified address.
1606 * Falls back to current->mempolicy or system default policy, as necessary.
1607 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1608 * count--added by the get_policy() vm_op, as appropriate--to protect against
1609 * freeing by another task. It is the caller's responsibility to free the
1610 * extra reference for shared policies.
1612 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1615 struct mempolicy *pol = __get_vma_policy(vma, addr);
1618 pol = get_task_policy(current);
1623 bool vma_policy_mof(struct vm_area_struct *vma)
1625 struct mempolicy *pol;
1627 if (vma->vm_ops && vma->vm_ops->get_policy) {
1630 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1631 if (pol && (pol->flags & MPOL_F_MOF))
1638 pol = vma->vm_policy;
1640 pol = get_task_policy(current);
1642 return pol->flags & MPOL_F_MOF;
1645 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1647 enum zone_type dynamic_policy_zone = policy_zone;
1649 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1652 * if policy->v.nodes has movable memory only,
1653 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1655 * policy->v.nodes is intersect with node_states[N_MEMORY].
1656 * so if the following test faile, it implies
1657 * policy->v.nodes has movable memory only.
1659 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1660 dynamic_policy_zone = ZONE_MOVABLE;
1662 return zone >= dynamic_policy_zone;
1666 * Return a nodemask representing a mempolicy for filtering nodes for
1669 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1671 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1672 if (unlikely(policy->mode == MPOL_BIND) &&
1673 apply_policy_zone(policy, gfp_zone(gfp)) &&
1674 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1675 return &policy->v.nodes;
1680 /* Return the node id preferred by the given mempolicy, or the given id */
1681 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1684 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1685 nd = policy->v.preferred_node;
1688 * __GFP_THISNODE shouldn't even be used with the bind policy
1689 * because we might easily break the expectation to stay on the
1690 * requested node and not break the policy.
1692 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1698 /* Do dynamic interleaving for a process */
1699 static unsigned interleave_nodes(struct mempolicy *policy)
1702 struct task_struct *me = current;
1704 next = next_node_in(me->il_prev, policy->v.nodes);
1705 if (next < MAX_NUMNODES)
1711 * Depending on the memory policy provide a node from which to allocate the
1714 unsigned int mempolicy_slab_node(void)
1716 struct mempolicy *policy;
1717 int node = numa_mem_id();
1722 policy = current->mempolicy;
1723 if (!policy || policy->flags & MPOL_F_LOCAL)
1726 switch (policy->mode) {
1727 case MPOL_PREFERRED:
1729 * handled MPOL_F_LOCAL above
1731 return policy->v.preferred_node;
1733 case MPOL_INTERLEAVE:
1734 return interleave_nodes(policy);
1740 * Follow bind policy behavior and start allocation at the
1743 struct zonelist *zonelist;
1744 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1745 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1746 z = first_zones_zonelist(zonelist, highest_zoneidx,
1748 return z->zone ? z->zone->node : node;
1757 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1758 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1759 * number of present nodes.
1761 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1763 unsigned nnodes = nodes_weight(pol->v.nodes);
1769 return numa_node_id();
1770 target = (unsigned int)n % nnodes;
1771 nid = first_node(pol->v.nodes);
1772 for (i = 0; i < target; i++)
1773 nid = next_node(nid, pol->v.nodes);
1777 /* Determine a node number for interleave */
1778 static inline unsigned interleave_nid(struct mempolicy *pol,
1779 struct vm_area_struct *vma, unsigned long addr, int shift)
1785 * for small pages, there is no difference between
1786 * shift and PAGE_SHIFT, so the bit-shift is safe.
1787 * for huge pages, since vm_pgoff is in units of small
1788 * pages, we need to shift off the always 0 bits to get
1791 BUG_ON(shift < PAGE_SHIFT);
1792 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1793 off += (addr - vma->vm_start) >> shift;
1794 return offset_il_node(pol, off);
1796 return interleave_nodes(pol);
1799 #ifdef CONFIG_HUGETLBFS
1801 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1802 * @vma: virtual memory area whose policy is sought
1803 * @addr: address in @vma for shared policy lookup and interleave policy
1804 * @gfp_flags: for requested zone
1805 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1806 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1808 * Returns a nid suitable for a huge page allocation and a pointer
1809 * to the struct mempolicy for conditional unref after allocation.
1810 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1811 * @nodemask for filtering the zonelist.
1813 * Must be protected by read_mems_allowed_begin()
1815 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1816 struct mempolicy **mpol, nodemask_t **nodemask)
1820 *mpol = get_vma_policy(vma, addr);
1821 *nodemask = NULL; /* assume !MPOL_BIND */
1823 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1824 nid = interleave_nid(*mpol, vma, addr,
1825 huge_page_shift(hstate_vma(vma)));
1827 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1828 if ((*mpol)->mode == MPOL_BIND)
1829 *nodemask = &(*mpol)->v.nodes;
1835 * init_nodemask_of_mempolicy
1837 * If the current task's mempolicy is "default" [NULL], return 'false'
1838 * to indicate default policy. Otherwise, extract the policy nodemask
1839 * for 'bind' or 'interleave' policy into the argument nodemask, or
1840 * initialize the argument nodemask to contain the single node for
1841 * 'preferred' or 'local' policy and return 'true' to indicate presence
1842 * of non-default mempolicy.
1844 * We don't bother with reference counting the mempolicy [mpol_get/put]
1845 * because the current task is examining it's own mempolicy and a task's
1846 * mempolicy is only ever changed by the task itself.
1848 * N.B., it is the caller's responsibility to free a returned nodemask.
1850 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1852 struct mempolicy *mempolicy;
1855 if (!(mask && current->mempolicy))
1859 mempolicy = current->mempolicy;
1860 switch (mempolicy->mode) {
1861 case MPOL_PREFERRED:
1862 if (mempolicy->flags & MPOL_F_LOCAL)
1863 nid = numa_node_id();
1865 nid = mempolicy->v.preferred_node;
1866 init_nodemask_of_node(mask, nid);
1871 case MPOL_INTERLEAVE:
1872 *mask = mempolicy->v.nodes;
1878 task_unlock(current);
1885 * mempolicy_nodemask_intersects
1887 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1888 * policy. Otherwise, check for intersection between mask and the policy
1889 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1890 * policy, always return true since it may allocate elsewhere on fallback.
1892 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1894 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1895 const nodemask_t *mask)
1897 struct mempolicy *mempolicy;
1903 mempolicy = tsk->mempolicy;
1907 switch (mempolicy->mode) {
1908 case MPOL_PREFERRED:
1910 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1911 * allocate from, they may fallback to other nodes when oom.
1912 * Thus, it's possible for tsk to have allocated memory from
1917 case MPOL_INTERLEAVE:
1918 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1928 /* Allocate a page in interleaved policy.
1929 Own path because it needs to do special accounting. */
1930 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1935 page = __alloc_pages(gfp, order, nid);
1936 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1937 if (!static_branch_likely(&vm_numa_stat_key))
1939 if (page && page_to_nid(page) == nid) {
1941 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1948 * alloc_pages_vma - Allocate a page for a VMA.
1951 * %GFP_USER user allocation.
1952 * %GFP_KERNEL kernel allocations,
1953 * %GFP_HIGHMEM highmem/user allocations,
1954 * %GFP_FS allocation should not call back into a file system.
1955 * %GFP_ATOMIC don't sleep.
1957 * @order:Order of the GFP allocation.
1958 * @vma: Pointer to VMA or NULL if not available.
1959 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1960 * @node: Which node to prefer for allocation (modulo policy).
1961 * @hugepage: for hugepages try only the preferred node if possible
1963 * This function allocates a page from the kernel page pool and applies
1964 * a NUMA policy associated with the VMA or the current process.
1965 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1966 * mm_struct of the VMA to prevent it from going away. Should be used for
1967 * all allocations for pages that will be mapped into user space. Returns
1968 * NULL when no page can be allocated.
1971 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1972 unsigned long addr, int node, bool hugepage)
1974 struct mempolicy *pol;
1979 pol = get_vma_policy(vma, addr);
1981 if (pol->mode == MPOL_INTERLEAVE) {
1984 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1986 page = alloc_page_interleave(gfp, order, nid);
1990 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1991 int hpage_node = node;
1994 * For hugepage allocation and non-interleave policy which
1995 * allows the current node (or other explicitly preferred
1996 * node) we only try to allocate from the current/preferred
1997 * node and don't fall back to other nodes, as the cost of
1998 * remote accesses would likely offset THP benefits.
2000 * If the policy is interleave, or does not allow the current
2001 * node in its nodemask, we allocate the standard way.
2003 if (pol->mode == MPOL_PREFERRED &&
2004 !(pol->flags & MPOL_F_LOCAL))
2005 hpage_node = pol->v.preferred_node;
2007 nmask = policy_nodemask(gfp, pol);
2008 if (!nmask || node_isset(hpage_node, *nmask)) {
2010 page = __alloc_pages_node(hpage_node,
2011 gfp | __GFP_THISNODE, order);
2016 nmask = policy_nodemask(gfp, pol);
2017 preferred_nid = policy_node(gfp, pol, node);
2018 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2025 * alloc_pages_current - Allocate pages.
2028 * %GFP_USER user allocation,
2029 * %GFP_KERNEL kernel allocation,
2030 * %GFP_HIGHMEM highmem allocation,
2031 * %GFP_FS don't call back into a file system.
2032 * %GFP_ATOMIC don't sleep.
2033 * @order: Power of two of allocation size in pages. 0 is a single page.
2035 * Allocate a page from the kernel page pool. When not in
2036 * interrupt context and apply the current process NUMA policy.
2037 * Returns NULL when no page can be allocated.
2039 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2041 struct mempolicy *pol = &default_policy;
2044 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2045 pol = get_task_policy(current);
2048 * No reference counting needed for current->mempolicy
2049 * nor system default_policy
2051 if (pol->mode == MPOL_INTERLEAVE)
2052 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2054 page = __alloc_pages_nodemask(gfp, order,
2055 policy_node(gfp, pol, numa_node_id()),
2056 policy_nodemask(gfp, pol));
2060 EXPORT_SYMBOL(alloc_pages_current);
2062 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2064 struct mempolicy *pol = mpol_dup(vma_policy(src));
2067 return PTR_ERR(pol);
2068 dst->vm_policy = pol;
2073 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2074 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2075 * with the mems_allowed returned by cpuset_mems_allowed(). This
2076 * keeps mempolicies cpuset relative after its cpuset moves. See
2077 * further kernel/cpuset.c update_nodemask().
2079 * current's mempolicy may be rebinded by the other task(the task that changes
2080 * cpuset's mems), so we needn't do rebind work for current task.
2083 /* Slow path of a mempolicy duplicate */
2084 struct mempolicy *__mpol_dup(struct mempolicy *old)
2086 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2089 return ERR_PTR(-ENOMEM);
2091 /* task's mempolicy is protected by alloc_lock */
2092 if (old == current->mempolicy) {
2095 task_unlock(current);
2099 if (current_cpuset_is_being_rebound()) {
2100 nodemask_t mems = cpuset_mems_allowed(current);
2101 mpol_rebind_policy(new, &mems);
2103 atomic_set(&new->refcnt, 1);
2107 /* Slow path of a mempolicy comparison */
2108 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2112 if (a->mode != b->mode)
2114 if (a->flags != b->flags)
2116 if (mpol_store_user_nodemask(a))
2117 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2123 case MPOL_INTERLEAVE:
2124 return !!nodes_equal(a->v.nodes, b->v.nodes);
2125 case MPOL_PREFERRED:
2126 return a->v.preferred_node == b->v.preferred_node;
2134 * Shared memory backing store policy support.
2136 * Remember policies even when nobody has shared memory mapped.
2137 * The policies are kept in Red-Black tree linked from the inode.
2138 * They are protected by the sp->lock rwlock, which should be held
2139 * for any accesses to the tree.
2143 * lookup first element intersecting start-end. Caller holds sp->lock for
2144 * reading or for writing
2146 static struct sp_node *
2147 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2149 struct rb_node *n = sp->root.rb_node;
2152 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2154 if (start >= p->end)
2156 else if (end <= p->start)
2164 struct sp_node *w = NULL;
2165 struct rb_node *prev = rb_prev(n);
2168 w = rb_entry(prev, struct sp_node, nd);
2169 if (w->end <= start)
2173 return rb_entry(n, struct sp_node, nd);
2177 * Insert a new shared policy into the list. Caller holds sp->lock for
2180 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2182 struct rb_node **p = &sp->root.rb_node;
2183 struct rb_node *parent = NULL;
2188 nd = rb_entry(parent, struct sp_node, nd);
2189 if (new->start < nd->start)
2191 else if (new->end > nd->end)
2192 p = &(*p)->rb_right;
2196 rb_link_node(&new->nd, parent, p);
2197 rb_insert_color(&new->nd, &sp->root);
2198 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2199 new->policy ? new->policy->mode : 0);
2202 /* Find shared policy intersecting idx */
2204 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2206 struct mempolicy *pol = NULL;
2209 if (!sp->root.rb_node)
2211 read_lock(&sp->lock);
2212 sn = sp_lookup(sp, idx, idx+1);
2214 mpol_get(sn->policy);
2217 read_unlock(&sp->lock);
2221 static void sp_free(struct sp_node *n)
2223 mpol_put(n->policy);
2224 kmem_cache_free(sn_cache, n);
2228 * mpol_misplaced - check whether current page node is valid in policy
2230 * @page: page to be checked
2231 * @vma: vm area where page mapped
2232 * @addr: virtual address where page mapped
2234 * Lookup current policy node id for vma,addr and "compare to" page's
2238 * -1 - not misplaced, page is in the right node
2239 * node - node id where the page should be
2241 * Policy determination "mimics" alloc_page_vma().
2242 * Called from fault path where we know the vma and faulting address.
2244 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2246 struct mempolicy *pol;
2248 int curnid = page_to_nid(page);
2249 unsigned long pgoff;
2250 int thiscpu = raw_smp_processor_id();
2251 int thisnid = cpu_to_node(thiscpu);
2255 pol = get_vma_policy(vma, addr);
2256 if (!(pol->flags & MPOL_F_MOF))
2259 switch (pol->mode) {
2260 case MPOL_INTERLEAVE:
2261 pgoff = vma->vm_pgoff;
2262 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2263 polnid = offset_il_node(pol, pgoff);
2266 case MPOL_PREFERRED:
2267 if (pol->flags & MPOL_F_LOCAL)
2268 polnid = numa_node_id();
2270 polnid = pol->v.preferred_node;
2276 * allows binding to multiple nodes.
2277 * use current page if in policy nodemask,
2278 * else select nearest allowed node, if any.
2279 * If no allowed nodes, use current [!misplaced].
2281 if (node_isset(curnid, pol->v.nodes))
2283 z = first_zones_zonelist(
2284 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2285 gfp_zone(GFP_HIGHUSER),
2287 polnid = z->zone->node;
2294 /* Migrate the page towards the node whose CPU is referencing it */
2295 if (pol->flags & MPOL_F_MORON) {
2298 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2302 if (curnid != polnid)
2311 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2312 * dropped after task->mempolicy is set to NULL so that any allocation done as
2313 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2316 void mpol_put_task_policy(struct task_struct *task)
2318 struct mempolicy *pol;
2321 pol = task->mempolicy;
2322 task->mempolicy = NULL;
2327 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2329 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2330 rb_erase(&n->nd, &sp->root);
2334 static void sp_node_init(struct sp_node *node, unsigned long start,
2335 unsigned long end, struct mempolicy *pol)
2337 node->start = start;
2342 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2343 struct mempolicy *pol)
2346 struct mempolicy *newpol;
2348 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2352 newpol = mpol_dup(pol);
2353 if (IS_ERR(newpol)) {
2354 kmem_cache_free(sn_cache, n);
2357 newpol->flags |= MPOL_F_SHARED;
2358 sp_node_init(n, start, end, newpol);
2363 /* Replace a policy range. */
2364 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2365 unsigned long end, struct sp_node *new)
2368 struct sp_node *n_new = NULL;
2369 struct mempolicy *mpol_new = NULL;
2373 write_lock(&sp->lock);
2374 n = sp_lookup(sp, start, end);
2375 /* Take care of old policies in the same range. */
2376 while (n && n->start < end) {
2377 struct rb_node *next = rb_next(&n->nd);
2378 if (n->start >= start) {
2384 /* Old policy spanning whole new range. */
2389 *mpol_new = *n->policy;
2390 atomic_set(&mpol_new->refcnt, 1);
2391 sp_node_init(n_new, end, n->end, mpol_new);
2393 sp_insert(sp, n_new);
2402 n = rb_entry(next, struct sp_node, nd);
2406 write_unlock(&sp->lock);
2413 kmem_cache_free(sn_cache, n_new);
2418 write_unlock(&sp->lock);
2420 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2423 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2430 * mpol_shared_policy_init - initialize shared policy for inode
2431 * @sp: pointer to inode shared policy
2432 * @mpol: struct mempolicy to install
2434 * Install non-NULL @mpol in inode's shared policy rb-tree.
2435 * On entry, the current task has a reference on a non-NULL @mpol.
2436 * This must be released on exit.
2437 * This is called at get_inode() calls and we can use GFP_KERNEL.
2439 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2443 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2444 rwlock_init(&sp->lock);
2447 struct vm_area_struct pvma;
2448 struct mempolicy *new;
2449 NODEMASK_SCRATCH(scratch);
2453 /* contextualize the tmpfs mount point mempolicy */
2454 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2456 goto free_scratch; /* no valid nodemask intersection */
2459 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2460 task_unlock(current);
2464 /* Create pseudo-vma that contains just the policy */
2465 memset(&pvma, 0, sizeof(struct vm_area_struct));
2466 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2467 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2470 mpol_put(new); /* drop initial ref */
2472 NODEMASK_SCRATCH_FREE(scratch);
2474 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2478 int mpol_set_shared_policy(struct shared_policy *info,
2479 struct vm_area_struct *vma, struct mempolicy *npol)
2482 struct sp_node *new = NULL;
2483 unsigned long sz = vma_pages(vma);
2485 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2487 sz, npol ? npol->mode : -1,
2488 npol ? npol->flags : -1,
2489 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2492 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2496 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2502 /* Free a backing policy store on inode delete. */
2503 void mpol_free_shared_policy(struct shared_policy *p)
2506 struct rb_node *next;
2508 if (!p->root.rb_node)
2510 write_lock(&p->lock);
2511 next = rb_first(&p->root);
2513 n = rb_entry(next, struct sp_node, nd);
2514 next = rb_next(&n->nd);
2517 write_unlock(&p->lock);
2520 #ifdef CONFIG_NUMA_BALANCING
2521 static int __initdata numabalancing_override;
2523 static void __init check_numabalancing_enable(void)
2525 bool numabalancing_default = false;
2527 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2528 numabalancing_default = true;
2530 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2531 if (numabalancing_override)
2532 set_numabalancing_state(numabalancing_override == 1);
2534 if (num_online_nodes() > 1 && !numabalancing_override) {
2535 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2536 numabalancing_default ? "Enabling" : "Disabling");
2537 set_numabalancing_state(numabalancing_default);
2541 static int __init setup_numabalancing(char *str)
2547 if (!strcmp(str, "enable")) {
2548 numabalancing_override = 1;
2550 } else if (!strcmp(str, "disable")) {
2551 numabalancing_override = -1;
2556 pr_warn("Unable to parse numa_balancing=\n");
2560 __setup("numa_balancing=", setup_numabalancing);
2562 static inline void __init check_numabalancing_enable(void)
2565 #endif /* CONFIG_NUMA_BALANCING */
2567 /* assumes fs == KERNEL_DS */
2568 void __init numa_policy_init(void)
2570 nodemask_t interleave_nodes;
2571 unsigned long largest = 0;
2572 int nid, prefer = 0;
2574 policy_cache = kmem_cache_create("numa_policy",
2575 sizeof(struct mempolicy),
2576 0, SLAB_PANIC, NULL);
2578 sn_cache = kmem_cache_create("shared_policy_node",
2579 sizeof(struct sp_node),
2580 0, SLAB_PANIC, NULL);
2582 for_each_node(nid) {
2583 preferred_node_policy[nid] = (struct mempolicy) {
2584 .refcnt = ATOMIC_INIT(1),
2585 .mode = MPOL_PREFERRED,
2586 .flags = MPOL_F_MOF | MPOL_F_MORON,
2587 .v = { .preferred_node = nid, },
2592 * Set interleaving policy for system init. Interleaving is only
2593 * enabled across suitably sized nodes (default is >= 16MB), or
2594 * fall back to the largest node if they're all smaller.
2596 nodes_clear(interleave_nodes);
2597 for_each_node_state(nid, N_MEMORY) {
2598 unsigned long total_pages = node_present_pages(nid);
2600 /* Preserve the largest node */
2601 if (largest < total_pages) {
2602 largest = total_pages;
2606 /* Interleave this node? */
2607 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2608 node_set(nid, interleave_nodes);
2611 /* All too small, use the largest */
2612 if (unlikely(nodes_empty(interleave_nodes)))
2613 node_set(prefer, interleave_nodes);
2615 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2616 pr_err("%s: interleaving failed\n", __func__);
2618 check_numabalancing_enable();
2621 /* Reset policy of current process to default */
2622 void numa_default_policy(void)
2624 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2628 * Parse and format mempolicy from/to strings
2632 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2634 static const char * const policy_modes[] =
2636 [MPOL_DEFAULT] = "default",
2637 [MPOL_PREFERRED] = "prefer",
2638 [MPOL_BIND] = "bind",
2639 [MPOL_INTERLEAVE] = "interleave",
2640 [MPOL_LOCAL] = "local",
2646 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2647 * @str: string containing mempolicy to parse
2648 * @mpol: pointer to struct mempolicy pointer, returned on success.
2651 * <mode>[=<flags>][:<nodelist>]
2653 * On success, returns 0, else 1
2655 int mpol_parse_str(char *str, struct mempolicy **mpol)
2657 struct mempolicy *new = NULL;
2658 unsigned short mode;
2659 unsigned short mode_flags;
2661 char *nodelist = strchr(str, ':');
2662 char *flags = strchr(str, '=');
2666 /* NUL-terminate mode or flags string */
2668 if (nodelist_parse(nodelist, nodes))
2670 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2676 *flags++ = '\0'; /* terminate mode string */
2678 for (mode = 0; mode < MPOL_MAX; mode++) {
2679 if (!strcmp(str, policy_modes[mode])) {
2683 if (mode >= MPOL_MAX)
2687 case MPOL_PREFERRED:
2689 * Insist on a nodelist of one node only
2692 char *rest = nodelist;
2693 while (isdigit(*rest))
2699 case MPOL_INTERLEAVE:
2701 * Default to online nodes with memory if no nodelist
2704 nodes = node_states[N_MEMORY];
2708 * Don't allow a nodelist; mpol_new() checks flags
2712 mode = MPOL_PREFERRED;
2716 * Insist on a empty nodelist
2723 * Insist on a nodelist
2732 * Currently, we only support two mutually exclusive
2735 if (!strcmp(flags, "static"))
2736 mode_flags |= MPOL_F_STATIC_NODES;
2737 else if (!strcmp(flags, "relative"))
2738 mode_flags |= MPOL_F_RELATIVE_NODES;
2743 new = mpol_new(mode, mode_flags, &nodes);
2748 * Save nodes for mpol_to_str() to show the tmpfs mount options
2749 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2751 if (mode != MPOL_PREFERRED)
2752 new->v.nodes = nodes;
2754 new->v.preferred_node = first_node(nodes);
2756 new->flags |= MPOL_F_LOCAL;
2759 * Save nodes for contextualization: this will be used to "clone"
2760 * the mempolicy in a specific context [cpuset] at a later time.
2762 new->w.user_nodemask = nodes;
2767 /* Restore string for error message */
2776 #endif /* CONFIG_TMPFS */
2779 * mpol_to_str - format a mempolicy structure for printing
2780 * @buffer: to contain formatted mempolicy string
2781 * @maxlen: length of @buffer
2782 * @pol: pointer to mempolicy to be formatted
2784 * Convert @pol into a string. If @buffer is too short, truncate the string.
2785 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2786 * longest flag, "relative", and to display at least a few node ids.
2788 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2791 nodemask_t nodes = NODE_MASK_NONE;
2792 unsigned short mode = MPOL_DEFAULT;
2793 unsigned short flags = 0;
2795 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2803 case MPOL_PREFERRED:
2804 if (flags & MPOL_F_LOCAL)
2807 node_set(pol->v.preferred_node, nodes);
2810 case MPOL_INTERLEAVE:
2811 nodes = pol->v.nodes;
2815 snprintf(p, maxlen, "unknown");
2819 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2821 if (flags & MPOL_MODE_FLAGS) {
2822 p += snprintf(p, buffer + maxlen - p, "=");
2825 * Currently, the only defined flags are mutually exclusive
2827 if (flags & MPOL_F_STATIC_NODES)
2828 p += snprintf(p, buffer + maxlen - p, "static");
2829 else if (flags & MPOL_F_RELATIVE_NODES)
2830 p += snprintf(p, buffer + maxlen - p, "relative");
2833 if (!nodes_empty(nodes))
2834 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2835 nodemask_pr_args(&nodes));