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/swap.h>
89 #include <linux/seq_file.h>
90 #include <linux/proc_fs.h>
91 #include <linux/migrate.h>
92 #include <linux/ksm.h>
93 #include <linux/rmap.h>
94 #include <linux/security.h>
95 #include <linux/syscalls.h>
96 #include <linux/ctype.h>
97 #include <linux/mm_inline.h>
98 #include <linux/mmu_notifier.h>
99 #include <linux/printk.h>
100 #include <linux/swapops.h>
102 #include <asm/tlbflush.h>
103 #include <linux/uaccess.h>
105 #include "internal.h"
108 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
109 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
111 static struct kmem_cache *policy_cache;
112 static struct kmem_cache *sn_cache;
114 /* Highest zone. An specific allocation for a zone below that is not
116 enum zone_type policy_zone = 0;
119 * run-time system-wide default policy => local allocation
121 static struct mempolicy default_policy = {
122 .refcnt = ATOMIC_INIT(1), /* never free it */
123 .mode = MPOL_PREFERRED,
124 .flags = MPOL_F_LOCAL,
127 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
129 struct mempolicy *get_task_policy(struct task_struct *p)
131 struct mempolicy *pol = p->mempolicy;
137 node = numa_node_id();
138 if (node != NUMA_NO_NODE) {
139 pol = &preferred_node_policy[node];
140 /* preferred_node_policy is not initialised early in boot */
145 return &default_policy;
148 static const struct mempolicy_operations {
149 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
150 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
151 } mpol_ops[MPOL_MAX];
153 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
155 return pol->flags & MPOL_MODE_FLAGS;
158 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
159 const nodemask_t *rel)
162 nodes_fold(tmp, *orig, nodes_weight(*rel));
163 nodes_onto(*ret, tmp, *rel);
166 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
168 if (nodes_empty(*nodes))
170 pol->v.nodes = *nodes;
174 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
177 pol->flags |= MPOL_F_LOCAL; /* local allocation */
178 else if (nodes_empty(*nodes))
179 return -EINVAL; /* no allowed nodes */
181 pol->v.preferred_node = first_node(*nodes);
185 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
187 if (nodes_empty(*nodes))
189 pol->v.nodes = *nodes;
194 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
195 * any, for the new policy. mpol_new() has already validated the nodes
196 * parameter with respect to the policy mode and flags. But, we need to
197 * handle an empty nodemask with MPOL_PREFERRED here.
199 * Must be called holding task's alloc_lock to protect task's mems_allowed
200 * and mempolicy. May also be called holding the mmap_semaphore for write.
202 static int mpol_set_nodemask(struct mempolicy *pol,
203 const nodemask_t *nodes, struct nodemask_scratch *nsc)
207 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
211 nodes_and(nsc->mask1,
212 cpuset_current_mems_allowed, node_states[N_MEMORY]);
215 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
216 nodes = NULL; /* explicit local allocation */
218 if (pol->flags & MPOL_F_RELATIVE_NODES)
219 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
221 nodes_and(nsc->mask2, *nodes, nsc->mask1);
223 if (mpol_store_user_nodemask(pol))
224 pol->w.user_nodemask = *nodes;
226 pol->w.cpuset_mems_allowed =
227 cpuset_current_mems_allowed;
231 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
233 ret = mpol_ops[pol->mode].create(pol, NULL);
238 * This function just creates a new policy, does some check and simple
239 * initialization. You must invoke mpol_set_nodemask() to set nodes.
241 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
244 struct mempolicy *policy;
246 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
247 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
249 if (mode == MPOL_DEFAULT) {
250 if (nodes && !nodes_empty(*nodes))
251 return ERR_PTR(-EINVAL);
257 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
258 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
259 * All other modes require a valid pointer to a non-empty nodemask.
261 if (mode == MPOL_PREFERRED) {
262 if (nodes_empty(*nodes)) {
263 if (((flags & MPOL_F_STATIC_NODES) ||
264 (flags & MPOL_F_RELATIVE_NODES)))
265 return ERR_PTR(-EINVAL);
267 } else if (mode == MPOL_LOCAL) {
268 if (!nodes_empty(*nodes) ||
269 (flags & MPOL_F_STATIC_NODES) ||
270 (flags & MPOL_F_RELATIVE_NODES))
271 return ERR_PTR(-EINVAL);
272 mode = MPOL_PREFERRED;
273 } else if (nodes_empty(*nodes))
274 return ERR_PTR(-EINVAL);
275 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
277 return ERR_PTR(-ENOMEM);
278 atomic_set(&policy->refcnt, 1);
280 policy->flags = flags;
285 /* Slow path of a mpol destructor. */
286 void __mpol_put(struct mempolicy *p)
288 if (!atomic_dec_and_test(&p->refcnt))
290 kmem_cache_free(policy_cache, p);
293 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
297 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
301 if (pol->flags & MPOL_F_STATIC_NODES)
302 nodes_and(tmp, pol->w.user_nodemask, *nodes);
303 else if (pol->flags & MPOL_F_RELATIVE_NODES)
304 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
306 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
308 pol->w.cpuset_mems_allowed = tmp;
311 if (nodes_empty(tmp))
317 static void mpol_rebind_preferred(struct mempolicy *pol,
318 const nodemask_t *nodes)
322 if (pol->flags & MPOL_F_STATIC_NODES) {
323 int node = first_node(pol->w.user_nodemask);
325 if (node_isset(node, *nodes)) {
326 pol->v.preferred_node = node;
327 pol->flags &= ~MPOL_F_LOCAL;
329 pol->flags |= MPOL_F_LOCAL;
330 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
331 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
332 pol->v.preferred_node = first_node(tmp);
333 } else if (!(pol->flags & MPOL_F_LOCAL)) {
334 pol->v.preferred_node = node_remap(pol->v.preferred_node,
335 pol->w.cpuset_mems_allowed,
337 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_sem. 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_sem during call.
377 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
379 struct vm_area_struct *vma;
381 down_write(&mm->mmap_sem);
382 for (vma = mm->mmap; vma; vma = vma->vm_next)
383 mpol_rebind_policy(vma->vm_policy, new);
384 up_write(&mm->mmap_sem);
387 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
389 .rebind = mpol_rebind_default,
391 [MPOL_INTERLEAVE] = {
392 .create = mpol_new_interleave,
393 .rebind = mpol_rebind_nodemask,
396 .create = mpol_new_preferred,
397 .rebind = mpol_rebind_preferred,
400 .create = mpol_new_bind,
401 .rebind = mpol_rebind_nodemask,
405 static void migrate_page_add(struct page *page, struct list_head *pagelist,
406 unsigned long flags);
409 struct list_head *pagelist;
412 struct vm_area_struct *prev;
416 * Check if the page's nid is in qp->nmask.
418 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
419 * in the invert of qp->nmask.
421 static inline bool queue_pages_required(struct page *page,
422 struct queue_pages *qp)
424 int nid = page_to_nid(page);
425 unsigned long flags = qp->flags;
427 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
430 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
431 unsigned long end, struct mm_walk *walk)
435 struct queue_pages *qp = walk->private;
438 if (unlikely(is_pmd_migration_entry(*pmd))) {
442 page = pmd_page(*pmd);
443 if (is_huge_zero_page(page)) {
445 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
448 if (!thp_migration_supported()) {
452 ret = split_huge_page(page);
457 if (!queue_pages_required(page, qp)) {
464 /* go to thp migration */
465 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
466 migrate_page_add(page, qp->pagelist, flags);
474 * Scan through pages checking if pages follow certain conditions,
475 * and move them to the pagelist if they do.
477 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
478 unsigned long end, struct mm_walk *walk)
480 struct vm_area_struct *vma = walk->vma;
482 struct queue_pages *qp = walk->private;
483 unsigned long flags = qp->flags;
488 ptl = pmd_trans_huge_lock(pmd, vma);
490 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
495 if (pmd_trans_unstable(pmd))
498 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
499 for (; addr != end; pte++, addr += PAGE_SIZE) {
500 if (!pte_present(*pte))
502 page = vm_normal_page(vma, addr, *pte);
506 * vm_normal_page() filters out zero pages, but there might
507 * still be PageReserved pages to skip, perhaps in a VDSO.
509 if (PageReserved(page))
511 if (!queue_pages_required(page, qp))
513 if (PageTransCompound(page) && !thp_migration_supported()) {
515 pte_unmap_unlock(pte, ptl);
517 ret = split_huge_page(page);
520 /* Failed to split -- skip. */
522 pte = pte_offset_map_lock(walk->mm, pmd,
529 migrate_page_add(page, qp->pagelist, flags);
531 pte_unmap_unlock(pte - 1, ptl);
536 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
537 unsigned long addr, unsigned long end,
538 struct mm_walk *walk)
540 #ifdef CONFIG_HUGETLB_PAGE
541 struct queue_pages *qp = walk->private;
542 unsigned long flags = qp->flags;
547 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
548 entry = huge_ptep_get(pte);
549 if (!pte_present(entry))
551 page = pte_page(entry);
552 if (!queue_pages_required(page, qp))
554 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
555 if (flags & (MPOL_MF_MOVE_ALL) ||
556 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
557 isolate_huge_page(page, qp->pagelist);
566 #ifdef CONFIG_NUMA_BALANCING
568 * This is used to mark a range of virtual addresses to be inaccessible.
569 * These are later cleared by a NUMA hinting fault. Depending on these
570 * faults, pages may be migrated for better NUMA placement.
572 * This is assuming that NUMA faults are handled using PROT_NONE. If
573 * an architecture makes a different choice, it will need further
574 * changes to the core.
576 unsigned long change_prot_numa(struct vm_area_struct *vma,
577 unsigned long addr, unsigned long end)
581 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
583 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
588 static unsigned long change_prot_numa(struct vm_area_struct *vma,
589 unsigned long addr, unsigned long end)
593 #endif /* CONFIG_NUMA_BALANCING */
595 static int queue_pages_test_walk(unsigned long start, unsigned long end,
596 struct mm_walk *walk)
598 struct vm_area_struct *vma = walk->vma;
599 struct queue_pages *qp = walk->private;
600 unsigned long endvma = vma->vm_end;
601 unsigned long flags = qp->flags;
603 if (!vma_migratable(vma))
608 if (vma->vm_start > start)
609 start = vma->vm_start;
611 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
612 if (!vma->vm_next && vma->vm_end < end)
614 if (qp->prev && qp->prev->vm_end < vma->vm_start)
620 if (flags & MPOL_MF_LAZY) {
621 /* Similar to task_numa_work, skip inaccessible VMAs */
622 if (!is_vm_hugetlb_page(vma) &&
623 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
624 !(vma->vm_flags & VM_MIXEDMAP))
625 change_prot_numa(vma, start, endvma);
629 /* queue pages from current vma */
630 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
636 * Walk through page tables and collect pages to be migrated.
638 * If pages found in a given range are on a set of nodes (determined by
639 * @nodes and @flags,) it's isolated and queued to the pagelist which is
640 * passed via @private.)
643 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
644 nodemask_t *nodes, unsigned long flags,
645 struct list_head *pagelist)
647 struct queue_pages qp = {
648 .pagelist = pagelist,
653 struct mm_walk queue_pages_walk = {
654 .hugetlb_entry = queue_pages_hugetlb,
655 .pmd_entry = queue_pages_pte_range,
656 .test_walk = queue_pages_test_walk,
661 return walk_page_range(start, end, &queue_pages_walk);
665 * Apply policy to a single VMA
666 * This must be called with the mmap_sem held for writing.
668 static int vma_replace_policy(struct vm_area_struct *vma,
669 struct mempolicy *pol)
672 struct mempolicy *old;
673 struct mempolicy *new;
675 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
676 vma->vm_start, vma->vm_end, vma->vm_pgoff,
677 vma->vm_ops, vma->vm_file,
678 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
684 if (vma->vm_ops && vma->vm_ops->set_policy) {
685 err = vma->vm_ops->set_policy(vma, new);
690 old = vma->vm_policy;
691 vma->vm_policy = new; /* protected by mmap_sem */
700 /* Step 2: apply policy to a range and do splits. */
701 static int mbind_range(struct mm_struct *mm, unsigned long start,
702 unsigned long end, struct mempolicy *new_pol)
704 struct vm_area_struct *next;
705 struct vm_area_struct *prev;
706 struct vm_area_struct *vma;
709 unsigned long vmstart;
712 vma = find_vma(mm, start);
713 if (!vma || vma->vm_start > start)
717 if (start > vma->vm_start)
720 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
722 vmstart = max(start, vma->vm_start);
723 vmend = min(end, vma->vm_end);
725 if (mpol_equal(vma_policy(vma), new_pol))
728 pgoff = vma->vm_pgoff +
729 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
730 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
731 vma->anon_vma, vma->vm_file, pgoff,
732 new_pol, vma->vm_userfaultfd_ctx);
736 if (mpol_equal(vma_policy(vma), new_pol))
738 /* vma_merge() joined vma && vma->next, case 8 */
741 if (vma->vm_start != vmstart) {
742 err = split_vma(vma->vm_mm, vma, vmstart, 1);
746 if (vma->vm_end != vmend) {
747 err = split_vma(vma->vm_mm, vma, vmend, 0);
752 err = vma_replace_policy(vma, new_pol);
761 /* Set the process memory policy */
762 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
765 struct mempolicy *new, *old;
766 NODEMASK_SCRATCH(scratch);
772 new = mpol_new(mode, flags, nodes);
779 ret = mpol_set_nodemask(new, nodes, scratch);
781 task_unlock(current);
785 old = current->mempolicy;
786 current->mempolicy = new;
787 if (new && new->mode == MPOL_INTERLEAVE)
788 current->il_prev = MAX_NUMNODES-1;
789 task_unlock(current);
793 NODEMASK_SCRATCH_FREE(scratch);
798 * Return nodemask for policy for get_mempolicy() query
800 * Called with task's alloc_lock held
802 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
805 if (p == &default_policy)
811 case MPOL_INTERLEAVE:
815 if (!(p->flags & MPOL_F_LOCAL))
816 node_set(p->v.preferred_node, *nodes);
817 /* else return empty node mask for local allocation */
824 static int lookup_node(unsigned long addr)
829 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
831 err = page_to_nid(p);
837 /* Retrieve NUMA policy */
838 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
839 unsigned long addr, unsigned long flags)
842 struct mm_struct *mm = current->mm;
843 struct vm_area_struct *vma = NULL;
844 struct mempolicy *pol = current->mempolicy;
847 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
850 if (flags & MPOL_F_MEMS_ALLOWED) {
851 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
853 *policy = 0; /* just so it's initialized */
855 *nmask = cpuset_current_mems_allowed;
856 task_unlock(current);
860 if (flags & MPOL_F_ADDR) {
862 * Do NOT fall back to task policy if the
863 * vma/shared policy at addr is NULL. We
864 * want to return MPOL_DEFAULT in this case.
866 down_read(&mm->mmap_sem);
867 vma = find_vma_intersection(mm, addr, addr+1);
869 up_read(&mm->mmap_sem);
872 if (vma->vm_ops && vma->vm_ops->get_policy)
873 pol = vma->vm_ops->get_policy(vma, addr);
875 pol = vma->vm_policy;
880 pol = &default_policy; /* indicates default behavior */
882 if (flags & MPOL_F_NODE) {
883 if (flags & MPOL_F_ADDR) {
884 err = lookup_node(addr);
888 } else if (pol == current->mempolicy &&
889 pol->mode == MPOL_INTERLEAVE) {
890 *policy = next_node_in(current->il_prev, pol->v.nodes);
896 *policy = pol == &default_policy ? MPOL_DEFAULT :
899 * Internal mempolicy flags must be masked off before exposing
900 * the policy to userspace.
902 *policy |= (pol->flags & MPOL_MODE_FLAGS);
907 if (mpol_store_user_nodemask(pol)) {
908 *nmask = pol->w.user_nodemask;
911 get_policy_nodemask(pol, nmask);
912 task_unlock(current);
919 up_read(¤t->mm->mmap_sem);
923 #ifdef CONFIG_MIGRATION
925 * page migration, thp tail pages can be passed.
927 static void migrate_page_add(struct page *page, struct list_head *pagelist,
930 struct page *head = compound_head(page);
932 * Avoid migrating a page that is shared with others.
934 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
935 if (!isolate_lru_page(head)) {
936 list_add_tail(&head->lru, pagelist);
937 mod_node_page_state(page_pgdat(head),
938 NR_ISOLATED_ANON + page_is_file_cache(head),
939 hpage_nr_pages(head));
944 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
947 return alloc_huge_page_node(page_hstate(compound_head(page)),
949 else if (thp_migration_supported() && PageTransHuge(page)) {
952 thp = alloc_pages_node(node,
953 (GFP_TRANSHUGE | __GFP_THISNODE),
957 prep_transhuge_page(thp);
960 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
965 * Migrate pages from one node to a target node.
966 * Returns error or the number of pages not migrated.
968 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
976 node_set(source, nmask);
979 * This does not "check" the range but isolates all pages that
980 * need migration. Between passing in the full user address
981 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
983 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
984 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
985 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
987 if (!list_empty(&pagelist)) {
988 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
989 MIGRATE_SYNC, MR_SYSCALL);
991 putback_movable_pages(&pagelist);
998 * Move pages between the two nodesets so as to preserve the physical
999 * layout as much as possible.
1001 * Returns the number of page that could not be moved.
1003 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1004 const nodemask_t *to, int flags)
1010 err = migrate_prep();
1014 down_read(&mm->mmap_sem);
1017 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1018 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1019 * bit in 'tmp', and return that <source, dest> pair for migration.
1020 * The pair of nodemasks 'to' and 'from' define the map.
1022 * If no pair of bits is found that way, fallback to picking some
1023 * pair of 'source' and 'dest' bits that are not the same. If the
1024 * 'source' and 'dest' bits are the same, this represents a node
1025 * that will be migrating to itself, so no pages need move.
1027 * If no bits are left in 'tmp', or if all remaining bits left
1028 * in 'tmp' correspond to the same bit in 'to', return false
1029 * (nothing left to migrate).
1031 * This lets us pick a pair of nodes to migrate between, such that
1032 * if possible the dest node is not already occupied by some other
1033 * source node, minimizing the risk of overloading the memory on a
1034 * node that would happen if we migrated incoming memory to a node
1035 * before migrating outgoing memory source that same node.
1037 * A single scan of tmp is sufficient. As we go, we remember the
1038 * most recent <s, d> pair that moved (s != d). If we find a pair
1039 * that not only moved, but what's better, moved to an empty slot
1040 * (d is not set in tmp), then we break out then, with that pair.
1041 * Otherwise when we finish scanning from_tmp, we at least have the
1042 * most recent <s, d> pair that moved. If we get all the way through
1043 * the scan of tmp without finding any node that moved, much less
1044 * moved to an empty node, then there is nothing left worth migrating.
1048 while (!nodes_empty(tmp)) {
1050 int source = NUMA_NO_NODE;
1053 for_each_node_mask(s, tmp) {
1056 * do_migrate_pages() tries to maintain the relative
1057 * node relationship of the pages established between
1058 * threads and memory areas.
1060 * However if the number of source nodes is not equal to
1061 * the number of destination nodes we can not preserve
1062 * this node relative relationship. In that case, skip
1063 * copying memory from a node that is in the destination
1066 * Example: [2,3,4] -> [3,4,5] moves everything.
1067 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1070 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1071 (node_isset(s, *to)))
1074 d = node_remap(s, *from, *to);
1078 source = s; /* Node moved. Memorize */
1081 /* dest not in remaining from nodes? */
1082 if (!node_isset(dest, tmp))
1085 if (source == NUMA_NO_NODE)
1088 node_clear(source, tmp);
1089 err = migrate_to_node(mm, source, dest, flags);
1095 up_read(&mm->mmap_sem);
1103 * Allocate a new page for page migration based on vma policy.
1104 * Start by assuming the page is mapped by the same vma as contains @start.
1105 * Search forward from there, if not. N.B., this assumes that the
1106 * list of pages handed to migrate_pages()--which is how we get here--
1107 * is in virtual address order.
1109 static struct page *new_page(struct page *page, unsigned long start, int **x)
1111 struct vm_area_struct *vma;
1112 unsigned long uninitialized_var(address);
1114 vma = find_vma(current->mm, start);
1116 address = page_address_in_vma(page, vma);
1117 if (address != -EFAULT)
1122 if (PageHuge(page)) {
1124 return alloc_huge_page_noerr(vma, address, 1);
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 if (nlongs > PAGE_SIZE/sizeof(long))
1294 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1295 if (get_user(t, nmask + k))
1297 if (k == nlongs - 1) {
1303 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1307 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1308 unsigned long valid_mask = endmask;
1310 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1311 if (get_user(t, nmask + nlongs - 1))
1317 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1319 nodes_addr(*nodes)[nlongs-1] &= endmask;
1323 /* Copy a kernel node mask to user space */
1324 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1327 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1328 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1330 if (copy > nbytes) {
1331 if (copy > PAGE_SIZE)
1333 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1337 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1340 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1341 unsigned long, mode, const unsigned long __user *, nmask,
1342 unsigned long, maxnode, unsigned, flags)
1346 unsigned short mode_flags;
1348 mode_flags = mode & MPOL_MODE_FLAGS;
1349 mode &= ~MPOL_MODE_FLAGS;
1350 if (mode >= MPOL_MAX)
1352 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1353 (mode_flags & MPOL_F_RELATIVE_NODES))
1355 err = get_nodes(&nodes, nmask, maxnode);
1358 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1361 /* Set the process memory policy */
1362 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1363 unsigned long, maxnode)
1367 unsigned short flags;
1369 flags = mode & MPOL_MODE_FLAGS;
1370 mode &= ~MPOL_MODE_FLAGS;
1371 if ((unsigned int)mode >= MPOL_MAX)
1373 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1375 err = get_nodes(&nodes, nmask, maxnode);
1378 return do_set_mempolicy(mode, flags, &nodes);
1381 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1382 const unsigned long __user *, old_nodes,
1383 const unsigned long __user *, new_nodes)
1385 const struct cred *cred = current_cred(), *tcred;
1386 struct mm_struct *mm = NULL;
1387 struct task_struct *task;
1388 nodemask_t task_nodes;
1392 NODEMASK_SCRATCH(scratch);
1397 old = &scratch->mask1;
1398 new = &scratch->mask2;
1400 err = get_nodes(old, old_nodes, maxnode);
1404 err = get_nodes(new, new_nodes, maxnode);
1408 /* Find the mm_struct */
1410 task = pid ? find_task_by_vpid(pid) : current;
1416 get_task_struct(task);
1421 * Check if this process has the right to modify the specified
1422 * process. The right exists if the process has administrative
1423 * capabilities, superuser privileges or the same
1424 * userid as the target process.
1426 tcred = __task_cred(task);
1427 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1428 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1429 !capable(CAP_SYS_NICE)) {
1436 task_nodes = cpuset_mems_allowed(task);
1437 /* Is the user allowed to access the target nodes? */
1438 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1443 task_nodes = cpuset_mems_allowed(current);
1444 nodes_and(*new, *new, task_nodes);
1445 if (nodes_empty(*new))
1448 nodes_and(*new, *new, node_states[N_MEMORY]);
1449 if (nodes_empty(*new))
1452 err = security_task_movememory(task);
1456 mm = get_task_mm(task);
1457 put_task_struct(task);
1464 err = do_migrate_pages(mm, old, new,
1465 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1469 NODEMASK_SCRATCH_FREE(scratch);
1474 put_task_struct(task);
1480 /* Retrieve NUMA policy */
1481 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1482 unsigned long __user *, nmask, unsigned long, maxnode,
1483 unsigned long, addr, unsigned long, flags)
1486 int uninitialized_var(pval);
1489 if (nmask != NULL && maxnode < MAX_NUMNODES)
1492 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1497 if (policy && put_user(pval, policy))
1501 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1506 #ifdef CONFIG_COMPAT
1508 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1509 compat_ulong_t __user *, nmask,
1510 compat_ulong_t, maxnode,
1511 compat_ulong_t, addr, compat_ulong_t, flags)
1514 unsigned long __user *nm = NULL;
1515 unsigned long nr_bits, alloc_size;
1516 DECLARE_BITMAP(bm, MAX_NUMNODES);
1518 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1519 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1522 nm = compat_alloc_user_space(alloc_size);
1524 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1526 if (!err && nmask) {
1527 unsigned long copy_size;
1528 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1529 err = copy_from_user(bm, nm, copy_size);
1530 /* ensure entire bitmap is zeroed */
1531 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1532 err |= compat_put_bitmap(nmask, bm, nr_bits);
1538 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1539 compat_ulong_t, maxnode)
1541 unsigned long __user *nm = NULL;
1542 unsigned long nr_bits, alloc_size;
1543 DECLARE_BITMAP(bm, MAX_NUMNODES);
1545 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1546 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1549 if (compat_get_bitmap(bm, nmask, nr_bits))
1551 nm = compat_alloc_user_space(alloc_size);
1552 if (copy_to_user(nm, bm, alloc_size))
1556 return sys_set_mempolicy(mode, nm, nr_bits+1);
1559 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1560 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1561 compat_ulong_t, maxnode, compat_ulong_t, flags)
1563 unsigned long __user *nm = NULL;
1564 unsigned long nr_bits, alloc_size;
1567 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1568 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1571 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1573 nm = compat_alloc_user_space(alloc_size);
1574 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1578 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1583 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1586 struct mempolicy *pol = NULL;
1589 if (vma->vm_ops && vma->vm_ops->get_policy) {
1590 pol = vma->vm_ops->get_policy(vma, addr);
1591 } else if (vma->vm_policy) {
1592 pol = vma->vm_policy;
1595 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1596 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1597 * count on these policies which will be dropped by
1598 * mpol_cond_put() later
1600 if (mpol_needs_cond_ref(pol))
1609 * get_vma_policy(@vma, @addr)
1610 * @vma: virtual memory area whose policy is sought
1611 * @addr: address in @vma for shared policy lookup
1613 * Returns effective policy for a VMA at specified address.
1614 * Falls back to current->mempolicy or system default policy, as necessary.
1615 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1616 * count--added by the get_policy() vm_op, as appropriate--to protect against
1617 * freeing by another task. It is the caller's responsibility to free the
1618 * extra reference for shared policies.
1620 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1623 struct mempolicy *pol = __get_vma_policy(vma, addr);
1626 pol = get_task_policy(current);
1631 bool vma_policy_mof(struct vm_area_struct *vma)
1633 struct mempolicy *pol;
1635 if (vma->vm_ops && vma->vm_ops->get_policy) {
1638 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1639 if (pol && (pol->flags & MPOL_F_MOF))
1646 pol = vma->vm_policy;
1648 pol = get_task_policy(current);
1650 return pol->flags & MPOL_F_MOF;
1653 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1655 enum zone_type dynamic_policy_zone = policy_zone;
1657 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1660 * if policy->v.nodes has movable memory only,
1661 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1663 * policy->v.nodes is intersect with node_states[N_MEMORY].
1664 * so if the following test faile, it implies
1665 * policy->v.nodes has movable memory only.
1667 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1668 dynamic_policy_zone = ZONE_MOVABLE;
1670 return zone >= dynamic_policy_zone;
1674 * Return a nodemask representing a mempolicy for filtering nodes for
1677 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1679 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1680 if (unlikely(policy->mode == MPOL_BIND) &&
1681 apply_policy_zone(policy, gfp_zone(gfp)) &&
1682 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1683 return &policy->v.nodes;
1688 /* Return the node id preferred by the given mempolicy, or the given id */
1689 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1692 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1693 nd = policy->v.preferred_node;
1696 * __GFP_THISNODE shouldn't even be used with the bind policy
1697 * because we might easily break the expectation to stay on the
1698 * requested node and not break the policy.
1700 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1706 /* Do dynamic interleaving for a process */
1707 static unsigned interleave_nodes(struct mempolicy *policy)
1710 struct task_struct *me = current;
1712 next = next_node_in(me->il_prev, policy->v.nodes);
1713 if (next < MAX_NUMNODES)
1719 * Depending on the memory policy provide a node from which to allocate the
1722 unsigned int mempolicy_slab_node(void)
1724 struct mempolicy *policy;
1725 int node = numa_mem_id();
1730 policy = current->mempolicy;
1731 if (!policy || policy->flags & MPOL_F_LOCAL)
1734 switch (policy->mode) {
1735 case MPOL_PREFERRED:
1737 * handled MPOL_F_LOCAL above
1739 return policy->v.preferred_node;
1741 case MPOL_INTERLEAVE:
1742 return interleave_nodes(policy);
1748 * Follow bind policy behavior and start allocation at the
1751 struct zonelist *zonelist;
1752 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1753 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1754 z = first_zones_zonelist(zonelist, highest_zoneidx,
1756 return z->zone ? z->zone->node : node;
1765 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1766 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1767 * number of present nodes.
1769 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1771 unsigned nnodes = nodes_weight(pol->v.nodes);
1777 return numa_node_id();
1778 target = (unsigned int)n % nnodes;
1779 nid = first_node(pol->v.nodes);
1780 for (i = 0; i < target; i++)
1781 nid = next_node(nid, pol->v.nodes);
1785 /* Determine a node number for interleave */
1786 static inline unsigned interleave_nid(struct mempolicy *pol,
1787 struct vm_area_struct *vma, unsigned long addr, int shift)
1793 * for small pages, there is no difference between
1794 * shift and PAGE_SHIFT, so the bit-shift is safe.
1795 * for huge pages, since vm_pgoff is in units of small
1796 * pages, we need to shift off the always 0 bits to get
1799 BUG_ON(shift < PAGE_SHIFT);
1800 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1801 off += (addr - vma->vm_start) >> shift;
1802 return offset_il_node(pol, off);
1804 return interleave_nodes(pol);
1807 #ifdef CONFIG_HUGETLBFS
1809 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1810 * @vma: virtual memory area whose policy is sought
1811 * @addr: address in @vma for shared policy lookup and interleave policy
1812 * @gfp_flags: for requested zone
1813 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1814 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1816 * Returns a nid suitable for a huge page allocation and a pointer
1817 * to the struct mempolicy for conditional unref after allocation.
1818 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1819 * @nodemask for filtering the zonelist.
1821 * Must be protected by read_mems_allowed_begin()
1823 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1824 struct mempolicy **mpol, nodemask_t **nodemask)
1828 *mpol = get_vma_policy(vma, addr);
1829 *nodemask = NULL; /* assume !MPOL_BIND */
1831 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1832 nid = interleave_nid(*mpol, vma, addr,
1833 huge_page_shift(hstate_vma(vma)));
1835 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1836 if ((*mpol)->mode == MPOL_BIND)
1837 *nodemask = &(*mpol)->v.nodes;
1843 * init_nodemask_of_mempolicy
1845 * If the current task's mempolicy is "default" [NULL], return 'false'
1846 * to indicate default policy. Otherwise, extract the policy nodemask
1847 * for 'bind' or 'interleave' policy into the argument nodemask, or
1848 * initialize the argument nodemask to contain the single node for
1849 * 'preferred' or 'local' policy and return 'true' to indicate presence
1850 * of non-default mempolicy.
1852 * We don't bother with reference counting the mempolicy [mpol_get/put]
1853 * because the current task is examining it's own mempolicy and a task's
1854 * mempolicy is only ever changed by the task itself.
1856 * N.B., it is the caller's responsibility to free a returned nodemask.
1858 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1860 struct mempolicy *mempolicy;
1863 if (!(mask && current->mempolicy))
1867 mempolicy = current->mempolicy;
1868 switch (mempolicy->mode) {
1869 case MPOL_PREFERRED:
1870 if (mempolicy->flags & MPOL_F_LOCAL)
1871 nid = numa_node_id();
1873 nid = mempolicy->v.preferred_node;
1874 init_nodemask_of_node(mask, nid);
1879 case MPOL_INTERLEAVE:
1880 *mask = mempolicy->v.nodes;
1886 task_unlock(current);
1893 * mempolicy_nodemask_intersects
1895 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1896 * policy. Otherwise, check for intersection between mask and the policy
1897 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1898 * policy, always return true since it may allocate elsewhere on fallback.
1900 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1902 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1903 const nodemask_t *mask)
1905 struct mempolicy *mempolicy;
1911 mempolicy = tsk->mempolicy;
1915 switch (mempolicy->mode) {
1916 case MPOL_PREFERRED:
1918 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1919 * allocate from, they may fallback to other nodes when oom.
1920 * Thus, it's possible for tsk to have allocated memory from
1925 case MPOL_INTERLEAVE:
1926 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1936 /* Allocate a page in interleaved policy.
1937 Own path because it needs to do special accounting. */
1938 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1943 page = __alloc_pages(gfp, order, nid);
1944 if (page && page_to_nid(page) == nid) {
1946 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1953 * alloc_pages_vma - Allocate a page for a VMA.
1956 * %GFP_USER user allocation.
1957 * %GFP_KERNEL kernel allocations,
1958 * %GFP_HIGHMEM highmem/user allocations,
1959 * %GFP_FS allocation should not call back into a file system.
1960 * %GFP_ATOMIC don't sleep.
1962 * @order:Order of the GFP allocation.
1963 * @vma: Pointer to VMA or NULL if not available.
1964 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1965 * @node: Which node to prefer for allocation (modulo policy).
1966 * @hugepage: for hugepages try only the preferred node if possible
1968 * This function allocates a page from the kernel page pool and applies
1969 * a NUMA policy associated with the VMA or the current process.
1970 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1971 * mm_struct of the VMA to prevent it from going away. Should be used for
1972 * all allocations for pages that will be mapped into user space. Returns
1973 * NULL when no page can be allocated.
1976 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1977 unsigned long addr, int node, bool hugepage)
1979 struct mempolicy *pol;
1984 pol = get_vma_policy(vma, addr);
1986 if (pol->mode == MPOL_INTERLEAVE) {
1989 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1991 page = alloc_page_interleave(gfp, order, nid);
1995 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1996 int hpage_node = node;
1999 * For hugepage allocation and non-interleave policy which
2000 * allows the current node (or other explicitly preferred
2001 * node) we only try to allocate from the current/preferred
2002 * node and don't fall back to other nodes, as the cost of
2003 * remote accesses would likely offset THP benefits.
2005 * If the policy is interleave, or does not allow the current
2006 * node in its nodemask, we allocate the standard way.
2008 if (pol->mode == MPOL_PREFERRED &&
2009 !(pol->flags & MPOL_F_LOCAL))
2010 hpage_node = pol->v.preferred_node;
2012 nmask = policy_nodemask(gfp, pol);
2013 if (!nmask || node_isset(hpage_node, *nmask)) {
2015 page = __alloc_pages_node(hpage_node,
2016 gfp | __GFP_THISNODE, order);
2021 nmask = policy_nodemask(gfp, pol);
2022 preferred_nid = policy_node(gfp, pol, node);
2023 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2030 * alloc_pages_current - Allocate pages.
2033 * %GFP_USER user allocation,
2034 * %GFP_KERNEL kernel allocation,
2035 * %GFP_HIGHMEM highmem allocation,
2036 * %GFP_FS don't call back into a file system.
2037 * %GFP_ATOMIC don't sleep.
2038 * @order: Power of two of allocation size in pages. 0 is a single page.
2040 * Allocate a page from the kernel page pool. When not in
2041 * interrupt context and apply the current process NUMA policy.
2042 * Returns NULL when no page can be allocated.
2044 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2046 struct mempolicy *pol = &default_policy;
2049 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2050 pol = get_task_policy(current);
2053 * No reference counting needed for current->mempolicy
2054 * nor system default_policy
2056 if (pol->mode == MPOL_INTERLEAVE)
2057 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2059 page = __alloc_pages_nodemask(gfp, order,
2060 policy_node(gfp, pol, numa_node_id()),
2061 policy_nodemask(gfp, pol));
2065 EXPORT_SYMBOL(alloc_pages_current);
2067 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2069 struct mempolicy *pol = mpol_dup(vma_policy(src));
2072 return PTR_ERR(pol);
2073 dst->vm_policy = pol;
2078 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2079 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2080 * with the mems_allowed returned by cpuset_mems_allowed(). This
2081 * keeps mempolicies cpuset relative after its cpuset moves. See
2082 * further kernel/cpuset.c update_nodemask().
2084 * current's mempolicy may be rebinded by the other task(the task that changes
2085 * cpuset's mems), so we needn't do rebind work for current task.
2088 /* Slow path of a mempolicy duplicate */
2089 struct mempolicy *__mpol_dup(struct mempolicy *old)
2091 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2094 return ERR_PTR(-ENOMEM);
2096 /* task's mempolicy is protected by alloc_lock */
2097 if (old == current->mempolicy) {
2100 task_unlock(current);
2104 if (current_cpuset_is_being_rebound()) {
2105 nodemask_t mems = cpuset_mems_allowed(current);
2106 mpol_rebind_policy(new, &mems);
2108 atomic_set(&new->refcnt, 1);
2112 /* Slow path of a mempolicy comparison */
2113 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2117 if (a->mode != b->mode)
2119 if (a->flags != b->flags)
2121 if (mpol_store_user_nodemask(a))
2122 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2128 case MPOL_INTERLEAVE:
2129 return !!nodes_equal(a->v.nodes, b->v.nodes);
2130 case MPOL_PREFERRED:
2131 return a->v.preferred_node == b->v.preferred_node;
2139 * Shared memory backing store policy support.
2141 * Remember policies even when nobody has shared memory mapped.
2142 * The policies are kept in Red-Black tree linked from the inode.
2143 * They are protected by the sp->lock rwlock, which should be held
2144 * for any accesses to the tree.
2148 * lookup first element intersecting start-end. Caller holds sp->lock for
2149 * reading or for writing
2151 static struct sp_node *
2152 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2154 struct rb_node *n = sp->root.rb_node;
2157 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2159 if (start >= p->end)
2161 else if (end <= p->start)
2169 struct sp_node *w = NULL;
2170 struct rb_node *prev = rb_prev(n);
2173 w = rb_entry(prev, struct sp_node, nd);
2174 if (w->end <= start)
2178 return rb_entry(n, struct sp_node, nd);
2182 * Insert a new shared policy into the list. Caller holds sp->lock for
2185 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2187 struct rb_node **p = &sp->root.rb_node;
2188 struct rb_node *parent = NULL;
2193 nd = rb_entry(parent, struct sp_node, nd);
2194 if (new->start < nd->start)
2196 else if (new->end > nd->end)
2197 p = &(*p)->rb_right;
2201 rb_link_node(&new->nd, parent, p);
2202 rb_insert_color(&new->nd, &sp->root);
2203 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2204 new->policy ? new->policy->mode : 0);
2207 /* Find shared policy intersecting idx */
2209 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2211 struct mempolicy *pol = NULL;
2214 if (!sp->root.rb_node)
2216 read_lock(&sp->lock);
2217 sn = sp_lookup(sp, idx, idx+1);
2219 mpol_get(sn->policy);
2222 read_unlock(&sp->lock);
2226 static void sp_free(struct sp_node *n)
2228 mpol_put(n->policy);
2229 kmem_cache_free(sn_cache, n);
2233 * mpol_misplaced - check whether current page node is valid in policy
2235 * @page: page to be checked
2236 * @vma: vm area where page mapped
2237 * @addr: virtual address where page mapped
2239 * Lookup current policy node id for vma,addr and "compare to" page's
2243 * -1 - not misplaced, page is in the right node
2244 * node - node id where the page should be
2246 * Policy determination "mimics" alloc_page_vma().
2247 * Called from fault path where we know the vma and faulting address.
2249 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2251 struct mempolicy *pol;
2253 int curnid = page_to_nid(page);
2254 unsigned long pgoff;
2255 int thiscpu = raw_smp_processor_id();
2256 int thisnid = cpu_to_node(thiscpu);
2260 pol = get_vma_policy(vma, addr);
2261 if (!(pol->flags & MPOL_F_MOF))
2264 switch (pol->mode) {
2265 case MPOL_INTERLEAVE:
2266 pgoff = vma->vm_pgoff;
2267 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2268 polnid = offset_il_node(pol, pgoff);
2271 case MPOL_PREFERRED:
2272 if (pol->flags & MPOL_F_LOCAL)
2273 polnid = numa_node_id();
2275 polnid = pol->v.preferred_node;
2281 * allows binding to multiple nodes.
2282 * use current page if in policy nodemask,
2283 * else select nearest allowed node, if any.
2284 * If no allowed nodes, use current [!misplaced].
2286 if (node_isset(curnid, pol->v.nodes))
2288 z = first_zones_zonelist(
2289 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2290 gfp_zone(GFP_HIGHUSER),
2292 polnid = z->zone->node;
2299 /* Migrate the page towards the node whose CPU is referencing it */
2300 if (pol->flags & MPOL_F_MORON) {
2303 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2307 if (curnid != polnid)
2316 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2317 * dropped after task->mempolicy is set to NULL so that any allocation done as
2318 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2321 void mpol_put_task_policy(struct task_struct *task)
2323 struct mempolicy *pol;
2326 pol = task->mempolicy;
2327 task->mempolicy = NULL;
2332 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2334 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2335 rb_erase(&n->nd, &sp->root);
2339 static void sp_node_init(struct sp_node *node, unsigned long start,
2340 unsigned long end, struct mempolicy *pol)
2342 node->start = start;
2347 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2348 struct mempolicy *pol)
2351 struct mempolicy *newpol;
2353 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2357 newpol = mpol_dup(pol);
2358 if (IS_ERR(newpol)) {
2359 kmem_cache_free(sn_cache, n);
2362 newpol->flags |= MPOL_F_SHARED;
2363 sp_node_init(n, start, end, newpol);
2368 /* Replace a policy range. */
2369 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2370 unsigned long end, struct sp_node *new)
2373 struct sp_node *n_new = NULL;
2374 struct mempolicy *mpol_new = NULL;
2378 write_lock(&sp->lock);
2379 n = sp_lookup(sp, start, end);
2380 /* Take care of old policies in the same range. */
2381 while (n && n->start < end) {
2382 struct rb_node *next = rb_next(&n->nd);
2383 if (n->start >= start) {
2389 /* Old policy spanning whole new range. */
2394 *mpol_new = *n->policy;
2395 atomic_set(&mpol_new->refcnt, 1);
2396 sp_node_init(n_new, end, n->end, mpol_new);
2398 sp_insert(sp, n_new);
2407 n = rb_entry(next, struct sp_node, nd);
2411 write_unlock(&sp->lock);
2418 kmem_cache_free(sn_cache, n_new);
2423 write_unlock(&sp->lock);
2425 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2428 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2435 * mpol_shared_policy_init - initialize shared policy for inode
2436 * @sp: pointer to inode shared policy
2437 * @mpol: struct mempolicy to install
2439 * Install non-NULL @mpol in inode's shared policy rb-tree.
2440 * On entry, the current task has a reference on a non-NULL @mpol.
2441 * This must be released on exit.
2442 * This is called at get_inode() calls and we can use GFP_KERNEL.
2444 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2448 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2449 rwlock_init(&sp->lock);
2452 struct vm_area_struct pvma;
2453 struct mempolicy *new;
2454 NODEMASK_SCRATCH(scratch);
2458 /* contextualize the tmpfs mount point mempolicy */
2459 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2461 goto free_scratch; /* no valid nodemask intersection */
2464 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2465 task_unlock(current);
2469 /* Create pseudo-vma that contains just the policy */
2470 memset(&pvma, 0, sizeof(struct vm_area_struct));
2471 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2472 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2475 mpol_put(new); /* drop initial ref */
2477 NODEMASK_SCRATCH_FREE(scratch);
2479 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2483 int mpol_set_shared_policy(struct shared_policy *info,
2484 struct vm_area_struct *vma, struct mempolicy *npol)
2487 struct sp_node *new = NULL;
2488 unsigned long sz = vma_pages(vma);
2490 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2492 sz, npol ? npol->mode : -1,
2493 npol ? npol->flags : -1,
2494 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2497 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2501 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2507 /* Free a backing policy store on inode delete. */
2508 void mpol_free_shared_policy(struct shared_policy *p)
2511 struct rb_node *next;
2513 if (!p->root.rb_node)
2515 write_lock(&p->lock);
2516 next = rb_first(&p->root);
2518 n = rb_entry(next, struct sp_node, nd);
2519 next = rb_next(&n->nd);
2522 write_unlock(&p->lock);
2525 #ifdef CONFIG_NUMA_BALANCING
2526 static int __initdata numabalancing_override;
2528 static void __init check_numabalancing_enable(void)
2530 bool numabalancing_default = false;
2532 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2533 numabalancing_default = true;
2535 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2536 if (numabalancing_override)
2537 set_numabalancing_state(numabalancing_override == 1);
2539 if (num_online_nodes() > 1 && !numabalancing_override) {
2540 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2541 numabalancing_default ? "Enabling" : "Disabling");
2542 set_numabalancing_state(numabalancing_default);
2546 static int __init setup_numabalancing(char *str)
2552 if (!strcmp(str, "enable")) {
2553 numabalancing_override = 1;
2555 } else if (!strcmp(str, "disable")) {
2556 numabalancing_override = -1;
2561 pr_warn("Unable to parse numa_balancing=\n");
2565 __setup("numa_balancing=", setup_numabalancing);
2567 static inline void __init check_numabalancing_enable(void)
2570 #endif /* CONFIG_NUMA_BALANCING */
2572 /* assumes fs == KERNEL_DS */
2573 void __init numa_policy_init(void)
2575 nodemask_t interleave_nodes;
2576 unsigned long largest = 0;
2577 int nid, prefer = 0;
2579 policy_cache = kmem_cache_create("numa_policy",
2580 sizeof(struct mempolicy),
2581 0, SLAB_PANIC, NULL);
2583 sn_cache = kmem_cache_create("shared_policy_node",
2584 sizeof(struct sp_node),
2585 0, SLAB_PANIC, NULL);
2587 for_each_node(nid) {
2588 preferred_node_policy[nid] = (struct mempolicy) {
2589 .refcnt = ATOMIC_INIT(1),
2590 .mode = MPOL_PREFERRED,
2591 .flags = MPOL_F_MOF | MPOL_F_MORON,
2592 .v = { .preferred_node = nid, },
2597 * Set interleaving policy for system init. Interleaving is only
2598 * enabled across suitably sized nodes (default is >= 16MB), or
2599 * fall back to the largest node if they're all smaller.
2601 nodes_clear(interleave_nodes);
2602 for_each_node_state(nid, N_MEMORY) {
2603 unsigned long total_pages = node_present_pages(nid);
2605 /* Preserve the largest node */
2606 if (largest < total_pages) {
2607 largest = total_pages;
2611 /* Interleave this node? */
2612 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2613 node_set(nid, interleave_nodes);
2616 /* All too small, use the largest */
2617 if (unlikely(nodes_empty(interleave_nodes)))
2618 node_set(prefer, interleave_nodes);
2620 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2621 pr_err("%s: interleaving failed\n", __func__);
2623 check_numabalancing_enable();
2626 /* Reset policy of current process to default */
2627 void numa_default_policy(void)
2629 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2633 * Parse and format mempolicy from/to strings
2637 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2639 static const char * const policy_modes[] =
2641 [MPOL_DEFAULT] = "default",
2642 [MPOL_PREFERRED] = "prefer",
2643 [MPOL_BIND] = "bind",
2644 [MPOL_INTERLEAVE] = "interleave",
2645 [MPOL_LOCAL] = "local",
2651 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2652 * @str: string containing mempolicy to parse
2653 * @mpol: pointer to struct mempolicy pointer, returned on success.
2656 * <mode>[=<flags>][:<nodelist>]
2658 * On success, returns 0, else 1
2660 int mpol_parse_str(char *str, struct mempolicy **mpol)
2662 struct mempolicy *new = NULL;
2663 unsigned short mode;
2664 unsigned short mode_flags;
2666 char *nodelist = strchr(str, ':');
2667 char *flags = strchr(str, '=');
2671 /* NUL-terminate mode or flags string */
2673 if (nodelist_parse(nodelist, nodes))
2675 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2681 *flags++ = '\0'; /* terminate mode string */
2683 for (mode = 0; mode < MPOL_MAX; mode++) {
2684 if (!strcmp(str, policy_modes[mode])) {
2688 if (mode >= MPOL_MAX)
2692 case MPOL_PREFERRED:
2694 * Insist on a nodelist of one node only
2697 char *rest = nodelist;
2698 while (isdigit(*rest))
2704 case MPOL_INTERLEAVE:
2706 * Default to online nodes with memory if no nodelist
2709 nodes = node_states[N_MEMORY];
2713 * Don't allow a nodelist; mpol_new() checks flags
2717 mode = MPOL_PREFERRED;
2721 * Insist on a empty nodelist
2728 * Insist on a nodelist
2737 * Currently, we only support two mutually exclusive
2740 if (!strcmp(flags, "static"))
2741 mode_flags |= MPOL_F_STATIC_NODES;
2742 else if (!strcmp(flags, "relative"))
2743 mode_flags |= MPOL_F_RELATIVE_NODES;
2748 new = mpol_new(mode, mode_flags, &nodes);
2753 * Save nodes for mpol_to_str() to show the tmpfs mount options
2754 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2756 if (mode != MPOL_PREFERRED)
2757 new->v.nodes = nodes;
2759 new->v.preferred_node = first_node(nodes);
2761 new->flags |= MPOL_F_LOCAL;
2764 * Save nodes for contextualization: this will be used to "clone"
2765 * the mempolicy in a specific context [cpuset] at a later time.
2767 new->w.user_nodemask = nodes;
2772 /* Restore string for error message */
2781 #endif /* CONFIG_TMPFS */
2784 * mpol_to_str - format a mempolicy structure for printing
2785 * @buffer: to contain formatted mempolicy string
2786 * @maxlen: length of @buffer
2787 * @pol: pointer to mempolicy to be formatted
2789 * Convert @pol into a string. If @buffer is too short, truncate the string.
2790 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2791 * longest flag, "relative", and to display at least a few node ids.
2793 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2796 nodemask_t nodes = NODE_MASK_NONE;
2797 unsigned short mode = MPOL_DEFAULT;
2798 unsigned short flags = 0;
2800 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2808 case MPOL_PREFERRED:
2809 if (flags & MPOL_F_LOCAL)
2812 node_set(pol->v.preferred_node, nodes);
2815 case MPOL_INTERLEAVE:
2816 nodes = pol->v.nodes;
2820 snprintf(p, maxlen, "unknown");
2824 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2826 if (flags & MPOL_MODE_FLAGS) {
2827 p += snprintf(p, buffer + maxlen - p, "=");
2830 * Currently, the only defined flags are mutually exclusive
2832 if (flags & MPOL_F_STATIC_NODES)
2833 p += snprintf(p, buffer + maxlen - p, "static");
2834 else if (flags & MPOL_F_RELATIVE_NODES)
2835 p += snprintf(p, buffer + maxlen - p, "relative");
2838 if (!nodes_empty(nodes))
2839 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2840 nodemask_pr_args(&nodes));