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 = *nodes;
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)
350 if (!pol || pol->mode == MPOL_LOCAL)
352 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
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);
431 * queue_pages_pmd() has three possible return values:
432 * 1 - pages are placed on the right node or queued successfully.
434 * -EIO - is migration entry or MPOL_MF_STRICT was specified and an existing
435 * page was already on a node that does not follow the policy.
437 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
438 unsigned long end, struct mm_walk *walk)
442 struct queue_pages *qp = walk->private;
445 if (unlikely(is_pmd_migration_entry(*pmd))) {
449 page = pmd_page(*pmd);
450 if (is_huge_zero_page(page)) {
452 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
455 if (!thp_migration_supported()) {
459 ret = split_huge_page(page);
464 if (!queue_pages_required(page, qp)) {
471 /* go to thp migration */
472 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
473 if (!vma_migratable(walk->vma)) {
478 migrate_page_add(page, qp->pagelist, flags);
488 * Scan through pages checking if pages follow certain conditions,
489 * and move them to the pagelist if they do.
491 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
492 unsigned long end, struct mm_walk *walk)
494 struct vm_area_struct *vma = walk->vma;
496 struct queue_pages *qp = walk->private;
497 unsigned long flags = qp->flags;
499 pte_t *pte, *mapped_pte;
502 ptl = pmd_trans_huge_lock(pmd, vma);
504 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
511 if (pmd_trans_unstable(pmd))
514 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
515 for (; addr != end; pte++, addr += PAGE_SIZE) {
516 if (!pte_present(*pte))
518 page = vm_normal_page(vma, addr, *pte);
522 * vm_normal_page() filters out zero pages, but there might
523 * still be PageReserved pages to skip, perhaps in a VDSO.
525 if (PageReserved(page))
527 if (!queue_pages_required(page, qp))
529 if (PageTransCompound(page) && !thp_migration_supported()) {
531 pte_unmap_unlock(pte, ptl);
533 ret = split_huge_page(page);
536 /* Failed to split -- skip. */
538 pte = pte_offset_map_lock(walk->mm, pmd,
545 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
546 if (!vma_migratable(vma))
548 migrate_page_add(page, qp->pagelist, flags);
552 pte_unmap_unlock(mapped_pte, ptl);
554 return addr != end ? -EIO : 0;
557 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
558 unsigned long addr, unsigned long end,
559 struct mm_walk *walk)
561 #ifdef CONFIG_HUGETLB_PAGE
562 struct queue_pages *qp = walk->private;
563 unsigned long flags = qp->flags;
568 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
569 entry = huge_ptep_get(pte);
570 if (!pte_present(entry))
572 page = pte_page(entry);
573 if (!queue_pages_required(page, qp))
575 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
576 if (flags & (MPOL_MF_MOVE_ALL) ||
577 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1 &&
578 !hugetlb_pmd_shared(pte)))
579 isolate_huge_page(page, qp->pagelist);
588 #ifdef CONFIG_NUMA_BALANCING
590 * This is used to mark a range of virtual addresses to be inaccessible.
591 * These are later cleared by a NUMA hinting fault. Depending on these
592 * faults, pages may be migrated for better NUMA placement.
594 * This is assuming that NUMA faults are handled using PROT_NONE. If
595 * an architecture makes a different choice, it will need further
596 * changes to the core.
598 unsigned long change_prot_numa(struct vm_area_struct *vma,
599 unsigned long addr, unsigned long end)
603 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
605 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
610 static unsigned long change_prot_numa(struct vm_area_struct *vma,
611 unsigned long addr, unsigned long end)
615 #endif /* CONFIG_NUMA_BALANCING */
617 static int queue_pages_test_walk(unsigned long start, unsigned long end,
618 struct mm_walk *walk)
620 struct vm_area_struct *vma = walk->vma;
621 struct queue_pages *qp = walk->private;
622 unsigned long endvma = vma->vm_end;
623 unsigned long flags = qp->flags;
626 * Need check MPOL_MF_STRICT to return -EIO if possible
627 * regardless of vma_migratable
629 if (!vma_migratable(vma) &&
630 !(flags & MPOL_MF_STRICT))
635 if (vma->vm_start > start)
636 start = vma->vm_start;
638 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
639 if (!vma->vm_next && vma->vm_end < end)
641 if (qp->prev && qp->prev->vm_end < vma->vm_start)
647 if (flags & MPOL_MF_LAZY) {
648 /* Similar to task_numa_work, skip inaccessible VMAs */
649 if (!is_vm_hugetlb_page(vma) &&
650 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
651 !(vma->vm_flags & VM_MIXEDMAP))
652 change_prot_numa(vma, start, endvma);
656 /* queue pages from current vma */
657 if (flags & MPOL_MF_VALID)
663 * Walk through page tables and collect pages to be migrated.
665 * If pages found in a given range are on a set of nodes (determined by
666 * @nodes and @flags,) it's isolated and queued to the pagelist which is
667 * passed via @private.)
670 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
671 nodemask_t *nodes, unsigned long flags,
672 struct list_head *pagelist)
674 struct queue_pages qp = {
675 .pagelist = pagelist,
680 struct mm_walk queue_pages_walk = {
681 .hugetlb_entry = queue_pages_hugetlb,
682 .pmd_entry = queue_pages_pte_range,
683 .test_walk = queue_pages_test_walk,
688 return walk_page_range(start, end, &queue_pages_walk);
692 * Apply policy to a single VMA
693 * This must be called with the mmap_sem held for writing.
695 static int vma_replace_policy(struct vm_area_struct *vma,
696 struct mempolicy *pol)
699 struct mempolicy *old;
700 struct mempolicy *new;
702 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
703 vma->vm_start, vma->vm_end, vma->vm_pgoff,
704 vma->vm_ops, vma->vm_file,
705 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
711 if (vma->vm_ops && vma->vm_ops->set_policy) {
712 err = vma->vm_ops->set_policy(vma, new);
717 old = vma->vm_policy;
718 vma->vm_policy = new; /* protected by mmap_sem */
727 /* Step 2: apply policy to a range and do splits. */
728 static int mbind_range(struct mm_struct *mm, unsigned long start,
729 unsigned long end, struct mempolicy *new_pol)
731 struct vm_area_struct *prev;
732 struct vm_area_struct *vma;
735 unsigned long vmstart;
738 vma = find_vma(mm, start);
739 if (!vma || vma->vm_start > start)
743 if (start > vma->vm_start)
746 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
747 vmstart = max(start, vma->vm_start);
748 vmend = min(end, vma->vm_end);
750 if (mpol_equal(vma_policy(vma), new_pol))
753 pgoff = vma->vm_pgoff +
754 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
755 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
756 vma->anon_vma, vma->vm_file, pgoff,
757 new_pol, vma->vm_userfaultfd_ctx);
762 if (vma->vm_start != vmstart) {
763 err = split_vma(vma->vm_mm, vma, vmstart, 1);
767 if (vma->vm_end != vmend) {
768 err = split_vma(vma->vm_mm, vma, vmend, 0);
773 err = vma_replace_policy(vma, new_pol);
782 /* Set the process memory policy */
783 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
786 struct mempolicy *new, *old;
787 NODEMASK_SCRATCH(scratch);
793 new = mpol_new(mode, flags, nodes);
800 ret = mpol_set_nodemask(new, nodes, scratch);
802 task_unlock(current);
806 old = current->mempolicy;
807 current->mempolicy = new;
808 if (new && new->mode == MPOL_INTERLEAVE)
809 current->il_prev = MAX_NUMNODES-1;
810 task_unlock(current);
814 NODEMASK_SCRATCH_FREE(scratch);
819 * Return nodemask for policy for get_mempolicy() query
821 * Called with task's alloc_lock held
823 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
826 if (p == &default_policy)
832 case MPOL_INTERLEAVE:
836 if (!(p->flags & MPOL_F_LOCAL))
837 node_set(p->v.preferred_node, *nodes);
838 /* else return empty node mask for local allocation */
845 static int lookup_node(unsigned long addr)
850 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
852 err = page_to_nid(p);
858 /* Retrieve NUMA policy */
859 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
860 unsigned long addr, unsigned long flags)
863 struct mm_struct *mm = current->mm;
864 struct vm_area_struct *vma = NULL;
865 struct mempolicy *pol = current->mempolicy;
868 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
871 if (flags & MPOL_F_MEMS_ALLOWED) {
872 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
874 *policy = 0; /* just so it's initialized */
876 *nmask = cpuset_current_mems_allowed;
877 task_unlock(current);
881 if (flags & MPOL_F_ADDR) {
883 * Do NOT fall back to task policy if the
884 * vma/shared policy at addr is NULL. We
885 * want to return MPOL_DEFAULT in this case.
887 down_read(&mm->mmap_sem);
888 vma = find_vma_intersection(mm, addr, addr+1);
890 up_read(&mm->mmap_sem);
893 if (vma->vm_ops && vma->vm_ops->get_policy)
894 pol = vma->vm_ops->get_policy(vma, addr);
896 pol = vma->vm_policy;
901 pol = &default_policy; /* indicates default behavior */
903 if (flags & MPOL_F_NODE) {
904 if (flags & MPOL_F_ADDR) {
905 err = lookup_node(addr);
909 } else if (pol == current->mempolicy &&
910 pol->mode == MPOL_INTERLEAVE) {
911 *policy = next_node_in(current->il_prev, pol->v.nodes);
917 *policy = pol == &default_policy ? MPOL_DEFAULT :
920 * Internal mempolicy flags must be masked off before exposing
921 * the policy to userspace.
923 *policy |= (pol->flags & MPOL_MODE_FLAGS);
928 if (mpol_store_user_nodemask(pol)) {
929 *nmask = pol->w.user_nodemask;
932 get_policy_nodemask(pol, nmask);
933 task_unlock(current);
940 up_read(¤t->mm->mmap_sem);
944 #ifdef CONFIG_MIGRATION
946 * page migration, thp tail pages can be passed.
948 static void migrate_page_add(struct page *page, struct list_head *pagelist,
951 struct page *head = compound_head(page);
953 * Avoid migrating a page that is shared with others.
955 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
956 if (!isolate_lru_page(head)) {
957 list_add_tail(&head->lru, pagelist);
958 mod_node_page_state(page_pgdat(head),
959 NR_ISOLATED_ANON + page_is_file_cache(head),
960 hpage_nr_pages(head));
965 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
968 return alloc_huge_page_node(page_hstate(compound_head(page)),
970 else if (thp_migration_supported() && PageTransHuge(page)) {
973 thp = alloc_pages_node(node,
974 (GFP_TRANSHUGE | __GFP_THISNODE),
978 prep_transhuge_page(thp);
981 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
986 * Migrate pages from one node to a target node.
987 * Returns error or the number of pages not migrated.
989 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
997 node_set(source, nmask);
1000 * This does not "check" the range but isolates all pages that
1001 * need migration. Between passing in the full user address
1002 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1004 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1005 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1006 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1008 if (!list_empty(&pagelist)) {
1009 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1010 MIGRATE_SYNC, MR_SYSCALL);
1012 putback_movable_pages(&pagelist);
1019 * Move pages between the two nodesets so as to preserve the physical
1020 * layout as much as possible.
1022 * Returns the number of page that could not be moved.
1024 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1025 const nodemask_t *to, int flags)
1031 err = migrate_prep();
1035 down_read(&mm->mmap_sem);
1038 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1039 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1040 * bit in 'tmp', and return that <source, dest> pair for migration.
1041 * The pair of nodemasks 'to' and 'from' define the map.
1043 * If no pair of bits is found that way, fallback to picking some
1044 * pair of 'source' and 'dest' bits that are not the same. If the
1045 * 'source' and 'dest' bits are the same, this represents a node
1046 * that will be migrating to itself, so no pages need move.
1048 * If no bits are left in 'tmp', or if all remaining bits left
1049 * in 'tmp' correspond to the same bit in 'to', return false
1050 * (nothing left to migrate).
1052 * This lets us pick a pair of nodes to migrate between, such that
1053 * if possible the dest node is not already occupied by some other
1054 * source node, minimizing the risk of overloading the memory on a
1055 * node that would happen if we migrated incoming memory to a node
1056 * before migrating outgoing memory source that same node.
1058 * A single scan of tmp is sufficient. As we go, we remember the
1059 * most recent <s, d> pair that moved (s != d). If we find a pair
1060 * that not only moved, but what's better, moved to an empty slot
1061 * (d is not set in tmp), then we break out then, with that pair.
1062 * Otherwise when we finish scanning from_tmp, we at least have the
1063 * most recent <s, d> pair that moved. If we get all the way through
1064 * the scan of tmp without finding any node that moved, much less
1065 * moved to an empty node, then there is nothing left worth migrating.
1069 while (!nodes_empty(tmp)) {
1071 int source = NUMA_NO_NODE;
1074 for_each_node_mask(s, tmp) {
1077 * do_migrate_pages() tries to maintain the relative
1078 * node relationship of the pages established between
1079 * threads and memory areas.
1081 * However if the number of source nodes is not equal to
1082 * the number of destination nodes we can not preserve
1083 * this node relative relationship. In that case, skip
1084 * copying memory from a node that is in the destination
1087 * Example: [2,3,4] -> [3,4,5] moves everything.
1088 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1091 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1092 (node_isset(s, *to)))
1095 d = node_remap(s, *from, *to);
1099 source = s; /* Node moved. Memorize */
1102 /* dest not in remaining from nodes? */
1103 if (!node_isset(dest, tmp))
1106 if (source == NUMA_NO_NODE)
1109 node_clear(source, tmp);
1110 err = migrate_to_node(mm, source, dest, flags);
1116 up_read(&mm->mmap_sem);
1124 * Allocate a new page for page migration based on vma policy.
1125 * Start by assuming the page is mapped by the same vma as contains @start.
1126 * Search forward from there, if not. N.B., this assumes that the
1127 * list of pages handed to migrate_pages()--which is how we get here--
1128 * is in virtual address order.
1130 static struct page *new_page(struct page *page, unsigned long start, int **x)
1132 struct vm_area_struct *vma;
1133 unsigned long uninitialized_var(address);
1135 vma = find_vma(current->mm, start);
1137 address = page_address_in_vma(page, vma);
1138 if (address != -EFAULT)
1143 if (PageHuge(page)) {
1145 return alloc_huge_page_noerr(vma, address, 1);
1146 } else if (thp_migration_supported() && PageTransHuge(page)) {
1149 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1153 prep_transhuge_page(thp);
1157 * if !vma, alloc_page_vma() will use task or system default policy
1159 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1164 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1165 unsigned long flags)
1169 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1170 const nodemask_t *to, int flags)
1175 static struct page *new_page(struct page *page, unsigned long start, int **x)
1181 static long do_mbind(unsigned long start, unsigned long len,
1182 unsigned short mode, unsigned short mode_flags,
1183 nodemask_t *nmask, unsigned long flags)
1185 struct mm_struct *mm = current->mm;
1186 struct mempolicy *new;
1189 LIST_HEAD(pagelist);
1191 if (flags & ~(unsigned long)MPOL_MF_VALID)
1193 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1196 if (start & ~PAGE_MASK)
1199 if (mode == MPOL_DEFAULT)
1200 flags &= ~MPOL_MF_STRICT;
1202 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1210 new = mpol_new(mode, mode_flags, nmask);
1212 return PTR_ERR(new);
1214 if (flags & MPOL_MF_LAZY)
1215 new->flags |= MPOL_F_MOF;
1218 * If we are using the default policy then operation
1219 * on discontinuous address spaces is okay after all
1222 flags |= MPOL_MF_DISCONTIG_OK;
1224 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1225 start, start + len, mode, mode_flags,
1226 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1228 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1230 err = migrate_prep();
1235 NODEMASK_SCRATCH(scratch);
1237 down_write(&mm->mmap_sem);
1239 err = mpol_set_nodemask(new, nmask, scratch);
1240 task_unlock(current);
1242 up_write(&mm->mmap_sem);
1245 NODEMASK_SCRATCH_FREE(scratch);
1250 err = queue_pages_range(mm, start, end, nmask,
1251 flags | MPOL_MF_INVERT, &pagelist);
1253 err = mbind_range(mm, start, end, new);
1258 if (!list_empty(&pagelist)) {
1259 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1260 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1261 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1263 putback_movable_pages(&pagelist);
1266 if (nr_failed && (flags & MPOL_MF_STRICT))
1269 putback_movable_pages(&pagelist);
1271 up_write(&mm->mmap_sem);
1278 * User space interface with variable sized bitmaps for nodelists.
1281 /* Copy a node mask from user space. */
1282 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1283 unsigned long maxnode)
1287 unsigned long nlongs;
1288 unsigned long endmask;
1291 nodes_clear(*nodes);
1292 if (maxnode == 0 || !nmask)
1294 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1297 nlongs = BITS_TO_LONGS(maxnode);
1298 if ((maxnode % BITS_PER_LONG) == 0)
1301 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1304 * When the user specified more nodes than supported just check
1305 * if the non supported part is all zero.
1307 * If maxnode have more longs than MAX_NUMNODES, check
1308 * the bits in that area first. And then go through to
1309 * check the rest bits which equal or bigger than MAX_NUMNODES.
1310 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1312 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1313 if (nlongs > PAGE_SIZE/sizeof(long))
1315 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1316 if (get_user(t, nmask + k))
1318 if (k == nlongs - 1) {
1324 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1328 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1329 unsigned long valid_mask = endmask;
1331 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1332 if (get_user(t, nmask + nlongs - 1))
1338 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1340 nodes_addr(*nodes)[nlongs-1] &= endmask;
1344 /* Copy a kernel node mask to user space */
1345 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1348 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1349 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1351 if (copy > nbytes) {
1352 if (copy > PAGE_SIZE)
1354 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1358 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1361 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1362 unsigned long, mode, const unsigned long __user *, nmask,
1363 unsigned long, maxnode, unsigned, flags)
1367 unsigned short mode_flags;
1369 mode_flags = mode & MPOL_MODE_FLAGS;
1370 mode &= ~MPOL_MODE_FLAGS;
1371 if (mode >= MPOL_MAX)
1373 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1374 (mode_flags & MPOL_F_RELATIVE_NODES))
1376 err = get_nodes(&nodes, nmask, maxnode);
1379 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1382 /* Set the process memory policy */
1383 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1384 unsigned long, maxnode)
1388 unsigned short flags;
1390 flags = mode & MPOL_MODE_FLAGS;
1391 mode &= ~MPOL_MODE_FLAGS;
1392 if ((unsigned int)mode >= MPOL_MAX)
1394 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1396 err = get_nodes(&nodes, nmask, maxnode);
1399 return do_set_mempolicy(mode, flags, &nodes);
1402 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1403 const unsigned long __user *, old_nodes,
1404 const unsigned long __user *, new_nodes)
1406 const struct cred *cred = current_cred(), *tcred;
1407 struct mm_struct *mm = NULL;
1408 struct task_struct *task;
1409 nodemask_t task_nodes;
1413 NODEMASK_SCRATCH(scratch);
1418 old = &scratch->mask1;
1419 new = &scratch->mask2;
1421 err = get_nodes(old, old_nodes, maxnode);
1425 err = get_nodes(new, new_nodes, maxnode);
1429 /* Find the mm_struct */
1431 task = pid ? find_task_by_vpid(pid) : current;
1437 get_task_struct(task);
1442 * Check if this process has the right to modify the specified
1443 * process. The right exists if the process has administrative
1444 * capabilities, superuser privileges or the same
1445 * userid as the target process.
1447 tcred = __task_cred(task);
1448 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1449 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1450 !capable(CAP_SYS_NICE)) {
1457 task_nodes = cpuset_mems_allowed(task);
1458 /* Is the user allowed to access the target nodes? */
1459 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1464 task_nodes = cpuset_mems_allowed(current);
1465 nodes_and(*new, *new, task_nodes);
1466 if (nodes_empty(*new))
1469 nodes_and(*new, *new, node_states[N_MEMORY]);
1470 if (nodes_empty(*new))
1473 err = security_task_movememory(task);
1477 mm = get_task_mm(task);
1478 put_task_struct(task);
1485 err = do_migrate_pages(mm, old, new,
1486 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1490 NODEMASK_SCRATCH_FREE(scratch);
1495 put_task_struct(task);
1501 /* Retrieve NUMA policy */
1502 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1503 unsigned long __user *, nmask, unsigned long, maxnode,
1504 unsigned long, addr, unsigned long, flags)
1507 int uninitialized_var(pval);
1510 if (nmask != NULL && maxnode < nr_node_ids)
1513 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1518 if (policy && put_user(pval, policy))
1522 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1527 #ifdef CONFIG_COMPAT
1529 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1530 compat_ulong_t __user *, nmask,
1531 compat_ulong_t, maxnode,
1532 compat_ulong_t, addr, compat_ulong_t, flags)
1535 unsigned long __user *nm = NULL;
1536 unsigned long nr_bits, alloc_size;
1537 DECLARE_BITMAP(bm, MAX_NUMNODES);
1539 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1540 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1543 nm = compat_alloc_user_space(alloc_size);
1545 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1547 if (!err && nmask) {
1548 unsigned long copy_size;
1549 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1550 err = copy_from_user(bm, nm, copy_size);
1551 /* ensure entire bitmap is zeroed */
1552 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1553 err |= compat_put_bitmap(nmask, bm, nr_bits);
1559 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1560 compat_ulong_t, maxnode)
1562 unsigned long __user *nm = NULL;
1563 unsigned long nr_bits, alloc_size;
1564 DECLARE_BITMAP(bm, MAX_NUMNODES);
1566 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1567 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1570 if (compat_get_bitmap(bm, nmask, nr_bits))
1572 nm = compat_alloc_user_space(alloc_size);
1573 if (copy_to_user(nm, bm, alloc_size))
1577 return sys_set_mempolicy(mode, nm, nr_bits+1);
1580 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1581 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1582 compat_ulong_t, maxnode, compat_ulong_t, flags)
1584 unsigned long __user *nm = NULL;
1585 unsigned long nr_bits, alloc_size;
1588 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1589 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1592 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1594 nm = compat_alloc_user_space(alloc_size);
1595 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1599 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1604 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1607 struct mempolicy *pol = NULL;
1610 if (vma->vm_ops && vma->vm_ops->get_policy) {
1611 pol = vma->vm_ops->get_policy(vma, addr);
1612 } else if (vma->vm_policy) {
1613 pol = vma->vm_policy;
1616 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1617 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1618 * count on these policies which will be dropped by
1619 * mpol_cond_put() later
1621 if (mpol_needs_cond_ref(pol))
1630 * get_vma_policy(@vma, @addr)
1631 * @vma: virtual memory area whose policy is sought
1632 * @addr: address in @vma for shared policy lookup
1634 * Returns effective policy for a VMA at specified address.
1635 * Falls back to current->mempolicy or system default policy, as necessary.
1636 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1637 * count--added by the get_policy() vm_op, as appropriate--to protect against
1638 * freeing by another task. It is the caller's responsibility to free the
1639 * extra reference for shared policies.
1641 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1644 struct mempolicy *pol = __get_vma_policy(vma, addr);
1647 pol = get_task_policy(current);
1652 bool vma_policy_mof(struct vm_area_struct *vma)
1654 struct mempolicy *pol;
1656 if (vma->vm_ops && vma->vm_ops->get_policy) {
1659 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1660 if (pol && (pol->flags & MPOL_F_MOF))
1667 pol = vma->vm_policy;
1669 pol = get_task_policy(current);
1671 return pol->flags & MPOL_F_MOF;
1674 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1676 enum zone_type dynamic_policy_zone = policy_zone;
1678 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1681 * if policy->v.nodes has movable memory only,
1682 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1684 * policy->v.nodes is intersect with node_states[N_MEMORY].
1685 * so if the following test faile, it implies
1686 * policy->v.nodes has movable memory only.
1688 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1689 dynamic_policy_zone = ZONE_MOVABLE;
1691 return zone >= dynamic_policy_zone;
1695 * Return a nodemask representing a mempolicy for filtering nodes for
1698 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1700 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1701 if (unlikely(policy->mode == MPOL_BIND) &&
1702 apply_policy_zone(policy, gfp_zone(gfp)) &&
1703 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1704 return &policy->v.nodes;
1709 /* Return the node id preferred by the given mempolicy, or the given id */
1710 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1713 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1714 nd = policy->v.preferred_node;
1717 * __GFP_THISNODE shouldn't even be used with the bind policy
1718 * because we might easily break the expectation to stay on the
1719 * requested node and not break the policy.
1721 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1727 /* Do dynamic interleaving for a process */
1728 static unsigned interleave_nodes(struct mempolicy *policy)
1731 struct task_struct *me = current;
1733 next = next_node_in(me->il_prev, policy->v.nodes);
1734 if (next < MAX_NUMNODES)
1740 * Depending on the memory policy provide a node from which to allocate the
1743 unsigned int mempolicy_slab_node(void)
1745 struct mempolicy *policy;
1746 int node = numa_mem_id();
1751 policy = current->mempolicy;
1752 if (!policy || policy->flags & MPOL_F_LOCAL)
1755 switch (policy->mode) {
1756 case MPOL_PREFERRED:
1758 * handled MPOL_F_LOCAL above
1760 return policy->v.preferred_node;
1762 case MPOL_INTERLEAVE:
1763 return interleave_nodes(policy);
1769 * Follow bind policy behavior and start allocation at the
1772 struct zonelist *zonelist;
1773 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1774 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1775 z = first_zones_zonelist(zonelist, highest_zoneidx,
1777 return z->zone ? z->zone->node : node;
1786 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1787 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1788 * number of present nodes.
1790 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1792 unsigned nnodes = nodes_weight(pol->v.nodes);
1798 return numa_node_id();
1799 target = (unsigned int)n % nnodes;
1800 nid = first_node(pol->v.nodes);
1801 for (i = 0; i < target; i++)
1802 nid = next_node(nid, pol->v.nodes);
1806 /* Determine a node number for interleave */
1807 static inline unsigned interleave_nid(struct mempolicy *pol,
1808 struct vm_area_struct *vma, unsigned long addr, int shift)
1814 * for small pages, there is no difference between
1815 * shift and PAGE_SHIFT, so the bit-shift is safe.
1816 * for huge pages, since vm_pgoff is in units of small
1817 * pages, we need to shift off the always 0 bits to get
1820 BUG_ON(shift < PAGE_SHIFT);
1821 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1822 off += (addr - vma->vm_start) >> shift;
1823 return offset_il_node(pol, off);
1825 return interleave_nodes(pol);
1828 #ifdef CONFIG_HUGETLBFS
1830 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1831 * @vma: virtual memory area whose policy is sought
1832 * @addr: address in @vma for shared policy lookup and interleave policy
1833 * @gfp_flags: for requested zone
1834 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1835 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1837 * Returns a nid suitable for a huge page allocation and a pointer
1838 * to the struct mempolicy for conditional unref after allocation.
1839 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1840 * @nodemask for filtering the zonelist.
1842 * Must be protected by read_mems_allowed_begin()
1844 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1845 struct mempolicy **mpol, nodemask_t **nodemask)
1849 *mpol = get_vma_policy(vma, addr);
1850 *nodemask = NULL; /* assume !MPOL_BIND */
1852 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1853 nid = interleave_nid(*mpol, vma, addr,
1854 huge_page_shift(hstate_vma(vma)));
1856 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1857 if ((*mpol)->mode == MPOL_BIND)
1858 *nodemask = &(*mpol)->v.nodes;
1864 * init_nodemask_of_mempolicy
1866 * If the current task's mempolicy is "default" [NULL], return 'false'
1867 * to indicate default policy. Otherwise, extract the policy nodemask
1868 * for 'bind' or 'interleave' policy into the argument nodemask, or
1869 * initialize the argument nodemask to contain the single node for
1870 * 'preferred' or 'local' policy and return 'true' to indicate presence
1871 * of non-default mempolicy.
1873 * We don't bother with reference counting the mempolicy [mpol_get/put]
1874 * because the current task is examining it's own mempolicy and a task's
1875 * mempolicy is only ever changed by the task itself.
1877 * N.B., it is the caller's responsibility to free a returned nodemask.
1879 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1881 struct mempolicy *mempolicy;
1884 if (!(mask && current->mempolicy))
1888 mempolicy = current->mempolicy;
1889 switch (mempolicy->mode) {
1890 case MPOL_PREFERRED:
1891 if (mempolicy->flags & MPOL_F_LOCAL)
1892 nid = numa_node_id();
1894 nid = mempolicy->v.preferred_node;
1895 init_nodemask_of_node(mask, nid);
1900 case MPOL_INTERLEAVE:
1901 *mask = mempolicy->v.nodes;
1907 task_unlock(current);
1914 * mempolicy_nodemask_intersects
1916 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1917 * policy. Otherwise, check for intersection between mask and the policy
1918 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1919 * policy, always return true since it may allocate elsewhere on fallback.
1921 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1923 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1924 const nodemask_t *mask)
1926 struct mempolicy *mempolicy;
1932 mempolicy = tsk->mempolicy;
1936 switch (mempolicy->mode) {
1937 case MPOL_PREFERRED:
1939 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1940 * allocate from, they may fallback to other nodes when oom.
1941 * Thus, it's possible for tsk to have allocated memory from
1946 case MPOL_INTERLEAVE:
1947 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1957 /* Allocate a page in interleaved policy.
1958 Own path because it needs to do special accounting. */
1959 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1964 page = __alloc_pages(gfp, order, nid);
1965 if (page && page_to_nid(page) == nid) {
1967 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1974 * alloc_pages_vma - Allocate a page for a VMA.
1977 * %GFP_USER user allocation.
1978 * %GFP_KERNEL kernel allocations,
1979 * %GFP_HIGHMEM highmem/user allocations,
1980 * %GFP_FS allocation should not call back into a file system.
1981 * %GFP_ATOMIC don't sleep.
1983 * @order:Order of the GFP allocation.
1984 * @vma: Pointer to VMA or NULL if not available.
1985 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1986 * @node: Which node to prefer for allocation (modulo policy).
1987 * @hugepage: for hugepages try only the preferred node if possible
1989 * This function allocates a page from the kernel page pool and applies
1990 * a NUMA policy associated with the VMA or the current process.
1991 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1992 * mm_struct of the VMA to prevent it from going away. Should be used for
1993 * all allocations for pages that will be mapped into user space. Returns
1994 * NULL when no page can be allocated.
1997 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1998 unsigned long addr, int node, bool hugepage)
2000 struct mempolicy *pol;
2005 pol = get_vma_policy(vma, addr);
2007 if (pol->mode == MPOL_INTERLEAVE) {
2010 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2012 page = alloc_page_interleave(gfp, order, nid);
2016 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2017 int hpage_node = node;
2020 * For hugepage allocation and non-interleave policy which
2021 * allows the current node (or other explicitly preferred
2022 * node) we only try to allocate from the current/preferred
2023 * node and don't fall back to other nodes, as the cost of
2024 * remote accesses would likely offset THP benefits.
2026 * If the policy is interleave, or does not allow the current
2027 * node in its nodemask, we allocate the standard way.
2029 if (pol->mode == MPOL_PREFERRED &&
2030 !(pol->flags & MPOL_F_LOCAL))
2031 hpage_node = pol->v.preferred_node;
2033 nmask = policy_nodemask(gfp, pol);
2034 if (!nmask || node_isset(hpage_node, *nmask)) {
2037 * We cannot invoke reclaim if __GFP_THISNODE
2038 * is set. Invoking reclaim with
2039 * __GFP_THISNODE set, would cause THP
2040 * allocations to trigger heavy swapping
2041 * despite there may be tons of free memory
2042 * (including potentially plenty of THP
2043 * already available in the buddy) on all the
2046 * At most we could invoke compaction when
2047 * __GFP_THISNODE is set (but we would need to
2048 * refrain from invoking reclaim even if
2049 * compaction returned COMPACT_SKIPPED because
2050 * there wasn't not enough memory to succeed
2051 * compaction). For now just avoid
2052 * __GFP_THISNODE instead of limiting the
2053 * allocation path to a strict and single
2054 * compaction invocation.
2056 * Supposedly if direct reclaim was enabled by
2057 * the caller, the app prefers THP regardless
2058 * of the node it comes from so this would be
2059 * more desiderable behavior than only
2060 * providing THP originated from the local
2061 * node in such case.
2063 if (!(gfp & __GFP_DIRECT_RECLAIM))
2064 gfp |= __GFP_THISNODE;
2065 page = __alloc_pages_node(hpage_node, gfp, order);
2070 nmask = policy_nodemask(gfp, pol);
2071 preferred_nid = policy_node(gfp, pol, node);
2072 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2079 * alloc_pages_current - Allocate pages.
2082 * %GFP_USER user allocation,
2083 * %GFP_KERNEL kernel allocation,
2084 * %GFP_HIGHMEM highmem allocation,
2085 * %GFP_FS don't call back into a file system.
2086 * %GFP_ATOMIC don't sleep.
2087 * @order: Power of two of allocation size in pages. 0 is a single page.
2089 * Allocate a page from the kernel page pool. When not in
2090 * interrupt context and apply the current process NUMA policy.
2091 * Returns NULL when no page can be allocated.
2093 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2095 struct mempolicy *pol = &default_policy;
2098 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2099 pol = get_task_policy(current);
2102 * No reference counting needed for current->mempolicy
2103 * nor system default_policy
2105 if (pol->mode == MPOL_INTERLEAVE)
2106 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2108 page = __alloc_pages_nodemask(gfp, order,
2109 policy_node(gfp, pol, numa_node_id()),
2110 policy_nodemask(gfp, pol));
2114 EXPORT_SYMBOL(alloc_pages_current);
2116 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2118 struct mempolicy *pol = mpol_dup(vma_policy(src));
2121 return PTR_ERR(pol);
2122 dst->vm_policy = pol;
2127 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2128 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2129 * with the mems_allowed returned by cpuset_mems_allowed(). This
2130 * keeps mempolicies cpuset relative after its cpuset moves. See
2131 * further kernel/cpuset.c update_nodemask().
2133 * current's mempolicy may be rebinded by the other task(the task that changes
2134 * cpuset's mems), so we needn't do rebind work for current task.
2137 /* Slow path of a mempolicy duplicate */
2138 struct mempolicy *__mpol_dup(struct mempolicy *old)
2140 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2143 return ERR_PTR(-ENOMEM);
2145 /* task's mempolicy is protected by alloc_lock */
2146 if (old == current->mempolicy) {
2149 task_unlock(current);
2153 if (current_cpuset_is_being_rebound()) {
2154 nodemask_t mems = cpuset_mems_allowed(current);
2155 mpol_rebind_policy(new, &mems);
2157 atomic_set(&new->refcnt, 1);
2161 /* Slow path of a mempolicy comparison */
2162 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2166 if (a->mode != b->mode)
2168 if (a->flags != b->flags)
2170 if (mpol_store_user_nodemask(a))
2171 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2177 case MPOL_INTERLEAVE:
2178 return !!nodes_equal(a->v.nodes, b->v.nodes);
2179 case MPOL_PREFERRED:
2180 /* a's ->flags is the same as b's */
2181 if (a->flags & MPOL_F_LOCAL)
2183 return a->v.preferred_node == b->v.preferred_node;
2191 * Shared memory backing store policy support.
2193 * Remember policies even when nobody has shared memory mapped.
2194 * The policies are kept in Red-Black tree linked from the inode.
2195 * They are protected by the sp->lock rwlock, which should be held
2196 * for any accesses to the tree.
2200 * lookup first element intersecting start-end. Caller holds sp->lock for
2201 * reading or for writing
2203 static struct sp_node *
2204 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2206 struct rb_node *n = sp->root.rb_node;
2209 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2211 if (start >= p->end)
2213 else if (end <= p->start)
2221 struct sp_node *w = NULL;
2222 struct rb_node *prev = rb_prev(n);
2225 w = rb_entry(prev, struct sp_node, nd);
2226 if (w->end <= start)
2230 return rb_entry(n, struct sp_node, nd);
2234 * Insert a new shared policy into the list. Caller holds sp->lock for
2237 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2239 struct rb_node **p = &sp->root.rb_node;
2240 struct rb_node *parent = NULL;
2245 nd = rb_entry(parent, struct sp_node, nd);
2246 if (new->start < nd->start)
2248 else if (new->end > nd->end)
2249 p = &(*p)->rb_right;
2253 rb_link_node(&new->nd, parent, p);
2254 rb_insert_color(&new->nd, &sp->root);
2255 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2256 new->policy ? new->policy->mode : 0);
2259 /* Find shared policy intersecting idx */
2261 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2263 struct mempolicy *pol = NULL;
2266 if (!sp->root.rb_node)
2268 read_lock(&sp->lock);
2269 sn = sp_lookup(sp, idx, idx+1);
2271 mpol_get(sn->policy);
2274 read_unlock(&sp->lock);
2278 static void sp_free(struct sp_node *n)
2280 mpol_put(n->policy);
2281 kmem_cache_free(sn_cache, n);
2285 * mpol_misplaced - check whether current page node is valid in policy
2287 * @page: page to be checked
2288 * @vma: vm area where page mapped
2289 * @addr: virtual address where page mapped
2291 * Lookup current policy node id for vma,addr and "compare to" page's
2295 * -1 - not misplaced, page is in the right node
2296 * node - node id where the page should be
2298 * Policy determination "mimics" alloc_page_vma().
2299 * Called from fault path where we know the vma and faulting address.
2301 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2303 struct mempolicy *pol;
2305 int curnid = page_to_nid(page);
2306 unsigned long pgoff;
2307 int thiscpu = raw_smp_processor_id();
2308 int thisnid = cpu_to_node(thiscpu);
2312 pol = get_vma_policy(vma, addr);
2313 if (!(pol->flags & MPOL_F_MOF))
2316 switch (pol->mode) {
2317 case MPOL_INTERLEAVE:
2318 pgoff = vma->vm_pgoff;
2319 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2320 polnid = offset_il_node(pol, pgoff);
2323 case MPOL_PREFERRED:
2324 if (pol->flags & MPOL_F_LOCAL)
2325 polnid = numa_node_id();
2327 polnid = pol->v.preferred_node;
2333 * allows binding to multiple nodes.
2334 * use current page if in policy nodemask,
2335 * else select nearest allowed node, if any.
2336 * If no allowed nodes, use current [!misplaced].
2338 if (node_isset(curnid, pol->v.nodes))
2340 z = first_zones_zonelist(
2341 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2342 gfp_zone(GFP_HIGHUSER),
2344 polnid = z->zone->node;
2351 /* Migrate the page towards the node whose CPU is referencing it */
2352 if (pol->flags & MPOL_F_MORON) {
2355 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2359 if (curnid != polnid)
2368 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2369 * dropped after task->mempolicy is set to NULL so that any allocation done as
2370 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2373 void mpol_put_task_policy(struct task_struct *task)
2375 struct mempolicy *pol;
2378 pol = task->mempolicy;
2379 task->mempolicy = NULL;
2384 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2386 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2387 rb_erase(&n->nd, &sp->root);
2391 static void sp_node_init(struct sp_node *node, unsigned long start,
2392 unsigned long end, struct mempolicy *pol)
2394 node->start = start;
2399 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2400 struct mempolicy *pol)
2403 struct mempolicy *newpol;
2405 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2409 newpol = mpol_dup(pol);
2410 if (IS_ERR(newpol)) {
2411 kmem_cache_free(sn_cache, n);
2414 newpol->flags |= MPOL_F_SHARED;
2415 sp_node_init(n, start, end, newpol);
2420 /* Replace a policy range. */
2421 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2422 unsigned long end, struct sp_node *new)
2425 struct sp_node *n_new = NULL;
2426 struct mempolicy *mpol_new = NULL;
2430 write_lock(&sp->lock);
2431 n = sp_lookup(sp, start, end);
2432 /* Take care of old policies in the same range. */
2433 while (n && n->start < end) {
2434 struct rb_node *next = rb_next(&n->nd);
2435 if (n->start >= start) {
2441 /* Old policy spanning whole new range. */
2446 *mpol_new = *n->policy;
2447 atomic_set(&mpol_new->refcnt, 1);
2448 sp_node_init(n_new, end, n->end, mpol_new);
2450 sp_insert(sp, n_new);
2459 n = rb_entry(next, struct sp_node, nd);
2463 write_unlock(&sp->lock);
2470 kmem_cache_free(sn_cache, n_new);
2475 write_unlock(&sp->lock);
2477 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2480 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2483 atomic_set(&mpol_new->refcnt, 1);
2488 * mpol_shared_policy_init - initialize shared policy for inode
2489 * @sp: pointer to inode shared policy
2490 * @mpol: struct mempolicy to install
2492 * Install non-NULL @mpol in inode's shared policy rb-tree.
2493 * On entry, the current task has a reference on a non-NULL @mpol.
2494 * This must be released on exit.
2495 * This is called at get_inode() calls and we can use GFP_KERNEL.
2497 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2501 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2502 rwlock_init(&sp->lock);
2505 struct vm_area_struct pvma;
2506 struct mempolicy *new;
2507 NODEMASK_SCRATCH(scratch);
2511 /* contextualize the tmpfs mount point mempolicy */
2512 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2514 goto free_scratch; /* no valid nodemask intersection */
2517 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2518 task_unlock(current);
2522 /* Create pseudo-vma that contains just the policy */
2523 memset(&pvma, 0, sizeof(struct vm_area_struct));
2524 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2525 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2528 mpol_put(new); /* drop initial ref */
2530 NODEMASK_SCRATCH_FREE(scratch);
2532 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2536 int mpol_set_shared_policy(struct shared_policy *info,
2537 struct vm_area_struct *vma, struct mempolicy *npol)
2540 struct sp_node *new = NULL;
2541 unsigned long sz = vma_pages(vma);
2543 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2545 sz, npol ? npol->mode : -1,
2546 npol ? npol->flags : -1,
2547 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2550 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2554 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2560 /* Free a backing policy store on inode delete. */
2561 void mpol_free_shared_policy(struct shared_policy *p)
2564 struct rb_node *next;
2566 if (!p->root.rb_node)
2568 write_lock(&p->lock);
2569 next = rb_first(&p->root);
2571 n = rb_entry(next, struct sp_node, nd);
2572 next = rb_next(&n->nd);
2575 write_unlock(&p->lock);
2578 #ifdef CONFIG_NUMA_BALANCING
2579 static int __initdata numabalancing_override;
2581 static void __init check_numabalancing_enable(void)
2583 bool numabalancing_default = false;
2585 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2586 numabalancing_default = true;
2588 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2589 if (numabalancing_override)
2590 set_numabalancing_state(numabalancing_override == 1);
2592 if (num_online_nodes() > 1 && !numabalancing_override) {
2593 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2594 numabalancing_default ? "Enabling" : "Disabling");
2595 set_numabalancing_state(numabalancing_default);
2599 static int __init setup_numabalancing(char *str)
2605 if (!strcmp(str, "enable")) {
2606 numabalancing_override = 1;
2608 } else if (!strcmp(str, "disable")) {
2609 numabalancing_override = -1;
2614 pr_warn("Unable to parse numa_balancing=\n");
2618 __setup("numa_balancing=", setup_numabalancing);
2620 static inline void __init check_numabalancing_enable(void)
2623 #endif /* CONFIG_NUMA_BALANCING */
2625 /* assumes fs == KERNEL_DS */
2626 void __init numa_policy_init(void)
2628 nodemask_t interleave_nodes;
2629 unsigned long largest = 0;
2630 int nid, prefer = 0;
2632 policy_cache = kmem_cache_create("numa_policy",
2633 sizeof(struct mempolicy),
2634 0, SLAB_PANIC, NULL);
2636 sn_cache = kmem_cache_create("shared_policy_node",
2637 sizeof(struct sp_node),
2638 0, SLAB_PANIC, NULL);
2640 for_each_node(nid) {
2641 preferred_node_policy[nid] = (struct mempolicy) {
2642 .refcnt = ATOMIC_INIT(1),
2643 .mode = MPOL_PREFERRED,
2644 .flags = MPOL_F_MOF | MPOL_F_MORON,
2645 .v = { .preferred_node = nid, },
2650 * Set interleaving policy for system init. Interleaving is only
2651 * enabled across suitably sized nodes (default is >= 16MB), or
2652 * fall back to the largest node if they're all smaller.
2654 nodes_clear(interleave_nodes);
2655 for_each_node_state(nid, N_MEMORY) {
2656 unsigned long total_pages = node_present_pages(nid);
2658 /* Preserve the largest node */
2659 if (largest < total_pages) {
2660 largest = total_pages;
2664 /* Interleave this node? */
2665 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2666 node_set(nid, interleave_nodes);
2669 /* All too small, use the largest */
2670 if (unlikely(nodes_empty(interleave_nodes)))
2671 node_set(prefer, interleave_nodes);
2673 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2674 pr_err("%s: interleaving failed\n", __func__);
2676 check_numabalancing_enable();
2679 /* Reset policy of current process to default */
2680 void numa_default_policy(void)
2682 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2686 * Parse and format mempolicy from/to strings
2690 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2692 static const char * const policy_modes[] =
2694 [MPOL_DEFAULT] = "default",
2695 [MPOL_PREFERRED] = "prefer",
2696 [MPOL_BIND] = "bind",
2697 [MPOL_INTERLEAVE] = "interleave",
2698 [MPOL_LOCAL] = "local",
2704 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2705 * @str: string containing mempolicy to parse
2706 * @mpol: pointer to struct mempolicy pointer, returned on success.
2709 * <mode>[=<flags>][:<nodelist>]
2711 * On success, returns 0, else 1
2713 int mpol_parse_str(char *str, struct mempolicy **mpol)
2715 struct mempolicy *new = NULL;
2716 unsigned short mode;
2717 unsigned short mode_flags;
2719 char *nodelist = strchr(str, ':');
2720 char *flags = strchr(str, '=');
2724 *flags++ = '\0'; /* terminate mode string */
2727 /* NUL-terminate mode or flags string */
2729 if (nodelist_parse(nodelist, nodes))
2731 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2736 for (mode = 0; mode < MPOL_MAX; mode++) {
2737 if (!strcmp(str, policy_modes[mode])) {
2741 if (mode >= MPOL_MAX)
2745 case MPOL_PREFERRED:
2747 * Insist on a nodelist of one node only, although later
2748 * we use first_node(nodes) to grab a single node, so here
2749 * nodelist (or nodes) cannot be empty.
2752 char *rest = nodelist;
2753 while (isdigit(*rest))
2757 if (nodes_empty(nodes))
2761 case MPOL_INTERLEAVE:
2763 * Default to online nodes with memory if no nodelist
2766 nodes = node_states[N_MEMORY];
2770 * Don't allow a nodelist; mpol_new() checks flags
2774 mode = MPOL_PREFERRED;
2778 * Insist on a empty nodelist
2785 * Insist on a nodelist
2794 * Currently, we only support two mutually exclusive
2797 if (!strcmp(flags, "static"))
2798 mode_flags |= MPOL_F_STATIC_NODES;
2799 else if (!strcmp(flags, "relative"))
2800 mode_flags |= MPOL_F_RELATIVE_NODES;
2805 new = mpol_new(mode, mode_flags, &nodes);
2810 * Save nodes for mpol_to_str() to show the tmpfs mount options
2811 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2813 if (mode != MPOL_PREFERRED)
2814 new->v.nodes = nodes;
2816 new->v.preferred_node = first_node(nodes);
2818 new->flags |= MPOL_F_LOCAL;
2821 * Save nodes for contextualization: this will be used to "clone"
2822 * the mempolicy in a specific context [cpuset] at a later time.
2824 new->w.user_nodemask = nodes;
2829 /* Restore string for error message */
2838 #endif /* CONFIG_TMPFS */
2841 * mpol_to_str - format a mempolicy structure for printing
2842 * @buffer: to contain formatted mempolicy string
2843 * @maxlen: length of @buffer
2844 * @pol: pointer to mempolicy to be formatted
2846 * Convert @pol into a string. If @buffer is too short, truncate the string.
2847 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2848 * longest flag, "relative", and to display at least a few node ids.
2850 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2853 nodemask_t nodes = NODE_MASK_NONE;
2854 unsigned short mode = MPOL_DEFAULT;
2855 unsigned short flags = 0;
2857 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2865 case MPOL_PREFERRED:
2866 if (flags & MPOL_F_LOCAL)
2869 node_set(pol->v.preferred_node, nodes);
2872 case MPOL_INTERLEAVE:
2873 nodes = pol->v.nodes;
2877 snprintf(p, maxlen, "unknown");
2881 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2883 if (flags & MPOL_MODE_FLAGS) {
2884 p += snprintf(p, buffer + maxlen - p, "=");
2887 * Currently, the only defined flags are mutually exclusive
2889 if (flags & MPOL_F_STATIC_NODES)
2890 p += snprintf(p, buffer + maxlen - p, "static");
2891 else if (flags & MPOL_F_RELATIVE_NODES)
2892 p += snprintf(p, buffer + maxlen - p, "relative");
2895 if (!nodes_empty(nodes))
2896 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2897 nodemask_pr_args(&nodes));