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 isolate_huge_page(page, qp->pagelist);
587 #ifdef CONFIG_NUMA_BALANCING
589 * This is used to mark a range of virtual addresses to be inaccessible.
590 * These are later cleared by a NUMA hinting fault. Depending on these
591 * faults, pages may be migrated for better NUMA placement.
593 * This is assuming that NUMA faults are handled using PROT_NONE. If
594 * an architecture makes a different choice, it will need further
595 * changes to the core.
597 unsigned long change_prot_numa(struct vm_area_struct *vma,
598 unsigned long addr, unsigned long end)
602 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
604 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
609 static unsigned long change_prot_numa(struct vm_area_struct *vma,
610 unsigned long addr, unsigned long end)
614 #endif /* CONFIG_NUMA_BALANCING */
616 static int queue_pages_test_walk(unsigned long start, unsigned long end,
617 struct mm_walk *walk)
619 struct vm_area_struct *vma = walk->vma;
620 struct queue_pages *qp = walk->private;
621 unsigned long endvma = vma->vm_end;
622 unsigned long flags = qp->flags;
625 * Need check MPOL_MF_STRICT to return -EIO if possible
626 * regardless of vma_migratable
628 if (!vma_migratable(vma) &&
629 !(flags & MPOL_MF_STRICT))
634 if (vma->vm_start > start)
635 start = vma->vm_start;
637 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
638 if (!vma->vm_next && vma->vm_end < end)
640 if (qp->prev && qp->prev->vm_end < vma->vm_start)
646 if (flags & MPOL_MF_LAZY) {
647 /* Similar to task_numa_work, skip inaccessible VMAs */
648 if (!is_vm_hugetlb_page(vma) &&
649 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
650 !(vma->vm_flags & VM_MIXEDMAP))
651 change_prot_numa(vma, start, endvma);
655 /* queue pages from current vma */
656 if (flags & MPOL_MF_VALID)
662 * Walk through page tables and collect pages to be migrated.
664 * If pages found in a given range are on a set of nodes (determined by
665 * @nodes and @flags,) it's isolated and queued to the pagelist which is
666 * passed via @private.)
669 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
670 nodemask_t *nodes, unsigned long flags,
671 struct list_head *pagelist)
673 struct queue_pages qp = {
674 .pagelist = pagelist,
679 struct mm_walk queue_pages_walk = {
680 .hugetlb_entry = queue_pages_hugetlb,
681 .pmd_entry = queue_pages_pte_range,
682 .test_walk = queue_pages_test_walk,
687 return walk_page_range(start, end, &queue_pages_walk);
691 * Apply policy to a single VMA
692 * This must be called with the mmap_sem held for writing.
694 static int vma_replace_policy(struct vm_area_struct *vma,
695 struct mempolicy *pol)
698 struct mempolicy *old;
699 struct mempolicy *new;
701 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
702 vma->vm_start, vma->vm_end, vma->vm_pgoff,
703 vma->vm_ops, vma->vm_file,
704 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
710 if (vma->vm_ops && vma->vm_ops->set_policy) {
711 err = vma->vm_ops->set_policy(vma, new);
716 old = vma->vm_policy;
717 vma->vm_policy = new; /* protected by mmap_sem */
726 /* Step 2: apply policy to a range and do splits. */
727 static int mbind_range(struct mm_struct *mm, unsigned long start,
728 unsigned long end, struct mempolicy *new_pol)
730 struct vm_area_struct *prev;
731 struct vm_area_struct *vma;
734 unsigned long vmstart;
737 vma = find_vma(mm, start);
738 if (!vma || vma->vm_start > start)
742 if (start > vma->vm_start)
745 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
746 vmstart = max(start, vma->vm_start);
747 vmend = min(end, vma->vm_end);
749 if (mpol_equal(vma_policy(vma), new_pol))
752 pgoff = vma->vm_pgoff +
753 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
754 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
755 vma->anon_vma, vma->vm_file, pgoff,
756 new_pol, vma->vm_userfaultfd_ctx);
761 if (vma->vm_start != vmstart) {
762 err = split_vma(vma->vm_mm, vma, vmstart, 1);
766 if (vma->vm_end != vmend) {
767 err = split_vma(vma->vm_mm, vma, vmend, 0);
772 err = vma_replace_policy(vma, new_pol);
781 /* Set the process memory policy */
782 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
785 struct mempolicy *new, *old;
786 NODEMASK_SCRATCH(scratch);
792 new = mpol_new(mode, flags, nodes);
799 ret = mpol_set_nodemask(new, nodes, scratch);
801 task_unlock(current);
805 old = current->mempolicy;
806 current->mempolicy = new;
807 if (new && new->mode == MPOL_INTERLEAVE)
808 current->il_prev = MAX_NUMNODES-1;
809 task_unlock(current);
813 NODEMASK_SCRATCH_FREE(scratch);
818 * Return nodemask for policy for get_mempolicy() query
820 * Called with task's alloc_lock held
822 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
825 if (p == &default_policy)
831 case MPOL_INTERLEAVE:
835 if (!(p->flags & MPOL_F_LOCAL))
836 node_set(p->v.preferred_node, *nodes);
837 /* else return empty node mask for local allocation */
844 static int lookup_node(unsigned long addr)
849 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
851 err = page_to_nid(p);
857 /* Retrieve NUMA policy */
858 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
859 unsigned long addr, unsigned long flags)
862 struct mm_struct *mm = current->mm;
863 struct vm_area_struct *vma = NULL;
864 struct mempolicy *pol = current->mempolicy;
867 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
870 if (flags & MPOL_F_MEMS_ALLOWED) {
871 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
873 *policy = 0; /* just so it's initialized */
875 *nmask = cpuset_current_mems_allowed;
876 task_unlock(current);
880 if (flags & MPOL_F_ADDR) {
882 * Do NOT fall back to task policy if the
883 * vma/shared policy at addr is NULL. We
884 * want to return MPOL_DEFAULT in this case.
886 down_read(&mm->mmap_sem);
887 vma = find_vma_intersection(mm, addr, addr+1);
889 up_read(&mm->mmap_sem);
892 if (vma->vm_ops && vma->vm_ops->get_policy)
893 pol = vma->vm_ops->get_policy(vma, addr);
895 pol = vma->vm_policy;
900 pol = &default_policy; /* indicates default behavior */
902 if (flags & MPOL_F_NODE) {
903 if (flags & MPOL_F_ADDR) {
904 err = lookup_node(addr);
908 } else if (pol == current->mempolicy &&
909 pol->mode == MPOL_INTERLEAVE) {
910 *policy = next_node_in(current->il_prev, pol->v.nodes);
916 *policy = pol == &default_policy ? MPOL_DEFAULT :
919 * Internal mempolicy flags must be masked off before exposing
920 * the policy to userspace.
922 *policy |= (pol->flags & MPOL_MODE_FLAGS);
927 if (mpol_store_user_nodemask(pol)) {
928 *nmask = pol->w.user_nodemask;
931 get_policy_nodemask(pol, nmask);
932 task_unlock(current);
939 up_read(¤t->mm->mmap_sem);
943 #ifdef CONFIG_MIGRATION
945 * page migration, thp tail pages can be passed.
947 static void migrate_page_add(struct page *page, struct list_head *pagelist,
950 struct page *head = compound_head(page);
952 * Avoid migrating a page that is shared with others.
954 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
955 if (!isolate_lru_page(head)) {
956 list_add_tail(&head->lru, pagelist);
957 mod_node_page_state(page_pgdat(head),
958 NR_ISOLATED_ANON + page_is_file_cache(head),
959 hpage_nr_pages(head));
964 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
967 return alloc_huge_page_node(page_hstate(compound_head(page)),
969 else if (thp_migration_supported() && PageTransHuge(page)) {
972 thp = alloc_pages_node(node,
973 (GFP_TRANSHUGE | __GFP_THISNODE),
977 prep_transhuge_page(thp);
980 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
985 * Migrate pages from one node to a target node.
986 * Returns error or the number of pages not migrated.
988 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
996 node_set(source, nmask);
999 * This does not "check" the range but isolates all pages that
1000 * need migration. Between passing in the full user address
1001 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1003 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1004 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1005 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1007 if (!list_empty(&pagelist)) {
1008 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1009 MIGRATE_SYNC, MR_SYSCALL);
1011 putback_movable_pages(&pagelist);
1018 * Move pages between the two nodesets so as to preserve the physical
1019 * layout as much as possible.
1021 * Returns the number of page that could not be moved.
1023 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1024 const nodemask_t *to, int flags)
1030 err = migrate_prep();
1034 down_read(&mm->mmap_sem);
1037 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1038 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1039 * bit in 'tmp', and return that <source, dest> pair for migration.
1040 * The pair of nodemasks 'to' and 'from' define the map.
1042 * If no pair of bits is found that way, fallback to picking some
1043 * pair of 'source' and 'dest' bits that are not the same. If the
1044 * 'source' and 'dest' bits are the same, this represents a node
1045 * that will be migrating to itself, so no pages need move.
1047 * If no bits are left in 'tmp', or if all remaining bits left
1048 * in 'tmp' correspond to the same bit in 'to', return false
1049 * (nothing left to migrate).
1051 * This lets us pick a pair of nodes to migrate between, such that
1052 * if possible the dest node is not already occupied by some other
1053 * source node, minimizing the risk of overloading the memory on a
1054 * node that would happen if we migrated incoming memory to a node
1055 * before migrating outgoing memory source that same node.
1057 * A single scan of tmp is sufficient. As we go, we remember the
1058 * most recent <s, d> pair that moved (s != d). If we find a pair
1059 * that not only moved, but what's better, moved to an empty slot
1060 * (d is not set in tmp), then we break out then, with that pair.
1061 * Otherwise when we finish scanning from_tmp, we at least have the
1062 * most recent <s, d> pair that moved. If we get all the way through
1063 * the scan of tmp without finding any node that moved, much less
1064 * moved to an empty node, then there is nothing left worth migrating.
1068 while (!nodes_empty(tmp)) {
1070 int source = NUMA_NO_NODE;
1073 for_each_node_mask(s, tmp) {
1076 * do_migrate_pages() tries to maintain the relative
1077 * node relationship of the pages established between
1078 * threads and memory areas.
1080 * However if the number of source nodes is not equal to
1081 * the number of destination nodes we can not preserve
1082 * this node relative relationship. In that case, skip
1083 * copying memory from a node that is in the destination
1086 * Example: [2,3,4] -> [3,4,5] moves everything.
1087 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1090 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1091 (node_isset(s, *to)))
1094 d = node_remap(s, *from, *to);
1098 source = s; /* Node moved. Memorize */
1101 /* dest not in remaining from nodes? */
1102 if (!node_isset(dest, tmp))
1105 if (source == NUMA_NO_NODE)
1108 node_clear(source, tmp);
1109 err = migrate_to_node(mm, source, dest, flags);
1115 up_read(&mm->mmap_sem);
1123 * Allocate a new page for page migration based on vma policy.
1124 * Start by assuming the page is mapped by the same vma as contains @start.
1125 * Search forward from there, if not. N.B., this assumes that the
1126 * list of pages handed to migrate_pages()--which is how we get here--
1127 * is in virtual address order.
1129 static struct page *new_page(struct page *page, unsigned long start, int **x)
1131 struct vm_area_struct *vma;
1132 unsigned long uninitialized_var(address);
1134 vma = find_vma(current->mm, start);
1136 address = page_address_in_vma(page, vma);
1137 if (address != -EFAULT)
1142 if (PageHuge(page)) {
1144 return alloc_huge_page_noerr(vma, address, 1);
1145 } else if (thp_migration_supported() && PageTransHuge(page)) {
1148 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1152 prep_transhuge_page(thp);
1156 * if !vma, alloc_page_vma() will use task or system default policy
1158 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1163 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1164 unsigned long flags)
1168 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1169 const nodemask_t *to, int flags)
1174 static struct page *new_page(struct page *page, unsigned long start, int **x)
1180 static long do_mbind(unsigned long start, unsigned long len,
1181 unsigned short mode, unsigned short mode_flags,
1182 nodemask_t *nmask, unsigned long flags)
1184 struct mm_struct *mm = current->mm;
1185 struct mempolicy *new;
1188 LIST_HEAD(pagelist);
1190 if (flags & ~(unsigned long)MPOL_MF_VALID)
1192 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1195 if (start & ~PAGE_MASK)
1198 if (mode == MPOL_DEFAULT)
1199 flags &= ~MPOL_MF_STRICT;
1201 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1209 new = mpol_new(mode, mode_flags, nmask);
1211 return PTR_ERR(new);
1213 if (flags & MPOL_MF_LAZY)
1214 new->flags |= MPOL_F_MOF;
1217 * If we are using the default policy then operation
1218 * on discontinuous address spaces is okay after all
1221 flags |= MPOL_MF_DISCONTIG_OK;
1223 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1224 start, start + len, mode, mode_flags,
1225 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1227 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1229 err = migrate_prep();
1234 NODEMASK_SCRATCH(scratch);
1236 down_write(&mm->mmap_sem);
1238 err = mpol_set_nodemask(new, nmask, scratch);
1239 task_unlock(current);
1241 up_write(&mm->mmap_sem);
1244 NODEMASK_SCRATCH_FREE(scratch);
1249 err = queue_pages_range(mm, start, end, nmask,
1250 flags | MPOL_MF_INVERT, &pagelist);
1252 err = mbind_range(mm, start, end, new);
1257 if (!list_empty(&pagelist)) {
1258 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1259 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1260 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1262 putback_movable_pages(&pagelist);
1265 if (nr_failed && (flags & MPOL_MF_STRICT))
1268 putback_movable_pages(&pagelist);
1270 up_write(&mm->mmap_sem);
1277 * User space interface with variable sized bitmaps for nodelists.
1280 /* Copy a node mask from user space. */
1281 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1282 unsigned long maxnode)
1286 unsigned long nlongs;
1287 unsigned long endmask;
1290 nodes_clear(*nodes);
1291 if (maxnode == 0 || !nmask)
1293 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1296 nlongs = BITS_TO_LONGS(maxnode);
1297 if ((maxnode % BITS_PER_LONG) == 0)
1300 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1303 * When the user specified more nodes than supported just check
1304 * if the non supported part is all zero.
1306 * If maxnode have more longs than MAX_NUMNODES, check
1307 * the bits in that area first. And then go through to
1308 * check the rest bits which equal or bigger than MAX_NUMNODES.
1309 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1311 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1312 if (nlongs > PAGE_SIZE/sizeof(long))
1314 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1315 if (get_user(t, nmask + k))
1317 if (k == nlongs - 1) {
1323 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1327 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1328 unsigned long valid_mask = endmask;
1330 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1331 if (get_user(t, nmask + nlongs - 1))
1337 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1339 nodes_addr(*nodes)[nlongs-1] &= endmask;
1343 /* Copy a kernel node mask to user space */
1344 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1347 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1348 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1350 if (copy > nbytes) {
1351 if (copy > PAGE_SIZE)
1353 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1357 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1360 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1361 unsigned long, mode, const unsigned long __user *, nmask,
1362 unsigned long, maxnode, unsigned, flags)
1366 unsigned short mode_flags;
1368 mode_flags = mode & MPOL_MODE_FLAGS;
1369 mode &= ~MPOL_MODE_FLAGS;
1370 if (mode >= MPOL_MAX)
1372 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1373 (mode_flags & MPOL_F_RELATIVE_NODES))
1375 err = get_nodes(&nodes, nmask, maxnode);
1378 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1381 /* Set the process memory policy */
1382 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1383 unsigned long, maxnode)
1387 unsigned short flags;
1389 flags = mode & MPOL_MODE_FLAGS;
1390 mode &= ~MPOL_MODE_FLAGS;
1391 if ((unsigned int)mode >= MPOL_MAX)
1393 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1395 err = get_nodes(&nodes, nmask, maxnode);
1398 return do_set_mempolicy(mode, flags, &nodes);
1401 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1402 const unsigned long __user *, old_nodes,
1403 const unsigned long __user *, new_nodes)
1405 const struct cred *cred = current_cred(), *tcred;
1406 struct mm_struct *mm = NULL;
1407 struct task_struct *task;
1408 nodemask_t task_nodes;
1412 NODEMASK_SCRATCH(scratch);
1417 old = &scratch->mask1;
1418 new = &scratch->mask2;
1420 err = get_nodes(old, old_nodes, maxnode);
1424 err = get_nodes(new, new_nodes, maxnode);
1428 /* Find the mm_struct */
1430 task = pid ? find_task_by_vpid(pid) : current;
1436 get_task_struct(task);
1441 * Check if this process has the right to modify the specified
1442 * process. The right exists if the process has administrative
1443 * capabilities, superuser privileges or the same
1444 * userid as the target process.
1446 tcred = __task_cred(task);
1447 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1448 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1449 !capable(CAP_SYS_NICE)) {
1456 task_nodes = cpuset_mems_allowed(task);
1457 /* Is the user allowed to access the target nodes? */
1458 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1463 task_nodes = cpuset_mems_allowed(current);
1464 nodes_and(*new, *new, task_nodes);
1465 if (nodes_empty(*new))
1468 nodes_and(*new, *new, node_states[N_MEMORY]);
1469 if (nodes_empty(*new))
1472 err = security_task_movememory(task);
1476 mm = get_task_mm(task);
1477 put_task_struct(task);
1484 err = do_migrate_pages(mm, old, new,
1485 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1489 NODEMASK_SCRATCH_FREE(scratch);
1494 put_task_struct(task);
1500 /* Retrieve NUMA policy */
1501 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1502 unsigned long __user *, nmask, unsigned long, maxnode,
1503 unsigned long, addr, unsigned long, flags)
1506 int uninitialized_var(pval);
1509 if (nmask != NULL && maxnode < nr_node_ids)
1512 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1517 if (policy && put_user(pval, policy))
1521 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1526 #ifdef CONFIG_COMPAT
1528 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1529 compat_ulong_t __user *, nmask,
1530 compat_ulong_t, maxnode,
1531 compat_ulong_t, addr, compat_ulong_t, flags)
1534 unsigned long __user *nm = NULL;
1535 unsigned long nr_bits, alloc_size;
1536 DECLARE_BITMAP(bm, MAX_NUMNODES);
1538 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1539 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1542 nm = compat_alloc_user_space(alloc_size);
1544 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1546 if (!err && nmask) {
1547 unsigned long copy_size;
1548 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1549 err = copy_from_user(bm, nm, copy_size);
1550 /* ensure entire bitmap is zeroed */
1551 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1552 err |= compat_put_bitmap(nmask, bm, nr_bits);
1558 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1559 compat_ulong_t, maxnode)
1561 unsigned long __user *nm = NULL;
1562 unsigned long nr_bits, alloc_size;
1563 DECLARE_BITMAP(bm, MAX_NUMNODES);
1565 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1566 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1569 if (compat_get_bitmap(bm, nmask, nr_bits))
1571 nm = compat_alloc_user_space(alloc_size);
1572 if (copy_to_user(nm, bm, alloc_size))
1576 return sys_set_mempolicy(mode, nm, nr_bits+1);
1579 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1580 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1581 compat_ulong_t, maxnode, compat_ulong_t, flags)
1583 unsigned long __user *nm = NULL;
1584 unsigned long nr_bits, alloc_size;
1587 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1588 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1591 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1593 nm = compat_alloc_user_space(alloc_size);
1594 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1598 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1603 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1606 struct mempolicy *pol = NULL;
1609 if (vma->vm_ops && vma->vm_ops->get_policy) {
1610 pol = vma->vm_ops->get_policy(vma, addr);
1611 } else if (vma->vm_policy) {
1612 pol = vma->vm_policy;
1615 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1616 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1617 * count on these policies which will be dropped by
1618 * mpol_cond_put() later
1620 if (mpol_needs_cond_ref(pol))
1629 * get_vma_policy(@vma, @addr)
1630 * @vma: virtual memory area whose policy is sought
1631 * @addr: address in @vma for shared policy lookup
1633 * Returns effective policy for a VMA at specified address.
1634 * Falls back to current->mempolicy or system default policy, as necessary.
1635 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1636 * count--added by the get_policy() vm_op, as appropriate--to protect against
1637 * freeing by another task. It is the caller's responsibility to free the
1638 * extra reference for shared policies.
1640 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1643 struct mempolicy *pol = __get_vma_policy(vma, addr);
1646 pol = get_task_policy(current);
1651 bool vma_policy_mof(struct vm_area_struct *vma)
1653 struct mempolicy *pol;
1655 if (vma->vm_ops && vma->vm_ops->get_policy) {
1658 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1659 if (pol && (pol->flags & MPOL_F_MOF))
1666 pol = vma->vm_policy;
1668 pol = get_task_policy(current);
1670 return pol->flags & MPOL_F_MOF;
1673 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1675 enum zone_type dynamic_policy_zone = policy_zone;
1677 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1680 * if policy->v.nodes has movable memory only,
1681 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1683 * policy->v.nodes is intersect with node_states[N_MEMORY].
1684 * so if the following test faile, it implies
1685 * policy->v.nodes has movable memory only.
1687 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1688 dynamic_policy_zone = ZONE_MOVABLE;
1690 return zone >= dynamic_policy_zone;
1694 * Return a nodemask representing a mempolicy for filtering nodes for
1697 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1699 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1700 if (unlikely(policy->mode == MPOL_BIND) &&
1701 apply_policy_zone(policy, gfp_zone(gfp)) &&
1702 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1703 return &policy->v.nodes;
1708 /* Return the node id preferred by the given mempolicy, or the given id */
1709 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1712 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1713 nd = policy->v.preferred_node;
1716 * __GFP_THISNODE shouldn't even be used with the bind policy
1717 * because we might easily break the expectation to stay on the
1718 * requested node and not break the policy.
1720 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1726 /* Do dynamic interleaving for a process */
1727 static unsigned interleave_nodes(struct mempolicy *policy)
1730 struct task_struct *me = current;
1732 next = next_node_in(me->il_prev, policy->v.nodes);
1733 if (next < MAX_NUMNODES)
1739 * Depending on the memory policy provide a node from which to allocate the
1742 unsigned int mempolicy_slab_node(void)
1744 struct mempolicy *policy;
1745 int node = numa_mem_id();
1750 policy = current->mempolicy;
1751 if (!policy || policy->flags & MPOL_F_LOCAL)
1754 switch (policy->mode) {
1755 case MPOL_PREFERRED:
1757 * handled MPOL_F_LOCAL above
1759 return policy->v.preferred_node;
1761 case MPOL_INTERLEAVE:
1762 return interleave_nodes(policy);
1768 * Follow bind policy behavior and start allocation at the
1771 struct zonelist *zonelist;
1772 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1773 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1774 z = first_zones_zonelist(zonelist, highest_zoneidx,
1776 return z->zone ? z->zone->node : node;
1785 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1786 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1787 * number of present nodes.
1789 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1791 unsigned nnodes = nodes_weight(pol->v.nodes);
1797 return numa_node_id();
1798 target = (unsigned int)n % nnodes;
1799 nid = first_node(pol->v.nodes);
1800 for (i = 0; i < target; i++)
1801 nid = next_node(nid, pol->v.nodes);
1805 /* Determine a node number for interleave */
1806 static inline unsigned interleave_nid(struct mempolicy *pol,
1807 struct vm_area_struct *vma, unsigned long addr, int shift)
1813 * for small pages, there is no difference between
1814 * shift and PAGE_SHIFT, so the bit-shift is safe.
1815 * for huge pages, since vm_pgoff is in units of small
1816 * pages, we need to shift off the always 0 bits to get
1819 BUG_ON(shift < PAGE_SHIFT);
1820 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1821 off += (addr - vma->vm_start) >> shift;
1822 return offset_il_node(pol, off);
1824 return interleave_nodes(pol);
1827 #ifdef CONFIG_HUGETLBFS
1829 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1830 * @vma: virtual memory area whose policy is sought
1831 * @addr: address in @vma for shared policy lookup and interleave policy
1832 * @gfp_flags: for requested zone
1833 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1834 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1836 * Returns a nid suitable for a huge page allocation and a pointer
1837 * to the struct mempolicy for conditional unref after allocation.
1838 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1839 * @nodemask for filtering the zonelist.
1841 * Must be protected by read_mems_allowed_begin()
1843 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1844 struct mempolicy **mpol, nodemask_t **nodemask)
1848 *mpol = get_vma_policy(vma, addr);
1849 *nodemask = NULL; /* assume !MPOL_BIND */
1851 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1852 nid = interleave_nid(*mpol, vma, addr,
1853 huge_page_shift(hstate_vma(vma)));
1855 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1856 if ((*mpol)->mode == MPOL_BIND)
1857 *nodemask = &(*mpol)->v.nodes;
1863 * init_nodemask_of_mempolicy
1865 * If the current task's mempolicy is "default" [NULL], return 'false'
1866 * to indicate default policy. Otherwise, extract the policy nodemask
1867 * for 'bind' or 'interleave' policy into the argument nodemask, or
1868 * initialize the argument nodemask to contain the single node for
1869 * 'preferred' or 'local' policy and return 'true' to indicate presence
1870 * of non-default mempolicy.
1872 * We don't bother with reference counting the mempolicy [mpol_get/put]
1873 * because the current task is examining it's own mempolicy and a task's
1874 * mempolicy is only ever changed by the task itself.
1876 * N.B., it is the caller's responsibility to free a returned nodemask.
1878 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1880 struct mempolicy *mempolicy;
1883 if (!(mask && current->mempolicy))
1887 mempolicy = current->mempolicy;
1888 switch (mempolicy->mode) {
1889 case MPOL_PREFERRED:
1890 if (mempolicy->flags & MPOL_F_LOCAL)
1891 nid = numa_node_id();
1893 nid = mempolicy->v.preferred_node;
1894 init_nodemask_of_node(mask, nid);
1899 case MPOL_INTERLEAVE:
1900 *mask = mempolicy->v.nodes;
1906 task_unlock(current);
1913 * mempolicy_nodemask_intersects
1915 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1916 * policy. Otherwise, check for intersection between mask and the policy
1917 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1918 * policy, always return true since it may allocate elsewhere on fallback.
1920 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1922 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1923 const nodemask_t *mask)
1925 struct mempolicy *mempolicy;
1931 mempolicy = tsk->mempolicy;
1935 switch (mempolicy->mode) {
1936 case MPOL_PREFERRED:
1938 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1939 * allocate from, they may fallback to other nodes when oom.
1940 * Thus, it's possible for tsk to have allocated memory from
1945 case MPOL_INTERLEAVE:
1946 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1956 /* Allocate a page in interleaved policy.
1957 Own path because it needs to do special accounting. */
1958 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1963 page = __alloc_pages(gfp, order, nid);
1964 if (page && page_to_nid(page) == nid) {
1966 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1973 * alloc_pages_vma - Allocate a page for a VMA.
1976 * %GFP_USER user allocation.
1977 * %GFP_KERNEL kernel allocations,
1978 * %GFP_HIGHMEM highmem/user allocations,
1979 * %GFP_FS allocation should not call back into a file system.
1980 * %GFP_ATOMIC don't sleep.
1982 * @order:Order of the GFP allocation.
1983 * @vma: Pointer to VMA or NULL if not available.
1984 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1985 * @node: Which node to prefer for allocation (modulo policy).
1986 * @hugepage: for hugepages try only the preferred node if possible
1988 * This function allocates a page from the kernel page pool and applies
1989 * a NUMA policy associated with the VMA or the current process.
1990 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1991 * mm_struct of the VMA to prevent it from going away. Should be used for
1992 * all allocations for pages that will be mapped into user space. Returns
1993 * NULL when no page can be allocated.
1996 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1997 unsigned long addr, int node, bool hugepage)
1999 struct mempolicy *pol;
2004 pol = get_vma_policy(vma, addr);
2006 if (pol->mode == MPOL_INTERLEAVE) {
2009 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2011 page = alloc_page_interleave(gfp, order, nid);
2015 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2016 int hpage_node = node;
2019 * For hugepage allocation and non-interleave policy which
2020 * allows the current node (or other explicitly preferred
2021 * node) we only try to allocate from the current/preferred
2022 * node and don't fall back to other nodes, as the cost of
2023 * remote accesses would likely offset THP benefits.
2025 * If the policy is interleave, or does not allow the current
2026 * node in its nodemask, we allocate the standard way.
2028 if (pol->mode == MPOL_PREFERRED &&
2029 !(pol->flags & MPOL_F_LOCAL))
2030 hpage_node = pol->v.preferred_node;
2032 nmask = policy_nodemask(gfp, pol);
2033 if (!nmask || node_isset(hpage_node, *nmask)) {
2036 * We cannot invoke reclaim if __GFP_THISNODE
2037 * is set. Invoking reclaim with
2038 * __GFP_THISNODE set, would cause THP
2039 * allocations to trigger heavy swapping
2040 * despite there may be tons of free memory
2041 * (including potentially plenty of THP
2042 * already available in the buddy) on all the
2045 * At most we could invoke compaction when
2046 * __GFP_THISNODE is set (but we would need to
2047 * refrain from invoking reclaim even if
2048 * compaction returned COMPACT_SKIPPED because
2049 * there wasn't not enough memory to succeed
2050 * compaction). For now just avoid
2051 * __GFP_THISNODE instead of limiting the
2052 * allocation path to a strict and single
2053 * compaction invocation.
2055 * Supposedly if direct reclaim was enabled by
2056 * the caller, the app prefers THP regardless
2057 * of the node it comes from so this would be
2058 * more desiderable behavior than only
2059 * providing THP originated from the local
2060 * node in such case.
2062 if (!(gfp & __GFP_DIRECT_RECLAIM))
2063 gfp |= __GFP_THISNODE;
2064 page = __alloc_pages_node(hpage_node, gfp, order);
2069 nmask = policy_nodemask(gfp, pol);
2070 preferred_nid = policy_node(gfp, pol, node);
2071 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2078 * alloc_pages_current - Allocate pages.
2081 * %GFP_USER user allocation,
2082 * %GFP_KERNEL kernel allocation,
2083 * %GFP_HIGHMEM highmem allocation,
2084 * %GFP_FS don't call back into a file system.
2085 * %GFP_ATOMIC don't sleep.
2086 * @order: Power of two of allocation size in pages. 0 is a single page.
2088 * Allocate a page from the kernel page pool. When not in
2089 * interrupt context and apply the current process NUMA policy.
2090 * Returns NULL when no page can be allocated.
2092 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2094 struct mempolicy *pol = &default_policy;
2097 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2098 pol = get_task_policy(current);
2101 * No reference counting needed for current->mempolicy
2102 * nor system default_policy
2104 if (pol->mode == MPOL_INTERLEAVE)
2105 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2107 page = __alloc_pages_nodemask(gfp, order,
2108 policy_node(gfp, pol, numa_node_id()),
2109 policy_nodemask(gfp, pol));
2113 EXPORT_SYMBOL(alloc_pages_current);
2115 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2117 struct mempolicy *pol = mpol_dup(vma_policy(src));
2120 return PTR_ERR(pol);
2121 dst->vm_policy = pol;
2126 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2127 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2128 * with the mems_allowed returned by cpuset_mems_allowed(). This
2129 * keeps mempolicies cpuset relative after its cpuset moves. See
2130 * further kernel/cpuset.c update_nodemask().
2132 * current's mempolicy may be rebinded by the other task(the task that changes
2133 * cpuset's mems), so we needn't do rebind work for current task.
2136 /* Slow path of a mempolicy duplicate */
2137 struct mempolicy *__mpol_dup(struct mempolicy *old)
2139 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2142 return ERR_PTR(-ENOMEM);
2144 /* task's mempolicy is protected by alloc_lock */
2145 if (old == current->mempolicy) {
2148 task_unlock(current);
2152 if (current_cpuset_is_being_rebound()) {
2153 nodemask_t mems = cpuset_mems_allowed(current);
2154 mpol_rebind_policy(new, &mems);
2156 atomic_set(&new->refcnt, 1);
2160 /* Slow path of a mempolicy comparison */
2161 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2165 if (a->mode != b->mode)
2167 if (a->flags != b->flags)
2169 if (mpol_store_user_nodemask(a))
2170 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2176 case MPOL_INTERLEAVE:
2177 return !!nodes_equal(a->v.nodes, b->v.nodes);
2178 case MPOL_PREFERRED:
2179 /* a's ->flags is the same as b's */
2180 if (a->flags & MPOL_F_LOCAL)
2182 return a->v.preferred_node == b->v.preferred_node;
2190 * Shared memory backing store policy support.
2192 * Remember policies even when nobody has shared memory mapped.
2193 * The policies are kept in Red-Black tree linked from the inode.
2194 * They are protected by the sp->lock rwlock, which should be held
2195 * for any accesses to the tree.
2199 * lookup first element intersecting start-end. Caller holds sp->lock for
2200 * reading or for writing
2202 static struct sp_node *
2203 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2205 struct rb_node *n = sp->root.rb_node;
2208 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2210 if (start >= p->end)
2212 else if (end <= p->start)
2220 struct sp_node *w = NULL;
2221 struct rb_node *prev = rb_prev(n);
2224 w = rb_entry(prev, struct sp_node, nd);
2225 if (w->end <= start)
2229 return rb_entry(n, struct sp_node, nd);
2233 * Insert a new shared policy into the list. Caller holds sp->lock for
2236 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2238 struct rb_node **p = &sp->root.rb_node;
2239 struct rb_node *parent = NULL;
2244 nd = rb_entry(parent, struct sp_node, nd);
2245 if (new->start < nd->start)
2247 else if (new->end > nd->end)
2248 p = &(*p)->rb_right;
2252 rb_link_node(&new->nd, parent, p);
2253 rb_insert_color(&new->nd, &sp->root);
2254 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2255 new->policy ? new->policy->mode : 0);
2258 /* Find shared policy intersecting idx */
2260 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2262 struct mempolicy *pol = NULL;
2265 if (!sp->root.rb_node)
2267 read_lock(&sp->lock);
2268 sn = sp_lookup(sp, idx, idx+1);
2270 mpol_get(sn->policy);
2273 read_unlock(&sp->lock);
2277 static void sp_free(struct sp_node *n)
2279 mpol_put(n->policy);
2280 kmem_cache_free(sn_cache, n);
2284 * mpol_misplaced - check whether current page node is valid in policy
2286 * @page: page to be checked
2287 * @vma: vm area where page mapped
2288 * @addr: virtual address where page mapped
2290 * Lookup current policy node id for vma,addr and "compare to" page's
2294 * -1 - not misplaced, page is in the right node
2295 * node - node id where the page should be
2297 * Policy determination "mimics" alloc_page_vma().
2298 * Called from fault path where we know the vma and faulting address.
2300 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2302 struct mempolicy *pol;
2304 int curnid = page_to_nid(page);
2305 unsigned long pgoff;
2306 int thiscpu = raw_smp_processor_id();
2307 int thisnid = cpu_to_node(thiscpu);
2311 pol = get_vma_policy(vma, addr);
2312 if (!(pol->flags & MPOL_F_MOF))
2315 switch (pol->mode) {
2316 case MPOL_INTERLEAVE:
2317 pgoff = vma->vm_pgoff;
2318 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2319 polnid = offset_il_node(pol, pgoff);
2322 case MPOL_PREFERRED:
2323 if (pol->flags & MPOL_F_LOCAL)
2324 polnid = numa_node_id();
2326 polnid = pol->v.preferred_node;
2332 * allows binding to multiple nodes.
2333 * use current page if in policy nodemask,
2334 * else select nearest allowed node, if any.
2335 * If no allowed nodes, use current [!misplaced].
2337 if (node_isset(curnid, pol->v.nodes))
2339 z = first_zones_zonelist(
2340 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2341 gfp_zone(GFP_HIGHUSER),
2343 polnid = z->zone->node;
2350 /* Migrate the page towards the node whose CPU is referencing it */
2351 if (pol->flags & MPOL_F_MORON) {
2354 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2358 if (curnid != polnid)
2367 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2368 * dropped after task->mempolicy is set to NULL so that any allocation done as
2369 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2372 void mpol_put_task_policy(struct task_struct *task)
2374 struct mempolicy *pol;
2377 pol = task->mempolicy;
2378 task->mempolicy = NULL;
2383 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2385 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2386 rb_erase(&n->nd, &sp->root);
2390 static void sp_node_init(struct sp_node *node, unsigned long start,
2391 unsigned long end, struct mempolicy *pol)
2393 node->start = start;
2398 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2399 struct mempolicy *pol)
2402 struct mempolicy *newpol;
2404 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2408 newpol = mpol_dup(pol);
2409 if (IS_ERR(newpol)) {
2410 kmem_cache_free(sn_cache, n);
2413 newpol->flags |= MPOL_F_SHARED;
2414 sp_node_init(n, start, end, newpol);
2419 /* Replace a policy range. */
2420 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2421 unsigned long end, struct sp_node *new)
2424 struct sp_node *n_new = NULL;
2425 struct mempolicy *mpol_new = NULL;
2429 write_lock(&sp->lock);
2430 n = sp_lookup(sp, start, end);
2431 /* Take care of old policies in the same range. */
2432 while (n && n->start < end) {
2433 struct rb_node *next = rb_next(&n->nd);
2434 if (n->start >= start) {
2440 /* Old policy spanning whole new range. */
2445 *mpol_new = *n->policy;
2446 atomic_set(&mpol_new->refcnt, 1);
2447 sp_node_init(n_new, end, n->end, mpol_new);
2449 sp_insert(sp, n_new);
2458 n = rb_entry(next, struct sp_node, nd);
2462 write_unlock(&sp->lock);
2469 kmem_cache_free(sn_cache, n_new);
2474 write_unlock(&sp->lock);
2476 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2479 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2482 atomic_set(&mpol_new->refcnt, 1);
2487 * mpol_shared_policy_init - initialize shared policy for inode
2488 * @sp: pointer to inode shared policy
2489 * @mpol: struct mempolicy to install
2491 * Install non-NULL @mpol in inode's shared policy rb-tree.
2492 * On entry, the current task has a reference on a non-NULL @mpol.
2493 * This must be released on exit.
2494 * This is called at get_inode() calls and we can use GFP_KERNEL.
2496 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2500 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2501 rwlock_init(&sp->lock);
2504 struct vm_area_struct pvma;
2505 struct mempolicy *new;
2506 NODEMASK_SCRATCH(scratch);
2510 /* contextualize the tmpfs mount point mempolicy */
2511 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2513 goto free_scratch; /* no valid nodemask intersection */
2516 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2517 task_unlock(current);
2521 /* Create pseudo-vma that contains just the policy */
2522 memset(&pvma, 0, sizeof(struct vm_area_struct));
2523 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2524 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2527 mpol_put(new); /* drop initial ref */
2529 NODEMASK_SCRATCH_FREE(scratch);
2531 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2535 int mpol_set_shared_policy(struct shared_policy *info,
2536 struct vm_area_struct *vma, struct mempolicy *npol)
2539 struct sp_node *new = NULL;
2540 unsigned long sz = vma_pages(vma);
2542 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2544 sz, npol ? npol->mode : -1,
2545 npol ? npol->flags : -1,
2546 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2549 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2553 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2559 /* Free a backing policy store on inode delete. */
2560 void mpol_free_shared_policy(struct shared_policy *p)
2563 struct rb_node *next;
2565 if (!p->root.rb_node)
2567 write_lock(&p->lock);
2568 next = rb_first(&p->root);
2570 n = rb_entry(next, struct sp_node, nd);
2571 next = rb_next(&n->nd);
2574 write_unlock(&p->lock);
2577 #ifdef CONFIG_NUMA_BALANCING
2578 static int __initdata numabalancing_override;
2580 static void __init check_numabalancing_enable(void)
2582 bool numabalancing_default = false;
2584 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2585 numabalancing_default = true;
2587 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2588 if (numabalancing_override)
2589 set_numabalancing_state(numabalancing_override == 1);
2591 if (num_online_nodes() > 1 && !numabalancing_override) {
2592 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2593 numabalancing_default ? "Enabling" : "Disabling");
2594 set_numabalancing_state(numabalancing_default);
2598 static int __init setup_numabalancing(char *str)
2604 if (!strcmp(str, "enable")) {
2605 numabalancing_override = 1;
2607 } else if (!strcmp(str, "disable")) {
2608 numabalancing_override = -1;
2613 pr_warn("Unable to parse numa_balancing=\n");
2617 __setup("numa_balancing=", setup_numabalancing);
2619 static inline void __init check_numabalancing_enable(void)
2622 #endif /* CONFIG_NUMA_BALANCING */
2624 /* assumes fs == KERNEL_DS */
2625 void __init numa_policy_init(void)
2627 nodemask_t interleave_nodes;
2628 unsigned long largest = 0;
2629 int nid, prefer = 0;
2631 policy_cache = kmem_cache_create("numa_policy",
2632 sizeof(struct mempolicy),
2633 0, SLAB_PANIC, NULL);
2635 sn_cache = kmem_cache_create("shared_policy_node",
2636 sizeof(struct sp_node),
2637 0, SLAB_PANIC, NULL);
2639 for_each_node(nid) {
2640 preferred_node_policy[nid] = (struct mempolicy) {
2641 .refcnt = ATOMIC_INIT(1),
2642 .mode = MPOL_PREFERRED,
2643 .flags = MPOL_F_MOF | MPOL_F_MORON,
2644 .v = { .preferred_node = nid, },
2649 * Set interleaving policy for system init. Interleaving is only
2650 * enabled across suitably sized nodes (default is >= 16MB), or
2651 * fall back to the largest node if they're all smaller.
2653 nodes_clear(interleave_nodes);
2654 for_each_node_state(nid, N_MEMORY) {
2655 unsigned long total_pages = node_present_pages(nid);
2657 /* Preserve the largest node */
2658 if (largest < total_pages) {
2659 largest = total_pages;
2663 /* Interleave this node? */
2664 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2665 node_set(nid, interleave_nodes);
2668 /* All too small, use the largest */
2669 if (unlikely(nodes_empty(interleave_nodes)))
2670 node_set(prefer, interleave_nodes);
2672 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2673 pr_err("%s: interleaving failed\n", __func__);
2675 check_numabalancing_enable();
2678 /* Reset policy of current process to default */
2679 void numa_default_policy(void)
2681 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2685 * Parse and format mempolicy from/to strings
2689 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2691 static const char * const policy_modes[] =
2693 [MPOL_DEFAULT] = "default",
2694 [MPOL_PREFERRED] = "prefer",
2695 [MPOL_BIND] = "bind",
2696 [MPOL_INTERLEAVE] = "interleave",
2697 [MPOL_LOCAL] = "local",
2703 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2704 * @str: string containing mempolicy to parse
2705 * @mpol: pointer to struct mempolicy pointer, returned on success.
2708 * <mode>[=<flags>][:<nodelist>]
2710 * On success, returns 0, else 1
2712 int mpol_parse_str(char *str, struct mempolicy **mpol)
2714 struct mempolicy *new = NULL;
2715 unsigned short mode;
2716 unsigned short mode_flags;
2718 char *nodelist = strchr(str, ':');
2719 char *flags = strchr(str, '=');
2723 *flags++ = '\0'; /* terminate mode string */
2726 /* NUL-terminate mode or flags string */
2728 if (nodelist_parse(nodelist, nodes))
2730 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2735 for (mode = 0; mode < MPOL_MAX; mode++) {
2736 if (!strcmp(str, policy_modes[mode])) {
2740 if (mode >= MPOL_MAX)
2744 case MPOL_PREFERRED:
2746 * Insist on a nodelist of one node only, although later
2747 * we use first_node(nodes) to grab a single node, so here
2748 * nodelist (or nodes) cannot be empty.
2751 char *rest = nodelist;
2752 while (isdigit(*rest))
2756 if (nodes_empty(nodes))
2760 case MPOL_INTERLEAVE:
2762 * Default to online nodes with memory if no nodelist
2765 nodes = node_states[N_MEMORY];
2769 * Don't allow a nodelist; mpol_new() checks flags
2773 mode = MPOL_PREFERRED;
2777 * Insist on a empty nodelist
2784 * Insist on a nodelist
2793 * Currently, we only support two mutually exclusive
2796 if (!strcmp(flags, "static"))
2797 mode_flags |= MPOL_F_STATIC_NODES;
2798 else if (!strcmp(flags, "relative"))
2799 mode_flags |= MPOL_F_RELATIVE_NODES;
2804 new = mpol_new(mode, mode_flags, &nodes);
2809 * Save nodes for mpol_to_str() to show the tmpfs mount options
2810 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2812 if (mode != MPOL_PREFERRED)
2813 new->v.nodes = nodes;
2815 new->v.preferred_node = first_node(nodes);
2817 new->flags |= MPOL_F_LOCAL;
2820 * Save nodes for contextualization: this will be used to "clone"
2821 * the mempolicy in a specific context [cpuset] at a later time.
2823 new->w.user_nodemask = nodes;
2828 /* Restore string for error message */
2837 #endif /* CONFIG_TMPFS */
2840 * mpol_to_str - format a mempolicy structure for printing
2841 * @buffer: to contain formatted mempolicy string
2842 * @maxlen: length of @buffer
2843 * @pol: pointer to mempolicy to be formatted
2845 * Convert @pol into a string. If @buffer is too short, truncate the string.
2846 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2847 * longest flag, "relative", and to display at least a few node ids.
2849 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2852 nodemask_t nodes = NODE_MASK_NONE;
2853 unsigned short mode = MPOL_DEFAULT;
2854 unsigned short flags = 0;
2856 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2864 case MPOL_PREFERRED:
2865 if (flags & MPOL_F_LOCAL)
2868 node_set(pol->v.preferred_node, nodes);
2871 case MPOL_INTERLEAVE:
2872 nodes = pol->v.nodes;
2876 snprintf(p, maxlen, "unknown");
2880 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2882 if (flags & MPOL_MODE_FLAGS) {
2883 p += snprintf(p, buffer + maxlen - p, "=");
2886 * Currently, the only defined flags are mutually exclusive
2888 if (flags & MPOL_F_STATIC_NODES)
2889 p += snprintf(p, buffer + maxlen - p, "static");
2890 else if (flags & MPOL_F_RELATIVE_NODES)
2891 p += snprintf(p, buffer + maxlen - p, "relative");
2894 if (!nodes_empty(nodes))
2895 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2896 nodemask_pr_args(&nodes));