1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple NUMA memory policy for the Linux kernel.
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/sched/mm.h>
77 #include <linux/sched/numa_balancing.h>
78 #include <linux/sched/task.h>
79 #include <linux/nodemask.h>
80 #include <linux/cpuset.h>
81 #include <linux/slab.h>
82 #include <linux/string.h>
83 #include <linux/export.h>
84 #include <linux/nsproxy.h>
85 #include <linux/interrupt.h>
86 #include <linux/init.h>
87 #include <linux/compat.h>
88 #include <linux/ptrace.h>
89 #include <linux/swap.h>
90 #include <linux/seq_file.h>
91 #include <linux/proc_fs.h>
92 #include <linux/migrate.h>
93 #include <linux/ksm.h>
94 #include <linux/rmap.h>
95 #include <linux/security.h>
96 #include <linux/syscalls.h>
97 #include <linux/ctype.h>
98 #include <linux/mm_inline.h>
99 #include <linux/mmu_notifier.h>
100 #include <linux/printk.h>
101 #include <linux/swapops.h>
103 #include <asm/tlbflush.h>
104 #include <linux/uaccess.h>
106 #include "internal.h"
109 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
110 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
112 static struct kmem_cache *policy_cache;
113 static struct kmem_cache *sn_cache;
115 /* Highest zone. An specific allocation for a zone below that is not
117 enum zone_type policy_zone = 0;
120 * run-time system-wide default policy => local allocation
122 static struct mempolicy default_policy = {
123 .refcnt = ATOMIC_INIT(1), /* never free it */
124 .mode = MPOL_PREFERRED,
125 .flags = MPOL_F_LOCAL,
128 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
130 struct mempolicy *get_task_policy(struct task_struct *p)
132 struct mempolicy *pol = p->mempolicy;
138 node = numa_node_id();
139 if (node != NUMA_NO_NODE) {
140 pol = &preferred_node_policy[node];
141 /* preferred_node_policy is not initialised early in boot */
146 return &default_policy;
149 static const struct mempolicy_operations {
150 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
151 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
152 } mpol_ops[MPOL_MAX];
154 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
156 return pol->flags & MPOL_MODE_FLAGS;
159 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
160 const nodemask_t *rel)
163 nodes_fold(tmp, *orig, nodes_weight(*rel));
164 nodes_onto(*ret, tmp, *rel);
167 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
169 if (nodes_empty(*nodes))
171 pol->v.nodes = *nodes;
175 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
178 pol->flags |= MPOL_F_LOCAL; /* local allocation */
179 else if (nodes_empty(*nodes))
180 return -EINVAL; /* no allowed nodes */
182 pol->v.preferred_node = first_node(*nodes);
186 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
188 if (nodes_empty(*nodes))
190 pol->v.nodes = *nodes;
195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
196 * any, for the new policy. mpol_new() has already validated the nodes
197 * parameter with respect to the policy mode and flags. But, we need to
198 * handle an empty nodemask with MPOL_PREFERRED here.
200 * Must be called holding task's alloc_lock to protect task's mems_allowed
201 * and mempolicy. May also be called holding the mmap_semaphore for write.
203 static int mpol_set_nodemask(struct mempolicy *pol,
204 const nodemask_t *nodes, struct nodemask_scratch *nsc)
208 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
212 nodes_and(nsc->mask1,
213 cpuset_current_mems_allowed, node_states[N_MEMORY]);
216 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
217 nodes = NULL; /* explicit local allocation */
219 if (pol->flags & MPOL_F_RELATIVE_NODES)
220 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
222 nodes_and(nsc->mask2, *nodes, nsc->mask1);
224 if (mpol_store_user_nodemask(pol))
225 pol->w.user_nodemask = *nodes;
227 pol->w.cpuset_mems_allowed =
228 cpuset_current_mems_allowed;
232 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
234 ret = mpol_ops[pol->mode].create(pol, NULL);
239 * This function just creates a new policy, does some check and simple
240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
242 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
245 struct mempolicy *policy;
247 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
248 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
250 if (mode == MPOL_DEFAULT) {
251 if (nodes && !nodes_empty(*nodes))
252 return ERR_PTR(-EINVAL);
258 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
259 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
260 * All other modes require a valid pointer to a non-empty nodemask.
262 if (mode == MPOL_PREFERRED) {
263 if (nodes_empty(*nodes)) {
264 if (((flags & MPOL_F_STATIC_NODES) ||
265 (flags & MPOL_F_RELATIVE_NODES)))
266 return ERR_PTR(-EINVAL);
268 } else if (mode == MPOL_LOCAL) {
269 if (!nodes_empty(*nodes) ||
270 (flags & MPOL_F_STATIC_NODES) ||
271 (flags & MPOL_F_RELATIVE_NODES))
272 return ERR_PTR(-EINVAL);
273 mode = MPOL_PREFERRED;
274 } else if (nodes_empty(*nodes))
275 return ERR_PTR(-EINVAL);
276 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
278 return ERR_PTR(-ENOMEM);
279 atomic_set(&policy->refcnt, 1);
281 policy->flags = flags;
286 /* Slow path of a mpol destructor. */
287 void __mpol_put(struct mempolicy *p)
289 if (!atomic_dec_and_test(&p->refcnt))
291 kmem_cache_free(policy_cache, p);
294 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
298 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
302 if (pol->flags & MPOL_F_STATIC_NODES)
303 nodes_and(tmp, pol->w.user_nodemask, *nodes);
304 else if (pol->flags & MPOL_F_RELATIVE_NODES)
305 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
307 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
309 pol->w.cpuset_mems_allowed = *nodes;
312 if (nodes_empty(tmp))
318 static void mpol_rebind_preferred(struct mempolicy *pol,
319 const nodemask_t *nodes)
323 if (pol->flags & MPOL_F_STATIC_NODES) {
324 int node = first_node(pol->w.user_nodemask);
326 if (node_isset(node, *nodes)) {
327 pol->v.preferred_node = node;
328 pol->flags &= ~MPOL_F_LOCAL;
330 pol->flags |= MPOL_F_LOCAL;
331 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
332 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
333 pol->v.preferred_node = first_node(tmp);
334 } else if (!(pol->flags & MPOL_F_LOCAL)) {
335 pol->v.preferred_node = node_remap(pol->v.preferred_node,
336 pol->w.cpuset_mems_allowed,
338 pol->w.cpuset_mems_allowed = *nodes;
343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
346 * policies are protected by task->mems_allowed_seq to prevent a premature
347 * OOM/allocation failure due to parallel nodemask modification.
349 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
353 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
354 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
357 mpol_ops[pol->mode].rebind(pol, newmask);
361 * Wrapper for mpol_rebind_policy() that just requires task
362 * pointer, and updates task mempolicy.
364 * Called with task's alloc_lock held.
367 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
369 mpol_rebind_policy(tsk->mempolicy, new);
373 * Rebind each vma in mm to new nodemask.
375 * Call holding a reference to mm. Takes mm->mmap_sem during call.
378 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
380 struct vm_area_struct *vma;
382 down_write(&mm->mmap_sem);
383 for (vma = mm->mmap; vma; vma = vma->vm_next)
384 mpol_rebind_policy(vma->vm_policy, new);
385 up_write(&mm->mmap_sem);
388 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
390 .rebind = mpol_rebind_default,
392 [MPOL_INTERLEAVE] = {
393 .create = mpol_new_interleave,
394 .rebind = mpol_rebind_nodemask,
397 .create = mpol_new_preferred,
398 .rebind = mpol_rebind_preferred,
401 .create = mpol_new_bind,
402 .rebind = mpol_rebind_nodemask,
406 static int migrate_page_add(struct page *page, struct list_head *pagelist,
407 unsigned long flags);
410 struct list_head *pagelist;
413 struct vm_area_struct *prev;
417 * Check if the page's nid is in qp->nmask.
419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
420 * in the invert of qp->nmask.
422 static inline bool queue_pages_required(struct page *page,
423 struct queue_pages *qp)
425 int nid = page_to_nid(page);
426 unsigned long flags = qp->flags;
428 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
432 * queue_pages_pmd() has four possible return values:
433 * 0 - pages are placed on the right node or queued successfully.
434 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
437 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
438 * existing page was already on a node that does not follow the
441 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
442 unsigned long end, struct mm_walk *walk)
446 struct queue_pages *qp = walk->private;
449 if (unlikely(is_pmd_migration_entry(*pmd))) {
453 page = pmd_page(*pmd);
454 if (is_huge_zero_page(page)) {
456 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
460 if (!queue_pages_required(page, qp))
464 /* go to thp migration */
465 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
466 if (!vma_migratable(walk->vma) ||
467 migrate_page_add(page, qp->pagelist, flags)) {
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
483 * queue_pages_pte_range() has three possible return values:
484 * 0 - pages are placed on the right node or queued successfully.
485 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
487 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
488 * on a node that does not follow the policy.
490 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
491 unsigned long end, struct mm_walk *walk)
493 struct vm_area_struct *vma = walk->vma;
495 struct queue_pages *qp = walk->private;
496 unsigned long flags = qp->flags;
498 bool has_unmovable = false;
502 ptl = pmd_trans_huge_lock(pmd, vma);
504 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
508 /* THP was split, fall through to pte walk */
510 if (pmd_trans_unstable(pmd))
513 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
514 for (; addr != end; pte++, addr += PAGE_SIZE) {
515 if (!pte_present(*pte))
517 page = vm_normal_page(vma, addr, *pte);
521 * vm_normal_page() filters out zero pages, but there might
522 * still be PageReserved pages to skip, perhaps in a VDSO.
524 if (PageReserved(page))
526 if (!queue_pages_required(page, qp))
528 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
529 /* MPOL_MF_STRICT must be specified if we get here */
530 if (!vma_migratable(vma)) {
531 has_unmovable = true;
536 * Do not abort immediately since there may be
537 * temporary off LRU pages in the range. Still
538 * need migrate other LRU pages.
540 if (migrate_page_add(page, qp->pagelist, flags))
541 has_unmovable = true;
545 pte_unmap_unlock(pte - 1, ptl);
551 return addr != end ? -EIO : 0;
554 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
555 unsigned long addr, unsigned long end,
556 struct mm_walk *walk)
558 #ifdef CONFIG_HUGETLB_PAGE
559 struct queue_pages *qp = walk->private;
560 unsigned long flags = qp->flags;
565 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
566 entry = huge_ptep_get(pte);
567 if (!pte_present(entry))
569 page = pte_page(entry);
570 if (!queue_pages_required(page, qp))
572 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
573 if (flags & (MPOL_MF_MOVE_ALL) ||
574 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
575 isolate_huge_page(page, qp->pagelist);
584 #ifdef CONFIG_NUMA_BALANCING
586 * This is used to mark a range of virtual addresses to be inaccessible.
587 * These are later cleared by a NUMA hinting fault. Depending on these
588 * faults, pages may be migrated for better NUMA placement.
590 * This is assuming that NUMA faults are handled using PROT_NONE. If
591 * an architecture makes a different choice, it will need further
592 * changes to the core.
594 unsigned long change_prot_numa(struct vm_area_struct *vma,
595 unsigned long addr, unsigned long end)
599 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
601 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
606 static unsigned long change_prot_numa(struct vm_area_struct *vma,
607 unsigned long addr, unsigned long end)
611 #endif /* CONFIG_NUMA_BALANCING */
613 static int queue_pages_test_walk(unsigned long start, unsigned long end,
614 struct mm_walk *walk)
616 struct vm_area_struct *vma = walk->vma;
617 struct queue_pages *qp = walk->private;
618 unsigned long endvma = vma->vm_end;
619 unsigned long flags = qp->flags;
622 * Need check MPOL_MF_STRICT to return -EIO if possible
623 * regardless of vma_migratable
625 if (!vma_migratable(vma) &&
626 !(flags & MPOL_MF_STRICT))
631 if (vma->vm_start > start)
632 start = vma->vm_start;
634 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
635 if (!vma->vm_next && vma->vm_end < end)
637 if (qp->prev && qp->prev->vm_end < vma->vm_start)
643 if (flags & MPOL_MF_LAZY) {
644 /* Similar to task_numa_work, skip inaccessible VMAs */
645 if (!is_vm_hugetlb_page(vma) &&
646 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
647 !(vma->vm_flags & VM_MIXEDMAP))
648 change_prot_numa(vma, start, endvma);
652 /* queue pages from current vma */
653 if (flags & MPOL_MF_VALID)
659 * Walk through page tables and collect pages to be migrated.
661 * If pages found in a given range are on a set of nodes (determined by
662 * @nodes and @flags,) it's isolated and queued to the pagelist which is
663 * passed via @private.
665 * queue_pages_range() has three possible return values:
666 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
668 * 0 - queue pages successfully or no misplaced page.
669 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
670 * memory range specified by nodemask and maxnode points outside
671 * your accessible address space (-EFAULT)
674 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
675 nodemask_t *nodes, unsigned long flags,
676 struct list_head *pagelist)
678 struct queue_pages qp = {
679 .pagelist = pagelist,
684 struct mm_walk queue_pages_walk = {
685 .hugetlb_entry = queue_pages_hugetlb,
686 .pmd_entry = queue_pages_pte_range,
687 .test_walk = queue_pages_test_walk,
692 return walk_page_range(start, end, &queue_pages_walk);
696 * Apply policy to a single VMA
697 * This must be called with the mmap_sem held for writing.
699 static int vma_replace_policy(struct vm_area_struct *vma,
700 struct mempolicy *pol)
703 struct mempolicy *old;
704 struct mempolicy *new;
706 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
707 vma->vm_start, vma->vm_end, vma->vm_pgoff,
708 vma->vm_ops, vma->vm_file,
709 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
715 if (vma->vm_ops && vma->vm_ops->set_policy) {
716 err = vma->vm_ops->set_policy(vma, new);
721 old = vma->vm_policy;
722 vma->vm_policy = new; /* protected by mmap_sem */
731 /* Step 2: apply policy to a range and do splits. */
732 static int mbind_range(struct mm_struct *mm, unsigned long start,
733 unsigned long end, struct mempolicy *new_pol)
735 struct vm_area_struct *next;
736 struct vm_area_struct *prev;
737 struct vm_area_struct *vma;
740 unsigned long vmstart;
743 vma = find_vma(mm, start);
744 if (!vma || vma->vm_start > start)
748 if (start > vma->vm_start)
751 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
753 vmstart = max(start, vma->vm_start);
754 vmend = min(end, vma->vm_end);
756 if (mpol_equal(vma_policy(vma), new_pol))
759 pgoff = vma->vm_pgoff +
760 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
761 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
762 vma->anon_vma, vma->vm_file, pgoff,
763 new_pol, vma->vm_userfaultfd_ctx);
767 if (mpol_equal(vma_policy(vma), new_pol))
769 /* vma_merge() joined vma && vma->next, case 8 */
772 if (vma->vm_start != vmstart) {
773 err = split_vma(vma->vm_mm, vma, vmstart, 1);
777 if (vma->vm_end != vmend) {
778 err = split_vma(vma->vm_mm, vma, vmend, 0);
783 err = vma_replace_policy(vma, new_pol);
792 /* Set the process memory policy */
793 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
796 struct mempolicy *new, *old;
797 NODEMASK_SCRATCH(scratch);
803 new = mpol_new(mode, flags, nodes);
810 ret = mpol_set_nodemask(new, nodes, scratch);
812 task_unlock(current);
816 old = current->mempolicy;
817 current->mempolicy = new;
818 if (new && new->mode == MPOL_INTERLEAVE)
819 current->il_prev = MAX_NUMNODES-1;
820 task_unlock(current);
824 NODEMASK_SCRATCH_FREE(scratch);
829 * Return nodemask for policy for get_mempolicy() query
831 * Called with task's alloc_lock held
833 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
836 if (p == &default_policy)
842 case MPOL_INTERLEAVE:
846 if (!(p->flags & MPOL_F_LOCAL))
847 node_set(p->v.preferred_node, *nodes);
848 /* else return empty node mask for local allocation */
855 static int lookup_node(struct mm_struct *mm, unsigned long addr)
861 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
863 err = page_to_nid(p);
867 up_read(&mm->mmap_sem);
871 /* Retrieve NUMA policy */
872 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
873 unsigned long addr, unsigned long flags)
876 struct mm_struct *mm = current->mm;
877 struct vm_area_struct *vma = NULL;
878 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
881 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
884 if (flags & MPOL_F_MEMS_ALLOWED) {
885 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
887 *policy = 0; /* just so it's initialized */
889 *nmask = cpuset_current_mems_allowed;
890 task_unlock(current);
894 if (flags & MPOL_F_ADDR) {
896 * Do NOT fall back to task policy if the
897 * vma/shared policy at addr is NULL. We
898 * want to return MPOL_DEFAULT in this case.
900 down_read(&mm->mmap_sem);
901 vma = find_vma_intersection(mm, addr, addr+1);
903 up_read(&mm->mmap_sem);
906 if (vma->vm_ops && vma->vm_ops->get_policy)
907 pol = vma->vm_ops->get_policy(vma, addr);
909 pol = vma->vm_policy;
914 pol = &default_policy; /* indicates default behavior */
916 if (flags & MPOL_F_NODE) {
917 if (flags & MPOL_F_ADDR) {
919 * Take a refcount on the mpol, lookup_node()
920 * wil drop the mmap_sem, so after calling
921 * lookup_node() only "pol" remains valid, "vma"
927 err = lookup_node(mm, addr);
931 } else if (pol == current->mempolicy &&
932 pol->mode == MPOL_INTERLEAVE) {
933 *policy = next_node_in(current->il_prev, pol->v.nodes);
939 *policy = pol == &default_policy ? MPOL_DEFAULT :
942 * Internal mempolicy flags must be masked off before exposing
943 * the policy to userspace.
945 *policy |= (pol->flags & MPOL_MODE_FLAGS);
950 if (mpol_store_user_nodemask(pol)) {
951 *nmask = pol->w.user_nodemask;
954 get_policy_nodemask(pol, nmask);
955 task_unlock(current);
962 up_read(&mm->mmap_sem);
964 mpol_put(pol_refcount);
968 #ifdef CONFIG_MIGRATION
970 * page migration, thp tail pages can be passed.
972 static int migrate_page_add(struct page *page, struct list_head *pagelist,
975 struct page *head = compound_head(page);
977 * Avoid migrating a page that is shared with others.
979 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
980 if (!isolate_lru_page(head)) {
981 list_add_tail(&head->lru, pagelist);
982 mod_node_page_state(page_pgdat(head),
983 NR_ISOLATED_ANON + page_is_file_cache(head),
984 hpage_nr_pages(head));
985 } else if (flags & MPOL_MF_STRICT) {
987 * Non-movable page may reach here. And, there may be
988 * temporary off LRU pages or non-LRU movable pages.
989 * Treat them as unmovable pages since they can't be
990 * isolated, so they can't be moved at the moment. It
991 * should return -EIO for this case too.
1000 /* page allocation callback for NUMA node migration */
1001 struct page *alloc_new_node_page(struct page *page, unsigned long node)
1004 return alloc_huge_page_node(page_hstate(compound_head(page)),
1006 else if (PageTransHuge(page)) {
1009 thp = alloc_pages_node(node,
1010 (GFP_TRANSHUGE | __GFP_THISNODE),
1014 prep_transhuge_page(thp);
1017 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1022 * Migrate pages from one node to a target node.
1023 * Returns error or the number of pages not migrated.
1025 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1029 LIST_HEAD(pagelist);
1033 node_set(source, nmask);
1036 * This does not "check" the range but isolates all pages that
1037 * need migration. Between passing in the full user address
1038 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1040 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1041 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1042 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1044 if (!list_empty(&pagelist)) {
1045 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1046 MIGRATE_SYNC, MR_SYSCALL);
1048 putback_movable_pages(&pagelist);
1055 * Move pages between the two nodesets so as to preserve the physical
1056 * layout as much as possible.
1058 * Returns the number of page that could not be moved.
1060 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1061 const nodemask_t *to, int flags)
1067 err = migrate_prep();
1071 down_read(&mm->mmap_sem);
1074 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1075 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1076 * bit in 'tmp', and return that <source, dest> pair for migration.
1077 * The pair of nodemasks 'to' and 'from' define the map.
1079 * If no pair of bits is found that way, fallback to picking some
1080 * pair of 'source' and 'dest' bits that are not the same. If the
1081 * 'source' and 'dest' bits are the same, this represents a node
1082 * that will be migrating to itself, so no pages need move.
1084 * If no bits are left in 'tmp', or if all remaining bits left
1085 * in 'tmp' correspond to the same bit in 'to', return false
1086 * (nothing left to migrate).
1088 * This lets us pick a pair of nodes to migrate between, such that
1089 * if possible the dest node is not already occupied by some other
1090 * source node, minimizing the risk of overloading the memory on a
1091 * node that would happen if we migrated incoming memory to a node
1092 * before migrating outgoing memory source that same node.
1094 * A single scan of tmp is sufficient. As we go, we remember the
1095 * most recent <s, d> pair that moved (s != d). If we find a pair
1096 * that not only moved, but what's better, moved to an empty slot
1097 * (d is not set in tmp), then we break out then, with that pair.
1098 * Otherwise when we finish scanning from_tmp, we at least have the
1099 * most recent <s, d> pair that moved. If we get all the way through
1100 * the scan of tmp without finding any node that moved, much less
1101 * moved to an empty node, then there is nothing left worth migrating.
1105 while (!nodes_empty(tmp)) {
1107 int source = NUMA_NO_NODE;
1110 for_each_node_mask(s, tmp) {
1113 * do_migrate_pages() tries to maintain the relative
1114 * node relationship of the pages established between
1115 * threads and memory areas.
1117 * However if the number of source nodes is not equal to
1118 * the number of destination nodes we can not preserve
1119 * this node relative relationship. In that case, skip
1120 * copying memory from a node that is in the destination
1123 * Example: [2,3,4] -> [3,4,5] moves everything.
1124 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1127 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1128 (node_isset(s, *to)))
1131 d = node_remap(s, *from, *to);
1135 source = s; /* Node moved. Memorize */
1138 /* dest not in remaining from nodes? */
1139 if (!node_isset(dest, tmp))
1142 if (source == NUMA_NO_NODE)
1145 node_clear(source, tmp);
1146 err = migrate_to_node(mm, source, dest, flags);
1152 up_read(&mm->mmap_sem);
1160 * Allocate a new page for page migration based on vma policy.
1161 * Start by assuming the page is mapped by the same vma as contains @start.
1162 * Search forward from there, if not. N.B., this assumes that the
1163 * list of pages handed to migrate_pages()--which is how we get here--
1164 * is in virtual address order.
1166 static struct page *new_page(struct page *page, unsigned long start)
1168 struct vm_area_struct *vma;
1169 unsigned long uninitialized_var(address);
1171 vma = find_vma(current->mm, start);
1173 address = page_address_in_vma(page, vma);
1174 if (address != -EFAULT)
1179 if (PageHuge(page)) {
1180 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1182 } else if (PageTransHuge(page)) {
1185 thp = alloc_pages_vma(GFP_TRANSHUGE, HPAGE_PMD_ORDER, vma,
1186 address, numa_node_id());
1189 prep_transhuge_page(thp);
1193 * if !vma, alloc_page_vma() will use task or system default policy
1195 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1200 static int migrate_page_add(struct page *page, struct list_head *pagelist,
1201 unsigned long flags)
1206 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1207 const nodemask_t *to, int flags)
1212 static struct page *new_page(struct page *page, unsigned long start)
1218 static long do_mbind(unsigned long start, unsigned long len,
1219 unsigned short mode, unsigned short mode_flags,
1220 nodemask_t *nmask, unsigned long flags)
1222 struct mm_struct *mm = current->mm;
1223 struct mempolicy *new;
1227 LIST_HEAD(pagelist);
1229 if (flags & ~(unsigned long)MPOL_MF_VALID)
1231 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1234 if (start & ~PAGE_MASK)
1237 if (mode == MPOL_DEFAULT)
1238 flags &= ~MPOL_MF_STRICT;
1240 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1248 new = mpol_new(mode, mode_flags, nmask);
1250 return PTR_ERR(new);
1252 if (flags & MPOL_MF_LAZY)
1253 new->flags |= MPOL_F_MOF;
1256 * If we are using the default policy then operation
1257 * on discontinuous address spaces is okay after all
1260 flags |= MPOL_MF_DISCONTIG_OK;
1262 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1263 start, start + len, mode, mode_flags,
1264 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1266 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1268 err = migrate_prep();
1273 NODEMASK_SCRATCH(scratch);
1275 down_write(&mm->mmap_sem);
1277 err = mpol_set_nodemask(new, nmask, scratch);
1278 task_unlock(current);
1280 up_write(&mm->mmap_sem);
1283 NODEMASK_SCRATCH_FREE(scratch);
1288 ret = queue_pages_range(mm, start, end, nmask,
1289 flags | MPOL_MF_INVERT, &pagelist);
1296 err = mbind_range(mm, start, end, new);
1301 if (!list_empty(&pagelist)) {
1302 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1303 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1304 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1306 putback_movable_pages(&pagelist);
1309 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1313 if (!list_empty(&pagelist))
1314 putback_movable_pages(&pagelist);
1317 up_write(&mm->mmap_sem);
1324 * User space interface with variable sized bitmaps for nodelists.
1327 /* Copy a node mask from user space. */
1328 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1329 unsigned long maxnode)
1333 unsigned long nlongs;
1334 unsigned long endmask;
1337 nodes_clear(*nodes);
1338 if (maxnode == 0 || !nmask)
1340 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1343 nlongs = BITS_TO_LONGS(maxnode);
1344 if ((maxnode % BITS_PER_LONG) == 0)
1347 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1350 * When the user specified more nodes than supported just check
1351 * if the non supported part is all zero.
1353 * If maxnode have more longs than MAX_NUMNODES, check
1354 * the bits in that area first. And then go through to
1355 * check the rest bits which equal or bigger than MAX_NUMNODES.
1356 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1358 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1359 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1360 if (get_user(t, nmask + k))
1362 if (k == nlongs - 1) {
1368 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1372 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1373 unsigned long valid_mask = endmask;
1375 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1376 if (get_user(t, nmask + nlongs - 1))
1382 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1384 nodes_addr(*nodes)[nlongs-1] &= endmask;
1388 /* Copy a kernel node mask to user space */
1389 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1392 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1393 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1395 if (copy > nbytes) {
1396 if (copy > PAGE_SIZE)
1398 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1402 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1405 static long kernel_mbind(unsigned long start, unsigned long len,
1406 unsigned long mode, const unsigned long __user *nmask,
1407 unsigned long maxnode, unsigned int flags)
1411 unsigned short mode_flags;
1413 mode_flags = mode & MPOL_MODE_FLAGS;
1414 mode &= ~MPOL_MODE_FLAGS;
1415 if (mode >= MPOL_MAX)
1417 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1418 (mode_flags & MPOL_F_RELATIVE_NODES))
1420 err = get_nodes(&nodes, nmask, maxnode);
1423 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1426 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1427 unsigned long, mode, const unsigned long __user *, nmask,
1428 unsigned long, maxnode, unsigned int, flags)
1430 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1433 /* Set the process memory policy */
1434 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1435 unsigned long maxnode)
1439 unsigned short flags;
1441 flags = mode & MPOL_MODE_FLAGS;
1442 mode &= ~MPOL_MODE_FLAGS;
1443 if ((unsigned int)mode >= MPOL_MAX)
1445 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1447 err = get_nodes(&nodes, nmask, maxnode);
1450 return do_set_mempolicy(mode, flags, &nodes);
1453 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1454 unsigned long, maxnode)
1456 return kernel_set_mempolicy(mode, nmask, maxnode);
1459 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1460 const unsigned long __user *old_nodes,
1461 const unsigned long __user *new_nodes)
1463 struct mm_struct *mm = NULL;
1464 struct task_struct *task;
1465 nodemask_t task_nodes;
1469 NODEMASK_SCRATCH(scratch);
1474 old = &scratch->mask1;
1475 new = &scratch->mask2;
1477 err = get_nodes(old, old_nodes, maxnode);
1481 err = get_nodes(new, new_nodes, maxnode);
1485 /* Find the mm_struct */
1487 task = pid ? find_task_by_vpid(pid) : current;
1493 get_task_struct(task);
1498 * Check if this process has the right to modify the specified process.
1499 * Use the regular "ptrace_may_access()" checks.
1501 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1508 task_nodes = cpuset_mems_allowed(task);
1509 /* Is the user allowed to access the target nodes? */
1510 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1515 task_nodes = cpuset_mems_allowed(current);
1516 nodes_and(*new, *new, task_nodes);
1517 if (nodes_empty(*new))
1520 nodes_and(*new, *new, node_states[N_MEMORY]);
1521 if (nodes_empty(*new))
1524 err = security_task_movememory(task);
1528 mm = get_task_mm(task);
1529 put_task_struct(task);
1536 err = do_migrate_pages(mm, old, new,
1537 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1541 NODEMASK_SCRATCH_FREE(scratch);
1546 put_task_struct(task);
1551 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1552 const unsigned long __user *, old_nodes,
1553 const unsigned long __user *, new_nodes)
1555 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1559 /* Retrieve NUMA policy */
1560 static int kernel_get_mempolicy(int __user *policy,
1561 unsigned long __user *nmask,
1562 unsigned long maxnode,
1564 unsigned long flags)
1567 int uninitialized_var(pval);
1570 if (nmask != NULL && maxnode < nr_node_ids)
1573 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1578 if (policy && put_user(pval, policy))
1582 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1587 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1588 unsigned long __user *, nmask, unsigned long, maxnode,
1589 unsigned long, addr, unsigned long, flags)
1591 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1594 #ifdef CONFIG_COMPAT
1596 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1597 compat_ulong_t __user *, nmask,
1598 compat_ulong_t, maxnode,
1599 compat_ulong_t, addr, compat_ulong_t, flags)
1602 unsigned long __user *nm = NULL;
1603 unsigned long nr_bits, alloc_size;
1604 DECLARE_BITMAP(bm, MAX_NUMNODES);
1606 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1607 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1610 nm = compat_alloc_user_space(alloc_size);
1612 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1614 if (!err && nmask) {
1615 unsigned long copy_size;
1616 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1617 err = copy_from_user(bm, nm, copy_size);
1618 /* ensure entire bitmap is zeroed */
1619 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1620 err |= compat_put_bitmap(nmask, bm, nr_bits);
1626 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1627 compat_ulong_t, maxnode)
1629 unsigned long __user *nm = NULL;
1630 unsigned long nr_bits, alloc_size;
1631 DECLARE_BITMAP(bm, MAX_NUMNODES);
1633 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1634 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1637 if (compat_get_bitmap(bm, nmask, nr_bits))
1639 nm = compat_alloc_user_space(alloc_size);
1640 if (copy_to_user(nm, bm, alloc_size))
1644 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1647 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1648 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1649 compat_ulong_t, maxnode, compat_ulong_t, flags)
1651 unsigned long __user *nm = NULL;
1652 unsigned long nr_bits, alloc_size;
1655 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1656 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1659 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1661 nm = compat_alloc_user_space(alloc_size);
1662 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1666 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1669 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1670 compat_ulong_t, maxnode,
1671 const compat_ulong_t __user *, old_nodes,
1672 const compat_ulong_t __user *, new_nodes)
1674 unsigned long __user *old = NULL;
1675 unsigned long __user *new = NULL;
1676 nodemask_t tmp_mask;
1677 unsigned long nr_bits;
1680 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1681 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1683 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1685 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1687 new = old + size / sizeof(unsigned long);
1688 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1692 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1695 new = compat_alloc_user_space(size);
1696 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1699 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1702 #endif /* CONFIG_COMPAT */
1704 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1707 struct mempolicy *pol = NULL;
1710 if (vma->vm_ops && vma->vm_ops->get_policy) {
1711 pol = vma->vm_ops->get_policy(vma, addr);
1712 } else if (vma->vm_policy) {
1713 pol = vma->vm_policy;
1716 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1717 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1718 * count on these policies which will be dropped by
1719 * mpol_cond_put() later
1721 if (mpol_needs_cond_ref(pol))
1730 * get_vma_policy(@vma, @addr)
1731 * @vma: virtual memory area whose policy is sought
1732 * @addr: address in @vma for shared policy lookup
1734 * Returns effective policy for a VMA at specified address.
1735 * Falls back to current->mempolicy or system default policy, as necessary.
1736 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1737 * count--added by the get_policy() vm_op, as appropriate--to protect against
1738 * freeing by another task. It is the caller's responsibility to free the
1739 * extra reference for shared policies.
1741 struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1744 struct mempolicy *pol = __get_vma_policy(vma, addr);
1747 pol = get_task_policy(current);
1752 bool vma_policy_mof(struct vm_area_struct *vma)
1754 struct mempolicy *pol;
1756 if (vma->vm_ops && vma->vm_ops->get_policy) {
1759 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1760 if (pol && (pol->flags & MPOL_F_MOF))
1767 pol = vma->vm_policy;
1769 pol = get_task_policy(current);
1771 return pol->flags & MPOL_F_MOF;
1774 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1776 enum zone_type dynamic_policy_zone = policy_zone;
1778 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1781 * if policy->v.nodes has movable memory only,
1782 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1784 * policy->v.nodes is intersect with node_states[N_MEMORY].
1785 * so if the following test faile, it implies
1786 * policy->v.nodes has movable memory only.
1788 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1789 dynamic_policy_zone = ZONE_MOVABLE;
1791 return zone >= dynamic_policy_zone;
1795 * Return a nodemask representing a mempolicy for filtering nodes for
1798 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1800 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1801 if (unlikely(policy->mode == MPOL_BIND) &&
1802 apply_policy_zone(policy, gfp_zone(gfp)) &&
1803 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1804 return &policy->v.nodes;
1809 /* Return the node id preferred by the given mempolicy, or the given id */
1810 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1813 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1814 nd = policy->v.preferred_node;
1817 * __GFP_THISNODE shouldn't even be used with the bind policy
1818 * because we might easily break the expectation to stay on the
1819 * requested node and not break the policy.
1821 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1827 /* Do dynamic interleaving for a process */
1828 static unsigned interleave_nodes(struct mempolicy *policy)
1831 struct task_struct *me = current;
1833 next = next_node_in(me->il_prev, policy->v.nodes);
1834 if (next < MAX_NUMNODES)
1840 * Depending on the memory policy provide a node from which to allocate the
1843 unsigned int mempolicy_slab_node(void)
1845 struct mempolicy *policy;
1846 int node = numa_mem_id();
1851 policy = current->mempolicy;
1852 if (!policy || policy->flags & MPOL_F_LOCAL)
1855 switch (policy->mode) {
1856 case MPOL_PREFERRED:
1858 * handled MPOL_F_LOCAL above
1860 return policy->v.preferred_node;
1862 case MPOL_INTERLEAVE:
1863 return interleave_nodes(policy);
1869 * Follow bind policy behavior and start allocation at the
1872 struct zonelist *zonelist;
1873 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1874 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1875 z = first_zones_zonelist(zonelist, highest_zoneidx,
1877 return z->zone ? zone_to_nid(z->zone) : node;
1886 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1887 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1888 * number of present nodes.
1890 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1892 unsigned nnodes = nodes_weight(pol->v.nodes);
1898 return numa_node_id();
1899 target = (unsigned int)n % nnodes;
1900 nid = first_node(pol->v.nodes);
1901 for (i = 0; i < target; i++)
1902 nid = next_node(nid, pol->v.nodes);
1906 /* Determine a node number for interleave */
1907 static inline unsigned interleave_nid(struct mempolicy *pol,
1908 struct vm_area_struct *vma, unsigned long addr, int shift)
1914 * for small pages, there is no difference between
1915 * shift and PAGE_SHIFT, so the bit-shift is safe.
1916 * for huge pages, since vm_pgoff is in units of small
1917 * pages, we need to shift off the always 0 bits to get
1920 BUG_ON(shift < PAGE_SHIFT);
1921 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1922 off += (addr - vma->vm_start) >> shift;
1923 return offset_il_node(pol, off);
1925 return interleave_nodes(pol);
1928 #ifdef CONFIG_HUGETLBFS
1930 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1931 * @vma: virtual memory area whose policy is sought
1932 * @addr: address in @vma for shared policy lookup and interleave policy
1933 * @gfp_flags: for requested zone
1934 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1935 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1937 * Returns a nid suitable for a huge page allocation and a pointer
1938 * to the struct mempolicy for conditional unref after allocation.
1939 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1940 * @nodemask for filtering the zonelist.
1942 * Must be protected by read_mems_allowed_begin()
1944 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1945 struct mempolicy **mpol, nodemask_t **nodemask)
1949 *mpol = get_vma_policy(vma, addr);
1950 *nodemask = NULL; /* assume !MPOL_BIND */
1952 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1953 nid = interleave_nid(*mpol, vma, addr,
1954 huge_page_shift(hstate_vma(vma)));
1956 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1957 if ((*mpol)->mode == MPOL_BIND)
1958 *nodemask = &(*mpol)->v.nodes;
1964 * init_nodemask_of_mempolicy
1966 * If the current task's mempolicy is "default" [NULL], return 'false'
1967 * to indicate default policy. Otherwise, extract the policy nodemask
1968 * for 'bind' or 'interleave' policy into the argument nodemask, or
1969 * initialize the argument nodemask to contain the single node for
1970 * 'preferred' or 'local' policy and return 'true' to indicate presence
1971 * of non-default mempolicy.
1973 * We don't bother with reference counting the mempolicy [mpol_get/put]
1974 * because the current task is examining it's own mempolicy and a task's
1975 * mempolicy is only ever changed by the task itself.
1977 * N.B., it is the caller's responsibility to free a returned nodemask.
1979 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1981 struct mempolicy *mempolicy;
1984 if (!(mask && current->mempolicy))
1988 mempolicy = current->mempolicy;
1989 switch (mempolicy->mode) {
1990 case MPOL_PREFERRED:
1991 if (mempolicy->flags & MPOL_F_LOCAL)
1992 nid = numa_node_id();
1994 nid = mempolicy->v.preferred_node;
1995 init_nodemask_of_node(mask, nid);
2000 case MPOL_INTERLEAVE:
2001 *mask = mempolicy->v.nodes;
2007 task_unlock(current);
2014 * mempolicy_nodemask_intersects
2016 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2017 * policy. Otherwise, check for intersection between mask and the policy
2018 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2019 * policy, always return true since it may allocate elsewhere on fallback.
2021 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2023 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2024 const nodemask_t *mask)
2026 struct mempolicy *mempolicy;
2032 mempolicy = tsk->mempolicy;
2036 switch (mempolicy->mode) {
2037 case MPOL_PREFERRED:
2039 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2040 * allocate from, they may fallback to other nodes when oom.
2041 * Thus, it's possible for tsk to have allocated memory from
2046 case MPOL_INTERLEAVE:
2047 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2057 /* Allocate a page in interleaved policy.
2058 Own path because it needs to do special accounting. */
2059 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2064 page = __alloc_pages(gfp, order, nid);
2065 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2066 if (!static_branch_likely(&vm_numa_stat_key))
2068 if (page && page_to_nid(page) == nid) {
2070 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2077 * alloc_pages_vma - Allocate a page for a VMA.
2080 * %GFP_USER user allocation.
2081 * %GFP_KERNEL kernel allocations,
2082 * %GFP_HIGHMEM highmem/user allocations,
2083 * %GFP_FS allocation should not call back into a file system.
2084 * %GFP_ATOMIC don't sleep.
2086 * @order:Order of the GFP allocation.
2087 * @vma: Pointer to VMA or NULL if not available.
2088 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2089 * @node: Which node to prefer for allocation (modulo policy).
2091 * This function allocates a page from the kernel page pool and applies
2092 * a NUMA policy associated with the VMA or the current process.
2093 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2094 * mm_struct of the VMA to prevent it from going away. Should be used for
2095 * all allocations for pages that will be mapped into user space. Returns
2096 * NULL when no page can be allocated.
2099 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2100 unsigned long addr, int node)
2102 struct mempolicy *pol;
2107 pol = get_vma_policy(vma, addr);
2109 if (pol->mode == MPOL_INTERLEAVE) {
2112 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2114 page = alloc_page_interleave(gfp, order, nid);
2118 nmask = policy_nodemask(gfp, pol);
2119 preferred_nid = policy_node(gfp, pol, node);
2120 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2125 EXPORT_SYMBOL(alloc_pages_vma);
2128 * alloc_pages_current - Allocate pages.
2131 * %GFP_USER user allocation,
2132 * %GFP_KERNEL kernel allocation,
2133 * %GFP_HIGHMEM highmem allocation,
2134 * %GFP_FS don't call back into a file system.
2135 * %GFP_ATOMIC don't sleep.
2136 * @order: Power of two of allocation size in pages. 0 is a single page.
2138 * Allocate a page from the kernel page pool. When not in
2139 * interrupt context and apply the current process NUMA policy.
2140 * Returns NULL when no page can be allocated.
2142 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2144 struct mempolicy *pol = &default_policy;
2147 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2148 pol = get_task_policy(current);
2151 * No reference counting needed for current->mempolicy
2152 * nor system default_policy
2154 if (pol->mode == MPOL_INTERLEAVE)
2155 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2157 page = __alloc_pages_nodemask(gfp, order,
2158 policy_node(gfp, pol, numa_node_id()),
2159 policy_nodemask(gfp, pol));
2163 EXPORT_SYMBOL(alloc_pages_current);
2165 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2167 struct mempolicy *pol = mpol_dup(vma_policy(src));
2170 return PTR_ERR(pol);
2171 dst->vm_policy = pol;
2176 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2177 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2178 * with the mems_allowed returned by cpuset_mems_allowed(). This
2179 * keeps mempolicies cpuset relative after its cpuset moves. See
2180 * further kernel/cpuset.c update_nodemask().
2182 * current's mempolicy may be rebinded by the other task(the task that changes
2183 * cpuset's mems), so we needn't do rebind work for current task.
2186 /* Slow path of a mempolicy duplicate */
2187 struct mempolicy *__mpol_dup(struct mempolicy *old)
2189 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2192 return ERR_PTR(-ENOMEM);
2194 /* task's mempolicy is protected by alloc_lock */
2195 if (old == current->mempolicy) {
2198 task_unlock(current);
2202 if (current_cpuset_is_being_rebound()) {
2203 nodemask_t mems = cpuset_mems_allowed(current);
2204 mpol_rebind_policy(new, &mems);
2206 atomic_set(&new->refcnt, 1);
2210 /* Slow path of a mempolicy comparison */
2211 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2215 if (a->mode != b->mode)
2217 if (a->flags != b->flags)
2219 if (mpol_store_user_nodemask(a))
2220 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2226 case MPOL_INTERLEAVE:
2227 return !!nodes_equal(a->v.nodes, b->v.nodes);
2228 case MPOL_PREFERRED:
2229 /* a's ->flags is the same as b's */
2230 if (a->flags & MPOL_F_LOCAL)
2232 return a->v.preferred_node == b->v.preferred_node;
2240 * Shared memory backing store policy support.
2242 * Remember policies even when nobody has shared memory mapped.
2243 * The policies are kept in Red-Black tree linked from the inode.
2244 * They are protected by the sp->lock rwlock, which should be held
2245 * for any accesses to the tree.
2249 * lookup first element intersecting start-end. Caller holds sp->lock for
2250 * reading or for writing
2252 static struct sp_node *
2253 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2255 struct rb_node *n = sp->root.rb_node;
2258 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2260 if (start >= p->end)
2262 else if (end <= p->start)
2270 struct sp_node *w = NULL;
2271 struct rb_node *prev = rb_prev(n);
2274 w = rb_entry(prev, struct sp_node, nd);
2275 if (w->end <= start)
2279 return rb_entry(n, struct sp_node, nd);
2283 * Insert a new shared policy into the list. Caller holds sp->lock for
2286 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2288 struct rb_node **p = &sp->root.rb_node;
2289 struct rb_node *parent = NULL;
2294 nd = rb_entry(parent, struct sp_node, nd);
2295 if (new->start < nd->start)
2297 else if (new->end > nd->end)
2298 p = &(*p)->rb_right;
2302 rb_link_node(&new->nd, parent, p);
2303 rb_insert_color(&new->nd, &sp->root);
2304 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2305 new->policy ? new->policy->mode : 0);
2308 /* Find shared policy intersecting idx */
2310 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2312 struct mempolicy *pol = NULL;
2315 if (!sp->root.rb_node)
2317 read_lock(&sp->lock);
2318 sn = sp_lookup(sp, idx, idx+1);
2320 mpol_get(sn->policy);
2323 read_unlock(&sp->lock);
2327 static void sp_free(struct sp_node *n)
2329 mpol_put(n->policy);
2330 kmem_cache_free(sn_cache, n);
2334 * mpol_misplaced - check whether current page node is valid in policy
2336 * @page: page to be checked
2337 * @vma: vm area where page mapped
2338 * @addr: virtual address where page mapped
2340 * Lookup current policy node id for vma,addr and "compare to" page's
2344 * -1 - not misplaced, page is in the right node
2345 * node - node id where the page should be
2347 * Policy determination "mimics" alloc_page_vma().
2348 * Called from fault path where we know the vma and faulting address.
2350 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2352 struct mempolicy *pol;
2354 int curnid = page_to_nid(page);
2355 unsigned long pgoff;
2356 int thiscpu = raw_smp_processor_id();
2357 int thisnid = cpu_to_node(thiscpu);
2358 int polnid = NUMA_NO_NODE;
2361 pol = get_vma_policy(vma, addr);
2362 if (!(pol->flags & MPOL_F_MOF))
2365 switch (pol->mode) {
2366 case MPOL_INTERLEAVE:
2367 pgoff = vma->vm_pgoff;
2368 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2369 polnid = offset_il_node(pol, pgoff);
2372 case MPOL_PREFERRED:
2373 if (pol->flags & MPOL_F_LOCAL)
2374 polnid = numa_node_id();
2376 polnid = pol->v.preferred_node;
2382 * allows binding to multiple nodes.
2383 * use current page if in policy nodemask,
2384 * else select nearest allowed node, if any.
2385 * If no allowed nodes, use current [!misplaced].
2387 if (node_isset(curnid, pol->v.nodes))
2389 z = first_zones_zonelist(
2390 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2391 gfp_zone(GFP_HIGHUSER),
2393 polnid = zone_to_nid(z->zone);
2400 /* Migrate the page towards the node whose CPU is referencing it */
2401 if (pol->flags & MPOL_F_MORON) {
2404 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2408 if (curnid != polnid)
2417 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2418 * dropped after task->mempolicy is set to NULL so that any allocation done as
2419 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2422 void mpol_put_task_policy(struct task_struct *task)
2424 struct mempolicy *pol;
2427 pol = task->mempolicy;
2428 task->mempolicy = NULL;
2433 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2435 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2436 rb_erase(&n->nd, &sp->root);
2440 static void sp_node_init(struct sp_node *node, unsigned long start,
2441 unsigned long end, struct mempolicy *pol)
2443 node->start = start;
2448 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2449 struct mempolicy *pol)
2452 struct mempolicy *newpol;
2454 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2458 newpol = mpol_dup(pol);
2459 if (IS_ERR(newpol)) {
2460 kmem_cache_free(sn_cache, n);
2463 newpol->flags |= MPOL_F_SHARED;
2464 sp_node_init(n, start, end, newpol);
2469 /* Replace a policy range. */
2470 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2471 unsigned long end, struct sp_node *new)
2474 struct sp_node *n_new = NULL;
2475 struct mempolicy *mpol_new = NULL;
2479 write_lock(&sp->lock);
2480 n = sp_lookup(sp, start, end);
2481 /* Take care of old policies in the same range. */
2482 while (n && n->start < end) {
2483 struct rb_node *next = rb_next(&n->nd);
2484 if (n->start >= start) {
2490 /* Old policy spanning whole new range. */
2495 *mpol_new = *n->policy;
2496 atomic_set(&mpol_new->refcnt, 1);
2497 sp_node_init(n_new, end, n->end, mpol_new);
2499 sp_insert(sp, n_new);
2508 n = rb_entry(next, struct sp_node, nd);
2512 write_unlock(&sp->lock);
2519 kmem_cache_free(sn_cache, n_new);
2524 write_unlock(&sp->lock);
2526 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2529 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2536 * mpol_shared_policy_init - initialize shared policy for inode
2537 * @sp: pointer to inode shared policy
2538 * @mpol: struct mempolicy to install
2540 * Install non-NULL @mpol in inode's shared policy rb-tree.
2541 * On entry, the current task has a reference on a non-NULL @mpol.
2542 * This must be released on exit.
2543 * This is called at get_inode() calls and we can use GFP_KERNEL.
2545 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2549 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2550 rwlock_init(&sp->lock);
2553 struct vm_area_struct pvma;
2554 struct mempolicy *new;
2555 NODEMASK_SCRATCH(scratch);
2559 /* contextualize the tmpfs mount point mempolicy */
2560 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2562 goto free_scratch; /* no valid nodemask intersection */
2565 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2566 task_unlock(current);
2570 /* Create pseudo-vma that contains just the policy */
2571 vma_init(&pvma, NULL);
2572 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2573 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2576 mpol_put(new); /* drop initial ref */
2578 NODEMASK_SCRATCH_FREE(scratch);
2580 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2584 int mpol_set_shared_policy(struct shared_policy *info,
2585 struct vm_area_struct *vma, struct mempolicy *npol)
2588 struct sp_node *new = NULL;
2589 unsigned long sz = vma_pages(vma);
2591 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2593 sz, npol ? npol->mode : -1,
2594 npol ? npol->flags : -1,
2595 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2598 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2602 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2608 /* Free a backing policy store on inode delete. */
2609 void mpol_free_shared_policy(struct shared_policy *p)
2612 struct rb_node *next;
2614 if (!p->root.rb_node)
2616 write_lock(&p->lock);
2617 next = rb_first(&p->root);
2619 n = rb_entry(next, struct sp_node, nd);
2620 next = rb_next(&n->nd);
2623 write_unlock(&p->lock);
2626 #ifdef CONFIG_NUMA_BALANCING
2627 static int __initdata numabalancing_override;
2629 static void __init check_numabalancing_enable(void)
2631 bool numabalancing_default = false;
2633 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2634 numabalancing_default = true;
2636 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2637 if (numabalancing_override)
2638 set_numabalancing_state(numabalancing_override == 1);
2640 if (num_online_nodes() > 1 && !numabalancing_override) {
2641 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2642 numabalancing_default ? "Enabling" : "Disabling");
2643 set_numabalancing_state(numabalancing_default);
2647 static int __init setup_numabalancing(char *str)
2653 if (!strcmp(str, "enable")) {
2654 numabalancing_override = 1;
2656 } else if (!strcmp(str, "disable")) {
2657 numabalancing_override = -1;
2662 pr_warn("Unable to parse numa_balancing=\n");
2666 __setup("numa_balancing=", setup_numabalancing);
2668 static inline void __init check_numabalancing_enable(void)
2671 #endif /* CONFIG_NUMA_BALANCING */
2673 /* assumes fs == KERNEL_DS */
2674 void __init numa_policy_init(void)
2676 nodemask_t interleave_nodes;
2677 unsigned long largest = 0;
2678 int nid, prefer = 0;
2680 policy_cache = kmem_cache_create("numa_policy",
2681 sizeof(struct mempolicy),
2682 0, SLAB_PANIC, NULL);
2684 sn_cache = kmem_cache_create("shared_policy_node",
2685 sizeof(struct sp_node),
2686 0, SLAB_PANIC, NULL);
2688 for_each_node(nid) {
2689 preferred_node_policy[nid] = (struct mempolicy) {
2690 .refcnt = ATOMIC_INIT(1),
2691 .mode = MPOL_PREFERRED,
2692 .flags = MPOL_F_MOF | MPOL_F_MORON,
2693 .v = { .preferred_node = nid, },
2698 * Set interleaving policy for system init. Interleaving is only
2699 * enabled across suitably sized nodes (default is >= 16MB), or
2700 * fall back to the largest node if they're all smaller.
2702 nodes_clear(interleave_nodes);
2703 for_each_node_state(nid, N_MEMORY) {
2704 unsigned long total_pages = node_present_pages(nid);
2706 /* Preserve the largest node */
2707 if (largest < total_pages) {
2708 largest = total_pages;
2712 /* Interleave this node? */
2713 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2714 node_set(nid, interleave_nodes);
2717 /* All too small, use the largest */
2718 if (unlikely(nodes_empty(interleave_nodes)))
2719 node_set(prefer, interleave_nodes);
2721 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2722 pr_err("%s: interleaving failed\n", __func__);
2724 check_numabalancing_enable();
2727 /* Reset policy of current process to default */
2728 void numa_default_policy(void)
2730 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2734 * Parse and format mempolicy from/to strings
2738 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2740 static const char * const policy_modes[] =
2742 [MPOL_DEFAULT] = "default",
2743 [MPOL_PREFERRED] = "prefer",
2744 [MPOL_BIND] = "bind",
2745 [MPOL_INTERLEAVE] = "interleave",
2746 [MPOL_LOCAL] = "local",
2752 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2753 * @str: string containing mempolicy to parse
2754 * @mpol: pointer to struct mempolicy pointer, returned on success.
2757 * <mode>[=<flags>][:<nodelist>]
2759 * On success, returns 0, else 1
2761 int mpol_parse_str(char *str, struct mempolicy **mpol)
2763 struct mempolicy *new = NULL;
2764 unsigned short mode_flags;
2766 char *nodelist = strchr(str, ':');
2767 char *flags = strchr(str, '=');
2771 /* NUL-terminate mode or flags string */
2773 if (nodelist_parse(nodelist, nodes))
2775 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2781 *flags++ = '\0'; /* terminate mode string */
2783 mode = match_string(policy_modes, MPOL_MAX, str);
2788 case MPOL_PREFERRED:
2790 * Insist on a nodelist of one node only
2793 char *rest = nodelist;
2794 while (isdigit(*rest))
2800 case MPOL_INTERLEAVE:
2802 * Default to online nodes with memory if no nodelist
2805 nodes = node_states[N_MEMORY];
2809 * Don't allow a nodelist; mpol_new() checks flags
2813 mode = MPOL_PREFERRED;
2817 * Insist on a empty nodelist
2824 * Insist on a nodelist
2833 * Currently, we only support two mutually exclusive
2836 if (!strcmp(flags, "static"))
2837 mode_flags |= MPOL_F_STATIC_NODES;
2838 else if (!strcmp(flags, "relative"))
2839 mode_flags |= MPOL_F_RELATIVE_NODES;
2844 new = mpol_new(mode, mode_flags, &nodes);
2849 * Save nodes for mpol_to_str() to show the tmpfs mount options
2850 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2852 if (mode != MPOL_PREFERRED)
2853 new->v.nodes = nodes;
2855 new->v.preferred_node = first_node(nodes);
2857 new->flags |= MPOL_F_LOCAL;
2860 * Save nodes for contextualization: this will be used to "clone"
2861 * the mempolicy in a specific context [cpuset] at a later time.
2863 new->w.user_nodemask = nodes;
2868 /* Restore string for error message */
2877 #endif /* CONFIG_TMPFS */
2880 * mpol_to_str - format a mempolicy structure for printing
2881 * @buffer: to contain formatted mempolicy string
2882 * @maxlen: length of @buffer
2883 * @pol: pointer to mempolicy to be formatted
2885 * Convert @pol into a string. If @buffer is too short, truncate the string.
2886 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2887 * longest flag, "relative", and to display at least a few node ids.
2889 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2892 nodemask_t nodes = NODE_MASK_NONE;
2893 unsigned short mode = MPOL_DEFAULT;
2894 unsigned short flags = 0;
2896 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2904 case MPOL_PREFERRED:
2905 if (flags & MPOL_F_LOCAL)
2908 node_set(pol->v.preferred_node, nodes);
2911 case MPOL_INTERLEAVE:
2912 nodes = pol->v.nodes;
2916 snprintf(p, maxlen, "unknown");
2920 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2922 if (flags & MPOL_MODE_FLAGS) {
2923 p += snprintf(p, buffer + maxlen - p, "=");
2926 * Currently, the only defined flags are mutually exclusive
2928 if (flags & MPOL_F_STATIC_NODES)
2929 p += snprintf(p, buffer + maxlen - p, "static");
2930 else if (flags & MPOL_F_RELATIVE_NODES)
2931 p += snprintf(p, buffer + maxlen - p, "relative");
2934 if (!nodes_empty(nodes))
2935 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2936 nodemask_pr_args(&nodes));