2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/sched/mm.h>
77 #include <linux/sched/numa_balancing.h>
78 #include <linux/sched/task.h>
79 #include <linux/nodemask.h>
80 #include <linux/cpuset.h>
81 #include <linux/slab.h>
82 #include <linux/string.h>
83 #include <linux/export.h>
84 #include <linux/nsproxy.h>
85 #include <linux/interrupt.h>
86 #include <linux/init.h>
87 #include <linux/compat.h>
88 #include <linux/ptrace.h>
89 #include <linux/swap.h>
90 #include <linux/seq_file.h>
91 #include <linux/proc_fs.h>
92 #include <linux/migrate.h>
93 #include <linux/ksm.h>
94 #include <linux/rmap.h>
95 #include <linux/security.h>
96 #include <linux/syscalls.h>
97 #include <linux/ctype.h>
98 #include <linux/mm_inline.h>
99 #include <linux/mmu_notifier.h>
100 #include <linux/printk.h>
101 #include <linux/swapops.h>
103 #include <asm/tlbflush.h>
104 #include <linux/uaccess.h>
106 #include "internal.h"
109 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
110 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
112 static struct kmem_cache *policy_cache;
113 static struct kmem_cache *sn_cache;
115 /* Highest zone. An specific allocation for a zone below that is not
117 enum zone_type policy_zone = 0;
120 * run-time system-wide default policy => local allocation
122 static struct mempolicy default_policy = {
123 .refcnt = ATOMIC_INIT(1), /* never free it */
124 .mode = MPOL_PREFERRED,
125 .flags = MPOL_F_LOCAL,
128 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
130 struct mempolicy *get_task_policy(struct task_struct *p)
132 struct mempolicy *pol = p->mempolicy;
138 node = numa_node_id();
139 if (node != NUMA_NO_NODE) {
140 pol = &preferred_node_policy[node];
141 /* preferred_node_policy is not initialised early in boot */
146 return &default_policy;
149 static const struct mempolicy_operations {
150 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
151 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
152 } mpol_ops[MPOL_MAX];
154 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
156 return pol->flags & MPOL_MODE_FLAGS;
159 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
160 const nodemask_t *rel)
163 nodes_fold(tmp, *orig, nodes_weight(*rel));
164 nodes_onto(*ret, tmp, *rel);
167 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
169 if (nodes_empty(*nodes))
171 pol->v.nodes = *nodes;
175 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
178 pol->flags |= MPOL_F_LOCAL; /* local allocation */
179 else if (nodes_empty(*nodes))
180 return -EINVAL; /* no allowed nodes */
182 pol->v.preferred_node = first_node(*nodes);
186 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
188 if (nodes_empty(*nodes))
190 pol->v.nodes = *nodes;
195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
196 * any, for the new policy. mpol_new() has already validated the nodes
197 * parameter with respect to the policy mode and flags. But, we need to
198 * handle an empty nodemask with MPOL_PREFERRED here.
200 * Must be called holding task's alloc_lock to protect task's mems_allowed
201 * and mempolicy. May also be called holding the mmap_semaphore for write.
203 static int mpol_set_nodemask(struct mempolicy *pol,
204 const nodemask_t *nodes, struct nodemask_scratch *nsc)
208 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
212 nodes_and(nsc->mask1,
213 cpuset_current_mems_allowed, node_states[N_MEMORY]);
216 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
217 nodes = NULL; /* explicit local allocation */
219 if (pol->flags & MPOL_F_RELATIVE_NODES)
220 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
222 nodes_and(nsc->mask2, *nodes, nsc->mask1);
224 if (mpol_store_user_nodemask(pol))
225 pol->w.user_nodemask = *nodes;
227 pol->w.cpuset_mems_allowed =
228 cpuset_current_mems_allowed;
232 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
234 ret = mpol_ops[pol->mode].create(pol, NULL);
239 * This function just creates a new policy, does some check and simple
240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
242 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
245 struct mempolicy *policy;
247 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
248 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
250 if (mode == MPOL_DEFAULT) {
251 if (nodes && !nodes_empty(*nodes))
252 return ERR_PTR(-EINVAL);
258 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
259 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
260 * All other modes require a valid pointer to a non-empty nodemask.
262 if (mode == MPOL_PREFERRED) {
263 if (nodes_empty(*nodes)) {
264 if (((flags & MPOL_F_STATIC_NODES) ||
265 (flags & MPOL_F_RELATIVE_NODES)))
266 return ERR_PTR(-EINVAL);
268 } else if (mode == MPOL_LOCAL) {
269 if (!nodes_empty(*nodes) ||
270 (flags & MPOL_F_STATIC_NODES) ||
271 (flags & MPOL_F_RELATIVE_NODES))
272 return ERR_PTR(-EINVAL);
273 mode = MPOL_PREFERRED;
274 } else if (nodes_empty(*nodes))
275 return ERR_PTR(-EINVAL);
276 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
278 return ERR_PTR(-ENOMEM);
279 atomic_set(&policy->refcnt, 1);
281 policy->flags = flags;
286 /* Slow path of a mpol destructor. */
287 void __mpol_put(struct mempolicy *p)
289 if (!atomic_dec_and_test(&p->refcnt))
291 kmem_cache_free(policy_cache, p);
294 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
298 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
302 if (pol->flags & MPOL_F_STATIC_NODES)
303 nodes_and(tmp, pol->w.user_nodemask, *nodes);
304 else if (pol->flags & MPOL_F_RELATIVE_NODES)
305 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
307 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
309 pol->w.cpuset_mems_allowed = tmp;
312 if (nodes_empty(tmp))
318 static void mpol_rebind_preferred(struct mempolicy *pol,
319 const nodemask_t *nodes)
323 if (pol->flags & MPOL_F_STATIC_NODES) {
324 int node = first_node(pol->w.user_nodemask);
326 if (node_isset(node, *nodes)) {
327 pol->v.preferred_node = node;
328 pol->flags &= ~MPOL_F_LOCAL;
330 pol->flags |= MPOL_F_LOCAL;
331 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
332 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
333 pol->v.preferred_node = first_node(tmp);
334 } else if (!(pol->flags & MPOL_F_LOCAL)) {
335 pol->v.preferred_node = node_remap(pol->v.preferred_node,
336 pol->w.cpuset_mems_allowed,
338 pol->w.cpuset_mems_allowed = *nodes;
343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
346 * policies are protected by task->mems_allowed_seq to prevent a premature
347 * OOM/allocation failure due to parallel nodemask modification.
349 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
353 if (!mpol_store_user_nodemask(pol) &&
354 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
357 mpol_ops[pol->mode].rebind(pol, newmask);
361 * Wrapper for mpol_rebind_policy() that just requires task
362 * pointer, and updates task mempolicy.
364 * Called with task's alloc_lock held.
367 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
369 mpol_rebind_policy(tsk->mempolicy, new);
373 * Rebind each vma in mm to new nodemask.
375 * Call holding a reference to mm. Takes mm->mmap_sem during call.
378 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
380 struct vm_area_struct *vma;
382 down_write(&mm->mmap_sem);
383 for (vma = mm->mmap; vma; vma = vma->vm_next)
384 mpol_rebind_policy(vma->vm_policy, new);
385 up_write(&mm->mmap_sem);
388 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
390 .rebind = mpol_rebind_default,
392 [MPOL_INTERLEAVE] = {
393 .create = mpol_new_interleave,
394 .rebind = mpol_rebind_nodemask,
397 .create = mpol_new_preferred,
398 .rebind = mpol_rebind_preferred,
401 .create = mpol_new_bind,
402 .rebind = mpol_rebind_nodemask,
406 static void migrate_page_add(struct page *page, struct list_head *pagelist,
407 unsigned long flags);
410 struct list_head *pagelist;
413 struct vm_area_struct *prev;
417 * Check if the page's nid is in qp->nmask.
419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
420 * in the invert of qp->nmask.
422 static inline bool queue_pages_required(struct page *page,
423 struct queue_pages *qp)
425 int nid = page_to_nid(page);
426 unsigned long flags = qp->flags;
428 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
431 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
432 unsigned long end, struct mm_walk *walk)
436 struct queue_pages *qp = walk->private;
439 if (unlikely(is_pmd_migration_entry(*pmd))) {
443 page = pmd_page(*pmd);
444 if (is_huge_zero_page(page)) {
446 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
449 if (!queue_pages_required(page, qp)) {
456 /* go to thp migration */
457 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
458 migrate_page_add(page, qp->pagelist, flags);
466 * Scan through pages checking if pages follow certain conditions,
467 * and move them to the pagelist if they do.
469 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
470 unsigned long end, struct mm_walk *walk)
472 struct vm_area_struct *vma = walk->vma;
474 struct queue_pages *qp = walk->private;
475 unsigned long flags = qp->flags;
480 ptl = pmd_trans_huge_lock(pmd, vma);
482 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
487 if (pmd_trans_unstable(pmd))
490 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
491 for (; addr != end; pte++, addr += PAGE_SIZE) {
492 if (!pte_present(*pte))
494 page = vm_normal_page(vma, addr, *pte);
498 * vm_normal_page() filters out zero pages, but there might
499 * still be PageReserved pages to skip, perhaps in a VDSO.
501 if (PageReserved(page))
503 if (!queue_pages_required(page, qp))
505 migrate_page_add(page, qp->pagelist, flags);
507 pte_unmap_unlock(pte - 1, ptl);
512 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
513 unsigned long addr, unsigned long end,
514 struct mm_walk *walk)
516 #ifdef CONFIG_HUGETLB_PAGE
517 struct queue_pages *qp = walk->private;
518 unsigned long flags = qp->flags;
523 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
524 entry = huge_ptep_get(pte);
525 if (!pte_present(entry))
527 page = pte_page(entry);
528 if (!queue_pages_required(page, qp))
530 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
531 if (flags & (MPOL_MF_MOVE_ALL) ||
532 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
533 isolate_huge_page(page, qp->pagelist);
542 #ifdef CONFIG_NUMA_BALANCING
544 * This is used to mark a range of virtual addresses to be inaccessible.
545 * These are later cleared by a NUMA hinting fault. Depending on these
546 * faults, pages may be migrated for better NUMA placement.
548 * This is assuming that NUMA faults are handled using PROT_NONE. If
549 * an architecture makes a different choice, it will need further
550 * changes to the core.
552 unsigned long change_prot_numa(struct vm_area_struct *vma,
553 unsigned long addr, unsigned long end)
557 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
559 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
564 static unsigned long change_prot_numa(struct vm_area_struct *vma,
565 unsigned long addr, unsigned long end)
569 #endif /* CONFIG_NUMA_BALANCING */
571 static int queue_pages_test_walk(unsigned long start, unsigned long end,
572 struct mm_walk *walk)
574 struct vm_area_struct *vma = walk->vma;
575 struct queue_pages *qp = walk->private;
576 unsigned long endvma = vma->vm_end;
577 unsigned long flags = qp->flags;
579 if (!vma_migratable(vma))
584 if (vma->vm_start > start)
585 start = vma->vm_start;
587 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
588 if (!vma->vm_next && vma->vm_end < end)
590 if (qp->prev && qp->prev->vm_end < vma->vm_start)
596 if (flags & MPOL_MF_LAZY) {
597 /* Similar to task_numa_work, skip inaccessible VMAs */
598 if (!is_vm_hugetlb_page(vma) &&
599 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
600 !(vma->vm_flags & VM_MIXEDMAP))
601 change_prot_numa(vma, start, endvma);
605 /* queue pages from current vma */
606 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
612 * Walk through page tables and collect pages to be migrated.
614 * If pages found in a given range are on a set of nodes (determined by
615 * @nodes and @flags,) it's isolated and queued to the pagelist which is
616 * passed via @private.)
619 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
620 nodemask_t *nodes, unsigned long flags,
621 struct list_head *pagelist)
623 struct queue_pages qp = {
624 .pagelist = pagelist,
629 struct mm_walk queue_pages_walk = {
630 .hugetlb_entry = queue_pages_hugetlb,
631 .pmd_entry = queue_pages_pte_range,
632 .test_walk = queue_pages_test_walk,
637 return walk_page_range(start, end, &queue_pages_walk);
641 * Apply policy to a single VMA
642 * This must be called with the mmap_sem held for writing.
644 static int vma_replace_policy(struct vm_area_struct *vma,
645 struct mempolicy *pol)
648 struct mempolicy *old;
649 struct mempolicy *new;
651 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
652 vma->vm_start, vma->vm_end, vma->vm_pgoff,
653 vma->vm_ops, vma->vm_file,
654 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
660 if (vma->vm_ops && vma->vm_ops->set_policy) {
661 err = vma->vm_ops->set_policy(vma, new);
666 old = vma->vm_policy;
667 vma->vm_policy = new; /* protected by mmap_sem */
676 /* Step 2: apply policy to a range and do splits. */
677 static int mbind_range(struct mm_struct *mm, unsigned long start,
678 unsigned long end, struct mempolicy *new_pol)
680 struct vm_area_struct *next;
681 struct vm_area_struct *prev;
682 struct vm_area_struct *vma;
685 unsigned long vmstart;
688 vma = find_vma(mm, start);
689 if (!vma || vma->vm_start > start)
693 if (start > vma->vm_start)
696 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
698 vmstart = max(start, vma->vm_start);
699 vmend = min(end, vma->vm_end);
701 if (mpol_equal(vma_policy(vma), new_pol))
704 pgoff = vma->vm_pgoff +
705 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
706 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
707 vma->anon_vma, vma->vm_file, pgoff,
708 new_pol, vma->vm_userfaultfd_ctx);
712 if (mpol_equal(vma_policy(vma), new_pol))
714 /* vma_merge() joined vma && vma->next, case 8 */
717 if (vma->vm_start != vmstart) {
718 err = split_vma(vma->vm_mm, vma, vmstart, 1);
722 if (vma->vm_end != vmend) {
723 err = split_vma(vma->vm_mm, vma, vmend, 0);
728 err = vma_replace_policy(vma, new_pol);
737 /* Set the process memory policy */
738 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
741 struct mempolicy *new, *old;
742 NODEMASK_SCRATCH(scratch);
748 new = mpol_new(mode, flags, nodes);
755 ret = mpol_set_nodemask(new, nodes, scratch);
757 task_unlock(current);
761 old = current->mempolicy;
762 current->mempolicy = new;
763 if (new && new->mode == MPOL_INTERLEAVE)
764 current->il_prev = MAX_NUMNODES-1;
765 task_unlock(current);
769 NODEMASK_SCRATCH_FREE(scratch);
774 * Return nodemask for policy for get_mempolicy() query
776 * Called with task's alloc_lock held
778 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
781 if (p == &default_policy)
787 case MPOL_INTERLEAVE:
791 if (!(p->flags & MPOL_F_LOCAL))
792 node_set(p->v.preferred_node, *nodes);
793 /* else return empty node mask for local allocation */
800 static int lookup_node(unsigned long addr)
805 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
807 err = page_to_nid(p);
813 /* Retrieve NUMA policy */
814 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
815 unsigned long addr, unsigned long flags)
818 struct mm_struct *mm = current->mm;
819 struct vm_area_struct *vma = NULL;
820 struct mempolicy *pol = current->mempolicy;
823 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
826 if (flags & MPOL_F_MEMS_ALLOWED) {
827 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
829 *policy = 0; /* just so it's initialized */
831 *nmask = cpuset_current_mems_allowed;
832 task_unlock(current);
836 if (flags & MPOL_F_ADDR) {
838 * Do NOT fall back to task policy if the
839 * vma/shared policy at addr is NULL. We
840 * want to return MPOL_DEFAULT in this case.
842 down_read(&mm->mmap_sem);
843 vma = find_vma_intersection(mm, addr, addr+1);
845 up_read(&mm->mmap_sem);
848 if (vma->vm_ops && vma->vm_ops->get_policy)
849 pol = vma->vm_ops->get_policy(vma, addr);
851 pol = vma->vm_policy;
856 pol = &default_policy; /* indicates default behavior */
858 if (flags & MPOL_F_NODE) {
859 if (flags & MPOL_F_ADDR) {
860 err = lookup_node(addr);
864 } else if (pol == current->mempolicy &&
865 pol->mode == MPOL_INTERLEAVE) {
866 *policy = next_node_in(current->il_prev, pol->v.nodes);
872 *policy = pol == &default_policy ? MPOL_DEFAULT :
875 * Internal mempolicy flags must be masked off before exposing
876 * the policy to userspace.
878 *policy |= (pol->flags & MPOL_MODE_FLAGS);
883 if (mpol_store_user_nodemask(pol)) {
884 *nmask = pol->w.user_nodemask;
887 get_policy_nodemask(pol, nmask);
888 task_unlock(current);
895 up_read(¤t->mm->mmap_sem);
899 #ifdef CONFIG_MIGRATION
901 * page migration, thp tail pages can be passed.
903 static void migrate_page_add(struct page *page, struct list_head *pagelist,
906 struct page *head = compound_head(page);
908 * Avoid migrating a page that is shared with others.
910 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
911 if (!isolate_lru_page(head)) {
912 list_add_tail(&head->lru, pagelist);
913 mod_node_page_state(page_pgdat(head),
914 NR_ISOLATED_ANON + page_is_file_cache(head),
915 hpage_nr_pages(head));
920 /* page allocation callback for NUMA node migration */
921 struct page *alloc_new_node_page(struct page *page, unsigned long node)
924 return alloc_huge_page_node(page_hstate(compound_head(page)),
926 else if (PageTransHuge(page)) {
929 thp = alloc_pages_node(node,
930 (GFP_TRANSHUGE | __GFP_THISNODE),
934 prep_transhuge_page(thp);
937 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
942 * Migrate pages from one node to a target node.
943 * Returns error or the number of pages not migrated.
945 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
953 node_set(source, nmask);
956 * This does not "check" the range but isolates all pages that
957 * need migration. Between passing in the full user address
958 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
960 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
961 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
962 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
964 if (!list_empty(&pagelist)) {
965 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
966 MIGRATE_SYNC, MR_SYSCALL);
968 putback_movable_pages(&pagelist);
975 * Move pages between the two nodesets so as to preserve the physical
976 * layout as much as possible.
978 * Returns the number of page that could not be moved.
980 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
981 const nodemask_t *to, int flags)
987 err = migrate_prep();
991 down_read(&mm->mmap_sem);
994 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
995 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
996 * bit in 'tmp', and return that <source, dest> pair for migration.
997 * The pair of nodemasks 'to' and 'from' define the map.
999 * If no pair of bits is found that way, fallback to picking some
1000 * pair of 'source' and 'dest' bits that are not the same. If the
1001 * 'source' and 'dest' bits are the same, this represents a node
1002 * that will be migrating to itself, so no pages need move.
1004 * If no bits are left in 'tmp', or if all remaining bits left
1005 * in 'tmp' correspond to the same bit in 'to', return false
1006 * (nothing left to migrate).
1008 * This lets us pick a pair of nodes to migrate between, such that
1009 * if possible the dest node is not already occupied by some other
1010 * source node, minimizing the risk of overloading the memory on a
1011 * node that would happen if we migrated incoming memory to a node
1012 * before migrating outgoing memory source that same node.
1014 * A single scan of tmp is sufficient. As we go, we remember the
1015 * most recent <s, d> pair that moved (s != d). If we find a pair
1016 * that not only moved, but what's better, moved to an empty slot
1017 * (d is not set in tmp), then we break out then, with that pair.
1018 * Otherwise when we finish scanning from_tmp, we at least have the
1019 * most recent <s, d> pair that moved. If we get all the way through
1020 * the scan of tmp without finding any node that moved, much less
1021 * moved to an empty node, then there is nothing left worth migrating.
1025 while (!nodes_empty(tmp)) {
1027 int source = NUMA_NO_NODE;
1030 for_each_node_mask(s, tmp) {
1033 * do_migrate_pages() tries to maintain the relative
1034 * node relationship of the pages established between
1035 * threads and memory areas.
1037 * However if the number of source nodes is not equal to
1038 * the number of destination nodes we can not preserve
1039 * this node relative relationship. In that case, skip
1040 * copying memory from a node that is in the destination
1043 * Example: [2,3,4] -> [3,4,5] moves everything.
1044 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1047 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1048 (node_isset(s, *to)))
1051 d = node_remap(s, *from, *to);
1055 source = s; /* Node moved. Memorize */
1058 /* dest not in remaining from nodes? */
1059 if (!node_isset(dest, tmp))
1062 if (source == NUMA_NO_NODE)
1065 node_clear(source, tmp);
1066 err = migrate_to_node(mm, source, dest, flags);
1072 up_read(&mm->mmap_sem);
1080 * Allocate a new page for page migration based on vma policy.
1081 * Start by assuming the page is mapped by the same vma as contains @start.
1082 * Search forward from there, if not. N.B., this assumes that the
1083 * list of pages handed to migrate_pages()--which is how we get here--
1084 * is in virtual address order.
1086 static struct page *new_page(struct page *page, unsigned long start)
1088 struct vm_area_struct *vma;
1089 unsigned long uninitialized_var(address);
1091 vma = find_vma(current->mm, start);
1093 address = page_address_in_vma(page, vma);
1094 if (address != -EFAULT)
1099 if (PageHuge(page)) {
1100 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1102 } else if (PageTransHuge(page)) {
1105 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1109 prep_transhuge_page(thp);
1113 * if !vma, alloc_page_vma() will use task or system default policy
1115 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1120 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1121 unsigned long flags)
1125 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1126 const nodemask_t *to, int flags)
1131 static struct page *new_page(struct page *page, unsigned long start)
1137 static long do_mbind(unsigned long start, unsigned long len,
1138 unsigned short mode, unsigned short mode_flags,
1139 nodemask_t *nmask, unsigned long flags)
1141 struct mm_struct *mm = current->mm;
1142 struct mempolicy *new;
1145 LIST_HEAD(pagelist);
1147 if (flags & ~(unsigned long)MPOL_MF_VALID)
1149 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1152 if (start & ~PAGE_MASK)
1155 if (mode == MPOL_DEFAULT)
1156 flags &= ~MPOL_MF_STRICT;
1158 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1166 new = mpol_new(mode, mode_flags, nmask);
1168 return PTR_ERR(new);
1170 if (flags & MPOL_MF_LAZY)
1171 new->flags |= MPOL_F_MOF;
1174 * If we are using the default policy then operation
1175 * on discontinuous address spaces is okay after all
1178 flags |= MPOL_MF_DISCONTIG_OK;
1180 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1181 start, start + len, mode, mode_flags,
1182 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1184 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1186 err = migrate_prep();
1191 NODEMASK_SCRATCH(scratch);
1193 down_write(&mm->mmap_sem);
1195 err = mpol_set_nodemask(new, nmask, scratch);
1196 task_unlock(current);
1198 up_write(&mm->mmap_sem);
1201 NODEMASK_SCRATCH_FREE(scratch);
1206 err = queue_pages_range(mm, start, end, nmask,
1207 flags | MPOL_MF_INVERT, &pagelist);
1209 err = mbind_range(mm, start, end, new);
1214 if (!list_empty(&pagelist)) {
1215 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1216 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1217 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1219 putback_movable_pages(&pagelist);
1222 if (nr_failed && (flags & MPOL_MF_STRICT))
1225 putback_movable_pages(&pagelist);
1227 up_write(&mm->mmap_sem);
1234 * User space interface with variable sized bitmaps for nodelists.
1237 /* Copy a node mask from user space. */
1238 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1239 unsigned long maxnode)
1243 unsigned long nlongs;
1244 unsigned long endmask;
1247 nodes_clear(*nodes);
1248 if (maxnode == 0 || !nmask)
1250 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1253 nlongs = BITS_TO_LONGS(maxnode);
1254 if ((maxnode % BITS_PER_LONG) == 0)
1257 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1260 * When the user specified more nodes than supported just check
1261 * if the non supported part is all zero.
1263 * If maxnode have more longs than MAX_NUMNODES, check
1264 * the bits in that area first. And then go through to
1265 * check the rest bits which equal or bigger than MAX_NUMNODES.
1266 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1268 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1269 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1270 if (get_user(t, nmask + k))
1272 if (k == nlongs - 1) {
1278 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1282 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1283 unsigned long valid_mask = endmask;
1285 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1286 if (get_user(t, nmask + nlongs - 1))
1292 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1294 nodes_addr(*nodes)[nlongs-1] &= endmask;
1298 /* Copy a kernel node mask to user space */
1299 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1302 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1303 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1305 if (copy > nbytes) {
1306 if (copy > PAGE_SIZE)
1308 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1312 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1315 static long kernel_mbind(unsigned long start, unsigned long len,
1316 unsigned long mode, const unsigned long __user *nmask,
1317 unsigned long maxnode, unsigned int flags)
1321 unsigned short mode_flags;
1323 mode_flags = mode & MPOL_MODE_FLAGS;
1324 mode &= ~MPOL_MODE_FLAGS;
1325 if (mode >= MPOL_MAX)
1327 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1328 (mode_flags & MPOL_F_RELATIVE_NODES))
1330 err = get_nodes(&nodes, nmask, maxnode);
1333 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1336 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1337 unsigned long, mode, const unsigned long __user *, nmask,
1338 unsigned long, maxnode, unsigned int, flags)
1340 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1343 /* Set the process memory policy */
1344 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1345 unsigned long maxnode)
1349 unsigned short flags;
1351 flags = mode & MPOL_MODE_FLAGS;
1352 mode &= ~MPOL_MODE_FLAGS;
1353 if ((unsigned int)mode >= MPOL_MAX)
1355 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1357 err = get_nodes(&nodes, nmask, maxnode);
1360 return do_set_mempolicy(mode, flags, &nodes);
1363 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1364 unsigned long, maxnode)
1366 return kernel_set_mempolicy(mode, nmask, maxnode);
1369 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1370 const unsigned long __user *old_nodes,
1371 const unsigned long __user *new_nodes)
1373 struct mm_struct *mm = NULL;
1374 struct task_struct *task;
1375 nodemask_t task_nodes;
1379 NODEMASK_SCRATCH(scratch);
1384 old = &scratch->mask1;
1385 new = &scratch->mask2;
1387 err = get_nodes(old, old_nodes, maxnode);
1391 err = get_nodes(new, new_nodes, maxnode);
1395 /* Find the mm_struct */
1397 task = pid ? find_task_by_vpid(pid) : current;
1403 get_task_struct(task);
1408 * Check if this process has the right to modify the specified process.
1409 * Use the regular "ptrace_may_access()" checks.
1411 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1418 task_nodes = cpuset_mems_allowed(task);
1419 /* Is the user allowed to access the target nodes? */
1420 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1425 task_nodes = cpuset_mems_allowed(current);
1426 nodes_and(*new, *new, task_nodes);
1427 if (nodes_empty(*new))
1430 nodes_and(*new, *new, node_states[N_MEMORY]);
1431 if (nodes_empty(*new))
1434 err = security_task_movememory(task);
1438 mm = get_task_mm(task);
1439 put_task_struct(task);
1446 err = do_migrate_pages(mm, old, new,
1447 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1451 NODEMASK_SCRATCH_FREE(scratch);
1456 put_task_struct(task);
1461 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1462 const unsigned long __user *, old_nodes,
1463 const unsigned long __user *, new_nodes)
1465 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1469 /* Retrieve NUMA policy */
1470 static int kernel_get_mempolicy(int __user *policy,
1471 unsigned long __user *nmask,
1472 unsigned long maxnode,
1474 unsigned long flags)
1477 int uninitialized_var(pval);
1480 if (nmask != NULL && maxnode < MAX_NUMNODES)
1483 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1488 if (policy && put_user(pval, policy))
1492 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1497 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1498 unsigned long __user *, nmask, unsigned long, maxnode,
1499 unsigned long, addr, unsigned long, flags)
1501 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1504 #ifdef CONFIG_COMPAT
1506 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1507 compat_ulong_t __user *, nmask,
1508 compat_ulong_t, maxnode,
1509 compat_ulong_t, addr, compat_ulong_t, flags)
1512 unsigned long __user *nm = NULL;
1513 unsigned long nr_bits, alloc_size;
1514 DECLARE_BITMAP(bm, MAX_NUMNODES);
1516 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1517 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1520 nm = compat_alloc_user_space(alloc_size);
1522 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1524 if (!err && nmask) {
1525 unsigned long copy_size;
1526 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1527 err = copy_from_user(bm, nm, copy_size);
1528 /* ensure entire bitmap is zeroed */
1529 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1530 err |= compat_put_bitmap(nmask, bm, nr_bits);
1536 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1537 compat_ulong_t, maxnode)
1539 unsigned long __user *nm = NULL;
1540 unsigned long nr_bits, alloc_size;
1541 DECLARE_BITMAP(bm, MAX_NUMNODES);
1543 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1544 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1547 if (compat_get_bitmap(bm, nmask, nr_bits))
1549 nm = compat_alloc_user_space(alloc_size);
1550 if (copy_to_user(nm, bm, alloc_size))
1554 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1557 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1558 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1559 compat_ulong_t, maxnode, compat_ulong_t, flags)
1561 unsigned long __user *nm = NULL;
1562 unsigned long nr_bits, alloc_size;
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(nodes_addr(bm), nmask, nr_bits))
1571 nm = compat_alloc_user_space(alloc_size);
1572 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1576 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1579 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1580 compat_ulong_t, maxnode,
1581 const compat_ulong_t __user *, old_nodes,
1582 const compat_ulong_t __user *, new_nodes)
1584 unsigned long __user *old = NULL;
1585 unsigned long __user *new = NULL;
1586 nodemask_t tmp_mask;
1587 unsigned long nr_bits;
1590 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1591 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1593 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1595 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1597 new = old + size / sizeof(unsigned long);
1598 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1602 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1605 new = compat_alloc_user_space(size);
1606 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1609 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1612 #endif /* CONFIG_COMPAT */
1614 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1617 struct mempolicy *pol = NULL;
1620 if (vma->vm_ops && vma->vm_ops->get_policy) {
1621 pol = vma->vm_ops->get_policy(vma, addr);
1622 } else if (vma->vm_policy) {
1623 pol = vma->vm_policy;
1626 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1627 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1628 * count on these policies which will be dropped by
1629 * mpol_cond_put() later
1631 if (mpol_needs_cond_ref(pol))
1640 * get_vma_policy(@vma, @addr)
1641 * @vma: virtual memory area whose policy is sought
1642 * @addr: address in @vma for shared policy lookup
1644 * Returns effective policy for a VMA at specified address.
1645 * Falls back to current->mempolicy or system default policy, as necessary.
1646 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1647 * count--added by the get_policy() vm_op, as appropriate--to protect against
1648 * freeing by another task. It is the caller's responsibility to free the
1649 * extra reference for shared policies.
1651 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1654 struct mempolicy *pol = __get_vma_policy(vma, addr);
1657 pol = get_task_policy(current);
1662 bool vma_policy_mof(struct vm_area_struct *vma)
1664 struct mempolicy *pol;
1666 if (vma->vm_ops && vma->vm_ops->get_policy) {
1669 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1670 if (pol && (pol->flags & MPOL_F_MOF))
1677 pol = vma->vm_policy;
1679 pol = get_task_policy(current);
1681 return pol->flags & MPOL_F_MOF;
1684 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1686 enum zone_type dynamic_policy_zone = policy_zone;
1688 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1691 * if policy->v.nodes has movable memory only,
1692 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1694 * policy->v.nodes is intersect with node_states[N_MEMORY].
1695 * so if the following test faile, it implies
1696 * policy->v.nodes has movable memory only.
1698 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1699 dynamic_policy_zone = ZONE_MOVABLE;
1701 return zone >= dynamic_policy_zone;
1705 * Return a nodemask representing a mempolicy for filtering nodes for
1708 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1710 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1711 if (unlikely(policy->mode == MPOL_BIND) &&
1712 apply_policy_zone(policy, gfp_zone(gfp)) &&
1713 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1714 return &policy->v.nodes;
1719 /* Return the node id preferred by the given mempolicy, or the given id */
1720 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1723 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1724 nd = policy->v.preferred_node;
1727 * __GFP_THISNODE shouldn't even be used with the bind policy
1728 * because we might easily break the expectation to stay on the
1729 * requested node and not break the policy.
1731 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1737 /* Do dynamic interleaving for a process */
1738 static unsigned interleave_nodes(struct mempolicy *policy)
1741 struct task_struct *me = current;
1743 next = next_node_in(me->il_prev, policy->v.nodes);
1744 if (next < MAX_NUMNODES)
1750 * Depending on the memory policy provide a node from which to allocate the
1753 unsigned int mempolicy_slab_node(void)
1755 struct mempolicy *policy;
1756 int node = numa_mem_id();
1761 policy = current->mempolicy;
1762 if (!policy || policy->flags & MPOL_F_LOCAL)
1765 switch (policy->mode) {
1766 case MPOL_PREFERRED:
1768 * handled MPOL_F_LOCAL above
1770 return policy->v.preferred_node;
1772 case MPOL_INTERLEAVE:
1773 return interleave_nodes(policy);
1779 * Follow bind policy behavior and start allocation at the
1782 struct zonelist *zonelist;
1783 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1784 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1785 z = first_zones_zonelist(zonelist, highest_zoneidx,
1787 return z->zone ? zone_to_nid(z->zone) : node;
1796 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1797 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1798 * number of present nodes.
1800 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1802 unsigned nnodes = nodes_weight(pol->v.nodes);
1808 return numa_node_id();
1809 target = (unsigned int)n % nnodes;
1810 nid = first_node(pol->v.nodes);
1811 for (i = 0; i < target; i++)
1812 nid = next_node(nid, pol->v.nodes);
1816 /* Determine a node number for interleave */
1817 static inline unsigned interleave_nid(struct mempolicy *pol,
1818 struct vm_area_struct *vma, unsigned long addr, int shift)
1824 * for small pages, there is no difference between
1825 * shift and PAGE_SHIFT, so the bit-shift is safe.
1826 * for huge pages, since vm_pgoff is in units of small
1827 * pages, we need to shift off the always 0 bits to get
1830 BUG_ON(shift < PAGE_SHIFT);
1831 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1832 off += (addr - vma->vm_start) >> shift;
1833 return offset_il_node(pol, off);
1835 return interleave_nodes(pol);
1838 #ifdef CONFIG_HUGETLBFS
1840 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1841 * @vma: virtual memory area whose policy is sought
1842 * @addr: address in @vma for shared policy lookup and interleave policy
1843 * @gfp_flags: for requested zone
1844 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1845 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1847 * Returns a nid suitable for a huge page allocation and a pointer
1848 * to the struct mempolicy for conditional unref after allocation.
1849 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1850 * @nodemask for filtering the zonelist.
1852 * Must be protected by read_mems_allowed_begin()
1854 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1855 struct mempolicy **mpol, nodemask_t **nodemask)
1859 *mpol = get_vma_policy(vma, addr);
1860 *nodemask = NULL; /* assume !MPOL_BIND */
1862 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1863 nid = interleave_nid(*mpol, vma, addr,
1864 huge_page_shift(hstate_vma(vma)));
1866 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1867 if ((*mpol)->mode == MPOL_BIND)
1868 *nodemask = &(*mpol)->v.nodes;
1874 * init_nodemask_of_mempolicy
1876 * If the current task's mempolicy is "default" [NULL], return 'false'
1877 * to indicate default policy. Otherwise, extract the policy nodemask
1878 * for 'bind' or 'interleave' policy into the argument nodemask, or
1879 * initialize the argument nodemask to contain the single node for
1880 * 'preferred' or 'local' policy and return 'true' to indicate presence
1881 * of non-default mempolicy.
1883 * We don't bother with reference counting the mempolicy [mpol_get/put]
1884 * because the current task is examining it's own mempolicy and a task's
1885 * mempolicy is only ever changed by the task itself.
1887 * N.B., it is the caller's responsibility to free a returned nodemask.
1889 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1891 struct mempolicy *mempolicy;
1894 if (!(mask && current->mempolicy))
1898 mempolicy = current->mempolicy;
1899 switch (mempolicy->mode) {
1900 case MPOL_PREFERRED:
1901 if (mempolicy->flags & MPOL_F_LOCAL)
1902 nid = numa_node_id();
1904 nid = mempolicy->v.preferred_node;
1905 init_nodemask_of_node(mask, nid);
1910 case MPOL_INTERLEAVE:
1911 *mask = mempolicy->v.nodes;
1917 task_unlock(current);
1924 * mempolicy_nodemask_intersects
1926 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1927 * policy. Otherwise, check for intersection between mask and the policy
1928 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1929 * policy, always return true since it may allocate elsewhere on fallback.
1931 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1933 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1934 const nodemask_t *mask)
1936 struct mempolicy *mempolicy;
1942 mempolicy = tsk->mempolicy;
1946 switch (mempolicy->mode) {
1947 case MPOL_PREFERRED:
1949 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1950 * allocate from, they may fallback to other nodes when oom.
1951 * Thus, it's possible for tsk to have allocated memory from
1956 case MPOL_INTERLEAVE:
1957 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1967 /* Allocate a page in interleaved policy.
1968 Own path because it needs to do special accounting. */
1969 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1974 page = __alloc_pages(gfp, order, nid);
1975 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1976 if (!static_branch_likely(&vm_numa_stat_key))
1978 if (page && page_to_nid(page) == nid) {
1980 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1987 * alloc_pages_vma - Allocate a page for a VMA.
1990 * %GFP_USER user allocation.
1991 * %GFP_KERNEL kernel allocations,
1992 * %GFP_HIGHMEM highmem/user allocations,
1993 * %GFP_FS allocation should not call back into a file system.
1994 * %GFP_ATOMIC don't sleep.
1996 * @order:Order of the GFP allocation.
1997 * @vma: Pointer to VMA or NULL if not available.
1998 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1999 * @node: Which node to prefer for allocation (modulo policy).
2000 * @hugepage: for hugepages try only the preferred node if possible
2002 * This function allocates a page from the kernel page pool and applies
2003 * a NUMA policy associated with the VMA or the current process.
2004 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2005 * mm_struct of the VMA to prevent it from going away. Should be used for
2006 * all allocations for pages that will be mapped into user space. Returns
2007 * NULL when no page can be allocated.
2010 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2011 unsigned long addr, int node, bool hugepage)
2013 struct mempolicy *pol;
2018 pol = get_vma_policy(vma, addr);
2020 if (pol->mode == MPOL_INTERLEAVE) {
2023 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2025 page = alloc_page_interleave(gfp, order, nid);
2029 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2030 int hpage_node = node;
2033 * For hugepage allocation and non-interleave policy which
2034 * allows the current node (or other explicitly preferred
2035 * node) we only try to allocate from the current/preferred
2036 * node and don't fall back to other nodes, as the cost of
2037 * remote accesses would likely offset THP benefits.
2039 * If the policy is interleave, or does not allow the current
2040 * node in its nodemask, we allocate the standard way.
2042 if (pol->mode == MPOL_PREFERRED &&
2043 !(pol->flags & MPOL_F_LOCAL))
2044 hpage_node = pol->v.preferred_node;
2046 nmask = policy_nodemask(gfp, pol);
2047 if (!nmask || node_isset(hpage_node, *nmask)) {
2050 * We cannot invoke reclaim if __GFP_THISNODE
2051 * is set. Invoking reclaim with
2052 * __GFP_THISNODE set, would cause THP
2053 * allocations to trigger heavy swapping
2054 * despite there may be tons of free memory
2055 * (including potentially plenty of THP
2056 * already available in the buddy) on all the
2059 * At most we could invoke compaction when
2060 * __GFP_THISNODE is set (but we would need to
2061 * refrain from invoking reclaim even if
2062 * compaction returned COMPACT_SKIPPED because
2063 * there wasn't not enough memory to succeed
2064 * compaction). For now just avoid
2065 * __GFP_THISNODE instead of limiting the
2066 * allocation path to a strict and single
2067 * compaction invocation.
2069 * Supposedly if direct reclaim was enabled by
2070 * the caller, the app prefers THP regardless
2071 * of the node it comes from so this would be
2072 * more desiderable behavior than only
2073 * providing THP originated from the local
2074 * node in such case.
2076 if (!(gfp & __GFP_DIRECT_RECLAIM))
2077 gfp |= __GFP_THISNODE;
2078 page = __alloc_pages_node(hpage_node, gfp, order);
2083 nmask = policy_nodemask(gfp, pol);
2084 preferred_nid = policy_node(gfp, pol, node);
2085 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2092 * alloc_pages_current - Allocate pages.
2095 * %GFP_USER user allocation,
2096 * %GFP_KERNEL kernel allocation,
2097 * %GFP_HIGHMEM highmem allocation,
2098 * %GFP_FS don't call back into a file system.
2099 * %GFP_ATOMIC don't sleep.
2100 * @order: Power of two of allocation size in pages. 0 is a single page.
2102 * Allocate a page from the kernel page pool. When not in
2103 * interrupt context and apply the current process NUMA policy.
2104 * Returns NULL when no page can be allocated.
2106 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2108 struct mempolicy *pol = &default_policy;
2111 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2112 pol = get_task_policy(current);
2115 * No reference counting needed for current->mempolicy
2116 * nor system default_policy
2118 if (pol->mode == MPOL_INTERLEAVE)
2119 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2121 page = __alloc_pages_nodemask(gfp, order,
2122 policy_node(gfp, pol, numa_node_id()),
2123 policy_nodemask(gfp, pol));
2127 EXPORT_SYMBOL(alloc_pages_current);
2129 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2131 struct mempolicy *pol = mpol_dup(vma_policy(src));
2134 return PTR_ERR(pol);
2135 dst->vm_policy = pol;
2140 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2141 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2142 * with the mems_allowed returned by cpuset_mems_allowed(). This
2143 * keeps mempolicies cpuset relative after its cpuset moves. See
2144 * further kernel/cpuset.c update_nodemask().
2146 * current's mempolicy may be rebinded by the other task(the task that changes
2147 * cpuset's mems), so we needn't do rebind work for current task.
2150 /* Slow path of a mempolicy duplicate */
2151 struct mempolicy *__mpol_dup(struct mempolicy *old)
2153 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2156 return ERR_PTR(-ENOMEM);
2158 /* task's mempolicy is protected by alloc_lock */
2159 if (old == current->mempolicy) {
2162 task_unlock(current);
2166 if (current_cpuset_is_being_rebound()) {
2167 nodemask_t mems = cpuset_mems_allowed(current);
2168 mpol_rebind_policy(new, &mems);
2170 atomic_set(&new->refcnt, 1);
2174 /* Slow path of a mempolicy comparison */
2175 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2179 if (a->mode != b->mode)
2181 if (a->flags != b->flags)
2183 if (mpol_store_user_nodemask(a))
2184 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2190 case MPOL_INTERLEAVE:
2191 return !!nodes_equal(a->v.nodes, b->v.nodes);
2192 case MPOL_PREFERRED:
2193 /* a's ->flags is the same as b's */
2194 if (a->flags & MPOL_F_LOCAL)
2196 return a->v.preferred_node == b->v.preferred_node;
2204 * Shared memory backing store policy support.
2206 * Remember policies even when nobody has shared memory mapped.
2207 * The policies are kept in Red-Black tree linked from the inode.
2208 * They are protected by the sp->lock rwlock, which should be held
2209 * for any accesses to the tree.
2213 * lookup first element intersecting start-end. Caller holds sp->lock for
2214 * reading or for writing
2216 static struct sp_node *
2217 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2219 struct rb_node *n = sp->root.rb_node;
2222 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2224 if (start >= p->end)
2226 else if (end <= p->start)
2234 struct sp_node *w = NULL;
2235 struct rb_node *prev = rb_prev(n);
2238 w = rb_entry(prev, struct sp_node, nd);
2239 if (w->end <= start)
2243 return rb_entry(n, struct sp_node, nd);
2247 * Insert a new shared policy into the list. Caller holds sp->lock for
2250 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2252 struct rb_node **p = &sp->root.rb_node;
2253 struct rb_node *parent = NULL;
2258 nd = rb_entry(parent, struct sp_node, nd);
2259 if (new->start < nd->start)
2261 else if (new->end > nd->end)
2262 p = &(*p)->rb_right;
2266 rb_link_node(&new->nd, parent, p);
2267 rb_insert_color(&new->nd, &sp->root);
2268 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2269 new->policy ? new->policy->mode : 0);
2272 /* Find shared policy intersecting idx */
2274 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2276 struct mempolicy *pol = NULL;
2279 if (!sp->root.rb_node)
2281 read_lock(&sp->lock);
2282 sn = sp_lookup(sp, idx, idx+1);
2284 mpol_get(sn->policy);
2287 read_unlock(&sp->lock);
2291 static void sp_free(struct sp_node *n)
2293 mpol_put(n->policy);
2294 kmem_cache_free(sn_cache, n);
2298 * mpol_misplaced - check whether current page node is valid in policy
2300 * @page: page to be checked
2301 * @vma: vm area where page mapped
2302 * @addr: virtual address where page mapped
2304 * Lookup current policy node id for vma,addr and "compare to" page's
2308 * -1 - not misplaced, page is in the right node
2309 * node - node id where the page should be
2311 * Policy determination "mimics" alloc_page_vma().
2312 * Called from fault path where we know the vma and faulting address.
2314 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2316 struct mempolicy *pol;
2318 int curnid = page_to_nid(page);
2319 unsigned long pgoff;
2320 int thiscpu = raw_smp_processor_id();
2321 int thisnid = cpu_to_node(thiscpu);
2325 pol = get_vma_policy(vma, addr);
2326 if (!(pol->flags & MPOL_F_MOF))
2329 switch (pol->mode) {
2330 case MPOL_INTERLEAVE:
2331 pgoff = vma->vm_pgoff;
2332 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2333 polnid = offset_il_node(pol, pgoff);
2336 case MPOL_PREFERRED:
2337 if (pol->flags & MPOL_F_LOCAL)
2338 polnid = numa_node_id();
2340 polnid = pol->v.preferred_node;
2346 * allows binding to multiple nodes.
2347 * use current page if in policy nodemask,
2348 * else select nearest allowed node, if any.
2349 * If no allowed nodes, use current [!misplaced].
2351 if (node_isset(curnid, pol->v.nodes))
2353 z = first_zones_zonelist(
2354 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2355 gfp_zone(GFP_HIGHUSER),
2357 polnid = zone_to_nid(z->zone);
2364 /* Migrate the page towards the node whose CPU is referencing it */
2365 if (pol->flags & MPOL_F_MORON) {
2368 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2372 if (curnid != polnid)
2381 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2382 * dropped after task->mempolicy is set to NULL so that any allocation done as
2383 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2386 void mpol_put_task_policy(struct task_struct *task)
2388 struct mempolicy *pol;
2391 pol = task->mempolicy;
2392 task->mempolicy = NULL;
2397 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2399 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2400 rb_erase(&n->nd, &sp->root);
2404 static void sp_node_init(struct sp_node *node, unsigned long start,
2405 unsigned long end, struct mempolicy *pol)
2407 node->start = start;
2412 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2413 struct mempolicy *pol)
2416 struct mempolicy *newpol;
2418 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2422 newpol = mpol_dup(pol);
2423 if (IS_ERR(newpol)) {
2424 kmem_cache_free(sn_cache, n);
2427 newpol->flags |= MPOL_F_SHARED;
2428 sp_node_init(n, start, end, newpol);
2433 /* Replace a policy range. */
2434 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2435 unsigned long end, struct sp_node *new)
2438 struct sp_node *n_new = NULL;
2439 struct mempolicy *mpol_new = NULL;
2443 write_lock(&sp->lock);
2444 n = sp_lookup(sp, start, end);
2445 /* Take care of old policies in the same range. */
2446 while (n && n->start < end) {
2447 struct rb_node *next = rb_next(&n->nd);
2448 if (n->start >= start) {
2454 /* Old policy spanning whole new range. */
2459 *mpol_new = *n->policy;
2460 atomic_set(&mpol_new->refcnt, 1);
2461 sp_node_init(n_new, end, n->end, mpol_new);
2463 sp_insert(sp, n_new);
2472 n = rb_entry(next, struct sp_node, nd);
2476 write_unlock(&sp->lock);
2483 kmem_cache_free(sn_cache, n_new);
2488 write_unlock(&sp->lock);
2490 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2493 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2500 * mpol_shared_policy_init - initialize shared policy for inode
2501 * @sp: pointer to inode shared policy
2502 * @mpol: struct mempolicy to install
2504 * Install non-NULL @mpol in inode's shared policy rb-tree.
2505 * On entry, the current task has a reference on a non-NULL @mpol.
2506 * This must be released on exit.
2507 * This is called at get_inode() calls and we can use GFP_KERNEL.
2509 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2513 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2514 rwlock_init(&sp->lock);
2517 struct vm_area_struct pvma;
2518 struct mempolicy *new;
2519 NODEMASK_SCRATCH(scratch);
2523 /* contextualize the tmpfs mount point mempolicy */
2524 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2526 goto free_scratch; /* no valid nodemask intersection */
2529 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2530 task_unlock(current);
2534 /* Create pseudo-vma that contains just the policy */
2535 vma_init(&pvma, NULL);
2536 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2537 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2540 mpol_put(new); /* drop initial ref */
2542 NODEMASK_SCRATCH_FREE(scratch);
2544 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2548 int mpol_set_shared_policy(struct shared_policy *info,
2549 struct vm_area_struct *vma, struct mempolicy *npol)
2552 struct sp_node *new = NULL;
2553 unsigned long sz = vma_pages(vma);
2555 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2557 sz, npol ? npol->mode : -1,
2558 npol ? npol->flags : -1,
2559 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2562 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2566 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2572 /* Free a backing policy store on inode delete. */
2573 void mpol_free_shared_policy(struct shared_policy *p)
2576 struct rb_node *next;
2578 if (!p->root.rb_node)
2580 write_lock(&p->lock);
2581 next = rb_first(&p->root);
2583 n = rb_entry(next, struct sp_node, nd);
2584 next = rb_next(&n->nd);
2587 write_unlock(&p->lock);
2590 #ifdef CONFIG_NUMA_BALANCING
2591 static int __initdata numabalancing_override;
2593 static void __init check_numabalancing_enable(void)
2595 bool numabalancing_default = false;
2597 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2598 numabalancing_default = true;
2600 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2601 if (numabalancing_override)
2602 set_numabalancing_state(numabalancing_override == 1);
2604 if (num_online_nodes() > 1 && !numabalancing_override) {
2605 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2606 numabalancing_default ? "Enabling" : "Disabling");
2607 set_numabalancing_state(numabalancing_default);
2611 static int __init setup_numabalancing(char *str)
2617 if (!strcmp(str, "enable")) {
2618 numabalancing_override = 1;
2620 } else if (!strcmp(str, "disable")) {
2621 numabalancing_override = -1;
2626 pr_warn("Unable to parse numa_balancing=\n");
2630 __setup("numa_balancing=", setup_numabalancing);
2632 static inline void __init check_numabalancing_enable(void)
2635 #endif /* CONFIG_NUMA_BALANCING */
2637 /* assumes fs == KERNEL_DS */
2638 void __init numa_policy_init(void)
2640 nodemask_t interleave_nodes;
2641 unsigned long largest = 0;
2642 int nid, prefer = 0;
2644 policy_cache = kmem_cache_create("numa_policy",
2645 sizeof(struct mempolicy),
2646 0, SLAB_PANIC, NULL);
2648 sn_cache = kmem_cache_create("shared_policy_node",
2649 sizeof(struct sp_node),
2650 0, SLAB_PANIC, NULL);
2652 for_each_node(nid) {
2653 preferred_node_policy[nid] = (struct mempolicy) {
2654 .refcnt = ATOMIC_INIT(1),
2655 .mode = MPOL_PREFERRED,
2656 .flags = MPOL_F_MOF | MPOL_F_MORON,
2657 .v = { .preferred_node = nid, },
2662 * Set interleaving policy for system init. Interleaving is only
2663 * enabled across suitably sized nodes (default is >= 16MB), or
2664 * fall back to the largest node if they're all smaller.
2666 nodes_clear(interleave_nodes);
2667 for_each_node_state(nid, N_MEMORY) {
2668 unsigned long total_pages = node_present_pages(nid);
2670 /* Preserve the largest node */
2671 if (largest < total_pages) {
2672 largest = total_pages;
2676 /* Interleave this node? */
2677 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2678 node_set(nid, interleave_nodes);
2681 /* All too small, use the largest */
2682 if (unlikely(nodes_empty(interleave_nodes)))
2683 node_set(prefer, interleave_nodes);
2685 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2686 pr_err("%s: interleaving failed\n", __func__);
2688 check_numabalancing_enable();
2691 /* Reset policy of current process to default */
2692 void numa_default_policy(void)
2694 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2698 * Parse and format mempolicy from/to strings
2702 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2704 static const char * const policy_modes[] =
2706 [MPOL_DEFAULT] = "default",
2707 [MPOL_PREFERRED] = "prefer",
2708 [MPOL_BIND] = "bind",
2709 [MPOL_INTERLEAVE] = "interleave",
2710 [MPOL_LOCAL] = "local",
2716 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2717 * @str: string containing mempolicy to parse
2718 * @mpol: pointer to struct mempolicy pointer, returned on success.
2721 * <mode>[=<flags>][:<nodelist>]
2723 * On success, returns 0, else 1
2725 int mpol_parse_str(char *str, struct mempolicy **mpol)
2727 struct mempolicy *new = NULL;
2728 unsigned short mode;
2729 unsigned short mode_flags;
2731 char *nodelist = strchr(str, ':');
2732 char *flags = strchr(str, '=');
2736 /* NUL-terminate mode or flags string */
2738 if (nodelist_parse(nodelist, nodes))
2740 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2746 *flags++ = '\0'; /* terminate mode string */
2748 for (mode = 0; mode < MPOL_MAX; mode++) {
2749 if (!strcmp(str, policy_modes[mode])) {
2753 if (mode >= MPOL_MAX)
2757 case MPOL_PREFERRED:
2759 * Insist on a nodelist of one node only
2762 char *rest = nodelist;
2763 while (isdigit(*rest))
2769 case MPOL_INTERLEAVE:
2771 * Default to online nodes with memory if no nodelist
2774 nodes = node_states[N_MEMORY];
2778 * Don't allow a nodelist; mpol_new() checks flags
2782 mode = MPOL_PREFERRED;
2786 * Insist on a empty nodelist
2793 * Insist on a nodelist
2802 * Currently, we only support two mutually exclusive
2805 if (!strcmp(flags, "static"))
2806 mode_flags |= MPOL_F_STATIC_NODES;
2807 else if (!strcmp(flags, "relative"))
2808 mode_flags |= MPOL_F_RELATIVE_NODES;
2813 new = mpol_new(mode, mode_flags, &nodes);
2818 * Save nodes for mpol_to_str() to show the tmpfs mount options
2819 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2821 if (mode != MPOL_PREFERRED)
2822 new->v.nodes = nodes;
2824 new->v.preferred_node = first_node(nodes);
2826 new->flags |= MPOL_F_LOCAL;
2829 * Save nodes for contextualization: this will be used to "clone"
2830 * the mempolicy in a specific context [cpuset] at a later time.
2832 new->w.user_nodemask = nodes;
2837 /* Restore string for error message */
2846 #endif /* CONFIG_TMPFS */
2849 * mpol_to_str - format a mempolicy structure for printing
2850 * @buffer: to contain formatted mempolicy string
2851 * @maxlen: length of @buffer
2852 * @pol: pointer to mempolicy to be formatted
2854 * Convert @pol into a string. If @buffer is too short, truncate the string.
2855 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2856 * longest flag, "relative", and to display at least a few node ids.
2858 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2861 nodemask_t nodes = NODE_MASK_NONE;
2862 unsigned short mode = MPOL_DEFAULT;
2863 unsigned short flags = 0;
2865 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2873 case MPOL_PREFERRED:
2874 if (flags & MPOL_F_LOCAL)
2877 node_set(pol->v.preferred_node, nodes);
2880 case MPOL_INTERLEAVE:
2881 nodes = pol->v.nodes;
2885 snprintf(p, maxlen, "unknown");
2889 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2891 if (flags & MPOL_MODE_FLAGS) {
2892 p += snprintf(p, buffer + maxlen - p, "=");
2895 * Currently, the only defined flags are mutually exclusive
2897 if (flags & MPOL_F_STATIC_NODES)
2898 p += snprintf(p, buffer + maxlen - p, "static");
2899 else if (flags & MPOL_F_RELATIVE_NODES)
2900 p += snprintf(p, buffer + maxlen - p, "relative");
2903 if (!nodes_empty(nodes))
2904 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2905 nodemask_pr_args(&nodes));