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(struct mm_struct *mm, unsigned long addr)
806 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
808 err = page_to_nid(p);
812 up_read(&mm->mmap_sem);
816 /* Retrieve NUMA policy */
817 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
818 unsigned long addr, unsigned long flags)
821 struct mm_struct *mm = current->mm;
822 struct vm_area_struct *vma = NULL;
823 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
826 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
829 if (flags & MPOL_F_MEMS_ALLOWED) {
830 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
832 *policy = 0; /* just so it's initialized */
834 *nmask = cpuset_current_mems_allowed;
835 task_unlock(current);
839 if (flags & MPOL_F_ADDR) {
841 * Do NOT fall back to task policy if the
842 * vma/shared policy at addr is NULL. We
843 * want to return MPOL_DEFAULT in this case.
845 down_read(&mm->mmap_sem);
846 vma = find_vma_intersection(mm, addr, addr+1);
848 up_read(&mm->mmap_sem);
851 if (vma->vm_ops && vma->vm_ops->get_policy)
852 pol = vma->vm_ops->get_policy(vma, addr);
854 pol = vma->vm_policy;
859 pol = &default_policy; /* indicates default behavior */
861 if (flags & MPOL_F_NODE) {
862 if (flags & MPOL_F_ADDR) {
864 * Take a refcount on the mpol, lookup_node()
865 * wil drop the mmap_sem, so after calling
866 * lookup_node() only "pol" remains valid, "vma"
872 err = lookup_node(mm, addr);
876 } else if (pol == current->mempolicy &&
877 pol->mode == MPOL_INTERLEAVE) {
878 *policy = next_node_in(current->il_prev, pol->v.nodes);
884 *policy = pol == &default_policy ? MPOL_DEFAULT :
887 * Internal mempolicy flags must be masked off before exposing
888 * the policy to userspace.
890 *policy |= (pol->flags & MPOL_MODE_FLAGS);
895 if (mpol_store_user_nodemask(pol)) {
896 *nmask = pol->w.user_nodemask;
899 get_policy_nodemask(pol, nmask);
900 task_unlock(current);
907 up_read(&mm->mmap_sem);
909 mpol_put(pol_refcount);
913 #ifdef CONFIG_MIGRATION
915 * page migration, thp tail pages can be passed.
917 static void migrate_page_add(struct page *page, struct list_head *pagelist,
920 struct page *head = compound_head(page);
922 * Avoid migrating a page that is shared with others.
924 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
925 if (!isolate_lru_page(head)) {
926 list_add_tail(&head->lru, pagelist);
927 mod_node_page_state(page_pgdat(head),
928 NR_ISOLATED_ANON + page_is_file_cache(head),
929 hpage_nr_pages(head));
934 /* page allocation callback for NUMA node migration */
935 struct page *alloc_new_node_page(struct page *page, unsigned long node)
938 return alloc_huge_page_node(page_hstate(compound_head(page)),
940 else if (PageTransHuge(page)) {
943 thp = alloc_pages_node(node,
944 (GFP_TRANSHUGE | __GFP_THISNODE),
948 prep_transhuge_page(thp);
951 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
956 * Migrate pages from one node to a target node.
957 * Returns error or the number of pages not migrated.
959 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
967 node_set(source, nmask);
970 * This does not "check" the range but isolates all pages that
971 * need migration. Between passing in the full user address
972 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
974 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
975 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
976 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
978 if (!list_empty(&pagelist)) {
979 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
980 MIGRATE_SYNC, MR_SYSCALL);
982 putback_movable_pages(&pagelist);
989 * Move pages between the two nodesets so as to preserve the physical
990 * layout as much as possible.
992 * Returns the number of page that could not be moved.
994 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
995 const nodemask_t *to, int flags)
1001 err = migrate_prep();
1005 down_read(&mm->mmap_sem);
1008 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1009 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1010 * bit in 'tmp', and return that <source, dest> pair for migration.
1011 * The pair of nodemasks 'to' and 'from' define the map.
1013 * If no pair of bits is found that way, fallback to picking some
1014 * pair of 'source' and 'dest' bits that are not the same. If the
1015 * 'source' and 'dest' bits are the same, this represents a node
1016 * that will be migrating to itself, so no pages need move.
1018 * If no bits are left in 'tmp', or if all remaining bits left
1019 * in 'tmp' correspond to the same bit in 'to', return false
1020 * (nothing left to migrate).
1022 * This lets us pick a pair of nodes to migrate between, such that
1023 * if possible the dest node is not already occupied by some other
1024 * source node, minimizing the risk of overloading the memory on a
1025 * node that would happen if we migrated incoming memory to a node
1026 * before migrating outgoing memory source that same node.
1028 * A single scan of tmp is sufficient. As we go, we remember the
1029 * most recent <s, d> pair that moved (s != d). If we find a pair
1030 * that not only moved, but what's better, moved to an empty slot
1031 * (d is not set in tmp), then we break out then, with that pair.
1032 * Otherwise when we finish scanning from_tmp, we at least have the
1033 * most recent <s, d> pair that moved. If we get all the way through
1034 * the scan of tmp without finding any node that moved, much less
1035 * moved to an empty node, then there is nothing left worth migrating.
1039 while (!nodes_empty(tmp)) {
1041 int source = NUMA_NO_NODE;
1044 for_each_node_mask(s, tmp) {
1047 * do_migrate_pages() tries to maintain the relative
1048 * node relationship of the pages established between
1049 * threads and memory areas.
1051 * However if the number of source nodes is not equal to
1052 * the number of destination nodes we can not preserve
1053 * this node relative relationship. In that case, skip
1054 * copying memory from a node that is in the destination
1057 * Example: [2,3,4] -> [3,4,5] moves everything.
1058 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1061 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1062 (node_isset(s, *to)))
1065 d = node_remap(s, *from, *to);
1069 source = s; /* Node moved. Memorize */
1072 /* dest not in remaining from nodes? */
1073 if (!node_isset(dest, tmp))
1076 if (source == NUMA_NO_NODE)
1079 node_clear(source, tmp);
1080 err = migrate_to_node(mm, source, dest, flags);
1086 up_read(&mm->mmap_sem);
1094 * Allocate a new page for page migration based on vma policy.
1095 * Start by assuming the page is mapped by the same vma as contains @start.
1096 * Search forward from there, if not. N.B., this assumes that the
1097 * list of pages handed to migrate_pages()--which is how we get here--
1098 * is in virtual address order.
1100 static struct page *new_page(struct page *page, unsigned long start)
1102 struct vm_area_struct *vma;
1103 unsigned long uninitialized_var(address);
1105 vma = find_vma(current->mm, start);
1107 address = page_address_in_vma(page, vma);
1108 if (address != -EFAULT)
1113 if (PageHuge(page)) {
1114 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1116 } else if (PageTransHuge(page)) {
1119 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1123 prep_transhuge_page(thp);
1127 * if !vma, alloc_page_vma() will use task or system default policy
1129 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1134 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1135 unsigned long flags)
1139 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1140 const nodemask_t *to, int flags)
1145 static struct page *new_page(struct page *page, unsigned long start)
1151 static long do_mbind(unsigned long start, unsigned long len,
1152 unsigned short mode, unsigned short mode_flags,
1153 nodemask_t *nmask, unsigned long flags)
1155 struct mm_struct *mm = current->mm;
1156 struct mempolicy *new;
1159 LIST_HEAD(pagelist);
1161 if (flags & ~(unsigned long)MPOL_MF_VALID)
1163 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1166 if (start & ~PAGE_MASK)
1169 if (mode == MPOL_DEFAULT)
1170 flags &= ~MPOL_MF_STRICT;
1172 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1180 new = mpol_new(mode, mode_flags, nmask);
1182 return PTR_ERR(new);
1184 if (flags & MPOL_MF_LAZY)
1185 new->flags |= MPOL_F_MOF;
1188 * If we are using the default policy then operation
1189 * on discontinuous address spaces is okay after all
1192 flags |= MPOL_MF_DISCONTIG_OK;
1194 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1195 start, start + len, mode, mode_flags,
1196 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1198 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1200 err = migrate_prep();
1205 NODEMASK_SCRATCH(scratch);
1207 down_write(&mm->mmap_sem);
1209 err = mpol_set_nodemask(new, nmask, scratch);
1210 task_unlock(current);
1212 up_write(&mm->mmap_sem);
1215 NODEMASK_SCRATCH_FREE(scratch);
1220 err = queue_pages_range(mm, start, end, nmask,
1221 flags | MPOL_MF_INVERT, &pagelist);
1223 err = mbind_range(mm, start, end, new);
1228 if (!list_empty(&pagelist)) {
1229 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1230 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1231 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1233 putback_movable_pages(&pagelist);
1236 if (nr_failed && (flags & MPOL_MF_STRICT))
1239 putback_movable_pages(&pagelist);
1241 up_write(&mm->mmap_sem);
1248 * User space interface with variable sized bitmaps for nodelists.
1251 /* Copy a node mask from user space. */
1252 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1253 unsigned long maxnode)
1257 unsigned long nlongs;
1258 unsigned long endmask;
1261 nodes_clear(*nodes);
1262 if (maxnode == 0 || !nmask)
1264 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1267 nlongs = BITS_TO_LONGS(maxnode);
1268 if ((maxnode % BITS_PER_LONG) == 0)
1271 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1274 * When the user specified more nodes than supported just check
1275 * if the non supported part is all zero.
1277 * If maxnode have more longs than MAX_NUMNODES, check
1278 * the bits in that area first. And then go through to
1279 * check the rest bits which equal or bigger than MAX_NUMNODES.
1280 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1282 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1283 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1284 if (get_user(t, nmask + k))
1286 if (k == nlongs - 1) {
1292 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1296 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1297 unsigned long valid_mask = endmask;
1299 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1300 if (get_user(t, nmask + nlongs - 1))
1306 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1308 nodes_addr(*nodes)[nlongs-1] &= endmask;
1312 /* Copy a kernel node mask to user space */
1313 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1316 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1317 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1319 if (copy > nbytes) {
1320 if (copy > PAGE_SIZE)
1322 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1326 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1329 static long kernel_mbind(unsigned long start, unsigned long len,
1330 unsigned long mode, const unsigned long __user *nmask,
1331 unsigned long maxnode, unsigned int flags)
1335 unsigned short mode_flags;
1337 mode_flags = mode & MPOL_MODE_FLAGS;
1338 mode &= ~MPOL_MODE_FLAGS;
1339 if (mode >= MPOL_MAX)
1341 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1342 (mode_flags & MPOL_F_RELATIVE_NODES))
1344 err = get_nodes(&nodes, nmask, maxnode);
1347 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1350 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1351 unsigned long, mode, const unsigned long __user *, nmask,
1352 unsigned long, maxnode, unsigned int, flags)
1354 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1357 /* Set the process memory policy */
1358 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1359 unsigned long maxnode)
1363 unsigned short flags;
1365 flags = mode & MPOL_MODE_FLAGS;
1366 mode &= ~MPOL_MODE_FLAGS;
1367 if ((unsigned int)mode >= MPOL_MAX)
1369 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1371 err = get_nodes(&nodes, nmask, maxnode);
1374 return do_set_mempolicy(mode, flags, &nodes);
1377 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1378 unsigned long, maxnode)
1380 return kernel_set_mempolicy(mode, nmask, maxnode);
1383 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1384 const unsigned long __user *old_nodes,
1385 const unsigned long __user *new_nodes)
1387 struct mm_struct *mm = NULL;
1388 struct task_struct *task;
1389 nodemask_t task_nodes;
1393 NODEMASK_SCRATCH(scratch);
1398 old = &scratch->mask1;
1399 new = &scratch->mask2;
1401 err = get_nodes(old, old_nodes, maxnode);
1405 err = get_nodes(new, new_nodes, maxnode);
1409 /* Find the mm_struct */
1411 task = pid ? find_task_by_vpid(pid) : current;
1417 get_task_struct(task);
1422 * Check if this process has the right to modify the specified process.
1423 * Use the regular "ptrace_may_access()" checks.
1425 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1432 task_nodes = cpuset_mems_allowed(task);
1433 /* Is the user allowed to access the target nodes? */
1434 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1439 task_nodes = cpuset_mems_allowed(current);
1440 nodes_and(*new, *new, task_nodes);
1441 if (nodes_empty(*new))
1444 nodes_and(*new, *new, node_states[N_MEMORY]);
1445 if (nodes_empty(*new))
1448 err = security_task_movememory(task);
1452 mm = get_task_mm(task);
1453 put_task_struct(task);
1460 err = do_migrate_pages(mm, old, new,
1461 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1465 NODEMASK_SCRATCH_FREE(scratch);
1470 put_task_struct(task);
1475 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1476 const unsigned long __user *, old_nodes,
1477 const unsigned long __user *, new_nodes)
1479 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1483 /* Retrieve NUMA policy */
1484 static int kernel_get_mempolicy(int __user *policy,
1485 unsigned long __user *nmask,
1486 unsigned long maxnode,
1488 unsigned long flags)
1491 int uninitialized_var(pval);
1494 if (nmask != NULL && maxnode < MAX_NUMNODES)
1497 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1502 if (policy && put_user(pval, policy))
1506 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1511 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1512 unsigned long __user *, nmask, unsigned long, maxnode,
1513 unsigned long, addr, unsigned long, flags)
1515 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1518 #ifdef CONFIG_COMPAT
1520 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1521 compat_ulong_t __user *, nmask,
1522 compat_ulong_t, maxnode,
1523 compat_ulong_t, addr, compat_ulong_t, flags)
1526 unsigned long __user *nm = NULL;
1527 unsigned long nr_bits, alloc_size;
1528 DECLARE_BITMAP(bm, MAX_NUMNODES);
1530 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1531 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1534 nm = compat_alloc_user_space(alloc_size);
1536 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1538 if (!err && nmask) {
1539 unsigned long copy_size;
1540 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1541 err = copy_from_user(bm, nm, copy_size);
1542 /* ensure entire bitmap is zeroed */
1543 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1544 err |= compat_put_bitmap(nmask, bm, nr_bits);
1550 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1551 compat_ulong_t, maxnode)
1553 unsigned long __user *nm = NULL;
1554 unsigned long nr_bits, alloc_size;
1555 DECLARE_BITMAP(bm, MAX_NUMNODES);
1557 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1558 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1561 if (compat_get_bitmap(bm, nmask, nr_bits))
1563 nm = compat_alloc_user_space(alloc_size);
1564 if (copy_to_user(nm, bm, alloc_size))
1568 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1571 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1572 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1573 compat_ulong_t, maxnode, compat_ulong_t, flags)
1575 unsigned long __user *nm = NULL;
1576 unsigned long nr_bits, alloc_size;
1579 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1580 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1583 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1585 nm = compat_alloc_user_space(alloc_size);
1586 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1590 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1593 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1594 compat_ulong_t, maxnode,
1595 const compat_ulong_t __user *, old_nodes,
1596 const compat_ulong_t __user *, new_nodes)
1598 unsigned long __user *old = NULL;
1599 unsigned long __user *new = NULL;
1600 nodemask_t tmp_mask;
1601 unsigned long nr_bits;
1604 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1605 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1607 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1609 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1611 new = old + size / sizeof(unsigned long);
1612 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1616 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1619 new = compat_alloc_user_space(size);
1620 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1623 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1626 #endif /* CONFIG_COMPAT */
1628 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1631 struct mempolicy *pol = NULL;
1634 if (vma->vm_ops && vma->vm_ops->get_policy) {
1635 pol = vma->vm_ops->get_policy(vma, addr);
1636 } else if (vma->vm_policy) {
1637 pol = vma->vm_policy;
1640 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1641 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1642 * count on these policies which will be dropped by
1643 * mpol_cond_put() later
1645 if (mpol_needs_cond_ref(pol))
1654 * get_vma_policy(@vma, @addr)
1655 * @vma: virtual memory area whose policy is sought
1656 * @addr: address in @vma for shared policy lookup
1658 * Returns effective policy for a VMA at specified address.
1659 * Falls back to current->mempolicy or system default policy, as necessary.
1660 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1661 * count--added by the get_policy() vm_op, as appropriate--to protect against
1662 * freeing by another task. It is the caller's responsibility to free the
1663 * extra reference for shared policies.
1665 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1668 struct mempolicy *pol = __get_vma_policy(vma, addr);
1671 pol = get_task_policy(current);
1676 bool vma_policy_mof(struct vm_area_struct *vma)
1678 struct mempolicy *pol;
1680 if (vma->vm_ops && vma->vm_ops->get_policy) {
1683 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1684 if (pol && (pol->flags & MPOL_F_MOF))
1691 pol = vma->vm_policy;
1693 pol = get_task_policy(current);
1695 return pol->flags & MPOL_F_MOF;
1698 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1700 enum zone_type dynamic_policy_zone = policy_zone;
1702 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1705 * if policy->v.nodes has movable memory only,
1706 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1708 * policy->v.nodes is intersect with node_states[N_MEMORY].
1709 * so if the following test faile, it implies
1710 * policy->v.nodes has movable memory only.
1712 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1713 dynamic_policy_zone = ZONE_MOVABLE;
1715 return zone >= dynamic_policy_zone;
1719 * Return a nodemask representing a mempolicy for filtering nodes for
1722 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1724 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1725 if (unlikely(policy->mode == MPOL_BIND) &&
1726 apply_policy_zone(policy, gfp_zone(gfp)) &&
1727 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1728 return &policy->v.nodes;
1733 /* Return the node id preferred by the given mempolicy, or the given id */
1734 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1737 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1738 nd = policy->v.preferred_node;
1741 * __GFP_THISNODE shouldn't even be used with the bind policy
1742 * because we might easily break the expectation to stay on the
1743 * requested node and not break the policy.
1745 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1751 /* Do dynamic interleaving for a process */
1752 static unsigned interleave_nodes(struct mempolicy *policy)
1755 struct task_struct *me = current;
1757 next = next_node_in(me->il_prev, policy->v.nodes);
1758 if (next < MAX_NUMNODES)
1764 * Depending on the memory policy provide a node from which to allocate the
1767 unsigned int mempolicy_slab_node(void)
1769 struct mempolicy *policy;
1770 int node = numa_mem_id();
1775 policy = current->mempolicy;
1776 if (!policy || policy->flags & MPOL_F_LOCAL)
1779 switch (policy->mode) {
1780 case MPOL_PREFERRED:
1782 * handled MPOL_F_LOCAL above
1784 return policy->v.preferred_node;
1786 case MPOL_INTERLEAVE:
1787 return interleave_nodes(policy);
1793 * Follow bind policy behavior and start allocation at the
1796 struct zonelist *zonelist;
1797 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1798 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1799 z = first_zones_zonelist(zonelist, highest_zoneidx,
1801 return z->zone ? zone_to_nid(z->zone) : node;
1810 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1811 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1812 * number of present nodes.
1814 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1816 unsigned nnodes = nodes_weight(pol->v.nodes);
1822 return numa_node_id();
1823 target = (unsigned int)n % nnodes;
1824 nid = first_node(pol->v.nodes);
1825 for (i = 0; i < target; i++)
1826 nid = next_node(nid, pol->v.nodes);
1830 /* Determine a node number for interleave */
1831 static inline unsigned interleave_nid(struct mempolicy *pol,
1832 struct vm_area_struct *vma, unsigned long addr, int shift)
1838 * for small pages, there is no difference between
1839 * shift and PAGE_SHIFT, so the bit-shift is safe.
1840 * for huge pages, since vm_pgoff is in units of small
1841 * pages, we need to shift off the always 0 bits to get
1844 BUG_ON(shift < PAGE_SHIFT);
1845 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1846 off += (addr - vma->vm_start) >> shift;
1847 return offset_il_node(pol, off);
1849 return interleave_nodes(pol);
1852 #ifdef CONFIG_HUGETLBFS
1854 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1855 * @vma: virtual memory area whose policy is sought
1856 * @addr: address in @vma for shared policy lookup and interleave policy
1857 * @gfp_flags: for requested zone
1858 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1859 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1861 * Returns a nid suitable for a huge page allocation and a pointer
1862 * to the struct mempolicy for conditional unref after allocation.
1863 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1864 * @nodemask for filtering the zonelist.
1866 * Must be protected by read_mems_allowed_begin()
1868 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1869 struct mempolicy **mpol, nodemask_t **nodemask)
1873 *mpol = get_vma_policy(vma, addr);
1874 *nodemask = NULL; /* assume !MPOL_BIND */
1876 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1877 nid = interleave_nid(*mpol, vma, addr,
1878 huge_page_shift(hstate_vma(vma)));
1880 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1881 if ((*mpol)->mode == MPOL_BIND)
1882 *nodemask = &(*mpol)->v.nodes;
1888 * init_nodemask_of_mempolicy
1890 * If the current task's mempolicy is "default" [NULL], return 'false'
1891 * to indicate default policy. Otherwise, extract the policy nodemask
1892 * for 'bind' or 'interleave' policy into the argument nodemask, or
1893 * initialize the argument nodemask to contain the single node for
1894 * 'preferred' or 'local' policy and return 'true' to indicate presence
1895 * of non-default mempolicy.
1897 * We don't bother with reference counting the mempolicy [mpol_get/put]
1898 * because the current task is examining it's own mempolicy and a task's
1899 * mempolicy is only ever changed by the task itself.
1901 * N.B., it is the caller's responsibility to free a returned nodemask.
1903 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1905 struct mempolicy *mempolicy;
1908 if (!(mask && current->mempolicy))
1912 mempolicy = current->mempolicy;
1913 switch (mempolicy->mode) {
1914 case MPOL_PREFERRED:
1915 if (mempolicy->flags & MPOL_F_LOCAL)
1916 nid = numa_node_id();
1918 nid = mempolicy->v.preferred_node;
1919 init_nodemask_of_node(mask, nid);
1924 case MPOL_INTERLEAVE:
1925 *mask = mempolicy->v.nodes;
1931 task_unlock(current);
1938 * mempolicy_nodemask_intersects
1940 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1941 * policy. Otherwise, check for intersection between mask and the policy
1942 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1943 * policy, always return true since it may allocate elsewhere on fallback.
1945 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1947 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1948 const nodemask_t *mask)
1950 struct mempolicy *mempolicy;
1956 mempolicy = tsk->mempolicy;
1960 switch (mempolicy->mode) {
1961 case MPOL_PREFERRED:
1963 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1964 * allocate from, they may fallback to other nodes when oom.
1965 * Thus, it's possible for tsk to have allocated memory from
1970 case MPOL_INTERLEAVE:
1971 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1981 /* Allocate a page in interleaved policy.
1982 Own path because it needs to do special accounting. */
1983 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1988 page = __alloc_pages(gfp, order, nid);
1989 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1990 if (!static_branch_likely(&vm_numa_stat_key))
1992 if (page && page_to_nid(page) == nid) {
1994 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2001 * alloc_pages_vma - Allocate a page for a VMA.
2004 * %GFP_USER user allocation.
2005 * %GFP_KERNEL kernel allocations,
2006 * %GFP_HIGHMEM highmem/user allocations,
2007 * %GFP_FS allocation should not call back into a file system.
2008 * %GFP_ATOMIC don't sleep.
2010 * @order:Order of the GFP allocation.
2011 * @vma: Pointer to VMA or NULL if not available.
2012 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2013 * @node: Which node to prefer for allocation (modulo policy).
2014 * @hugepage: for hugepages try only the preferred node if possible
2016 * This function allocates a page from the kernel page pool and applies
2017 * a NUMA policy associated with the VMA or the current process.
2018 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2019 * mm_struct of the VMA to prevent it from going away. Should be used for
2020 * all allocations for pages that will be mapped into user space. Returns
2021 * NULL when no page can be allocated.
2024 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2025 unsigned long addr, int node, bool hugepage)
2027 struct mempolicy *pol;
2032 pol = get_vma_policy(vma, addr);
2034 if (pol->mode == MPOL_INTERLEAVE) {
2037 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2039 page = alloc_page_interleave(gfp, order, nid);
2043 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2044 int hpage_node = node;
2047 * For hugepage allocation and non-interleave policy which
2048 * allows the current node (or other explicitly preferred
2049 * node) we only try to allocate from the current/preferred
2050 * node and don't fall back to other nodes, as the cost of
2051 * remote accesses would likely offset THP benefits.
2053 * If the policy is interleave, or does not allow the current
2054 * node in its nodemask, we allocate the standard way.
2056 if (pol->mode == MPOL_PREFERRED &&
2057 !(pol->flags & MPOL_F_LOCAL))
2058 hpage_node = pol->v.preferred_node;
2060 nmask = policy_nodemask(gfp, pol);
2061 if (!nmask || node_isset(hpage_node, *nmask)) {
2064 * We cannot invoke reclaim if __GFP_THISNODE
2065 * is set. Invoking reclaim with
2066 * __GFP_THISNODE set, would cause THP
2067 * allocations to trigger heavy swapping
2068 * despite there may be tons of free memory
2069 * (including potentially plenty of THP
2070 * already available in the buddy) on all the
2073 * At most we could invoke compaction when
2074 * __GFP_THISNODE is set (but we would need to
2075 * refrain from invoking reclaim even if
2076 * compaction returned COMPACT_SKIPPED because
2077 * there wasn't not enough memory to succeed
2078 * compaction). For now just avoid
2079 * __GFP_THISNODE instead of limiting the
2080 * allocation path to a strict and single
2081 * compaction invocation.
2083 * Supposedly if direct reclaim was enabled by
2084 * the caller, the app prefers THP regardless
2085 * of the node it comes from so this would be
2086 * more desiderable behavior than only
2087 * providing THP originated from the local
2088 * node in such case.
2090 if (!(gfp & __GFP_DIRECT_RECLAIM))
2091 gfp |= __GFP_THISNODE;
2092 page = __alloc_pages_node(hpage_node, gfp, order);
2097 nmask = policy_nodemask(gfp, pol);
2098 preferred_nid = policy_node(gfp, pol, node);
2099 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2106 * alloc_pages_current - Allocate pages.
2109 * %GFP_USER user allocation,
2110 * %GFP_KERNEL kernel allocation,
2111 * %GFP_HIGHMEM highmem allocation,
2112 * %GFP_FS don't call back into a file system.
2113 * %GFP_ATOMIC don't sleep.
2114 * @order: Power of two of allocation size in pages. 0 is a single page.
2116 * Allocate a page from the kernel page pool. When not in
2117 * interrupt context and apply the current process NUMA policy.
2118 * Returns NULL when no page can be allocated.
2120 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2122 struct mempolicy *pol = &default_policy;
2125 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2126 pol = get_task_policy(current);
2129 * No reference counting needed for current->mempolicy
2130 * nor system default_policy
2132 if (pol->mode == MPOL_INTERLEAVE)
2133 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2135 page = __alloc_pages_nodemask(gfp, order,
2136 policy_node(gfp, pol, numa_node_id()),
2137 policy_nodemask(gfp, pol));
2141 EXPORT_SYMBOL(alloc_pages_current);
2143 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2145 struct mempolicy *pol = mpol_dup(vma_policy(src));
2148 return PTR_ERR(pol);
2149 dst->vm_policy = pol;
2154 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2155 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2156 * with the mems_allowed returned by cpuset_mems_allowed(). This
2157 * keeps mempolicies cpuset relative after its cpuset moves. See
2158 * further kernel/cpuset.c update_nodemask().
2160 * current's mempolicy may be rebinded by the other task(the task that changes
2161 * cpuset's mems), so we needn't do rebind work for current task.
2164 /* Slow path of a mempolicy duplicate */
2165 struct mempolicy *__mpol_dup(struct mempolicy *old)
2167 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2170 return ERR_PTR(-ENOMEM);
2172 /* task's mempolicy is protected by alloc_lock */
2173 if (old == current->mempolicy) {
2176 task_unlock(current);
2180 if (current_cpuset_is_being_rebound()) {
2181 nodemask_t mems = cpuset_mems_allowed(current);
2182 mpol_rebind_policy(new, &mems);
2184 atomic_set(&new->refcnt, 1);
2188 /* Slow path of a mempolicy comparison */
2189 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2193 if (a->mode != b->mode)
2195 if (a->flags != b->flags)
2197 if (mpol_store_user_nodemask(a))
2198 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2204 case MPOL_INTERLEAVE:
2205 return !!nodes_equal(a->v.nodes, b->v.nodes);
2206 case MPOL_PREFERRED:
2207 /* a's ->flags is the same as b's */
2208 if (a->flags & MPOL_F_LOCAL)
2210 return a->v.preferred_node == b->v.preferred_node;
2218 * Shared memory backing store policy support.
2220 * Remember policies even when nobody has shared memory mapped.
2221 * The policies are kept in Red-Black tree linked from the inode.
2222 * They are protected by the sp->lock rwlock, which should be held
2223 * for any accesses to the tree.
2227 * lookup first element intersecting start-end. Caller holds sp->lock for
2228 * reading or for writing
2230 static struct sp_node *
2231 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2233 struct rb_node *n = sp->root.rb_node;
2236 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2238 if (start >= p->end)
2240 else if (end <= p->start)
2248 struct sp_node *w = NULL;
2249 struct rb_node *prev = rb_prev(n);
2252 w = rb_entry(prev, struct sp_node, nd);
2253 if (w->end <= start)
2257 return rb_entry(n, struct sp_node, nd);
2261 * Insert a new shared policy into the list. Caller holds sp->lock for
2264 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2266 struct rb_node **p = &sp->root.rb_node;
2267 struct rb_node *parent = NULL;
2272 nd = rb_entry(parent, struct sp_node, nd);
2273 if (new->start < nd->start)
2275 else if (new->end > nd->end)
2276 p = &(*p)->rb_right;
2280 rb_link_node(&new->nd, parent, p);
2281 rb_insert_color(&new->nd, &sp->root);
2282 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2283 new->policy ? new->policy->mode : 0);
2286 /* Find shared policy intersecting idx */
2288 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2290 struct mempolicy *pol = NULL;
2293 if (!sp->root.rb_node)
2295 read_lock(&sp->lock);
2296 sn = sp_lookup(sp, idx, idx+1);
2298 mpol_get(sn->policy);
2301 read_unlock(&sp->lock);
2305 static void sp_free(struct sp_node *n)
2307 mpol_put(n->policy);
2308 kmem_cache_free(sn_cache, n);
2312 * mpol_misplaced - check whether current page node is valid in policy
2314 * @page: page to be checked
2315 * @vma: vm area where page mapped
2316 * @addr: virtual address where page mapped
2318 * Lookup current policy node id for vma,addr and "compare to" page's
2322 * -1 - not misplaced, page is in the right node
2323 * node - node id where the page should be
2325 * Policy determination "mimics" alloc_page_vma().
2326 * Called from fault path where we know the vma and faulting address.
2328 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2330 struct mempolicy *pol;
2332 int curnid = page_to_nid(page);
2333 unsigned long pgoff;
2334 int thiscpu = raw_smp_processor_id();
2335 int thisnid = cpu_to_node(thiscpu);
2339 pol = get_vma_policy(vma, addr);
2340 if (!(pol->flags & MPOL_F_MOF))
2343 switch (pol->mode) {
2344 case MPOL_INTERLEAVE:
2345 pgoff = vma->vm_pgoff;
2346 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2347 polnid = offset_il_node(pol, pgoff);
2350 case MPOL_PREFERRED:
2351 if (pol->flags & MPOL_F_LOCAL)
2352 polnid = numa_node_id();
2354 polnid = pol->v.preferred_node;
2360 * allows binding to multiple nodes.
2361 * use current page if in policy nodemask,
2362 * else select nearest allowed node, if any.
2363 * If no allowed nodes, use current [!misplaced].
2365 if (node_isset(curnid, pol->v.nodes))
2367 z = first_zones_zonelist(
2368 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2369 gfp_zone(GFP_HIGHUSER),
2371 polnid = zone_to_nid(z->zone);
2378 /* Migrate the page towards the node whose CPU is referencing it */
2379 if (pol->flags & MPOL_F_MORON) {
2382 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2386 if (curnid != polnid)
2395 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2396 * dropped after task->mempolicy is set to NULL so that any allocation done as
2397 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2400 void mpol_put_task_policy(struct task_struct *task)
2402 struct mempolicy *pol;
2405 pol = task->mempolicy;
2406 task->mempolicy = NULL;
2411 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2413 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2414 rb_erase(&n->nd, &sp->root);
2418 static void sp_node_init(struct sp_node *node, unsigned long start,
2419 unsigned long end, struct mempolicy *pol)
2421 node->start = start;
2426 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2427 struct mempolicy *pol)
2430 struct mempolicy *newpol;
2432 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2436 newpol = mpol_dup(pol);
2437 if (IS_ERR(newpol)) {
2438 kmem_cache_free(sn_cache, n);
2441 newpol->flags |= MPOL_F_SHARED;
2442 sp_node_init(n, start, end, newpol);
2447 /* Replace a policy range. */
2448 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2449 unsigned long end, struct sp_node *new)
2452 struct sp_node *n_new = NULL;
2453 struct mempolicy *mpol_new = NULL;
2457 write_lock(&sp->lock);
2458 n = sp_lookup(sp, start, end);
2459 /* Take care of old policies in the same range. */
2460 while (n && n->start < end) {
2461 struct rb_node *next = rb_next(&n->nd);
2462 if (n->start >= start) {
2468 /* Old policy spanning whole new range. */
2473 *mpol_new = *n->policy;
2474 atomic_set(&mpol_new->refcnt, 1);
2475 sp_node_init(n_new, end, n->end, mpol_new);
2477 sp_insert(sp, n_new);
2486 n = rb_entry(next, struct sp_node, nd);
2490 write_unlock(&sp->lock);
2497 kmem_cache_free(sn_cache, n_new);
2502 write_unlock(&sp->lock);
2504 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2507 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2514 * mpol_shared_policy_init - initialize shared policy for inode
2515 * @sp: pointer to inode shared policy
2516 * @mpol: struct mempolicy to install
2518 * Install non-NULL @mpol in inode's shared policy rb-tree.
2519 * On entry, the current task has a reference on a non-NULL @mpol.
2520 * This must be released on exit.
2521 * This is called at get_inode() calls and we can use GFP_KERNEL.
2523 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2527 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2528 rwlock_init(&sp->lock);
2531 struct vm_area_struct pvma;
2532 struct mempolicy *new;
2533 NODEMASK_SCRATCH(scratch);
2537 /* contextualize the tmpfs mount point mempolicy */
2538 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2540 goto free_scratch; /* no valid nodemask intersection */
2543 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2544 task_unlock(current);
2548 /* Create pseudo-vma that contains just the policy */
2549 vma_init(&pvma, NULL);
2550 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2551 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2554 mpol_put(new); /* drop initial ref */
2556 NODEMASK_SCRATCH_FREE(scratch);
2558 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2562 int mpol_set_shared_policy(struct shared_policy *info,
2563 struct vm_area_struct *vma, struct mempolicy *npol)
2566 struct sp_node *new = NULL;
2567 unsigned long sz = vma_pages(vma);
2569 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2571 sz, npol ? npol->mode : -1,
2572 npol ? npol->flags : -1,
2573 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2576 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2580 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2586 /* Free a backing policy store on inode delete. */
2587 void mpol_free_shared_policy(struct shared_policy *p)
2590 struct rb_node *next;
2592 if (!p->root.rb_node)
2594 write_lock(&p->lock);
2595 next = rb_first(&p->root);
2597 n = rb_entry(next, struct sp_node, nd);
2598 next = rb_next(&n->nd);
2601 write_unlock(&p->lock);
2604 #ifdef CONFIG_NUMA_BALANCING
2605 static int __initdata numabalancing_override;
2607 static void __init check_numabalancing_enable(void)
2609 bool numabalancing_default = false;
2611 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2612 numabalancing_default = true;
2614 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2615 if (numabalancing_override)
2616 set_numabalancing_state(numabalancing_override == 1);
2618 if (num_online_nodes() > 1 && !numabalancing_override) {
2619 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2620 numabalancing_default ? "Enabling" : "Disabling");
2621 set_numabalancing_state(numabalancing_default);
2625 static int __init setup_numabalancing(char *str)
2631 if (!strcmp(str, "enable")) {
2632 numabalancing_override = 1;
2634 } else if (!strcmp(str, "disable")) {
2635 numabalancing_override = -1;
2640 pr_warn("Unable to parse numa_balancing=\n");
2644 __setup("numa_balancing=", setup_numabalancing);
2646 static inline void __init check_numabalancing_enable(void)
2649 #endif /* CONFIG_NUMA_BALANCING */
2651 /* assumes fs == KERNEL_DS */
2652 void __init numa_policy_init(void)
2654 nodemask_t interleave_nodes;
2655 unsigned long largest = 0;
2656 int nid, prefer = 0;
2658 policy_cache = kmem_cache_create("numa_policy",
2659 sizeof(struct mempolicy),
2660 0, SLAB_PANIC, NULL);
2662 sn_cache = kmem_cache_create("shared_policy_node",
2663 sizeof(struct sp_node),
2664 0, SLAB_PANIC, NULL);
2666 for_each_node(nid) {
2667 preferred_node_policy[nid] = (struct mempolicy) {
2668 .refcnt = ATOMIC_INIT(1),
2669 .mode = MPOL_PREFERRED,
2670 .flags = MPOL_F_MOF | MPOL_F_MORON,
2671 .v = { .preferred_node = nid, },
2676 * Set interleaving policy for system init. Interleaving is only
2677 * enabled across suitably sized nodes (default is >= 16MB), or
2678 * fall back to the largest node if they're all smaller.
2680 nodes_clear(interleave_nodes);
2681 for_each_node_state(nid, N_MEMORY) {
2682 unsigned long total_pages = node_present_pages(nid);
2684 /* Preserve the largest node */
2685 if (largest < total_pages) {
2686 largest = total_pages;
2690 /* Interleave this node? */
2691 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2692 node_set(nid, interleave_nodes);
2695 /* All too small, use the largest */
2696 if (unlikely(nodes_empty(interleave_nodes)))
2697 node_set(prefer, interleave_nodes);
2699 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2700 pr_err("%s: interleaving failed\n", __func__);
2702 check_numabalancing_enable();
2705 /* Reset policy of current process to default */
2706 void numa_default_policy(void)
2708 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2712 * Parse and format mempolicy from/to strings
2716 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2718 static const char * const policy_modes[] =
2720 [MPOL_DEFAULT] = "default",
2721 [MPOL_PREFERRED] = "prefer",
2722 [MPOL_BIND] = "bind",
2723 [MPOL_INTERLEAVE] = "interleave",
2724 [MPOL_LOCAL] = "local",
2730 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2731 * @str: string containing mempolicy to parse
2732 * @mpol: pointer to struct mempolicy pointer, returned on success.
2735 * <mode>[=<flags>][:<nodelist>]
2737 * On success, returns 0, else 1
2739 int mpol_parse_str(char *str, struct mempolicy **mpol)
2741 struct mempolicy *new = NULL;
2742 unsigned short mode_flags;
2744 char *nodelist = strchr(str, ':');
2745 char *flags = strchr(str, '=');
2749 /* NUL-terminate mode or flags string */
2751 if (nodelist_parse(nodelist, nodes))
2753 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2759 *flags++ = '\0'; /* terminate mode string */
2761 mode = match_string(policy_modes, MPOL_MAX, str);
2766 case MPOL_PREFERRED:
2768 * Insist on a nodelist of one node only
2771 char *rest = nodelist;
2772 while (isdigit(*rest))
2778 case MPOL_INTERLEAVE:
2780 * Default to online nodes with memory if no nodelist
2783 nodes = node_states[N_MEMORY];
2787 * Don't allow a nodelist; mpol_new() checks flags
2791 mode = MPOL_PREFERRED;
2795 * Insist on a empty nodelist
2802 * Insist on a nodelist
2811 * Currently, we only support two mutually exclusive
2814 if (!strcmp(flags, "static"))
2815 mode_flags |= MPOL_F_STATIC_NODES;
2816 else if (!strcmp(flags, "relative"))
2817 mode_flags |= MPOL_F_RELATIVE_NODES;
2822 new = mpol_new(mode, mode_flags, &nodes);
2827 * Save nodes for mpol_to_str() to show the tmpfs mount options
2828 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2830 if (mode != MPOL_PREFERRED)
2831 new->v.nodes = nodes;
2833 new->v.preferred_node = first_node(nodes);
2835 new->flags |= MPOL_F_LOCAL;
2838 * Save nodes for contextualization: this will be used to "clone"
2839 * the mempolicy in a specific context [cpuset] at a later time.
2841 new->w.user_nodemask = nodes;
2846 /* Restore string for error message */
2855 #endif /* CONFIG_TMPFS */
2858 * mpol_to_str - format a mempolicy structure for printing
2859 * @buffer: to contain formatted mempolicy string
2860 * @maxlen: length of @buffer
2861 * @pol: pointer to mempolicy to be formatted
2863 * Convert @pol into a string. If @buffer is too short, truncate the string.
2864 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2865 * longest flag, "relative", and to display at least a few node ids.
2867 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2870 nodemask_t nodes = NODE_MASK_NONE;
2871 unsigned short mode = MPOL_DEFAULT;
2872 unsigned short flags = 0;
2874 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2882 case MPOL_PREFERRED:
2883 if (flags & MPOL_F_LOCAL)
2886 node_set(pol->v.preferred_node, nodes);
2889 case MPOL_INTERLEAVE:
2890 nodes = pol->v.nodes;
2894 snprintf(p, maxlen, "unknown");
2898 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2900 if (flags & MPOL_MODE_FLAGS) {
2901 p += snprintf(p, buffer + maxlen - p, "=");
2904 * Currently, the only defined flags are mutually exclusive
2906 if (flags & MPOL_F_STATIC_NODES)
2907 p += snprintf(p, buffer + maxlen - p, "static");
2908 else if (flags & MPOL_F_RELATIVE_NODES)
2909 p += snprintf(p, buffer + maxlen - p, "relative");
2912 if (!nodes_empty(nodes))
2913 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2914 nodemask_pr_args(&nodes));