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/nodemask.h>
79 #include <linux/cpuset.h>
80 #include <linux/slab.h>
81 #include <linux/string.h>
82 #include <linux/export.h>
83 #include <linux/nsproxy.h>
84 #include <linux/interrupt.h>
85 #include <linux/init.h>
86 #include <linux/compat.h>
87 #include <linux/swap.h>
88 #include <linux/seq_file.h>
89 #include <linux/proc_fs.h>
90 #include <linux/migrate.h>
91 #include <linux/ksm.h>
92 #include <linux/rmap.h>
93 #include <linux/security.h>
94 #include <linux/syscalls.h>
95 #include <linux/ctype.h>
96 #include <linux/mm_inline.h>
97 #include <linux/mmu_notifier.h>
98 #include <linux/printk.h>
100 #include <asm/tlbflush.h>
101 #include <linux/uaccess.h>
103 #include "internal.h"
106 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
107 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
109 static struct kmem_cache *policy_cache;
110 static struct kmem_cache *sn_cache;
112 /* Highest zone. An specific allocation for a zone below that is not
114 enum zone_type policy_zone = 0;
117 * run-time system-wide default policy => local allocation
119 static struct mempolicy default_policy = {
120 .refcnt = ATOMIC_INIT(1), /* never free it */
121 .mode = MPOL_PREFERRED,
122 .flags = MPOL_F_LOCAL,
125 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
127 struct mempolicy *get_task_policy(struct task_struct *p)
129 struct mempolicy *pol = p->mempolicy;
135 node = numa_node_id();
136 if (node != NUMA_NO_NODE) {
137 pol = &preferred_node_policy[node];
138 /* preferred_node_policy is not initialised early in boot */
143 return &default_policy;
146 static const struct mempolicy_operations {
147 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
149 * If read-side task has no lock to protect task->mempolicy, write-side
150 * task will rebind the task->mempolicy by two step. The first step is
151 * setting all the newly nodes, and the second step is cleaning all the
152 * disallowed nodes. In this way, we can avoid finding no node to alloc
154 * If we have a lock to protect task->mempolicy in read-side, we do
158 * MPOL_REBIND_ONCE - do rebind work at once
159 * MPOL_REBIND_STEP1 - set all the newly nodes
160 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
162 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
163 enum mpol_rebind_step step);
164 } mpol_ops[MPOL_MAX];
166 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
168 return pol->flags & MPOL_MODE_FLAGS;
171 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
172 const nodemask_t *rel)
175 nodes_fold(tmp, *orig, nodes_weight(*rel));
176 nodes_onto(*ret, tmp, *rel);
179 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
181 if (nodes_empty(*nodes))
183 pol->v.nodes = *nodes;
187 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
190 pol->flags |= MPOL_F_LOCAL; /* local allocation */
191 else if (nodes_empty(*nodes))
192 return -EINVAL; /* no allowed nodes */
194 pol->v.preferred_node = first_node(*nodes);
198 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
200 if (nodes_empty(*nodes))
202 pol->v.nodes = *nodes;
207 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
208 * any, for the new policy. mpol_new() has already validated the nodes
209 * parameter with respect to the policy mode and flags. But, we need to
210 * handle an empty nodemask with MPOL_PREFERRED here.
212 * Must be called holding task's alloc_lock to protect task's mems_allowed
213 * and mempolicy. May also be called holding the mmap_semaphore for write.
215 static int mpol_set_nodemask(struct mempolicy *pol,
216 const nodemask_t *nodes, struct nodemask_scratch *nsc)
220 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
224 nodes_and(nsc->mask1,
225 cpuset_current_mems_allowed, node_states[N_MEMORY]);
228 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
229 nodes = NULL; /* explicit local allocation */
231 if (pol->flags & MPOL_F_RELATIVE_NODES)
232 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
234 nodes_and(nsc->mask2, *nodes, nsc->mask1);
236 if (mpol_store_user_nodemask(pol))
237 pol->w.user_nodemask = *nodes;
239 pol->w.cpuset_mems_allowed =
240 cpuset_current_mems_allowed;
244 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
246 ret = mpol_ops[pol->mode].create(pol, NULL);
251 * This function just creates a new policy, does some check and simple
252 * initialization. You must invoke mpol_set_nodemask() to set nodes.
254 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
257 struct mempolicy *policy;
259 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
260 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
262 if (mode == MPOL_DEFAULT) {
263 if (nodes && !nodes_empty(*nodes))
264 return ERR_PTR(-EINVAL);
270 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
271 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
272 * All other modes require a valid pointer to a non-empty nodemask.
274 if (mode == MPOL_PREFERRED) {
275 if (nodes_empty(*nodes)) {
276 if (((flags & MPOL_F_STATIC_NODES) ||
277 (flags & MPOL_F_RELATIVE_NODES)))
278 return ERR_PTR(-EINVAL);
280 } else if (mode == MPOL_LOCAL) {
281 if (!nodes_empty(*nodes) ||
282 (flags & MPOL_F_STATIC_NODES) ||
283 (flags & MPOL_F_RELATIVE_NODES))
284 return ERR_PTR(-EINVAL);
285 mode = MPOL_PREFERRED;
286 } else if (nodes_empty(*nodes))
287 return ERR_PTR(-EINVAL);
288 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
290 return ERR_PTR(-ENOMEM);
291 atomic_set(&policy->refcnt, 1);
293 policy->flags = flags;
298 /* Slow path of a mpol destructor. */
299 void __mpol_put(struct mempolicy *p)
301 if (!atomic_dec_and_test(&p->refcnt))
303 kmem_cache_free(policy_cache, p);
306 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
307 enum mpol_rebind_step step)
313 * MPOL_REBIND_ONCE - do rebind work at once
314 * MPOL_REBIND_STEP1 - set all the newly nodes
315 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
317 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
318 enum mpol_rebind_step step)
322 if (pol->flags & MPOL_F_STATIC_NODES)
323 nodes_and(tmp, pol->w.user_nodemask, *nodes);
324 else if (pol->flags & MPOL_F_RELATIVE_NODES)
325 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
328 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
331 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
332 nodes_remap(tmp, pol->v.nodes,
333 pol->w.cpuset_mems_allowed, *nodes);
334 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
335 } else if (step == MPOL_REBIND_STEP2) {
336 tmp = pol->w.cpuset_mems_allowed;
337 pol->w.cpuset_mems_allowed = *nodes;
342 if (nodes_empty(tmp))
345 if (step == MPOL_REBIND_STEP1)
346 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
347 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
352 if (!node_isset(current->il_next, tmp)) {
353 current->il_next = next_node_in(current->il_next, tmp);
354 if (current->il_next >= MAX_NUMNODES)
355 current->il_next = numa_node_id();
359 static void mpol_rebind_preferred(struct mempolicy *pol,
360 const nodemask_t *nodes,
361 enum mpol_rebind_step step)
365 if (pol->flags & MPOL_F_STATIC_NODES) {
366 int node = first_node(pol->w.user_nodemask);
368 if (node_isset(node, *nodes)) {
369 pol->v.preferred_node = node;
370 pol->flags &= ~MPOL_F_LOCAL;
372 pol->flags |= MPOL_F_LOCAL;
373 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
374 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
375 pol->v.preferred_node = first_node(tmp);
376 } else if (!(pol->flags & MPOL_F_LOCAL)) {
377 pol->v.preferred_node = node_remap(pol->v.preferred_node,
378 pol->w.cpuset_mems_allowed,
380 pol->w.cpuset_mems_allowed = *nodes;
385 * mpol_rebind_policy - Migrate a policy to a different set of nodes
387 * If read-side task has no lock to protect task->mempolicy, write-side
388 * task will rebind the task->mempolicy by two step. The first step is
389 * setting all the newly nodes, and the second step is cleaning all the
390 * disallowed nodes. In this way, we can avoid finding no node to alloc
392 * If we have a lock to protect task->mempolicy in read-side, we do
396 * MPOL_REBIND_ONCE - do rebind work at once
397 * MPOL_REBIND_STEP1 - set all the newly nodes
398 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
400 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
401 enum mpol_rebind_step step)
405 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
406 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
409 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
412 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
415 if (step == MPOL_REBIND_STEP1)
416 pol->flags |= MPOL_F_REBINDING;
417 else if (step == MPOL_REBIND_STEP2)
418 pol->flags &= ~MPOL_F_REBINDING;
419 else if (step >= MPOL_REBIND_NSTEP)
422 mpol_ops[pol->mode].rebind(pol, newmask, step);
426 * Wrapper for mpol_rebind_policy() that just requires task
427 * pointer, and updates task mempolicy.
429 * Called with task's alloc_lock held.
432 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
433 enum mpol_rebind_step step)
435 mpol_rebind_policy(tsk->mempolicy, new, step);
439 * Rebind each vma in mm to new nodemask.
441 * Call holding a reference to mm. Takes mm->mmap_sem during call.
444 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
446 struct vm_area_struct *vma;
448 down_write(&mm->mmap_sem);
449 for (vma = mm->mmap; vma; vma = vma->vm_next)
450 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
451 up_write(&mm->mmap_sem);
454 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
456 .rebind = mpol_rebind_default,
458 [MPOL_INTERLEAVE] = {
459 .create = mpol_new_interleave,
460 .rebind = mpol_rebind_nodemask,
463 .create = mpol_new_preferred,
464 .rebind = mpol_rebind_preferred,
467 .create = mpol_new_bind,
468 .rebind = mpol_rebind_nodemask,
472 static void migrate_page_add(struct page *page, struct list_head *pagelist,
473 unsigned long flags);
476 struct list_head *pagelist;
479 struct vm_area_struct *prev;
483 * Scan through pages checking if pages follow certain conditions,
484 * and move them to the pagelist if they do.
486 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
487 unsigned long end, struct mm_walk *walk)
489 struct vm_area_struct *vma = walk->vma;
491 struct queue_pages *qp = walk->private;
492 unsigned long flags = qp->flags;
497 if (pmd_trans_huge(*pmd)) {
498 ptl = pmd_lock(walk->mm, pmd);
499 if (pmd_trans_huge(*pmd)) {
500 page = pmd_page(*pmd);
501 if (is_huge_zero_page(page)) {
503 __split_huge_pmd(vma, pmd, addr, false, NULL);
508 ret = split_huge_page(page);
519 if (pmd_trans_unstable(pmd))
522 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
523 for (; addr != end; pte++, addr += PAGE_SIZE) {
524 if (!pte_present(*pte))
526 page = vm_normal_page(vma, addr, *pte);
530 * vm_normal_page() filters out zero pages, but there might
531 * still be PageReserved pages to skip, perhaps in a VDSO.
533 if (PageReserved(page))
535 nid = page_to_nid(page);
536 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
538 if (PageTransCompound(page)) {
540 pte_unmap_unlock(pte, ptl);
542 ret = split_huge_page(page);
545 /* Failed to split -- skip. */
547 pte = pte_offset_map_lock(walk->mm, pmd,
554 migrate_page_add(page, qp->pagelist, flags);
556 pte_unmap_unlock(pte - 1, ptl);
561 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
562 unsigned long addr, unsigned long end,
563 struct mm_walk *walk)
565 #ifdef CONFIG_HUGETLB_PAGE
566 struct queue_pages *qp = walk->private;
567 unsigned long flags = qp->flags;
573 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
574 entry = huge_ptep_get(pte);
575 if (!pte_present(entry))
577 page = pte_page(entry);
578 nid = page_to_nid(page);
579 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
581 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
582 if (flags & (MPOL_MF_MOVE_ALL) ||
583 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
584 isolate_huge_page(page, qp->pagelist);
593 #ifdef CONFIG_NUMA_BALANCING
595 * This is used to mark a range of virtual addresses to be inaccessible.
596 * These are later cleared by a NUMA hinting fault. Depending on these
597 * faults, pages may be migrated for better NUMA placement.
599 * This is assuming that NUMA faults are handled using PROT_NONE. If
600 * an architecture makes a different choice, it will need further
601 * changes to the core.
603 unsigned long change_prot_numa(struct vm_area_struct *vma,
604 unsigned long addr, unsigned long end)
608 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
610 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
615 static unsigned long change_prot_numa(struct vm_area_struct *vma,
616 unsigned long addr, unsigned long end)
620 #endif /* CONFIG_NUMA_BALANCING */
622 static int queue_pages_test_walk(unsigned long start, unsigned long end,
623 struct mm_walk *walk)
625 struct vm_area_struct *vma = walk->vma;
626 struct queue_pages *qp = walk->private;
627 unsigned long endvma = vma->vm_end;
628 unsigned long flags = qp->flags;
630 if (!vma_migratable(vma))
635 if (vma->vm_start > start)
636 start = vma->vm_start;
638 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
639 if (!vma->vm_next && vma->vm_end < end)
641 if (qp->prev && qp->prev->vm_end < vma->vm_start)
647 if (flags & MPOL_MF_LAZY) {
648 /* Similar to task_numa_work, skip inaccessible VMAs */
649 if (!is_vm_hugetlb_page(vma) &&
650 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
651 !(vma->vm_flags & VM_MIXEDMAP))
652 change_prot_numa(vma, start, endvma);
656 /* queue pages from current vma */
657 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
663 * Walk through page tables and collect pages to be migrated.
665 * If pages found in a given range are on a set of nodes (determined by
666 * @nodes and @flags,) it's isolated and queued to the pagelist which is
667 * passed via @private.)
670 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
671 nodemask_t *nodes, unsigned long flags,
672 struct list_head *pagelist)
674 struct queue_pages qp = {
675 .pagelist = pagelist,
680 struct mm_walk queue_pages_walk = {
681 .hugetlb_entry = queue_pages_hugetlb,
682 .pmd_entry = queue_pages_pte_range,
683 .test_walk = queue_pages_test_walk,
688 return walk_page_range(start, end, &queue_pages_walk);
692 * Apply policy to a single VMA
693 * This must be called with the mmap_sem held for writing.
695 static int vma_replace_policy(struct vm_area_struct *vma,
696 struct mempolicy *pol)
699 struct mempolicy *old;
700 struct mempolicy *new;
702 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
703 vma->vm_start, vma->vm_end, vma->vm_pgoff,
704 vma->vm_ops, vma->vm_file,
705 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
711 if (vma->vm_ops && vma->vm_ops->set_policy) {
712 err = vma->vm_ops->set_policy(vma, new);
717 old = vma->vm_policy;
718 vma->vm_policy = new; /* protected by mmap_sem */
727 /* Step 2: apply policy to a range and do splits. */
728 static int mbind_range(struct mm_struct *mm, unsigned long start,
729 unsigned long end, struct mempolicy *new_pol)
731 struct vm_area_struct *next;
732 struct vm_area_struct *prev;
733 struct vm_area_struct *vma;
736 unsigned long vmstart;
739 vma = find_vma(mm, start);
740 if (!vma || vma->vm_start > start)
744 if (start > vma->vm_start)
747 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
749 vmstart = max(start, vma->vm_start);
750 vmend = min(end, vma->vm_end);
752 if (mpol_equal(vma_policy(vma), new_pol))
755 pgoff = vma->vm_pgoff +
756 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
757 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
758 vma->anon_vma, vma->vm_file, pgoff,
759 new_pol, vma->vm_userfaultfd_ctx);
763 if (mpol_equal(vma_policy(vma), new_pol))
765 /* vma_merge() joined vma && vma->next, case 8 */
768 if (vma->vm_start != vmstart) {
769 err = split_vma(vma->vm_mm, vma, vmstart, 1);
773 if (vma->vm_end != vmend) {
774 err = split_vma(vma->vm_mm, vma, vmend, 0);
779 err = vma_replace_policy(vma, new_pol);
788 /* Set the process memory policy */
789 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
792 struct mempolicy *new, *old;
793 NODEMASK_SCRATCH(scratch);
799 new = mpol_new(mode, flags, nodes);
806 ret = mpol_set_nodemask(new, nodes, scratch);
808 task_unlock(current);
812 old = current->mempolicy;
813 current->mempolicy = new;
814 if (new && new->mode == MPOL_INTERLEAVE &&
815 nodes_weight(new->v.nodes))
816 current->il_next = first_node(new->v.nodes);
817 task_unlock(current);
821 NODEMASK_SCRATCH_FREE(scratch);
826 * Return nodemask for policy for get_mempolicy() query
828 * Called with task's alloc_lock held
830 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
833 if (p == &default_policy)
839 case MPOL_INTERLEAVE:
843 if (!(p->flags & MPOL_F_LOCAL))
844 node_set(p->v.preferred_node, *nodes);
845 /* else return empty node mask for local allocation */
852 static int lookup_node(unsigned long addr)
857 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
859 err = page_to_nid(p);
865 /* Retrieve NUMA policy */
866 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
867 unsigned long addr, unsigned long flags)
870 struct mm_struct *mm = current->mm;
871 struct vm_area_struct *vma = NULL;
872 struct mempolicy *pol = current->mempolicy;
875 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
878 if (flags & MPOL_F_MEMS_ALLOWED) {
879 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
881 *policy = 0; /* just so it's initialized */
883 *nmask = cpuset_current_mems_allowed;
884 task_unlock(current);
888 if (flags & MPOL_F_ADDR) {
890 * Do NOT fall back to task policy if the
891 * vma/shared policy at addr is NULL. We
892 * want to return MPOL_DEFAULT in this case.
894 down_read(&mm->mmap_sem);
895 vma = find_vma_intersection(mm, addr, addr+1);
897 up_read(&mm->mmap_sem);
900 if (vma->vm_ops && vma->vm_ops->get_policy)
901 pol = vma->vm_ops->get_policy(vma, addr);
903 pol = vma->vm_policy;
908 pol = &default_policy; /* indicates default behavior */
910 if (flags & MPOL_F_NODE) {
911 if (flags & MPOL_F_ADDR) {
912 err = lookup_node(addr);
916 } else if (pol == current->mempolicy &&
917 pol->mode == MPOL_INTERLEAVE) {
918 *policy = current->il_next;
924 *policy = pol == &default_policy ? MPOL_DEFAULT :
927 * Internal mempolicy flags must be masked off before exposing
928 * the policy to userspace.
930 *policy |= (pol->flags & MPOL_MODE_FLAGS);
934 up_read(¤t->mm->mmap_sem);
940 if (mpol_store_user_nodemask(pol)) {
941 *nmask = pol->w.user_nodemask;
944 get_policy_nodemask(pol, nmask);
945 task_unlock(current);
952 up_read(¤t->mm->mmap_sem);
956 #ifdef CONFIG_MIGRATION
960 static void migrate_page_add(struct page *page, struct list_head *pagelist,
964 * Avoid migrating a page that is shared with others.
966 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
967 if (!isolate_lru_page(page)) {
968 list_add_tail(&page->lru, pagelist);
969 inc_node_page_state(page, NR_ISOLATED_ANON +
970 page_is_file_cache(page));
975 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
978 return alloc_huge_page_node(page_hstate(compound_head(page)),
981 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
986 * Migrate pages from one node to a target node.
987 * Returns error or the number of pages not migrated.
989 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
997 node_set(source, nmask);
1000 * This does not "check" the range but isolates all pages that
1001 * need migration. Between passing in the full user address
1002 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1004 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1005 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1006 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1008 if (!list_empty(&pagelist)) {
1009 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1010 MIGRATE_SYNC, MR_SYSCALL);
1012 putback_movable_pages(&pagelist);
1019 * Move pages between the two nodesets so as to preserve the physical
1020 * layout as much as possible.
1022 * Returns the number of page that could not be moved.
1024 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1025 const nodemask_t *to, int flags)
1031 err = migrate_prep();
1035 down_read(&mm->mmap_sem);
1038 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1039 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1040 * bit in 'tmp', and return that <source, dest> pair for migration.
1041 * The pair of nodemasks 'to' and 'from' define the map.
1043 * If no pair of bits is found that way, fallback to picking some
1044 * pair of 'source' and 'dest' bits that are not the same. If the
1045 * 'source' and 'dest' bits are the same, this represents a node
1046 * that will be migrating to itself, so no pages need move.
1048 * If no bits are left in 'tmp', or if all remaining bits left
1049 * in 'tmp' correspond to the same bit in 'to', return false
1050 * (nothing left to migrate).
1052 * This lets us pick a pair of nodes to migrate between, such that
1053 * if possible the dest node is not already occupied by some other
1054 * source node, minimizing the risk of overloading the memory on a
1055 * node that would happen if we migrated incoming memory to a node
1056 * before migrating outgoing memory source that same node.
1058 * A single scan of tmp is sufficient. As we go, we remember the
1059 * most recent <s, d> pair that moved (s != d). If we find a pair
1060 * that not only moved, but what's better, moved to an empty slot
1061 * (d is not set in tmp), then we break out then, with that pair.
1062 * Otherwise when we finish scanning from_tmp, we at least have the
1063 * most recent <s, d> pair that moved. If we get all the way through
1064 * the scan of tmp without finding any node that moved, much less
1065 * moved to an empty node, then there is nothing left worth migrating.
1069 while (!nodes_empty(tmp)) {
1071 int source = NUMA_NO_NODE;
1074 for_each_node_mask(s, tmp) {
1077 * do_migrate_pages() tries to maintain the relative
1078 * node relationship of the pages established between
1079 * threads and memory areas.
1081 * However if the number of source nodes is not equal to
1082 * the number of destination nodes we can not preserve
1083 * this node relative relationship. In that case, skip
1084 * copying memory from a node that is in the destination
1087 * Example: [2,3,4] -> [3,4,5] moves everything.
1088 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1091 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1092 (node_isset(s, *to)))
1095 d = node_remap(s, *from, *to);
1099 source = s; /* Node moved. Memorize */
1102 /* dest not in remaining from nodes? */
1103 if (!node_isset(dest, tmp))
1106 if (source == NUMA_NO_NODE)
1109 node_clear(source, tmp);
1110 err = migrate_to_node(mm, source, dest, flags);
1116 up_read(&mm->mmap_sem);
1124 * Allocate a new page for page migration based on vma policy.
1125 * Start by assuming the page is mapped by the same vma as contains @start.
1126 * Search forward from there, if not. N.B., this assumes that the
1127 * list of pages handed to migrate_pages()--which is how we get here--
1128 * is in virtual address order.
1130 static struct page *new_page(struct page *page, unsigned long start, int **x)
1132 struct vm_area_struct *vma;
1133 unsigned long uninitialized_var(address);
1135 vma = find_vma(current->mm, start);
1137 address = page_address_in_vma(page, vma);
1138 if (address != -EFAULT)
1143 if (PageHuge(page)) {
1145 return alloc_huge_page_noerr(vma, address, 1);
1148 * if !vma, alloc_page_vma() will use task or system default policy
1150 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1154 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1155 unsigned long flags)
1159 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1160 const nodemask_t *to, int flags)
1165 static struct page *new_page(struct page *page, unsigned long start, int **x)
1171 static long do_mbind(unsigned long start, unsigned long len,
1172 unsigned short mode, unsigned short mode_flags,
1173 nodemask_t *nmask, unsigned long flags)
1175 struct mm_struct *mm = current->mm;
1176 struct mempolicy *new;
1179 LIST_HEAD(pagelist);
1181 if (flags & ~(unsigned long)MPOL_MF_VALID)
1183 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1186 if (start & ~PAGE_MASK)
1189 if (mode == MPOL_DEFAULT)
1190 flags &= ~MPOL_MF_STRICT;
1192 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1200 new = mpol_new(mode, mode_flags, nmask);
1202 return PTR_ERR(new);
1204 if (flags & MPOL_MF_LAZY)
1205 new->flags |= MPOL_F_MOF;
1208 * If we are using the default policy then operation
1209 * on discontinuous address spaces is okay after all
1212 flags |= MPOL_MF_DISCONTIG_OK;
1214 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1215 start, start + len, mode, mode_flags,
1216 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1218 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1220 err = migrate_prep();
1225 NODEMASK_SCRATCH(scratch);
1227 down_write(&mm->mmap_sem);
1229 err = mpol_set_nodemask(new, nmask, scratch);
1230 task_unlock(current);
1232 up_write(&mm->mmap_sem);
1235 NODEMASK_SCRATCH_FREE(scratch);
1240 err = queue_pages_range(mm, start, end, nmask,
1241 flags | MPOL_MF_INVERT, &pagelist);
1243 err = mbind_range(mm, start, end, new);
1248 if (!list_empty(&pagelist)) {
1249 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1250 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1251 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1253 putback_movable_pages(&pagelist);
1256 if (nr_failed && (flags & MPOL_MF_STRICT))
1259 putback_movable_pages(&pagelist);
1261 up_write(&mm->mmap_sem);
1268 * User space interface with variable sized bitmaps for nodelists.
1271 /* Copy a node mask from user space. */
1272 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1273 unsigned long maxnode)
1276 unsigned long nlongs;
1277 unsigned long endmask;
1280 nodes_clear(*nodes);
1281 if (maxnode == 0 || !nmask)
1283 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1286 nlongs = BITS_TO_LONGS(maxnode);
1287 if ((maxnode % BITS_PER_LONG) == 0)
1290 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1292 /* When the user specified more nodes than supported just check
1293 if the non supported part is all zero. */
1294 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1295 if (nlongs > PAGE_SIZE/sizeof(long))
1297 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1299 if (get_user(t, nmask + k))
1301 if (k == nlongs - 1) {
1307 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1311 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1313 nodes_addr(*nodes)[nlongs-1] &= endmask;
1317 /* Copy a kernel node mask to user space */
1318 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1321 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1322 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1324 if (copy > nbytes) {
1325 if (copy > PAGE_SIZE)
1327 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1331 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1334 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1335 unsigned long, mode, const unsigned long __user *, nmask,
1336 unsigned long, maxnode, unsigned, flags)
1340 unsigned short mode_flags;
1342 mode_flags = mode & MPOL_MODE_FLAGS;
1343 mode &= ~MPOL_MODE_FLAGS;
1344 if (mode >= MPOL_MAX)
1346 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1347 (mode_flags & MPOL_F_RELATIVE_NODES))
1349 err = get_nodes(&nodes, nmask, maxnode);
1352 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1355 /* Set the process memory policy */
1356 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1357 unsigned long, maxnode)
1361 unsigned short flags;
1363 flags = mode & MPOL_MODE_FLAGS;
1364 mode &= ~MPOL_MODE_FLAGS;
1365 if ((unsigned int)mode >= MPOL_MAX)
1367 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1369 err = get_nodes(&nodes, nmask, maxnode);
1372 return do_set_mempolicy(mode, flags, &nodes);
1375 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1376 const unsigned long __user *, old_nodes,
1377 const unsigned long __user *, new_nodes)
1379 const struct cred *cred = current_cred(), *tcred;
1380 struct mm_struct *mm = NULL;
1381 struct task_struct *task;
1382 nodemask_t task_nodes;
1386 NODEMASK_SCRATCH(scratch);
1391 old = &scratch->mask1;
1392 new = &scratch->mask2;
1394 err = get_nodes(old, old_nodes, maxnode);
1398 err = get_nodes(new, new_nodes, maxnode);
1402 /* Find the mm_struct */
1404 task = pid ? find_task_by_vpid(pid) : current;
1410 get_task_struct(task);
1415 * Check if this process has the right to modify the specified
1416 * process. The right exists if the process has administrative
1417 * capabilities, superuser privileges or the same
1418 * userid as the target process.
1420 tcred = __task_cred(task);
1421 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1422 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1423 !capable(CAP_SYS_NICE)) {
1430 task_nodes = cpuset_mems_allowed(task);
1431 /* Is the user allowed to access the target nodes? */
1432 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1437 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1442 err = security_task_movememory(task);
1446 mm = get_task_mm(task);
1447 put_task_struct(task);
1454 err = do_migrate_pages(mm, old, new,
1455 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1459 NODEMASK_SCRATCH_FREE(scratch);
1464 put_task_struct(task);
1470 /* Retrieve NUMA policy */
1471 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1472 unsigned long __user *, nmask, unsigned long, maxnode,
1473 unsigned long, addr, unsigned long, flags)
1476 int uninitialized_var(pval);
1479 if (nmask != NULL && maxnode < MAX_NUMNODES)
1482 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1487 if (policy && put_user(pval, policy))
1491 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1496 #ifdef CONFIG_COMPAT
1498 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1499 compat_ulong_t __user *, nmask,
1500 compat_ulong_t, maxnode,
1501 compat_ulong_t, addr, compat_ulong_t, flags)
1504 unsigned long __user *nm = NULL;
1505 unsigned long nr_bits, alloc_size;
1506 DECLARE_BITMAP(bm, MAX_NUMNODES);
1508 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1509 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1512 nm = compat_alloc_user_space(alloc_size);
1514 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1516 if (!err && nmask) {
1517 unsigned long copy_size;
1518 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1519 err = copy_from_user(bm, nm, copy_size);
1520 /* ensure entire bitmap is zeroed */
1521 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1522 err |= compat_put_bitmap(nmask, bm, nr_bits);
1528 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1529 compat_ulong_t, maxnode)
1532 unsigned long __user *nm = NULL;
1533 unsigned long nr_bits, alloc_size;
1534 DECLARE_BITMAP(bm, MAX_NUMNODES);
1536 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1537 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1540 err = compat_get_bitmap(bm, nmask, nr_bits);
1541 nm = compat_alloc_user_space(alloc_size);
1542 err |= copy_to_user(nm, bm, alloc_size);
1548 return sys_set_mempolicy(mode, nm, nr_bits+1);
1551 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1552 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1553 compat_ulong_t, maxnode, compat_ulong_t, flags)
1556 unsigned long __user *nm = NULL;
1557 unsigned long nr_bits, alloc_size;
1560 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1561 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1564 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1565 nm = compat_alloc_user_space(alloc_size);
1566 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1572 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1577 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1580 struct mempolicy *pol = NULL;
1583 if (vma->vm_ops && vma->vm_ops->get_policy) {
1584 pol = vma->vm_ops->get_policy(vma, addr);
1585 } else if (vma->vm_policy) {
1586 pol = vma->vm_policy;
1589 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1590 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1591 * count on these policies which will be dropped by
1592 * mpol_cond_put() later
1594 if (mpol_needs_cond_ref(pol))
1603 * get_vma_policy(@vma, @addr)
1604 * @vma: virtual memory area whose policy is sought
1605 * @addr: address in @vma for shared policy lookup
1607 * Returns effective policy for a VMA at specified address.
1608 * Falls back to current->mempolicy or system default policy, as necessary.
1609 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1610 * count--added by the get_policy() vm_op, as appropriate--to protect against
1611 * freeing by another task. It is the caller's responsibility to free the
1612 * extra reference for shared policies.
1614 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1617 struct mempolicy *pol = __get_vma_policy(vma, addr);
1620 pol = get_task_policy(current);
1625 bool vma_policy_mof(struct vm_area_struct *vma)
1627 struct mempolicy *pol;
1629 if (vma->vm_ops && vma->vm_ops->get_policy) {
1632 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1633 if (pol && (pol->flags & MPOL_F_MOF))
1640 pol = vma->vm_policy;
1642 pol = get_task_policy(current);
1644 return pol->flags & MPOL_F_MOF;
1647 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1649 enum zone_type dynamic_policy_zone = policy_zone;
1651 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1654 * if policy->v.nodes has movable memory only,
1655 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1657 * policy->v.nodes is intersect with node_states[N_MEMORY].
1658 * so if the following test faile, it implies
1659 * policy->v.nodes has movable memory only.
1661 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1662 dynamic_policy_zone = ZONE_MOVABLE;
1664 return zone >= dynamic_policy_zone;
1668 * Return a nodemask representing a mempolicy for filtering nodes for
1671 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1673 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1674 if (unlikely(policy->mode == MPOL_BIND) &&
1675 apply_policy_zone(policy, gfp_zone(gfp)) &&
1676 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1677 return &policy->v.nodes;
1682 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1683 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1686 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1687 nd = policy->v.preferred_node;
1690 * __GFP_THISNODE shouldn't even be used with the bind policy
1691 * because we might easily break the expectation to stay on the
1692 * requested node and not break the policy.
1694 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1697 return node_zonelist(nd, gfp);
1700 /* Do dynamic interleaving for a process */
1701 static unsigned interleave_nodes(struct mempolicy *policy)
1704 struct task_struct *me = current;
1707 next = next_node_in(nid, policy->v.nodes);
1708 if (next < MAX_NUMNODES)
1714 * Depending on the memory policy provide a node from which to allocate the
1717 unsigned int mempolicy_slab_node(void)
1719 struct mempolicy *policy;
1720 int node = numa_mem_id();
1725 policy = current->mempolicy;
1726 if (!policy || policy->flags & MPOL_F_LOCAL)
1729 switch (policy->mode) {
1730 case MPOL_PREFERRED:
1732 * handled MPOL_F_LOCAL above
1734 return policy->v.preferred_node;
1736 case MPOL_INTERLEAVE:
1737 return interleave_nodes(policy);
1743 * Follow bind policy behavior and start allocation at the
1746 struct zonelist *zonelist;
1747 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1748 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1749 z = first_zones_zonelist(zonelist, highest_zoneidx,
1751 return z->zone ? z->zone->node : node;
1760 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1761 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1762 * number of present nodes.
1764 static unsigned offset_il_node(struct mempolicy *pol,
1765 struct vm_area_struct *vma, unsigned long n)
1767 unsigned nnodes = nodes_weight(pol->v.nodes);
1773 return numa_node_id();
1774 target = (unsigned int)n % nnodes;
1775 nid = first_node(pol->v.nodes);
1776 for (i = 0; i < target; i++)
1777 nid = next_node(nid, pol->v.nodes);
1781 /* Determine a node number for interleave */
1782 static inline unsigned interleave_nid(struct mempolicy *pol,
1783 struct vm_area_struct *vma, unsigned long addr, int shift)
1789 * for small pages, there is no difference between
1790 * shift and PAGE_SHIFT, so the bit-shift is safe.
1791 * for huge pages, since vm_pgoff is in units of small
1792 * pages, we need to shift off the always 0 bits to get
1795 BUG_ON(shift < PAGE_SHIFT);
1796 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1797 off += (addr - vma->vm_start) >> shift;
1798 return offset_il_node(pol, vma, off);
1800 return interleave_nodes(pol);
1803 #ifdef CONFIG_HUGETLBFS
1805 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1806 * @vma: virtual memory area whose policy is sought
1807 * @addr: address in @vma for shared policy lookup and interleave policy
1808 * @gfp_flags: for requested zone
1809 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1810 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1812 * Returns a zonelist suitable for a huge page allocation and a pointer
1813 * to the struct mempolicy for conditional unref after allocation.
1814 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1815 * @nodemask for filtering the zonelist.
1817 * Must be protected by read_mems_allowed_begin()
1819 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1820 gfp_t gfp_flags, struct mempolicy **mpol,
1821 nodemask_t **nodemask)
1823 struct zonelist *zl;
1825 *mpol = get_vma_policy(vma, addr);
1826 *nodemask = NULL; /* assume !MPOL_BIND */
1828 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1829 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1830 huge_page_shift(hstate_vma(vma))), gfp_flags);
1832 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1833 if ((*mpol)->mode == MPOL_BIND)
1834 *nodemask = &(*mpol)->v.nodes;
1840 * init_nodemask_of_mempolicy
1842 * If the current task's mempolicy is "default" [NULL], return 'false'
1843 * to indicate default policy. Otherwise, extract the policy nodemask
1844 * for 'bind' or 'interleave' policy into the argument nodemask, or
1845 * initialize the argument nodemask to contain the single node for
1846 * 'preferred' or 'local' policy and return 'true' to indicate presence
1847 * of non-default mempolicy.
1849 * We don't bother with reference counting the mempolicy [mpol_get/put]
1850 * because the current task is examining it's own mempolicy and a task's
1851 * mempolicy is only ever changed by the task itself.
1853 * N.B., it is the caller's responsibility to free a returned nodemask.
1855 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1857 struct mempolicy *mempolicy;
1860 if (!(mask && current->mempolicy))
1864 mempolicy = current->mempolicy;
1865 switch (mempolicy->mode) {
1866 case MPOL_PREFERRED:
1867 if (mempolicy->flags & MPOL_F_LOCAL)
1868 nid = numa_node_id();
1870 nid = mempolicy->v.preferred_node;
1871 init_nodemask_of_node(mask, nid);
1876 case MPOL_INTERLEAVE:
1877 *mask = mempolicy->v.nodes;
1883 task_unlock(current);
1890 * mempolicy_nodemask_intersects
1892 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1893 * policy. Otherwise, check for intersection between mask and the policy
1894 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1895 * policy, always return true since it may allocate elsewhere on fallback.
1897 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1899 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1900 const nodemask_t *mask)
1902 struct mempolicy *mempolicy;
1908 mempolicy = tsk->mempolicy;
1912 switch (mempolicy->mode) {
1913 case MPOL_PREFERRED:
1915 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1916 * allocate from, they may fallback to other nodes when oom.
1917 * Thus, it's possible for tsk to have allocated memory from
1922 case MPOL_INTERLEAVE:
1923 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1933 /* Allocate a page in interleaved policy.
1934 Own path because it needs to do special accounting. */
1935 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1938 struct zonelist *zl;
1941 zl = node_zonelist(nid, gfp);
1942 page = __alloc_pages(gfp, order, zl);
1943 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1944 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1949 * alloc_pages_vma - Allocate a page for a VMA.
1952 * %GFP_USER user allocation.
1953 * %GFP_KERNEL kernel allocations,
1954 * %GFP_HIGHMEM highmem/user allocations,
1955 * %GFP_FS allocation should not call back into a file system.
1956 * %GFP_ATOMIC don't sleep.
1958 * @order:Order of the GFP allocation.
1959 * @vma: Pointer to VMA or NULL if not available.
1960 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1961 * @node: Which node to prefer for allocation (modulo policy).
1962 * @hugepage: for hugepages try only the preferred node if possible
1964 * This function allocates a page from the kernel page pool and applies
1965 * a NUMA policy associated with the VMA or the current process.
1966 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1967 * mm_struct of the VMA to prevent it from going away. Should be used for
1968 * all allocations for pages that will be mapped into user space. Returns
1969 * NULL when no page can be allocated.
1972 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1973 unsigned long addr, int node, bool hugepage)
1975 struct mempolicy *pol;
1977 unsigned int cpuset_mems_cookie;
1978 struct zonelist *zl;
1982 pol = get_vma_policy(vma, addr);
1983 cpuset_mems_cookie = read_mems_allowed_begin();
1985 if (pol->mode == MPOL_INTERLEAVE) {
1988 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1990 page = alloc_page_interleave(gfp, order, nid);
1994 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1995 int hpage_node = node;
1998 * For hugepage allocation and non-interleave policy which
1999 * allows the current node (or other explicitly preferred
2000 * node) we only try to allocate from the current/preferred
2001 * node and don't fall back to other nodes, as the cost of
2002 * remote accesses would likely offset THP benefits.
2004 * If the policy is interleave, or does not allow the current
2005 * node in its nodemask, we allocate the standard way.
2007 if (pol->mode == MPOL_PREFERRED &&
2008 !(pol->flags & MPOL_F_LOCAL))
2009 hpage_node = pol->v.preferred_node;
2011 nmask = policy_nodemask(gfp, pol);
2012 if (!nmask || node_isset(hpage_node, *nmask)) {
2014 page = __alloc_pages_node(hpage_node,
2015 gfp | __GFP_THISNODE, order);
2020 nmask = policy_nodemask(gfp, pol);
2021 zl = policy_zonelist(gfp, pol, node);
2022 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2025 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2031 * alloc_pages_current - Allocate pages.
2034 * %GFP_USER user allocation,
2035 * %GFP_KERNEL kernel allocation,
2036 * %GFP_HIGHMEM highmem allocation,
2037 * %GFP_FS don't call back into a file system.
2038 * %GFP_ATOMIC don't sleep.
2039 * @order: Power of two of allocation size in pages. 0 is a single page.
2041 * Allocate a page from the kernel page pool. When not in
2042 * interrupt context and apply the current process NUMA policy.
2043 * Returns NULL when no page can be allocated.
2045 * Don't call cpuset_update_task_memory_state() unless
2046 * 1) it's ok to take cpuset_sem (can WAIT), and
2047 * 2) allocating for current task (not interrupt).
2049 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2051 struct mempolicy *pol = &default_policy;
2053 unsigned int cpuset_mems_cookie;
2055 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2056 pol = get_task_policy(current);
2059 cpuset_mems_cookie = read_mems_allowed_begin();
2062 * No reference counting needed for current->mempolicy
2063 * nor system default_policy
2065 if (pol->mode == MPOL_INTERLEAVE)
2066 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2068 page = __alloc_pages_nodemask(gfp, order,
2069 policy_zonelist(gfp, pol, numa_node_id()),
2070 policy_nodemask(gfp, pol));
2072 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2077 EXPORT_SYMBOL(alloc_pages_current);
2079 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2081 struct mempolicy *pol = mpol_dup(vma_policy(src));
2084 return PTR_ERR(pol);
2085 dst->vm_policy = pol;
2090 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2091 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2092 * with the mems_allowed returned by cpuset_mems_allowed(). This
2093 * keeps mempolicies cpuset relative after its cpuset moves. See
2094 * further kernel/cpuset.c update_nodemask().
2096 * current's mempolicy may be rebinded by the other task(the task that changes
2097 * cpuset's mems), so we needn't do rebind work for current task.
2100 /* Slow path of a mempolicy duplicate */
2101 struct mempolicy *__mpol_dup(struct mempolicy *old)
2103 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2106 return ERR_PTR(-ENOMEM);
2108 /* task's mempolicy is protected by alloc_lock */
2109 if (old == current->mempolicy) {
2112 task_unlock(current);
2116 if (current_cpuset_is_being_rebound()) {
2117 nodemask_t mems = cpuset_mems_allowed(current);
2118 if (new->flags & MPOL_F_REBINDING)
2119 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2121 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2123 atomic_set(&new->refcnt, 1);
2127 /* Slow path of a mempolicy comparison */
2128 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2132 if (a->mode != b->mode)
2134 if (a->flags != b->flags)
2136 if (mpol_store_user_nodemask(a))
2137 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2143 case MPOL_INTERLEAVE:
2144 return !!nodes_equal(a->v.nodes, b->v.nodes);
2145 case MPOL_PREFERRED:
2146 return a->v.preferred_node == b->v.preferred_node;
2154 * Shared memory backing store policy support.
2156 * Remember policies even when nobody has shared memory mapped.
2157 * The policies are kept in Red-Black tree linked from the inode.
2158 * They are protected by the sp->lock rwlock, which should be held
2159 * for any accesses to the tree.
2163 * lookup first element intersecting start-end. Caller holds sp->lock for
2164 * reading or for writing
2166 static struct sp_node *
2167 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2169 struct rb_node *n = sp->root.rb_node;
2172 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2174 if (start >= p->end)
2176 else if (end <= p->start)
2184 struct sp_node *w = NULL;
2185 struct rb_node *prev = rb_prev(n);
2188 w = rb_entry(prev, struct sp_node, nd);
2189 if (w->end <= start)
2193 return rb_entry(n, struct sp_node, nd);
2197 * Insert a new shared policy into the list. Caller holds sp->lock for
2200 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2202 struct rb_node **p = &sp->root.rb_node;
2203 struct rb_node *parent = NULL;
2208 nd = rb_entry(parent, struct sp_node, nd);
2209 if (new->start < nd->start)
2211 else if (new->end > nd->end)
2212 p = &(*p)->rb_right;
2216 rb_link_node(&new->nd, parent, p);
2217 rb_insert_color(&new->nd, &sp->root);
2218 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2219 new->policy ? new->policy->mode : 0);
2222 /* Find shared policy intersecting idx */
2224 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2226 struct mempolicy *pol = NULL;
2229 if (!sp->root.rb_node)
2231 read_lock(&sp->lock);
2232 sn = sp_lookup(sp, idx, idx+1);
2234 mpol_get(sn->policy);
2237 read_unlock(&sp->lock);
2241 static void sp_free(struct sp_node *n)
2243 mpol_put(n->policy);
2244 kmem_cache_free(sn_cache, n);
2248 * mpol_misplaced - check whether current page node is valid in policy
2250 * @page: page to be checked
2251 * @vma: vm area where page mapped
2252 * @addr: virtual address where page mapped
2254 * Lookup current policy node id for vma,addr and "compare to" page's
2258 * -1 - not misplaced, page is in the right node
2259 * node - node id where the page should be
2261 * Policy determination "mimics" alloc_page_vma().
2262 * Called from fault path where we know the vma and faulting address.
2264 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2266 struct mempolicy *pol;
2268 int curnid = page_to_nid(page);
2269 unsigned long pgoff;
2270 int thiscpu = raw_smp_processor_id();
2271 int thisnid = cpu_to_node(thiscpu);
2277 pol = get_vma_policy(vma, addr);
2278 if (!(pol->flags & MPOL_F_MOF))
2281 switch (pol->mode) {
2282 case MPOL_INTERLEAVE:
2283 BUG_ON(addr >= vma->vm_end);
2284 BUG_ON(addr < vma->vm_start);
2286 pgoff = vma->vm_pgoff;
2287 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2288 polnid = offset_il_node(pol, vma, pgoff);
2291 case MPOL_PREFERRED:
2292 if (pol->flags & MPOL_F_LOCAL)
2293 polnid = numa_node_id();
2295 polnid = pol->v.preferred_node;
2301 * allows binding to multiple nodes.
2302 * use current page if in policy nodemask,
2303 * else select nearest allowed node, if any.
2304 * If no allowed nodes, use current [!misplaced].
2306 if (node_isset(curnid, pol->v.nodes))
2308 z = first_zones_zonelist(
2309 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2310 gfp_zone(GFP_HIGHUSER),
2312 polnid = z->zone->node;
2319 /* Migrate the page towards the node whose CPU is referencing it */
2320 if (pol->flags & MPOL_F_MORON) {
2323 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2327 if (curnid != polnid)
2336 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2337 * dropped after task->mempolicy is set to NULL so that any allocation done as
2338 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2341 void mpol_put_task_policy(struct task_struct *task)
2343 struct mempolicy *pol;
2346 pol = task->mempolicy;
2347 task->mempolicy = NULL;
2352 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2354 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2355 rb_erase(&n->nd, &sp->root);
2359 static void sp_node_init(struct sp_node *node, unsigned long start,
2360 unsigned long end, struct mempolicy *pol)
2362 node->start = start;
2367 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2368 struct mempolicy *pol)
2371 struct mempolicy *newpol;
2373 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2377 newpol = mpol_dup(pol);
2378 if (IS_ERR(newpol)) {
2379 kmem_cache_free(sn_cache, n);
2382 newpol->flags |= MPOL_F_SHARED;
2383 sp_node_init(n, start, end, newpol);
2388 /* Replace a policy range. */
2389 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2390 unsigned long end, struct sp_node *new)
2393 struct sp_node *n_new = NULL;
2394 struct mempolicy *mpol_new = NULL;
2398 write_lock(&sp->lock);
2399 n = sp_lookup(sp, start, end);
2400 /* Take care of old policies in the same range. */
2401 while (n && n->start < end) {
2402 struct rb_node *next = rb_next(&n->nd);
2403 if (n->start >= start) {
2409 /* Old policy spanning whole new range. */
2414 *mpol_new = *n->policy;
2415 atomic_set(&mpol_new->refcnt, 1);
2416 sp_node_init(n_new, end, n->end, mpol_new);
2418 sp_insert(sp, n_new);
2427 n = rb_entry(next, struct sp_node, nd);
2431 write_unlock(&sp->lock);
2438 kmem_cache_free(sn_cache, n_new);
2443 write_unlock(&sp->lock);
2445 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2448 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2455 * mpol_shared_policy_init - initialize shared policy for inode
2456 * @sp: pointer to inode shared policy
2457 * @mpol: struct mempolicy to install
2459 * Install non-NULL @mpol in inode's shared policy rb-tree.
2460 * On entry, the current task has a reference on a non-NULL @mpol.
2461 * This must be released on exit.
2462 * This is called at get_inode() calls and we can use GFP_KERNEL.
2464 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2468 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2469 rwlock_init(&sp->lock);
2472 struct vm_area_struct pvma;
2473 struct mempolicy *new;
2474 NODEMASK_SCRATCH(scratch);
2478 /* contextualize the tmpfs mount point mempolicy */
2479 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2481 goto free_scratch; /* no valid nodemask intersection */
2484 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2485 task_unlock(current);
2489 /* Create pseudo-vma that contains just the policy */
2490 memset(&pvma, 0, sizeof(struct vm_area_struct));
2491 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2492 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2495 mpol_put(new); /* drop initial ref */
2497 NODEMASK_SCRATCH_FREE(scratch);
2499 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2503 int mpol_set_shared_policy(struct shared_policy *info,
2504 struct vm_area_struct *vma, struct mempolicy *npol)
2507 struct sp_node *new = NULL;
2508 unsigned long sz = vma_pages(vma);
2510 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2512 sz, npol ? npol->mode : -1,
2513 npol ? npol->flags : -1,
2514 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2517 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2521 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2527 /* Free a backing policy store on inode delete. */
2528 void mpol_free_shared_policy(struct shared_policy *p)
2531 struct rb_node *next;
2533 if (!p->root.rb_node)
2535 write_lock(&p->lock);
2536 next = rb_first(&p->root);
2538 n = rb_entry(next, struct sp_node, nd);
2539 next = rb_next(&n->nd);
2542 write_unlock(&p->lock);
2545 #ifdef CONFIG_NUMA_BALANCING
2546 static int __initdata numabalancing_override;
2548 static void __init check_numabalancing_enable(void)
2550 bool numabalancing_default = false;
2552 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2553 numabalancing_default = true;
2555 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2556 if (numabalancing_override)
2557 set_numabalancing_state(numabalancing_override == 1);
2559 if (num_online_nodes() > 1 && !numabalancing_override) {
2560 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2561 numabalancing_default ? "Enabling" : "Disabling");
2562 set_numabalancing_state(numabalancing_default);
2566 static int __init setup_numabalancing(char *str)
2572 if (!strcmp(str, "enable")) {
2573 numabalancing_override = 1;
2575 } else if (!strcmp(str, "disable")) {
2576 numabalancing_override = -1;
2581 pr_warn("Unable to parse numa_balancing=\n");
2585 __setup("numa_balancing=", setup_numabalancing);
2587 static inline void __init check_numabalancing_enable(void)
2590 #endif /* CONFIG_NUMA_BALANCING */
2592 /* assumes fs == KERNEL_DS */
2593 void __init numa_policy_init(void)
2595 nodemask_t interleave_nodes;
2596 unsigned long largest = 0;
2597 int nid, prefer = 0;
2599 policy_cache = kmem_cache_create("numa_policy",
2600 sizeof(struct mempolicy),
2601 0, SLAB_PANIC, NULL);
2603 sn_cache = kmem_cache_create("shared_policy_node",
2604 sizeof(struct sp_node),
2605 0, SLAB_PANIC, NULL);
2607 for_each_node(nid) {
2608 preferred_node_policy[nid] = (struct mempolicy) {
2609 .refcnt = ATOMIC_INIT(1),
2610 .mode = MPOL_PREFERRED,
2611 .flags = MPOL_F_MOF | MPOL_F_MORON,
2612 .v = { .preferred_node = nid, },
2617 * Set interleaving policy for system init. Interleaving is only
2618 * enabled across suitably sized nodes (default is >= 16MB), or
2619 * fall back to the largest node if they're all smaller.
2621 nodes_clear(interleave_nodes);
2622 for_each_node_state(nid, N_MEMORY) {
2623 unsigned long total_pages = node_present_pages(nid);
2625 /* Preserve the largest node */
2626 if (largest < total_pages) {
2627 largest = total_pages;
2631 /* Interleave this node? */
2632 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2633 node_set(nid, interleave_nodes);
2636 /* All too small, use the largest */
2637 if (unlikely(nodes_empty(interleave_nodes)))
2638 node_set(prefer, interleave_nodes);
2640 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2641 pr_err("%s: interleaving failed\n", __func__);
2643 check_numabalancing_enable();
2646 /* Reset policy of current process to default */
2647 void numa_default_policy(void)
2649 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2653 * Parse and format mempolicy from/to strings
2657 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2659 static const char * const policy_modes[] =
2661 [MPOL_DEFAULT] = "default",
2662 [MPOL_PREFERRED] = "prefer",
2663 [MPOL_BIND] = "bind",
2664 [MPOL_INTERLEAVE] = "interleave",
2665 [MPOL_LOCAL] = "local",
2671 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2672 * @str: string containing mempolicy to parse
2673 * @mpol: pointer to struct mempolicy pointer, returned on success.
2676 * <mode>[=<flags>][:<nodelist>]
2678 * On success, returns 0, else 1
2680 int mpol_parse_str(char *str, struct mempolicy **mpol)
2682 struct mempolicy *new = NULL;
2683 unsigned short mode;
2684 unsigned short mode_flags;
2686 char *nodelist = strchr(str, ':');
2687 char *flags = strchr(str, '=');
2691 /* NUL-terminate mode or flags string */
2693 if (nodelist_parse(nodelist, nodes))
2695 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2701 *flags++ = '\0'; /* terminate mode string */
2703 for (mode = 0; mode < MPOL_MAX; mode++) {
2704 if (!strcmp(str, policy_modes[mode])) {
2708 if (mode >= MPOL_MAX)
2712 case MPOL_PREFERRED:
2714 * Insist on a nodelist of one node only
2717 char *rest = nodelist;
2718 while (isdigit(*rest))
2724 case MPOL_INTERLEAVE:
2726 * Default to online nodes with memory if no nodelist
2729 nodes = node_states[N_MEMORY];
2733 * Don't allow a nodelist; mpol_new() checks flags
2737 mode = MPOL_PREFERRED;
2741 * Insist on a empty nodelist
2748 * Insist on a nodelist
2757 * Currently, we only support two mutually exclusive
2760 if (!strcmp(flags, "static"))
2761 mode_flags |= MPOL_F_STATIC_NODES;
2762 else if (!strcmp(flags, "relative"))
2763 mode_flags |= MPOL_F_RELATIVE_NODES;
2768 new = mpol_new(mode, mode_flags, &nodes);
2773 * Save nodes for mpol_to_str() to show the tmpfs mount options
2774 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2776 if (mode != MPOL_PREFERRED)
2777 new->v.nodes = nodes;
2779 new->v.preferred_node = first_node(nodes);
2781 new->flags |= MPOL_F_LOCAL;
2784 * Save nodes for contextualization: this will be used to "clone"
2785 * the mempolicy in a specific context [cpuset] at a later time.
2787 new->w.user_nodemask = nodes;
2792 /* Restore string for error message */
2801 #endif /* CONFIG_TMPFS */
2804 * mpol_to_str - format a mempolicy structure for printing
2805 * @buffer: to contain formatted mempolicy string
2806 * @maxlen: length of @buffer
2807 * @pol: pointer to mempolicy to be formatted
2809 * Convert @pol into a string. If @buffer is too short, truncate the string.
2810 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2811 * longest flag, "relative", and to display at least a few node ids.
2813 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2816 nodemask_t nodes = NODE_MASK_NONE;
2817 unsigned short mode = MPOL_DEFAULT;
2818 unsigned short flags = 0;
2820 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2828 case MPOL_PREFERRED:
2829 if (flags & MPOL_F_LOCAL)
2832 node_set(pol->v.preferred_node, nodes);
2835 case MPOL_INTERLEAVE:
2836 nodes = pol->v.nodes;
2840 snprintf(p, maxlen, "unknown");
2844 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2846 if (flags & MPOL_MODE_FLAGS) {
2847 p += snprintf(p, buffer + maxlen - p, "=");
2850 * Currently, the only defined flags are mutually exclusive
2852 if (flags & MPOL_F_STATIC_NODES)
2853 p += snprintf(p, buffer + maxlen - p, "static");
2854 else if (flags & MPOL_F_RELATIVE_NODES)
2855 p += snprintf(p, buffer + maxlen - p, "relative");
2858 if (!nodes_empty(nodes))
2859 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2860 nodemask_pr_args(&nodes));