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/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/export.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/ksm.h>
90 #include <linux/rmap.h>
91 #include <linux/security.h>
92 #include <linux/syscalls.h>
93 #include <linux/ctype.h>
94 #include <linux/mm_inline.h>
95 #include <linux/mmu_notifier.h>
96 #include <linux/printk.h>
98 #include <asm/tlbflush.h>
99 #include <asm/uaccess.h>
101 #include "internal.h"
104 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
105 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
107 static struct kmem_cache *policy_cache;
108 static struct kmem_cache *sn_cache;
110 /* Highest zone. An specific allocation for a zone below that is not
112 enum zone_type policy_zone = 0;
115 * run-time system-wide default policy => local allocation
117 static struct mempolicy default_policy = {
118 .refcnt = ATOMIC_INIT(1), /* never free it */
119 .mode = MPOL_PREFERRED,
120 .flags = MPOL_F_LOCAL,
123 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
125 struct mempolicy *get_task_policy(struct task_struct *p)
127 struct mempolicy *pol = p->mempolicy;
133 node = numa_node_id();
134 if (node != NUMA_NO_NODE) {
135 pol = &preferred_node_policy[node];
136 /* preferred_node_policy is not initialised early in boot */
141 return &default_policy;
144 static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * If we have a lock to protect task->mempolicy in read-side, we do
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
162 } mpol_ops[MPOL_MAX];
164 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
166 return pol->flags & MPOL_MODE_FLAGS;
169 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
170 const nodemask_t *rel)
173 nodes_fold(tmp, *orig, nodes_weight(*rel));
174 nodes_onto(*ret, tmp, *rel);
177 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
179 if (nodes_empty(*nodes))
181 pol->v.nodes = *nodes;
185 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
188 pol->flags |= MPOL_F_LOCAL; /* local allocation */
189 else if (nodes_empty(*nodes))
190 return -EINVAL; /* no allowed nodes */
192 pol->v.preferred_node = first_node(*nodes);
196 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
198 if (nodes_empty(*nodes))
200 pol->v.nodes = *nodes;
205 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
206 * any, for the new policy. mpol_new() has already validated the nodes
207 * parameter with respect to the policy mode and flags. But, we need to
208 * handle an empty nodemask with MPOL_PREFERRED here.
210 * Must be called holding task's alloc_lock to protect task's mems_allowed
211 * and mempolicy. May also be called holding the mmap_semaphore for write.
213 static int mpol_set_nodemask(struct mempolicy *pol,
214 const nodemask_t *nodes, struct nodemask_scratch *nsc)
218 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
222 nodes_and(nsc->mask1,
223 cpuset_current_mems_allowed, node_states[N_MEMORY]);
226 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
227 nodes = NULL; /* explicit local allocation */
229 if (pol->flags & MPOL_F_RELATIVE_NODES)
230 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
232 nodes_and(nsc->mask2, *nodes, nsc->mask1);
234 if (mpol_store_user_nodemask(pol))
235 pol->w.user_nodemask = *nodes;
237 pol->w.cpuset_mems_allowed =
238 cpuset_current_mems_allowed;
242 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
244 ret = mpol_ops[pol->mode].create(pol, NULL);
249 * This function just creates a new policy, does some check and simple
250 * initialization. You must invoke mpol_set_nodemask() to set nodes.
252 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
255 struct mempolicy *policy;
257 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
258 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
260 if (mode == MPOL_DEFAULT) {
261 if (nodes && !nodes_empty(*nodes))
262 return ERR_PTR(-EINVAL);
268 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
269 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
270 * All other modes require a valid pointer to a non-empty nodemask.
272 if (mode == MPOL_PREFERRED) {
273 if (nodes_empty(*nodes)) {
274 if (((flags & MPOL_F_STATIC_NODES) ||
275 (flags & MPOL_F_RELATIVE_NODES)))
276 return ERR_PTR(-EINVAL);
278 } else if (mode == MPOL_LOCAL) {
279 if (!nodes_empty(*nodes))
280 return ERR_PTR(-EINVAL);
281 mode = MPOL_PREFERRED;
282 } else if (nodes_empty(*nodes))
283 return ERR_PTR(-EINVAL);
284 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
286 return ERR_PTR(-ENOMEM);
287 atomic_set(&policy->refcnt, 1);
289 policy->flags = flags;
294 /* Slow path of a mpol destructor. */
295 void __mpol_put(struct mempolicy *p)
297 if (!atomic_dec_and_test(&p->refcnt))
299 kmem_cache_free(policy_cache, p);
302 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
303 enum mpol_rebind_step step)
309 * MPOL_REBIND_ONCE - do rebind work at once
310 * MPOL_REBIND_STEP1 - set all the newly nodes
311 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
313 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
314 enum mpol_rebind_step step)
318 if (pol->flags & MPOL_F_STATIC_NODES)
319 nodes_and(tmp, pol->w.user_nodemask, *nodes);
320 else if (pol->flags & MPOL_F_RELATIVE_NODES)
321 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
324 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
327 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
328 nodes_remap(tmp, pol->v.nodes,
329 pol->w.cpuset_mems_allowed, *nodes);
330 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
331 } else if (step == MPOL_REBIND_STEP2) {
332 tmp = pol->w.cpuset_mems_allowed;
333 pol->w.cpuset_mems_allowed = *nodes;
338 if (nodes_empty(tmp))
341 if (step == MPOL_REBIND_STEP1)
342 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
343 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
348 if (!node_isset(current->il_next, tmp)) {
349 current->il_next = next_node_in(current->il_next, tmp);
350 if (current->il_next >= MAX_NUMNODES)
351 current->il_next = numa_node_id();
355 static void mpol_rebind_preferred(struct mempolicy *pol,
356 const nodemask_t *nodes,
357 enum mpol_rebind_step step)
361 if (pol->flags & MPOL_F_STATIC_NODES) {
362 int node = first_node(pol->w.user_nodemask);
364 if (node_isset(node, *nodes)) {
365 pol->v.preferred_node = node;
366 pol->flags &= ~MPOL_F_LOCAL;
368 pol->flags |= MPOL_F_LOCAL;
369 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
370 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
371 pol->v.preferred_node = first_node(tmp);
372 } else if (!(pol->flags & MPOL_F_LOCAL)) {
373 pol->v.preferred_node = node_remap(pol->v.preferred_node,
374 pol->w.cpuset_mems_allowed,
376 pol->w.cpuset_mems_allowed = *nodes;
381 * mpol_rebind_policy - Migrate a policy to a different set of nodes
383 * If read-side task has no lock to protect task->mempolicy, write-side
384 * task will rebind the task->mempolicy by two step. The first step is
385 * setting all the newly nodes, and the second step is cleaning all the
386 * disallowed nodes. In this way, we can avoid finding no node to alloc
388 * If we have a lock to protect task->mempolicy in read-side, we do
392 * MPOL_REBIND_ONCE - do rebind work at once
393 * MPOL_REBIND_STEP1 - set all the newly nodes
394 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
396 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
397 enum mpol_rebind_step step)
399 if (!pol || pol->mode == MPOL_LOCAL)
401 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
402 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
405 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
408 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
411 if (step == MPOL_REBIND_STEP1)
412 pol->flags |= MPOL_F_REBINDING;
413 else if (step == MPOL_REBIND_STEP2)
414 pol->flags &= ~MPOL_F_REBINDING;
415 else if (step >= MPOL_REBIND_NSTEP)
418 mpol_ops[pol->mode].rebind(pol, newmask, step);
422 * Wrapper for mpol_rebind_policy() that just requires task
423 * pointer, and updates task mempolicy.
425 * Called with task's alloc_lock held.
428 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
429 enum mpol_rebind_step step)
431 mpol_rebind_policy(tsk->mempolicy, new, step);
435 * Rebind each vma in mm to new nodemask.
437 * Call holding a reference to mm. Takes mm->mmap_sem during call.
440 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
442 struct vm_area_struct *vma;
444 down_write(&mm->mmap_sem);
445 for (vma = mm->mmap; vma; vma = vma->vm_next)
446 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
447 up_write(&mm->mmap_sem);
450 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
452 .rebind = mpol_rebind_default,
454 [MPOL_INTERLEAVE] = {
455 .create = mpol_new_interleave,
456 .rebind = mpol_rebind_nodemask,
459 .create = mpol_new_preferred,
460 .rebind = mpol_rebind_preferred,
463 .create = mpol_new_bind,
464 .rebind = mpol_rebind_nodemask,
468 static void migrate_page_add(struct page *page, struct list_head *pagelist,
469 unsigned long flags);
472 struct list_head *pagelist;
475 struct vm_area_struct *prev;
479 * Scan through pages checking if pages follow certain conditions,
480 * and move them to the pagelist if they do.
482 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
483 unsigned long end, struct mm_walk *walk)
485 struct vm_area_struct *vma = walk->vma;
487 struct queue_pages *qp = walk->private;
488 unsigned long flags = qp->flags;
490 pte_t *pte, *mapped_pte;
493 if (pmd_trans_huge(*pmd)) {
494 ptl = pmd_lock(walk->mm, pmd);
495 if (pmd_trans_huge(*pmd)) {
496 page = pmd_page(*pmd);
497 if (is_huge_zero_page(page)) {
499 split_huge_pmd(vma, pmd, addr);
504 ret = split_huge_page(page);
515 if (pmd_trans_unstable(pmd))
518 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
519 for (; addr != end; pte++, addr += PAGE_SIZE) {
520 if (!pte_present(*pte))
522 page = vm_normal_page(vma, addr, *pte);
526 * vm_normal_page() filters out zero pages, but there might
527 * still be PageReserved pages to skip, perhaps in a VDSO.
529 if (PageReserved(page))
531 nid = page_to_nid(page);
532 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
534 if (PageTransCompound(page)) {
536 pte_unmap_unlock(pte, ptl);
538 ret = split_huge_page(page);
541 /* Failed to split -- skip. */
543 pte = pte_offset_map_lock(walk->mm, pmd,
550 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
551 if (!vma_migratable(vma))
553 migrate_page_add(page, qp->pagelist, flags);
557 pte_unmap_unlock(mapped_pte, ptl);
559 return addr != end ? -EIO : 0;
562 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
563 unsigned long addr, unsigned long end,
564 struct mm_walk *walk)
566 #ifdef CONFIG_HUGETLB_PAGE
567 struct queue_pages *qp = walk->private;
568 unsigned long flags = qp->flags;
574 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
575 entry = huge_ptep_get(pte);
576 if (!pte_present(entry))
578 page = pte_page(entry);
579 nid = page_to_nid(page);
580 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
582 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
583 if (flags & (MPOL_MF_MOVE_ALL) ||
584 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
585 isolate_huge_page(page, qp->pagelist);
594 #ifdef CONFIG_NUMA_BALANCING
596 * This is used to mark a range of virtual addresses to be inaccessible.
597 * These are later cleared by a NUMA hinting fault. Depending on these
598 * faults, pages may be migrated for better NUMA placement.
600 * This is assuming that NUMA faults are handled using PROT_NONE. If
601 * an architecture makes a different choice, it will need further
602 * changes to the core.
604 unsigned long change_prot_numa(struct vm_area_struct *vma,
605 unsigned long addr, unsigned long end)
609 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
611 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
616 static unsigned long change_prot_numa(struct vm_area_struct *vma,
617 unsigned long addr, unsigned long end)
621 #endif /* CONFIG_NUMA_BALANCING */
623 static int queue_pages_test_walk(unsigned long start, unsigned long end,
624 struct mm_walk *walk)
626 struct vm_area_struct *vma = walk->vma;
627 struct queue_pages *qp = walk->private;
628 unsigned long endvma = vma->vm_end;
629 unsigned long flags = qp->flags;
632 * Need check MPOL_MF_STRICT to return -EIO if possible
633 * regardless of vma_migratable
635 if (!vma_migratable(vma) &&
636 !(flags & MPOL_MF_STRICT))
641 if (vma->vm_start > start)
642 start = vma->vm_start;
644 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
645 if (!vma->vm_next && vma->vm_end < end)
647 if (qp->prev && qp->prev->vm_end < vma->vm_start)
653 if (flags & MPOL_MF_LAZY) {
654 /* Similar to task_numa_work, skip inaccessible VMAs */
655 if (!is_vm_hugetlb_page(vma) &&
656 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
657 !(vma->vm_flags & VM_MIXEDMAP))
658 change_prot_numa(vma, start, endvma);
662 /* queue pages from current vma */
663 if (flags & MPOL_MF_VALID)
669 * Walk through page tables and collect pages to be migrated.
671 * If pages found in a given range are on a set of nodes (determined by
672 * @nodes and @flags,) it's isolated and queued to the pagelist which is
673 * passed via @private.)
676 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
677 nodemask_t *nodes, unsigned long flags,
678 struct list_head *pagelist)
680 struct queue_pages qp = {
681 .pagelist = pagelist,
686 struct mm_walk queue_pages_walk = {
687 .hugetlb_entry = queue_pages_hugetlb,
688 .pmd_entry = queue_pages_pte_range,
689 .test_walk = queue_pages_test_walk,
694 return walk_page_range(start, end, &queue_pages_walk);
698 * Apply policy to a single VMA
699 * This must be called with the mmap_sem held for writing.
701 static int vma_replace_policy(struct vm_area_struct *vma,
702 struct mempolicy *pol)
705 struct mempolicy *old;
706 struct mempolicy *new;
708 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
709 vma->vm_start, vma->vm_end, vma->vm_pgoff,
710 vma->vm_ops, vma->vm_file,
711 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
717 if (vma->vm_ops && vma->vm_ops->set_policy) {
718 err = vma->vm_ops->set_policy(vma, new);
723 old = vma->vm_policy;
724 vma->vm_policy = new; /* protected by mmap_sem */
733 /* Step 2: apply policy to a range and do splits. */
734 static int mbind_range(struct mm_struct *mm, unsigned long start,
735 unsigned long end, struct mempolicy *new_pol)
737 struct vm_area_struct *prev;
738 struct vm_area_struct *vma;
741 unsigned long vmstart;
744 vma = find_vma(mm, start);
745 if (!vma || vma->vm_start > start)
749 if (start > vma->vm_start)
752 for (; vma && vma->vm_start < end; prev = vma, vma = vma->vm_next) {
753 vmstart = max(start, vma->vm_start);
754 vmend = min(end, vma->vm_end);
756 if (mpol_equal(vma_policy(vma), new_pol))
759 pgoff = vma->vm_pgoff +
760 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
761 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
762 vma->anon_vma, vma->vm_file, pgoff,
763 new_pol, vma->vm_userfaultfd_ctx);
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);
935 if (mpol_store_user_nodemask(pol)) {
936 *nmask = pol->w.user_nodemask;
939 get_policy_nodemask(pol, nmask);
940 task_unlock(current);
947 up_read(¤t->mm->mmap_sem);
951 #ifdef CONFIG_MIGRATION
955 static void migrate_page_add(struct page *page, struct list_head *pagelist,
959 * Avoid migrating a page that is shared with others.
961 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
962 if (!isolate_lru_page(page)) {
963 list_add_tail(&page->lru, pagelist);
964 inc_node_page_state(page, NR_ISOLATED_ANON +
965 page_is_file_cache(page));
970 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
973 return alloc_huge_page_node(page_hstate(compound_head(page)),
976 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
981 * Migrate pages from one node to a target node.
982 * Returns error or the number of pages not migrated.
984 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
992 node_set(source, nmask);
995 * This does not "check" the range but isolates all pages that
996 * need migration. Between passing in the full user address
997 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
999 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1000 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1001 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1003 if (!list_empty(&pagelist)) {
1004 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1005 MIGRATE_SYNC, MR_SYSCALL);
1007 putback_movable_pages(&pagelist);
1014 * Move pages between the two nodesets so as to preserve the physical
1015 * layout as much as possible.
1017 * Returns the number of page that could not be moved.
1019 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1020 const nodemask_t *to, int flags)
1026 err = migrate_prep();
1030 down_read(&mm->mmap_sem);
1033 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1034 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1035 * bit in 'tmp', and return that <source, dest> pair for migration.
1036 * The pair of nodemasks 'to' and 'from' define the map.
1038 * If no pair of bits is found that way, fallback to picking some
1039 * pair of 'source' and 'dest' bits that are not the same. If the
1040 * 'source' and 'dest' bits are the same, this represents a node
1041 * that will be migrating to itself, so no pages need move.
1043 * If no bits are left in 'tmp', or if all remaining bits left
1044 * in 'tmp' correspond to the same bit in 'to', return false
1045 * (nothing left to migrate).
1047 * This lets us pick a pair of nodes to migrate between, such that
1048 * if possible the dest node is not already occupied by some other
1049 * source node, minimizing the risk of overloading the memory on a
1050 * node that would happen if we migrated incoming memory to a node
1051 * before migrating outgoing memory source that same node.
1053 * A single scan of tmp is sufficient. As we go, we remember the
1054 * most recent <s, d> pair that moved (s != d). If we find a pair
1055 * that not only moved, but what's better, moved to an empty slot
1056 * (d is not set in tmp), then we break out then, with that pair.
1057 * Otherwise when we finish scanning from_tmp, we at least have the
1058 * most recent <s, d> pair that moved. If we get all the way through
1059 * the scan of tmp without finding any node that moved, much less
1060 * moved to an empty node, then there is nothing left worth migrating.
1064 while (!nodes_empty(tmp)) {
1066 int source = NUMA_NO_NODE;
1069 for_each_node_mask(s, tmp) {
1072 * do_migrate_pages() tries to maintain the relative
1073 * node relationship of the pages established between
1074 * threads and memory areas.
1076 * However if the number of source nodes is not equal to
1077 * the number of destination nodes we can not preserve
1078 * this node relative relationship. In that case, skip
1079 * copying memory from a node that is in the destination
1082 * Example: [2,3,4] -> [3,4,5] moves everything.
1083 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1086 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1087 (node_isset(s, *to)))
1090 d = node_remap(s, *from, *to);
1094 source = s; /* Node moved. Memorize */
1097 /* dest not in remaining from nodes? */
1098 if (!node_isset(dest, tmp))
1101 if (source == NUMA_NO_NODE)
1104 node_clear(source, tmp);
1105 err = migrate_to_node(mm, source, dest, flags);
1111 up_read(&mm->mmap_sem);
1119 * Allocate a new page for page migration based on vma policy.
1120 * Start by assuming the page is mapped by the same vma as contains @start.
1121 * Search forward from there, if not. N.B., this assumes that the
1122 * list of pages handed to migrate_pages()--which is how we get here--
1123 * is in virtual address order.
1125 static struct page *new_page(struct page *page, unsigned long start, int **x)
1127 struct vm_area_struct *vma;
1128 unsigned long uninitialized_var(address);
1130 vma = find_vma(current->mm, start);
1132 address = page_address_in_vma(page, vma);
1133 if (address != -EFAULT)
1138 if (PageHuge(page)) {
1140 return alloc_huge_page_noerr(vma, address, 1);
1143 * if !vma, alloc_page_vma() will use task or system default policy
1145 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1149 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1150 unsigned long flags)
1154 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1155 const nodemask_t *to, int flags)
1160 static struct page *new_page(struct page *page, unsigned long start, int **x)
1166 static long do_mbind(unsigned long start, unsigned long len,
1167 unsigned short mode, unsigned short mode_flags,
1168 nodemask_t *nmask, unsigned long flags)
1170 struct mm_struct *mm = current->mm;
1171 struct mempolicy *new;
1174 LIST_HEAD(pagelist);
1176 if (flags & ~(unsigned long)MPOL_MF_VALID)
1178 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1181 if (start & ~PAGE_MASK)
1184 if (mode == MPOL_DEFAULT)
1185 flags &= ~MPOL_MF_STRICT;
1187 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1195 new = mpol_new(mode, mode_flags, nmask);
1197 return PTR_ERR(new);
1199 if (flags & MPOL_MF_LAZY)
1200 new->flags |= MPOL_F_MOF;
1203 * If we are using the default policy then operation
1204 * on discontinuous address spaces is okay after all
1207 flags |= MPOL_MF_DISCONTIG_OK;
1209 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1210 start, start + len, mode, mode_flags,
1211 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1213 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1215 err = migrate_prep();
1220 NODEMASK_SCRATCH(scratch);
1222 down_write(&mm->mmap_sem);
1224 err = mpol_set_nodemask(new, nmask, scratch);
1225 task_unlock(current);
1227 up_write(&mm->mmap_sem);
1230 NODEMASK_SCRATCH_FREE(scratch);
1235 err = queue_pages_range(mm, start, end, nmask,
1236 flags | MPOL_MF_INVERT, &pagelist);
1238 err = mbind_range(mm, start, end, new);
1243 if (!list_empty(&pagelist)) {
1244 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1245 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1246 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1248 putback_movable_pages(&pagelist);
1251 if (nr_failed && (flags & MPOL_MF_STRICT))
1254 putback_movable_pages(&pagelist);
1256 up_write(&mm->mmap_sem);
1263 * User space interface with variable sized bitmaps for nodelists.
1266 /* Copy a node mask from user space. */
1267 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1268 unsigned long maxnode)
1272 unsigned long nlongs;
1273 unsigned long endmask;
1276 nodes_clear(*nodes);
1277 if (maxnode == 0 || !nmask)
1279 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1282 nlongs = BITS_TO_LONGS(maxnode);
1283 if ((maxnode % BITS_PER_LONG) == 0)
1286 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1289 * When the user specified more nodes than supported just check
1290 * if the non supported part is all zero.
1292 * If maxnode have more longs than MAX_NUMNODES, check
1293 * the bits in that area first. And then go through to
1294 * check the rest bits which equal or bigger than MAX_NUMNODES.
1295 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1297 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1298 if (nlongs > PAGE_SIZE/sizeof(long))
1300 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1301 if (get_user(t, nmask + k))
1303 if (k == nlongs - 1) {
1309 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1313 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1314 unsigned long valid_mask = endmask;
1316 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1317 if (get_user(t, nmask + nlongs - 1))
1323 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1325 nodes_addr(*nodes)[nlongs-1] &= endmask;
1329 /* Copy a kernel node mask to user space */
1330 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1333 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1334 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1336 if (copy > nbytes) {
1337 if (copy > PAGE_SIZE)
1339 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1343 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1346 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1347 unsigned long, mode, const unsigned long __user *, nmask,
1348 unsigned long, maxnode, unsigned, flags)
1352 unsigned short mode_flags;
1354 mode_flags = mode & MPOL_MODE_FLAGS;
1355 mode &= ~MPOL_MODE_FLAGS;
1356 if (mode >= MPOL_MAX)
1358 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1359 (mode_flags & MPOL_F_RELATIVE_NODES))
1361 err = get_nodes(&nodes, nmask, maxnode);
1364 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1367 /* Set the process memory policy */
1368 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1369 unsigned long, maxnode)
1373 unsigned short flags;
1375 flags = mode & MPOL_MODE_FLAGS;
1376 mode &= ~MPOL_MODE_FLAGS;
1377 if ((unsigned int)mode >= MPOL_MAX)
1379 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1381 err = get_nodes(&nodes, nmask, maxnode);
1384 return do_set_mempolicy(mode, flags, &nodes);
1387 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1388 const unsigned long __user *, old_nodes,
1389 const unsigned long __user *, new_nodes)
1391 const struct cred *cred = current_cred(), *tcred;
1392 struct mm_struct *mm = NULL;
1393 struct task_struct *task;
1394 nodemask_t task_nodes;
1398 NODEMASK_SCRATCH(scratch);
1403 old = &scratch->mask1;
1404 new = &scratch->mask2;
1406 err = get_nodes(old, old_nodes, maxnode);
1410 err = get_nodes(new, new_nodes, maxnode);
1414 /* Find the mm_struct */
1416 task = pid ? find_task_by_vpid(pid) : current;
1422 get_task_struct(task);
1427 * Check if this process has the right to modify the specified
1428 * process. The right exists if the process has administrative
1429 * capabilities, superuser privileges or the same
1430 * userid as the target process.
1432 tcred = __task_cred(task);
1433 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1434 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1435 !capable(CAP_SYS_NICE)) {
1442 task_nodes = cpuset_mems_allowed(task);
1443 /* Is the user allowed to access the target nodes? */
1444 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1449 task_nodes = cpuset_mems_allowed(current);
1450 nodes_and(*new, *new, task_nodes);
1451 if (nodes_empty(*new))
1454 nodes_and(*new, *new, node_states[N_MEMORY]);
1455 if (nodes_empty(*new))
1458 err = security_task_movememory(task);
1462 mm = get_task_mm(task);
1463 put_task_struct(task);
1470 err = do_migrate_pages(mm, old, new,
1471 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1475 NODEMASK_SCRATCH_FREE(scratch);
1480 put_task_struct(task);
1486 /* Retrieve NUMA policy */
1487 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1488 unsigned long __user *, nmask, unsigned long, maxnode,
1489 unsigned long, addr, unsigned long, flags)
1492 int uninitialized_var(pval);
1495 if (nmask != NULL && maxnode < nr_node_ids)
1498 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1503 if (policy && put_user(pval, policy))
1507 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1512 #ifdef CONFIG_COMPAT
1514 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1515 compat_ulong_t __user *, nmask,
1516 compat_ulong_t, maxnode,
1517 compat_ulong_t, addr, compat_ulong_t, flags)
1520 unsigned long __user *nm = NULL;
1521 unsigned long nr_bits, alloc_size;
1522 DECLARE_BITMAP(bm, MAX_NUMNODES);
1524 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1525 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1528 nm = compat_alloc_user_space(alloc_size);
1530 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1532 if (!err && nmask) {
1533 unsigned long copy_size;
1534 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1535 err = copy_from_user(bm, nm, copy_size);
1536 /* ensure entire bitmap is zeroed */
1537 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1538 err |= compat_put_bitmap(nmask, bm, nr_bits);
1544 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1545 compat_ulong_t, maxnode)
1547 unsigned long __user *nm = NULL;
1548 unsigned long nr_bits, alloc_size;
1549 DECLARE_BITMAP(bm, MAX_NUMNODES);
1551 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1552 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1555 if (compat_get_bitmap(bm, nmask, nr_bits))
1557 nm = compat_alloc_user_space(alloc_size);
1558 if (copy_to_user(nm, bm, alloc_size))
1562 return sys_set_mempolicy(mode, nm, nr_bits+1);
1565 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1566 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1567 compat_ulong_t, maxnode, compat_ulong_t, flags)
1569 unsigned long __user *nm = NULL;
1570 unsigned long nr_bits, alloc_size;
1573 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1574 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1577 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1579 nm = compat_alloc_user_space(alloc_size);
1580 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1584 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1589 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1592 struct mempolicy *pol = NULL;
1595 if (vma->vm_ops && vma->vm_ops->get_policy) {
1596 pol = vma->vm_ops->get_policy(vma, addr);
1597 } else if (vma->vm_policy) {
1598 pol = vma->vm_policy;
1601 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1602 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1603 * count on these policies which will be dropped by
1604 * mpol_cond_put() later
1606 if (mpol_needs_cond_ref(pol))
1615 * get_vma_policy(@vma, @addr)
1616 * @vma: virtual memory area whose policy is sought
1617 * @addr: address in @vma for shared policy lookup
1619 * Returns effective policy for a VMA at specified address.
1620 * Falls back to current->mempolicy or system default policy, as necessary.
1621 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1622 * count--added by the get_policy() vm_op, as appropriate--to protect against
1623 * freeing by another task. It is the caller's responsibility to free the
1624 * extra reference for shared policies.
1626 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1629 struct mempolicy *pol = __get_vma_policy(vma, addr);
1632 pol = get_task_policy(current);
1637 bool vma_policy_mof(struct vm_area_struct *vma)
1639 struct mempolicy *pol;
1641 if (vma->vm_ops && vma->vm_ops->get_policy) {
1644 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1645 if (pol && (pol->flags & MPOL_F_MOF))
1652 pol = vma->vm_policy;
1654 pol = get_task_policy(current);
1656 return pol->flags & MPOL_F_MOF;
1659 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1661 enum zone_type dynamic_policy_zone = policy_zone;
1663 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1666 * if policy->v.nodes has movable memory only,
1667 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1669 * policy->v.nodes is intersect with node_states[N_MEMORY].
1670 * so if the following test faile, it implies
1671 * policy->v.nodes has movable memory only.
1673 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1674 dynamic_policy_zone = ZONE_MOVABLE;
1676 return zone >= dynamic_policy_zone;
1680 * Return a nodemask representing a mempolicy for filtering nodes for
1683 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1685 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1686 if (unlikely(policy->mode == MPOL_BIND) &&
1687 apply_policy_zone(policy, gfp_zone(gfp)) &&
1688 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1689 return &policy->v.nodes;
1694 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1695 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1698 switch (policy->mode) {
1699 case MPOL_PREFERRED:
1700 if (!(policy->flags & MPOL_F_LOCAL))
1701 nd = policy->v.preferred_node;
1705 * Normally, MPOL_BIND allocations are node-local within the
1706 * allowed nodemask. However, if __GFP_THISNODE is set and the
1707 * current node isn't part of the mask, we use the zonelist for
1708 * the first node in the mask instead.
1710 if (unlikely(gfp & __GFP_THISNODE) &&
1711 unlikely(!node_isset(nd, policy->v.nodes)))
1712 nd = first_node(policy->v.nodes);
1717 return node_zonelist(nd, gfp);
1720 /* Do dynamic interleaving for a process */
1721 static unsigned interleave_nodes(struct mempolicy *policy)
1724 struct task_struct *me = current;
1727 next = next_node_in(nid, policy->v.nodes);
1728 if (next < MAX_NUMNODES)
1734 * Depending on the memory policy provide a node from which to allocate the
1737 unsigned int mempolicy_slab_node(void)
1739 struct mempolicy *policy;
1740 int node = numa_mem_id();
1745 policy = current->mempolicy;
1746 if (!policy || policy->flags & MPOL_F_LOCAL)
1749 switch (policy->mode) {
1750 case MPOL_PREFERRED:
1752 * handled MPOL_F_LOCAL above
1754 return policy->v.preferred_node;
1756 case MPOL_INTERLEAVE:
1757 return interleave_nodes(policy);
1763 * Follow bind policy behavior and start allocation at the
1766 struct zonelist *zonelist;
1767 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1768 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1769 z = first_zones_zonelist(zonelist, highest_zoneidx,
1771 return z->zone ? z->zone->node : node;
1780 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1781 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1782 * number of present nodes.
1784 static unsigned offset_il_node(struct mempolicy *pol,
1785 struct vm_area_struct *vma, unsigned long n)
1787 unsigned nnodes = nodes_weight(pol->v.nodes);
1793 return numa_node_id();
1794 target = (unsigned int)n % nnodes;
1795 nid = first_node(pol->v.nodes);
1796 for (i = 0; i < target; i++)
1797 nid = next_node(nid, pol->v.nodes);
1801 /* Determine a node number for interleave */
1802 static inline unsigned interleave_nid(struct mempolicy *pol,
1803 struct vm_area_struct *vma, unsigned long addr, int shift)
1809 * for small pages, there is no difference between
1810 * shift and PAGE_SHIFT, so the bit-shift is safe.
1811 * for huge pages, since vm_pgoff is in units of small
1812 * pages, we need to shift off the always 0 bits to get
1815 BUG_ON(shift < PAGE_SHIFT);
1816 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1817 off += (addr - vma->vm_start) >> shift;
1818 return offset_il_node(pol, vma, off);
1820 return interleave_nodes(pol);
1823 #ifdef CONFIG_HUGETLBFS
1825 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1826 * @vma: virtual memory area whose policy is sought
1827 * @addr: address in @vma for shared policy lookup and interleave policy
1828 * @gfp_flags: for requested zone
1829 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1830 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1832 * Returns a zonelist suitable for a huge page allocation and a pointer
1833 * to the struct mempolicy for conditional unref after allocation.
1834 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1835 * @nodemask for filtering the zonelist.
1837 * Must be protected by read_mems_allowed_begin()
1839 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1840 gfp_t gfp_flags, struct mempolicy **mpol,
1841 nodemask_t **nodemask)
1843 struct zonelist *zl;
1845 *mpol = get_vma_policy(vma, addr);
1846 *nodemask = NULL; /* assume !MPOL_BIND */
1848 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1849 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1850 huge_page_shift(hstate_vma(vma))), gfp_flags);
1852 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1853 if ((*mpol)->mode == MPOL_BIND)
1854 *nodemask = &(*mpol)->v.nodes;
1860 * init_nodemask_of_mempolicy
1862 * If the current task's mempolicy is "default" [NULL], return 'false'
1863 * to indicate default policy. Otherwise, extract the policy nodemask
1864 * for 'bind' or 'interleave' policy into the argument nodemask, or
1865 * initialize the argument nodemask to contain the single node for
1866 * 'preferred' or 'local' policy and return 'true' to indicate presence
1867 * of non-default mempolicy.
1869 * We don't bother with reference counting the mempolicy [mpol_get/put]
1870 * because the current task is examining it's own mempolicy and a task's
1871 * mempolicy is only ever changed by the task itself.
1873 * N.B., it is the caller's responsibility to free a returned nodemask.
1875 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1877 struct mempolicy *mempolicy;
1880 if (!(mask && current->mempolicy))
1884 mempolicy = current->mempolicy;
1885 switch (mempolicy->mode) {
1886 case MPOL_PREFERRED:
1887 if (mempolicy->flags & MPOL_F_LOCAL)
1888 nid = numa_node_id();
1890 nid = mempolicy->v.preferred_node;
1891 init_nodemask_of_node(mask, nid);
1896 case MPOL_INTERLEAVE:
1897 *mask = mempolicy->v.nodes;
1903 task_unlock(current);
1910 * mempolicy_nodemask_intersects
1912 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1913 * policy. Otherwise, check for intersection between mask and the policy
1914 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1915 * policy, always return true since it may allocate elsewhere on fallback.
1917 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1919 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1920 const nodemask_t *mask)
1922 struct mempolicy *mempolicy;
1928 mempolicy = tsk->mempolicy;
1932 switch (mempolicy->mode) {
1933 case MPOL_PREFERRED:
1935 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1936 * allocate from, they may fallback to other nodes when oom.
1937 * Thus, it's possible for tsk to have allocated memory from
1942 case MPOL_INTERLEAVE:
1943 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1953 /* Allocate a page in interleaved policy.
1954 Own path because it needs to do special accounting. */
1955 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1958 struct zonelist *zl;
1961 zl = node_zonelist(nid, gfp);
1962 page = __alloc_pages(gfp, order, zl);
1963 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1964 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1969 * alloc_pages_vma - Allocate a page for a VMA.
1972 * %GFP_USER user allocation.
1973 * %GFP_KERNEL kernel allocations,
1974 * %GFP_HIGHMEM highmem/user allocations,
1975 * %GFP_FS allocation should not call back into a file system.
1976 * %GFP_ATOMIC don't sleep.
1978 * @order:Order of the GFP allocation.
1979 * @vma: Pointer to VMA or NULL if not available.
1980 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1981 * @node: Which node to prefer for allocation (modulo policy).
1982 * @hugepage: for hugepages try only the preferred node if possible
1984 * This function allocates a page from the kernel page pool and applies
1985 * a NUMA policy associated with the VMA or the current process.
1986 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1987 * mm_struct of the VMA to prevent it from going away. Should be used for
1988 * all allocations for pages that will be mapped into user space. Returns
1989 * NULL when no page can be allocated.
1992 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1993 unsigned long addr, int node, bool hugepage)
1995 struct mempolicy *pol;
1997 unsigned int cpuset_mems_cookie;
1998 struct zonelist *zl;
2002 pol = get_vma_policy(vma, addr);
2003 cpuset_mems_cookie = read_mems_allowed_begin();
2005 if (pol->mode == MPOL_INTERLEAVE) {
2008 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2010 page = alloc_page_interleave(gfp, order, nid);
2014 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2015 int hpage_node = node;
2018 * For hugepage allocation and non-interleave policy which
2019 * allows the current node (or other explicitly preferred
2020 * node) we only try to allocate from the current/preferred
2021 * node and don't fall back to other nodes, as the cost of
2022 * remote accesses would likely offset THP benefits.
2024 * If the policy is interleave, or does not allow the current
2025 * node in its nodemask, we allocate the standard way.
2027 if (pol->mode == MPOL_PREFERRED &&
2028 !(pol->flags & MPOL_F_LOCAL))
2029 hpage_node = pol->v.preferred_node;
2031 nmask = policy_nodemask(gfp, pol);
2032 if (!nmask || node_isset(hpage_node, *nmask)) {
2035 * We cannot invoke reclaim if __GFP_THISNODE
2036 * is set. Invoking reclaim with
2037 * __GFP_THISNODE set, would cause THP
2038 * allocations to trigger heavy swapping
2039 * despite there may be tons of free memory
2040 * (including potentially plenty of THP
2041 * already available in the buddy) on all the
2044 * At most we could invoke compaction when
2045 * __GFP_THISNODE is set (but we would need to
2046 * refrain from invoking reclaim even if
2047 * compaction returned COMPACT_SKIPPED because
2048 * there wasn't not enough memory to succeed
2049 * compaction). For now just avoid
2050 * __GFP_THISNODE instead of limiting the
2051 * allocation path to a strict and single
2052 * compaction invocation.
2054 * Supposedly if direct reclaim was enabled by
2055 * the caller, the app prefers THP regardless
2056 * of the node it comes from so this would be
2057 * more desiderable behavior than only
2058 * providing THP originated from the local
2059 * node in such case.
2061 if (!(gfp & __GFP_DIRECT_RECLAIM))
2062 gfp |= __GFP_THISNODE;
2063 page = __alloc_pages_node(hpage_node, gfp, order);
2068 nmask = policy_nodemask(gfp, pol);
2069 zl = policy_zonelist(gfp, pol, node);
2070 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2073 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2079 * alloc_pages_current - Allocate pages.
2082 * %GFP_USER user allocation,
2083 * %GFP_KERNEL kernel allocation,
2084 * %GFP_HIGHMEM highmem allocation,
2085 * %GFP_FS don't call back into a file system.
2086 * %GFP_ATOMIC don't sleep.
2087 * @order: Power of two of allocation size in pages. 0 is a single page.
2089 * Allocate a page from the kernel page pool. When not in
2090 * interrupt context and apply the current process NUMA policy.
2091 * Returns NULL when no page can be allocated.
2093 * Don't call cpuset_update_task_memory_state() unless
2094 * 1) it's ok to take cpuset_sem (can WAIT), and
2095 * 2) allocating for current task (not interrupt).
2097 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2099 struct mempolicy *pol = &default_policy;
2101 unsigned int cpuset_mems_cookie;
2103 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2104 pol = get_task_policy(current);
2107 cpuset_mems_cookie = read_mems_allowed_begin();
2110 * No reference counting needed for current->mempolicy
2111 * nor system default_policy
2113 if (pol->mode == MPOL_INTERLEAVE)
2114 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2116 page = __alloc_pages_nodemask(gfp, order,
2117 policy_zonelist(gfp, pol, numa_node_id()),
2118 policy_nodemask(gfp, pol));
2120 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2125 EXPORT_SYMBOL(alloc_pages_current);
2127 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2129 struct mempolicy *pol = mpol_dup(vma_policy(src));
2132 return PTR_ERR(pol);
2133 dst->vm_policy = pol;
2138 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2139 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2140 * with the mems_allowed returned by cpuset_mems_allowed(). This
2141 * keeps mempolicies cpuset relative after its cpuset moves. See
2142 * further kernel/cpuset.c update_nodemask().
2144 * current's mempolicy may be rebinded by the other task(the task that changes
2145 * cpuset's mems), so we needn't do rebind work for current task.
2148 /* Slow path of a mempolicy duplicate */
2149 struct mempolicy *__mpol_dup(struct mempolicy *old)
2151 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2154 return ERR_PTR(-ENOMEM);
2156 /* task's mempolicy is protected by alloc_lock */
2157 if (old == current->mempolicy) {
2160 task_unlock(current);
2164 if (current_cpuset_is_being_rebound()) {
2165 nodemask_t mems = cpuset_mems_allowed(current);
2166 if (new->flags & MPOL_F_REBINDING)
2167 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2169 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2171 atomic_set(&new->refcnt, 1);
2175 /* Slow path of a mempolicy comparison */
2176 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2180 if (a->mode != b->mode)
2182 if (a->flags != b->flags)
2184 if (mpol_store_user_nodemask(a))
2185 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2191 case MPOL_INTERLEAVE:
2192 return !!nodes_equal(a->v.nodes, b->v.nodes);
2193 case MPOL_PREFERRED:
2194 /* a's ->flags is the same as b's */
2195 if (a->flags & MPOL_F_LOCAL)
2197 return a->v.preferred_node == b->v.preferred_node;
2205 * Shared memory backing store policy support.
2207 * Remember policies even when nobody has shared memory mapped.
2208 * The policies are kept in Red-Black tree linked from the inode.
2209 * They are protected by the sp->lock rwlock, which should be held
2210 * for any accesses to the tree.
2214 * lookup first element intersecting start-end. Caller holds sp->lock for
2215 * reading or for writing
2217 static struct sp_node *
2218 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2220 struct rb_node *n = sp->root.rb_node;
2223 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2225 if (start >= p->end)
2227 else if (end <= p->start)
2235 struct sp_node *w = NULL;
2236 struct rb_node *prev = rb_prev(n);
2239 w = rb_entry(prev, struct sp_node, nd);
2240 if (w->end <= start)
2244 return rb_entry(n, struct sp_node, nd);
2248 * Insert a new shared policy into the list. Caller holds sp->lock for
2251 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2253 struct rb_node **p = &sp->root.rb_node;
2254 struct rb_node *parent = NULL;
2259 nd = rb_entry(parent, struct sp_node, nd);
2260 if (new->start < nd->start)
2262 else if (new->end > nd->end)
2263 p = &(*p)->rb_right;
2267 rb_link_node(&new->nd, parent, p);
2268 rb_insert_color(&new->nd, &sp->root);
2269 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2270 new->policy ? new->policy->mode : 0);
2273 /* Find shared policy intersecting idx */
2275 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2277 struct mempolicy *pol = NULL;
2280 if (!sp->root.rb_node)
2282 read_lock(&sp->lock);
2283 sn = sp_lookup(sp, idx, idx+1);
2285 mpol_get(sn->policy);
2288 read_unlock(&sp->lock);
2292 static void sp_free(struct sp_node *n)
2294 mpol_put(n->policy);
2295 kmem_cache_free(sn_cache, n);
2299 * mpol_misplaced - check whether current page node is valid in policy
2301 * @page: page to be checked
2302 * @vma: vm area where page mapped
2303 * @addr: virtual address where page mapped
2305 * Lookup current policy node id for vma,addr and "compare to" page's
2309 * -1 - not misplaced, page is in the right node
2310 * node - node id where the page should be
2312 * Policy determination "mimics" alloc_page_vma().
2313 * Called from fault path where we know the vma and faulting address.
2315 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2317 struct mempolicy *pol;
2319 int curnid = page_to_nid(page);
2320 unsigned long pgoff;
2321 int thiscpu = raw_smp_processor_id();
2322 int thisnid = cpu_to_node(thiscpu);
2328 pol = get_vma_policy(vma, addr);
2329 if (!(pol->flags & MPOL_F_MOF))
2332 switch (pol->mode) {
2333 case MPOL_INTERLEAVE:
2334 BUG_ON(addr >= vma->vm_end);
2335 BUG_ON(addr < vma->vm_start);
2337 pgoff = vma->vm_pgoff;
2338 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2339 polnid = offset_il_node(pol, vma, pgoff);
2342 case MPOL_PREFERRED:
2343 if (pol->flags & MPOL_F_LOCAL)
2344 polnid = numa_node_id();
2346 polnid = pol->v.preferred_node;
2352 * allows binding to multiple nodes.
2353 * use current page if in policy nodemask,
2354 * else select nearest allowed node, if any.
2355 * If no allowed nodes, use current [!misplaced].
2357 if (node_isset(curnid, pol->v.nodes))
2359 z = first_zones_zonelist(
2360 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2361 gfp_zone(GFP_HIGHUSER),
2363 polnid = z->zone->node;
2370 /* Migrate the page towards the node whose CPU is referencing it */
2371 if (pol->flags & MPOL_F_MORON) {
2374 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2378 if (curnid != polnid)
2387 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2388 * dropped after task->mempolicy is set to NULL so that any allocation done as
2389 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2392 void mpol_put_task_policy(struct task_struct *task)
2394 struct mempolicy *pol;
2397 pol = task->mempolicy;
2398 task->mempolicy = NULL;
2403 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2405 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2406 rb_erase(&n->nd, &sp->root);
2410 static void sp_node_init(struct sp_node *node, unsigned long start,
2411 unsigned long end, struct mempolicy *pol)
2413 node->start = start;
2418 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2419 struct mempolicy *pol)
2422 struct mempolicy *newpol;
2424 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2428 newpol = mpol_dup(pol);
2429 if (IS_ERR(newpol)) {
2430 kmem_cache_free(sn_cache, n);
2433 newpol->flags |= MPOL_F_SHARED;
2434 sp_node_init(n, start, end, newpol);
2439 /* Replace a policy range. */
2440 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2441 unsigned long end, struct sp_node *new)
2444 struct sp_node *n_new = NULL;
2445 struct mempolicy *mpol_new = NULL;
2449 write_lock(&sp->lock);
2450 n = sp_lookup(sp, start, end);
2451 /* Take care of old policies in the same range. */
2452 while (n && n->start < end) {
2453 struct rb_node *next = rb_next(&n->nd);
2454 if (n->start >= start) {
2460 /* Old policy spanning whole new range. */
2465 *mpol_new = *n->policy;
2466 atomic_set(&mpol_new->refcnt, 1);
2467 sp_node_init(n_new, end, n->end, mpol_new);
2469 sp_insert(sp, n_new);
2478 n = rb_entry(next, struct sp_node, nd);
2482 write_unlock(&sp->lock);
2489 kmem_cache_free(sn_cache, n_new);
2494 write_unlock(&sp->lock);
2496 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2499 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2502 atomic_set(&mpol_new->refcnt, 1);
2507 * mpol_shared_policy_init - initialize shared policy for inode
2508 * @sp: pointer to inode shared policy
2509 * @mpol: struct mempolicy to install
2511 * Install non-NULL @mpol in inode's shared policy rb-tree.
2512 * On entry, the current task has a reference on a non-NULL @mpol.
2513 * This must be released on exit.
2514 * This is called at get_inode() calls and we can use GFP_KERNEL.
2516 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2520 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2521 rwlock_init(&sp->lock);
2524 struct vm_area_struct pvma;
2525 struct mempolicy *new;
2526 NODEMASK_SCRATCH(scratch);
2530 /* contextualize the tmpfs mount point mempolicy */
2531 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2533 goto free_scratch; /* no valid nodemask intersection */
2536 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2537 task_unlock(current);
2541 /* Create pseudo-vma that contains just the policy */
2542 memset(&pvma, 0, sizeof(struct vm_area_struct));
2543 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2544 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2547 mpol_put(new); /* drop initial ref */
2549 NODEMASK_SCRATCH_FREE(scratch);
2551 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2555 int mpol_set_shared_policy(struct shared_policy *info,
2556 struct vm_area_struct *vma, struct mempolicy *npol)
2559 struct sp_node *new = NULL;
2560 unsigned long sz = vma_pages(vma);
2562 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2564 sz, npol ? npol->mode : -1,
2565 npol ? npol->flags : -1,
2566 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2569 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2573 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2579 /* Free a backing policy store on inode delete. */
2580 void mpol_free_shared_policy(struct shared_policy *p)
2583 struct rb_node *next;
2585 if (!p->root.rb_node)
2587 write_lock(&p->lock);
2588 next = rb_first(&p->root);
2590 n = rb_entry(next, struct sp_node, nd);
2591 next = rb_next(&n->nd);
2594 write_unlock(&p->lock);
2597 #ifdef CONFIG_NUMA_BALANCING
2598 static int __initdata numabalancing_override;
2600 static void __init check_numabalancing_enable(void)
2602 bool numabalancing_default = false;
2604 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2605 numabalancing_default = true;
2607 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2608 if (numabalancing_override)
2609 set_numabalancing_state(numabalancing_override == 1);
2611 if (num_online_nodes() > 1 && !numabalancing_override) {
2612 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2613 numabalancing_default ? "Enabling" : "Disabling");
2614 set_numabalancing_state(numabalancing_default);
2618 static int __init setup_numabalancing(char *str)
2624 if (!strcmp(str, "enable")) {
2625 numabalancing_override = 1;
2627 } else if (!strcmp(str, "disable")) {
2628 numabalancing_override = -1;
2633 pr_warn("Unable to parse numa_balancing=\n");
2637 __setup("numa_balancing=", setup_numabalancing);
2639 static inline void __init check_numabalancing_enable(void)
2642 #endif /* CONFIG_NUMA_BALANCING */
2644 /* assumes fs == KERNEL_DS */
2645 void __init numa_policy_init(void)
2647 nodemask_t interleave_nodes;
2648 unsigned long largest = 0;
2649 int nid, prefer = 0;
2651 policy_cache = kmem_cache_create("numa_policy",
2652 sizeof(struct mempolicy),
2653 0, SLAB_PANIC, NULL);
2655 sn_cache = kmem_cache_create("shared_policy_node",
2656 sizeof(struct sp_node),
2657 0, SLAB_PANIC, NULL);
2659 for_each_node(nid) {
2660 preferred_node_policy[nid] = (struct mempolicy) {
2661 .refcnt = ATOMIC_INIT(1),
2662 .mode = MPOL_PREFERRED,
2663 .flags = MPOL_F_MOF | MPOL_F_MORON,
2664 .v = { .preferred_node = nid, },
2669 * Set interleaving policy for system init. Interleaving is only
2670 * enabled across suitably sized nodes (default is >= 16MB), or
2671 * fall back to the largest node if they're all smaller.
2673 nodes_clear(interleave_nodes);
2674 for_each_node_state(nid, N_MEMORY) {
2675 unsigned long total_pages = node_present_pages(nid);
2677 /* Preserve the largest node */
2678 if (largest < total_pages) {
2679 largest = total_pages;
2683 /* Interleave this node? */
2684 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2685 node_set(nid, interleave_nodes);
2688 /* All too small, use the largest */
2689 if (unlikely(nodes_empty(interleave_nodes)))
2690 node_set(prefer, interleave_nodes);
2692 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2693 pr_err("%s: interleaving failed\n", __func__);
2695 check_numabalancing_enable();
2698 /* Reset policy of current process to default */
2699 void numa_default_policy(void)
2701 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2705 * Parse and format mempolicy from/to strings
2709 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2711 static const char * const policy_modes[] =
2713 [MPOL_DEFAULT] = "default",
2714 [MPOL_PREFERRED] = "prefer",
2715 [MPOL_BIND] = "bind",
2716 [MPOL_INTERLEAVE] = "interleave",
2717 [MPOL_LOCAL] = "local",
2723 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2724 * @str: string containing mempolicy to parse
2725 * @mpol: pointer to struct mempolicy pointer, returned on success.
2728 * <mode>[=<flags>][:<nodelist>]
2730 * On success, returns 0, else 1
2732 int mpol_parse_str(char *str, struct mempolicy **mpol)
2734 struct mempolicy *new = NULL;
2735 unsigned short mode;
2736 unsigned short mode_flags;
2738 char *nodelist = strchr(str, ':');
2739 char *flags = strchr(str, '=');
2743 *flags++ = '\0'; /* terminate mode string */
2746 /* NUL-terminate mode or flags string */
2748 if (nodelist_parse(nodelist, nodes))
2750 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2755 for (mode = 0; mode < MPOL_MAX; mode++) {
2756 if (!strcmp(str, policy_modes[mode])) {
2760 if (mode >= MPOL_MAX)
2764 case MPOL_PREFERRED:
2766 * Insist on a nodelist of one node only, although later
2767 * we use first_node(nodes) to grab a single node, so here
2768 * nodelist (or nodes) cannot be empty.
2771 char *rest = nodelist;
2772 while (isdigit(*rest))
2776 if (nodes_empty(nodes))
2780 case MPOL_INTERLEAVE:
2782 * Default to online nodes with memory if no nodelist
2785 nodes = node_states[N_MEMORY];
2789 * Don't allow a nodelist; mpol_new() checks flags
2793 mode = MPOL_PREFERRED;
2797 * Insist on a empty nodelist
2804 * Insist on a nodelist
2813 * Currently, we only support two mutually exclusive
2816 if (!strcmp(flags, "static"))
2817 mode_flags |= MPOL_F_STATIC_NODES;
2818 else if (!strcmp(flags, "relative"))
2819 mode_flags |= MPOL_F_RELATIVE_NODES;
2824 new = mpol_new(mode, mode_flags, &nodes);
2829 * Save nodes for mpol_to_str() to show the tmpfs mount options
2830 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2832 if (mode != MPOL_PREFERRED)
2833 new->v.nodes = nodes;
2835 new->v.preferred_node = first_node(nodes);
2837 new->flags |= MPOL_F_LOCAL;
2840 * Save nodes for contextualization: this will be used to "clone"
2841 * the mempolicy in a specific context [cpuset] at a later time.
2843 new->w.user_nodemask = nodes;
2848 /* Restore string for error message */
2857 #endif /* CONFIG_TMPFS */
2860 * mpol_to_str - format a mempolicy structure for printing
2861 * @buffer: to contain formatted mempolicy string
2862 * @maxlen: length of @buffer
2863 * @pol: pointer to mempolicy to be formatted
2865 * Convert @pol into a string. If @buffer is too short, truncate the string.
2866 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2867 * longest flag, "relative", and to display at least a few node ids.
2869 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2872 nodemask_t nodes = NODE_MASK_NONE;
2873 unsigned short mode = MPOL_DEFAULT;
2874 unsigned short flags = 0;
2876 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2884 case MPOL_PREFERRED:
2885 if (flags & MPOL_F_LOCAL)
2888 node_set(pol->v.preferred_node, nodes);
2891 case MPOL_INTERLEAVE:
2892 nodes = pol->v.nodes;
2896 snprintf(p, maxlen, "unknown");
2900 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2902 if (flags & MPOL_MODE_FLAGS) {
2903 p += snprintf(p, buffer + maxlen - p, "=");
2906 * Currently, the only defined flags are mutually exclusive
2908 if (flags & MPOL_F_STATIC_NODES)
2909 p += snprintf(p, buffer + maxlen - p, "static");
2910 else if (flags & MPOL_F_RELATIVE_NODES)
2911 p += snprintf(p, buffer + maxlen - p, "relative");
2914 if (!nodes_empty(nodes))
2915 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2916 nodemask_pr_args(&nodes));