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 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
93 #include <linux/mmu_notifier.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
97 #include <linux/random.h>
102 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
105 static struct kmem_cache *policy_cache;
106 static struct kmem_cache *sn_cache;
108 /* Highest zone. An specific allocation for a zone below that is not
110 enum zone_type policy_zone = 0;
113 * run-time system-wide default policy => local allocation
115 static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
121 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
123 static struct mempolicy *get_task_policy(struct task_struct *p)
125 struct mempolicy *pol = p->mempolicy;
128 int node = numa_node_id();
130 if (node != NUMA_NO_NODE) {
131 pol = &preferred_node_policy[node];
133 * preferred_node_policy is not initialised early in
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 /* Check that the nodemask contains at least one populated zone */
165 static int is_valid_nodemask(const nodemask_t *nodemask)
167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
170 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
172 return pol->flags & MPOL_MODE_FLAGS;
175 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
183 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
185 if (nodes_empty(*nodes))
187 pol->v.nodes = *nodes;
191 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
198 pol->v.preferred_node = first_node(*nodes);
202 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
204 if (!is_valid_nodemask(nodes))
206 pol->v.nodes = *nodes;
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
219 static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
228 nodes_and(nsc->mask1,
229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
250 ret = mpol_ops[pol->mode].create(pol, NULL);
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
295 policy->flags = flags;
300 /* Slow path of a mpol destructor. */
301 void __mpol_put(struct mempolicy *p)
303 if (!atomic_dec_and_test(&p->refcnt))
305 kmem_cache_free(policy_cache, p);
308 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
319 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
344 if (nodes_empty(tmp))
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
363 static void mpol_rebind_preferred(struct mempolicy *pol,
364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
372 if (node_isset(node, *nodes)) {
373 pol->v.preferred_node = node;
374 pol->flags &= ~MPOL_F_LOCAL;
376 pol->flags |= MPOL_F_LOCAL;
377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
384 pol->w.cpuset_mems_allowed = *nodes;
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
396 * If we have a lock to protect task->mempolicy in read-side, we do
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
404 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
433 * Called with task's alloc_lock held.
436 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
439 mpol_rebind_policy(tsk->mempolicy, new, step);
443 * Rebind each vma in mm to new nodemask.
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
448 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
450 struct vm_area_struct *vma;
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
455 up_write(&mm->mmap_sem);
458 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
460 .rebind = mpol_rebind_default,
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
476 static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
483 static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
497 if (!pte_present(*pte))
499 page = vm_normal_page(vma, addr, *pte);
503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
506 if (PageReserved(page))
508 nid = page_to_nid(page);
509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
513 migrate_page_add(page, private, flags);
516 } while (pte++, addr += PAGE_SIZE, addr != end);
517 pte_unmap_unlock(orig_pte, ptl);
521 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
525 #ifdef CONFIG_HUGETLB_PAGE
530 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
531 page = pte_page(huge_ptep_get((pte_t *)pmd));
532 nid = page_to_nid(page);
533 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
535 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
536 if (flags & (MPOL_MF_MOVE_ALL) ||
537 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
538 isolate_huge_page(page, private);
546 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
547 unsigned long addr, unsigned long end,
548 const nodemask_t *nodes, unsigned long flags,
554 pmd = pmd_offset(pud, addr);
556 next = pmd_addr_end(addr, end);
557 if (!pmd_present(*pmd))
559 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
560 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
564 split_huge_page_pmd(vma, addr, pmd);
565 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
567 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
570 } while (pmd++, addr = next, addr != end);
574 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
575 unsigned long addr, unsigned long end,
576 const nodemask_t *nodes, unsigned long flags,
582 pud = pud_offset(pgd, addr);
584 next = pud_addr_end(addr, end);
585 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
587 if (pud_none_or_clear_bad(pud))
589 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
592 } while (pud++, addr = next, addr != end);
596 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
597 unsigned long addr, unsigned long end,
598 const nodemask_t *nodes, unsigned long flags,
604 pgd = pgd_offset(vma->vm_mm, addr);
606 next = pgd_addr_end(addr, end);
607 if (pgd_none_or_clear_bad(pgd))
609 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
612 } while (pgd++, addr = next, addr != end);
616 #ifdef CONFIG_NUMA_BALANCING
618 * This is used to mark a range of virtual addresses to be inaccessible.
619 * These are later cleared by a NUMA hinting fault. Depending on these
620 * faults, pages may be migrated for better NUMA placement.
622 * This is assuming that NUMA faults are handled using PROT_NONE. If
623 * an architecture makes a different choice, it will need further
624 * changes to the core.
626 unsigned long change_prot_numa(struct vm_area_struct *vma,
627 unsigned long addr, unsigned long end)
631 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
633 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
638 static unsigned long change_prot_numa(struct vm_area_struct *vma,
639 unsigned long addr, unsigned long end)
643 #endif /* CONFIG_NUMA_BALANCING */
646 * Walk through page tables and collect pages to be migrated.
648 * If pages found in a given range are on a set of nodes (determined by
649 * @nodes and @flags,) it's isolated and queued to the pagelist which is
650 * passed via @private.)
652 static struct vm_area_struct *
653 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
654 const nodemask_t *nodes, unsigned long flags, void *private)
657 struct vm_area_struct *first, *vma, *prev;
660 first = find_vma(mm, start);
662 return ERR_PTR(-EFAULT);
664 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
665 unsigned long endvma = vma->vm_end;
669 if (vma->vm_start > start)
670 start = vma->vm_start;
672 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
673 if (!vma->vm_next && vma->vm_end < end)
674 return ERR_PTR(-EFAULT);
675 if (prev && prev->vm_end < vma->vm_start)
676 return ERR_PTR(-EFAULT);
679 if (flags & MPOL_MF_LAZY) {
680 change_prot_numa(vma, start, endvma);
684 if ((flags & MPOL_MF_STRICT) ||
685 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
686 vma_migratable(vma))) {
688 err = queue_pages_pgd_range(vma, start, endvma, nodes,
691 first = ERR_PTR(err);
702 * Apply policy to a single VMA
703 * This must be called with the mmap_sem held for writing.
705 static int vma_replace_policy(struct vm_area_struct *vma,
706 struct mempolicy *pol)
709 struct mempolicy *old;
710 struct mempolicy *new;
712 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
713 vma->vm_start, vma->vm_end, vma->vm_pgoff,
714 vma->vm_ops, vma->vm_file,
715 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
721 if (vma->vm_ops && vma->vm_ops->set_policy) {
722 err = vma->vm_ops->set_policy(vma, new);
727 old = vma->vm_policy;
728 vma->vm_policy = new; /* protected by mmap_sem */
737 /* Step 2: apply policy to a range and do splits. */
738 static int mbind_range(struct mm_struct *mm, unsigned long start,
739 unsigned long end, struct mempolicy *new_pol)
741 struct vm_area_struct *next;
742 struct vm_area_struct *prev;
743 struct vm_area_struct *vma;
746 unsigned long vmstart;
749 vma = find_vma(mm, start);
750 if (!vma || vma->vm_start > start)
754 if (start > vma->vm_start)
757 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
759 vmstart = max(start, vma->vm_start);
760 vmend = min(end, vma->vm_end);
762 if (mpol_equal(vma_policy(vma), new_pol))
765 pgoff = vma->vm_pgoff +
766 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
767 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
768 vma->anon_vma, vma->vm_file, pgoff,
773 if (mpol_equal(vma_policy(vma), new_pol))
775 /* vma_merge() joined vma && vma->next, case 8 */
778 if (vma->vm_start != vmstart) {
779 err = split_vma(vma->vm_mm, vma, vmstart, 1);
783 if (vma->vm_end != vmend) {
784 err = split_vma(vma->vm_mm, vma, vmend, 0);
789 err = vma_replace_policy(vma, new_pol);
799 * Update task->flags PF_MEMPOLICY bit: set iff non-default
800 * mempolicy. Allows more rapid checking of this (combined perhaps
801 * with other PF_* flag bits) on memory allocation hot code paths.
803 * If called from outside this file, the task 'p' should -only- be
804 * a newly forked child not yet visible on the task list, because
805 * manipulating the task flags of a visible task is not safe.
807 * The above limitation is why this routine has the funny name
808 * mpol_fix_fork_child_flag().
810 * It is also safe to call this with a task pointer of current,
811 * which the static wrapper mpol_set_task_struct_flag() does,
812 * for use within this file.
815 void mpol_fix_fork_child_flag(struct task_struct *p)
818 p->flags |= PF_MEMPOLICY;
820 p->flags &= ~PF_MEMPOLICY;
823 static void mpol_set_task_struct_flag(void)
825 mpol_fix_fork_child_flag(current);
828 /* Set the process memory policy */
829 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
832 struct mempolicy *new, *old;
833 struct mm_struct *mm = current->mm;
834 NODEMASK_SCRATCH(scratch);
840 new = mpol_new(mode, flags, nodes);
846 * prevent changing our mempolicy while show_numa_maps()
848 * Note: do_set_mempolicy() can be called at init time
852 down_write(&mm->mmap_sem);
854 ret = mpol_set_nodemask(new, nodes, scratch);
856 task_unlock(current);
858 up_write(&mm->mmap_sem);
862 old = current->mempolicy;
863 current->mempolicy = new;
864 mpol_set_task_struct_flag();
865 if (new && new->mode == MPOL_INTERLEAVE &&
866 nodes_weight(new->v.nodes))
867 current->il_next = first_node(new->v.nodes);
868 task_unlock(current);
870 up_write(&mm->mmap_sem);
875 NODEMASK_SCRATCH_FREE(scratch);
880 * Return nodemask for policy for get_mempolicy() query
882 * Called with task's alloc_lock held
884 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
887 if (p == &default_policy)
893 case MPOL_INTERLEAVE:
897 if (!(p->flags & MPOL_F_LOCAL))
898 node_set(p->v.preferred_node, *nodes);
899 /* else return empty node mask for local allocation */
906 static int lookup_node(struct mm_struct *mm, unsigned long addr)
911 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
913 err = page_to_nid(p);
919 /* Retrieve NUMA policy */
920 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
921 unsigned long addr, unsigned long flags)
924 struct mm_struct *mm = current->mm;
925 struct vm_area_struct *vma = NULL;
926 struct mempolicy *pol = current->mempolicy;
929 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
932 if (flags & MPOL_F_MEMS_ALLOWED) {
933 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
935 *policy = 0; /* just so it's initialized */
937 *nmask = cpuset_current_mems_allowed;
938 task_unlock(current);
942 if (flags & MPOL_F_ADDR) {
944 * Do NOT fall back to task policy if the
945 * vma/shared policy at addr is NULL. We
946 * want to return MPOL_DEFAULT in this case.
948 down_read(&mm->mmap_sem);
949 vma = find_vma_intersection(mm, addr, addr+1);
951 up_read(&mm->mmap_sem);
954 if (vma->vm_ops && vma->vm_ops->get_policy)
955 pol = vma->vm_ops->get_policy(vma, addr);
957 pol = vma->vm_policy;
962 pol = &default_policy; /* indicates default behavior */
964 if (flags & MPOL_F_NODE) {
965 if (flags & MPOL_F_ADDR) {
966 err = lookup_node(mm, addr);
970 } else if (pol == current->mempolicy &&
971 pol->mode == MPOL_INTERLEAVE) {
972 *policy = current->il_next;
978 *policy = pol == &default_policy ? MPOL_DEFAULT :
981 * Internal mempolicy flags must be masked off before exposing
982 * the policy to userspace.
984 *policy |= (pol->flags & MPOL_MODE_FLAGS);
988 up_read(¤t->mm->mmap_sem);
994 if (mpol_store_user_nodemask(pol)) {
995 *nmask = pol->w.user_nodemask;
998 get_policy_nodemask(pol, nmask);
999 task_unlock(current);
1006 up_read(¤t->mm->mmap_sem);
1010 #ifdef CONFIG_MIGRATION
1014 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1015 unsigned long flags)
1018 * Avoid migrating a page that is shared with others.
1020 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1021 if (!isolate_lru_page(page)) {
1022 list_add_tail(&page->lru, pagelist);
1023 inc_zone_page_state(page, NR_ISOLATED_ANON +
1024 page_is_file_cache(page));
1029 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1032 return alloc_huge_page_node(page_hstate(compound_head(page)),
1035 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1039 * Migrate pages from one node to a target node.
1040 * Returns error or the number of pages not migrated.
1042 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1046 LIST_HEAD(pagelist);
1050 node_set(source, nmask);
1053 * This does not "check" the range but isolates all pages that
1054 * need migration. Between passing in the full user address
1055 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1057 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1058 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1059 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1061 if (!list_empty(&pagelist)) {
1062 err = migrate_pages(&pagelist, new_node_page, dest,
1063 MIGRATE_SYNC, MR_SYSCALL);
1065 putback_movable_pages(&pagelist);
1072 * Move pages between the two nodesets so as to preserve the physical
1073 * layout as much as possible.
1075 * Returns the number of page that could not be moved.
1077 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1078 const nodemask_t *to, int flags)
1084 err = migrate_prep();
1088 down_read(&mm->mmap_sem);
1090 err = migrate_vmas(mm, from, to, flags);
1095 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1096 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1097 * bit in 'tmp', and return that <source, dest> pair for migration.
1098 * The pair of nodemasks 'to' and 'from' define the map.
1100 * If no pair of bits is found that way, fallback to picking some
1101 * pair of 'source' and 'dest' bits that are not the same. If the
1102 * 'source' and 'dest' bits are the same, this represents a node
1103 * that will be migrating to itself, so no pages need move.
1105 * If no bits are left in 'tmp', or if all remaining bits left
1106 * in 'tmp' correspond to the same bit in 'to', return false
1107 * (nothing left to migrate).
1109 * This lets us pick a pair of nodes to migrate between, such that
1110 * if possible the dest node is not already occupied by some other
1111 * source node, minimizing the risk of overloading the memory on a
1112 * node that would happen if we migrated incoming memory to a node
1113 * before migrating outgoing memory source that same node.
1115 * A single scan of tmp is sufficient. As we go, we remember the
1116 * most recent <s, d> pair that moved (s != d). If we find a pair
1117 * that not only moved, but what's better, moved to an empty slot
1118 * (d is not set in tmp), then we break out then, with that pair.
1119 * Otherwise when we finish scanning from_tmp, we at least have the
1120 * most recent <s, d> pair that moved. If we get all the way through
1121 * the scan of tmp without finding any node that moved, much less
1122 * moved to an empty node, then there is nothing left worth migrating.
1126 while (!nodes_empty(tmp)) {
1128 int source = NUMA_NO_NODE;
1131 for_each_node_mask(s, tmp) {
1134 * do_migrate_pages() tries to maintain the relative
1135 * node relationship of the pages established between
1136 * threads and memory areas.
1138 * However if the number of source nodes is not equal to
1139 * the number of destination nodes we can not preserve
1140 * this node relative relationship. In that case, skip
1141 * copying memory from a node that is in the destination
1144 * Example: [2,3,4] -> [3,4,5] moves everything.
1145 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1148 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1149 (node_isset(s, *to)))
1152 d = node_remap(s, *from, *to);
1156 source = s; /* Node moved. Memorize */
1159 /* dest not in remaining from nodes? */
1160 if (!node_isset(dest, tmp))
1163 if (source == NUMA_NO_NODE)
1166 node_clear(source, tmp);
1167 err = migrate_to_node(mm, source, dest, flags);
1174 up_read(&mm->mmap_sem);
1182 * Allocate a new page for page migration based on vma policy.
1183 * Start assuming that page is mapped by vma pointed to by @private.
1184 * Search forward from there, if not. N.B., this assumes that the
1185 * list of pages handed to migrate_pages()--which is how we get here--
1186 * is in virtual address order.
1188 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1190 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1191 unsigned long uninitialized_var(address);
1194 address = page_address_in_vma(page, vma);
1195 if (address != -EFAULT)
1200 if (PageHuge(page)) {
1202 return alloc_huge_page_noerr(vma, address, 1);
1207 * if !vma, alloc_page_vma() will use task or system default policy
1209 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1213 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1214 unsigned long flags)
1218 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1219 const nodemask_t *to, int flags)
1224 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1230 static long do_mbind(unsigned long start, unsigned long len,
1231 unsigned short mode, unsigned short mode_flags,
1232 nodemask_t *nmask, unsigned long flags)
1234 struct vm_area_struct *vma;
1235 struct mm_struct *mm = current->mm;
1236 struct mempolicy *new;
1239 LIST_HEAD(pagelist);
1241 if (flags & ~(unsigned long)MPOL_MF_VALID)
1243 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1246 if (start & ~PAGE_MASK)
1249 if (mode == MPOL_DEFAULT)
1250 flags &= ~MPOL_MF_STRICT;
1252 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1260 new = mpol_new(mode, mode_flags, nmask);
1262 return PTR_ERR(new);
1264 if (flags & MPOL_MF_LAZY)
1265 new->flags |= MPOL_F_MOF;
1268 * If we are using the default policy then operation
1269 * on discontinuous address spaces is okay after all
1272 flags |= MPOL_MF_DISCONTIG_OK;
1274 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1275 start, start + len, mode, mode_flags,
1276 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1278 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1280 err = migrate_prep();
1285 NODEMASK_SCRATCH(scratch);
1287 down_write(&mm->mmap_sem);
1289 err = mpol_set_nodemask(new, nmask, scratch);
1290 task_unlock(current);
1292 up_write(&mm->mmap_sem);
1295 NODEMASK_SCRATCH_FREE(scratch);
1300 vma = queue_pages_range(mm, start, end, nmask,
1301 flags | MPOL_MF_INVERT, &pagelist);
1303 err = PTR_ERR(vma); /* maybe ... */
1305 err = mbind_range(mm, start, end, new);
1310 if (!list_empty(&pagelist)) {
1311 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1312 nr_failed = migrate_pages(&pagelist, new_vma_page,
1314 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1316 putback_movable_pages(&pagelist);
1319 if (nr_failed && (flags & MPOL_MF_STRICT))
1322 putback_movable_pages(&pagelist);
1324 up_write(&mm->mmap_sem);
1331 * User space interface with variable sized bitmaps for nodelists.
1334 /* Copy a node mask from user space. */
1335 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1336 unsigned long maxnode)
1339 unsigned long nlongs;
1340 unsigned long endmask;
1343 nodes_clear(*nodes);
1344 if (maxnode == 0 || !nmask)
1346 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1349 nlongs = BITS_TO_LONGS(maxnode);
1350 if ((maxnode % BITS_PER_LONG) == 0)
1353 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1355 /* When the user specified more nodes than supported just check
1356 if the non supported part is all zero. */
1357 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1358 if (nlongs > PAGE_SIZE/sizeof(long))
1360 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1362 if (get_user(t, nmask + k))
1364 if (k == nlongs - 1) {
1370 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1374 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1376 nodes_addr(*nodes)[nlongs-1] &= endmask;
1380 /* Copy a kernel node mask to user space */
1381 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1384 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1385 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1387 if (copy > nbytes) {
1388 if (copy > PAGE_SIZE)
1390 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1394 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1397 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1398 unsigned long, mode, unsigned long __user *, nmask,
1399 unsigned long, maxnode, unsigned, flags)
1403 unsigned short mode_flags;
1405 mode_flags = mode & MPOL_MODE_FLAGS;
1406 mode &= ~MPOL_MODE_FLAGS;
1407 if (mode >= MPOL_MAX)
1409 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1410 (mode_flags & MPOL_F_RELATIVE_NODES))
1412 err = get_nodes(&nodes, nmask, maxnode);
1415 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1418 /* Set the process memory policy */
1419 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1420 unsigned long, maxnode)
1424 unsigned short flags;
1426 flags = mode & MPOL_MODE_FLAGS;
1427 mode &= ~MPOL_MODE_FLAGS;
1428 if ((unsigned int)mode >= MPOL_MAX)
1430 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1432 err = get_nodes(&nodes, nmask, maxnode);
1435 return do_set_mempolicy(mode, flags, &nodes);
1438 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1439 const unsigned long __user *, old_nodes,
1440 const unsigned long __user *, new_nodes)
1442 const struct cred *cred = current_cred(), *tcred;
1443 struct mm_struct *mm = NULL;
1444 struct task_struct *task;
1445 nodemask_t task_nodes;
1449 NODEMASK_SCRATCH(scratch);
1454 old = &scratch->mask1;
1455 new = &scratch->mask2;
1457 err = get_nodes(old, old_nodes, maxnode);
1461 err = get_nodes(new, new_nodes, maxnode);
1465 /* Find the mm_struct */
1467 task = pid ? find_task_by_vpid(pid) : current;
1473 get_task_struct(task);
1478 * Check if this process has the right to modify the specified
1479 * process. The right exists if the process has administrative
1480 * capabilities, superuser privileges or the same
1481 * userid as the target process.
1483 tcred = __task_cred(task);
1484 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1485 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1486 !capable(CAP_SYS_NICE)) {
1493 task_nodes = cpuset_mems_allowed(task);
1494 /* Is the user allowed to access the target nodes? */
1495 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1500 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1505 err = security_task_movememory(task);
1509 mm = get_task_mm(task);
1510 put_task_struct(task);
1517 err = do_migrate_pages(mm, old, new,
1518 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1522 NODEMASK_SCRATCH_FREE(scratch);
1527 put_task_struct(task);
1533 /* Retrieve NUMA policy */
1534 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1535 unsigned long __user *, nmask, unsigned long, maxnode,
1536 unsigned long, addr, unsigned long, flags)
1539 int uninitialized_var(pval);
1542 if (nmask != NULL && maxnode < MAX_NUMNODES)
1545 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1550 if (policy && put_user(pval, policy))
1554 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1559 #ifdef CONFIG_COMPAT
1561 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1562 compat_ulong_t __user *nmask,
1563 compat_ulong_t maxnode,
1564 compat_ulong_t addr, compat_ulong_t flags)
1567 unsigned long __user *nm = NULL;
1568 unsigned long nr_bits, alloc_size;
1569 DECLARE_BITMAP(bm, MAX_NUMNODES);
1571 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1572 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1575 nm = compat_alloc_user_space(alloc_size);
1577 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1579 if (!err && nmask) {
1580 unsigned long copy_size;
1581 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1582 err = copy_from_user(bm, nm, copy_size);
1583 /* ensure entire bitmap is zeroed */
1584 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1585 err |= compat_put_bitmap(nmask, bm, nr_bits);
1591 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1592 compat_ulong_t maxnode)
1595 unsigned long __user *nm = NULL;
1596 unsigned long nr_bits, alloc_size;
1597 DECLARE_BITMAP(bm, MAX_NUMNODES);
1599 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1600 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1603 err = compat_get_bitmap(bm, nmask, nr_bits);
1604 nm = compat_alloc_user_space(alloc_size);
1605 err |= copy_to_user(nm, bm, alloc_size);
1611 return sys_set_mempolicy(mode, nm, nr_bits+1);
1614 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1615 compat_ulong_t mode, compat_ulong_t __user *nmask,
1616 compat_ulong_t maxnode, compat_ulong_t flags)
1619 unsigned long __user *nm = NULL;
1620 unsigned long nr_bits, alloc_size;
1623 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1624 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1627 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1628 nm = compat_alloc_user_space(alloc_size);
1629 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1635 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1641 * get_vma_policy(@task, @vma, @addr)
1642 * @task - task for fallback if vma policy == default
1643 * @vma - virtual memory area whose policy is sought
1644 * @addr - address in @vma for shared policy lookup
1646 * Returns effective policy for a VMA at specified address.
1647 * Falls back to @task or system default policy, as necessary.
1648 * Current or other task's task mempolicy and non-shared vma policies must be
1649 * protected by task_lock(task) by the caller.
1650 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1651 * count--added by the get_policy() vm_op, as appropriate--to protect against
1652 * freeing by another task. It is the caller's responsibility to free the
1653 * extra reference for shared policies.
1655 struct mempolicy *get_vma_policy(struct task_struct *task,
1656 struct vm_area_struct *vma, unsigned long addr)
1658 struct mempolicy *pol = get_task_policy(task);
1661 if (vma->vm_ops && vma->vm_ops->get_policy) {
1662 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1666 } else if (vma->vm_policy) {
1667 pol = vma->vm_policy;
1670 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1671 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1672 * count on these policies which will be dropped by
1673 * mpol_cond_put() later
1675 if (mpol_needs_cond_ref(pol))
1680 pol = &default_policy;
1684 bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1686 struct mempolicy *pol = get_task_policy(task);
1688 if (vma->vm_ops && vma->vm_ops->get_policy) {
1691 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1692 if (pol && (pol->flags & MPOL_F_MOF))
1697 } else if (vma->vm_policy) {
1698 pol = vma->vm_policy;
1703 return default_policy.flags & MPOL_F_MOF;
1705 return pol->flags & MPOL_F_MOF;
1708 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1710 enum zone_type dynamic_policy_zone = policy_zone;
1712 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1715 * if policy->v.nodes has movable memory only,
1716 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1718 * policy->v.nodes is intersect with node_states[N_MEMORY].
1719 * so if the following test faile, it implies
1720 * policy->v.nodes has movable memory only.
1722 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1723 dynamic_policy_zone = ZONE_MOVABLE;
1725 return zone >= dynamic_policy_zone;
1729 * Return a nodemask representing a mempolicy for filtering nodes for
1732 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1734 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1735 if (unlikely(policy->mode == MPOL_BIND) &&
1736 apply_policy_zone(policy, gfp_zone(gfp)) &&
1737 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1738 return &policy->v.nodes;
1743 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1744 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1747 switch (policy->mode) {
1748 case MPOL_PREFERRED:
1749 if (!(policy->flags & MPOL_F_LOCAL))
1750 nd = policy->v.preferred_node;
1754 * Normally, MPOL_BIND allocations are node-local within the
1755 * allowed nodemask. However, if __GFP_THISNODE is set and the
1756 * current node isn't part of the mask, we use the zonelist for
1757 * the first node in the mask instead.
1759 if (unlikely(gfp & __GFP_THISNODE) &&
1760 unlikely(!node_isset(nd, policy->v.nodes)))
1761 nd = first_node(policy->v.nodes);
1766 return node_zonelist(nd, gfp);
1769 /* Do dynamic interleaving for a process */
1770 static unsigned interleave_nodes(struct mempolicy *policy)
1773 struct task_struct *me = current;
1776 next = next_node(nid, policy->v.nodes);
1777 if (next >= MAX_NUMNODES)
1778 next = first_node(policy->v.nodes);
1779 if (next < MAX_NUMNODES)
1785 * Depending on the memory policy provide a node from which to allocate the
1787 * @policy must be protected by freeing by the caller. If @policy is
1788 * the current task's mempolicy, this protection is implicit, as only the
1789 * task can change it's policy. The system default policy requires no
1792 unsigned slab_node(void)
1794 struct mempolicy *policy;
1797 return numa_node_id();
1799 policy = current->mempolicy;
1800 if (!policy || policy->flags & MPOL_F_LOCAL)
1801 return numa_node_id();
1803 switch (policy->mode) {
1804 case MPOL_PREFERRED:
1806 * handled MPOL_F_LOCAL above
1808 return policy->v.preferred_node;
1810 case MPOL_INTERLEAVE:
1811 return interleave_nodes(policy);
1815 * Follow bind policy behavior and start allocation at the
1818 struct zonelist *zonelist;
1820 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1821 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1822 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1825 return zone ? zone->node : numa_node_id();
1833 /* Do static interleaving for a VMA with known offset. */
1834 static unsigned offset_il_node(struct mempolicy *pol,
1835 struct vm_area_struct *vma, unsigned long off)
1837 unsigned nnodes = nodes_weight(pol->v.nodes);
1840 int nid = NUMA_NO_NODE;
1843 return numa_node_id();
1844 target = (unsigned int)off % nnodes;
1847 nid = next_node(nid, pol->v.nodes);
1849 } while (c <= target);
1853 /* Determine a node number for interleave */
1854 static inline unsigned interleave_nid(struct mempolicy *pol,
1855 struct vm_area_struct *vma, unsigned long addr, int shift)
1861 * for small pages, there is no difference between
1862 * shift and PAGE_SHIFT, so the bit-shift is safe.
1863 * for huge pages, since vm_pgoff is in units of small
1864 * pages, we need to shift off the always 0 bits to get
1867 BUG_ON(shift < PAGE_SHIFT);
1868 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1869 off += (addr - vma->vm_start) >> shift;
1870 return offset_il_node(pol, vma, off);
1872 return interleave_nodes(pol);
1876 * Return the bit number of a random bit set in the nodemask.
1877 * (returns NUMA_NO_NODE if nodemask is empty)
1879 int node_random(const nodemask_t *maskp)
1881 int w, bit = NUMA_NO_NODE;
1883 w = nodes_weight(*maskp);
1885 bit = bitmap_ord_to_pos(maskp->bits,
1886 get_random_int() % w, MAX_NUMNODES);
1890 #ifdef CONFIG_HUGETLBFS
1892 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1893 * @vma = virtual memory area whose policy is sought
1894 * @addr = address in @vma for shared policy lookup and interleave policy
1895 * @gfp_flags = for requested zone
1896 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1897 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1899 * Returns a zonelist suitable for a huge page allocation and a pointer
1900 * to the struct mempolicy for conditional unref after allocation.
1901 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1902 * @nodemask for filtering the zonelist.
1904 * Must be protected by get_mems_allowed()
1906 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1907 gfp_t gfp_flags, struct mempolicy **mpol,
1908 nodemask_t **nodemask)
1910 struct zonelist *zl;
1912 *mpol = get_vma_policy(current, vma, addr);
1913 *nodemask = NULL; /* assume !MPOL_BIND */
1915 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1916 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1917 huge_page_shift(hstate_vma(vma))), gfp_flags);
1919 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1920 if ((*mpol)->mode == MPOL_BIND)
1921 *nodemask = &(*mpol)->v.nodes;
1927 * init_nodemask_of_mempolicy
1929 * If the current task's mempolicy is "default" [NULL], return 'false'
1930 * to indicate default policy. Otherwise, extract the policy nodemask
1931 * for 'bind' or 'interleave' policy into the argument nodemask, or
1932 * initialize the argument nodemask to contain the single node for
1933 * 'preferred' or 'local' policy and return 'true' to indicate presence
1934 * of non-default mempolicy.
1936 * We don't bother with reference counting the mempolicy [mpol_get/put]
1937 * because the current task is examining it's own mempolicy and a task's
1938 * mempolicy is only ever changed by the task itself.
1940 * N.B., it is the caller's responsibility to free a returned nodemask.
1942 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1944 struct mempolicy *mempolicy;
1947 if (!(mask && current->mempolicy))
1951 mempolicy = current->mempolicy;
1952 switch (mempolicy->mode) {
1953 case MPOL_PREFERRED:
1954 if (mempolicy->flags & MPOL_F_LOCAL)
1955 nid = numa_node_id();
1957 nid = mempolicy->v.preferred_node;
1958 init_nodemask_of_node(mask, nid);
1963 case MPOL_INTERLEAVE:
1964 *mask = mempolicy->v.nodes;
1970 task_unlock(current);
1977 * mempolicy_nodemask_intersects
1979 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1980 * policy. Otherwise, check for intersection between mask and the policy
1981 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1982 * policy, always return true since it may allocate elsewhere on fallback.
1984 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1986 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1987 const nodemask_t *mask)
1989 struct mempolicy *mempolicy;
1995 mempolicy = tsk->mempolicy;
1999 switch (mempolicy->mode) {
2000 case MPOL_PREFERRED:
2002 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2003 * allocate from, they may fallback to other nodes when oom.
2004 * Thus, it's possible for tsk to have allocated memory from
2009 case MPOL_INTERLEAVE:
2010 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2020 /* Allocate a page in interleaved policy.
2021 Own path because it needs to do special accounting. */
2022 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2025 struct zonelist *zl;
2028 zl = node_zonelist(nid, gfp);
2029 page = __alloc_pages(gfp, order, zl);
2030 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
2031 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2036 * alloc_pages_vma - Allocate a page for a VMA.
2039 * %GFP_USER user allocation.
2040 * %GFP_KERNEL kernel allocations,
2041 * %GFP_HIGHMEM highmem/user allocations,
2042 * %GFP_FS allocation should not call back into a file system.
2043 * %GFP_ATOMIC don't sleep.
2045 * @order:Order of the GFP allocation.
2046 * @vma: Pointer to VMA or NULL if not available.
2047 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2049 * This function allocates a page from the kernel page pool and applies
2050 * a NUMA policy associated with the VMA or the current process.
2051 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2052 * mm_struct of the VMA to prevent it from going away. Should be used for
2053 * all allocations for pages that will be mapped into
2054 * user space. Returns NULL when no page can be allocated.
2056 * Should be called with the mm_sem of the vma hold.
2059 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2060 unsigned long addr, int node)
2062 struct mempolicy *pol;
2064 unsigned int cpuset_mems_cookie;
2067 pol = get_vma_policy(current, vma, addr);
2068 cpuset_mems_cookie = get_mems_allowed();
2070 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2073 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2075 page = alloc_page_interleave(gfp, order, nid);
2076 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2081 page = __alloc_pages_nodemask(gfp, order,
2082 policy_zonelist(gfp, pol, node),
2083 policy_nodemask(gfp, pol));
2084 if (unlikely(mpol_needs_cond_ref(pol)))
2086 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2092 * alloc_pages_current - Allocate pages.
2095 * %GFP_USER user allocation,
2096 * %GFP_KERNEL kernel allocation,
2097 * %GFP_HIGHMEM highmem allocation,
2098 * %GFP_FS don't call back into a file system.
2099 * %GFP_ATOMIC don't sleep.
2100 * @order: Power of two of allocation size in pages. 0 is a single page.
2102 * Allocate a page from the kernel page pool. When not in
2103 * interrupt context and apply the current process NUMA policy.
2104 * Returns NULL when no page can be allocated.
2106 * Don't call cpuset_update_task_memory_state() unless
2107 * 1) it's ok to take cpuset_sem (can WAIT), and
2108 * 2) allocating for current task (not interrupt).
2110 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2112 struct mempolicy *pol = get_task_policy(current);
2114 unsigned int cpuset_mems_cookie;
2116 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2117 pol = &default_policy;
2120 cpuset_mems_cookie = get_mems_allowed();
2123 * No reference counting needed for current->mempolicy
2124 * nor system default_policy
2126 if (pol->mode == MPOL_INTERLEAVE)
2127 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2129 page = __alloc_pages_nodemask(gfp, order,
2130 policy_zonelist(gfp, pol, numa_node_id()),
2131 policy_nodemask(gfp, pol));
2133 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2138 EXPORT_SYMBOL(alloc_pages_current);
2140 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2142 struct mempolicy *pol = mpol_dup(vma_policy(src));
2145 return PTR_ERR(pol);
2146 dst->vm_policy = pol;
2151 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2152 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2153 * with the mems_allowed returned by cpuset_mems_allowed(). This
2154 * keeps mempolicies cpuset relative after its cpuset moves. See
2155 * further kernel/cpuset.c update_nodemask().
2157 * current's mempolicy may be rebinded by the other task(the task that changes
2158 * cpuset's mems), so we needn't do rebind work for current task.
2161 /* Slow path of a mempolicy duplicate */
2162 struct mempolicy *__mpol_dup(struct mempolicy *old)
2164 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2167 return ERR_PTR(-ENOMEM);
2169 /* task's mempolicy is protected by alloc_lock */
2170 if (old == current->mempolicy) {
2173 task_unlock(current);
2178 if (current_cpuset_is_being_rebound()) {
2179 nodemask_t mems = cpuset_mems_allowed(current);
2180 if (new->flags & MPOL_F_REBINDING)
2181 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2183 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2186 atomic_set(&new->refcnt, 1);
2190 /* Slow path of a mempolicy comparison */
2191 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2195 if (a->mode != b->mode)
2197 if (a->flags != b->flags)
2199 if (mpol_store_user_nodemask(a))
2200 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2206 case MPOL_INTERLEAVE:
2207 return !!nodes_equal(a->v.nodes, b->v.nodes);
2208 case MPOL_PREFERRED:
2209 return a->v.preferred_node == b->v.preferred_node;
2217 * Shared memory backing store policy support.
2219 * Remember policies even when nobody has shared memory mapped.
2220 * The policies are kept in Red-Black tree linked from the inode.
2221 * They are protected by the sp->lock spinlock, which should be held
2222 * for any accesses to the tree.
2225 /* lookup first element intersecting start-end */
2226 /* Caller holds sp->lock */
2227 static struct sp_node *
2228 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2230 struct rb_node *n = sp->root.rb_node;
2233 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2235 if (start >= p->end)
2237 else if (end <= p->start)
2245 struct sp_node *w = NULL;
2246 struct rb_node *prev = rb_prev(n);
2249 w = rb_entry(prev, struct sp_node, nd);
2250 if (w->end <= start)
2254 return rb_entry(n, struct sp_node, nd);
2257 /* Insert a new shared policy into the list. */
2258 /* Caller holds sp->lock */
2259 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2261 struct rb_node **p = &sp->root.rb_node;
2262 struct rb_node *parent = NULL;
2267 nd = rb_entry(parent, struct sp_node, nd);
2268 if (new->start < nd->start)
2270 else if (new->end > nd->end)
2271 p = &(*p)->rb_right;
2275 rb_link_node(&new->nd, parent, p);
2276 rb_insert_color(&new->nd, &sp->root);
2277 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2278 new->policy ? new->policy->mode : 0);
2281 /* Find shared policy intersecting idx */
2283 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2285 struct mempolicy *pol = NULL;
2288 if (!sp->root.rb_node)
2290 spin_lock(&sp->lock);
2291 sn = sp_lookup(sp, idx, idx+1);
2293 mpol_get(sn->policy);
2296 spin_unlock(&sp->lock);
2300 static void sp_free(struct sp_node *n)
2302 mpol_put(n->policy);
2303 kmem_cache_free(sn_cache, n);
2306 #ifdef CONFIG_NUMA_BALANCING
2307 static bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2309 /* Never defer a private fault */
2310 if (cpupid_match_pid(p, last_cpupid))
2313 if (p->numa_migrate_deferred) {
2314 p->numa_migrate_deferred--;
2320 static inline void defer_numa_migrate(struct task_struct *p)
2322 p->numa_migrate_deferred = sysctl_numa_balancing_migrate_deferred;
2325 static inline bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2330 static inline void defer_numa_migrate(struct task_struct *p)
2333 #endif /* CONFIG_NUMA_BALANCING */
2336 * mpol_misplaced - check whether current page node is valid in policy
2338 * @page - page to be checked
2339 * @vma - vm area where page mapped
2340 * @addr - virtual address where page mapped
2342 * Lookup current policy node id for vma,addr and "compare to" page's
2346 * -1 - not misplaced, page is in the right node
2347 * node - node id where the page should be
2349 * Policy determination "mimics" alloc_page_vma().
2350 * Called from fault path where we know the vma and faulting address.
2352 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2354 struct mempolicy *pol;
2356 int curnid = page_to_nid(page);
2357 unsigned long pgoff;
2358 int thiscpu = raw_smp_processor_id();
2359 int thisnid = cpu_to_node(thiscpu);
2365 pol = get_vma_policy(current, vma, addr);
2366 if (!(pol->flags & MPOL_F_MOF))
2369 switch (pol->mode) {
2370 case MPOL_INTERLEAVE:
2371 BUG_ON(addr >= vma->vm_end);
2372 BUG_ON(addr < vma->vm_start);
2374 pgoff = vma->vm_pgoff;
2375 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2376 polnid = offset_il_node(pol, vma, pgoff);
2379 case MPOL_PREFERRED:
2380 if (pol->flags & MPOL_F_LOCAL)
2381 polnid = numa_node_id();
2383 polnid = pol->v.preferred_node;
2388 * allows binding to multiple nodes.
2389 * use current page if in policy nodemask,
2390 * else select nearest allowed node, if any.
2391 * If no allowed nodes, use current [!misplaced].
2393 if (node_isset(curnid, pol->v.nodes))
2395 (void)first_zones_zonelist(
2396 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2397 gfp_zone(GFP_HIGHUSER),
2398 &pol->v.nodes, &zone);
2399 polnid = zone->node;
2406 /* Migrate the page towards the node whose CPU is referencing it */
2407 if (pol->flags & MPOL_F_MORON) {
2412 this_cpupid = cpu_pid_to_cpupid(thiscpu, current->pid);
2415 * Multi-stage node selection is used in conjunction
2416 * with a periodic migration fault to build a temporal
2417 * task<->page relation. By using a two-stage filter we
2418 * remove short/unlikely relations.
2420 * Using P(p) ~ n_p / n_t as per frequentist
2421 * probability, we can equate a task's usage of a
2422 * particular page (n_p) per total usage of this
2423 * page (n_t) (in a given time-span) to a probability.
2425 * Our periodic faults will sample this probability and
2426 * getting the same result twice in a row, given these
2427 * samples are fully independent, is then given by
2428 * P(n)^2, provided our sample period is sufficiently
2429 * short compared to the usage pattern.
2431 * This quadric squishes small probabilities, making
2432 * it less likely we act on an unlikely task<->page
2435 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
2436 if (!cpupid_pid_unset(last_cpupid) && cpupid_to_nid(last_cpupid) != thisnid) {
2438 /* See sysctl_numa_balancing_migrate_deferred comment */
2439 if (!cpupid_match_pid(current, last_cpupid))
2440 defer_numa_migrate(current);
2446 * The quadratic filter above reduces extraneous migration
2447 * of shared pages somewhat. This code reduces it even more,
2448 * reducing the overhead of page migrations of shared pages.
2449 * This makes workloads with shared pages rely more on
2450 * "move task near its memory", and less on "move memory
2451 * towards its task", which is exactly what we want.
2453 if (numa_migrate_deferred(current, last_cpupid))
2457 if (curnid != polnid)
2465 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2467 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2468 rb_erase(&n->nd, &sp->root);
2472 static void sp_node_init(struct sp_node *node, unsigned long start,
2473 unsigned long end, struct mempolicy *pol)
2475 node->start = start;
2480 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2481 struct mempolicy *pol)
2484 struct mempolicy *newpol;
2486 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2490 newpol = mpol_dup(pol);
2491 if (IS_ERR(newpol)) {
2492 kmem_cache_free(sn_cache, n);
2495 newpol->flags |= MPOL_F_SHARED;
2496 sp_node_init(n, start, end, newpol);
2501 /* Replace a policy range. */
2502 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2503 unsigned long end, struct sp_node *new)
2506 struct sp_node *n_new = NULL;
2507 struct mempolicy *mpol_new = NULL;
2511 spin_lock(&sp->lock);
2512 n = sp_lookup(sp, start, end);
2513 /* Take care of old policies in the same range. */
2514 while (n && n->start < end) {
2515 struct rb_node *next = rb_next(&n->nd);
2516 if (n->start >= start) {
2522 /* Old policy spanning whole new range. */
2527 *mpol_new = *n->policy;
2528 atomic_set(&mpol_new->refcnt, 1);
2529 sp_node_init(n_new, end, n->end, mpol_new);
2531 sp_insert(sp, n_new);
2540 n = rb_entry(next, struct sp_node, nd);
2544 spin_unlock(&sp->lock);
2551 kmem_cache_free(sn_cache, n_new);
2556 spin_unlock(&sp->lock);
2558 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2561 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2568 * mpol_shared_policy_init - initialize shared policy for inode
2569 * @sp: pointer to inode shared policy
2570 * @mpol: struct mempolicy to install
2572 * Install non-NULL @mpol in inode's shared policy rb-tree.
2573 * On entry, the current task has a reference on a non-NULL @mpol.
2574 * This must be released on exit.
2575 * This is called at get_inode() calls and we can use GFP_KERNEL.
2577 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2581 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2582 spin_lock_init(&sp->lock);
2585 struct vm_area_struct pvma;
2586 struct mempolicy *new;
2587 NODEMASK_SCRATCH(scratch);
2591 /* contextualize the tmpfs mount point mempolicy */
2592 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2594 goto free_scratch; /* no valid nodemask intersection */
2597 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2598 task_unlock(current);
2602 /* Create pseudo-vma that contains just the policy */
2603 memset(&pvma, 0, sizeof(struct vm_area_struct));
2604 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2605 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2608 mpol_put(new); /* drop initial ref */
2610 NODEMASK_SCRATCH_FREE(scratch);
2612 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2616 int mpol_set_shared_policy(struct shared_policy *info,
2617 struct vm_area_struct *vma, struct mempolicy *npol)
2620 struct sp_node *new = NULL;
2621 unsigned long sz = vma_pages(vma);
2623 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2625 sz, npol ? npol->mode : -1,
2626 npol ? npol->flags : -1,
2627 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2630 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2634 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2640 /* Free a backing policy store on inode delete. */
2641 void mpol_free_shared_policy(struct shared_policy *p)
2644 struct rb_node *next;
2646 if (!p->root.rb_node)
2648 spin_lock(&p->lock);
2649 next = rb_first(&p->root);
2651 n = rb_entry(next, struct sp_node, nd);
2652 next = rb_next(&n->nd);
2655 spin_unlock(&p->lock);
2658 #ifdef CONFIG_NUMA_BALANCING
2659 static bool __initdata numabalancing_override;
2661 static void __init check_numabalancing_enable(void)
2663 bool numabalancing_default = false;
2665 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2666 numabalancing_default = true;
2668 if (nr_node_ids > 1 && !numabalancing_override) {
2669 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2670 "Configure with numa_balancing= or sysctl");
2671 set_numabalancing_state(numabalancing_default);
2675 static int __init setup_numabalancing(char *str)
2680 numabalancing_override = true;
2682 if (!strcmp(str, "enable")) {
2683 set_numabalancing_state(true);
2685 } else if (!strcmp(str, "disable")) {
2686 set_numabalancing_state(false);
2691 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2695 __setup("numa_balancing=", setup_numabalancing);
2697 static inline void __init check_numabalancing_enable(void)
2700 #endif /* CONFIG_NUMA_BALANCING */
2702 /* assumes fs == KERNEL_DS */
2703 void __init numa_policy_init(void)
2705 nodemask_t interleave_nodes;
2706 unsigned long largest = 0;
2707 int nid, prefer = 0;
2709 policy_cache = kmem_cache_create("numa_policy",
2710 sizeof(struct mempolicy),
2711 0, SLAB_PANIC, NULL);
2713 sn_cache = kmem_cache_create("shared_policy_node",
2714 sizeof(struct sp_node),
2715 0, SLAB_PANIC, NULL);
2717 for_each_node(nid) {
2718 preferred_node_policy[nid] = (struct mempolicy) {
2719 .refcnt = ATOMIC_INIT(1),
2720 .mode = MPOL_PREFERRED,
2721 .flags = MPOL_F_MOF | MPOL_F_MORON,
2722 .v = { .preferred_node = nid, },
2727 * Set interleaving policy for system init. Interleaving is only
2728 * enabled across suitably sized nodes (default is >= 16MB), or
2729 * fall back to the largest node if they're all smaller.
2731 nodes_clear(interleave_nodes);
2732 for_each_node_state(nid, N_MEMORY) {
2733 unsigned long total_pages = node_present_pages(nid);
2735 /* Preserve the largest node */
2736 if (largest < total_pages) {
2737 largest = total_pages;
2741 /* Interleave this node? */
2742 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2743 node_set(nid, interleave_nodes);
2746 /* All too small, use the largest */
2747 if (unlikely(nodes_empty(interleave_nodes)))
2748 node_set(prefer, interleave_nodes);
2750 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2751 printk("numa_policy_init: interleaving failed\n");
2753 check_numabalancing_enable();
2756 /* Reset policy of current process to default */
2757 void numa_default_policy(void)
2759 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2763 * Parse and format mempolicy from/to strings
2767 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2769 static const char * const policy_modes[] =
2771 [MPOL_DEFAULT] = "default",
2772 [MPOL_PREFERRED] = "prefer",
2773 [MPOL_BIND] = "bind",
2774 [MPOL_INTERLEAVE] = "interleave",
2775 [MPOL_LOCAL] = "local",
2781 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2782 * @str: string containing mempolicy to parse
2783 * @mpol: pointer to struct mempolicy pointer, returned on success.
2786 * <mode>[=<flags>][:<nodelist>]
2788 * On success, returns 0, else 1
2790 int mpol_parse_str(char *str, struct mempolicy **mpol)
2792 struct mempolicy *new = NULL;
2793 unsigned short mode;
2794 unsigned short mode_flags;
2796 char *nodelist = strchr(str, ':');
2797 char *flags = strchr(str, '=');
2801 /* NUL-terminate mode or flags string */
2803 if (nodelist_parse(nodelist, nodes))
2805 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2811 *flags++ = '\0'; /* terminate mode string */
2813 for (mode = 0; mode < MPOL_MAX; mode++) {
2814 if (!strcmp(str, policy_modes[mode])) {
2818 if (mode >= MPOL_MAX)
2822 case MPOL_PREFERRED:
2824 * Insist on a nodelist of one node only
2827 char *rest = nodelist;
2828 while (isdigit(*rest))
2834 case MPOL_INTERLEAVE:
2836 * Default to online nodes with memory if no nodelist
2839 nodes = node_states[N_MEMORY];
2843 * Don't allow a nodelist; mpol_new() checks flags
2847 mode = MPOL_PREFERRED;
2851 * Insist on a empty nodelist
2858 * Insist on a nodelist
2867 * Currently, we only support two mutually exclusive
2870 if (!strcmp(flags, "static"))
2871 mode_flags |= MPOL_F_STATIC_NODES;
2872 else if (!strcmp(flags, "relative"))
2873 mode_flags |= MPOL_F_RELATIVE_NODES;
2878 new = mpol_new(mode, mode_flags, &nodes);
2883 * Save nodes for mpol_to_str() to show the tmpfs mount options
2884 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2886 if (mode != MPOL_PREFERRED)
2887 new->v.nodes = nodes;
2889 new->v.preferred_node = first_node(nodes);
2891 new->flags |= MPOL_F_LOCAL;
2894 * Save nodes for contextualization: this will be used to "clone"
2895 * the mempolicy in a specific context [cpuset] at a later time.
2897 new->w.user_nodemask = nodes;
2902 /* Restore string for error message */
2911 #endif /* CONFIG_TMPFS */
2914 * mpol_to_str - format a mempolicy structure for printing
2915 * @buffer: to contain formatted mempolicy string
2916 * @maxlen: length of @buffer
2917 * @pol: pointer to mempolicy to be formatted
2919 * Convert @pol into a string. If @buffer is too short, truncate the string.
2920 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2921 * longest flag, "relative", and to display at least a few node ids.
2923 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2926 nodemask_t nodes = NODE_MASK_NONE;
2927 unsigned short mode = MPOL_DEFAULT;
2928 unsigned short flags = 0;
2930 if (pol && pol != &default_policy) {
2938 case MPOL_PREFERRED:
2939 if (flags & MPOL_F_LOCAL)
2942 node_set(pol->v.preferred_node, nodes);
2945 case MPOL_INTERLEAVE:
2946 nodes = pol->v.nodes;
2950 snprintf(p, maxlen, "unknown");
2954 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2956 if (flags & MPOL_MODE_FLAGS) {
2957 p += snprintf(p, buffer + maxlen - p, "=");
2960 * Currently, the only defined flags are mutually exclusive
2962 if (flags & MPOL_F_STATIC_NODES)
2963 p += snprintf(p, buffer + maxlen - p, "static");
2964 else if (flags & MPOL_F_RELATIVE_NODES)
2965 p += snprintf(p, buffer + maxlen - p, "relative");
2968 if (!nodes_empty(nodes)) {
2969 p += snprintf(p, buffer + maxlen - p, ":");
2970 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);