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
531 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
532 entry = huge_ptep_get((pte_t *)pmd);
533 if (!pte_present(entry))
535 page = pte_page(entry);
536 nid = page_to_nid(page);
537 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
539 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
540 if (flags & (MPOL_MF_MOVE_ALL) ||
541 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
542 isolate_huge_page(page, private);
550 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
551 unsigned long addr, unsigned long end,
552 const nodemask_t *nodes, unsigned long flags,
558 pmd = pmd_offset(pud, addr);
560 next = pmd_addr_end(addr, end);
561 if (!pmd_present(*pmd))
563 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
564 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
568 split_huge_page_pmd(vma, addr, pmd);
569 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
571 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
574 } while (pmd++, addr = next, addr != end);
578 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
579 unsigned long addr, unsigned long end,
580 const nodemask_t *nodes, unsigned long flags,
586 pud = pud_offset(pgd, addr);
588 next = pud_addr_end(addr, end);
589 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
591 if (pud_none_or_clear_bad(pud))
593 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
596 } while (pud++, addr = next, addr != end);
600 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
601 unsigned long addr, unsigned long end,
602 const nodemask_t *nodes, unsigned long flags,
608 pgd = pgd_offset(vma->vm_mm, addr);
610 next = pgd_addr_end(addr, end);
611 if (pgd_none_or_clear_bad(pgd))
613 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
616 } while (pgd++, addr = next, addr != end);
620 #ifdef CONFIG_NUMA_BALANCING
622 * This is used to mark a range of virtual addresses to be inaccessible.
623 * These are later cleared by a NUMA hinting fault. Depending on these
624 * faults, pages may be migrated for better NUMA placement.
626 * This is assuming that NUMA faults are handled using PROT_NONE. If
627 * an architecture makes a different choice, it will need further
628 * changes to the core.
630 unsigned long change_prot_numa(struct vm_area_struct *vma,
631 unsigned long addr, unsigned long end)
635 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
637 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
642 static unsigned long change_prot_numa(struct vm_area_struct *vma,
643 unsigned long addr, unsigned long end)
647 #endif /* CONFIG_NUMA_BALANCING */
650 * Walk through page tables and collect pages to be migrated.
652 * If pages found in a given range are on a set of nodes (determined by
653 * @nodes and @flags,) it's isolated and queued to the pagelist which is
654 * passed via @private.)
656 static struct vm_area_struct *
657 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
658 const nodemask_t *nodes, unsigned long flags, void *private)
661 struct vm_area_struct *first, *vma, *prev;
664 first = find_vma(mm, start);
666 return ERR_PTR(-EFAULT);
668 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
669 unsigned long endvma = vma->vm_end;
673 if (vma->vm_start > start)
674 start = vma->vm_start;
676 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
677 if (!vma->vm_next && vma->vm_end < end)
678 return ERR_PTR(-EFAULT);
679 if (prev && prev->vm_end < vma->vm_start)
680 return ERR_PTR(-EFAULT);
683 if (flags & MPOL_MF_LAZY) {
684 change_prot_numa(vma, start, endvma);
688 if ((flags & MPOL_MF_STRICT) ||
689 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
690 vma_migratable(vma))) {
692 err = queue_pages_pgd_range(vma, start, endvma, nodes,
695 first = ERR_PTR(err);
706 * Apply policy to a single VMA
707 * This must be called with the mmap_sem held for writing.
709 static int vma_replace_policy(struct vm_area_struct *vma,
710 struct mempolicy *pol)
713 struct mempolicy *old;
714 struct mempolicy *new;
716 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
717 vma->vm_start, vma->vm_end, vma->vm_pgoff,
718 vma->vm_ops, vma->vm_file,
719 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
725 if (vma->vm_ops && vma->vm_ops->set_policy) {
726 err = vma->vm_ops->set_policy(vma, new);
731 old = vma->vm_policy;
732 vma->vm_policy = new; /* protected by mmap_sem */
741 /* Step 2: apply policy to a range and do splits. */
742 static int mbind_range(struct mm_struct *mm, unsigned long start,
743 unsigned long end, struct mempolicy *new_pol)
745 struct vm_area_struct *next;
746 struct vm_area_struct *prev;
747 struct vm_area_struct *vma;
750 unsigned long vmstart;
753 vma = find_vma(mm, start);
754 if (!vma || vma->vm_start > start)
758 if (start > vma->vm_start)
761 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
763 vmstart = max(start, vma->vm_start);
764 vmend = min(end, vma->vm_end);
766 if (mpol_equal(vma_policy(vma), new_pol))
769 pgoff = vma->vm_pgoff +
770 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
771 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
772 vma->anon_vma, vma->vm_file, pgoff,
777 if (mpol_equal(vma_policy(vma), new_pol))
779 /* vma_merge() joined vma && vma->next, case 8 */
782 if (vma->vm_start != vmstart) {
783 err = split_vma(vma->vm_mm, vma, vmstart, 1);
787 if (vma->vm_end != vmend) {
788 err = split_vma(vma->vm_mm, vma, vmend, 0);
793 err = vma_replace_policy(vma, new_pol);
803 * Update task->flags PF_MEMPOLICY bit: set iff non-default
804 * mempolicy. Allows more rapid checking of this (combined perhaps
805 * with other PF_* flag bits) on memory allocation hot code paths.
807 * If called from outside this file, the task 'p' should -only- be
808 * a newly forked child not yet visible on the task list, because
809 * manipulating the task flags of a visible task is not safe.
811 * The above limitation is why this routine has the funny name
812 * mpol_fix_fork_child_flag().
814 * It is also safe to call this with a task pointer of current,
815 * which the static wrapper mpol_set_task_struct_flag() does,
816 * for use within this file.
819 void mpol_fix_fork_child_flag(struct task_struct *p)
822 p->flags |= PF_MEMPOLICY;
824 p->flags &= ~PF_MEMPOLICY;
827 static void mpol_set_task_struct_flag(void)
829 mpol_fix_fork_child_flag(current);
832 /* Set the process memory policy */
833 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
836 struct mempolicy *new, *old;
837 struct mm_struct *mm = current->mm;
838 NODEMASK_SCRATCH(scratch);
844 new = mpol_new(mode, flags, nodes);
850 * prevent changing our mempolicy while show_numa_maps()
852 * Note: do_set_mempolicy() can be called at init time
856 down_write(&mm->mmap_sem);
858 ret = mpol_set_nodemask(new, nodes, scratch);
860 task_unlock(current);
862 up_write(&mm->mmap_sem);
866 old = current->mempolicy;
867 current->mempolicy = new;
868 mpol_set_task_struct_flag();
869 if (new && new->mode == MPOL_INTERLEAVE &&
870 nodes_weight(new->v.nodes))
871 current->il_next = first_node(new->v.nodes);
872 task_unlock(current);
874 up_write(&mm->mmap_sem);
879 NODEMASK_SCRATCH_FREE(scratch);
884 * Return nodemask for policy for get_mempolicy() query
886 * Called with task's alloc_lock held
888 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
891 if (p == &default_policy)
897 case MPOL_INTERLEAVE:
901 if (!(p->flags & MPOL_F_LOCAL))
902 node_set(p->v.preferred_node, *nodes);
903 /* else return empty node mask for local allocation */
910 static int lookup_node(struct mm_struct *mm, unsigned long addr)
915 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
917 err = page_to_nid(p);
923 /* Retrieve NUMA policy */
924 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
925 unsigned long addr, unsigned long flags)
928 struct mm_struct *mm = current->mm;
929 struct vm_area_struct *vma = NULL;
930 struct mempolicy *pol = current->mempolicy;
933 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
936 if (flags & MPOL_F_MEMS_ALLOWED) {
937 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
939 *policy = 0; /* just so it's initialized */
941 *nmask = cpuset_current_mems_allowed;
942 task_unlock(current);
946 if (flags & MPOL_F_ADDR) {
948 * Do NOT fall back to task policy if the
949 * vma/shared policy at addr is NULL. We
950 * want to return MPOL_DEFAULT in this case.
952 down_read(&mm->mmap_sem);
953 vma = find_vma_intersection(mm, addr, addr+1);
955 up_read(&mm->mmap_sem);
958 if (vma->vm_ops && vma->vm_ops->get_policy)
959 pol = vma->vm_ops->get_policy(vma, addr);
961 pol = vma->vm_policy;
966 pol = &default_policy; /* indicates default behavior */
968 if (flags & MPOL_F_NODE) {
969 if (flags & MPOL_F_ADDR) {
970 err = lookup_node(mm, addr);
974 } else if (pol == current->mempolicy &&
975 pol->mode == MPOL_INTERLEAVE) {
976 *policy = current->il_next;
982 *policy = pol == &default_policy ? MPOL_DEFAULT :
985 * Internal mempolicy flags must be masked off before exposing
986 * the policy to userspace.
988 *policy |= (pol->flags & MPOL_MODE_FLAGS);
992 up_read(¤t->mm->mmap_sem);
998 if (mpol_store_user_nodemask(pol)) {
999 *nmask = pol->w.user_nodemask;
1002 get_policy_nodemask(pol, nmask);
1003 task_unlock(current);
1010 up_read(¤t->mm->mmap_sem);
1014 #ifdef CONFIG_MIGRATION
1018 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1019 unsigned long flags)
1022 * Avoid migrating a page that is shared with others.
1024 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1025 if (!isolate_lru_page(page)) {
1026 list_add_tail(&page->lru, pagelist);
1027 inc_zone_page_state(page, NR_ISOLATED_ANON +
1028 page_is_file_cache(page));
1033 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1036 return alloc_huge_page_node(page_hstate(compound_head(page)),
1039 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1043 * Migrate pages from one node to a target node.
1044 * Returns error or the number of pages not migrated.
1046 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1050 LIST_HEAD(pagelist);
1054 node_set(source, nmask);
1057 * This does not "check" the range but isolates all pages that
1058 * need migration. Between passing in the full user address
1059 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1061 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1062 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1063 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1065 if (!list_empty(&pagelist)) {
1066 err = migrate_pages(&pagelist, new_node_page, dest,
1067 MIGRATE_SYNC, MR_SYSCALL);
1069 putback_movable_pages(&pagelist);
1076 * Move pages between the two nodesets so as to preserve the physical
1077 * layout as much as possible.
1079 * Returns the number of page that could not be moved.
1081 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1082 const nodemask_t *to, int flags)
1088 err = migrate_prep();
1092 down_read(&mm->mmap_sem);
1094 err = migrate_vmas(mm, from, to, flags);
1099 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1100 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1101 * bit in 'tmp', and return that <source, dest> pair for migration.
1102 * The pair of nodemasks 'to' and 'from' define the map.
1104 * If no pair of bits is found that way, fallback to picking some
1105 * pair of 'source' and 'dest' bits that are not the same. If the
1106 * 'source' and 'dest' bits are the same, this represents a node
1107 * that will be migrating to itself, so no pages need move.
1109 * If no bits are left in 'tmp', or if all remaining bits left
1110 * in 'tmp' correspond to the same bit in 'to', return false
1111 * (nothing left to migrate).
1113 * This lets us pick a pair of nodes to migrate between, such that
1114 * if possible the dest node is not already occupied by some other
1115 * source node, minimizing the risk of overloading the memory on a
1116 * node that would happen if we migrated incoming memory to a node
1117 * before migrating outgoing memory source that same node.
1119 * A single scan of tmp is sufficient. As we go, we remember the
1120 * most recent <s, d> pair that moved (s != d). If we find a pair
1121 * that not only moved, but what's better, moved to an empty slot
1122 * (d is not set in tmp), then we break out then, with that pair.
1123 * Otherwise when we finish scanning from_tmp, we at least have the
1124 * most recent <s, d> pair that moved. If we get all the way through
1125 * the scan of tmp without finding any node that moved, much less
1126 * moved to an empty node, then there is nothing left worth migrating.
1130 while (!nodes_empty(tmp)) {
1132 int source = NUMA_NO_NODE;
1135 for_each_node_mask(s, tmp) {
1138 * do_migrate_pages() tries to maintain the relative
1139 * node relationship of the pages established between
1140 * threads and memory areas.
1142 * However if the number of source nodes is not equal to
1143 * the number of destination nodes we can not preserve
1144 * this node relative relationship. In that case, skip
1145 * copying memory from a node that is in the destination
1148 * Example: [2,3,4] -> [3,4,5] moves everything.
1149 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1152 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1153 (node_isset(s, *to)))
1156 d = node_remap(s, *from, *to);
1160 source = s; /* Node moved. Memorize */
1163 /* dest not in remaining from nodes? */
1164 if (!node_isset(dest, tmp))
1167 if (source == NUMA_NO_NODE)
1170 node_clear(source, tmp);
1171 err = migrate_to_node(mm, source, dest, flags);
1178 up_read(&mm->mmap_sem);
1186 * Allocate a new page for page migration based on vma policy.
1187 * Start assuming that page is mapped by vma pointed to by @private.
1188 * Search forward from there, if not. N.B., this assumes that the
1189 * list of pages handed to migrate_pages()--which is how we get here--
1190 * is in virtual address order.
1192 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1194 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1195 unsigned long uninitialized_var(address);
1198 address = page_address_in_vma(page, vma);
1199 if (address != -EFAULT)
1204 if (PageHuge(page)) {
1206 return alloc_huge_page_noerr(vma, address, 1);
1209 * if !vma, alloc_page_vma() will use task or system default policy
1211 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1215 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1216 unsigned long flags)
1220 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1221 const nodemask_t *to, int flags)
1226 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1232 static long do_mbind(unsigned long start, unsigned long len,
1233 unsigned short mode, unsigned short mode_flags,
1234 nodemask_t *nmask, unsigned long flags)
1236 struct vm_area_struct *vma;
1237 struct mm_struct *mm = current->mm;
1238 struct mempolicy *new;
1241 LIST_HEAD(pagelist);
1243 if (flags & ~(unsigned long)MPOL_MF_VALID)
1245 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1248 if (start & ~PAGE_MASK)
1251 if (mode == MPOL_DEFAULT)
1252 flags &= ~MPOL_MF_STRICT;
1254 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1262 new = mpol_new(mode, mode_flags, nmask);
1264 return PTR_ERR(new);
1266 if (flags & MPOL_MF_LAZY)
1267 new->flags |= MPOL_F_MOF;
1270 * If we are using the default policy then operation
1271 * on discontinuous address spaces is okay after all
1274 flags |= MPOL_MF_DISCONTIG_OK;
1276 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1277 start, start + len, mode, mode_flags,
1278 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1280 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1282 err = migrate_prep();
1287 NODEMASK_SCRATCH(scratch);
1289 down_write(&mm->mmap_sem);
1291 err = mpol_set_nodemask(new, nmask, scratch);
1292 task_unlock(current);
1294 up_write(&mm->mmap_sem);
1297 NODEMASK_SCRATCH_FREE(scratch);
1302 vma = queue_pages_range(mm, start, end, nmask,
1303 flags | MPOL_MF_INVERT, &pagelist);
1305 err = PTR_ERR(vma); /* maybe ... */
1307 err = mbind_range(mm, start, end, new);
1312 if (!list_empty(&pagelist)) {
1313 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1314 nr_failed = migrate_pages(&pagelist, new_vma_page,
1316 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1318 putback_movable_pages(&pagelist);
1321 if (nr_failed && (flags & MPOL_MF_STRICT))
1324 putback_movable_pages(&pagelist);
1326 up_write(&mm->mmap_sem);
1333 * User space interface with variable sized bitmaps for nodelists.
1336 /* Copy a node mask from user space. */
1337 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1338 unsigned long maxnode)
1341 unsigned long nlongs;
1342 unsigned long endmask;
1345 nodes_clear(*nodes);
1346 if (maxnode == 0 || !nmask)
1348 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1351 nlongs = BITS_TO_LONGS(maxnode);
1352 if ((maxnode % BITS_PER_LONG) == 0)
1355 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1357 /* When the user specified more nodes than supported just check
1358 if the non supported part is all zero. */
1359 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1360 if (nlongs > PAGE_SIZE/sizeof(long))
1362 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1364 if (get_user(t, nmask + k))
1366 if (k == nlongs - 1) {
1372 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1376 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1378 nodes_addr(*nodes)[nlongs-1] &= endmask;
1382 /* Copy a kernel node mask to user space */
1383 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1386 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1387 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1389 if (copy > nbytes) {
1390 if (copy > PAGE_SIZE)
1392 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1396 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1399 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1400 unsigned long, mode, unsigned long __user *, nmask,
1401 unsigned long, maxnode, unsigned, flags)
1405 unsigned short mode_flags;
1407 mode_flags = mode & MPOL_MODE_FLAGS;
1408 mode &= ~MPOL_MODE_FLAGS;
1409 if (mode >= MPOL_MAX)
1411 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1412 (mode_flags & MPOL_F_RELATIVE_NODES))
1414 err = get_nodes(&nodes, nmask, maxnode);
1417 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1420 /* Set the process memory policy */
1421 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1422 unsigned long, maxnode)
1426 unsigned short flags;
1428 flags = mode & MPOL_MODE_FLAGS;
1429 mode &= ~MPOL_MODE_FLAGS;
1430 if ((unsigned int)mode >= MPOL_MAX)
1432 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1434 err = get_nodes(&nodes, nmask, maxnode);
1437 return do_set_mempolicy(mode, flags, &nodes);
1440 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1441 const unsigned long __user *, old_nodes,
1442 const unsigned long __user *, new_nodes)
1444 const struct cred *cred = current_cred(), *tcred;
1445 struct mm_struct *mm = NULL;
1446 struct task_struct *task;
1447 nodemask_t task_nodes;
1451 NODEMASK_SCRATCH(scratch);
1456 old = &scratch->mask1;
1457 new = &scratch->mask2;
1459 err = get_nodes(old, old_nodes, maxnode);
1463 err = get_nodes(new, new_nodes, maxnode);
1467 /* Find the mm_struct */
1469 task = pid ? find_task_by_vpid(pid) : current;
1475 get_task_struct(task);
1480 * Check if this process has the right to modify the specified
1481 * process. The right exists if the process has administrative
1482 * capabilities, superuser privileges or the same
1483 * userid as the target process.
1485 tcred = __task_cred(task);
1486 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1487 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1488 !capable(CAP_SYS_NICE)) {
1495 task_nodes = cpuset_mems_allowed(task);
1496 /* Is the user allowed to access the target nodes? */
1497 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1502 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1507 err = security_task_movememory(task);
1511 mm = get_task_mm(task);
1512 put_task_struct(task);
1519 err = do_migrate_pages(mm, old, new,
1520 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1524 NODEMASK_SCRATCH_FREE(scratch);
1529 put_task_struct(task);
1535 /* Retrieve NUMA policy */
1536 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1537 unsigned long __user *, nmask, unsigned long, maxnode,
1538 unsigned long, addr, unsigned long, flags)
1541 int uninitialized_var(pval);
1544 if (nmask != NULL && maxnode < MAX_NUMNODES)
1547 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1552 if (policy && put_user(pval, policy))
1556 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1561 #ifdef CONFIG_COMPAT
1563 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1564 compat_ulong_t __user *nmask,
1565 compat_ulong_t maxnode,
1566 compat_ulong_t addr, compat_ulong_t flags)
1569 unsigned long __user *nm = NULL;
1570 unsigned long nr_bits, alloc_size;
1571 DECLARE_BITMAP(bm, MAX_NUMNODES);
1573 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1574 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1577 nm = compat_alloc_user_space(alloc_size);
1579 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1581 if (!err && nmask) {
1582 unsigned long copy_size;
1583 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1584 err = copy_from_user(bm, nm, copy_size);
1585 /* ensure entire bitmap is zeroed */
1586 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1587 err |= compat_put_bitmap(nmask, bm, nr_bits);
1593 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1594 compat_ulong_t maxnode)
1597 unsigned long __user *nm = NULL;
1598 unsigned long nr_bits, alloc_size;
1599 DECLARE_BITMAP(bm, MAX_NUMNODES);
1601 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1602 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1605 err = compat_get_bitmap(bm, nmask, nr_bits);
1606 nm = compat_alloc_user_space(alloc_size);
1607 err |= copy_to_user(nm, bm, alloc_size);
1613 return sys_set_mempolicy(mode, nm, nr_bits+1);
1616 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1617 compat_ulong_t mode, compat_ulong_t __user *nmask,
1618 compat_ulong_t maxnode, compat_ulong_t flags)
1621 unsigned long __user *nm = NULL;
1622 unsigned long nr_bits, alloc_size;
1625 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1626 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1629 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1630 nm = compat_alloc_user_space(alloc_size);
1631 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1637 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1643 * get_vma_policy(@task, @vma, @addr)
1644 * @task - task for fallback if vma policy == default
1645 * @vma - virtual memory area whose policy is sought
1646 * @addr - address in @vma for shared policy lookup
1648 * Returns effective policy for a VMA at specified address.
1649 * Falls back to @task or system default policy, as necessary.
1650 * Current or other task's task mempolicy and non-shared vma policies must be
1651 * protected by task_lock(task) by the caller.
1652 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1653 * count--added by the get_policy() vm_op, as appropriate--to protect against
1654 * freeing by another task. It is the caller's responsibility to free the
1655 * extra reference for shared policies.
1657 struct mempolicy *get_vma_policy(struct task_struct *task,
1658 struct vm_area_struct *vma, unsigned long addr)
1660 struct mempolicy *pol = get_task_policy(task);
1663 if (vma->vm_ops && vma->vm_ops->get_policy) {
1664 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1668 } else if (vma->vm_policy) {
1669 pol = vma->vm_policy;
1672 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1673 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1674 * count on these policies which will be dropped by
1675 * mpol_cond_put() later
1677 if (mpol_needs_cond_ref(pol))
1682 pol = &default_policy;
1686 bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1688 struct mempolicy *pol = get_task_policy(task);
1690 if (vma->vm_ops && vma->vm_ops->get_policy) {
1693 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1694 if (pol && (pol->flags & MPOL_F_MOF))
1699 } else if (vma->vm_policy) {
1700 pol = vma->vm_policy;
1705 return default_policy.flags & MPOL_F_MOF;
1707 return pol->flags & MPOL_F_MOF;
1710 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1712 enum zone_type dynamic_policy_zone = policy_zone;
1714 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1717 * if policy->v.nodes has movable memory only,
1718 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1720 * policy->v.nodes is intersect with node_states[N_MEMORY].
1721 * so if the following test faile, it implies
1722 * policy->v.nodes has movable memory only.
1724 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1725 dynamic_policy_zone = ZONE_MOVABLE;
1727 return zone >= dynamic_policy_zone;
1731 * Return a nodemask representing a mempolicy for filtering nodes for
1734 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1736 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1737 if (unlikely(policy->mode == MPOL_BIND) &&
1738 apply_policy_zone(policy, gfp_zone(gfp)) &&
1739 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1740 return &policy->v.nodes;
1745 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1746 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1749 switch (policy->mode) {
1750 case MPOL_PREFERRED:
1751 if (!(policy->flags & MPOL_F_LOCAL))
1752 nd = policy->v.preferred_node;
1756 * Normally, MPOL_BIND allocations are node-local within the
1757 * allowed nodemask. However, if __GFP_THISNODE is set and the
1758 * current node isn't part of the mask, we use the zonelist for
1759 * the first node in the mask instead.
1761 if (unlikely(gfp & __GFP_THISNODE) &&
1762 unlikely(!node_isset(nd, policy->v.nodes)))
1763 nd = first_node(policy->v.nodes);
1768 return node_zonelist(nd, gfp);
1771 /* Do dynamic interleaving for a process */
1772 static unsigned interleave_nodes(struct mempolicy *policy)
1775 struct task_struct *me = current;
1778 next = next_node(nid, policy->v.nodes);
1779 if (next >= MAX_NUMNODES)
1780 next = first_node(policy->v.nodes);
1781 if (next < MAX_NUMNODES)
1787 * Depending on the memory policy provide a node from which to allocate the
1789 * @policy must be protected by freeing by the caller. If @policy is
1790 * the current task's mempolicy, this protection is implicit, as only the
1791 * task can change it's policy. The system default policy requires no
1794 unsigned slab_node(void)
1796 struct mempolicy *policy;
1799 return numa_node_id();
1801 policy = current->mempolicy;
1802 if (!policy || policy->flags & MPOL_F_LOCAL)
1803 return numa_node_id();
1805 switch (policy->mode) {
1806 case MPOL_PREFERRED:
1808 * handled MPOL_F_LOCAL above
1810 return policy->v.preferred_node;
1812 case MPOL_INTERLEAVE:
1813 return interleave_nodes(policy);
1817 * Follow bind policy behavior and start allocation at the
1820 struct zonelist *zonelist;
1822 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1823 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1824 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1827 return zone ? zone->node : numa_node_id();
1835 /* Do static interleaving for a VMA with known offset. */
1836 static unsigned offset_il_node(struct mempolicy *pol,
1837 struct vm_area_struct *vma, unsigned long off)
1839 unsigned nnodes = nodes_weight(pol->v.nodes);
1842 int nid = NUMA_NO_NODE;
1845 return numa_node_id();
1846 target = (unsigned int)off % nnodes;
1849 nid = next_node(nid, pol->v.nodes);
1851 } while (c <= target);
1855 /* Determine a node number for interleave */
1856 static inline unsigned interleave_nid(struct mempolicy *pol,
1857 struct vm_area_struct *vma, unsigned long addr, int shift)
1863 * for small pages, there is no difference between
1864 * shift and PAGE_SHIFT, so the bit-shift is safe.
1865 * for huge pages, since vm_pgoff is in units of small
1866 * pages, we need to shift off the always 0 bits to get
1869 BUG_ON(shift < PAGE_SHIFT);
1870 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1871 off += (addr - vma->vm_start) >> shift;
1872 return offset_il_node(pol, vma, off);
1874 return interleave_nodes(pol);
1878 * Return the bit number of a random bit set in the nodemask.
1879 * (returns NUMA_NO_NODE if nodemask is empty)
1881 int node_random(const nodemask_t *maskp)
1883 int w, bit = NUMA_NO_NODE;
1885 w = nodes_weight(*maskp);
1887 bit = bitmap_ord_to_pos(maskp->bits,
1888 get_random_int() % w, MAX_NUMNODES);
1892 #ifdef CONFIG_HUGETLBFS
1894 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1895 * @vma = virtual memory area whose policy is sought
1896 * @addr = address in @vma for shared policy lookup and interleave policy
1897 * @gfp_flags = for requested zone
1898 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1899 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1901 * Returns a zonelist suitable for a huge page allocation and a pointer
1902 * to the struct mempolicy for conditional unref after allocation.
1903 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1904 * @nodemask for filtering the zonelist.
1906 * Must be protected by get_mems_allowed()
1908 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1909 gfp_t gfp_flags, struct mempolicy **mpol,
1910 nodemask_t **nodemask)
1912 struct zonelist *zl;
1914 *mpol = get_vma_policy(current, vma, addr);
1915 *nodemask = NULL; /* assume !MPOL_BIND */
1917 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1918 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1919 huge_page_shift(hstate_vma(vma))), gfp_flags);
1921 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1922 if ((*mpol)->mode == MPOL_BIND)
1923 *nodemask = &(*mpol)->v.nodes;
1929 * init_nodemask_of_mempolicy
1931 * If the current task's mempolicy is "default" [NULL], return 'false'
1932 * to indicate default policy. Otherwise, extract the policy nodemask
1933 * for 'bind' or 'interleave' policy into the argument nodemask, or
1934 * initialize the argument nodemask to contain the single node for
1935 * 'preferred' or 'local' policy and return 'true' to indicate presence
1936 * of non-default mempolicy.
1938 * We don't bother with reference counting the mempolicy [mpol_get/put]
1939 * because the current task is examining it's own mempolicy and a task's
1940 * mempolicy is only ever changed by the task itself.
1942 * N.B., it is the caller's responsibility to free a returned nodemask.
1944 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1946 struct mempolicy *mempolicy;
1949 if (!(mask && current->mempolicy))
1953 mempolicy = current->mempolicy;
1954 switch (mempolicy->mode) {
1955 case MPOL_PREFERRED:
1956 if (mempolicy->flags & MPOL_F_LOCAL)
1957 nid = numa_node_id();
1959 nid = mempolicy->v.preferred_node;
1960 init_nodemask_of_node(mask, nid);
1965 case MPOL_INTERLEAVE:
1966 *mask = mempolicy->v.nodes;
1972 task_unlock(current);
1979 * mempolicy_nodemask_intersects
1981 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1982 * policy. Otherwise, check for intersection between mask and the policy
1983 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1984 * policy, always return true since it may allocate elsewhere on fallback.
1986 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1988 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1989 const nodemask_t *mask)
1991 struct mempolicy *mempolicy;
1997 mempolicy = tsk->mempolicy;
2001 switch (mempolicy->mode) {
2002 case MPOL_PREFERRED:
2004 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2005 * allocate from, they may fallback to other nodes when oom.
2006 * Thus, it's possible for tsk to have allocated memory from
2011 case MPOL_INTERLEAVE:
2012 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2022 /* Allocate a page in interleaved policy.
2023 Own path because it needs to do special accounting. */
2024 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2027 struct zonelist *zl;
2030 zl = node_zonelist(nid, gfp);
2031 page = __alloc_pages(gfp, order, zl);
2032 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
2033 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2038 * alloc_pages_vma - Allocate a page for a VMA.
2041 * %GFP_USER user allocation.
2042 * %GFP_KERNEL kernel allocations,
2043 * %GFP_HIGHMEM highmem/user allocations,
2044 * %GFP_FS allocation should not call back into a file system.
2045 * %GFP_ATOMIC don't sleep.
2047 * @order:Order of the GFP allocation.
2048 * @vma: Pointer to VMA or NULL if not available.
2049 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2051 * This function allocates a page from the kernel page pool and applies
2052 * a NUMA policy associated with the VMA or the current process.
2053 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2054 * mm_struct of the VMA to prevent it from going away. Should be used for
2055 * all allocations for pages that will be mapped into
2056 * user space. Returns NULL when no page can be allocated.
2058 * Should be called with the mm_sem of the vma hold.
2061 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2062 unsigned long addr, int node)
2064 struct mempolicy *pol;
2066 unsigned int cpuset_mems_cookie;
2069 pol = get_vma_policy(current, vma, addr);
2070 cpuset_mems_cookie = get_mems_allowed();
2072 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2075 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2077 page = alloc_page_interleave(gfp, order, nid);
2078 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2083 page = __alloc_pages_nodemask(gfp, order,
2084 policy_zonelist(gfp, pol, node),
2085 policy_nodemask(gfp, pol));
2086 if (unlikely(mpol_needs_cond_ref(pol)))
2088 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2094 * alloc_pages_current - Allocate pages.
2097 * %GFP_USER user allocation,
2098 * %GFP_KERNEL kernel allocation,
2099 * %GFP_HIGHMEM highmem allocation,
2100 * %GFP_FS don't call back into a file system.
2101 * %GFP_ATOMIC don't sleep.
2102 * @order: Power of two of allocation size in pages. 0 is a single page.
2104 * Allocate a page from the kernel page pool. When not in
2105 * interrupt context and apply the current process NUMA policy.
2106 * Returns NULL when no page can be allocated.
2108 * Don't call cpuset_update_task_memory_state() unless
2109 * 1) it's ok to take cpuset_sem (can WAIT), and
2110 * 2) allocating for current task (not interrupt).
2112 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2114 struct mempolicy *pol = get_task_policy(current);
2116 unsigned int cpuset_mems_cookie;
2118 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2119 pol = &default_policy;
2122 cpuset_mems_cookie = get_mems_allowed();
2125 * No reference counting needed for current->mempolicy
2126 * nor system default_policy
2128 if (pol->mode == MPOL_INTERLEAVE)
2129 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2131 page = __alloc_pages_nodemask(gfp, order,
2132 policy_zonelist(gfp, pol, numa_node_id()),
2133 policy_nodemask(gfp, pol));
2135 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2140 EXPORT_SYMBOL(alloc_pages_current);
2142 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2144 struct mempolicy *pol = mpol_dup(vma_policy(src));
2147 return PTR_ERR(pol);
2148 dst->vm_policy = pol;
2153 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2154 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2155 * with the mems_allowed returned by cpuset_mems_allowed(). This
2156 * keeps mempolicies cpuset relative after its cpuset moves. See
2157 * further kernel/cpuset.c update_nodemask().
2159 * current's mempolicy may be rebinded by the other task(the task that changes
2160 * cpuset's mems), so we needn't do rebind work for current task.
2163 /* Slow path of a mempolicy duplicate */
2164 struct mempolicy *__mpol_dup(struct mempolicy *old)
2166 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2169 return ERR_PTR(-ENOMEM);
2171 /* task's mempolicy is protected by alloc_lock */
2172 if (old == current->mempolicy) {
2175 task_unlock(current);
2180 if (current_cpuset_is_being_rebound()) {
2181 nodemask_t mems = cpuset_mems_allowed(current);
2182 if (new->flags & MPOL_F_REBINDING)
2183 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2185 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2188 atomic_set(&new->refcnt, 1);
2192 /* Slow path of a mempolicy comparison */
2193 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2197 if (a->mode != b->mode)
2199 if (a->flags != b->flags)
2201 if (mpol_store_user_nodemask(a))
2202 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2208 case MPOL_INTERLEAVE:
2209 return !!nodes_equal(a->v.nodes, b->v.nodes);
2210 case MPOL_PREFERRED:
2211 return a->v.preferred_node == b->v.preferred_node;
2219 * Shared memory backing store policy support.
2221 * Remember policies even when nobody has shared memory mapped.
2222 * The policies are kept in Red-Black tree linked from the inode.
2223 * They are protected by the sp->lock spinlock, which should be held
2224 * for any accesses to the tree.
2227 /* lookup first element intersecting start-end */
2228 /* Caller holds sp->lock */
2229 static struct sp_node *
2230 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2232 struct rb_node *n = sp->root.rb_node;
2235 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2237 if (start >= p->end)
2239 else if (end <= p->start)
2247 struct sp_node *w = NULL;
2248 struct rb_node *prev = rb_prev(n);
2251 w = rb_entry(prev, struct sp_node, nd);
2252 if (w->end <= start)
2256 return rb_entry(n, struct sp_node, nd);
2259 /* Insert a new shared policy into the list. */
2260 /* Caller holds sp->lock */
2261 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2263 struct rb_node **p = &sp->root.rb_node;
2264 struct rb_node *parent = NULL;
2269 nd = rb_entry(parent, struct sp_node, nd);
2270 if (new->start < nd->start)
2272 else if (new->end > nd->end)
2273 p = &(*p)->rb_right;
2277 rb_link_node(&new->nd, parent, p);
2278 rb_insert_color(&new->nd, &sp->root);
2279 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2280 new->policy ? new->policy->mode : 0);
2283 /* Find shared policy intersecting idx */
2285 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2287 struct mempolicy *pol = NULL;
2290 if (!sp->root.rb_node)
2292 spin_lock(&sp->lock);
2293 sn = sp_lookup(sp, idx, idx+1);
2295 mpol_get(sn->policy);
2298 spin_unlock(&sp->lock);
2302 static void sp_free(struct sp_node *n)
2304 mpol_put(n->policy);
2305 kmem_cache_free(sn_cache, n);
2308 #ifdef CONFIG_NUMA_BALANCING
2309 static bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2311 /* Never defer a private fault */
2312 if (cpupid_match_pid(p, last_cpupid))
2315 if (p->numa_migrate_deferred) {
2316 p->numa_migrate_deferred--;
2322 static inline void defer_numa_migrate(struct task_struct *p)
2324 p->numa_migrate_deferred = sysctl_numa_balancing_migrate_deferred;
2327 static inline bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2332 static inline void defer_numa_migrate(struct task_struct *p)
2335 #endif /* CONFIG_NUMA_BALANCING */
2338 * mpol_misplaced - check whether current page node is valid in policy
2340 * @page - page to be checked
2341 * @vma - vm area where page mapped
2342 * @addr - virtual address where page mapped
2344 * Lookup current policy node id for vma,addr and "compare to" page's
2348 * -1 - not misplaced, page is in the right node
2349 * node - node id where the page should be
2351 * Policy determination "mimics" alloc_page_vma().
2352 * Called from fault path where we know the vma and faulting address.
2354 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2356 struct mempolicy *pol;
2358 int curnid = page_to_nid(page);
2359 unsigned long pgoff;
2360 int thiscpu = raw_smp_processor_id();
2361 int thisnid = cpu_to_node(thiscpu);
2367 pol = get_vma_policy(current, vma, addr);
2368 if (!(pol->flags & MPOL_F_MOF))
2371 switch (pol->mode) {
2372 case MPOL_INTERLEAVE:
2373 BUG_ON(addr >= vma->vm_end);
2374 BUG_ON(addr < vma->vm_start);
2376 pgoff = vma->vm_pgoff;
2377 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2378 polnid = offset_il_node(pol, vma, pgoff);
2381 case MPOL_PREFERRED:
2382 if (pol->flags & MPOL_F_LOCAL)
2383 polnid = numa_node_id();
2385 polnid = pol->v.preferred_node;
2390 * allows binding to multiple nodes.
2391 * use current page if in policy nodemask,
2392 * else select nearest allowed node, if any.
2393 * If no allowed nodes, use current [!misplaced].
2395 if (node_isset(curnid, pol->v.nodes))
2397 (void)first_zones_zonelist(
2398 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2399 gfp_zone(GFP_HIGHUSER),
2400 &pol->v.nodes, &zone);
2401 polnid = zone->node;
2408 /* Migrate the page towards the node whose CPU is referencing it */
2409 if (pol->flags & MPOL_F_MORON) {
2414 this_cpupid = cpu_pid_to_cpupid(thiscpu, current->pid);
2417 * Multi-stage node selection is used in conjunction
2418 * with a periodic migration fault to build a temporal
2419 * task<->page relation. By using a two-stage filter we
2420 * remove short/unlikely relations.
2422 * Using P(p) ~ n_p / n_t as per frequentist
2423 * probability, we can equate a task's usage of a
2424 * particular page (n_p) per total usage of this
2425 * page (n_t) (in a given time-span) to a probability.
2427 * Our periodic faults will sample this probability and
2428 * getting the same result twice in a row, given these
2429 * samples are fully independent, is then given by
2430 * P(n)^2, provided our sample period is sufficiently
2431 * short compared to the usage pattern.
2433 * This quadric squishes small probabilities, making
2434 * it less likely we act on an unlikely task<->page
2437 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
2438 if (!cpupid_pid_unset(last_cpupid) && cpupid_to_nid(last_cpupid) != thisnid) {
2440 /* See sysctl_numa_balancing_migrate_deferred comment */
2441 if (!cpupid_match_pid(current, last_cpupid))
2442 defer_numa_migrate(current);
2448 * The quadratic filter above reduces extraneous migration
2449 * of shared pages somewhat. This code reduces it even more,
2450 * reducing the overhead of page migrations of shared pages.
2451 * This makes workloads with shared pages rely more on
2452 * "move task near its memory", and less on "move memory
2453 * towards its task", which is exactly what we want.
2455 if (numa_migrate_deferred(current, last_cpupid))
2459 if (curnid != polnid)
2467 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2469 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2470 rb_erase(&n->nd, &sp->root);
2474 static void sp_node_init(struct sp_node *node, unsigned long start,
2475 unsigned long end, struct mempolicy *pol)
2477 node->start = start;
2482 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2483 struct mempolicy *pol)
2486 struct mempolicy *newpol;
2488 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2492 newpol = mpol_dup(pol);
2493 if (IS_ERR(newpol)) {
2494 kmem_cache_free(sn_cache, n);
2497 newpol->flags |= MPOL_F_SHARED;
2498 sp_node_init(n, start, end, newpol);
2503 /* Replace a policy range. */
2504 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2505 unsigned long end, struct sp_node *new)
2508 struct sp_node *n_new = NULL;
2509 struct mempolicy *mpol_new = NULL;
2513 spin_lock(&sp->lock);
2514 n = sp_lookup(sp, start, end);
2515 /* Take care of old policies in the same range. */
2516 while (n && n->start < end) {
2517 struct rb_node *next = rb_next(&n->nd);
2518 if (n->start >= start) {
2524 /* Old policy spanning whole new range. */
2529 *mpol_new = *n->policy;
2530 atomic_set(&mpol_new->refcnt, 1);
2531 sp_node_init(n_new, end, n->end, mpol_new);
2533 sp_insert(sp, n_new);
2542 n = rb_entry(next, struct sp_node, nd);
2546 spin_unlock(&sp->lock);
2553 kmem_cache_free(sn_cache, n_new);
2558 spin_unlock(&sp->lock);
2560 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2563 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2570 * mpol_shared_policy_init - initialize shared policy for inode
2571 * @sp: pointer to inode shared policy
2572 * @mpol: struct mempolicy to install
2574 * Install non-NULL @mpol in inode's shared policy rb-tree.
2575 * On entry, the current task has a reference on a non-NULL @mpol.
2576 * This must be released on exit.
2577 * This is called at get_inode() calls and we can use GFP_KERNEL.
2579 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2583 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2584 spin_lock_init(&sp->lock);
2587 struct vm_area_struct pvma;
2588 struct mempolicy *new;
2589 NODEMASK_SCRATCH(scratch);
2593 /* contextualize the tmpfs mount point mempolicy */
2594 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2596 goto free_scratch; /* no valid nodemask intersection */
2599 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2600 task_unlock(current);
2604 /* Create pseudo-vma that contains just the policy */
2605 memset(&pvma, 0, sizeof(struct vm_area_struct));
2606 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2607 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2610 mpol_put(new); /* drop initial ref */
2612 NODEMASK_SCRATCH_FREE(scratch);
2614 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2618 int mpol_set_shared_policy(struct shared_policy *info,
2619 struct vm_area_struct *vma, struct mempolicy *npol)
2622 struct sp_node *new = NULL;
2623 unsigned long sz = vma_pages(vma);
2625 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2627 sz, npol ? npol->mode : -1,
2628 npol ? npol->flags : -1,
2629 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2632 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2636 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2642 /* Free a backing policy store on inode delete. */
2643 void mpol_free_shared_policy(struct shared_policy *p)
2646 struct rb_node *next;
2648 if (!p->root.rb_node)
2650 spin_lock(&p->lock);
2651 next = rb_first(&p->root);
2653 n = rb_entry(next, struct sp_node, nd);
2654 next = rb_next(&n->nd);
2657 spin_unlock(&p->lock);
2660 #ifdef CONFIG_NUMA_BALANCING
2661 static int __initdata numabalancing_override;
2663 static void __init check_numabalancing_enable(void)
2665 bool numabalancing_default = false;
2667 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2668 numabalancing_default = true;
2670 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2671 if (numabalancing_override)
2672 set_numabalancing_state(numabalancing_override == 1);
2674 if (nr_node_ids > 1 && !numabalancing_override) {
2675 pr_info("%s automatic NUMA balancing. "
2676 "Configure with numa_balancing= or the "
2677 "kernel.numa_balancing sysctl",
2678 numabalancing_default ? "Enabling" : "Disabling");
2679 set_numabalancing_state(numabalancing_default);
2683 static int __init setup_numabalancing(char *str)
2689 if (!strcmp(str, "enable")) {
2690 numabalancing_override = 1;
2692 } else if (!strcmp(str, "disable")) {
2693 numabalancing_override = -1;
2698 pr_warn("Unable to parse numa_balancing=\n");
2702 __setup("numa_balancing=", setup_numabalancing);
2704 static inline void __init check_numabalancing_enable(void)
2707 #endif /* CONFIG_NUMA_BALANCING */
2709 /* assumes fs == KERNEL_DS */
2710 void __init numa_policy_init(void)
2712 nodemask_t interleave_nodes;
2713 unsigned long largest = 0;
2714 int nid, prefer = 0;
2716 policy_cache = kmem_cache_create("numa_policy",
2717 sizeof(struct mempolicy),
2718 0, SLAB_PANIC, NULL);
2720 sn_cache = kmem_cache_create("shared_policy_node",
2721 sizeof(struct sp_node),
2722 0, SLAB_PANIC, NULL);
2724 for_each_node(nid) {
2725 preferred_node_policy[nid] = (struct mempolicy) {
2726 .refcnt = ATOMIC_INIT(1),
2727 .mode = MPOL_PREFERRED,
2728 .flags = MPOL_F_MOF | MPOL_F_MORON,
2729 .v = { .preferred_node = nid, },
2734 * Set interleaving policy for system init. Interleaving is only
2735 * enabled across suitably sized nodes (default is >= 16MB), or
2736 * fall back to the largest node if they're all smaller.
2738 nodes_clear(interleave_nodes);
2739 for_each_node_state(nid, N_MEMORY) {
2740 unsigned long total_pages = node_present_pages(nid);
2742 /* Preserve the largest node */
2743 if (largest < total_pages) {
2744 largest = total_pages;
2748 /* Interleave this node? */
2749 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2750 node_set(nid, interleave_nodes);
2753 /* All too small, use the largest */
2754 if (unlikely(nodes_empty(interleave_nodes)))
2755 node_set(prefer, interleave_nodes);
2757 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2758 printk("numa_policy_init: interleaving failed\n");
2760 check_numabalancing_enable();
2763 /* Reset policy of current process to default */
2764 void numa_default_policy(void)
2766 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2770 * Parse and format mempolicy from/to strings
2774 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2776 static const char * const policy_modes[] =
2778 [MPOL_DEFAULT] = "default",
2779 [MPOL_PREFERRED] = "prefer",
2780 [MPOL_BIND] = "bind",
2781 [MPOL_INTERLEAVE] = "interleave",
2782 [MPOL_LOCAL] = "local",
2788 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2789 * @str: string containing mempolicy to parse
2790 * @mpol: pointer to struct mempolicy pointer, returned on success.
2793 * <mode>[=<flags>][:<nodelist>]
2795 * On success, returns 0, else 1
2797 int mpol_parse_str(char *str, struct mempolicy **mpol)
2799 struct mempolicy *new = NULL;
2800 unsigned short mode;
2801 unsigned short mode_flags;
2803 char *nodelist = strchr(str, ':');
2804 char *flags = strchr(str, '=');
2808 /* NUL-terminate mode or flags string */
2810 if (nodelist_parse(nodelist, nodes))
2812 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2818 *flags++ = '\0'; /* terminate mode string */
2820 for (mode = 0; mode < MPOL_MAX; mode++) {
2821 if (!strcmp(str, policy_modes[mode])) {
2825 if (mode >= MPOL_MAX)
2829 case MPOL_PREFERRED:
2831 * Insist on a nodelist of one node only
2834 char *rest = nodelist;
2835 while (isdigit(*rest))
2841 case MPOL_INTERLEAVE:
2843 * Default to online nodes with memory if no nodelist
2846 nodes = node_states[N_MEMORY];
2850 * Don't allow a nodelist; mpol_new() checks flags
2854 mode = MPOL_PREFERRED;
2858 * Insist on a empty nodelist
2865 * Insist on a nodelist
2874 * Currently, we only support two mutually exclusive
2877 if (!strcmp(flags, "static"))
2878 mode_flags |= MPOL_F_STATIC_NODES;
2879 else if (!strcmp(flags, "relative"))
2880 mode_flags |= MPOL_F_RELATIVE_NODES;
2885 new = mpol_new(mode, mode_flags, &nodes);
2890 * Save nodes for mpol_to_str() to show the tmpfs mount options
2891 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2893 if (mode != MPOL_PREFERRED)
2894 new->v.nodes = nodes;
2896 new->v.preferred_node = first_node(nodes);
2898 new->flags |= MPOL_F_LOCAL;
2901 * Save nodes for contextualization: this will be used to "clone"
2902 * the mempolicy in a specific context [cpuset] at a later time.
2904 new->w.user_nodemask = nodes;
2909 /* Restore string for error message */
2918 #endif /* CONFIG_TMPFS */
2921 * mpol_to_str - format a mempolicy structure for printing
2922 * @buffer: to contain formatted mempolicy string
2923 * @maxlen: length of @buffer
2924 * @pol: pointer to mempolicy to be formatted
2926 * Convert @pol into a string. If @buffer is too short, truncate the string.
2927 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2928 * longest flag, "relative", and to display at least a few node ids.
2930 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2933 nodemask_t nodes = NODE_MASK_NONE;
2934 unsigned short mode = MPOL_DEFAULT;
2935 unsigned short flags = 0;
2937 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2945 case MPOL_PREFERRED:
2946 if (flags & MPOL_F_LOCAL)
2949 node_set(pol->v.preferred_node, nodes);
2952 case MPOL_INTERLEAVE:
2953 nodes = pol->v.nodes;
2957 snprintf(p, maxlen, "unknown");
2961 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2963 if (flags & MPOL_MODE_FLAGS) {
2964 p += snprintf(p, buffer + maxlen - p, "=");
2967 * Currently, the only defined flags are mutually exclusive
2969 if (flags & MPOL_F_STATIC_NODES)
2970 p += snprintf(p, buffer + maxlen - p, "static");
2971 else if (flags & MPOL_F_RELATIVE_NODES)
2972 p += snprintf(p, buffer + maxlen - p, "relative");
2975 if (!nodes_empty(nodes)) {
2976 p += snprintf(p, buffer + maxlen - p, ":");
2977 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);