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);
479 /* Scan through pages checking if pages follow certain conditions. */
480 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
481 unsigned long addr, unsigned long end,
482 const nodemask_t *nodes, unsigned long flags,
489 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
494 if (!pte_present(*pte))
496 page = vm_normal_page(vma, addr, *pte);
500 * vm_normal_page() filters out zero pages, but there might
501 * still be PageReserved pages to skip, perhaps in a VDSO.
503 if (PageReserved(page))
505 nid = page_to_nid(page);
506 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
509 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
510 migrate_page_add(page, private, flags);
513 } while (pte++, addr += PAGE_SIZE, addr != end);
514 pte_unmap_unlock(orig_pte, ptl);
518 static void check_hugetlb_pmd_range(struct vm_area_struct *vma, pmd_t *pmd,
519 const nodemask_t *nodes, unsigned long flags,
522 #ifdef CONFIG_HUGETLB_PAGE
526 spin_lock(&vma->vm_mm->page_table_lock);
527 page = pte_page(huge_ptep_get((pte_t *)pmd));
528 nid = page_to_nid(page);
529 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
531 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
532 if (flags & (MPOL_MF_MOVE_ALL) ||
533 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
534 isolate_huge_page(page, private);
536 spin_unlock(&vma->vm_mm->page_table_lock);
542 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
543 unsigned long addr, unsigned long end,
544 const nodemask_t *nodes, unsigned long flags,
550 pmd = pmd_offset(pud, addr);
552 next = pmd_addr_end(addr, end);
553 if (!pmd_present(*pmd))
555 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
556 check_hugetlb_pmd_range(vma, pmd, nodes,
560 split_huge_page_pmd(vma, addr, pmd);
561 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
563 if (check_pte_range(vma, pmd, addr, next, nodes,
566 } while (pmd++, addr = next, addr != end);
570 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
571 unsigned long addr, unsigned long end,
572 const nodemask_t *nodes, unsigned long flags,
578 pud = pud_offset(pgd, addr);
580 next = pud_addr_end(addr, end);
581 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
583 if (pud_none_or_clear_bad(pud))
585 if (check_pmd_range(vma, pud, addr, next, nodes,
588 } while (pud++, addr = next, addr != end);
592 static inline int check_pgd_range(struct vm_area_struct *vma,
593 unsigned long addr, unsigned long end,
594 const nodemask_t *nodes, unsigned long flags,
600 pgd = pgd_offset(vma->vm_mm, addr);
602 next = pgd_addr_end(addr, end);
603 if (pgd_none_or_clear_bad(pgd))
605 if (check_pud_range(vma, pgd, addr, next, nodes,
608 } while (pgd++, addr = next, addr != end);
612 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
614 * This is used to mark a range of virtual addresses to be inaccessible.
615 * These are later cleared by a NUMA hinting fault. Depending on these
616 * faults, pages may be migrated for better NUMA placement.
618 * This is assuming that NUMA faults are handled using PROT_NONE. If
619 * an architecture makes a different choice, it will need further
620 * changes to the core.
622 unsigned long change_prot_numa(struct vm_area_struct *vma,
623 unsigned long addr, unsigned long end)
626 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
628 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
630 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
635 static unsigned long change_prot_numa(struct vm_area_struct *vma,
636 unsigned long addr, unsigned long end)
640 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
643 * Check if all pages in a range are on a set of nodes.
644 * If pagelist != NULL then isolate pages from the LRU and
645 * put them on the pagelist.
647 static struct vm_area_struct *
648 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
649 const nodemask_t *nodes, unsigned long flags, void *private)
652 struct vm_area_struct *first, *vma, *prev;
655 first = find_vma(mm, start);
657 return ERR_PTR(-EFAULT);
659 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
660 unsigned long endvma = vma->vm_end;
664 if (vma->vm_start > start)
665 start = vma->vm_start;
667 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
668 if (!vma->vm_next && vma->vm_end < end)
669 return ERR_PTR(-EFAULT);
670 if (prev && prev->vm_end < vma->vm_start)
671 return ERR_PTR(-EFAULT);
674 if (flags & MPOL_MF_LAZY) {
675 change_prot_numa(vma, start, endvma);
679 if ((flags & MPOL_MF_STRICT) ||
680 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
681 vma_migratable(vma))) {
683 err = check_pgd_range(vma, start, endvma, nodes,
686 first = ERR_PTR(err);
697 * Apply policy to a single VMA
698 * This must be called with the mmap_sem held for writing.
700 static int vma_replace_policy(struct vm_area_struct *vma,
701 struct mempolicy *pol)
704 struct mempolicy *old;
705 struct mempolicy *new;
707 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
708 vma->vm_start, vma->vm_end, vma->vm_pgoff,
709 vma->vm_ops, vma->vm_file,
710 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
716 if (vma->vm_ops && vma->vm_ops->set_policy) {
717 err = vma->vm_ops->set_policy(vma, new);
722 old = vma->vm_policy;
723 vma->vm_policy = new; /* protected by mmap_sem */
732 /* Step 2: apply policy to a range and do splits. */
733 static int mbind_range(struct mm_struct *mm, unsigned long start,
734 unsigned long end, struct mempolicy *new_pol)
736 struct vm_area_struct *next;
737 struct vm_area_struct *prev;
738 struct vm_area_struct *vma;
741 unsigned long vmstart;
744 vma = find_vma(mm, start);
745 if (!vma || vma->vm_start > start)
749 if (start > vma->vm_start)
752 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
754 vmstart = max(start, vma->vm_start);
755 vmend = min(end, vma->vm_end);
757 if (mpol_equal(vma_policy(vma), new_pol))
760 pgoff = vma->vm_pgoff +
761 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
762 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
763 vma->anon_vma, vma->vm_file, pgoff,
768 if (mpol_equal(vma_policy(vma), new_pol))
770 /* vma_merge() joined vma && vma->next, case 8 */
773 if (vma->vm_start != vmstart) {
774 err = split_vma(vma->vm_mm, vma, vmstart, 1);
778 if (vma->vm_end != vmend) {
779 err = split_vma(vma->vm_mm, vma, vmend, 0);
784 err = vma_replace_policy(vma, new_pol);
794 * Update task->flags PF_MEMPOLICY bit: set iff non-default
795 * mempolicy. Allows more rapid checking of this (combined perhaps
796 * with other PF_* flag bits) on memory allocation hot code paths.
798 * If called from outside this file, the task 'p' should -only- be
799 * a newly forked child not yet visible on the task list, because
800 * manipulating the task flags of a visible task is not safe.
802 * The above limitation is why this routine has the funny name
803 * mpol_fix_fork_child_flag().
805 * It is also safe to call this with a task pointer of current,
806 * which the static wrapper mpol_set_task_struct_flag() does,
807 * for use within this file.
810 void mpol_fix_fork_child_flag(struct task_struct *p)
813 p->flags |= PF_MEMPOLICY;
815 p->flags &= ~PF_MEMPOLICY;
818 static void mpol_set_task_struct_flag(void)
820 mpol_fix_fork_child_flag(current);
823 /* Set the process memory policy */
824 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
827 struct mempolicy *new, *old;
828 struct mm_struct *mm = current->mm;
829 NODEMASK_SCRATCH(scratch);
835 new = mpol_new(mode, flags, nodes);
841 * prevent changing our mempolicy while show_numa_maps()
843 * Note: do_set_mempolicy() can be called at init time
847 down_write(&mm->mmap_sem);
849 ret = mpol_set_nodemask(new, nodes, scratch);
851 task_unlock(current);
853 up_write(&mm->mmap_sem);
857 old = current->mempolicy;
858 current->mempolicy = new;
859 mpol_set_task_struct_flag();
860 if (new && new->mode == MPOL_INTERLEAVE &&
861 nodes_weight(new->v.nodes))
862 current->il_next = first_node(new->v.nodes);
863 task_unlock(current);
865 up_write(&mm->mmap_sem);
870 NODEMASK_SCRATCH_FREE(scratch);
875 * Return nodemask for policy for get_mempolicy() query
877 * Called with task's alloc_lock held
879 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
882 if (p == &default_policy)
888 case MPOL_INTERLEAVE:
892 if (!(p->flags & MPOL_F_LOCAL))
893 node_set(p->v.preferred_node, *nodes);
894 /* else return empty node mask for local allocation */
901 static int lookup_node(struct mm_struct *mm, unsigned long addr)
906 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
908 err = page_to_nid(p);
914 /* Retrieve NUMA policy */
915 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
916 unsigned long addr, unsigned long flags)
919 struct mm_struct *mm = current->mm;
920 struct vm_area_struct *vma = NULL;
921 struct mempolicy *pol = current->mempolicy;
924 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
927 if (flags & MPOL_F_MEMS_ALLOWED) {
928 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
930 *policy = 0; /* just so it's initialized */
932 *nmask = cpuset_current_mems_allowed;
933 task_unlock(current);
937 if (flags & MPOL_F_ADDR) {
939 * Do NOT fall back to task policy if the
940 * vma/shared policy at addr is NULL. We
941 * want to return MPOL_DEFAULT in this case.
943 down_read(&mm->mmap_sem);
944 vma = find_vma_intersection(mm, addr, addr+1);
946 up_read(&mm->mmap_sem);
949 if (vma->vm_ops && vma->vm_ops->get_policy)
950 pol = vma->vm_ops->get_policy(vma, addr);
952 pol = vma->vm_policy;
957 pol = &default_policy; /* indicates default behavior */
959 if (flags & MPOL_F_NODE) {
960 if (flags & MPOL_F_ADDR) {
961 err = lookup_node(mm, addr);
965 } else if (pol == current->mempolicy &&
966 pol->mode == MPOL_INTERLEAVE) {
967 *policy = current->il_next;
973 *policy = pol == &default_policy ? MPOL_DEFAULT :
976 * Internal mempolicy flags must be masked off before exposing
977 * the policy to userspace.
979 *policy |= (pol->flags & MPOL_MODE_FLAGS);
983 up_read(¤t->mm->mmap_sem);
989 if (mpol_store_user_nodemask(pol)) {
990 *nmask = pol->w.user_nodemask;
993 get_policy_nodemask(pol, nmask);
994 task_unlock(current);
1001 up_read(¤t->mm->mmap_sem);
1005 #ifdef CONFIG_MIGRATION
1009 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1010 unsigned long flags)
1013 * Avoid migrating a page that is shared with others.
1015 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1016 if (!isolate_lru_page(page)) {
1017 list_add_tail(&page->lru, pagelist);
1018 inc_zone_page_state(page, NR_ISOLATED_ANON +
1019 page_is_file_cache(page));
1024 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1027 return alloc_huge_page_node(page_hstate(compound_head(page)),
1030 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1034 * Migrate pages from one node to a target node.
1035 * Returns error or the number of pages not migrated.
1037 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1041 LIST_HEAD(pagelist);
1045 node_set(source, nmask);
1048 * This does not "check" the range but isolates all pages that
1049 * need migration. Between passing in the full user address
1050 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1052 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1053 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1054 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1056 if (!list_empty(&pagelist)) {
1057 err = migrate_pages(&pagelist, new_node_page, dest,
1058 MIGRATE_SYNC, MR_SYSCALL);
1060 putback_movable_pages(&pagelist);
1067 * Move pages between the two nodesets so as to preserve the physical
1068 * layout as much as possible.
1070 * Returns the number of page that could not be moved.
1072 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1073 const nodemask_t *to, int flags)
1079 err = migrate_prep();
1083 down_read(&mm->mmap_sem);
1085 err = migrate_vmas(mm, from, to, flags);
1090 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1091 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1092 * bit in 'tmp', and return that <source, dest> pair for migration.
1093 * The pair of nodemasks 'to' and 'from' define the map.
1095 * If no pair of bits is found that way, fallback to picking some
1096 * pair of 'source' and 'dest' bits that are not the same. If the
1097 * 'source' and 'dest' bits are the same, this represents a node
1098 * that will be migrating to itself, so no pages need move.
1100 * If no bits are left in 'tmp', or if all remaining bits left
1101 * in 'tmp' correspond to the same bit in 'to', return false
1102 * (nothing left to migrate).
1104 * This lets us pick a pair of nodes to migrate between, such that
1105 * if possible the dest node is not already occupied by some other
1106 * source node, minimizing the risk of overloading the memory on a
1107 * node that would happen if we migrated incoming memory to a node
1108 * before migrating outgoing memory source that same node.
1110 * A single scan of tmp is sufficient. As we go, we remember the
1111 * most recent <s, d> pair that moved (s != d). If we find a pair
1112 * that not only moved, but what's better, moved to an empty slot
1113 * (d is not set in tmp), then we break out then, with that pair.
1114 * Otherwise when we finish scanning from_tmp, we at least have the
1115 * most recent <s, d> pair that moved. If we get all the way through
1116 * the scan of tmp without finding any node that moved, much less
1117 * moved to an empty node, then there is nothing left worth migrating.
1121 while (!nodes_empty(tmp)) {
1126 for_each_node_mask(s, tmp) {
1129 * do_migrate_pages() tries to maintain the relative
1130 * node relationship of the pages established between
1131 * threads and memory areas.
1133 * However if the number of source nodes is not equal to
1134 * the number of destination nodes we can not preserve
1135 * this node relative relationship. In that case, skip
1136 * copying memory from a node that is in the destination
1139 * Example: [2,3,4] -> [3,4,5] moves everything.
1140 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1143 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1144 (node_isset(s, *to)))
1147 d = node_remap(s, *from, *to);
1151 source = s; /* Node moved. Memorize */
1154 /* dest not in remaining from nodes? */
1155 if (!node_isset(dest, tmp))
1161 node_clear(source, tmp);
1162 err = migrate_to_node(mm, source, dest, flags);
1169 up_read(&mm->mmap_sem);
1177 * Allocate a new page for page migration based on vma policy.
1178 * Start assuming that page is mapped by vma pointed to by @private.
1179 * Search forward from there, if not. N.B., this assumes that the
1180 * list of pages handed to migrate_pages()--which is how we get here--
1181 * is in virtual address order.
1183 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1185 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1186 unsigned long uninitialized_var(address);
1189 address = page_address_in_vma(page, vma);
1190 if (address != -EFAULT)
1196 * if !vma, alloc_page_vma() will use task or system default policy
1198 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1202 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1203 unsigned long flags)
1207 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1208 const nodemask_t *to, int flags)
1213 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1219 static long do_mbind(unsigned long start, unsigned long len,
1220 unsigned short mode, unsigned short mode_flags,
1221 nodemask_t *nmask, unsigned long flags)
1223 struct vm_area_struct *vma;
1224 struct mm_struct *mm = current->mm;
1225 struct mempolicy *new;
1228 LIST_HEAD(pagelist);
1230 if (flags & ~(unsigned long)MPOL_MF_VALID)
1232 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1235 if (start & ~PAGE_MASK)
1238 if (mode == MPOL_DEFAULT)
1239 flags &= ~MPOL_MF_STRICT;
1241 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1249 new = mpol_new(mode, mode_flags, nmask);
1251 return PTR_ERR(new);
1253 if (flags & MPOL_MF_LAZY)
1254 new->flags |= MPOL_F_MOF;
1257 * If we are using the default policy then operation
1258 * on discontinuous address spaces is okay after all
1261 flags |= MPOL_MF_DISCONTIG_OK;
1263 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1264 start, start + len, mode, mode_flags,
1265 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1267 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1269 err = migrate_prep();
1274 NODEMASK_SCRATCH(scratch);
1276 down_write(&mm->mmap_sem);
1278 err = mpol_set_nodemask(new, nmask, scratch);
1279 task_unlock(current);
1281 up_write(&mm->mmap_sem);
1284 NODEMASK_SCRATCH_FREE(scratch);
1289 vma = check_range(mm, start, end, nmask,
1290 flags | MPOL_MF_INVERT, &pagelist);
1292 err = PTR_ERR(vma); /* maybe ... */
1294 err = mbind_range(mm, start, end, new);
1299 if (!list_empty(&pagelist)) {
1300 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1301 nr_failed = migrate_pages(&pagelist, new_vma_page,
1303 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1305 putback_lru_pages(&pagelist);
1308 if (nr_failed && (flags & MPOL_MF_STRICT))
1311 putback_lru_pages(&pagelist);
1313 up_write(&mm->mmap_sem);
1320 * User space interface with variable sized bitmaps for nodelists.
1323 /* Copy a node mask from user space. */
1324 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1325 unsigned long maxnode)
1328 unsigned long nlongs;
1329 unsigned long endmask;
1332 nodes_clear(*nodes);
1333 if (maxnode == 0 || !nmask)
1335 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1338 nlongs = BITS_TO_LONGS(maxnode);
1339 if ((maxnode % BITS_PER_LONG) == 0)
1342 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1344 /* When the user specified more nodes than supported just check
1345 if the non supported part is all zero. */
1346 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1347 if (nlongs > PAGE_SIZE/sizeof(long))
1349 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1351 if (get_user(t, nmask + k))
1353 if (k == nlongs - 1) {
1359 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1363 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1365 nodes_addr(*nodes)[nlongs-1] &= endmask;
1369 /* Copy a kernel node mask to user space */
1370 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1373 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1374 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1376 if (copy > nbytes) {
1377 if (copy > PAGE_SIZE)
1379 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1383 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1386 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1387 unsigned long, mode, unsigned long __user *, nmask,
1388 unsigned long, maxnode, unsigned, flags)
1392 unsigned short mode_flags;
1394 mode_flags = mode & MPOL_MODE_FLAGS;
1395 mode &= ~MPOL_MODE_FLAGS;
1396 if (mode >= MPOL_MAX)
1398 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1399 (mode_flags & MPOL_F_RELATIVE_NODES))
1401 err = get_nodes(&nodes, nmask, maxnode);
1404 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1407 /* Set the process memory policy */
1408 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1409 unsigned long, maxnode)
1413 unsigned short flags;
1415 flags = mode & MPOL_MODE_FLAGS;
1416 mode &= ~MPOL_MODE_FLAGS;
1417 if ((unsigned int)mode >= MPOL_MAX)
1419 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1421 err = get_nodes(&nodes, nmask, maxnode);
1424 return do_set_mempolicy(mode, flags, &nodes);
1427 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1428 const unsigned long __user *, old_nodes,
1429 const unsigned long __user *, new_nodes)
1431 const struct cred *cred = current_cred(), *tcred;
1432 struct mm_struct *mm = NULL;
1433 struct task_struct *task;
1434 nodemask_t task_nodes;
1438 NODEMASK_SCRATCH(scratch);
1443 old = &scratch->mask1;
1444 new = &scratch->mask2;
1446 err = get_nodes(old, old_nodes, maxnode);
1450 err = get_nodes(new, new_nodes, maxnode);
1454 /* Find the mm_struct */
1456 task = pid ? find_task_by_vpid(pid) : current;
1462 get_task_struct(task);
1467 * Check if this process has the right to modify the specified
1468 * process. The right exists if the process has administrative
1469 * capabilities, superuser privileges or the same
1470 * userid as the target process.
1472 tcred = __task_cred(task);
1473 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1474 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1475 !capable(CAP_SYS_NICE)) {
1482 task_nodes = cpuset_mems_allowed(task);
1483 /* Is the user allowed to access the target nodes? */
1484 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1489 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1494 err = security_task_movememory(task);
1498 mm = get_task_mm(task);
1499 put_task_struct(task);
1506 err = do_migrate_pages(mm, old, new,
1507 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1511 NODEMASK_SCRATCH_FREE(scratch);
1516 put_task_struct(task);
1522 /* Retrieve NUMA policy */
1523 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1524 unsigned long __user *, nmask, unsigned long, maxnode,
1525 unsigned long, addr, unsigned long, flags)
1528 int uninitialized_var(pval);
1531 if (nmask != NULL && maxnode < MAX_NUMNODES)
1534 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1539 if (policy && put_user(pval, policy))
1543 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1548 #ifdef CONFIG_COMPAT
1550 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1551 compat_ulong_t __user *nmask,
1552 compat_ulong_t maxnode,
1553 compat_ulong_t addr, compat_ulong_t flags)
1556 unsigned long __user *nm = NULL;
1557 unsigned long nr_bits, alloc_size;
1558 DECLARE_BITMAP(bm, MAX_NUMNODES);
1560 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1561 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1564 nm = compat_alloc_user_space(alloc_size);
1566 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1568 if (!err && nmask) {
1569 unsigned long copy_size;
1570 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1571 err = copy_from_user(bm, nm, copy_size);
1572 /* ensure entire bitmap is zeroed */
1573 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1574 err |= compat_put_bitmap(nmask, bm, nr_bits);
1580 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1581 compat_ulong_t maxnode)
1584 unsigned long __user *nm = NULL;
1585 unsigned long nr_bits, alloc_size;
1586 DECLARE_BITMAP(bm, MAX_NUMNODES);
1588 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1589 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1592 err = compat_get_bitmap(bm, nmask, nr_bits);
1593 nm = compat_alloc_user_space(alloc_size);
1594 err |= copy_to_user(nm, bm, alloc_size);
1600 return sys_set_mempolicy(mode, nm, nr_bits+1);
1603 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1604 compat_ulong_t mode, compat_ulong_t __user *nmask,
1605 compat_ulong_t maxnode, compat_ulong_t flags)
1608 unsigned long __user *nm = NULL;
1609 unsigned long nr_bits, alloc_size;
1612 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1613 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1616 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1617 nm = compat_alloc_user_space(alloc_size);
1618 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1624 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1630 * get_vma_policy(@task, @vma, @addr)
1631 * @task - task for fallback if vma policy == default
1632 * @vma - virtual memory area whose policy is sought
1633 * @addr - address in @vma for shared policy lookup
1635 * Returns effective policy for a VMA at specified address.
1636 * Falls back to @task or system default policy, as necessary.
1637 * Current or other task's task mempolicy and non-shared vma policies must be
1638 * protected by task_lock(task) by the caller.
1639 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1640 * count--added by the get_policy() vm_op, as appropriate--to protect against
1641 * freeing by another task. It is the caller's responsibility to free the
1642 * extra reference for shared policies.
1644 struct mempolicy *get_vma_policy(struct task_struct *task,
1645 struct vm_area_struct *vma, unsigned long addr)
1647 struct mempolicy *pol = get_task_policy(task);
1650 if (vma->vm_ops && vma->vm_ops->get_policy) {
1651 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1655 } else if (vma->vm_policy) {
1656 pol = vma->vm_policy;
1659 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1660 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1661 * count on these policies which will be dropped by
1662 * mpol_cond_put() later
1664 if (mpol_needs_cond_ref(pol))
1669 pol = &default_policy;
1673 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1675 enum zone_type dynamic_policy_zone = policy_zone;
1677 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1680 * if policy->v.nodes has movable memory only,
1681 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1683 * policy->v.nodes is intersect with node_states[N_MEMORY].
1684 * so if the following test faile, it implies
1685 * policy->v.nodes has movable memory only.
1687 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1688 dynamic_policy_zone = ZONE_MOVABLE;
1690 return zone >= dynamic_policy_zone;
1694 * Return a nodemask representing a mempolicy for filtering nodes for
1697 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1699 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1700 if (unlikely(policy->mode == MPOL_BIND) &&
1701 apply_policy_zone(policy, gfp_zone(gfp)) &&
1702 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1703 return &policy->v.nodes;
1708 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1709 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1712 switch (policy->mode) {
1713 case MPOL_PREFERRED:
1714 if (!(policy->flags & MPOL_F_LOCAL))
1715 nd = policy->v.preferred_node;
1719 * Normally, MPOL_BIND allocations are node-local within the
1720 * allowed nodemask. However, if __GFP_THISNODE is set and the
1721 * current node isn't part of the mask, we use the zonelist for
1722 * the first node in the mask instead.
1724 if (unlikely(gfp & __GFP_THISNODE) &&
1725 unlikely(!node_isset(nd, policy->v.nodes)))
1726 nd = first_node(policy->v.nodes);
1731 return node_zonelist(nd, gfp);
1734 /* Do dynamic interleaving for a process */
1735 static unsigned interleave_nodes(struct mempolicy *policy)
1738 struct task_struct *me = current;
1741 next = next_node(nid, policy->v.nodes);
1742 if (next >= MAX_NUMNODES)
1743 next = first_node(policy->v.nodes);
1744 if (next < MAX_NUMNODES)
1750 * Depending on the memory policy provide a node from which to allocate the
1752 * @policy must be protected by freeing by the caller. If @policy is
1753 * the current task's mempolicy, this protection is implicit, as only the
1754 * task can change it's policy. The system default policy requires no
1757 unsigned slab_node(void)
1759 struct mempolicy *policy;
1762 return numa_node_id();
1764 policy = current->mempolicy;
1765 if (!policy || policy->flags & MPOL_F_LOCAL)
1766 return numa_node_id();
1768 switch (policy->mode) {
1769 case MPOL_PREFERRED:
1771 * handled MPOL_F_LOCAL above
1773 return policy->v.preferred_node;
1775 case MPOL_INTERLEAVE:
1776 return interleave_nodes(policy);
1780 * Follow bind policy behavior and start allocation at the
1783 struct zonelist *zonelist;
1785 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1786 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1787 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1790 return zone ? zone->node : numa_node_id();
1798 /* Do static interleaving for a VMA with known offset. */
1799 static unsigned offset_il_node(struct mempolicy *pol,
1800 struct vm_area_struct *vma, unsigned long off)
1802 unsigned nnodes = nodes_weight(pol->v.nodes);
1808 return numa_node_id();
1809 target = (unsigned int)off % nnodes;
1812 nid = next_node(nid, pol->v.nodes);
1814 } while (c <= target);
1818 /* Determine a node number for interleave */
1819 static inline unsigned interleave_nid(struct mempolicy *pol,
1820 struct vm_area_struct *vma, unsigned long addr, int shift)
1826 * for small pages, there is no difference between
1827 * shift and PAGE_SHIFT, so the bit-shift is safe.
1828 * for huge pages, since vm_pgoff is in units of small
1829 * pages, we need to shift off the always 0 bits to get
1832 BUG_ON(shift < PAGE_SHIFT);
1833 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1834 off += (addr - vma->vm_start) >> shift;
1835 return offset_il_node(pol, vma, off);
1837 return interleave_nodes(pol);
1841 * Return the bit number of a random bit set in the nodemask.
1842 * (returns -1 if nodemask is empty)
1844 int node_random(const nodemask_t *maskp)
1848 w = nodes_weight(*maskp);
1850 bit = bitmap_ord_to_pos(maskp->bits,
1851 get_random_int() % w, MAX_NUMNODES);
1855 #ifdef CONFIG_HUGETLBFS
1857 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1858 * @vma = virtual memory area whose policy is sought
1859 * @addr = address in @vma for shared policy lookup and interleave policy
1860 * @gfp_flags = for requested zone
1861 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1862 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1864 * Returns a zonelist suitable for a huge page allocation and a pointer
1865 * to the struct mempolicy for conditional unref after allocation.
1866 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1867 * @nodemask for filtering the zonelist.
1869 * Must be protected by get_mems_allowed()
1871 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1872 gfp_t gfp_flags, struct mempolicy **mpol,
1873 nodemask_t **nodemask)
1875 struct zonelist *zl;
1877 *mpol = get_vma_policy(current, vma, addr);
1878 *nodemask = NULL; /* assume !MPOL_BIND */
1880 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1881 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1882 huge_page_shift(hstate_vma(vma))), gfp_flags);
1884 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1885 if ((*mpol)->mode == MPOL_BIND)
1886 *nodemask = &(*mpol)->v.nodes;
1892 * init_nodemask_of_mempolicy
1894 * If the current task's mempolicy is "default" [NULL], return 'false'
1895 * to indicate default policy. Otherwise, extract the policy nodemask
1896 * for 'bind' or 'interleave' policy into the argument nodemask, or
1897 * initialize the argument nodemask to contain the single node for
1898 * 'preferred' or 'local' policy and return 'true' to indicate presence
1899 * of non-default mempolicy.
1901 * We don't bother with reference counting the mempolicy [mpol_get/put]
1902 * because the current task is examining it's own mempolicy and a task's
1903 * mempolicy is only ever changed by the task itself.
1905 * N.B., it is the caller's responsibility to free a returned nodemask.
1907 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1909 struct mempolicy *mempolicy;
1912 if (!(mask && current->mempolicy))
1916 mempolicy = current->mempolicy;
1917 switch (mempolicy->mode) {
1918 case MPOL_PREFERRED:
1919 if (mempolicy->flags & MPOL_F_LOCAL)
1920 nid = numa_node_id();
1922 nid = mempolicy->v.preferred_node;
1923 init_nodemask_of_node(mask, nid);
1928 case MPOL_INTERLEAVE:
1929 *mask = mempolicy->v.nodes;
1935 task_unlock(current);
1942 * mempolicy_nodemask_intersects
1944 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1945 * policy. Otherwise, check for intersection between mask and the policy
1946 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1947 * policy, always return true since it may allocate elsewhere on fallback.
1949 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1951 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1952 const nodemask_t *mask)
1954 struct mempolicy *mempolicy;
1960 mempolicy = tsk->mempolicy;
1964 switch (mempolicy->mode) {
1965 case MPOL_PREFERRED:
1967 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1968 * allocate from, they may fallback to other nodes when oom.
1969 * Thus, it's possible for tsk to have allocated memory from
1974 case MPOL_INTERLEAVE:
1975 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1985 /* Allocate a page in interleaved policy.
1986 Own path because it needs to do special accounting. */
1987 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1990 struct zonelist *zl;
1993 zl = node_zonelist(nid, gfp);
1994 page = __alloc_pages(gfp, order, zl);
1995 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1996 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2001 * alloc_pages_vma - Allocate a page for a VMA.
2004 * %GFP_USER user allocation.
2005 * %GFP_KERNEL kernel allocations,
2006 * %GFP_HIGHMEM highmem/user allocations,
2007 * %GFP_FS allocation should not call back into a file system.
2008 * %GFP_ATOMIC don't sleep.
2010 * @order:Order of the GFP allocation.
2011 * @vma: Pointer to VMA or NULL if not available.
2012 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2014 * This function allocates a page from the kernel page pool and applies
2015 * a NUMA policy associated with the VMA or the current process.
2016 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2017 * mm_struct of the VMA to prevent it from going away. Should be used for
2018 * all allocations for pages that will be mapped into
2019 * user space. Returns NULL when no page can be allocated.
2021 * Should be called with the mm_sem of the vma hold.
2024 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2025 unsigned long addr, int node)
2027 struct mempolicy *pol;
2029 unsigned int cpuset_mems_cookie;
2032 pol = get_vma_policy(current, vma, addr);
2033 cpuset_mems_cookie = get_mems_allowed();
2035 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2038 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2040 page = alloc_page_interleave(gfp, order, nid);
2041 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2046 page = __alloc_pages_nodemask(gfp, order,
2047 policy_zonelist(gfp, pol, node),
2048 policy_nodemask(gfp, pol));
2049 if (unlikely(mpol_needs_cond_ref(pol)))
2051 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2057 * alloc_pages_current - Allocate pages.
2060 * %GFP_USER user allocation,
2061 * %GFP_KERNEL kernel allocation,
2062 * %GFP_HIGHMEM highmem allocation,
2063 * %GFP_FS don't call back into a file system.
2064 * %GFP_ATOMIC don't sleep.
2065 * @order: Power of two of allocation size in pages. 0 is a single page.
2067 * Allocate a page from the kernel page pool. When not in
2068 * interrupt context and apply the current process NUMA policy.
2069 * Returns NULL when no page can be allocated.
2071 * Don't call cpuset_update_task_memory_state() unless
2072 * 1) it's ok to take cpuset_sem (can WAIT), and
2073 * 2) allocating for current task (not interrupt).
2075 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2077 struct mempolicy *pol = get_task_policy(current);
2079 unsigned int cpuset_mems_cookie;
2081 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2082 pol = &default_policy;
2085 cpuset_mems_cookie = get_mems_allowed();
2088 * No reference counting needed for current->mempolicy
2089 * nor system default_policy
2091 if (pol->mode == MPOL_INTERLEAVE)
2092 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2094 page = __alloc_pages_nodemask(gfp, order,
2095 policy_zonelist(gfp, pol, numa_node_id()),
2096 policy_nodemask(gfp, pol));
2098 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2103 EXPORT_SYMBOL(alloc_pages_current);
2105 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2107 struct mempolicy *pol = mpol_dup(vma_policy(src));
2110 return PTR_ERR(pol);
2111 dst->vm_policy = pol;
2116 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2117 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2118 * with the mems_allowed returned by cpuset_mems_allowed(). This
2119 * keeps mempolicies cpuset relative after its cpuset moves. See
2120 * further kernel/cpuset.c update_nodemask().
2122 * current's mempolicy may be rebinded by the other task(the task that changes
2123 * cpuset's mems), so we needn't do rebind work for current task.
2126 /* Slow path of a mempolicy duplicate */
2127 struct mempolicy *__mpol_dup(struct mempolicy *old)
2129 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2132 return ERR_PTR(-ENOMEM);
2134 /* task's mempolicy is protected by alloc_lock */
2135 if (old == current->mempolicy) {
2138 task_unlock(current);
2143 if (current_cpuset_is_being_rebound()) {
2144 nodemask_t mems = cpuset_mems_allowed(current);
2145 if (new->flags & MPOL_F_REBINDING)
2146 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2148 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2151 atomic_set(&new->refcnt, 1);
2155 /* Slow path of a mempolicy comparison */
2156 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2160 if (a->mode != b->mode)
2162 if (a->flags != b->flags)
2164 if (mpol_store_user_nodemask(a))
2165 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2171 case MPOL_INTERLEAVE:
2172 return !!nodes_equal(a->v.nodes, b->v.nodes);
2173 case MPOL_PREFERRED:
2174 return a->v.preferred_node == b->v.preferred_node;
2182 * Shared memory backing store policy support.
2184 * Remember policies even when nobody has shared memory mapped.
2185 * The policies are kept in Red-Black tree linked from the inode.
2186 * They are protected by the sp->lock spinlock, which should be held
2187 * for any accesses to the tree.
2190 /* lookup first element intersecting start-end */
2191 /* Caller holds sp->lock */
2192 static struct sp_node *
2193 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2195 struct rb_node *n = sp->root.rb_node;
2198 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2200 if (start >= p->end)
2202 else if (end <= p->start)
2210 struct sp_node *w = NULL;
2211 struct rb_node *prev = rb_prev(n);
2214 w = rb_entry(prev, struct sp_node, nd);
2215 if (w->end <= start)
2219 return rb_entry(n, struct sp_node, nd);
2222 /* Insert a new shared policy into the list. */
2223 /* Caller holds sp->lock */
2224 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2226 struct rb_node **p = &sp->root.rb_node;
2227 struct rb_node *parent = NULL;
2232 nd = rb_entry(parent, struct sp_node, nd);
2233 if (new->start < nd->start)
2235 else if (new->end > nd->end)
2236 p = &(*p)->rb_right;
2240 rb_link_node(&new->nd, parent, p);
2241 rb_insert_color(&new->nd, &sp->root);
2242 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2243 new->policy ? new->policy->mode : 0);
2246 /* Find shared policy intersecting idx */
2248 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2250 struct mempolicy *pol = NULL;
2253 if (!sp->root.rb_node)
2255 spin_lock(&sp->lock);
2256 sn = sp_lookup(sp, idx, idx+1);
2258 mpol_get(sn->policy);
2261 spin_unlock(&sp->lock);
2265 static void sp_free(struct sp_node *n)
2267 mpol_put(n->policy);
2268 kmem_cache_free(sn_cache, n);
2272 * mpol_misplaced - check whether current page node is valid in policy
2274 * @page - page to be checked
2275 * @vma - vm area where page mapped
2276 * @addr - virtual address where page mapped
2278 * Lookup current policy node id for vma,addr and "compare to" page's
2282 * -1 - not misplaced, page is in the right node
2283 * node - node id where the page should be
2285 * Policy determination "mimics" alloc_page_vma().
2286 * Called from fault path where we know the vma and faulting address.
2288 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2290 struct mempolicy *pol;
2292 int curnid = page_to_nid(page);
2293 unsigned long pgoff;
2299 pol = get_vma_policy(current, vma, addr);
2300 if (!(pol->flags & MPOL_F_MOF))
2303 switch (pol->mode) {
2304 case MPOL_INTERLEAVE:
2305 BUG_ON(addr >= vma->vm_end);
2306 BUG_ON(addr < vma->vm_start);
2308 pgoff = vma->vm_pgoff;
2309 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2310 polnid = offset_il_node(pol, vma, pgoff);
2313 case MPOL_PREFERRED:
2314 if (pol->flags & MPOL_F_LOCAL)
2315 polnid = numa_node_id();
2317 polnid = pol->v.preferred_node;
2322 * allows binding to multiple nodes.
2323 * use current page if in policy nodemask,
2324 * else select nearest allowed node, if any.
2325 * If no allowed nodes, use current [!misplaced].
2327 if (node_isset(curnid, pol->v.nodes))
2329 (void)first_zones_zonelist(
2330 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2331 gfp_zone(GFP_HIGHUSER),
2332 &pol->v.nodes, &zone);
2333 polnid = zone->node;
2340 /* Migrate the page towards the node whose CPU is referencing it */
2341 if (pol->flags & MPOL_F_MORON) {
2344 polnid = numa_node_id();
2347 * Multi-stage node selection is used in conjunction
2348 * with a periodic migration fault to build a temporal
2349 * task<->page relation. By using a two-stage filter we
2350 * remove short/unlikely relations.
2352 * Using P(p) ~ n_p / n_t as per frequentist
2353 * probability, we can equate a task's usage of a
2354 * particular page (n_p) per total usage of this
2355 * page (n_t) (in a given time-span) to a probability.
2357 * Our periodic faults will sample this probability and
2358 * getting the same result twice in a row, given these
2359 * samples are fully independent, is then given by
2360 * P(n)^2, provided our sample period is sufficiently
2361 * short compared to the usage pattern.
2363 * This quadric squishes small probabilities, making
2364 * it less likely we act on an unlikely task<->page
2367 last_nid = page_nid_xchg_last(page, polnid);
2368 if (last_nid != polnid)
2372 if (curnid != polnid)
2380 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2382 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2383 rb_erase(&n->nd, &sp->root);
2387 static void sp_node_init(struct sp_node *node, unsigned long start,
2388 unsigned long end, struct mempolicy *pol)
2390 node->start = start;
2395 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2396 struct mempolicy *pol)
2399 struct mempolicy *newpol;
2401 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2405 newpol = mpol_dup(pol);
2406 if (IS_ERR(newpol)) {
2407 kmem_cache_free(sn_cache, n);
2410 newpol->flags |= MPOL_F_SHARED;
2411 sp_node_init(n, start, end, newpol);
2416 /* Replace a policy range. */
2417 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2418 unsigned long end, struct sp_node *new)
2421 struct sp_node *n_new = NULL;
2422 struct mempolicy *mpol_new = NULL;
2426 spin_lock(&sp->lock);
2427 n = sp_lookup(sp, start, end);
2428 /* Take care of old policies in the same range. */
2429 while (n && n->start < end) {
2430 struct rb_node *next = rb_next(&n->nd);
2431 if (n->start >= start) {
2437 /* Old policy spanning whole new range. */
2442 *mpol_new = *n->policy;
2443 atomic_set(&mpol_new->refcnt, 1);
2444 sp_node_init(n_new, end, n->end, mpol_new);
2446 sp_insert(sp, n_new);
2455 n = rb_entry(next, struct sp_node, nd);
2459 spin_unlock(&sp->lock);
2466 kmem_cache_free(sn_cache, n_new);
2471 spin_unlock(&sp->lock);
2473 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2476 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2483 * mpol_shared_policy_init - initialize shared policy for inode
2484 * @sp: pointer to inode shared policy
2485 * @mpol: struct mempolicy to install
2487 * Install non-NULL @mpol in inode's shared policy rb-tree.
2488 * On entry, the current task has a reference on a non-NULL @mpol.
2489 * This must be released on exit.
2490 * This is called at get_inode() calls and we can use GFP_KERNEL.
2492 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2496 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2497 spin_lock_init(&sp->lock);
2500 struct vm_area_struct pvma;
2501 struct mempolicy *new;
2502 NODEMASK_SCRATCH(scratch);
2506 /* contextualize the tmpfs mount point mempolicy */
2507 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2509 goto free_scratch; /* no valid nodemask intersection */
2512 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2513 task_unlock(current);
2517 /* Create pseudo-vma that contains just the policy */
2518 memset(&pvma, 0, sizeof(struct vm_area_struct));
2519 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2520 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2523 mpol_put(new); /* drop initial ref */
2525 NODEMASK_SCRATCH_FREE(scratch);
2527 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2531 int mpol_set_shared_policy(struct shared_policy *info,
2532 struct vm_area_struct *vma, struct mempolicy *npol)
2535 struct sp_node *new = NULL;
2536 unsigned long sz = vma_pages(vma);
2538 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2540 sz, npol ? npol->mode : -1,
2541 npol ? npol->flags : -1,
2542 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2545 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2549 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2555 /* Free a backing policy store on inode delete. */
2556 void mpol_free_shared_policy(struct shared_policy *p)
2559 struct rb_node *next;
2561 if (!p->root.rb_node)
2563 spin_lock(&p->lock);
2564 next = rb_first(&p->root);
2566 n = rb_entry(next, struct sp_node, nd);
2567 next = rb_next(&n->nd);
2570 spin_unlock(&p->lock);
2573 #ifdef CONFIG_NUMA_BALANCING
2574 static bool __initdata numabalancing_override;
2576 static void __init check_numabalancing_enable(void)
2578 bool numabalancing_default = false;
2580 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2581 numabalancing_default = true;
2583 if (nr_node_ids > 1 && !numabalancing_override) {
2584 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2585 "Configure with numa_balancing= or sysctl");
2586 set_numabalancing_state(numabalancing_default);
2590 static int __init setup_numabalancing(char *str)
2595 numabalancing_override = true;
2597 if (!strcmp(str, "enable")) {
2598 set_numabalancing_state(true);
2600 } else if (!strcmp(str, "disable")) {
2601 set_numabalancing_state(false);
2606 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2610 __setup("numa_balancing=", setup_numabalancing);
2612 static inline void __init check_numabalancing_enable(void)
2615 #endif /* CONFIG_NUMA_BALANCING */
2617 /* assumes fs == KERNEL_DS */
2618 void __init numa_policy_init(void)
2620 nodemask_t interleave_nodes;
2621 unsigned long largest = 0;
2622 int nid, prefer = 0;
2624 policy_cache = kmem_cache_create("numa_policy",
2625 sizeof(struct mempolicy),
2626 0, SLAB_PANIC, NULL);
2628 sn_cache = kmem_cache_create("shared_policy_node",
2629 sizeof(struct sp_node),
2630 0, SLAB_PANIC, NULL);
2632 for_each_node(nid) {
2633 preferred_node_policy[nid] = (struct mempolicy) {
2634 .refcnt = ATOMIC_INIT(1),
2635 .mode = MPOL_PREFERRED,
2636 .flags = MPOL_F_MOF | MPOL_F_MORON,
2637 .v = { .preferred_node = nid, },
2642 * Set interleaving policy for system init. Interleaving is only
2643 * enabled across suitably sized nodes (default is >= 16MB), or
2644 * fall back to the largest node if they're all smaller.
2646 nodes_clear(interleave_nodes);
2647 for_each_node_state(nid, N_MEMORY) {
2648 unsigned long total_pages = node_present_pages(nid);
2650 /* Preserve the largest node */
2651 if (largest < total_pages) {
2652 largest = total_pages;
2656 /* Interleave this node? */
2657 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2658 node_set(nid, interleave_nodes);
2661 /* All too small, use the largest */
2662 if (unlikely(nodes_empty(interleave_nodes)))
2663 node_set(prefer, interleave_nodes);
2665 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2666 printk("numa_policy_init: interleaving failed\n");
2668 check_numabalancing_enable();
2671 /* Reset policy of current process to default */
2672 void numa_default_policy(void)
2674 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2678 * Parse and format mempolicy from/to strings
2682 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2684 static const char * const policy_modes[] =
2686 [MPOL_DEFAULT] = "default",
2687 [MPOL_PREFERRED] = "prefer",
2688 [MPOL_BIND] = "bind",
2689 [MPOL_INTERLEAVE] = "interleave",
2690 [MPOL_LOCAL] = "local",
2696 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2697 * @str: string containing mempolicy to parse
2698 * @mpol: pointer to struct mempolicy pointer, returned on success.
2701 * <mode>[=<flags>][:<nodelist>]
2703 * On success, returns 0, else 1
2705 int mpol_parse_str(char *str, struct mempolicy **mpol)
2707 struct mempolicy *new = NULL;
2708 unsigned short mode;
2709 unsigned short mode_flags;
2711 char *nodelist = strchr(str, ':');
2712 char *flags = strchr(str, '=');
2716 /* NUL-terminate mode or flags string */
2718 if (nodelist_parse(nodelist, nodes))
2720 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2726 *flags++ = '\0'; /* terminate mode string */
2728 for (mode = 0; mode < MPOL_MAX; mode++) {
2729 if (!strcmp(str, policy_modes[mode])) {
2733 if (mode >= MPOL_MAX)
2737 case MPOL_PREFERRED:
2739 * Insist on a nodelist of one node only
2742 char *rest = nodelist;
2743 while (isdigit(*rest))
2749 case MPOL_INTERLEAVE:
2751 * Default to online nodes with memory if no nodelist
2754 nodes = node_states[N_MEMORY];
2758 * Don't allow a nodelist; mpol_new() checks flags
2762 mode = MPOL_PREFERRED;
2766 * Insist on a empty nodelist
2773 * Insist on a nodelist
2782 * Currently, we only support two mutually exclusive
2785 if (!strcmp(flags, "static"))
2786 mode_flags |= MPOL_F_STATIC_NODES;
2787 else if (!strcmp(flags, "relative"))
2788 mode_flags |= MPOL_F_RELATIVE_NODES;
2793 new = mpol_new(mode, mode_flags, &nodes);
2798 * Save nodes for mpol_to_str() to show the tmpfs mount options
2799 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2801 if (mode != MPOL_PREFERRED)
2802 new->v.nodes = nodes;
2804 new->v.preferred_node = first_node(nodes);
2806 new->flags |= MPOL_F_LOCAL;
2809 * Save nodes for contextualization: this will be used to "clone"
2810 * the mempolicy in a specific context [cpuset] at a later time.
2812 new->w.user_nodemask = nodes;
2817 /* Restore string for error message */
2826 #endif /* CONFIG_TMPFS */
2829 * mpol_to_str - format a mempolicy structure for printing
2830 * @buffer: to contain formatted mempolicy string
2831 * @maxlen: length of @buffer
2832 * @pol: pointer to mempolicy to be formatted
2834 * Convert a mempolicy into a string.
2835 * Returns the number of characters in buffer (if positive)
2836 * or an error (negative)
2838 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2843 unsigned short mode;
2844 unsigned short flags = pol ? pol->flags : 0;
2847 * Sanity check: room for longest mode, flag and some nodes
2849 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2851 if (!pol || pol == &default_policy)
2852 mode = MPOL_DEFAULT;
2861 case MPOL_PREFERRED:
2863 if (flags & MPOL_F_LOCAL)
2866 node_set(pol->v.preferred_node, nodes);
2871 case MPOL_INTERLEAVE:
2872 nodes = pol->v.nodes;
2879 l = strlen(policy_modes[mode]);
2880 if (buffer + maxlen < p + l + 1)
2883 strcpy(p, policy_modes[mode]);
2886 if (flags & MPOL_MODE_FLAGS) {
2887 if (buffer + maxlen < p + 2)
2892 * Currently, the only defined flags are mutually exclusive
2894 if (flags & MPOL_F_STATIC_NODES)
2895 p += snprintf(p, buffer + maxlen - p, "static");
2896 else if (flags & MPOL_F_RELATIVE_NODES)
2897 p += snprintf(p, buffer + maxlen - p, "relative");
2900 if (!nodes_empty(nodes)) {
2901 if (buffer + maxlen < p + 2)
2904 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);