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 return alloc_huge_page_noerr(vma, address, 1);
1198 * if !vma, alloc_page_vma() will use task or system default policy
1200 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1204 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1205 unsigned long flags)
1209 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1210 const nodemask_t *to, int flags)
1215 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1221 static long do_mbind(unsigned long start, unsigned long len,
1222 unsigned short mode, unsigned short mode_flags,
1223 nodemask_t *nmask, unsigned long flags)
1225 struct vm_area_struct *vma;
1226 struct mm_struct *mm = current->mm;
1227 struct mempolicy *new;
1230 LIST_HEAD(pagelist);
1232 if (flags & ~(unsigned long)MPOL_MF_VALID)
1234 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1237 if (start & ~PAGE_MASK)
1240 if (mode == MPOL_DEFAULT)
1241 flags &= ~MPOL_MF_STRICT;
1243 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1251 new = mpol_new(mode, mode_flags, nmask);
1253 return PTR_ERR(new);
1255 if (flags & MPOL_MF_LAZY)
1256 new->flags |= MPOL_F_MOF;
1259 * If we are using the default policy then operation
1260 * on discontinuous address spaces is okay after all
1263 flags |= MPOL_MF_DISCONTIG_OK;
1265 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1266 start, start + len, mode, mode_flags,
1267 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1269 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1271 err = migrate_prep();
1276 NODEMASK_SCRATCH(scratch);
1278 down_write(&mm->mmap_sem);
1280 err = mpol_set_nodemask(new, nmask, scratch);
1281 task_unlock(current);
1283 up_write(&mm->mmap_sem);
1286 NODEMASK_SCRATCH_FREE(scratch);
1291 vma = check_range(mm, start, end, nmask,
1292 flags | MPOL_MF_INVERT, &pagelist);
1294 err = PTR_ERR(vma); /* maybe ... */
1296 err = mbind_range(mm, start, end, new);
1301 if (!list_empty(&pagelist)) {
1302 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1303 nr_failed = migrate_pages(&pagelist, new_vma_page,
1305 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1307 putback_movable_pages(&pagelist);
1310 if (nr_failed && (flags & MPOL_MF_STRICT))
1313 putback_lru_pages(&pagelist);
1315 up_write(&mm->mmap_sem);
1322 * User space interface with variable sized bitmaps for nodelists.
1325 /* Copy a node mask from user space. */
1326 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1327 unsigned long maxnode)
1330 unsigned long nlongs;
1331 unsigned long endmask;
1334 nodes_clear(*nodes);
1335 if (maxnode == 0 || !nmask)
1337 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1340 nlongs = BITS_TO_LONGS(maxnode);
1341 if ((maxnode % BITS_PER_LONG) == 0)
1344 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1346 /* When the user specified more nodes than supported just check
1347 if the non supported part is all zero. */
1348 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1349 if (nlongs > PAGE_SIZE/sizeof(long))
1351 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1353 if (get_user(t, nmask + k))
1355 if (k == nlongs - 1) {
1361 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1365 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1367 nodes_addr(*nodes)[nlongs-1] &= endmask;
1371 /* Copy a kernel node mask to user space */
1372 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1375 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1376 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1378 if (copy > nbytes) {
1379 if (copy > PAGE_SIZE)
1381 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1385 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1388 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1389 unsigned long, mode, unsigned long __user *, nmask,
1390 unsigned long, maxnode, unsigned, flags)
1394 unsigned short mode_flags;
1396 mode_flags = mode & MPOL_MODE_FLAGS;
1397 mode &= ~MPOL_MODE_FLAGS;
1398 if (mode >= MPOL_MAX)
1400 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1401 (mode_flags & MPOL_F_RELATIVE_NODES))
1403 err = get_nodes(&nodes, nmask, maxnode);
1406 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1409 /* Set the process memory policy */
1410 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1411 unsigned long, maxnode)
1415 unsigned short flags;
1417 flags = mode & MPOL_MODE_FLAGS;
1418 mode &= ~MPOL_MODE_FLAGS;
1419 if ((unsigned int)mode >= MPOL_MAX)
1421 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1423 err = get_nodes(&nodes, nmask, maxnode);
1426 return do_set_mempolicy(mode, flags, &nodes);
1429 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1430 const unsigned long __user *, old_nodes,
1431 const unsigned long __user *, new_nodes)
1433 const struct cred *cred = current_cred(), *tcred;
1434 struct mm_struct *mm = NULL;
1435 struct task_struct *task;
1436 nodemask_t task_nodes;
1440 NODEMASK_SCRATCH(scratch);
1445 old = &scratch->mask1;
1446 new = &scratch->mask2;
1448 err = get_nodes(old, old_nodes, maxnode);
1452 err = get_nodes(new, new_nodes, maxnode);
1456 /* Find the mm_struct */
1458 task = pid ? find_task_by_vpid(pid) : current;
1464 get_task_struct(task);
1469 * Check if this process has the right to modify the specified
1470 * process. The right exists if the process has administrative
1471 * capabilities, superuser privileges or the same
1472 * userid as the target process.
1474 tcred = __task_cred(task);
1475 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1476 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1477 !capable(CAP_SYS_NICE)) {
1484 task_nodes = cpuset_mems_allowed(task);
1485 /* Is the user allowed to access the target nodes? */
1486 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1491 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1496 err = security_task_movememory(task);
1500 mm = get_task_mm(task);
1501 put_task_struct(task);
1508 err = do_migrate_pages(mm, old, new,
1509 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1513 NODEMASK_SCRATCH_FREE(scratch);
1518 put_task_struct(task);
1524 /* Retrieve NUMA policy */
1525 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1526 unsigned long __user *, nmask, unsigned long, maxnode,
1527 unsigned long, addr, unsigned long, flags)
1530 int uninitialized_var(pval);
1533 if (nmask != NULL && maxnode < MAX_NUMNODES)
1536 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1541 if (policy && put_user(pval, policy))
1545 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1550 #ifdef CONFIG_COMPAT
1552 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1553 compat_ulong_t __user *nmask,
1554 compat_ulong_t maxnode,
1555 compat_ulong_t addr, compat_ulong_t flags)
1558 unsigned long __user *nm = NULL;
1559 unsigned long nr_bits, alloc_size;
1560 DECLARE_BITMAP(bm, MAX_NUMNODES);
1562 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1563 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1566 nm = compat_alloc_user_space(alloc_size);
1568 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1570 if (!err && nmask) {
1571 unsigned long copy_size;
1572 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1573 err = copy_from_user(bm, nm, copy_size);
1574 /* ensure entire bitmap is zeroed */
1575 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1576 err |= compat_put_bitmap(nmask, bm, nr_bits);
1582 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1583 compat_ulong_t maxnode)
1586 unsigned long __user *nm = NULL;
1587 unsigned long nr_bits, alloc_size;
1588 DECLARE_BITMAP(bm, MAX_NUMNODES);
1590 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1591 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1594 err = compat_get_bitmap(bm, nmask, nr_bits);
1595 nm = compat_alloc_user_space(alloc_size);
1596 err |= copy_to_user(nm, bm, alloc_size);
1602 return sys_set_mempolicy(mode, nm, nr_bits+1);
1605 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1606 compat_ulong_t mode, compat_ulong_t __user *nmask,
1607 compat_ulong_t maxnode, compat_ulong_t flags)
1610 unsigned long __user *nm = NULL;
1611 unsigned long nr_bits, alloc_size;
1614 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1615 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1618 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1619 nm = compat_alloc_user_space(alloc_size);
1620 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1626 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1632 * get_vma_policy(@task, @vma, @addr)
1633 * @task - task for fallback if vma policy == default
1634 * @vma - virtual memory area whose policy is sought
1635 * @addr - address in @vma for shared policy lookup
1637 * Returns effective policy for a VMA at specified address.
1638 * Falls back to @task or system default policy, as necessary.
1639 * Current or other task's task mempolicy and non-shared vma policies must be
1640 * protected by task_lock(task) by the caller.
1641 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1642 * count--added by the get_policy() vm_op, as appropriate--to protect against
1643 * freeing by another task. It is the caller's responsibility to free the
1644 * extra reference for shared policies.
1646 struct mempolicy *get_vma_policy(struct task_struct *task,
1647 struct vm_area_struct *vma, unsigned long addr)
1649 struct mempolicy *pol = get_task_policy(task);
1652 if (vma->vm_ops && vma->vm_ops->get_policy) {
1653 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1657 } else if (vma->vm_policy) {
1658 pol = vma->vm_policy;
1661 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1662 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1663 * count on these policies which will be dropped by
1664 * mpol_cond_put() later
1666 if (mpol_needs_cond_ref(pol))
1671 pol = &default_policy;
1675 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1677 enum zone_type dynamic_policy_zone = policy_zone;
1679 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1682 * if policy->v.nodes has movable memory only,
1683 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1685 * policy->v.nodes is intersect with node_states[N_MEMORY].
1686 * so if the following test faile, it implies
1687 * policy->v.nodes has movable memory only.
1689 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1690 dynamic_policy_zone = ZONE_MOVABLE;
1692 return zone >= dynamic_policy_zone;
1696 * Return a nodemask representing a mempolicy for filtering nodes for
1699 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1701 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1702 if (unlikely(policy->mode == MPOL_BIND) &&
1703 apply_policy_zone(policy, gfp_zone(gfp)) &&
1704 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1705 return &policy->v.nodes;
1710 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1711 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1714 switch (policy->mode) {
1715 case MPOL_PREFERRED:
1716 if (!(policy->flags & MPOL_F_LOCAL))
1717 nd = policy->v.preferred_node;
1721 * Normally, MPOL_BIND allocations are node-local within the
1722 * allowed nodemask. However, if __GFP_THISNODE is set and the
1723 * current node isn't part of the mask, we use the zonelist for
1724 * the first node in the mask instead.
1726 if (unlikely(gfp & __GFP_THISNODE) &&
1727 unlikely(!node_isset(nd, policy->v.nodes)))
1728 nd = first_node(policy->v.nodes);
1733 return node_zonelist(nd, gfp);
1736 /* Do dynamic interleaving for a process */
1737 static unsigned interleave_nodes(struct mempolicy *policy)
1740 struct task_struct *me = current;
1743 next = next_node(nid, policy->v.nodes);
1744 if (next >= MAX_NUMNODES)
1745 next = first_node(policy->v.nodes);
1746 if (next < MAX_NUMNODES)
1752 * Depending on the memory policy provide a node from which to allocate the
1754 * @policy must be protected by freeing by the caller. If @policy is
1755 * the current task's mempolicy, this protection is implicit, as only the
1756 * task can change it's policy. The system default policy requires no
1759 unsigned slab_node(void)
1761 struct mempolicy *policy;
1764 return numa_node_id();
1766 policy = current->mempolicy;
1767 if (!policy || policy->flags & MPOL_F_LOCAL)
1768 return numa_node_id();
1770 switch (policy->mode) {
1771 case MPOL_PREFERRED:
1773 * handled MPOL_F_LOCAL above
1775 return policy->v.preferred_node;
1777 case MPOL_INTERLEAVE:
1778 return interleave_nodes(policy);
1782 * Follow bind policy behavior and start allocation at the
1785 struct zonelist *zonelist;
1787 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1788 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1789 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1792 return zone ? zone->node : numa_node_id();
1800 /* Do static interleaving for a VMA with known offset. */
1801 static unsigned offset_il_node(struct mempolicy *pol,
1802 struct vm_area_struct *vma, unsigned long off)
1804 unsigned nnodes = nodes_weight(pol->v.nodes);
1810 return numa_node_id();
1811 target = (unsigned int)off % nnodes;
1814 nid = next_node(nid, pol->v.nodes);
1816 } while (c <= target);
1820 /* Determine a node number for interleave */
1821 static inline unsigned interleave_nid(struct mempolicy *pol,
1822 struct vm_area_struct *vma, unsigned long addr, int shift)
1828 * for small pages, there is no difference between
1829 * shift and PAGE_SHIFT, so the bit-shift is safe.
1830 * for huge pages, since vm_pgoff is in units of small
1831 * pages, we need to shift off the always 0 bits to get
1834 BUG_ON(shift < PAGE_SHIFT);
1835 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1836 off += (addr - vma->vm_start) >> shift;
1837 return offset_il_node(pol, vma, off);
1839 return interleave_nodes(pol);
1843 * Return the bit number of a random bit set in the nodemask.
1844 * (returns -1 if nodemask is empty)
1846 int node_random(const nodemask_t *maskp)
1850 w = nodes_weight(*maskp);
1852 bit = bitmap_ord_to_pos(maskp->bits,
1853 get_random_int() % w, MAX_NUMNODES);
1857 #ifdef CONFIG_HUGETLBFS
1859 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1860 * @vma = virtual memory area whose policy is sought
1861 * @addr = address in @vma for shared policy lookup and interleave policy
1862 * @gfp_flags = for requested zone
1863 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1864 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1866 * Returns a zonelist suitable for a huge page allocation and a pointer
1867 * to the struct mempolicy for conditional unref after allocation.
1868 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1869 * @nodemask for filtering the zonelist.
1871 * Must be protected by get_mems_allowed()
1873 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1874 gfp_t gfp_flags, struct mempolicy **mpol,
1875 nodemask_t **nodemask)
1877 struct zonelist *zl;
1879 *mpol = get_vma_policy(current, vma, addr);
1880 *nodemask = NULL; /* assume !MPOL_BIND */
1882 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1883 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1884 huge_page_shift(hstate_vma(vma))), gfp_flags);
1886 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1887 if ((*mpol)->mode == MPOL_BIND)
1888 *nodemask = &(*mpol)->v.nodes;
1894 * init_nodemask_of_mempolicy
1896 * If the current task's mempolicy is "default" [NULL], return 'false'
1897 * to indicate default policy. Otherwise, extract the policy nodemask
1898 * for 'bind' or 'interleave' policy into the argument nodemask, or
1899 * initialize the argument nodemask to contain the single node for
1900 * 'preferred' or 'local' policy and return 'true' to indicate presence
1901 * of non-default mempolicy.
1903 * We don't bother with reference counting the mempolicy [mpol_get/put]
1904 * because the current task is examining it's own mempolicy and a task's
1905 * mempolicy is only ever changed by the task itself.
1907 * N.B., it is the caller's responsibility to free a returned nodemask.
1909 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1911 struct mempolicy *mempolicy;
1914 if (!(mask && current->mempolicy))
1918 mempolicy = current->mempolicy;
1919 switch (mempolicy->mode) {
1920 case MPOL_PREFERRED:
1921 if (mempolicy->flags & MPOL_F_LOCAL)
1922 nid = numa_node_id();
1924 nid = mempolicy->v.preferred_node;
1925 init_nodemask_of_node(mask, nid);
1930 case MPOL_INTERLEAVE:
1931 *mask = mempolicy->v.nodes;
1937 task_unlock(current);
1944 * mempolicy_nodemask_intersects
1946 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1947 * policy. Otherwise, check for intersection between mask and the policy
1948 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1949 * policy, always return true since it may allocate elsewhere on fallback.
1951 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1953 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1954 const nodemask_t *mask)
1956 struct mempolicy *mempolicy;
1962 mempolicy = tsk->mempolicy;
1966 switch (mempolicy->mode) {
1967 case MPOL_PREFERRED:
1969 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1970 * allocate from, they may fallback to other nodes when oom.
1971 * Thus, it's possible for tsk to have allocated memory from
1976 case MPOL_INTERLEAVE:
1977 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1987 /* Allocate a page in interleaved policy.
1988 Own path because it needs to do special accounting. */
1989 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1992 struct zonelist *zl;
1995 zl = node_zonelist(nid, gfp);
1996 page = __alloc_pages(gfp, order, zl);
1997 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1998 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2003 * alloc_pages_vma - Allocate a page for a VMA.
2006 * %GFP_USER user allocation.
2007 * %GFP_KERNEL kernel allocations,
2008 * %GFP_HIGHMEM highmem/user allocations,
2009 * %GFP_FS allocation should not call back into a file system.
2010 * %GFP_ATOMIC don't sleep.
2012 * @order:Order of the GFP allocation.
2013 * @vma: Pointer to VMA or NULL if not available.
2014 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2016 * This function allocates a page from the kernel page pool and applies
2017 * a NUMA policy associated with the VMA or the current process.
2018 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2019 * mm_struct of the VMA to prevent it from going away. Should be used for
2020 * all allocations for pages that will be mapped into
2021 * user space. Returns NULL when no page can be allocated.
2023 * Should be called with the mm_sem of the vma hold.
2026 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2027 unsigned long addr, int node)
2029 struct mempolicy *pol;
2031 unsigned int cpuset_mems_cookie;
2034 pol = get_vma_policy(current, vma, addr);
2035 cpuset_mems_cookie = get_mems_allowed();
2037 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2040 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2042 page = alloc_page_interleave(gfp, order, nid);
2043 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2048 page = __alloc_pages_nodemask(gfp, order,
2049 policy_zonelist(gfp, pol, node),
2050 policy_nodemask(gfp, pol));
2051 if (unlikely(mpol_needs_cond_ref(pol)))
2053 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2059 * alloc_pages_current - Allocate pages.
2062 * %GFP_USER user allocation,
2063 * %GFP_KERNEL kernel allocation,
2064 * %GFP_HIGHMEM highmem allocation,
2065 * %GFP_FS don't call back into a file system.
2066 * %GFP_ATOMIC don't sleep.
2067 * @order: Power of two of allocation size in pages. 0 is a single page.
2069 * Allocate a page from the kernel page pool. When not in
2070 * interrupt context and apply the current process NUMA policy.
2071 * Returns NULL when no page can be allocated.
2073 * Don't call cpuset_update_task_memory_state() unless
2074 * 1) it's ok to take cpuset_sem (can WAIT), and
2075 * 2) allocating for current task (not interrupt).
2077 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2079 struct mempolicy *pol = get_task_policy(current);
2081 unsigned int cpuset_mems_cookie;
2083 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2084 pol = &default_policy;
2087 cpuset_mems_cookie = get_mems_allowed();
2090 * No reference counting needed for current->mempolicy
2091 * nor system default_policy
2093 if (pol->mode == MPOL_INTERLEAVE)
2094 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2096 page = __alloc_pages_nodemask(gfp, order,
2097 policy_zonelist(gfp, pol, numa_node_id()),
2098 policy_nodemask(gfp, pol));
2100 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2105 EXPORT_SYMBOL(alloc_pages_current);
2107 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2109 struct mempolicy *pol = mpol_dup(vma_policy(src));
2112 return PTR_ERR(pol);
2113 dst->vm_policy = pol;
2118 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2119 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2120 * with the mems_allowed returned by cpuset_mems_allowed(). This
2121 * keeps mempolicies cpuset relative after its cpuset moves. See
2122 * further kernel/cpuset.c update_nodemask().
2124 * current's mempolicy may be rebinded by the other task(the task that changes
2125 * cpuset's mems), so we needn't do rebind work for current task.
2128 /* Slow path of a mempolicy duplicate */
2129 struct mempolicy *__mpol_dup(struct mempolicy *old)
2131 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2134 return ERR_PTR(-ENOMEM);
2136 /* task's mempolicy is protected by alloc_lock */
2137 if (old == current->mempolicy) {
2140 task_unlock(current);
2145 if (current_cpuset_is_being_rebound()) {
2146 nodemask_t mems = cpuset_mems_allowed(current);
2147 if (new->flags & MPOL_F_REBINDING)
2148 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2150 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2153 atomic_set(&new->refcnt, 1);
2157 /* Slow path of a mempolicy comparison */
2158 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2162 if (a->mode != b->mode)
2164 if (a->flags != b->flags)
2166 if (mpol_store_user_nodemask(a))
2167 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2173 case MPOL_INTERLEAVE:
2174 return !!nodes_equal(a->v.nodes, b->v.nodes);
2175 case MPOL_PREFERRED:
2176 return a->v.preferred_node == b->v.preferred_node;
2184 * Shared memory backing store policy support.
2186 * Remember policies even when nobody has shared memory mapped.
2187 * The policies are kept in Red-Black tree linked from the inode.
2188 * They are protected by the sp->lock spinlock, which should be held
2189 * for any accesses to the tree.
2192 /* lookup first element intersecting start-end */
2193 /* Caller holds sp->lock */
2194 static struct sp_node *
2195 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2197 struct rb_node *n = sp->root.rb_node;
2200 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2202 if (start >= p->end)
2204 else if (end <= p->start)
2212 struct sp_node *w = NULL;
2213 struct rb_node *prev = rb_prev(n);
2216 w = rb_entry(prev, struct sp_node, nd);
2217 if (w->end <= start)
2221 return rb_entry(n, struct sp_node, nd);
2224 /* Insert a new shared policy into the list. */
2225 /* Caller holds sp->lock */
2226 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2228 struct rb_node **p = &sp->root.rb_node;
2229 struct rb_node *parent = NULL;
2234 nd = rb_entry(parent, struct sp_node, nd);
2235 if (new->start < nd->start)
2237 else if (new->end > nd->end)
2238 p = &(*p)->rb_right;
2242 rb_link_node(&new->nd, parent, p);
2243 rb_insert_color(&new->nd, &sp->root);
2244 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2245 new->policy ? new->policy->mode : 0);
2248 /* Find shared policy intersecting idx */
2250 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2252 struct mempolicy *pol = NULL;
2255 if (!sp->root.rb_node)
2257 spin_lock(&sp->lock);
2258 sn = sp_lookup(sp, idx, idx+1);
2260 mpol_get(sn->policy);
2263 spin_unlock(&sp->lock);
2267 static void sp_free(struct sp_node *n)
2269 mpol_put(n->policy);
2270 kmem_cache_free(sn_cache, n);
2274 * mpol_misplaced - check whether current page node is valid in policy
2276 * @page - page to be checked
2277 * @vma - vm area where page mapped
2278 * @addr - virtual address where page mapped
2280 * Lookup current policy node id for vma,addr and "compare to" page's
2284 * -1 - not misplaced, page is in the right node
2285 * node - node id where the page should be
2287 * Policy determination "mimics" alloc_page_vma().
2288 * Called from fault path where we know the vma and faulting address.
2290 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2292 struct mempolicy *pol;
2294 int curnid = page_to_nid(page);
2295 unsigned long pgoff;
2301 pol = get_vma_policy(current, vma, addr);
2302 if (!(pol->flags & MPOL_F_MOF))
2305 switch (pol->mode) {
2306 case MPOL_INTERLEAVE:
2307 BUG_ON(addr >= vma->vm_end);
2308 BUG_ON(addr < vma->vm_start);
2310 pgoff = vma->vm_pgoff;
2311 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2312 polnid = offset_il_node(pol, vma, pgoff);
2315 case MPOL_PREFERRED:
2316 if (pol->flags & MPOL_F_LOCAL)
2317 polnid = numa_node_id();
2319 polnid = pol->v.preferred_node;
2324 * allows binding to multiple nodes.
2325 * use current page if in policy nodemask,
2326 * else select nearest allowed node, if any.
2327 * If no allowed nodes, use current [!misplaced].
2329 if (node_isset(curnid, pol->v.nodes))
2331 (void)first_zones_zonelist(
2332 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2333 gfp_zone(GFP_HIGHUSER),
2334 &pol->v.nodes, &zone);
2335 polnid = zone->node;
2342 /* Migrate the page towards the node whose CPU is referencing it */
2343 if (pol->flags & MPOL_F_MORON) {
2346 polnid = numa_node_id();
2349 * Multi-stage node selection is used in conjunction
2350 * with a periodic migration fault to build a temporal
2351 * task<->page relation. By using a two-stage filter we
2352 * remove short/unlikely relations.
2354 * Using P(p) ~ n_p / n_t as per frequentist
2355 * probability, we can equate a task's usage of a
2356 * particular page (n_p) per total usage of this
2357 * page (n_t) (in a given time-span) to a probability.
2359 * Our periodic faults will sample this probability and
2360 * getting the same result twice in a row, given these
2361 * samples are fully independent, is then given by
2362 * P(n)^2, provided our sample period is sufficiently
2363 * short compared to the usage pattern.
2365 * This quadric squishes small probabilities, making
2366 * it less likely we act on an unlikely task<->page
2369 last_nid = page_nid_xchg_last(page, polnid);
2370 if (last_nid != polnid)
2374 if (curnid != polnid)
2382 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2384 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2385 rb_erase(&n->nd, &sp->root);
2389 static void sp_node_init(struct sp_node *node, unsigned long start,
2390 unsigned long end, struct mempolicy *pol)
2392 node->start = start;
2397 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2398 struct mempolicy *pol)
2401 struct mempolicy *newpol;
2403 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2407 newpol = mpol_dup(pol);
2408 if (IS_ERR(newpol)) {
2409 kmem_cache_free(sn_cache, n);
2412 newpol->flags |= MPOL_F_SHARED;
2413 sp_node_init(n, start, end, newpol);
2418 /* Replace a policy range. */
2419 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2420 unsigned long end, struct sp_node *new)
2423 struct sp_node *n_new = NULL;
2424 struct mempolicy *mpol_new = NULL;
2428 spin_lock(&sp->lock);
2429 n = sp_lookup(sp, start, end);
2430 /* Take care of old policies in the same range. */
2431 while (n && n->start < end) {
2432 struct rb_node *next = rb_next(&n->nd);
2433 if (n->start >= start) {
2439 /* Old policy spanning whole new range. */
2444 *mpol_new = *n->policy;
2445 atomic_set(&mpol_new->refcnt, 1);
2446 sp_node_init(n_new, end, n->end, mpol_new);
2448 sp_insert(sp, n_new);
2457 n = rb_entry(next, struct sp_node, nd);
2461 spin_unlock(&sp->lock);
2468 kmem_cache_free(sn_cache, n_new);
2473 spin_unlock(&sp->lock);
2475 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2478 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2485 * mpol_shared_policy_init - initialize shared policy for inode
2486 * @sp: pointer to inode shared policy
2487 * @mpol: struct mempolicy to install
2489 * Install non-NULL @mpol in inode's shared policy rb-tree.
2490 * On entry, the current task has a reference on a non-NULL @mpol.
2491 * This must be released on exit.
2492 * This is called at get_inode() calls and we can use GFP_KERNEL.
2494 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2498 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2499 spin_lock_init(&sp->lock);
2502 struct vm_area_struct pvma;
2503 struct mempolicy *new;
2504 NODEMASK_SCRATCH(scratch);
2508 /* contextualize the tmpfs mount point mempolicy */
2509 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2511 goto free_scratch; /* no valid nodemask intersection */
2514 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2515 task_unlock(current);
2519 /* Create pseudo-vma that contains just the policy */
2520 memset(&pvma, 0, sizeof(struct vm_area_struct));
2521 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2522 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2525 mpol_put(new); /* drop initial ref */
2527 NODEMASK_SCRATCH_FREE(scratch);
2529 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2533 int mpol_set_shared_policy(struct shared_policy *info,
2534 struct vm_area_struct *vma, struct mempolicy *npol)
2537 struct sp_node *new = NULL;
2538 unsigned long sz = vma_pages(vma);
2540 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2542 sz, npol ? npol->mode : -1,
2543 npol ? npol->flags : -1,
2544 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2547 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2551 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2557 /* Free a backing policy store on inode delete. */
2558 void mpol_free_shared_policy(struct shared_policy *p)
2561 struct rb_node *next;
2563 if (!p->root.rb_node)
2565 spin_lock(&p->lock);
2566 next = rb_first(&p->root);
2568 n = rb_entry(next, struct sp_node, nd);
2569 next = rb_next(&n->nd);
2572 spin_unlock(&p->lock);
2575 #ifdef CONFIG_NUMA_BALANCING
2576 static bool __initdata numabalancing_override;
2578 static void __init check_numabalancing_enable(void)
2580 bool numabalancing_default = false;
2582 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2583 numabalancing_default = true;
2585 if (nr_node_ids > 1 && !numabalancing_override) {
2586 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2587 "Configure with numa_balancing= or sysctl");
2588 set_numabalancing_state(numabalancing_default);
2592 static int __init setup_numabalancing(char *str)
2597 numabalancing_override = true;
2599 if (!strcmp(str, "enable")) {
2600 set_numabalancing_state(true);
2602 } else if (!strcmp(str, "disable")) {
2603 set_numabalancing_state(false);
2608 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2612 __setup("numa_balancing=", setup_numabalancing);
2614 static inline void __init check_numabalancing_enable(void)
2617 #endif /* CONFIG_NUMA_BALANCING */
2619 /* assumes fs == KERNEL_DS */
2620 void __init numa_policy_init(void)
2622 nodemask_t interleave_nodes;
2623 unsigned long largest = 0;
2624 int nid, prefer = 0;
2626 policy_cache = kmem_cache_create("numa_policy",
2627 sizeof(struct mempolicy),
2628 0, SLAB_PANIC, NULL);
2630 sn_cache = kmem_cache_create("shared_policy_node",
2631 sizeof(struct sp_node),
2632 0, SLAB_PANIC, NULL);
2634 for_each_node(nid) {
2635 preferred_node_policy[nid] = (struct mempolicy) {
2636 .refcnt = ATOMIC_INIT(1),
2637 .mode = MPOL_PREFERRED,
2638 .flags = MPOL_F_MOF | MPOL_F_MORON,
2639 .v = { .preferred_node = nid, },
2644 * Set interleaving policy for system init. Interleaving is only
2645 * enabled across suitably sized nodes (default is >= 16MB), or
2646 * fall back to the largest node if they're all smaller.
2648 nodes_clear(interleave_nodes);
2649 for_each_node_state(nid, N_MEMORY) {
2650 unsigned long total_pages = node_present_pages(nid);
2652 /* Preserve the largest node */
2653 if (largest < total_pages) {
2654 largest = total_pages;
2658 /* Interleave this node? */
2659 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2660 node_set(nid, interleave_nodes);
2663 /* All too small, use the largest */
2664 if (unlikely(nodes_empty(interleave_nodes)))
2665 node_set(prefer, interleave_nodes);
2667 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2668 printk("numa_policy_init: interleaving failed\n");
2670 check_numabalancing_enable();
2673 /* Reset policy of current process to default */
2674 void numa_default_policy(void)
2676 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2680 * Parse and format mempolicy from/to strings
2684 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2686 static const char * const policy_modes[] =
2688 [MPOL_DEFAULT] = "default",
2689 [MPOL_PREFERRED] = "prefer",
2690 [MPOL_BIND] = "bind",
2691 [MPOL_INTERLEAVE] = "interleave",
2692 [MPOL_LOCAL] = "local",
2698 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2699 * @str: string containing mempolicy to parse
2700 * @mpol: pointer to struct mempolicy pointer, returned on success.
2703 * <mode>[=<flags>][:<nodelist>]
2705 * On success, returns 0, else 1
2707 int mpol_parse_str(char *str, struct mempolicy **mpol)
2709 struct mempolicy *new = NULL;
2710 unsigned short mode;
2711 unsigned short mode_flags;
2713 char *nodelist = strchr(str, ':');
2714 char *flags = strchr(str, '=');
2718 /* NUL-terminate mode or flags string */
2720 if (nodelist_parse(nodelist, nodes))
2722 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2728 *flags++ = '\0'; /* terminate mode string */
2730 for (mode = 0; mode < MPOL_MAX; mode++) {
2731 if (!strcmp(str, policy_modes[mode])) {
2735 if (mode >= MPOL_MAX)
2739 case MPOL_PREFERRED:
2741 * Insist on a nodelist of one node only
2744 char *rest = nodelist;
2745 while (isdigit(*rest))
2751 case MPOL_INTERLEAVE:
2753 * Default to online nodes with memory if no nodelist
2756 nodes = node_states[N_MEMORY];
2760 * Don't allow a nodelist; mpol_new() checks flags
2764 mode = MPOL_PREFERRED;
2768 * Insist on a empty nodelist
2775 * Insist on a nodelist
2784 * Currently, we only support two mutually exclusive
2787 if (!strcmp(flags, "static"))
2788 mode_flags |= MPOL_F_STATIC_NODES;
2789 else if (!strcmp(flags, "relative"))
2790 mode_flags |= MPOL_F_RELATIVE_NODES;
2795 new = mpol_new(mode, mode_flags, &nodes);
2800 * Save nodes for mpol_to_str() to show the tmpfs mount options
2801 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2803 if (mode != MPOL_PREFERRED)
2804 new->v.nodes = nodes;
2806 new->v.preferred_node = first_node(nodes);
2808 new->flags |= MPOL_F_LOCAL;
2811 * Save nodes for contextualization: this will be used to "clone"
2812 * the mempolicy in a specific context [cpuset] at a later time.
2814 new->w.user_nodemask = nodes;
2819 /* Restore string for error message */
2828 #endif /* CONFIG_TMPFS */
2831 * mpol_to_str - format a mempolicy structure for printing
2832 * @buffer: to contain formatted mempolicy string
2833 * @maxlen: length of @buffer
2834 * @pol: pointer to mempolicy to be formatted
2836 * Convert a mempolicy into a string.
2837 * Returns the number of characters in buffer (if positive)
2838 * or an error (negative)
2840 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2845 unsigned short mode;
2846 unsigned short flags = pol ? pol->flags : 0;
2849 * Sanity check: room for longest mode, flag and some nodes
2851 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2853 if (!pol || pol == &default_policy)
2854 mode = MPOL_DEFAULT;
2863 case MPOL_PREFERRED:
2865 if (flags & MPOL_F_LOCAL)
2868 node_set(pol->v.preferred_node, nodes);
2873 case MPOL_INTERLEAVE:
2874 nodes = pol->v.nodes;
2881 l = strlen(policy_modes[mode]);
2882 if (buffer + maxlen < p + l + 1)
2885 strcpy(p, policy_modes[mode]);
2888 if (flags & MPOL_MODE_FLAGS) {
2889 if (buffer + maxlen < p + 2)
2894 * Currently, the only defined flags are mutually exclusive
2896 if (flags & MPOL_F_STATIC_NODES)
2897 p += snprintf(p, buffer + maxlen - p, "static");
2898 else if (flags & MPOL_F_RELATIVE_NODES)
2899 p += snprintf(p, buffer + maxlen - p, "relative");
2902 if (!nodes_empty(nodes)) {
2903 if (buffer + maxlen < p + 2)
2906 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);