2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
93 #include <linux/mmu_notifier.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
97 #include <linux/random.h>
102 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
105 static struct kmem_cache *policy_cache;
106 static struct kmem_cache *sn_cache;
108 /* Highest zone. An specific allocation for a zone below that is not
110 enum zone_type policy_zone = 0;
113 * run-time system-wide default policy => local allocation
115 static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
121 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
123 static struct mempolicy *get_task_policy(struct task_struct *p)
125 struct mempolicy *pol = p->mempolicy;
128 int node = numa_node_id();
130 if (node != NUMA_NO_NODE) {
131 pol = &preferred_node_policy[node];
133 * preferred_node_policy is not initialised early in
144 static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * If we have a lock to protect task->mempolicy in read-side, we do
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
162 } mpol_ops[MPOL_MAX];
164 /* Check that the nodemask contains at least one populated zone */
165 static int is_valid_nodemask(const nodemask_t *nodemask)
167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
170 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
172 return pol->flags & MPOL_MODE_FLAGS;
175 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
183 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
185 if (nodes_empty(*nodes))
187 pol->v.nodes = *nodes;
191 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
198 pol->v.preferred_node = first_node(*nodes);
202 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
204 if (!is_valid_nodemask(nodes))
206 pol->v.nodes = *nodes;
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
219 static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
228 nodes_and(nsc->mask1,
229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
250 ret = mpol_ops[pol->mode].create(pol, NULL);
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
295 policy->flags = flags;
300 /* Slow path of a mpol destructor. */
301 void __mpol_put(struct mempolicy *p)
303 if (!atomic_dec_and_test(&p->refcnt))
305 kmem_cache_free(policy_cache, p);
308 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
319 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
344 if (nodes_empty(tmp))
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
363 static void mpol_rebind_preferred(struct mempolicy *pol,
364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
372 if (node_isset(node, *nodes)) {
373 pol->v.preferred_node = node;
374 pol->flags &= ~MPOL_F_LOCAL;
376 pol->flags |= MPOL_F_LOCAL;
377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
384 pol->w.cpuset_mems_allowed = *nodes;
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
396 * If we have a lock to protect task->mempolicy in read-side, we do
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
404 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
433 * Called with task's alloc_lock held.
436 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
439 mpol_rebind_policy(tsk->mempolicy, new, step);
443 * Rebind each vma in mm to new nodemask.
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
448 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
450 struct vm_area_struct *vma;
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
455 up_write(&mm->mmap_sem);
458 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
460 .rebind = mpol_rebind_default,
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
476 static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
483 static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
497 if (!pte_present(*pte))
499 page = vm_normal_page(vma, addr, *pte);
503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
506 if (PageReserved(page))
508 nid = page_to_nid(page);
509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
513 migrate_page_add(page, private, flags);
516 } while (pte++, addr += PAGE_SIZE, addr != end);
517 pte_unmap_unlock(orig_pte, ptl);
521 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
525 #ifdef CONFIG_HUGETLB_PAGE
529 spin_lock(&vma->vm_mm->page_table_lock);
530 page = pte_page(huge_ptep_get((pte_t *)pmd));
531 nid = page_to_nid(page);
532 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
534 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
535 if (flags & (MPOL_MF_MOVE_ALL) ||
536 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
537 isolate_huge_page(page, private);
539 spin_unlock(&vma->vm_mm->page_table_lock);
545 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
546 unsigned long addr, unsigned long end,
547 const nodemask_t *nodes, unsigned long flags,
553 pmd = pmd_offset(pud, addr);
555 next = pmd_addr_end(addr, end);
556 if (!pmd_present(*pmd))
558 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
559 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
563 split_huge_page_pmd(vma, addr, pmd);
564 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
566 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
569 } while (pmd++, addr = next, addr != end);
573 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
574 unsigned long addr, unsigned long end,
575 const nodemask_t *nodes, unsigned long flags,
581 pud = pud_offset(pgd, addr);
583 next = pud_addr_end(addr, end);
584 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
586 if (pud_none_or_clear_bad(pud))
588 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
591 } while (pud++, addr = next, addr != end);
595 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
596 unsigned long addr, unsigned long end,
597 const nodemask_t *nodes, unsigned long flags,
603 pgd = pgd_offset(vma->vm_mm, addr);
605 next = pgd_addr_end(addr, end);
606 if (pgd_none_or_clear_bad(pgd))
608 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
611 } while (pgd++, addr = next, addr != end);
615 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
617 * This is used to mark a range of virtual addresses to be inaccessible.
618 * These are later cleared by a NUMA hinting fault. Depending on these
619 * faults, pages may be migrated for better NUMA placement.
621 * This is assuming that NUMA faults are handled using PROT_NONE. If
622 * an architecture makes a different choice, it will need further
623 * changes to the core.
625 unsigned long change_prot_numa(struct vm_area_struct *vma,
626 unsigned long addr, unsigned long end)
629 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
631 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
633 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
638 static unsigned long change_prot_numa(struct vm_area_struct *vma,
639 unsigned long addr, unsigned long end)
643 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
646 * Walk through page tables and collect pages to be migrated.
648 * If pages found in a given range are on a set of nodes (determined by
649 * @nodes and @flags,) it's isolated and queued to the pagelist which is
650 * passed via @private.)
652 static struct vm_area_struct *
653 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
654 const nodemask_t *nodes, unsigned long flags, void *private)
657 struct vm_area_struct *first, *vma, *prev;
660 first = find_vma(mm, start);
662 return ERR_PTR(-EFAULT);
664 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
665 unsigned long endvma = vma->vm_end;
669 if (vma->vm_start > start)
670 start = vma->vm_start;
672 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
673 if (!vma->vm_next && vma->vm_end < end)
674 return ERR_PTR(-EFAULT);
675 if (prev && prev->vm_end < vma->vm_start)
676 return ERR_PTR(-EFAULT);
679 if (flags & MPOL_MF_LAZY) {
680 change_prot_numa(vma, start, endvma);
684 if ((flags & MPOL_MF_STRICT) ||
685 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
686 vma_migratable(vma))) {
688 err = queue_pages_pgd_range(vma, start, endvma, nodes,
691 first = ERR_PTR(err);
702 * Apply policy to a single VMA
703 * This must be called with the mmap_sem held for writing.
705 static int vma_replace_policy(struct vm_area_struct *vma,
706 struct mempolicy *pol)
709 struct mempolicy *old;
710 struct mempolicy *new;
712 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
713 vma->vm_start, vma->vm_end, vma->vm_pgoff,
714 vma->vm_ops, vma->vm_file,
715 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
721 if (vma->vm_ops && vma->vm_ops->set_policy) {
722 err = vma->vm_ops->set_policy(vma, new);
727 old = vma->vm_policy;
728 vma->vm_policy = new; /* protected by mmap_sem */
737 /* Step 2: apply policy to a range and do splits. */
738 static int mbind_range(struct mm_struct *mm, unsigned long start,
739 unsigned long end, struct mempolicy *new_pol)
741 struct vm_area_struct *next;
742 struct vm_area_struct *prev;
743 struct vm_area_struct *vma;
746 unsigned long vmstart;
749 vma = find_vma(mm, start);
750 if (!vma || vma->vm_start > start)
754 if (start > vma->vm_start)
757 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
759 vmstart = max(start, vma->vm_start);
760 vmend = min(end, vma->vm_end);
762 if (mpol_equal(vma_policy(vma), new_pol))
765 pgoff = vma->vm_pgoff +
766 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
767 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
768 vma->anon_vma, vma->vm_file, pgoff,
773 if (mpol_equal(vma_policy(vma), new_pol))
775 /* vma_merge() joined vma && vma->next, case 8 */
778 if (vma->vm_start != vmstart) {
779 err = split_vma(vma->vm_mm, vma, vmstart, 1);
783 if (vma->vm_end != vmend) {
784 err = split_vma(vma->vm_mm, vma, vmend, 0);
789 err = vma_replace_policy(vma, new_pol);
799 * Update task->flags PF_MEMPOLICY bit: set iff non-default
800 * mempolicy. Allows more rapid checking of this (combined perhaps
801 * with other PF_* flag bits) on memory allocation hot code paths.
803 * If called from outside this file, the task 'p' should -only- be
804 * a newly forked child not yet visible on the task list, because
805 * manipulating the task flags of a visible task is not safe.
807 * The above limitation is why this routine has the funny name
808 * mpol_fix_fork_child_flag().
810 * It is also safe to call this with a task pointer of current,
811 * which the static wrapper mpol_set_task_struct_flag() does,
812 * for use within this file.
815 void mpol_fix_fork_child_flag(struct task_struct *p)
818 p->flags |= PF_MEMPOLICY;
820 p->flags &= ~PF_MEMPOLICY;
823 static void mpol_set_task_struct_flag(void)
825 mpol_fix_fork_child_flag(current);
828 /* Set the process memory policy */
829 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
832 struct mempolicy *new, *old;
833 struct mm_struct *mm = current->mm;
834 NODEMASK_SCRATCH(scratch);
840 new = mpol_new(mode, flags, nodes);
846 * prevent changing our mempolicy while show_numa_maps()
848 * Note: do_set_mempolicy() can be called at init time
852 down_write(&mm->mmap_sem);
854 ret = mpol_set_nodemask(new, nodes, scratch);
856 task_unlock(current);
858 up_write(&mm->mmap_sem);
862 old = current->mempolicy;
863 current->mempolicy = new;
864 mpol_set_task_struct_flag();
865 if (new && new->mode == MPOL_INTERLEAVE &&
866 nodes_weight(new->v.nodes))
867 current->il_next = first_node(new->v.nodes);
868 task_unlock(current);
870 up_write(&mm->mmap_sem);
875 NODEMASK_SCRATCH_FREE(scratch);
880 * Return nodemask for policy for get_mempolicy() query
882 * Called with task's alloc_lock held
884 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
887 if (p == &default_policy)
893 case MPOL_INTERLEAVE:
897 if (!(p->flags & MPOL_F_LOCAL))
898 node_set(p->v.preferred_node, *nodes);
899 /* else return empty node mask for local allocation */
906 static int lookup_node(struct mm_struct *mm, unsigned long addr)
911 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
913 err = page_to_nid(p);
919 /* Retrieve NUMA policy */
920 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
921 unsigned long addr, unsigned long flags)
924 struct mm_struct *mm = current->mm;
925 struct vm_area_struct *vma = NULL;
926 struct mempolicy *pol = current->mempolicy;
929 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
932 if (flags & MPOL_F_MEMS_ALLOWED) {
933 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
935 *policy = 0; /* just so it's initialized */
937 *nmask = cpuset_current_mems_allowed;
938 task_unlock(current);
942 if (flags & MPOL_F_ADDR) {
944 * Do NOT fall back to task policy if the
945 * vma/shared policy at addr is NULL. We
946 * want to return MPOL_DEFAULT in this case.
948 down_read(&mm->mmap_sem);
949 vma = find_vma_intersection(mm, addr, addr+1);
951 up_read(&mm->mmap_sem);
954 if (vma->vm_ops && vma->vm_ops->get_policy)
955 pol = vma->vm_ops->get_policy(vma, addr);
957 pol = vma->vm_policy;
962 pol = &default_policy; /* indicates default behavior */
964 if (flags & MPOL_F_NODE) {
965 if (flags & MPOL_F_ADDR) {
966 err = lookup_node(mm, addr);
970 } else if (pol == current->mempolicy &&
971 pol->mode == MPOL_INTERLEAVE) {
972 *policy = current->il_next;
978 *policy = pol == &default_policy ? MPOL_DEFAULT :
981 * Internal mempolicy flags must be masked off before exposing
982 * the policy to userspace.
984 *policy |= (pol->flags & MPOL_MODE_FLAGS);
988 up_read(¤t->mm->mmap_sem);
994 if (mpol_store_user_nodemask(pol)) {
995 *nmask = pol->w.user_nodemask;
998 get_policy_nodemask(pol, nmask);
999 task_unlock(current);
1006 up_read(¤t->mm->mmap_sem);
1010 #ifdef CONFIG_MIGRATION
1014 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1015 unsigned long flags)
1018 * Avoid migrating a page that is shared with others.
1020 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1021 if (!isolate_lru_page(page)) {
1022 list_add_tail(&page->lru, pagelist);
1023 inc_zone_page_state(page, NR_ISOLATED_ANON +
1024 page_is_file_cache(page));
1029 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1032 return alloc_huge_page_node(page_hstate(compound_head(page)),
1035 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1039 * Migrate pages from one node to a target node.
1040 * Returns error or the number of pages not migrated.
1042 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1046 LIST_HEAD(pagelist);
1050 node_set(source, nmask);
1053 * This does not "check" the range but isolates all pages that
1054 * need migration. Between passing in the full user address
1055 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1057 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1058 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1059 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1061 if (!list_empty(&pagelist)) {
1062 err = migrate_pages(&pagelist, new_node_page, dest,
1063 MIGRATE_SYNC, MR_SYSCALL);
1065 putback_movable_pages(&pagelist);
1072 * Move pages between the two nodesets so as to preserve the physical
1073 * layout as much as possible.
1075 * Returns the number of page that could not be moved.
1077 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1078 const nodemask_t *to, int flags)
1084 err = migrate_prep();
1088 down_read(&mm->mmap_sem);
1090 err = migrate_vmas(mm, from, to, flags);
1095 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1096 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1097 * bit in 'tmp', and return that <source, dest> pair for migration.
1098 * The pair of nodemasks 'to' and 'from' define the map.
1100 * If no pair of bits is found that way, fallback to picking some
1101 * pair of 'source' and 'dest' bits that are not the same. If the
1102 * 'source' and 'dest' bits are the same, this represents a node
1103 * that will be migrating to itself, so no pages need move.
1105 * If no bits are left in 'tmp', or if all remaining bits left
1106 * in 'tmp' correspond to the same bit in 'to', return false
1107 * (nothing left to migrate).
1109 * This lets us pick a pair of nodes to migrate between, such that
1110 * if possible the dest node is not already occupied by some other
1111 * source node, minimizing the risk of overloading the memory on a
1112 * node that would happen if we migrated incoming memory to a node
1113 * before migrating outgoing memory source that same node.
1115 * A single scan of tmp is sufficient. As we go, we remember the
1116 * most recent <s, d> pair that moved (s != d). If we find a pair
1117 * that not only moved, but what's better, moved to an empty slot
1118 * (d is not set in tmp), then we break out then, with that pair.
1119 * Otherwise when we finish scanning from_tmp, we at least have the
1120 * most recent <s, d> pair that moved. If we get all the way through
1121 * the scan of tmp without finding any node that moved, much less
1122 * moved to an empty node, then there is nothing left worth migrating.
1126 while (!nodes_empty(tmp)) {
1131 for_each_node_mask(s, tmp) {
1134 * do_migrate_pages() tries to maintain the relative
1135 * node relationship of the pages established between
1136 * threads and memory areas.
1138 * However if the number of source nodes is not equal to
1139 * the number of destination nodes we can not preserve
1140 * this node relative relationship. In that case, skip
1141 * copying memory from a node that is in the destination
1144 * Example: [2,3,4] -> [3,4,5] moves everything.
1145 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1148 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1149 (node_isset(s, *to)))
1152 d = node_remap(s, *from, *to);
1156 source = s; /* Node moved. Memorize */
1159 /* dest not in remaining from nodes? */
1160 if (!node_isset(dest, tmp))
1166 node_clear(source, tmp);
1167 err = migrate_to_node(mm, source, dest, flags);
1174 up_read(&mm->mmap_sem);
1182 * Allocate a new page for page migration based on vma policy.
1183 * Start assuming that page is mapped by vma pointed to by @private.
1184 * Search forward from there, if not. N.B., this assumes that the
1185 * list of pages handed to migrate_pages()--which is how we get here--
1186 * is in virtual address order.
1188 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1190 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1191 unsigned long uninitialized_var(address);
1194 address = page_address_in_vma(page, vma);
1195 if (address != -EFAULT)
1201 return alloc_huge_page_noerr(vma, address, 1);
1203 * if !vma, alloc_page_vma() will use task or system default policy
1205 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1209 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1210 unsigned long flags)
1214 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1215 const nodemask_t *to, int flags)
1220 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1226 static long do_mbind(unsigned long start, unsigned long len,
1227 unsigned short mode, unsigned short mode_flags,
1228 nodemask_t *nmask, unsigned long flags)
1230 struct vm_area_struct *vma;
1231 struct mm_struct *mm = current->mm;
1232 struct mempolicy *new;
1235 LIST_HEAD(pagelist);
1237 if (flags & ~(unsigned long)MPOL_MF_VALID)
1239 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1242 if (start & ~PAGE_MASK)
1245 if (mode == MPOL_DEFAULT)
1246 flags &= ~MPOL_MF_STRICT;
1248 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1256 new = mpol_new(mode, mode_flags, nmask);
1258 return PTR_ERR(new);
1260 if (flags & MPOL_MF_LAZY)
1261 new->flags |= MPOL_F_MOF;
1264 * If we are using the default policy then operation
1265 * on discontinuous address spaces is okay after all
1268 flags |= MPOL_MF_DISCONTIG_OK;
1270 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1271 start, start + len, mode, mode_flags,
1272 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1274 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1276 err = migrate_prep();
1281 NODEMASK_SCRATCH(scratch);
1283 down_write(&mm->mmap_sem);
1285 err = mpol_set_nodemask(new, nmask, scratch);
1286 task_unlock(current);
1288 up_write(&mm->mmap_sem);
1291 NODEMASK_SCRATCH_FREE(scratch);
1296 vma = queue_pages_range(mm, start, end, nmask,
1297 flags | MPOL_MF_INVERT, &pagelist);
1299 err = PTR_ERR(vma); /* maybe ... */
1301 err = mbind_range(mm, start, end, new);
1306 if (!list_empty(&pagelist)) {
1307 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1308 nr_failed = migrate_pages(&pagelist, new_vma_page,
1310 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1312 putback_movable_pages(&pagelist);
1315 if (nr_failed && (flags & MPOL_MF_STRICT))
1318 putback_lru_pages(&pagelist);
1320 up_write(&mm->mmap_sem);
1327 * User space interface with variable sized bitmaps for nodelists.
1330 /* Copy a node mask from user space. */
1331 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1332 unsigned long maxnode)
1335 unsigned long nlongs;
1336 unsigned long endmask;
1339 nodes_clear(*nodes);
1340 if (maxnode == 0 || !nmask)
1342 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1345 nlongs = BITS_TO_LONGS(maxnode);
1346 if ((maxnode % BITS_PER_LONG) == 0)
1349 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1351 /* When the user specified more nodes than supported just check
1352 if the non supported part is all zero. */
1353 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1354 if (nlongs > PAGE_SIZE/sizeof(long))
1356 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1358 if (get_user(t, nmask + k))
1360 if (k == nlongs - 1) {
1366 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1370 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1372 nodes_addr(*nodes)[nlongs-1] &= endmask;
1376 /* Copy a kernel node mask to user space */
1377 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1380 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1381 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1383 if (copy > nbytes) {
1384 if (copy > PAGE_SIZE)
1386 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1390 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1393 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1394 unsigned long, mode, unsigned long __user *, nmask,
1395 unsigned long, maxnode, unsigned, flags)
1399 unsigned short mode_flags;
1401 mode_flags = mode & MPOL_MODE_FLAGS;
1402 mode &= ~MPOL_MODE_FLAGS;
1403 if (mode >= MPOL_MAX)
1405 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1406 (mode_flags & MPOL_F_RELATIVE_NODES))
1408 err = get_nodes(&nodes, nmask, maxnode);
1411 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1414 /* Set the process memory policy */
1415 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1416 unsigned long, maxnode)
1420 unsigned short flags;
1422 flags = mode & MPOL_MODE_FLAGS;
1423 mode &= ~MPOL_MODE_FLAGS;
1424 if ((unsigned int)mode >= MPOL_MAX)
1426 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1428 err = get_nodes(&nodes, nmask, maxnode);
1431 return do_set_mempolicy(mode, flags, &nodes);
1434 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1435 const unsigned long __user *, old_nodes,
1436 const unsigned long __user *, new_nodes)
1438 const struct cred *cred = current_cred(), *tcred;
1439 struct mm_struct *mm = NULL;
1440 struct task_struct *task;
1441 nodemask_t task_nodes;
1445 NODEMASK_SCRATCH(scratch);
1450 old = &scratch->mask1;
1451 new = &scratch->mask2;
1453 err = get_nodes(old, old_nodes, maxnode);
1457 err = get_nodes(new, new_nodes, maxnode);
1461 /* Find the mm_struct */
1463 task = pid ? find_task_by_vpid(pid) : current;
1469 get_task_struct(task);
1474 * Check if this process has the right to modify the specified
1475 * process. The right exists if the process has administrative
1476 * capabilities, superuser privileges or the same
1477 * userid as the target process.
1479 tcred = __task_cred(task);
1480 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1481 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1482 !capable(CAP_SYS_NICE)) {
1489 task_nodes = cpuset_mems_allowed(task);
1490 /* Is the user allowed to access the target nodes? */
1491 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1496 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1501 err = security_task_movememory(task);
1505 mm = get_task_mm(task);
1506 put_task_struct(task);
1513 err = do_migrate_pages(mm, old, new,
1514 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1518 NODEMASK_SCRATCH_FREE(scratch);
1523 put_task_struct(task);
1529 /* Retrieve NUMA policy */
1530 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1531 unsigned long __user *, nmask, unsigned long, maxnode,
1532 unsigned long, addr, unsigned long, flags)
1535 int uninitialized_var(pval);
1538 if (nmask != NULL && maxnode < MAX_NUMNODES)
1541 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1546 if (policy && put_user(pval, policy))
1550 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1555 #ifdef CONFIG_COMPAT
1557 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1558 compat_ulong_t __user *nmask,
1559 compat_ulong_t maxnode,
1560 compat_ulong_t addr, compat_ulong_t flags)
1563 unsigned long __user *nm = NULL;
1564 unsigned long nr_bits, alloc_size;
1565 DECLARE_BITMAP(bm, MAX_NUMNODES);
1567 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1568 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1571 nm = compat_alloc_user_space(alloc_size);
1573 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1575 if (!err && nmask) {
1576 unsigned long copy_size;
1577 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1578 err = copy_from_user(bm, nm, copy_size);
1579 /* ensure entire bitmap is zeroed */
1580 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1581 err |= compat_put_bitmap(nmask, bm, nr_bits);
1587 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1588 compat_ulong_t maxnode)
1591 unsigned long __user *nm = NULL;
1592 unsigned long nr_bits, alloc_size;
1593 DECLARE_BITMAP(bm, MAX_NUMNODES);
1595 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1596 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1599 err = compat_get_bitmap(bm, nmask, nr_bits);
1600 nm = compat_alloc_user_space(alloc_size);
1601 err |= copy_to_user(nm, bm, alloc_size);
1607 return sys_set_mempolicy(mode, nm, nr_bits+1);
1610 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1611 compat_ulong_t mode, compat_ulong_t __user *nmask,
1612 compat_ulong_t maxnode, compat_ulong_t flags)
1615 unsigned long __user *nm = NULL;
1616 unsigned long nr_bits, alloc_size;
1619 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1620 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1623 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1624 nm = compat_alloc_user_space(alloc_size);
1625 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1631 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1637 * get_vma_policy(@task, @vma, @addr)
1638 * @task - task for fallback if vma policy == default
1639 * @vma - virtual memory area whose policy is sought
1640 * @addr - address in @vma for shared policy lookup
1642 * Returns effective policy for a VMA at specified address.
1643 * Falls back to @task or system default policy, as necessary.
1644 * Current or other task's task mempolicy and non-shared vma policies must be
1645 * protected by task_lock(task) by the caller.
1646 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1647 * count--added by the get_policy() vm_op, as appropriate--to protect against
1648 * freeing by another task. It is the caller's responsibility to free the
1649 * extra reference for shared policies.
1651 struct mempolicy *get_vma_policy(struct task_struct *task,
1652 struct vm_area_struct *vma, unsigned long addr)
1654 struct mempolicy *pol = get_task_policy(task);
1657 if (vma->vm_ops && vma->vm_ops->get_policy) {
1658 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1662 } else if (vma->vm_policy) {
1663 pol = vma->vm_policy;
1666 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1667 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1668 * count on these policies which will be dropped by
1669 * mpol_cond_put() later
1671 if (mpol_needs_cond_ref(pol))
1676 pol = &default_policy;
1680 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1682 enum zone_type dynamic_policy_zone = policy_zone;
1684 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1687 * if policy->v.nodes has movable memory only,
1688 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1690 * policy->v.nodes is intersect with node_states[N_MEMORY].
1691 * so if the following test faile, it implies
1692 * policy->v.nodes has movable memory only.
1694 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1695 dynamic_policy_zone = ZONE_MOVABLE;
1697 return zone >= dynamic_policy_zone;
1701 * Return a nodemask representing a mempolicy for filtering nodes for
1704 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1706 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1707 if (unlikely(policy->mode == MPOL_BIND) &&
1708 apply_policy_zone(policy, gfp_zone(gfp)) &&
1709 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1710 return &policy->v.nodes;
1715 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1716 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1719 switch (policy->mode) {
1720 case MPOL_PREFERRED:
1721 if (!(policy->flags & MPOL_F_LOCAL))
1722 nd = policy->v.preferred_node;
1726 * Normally, MPOL_BIND allocations are node-local within the
1727 * allowed nodemask. However, if __GFP_THISNODE is set and the
1728 * current node isn't part of the mask, we use the zonelist for
1729 * the first node in the mask instead.
1731 if (unlikely(gfp & __GFP_THISNODE) &&
1732 unlikely(!node_isset(nd, policy->v.nodes)))
1733 nd = first_node(policy->v.nodes);
1738 return node_zonelist(nd, gfp);
1741 /* Do dynamic interleaving for a process */
1742 static unsigned interleave_nodes(struct mempolicy *policy)
1745 struct task_struct *me = current;
1748 next = next_node(nid, policy->v.nodes);
1749 if (next >= MAX_NUMNODES)
1750 next = first_node(policy->v.nodes);
1751 if (next < MAX_NUMNODES)
1757 * Depending on the memory policy provide a node from which to allocate the
1759 * @policy must be protected by freeing by the caller. If @policy is
1760 * the current task's mempolicy, this protection is implicit, as only the
1761 * task can change it's policy. The system default policy requires no
1764 unsigned slab_node(void)
1766 struct mempolicy *policy;
1769 return numa_node_id();
1771 policy = current->mempolicy;
1772 if (!policy || policy->flags & MPOL_F_LOCAL)
1773 return numa_node_id();
1775 switch (policy->mode) {
1776 case MPOL_PREFERRED:
1778 * handled MPOL_F_LOCAL above
1780 return policy->v.preferred_node;
1782 case MPOL_INTERLEAVE:
1783 return interleave_nodes(policy);
1787 * Follow bind policy behavior and start allocation at the
1790 struct zonelist *zonelist;
1792 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1793 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1794 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1797 return zone ? zone->node : numa_node_id();
1805 /* Do static interleaving for a VMA with known offset. */
1806 static unsigned offset_il_node(struct mempolicy *pol,
1807 struct vm_area_struct *vma, unsigned long off)
1809 unsigned nnodes = nodes_weight(pol->v.nodes);
1815 return numa_node_id();
1816 target = (unsigned int)off % nnodes;
1819 nid = next_node(nid, pol->v.nodes);
1821 } while (c <= target);
1825 /* Determine a node number for interleave */
1826 static inline unsigned interleave_nid(struct mempolicy *pol,
1827 struct vm_area_struct *vma, unsigned long addr, int shift)
1833 * for small pages, there is no difference between
1834 * shift and PAGE_SHIFT, so the bit-shift is safe.
1835 * for huge pages, since vm_pgoff is in units of small
1836 * pages, we need to shift off the always 0 bits to get
1839 BUG_ON(shift < PAGE_SHIFT);
1840 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1841 off += (addr - vma->vm_start) >> shift;
1842 return offset_il_node(pol, vma, off);
1844 return interleave_nodes(pol);
1848 * Return the bit number of a random bit set in the nodemask.
1849 * (returns -1 if nodemask is empty)
1851 int node_random(const nodemask_t *maskp)
1855 w = nodes_weight(*maskp);
1857 bit = bitmap_ord_to_pos(maskp->bits,
1858 get_random_int() % w, MAX_NUMNODES);
1862 #ifdef CONFIG_HUGETLBFS
1864 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1865 * @vma = virtual memory area whose policy is sought
1866 * @addr = address in @vma for shared policy lookup and interleave policy
1867 * @gfp_flags = for requested zone
1868 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1869 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1871 * Returns a zonelist suitable for a huge page allocation and a pointer
1872 * to the struct mempolicy for conditional unref after allocation.
1873 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1874 * @nodemask for filtering the zonelist.
1876 * Must be protected by get_mems_allowed()
1878 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1879 gfp_t gfp_flags, struct mempolicy **mpol,
1880 nodemask_t **nodemask)
1882 struct zonelist *zl;
1884 *mpol = get_vma_policy(current, vma, addr);
1885 *nodemask = NULL; /* assume !MPOL_BIND */
1887 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1888 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1889 huge_page_shift(hstate_vma(vma))), gfp_flags);
1891 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1892 if ((*mpol)->mode == MPOL_BIND)
1893 *nodemask = &(*mpol)->v.nodes;
1899 * init_nodemask_of_mempolicy
1901 * If the current task's mempolicy is "default" [NULL], return 'false'
1902 * to indicate default policy. Otherwise, extract the policy nodemask
1903 * for 'bind' or 'interleave' policy into the argument nodemask, or
1904 * initialize the argument nodemask to contain the single node for
1905 * 'preferred' or 'local' policy and return 'true' to indicate presence
1906 * of non-default mempolicy.
1908 * We don't bother with reference counting the mempolicy [mpol_get/put]
1909 * because the current task is examining it's own mempolicy and a task's
1910 * mempolicy is only ever changed by the task itself.
1912 * N.B., it is the caller's responsibility to free a returned nodemask.
1914 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1916 struct mempolicy *mempolicy;
1919 if (!(mask && current->mempolicy))
1923 mempolicy = current->mempolicy;
1924 switch (mempolicy->mode) {
1925 case MPOL_PREFERRED:
1926 if (mempolicy->flags & MPOL_F_LOCAL)
1927 nid = numa_node_id();
1929 nid = mempolicy->v.preferred_node;
1930 init_nodemask_of_node(mask, nid);
1935 case MPOL_INTERLEAVE:
1936 *mask = mempolicy->v.nodes;
1942 task_unlock(current);
1949 * mempolicy_nodemask_intersects
1951 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1952 * policy. Otherwise, check for intersection between mask and the policy
1953 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1954 * policy, always return true since it may allocate elsewhere on fallback.
1956 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1958 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1959 const nodemask_t *mask)
1961 struct mempolicy *mempolicy;
1967 mempolicy = tsk->mempolicy;
1971 switch (mempolicy->mode) {
1972 case MPOL_PREFERRED:
1974 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1975 * allocate from, they may fallback to other nodes when oom.
1976 * Thus, it's possible for tsk to have allocated memory from
1981 case MPOL_INTERLEAVE:
1982 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1992 /* Allocate a page in interleaved policy.
1993 Own path because it needs to do special accounting. */
1994 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1997 struct zonelist *zl;
2000 zl = node_zonelist(nid, gfp);
2001 page = __alloc_pages(gfp, order, zl);
2002 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
2003 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2008 * alloc_pages_vma - Allocate a page for a VMA.
2011 * %GFP_USER user allocation.
2012 * %GFP_KERNEL kernel allocations,
2013 * %GFP_HIGHMEM highmem/user allocations,
2014 * %GFP_FS allocation should not call back into a file system.
2015 * %GFP_ATOMIC don't sleep.
2017 * @order:Order of the GFP allocation.
2018 * @vma: Pointer to VMA or NULL if not available.
2019 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2021 * This function allocates a page from the kernel page pool and applies
2022 * a NUMA policy associated with the VMA or the current process.
2023 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2024 * mm_struct of the VMA to prevent it from going away. Should be used for
2025 * all allocations for pages that will be mapped into
2026 * user space. Returns NULL when no page can be allocated.
2028 * Should be called with the mm_sem of the vma hold.
2031 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2032 unsigned long addr, int node)
2034 struct mempolicy *pol;
2036 unsigned int cpuset_mems_cookie;
2039 pol = get_vma_policy(current, vma, addr);
2040 cpuset_mems_cookie = get_mems_allowed();
2042 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2045 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2047 page = alloc_page_interleave(gfp, order, nid);
2048 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2053 page = __alloc_pages_nodemask(gfp, order,
2054 policy_zonelist(gfp, pol, node),
2055 policy_nodemask(gfp, pol));
2056 if (unlikely(mpol_needs_cond_ref(pol)))
2058 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2064 * alloc_pages_current - Allocate pages.
2067 * %GFP_USER user allocation,
2068 * %GFP_KERNEL kernel allocation,
2069 * %GFP_HIGHMEM highmem allocation,
2070 * %GFP_FS don't call back into a file system.
2071 * %GFP_ATOMIC don't sleep.
2072 * @order: Power of two of allocation size in pages. 0 is a single page.
2074 * Allocate a page from the kernel page pool. When not in
2075 * interrupt context and apply the current process NUMA policy.
2076 * Returns NULL when no page can be allocated.
2078 * Don't call cpuset_update_task_memory_state() unless
2079 * 1) it's ok to take cpuset_sem (can WAIT), and
2080 * 2) allocating for current task (not interrupt).
2082 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2084 struct mempolicy *pol = get_task_policy(current);
2086 unsigned int cpuset_mems_cookie;
2088 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2089 pol = &default_policy;
2092 cpuset_mems_cookie = get_mems_allowed();
2095 * No reference counting needed for current->mempolicy
2096 * nor system default_policy
2098 if (pol->mode == MPOL_INTERLEAVE)
2099 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2101 page = __alloc_pages_nodemask(gfp, order,
2102 policy_zonelist(gfp, pol, numa_node_id()),
2103 policy_nodemask(gfp, pol));
2105 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2110 EXPORT_SYMBOL(alloc_pages_current);
2112 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2114 struct mempolicy *pol = mpol_dup(vma_policy(src));
2117 return PTR_ERR(pol);
2118 dst->vm_policy = pol;
2123 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2124 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2125 * with the mems_allowed returned by cpuset_mems_allowed(). This
2126 * keeps mempolicies cpuset relative after its cpuset moves. See
2127 * further kernel/cpuset.c update_nodemask().
2129 * current's mempolicy may be rebinded by the other task(the task that changes
2130 * cpuset's mems), so we needn't do rebind work for current task.
2133 /* Slow path of a mempolicy duplicate */
2134 struct mempolicy *__mpol_dup(struct mempolicy *old)
2136 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2139 return ERR_PTR(-ENOMEM);
2141 /* task's mempolicy is protected by alloc_lock */
2142 if (old == current->mempolicy) {
2145 task_unlock(current);
2150 if (current_cpuset_is_being_rebound()) {
2151 nodemask_t mems = cpuset_mems_allowed(current);
2152 if (new->flags & MPOL_F_REBINDING)
2153 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2155 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2158 atomic_set(&new->refcnt, 1);
2162 /* Slow path of a mempolicy comparison */
2163 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2167 if (a->mode != b->mode)
2169 if (a->flags != b->flags)
2171 if (mpol_store_user_nodemask(a))
2172 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2178 case MPOL_INTERLEAVE:
2179 return !!nodes_equal(a->v.nodes, b->v.nodes);
2180 case MPOL_PREFERRED:
2181 return a->v.preferred_node == b->v.preferred_node;
2189 * Shared memory backing store policy support.
2191 * Remember policies even when nobody has shared memory mapped.
2192 * The policies are kept in Red-Black tree linked from the inode.
2193 * They are protected by the sp->lock spinlock, which should be held
2194 * for any accesses to the tree.
2197 /* lookup first element intersecting start-end */
2198 /* Caller holds sp->lock */
2199 static struct sp_node *
2200 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2202 struct rb_node *n = sp->root.rb_node;
2205 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2207 if (start >= p->end)
2209 else if (end <= p->start)
2217 struct sp_node *w = NULL;
2218 struct rb_node *prev = rb_prev(n);
2221 w = rb_entry(prev, struct sp_node, nd);
2222 if (w->end <= start)
2226 return rb_entry(n, struct sp_node, nd);
2229 /* Insert a new shared policy into the list. */
2230 /* Caller holds sp->lock */
2231 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2233 struct rb_node **p = &sp->root.rb_node;
2234 struct rb_node *parent = NULL;
2239 nd = rb_entry(parent, struct sp_node, nd);
2240 if (new->start < nd->start)
2242 else if (new->end > nd->end)
2243 p = &(*p)->rb_right;
2247 rb_link_node(&new->nd, parent, p);
2248 rb_insert_color(&new->nd, &sp->root);
2249 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2250 new->policy ? new->policy->mode : 0);
2253 /* Find shared policy intersecting idx */
2255 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2257 struct mempolicy *pol = NULL;
2260 if (!sp->root.rb_node)
2262 spin_lock(&sp->lock);
2263 sn = sp_lookup(sp, idx, idx+1);
2265 mpol_get(sn->policy);
2268 spin_unlock(&sp->lock);
2272 static void sp_free(struct sp_node *n)
2274 mpol_put(n->policy);
2275 kmem_cache_free(sn_cache, n);
2279 * mpol_misplaced - check whether current page node is valid in policy
2281 * @page - page to be checked
2282 * @vma - vm area where page mapped
2283 * @addr - virtual address where page mapped
2285 * Lookup current policy node id for vma,addr and "compare to" page's
2289 * -1 - not misplaced, page is in the right node
2290 * node - node id where the page should be
2292 * Policy determination "mimics" alloc_page_vma().
2293 * Called from fault path where we know the vma and faulting address.
2295 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2297 struct mempolicy *pol;
2299 int curnid = page_to_nid(page);
2300 unsigned long pgoff;
2306 pol = get_vma_policy(current, vma, addr);
2307 if (!(pol->flags & MPOL_F_MOF))
2310 switch (pol->mode) {
2311 case MPOL_INTERLEAVE:
2312 BUG_ON(addr >= vma->vm_end);
2313 BUG_ON(addr < vma->vm_start);
2315 pgoff = vma->vm_pgoff;
2316 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2317 polnid = offset_il_node(pol, vma, pgoff);
2320 case MPOL_PREFERRED:
2321 if (pol->flags & MPOL_F_LOCAL)
2322 polnid = numa_node_id();
2324 polnid = pol->v.preferred_node;
2329 * allows binding to multiple nodes.
2330 * use current page if in policy nodemask,
2331 * else select nearest allowed node, if any.
2332 * If no allowed nodes, use current [!misplaced].
2334 if (node_isset(curnid, pol->v.nodes))
2336 (void)first_zones_zonelist(
2337 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2338 gfp_zone(GFP_HIGHUSER),
2339 &pol->v.nodes, &zone);
2340 polnid = zone->node;
2347 /* Migrate the page towards the node whose CPU is referencing it */
2348 if (pol->flags & MPOL_F_MORON) {
2351 polnid = numa_node_id();
2354 * Multi-stage node selection is used in conjunction
2355 * with a periodic migration fault to build a temporal
2356 * task<->page relation. By using a two-stage filter we
2357 * remove short/unlikely relations.
2359 * Using P(p) ~ n_p / n_t as per frequentist
2360 * probability, we can equate a task's usage of a
2361 * particular page (n_p) per total usage of this
2362 * page (n_t) (in a given time-span) to a probability.
2364 * Our periodic faults will sample this probability and
2365 * getting the same result twice in a row, given these
2366 * samples are fully independent, is then given by
2367 * P(n)^2, provided our sample period is sufficiently
2368 * short compared to the usage pattern.
2370 * This quadric squishes small probabilities, making
2371 * it less likely we act on an unlikely task<->page
2374 last_nid = page_nid_xchg_last(page, polnid);
2375 if (last_nid != polnid)
2379 if (curnid != polnid)
2387 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2389 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2390 rb_erase(&n->nd, &sp->root);
2394 static void sp_node_init(struct sp_node *node, unsigned long start,
2395 unsigned long end, struct mempolicy *pol)
2397 node->start = start;
2402 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2403 struct mempolicy *pol)
2406 struct mempolicy *newpol;
2408 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2412 newpol = mpol_dup(pol);
2413 if (IS_ERR(newpol)) {
2414 kmem_cache_free(sn_cache, n);
2417 newpol->flags |= MPOL_F_SHARED;
2418 sp_node_init(n, start, end, newpol);
2423 /* Replace a policy range. */
2424 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2425 unsigned long end, struct sp_node *new)
2428 struct sp_node *n_new = NULL;
2429 struct mempolicy *mpol_new = NULL;
2433 spin_lock(&sp->lock);
2434 n = sp_lookup(sp, start, end);
2435 /* Take care of old policies in the same range. */
2436 while (n && n->start < end) {
2437 struct rb_node *next = rb_next(&n->nd);
2438 if (n->start >= start) {
2444 /* Old policy spanning whole new range. */
2449 *mpol_new = *n->policy;
2450 atomic_set(&mpol_new->refcnt, 1);
2451 sp_node_init(n_new, end, n->end, mpol_new);
2453 sp_insert(sp, n_new);
2462 n = rb_entry(next, struct sp_node, nd);
2466 spin_unlock(&sp->lock);
2473 kmem_cache_free(sn_cache, n_new);
2478 spin_unlock(&sp->lock);
2480 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2483 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2490 * mpol_shared_policy_init - initialize shared policy for inode
2491 * @sp: pointer to inode shared policy
2492 * @mpol: struct mempolicy to install
2494 * Install non-NULL @mpol in inode's shared policy rb-tree.
2495 * On entry, the current task has a reference on a non-NULL @mpol.
2496 * This must be released on exit.
2497 * This is called at get_inode() calls and we can use GFP_KERNEL.
2499 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2503 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2504 spin_lock_init(&sp->lock);
2507 struct vm_area_struct pvma;
2508 struct mempolicy *new;
2509 NODEMASK_SCRATCH(scratch);
2513 /* contextualize the tmpfs mount point mempolicy */
2514 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2516 goto free_scratch; /* no valid nodemask intersection */
2519 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2520 task_unlock(current);
2524 /* Create pseudo-vma that contains just the policy */
2525 memset(&pvma, 0, sizeof(struct vm_area_struct));
2526 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2527 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2530 mpol_put(new); /* drop initial ref */
2532 NODEMASK_SCRATCH_FREE(scratch);
2534 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2538 int mpol_set_shared_policy(struct shared_policy *info,
2539 struct vm_area_struct *vma, struct mempolicy *npol)
2542 struct sp_node *new = NULL;
2543 unsigned long sz = vma_pages(vma);
2545 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2547 sz, npol ? npol->mode : -1,
2548 npol ? npol->flags : -1,
2549 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2552 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2556 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2562 /* Free a backing policy store on inode delete. */
2563 void mpol_free_shared_policy(struct shared_policy *p)
2566 struct rb_node *next;
2568 if (!p->root.rb_node)
2570 spin_lock(&p->lock);
2571 next = rb_first(&p->root);
2573 n = rb_entry(next, struct sp_node, nd);
2574 next = rb_next(&n->nd);
2577 spin_unlock(&p->lock);
2580 #ifdef CONFIG_NUMA_BALANCING
2581 static bool __initdata numabalancing_override;
2583 static void __init check_numabalancing_enable(void)
2585 bool numabalancing_default = false;
2587 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2588 numabalancing_default = true;
2590 if (nr_node_ids > 1 && !numabalancing_override) {
2591 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2592 "Configure with numa_balancing= or sysctl");
2593 set_numabalancing_state(numabalancing_default);
2597 static int __init setup_numabalancing(char *str)
2602 numabalancing_override = true;
2604 if (!strcmp(str, "enable")) {
2605 set_numabalancing_state(true);
2607 } else if (!strcmp(str, "disable")) {
2608 set_numabalancing_state(false);
2613 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2617 __setup("numa_balancing=", setup_numabalancing);
2619 static inline void __init check_numabalancing_enable(void)
2622 #endif /* CONFIG_NUMA_BALANCING */
2624 /* assumes fs == KERNEL_DS */
2625 void __init numa_policy_init(void)
2627 nodemask_t interleave_nodes;
2628 unsigned long largest = 0;
2629 int nid, prefer = 0;
2631 policy_cache = kmem_cache_create("numa_policy",
2632 sizeof(struct mempolicy),
2633 0, SLAB_PANIC, NULL);
2635 sn_cache = kmem_cache_create("shared_policy_node",
2636 sizeof(struct sp_node),
2637 0, SLAB_PANIC, NULL);
2639 for_each_node(nid) {
2640 preferred_node_policy[nid] = (struct mempolicy) {
2641 .refcnt = ATOMIC_INIT(1),
2642 .mode = MPOL_PREFERRED,
2643 .flags = MPOL_F_MOF | MPOL_F_MORON,
2644 .v = { .preferred_node = nid, },
2649 * Set interleaving policy for system init. Interleaving is only
2650 * enabled across suitably sized nodes (default is >= 16MB), or
2651 * fall back to the largest node if they're all smaller.
2653 nodes_clear(interleave_nodes);
2654 for_each_node_state(nid, N_MEMORY) {
2655 unsigned long total_pages = node_present_pages(nid);
2657 /* Preserve the largest node */
2658 if (largest < total_pages) {
2659 largest = total_pages;
2663 /* Interleave this node? */
2664 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2665 node_set(nid, interleave_nodes);
2668 /* All too small, use the largest */
2669 if (unlikely(nodes_empty(interleave_nodes)))
2670 node_set(prefer, interleave_nodes);
2672 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2673 printk("numa_policy_init: interleaving failed\n");
2675 check_numabalancing_enable();
2678 /* Reset policy of current process to default */
2679 void numa_default_policy(void)
2681 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2685 * Parse and format mempolicy from/to strings
2689 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2691 static const char * const policy_modes[] =
2693 [MPOL_DEFAULT] = "default",
2694 [MPOL_PREFERRED] = "prefer",
2695 [MPOL_BIND] = "bind",
2696 [MPOL_INTERLEAVE] = "interleave",
2697 [MPOL_LOCAL] = "local",
2703 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2704 * @str: string containing mempolicy to parse
2705 * @mpol: pointer to struct mempolicy pointer, returned on success.
2708 * <mode>[=<flags>][:<nodelist>]
2710 * On success, returns 0, else 1
2712 int mpol_parse_str(char *str, struct mempolicy **mpol)
2714 struct mempolicy *new = NULL;
2715 unsigned short mode;
2716 unsigned short mode_flags;
2718 char *nodelist = strchr(str, ':');
2719 char *flags = strchr(str, '=');
2723 /* NUL-terminate mode or flags string */
2725 if (nodelist_parse(nodelist, nodes))
2727 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2733 *flags++ = '\0'; /* terminate mode string */
2735 for (mode = 0; mode < MPOL_MAX; mode++) {
2736 if (!strcmp(str, policy_modes[mode])) {
2740 if (mode >= MPOL_MAX)
2744 case MPOL_PREFERRED:
2746 * Insist on a nodelist of one node only
2749 char *rest = nodelist;
2750 while (isdigit(*rest))
2756 case MPOL_INTERLEAVE:
2758 * Default to online nodes with memory if no nodelist
2761 nodes = node_states[N_MEMORY];
2765 * Don't allow a nodelist; mpol_new() checks flags
2769 mode = MPOL_PREFERRED;
2773 * Insist on a empty nodelist
2780 * Insist on a nodelist
2789 * Currently, we only support two mutually exclusive
2792 if (!strcmp(flags, "static"))
2793 mode_flags |= MPOL_F_STATIC_NODES;
2794 else if (!strcmp(flags, "relative"))
2795 mode_flags |= MPOL_F_RELATIVE_NODES;
2800 new = mpol_new(mode, mode_flags, &nodes);
2805 * Save nodes for mpol_to_str() to show the tmpfs mount options
2806 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2808 if (mode != MPOL_PREFERRED)
2809 new->v.nodes = nodes;
2811 new->v.preferred_node = first_node(nodes);
2813 new->flags |= MPOL_F_LOCAL;
2816 * Save nodes for contextualization: this will be used to "clone"
2817 * the mempolicy in a specific context [cpuset] at a later time.
2819 new->w.user_nodemask = nodes;
2824 /* Restore string for error message */
2833 #endif /* CONFIG_TMPFS */
2836 * mpol_to_str - format a mempolicy structure for printing
2837 * @buffer: to contain formatted mempolicy string
2838 * @maxlen: length of @buffer
2839 * @pol: pointer to mempolicy to be formatted
2841 * Convert a mempolicy into a string.
2842 * Returns the number of characters in buffer (if positive)
2843 * or an error (negative)
2845 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2850 unsigned short mode;
2851 unsigned short flags = pol ? pol->flags : 0;
2854 * Sanity check: room for longest mode, flag and some nodes
2856 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2858 if (!pol || pol == &default_policy)
2859 mode = MPOL_DEFAULT;
2868 case MPOL_PREFERRED:
2870 if (flags & MPOL_F_LOCAL)
2873 node_set(pol->v.preferred_node, nodes);
2878 case MPOL_INTERLEAVE:
2879 nodes = pol->v.nodes;
2886 l = strlen(policy_modes[mode]);
2887 if (buffer + maxlen < p + l + 1)
2890 strcpy(p, policy_modes[mode]);
2893 if (flags & MPOL_MODE_FLAGS) {
2894 if (buffer + maxlen < p + 2)
2899 * Currently, the only defined flags are mutually exclusive
2901 if (flags & MPOL_F_STATIC_NODES)
2902 p += snprintf(p, buffer + maxlen - p, "static");
2903 else if (flags & MPOL_F_RELATIVE_NODES)
2904 p += snprintf(p, buffer + maxlen - p, "relative");
2907 if (!nodes_empty(nodes)) {
2908 if (buffer + maxlen < p + 2)
2911 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);