2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
93 #include <linux/mmu_notifier.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
97 #include <linux/random.h>
102 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
105 static struct kmem_cache *policy_cache;
106 static struct kmem_cache *sn_cache;
108 /* Highest zone. An specific allocation for a zone below that is not
110 enum zone_type policy_zone = 0;
113 * run-time system-wide default policy => local allocation
115 static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
121 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
123 static struct mempolicy *get_task_policy(struct task_struct *p)
125 struct mempolicy *pol = p->mempolicy;
128 int node = numa_node_id();
130 if (node != NUMA_NO_NODE) {
131 pol = &preferred_node_policy[node];
133 * preferred_node_policy is not initialised early in
144 static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * If we have a lock to protect task->mempolicy in read-side, we do
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
162 } mpol_ops[MPOL_MAX];
164 /* Check that the nodemask contains at least one populated zone */
165 static int is_valid_nodemask(const nodemask_t *nodemask)
167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
170 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
172 return pol->flags & MPOL_MODE_FLAGS;
175 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
183 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
185 if (nodes_empty(*nodes))
187 pol->v.nodes = *nodes;
191 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
198 pol->v.preferred_node = first_node(*nodes);
202 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
204 if (!is_valid_nodemask(nodes))
206 pol->v.nodes = *nodes;
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
219 static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
228 nodes_and(nsc->mask1,
229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
250 ret = mpol_ops[pol->mode].create(pol, NULL);
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
261 struct mempolicy *policy;
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
295 policy->flags = flags;
300 /* Slow path of a mpol destructor. */
301 void __mpol_put(struct mempolicy *p)
303 if (!atomic_dec_and_test(&p->refcnt))
305 kmem_cache_free(policy_cache, p);
308 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
319 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
344 if (nodes_empty(tmp))
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
363 static void mpol_rebind_preferred(struct mempolicy *pol,
364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
372 if (node_isset(node, *nodes)) {
373 pol->v.preferred_node = node;
374 pol->flags &= ~MPOL_F_LOCAL;
376 pol->flags |= MPOL_F_LOCAL;
377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
384 pol->w.cpuset_mems_allowed = *nodes;
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
396 * If we have a lock to protect task->mempolicy in read-side, we do
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
404 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
433 * Called with task's alloc_lock held.
436 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
439 mpol_rebind_policy(tsk->mempolicy, new, step);
443 * Rebind each vma in mm to new nodemask.
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
448 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
450 struct vm_area_struct *vma;
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
455 up_write(&mm->mmap_sem);
458 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
460 .rebind = mpol_rebind_default,
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
476 static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
483 static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
497 if (!pte_present(*pte))
499 page = vm_normal_page(vma, addr, *pte);
503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
506 if (PageReserved(page))
508 nid = page_to_nid(page);
509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
513 migrate_page_add(page, private, flags);
516 } while (pte++, addr += PAGE_SIZE, addr != end);
517 pte_unmap_unlock(orig_pte, ptl);
521 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
525 #ifdef CONFIG_HUGETLB_PAGE
531 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
532 entry = huge_ptep_get((pte_t *)pmd);
533 if (!pte_present(entry))
535 page = pte_page(entry);
536 nid = page_to_nid(page);
537 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
539 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
540 if (flags & (MPOL_MF_MOVE_ALL) ||
541 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
542 isolate_huge_page(page, private);
550 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
551 unsigned long addr, unsigned long end,
552 const nodemask_t *nodes, unsigned long flags,
558 pmd = pmd_offset(pud, addr);
560 next = pmd_addr_end(addr, end);
561 if (!pmd_present(*pmd))
563 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
564 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
568 split_huge_page_pmd(vma, addr, pmd);
569 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
571 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
574 } while (pmd++, addr = next, addr != end);
578 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
579 unsigned long addr, unsigned long end,
580 const nodemask_t *nodes, unsigned long flags,
586 pud = pud_offset(pgd, addr);
588 next = pud_addr_end(addr, end);
589 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
591 if (pud_none_or_clear_bad(pud))
593 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
596 } while (pud++, addr = next, addr != end);
600 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
601 unsigned long addr, unsigned long end,
602 const nodemask_t *nodes, unsigned long flags,
608 pgd = pgd_offset(vma->vm_mm, addr);
610 next = pgd_addr_end(addr, end);
611 if (pgd_none_or_clear_bad(pgd))
613 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
616 } while (pgd++, addr = next, addr != end);
620 #ifdef CONFIG_NUMA_BALANCING
622 * This is used to mark a range of virtual addresses to be inaccessible.
623 * These are later cleared by a NUMA hinting fault. Depending on these
624 * faults, pages may be migrated for better NUMA placement.
626 * This is assuming that NUMA faults are handled using PROT_NONE. If
627 * an architecture makes a different choice, it will need further
628 * changes to the core.
630 unsigned long change_prot_numa(struct vm_area_struct *vma,
631 unsigned long addr, unsigned long end)
635 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
637 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
642 static unsigned long change_prot_numa(struct vm_area_struct *vma,
643 unsigned long addr, unsigned long end)
647 #endif /* CONFIG_NUMA_BALANCING */
650 * Walk through page tables and collect pages to be migrated.
652 * If pages found in a given range are on a set of nodes (determined by
653 * @nodes and @flags,) it's isolated and queued to the pagelist which is
654 * passed via @private.)
656 static struct vm_area_struct *
657 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
658 const nodemask_t *nodes, unsigned long flags, void *private)
661 struct vm_area_struct *first, *vma, *prev;
664 first = find_vma(mm, start);
666 return ERR_PTR(-EFAULT);
668 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
669 unsigned long endvma = vma->vm_end;
673 if (vma->vm_start > start)
674 start = vma->vm_start;
676 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
677 if (!vma->vm_next && vma->vm_end < end)
678 return ERR_PTR(-EFAULT);
679 if (prev && prev->vm_end < vma->vm_start)
680 return ERR_PTR(-EFAULT);
683 if (flags & MPOL_MF_LAZY) {
684 change_prot_numa(vma, start, endvma);
688 if ((flags & MPOL_MF_STRICT) ||
689 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
690 vma_migratable(vma))) {
692 err = queue_pages_pgd_range(vma, start, endvma, nodes,
695 first = ERR_PTR(err);
706 * Apply policy to a single VMA
707 * This must be called with the mmap_sem held for writing.
709 static int vma_replace_policy(struct vm_area_struct *vma,
710 struct mempolicy *pol)
713 struct mempolicy *old;
714 struct mempolicy *new;
716 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
717 vma->vm_start, vma->vm_end, vma->vm_pgoff,
718 vma->vm_ops, vma->vm_file,
719 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
725 if (vma->vm_ops && vma->vm_ops->set_policy) {
726 err = vma->vm_ops->set_policy(vma, new);
731 old = vma->vm_policy;
732 vma->vm_policy = new; /* protected by mmap_sem */
741 /* Step 2: apply policy to a range and do splits. */
742 static int mbind_range(struct mm_struct *mm, unsigned long start,
743 unsigned long end, struct mempolicy *new_pol)
745 struct vm_area_struct *next;
746 struct vm_area_struct *prev;
747 struct vm_area_struct *vma;
750 unsigned long vmstart;
753 vma = find_vma(mm, start);
754 if (!vma || vma->vm_start > start)
758 if (start > vma->vm_start)
761 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
763 vmstart = max(start, vma->vm_start);
764 vmend = min(end, vma->vm_end);
766 if (mpol_equal(vma_policy(vma), new_pol))
769 pgoff = vma->vm_pgoff +
770 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
771 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
772 vma->anon_vma, vma->vm_file, pgoff,
777 if (mpol_equal(vma_policy(vma), new_pol))
779 /* vma_merge() joined vma && vma->next, case 8 */
782 if (vma->vm_start != vmstart) {
783 err = split_vma(vma->vm_mm, vma, vmstart, 1);
787 if (vma->vm_end != vmend) {
788 err = split_vma(vma->vm_mm, vma, vmend, 0);
793 err = vma_replace_policy(vma, new_pol);
802 /* Set the process memory policy */
803 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
806 struct mempolicy *new, *old;
807 struct mm_struct *mm = current->mm;
808 NODEMASK_SCRATCH(scratch);
814 new = mpol_new(mode, flags, nodes);
820 * prevent changing our mempolicy while show_numa_maps()
822 * Note: do_set_mempolicy() can be called at init time
826 down_write(&mm->mmap_sem);
828 ret = mpol_set_nodemask(new, nodes, scratch);
830 task_unlock(current);
832 up_write(&mm->mmap_sem);
836 old = current->mempolicy;
837 current->mempolicy = new;
838 if (new && new->mode == MPOL_INTERLEAVE &&
839 nodes_weight(new->v.nodes))
840 current->il_next = first_node(new->v.nodes);
841 task_unlock(current);
843 up_write(&mm->mmap_sem);
848 NODEMASK_SCRATCH_FREE(scratch);
853 * Return nodemask for policy for get_mempolicy() query
855 * Called with task's alloc_lock held
857 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
860 if (p == &default_policy)
866 case MPOL_INTERLEAVE:
870 if (!(p->flags & MPOL_F_LOCAL))
871 node_set(p->v.preferred_node, *nodes);
872 /* else return empty node mask for local allocation */
879 static int lookup_node(struct mm_struct *mm, unsigned long addr)
884 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
886 err = page_to_nid(p);
892 /* Retrieve NUMA policy */
893 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
894 unsigned long addr, unsigned long flags)
897 struct mm_struct *mm = current->mm;
898 struct vm_area_struct *vma = NULL;
899 struct mempolicy *pol = current->mempolicy;
902 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
905 if (flags & MPOL_F_MEMS_ALLOWED) {
906 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
908 *policy = 0; /* just so it's initialized */
910 *nmask = cpuset_current_mems_allowed;
911 task_unlock(current);
915 if (flags & MPOL_F_ADDR) {
917 * Do NOT fall back to task policy if the
918 * vma/shared policy at addr is NULL. We
919 * want to return MPOL_DEFAULT in this case.
921 down_read(&mm->mmap_sem);
922 vma = find_vma_intersection(mm, addr, addr+1);
924 up_read(&mm->mmap_sem);
927 if (vma->vm_ops && vma->vm_ops->get_policy)
928 pol = vma->vm_ops->get_policy(vma, addr);
930 pol = vma->vm_policy;
935 pol = &default_policy; /* indicates default behavior */
937 if (flags & MPOL_F_NODE) {
938 if (flags & MPOL_F_ADDR) {
939 err = lookup_node(mm, addr);
943 } else if (pol == current->mempolicy &&
944 pol->mode == MPOL_INTERLEAVE) {
945 *policy = current->il_next;
951 *policy = pol == &default_policy ? MPOL_DEFAULT :
954 * Internal mempolicy flags must be masked off before exposing
955 * the policy to userspace.
957 *policy |= (pol->flags & MPOL_MODE_FLAGS);
961 up_read(¤t->mm->mmap_sem);
967 if (mpol_store_user_nodemask(pol)) {
968 *nmask = pol->w.user_nodemask;
971 get_policy_nodemask(pol, nmask);
972 task_unlock(current);
979 up_read(¤t->mm->mmap_sem);
983 #ifdef CONFIG_MIGRATION
987 static void migrate_page_add(struct page *page, struct list_head *pagelist,
991 * Avoid migrating a page that is shared with others.
993 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
994 if (!isolate_lru_page(page)) {
995 list_add_tail(&page->lru, pagelist);
996 inc_zone_page_state(page, NR_ISOLATED_ANON +
997 page_is_file_cache(page));
1002 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1005 return alloc_huge_page_node(page_hstate(compound_head(page)),
1008 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1012 * Migrate pages from one node to a target node.
1013 * Returns error or the number of pages not migrated.
1015 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1019 LIST_HEAD(pagelist);
1023 node_set(source, nmask);
1026 * This does not "check" the range but isolates all pages that
1027 * need migration. Between passing in the full user address
1028 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1030 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1031 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1032 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1034 if (!list_empty(&pagelist)) {
1035 err = migrate_pages(&pagelist, new_node_page, dest,
1036 MIGRATE_SYNC, MR_SYSCALL);
1038 putback_movable_pages(&pagelist);
1045 * Move pages between the two nodesets so as to preserve the physical
1046 * layout as much as possible.
1048 * Returns the number of page that could not be moved.
1050 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1051 const nodemask_t *to, int flags)
1057 err = migrate_prep();
1061 down_read(&mm->mmap_sem);
1063 err = migrate_vmas(mm, from, to, flags);
1068 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1069 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1070 * bit in 'tmp', and return that <source, dest> pair for migration.
1071 * The pair of nodemasks 'to' and 'from' define the map.
1073 * If no pair of bits is found that way, fallback to picking some
1074 * pair of 'source' and 'dest' bits that are not the same. If the
1075 * 'source' and 'dest' bits are the same, this represents a node
1076 * that will be migrating to itself, so no pages need move.
1078 * If no bits are left in 'tmp', or if all remaining bits left
1079 * in 'tmp' correspond to the same bit in 'to', return false
1080 * (nothing left to migrate).
1082 * This lets us pick a pair of nodes to migrate between, such that
1083 * if possible the dest node is not already occupied by some other
1084 * source node, minimizing the risk of overloading the memory on a
1085 * node that would happen if we migrated incoming memory to a node
1086 * before migrating outgoing memory source that same node.
1088 * A single scan of tmp is sufficient. As we go, we remember the
1089 * most recent <s, d> pair that moved (s != d). If we find a pair
1090 * that not only moved, but what's better, moved to an empty slot
1091 * (d is not set in tmp), then we break out then, with that pair.
1092 * Otherwise when we finish scanning from_tmp, we at least have the
1093 * most recent <s, d> pair that moved. If we get all the way through
1094 * the scan of tmp without finding any node that moved, much less
1095 * moved to an empty node, then there is nothing left worth migrating.
1099 while (!nodes_empty(tmp)) {
1101 int source = NUMA_NO_NODE;
1104 for_each_node_mask(s, tmp) {
1107 * do_migrate_pages() tries to maintain the relative
1108 * node relationship of the pages established between
1109 * threads and memory areas.
1111 * However if the number of source nodes is not equal to
1112 * the number of destination nodes we can not preserve
1113 * this node relative relationship. In that case, skip
1114 * copying memory from a node that is in the destination
1117 * Example: [2,3,4] -> [3,4,5] moves everything.
1118 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1121 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1122 (node_isset(s, *to)))
1125 d = node_remap(s, *from, *to);
1129 source = s; /* Node moved. Memorize */
1132 /* dest not in remaining from nodes? */
1133 if (!node_isset(dest, tmp))
1136 if (source == NUMA_NO_NODE)
1139 node_clear(source, tmp);
1140 err = migrate_to_node(mm, source, dest, flags);
1147 up_read(&mm->mmap_sem);
1155 * Allocate a new page for page migration based on vma policy.
1156 * Start assuming that page is mapped by vma pointed to by @private.
1157 * Search forward from there, if not. N.B., this assumes that the
1158 * list of pages handed to migrate_pages()--which is how we get here--
1159 * is in virtual address order.
1161 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1163 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1164 unsigned long uninitialized_var(address);
1167 address = page_address_in_vma(page, vma);
1168 if (address != -EFAULT)
1173 if (PageHuge(page)) {
1175 return alloc_huge_page_noerr(vma, address, 1);
1178 * if !vma, alloc_page_vma() will use task or system default policy
1180 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1184 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1185 unsigned long flags)
1189 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1190 const nodemask_t *to, int flags)
1195 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1201 static long do_mbind(unsigned long start, unsigned long len,
1202 unsigned short mode, unsigned short mode_flags,
1203 nodemask_t *nmask, unsigned long flags)
1205 struct vm_area_struct *vma;
1206 struct mm_struct *mm = current->mm;
1207 struct mempolicy *new;
1210 LIST_HEAD(pagelist);
1212 if (flags & ~(unsigned long)MPOL_MF_VALID)
1214 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1217 if (start & ~PAGE_MASK)
1220 if (mode == MPOL_DEFAULT)
1221 flags &= ~MPOL_MF_STRICT;
1223 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1231 new = mpol_new(mode, mode_flags, nmask);
1233 return PTR_ERR(new);
1235 if (flags & MPOL_MF_LAZY)
1236 new->flags |= MPOL_F_MOF;
1239 * If we are using the default policy then operation
1240 * on discontinuous address spaces is okay after all
1243 flags |= MPOL_MF_DISCONTIG_OK;
1245 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1246 start, start + len, mode, mode_flags,
1247 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1249 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1251 err = migrate_prep();
1256 NODEMASK_SCRATCH(scratch);
1258 down_write(&mm->mmap_sem);
1260 err = mpol_set_nodemask(new, nmask, scratch);
1261 task_unlock(current);
1263 up_write(&mm->mmap_sem);
1266 NODEMASK_SCRATCH_FREE(scratch);
1271 vma = queue_pages_range(mm, start, end, nmask,
1272 flags | MPOL_MF_INVERT, &pagelist);
1274 err = PTR_ERR(vma); /* maybe ... */
1276 err = mbind_range(mm, start, end, new);
1281 if (!list_empty(&pagelist)) {
1282 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1283 nr_failed = migrate_pages(&pagelist, new_vma_page,
1285 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1287 putback_movable_pages(&pagelist);
1290 if (nr_failed && (flags & MPOL_MF_STRICT))
1293 putback_movable_pages(&pagelist);
1295 up_write(&mm->mmap_sem);
1302 * User space interface with variable sized bitmaps for nodelists.
1305 /* Copy a node mask from user space. */
1306 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1307 unsigned long maxnode)
1310 unsigned long nlongs;
1311 unsigned long endmask;
1314 nodes_clear(*nodes);
1315 if (maxnode == 0 || !nmask)
1317 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1320 nlongs = BITS_TO_LONGS(maxnode);
1321 if ((maxnode % BITS_PER_LONG) == 0)
1324 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1326 /* When the user specified more nodes than supported just check
1327 if the non supported part is all zero. */
1328 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1329 if (nlongs > PAGE_SIZE/sizeof(long))
1331 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1333 if (get_user(t, nmask + k))
1335 if (k == nlongs - 1) {
1341 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1345 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1347 nodes_addr(*nodes)[nlongs-1] &= endmask;
1351 /* Copy a kernel node mask to user space */
1352 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1355 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1356 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1358 if (copy > nbytes) {
1359 if (copy > PAGE_SIZE)
1361 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1365 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1368 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1369 unsigned long, mode, unsigned long __user *, nmask,
1370 unsigned long, maxnode, unsigned, flags)
1374 unsigned short mode_flags;
1376 mode_flags = mode & MPOL_MODE_FLAGS;
1377 mode &= ~MPOL_MODE_FLAGS;
1378 if (mode >= MPOL_MAX)
1380 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1381 (mode_flags & MPOL_F_RELATIVE_NODES))
1383 err = get_nodes(&nodes, nmask, maxnode);
1386 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1389 /* Set the process memory policy */
1390 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1391 unsigned long, maxnode)
1395 unsigned short flags;
1397 flags = mode & MPOL_MODE_FLAGS;
1398 mode &= ~MPOL_MODE_FLAGS;
1399 if ((unsigned int)mode >= MPOL_MAX)
1401 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1403 err = get_nodes(&nodes, nmask, maxnode);
1406 return do_set_mempolicy(mode, flags, &nodes);
1409 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1410 const unsigned long __user *, old_nodes,
1411 const unsigned long __user *, new_nodes)
1413 const struct cred *cred = current_cred(), *tcred;
1414 struct mm_struct *mm = NULL;
1415 struct task_struct *task;
1416 nodemask_t task_nodes;
1420 NODEMASK_SCRATCH(scratch);
1425 old = &scratch->mask1;
1426 new = &scratch->mask2;
1428 err = get_nodes(old, old_nodes, maxnode);
1432 err = get_nodes(new, new_nodes, maxnode);
1436 /* Find the mm_struct */
1438 task = pid ? find_task_by_vpid(pid) : current;
1444 get_task_struct(task);
1449 * Check if this process has the right to modify the specified
1450 * process. The right exists if the process has administrative
1451 * capabilities, superuser privileges or the same
1452 * userid as the target process.
1454 tcred = __task_cred(task);
1455 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1456 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1457 !capable(CAP_SYS_NICE)) {
1464 task_nodes = cpuset_mems_allowed(task);
1465 /* Is the user allowed to access the target nodes? */
1466 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1471 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1476 err = security_task_movememory(task);
1480 mm = get_task_mm(task);
1481 put_task_struct(task);
1488 err = do_migrate_pages(mm, old, new,
1489 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1493 NODEMASK_SCRATCH_FREE(scratch);
1498 put_task_struct(task);
1504 /* Retrieve NUMA policy */
1505 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1506 unsigned long __user *, nmask, unsigned long, maxnode,
1507 unsigned long, addr, unsigned long, flags)
1510 int uninitialized_var(pval);
1513 if (nmask != NULL && maxnode < MAX_NUMNODES)
1516 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1521 if (policy && put_user(pval, policy))
1525 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1530 #ifdef CONFIG_COMPAT
1532 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1533 compat_ulong_t __user *, nmask,
1534 compat_ulong_t, maxnode,
1535 compat_ulong_t, addr, compat_ulong_t, flags)
1538 unsigned long __user *nm = NULL;
1539 unsigned long nr_bits, alloc_size;
1540 DECLARE_BITMAP(bm, MAX_NUMNODES);
1542 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1543 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1546 nm = compat_alloc_user_space(alloc_size);
1548 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1550 if (!err && nmask) {
1551 unsigned long copy_size;
1552 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1553 err = copy_from_user(bm, nm, copy_size);
1554 /* ensure entire bitmap is zeroed */
1555 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1556 err |= compat_put_bitmap(nmask, bm, nr_bits);
1562 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1563 compat_ulong_t, maxnode)
1566 unsigned long __user *nm = NULL;
1567 unsigned long nr_bits, alloc_size;
1568 DECLARE_BITMAP(bm, MAX_NUMNODES);
1570 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1571 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1574 err = compat_get_bitmap(bm, nmask, nr_bits);
1575 nm = compat_alloc_user_space(alloc_size);
1576 err |= copy_to_user(nm, bm, alloc_size);
1582 return sys_set_mempolicy(mode, nm, nr_bits+1);
1585 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1586 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1587 compat_ulong_t, maxnode, compat_ulong_t, flags)
1590 unsigned long __user *nm = NULL;
1591 unsigned long nr_bits, alloc_size;
1594 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1595 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1598 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1599 nm = compat_alloc_user_space(alloc_size);
1600 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1606 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1612 * get_vma_policy(@task, @vma, @addr)
1613 * @task - task for fallback if vma policy == default
1614 * @vma - virtual memory area whose policy is sought
1615 * @addr - address in @vma for shared policy lookup
1617 * Returns effective policy for a VMA at specified address.
1618 * Falls back to @task or system default policy, as necessary.
1619 * Current or other task's task mempolicy and non-shared vma policies must be
1620 * protected by task_lock(task) by the caller.
1621 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1622 * count--added by the get_policy() vm_op, as appropriate--to protect against
1623 * freeing by another task. It is the caller's responsibility to free the
1624 * extra reference for shared policies.
1626 struct mempolicy *get_vma_policy(struct task_struct *task,
1627 struct vm_area_struct *vma, unsigned long addr)
1629 struct mempolicy *pol = get_task_policy(task);
1632 if (vma->vm_ops && vma->vm_ops->get_policy) {
1633 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1637 } else if (vma->vm_policy) {
1638 pol = vma->vm_policy;
1641 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1642 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1643 * count on these policies which will be dropped by
1644 * mpol_cond_put() later
1646 if (mpol_needs_cond_ref(pol))
1651 pol = &default_policy;
1655 bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1657 struct mempolicy *pol = get_task_policy(task);
1659 if (vma->vm_ops && vma->vm_ops->get_policy) {
1662 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1663 if (pol && (pol->flags & MPOL_F_MOF))
1668 } else if (vma->vm_policy) {
1669 pol = vma->vm_policy;
1674 return default_policy.flags & MPOL_F_MOF;
1676 return pol->flags & MPOL_F_MOF;
1679 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1681 enum zone_type dynamic_policy_zone = policy_zone;
1683 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1686 * if policy->v.nodes has movable memory only,
1687 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1689 * policy->v.nodes is intersect with node_states[N_MEMORY].
1690 * so if the following test faile, it implies
1691 * policy->v.nodes has movable memory only.
1693 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1694 dynamic_policy_zone = ZONE_MOVABLE;
1696 return zone >= dynamic_policy_zone;
1700 * Return a nodemask representing a mempolicy for filtering nodes for
1703 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1705 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1706 if (unlikely(policy->mode == MPOL_BIND) &&
1707 apply_policy_zone(policy, gfp_zone(gfp)) &&
1708 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1709 return &policy->v.nodes;
1714 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1715 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1718 switch (policy->mode) {
1719 case MPOL_PREFERRED:
1720 if (!(policy->flags & MPOL_F_LOCAL))
1721 nd = policy->v.preferred_node;
1725 * Normally, MPOL_BIND allocations are node-local within the
1726 * allowed nodemask. However, if __GFP_THISNODE is set and the
1727 * current node isn't part of the mask, we use the zonelist for
1728 * the first node in the mask instead.
1730 if (unlikely(gfp & __GFP_THISNODE) &&
1731 unlikely(!node_isset(nd, policy->v.nodes)))
1732 nd = first_node(policy->v.nodes);
1737 return node_zonelist(nd, gfp);
1740 /* Do dynamic interleaving for a process */
1741 static unsigned interleave_nodes(struct mempolicy *policy)
1744 struct task_struct *me = current;
1747 next = next_node(nid, policy->v.nodes);
1748 if (next >= MAX_NUMNODES)
1749 next = first_node(policy->v.nodes);
1750 if (next < MAX_NUMNODES)
1756 * Depending on the memory policy provide a node from which to allocate the
1759 unsigned int mempolicy_slab_node(void)
1761 struct mempolicy *policy;
1762 int node = numa_mem_id();
1767 policy = current->mempolicy;
1768 if (!policy || policy->flags & MPOL_F_LOCAL)
1771 switch (policy->mode) {
1772 case MPOL_PREFERRED:
1774 * handled MPOL_F_LOCAL above
1776 return policy->v.preferred_node;
1778 case MPOL_INTERLEAVE:
1779 return interleave_nodes(policy);
1783 * Follow bind policy behavior and start allocation at the
1786 struct zonelist *zonelist;
1788 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1789 zonelist = &NODE_DATA(node)->node_zonelists[0];
1790 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1793 return zone ? zone->node : node;
1801 /* Do static interleaving for a VMA with known offset. */
1802 static unsigned offset_il_node(struct mempolicy *pol,
1803 struct vm_area_struct *vma, unsigned long off)
1805 unsigned nnodes = nodes_weight(pol->v.nodes);
1808 int nid = NUMA_NO_NODE;
1811 return numa_node_id();
1812 target = (unsigned int)off % nnodes;
1815 nid = next_node(nid, pol->v.nodes);
1817 } while (c <= target);
1821 /* Determine a node number for interleave */
1822 static inline unsigned interleave_nid(struct mempolicy *pol,
1823 struct vm_area_struct *vma, unsigned long addr, int shift)
1829 * for small pages, there is no difference between
1830 * shift and PAGE_SHIFT, so the bit-shift is safe.
1831 * for huge pages, since vm_pgoff is in units of small
1832 * pages, we need to shift off the always 0 bits to get
1835 BUG_ON(shift < PAGE_SHIFT);
1836 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1837 off += (addr - vma->vm_start) >> shift;
1838 return offset_il_node(pol, vma, off);
1840 return interleave_nodes(pol);
1844 * Return the bit number of a random bit set in the nodemask.
1845 * (returns NUMA_NO_NODE if nodemask is empty)
1847 int node_random(const nodemask_t *maskp)
1849 int w, bit = NUMA_NO_NODE;
1851 w = nodes_weight(*maskp);
1853 bit = bitmap_ord_to_pos(maskp->bits,
1854 get_random_int() % w, MAX_NUMNODES);
1858 #ifdef CONFIG_HUGETLBFS
1860 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1861 * @vma = virtual memory area whose policy is sought
1862 * @addr = address in @vma for shared policy lookup and interleave policy
1863 * @gfp_flags = for requested zone
1864 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1865 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1867 * Returns a zonelist suitable for a huge page allocation and a pointer
1868 * to the struct mempolicy for conditional unref after allocation.
1869 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1870 * @nodemask for filtering the zonelist.
1872 * Must be protected by read_mems_allowed_begin()
1874 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1875 gfp_t gfp_flags, struct mempolicy **mpol,
1876 nodemask_t **nodemask)
1878 struct zonelist *zl;
1880 *mpol = get_vma_policy(current, vma, addr);
1881 *nodemask = NULL; /* assume !MPOL_BIND */
1883 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1884 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1885 huge_page_shift(hstate_vma(vma))), gfp_flags);
1887 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1888 if ((*mpol)->mode == MPOL_BIND)
1889 *nodemask = &(*mpol)->v.nodes;
1895 * init_nodemask_of_mempolicy
1897 * If the current task's mempolicy is "default" [NULL], return 'false'
1898 * to indicate default policy. Otherwise, extract the policy nodemask
1899 * for 'bind' or 'interleave' policy into the argument nodemask, or
1900 * initialize the argument nodemask to contain the single node for
1901 * 'preferred' or 'local' policy and return 'true' to indicate presence
1902 * of non-default mempolicy.
1904 * We don't bother with reference counting the mempolicy [mpol_get/put]
1905 * because the current task is examining it's own mempolicy and a task's
1906 * mempolicy is only ever changed by the task itself.
1908 * N.B., it is the caller's responsibility to free a returned nodemask.
1910 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1912 struct mempolicy *mempolicy;
1915 if (!(mask && current->mempolicy))
1919 mempolicy = current->mempolicy;
1920 switch (mempolicy->mode) {
1921 case MPOL_PREFERRED:
1922 if (mempolicy->flags & MPOL_F_LOCAL)
1923 nid = numa_node_id();
1925 nid = mempolicy->v.preferred_node;
1926 init_nodemask_of_node(mask, nid);
1931 case MPOL_INTERLEAVE:
1932 *mask = mempolicy->v.nodes;
1938 task_unlock(current);
1945 * mempolicy_nodemask_intersects
1947 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1948 * policy. Otherwise, check for intersection between mask and the policy
1949 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1950 * policy, always return true since it may allocate elsewhere on fallback.
1952 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1954 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1955 const nodemask_t *mask)
1957 struct mempolicy *mempolicy;
1963 mempolicy = tsk->mempolicy;
1967 switch (mempolicy->mode) {
1968 case MPOL_PREFERRED:
1970 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1971 * allocate from, they may fallback to other nodes when oom.
1972 * Thus, it's possible for tsk to have allocated memory from
1977 case MPOL_INTERLEAVE:
1978 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1988 /* Allocate a page in interleaved policy.
1989 Own path because it needs to do special accounting. */
1990 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1993 struct zonelist *zl;
1996 zl = node_zonelist(nid, gfp);
1997 page = __alloc_pages(gfp, order, zl);
1998 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1999 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2004 * alloc_pages_vma - Allocate a page for a VMA.
2007 * %GFP_USER user allocation.
2008 * %GFP_KERNEL kernel allocations,
2009 * %GFP_HIGHMEM highmem/user allocations,
2010 * %GFP_FS allocation should not call back into a file system.
2011 * %GFP_ATOMIC don't sleep.
2013 * @order:Order of the GFP allocation.
2014 * @vma: Pointer to VMA or NULL if not available.
2015 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2017 * This function allocates a page from the kernel page pool and applies
2018 * a NUMA policy associated with the VMA or the current process.
2019 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2020 * mm_struct of the VMA to prevent it from going away. Should be used for
2021 * all allocations for pages that will be mapped into
2022 * user space. Returns NULL when no page can be allocated.
2024 * Should be called with the mm_sem of the vma hold.
2027 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2028 unsigned long addr, int node)
2030 struct mempolicy *pol;
2032 unsigned int cpuset_mems_cookie;
2035 pol = get_vma_policy(current, vma, addr);
2036 cpuset_mems_cookie = read_mems_allowed_begin();
2038 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2041 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2043 page = alloc_page_interleave(gfp, order, nid);
2044 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2049 page = __alloc_pages_nodemask(gfp, order,
2050 policy_zonelist(gfp, pol, node),
2051 policy_nodemask(gfp, pol));
2052 if (unlikely(mpol_needs_cond_ref(pol)))
2054 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2060 * alloc_pages_current - Allocate pages.
2063 * %GFP_USER user allocation,
2064 * %GFP_KERNEL kernel allocation,
2065 * %GFP_HIGHMEM highmem allocation,
2066 * %GFP_FS don't call back into a file system.
2067 * %GFP_ATOMIC don't sleep.
2068 * @order: Power of two of allocation size in pages. 0 is a single page.
2070 * Allocate a page from the kernel page pool. When not in
2071 * interrupt context and apply the current process NUMA policy.
2072 * Returns NULL when no page can be allocated.
2074 * Don't call cpuset_update_task_memory_state() unless
2075 * 1) it's ok to take cpuset_sem (can WAIT), and
2076 * 2) allocating for current task (not interrupt).
2078 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2080 struct mempolicy *pol = get_task_policy(current);
2082 unsigned int cpuset_mems_cookie;
2084 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2085 pol = &default_policy;
2088 cpuset_mems_cookie = read_mems_allowed_begin();
2091 * No reference counting needed for current->mempolicy
2092 * nor system default_policy
2094 if (pol->mode == MPOL_INTERLEAVE)
2095 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2097 page = __alloc_pages_nodemask(gfp, order,
2098 policy_zonelist(gfp, pol, numa_node_id()),
2099 policy_nodemask(gfp, pol));
2101 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2106 EXPORT_SYMBOL(alloc_pages_current);
2108 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2110 struct mempolicy *pol = mpol_dup(vma_policy(src));
2113 return PTR_ERR(pol);
2114 dst->vm_policy = pol;
2119 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2120 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2121 * with the mems_allowed returned by cpuset_mems_allowed(). This
2122 * keeps mempolicies cpuset relative after its cpuset moves. See
2123 * further kernel/cpuset.c update_nodemask().
2125 * current's mempolicy may be rebinded by the other task(the task that changes
2126 * cpuset's mems), so we needn't do rebind work for current task.
2129 /* Slow path of a mempolicy duplicate */
2130 struct mempolicy *__mpol_dup(struct mempolicy *old)
2132 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2135 return ERR_PTR(-ENOMEM);
2137 /* task's mempolicy is protected by alloc_lock */
2138 if (old == current->mempolicy) {
2141 task_unlock(current);
2146 if (current_cpuset_is_being_rebound()) {
2147 nodemask_t mems = cpuset_mems_allowed(current);
2148 if (new->flags & MPOL_F_REBINDING)
2149 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2151 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2154 atomic_set(&new->refcnt, 1);
2158 /* Slow path of a mempolicy comparison */
2159 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2163 if (a->mode != b->mode)
2165 if (a->flags != b->flags)
2167 if (mpol_store_user_nodemask(a))
2168 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2174 case MPOL_INTERLEAVE:
2175 return !!nodes_equal(a->v.nodes, b->v.nodes);
2176 case MPOL_PREFERRED:
2177 return a->v.preferred_node == b->v.preferred_node;
2185 * Shared memory backing store policy support.
2187 * Remember policies even when nobody has shared memory mapped.
2188 * The policies are kept in Red-Black tree linked from the inode.
2189 * They are protected by the sp->lock spinlock, which should be held
2190 * for any accesses to the tree.
2193 /* lookup first element intersecting start-end */
2194 /* Caller holds sp->lock */
2195 static struct sp_node *
2196 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2198 struct rb_node *n = sp->root.rb_node;
2201 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2203 if (start >= p->end)
2205 else if (end <= p->start)
2213 struct sp_node *w = NULL;
2214 struct rb_node *prev = rb_prev(n);
2217 w = rb_entry(prev, struct sp_node, nd);
2218 if (w->end <= start)
2222 return rb_entry(n, struct sp_node, nd);
2225 /* Insert a new shared policy into the list. */
2226 /* Caller holds sp->lock */
2227 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2229 struct rb_node **p = &sp->root.rb_node;
2230 struct rb_node *parent = NULL;
2235 nd = rb_entry(parent, struct sp_node, nd);
2236 if (new->start < nd->start)
2238 else if (new->end > nd->end)
2239 p = &(*p)->rb_right;
2243 rb_link_node(&new->nd, parent, p);
2244 rb_insert_color(&new->nd, &sp->root);
2245 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2246 new->policy ? new->policy->mode : 0);
2249 /* Find shared policy intersecting idx */
2251 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2253 struct mempolicy *pol = NULL;
2256 if (!sp->root.rb_node)
2258 spin_lock(&sp->lock);
2259 sn = sp_lookup(sp, idx, idx+1);
2261 mpol_get(sn->policy);
2264 spin_unlock(&sp->lock);
2268 static void sp_free(struct sp_node *n)
2270 mpol_put(n->policy);
2271 kmem_cache_free(sn_cache, n);
2275 * mpol_misplaced - check whether current page node is valid in policy
2277 * @page - page to be checked
2278 * @vma - vm area where page mapped
2279 * @addr - virtual address where page mapped
2281 * Lookup current policy node id for vma,addr and "compare to" page's
2285 * -1 - not misplaced, page is in the right node
2286 * node - node id where the page should be
2288 * Policy determination "mimics" alloc_page_vma().
2289 * Called from fault path where we know the vma and faulting address.
2291 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2293 struct mempolicy *pol;
2295 int curnid = page_to_nid(page);
2296 unsigned long pgoff;
2297 int thiscpu = raw_smp_processor_id();
2298 int thisnid = cpu_to_node(thiscpu);
2304 pol = get_vma_policy(current, vma, addr);
2305 if (!(pol->flags & MPOL_F_MOF))
2308 switch (pol->mode) {
2309 case MPOL_INTERLEAVE:
2310 BUG_ON(addr >= vma->vm_end);
2311 BUG_ON(addr < vma->vm_start);
2313 pgoff = vma->vm_pgoff;
2314 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2315 polnid = offset_il_node(pol, vma, pgoff);
2318 case MPOL_PREFERRED:
2319 if (pol->flags & MPOL_F_LOCAL)
2320 polnid = numa_node_id();
2322 polnid = pol->v.preferred_node;
2327 * allows binding to multiple nodes.
2328 * use current page if in policy nodemask,
2329 * else select nearest allowed node, if any.
2330 * If no allowed nodes, use current [!misplaced].
2332 if (node_isset(curnid, pol->v.nodes))
2334 (void)first_zones_zonelist(
2335 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2336 gfp_zone(GFP_HIGHUSER),
2337 &pol->v.nodes, &zone);
2338 polnid = zone->node;
2345 /* Migrate the page towards the node whose CPU is referencing it */
2346 if (pol->flags & MPOL_F_MORON) {
2349 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2353 if (curnid != polnid)
2361 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2363 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2364 rb_erase(&n->nd, &sp->root);
2368 static void sp_node_init(struct sp_node *node, unsigned long start,
2369 unsigned long end, struct mempolicy *pol)
2371 node->start = start;
2376 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2377 struct mempolicy *pol)
2380 struct mempolicy *newpol;
2382 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2386 newpol = mpol_dup(pol);
2387 if (IS_ERR(newpol)) {
2388 kmem_cache_free(sn_cache, n);
2391 newpol->flags |= MPOL_F_SHARED;
2392 sp_node_init(n, start, end, newpol);
2397 /* Replace a policy range. */
2398 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2399 unsigned long end, struct sp_node *new)
2402 struct sp_node *n_new = NULL;
2403 struct mempolicy *mpol_new = NULL;
2407 spin_lock(&sp->lock);
2408 n = sp_lookup(sp, start, end);
2409 /* Take care of old policies in the same range. */
2410 while (n && n->start < end) {
2411 struct rb_node *next = rb_next(&n->nd);
2412 if (n->start >= start) {
2418 /* Old policy spanning whole new range. */
2423 *mpol_new = *n->policy;
2424 atomic_set(&mpol_new->refcnt, 1);
2425 sp_node_init(n_new, end, n->end, mpol_new);
2427 sp_insert(sp, n_new);
2436 n = rb_entry(next, struct sp_node, nd);
2440 spin_unlock(&sp->lock);
2447 kmem_cache_free(sn_cache, n_new);
2452 spin_unlock(&sp->lock);
2454 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2457 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2464 * mpol_shared_policy_init - initialize shared policy for inode
2465 * @sp: pointer to inode shared policy
2466 * @mpol: struct mempolicy to install
2468 * Install non-NULL @mpol in inode's shared policy rb-tree.
2469 * On entry, the current task has a reference on a non-NULL @mpol.
2470 * This must be released on exit.
2471 * This is called at get_inode() calls and we can use GFP_KERNEL.
2473 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2477 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2478 spin_lock_init(&sp->lock);
2481 struct vm_area_struct pvma;
2482 struct mempolicy *new;
2483 NODEMASK_SCRATCH(scratch);
2487 /* contextualize the tmpfs mount point mempolicy */
2488 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2490 goto free_scratch; /* no valid nodemask intersection */
2493 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2494 task_unlock(current);
2498 /* Create pseudo-vma that contains just the policy */
2499 memset(&pvma, 0, sizeof(struct vm_area_struct));
2500 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2501 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2504 mpol_put(new); /* drop initial ref */
2506 NODEMASK_SCRATCH_FREE(scratch);
2508 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2512 int mpol_set_shared_policy(struct shared_policy *info,
2513 struct vm_area_struct *vma, struct mempolicy *npol)
2516 struct sp_node *new = NULL;
2517 unsigned long sz = vma_pages(vma);
2519 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2521 sz, npol ? npol->mode : -1,
2522 npol ? npol->flags : -1,
2523 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2526 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2530 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2536 /* Free a backing policy store on inode delete. */
2537 void mpol_free_shared_policy(struct shared_policy *p)
2540 struct rb_node *next;
2542 if (!p->root.rb_node)
2544 spin_lock(&p->lock);
2545 next = rb_first(&p->root);
2547 n = rb_entry(next, struct sp_node, nd);
2548 next = rb_next(&n->nd);
2551 spin_unlock(&p->lock);
2554 #ifdef CONFIG_NUMA_BALANCING
2555 static int __initdata numabalancing_override;
2557 static void __init check_numabalancing_enable(void)
2559 bool numabalancing_default = false;
2561 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2562 numabalancing_default = true;
2564 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2565 if (numabalancing_override)
2566 set_numabalancing_state(numabalancing_override == 1);
2568 if (nr_node_ids > 1 && !numabalancing_override) {
2569 pr_info("%s automatic NUMA balancing. "
2570 "Configure with numa_balancing= or the "
2571 "kernel.numa_balancing sysctl",
2572 numabalancing_default ? "Enabling" : "Disabling");
2573 set_numabalancing_state(numabalancing_default);
2577 static int __init setup_numabalancing(char *str)
2583 if (!strcmp(str, "enable")) {
2584 numabalancing_override = 1;
2586 } else if (!strcmp(str, "disable")) {
2587 numabalancing_override = -1;
2592 pr_warn("Unable to parse numa_balancing=\n");
2596 __setup("numa_balancing=", setup_numabalancing);
2598 static inline void __init check_numabalancing_enable(void)
2601 #endif /* CONFIG_NUMA_BALANCING */
2603 /* assumes fs == KERNEL_DS */
2604 void __init numa_policy_init(void)
2606 nodemask_t interleave_nodes;
2607 unsigned long largest = 0;
2608 int nid, prefer = 0;
2610 policy_cache = kmem_cache_create("numa_policy",
2611 sizeof(struct mempolicy),
2612 0, SLAB_PANIC, NULL);
2614 sn_cache = kmem_cache_create("shared_policy_node",
2615 sizeof(struct sp_node),
2616 0, SLAB_PANIC, NULL);
2618 for_each_node(nid) {
2619 preferred_node_policy[nid] = (struct mempolicy) {
2620 .refcnt = ATOMIC_INIT(1),
2621 .mode = MPOL_PREFERRED,
2622 .flags = MPOL_F_MOF | MPOL_F_MORON,
2623 .v = { .preferred_node = nid, },
2628 * Set interleaving policy for system init. Interleaving is only
2629 * enabled across suitably sized nodes (default is >= 16MB), or
2630 * fall back to the largest node if they're all smaller.
2632 nodes_clear(interleave_nodes);
2633 for_each_node_state(nid, N_MEMORY) {
2634 unsigned long total_pages = node_present_pages(nid);
2636 /* Preserve the largest node */
2637 if (largest < total_pages) {
2638 largest = total_pages;
2642 /* Interleave this node? */
2643 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2644 node_set(nid, interleave_nodes);
2647 /* All too small, use the largest */
2648 if (unlikely(nodes_empty(interleave_nodes)))
2649 node_set(prefer, interleave_nodes);
2651 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2652 printk("numa_policy_init: interleaving failed\n");
2654 check_numabalancing_enable();
2657 /* Reset policy of current process to default */
2658 void numa_default_policy(void)
2660 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2664 * Parse and format mempolicy from/to strings
2668 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2670 static const char * const policy_modes[] =
2672 [MPOL_DEFAULT] = "default",
2673 [MPOL_PREFERRED] = "prefer",
2674 [MPOL_BIND] = "bind",
2675 [MPOL_INTERLEAVE] = "interleave",
2676 [MPOL_LOCAL] = "local",
2682 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2683 * @str: string containing mempolicy to parse
2684 * @mpol: pointer to struct mempolicy pointer, returned on success.
2687 * <mode>[=<flags>][:<nodelist>]
2689 * On success, returns 0, else 1
2691 int mpol_parse_str(char *str, struct mempolicy **mpol)
2693 struct mempolicy *new = NULL;
2694 unsigned short mode;
2695 unsigned short mode_flags;
2697 char *nodelist = strchr(str, ':');
2698 char *flags = strchr(str, '=');
2702 /* NUL-terminate mode or flags string */
2704 if (nodelist_parse(nodelist, nodes))
2706 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2712 *flags++ = '\0'; /* terminate mode string */
2714 for (mode = 0; mode < MPOL_MAX; mode++) {
2715 if (!strcmp(str, policy_modes[mode])) {
2719 if (mode >= MPOL_MAX)
2723 case MPOL_PREFERRED:
2725 * Insist on a nodelist of one node only
2728 char *rest = nodelist;
2729 while (isdigit(*rest))
2735 case MPOL_INTERLEAVE:
2737 * Default to online nodes with memory if no nodelist
2740 nodes = node_states[N_MEMORY];
2744 * Don't allow a nodelist; mpol_new() checks flags
2748 mode = MPOL_PREFERRED;
2752 * Insist on a empty nodelist
2759 * Insist on a nodelist
2768 * Currently, we only support two mutually exclusive
2771 if (!strcmp(flags, "static"))
2772 mode_flags |= MPOL_F_STATIC_NODES;
2773 else if (!strcmp(flags, "relative"))
2774 mode_flags |= MPOL_F_RELATIVE_NODES;
2779 new = mpol_new(mode, mode_flags, &nodes);
2784 * Save nodes for mpol_to_str() to show the tmpfs mount options
2785 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2787 if (mode != MPOL_PREFERRED)
2788 new->v.nodes = nodes;
2790 new->v.preferred_node = first_node(nodes);
2792 new->flags |= MPOL_F_LOCAL;
2795 * Save nodes for contextualization: this will be used to "clone"
2796 * the mempolicy in a specific context [cpuset] at a later time.
2798 new->w.user_nodemask = nodes;
2803 /* Restore string for error message */
2812 #endif /* CONFIG_TMPFS */
2815 * mpol_to_str - format a mempolicy structure for printing
2816 * @buffer: to contain formatted mempolicy string
2817 * @maxlen: length of @buffer
2818 * @pol: pointer to mempolicy to be formatted
2820 * Convert @pol into a string. If @buffer is too short, truncate the string.
2821 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2822 * longest flag, "relative", and to display at least a few node ids.
2824 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2827 nodemask_t nodes = NODE_MASK_NONE;
2828 unsigned short mode = MPOL_DEFAULT;
2829 unsigned short flags = 0;
2831 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2839 case MPOL_PREFERRED:
2840 if (flags & MPOL_F_LOCAL)
2843 node_set(pol->v.preferred_node, nodes);
2846 case MPOL_INTERLEAVE:
2847 nodes = pol->v.nodes;
2851 snprintf(p, maxlen, "unknown");
2855 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2857 if (flags & MPOL_MODE_FLAGS) {
2858 p += snprintf(p, buffer + maxlen - p, "=");
2861 * Currently, the only defined flags are mutually exclusive
2863 if (flags & MPOL_F_STATIC_NODES)
2864 p += snprintf(p, buffer + maxlen - p, "static");
2865 else if (flags & MPOL_F_RELATIVE_NODES)
2866 p += snprintf(p, buffer + maxlen - p, "relative");
2869 if (!nodes_empty(nodes)) {
2870 p += snprintf(p, buffer + maxlen - p, ":");
2871 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);