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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/export.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/ksm.h>
90 #include <linux/rmap.h>
91 #include <linux/security.h>
92 #include <linux/syscalls.h>
93 #include <linux/ctype.h>
94 #include <linux/mm_inline.h>
95 #include <linux/mmu_notifier.h>
96 #include <linux/printk.h>
98 #include <asm/tlbflush.h>
99 #include <asm/uaccess.h>
100 #include <linux/random.h>
102 #include "internal.h"
105 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
106 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
108 static struct kmem_cache *policy_cache;
109 static struct kmem_cache *sn_cache;
111 /* Highest zone. An specific allocation for a zone below that is not
113 enum zone_type policy_zone = 0;
116 * run-time system-wide default policy => local allocation
118 static struct mempolicy default_policy = {
119 .refcnt = ATOMIC_INIT(1), /* never free it */
120 .mode = MPOL_PREFERRED,
121 .flags = MPOL_F_LOCAL,
124 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
126 struct mempolicy *get_task_policy(struct task_struct *p)
128 struct mempolicy *pol = p->mempolicy;
134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
142 return &default_policy;
145 static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
153 * If we have a lock to protect task->mempolicy in read-side, we do
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
163 } mpol_ops[MPOL_MAX];
165 /* Check that the nodemask contains at least one populated zone */
166 static int is_valid_nodemask(const nodemask_t *nodemask)
168 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
171 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
173 return pol->flags & MPOL_MODE_FLAGS;
176 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
177 const nodemask_t *rel)
180 nodes_fold(tmp, *orig, nodes_weight(*rel));
181 nodes_onto(*ret, tmp, *rel);
184 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
186 if (nodes_empty(*nodes))
188 pol->v.nodes = *nodes;
192 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
195 pol->flags |= MPOL_F_LOCAL; /* local allocation */
196 else if (nodes_empty(*nodes))
197 return -EINVAL; /* no allowed nodes */
199 pol->v.preferred_node = first_node(*nodes);
203 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
205 if (!is_valid_nodemask(nodes))
207 pol->v.nodes = *nodes;
212 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
213 * any, for the new policy. mpol_new() has already validated the nodes
214 * parameter with respect to the policy mode and flags. But, we need to
215 * handle an empty nodemask with MPOL_PREFERRED here.
217 * Must be called holding task's alloc_lock to protect task's mems_allowed
218 * and mempolicy. May also be called holding the mmap_semaphore for write.
220 static int mpol_set_nodemask(struct mempolicy *pol,
221 const nodemask_t *nodes, struct nodemask_scratch *nsc)
225 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
229 nodes_and(nsc->mask1,
230 cpuset_current_mems_allowed, node_states[N_MEMORY]);
233 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
234 nodes = NULL; /* explicit local allocation */
236 if (pol->flags & MPOL_F_RELATIVE_NODES)
237 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
239 nodes_and(nsc->mask2, *nodes, nsc->mask1);
241 if (mpol_store_user_nodemask(pol))
242 pol->w.user_nodemask = *nodes;
244 pol->w.cpuset_mems_allowed =
245 cpuset_current_mems_allowed;
249 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
251 ret = mpol_ops[pol->mode].create(pol, NULL);
256 * This function just creates a new policy, does some check and simple
257 * initialization. You must invoke mpol_set_nodemask() to set nodes.
259 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
262 struct mempolicy *policy;
264 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
265 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
267 if (mode == MPOL_DEFAULT) {
268 if (nodes && !nodes_empty(*nodes))
269 return ERR_PTR(-EINVAL);
275 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
276 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
277 * All other modes require a valid pointer to a non-empty nodemask.
279 if (mode == MPOL_PREFERRED) {
280 if (nodes_empty(*nodes)) {
281 if (((flags & MPOL_F_STATIC_NODES) ||
282 (flags & MPOL_F_RELATIVE_NODES)))
283 return ERR_PTR(-EINVAL);
285 } else if (mode == MPOL_LOCAL) {
286 if (!nodes_empty(*nodes))
287 return ERR_PTR(-EINVAL);
288 mode = MPOL_PREFERRED;
289 } else if (nodes_empty(*nodes))
290 return ERR_PTR(-EINVAL);
291 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
293 return ERR_PTR(-ENOMEM);
294 atomic_set(&policy->refcnt, 1);
296 policy->flags = flags;
301 /* Slow path of a mpol destructor. */
302 void __mpol_put(struct mempolicy *p)
304 if (!atomic_dec_and_test(&p->refcnt))
306 kmem_cache_free(policy_cache, p);
309 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
310 enum mpol_rebind_step step)
316 * MPOL_REBIND_ONCE - do rebind work at once
317 * MPOL_REBIND_STEP1 - set all the newly nodes
318 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
320 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
321 enum mpol_rebind_step step)
325 if (pol->flags & MPOL_F_STATIC_NODES)
326 nodes_and(tmp, pol->w.user_nodemask, *nodes);
327 else if (pol->flags & MPOL_F_RELATIVE_NODES)
328 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
331 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
334 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
335 nodes_remap(tmp, pol->v.nodes,
336 pol->w.cpuset_mems_allowed, *nodes);
337 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
338 } else if (step == MPOL_REBIND_STEP2) {
339 tmp = pol->w.cpuset_mems_allowed;
340 pol->w.cpuset_mems_allowed = *nodes;
345 if (nodes_empty(tmp))
348 if (step == MPOL_REBIND_STEP1)
349 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
350 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
355 if (!node_isset(current->il_next, tmp)) {
356 current->il_next = next_node(current->il_next, tmp);
357 if (current->il_next >= MAX_NUMNODES)
358 current->il_next = first_node(tmp);
359 if (current->il_next >= MAX_NUMNODES)
360 current->il_next = numa_node_id();
364 static void mpol_rebind_preferred(struct mempolicy *pol,
365 const nodemask_t *nodes,
366 enum mpol_rebind_step step)
370 if (pol->flags & MPOL_F_STATIC_NODES) {
371 int node = first_node(pol->w.user_nodemask);
373 if (node_isset(node, *nodes)) {
374 pol->v.preferred_node = node;
375 pol->flags &= ~MPOL_F_LOCAL;
377 pol->flags |= MPOL_F_LOCAL;
378 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
379 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
380 pol->v.preferred_node = first_node(tmp);
381 } else if (!(pol->flags & MPOL_F_LOCAL)) {
382 pol->v.preferred_node = node_remap(pol->v.preferred_node,
383 pol->w.cpuset_mems_allowed,
385 pol->w.cpuset_mems_allowed = *nodes;
390 * mpol_rebind_policy - Migrate a policy to a different set of nodes
392 * If read-side task has no lock to protect task->mempolicy, write-side
393 * task will rebind the task->mempolicy by two step. The first step is
394 * setting all the newly nodes, and the second step is cleaning all the
395 * disallowed nodes. In this way, we can avoid finding no node to alloc
397 * If we have a lock to protect task->mempolicy in read-side, we do
401 * MPOL_REBIND_ONCE - do rebind work at once
402 * MPOL_REBIND_STEP1 - set all the newly nodes
403 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
405 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
406 enum mpol_rebind_step step)
410 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
411 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
414 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
417 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
420 if (step == MPOL_REBIND_STEP1)
421 pol->flags |= MPOL_F_REBINDING;
422 else if (step == MPOL_REBIND_STEP2)
423 pol->flags &= ~MPOL_F_REBINDING;
424 else if (step >= MPOL_REBIND_NSTEP)
427 mpol_ops[pol->mode].rebind(pol, newmask, step);
431 * Wrapper for mpol_rebind_policy() that just requires task
432 * pointer, and updates task mempolicy.
434 * Called with task's alloc_lock held.
437 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
438 enum mpol_rebind_step step)
440 mpol_rebind_policy(tsk->mempolicy, new, step);
444 * Rebind each vma in mm to new nodemask.
446 * Call holding a reference to mm. Takes mm->mmap_sem during call.
449 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
451 struct vm_area_struct *vma;
453 down_write(&mm->mmap_sem);
454 for (vma = mm->mmap; vma; vma = vma->vm_next)
455 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
456 up_write(&mm->mmap_sem);
459 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
461 .rebind = mpol_rebind_default,
463 [MPOL_INTERLEAVE] = {
464 .create = mpol_new_interleave,
465 .rebind = mpol_rebind_nodemask,
468 .create = mpol_new_preferred,
469 .rebind = mpol_rebind_preferred,
472 .create = mpol_new_bind,
473 .rebind = mpol_rebind_nodemask,
477 static void migrate_page_add(struct page *page, struct list_head *pagelist,
478 unsigned long flags);
481 * Scan through pages checking if pages follow certain conditions,
482 * and move them to the pagelist if they do.
484 static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
485 unsigned long addr, unsigned long end,
486 const nodemask_t *nodes, unsigned long flags,
493 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
498 if (!pte_present(*pte))
500 page = vm_normal_page(vma, addr, *pte);
504 * vm_normal_page() filters out zero pages, but there might
505 * still be PageReserved pages to skip, perhaps in a VDSO.
507 if (PageReserved(page))
509 nid = page_to_nid(page);
510 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
513 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
514 migrate_page_add(page, private, flags);
517 } while (pte++, addr += PAGE_SIZE, addr != end);
518 pte_unmap_unlock(orig_pte, ptl);
522 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
523 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
526 #ifdef CONFIG_HUGETLB_PAGE
532 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
533 entry = huge_ptep_get((pte_t *)pmd);
534 if (!pte_present(entry))
536 page = pte_page(entry);
537 nid = page_to_nid(page);
538 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
540 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
541 if (flags & (MPOL_MF_MOVE_ALL) ||
542 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
543 isolate_huge_page(page, private);
551 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
552 unsigned long addr, unsigned long end,
553 const nodemask_t *nodes, unsigned long flags,
559 pmd = pmd_offset(pud, addr);
561 next = pmd_addr_end(addr, end);
562 if (!pmd_present(*pmd))
564 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
565 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
569 split_huge_page_pmd(vma, addr, pmd);
570 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
572 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
575 } while (pmd++, addr = next, addr != end);
579 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
580 unsigned long addr, unsigned long end,
581 const nodemask_t *nodes, unsigned long flags,
587 pud = pud_offset(pgd, addr);
589 next = pud_addr_end(addr, end);
590 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
592 if (pud_none_or_clear_bad(pud))
594 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
597 } while (pud++, addr = next, addr != end);
601 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
602 unsigned long addr, unsigned long end,
603 const nodemask_t *nodes, unsigned long flags,
609 pgd = pgd_offset(vma->vm_mm, addr);
611 next = pgd_addr_end(addr, end);
612 if (pgd_none_or_clear_bad(pgd))
614 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
617 } while (pgd++, addr = next, addr != end);
621 #ifdef CONFIG_NUMA_BALANCING
623 * This is used to mark a range of virtual addresses to be inaccessible.
624 * These are later cleared by a NUMA hinting fault. Depending on these
625 * faults, pages may be migrated for better NUMA placement.
627 * This is assuming that NUMA faults are handled using PROT_NONE. If
628 * an architecture makes a different choice, it will need further
629 * changes to the core.
631 unsigned long change_prot_numa(struct vm_area_struct *vma,
632 unsigned long addr, unsigned long end)
636 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
638 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
643 static unsigned long change_prot_numa(struct vm_area_struct *vma,
644 unsigned long addr, unsigned long end)
648 #endif /* CONFIG_NUMA_BALANCING */
651 * Walk through page tables and collect pages to be migrated.
653 * If pages found in a given range are on a set of nodes (determined by
654 * @nodes and @flags,) it's isolated and queued to the pagelist which is
655 * passed via @private.)
658 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
659 const nodemask_t *nodes, unsigned long flags, void *private)
662 struct vm_area_struct *vma, *prev;
664 vma = find_vma(mm, start);
668 for (; 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)
679 if (prev && prev->vm_end < vma->vm_start)
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,
704 * Apply policy to a single VMA
705 * This must be called with the mmap_sem held for writing.
707 static int vma_replace_policy(struct vm_area_struct *vma,
708 struct mempolicy *pol)
711 struct mempolicy *old;
712 struct mempolicy *new;
714 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
715 vma->vm_start, vma->vm_end, vma->vm_pgoff,
716 vma->vm_ops, vma->vm_file,
717 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
723 if (vma->vm_ops && vma->vm_ops->set_policy) {
724 err = vma->vm_ops->set_policy(vma, new);
729 old = vma->vm_policy;
730 vma->vm_policy = new; /* protected by mmap_sem */
739 /* Step 2: apply policy to a range and do splits. */
740 static int mbind_range(struct mm_struct *mm, unsigned long start,
741 unsigned long end, struct mempolicy *new_pol)
743 struct vm_area_struct *next;
744 struct vm_area_struct *prev;
745 struct vm_area_struct *vma;
748 unsigned long vmstart;
751 vma = find_vma(mm, start);
752 if (!vma || vma->vm_start > start)
756 if (start > vma->vm_start)
759 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
761 vmstart = max(start, vma->vm_start);
762 vmend = min(end, vma->vm_end);
764 if (mpol_equal(vma_policy(vma), new_pol))
767 pgoff = vma->vm_pgoff +
768 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
769 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
770 vma->anon_vma, vma->vm_file, pgoff,
775 if (mpol_equal(vma_policy(vma), new_pol))
777 /* vma_merge() joined vma && vma->next, case 8 */
780 if (vma->vm_start != vmstart) {
781 err = split_vma(vma->vm_mm, vma, vmstart, 1);
785 if (vma->vm_end != vmend) {
786 err = split_vma(vma->vm_mm, vma, vmend, 0);
791 err = vma_replace_policy(vma, new_pol);
800 /* Set the process memory policy */
801 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
804 struct mempolicy *new, *old;
805 NODEMASK_SCRATCH(scratch);
811 new = mpol_new(mode, flags, nodes);
818 ret = mpol_set_nodemask(new, nodes, scratch);
820 task_unlock(current);
824 old = current->mempolicy;
825 current->mempolicy = new;
826 if (new && new->mode == MPOL_INTERLEAVE &&
827 nodes_weight(new->v.nodes))
828 current->il_next = first_node(new->v.nodes);
829 task_unlock(current);
833 NODEMASK_SCRATCH_FREE(scratch);
838 * Return nodemask for policy for get_mempolicy() query
840 * Called with task's alloc_lock held
842 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
845 if (p == &default_policy)
851 case MPOL_INTERLEAVE:
855 if (!(p->flags & MPOL_F_LOCAL))
856 node_set(p->v.preferred_node, *nodes);
857 /* else return empty node mask for local allocation */
864 static int lookup_node(struct mm_struct *mm, unsigned long addr)
869 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
871 err = page_to_nid(p);
877 /* Retrieve NUMA policy */
878 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
879 unsigned long addr, unsigned long flags)
882 struct mm_struct *mm = current->mm;
883 struct vm_area_struct *vma = NULL;
884 struct mempolicy *pol = current->mempolicy;
887 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
890 if (flags & MPOL_F_MEMS_ALLOWED) {
891 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
893 *policy = 0; /* just so it's initialized */
895 *nmask = cpuset_current_mems_allowed;
896 task_unlock(current);
900 if (flags & MPOL_F_ADDR) {
902 * Do NOT fall back to task policy if the
903 * vma/shared policy at addr is NULL. We
904 * want to return MPOL_DEFAULT in this case.
906 down_read(&mm->mmap_sem);
907 vma = find_vma_intersection(mm, addr, addr+1);
909 up_read(&mm->mmap_sem);
912 if (vma->vm_ops && vma->vm_ops->get_policy)
913 pol = vma->vm_ops->get_policy(vma, addr);
915 pol = vma->vm_policy;
920 pol = &default_policy; /* indicates default behavior */
922 if (flags & MPOL_F_NODE) {
923 if (flags & MPOL_F_ADDR) {
924 err = lookup_node(mm, addr);
928 } else if (pol == current->mempolicy &&
929 pol->mode == MPOL_INTERLEAVE) {
930 *policy = current->il_next;
936 *policy = pol == &default_policy ? MPOL_DEFAULT :
939 * Internal mempolicy flags must be masked off before exposing
940 * the policy to userspace.
942 *policy |= (pol->flags & MPOL_MODE_FLAGS);
946 up_read(¤t->mm->mmap_sem);
952 if (mpol_store_user_nodemask(pol)) {
953 *nmask = pol->w.user_nodemask;
956 get_policy_nodemask(pol, nmask);
957 task_unlock(current);
964 up_read(¤t->mm->mmap_sem);
968 #ifdef CONFIG_MIGRATION
972 static void migrate_page_add(struct page *page, struct list_head *pagelist,
976 * Avoid migrating a page that is shared with others.
978 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
979 if (!isolate_lru_page(page)) {
980 list_add_tail(&page->lru, pagelist);
981 inc_zone_page_state(page, NR_ISOLATED_ANON +
982 page_is_file_cache(page));
987 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
990 return alloc_huge_page_node(page_hstate(compound_head(page)),
993 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
997 * Migrate pages from one node to a target node.
998 * Returns error or the number of pages not migrated.
1000 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1004 LIST_HEAD(pagelist);
1008 node_set(source, nmask);
1011 * This does not "check" the range but isolates all pages that
1012 * need migration. Between passing in the full user address
1013 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1015 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1016 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1017 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1019 if (!list_empty(&pagelist)) {
1020 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1021 MIGRATE_SYNC, MR_SYSCALL);
1023 putback_movable_pages(&pagelist);
1030 * Move pages between the two nodesets so as to preserve the physical
1031 * layout as much as possible.
1033 * Returns the number of page that could not be moved.
1035 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1036 const nodemask_t *to, int flags)
1042 err = migrate_prep();
1046 down_read(&mm->mmap_sem);
1048 err = migrate_vmas(mm, from, to, flags);
1053 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1054 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1055 * bit in 'tmp', and return that <source, dest> pair for migration.
1056 * The pair of nodemasks 'to' and 'from' define the map.
1058 * If no pair of bits is found that way, fallback to picking some
1059 * pair of 'source' and 'dest' bits that are not the same. If the
1060 * 'source' and 'dest' bits are the same, this represents a node
1061 * that will be migrating to itself, so no pages need move.
1063 * If no bits are left in 'tmp', or if all remaining bits left
1064 * in 'tmp' correspond to the same bit in 'to', return false
1065 * (nothing left to migrate).
1067 * This lets us pick a pair of nodes to migrate between, such that
1068 * if possible the dest node is not already occupied by some other
1069 * source node, minimizing the risk of overloading the memory on a
1070 * node that would happen if we migrated incoming memory to a node
1071 * before migrating outgoing memory source that same node.
1073 * A single scan of tmp is sufficient. As we go, we remember the
1074 * most recent <s, d> pair that moved (s != d). If we find a pair
1075 * that not only moved, but what's better, moved to an empty slot
1076 * (d is not set in tmp), then we break out then, with that pair.
1077 * Otherwise when we finish scanning from_tmp, we at least have the
1078 * most recent <s, d> pair that moved. If we get all the way through
1079 * the scan of tmp without finding any node that moved, much less
1080 * moved to an empty node, then there is nothing left worth migrating.
1084 while (!nodes_empty(tmp)) {
1086 int source = NUMA_NO_NODE;
1089 for_each_node_mask(s, tmp) {
1092 * do_migrate_pages() tries to maintain the relative
1093 * node relationship of the pages established between
1094 * threads and memory areas.
1096 * However if the number of source nodes is not equal to
1097 * the number of destination nodes we can not preserve
1098 * this node relative relationship. In that case, skip
1099 * copying memory from a node that is in the destination
1102 * Example: [2,3,4] -> [3,4,5] moves everything.
1103 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1106 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1107 (node_isset(s, *to)))
1110 d = node_remap(s, *from, *to);
1114 source = s; /* Node moved. Memorize */
1117 /* dest not in remaining from nodes? */
1118 if (!node_isset(dest, tmp))
1121 if (source == NUMA_NO_NODE)
1124 node_clear(source, tmp);
1125 err = migrate_to_node(mm, source, dest, flags);
1132 up_read(&mm->mmap_sem);
1140 * Allocate a new page for page migration based on vma policy.
1141 * Start by assuming the page is mapped by the same vma as contains @start.
1142 * Search forward from there, if not. N.B., this assumes that the
1143 * list of pages handed to migrate_pages()--which is how we get here--
1144 * is in virtual address order.
1146 static struct page *new_page(struct page *page, unsigned long start, int **x)
1148 struct vm_area_struct *vma;
1149 unsigned long uninitialized_var(address);
1151 vma = find_vma(current->mm, start);
1153 address = page_address_in_vma(page, vma);
1154 if (address != -EFAULT)
1159 if (PageHuge(page)) {
1161 return alloc_huge_page_noerr(vma, address, 1);
1164 * if !vma, alloc_page_vma() will use task or system default policy
1166 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1170 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1171 unsigned long flags)
1175 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1176 const nodemask_t *to, int flags)
1181 static struct page *new_page(struct page *page, unsigned long start, int **x)
1187 static long do_mbind(unsigned long start, unsigned long len,
1188 unsigned short mode, unsigned short mode_flags,
1189 nodemask_t *nmask, unsigned long flags)
1191 struct mm_struct *mm = current->mm;
1192 struct mempolicy *new;
1195 LIST_HEAD(pagelist);
1197 if (flags & ~(unsigned long)MPOL_MF_VALID)
1199 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1202 if (start & ~PAGE_MASK)
1205 if (mode == MPOL_DEFAULT)
1206 flags &= ~MPOL_MF_STRICT;
1208 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1216 new = mpol_new(mode, mode_flags, nmask);
1218 return PTR_ERR(new);
1220 if (flags & MPOL_MF_LAZY)
1221 new->flags |= MPOL_F_MOF;
1224 * If we are using the default policy then operation
1225 * on discontinuous address spaces is okay after all
1228 flags |= MPOL_MF_DISCONTIG_OK;
1230 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1231 start, start + len, mode, mode_flags,
1232 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1234 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1236 err = migrate_prep();
1241 NODEMASK_SCRATCH(scratch);
1243 down_write(&mm->mmap_sem);
1245 err = mpol_set_nodemask(new, nmask, scratch);
1246 task_unlock(current);
1248 up_write(&mm->mmap_sem);
1251 NODEMASK_SCRATCH_FREE(scratch);
1256 err = queue_pages_range(mm, start, end, nmask,
1257 flags | MPOL_MF_INVERT, &pagelist);
1259 err = mbind_range(mm, start, end, new);
1264 if (!list_empty(&pagelist)) {
1265 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1266 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1267 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1269 putback_movable_pages(&pagelist);
1272 if (nr_failed && (flags & MPOL_MF_STRICT))
1275 putback_movable_pages(&pagelist);
1277 up_write(&mm->mmap_sem);
1284 * User space interface with variable sized bitmaps for nodelists.
1287 /* Copy a node mask from user space. */
1288 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1289 unsigned long maxnode)
1292 unsigned long nlongs;
1293 unsigned long endmask;
1296 nodes_clear(*nodes);
1297 if (maxnode == 0 || !nmask)
1299 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1302 nlongs = BITS_TO_LONGS(maxnode);
1303 if ((maxnode % BITS_PER_LONG) == 0)
1306 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1308 /* When the user specified more nodes than supported just check
1309 if the non supported part is all zero. */
1310 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1311 if (nlongs > PAGE_SIZE/sizeof(long))
1313 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1315 if (get_user(t, nmask + k))
1317 if (k == nlongs - 1) {
1323 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1327 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1329 nodes_addr(*nodes)[nlongs-1] &= endmask;
1333 /* Copy a kernel node mask to user space */
1334 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1337 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1338 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1340 if (copy > nbytes) {
1341 if (copy > PAGE_SIZE)
1343 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1347 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1350 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1351 unsigned long, mode, const unsigned long __user *, nmask,
1352 unsigned long, maxnode, unsigned, flags)
1356 unsigned short mode_flags;
1358 mode_flags = mode & MPOL_MODE_FLAGS;
1359 mode &= ~MPOL_MODE_FLAGS;
1360 if (mode >= MPOL_MAX)
1362 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1363 (mode_flags & MPOL_F_RELATIVE_NODES))
1365 err = get_nodes(&nodes, nmask, maxnode);
1368 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1371 /* Set the process memory policy */
1372 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1373 unsigned long, maxnode)
1377 unsigned short flags;
1379 flags = mode & MPOL_MODE_FLAGS;
1380 mode &= ~MPOL_MODE_FLAGS;
1381 if ((unsigned int)mode >= MPOL_MAX)
1383 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1385 err = get_nodes(&nodes, nmask, maxnode);
1388 return do_set_mempolicy(mode, flags, &nodes);
1391 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1392 const unsigned long __user *, old_nodes,
1393 const unsigned long __user *, new_nodes)
1395 const struct cred *cred = current_cred(), *tcred;
1396 struct mm_struct *mm = NULL;
1397 struct task_struct *task;
1398 nodemask_t task_nodes;
1402 NODEMASK_SCRATCH(scratch);
1407 old = &scratch->mask1;
1408 new = &scratch->mask2;
1410 err = get_nodes(old, old_nodes, maxnode);
1414 err = get_nodes(new, new_nodes, maxnode);
1418 /* Find the mm_struct */
1420 task = pid ? find_task_by_vpid(pid) : current;
1426 get_task_struct(task);
1431 * Check if this process has the right to modify the specified
1432 * process. The right exists if the process has administrative
1433 * capabilities, superuser privileges or the same
1434 * userid as the target process.
1436 tcred = __task_cred(task);
1437 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1438 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1439 !capable(CAP_SYS_NICE)) {
1446 task_nodes = cpuset_mems_allowed(task);
1447 /* Is the user allowed to access the target nodes? */
1448 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1453 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1458 err = security_task_movememory(task);
1462 mm = get_task_mm(task);
1463 put_task_struct(task);
1470 err = do_migrate_pages(mm, old, new,
1471 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1475 NODEMASK_SCRATCH_FREE(scratch);
1480 put_task_struct(task);
1486 /* Retrieve NUMA policy */
1487 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1488 unsigned long __user *, nmask, unsigned long, maxnode,
1489 unsigned long, addr, unsigned long, flags)
1492 int uninitialized_var(pval);
1495 if (nmask != NULL && maxnode < MAX_NUMNODES)
1498 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1503 if (policy && put_user(pval, policy))
1507 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1512 #ifdef CONFIG_COMPAT
1514 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1515 compat_ulong_t __user *, nmask,
1516 compat_ulong_t, maxnode,
1517 compat_ulong_t, addr, compat_ulong_t, flags)
1520 unsigned long __user *nm = NULL;
1521 unsigned long nr_bits, alloc_size;
1522 DECLARE_BITMAP(bm, MAX_NUMNODES);
1524 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1525 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1528 nm = compat_alloc_user_space(alloc_size);
1530 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1532 if (!err && nmask) {
1533 unsigned long copy_size;
1534 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1535 err = copy_from_user(bm, nm, copy_size);
1536 /* ensure entire bitmap is zeroed */
1537 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1538 err |= compat_put_bitmap(nmask, bm, nr_bits);
1544 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1545 compat_ulong_t, maxnode)
1548 unsigned long __user *nm = NULL;
1549 unsigned long nr_bits, alloc_size;
1550 DECLARE_BITMAP(bm, MAX_NUMNODES);
1552 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1553 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1556 err = compat_get_bitmap(bm, nmask, nr_bits);
1557 nm = compat_alloc_user_space(alloc_size);
1558 err |= copy_to_user(nm, bm, alloc_size);
1564 return sys_set_mempolicy(mode, nm, nr_bits+1);
1567 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1568 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1569 compat_ulong_t, maxnode, compat_ulong_t, flags)
1572 unsigned long __user *nm = NULL;
1573 unsigned long nr_bits, alloc_size;
1576 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1577 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1580 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1581 nm = compat_alloc_user_space(alloc_size);
1582 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1588 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1593 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1596 struct mempolicy *pol = NULL;
1599 if (vma->vm_ops && vma->vm_ops->get_policy) {
1600 pol = vma->vm_ops->get_policy(vma, addr);
1601 } else if (vma->vm_policy) {
1602 pol = vma->vm_policy;
1605 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1606 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1607 * count on these policies which will be dropped by
1608 * mpol_cond_put() later
1610 if (mpol_needs_cond_ref(pol))
1619 * get_vma_policy(@vma, @addr)
1620 * @vma: virtual memory area whose policy is sought
1621 * @addr: address in @vma for shared policy lookup
1623 * Returns effective policy for a VMA at specified address.
1624 * Falls back to current->mempolicy or system default policy, as necessary.
1625 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1626 * count--added by the get_policy() vm_op, as appropriate--to protect against
1627 * freeing by another task. It is the caller's responsibility to free the
1628 * extra reference for shared policies.
1630 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1633 struct mempolicy *pol = __get_vma_policy(vma, addr);
1636 pol = get_task_policy(current);
1641 bool vma_policy_mof(struct vm_area_struct *vma)
1643 struct mempolicy *pol;
1645 if (vma->vm_ops && vma->vm_ops->get_policy) {
1648 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1649 if (pol && (pol->flags & MPOL_F_MOF))
1656 pol = vma->vm_policy;
1658 pol = get_task_policy(current);
1660 return pol->flags & MPOL_F_MOF;
1663 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1665 enum zone_type dynamic_policy_zone = policy_zone;
1667 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1670 * if policy->v.nodes has movable memory only,
1671 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1673 * policy->v.nodes is intersect with node_states[N_MEMORY].
1674 * so if the following test faile, it implies
1675 * policy->v.nodes has movable memory only.
1677 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1678 dynamic_policy_zone = ZONE_MOVABLE;
1680 return zone >= dynamic_policy_zone;
1684 * Return a nodemask representing a mempolicy for filtering nodes for
1687 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1689 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1690 if (unlikely(policy->mode == MPOL_BIND) &&
1691 apply_policy_zone(policy, gfp_zone(gfp)) &&
1692 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1693 return &policy->v.nodes;
1698 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1699 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1702 switch (policy->mode) {
1703 case MPOL_PREFERRED:
1704 if (!(policy->flags & MPOL_F_LOCAL))
1705 nd = policy->v.preferred_node;
1709 * Normally, MPOL_BIND allocations are node-local within the
1710 * allowed nodemask. However, if __GFP_THISNODE is set and the
1711 * current node isn't part of the mask, we use the zonelist for
1712 * the first node in the mask instead.
1714 if (unlikely(gfp & __GFP_THISNODE) &&
1715 unlikely(!node_isset(nd, policy->v.nodes)))
1716 nd = first_node(policy->v.nodes);
1721 return node_zonelist(nd, gfp);
1724 /* Do dynamic interleaving for a process */
1725 static unsigned interleave_nodes(struct mempolicy *policy)
1728 struct task_struct *me = current;
1731 next = next_node(nid, policy->v.nodes);
1732 if (next >= MAX_NUMNODES)
1733 next = first_node(policy->v.nodes);
1734 if (next < MAX_NUMNODES)
1740 * Depending on the memory policy provide a node from which to allocate the
1743 unsigned int mempolicy_slab_node(void)
1745 struct mempolicy *policy;
1746 int node = numa_mem_id();
1751 policy = current->mempolicy;
1752 if (!policy || policy->flags & MPOL_F_LOCAL)
1755 switch (policy->mode) {
1756 case MPOL_PREFERRED:
1758 * handled MPOL_F_LOCAL above
1760 return policy->v.preferred_node;
1762 case MPOL_INTERLEAVE:
1763 return interleave_nodes(policy);
1767 * Follow bind policy behavior and start allocation at the
1770 struct zonelist *zonelist;
1772 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1773 zonelist = &NODE_DATA(node)->node_zonelists[0];
1774 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1777 return zone ? zone->node : node;
1785 /* Do static interleaving for a VMA with known offset. */
1786 static unsigned offset_il_node(struct mempolicy *pol,
1787 struct vm_area_struct *vma, unsigned long off)
1789 unsigned nnodes = nodes_weight(pol->v.nodes);
1792 int nid = NUMA_NO_NODE;
1795 return numa_node_id();
1796 target = (unsigned int)off % nnodes;
1799 nid = next_node(nid, pol->v.nodes);
1801 } while (c <= target);
1805 /* Determine a node number for interleave */
1806 static inline unsigned interleave_nid(struct mempolicy *pol,
1807 struct vm_area_struct *vma, unsigned long addr, int shift)
1813 * for small pages, there is no difference between
1814 * shift and PAGE_SHIFT, so the bit-shift is safe.
1815 * for huge pages, since vm_pgoff is in units of small
1816 * pages, we need to shift off the always 0 bits to get
1819 BUG_ON(shift < PAGE_SHIFT);
1820 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1821 off += (addr - vma->vm_start) >> shift;
1822 return offset_il_node(pol, vma, off);
1824 return interleave_nodes(pol);
1828 * Return the bit number of a random bit set in the nodemask.
1829 * (returns NUMA_NO_NODE if nodemask is empty)
1831 int node_random(const nodemask_t *maskp)
1833 int w, bit = NUMA_NO_NODE;
1835 w = nodes_weight(*maskp);
1837 bit = bitmap_ord_to_pos(maskp->bits,
1838 get_random_int() % w, MAX_NUMNODES);
1842 #ifdef CONFIG_HUGETLBFS
1844 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1845 * @vma: virtual memory area whose policy is sought
1846 * @addr: address in @vma for shared policy lookup and interleave policy
1847 * @gfp_flags: for requested zone
1848 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1849 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1851 * Returns a zonelist suitable for a huge page allocation and a pointer
1852 * to the struct mempolicy for conditional unref after allocation.
1853 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1854 * @nodemask for filtering the zonelist.
1856 * Must be protected by read_mems_allowed_begin()
1858 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1859 gfp_t gfp_flags, struct mempolicy **mpol,
1860 nodemask_t **nodemask)
1862 struct zonelist *zl;
1864 *mpol = get_vma_policy(vma, addr);
1865 *nodemask = NULL; /* assume !MPOL_BIND */
1867 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1868 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1869 huge_page_shift(hstate_vma(vma))), gfp_flags);
1871 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1872 if ((*mpol)->mode == MPOL_BIND)
1873 *nodemask = &(*mpol)->v.nodes;
1879 * init_nodemask_of_mempolicy
1881 * If the current task's mempolicy is "default" [NULL], return 'false'
1882 * to indicate default policy. Otherwise, extract the policy nodemask
1883 * for 'bind' or 'interleave' policy into the argument nodemask, or
1884 * initialize the argument nodemask to contain the single node for
1885 * 'preferred' or 'local' policy and return 'true' to indicate presence
1886 * of non-default mempolicy.
1888 * We don't bother with reference counting the mempolicy [mpol_get/put]
1889 * because the current task is examining it's own mempolicy and a task's
1890 * mempolicy is only ever changed by the task itself.
1892 * N.B., it is the caller's responsibility to free a returned nodemask.
1894 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1896 struct mempolicy *mempolicy;
1899 if (!(mask && current->mempolicy))
1903 mempolicy = current->mempolicy;
1904 switch (mempolicy->mode) {
1905 case MPOL_PREFERRED:
1906 if (mempolicy->flags & MPOL_F_LOCAL)
1907 nid = numa_node_id();
1909 nid = mempolicy->v.preferred_node;
1910 init_nodemask_of_node(mask, nid);
1915 case MPOL_INTERLEAVE:
1916 *mask = mempolicy->v.nodes;
1922 task_unlock(current);
1929 * mempolicy_nodemask_intersects
1931 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1932 * policy. Otherwise, check for intersection between mask and the policy
1933 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1934 * policy, always return true since it may allocate elsewhere on fallback.
1936 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1938 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1939 const nodemask_t *mask)
1941 struct mempolicy *mempolicy;
1947 mempolicy = tsk->mempolicy;
1951 switch (mempolicy->mode) {
1952 case MPOL_PREFERRED:
1954 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1955 * allocate from, they may fallback to other nodes when oom.
1956 * Thus, it's possible for tsk to have allocated memory from
1961 case MPOL_INTERLEAVE:
1962 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1972 /* Allocate a page in interleaved policy.
1973 Own path because it needs to do special accounting. */
1974 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1977 struct zonelist *zl;
1980 zl = node_zonelist(nid, gfp);
1981 page = __alloc_pages(gfp, order, zl);
1982 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1983 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1988 * alloc_pages_vma - Allocate a page for a VMA.
1991 * %GFP_USER user allocation.
1992 * %GFP_KERNEL kernel allocations,
1993 * %GFP_HIGHMEM highmem/user allocations,
1994 * %GFP_FS allocation should not call back into a file system.
1995 * %GFP_ATOMIC don't sleep.
1997 * @order:Order of the GFP allocation.
1998 * @vma: Pointer to VMA or NULL if not available.
1999 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2001 * This function allocates a page from the kernel page pool and applies
2002 * a NUMA policy associated with the VMA or the current process.
2003 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2004 * mm_struct of the VMA to prevent it from going away. Should be used for
2005 * all allocations for pages that will be mapped into
2006 * user space. Returns NULL when no page can be allocated.
2008 * Should be called with the mm_sem of the vma hold.
2011 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2012 unsigned long addr, int node)
2014 struct mempolicy *pol;
2016 unsigned int cpuset_mems_cookie;
2019 pol = get_vma_policy(vma, addr);
2020 cpuset_mems_cookie = read_mems_allowed_begin();
2022 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2025 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2027 page = alloc_page_interleave(gfp, order, nid);
2028 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2033 page = __alloc_pages_nodemask(gfp, order,
2034 policy_zonelist(gfp, pol, node),
2035 policy_nodemask(gfp, pol));
2037 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2043 * alloc_pages_current - Allocate pages.
2046 * %GFP_USER user allocation,
2047 * %GFP_KERNEL kernel allocation,
2048 * %GFP_HIGHMEM highmem allocation,
2049 * %GFP_FS don't call back into a file system.
2050 * %GFP_ATOMIC don't sleep.
2051 * @order: Power of two of allocation size in pages. 0 is a single page.
2053 * Allocate a page from the kernel page pool. When not in
2054 * interrupt context and apply the current process NUMA policy.
2055 * Returns NULL when no page can be allocated.
2057 * Don't call cpuset_update_task_memory_state() unless
2058 * 1) it's ok to take cpuset_sem (can WAIT), and
2059 * 2) allocating for current task (not interrupt).
2061 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2063 struct mempolicy *pol = &default_policy;
2065 unsigned int cpuset_mems_cookie;
2067 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2068 pol = get_task_policy(current);
2071 cpuset_mems_cookie = read_mems_allowed_begin();
2074 * No reference counting needed for current->mempolicy
2075 * nor system default_policy
2077 if (pol->mode == MPOL_INTERLEAVE)
2078 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2080 page = __alloc_pages_nodemask(gfp, order,
2081 policy_zonelist(gfp, pol, numa_node_id()),
2082 policy_nodemask(gfp, pol));
2084 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2089 EXPORT_SYMBOL(alloc_pages_current);
2091 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2093 struct mempolicy *pol = mpol_dup(vma_policy(src));
2096 return PTR_ERR(pol);
2097 dst->vm_policy = pol;
2102 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2103 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2104 * with the mems_allowed returned by cpuset_mems_allowed(). This
2105 * keeps mempolicies cpuset relative after its cpuset moves. See
2106 * further kernel/cpuset.c update_nodemask().
2108 * current's mempolicy may be rebinded by the other task(the task that changes
2109 * cpuset's mems), so we needn't do rebind work for current task.
2112 /* Slow path of a mempolicy duplicate */
2113 struct mempolicy *__mpol_dup(struct mempolicy *old)
2115 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2118 return ERR_PTR(-ENOMEM);
2120 /* task's mempolicy is protected by alloc_lock */
2121 if (old == current->mempolicy) {
2124 task_unlock(current);
2128 if (current_cpuset_is_being_rebound()) {
2129 nodemask_t mems = cpuset_mems_allowed(current);
2130 if (new->flags & MPOL_F_REBINDING)
2131 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2133 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2135 atomic_set(&new->refcnt, 1);
2139 /* Slow path of a mempolicy comparison */
2140 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2144 if (a->mode != b->mode)
2146 if (a->flags != b->flags)
2148 if (mpol_store_user_nodemask(a))
2149 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2155 case MPOL_INTERLEAVE:
2156 return !!nodes_equal(a->v.nodes, b->v.nodes);
2157 case MPOL_PREFERRED:
2158 return a->v.preferred_node == b->v.preferred_node;
2166 * Shared memory backing store policy support.
2168 * Remember policies even when nobody has shared memory mapped.
2169 * The policies are kept in Red-Black tree linked from the inode.
2170 * They are protected by the sp->lock spinlock, which should be held
2171 * for any accesses to the tree.
2174 /* lookup first element intersecting start-end */
2175 /* Caller holds sp->lock */
2176 static struct sp_node *
2177 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2179 struct rb_node *n = sp->root.rb_node;
2182 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2184 if (start >= p->end)
2186 else if (end <= p->start)
2194 struct sp_node *w = NULL;
2195 struct rb_node *prev = rb_prev(n);
2198 w = rb_entry(prev, struct sp_node, nd);
2199 if (w->end <= start)
2203 return rb_entry(n, struct sp_node, nd);
2206 /* Insert a new shared policy into the list. */
2207 /* Caller holds sp->lock */
2208 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2210 struct rb_node **p = &sp->root.rb_node;
2211 struct rb_node *parent = NULL;
2216 nd = rb_entry(parent, struct sp_node, nd);
2217 if (new->start < nd->start)
2219 else if (new->end > nd->end)
2220 p = &(*p)->rb_right;
2224 rb_link_node(&new->nd, parent, p);
2225 rb_insert_color(&new->nd, &sp->root);
2226 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2227 new->policy ? new->policy->mode : 0);
2230 /* Find shared policy intersecting idx */
2232 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2234 struct mempolicy *pol = NULL;
2237 if (!sp->root.rb_node)
2239 spin_lock(&sp->lock);
2240 sn = sp_lookup(sp, idx, idx+1);
2242 mpol_get(sn->policy);
2245 spin_unlock(&sp->lock);
2249 static void sp_free(struct sp_node *n)
2251 mpol_put(n->policy);
2252 kmem_cache_free(sn_cache, n);
2256 * mpol_misplaced - check whether current page node is valid in policy
2258 * @page: page to be checked
2259 * @vma: vm area where page mapped
2260 * @addr: virtual address where page mapped
2262 * Lookup current policy node id for vma,addr and "compare to" page's
2266 * -1 - not misplaced, page is in the right node
2267 * node - node id where the page should be
2269 * Policy determination "mimics" alloc_page_vma().
2270 * Called from fault path where we know the vma and faulting address.
2272 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2274 struct mempolicy *pol;
2276 int curnid = page_to_nid(page);
2277 unsigned long pgoff;
2278 int thiscpu = raw_smp_processor_id();
2279 int thisnid = cpu_to_node(thiscpu);
2285 pol = get_vma_policy(vma, addr);
2286 if (!(pol->flags & MPOL_F_MOF))
2289 switch (pol->mode) {
2290 case MPOL_INTERLEAVE:
2291 BUG_ON(addr >= vma->vm_end);
2292 BUG_ON(addr < vma->vm_start);
2294 pgoff = vma->vm_pgoff;
2295 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2296 polnid = offset_il_node(pol, vma, pgoff);
2299 case MPOL_PREFERRED:
2300 if (pol->flags & MPOL_F_LOCAL)
2301 polnid = numa_node_id();
2303 polnid = pol->v.preferred_node;
2308 * allows binding to multiple nodes.
2309 * use current page if in policy nodemask,
2310 * else select nearest allowed node, if any.
2311 * If no allowed nodes, use current [!misplaced].
2313 if (node_isset(curnid, pol->v.nodes))
2315 (void)first_zones_zonelist(
2316 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2317 gfp_zone(GFP_HIGHUSER),
2318 &pol->v.nodes, &zone);
2319 polnid = zone->node;
2326 /* Migrate the page towards the node whose CPU is referencing it */
2327 if (pol->flags & MPOL_F_MORON) {
2330 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2334 if (curnid != polnid)
2342 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2344 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2345 rb_erase(&n->nd, &sp->root);
2349 static void sp_node_init(struct sp_node *node, unsigned long start,
2350 unsigned long end, struct mempolicy *pol)
2352 node->start = start;
2357 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2358 struct mempolicy *pol)
2361 struct mempolicy *newpol;
2363 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2367 newpol = mpol_dup(pol);
2368 if (IS_ERR(newpol)) {
2369 kmem_cache_free(sn_cache, n);
2372 newpol->flags |= MPOL_F_SHARED;
2373 sp_node_init(n, start, end, newpol);
2378 /* Replace a policy range. */
2379 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2380 unsigned long end, struct sp_node *new)
2383 struct sp_node *n_new = NULL;
2384 struct mempolicy *mpol_new = NULL;
2388 spin_lock(&sp->lock);
2389 n = sp_lookup(sp, start, end);
2390 /* Take care of old policies in the same range. */
2391 while (n && n->start < end) {
2392 struct rb_node *next = rb_next(&n->nd);
2393 if (n->start >= start) {
2399 /* Old policy spanning whole new range. */
2404 *mpol_new = *n->policy;
2405 atomic_set(&mpol_new->refcnt, 1);
2406 sp_node_init(n_new, end, n->end, mpol_new);
2408 sp_insert(sp, n_new);
2417 n = rb_entry(next, struct sp_node, nd);
2421 spin_unlock(&sp->lock);
2428 kmem_cache_free(sn_cache, n_new);
2433 spin_unlock(&sp->lock);
2435 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2438 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2445 * mpol_shared_policy_init - initialize shared policy for inode
2446 * @sp: pointer to inode shared policy
2447 * @mpol: struct mempolicy to install
2449 * Install non-NULL @mpol in inode's shared policy rb-tree.
2450 * On entry, the current task has a reference on a non-NULL @mpol.
2451 * This must be released on exit.
2452 * This is called at get_inode() calls and we can use GFP_KERNEL.
2454 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2458 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2459 spin_lock_init(&sp->lock);
2462 struct vm_area_struct pvma;
2463 struct mempolicy *new;
2464 NODEMASK_SCRATCH(scratch);
2468 /* contextualize the tmpfs mount point mempolicy */
2469 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2471 goto free_scratch; /* no valid nodemask intersection */
2474 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2475 task_unlock(current);
2479 /* Create pseudo-vma that contains just the policy */
2480 memset(&pvma, 0, sizeof(struct vm_area_struct));
2481 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2482 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2485 mpol_put(new); /* drop initial ref */
2487 NODEMASK_SCRATCH_FREE(scratch);
2489 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2493 int mpol_set_shared_policy(struct shared_policy *info,
2494 struct vm_area_struct *vma, struct mempolicy *npol)
2497 struct sp_node *new = NULL;
2498 unsigned long sz = vma_pages(vma);
2500 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2502 sz, npol ? npol->mode : -1,
2503 npol ? npol->flags : -1,
2504 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2507 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2511 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2517 /* Free a backing policy store on inode delete. */
2518 void mpol_free_shared_policy(struct shared_policy *p)
2521 struct rb_node *next;
2523 if (!p->root.rb_node)
2525 spin_lock(&p->lock);
2526 next = rb_first(&p->root);
2528 n = rb_entry(next, struct sp_node, nd);
2529 next = rb_next(&n->nd);
2532 spin_unlock(&p->lock);
2535 #ifdef CONFIG_NUMA_BALANCING
2536 static int __initdata numabalancing_override;
2538 static void __init check_numabalancing_enable(void)
2540 bool numabalancing_default = false;
2542 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2543 numabalancing_default = true;
2545 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2546 if (numabalancing_override)
2547 set_numabalancing_state(numabalancing_override == 1);
2549 if (nr_node_ids > 1 && !numabalancing_override) {
2550 pr_info("%s automatic NUMA balancing. "
2551 "Configure with numa_balancing= or the "
2552 "kernel.numa_balancing sysctl",
2553 numabalancing_default ? "Enabling" : "Disabling");
2554 set_numabalancing_state(numabalancing_default);
2558 static int __init setup_numabalancing(char *str)
2564 if (!strcmp(str, "enable")) {
2565 numabalancing_override = 1;
2567 } else if (!strcmp(str, "disable")) {
2568 numabalancing_override = -1;
2573 pr_warn("Unable to parse numa_balancing=\n");
2577 __setup("numa_balancing=", setup_numabalancing);
2579 static inline void __init check_numabalancing_enable(void)
2582 #endif /* CONFIG_NUMA_BALANCING */
2584 /* assumes fs == KERNEL_DS */
2585 void __init numa_policy_init(void)
2587 nodemask_t interleave_nodes;
2588 unsigned long largest = 0;
2589 int nid, prefer = 0;
2591 policy_cache = kmem_cache_create("numa_policy",
2592 sizeof(struct mempolicy),
2593 0, SLAB_PANIC, NULL);
2595 sn_cache = kmem_cache_create("shared_policy_node",
2596 sizeof(struct sp_node),
2597 0, SLAB_PANIC, NULL);
2599 for_each_node(nid) {
2600 preferred_node_policy[nid] = (struct mempolicy) {
2601 .refcnt = ATOMIC_INIT(1),
2602 .mode = MPOL_PREFERRED,
2603 .flags = MPOL_F_MOF | MPOL_F_MORON,
2604 .v = { .preferred_node = nid, },
2609 * Set interleaving policy for system init. Interleaving is only
2610 * enabled across suitably sized nodes (default is >= 16MB), or
2611 * fall back to the largest node if they're all smaller.
2613 nodes_clear(interleave_nodes);
2614 for_each_node_state(nid, N_MEMORY) {
2615 unsigned long total_pages = node_present_pages(nid);
2617 /* Preserve the largest node */
2618 if (largest < total_pages) {
2619 largest = total_pages;
2623 /* Interleave this node? */
2624 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2625 node_set(nid, interleave_nodes);
2628 /* All too small, use the largest */
2629 if (unlikely(nodes_empty(interleave_nodes)))
2630 node_set(prefer, interleave_nodes);
2632 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2633 pr_err("%s: interleaving failed\n", __func__);
2635 check_numabalancing_enable();
2638 /* Reset policy of current process to default */
2639 void numa_default_policy(void)
2641 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2645 * Parse and format mempolicy from/to strings
2649 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2651 static const char * const policy_modes[] =
2653 [MPOL_DEFAULT] = "default",
2654 [MPOL_PREFERRED] = "prefer",
2655 [MPOL_BIND] = "bind",
2656 [MPOL_INTERLEAVE] = "interleave",
2657 [MPOL_LOCAL] = "local",
2663 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2664 * @str: string containing mempolicy to parse
2665 * @mpol: pointer to struct mempolicy pointer, returned on success.
2668 * <mode>[=<flags>][:<nodelist>]
2670 * On success, returns 0, else 1
2672 int mpol_parse_str(char *str, struct mempolicy **mpol)
2674 struct mempolicy *new = NULL;
2675 unsigned short mode;
2676 unsigned short mode_flags;
2678 char *nodelist = strchr(str, ':');
2679 char *flags = strchr(str, '=');
2683 /* NUL-terminate mode or flags string */
2685 if (nodelist_parse(nodelist, nodes))
2687 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2693 *flags++ = '\0'; /* terminate mode string */
2695 for (mode = 0; mode < MPOL_MAX; mode++) {
2696 if (!strcmp(str, policy_modes[mode])) {
2700 if (mode >= MPOL_MAX)
2704 case MPOL_PREFERRED:
2706 * Insist on a nodelist of one node only
2709 char *rest = nodelist;
2710 while (isdigit(*rest))
2716 case MPOL_INTERLEAVE:
2718 * Default to online nodes with memory if no nodelist
2721 nodes = node_states[N_MEMORY];
2725 * Don't allow a nodelist; mpol_new() checks flags
2729 mode = MPOL_PREFERRED;
2733 * Insist on a empty nodelist
2740 * Insist on a nodelist
2749 * Currently, we only support two mutually exclusive
2752 if (!strcmp(flags, "static"))
2753 mode_flags |= MPOL_F_STATIC_NODES;
2754 else if (!strcmp(flags, "relative"))
2755 mode_flags |= MPOL_F_RELATIVE_NODES;
2760 new = mpol_new(mode, mode_flags, &nodes);
2765 * Save nodes for mpol_to_str() to show the tmpfs mount options
2766 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2768 if (mode != MPOL_PREFERRED)
2769 new->v.nodes = nodes;
2771 new->v.preferred_node = first_node(nodes);
2773 new->flags |= MPOL_F_LOCAL;
2776 * Save nodes for contextualization: this will be used to "clone"
2777 * the mempolicy in a specific context [cpuset] at a later time.
2779 new->w.user_nodemask = nodes;
2784 /* Restore string for error message */
2793 #endif /* CONFIG_TMPFS */
2796 * mpol_to_str - format a mempolicy structure for printing
2797 * @buffer: to contain formatted mempolicy string
2798 * @maxlen: length of @buffer
2799 * @pol: pointer to mempolicy to be formatted
2801 * Convert @pol into a string. If @buffer is too short, truncate the string.
2802 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2803 * longest flag, "relative", and to display at least a few node ids.
2805 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2808 nodemask_t nodes = NODE_MASK_NONE;
2809 unsigned short mode = MPOL_DEFAULT;
2810 unsigned short flags = 0;
2812 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2820 case MPOL_PREFERRED:
2821 if (flags & MPOL_F_LOCAL)
2824 node_set(pol->v.preferred_node, nodes);
2827 case MPOL_INTERLEAVE:
2828 nodes = pol->v.nodes;
2832 snprintf(p, maxlen, "unknown");
2836 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2838 if (flags & MPOL_MODE_FLAGS) {
2839 p += snprintf(p, buffer + maxlen - p, "=");
2842 * Currently, the only defined flags are mutually exclusive
2844 if (flags & MPOL_F_STATIC_NODES)
2845 p += snprintf(p, buffer + maxlen - p, "static");
2846 else if (flags & MPOL_F_RELATIVE_NODES)
2847 p += snprintf(p, buffer + maxlen - p, "relative");
2850 if (!nodes_empty(nodes)) {
2851 p += snprintf(p, buffer + maxlen - p, ":");
2852 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);