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 static 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 struct mm_struct *mm = current->mm;
806 NODEMASK_SCRATCH(scratch);
812 new = mpol_new(mode, flags, nodes);
818 * prevent changing our mempolicy while show_numa_maps()
820 * Note: do_set_mempolicy() can be called at init time
824 down_write(&mm->mmap_sem);
826 ret = mpol_set_nodemask(new, nodes, scratch);
828 task_unlock(current);
830 up_write(&mm->mmap_sem);
834 old = current->mempolicy;
835 current->mempolicy = new;
836 if (new && new->mode == MPOL_INTERLEAVE &&
837 nodes_weight(new->v.nodes))
838 current->il_next = first_node(new->v.nodes);
839 task_unlock(current);
841 up_write(&mm->mmap_sem);
846 NODEMASK_SCRATCH_FREE(scratch);
851 * Return nodemask for policy for get_mempolicy() query
853 * Called with task's alloc_lock held
855 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
858 if (p == &default_policy)
864 case MPOL_INTERLEAVE:
868 if (!(p->flags & MPOL_F_LOCAL))
869 node_set(p->v.preferred_node, *nodes);
870 /* else return empty node mask for local allocation */
877 static int lookup_node(struct mm_struct *mm, unsigned long addr)
882 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
884 err = page_to_nid(p);
890 /* Retrieve NUMA policy */
891 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
892 unsigned long addr, unsigned long flags)
895 struct mm_struct *mm = current->mm;
896 struct vm_area_struct *vma = NULL;
897 struct mempolicy *pol = current->mempolicy;
900 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
903 if (flags & MPOL_F_MEMS_ALLOWED) {
904 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
906 *policy = 0; /* just so it's initialized */
908 *nmask = cpuset_current_mems_allowed;
909 task_unlock(current);
913 if (flags & MPOL_F_ADDR) {
915 * Do NOT fall back to task policy if the
916 * vma/shared policy at addr is NULL. We
917 * want to return MPOL_DEFAULT in this case.
919 down_read(&mm->mmap_sem);
920 vma = find_vma_intersection(mm, addr, addr+1);
922 up_read(&mm->mmap_sem);
925 if (vma->vm_ops && vma->vm_ops->get_policy)
926 pol = vma->vm_ops->get_policy(vma, addr);
928 pol = vma->vm_policy;
933 pol = &default_policy; /* indicates default behavior */
935 if (flags & MPOL_F_NODE) {
936 if (flags & MPOL_F_ADDR) {
937 err = lookup_node(mm, addr);
941 } else if (pol == current->mempolicy &&
942 pol->mode == MPOL_INTERLEAVE) {
943 *policy = current->il_next;
949 *policy = pol == &default_policy ? MPOL_DEFAULT :
952 * Internal mempolicy flags must be masked off before exposing
953 * the policy to userspace.
955 *policy |= (pol->flags & MPOL_MODE_FLAGS);
959 up_read(¤t->mm->mmap_sem);
965 if (mpol_store_user_nodemask(pol)) {
966 *nmask = pol->w.user_nodemask;
969 get_policy_nodemask(pol, nmask);
970 task_unlock(current);
977 up_read(¤t->mm->mmap_sem);
981 #ifdef CONFIG_MIGRATION
985 static void migrate_page_add(struct page *page, struct list_head *pagelist,
989 * Avoid migrating a page that is shared with others.
991 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
992 if (!isolate_lru_page(page)) {
993 list_add_tail(&page->lru, pagelist);
994 inc_zone_page_state(page, NR_ISOLATED_ANON +
995 page_is_file_cache(page));
1000 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1003 return alloc_huge_page_node(page_hstate(compound_head(page)),
1006 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1010 * Migrate pages from one node to a target node.
1011 * Returns error or the number of pages not migrated.
1013 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1017 LIST_HEAD(pagelist);
1021 node_set(source, nmask);
1024 * This does not "check" the range but isolates all pages that
1025 * need migration. Between passing in the full user address
1026 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1028 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1029 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1030 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1032 if (!list_empty(&pagelist)) {
1033 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1034 MIGRATE_SYNC, MR_SYSCALL);
1036 putback_movable_pages(&pagelist);
1043 * Move pages between the two nodesets so as to preserve the physical
1044 * layout as much as possible.
1046 * Returns the number of page that could not be moved.
1048 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1049 const nodemask_t *to, int flags)
1055 err = migrate_prep();
1059 down_read(&mm->mmap_sem);
1061 err = migrate_vmas(mm, from, to, flags);
1066 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1067 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1068 * bit in 'tmp', and return that <source, dest> pair for migration.
1069 * The pair of nodemasks 'to' and 'from' define the map.
1071 * If no pair of bits is found that way, fallback to picking some
1072 * pair of 'source' and 'dest' bits that are not the same. If the
1073 * 'source' and 'dest' bits are the same, this represents a node
1074 * that will be migrating to itself, so no pages need move.
1076 * If no bits are left in 'tmp', or if all remaining bits left
1077 * in 'tmp' correspond to the same bit in 'to', return false
1078 * (nothing left to migrate).
1080 * This lets us pick a pair of nodes to migrate between, such that
1081 * if possible the dest node is not already occupied by some other
1082 * source node, minimizing the risk of overloading the memory on a
1083 * node that would happen if we migrated incoming memory to a node
1084 * before migrating outgoing memory source that same node.
1086 * A single scan of tmp is sufficient. As we go, we remember the
1087 * most recent <s, d> pair that moved (s != d). If we find a pair
1088 * that not only moved, but what's better, moved to an empty slot
1089 * (d is not set in tmp), then we break out then, with that pair.
1090 * Otherwise when we finish scanning from_tmp, we at least have the
1091 * most recent <s, d> pair that moved. If we get all the way through
1092 * the scan of tmp without finding any node that moved, much less
1093 * moved to an empty node, then there is nothing left worth migrating.
1097 while (!nodes_empty(tmp)) {
1099 int source = NUMA_NO_NODE;
1102 for_each_node_mask(s, tmp) {
1105 * do_migrate_pages() tries to maintain the relative
1106 * node relationship of the pages established between
1107 * threads and memory areas.
1109 * However if the number of source nodes is not equal to
1110 * the number of destination nodes we can not preserve
1111 * this node relative relationship. In that case, skip
1112 * copying memory from a node that is in the destination
1115 * Example: [2,3,4] -> [3,4,5] moves everything.
1116 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1119 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1120 (node_isset(s, *to)))
1123 d = node_remap(s, *from, *to);
1127 source = s; /* Node moved. Memorize */
1130 /* dest not in remaining from nodes? */
1131 if (!node_isset(dest, tmp))
1134 if (source == NUMA_NO_NODE)
1137 node_clear(source, tmp);
1138 err = migrate_to_node(mm, source, dest, flags);
1145 up_read(&mm->mmap_sem);
1153 * Allocate a new page for page migration based on vma policy.
1154 * Start by assuming the page is mapped by the same vma as contains @start.
1155 * Search forward from there, if not. N.B., this assumes that the
1156 * list of pages handed to migrate_pages()--which is how we get here--
1157 * is in virtual address order.
1159 static struct page *new_page(struct page *page, unsigned long start, int **x)
1161 struct vm_area_struct *vma;
1162 unsigned long uninitialized_var(address);
1164 vma = find_vma(current->mm, start);
1166 address = page_address_in_vma(page, vma);
1167 if (address != -EFAULT)
1172 if (PageHuge(page)) {
1174 return alloc_huge_page_noerr(vma, address, 1);
1177 * if !vma, alloc_page_vma() will use task or system default policy
1179 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1183 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1184 unsigned long flags)
1188 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1189 const nodemask_t *to, int flags)
1194 static struct page *new_page(struct page *page, unsigned long start, int **x)
1200 static long do_mbind(unsigned long start, unsigned long len,
1201 unsigned short mode, unsigned short mode_flags,
1202 nodemask_t *nmask, unsigned long flags)
1204 struct mm_struct *mm = current->mm;
1205 struct mempolicy *new;
1208 LIST_HEAD(pagelist);
1210 if (flags & ~(unsigned long)MPOL_MF_VALID)
1212 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1215 if (start & ~PAGE_MASK)
1218 if (mode == MPOL_DEFAULT)
1219 flags &= ~MPOL_MF_STRICT;
1221 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1229 new = mpol_new(mode, mode_flags, nmask);
1231 return PTR_ERR(new);
1233 if (flags & MPOL_MF_LAZY)
1234 new->flags |= MPOL_F_MOF;
1237 * If we are using the default policy then operation
1238 * on discontinuous address spaces is okay after all
1241 flags |= MPOL_MF_DISCONTIG_OK;
1243 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1244 start, start + len, mode, mode_flags,
1245 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1247 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1249 err = migrate_prep();
1254 NODEMASK_SCRATCH(scratch);
1256 down_write(&mm->mmap_sem);
1258 err = mpol_set_nodemask(new, nmask, scratch);
1259 task_unlock(current);
1261 up_write(&mm->mmap_sem);
1264 NODEMASK_SCRATCH_FREE(scratch);
1269 err = queue_pages_range(mm, start, end, nmask,
1270 flags | MPOL_MF_INVERT, &pagelist);
1272 err = mbind_range(mm, start, end, new);
1277 if (!list_empty(&pagelist)) {
1278 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1279 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1280 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1282 putback_movable_pages(&pagelist);
1285 if (nr_failed && (flags & MPOL_MF_STRICT))
1288 putback_movable_pages(&pagelist);
1290 up_write(&mm->mmap_sem);
1297 * User space interface with variable sized bitmaps for nodelists.
1300 /* Copy a node mask from user space. */
1301 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1302 unsigned long maxnode)
1305 unsigned long nlongs;
1306 unsigned long endmask;
1309 nodes_clear(*nodes);
1310 if (maxnode == 0 || !nmask)
1312 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1315 nlongs = BITS_TO_LONGS(maxnode);
1316 if ((maxnode % BITS_PER_LONG) == 0)
1319 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1321 /* When the user specified more nodes than supported just check
1322 if the non supported part is all zero. */
1323 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1324 if (nlongs > PAGE_SIZE/sizeof(long))
1326 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1328 if (get_user(t, nmask + k))
1330 if (k == nlongs - 1) {
1336 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1340 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1342 nodes_addr(*nodes)[nlongs-1] &= endmask;
1346 /* Copy a kernel node mask to user space */
1347 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1350 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1351 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1353 if (copy > nbytes) {
1354 if (copy > PAGE_SIZE)
1356 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1360 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1363 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1364 unsigned long, mode, const unsigned long __user *, nmask,
1365 unsigned long, maxnode, unsigned, flags)
1369 unsigned short mode_flags;
1371 mode_flags = mode & MPOL_MODE_FLAGS;
1372 mode &= ~MPOL_MODE_FLAGS;
1373 if (mode >= MPOL_MAX)
1375 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1376 (mode_flags & MPOL_F_RELATIVE_NODES))
1378 err = get_nodes(&nodes, nmask, maxnode);
1381 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1384 /* Set the process memory policy */
1385 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1386 unsigned long, maxnode)
1390 unsigned short flags;
1392 flags = mode & MPOL_MODE_FLAGS;
1393 mode &= ~MPOL_MODE_FLAGS;
1394 if ((unsigned int)mode >= MPOL_MAX)
1396 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1398 err = get_nodes(&nodes, nmask, maxnode);
1401 return do_set_mempolicy(mode, flags, &nodes);
1404 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1405 const unsigned long __user *, old_nodes,
1406 const unsigned long __user *, new_nodes)
1408 const struct cred *cred = current_cred(), *tcred;
1409 struct mm_struct *mm = NULL;
1410 struct task_struct *task;
1411 nodemask_t task_nodes;
1415 NODEMASK_SCRATCH(scratch);
1420 old = &scratch->mask1;
1421 new = &scratch->mask2;
1423 err = get_nodes(old, old_nodes, maxnode);
1427 err = get_nodes(new, new_nodes, maxnode);
1431 /* Find the mm_struct */
1433 task = pid ? find_task_by_vpid(pid) : current;
1439 get_task_struct(task);
1444 * Check if this process has the right to modify the specified
1445 * process. The right exists if the process has administrative
1446 * capabilities, superuser privileges or the same
1447 * userid as the target process.
1449 tcred = __task_cred(task);
1450 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1451 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1452 !capable(CAP_SYS_NICE)) {
1459 task_nodes = cpuset_mems_allowed(task);
1460 /* Is the user allowed to access the target nodes? */
1461 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1466 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1471 err = security_task_movememory(task);
1475 mm = get_task_mm(task);
1476 put_task_struct(task);
1483 err = do_migrate_pages(mm, old, new,
1484 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1488 NODEMASK_SCRATCH_FREE(scratch);
1493 put_task_struct(task);
1499 /* Retrieve NUMA policy */
1500 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1501 unsigned long __user *, nmask, unsigned long, maxnode,
1502 unsigned long, addr, unsigned long, flags)
1505 int uninitialized_var(pval);
1508 if (nmask != NULL && maxnode < MAX_NUMNODES)
1511 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1516 if (policy && put_user(pval, policy))
1520 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1525 #ifdef CONFIG_COMPAT
1527 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1528 compat_ulong_t __user *, nmask,
1529 compat_ulong_t, maxnode,
1530 compat_ulong_t, addr, compat_ulong_t, flags)
1533 unsigned long __user *nm = NULL;
1534 unsigned long nr_bits, alloc_size;
1535 DECLARE_BITMAP(bm, MAX_NUMNODES);
1537 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1538 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1541 nm = compat_alloc_user_space(alloc_size);
1543 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1545 if (!err && nmask) {
1546 unsigned long copy_size;
1547 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1548 err = copy_from_user(bm, nm, copy_size);
1549 /* ensure entire bitmap is zeroed */
1550 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1551 err |= compat_put_bitmap(nmask, bm, nr_bits);
1557 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1558 compat_ulong_t, maxnode)
1561 unsigned long __user *nm = NULL;
1562 unsigned long nr_bits, alloc_size;
1563 DECLARE_BITMAP(bm, MAX_NUMNODES);
1565 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1566 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1569 err = compat_get_bitmap(bm, nmask, nr_bits);
1570 nm = compat_alloc_user_space(alloc_size);
1571 err |= copy_to_user(nm, bm, alloc_size);
1577 return sys_set_mempolicy(mode, nm, nr_bits+1);
1580 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1581 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1582 compat_ulong_t, maxnode, compat_ulong_t, flags)
1585 unsigned long __user *nm = NULL;
1586 unsigned long nr_bits, alloc_size;
1589 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1590 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1593 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1594 nm = compat_alloc_user_space(alloc_size);
1595 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1601 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1607 * get_vma_policy(@task, @vma, @addr)
1608 * @task: task for fallback if vma policy == default
1609 * @vma: virtual memory area whose policy is sought
1610 * @addr: address in @vma for shared policy lookup
1612 * Returns effective policy for a VMA at specified address.
1613 * Falls back to @task or system default policy, as necessary.
1614 * Current or other task's task mempolicy and non-shared vma policies must be
1615 * protected by task_lock(task) by the caller.
1616 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1617 * count--added by the get_policy() vm_op, as appropriate--to protect against
1618 * freeing by another task. It is the caller's responsibility to free the
1619 * extra reference for shared policies.
1621 struct mempolicy *get_vma_policy(struct task_struct *task,
1622 struct vm_area_struct *vma, unsigned long addr)
1624 struct mempolicy *pol = get_task_policy(task);
1627 if (vma->vm_ops && vma->vm_ops->get_policy) {
1628 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1632 } else if (vma->vm_policy) {
1633 pol = vma->vm_policy;
1636 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1637 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1638 * count on these policies which will be dropped by
1639 * mpol_cond_put() later
1641 if (mpol_needs_cond_ref(pol))
1649 bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1651 struct mempolicy *pol = get_task_policy(task);
1654 if (vma->vm_ops && vma->vm_ops->get_policy) {
1657 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1658 if (pol && (pol->flags & MPOL_F_MOF))
1663 } else if (vma->vm_policy) {
1664 pol = vma->vm_policy;
1668 return pol->flags & MPOL_F_MOF;
1671 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1673 enum zone_type dynamic_policy_zone = policy_zone;
1675 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1678 * if policy->v.nodes has movable memory only,
1679 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1681 * policy->v.nodes is intersect with node_states[N_MEMORY].
1682 * so if the following test faile, it implies
1683 * policy->v.nodes has movable memory only.
1685 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1686 dynamic_policy_zone = ZONE_MOVABLE;
1688 return zone >= dynamic_policy_zone;
1692 * Return a nodemask representing a mempolicy for filtering nodes for
1695 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1697 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1698 if (unlikely(policy->mode == MPOL_BIND) &&
1699 apply_policy_zone(policy, gfp_zone(gfp)) &&
1700 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1701 return &policy->v.nodes;
1706 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1707 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1710 switch (policy->mode) {
1711 case MPOL_PREFERRED:
1712 if (!(policy->flags & MPOL_F_LOCAL))
1713 nd = policy->v.preferred_node;
1717 * Normally, MPOL_BIND allocations are node-local within the
1718 * allowed nodemask. However, if __GFP_THISNODE is set and the
1719 * current node isn't part of the mask, we use the zonelist for
1720 * the first node in the mask instead.
1722 if (unlikely(gfp & __GFP_THISNODE) &&
1723 unlikely(!node_isset(nd, policy->v.nodes)))
1724 nd = first_node(policy->v.nodes);
1729 return node_zonelist(nd, gfp);
1732 /* Do dynamic interleaving for a process */
1733 static unsigned interleave_nodes(struct mempolicy *policy)
1736 struct task_struct *me = current;
1739 next = next_node(nid, policy->v.nodes);
1740 if (next >= MAX_NUMNODES)
1741 next = first_node(policy->v.nodes);
1742 if (next < MAX_NUMNODES)
1748 * Depending on the memory policy provide a node from which to allocate the
1751 unsigned int mempolicy_slab_node(void)
1753 struct mempolicy *policy;
1754 int node = numa_mem_id();
1759 policy = current->mempolicy;
1760 if (!policy || policy->flags & MPOL_F_LOCAL)
1763 switch (policy->mode) {
1764 case MPOL_PREFERRED:
1766 * handled MPOL_F_LOCAL above
1768 return policy->v.preferred_node;
1770 case MPOL_INTERLEAVE:
1771 return interleave_nodes(policy);
1775 * Follow bind policy behavior and start allocation at the
1778 struct zonelist *zonelist;
1780 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1781 zonelist = &NODE_DATA(node)->node_zonelists[0];
1782 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1785 return zone ? zone->node : node;
1793 /* Do static interleaving for a VMA with known offset. */
1794 static unsigned offset_il_node(struct mempolicy *pol,
1795 struct vm_area_struct *vma, unsigned long off)
1797 unsigned nnodes = nodes_weight(pol->v.nodes);
1800 int nid = NUMA_NO_NODE;
1803 return numa_node_id();
1804 target = (unsigned int)off % nnodes;
1807 nid = next_node(nid, pol->v.nodes);
1809 } while (c <= target);
1813 /* Determine a node number for interleave */
1814 static inline unsigned interleave_nid(struct mempolicy *pol,
1815 struct vm_area_struct *vma, unsigned long addr, int shift)
1821 * for small pages, there is no difference between
1822 * shift and PAGE_SHIFT, so the bit-shift is safe.
1823 * for huge pages, since vm_pgoff is in units of small
1824 * pages, we need to shift off the always 0 bits to get
1827 BUG_ON(shift < PAGE_SHIFT);
1828 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1829 off += (addr - vma->vm_start) >> shift;
1830 return offset_il_node(pol, vma, off);
1832 return interleave_nodes(pol);
1836 * Return the bit number of a random bit set in the nodemask.
1837 * (returns NUMA_NO_NODE if nodemask is empty)
1839 int node_random(const nodemask_t *maskp)
1841 int w, bit = NUMA_NO_NODE;
1843 w = nodes_weight(*maskp);
1845 bit = bitmap_ord_to_pos(maskp->bits,
1846 get_random_int() % w, MAX_NUMNODES);
1850 #ifdef CONFIG_HUGETLBFS
1852 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1853 * @vma: virtual memory area whose policy is sought
1854 * @addr: address in @vma for shared policy lookup and interleave policy
1855 * @gfp_flags: for requested zone
1856 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1857 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1859 * Returns a zonelist suitable for a huge page allocation and a pointer
1860 * to the struct mempolicy for conditional unref after allocation.
1861 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1862 * @nodemask for filtering the zonelist.
1864 * Must be protected by read_mems_allowed_begin()
1866 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1867 gfp_t gfp_flags, struct mempolicy **mpol,
1868 nodemask_t **nodemask)
1870 struct zonelist *zl;
1872 *mpol = get_vma_policy(current, vma, addr);
1873 *nodemask = NULL; /* assume !MPOL_BIND */
1875 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1876 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1877 huge_page_shift(hstate_vma(vma))), gfp_flags);
1879 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1880 if ((*mpol)->mode == MPOL_BIND)
1881 *nodemask = &(*mpol)->v.nodes;
1887 * init_nodemask_of_mempolicy
1889 * If the current task's mempolicy is "default" [NULL], return 'false'
1890 * to indicate default policy. Otherwise, extract the policy nodemask
1891 * for 'bind' or 'interleave' policy into the argument nodemask, or
1892 * initialize the argument nodemask to contain the single node for
1893 * 'preferred' or 'local' policy and return 'true' to indicate presence
1894 * of non-default mempolicy.
1896 * We don't bother with reference counting the mempolicy [mpol_get/put]
1897 * because the current task is examining it's own mempolicy and a task's
1898 * mempolicy is only ever changed by the task itself.
1900 * N.B., it is the caller's responsibility to free a returned nodemask.
1902 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1904 struct mempolicy *mempolicy;
1907 if (!(mask && current->mempolicy))
1911 mempolicy = current->mempolicy;
1912 switch (mempolicy->mode) {
1913 case MPOL_PREFERRED:
1914 if (mempolicy->flags & MPOL_F_LOCAL)
1915 nid = numa_node_id();
1917 nid = mempolicy->v.preferred_node;
1918 init_nodemask_of_node(mask, nid);
1923 case MPOL_INTERLEAVE:
1924 *mask = mempolicy->v.nodes;
1930 task_unlock(current);
1937 * mempolicy_nodemask_intersects
1939 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1940 * policy. Otherwise, check for intersection between mask and the policy
1941 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1942 * policy, always return true since it may allocate elsewhere on fallback.
1944 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1946 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1947 const nodemask_t *mask)
1949 struct mempolicy *mempolicy;
1955 mempolicy = tsk->mempolicy;
1959 switch (mempolicy->mode) {
1960 case MPOL_PREFERRED:
1962 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1963 * allocate from, they may fallback to other nodes when oom.
1964 * Thus, it's possible for tsk to have allocated memory from
1969 case MPOL_INTERLEAVE:
1970 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1980 /* Allocate a page in interleaved policy.
1981 Own path because it needs to do special accounting. */
1982 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1985 struct zonelist *zl;
1988 zl = node_zonelist(nid, gfp);
1989 page = __alloc_pages(gfp, order, zl);
1990 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1991 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1996 * alloc_pages_vma - Allocate a page for a VMA.
1999 * %GFP_USER user allocation.
2000 * %GFP_KERNEL kernel allocations,
2001 * %GFP_HIGHMEM highmem/user allocations,
2002 * %GFP_FS allocation should not call back into a file system.
2003 * %GFP_ATOMIC don't sleep.
2005 * @order:Order of the GFP allocation.
2006 * @vma: Pointer to VMA or NULL if not available.
2007 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2009 * This function allocates a page from the kernel page pool and applies
2010 * a NUMA policy associated with the VMA or the current process.
2011 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2012 * mm_struct of the VMA to prevent it from going away. Should be used for
2013 * all allocations for pages that will be mapped into
2014 * user space. Returns NULL when no page can be allocated.
2016 * Should be called with the mm_sem of the vma hold.
2019 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2020 unsigned long addr, int node)
2022 struct mempolicy *pol;
2024 unsigned int cpuset_mems_cookie;
2027 pol = get_vma_policy(current, vma, addr);
2028 cpuset_mems_cookie = read_mems_allowed_begin();
2030 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2033 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2035 page = alloc_page_interleave(gfp, order, nid);
2036 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2041 page = __alloc_pages_nodemask(gfp, order,
2042 policy_zonelist(gfp, pol, node),
2043 policy_nodemask(gfp, pol));
2045 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2051 * alloc_pages_current - Allocate pages.
2054 * %GFP_USER user allocation,
2055 * %GFP_KERNEL kernel allocation,
2056 * %GFP_HIGHMEM highmem allocation,
2057 * %GFP_FS don't call back into a file system.
2058 * %GFP_ATOMIC don't sleep.
2059 * @order: Power of two of allocation size in pages. 0 is a single page.
2061 * Allocate a page from the kernel page pool. When not in
2062 * interrupt context and apply the current process NUMA policy.
2063 * Returns NULL when no page can be allocated.
2065 * Don't call cpuset_update_task_memory_state() unless
2066 * 1) it's ok to take cpuset_sem (can WAIT), and
2067 * 2) allocating for current task (not interrupt).
2069 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2071 struct mempolicy *pol = get_task_policy(current);
2073 unsigned int cpuset_mems_cookie;
2075 if (in_interrupt() || (gfp & __GFP_THISNODE))
2076 pol = &default_policy;
2079 cpuset_mems_cookie = read_mems_allowed_begin();
2082 * No reference counting needed for current->mempolicy
2083 * nor system default_policy
2085 if (pol->mode == MPOL_INTERLEAVE)
2086 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2088 page = __alloc_pages_nodemask(gfp, order,
2089 policy_zonelist(gfp, pol, numa_node_id()),
2090 policy_nodemask(gfp, pol));
2092 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2097 EXPORT_SYMBOL(alloc_pages_current);
2099 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2101 struct mempolicy *pol = mpol_dup(vma_policy(src));
2104 return PTR_ERR(pol);
2105 dst->vm_policy = pol;
2110 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2111 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2112 * with the mems_allowed returned by cpuset_mems_allowed(). This
2113 * keeps mempolicies cpuset relative after its cpuset moves. See
2114 * further kernel/cpuset.c update_nodemask().
2116 * current's mempolicy may be rebinded by the other task(the task that changes
2117 * cpuset's mems), so we needn't do rebind work for current task.
2120 /* Slow path of a mempolicy duplicate */
2121 struct mempolicy *__mpol_dup(struct mempolicy *old)
2123 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2126 return ERR_PTR(-ENOMEM);
2128 /* task's mempolicy is protected by alloc_lock */
2129 if (old == current->mempolicy) {
2132 task_unlock(current);
2136 if (current_cpuset_is_being_rebound()) {
2137 nodemask_t mems = cpuset_mems_allowed(current);
2138 if (new->flags & MPOL_F_REBINDING)
2139 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2141 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2143 atomic_set(&new->refcnt, 1);
2147 /* Slow path of a mempolicy comparison */
2148 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2152 if (a->mode != b->mode)
2154 if (a->flags != b->flags)
2156 if (mpol_store_user_nodemask(a))
2157 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2163 case MPOL_INTERLEAVE:
2164 return !!nodes_equal(a->v.nodes, b->v.nodes);
2165 case MPOL_PREFERRED:
2166 return a->v.preferred_node == b->v.preferred_node;
2174 * Shared memory backing store policy support.
2176 * Remember policies even when nobody has shared memory mapped.
2177 * The policies are kept in Red-Black tree linked from the inode.
2178 * They are protected by the sp->lock spinlock, which should be held
2179 * for any accesses to the tree.
2182 /* lookup first element intersecting start-end */
2183 /* Caller holds sp->lock */
2184 static struct sp_node *
2185 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2187 struct rb_node *n = sp->root.rb_node;
2190 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2192 if (start >= p->end)
2194 else if (end <= p->start)
2202 struct sp_node *w = NULL;
2203 struct rb_node *prev = rb_prev(n);
2206 w = rb_entry(prev, struct sp_node, nd);
2207 if (w->end <= start)
2211 return rb_entry(n, struct sp_node, nd);
2214 /* Insert a new shared policy into the list. */
2215 /* Caller holds sp->lock */
2216 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2218 struct rb_node **p = &sp->root.rb_node;
2219 struct rb_node *parent = NULL;
2224 nd = rb_entry(parent, struct sp_node, nd);
2225 if (new->start < nd->start)
2227 else if (new->end > nd->end)
2228 p = &(*p)->rb_right;
2232 rb_link_node(&new->nd, parent, p);
2233 rb_insert_color(&new->nd, &sp->root);
2234 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2235 new->policy ? new->policy->mode : 0);
2238 /* Find shared policy intersecting idx */
2240 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2242 struct mempolicy *pol = NULL;
2245 if (!sp->root.rb_node)
2247 spin_lock(&sp->lock);
2248 sn = sp_lookup(sp, idx, idx+1);
2250 mpol_get(sn->policy);
2253 spin_unlock(&sp->lock);
2257 static void sp_free(struct sp_node *n)
2259 mpol_put(n->policy);
2260 kmem_cache_free(sn_cache, n);
2264 * mpol_misplaced - check whether current page node is valid in policy
2266 * @page: page to be checked
2267 * @vma: vm area where page mapped
2268 * @addr: virtual address where page mapped
2270 * Lookup current policy node id for vma,addr and "compare to" page's
2274 * -1 - not misplaced, page is in the right node
2275 * node - node id where the page should be
2277 * Policy determination "mimics" alloc_page_vma().
2278 * Called from fault path where we know the vma and faulting address.
2280 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2282 struct mempolicy *pol;
2284 int curnid = page_to_nid(page);
2285 unsigned long pgoff;
2286 int thiscpu = raw_smp_processor_id();
2287 int thisnid = cpu_to_node(thiscpu);
2293 pol = get_vma_policy(current, vma, addr);
2294 if (!(pol->flags & MPOL_F_MOF))
2297 switch (pol->mode) {
2298 case MPOL_INTERLEAVE:
2299 BUG_ON(addr >= vma->vm_end);
2300 BUG_ON(addr < vma->vm_start);
2302 pgoff = vma->vm_pgoff;
2303 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2304 polnid = offset_il_node(pol, vma, pgoff);
2307 case MPOL_PREFERRED:
2308 if (pol->flags & MPOL_F_LOCAL)
2309 polnid = numa_node_id();
2311 polnid = pol->v.preferred_node;
2316 * allows binding to multiple nodes.
2317 * use current page if in policy nodemask,
2318 * else select nearest allowed node, if any.
2319 * If no allowed nodes, use current [!misplaced].
2321 if (node_isset(curnid, pol->v.nodes))
2323 (void)first_zones_zonelist(
2324 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2325 gfp_zone(GFP_HIGHUSER),
2326 &pol->v.nodes, &zone);
2327 polnid = zone->node;
2334 /* Migrate the page towards the node whose CPU is referencing it */
2335 if (pol->flags & MPOL_F_MORON) {
2338 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2342 if (curnid != polnid)
2350 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2352 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2353 rb_erase(&n->nd, &sp->root);
2357 static void sp_node_init(struct sp_node *node, unsigned long start,
2358 unsigned long end, struct mempolicy *pol)
2360 node->start = start;
2365 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2366 struct mempolicy *pol)
2369 struct mempolicy *newpol;
2371 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2375 newpol = mpol_dup(pol);
2376 if (IS_ERR(newpol)) {
2377 kmem_cache_free(sn_cache, n);
2380 newpol->flags |= MPOL_F_SHARED;
2381 sp_node_init(n, start, end, newpol);
2386 /* Replace a policy range. */
2387 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2388 unsigned long end, struct sp_node *new)
2391 struct sp_node *n_new = NULL;
2392 struct mempolicy *mpol_new = NULL;
2396 spin_lock(&sp->lock);
2397 n = sp_lookup(sp, start, end);
2398 /* Take care of old policies in the same range. */
2399 while (n && n->start < end) {
2400 struct rb_node *next = rb_next(&n->nd);
2401 if (n->start >= start) {
2407 /* Old policy spanning whole new range. */
2412 *mpol_new = *n->policy;
2413 atomic_set(&mpol_new->refcnt, 1);
2414 sp_node_init(n_new, end, n->end, mpol_new);
2416 sp_insert(sp, n_new);
2425 n = rb_entry(next, struct sp_node, nd);
2429 spin_unlock(&sp->lock);
2436 kmem_cache_free(sn_cache, n_new);
2441 spin_unlock(&sp->lock);
2443 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2446 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2453 * mpol_shared_policy_init - initialize shared policy for inode
2454 * @sp: pointer to inode shared policy
2455 * @mpol: struct mempolicy to install
2457 * Install non-NULL @mpol in inode's shared policy rb-tree.
2458 * On entry, the current task has a reference on a non-NULL @mpol.
2459 * This must be released on exit.
2460 * This is called at get_inode() calls and we can use GFP_KERNEL.
2462 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2466 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2467 spin_lock_init(&sp->lock);
2470 struct vm_area_struct pvma;
2471 struct mempolicy *new;
2472 NODEMASK_SCRATCH(scratch);
2476 /* contextualize the tmpfs mount point mempolicy */
2477 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2479 goto free_scratch; /* no valid nodemask intersection */
2482 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2483 task_unlock(current);
2487 /* Create pseudo-vma that contains just the policy */
2488 memset(&pvma, 0, sizeof(struct vm_area_struct));
2489 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2490 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2493 mpol_put(new); /* drop initial ref */
2495 NODEMASK_SCRATCH_FREE(scratch);
2497 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2501 int mpol_set_shared_policy(struct shared_policy *info,
2502 struct vm_area_struct *vma, struct mempolicy *npol)
2505 struct sp_node *new = NULL;
2506 unsigned long sz = vma_pages(vma);
2508 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2510 sz, npol ? npol->mode : -1,
2511 npol ? npol->flags : -1,
2512 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2515 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2519 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2525 /* Free a backing policy store on inode delete. */
2526 void mpol_free_shared_policy(struct shared_policy *p)
2529 struct rb_node *next;
2531 if (!p->root.rb_node)
2533 spin_lock(&p->lock);
2534 next = rb_first(&p->root);
2536 n = rb_entry(next, struct sp_node, nd);
2537 next = rb_next(&n->nd);
2540 spin_unlock(&p->lock);
2543 #ifdef CONFIG_NUMA_BALANCING
2544 static int __initdata numabalancing_override;
2546 static void __init check_numabalancing_enable(void)
2548 bool numabalancing_default = false;
2550 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2551 numabalancing_default = true;
2553 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2554 if (numabalancing_override)
2555 set_numabalancing_state(numabalancing_override == 1);
2557 if (nr_node_ids > 1 && !numabalancing_override) {
2558 pr_info("%s automatic NUMA balancing. "
2559 "Configure with numa_balancing= or the "
2560 "kernel.numa_balancing sysctl",
2561 numabalancing_default ? "Enabling" : "Disabling");
2562 set_numabalancing_state(numabalancing_default);
2566 static int __init setup_numabalancing(char *str)
2572 if (!strcmp(str, "enable")) {
2573 numabalancing_override = 1;
2575 } else if (!strcmp(str, "disable")) {
2576 numabalancing_override = -1;
2581 pr_warn("Unable to parse numa_balancing=\n");
2585 __setup("numa_balancing=", setup_numabalancing);
2587 static inline void __init check_numabalancing_enable(void)
2590 #endif /* CONFIG_NUMA_BALANCING */
2592 /* assumes fs == KERNEL_DS */
2593 void __init numa_policy_init(void)
2595 nodemask_t interleave_nodes;
2596 unsigned long largest = 0;
2597 int nid, prefer = 0;
2599 policy_cache = kmem_cache_create("numa_policy",
2600 sizeof(struct mempolicy),
2601 0, SLAB_PANIC, NULL);
2603 sn_cache = kmem_cache_create("shared_policy_node",
2604 sizeof(struct sp_node),
2605 0, SLAB_PANIC, NULL);
2607 for_each_node(nid) {
2608 preferred_node_policy[nid] = (struct mempolicy) {
2609 .refcnt = ATOMIC_INIT(1),
2610 .mode = MPOL_PREFERRED,
2611 .flags = MPOL_F_MOF | MPOL_F_MORON,
2612 .v = { .preferred_node = nid, },
2617 * Set interleaving policy for system init. Interleaving is only
2618 * enabled across suitably sized nodes (default is >= 16MB), or
2619 * fall back to the largest node if they're all smaller.
2621 nodes_clear(interleave_nodes);
2622 for_each_node_state(nid, N_MEMORY) {
2623 unsigned long total_pages = node_present_pages(nid);
2625 /* Preserve the largest node */
2626 if (largest < total_pages) {
2627 largest = total_pages;
2631 /* Interleave this node? */
2632 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2633 node_set(nid, interleave_nodes);
2636 /* All too small, use the largest */
2637 if (unlikely(nodes_empty(interleave_nodes)))
2638 node_set(prefer, interleave_nodes);
2640 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2641 pr_err("%s: interleaving failed\n", __func__);
2643 check_numabalancing_enable();
2646 /* Reset policy of current process to default */
2647 void numa_default_policy(void)
2649 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2653 * Parse and format mempolicy from/to strings
2657 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2659 static const char * const policy_modes[] =
2661 [MPOL_DEFAULT] = "default",
2662 [MPOL_PREFERRED] = "prefer",
2663 [MPOL_BIND] = "bind",
2664 [MPOL_INTERLEAVE] = "interleave",
2665 [MPOL_LOCAL] = "local",
2671 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2672 * @str: string containing mempolicy to parse
2673 * @mpol: pointer to struct mempolicy pointer, returned on success.
2676 * <mode>[=<flags>][:<nodelist>]
2678 * On success, returns 0, else 1
2680 int mpol_parse_str(char *str, struct mempolicy **mpol)
2682 struct mempolicy *new = NULL;
2683 unsigned short mode;
2684 unsigned short mode_flags;
2686 char *nodelist = strchr(str, ':');
2687 char *flags = strchr(str, '=');
2691 /* NUL-terminate mode or flags string */
2693 if (nodelist_parse(nodelist, nodes))
2695 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2701 *flags++ = '\0'; /* terminate mode string */
2703 for (mode = 0; mode < MPOL_MAX; mode++) {
2704 if (!strcmp(str, policy_modes[mode])) {
2708 if (mode >= MPOL_MAX)
2712 case MPOL_PREFERRED:
2714 * Insist on a nodelist of one node only
2717 char *rest = nodelist;
2718 while (isdigit(*rest))
2724 case MPOL_INTERLEAVE:
2726 * Default to online nodes with memory if no nodelist
2729 nodes = node_states[N_MEMORY];
2733 * Don't allow a nodelist; mpol_new() checks flags
2737 mode = MPOL_PREFERRED;
2741 * Insist on a empty nodelist
2748 * Insist on a nodelist
2757 * Currently, we only support two mutually exclusive
2760 if (!strcmp(flags, "static"))
2761 mode_flags |= MPOL_F_STATIC_NODES;
2762 else if (!strcmp(flags, "relative"))
2763 mode_flags |= MPOL_F_RELATIVE_NODES;
2768 new = mpol_new(mode, mode_flags, &nodes);
2773 * Save nodes for mpol_to_str() to show the tmpfs mount options
2774 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2776 if (mode != MPOL_PREFERRED)
2777 new->v.nodes = nodes;
2779 new->v.preferred_node = first_node(nodes);
2781 new->flags |= MPOL_F_LOCAL;
2784 * Save nodes for contextualization: this will be used to "clone"
2785 * the mempolicy in a specific context [cpuset] at a later time.
2787 new->w.user_nodemask = nodes;
2792 /* Restore string for error message */
2801 #endif /* CONFIG_TMPFS */
2804 * mpol_to_str - format a mempolicy structure for printing
2805 * @buffer: to contain formatted mempolicy string
2806 * @maxlen: length of @buffer
2807 * @pol: pointer to mempolicy to be formatted
2809 * Convert @pol into a string. If @buffer is too short, truncate the string.
2810 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2811 * longest flag, "relative", and to display at least a few node ids.
2813 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2816 nodemask_t nodes = NODE_MASK_NONE;
2817 unsigned short mode = MPOL_DEFAULT;
2818 unsigned short flags = 0;
2820 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2828 case MPOL_PREFERRED:
2829 if (flags & MPOL_F_LOCAL)
2832 node_set(pol->v.preferred_node, nodes);
2835 case MPOL_INTERLEAVE:
2836 nodes = pol->v.nodes;
2840 snprintf(p, maxlen, "unknown");
2844 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2846 if (flags & MPOL_MODE_FLAGS) {
2847 p += snprintf(p, buffer + maxlen - p, "=");
2850 * Currently, the only defined flags are mutually exclusive
2852 if (flags & MPOL_F_STATIC_NODES)
2853 p += snprintf(p, buffer + maxlen - p, "static");
2854 else if (flags & MPOL_F_RELATIVE_NODES)
2855 p += snprintf(p, buffer + maxlen - p, "relative");
2858 if (!nodes_empty(nodes)) {
2859 p += snprintf(p, buffer + maxlen - p, ":");
2860 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);