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 node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/gfp.h>
77 #include <linux/slab.h>
78 #include <linux/string.h>
79 #include <linux/module.h>
80 #include <linux/nsproxy.h>
81 #include <linux/interrupt.h>
82 #include <linux/init.h>
83 #include <linux/compat.h>
84 #include <linux/swap.h>
85 #include <linux/seq_file.h>
86 #include <linux/proc_fs.h>
87 #include <linux/migrate.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
92 #include <asm/tlbflush.h>
93 #include <asm/uaccess.h>
96 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
97 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
98 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
100 static struct kmem_cache
*policy_cache
;
101 static struct kmem_cache
*sn_cache
;
103 /* Highest zone. An specific allocation for a zone below that is not
105 enum zone_type policy_zone
= 0;
108 * run-time system-wide default policy => local allocation
110 struct mempolicy default_policy
= {
111 .refcnt
= ATOMIC_INIT(1), /* never free it */
112 .mode
= MPOL_PREFERRED
,
113 .flags
= MPOL_F_LOCAL
,
116 static const struct mempolicy_operations
{
117 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
118 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
119 } mpol_ops
[MPOL_MAX
];
121 /* Check that the nodemask contains at least one populated zone */
122 static int is_valid_nodemask(const nodemask_t
*nodemask
)
126 /* Check that there is something useful in this mask */
129 for_each_node_mask(nd
, *nodemask
) {
132 for (k
= 0; k
<= policy_zone
; k
++) {
133 z
= &NODE_DATA(nd
)->node_zones
[k
];
134 if (z
->present_pages
> 0)
142 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
144 return pol
->flags
& (MPOL_F_STATIC_NODES
| MPOL_F_RELATIVE_NODES
);
147 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
148 const nodemask_t
*rel
)
151 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
152 nodes_onto(*ret
, tmp
, *rel
);
155 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
157 if (nodes_empty(*nodes
))
159 pol
->v
.nodes
= *nodes
;
163 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
166 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
167 else if (nodes_empty(*nodes
))
168 return -EINVAL
; /* no allowed nodes */
170 pol
->v
.preferred_node
= first_node(*nodes
);
174 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
176 if (!is_valid_nodemask(nodes
))
178 pol
->v
.nodes
= *nodes
;
182 /* Create a new policy */
183 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
186 struct mempolicy
*policy
;
187 nodemask_t cpuset_context_nmask
;
190 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
191 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : -1);
193 if (mode
== MPOL_DEFAULT
) {
194 if (nodes
&& !nodes_empty(*nodes
))
195 return ERR_PTR(-EINVAL
);
196 return NULL
; /* simply delete any existing policy */
201 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
202 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
203 * All other modes require a valid pointer to a non-empty nodemask.
205 if (mode
== MPOL_PREFERRED
) {
206 if (nodes_empty(*nodes
)) {
207 if (((flags
& MPOL_F_STATIC_NODES
) ||
208 (flags
& MPOL_F_RELATIVE_NODES
)))
209 return ERR_PTR(-EINVAL
);
210 nodes
= NULL
; /* flag local alloc */
212 } else if (nodes_empty(*nodes
))
213 return ERR_PTR(-EINVAL
);
214 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
216 return ERR_PTR(-ENOMEM
);
217 atomic_set(&policy
->refcnt
, 1);
219 policy
->flags
= flags
;
223 * cpuset related setup doesn't apply to local allocation
225 cpuset_update_task_memory_state();
226 if (flags
& MPOL_F_RELATIVE_NODES
)
227 mpol_relative_nodemask(&cpuset_context_nmask
, nodes
,
228 &cpuset_current_mems_allowed
);
230 nodes_and(cpuset_context_nmask
, *nodes
,
231 cpuset_current_mems_allowed
);
232 if (mpol_store_user_nodemask(policy
))
233 policy
->w
.user_nodemask
= *nodes
;
235 policy
->w
.cpuset_mems_allowed
=
236 cpuset_mems_allowed(current
);
239 ret
= mpol_ops
[mode
].create(policy
,
240 nodes
? &cpuset_context_nmask
: NULL
);
242 kmem_cache_free(policy_cache
, policy
);
248 /* Slow path of a mpol destructor. */
249 void __mpol_put(struct mempolicy
*p
)
251 if (!atomic_dec_and_test(&p
->refcnt
))
253 kmem_cache_free(policy_cache
, p
);
256 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
260 static void mpol_rebind_nodemask(struct mempolicy
*pol
,
261 const nodemask_t
*nodes
)
265 if (pol
->flags
& MPOL_F_STATIC_NODES
)
266 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
267 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
268 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
270 nodes_remap(tmp
, pol
->v
.nodes
, pol
->w
.cpuset_mems_allowed
,
272 pol
->w
.cpuset_mems_allowed
= *nodes
;
276 if (!node_isset(current
->il_next
, tmp
)) {
277 current
->il_next
= next_node(current
->il_next
, tmp
);
278 if (current
->il_next
>= MAX_NUMNODES
)
279 current
->il_next
= first_node(tmp
);
280 if (current
->il_next
>= MAX_NUMNODES
)
281 current
->il_next
= numa_node_id();
285 static void mpol_rebind_preferred(struct mempolicy
*pol
,
286 const nodemask_t
*nodes
)
290 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
291 int node
= first_node(pol
->w
.user_nodemask
);
293 if (node_isset(node
, *nodes
)) {
294 pol
->v
.preferred_node
= node
;
295 pol
->flags
&= ~MPOL_F_LOCAL
;
297 pol
->flags
|= MPOL_F_LOCAL
;
298 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
299 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
300 pol
->v
.preferred_node
= first_node(tmp
);
301 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
302 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
303 pol
->w
.cpuset_mems_allowed
,
305 pol
->w
.cpuset_mems_allowed
= *nodes
;
309 /* Migrate a policy to a different set of nodes */
310 static void mpol_rebind_policy(struct mempolicy
*pol
,
311 const nodemask_t
*newmask
)
315 if (!mpol_store_user_nodemask(pol
) &&
316 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
318 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
322 * Wrapper for mpol_rebind_policy() that just requires task
323 * pointer, and updates task mempolicy.
326 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
328 mpol_rebind_policy(tsk
->mempolicy
, new);
332 * Rebind each vma in mm to new nodemask.
334 * Call holding a reference to mm. Takes mm->mmap_sem during call.
337 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
339 struct vm_area_struct
*vma
;
341 down_write(&mm
->mmap_sem
);
342 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
343 mpol_rebind_policy(vma
->vm_policy
, new);
344 up_write(&mm
->mmap_sem
);
347 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
349 .rebind
= mpol_rebind_default
,
351 [MPOL_INTERLEAVE
] = {
352 .create
= mpol_new_interleave
,
353 .rebind
= mpol_rebind_nodemask
,
356 .create
= mpol_new_preferred
,
357 .rebind
= mpol_rebind_preferred
,
360 .create
= mpol_new_bind
,
361 .rebind
= mpol_rebind_nodemask
,
365 static void gather_stats(struct page
*, void *, int pte_dirty
);
366 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
367 unsigned long flags
);
369 /* Scan through pages checking if pages follow certain conditions. */
370 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
371 unsigned long addr
, unsigned long end
,
372 const nodemask_t
*nodes
, unsigned long flags
,
379 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
384 if (!pte_present(*pte
))
386 page
= vm_normal_page(vma
, addr
, *pte
);
390 * The check for PageReserved here is important to avoid
391 * handling zero pages and other pages that may have been
392 * marked special by the system.
394 * If the PageReserved would not be checked here then f.e.
395 * the location of the zero page could have an influence
396 * on MPOL_MF_STRICT, zero pages would be counted for
397 * the per node stats, and there would be useless attempts
398 * to put zero pages on the migration list.
400 if (PageReserved(page
))
402 nid
= page_to_nid(page
);
403 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
406 if (flags
& MPOL_MF_STATS
)
407 gather_stats(page
, private, pte_dirty(*pte
));
408 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
409 migrate_page_add(page
, private, flags
);
412 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
413 pte_unmap_unlock(orig_pte
, ptl
);
417 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
418 unsigned long addr
, unsigned long end
,
419 const nodemask_t
*nodes
, unsigned long flags
,
425 pmd
= pmd_offset(pud
, addr
);
427 next
= pmd_addr_end(addr
, end
);
428 if (pmd_none_or_clear_bad(pmd
))
430 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
433 } while (pmd
++, addr
= next
, addr
!= end
);
437 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
438 unsigned long addr
, unsigned long end
,
439 const nodemask_t
*nodes
, unsigned long flags
,
445 pud
= pud_offset(pgd
, addr
);
447 next
= pud_addr_end(addr
, end
);
448 if (pud_none_or_clear_bad(pud
))
450 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
453 } while (pud
++, addr
= next
, addr
!= end
);
457 static inline int check_pgd_range(struct vm_area_struct
*vma
,
458 unsigned long addr
, unsigned long end
,
459 const nodemask_t
*nodes
, unsigned long flags
,
465 pgd
= pgd_offset(vma
->vm_mm
, addr
);
467 next
= pgd_addr_end(addr
, end
);
468 if (pgd_none_or_clear_bad(pgd
))
470 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
473 } while (pgd
++, addr
= next
, addr
!= end
);
478 * Check if all pages in a range are on a set of nodes.
479 * If pagelist != NULL then isolate pages from the LRU and
480 * put them on the pagelist.
482 static struct vm_area_struct
*
483 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
484 const nodemask_t
*nodes
, unsigned long flags
, void *private)
487 struct vm_area_struct
*first
, *vma
, *prev
;
489 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
491 err
= migrate_prep();
496 first
= find_vma(mm
, start
);
498 return ERR_PTR(-EFAULT
);
500 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
501 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
502 if (!vma
->vm_next
&& vma
->vm_end
< end
)
503 return ERR_PTR(-EFAULT
);
504 if (prev
&& prev
->vm_end
< vma
->vm_start
)
505 return ERR_PTR(-EFAULT
);
507 if (!is_vm_hugetlb_page(vma
) &&
508 ((flags
& MPOL_MF_STRICT
) ||
509 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
510 vma_migratable(vma
)))) {
511 unsigned long endvma
= vma
->vm_end
;
515 if (vma
->vm_start
> start
)
516 start
= vma
->vm_start
;
517 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
520 first
= ERR_PTR(err
);
529 /* Apply policy to a single VMA */
530 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
533 struct mempolicy
*old
= vma
->vm_policy
;
535 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
536 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
537 vma
->vm_ops
, vma
->vm_file
,
538 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
540 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
541 err
= vma
->vm_ops
->set_policy(vma
, new);
544 vma
->vm_policy
= new;
550 /* Step 2: apply policy to a range and do splits. */
551 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
552 unsigned long end
, struct mempolicy
*new)
554 struct vm_area_struct
*next
;
558 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
560 if (vma
->vm_start
< start
)
561 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
562 if (!err
&& vma
->vm_end
> end
)
563 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
565 err
= policy_vma(vma
, new);
573 * Update task->flags PF_MEMPOLICY bit: set iff non-default
574 * mempolicy. Allows more rapid checking of this (combined perhaps
575 * with other PF_* flag bits) on memory allocation hot code paths.
577 * If called from outside this file, the task 'p' should -only- be
578 * a newly forked child not yet visible on the task list, because
579 * manipulating the task flags of a visible task is not safe.
581 * The above limitation is why this routine has the funny name
582 * mpol_fix_fork_child_flag().
584 * It is also safe to call this with a task pointer of current,
585 * which the static wrapper mpol_set_task_struct_flag() does,
586 * for use within this file.
589 void mpol_fix_fork_child_flag(struct task_struct
*p
)
592 p
->flags
|= PF_MEMPOLICY
;
594 p
->flags
&= ~PF_MEMPOLICY
;
597 static void mpol_set_task_struct_flag(void)
599 mpol_fix_fork_child_flag(current
);
602 /* Set the process memory policy */
603 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
606 struct mempolicy
*new;
607 struct mm_struct
*mm
= current
->mm
;
609 new = mpol_new(mode
, flags
, nodes
);
614 * prevent changing our mempolicy while show_numa_maps()
616 * Note: do_set_mempolicy() can be called at init time
620 down_write(&mm
->mmap_sem
);
621 mpol_put(current
->mempolicy
);
622 current
->mempolicy
= new;
623 mpol_set_task_struct_flag();
624 if (new && new->mode
== MPOL_INTERLEAVE
&&
625 nodes_weight(new->v
.nodes
))
626 current
->il_next
= first_node(new->v
.nodes
);
628 up_write(&mm
->mmap_sem
);
634 * Return nodemask for policy for get_mempolicy() query
636 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
639 if (p
== &default_policy
)
645 case MPOL_INTERLEAVE
:
649 if (!(p
->flags
& MPOL_F_LOCAL
))
650 node_set(p
->v
.preferred_node
, *nodes
);
651 /* else return empty node mask for local allocation */
658 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
663 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
665 err
= page_to_nid(p
);
671 /* Retrieve NUMA policy */
672 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
673 unsigned long addr
, unsigned long flags
)
676 struct mm_struct
*mm
= current
->mm
;
677 struct vm_area_struct
*vma
= NULL
;
678 struct mempolicy
*pol
= current
->mempolicy
;
680 cpuset_update_task_memory_state();
682 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
685 if (flags
& MPOL_F_MEMS_ALLOWED
) {
686 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
688 *policy
= 0; /* just so it's initialized */
689 *nmask
= cpuset_current_mems_allowed
;
693 if (flags
& MPOL_F_ADDR
) {
695 * Do NOT fall back to task policy if the
696 * vma/shared policy at addr is NULL. We
697 * want to return MPOL_DEFAULT in this case.
699 down_read(&mm
->mmap_sem
);
700 vma
= find_vma_intersection(mm
, addr
, addr
+1);
702 up_read(&mm
->mmap_sem
);
705 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
706 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
708 pol
= vma
->vm_policy
;
713 pol
= &default_policy
; /* indicates default behavior */
715 if (flags
& MPOL_F_NODE
) {
716 if (flags
& MPOL_F_ADDR
) {
717 err
= lookup_node(mm
, addr
);
721 } else if (pol
== current
->mempolicy
&&
722 pol
->mode
== MPOL_INTERLEAVE
) {
723 *policy
= current
->il_next
;
729 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
731 *policy
|= pol
->flags
;
735 up_read(¤t
->mm
->mmap_sem
);
741 get_policy_nodemask(pol
, nmask
);
746 up_read(¤t
->mm
->mmap_sem
);
750 #ifdef CONFIG_MIGRATION
754 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
758 * Avoid migrating a page that is shared with others.
760 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1)
761 isolate_lru_page(page
, pagelist
);
764 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
766 return alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
770 * Migrate pages from one node to a target node.
771 * Returns error or the number of pages not migrated.
773 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
781 node_set(source
, nmask
);
783 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nmask
,
784 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
786 if (!list_empty(&pagelist
))
787 err
= migrate_pages(&pagelist
, new_node_page
, dest
);
793 * Move pages between the two nodesets so as to preserve the physical
794 * layout as much as possible.
796 * Returns the number of page that could not be moved.
798 int do_migrate_pages(struct mm_struct
*mm
,
799 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
806 down_read(&mm
->mmap_sem
);
808 err
= migrate_vmas(mm
, from_nodes
, to_nodes
, flags
);
813 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
814 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
815 * bit in 'tmp', and return that <source, dest> pair for migration.
816 * The pair of nodemasks 'to' and 'from' define the map.
818 * If no pair of bits is found that way, fallback to picking some
819 * pair of 'source' and 'dest' bits that are not the same. If the
820 * 'source' and 'dest' bits are the same, this represents a node
821 * that will be migrating to itself, so no pages need move.
823 * If no bits are left in 'tmp', or if all remaining bits left
824 * in 'tmp' correspond to the same bit in 'to', return false
825 * (nothing left to migrate).
827 * This lets us pick a pair of nodes to migrate between, such that
828 * if possible the dest node is not already occupied by some other
829 * source node, minimizing the risk of overloading the memory on a
830 * node that would happen if we migrated incoming memory to a node
831 * before migrating outgoing memory source that same node.
833 * A single scan of tmp is sufficient. As we go, we remember the
834 * most recent <s, d> pair that moved (s != d). If we find a pair
835 * that not only moved, but what's better, moved to an empty slot
836 * (d is not set in tmp), then we break out then, with that pair.
837 * Otherwise when we finish scannng from_tmp, we at least have the
838 * most recent <s, d> pair that moved. If we get all the way through
839 * the scan of tmp without finding any node that moved, much less
840 * moved to an empty node, then there is nothing left worth migrating.
844 while (!nodes_empty(tmp
)) {
849 for_each_node_mask(s
, tmp
) {
850 d
= node_remap(s
, *from_nodes
, *to_nodes
);
854 source
= s
; /* Node moved. Memorize */
857 /* dest not in remaining from nodes? */
858 if (!node_isset(dest
, tmp
))
864 node_clear(source
, tmp
);
865 err
= migrate_to_node(mm
, source
, dest
, flags
);
872 up_read(&mm
->mmap_sem
);
880 * Allocate a new page for page migration based on vma policy.
881 * Start assuming that page is mapped by vma pointed to by @private.
882 * Search forward from there, if not. N.B., this assumes that the
883 * list of pages handed to migrate_pages()--which is how we get here--
884 * is in virtual address order.
886 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
888 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
889 unsigned long uninitialized_var(address
);
892 address
= page_address_in_vma(page
, vma
);
893 if (address
!= -EFAULT
)
899 * if !vma, alloc_page_vma() will use task or system default policy
901 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
905 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
910 int do_migrate_pages(struct mm_struct
*mm
,
911 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
916 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
922 static long do_mbind(unsigned long start
, unsigned long len
,
923 unsigned short mode
, unsigned short mode_flags
,
924 nodemask_t
*nmask
, unsigned long flags
)
926 struct vm_area_struct
*vma
;
927 struct mm_struct
*mm
= current
->mm
;
928 struct mempolicy
*new;
933 if (flags
& ~(unsigned long)(MPOL_MF_STRICT
|
934 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
936 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
939 if (start
& ~PAGE_MASK
)
942 if (mode
== MPOL_DEFAULT
)
943 flags
&= ~MPOL_MF_STRICT
;
945 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
953 new = mpol_new(mode
, mode_flags
, nmask
);
958 * If we are using the default policy then operation
959 * on discontinuous address spaces is okay after all
962 flags
|= MPOL_MF_DISCONTIG_OK
;
964 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
965 start
, start
+ len
, mode
, mode_flags
,
966 nmask
? nodes_addr(*nmask
)[0] : -1);
968 down_write(&mm
->mmap_sem
);
969 vma
= check_range(mm
, start
, end
, nmask
,
970 flags
| MPOL_MF_INVERT
, &pagelist
);
976 err
= mbind_range(vma
, start
, end
, new);
978 if (!list_empty(&pagelist
))
979 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
982 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
986 up_write(&mm
->mmap_sem
);
992 * User space interface with variable sized bitmaps for nodelists.
995 /* Copy a node mask from user space. */
996 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
997 unsigned long maxnode
)
1000 unsigned long nlongs
;
1001 unsigned long endmask
;
1004 nodes_clear(*nodes
);
1005 if (maxnode
== 0 || !nmask
)
1007 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1010 nlongs
= BITS_TO_LONGS(maxnode
);
1011 if ((maxnode
% BITS_PER_LONG
) == 0)
1014 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1016 /* When the user specified more nodes than supported just check
1017 if the non supported part is all zero. */
1018 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1019 if (nlongs
> PAGE_SIZE
/sizeof(long))
1021 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1023 if (get_user(t
, nmask
+ k
))
1025 if (k
== nlongs
- 1) {
1031 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1035 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1037 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1041 /* Copy a kernel node mask to user space */
1042 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1045 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1046 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1048 if (copy
> nbytes
) {
1049 if (copy
> PAGE_SIZE
)
1051 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1055 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1058 asmlinkage
long sys_mbind(unsigned long start
, unsigned long len
,
1060 unsigned long __user
*nmask
, unsigned long maxnode
,
1065 unsigned short mode_flags
;
1067 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1068 mode
&= ~MPOL_MODE_FLAGS
;
1069 if (mode
>= MPOL_MAX
)
1071 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1072 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1074 err
= get_nodes(&nodes
, nmask
, maxnode
);
1077 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1080 /* Set the process memory policy */
1081 asmlinkage
long sys_set_mempolicy(int mode
, unsigned long __user
*nmask
,
1082 unsigned long maxnode
)
1086 unsigned short flags
;
1088 flags
= mode
& MPOL_MODE_FLAGS
;
1089 mode
&= ~MPOL_MODE_FLAGS
;
1090 if ((unsigned int)mode
>= MPOL_MAX
)
1092 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1094 err
= get_nodes(&nodes
, nmask
, maxnode
);
1097 return do_set_mempolicy(mode
, flags
, &nodes
);
1100 asmlinkage
long sys_migrate_pages(pid_t pid
, unsigned long maxnode
,
1101 const unsigned long __user
*old_nodes
,
1102 const unsigned long __user
*new_nodes
)
1104 struct mm_struct
*mm
;
1105 struct task_struct
*task
;
1108 nodemask_t task_nodes
;
1111 err
= get_nodes(&old
, old_nodes
, maxnode
);
1115 err
= get_nodes(&new, new_nodes
, maxnode
);
1119 /* Find the mm_struct */
1120 read_lock(&tasklist_lock
);
1121 task
= pid
? find_task_by_vpid(pid
) : current
;
1123 read_unlock(&tasklist_lock
);
1126 mm
= get_task_mm(task
);
1127 read_unlock(&tasklist_lock
);
1133 * Check if this process has the right to modify the specified
1134 * process. The right exists if the process has administrative
1135 * capabilities, superuser privileges or the same
1136 * userid as the target process.
1138 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
1139 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
1140 !capable(CAP_SYS_NICE
)) {
1145 task_nodes
= cpuset_mems_allowed(task
);
1146 /* Is the user allowed to access the target nodes? */
1147 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1152 if (!nodes_subset(new, node_states
[N_HIGH_MEMORY
])) {
1157 err
= security_task_movememory(task
);
1161 err
= do_migrate_pages(mm
, &old
, &new,
1162 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1169 /* Retrieve NUMA policy */
1170 asmlinkage
long sys_get_mempolicy(int __user
*policy
,
1171 unsigned long __user
*nmask
,
1172 unsigned long maxnode
,
1173 unsigned long addr
, unsigned long flags
)
1176 int uninitialized_var(pval
);
1179 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1182 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1187 if (policy
&& put_user(pval
, policy
))
1191 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1196 #ifdef CONFIG_COMPAT
1198 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1199 compat_ulong_t __user
*nmask
,
1200 compat_ulong_t maxnode
,
1201 compat_ulong_t addr
, compat_ulong_t flags
)
1204 unsigned long __user
*nm
= NULL
;
1205 unsigned long nr_bits
, alloc_size
;
1206 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1208 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1209 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1212 nm
= compat_alloc_user_space(alloc_size
);
1214 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1216 if (!err
&& nmask
) {
1217 err
= copy_from_user(bm
, nm
, alloc_size
);
1218 /* ensure entire bitmap is zeroed */
1219 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1220 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1226 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1227 compat_ulong_t maxnode
)
1230 unsigned long __user
*nm
= NULL
;
1231 unsigned long nr_bits
, alloc_size
;
1232 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1234 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1235 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1238 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1239 nm
= compat_alloc_user_space(alloc_size
);
1240 err
|= copy_to_user(nm
, bm
, alloc_size
);
1246 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1249 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1250 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1251 compat_ulong_t maxnode
, compat_ulong_t flags
)
1254 unsigned long __user
*nm
= NULL
;
1255 unsigned long nr_bits
, alloc_size
;
1258 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1259 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1262 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1263 nm
= compat_alloc_user_space(alloc_size
);
1264 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1270 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1276 * get_vma_policy(@task, @vma, @addr)
1277 * @task - task for fallback if vma policy == default
1278 * @vma - virtual memory area whose policy is sought
1279 * @addr - address in @vma for shared policy lookup
1281 * Returns effective policy for a VMA at specified address.
1282 * Falls back to @task or system default policy, as necessary.
1283 * Current or other task's task mempolicy and non-shared vma policies
1284 * are protected by the task's mmap_sem, which must be held for read by
1286 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1287 * count--added by the get_policy() vm_op, as appropriate--to protect against
1288 * freeing by another task. It is the caller's responsibility to free the
1289 * extra reference for shared policies.
1291 static struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1292 struct vm_area_struct
*vma
, unsigned long addr
)
1294 struct mempolicy
*pol
= task
->mempolicy
;
1297 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1298 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1302 } else if (vma
->vm_policy
)
1303 pol
= vma
->vm_policy
;
1306 pol
= &default_policy
;
1311 * Return a nodemask representing a mempolicy for filtering nodes for
1314 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1316 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1317 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1318 gfp_zone(gfp
) >= policy_zone
&&
1319 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1320 return &policy
->v
.nodes
;
1325 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1326 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
)
1328 int nd
= numa_node_id();
1330 switch (policy
->mode
) {
1331 case MPOL_PREFERRED
:
1332 if (!(policy
->flags
& MPOL_F_LOCAL
))
1333 nd
= policy
->v
.preferred_node
;
1337 * Normally, MPOL_BIND allocations are node-local within the
1338 * allowed nodemask. However, if __GFP_THISNODE is set and the
1339 * current node is part of the mask, we use the zonelist for
1340 * the first node in the mask instead.
1342 if (unlikely(gfp
& __GFP_THISNODE
) &&
1343 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1344 nd
= first_node(policy
->v
.nodes
);
1346 case MPOL_INTERLEAVE
: /* should not happen */
1351 return node_zonelist(nd
, gfp
);
1354 /* Do dynamic interleaving for a process */
1355 static unsigned interleave_nodes(struct mempolicy
*policy
)
1358 struct task_struct
*me
= current
;
1361 next
= next_node(nid
, policy
->v
.nodes
);
1362 if (next
>= MAX_NUMNODES
)
1363 next
= first_node(policy
->v
.nodes
);
1364 if (next
< MAX_NUMNODES
)
1370 * Depending on the memory policy provide a node from which to allocate the
1372 * @policy must be protected by freeing by the caller. If @policy is
1373 * the current task's mempolicy, this protection is implicit, as only the
1374 * task can change it's policy. The system default policy requires no
1377 unsigned slab_node(struct mempolicy
*policy
)
1379 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1380 return numa_node_id();
1382 switch (policy
->mode
) {
1383 case MPOL_PREFERRED
:
1385 * handled MPOL_F_LOCAL above
1387 return policy
->v
.preferred_node
;
1389 case MPOL_INTERLEAVE
:
1390 return interleave_nodes(policy
);
1394 * Follow bind policy behavior and start allocation at the
1397 struct zonelist
*zonelist
;
1399 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1400 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1401 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1412 /* Do static interleaving for a VMA with known offset. */
1413 static unsigned offset_il_node(struct mempolicy
*pol
,
1414 struct vm_area_struct
*vma
, unsigned long off
)
1416 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1422 return numa_node_id();
1423 target
= (unsigned int)off
% nnodes
;
1426 nid
= next_node(nid
, pol
->v
.nodes
);
1428 } while (c
<= target
);
1432 /* Determine a node number for interleave */
1433 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1434 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1440 * for small pages, there is no difference between
1441 * shift and PAGE_SHIFT, so the bit-shift is safe.
1442 * for huge pages, since vm_pgoff is in units of small
1443 * pages, we need to shift off the always 0 bits to get
1446 BUG_ON(shift
< PAGE_SHIFT
);
1447 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1448 off
+= (addr
- vma
->vm_start
) >> shift
;
1449 return offset_il_node(pol
, vma
, off
);
1451 return interleave_nodes(pol
);
1454 #ifdef CONFIG_HUGETLBFS
1456 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1457 * @vma = virtual memory area whose policy is sought
1458 * @addr = address in @vma for shared policy lookup and interleave policy
1459 * @gfp_flags = for requested zone
1460 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1461 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1463 * Returns a zonelist suitable for a huge page allocation and a pointer
1464 * to the struct mempolicy for conditional unref after allocation.
1465 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1466 * @nodemask for filtering the zonelist.
1468 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1469 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1470 nodemask_t
**nodemask
)
1472 struct zonelist
*zl
;
1474 *mpol
= get_vma_policy(current
, vma
, addr
);
1475 *nodemask
= NULL
; /* assume !MPOL_BIND */
1477 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1478 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1479 HPAGE_SHIFT
), gfp_flags
);
1481 zl
= policy_zonelist(gfp_flags
, *mpol
);
1482 if ((*mpol
)->mode
== MPOL_BIND
)
1483 *nodemask
= &(*mpol
)->v
.nodes
;
1489 /* Allocate a page in interleaved policy.
1490 Own path because it needs to do special accounting. */
1491 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1494 struct zonelist
*zl
;
1497 zl
= node_zonelist(nid
, gfp
);
1498 page
= __alloc_pages(gfp
, order
, zl
);
1499 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1500 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1505 * alloc_page_vma - Allocate a page for a VMA.
1508 * %GFP_USER user allocation.
1509 * %GFP_KERNEL kernel allocations,
1510 * %GFP_HIGHMEM highmem/user allocations,
1511 * %GFP_FS allocation should not call back into a file system.
1512 * %GFP_ATOMIC don't sleep.
1514 * @vma: Pointer to VMA or NULL if not available.
1515 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1517 * This function allocates a page from the kernel page pool and applies
1518 * a NUMA policy associated with the VMA or the current process.
1519 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1520 * mm_struct of the VMA to prevent it from going away. Should be used for
1521 * all allocations for pages that will be mapped into
1522 * user space. Returns NULL when no page can be allocated.
1524 * Should be called with the mm_sem of the vma hold.
1527 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1529 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1530 struct zonelist
*zl
;
1532 cpuset_update_task_memory_state();
1534 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
1537 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1539 return alloc_page_interleave(gfp
, 0, nid
);
1541 zl
= policy_zonelist(gfp
, pol
);
1542 if (unlikely(mpol_needs_cond_ref(pol
))) {
1544 * slow path: ref counted shared policy
1546 struct page
*page
= __alloc_pages_nodemask(gfp
, 0,
1547 zl
, policy_nodemask(gfp
, pol
));
1552 * fast path: default or task policy
1554 return __alloc_pages_nodemask(gfp
, 0, zl
, policy_nodemask(gfp
, pol
));
1558 * alloc_pages_current - Allocate pages.
1561 * %GFP_USER user allocation,
1562 * %GFP_KERNEL kernel allocation,
1563 * %GFP_HIGHMEM highmem allocation,
1564 * %GFP_FS don't call back into a file system.
1565 * %GFP_ATOMIC don't sleep.
1566 * @order: Power of two of allocation size in pages. 0 is a single page.
1568 * Allocate a page from the kernel page pool. When not in
1569 * interrupt context and apply the current process NUMA policy.
1570 * Returns NULL when no page can be allocated.
1572 * Don't call cpuset_update_task_memory_state() unless
1573 * 1) it's ok to take cpuset_sem (can WAIT), and
1574 * 2) allocating for current task (not interrupt).
1576 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1578 struct mempolicy
*pol
= current
->mempolicy
;
1580 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1581 cpuset_update_task_memory_state();
1582 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
1583 pol
= &default_policy
;
1586 * No reference counting needed for current->mempolicy
1587 * nor system default_policy
1589 if (pol
->mode
== MPOL_INTERLEAVE
)
1590 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1591 return __alloc_pages_nodemask(gfp
, order
,
1592 policy_zonelist(gfp
, pol
), policy_nodemask(gfp
, pol
));
1594 EXPORT_SYMBOL(alloc_pages_current
);
1597 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1598 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1599 * with the mems_allowed returned by cpuset_mems_allowed(). This
1600 * keeps mempolicies cpuset relative after its cpuset moves. See
1601 * further kernel/cpuset.c update_nodemask().
1604 /* Slow path of a mempolicy duplicate */
1605 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
1607 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1610 return ERR_PTR(-ENOMEM
);
1611 if (current_cpuset_is_being_rebound()) {
1612 nodemask_t mems
= cpuset_mems_allowed(current
);
1613 mpol_rebind_policy(old
, &mems
);
1616 atomic_set(&new->refcnt
, 1);
1621 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1622 * eliminate the * MPOL_F_* flags that require conditional ref and
1623 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1624 * after return. Use the returned value.
1626 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1627 * policy lookup, even if the policy needs/has extra ref on lookup.
1628 * shmem_readahead needs this.
1630 struct mempolicy
*__mpol_cond_copy(struct mempolicy
*tompol
,
1631 struct mempolicy
*frompol
)
1633 if (!mpol_needs_cond_ref(frompol
))
1637 tompol
->flags
&= ~MPOL_F_SHARED
; /* copy doesn't need unref */
1638 __mpol_put(frompol
);
1642 static int mpol_match_intent(const struct mempolicy
*a
,
1643 const struct mempolicy
*b
)
1645 if (a
->flags
!= b
->flags
)
1647 if (!mpol_store_user_nodemask(a
))
1649 return nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
);
1652 /* Slow path of a mempolicy comparison */
1653 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1657 if (a
->mode
!= b
->mode
)
1659 if (a
->mode
!= MPOL_DEFAULT
&& !mpol_match_intent(a
, b
))
1664 case MPOL_INTERLEAVE
:
1665 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1666 case MPOL_PREFERRED
:
1667 return a
->v
.preferred_node
== b
->v
.preferred_node
&&
1668 a
->flags
== b
->flags
;
1676 * Shared memory backing store policy support.
1678 * Remember policies even when nobody has shared memory mapped.
1679 * The policies are kept in Red-Black tree linked from the inode.
1680 * They are protected by the sp->lock spinlock, which should be held
1681 * for any accesses to the tree.
1684 /* lookup first element intersecting start-end */
1685 /* Caller holds sp->lock */
1686 static struct sp_node
*
1687 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1689 struct rb_node
*n
= sp
->root
.rb_node
;
1692 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1694 if (start
>= p
->end
)
1696 else if (end
<= p
->start
)
1704 struct sp_node
*w
= NULL
;
1705 struct rb_node
*prev
= rb_prev(n
);
1708 w
= rb_entry(prev
, struct sp_node
, nd
);
1709 if (w
->end
<= start
)
1713 return rb_entry(n
, struct sp_node
, nd
);
1716 /* Insert a new shared policy into the list. */
1717 /* Caller holds sp->lock */
1718 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1720 struct rb_node
**p
= &sp
->root
.rb_node
;
1721 struct rb_node
*parent
= NULL
;
1726 nd
= rb_entry(parent
, struct sp_node
, nd
);
1727 if (new->start
< nd
->start
)
1729 else if (new->end
> nd
->end
)
1730 p
= &(*p
)->rb_right
;
1734 rb_link_node(&new->nd
, parent
, p
);
1735 rb_insert_color(&new->nd
, &sp
->root
);
1736 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
1737 new->policy
? new->policy
->mode
: 0);
1740 /* Find shared policy intersecting idx */
1742 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1744 struct mempolicy
*pol
= NULL
;
1747 if (!sp
->root
.rb_node
)
1749 spin_lock(&sp
->lock
);
1750 sn
= sp_lookup(sp
, idx
, idx
+1);
1752 mpol_get(sn
->policy
);
1755 spin_unlock(&sp
->lock
);
1759 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1761 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
1762 rb_erase(&n
->nd
, &sp
->root
);
1763 mpol_put(n
->policy
);
1764 kmem_cache_free(sn_cache
, n
);
1767 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
1768 struct mempolicy
*pol
)
1770 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1777 pol
->flags
|= MPOL_F_SHARED
; /* for unref */
1782 /* Replace a policy range. */
1783 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1784 unsigned long end
, struct sp_node
*new)
1786 struct sp_node
*n
, *new2
= NULL
;
1789 spin_lock(&sp
->lock
);
1790 n
= sp_lookup(sp
, start
, end
);
1791 /* Take care of old policies in the same range. */
1792 while (n
&& n
->start
< end
) {
1793 struct rb_node
*next
= rb_next(&n
->nd
);
1794 if (n
->start
>= start
) {
1800 /* Old policy spanning whole new range. */
1803 spin_unlock(&sp
->lock
);
1804 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1810 sp_insert(sp
, new2
);
1818 n
= rb_entry(next
, struct sp_node
, nd
);
1822 spin_unlock(&sp
->lock
);
1824 mpol_put(new2
->policy
);
1825 kmem_cache_free(sn_cache
, new2
);
1830 void mpol_shared_policy_init(struct shared_policy
*info
, unsigned short policy
,
1831 unsigned short flags
, nodemask_t
*policy_nodes
)
1833 info
->root
= RB_ROOT
;
1834 spin_lock_init(&info
->lock
);
1836 if (policy
!= MPOL_DEFAULT
) {
1837 struct mempolicy
*newpol
;
1839 /* Falls back to NULL policy [MPOL_DEFAULT] on any error */
1840 newpol
= mpol_new(policy
, flags
, policy_nodes
);
1841 if (!IS_ERR(newpol
)) {
1842 /* Create pseudo-vma that contains just the policy */
1843 struct vm_area_struct pvma
;
1845 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
1846 /* Policy covers entire file */
1847 pvma
.vm_end
= TASK_SIZE
;
1848 mpol_set_shared_policy(info
, &pvma
, newpol
);
1854 int mpol_set_shared_policy(struct shared_policy
*info
,
1855 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1858 struct sp_node
*new = NULL
;
1859 unsigned long sz
= vma_pages(vma
);
1861 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1863 sz
, npol
? npol
->mode
: -1,
1864 npol
? npol
->flags
: -1,
1865 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1868 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1872 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1874 kmem_cache_free(sn_cache
, new);
1878 /* Free a backing policy store on inode delete. */
1879 void mpol_free_shared_policy(struct shared_policy
*p
)
1882 struct rb_node
*next
;
1884 if (!p
->root
.rb_node
)
1886 spin_lock(&p
->lock
);
1887 next
= rb_first(&p
->root
);
1889 n
= rb_entry(next
, struct sp_node
, nd
);
1890 next
= rb_next(&n
->nd
);
1891 rb_erase(&n
->nd
, &p
->root
);
1892 mpol_put(n
->policy
);
1893 kmem_cache_free(sn_cache
, n
);
1895 spin_unlock(&p
->lock
);
1898 /* assumes fs == KERNEL_DS */
1899 void __init
numa_policy_init(void)
1901 nodemask_t interleave_nodes
;
1902 unsigned long largest
= 0;
1903 int nid
, prefer
= 0;
1905 policy_cache
= kmem_cache_create("numa_policy",
1906 sizeof(struct mempolicy
),
1907 0, SLAB_PANIC
, NULL
);
1909 sn_cache
= kmem_cache_create("shared_policy_node",
1910 sizeof(struct sp_node
),
1911 0, SLAB_PANIC
, NULL
);
1914 * Set interleaving policy for system init. Interleaving is only
1915 * enabled across suitably sized nodes (default is >= 16MB), or
1916 * fall back to the largest node if they're all smaller.
1918 nodes_clear(interleave_nodes
);
1919 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1920 unsigned long total_pages
= node_present_pages(nid
);
1922 /* Preserve the largest node */
1923 if (largest
< total_pages
) {
1924 largest
= total_pages
;
1928 /* Interleave this node? */
1929 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
1930 node_set(nid
, interleave_nodes
);
1933 /* All too small, use the largest */
1934 if (unlikely(nodes_empty(interleave_nodes
)))
1935 node_set(prefer
, interleave_nodes
);
1937 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
1938 printk("numa_policy_init: interleaving failed\n");
1941 /* Reset policy of current process to default */
1942 void numa_default_policy(void)
1944 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
1948 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
1949 * Used only for mpol_to_str()
1951 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
1952 static const char * const policy_types
[] =
1953 { "default", "prefer", "bind", "interleave", "local" };
1956 * Convert a mempolicy into a string.
1957 * Returns the number of characters in buffer (if positive)
1958 * or an error (negative)
1960 static inline int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
1965 unsigned short mode
;
1966 unsigned short flags
= pol
? pol
->flags
: 0;
1969 * Sanity check: room for longest mode, flag and some nodes
1971 VM_BUG_ON(maxlen
< strlen("interleave") + strlen("relative") + 16);
1973 if (!pol
|| pol
== &default_policy
)
1974 mode
= MPOL_DEFAULT
;
1983 case MPOL_PREFERRED
:
1985 if (flags
& MPOL_F_LOCAL
)
1986 mode
= MPOL_LOCAL
; /* pseudo-policy */
1988 node_set(pol
->v
.preferred_node
, nodes
);
1993 case MPOL_INTERLEAVE
:
1994 nodes
= pol
->v
.nodes
;
2001 l
= strlen(policy_types
[mode
]);
2002 if (buffer
+ maxlen
< p
+ l
+ 1)
2005 strcpy(p
, policy_types
[mode
]);
2008 if (flags
& MPOL_MODE_FLAGS
) {
2009 if (buffer
+ maxlen
< p
+ 2)
2014 * Currently, the only defined flags are mutually exclusive
2016 if (flags
& MPOL_F_STATIC_NODES
)
2017 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2018 else if (flags
& MPOL_F_RELATIVE_NODES
)
2019 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2022 if (!nodes_empty(nodes
)) {
2023 if (buffer
+ maxlen
< p
+ 2)
2026 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
2032 unsigned long pages
;
2034 unsigned long active
;
2035 unsigned long writeback
;
2036 unsigned long mapcount_max
;
2037 unsigned long dirty
;
2038 unsigned long swapcache
;
2039 unsigned long node
[MAX_NUMNODES
];
2042 static void gather_stats(struct page
*page
, void *private, int pte_dirty
)
2044 struct numa_maps
*md
= private;
2045 int count
= page_mapcount(page
);
2048 if (pte_dirty
|| PageDirty(page
))
2051 if (PageSwapCache(page
))
2054 if (PageActive(page
))
2057 if (PageWriteback(page
))
2063 if (count
> md
->mapcount_max
)
2064 md
->mapcount_max
= count
;
2066 md
->node
[page_to_nid(page
)]++;
2069 #ifdef CONFIG_HUGETLB_PAGE
2070 static void check_huge_range(struct vm_area_struct
*vma
,
2071 unsigned long start
, unsigned long end
,
2072 struct numa_maps
*md
)
2077 for (addr
= start
; addr
< end
; addr
+= HPAGE_SIZE
) {
2078 pte_t
*ptep
= huge_pte_offset(vma
->vm_mm
, addr
& HPAGE_MASK
);
2088 page
= pte_page(pte
);
2092 gather_stats(page
, md
, pte_dirty(*ptep
));
2096 static inline void check_huge_range(struct vm_area_struct
*vma
,
2097 unsigned long start
, unsigned long end
,
2098 struct numa_maps
*md
)
2104 * Display pages allocated per node and memory policy via /proc.
2106 int show_numa_map(struct seq_file
*m
, void *v
)
2108 struct proc_maps_private
*priv
= m
->private;
2109 struct vm_area_struct
*vma
= v
;
2110 struct numa_maps
*md
;
2111 struct file
*file
= vma
->vm_file
;
2112 struct mm_struct
*mm
= vma
->vm_mm
;
2113 struct mempolicy
*pol
;
2120 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
2124 pol
= get_vma_policy(priv
->task
, vma
, vma
->vm_start
);
2125 mpol_to_str(buffer
, sizeof(buffer
), pol
);
2128 seq_printf(m
, "%08lx %s", vma
->vm_start
, buffer
);
2131 seq_printf(m
, " file=");
2132 seq_path(m
, &file
->f_path
, "\n\t= ");
2133 } else if (vma
->vm_start
<= mm
->brk
&& vma
->vm_end
>= mm
->start_brk
) {
2134 seq_printf(m
, " heap");
2135 } else if (vma
->vm_start
<= mm
->start_stack
&&
2136 vma
->vm_end
>= mm
->start_stack
) {
2137 seq_printf(m
, " stack");
2140 if (is_vm_hugetlb_page(vma
)) {
2141 check_huge_range(vma
, vma
->vm_start
, vma
->vm_end
, md
);
2142 seq_printf(m
, " huge");
2144 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
2145 &node_states
[N_HIGH_MEMORY
], MPOL_MF_STATS
, md
);
2152 seq_printf(m
," anon=%lu",md
->anon
);
2155 seq_printf(m
," dirty=%lu",md
->dirty
);
2157 if (md
->pages
!= md
->anon
&& md
->pages
!= md
->dirty
)
2158 seq_printf(m
, " mapped=%lu", md
->pages
);
2160 if (md
->mapcount_max
> 1)
2161 seq_printf(m
, " mapmax=%lu", md
->mapcount_max
);
2164 seq_printf(m
," swapcache=%lu", md
->swapcache
);
2166 if (md
->active
< md
->pages
&& !is_vm_hugetlb_page(vma
))
2167 seq_printf(m
," active=%lu", md
->active
);
2170 seq_printf(m
," writeback=%lu", md
->writeback
);
2172 for_each_node_state(n
, N_HIGH_MEMORY
)
2174 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
2179 if (m
->count
< m
->size
)
2180 m
->version
= (vma
!= priv
->tail_vma
) ? vma
->vm_start
: 0;