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.
66 could replace all the switch()es with a mempolicy_ops structure.
69 #include <linux/mempolicy.h>
71 #include <linux/highmem.h>
72 #include <linux/hugetlb.h>
73 #include <linux/kernel.h>
74 #include <linux/sched.h>
75 #include <linux/nodemask.h>
76 #include <linux/cpuset.h>
77 #include <linux/gfp.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/module.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/rmap.h>
90 #include <linux/security.h>
91 #include <linux/syscalls.h>
93 #include <asm/tlbflush.h>
94 #include <asm/uaccess.h>
97 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
98 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
99 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
101 static struct kmem_cache
*policy_cache
;
102 static struct kmem_cache
*sn_cache
;
104 /* Highest zone. An specific allocation for a zone below that is not
106 enum zone_type policy_zone
= 0;
108 struct mempolicy default_policy
= {
109 .refcnt
= ATOMIC_INIT(1), /* never free it */
110 .policy
= MPOL_DEFAULT
,
113 static void mpol_rebind_policy(struct mempolicy
*pol
,
114 const nodemask_t
*newmask
);
116 /* Check that the nodemask contains at least one populated zone */
117 static int is_valid_nodemask(nodemask_t
*nodemask
)
121 /* Check that there is something useful in this mask */
124 for_each_node_mask(nd
, *nodemask
) {
127 for (k
= 0; k
<= policy_zone
; k
++) {
128 z
= &NODE_DATA(nd
)->node_zones
[k
];
129 if (z
->present_pages
> 0)
137 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
139 return pol
->flags
& (MPOL_F_STATIC_NODES
| MPOL_F_RELATIVE_NODES
);
142 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
143 const nodemask_t
*rel
)
146 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
147 nodes_onto(*ret
, tmp
, *rel
);
150 /* Create a new policy */
151 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
154 struct mempolicy
*policy
;
155 nodemask_t cpuset_context_nmask
;
157 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
158 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : -1);
160 if (mode
== MPOL_DEFAULT
)
161 return (nodes
&& nodes_weight(*nodes
)) ? ERR_PTR(-EINVAL
) :
163 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
165 return ERR_PTR(-ENOMEM
);
166 atomic_set(&policy
->refcnt
, 1);
167 cpuset_update_task_memory_state();
168 if (flags
& MPOL_F_RELATIVE_NODES
)
169 mpol_relative_nodemask(&cpuset_context_nmask
, nodes
,
170 &cpuset_current_mems_allowed
);
172 nodes_and(cpuset_context_nmask
, *nodes
,
173 cpuset_current_mems_allowed
);
175 case MPOL_INTERLEAVE
:
176 if (nodes_empty(*nodes
) || nodes_empty(cpuset_context_nmask
))
178 policy
->v
.nodes
= cpuset_context_nmask
;
181 policy
->v
.preferred_node
= first_node(cpuset_context_nmask
);
182 if (policy
->v
.preferred_node
>= MAX_NUMNODES
)
186 if (!is_valid_nodemask(&cpuset_context_nmask
))
188 policy
->v
.nodes
= cpuset_context_nmask
;
193 policy
->policy
= mode
;
194 policy
->flags
= flags
;
195 if (mpol_store_user_nodemask(policy
))
196 policy
->w
.user_nodemask
= *nodes
;
198 policy
->w
.cpuset_mems_allowed
= cpuset_mems_allowed(current
);
202 kmem_cache_free(policy_cache
, policy
);
203 return ERR_PTR(-EINVAL
);
206 static void gather_stats(struct page
*, void *, int pte_dirty
);
207 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
208 unsigned long flags
);
210 /* Scan through pages checking if pages follow certain conditions. */
211 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
212 unsigned long addr
, unsigned long end
,
213 const nodemask_t
*nodes
, unsigned long flags
,
220 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
225 if (!pte_present(*pte
))
227 page
= vm_normal_page(vma
, addr
, *pte
);
231 * The check for PageReserved here is important to avoid
232 * handling zero pages and other pages that may have been
233 * marked special by the system.
235 * If the PageReserved would not be checked here then f.e.
236 * the location of the zero page could have an influence
237 * on MPOL_MF_STRICT, zero pages would be counted for
238 * the per node stats, and there would be useless attempts
239 * to put zero pages on the migration list.
241 if (PageReserved(page
))
243 nid
= page_to_nid(page
);
244 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
247 if (flags
& MPOL_MF_STATS
)
248 gather_stats(page
, private, pte_dirty(*pte
));
249 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
250 migrate_page_add(page
, private, flags
);
253 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
254 pte_unmap_unlock(orig_pte
, ptl
);
258 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
259 unsigned long addr
, unsigned long end
,
260 const nodemask_t
*nodes
, unsigned long flags
,
266 pmd
= pmd_offset(pud
, addr
);
268 next
= pmd_addr_end(addr
, end
);
269 if (pmd_none_or_clear_bad(pmd
))
271 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
274 } while (pmd
++, addr
= next
, addr
!= end
);
278 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
279 unsigned long addr
, unsigned long end
,
280 const nodemask_t
*nodes
, unsigned long flags
,
286 pud
= pud_offset(pgd
, addr
);
288 next
= pud_addr_end(addr
, end
);
289 if (pud_none_or_clear_bad(pud
))
291 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
294 } while (pud
++, addr
= next
, addr
!= end
);
298 static inline int check_pgd_range(struct vm_area_struct
*vma
,
299 unsigned long addr
, unsigned long end
,
300 const nodemask_t
*nodes
, unsigned long flags
,
306 pgd
= pgd_offset(vma
->vm_mm
, addr
);
308 next
= pgd_addr_end(addr
, end
);
309 if (pgd_none_or_clear_bad(pgd
))
311 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
314 } while (pgd
++, addr
= next
, addr
!= end
);
319 * Check if all pages in a range are on a set of nodes.
320 * If pagelist != NULL then isolate pages from the LRU and
321 * put them on the pagelist.
323 static struct vm_area_struct
*
324 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
325 const nodemask_t
*nodes
, unsigned long flags
, void *private)
328 struct vm_area_struct
*first
, *vma
, *prev
;
330 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
332 err
= migrate_prep();
337 first
= find_vma(mm
, start
);
339 return ERR_PTR(-EFAULT
);
341 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
342 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
343 if (!vma
->vm_next
&& vma
->vm_end
< end
)
344 return ERR_PTR(-EFAULT
);
345 if (prev
&& prev
->vm_end
< vma
->vm_start
)
346 return ERR_PTR(-EFAULT
);
348 if (!is_vm_hugetlb_page(vma
) &&
349 ((flags
& MPOL_MF_STRICT
) ||
350 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
351 vma_migratable(vma
)))) {
352 unsigned long endvma
= vma
->vm_end
;
356 if (vma
->vm_start
> start
)
357 start
= vma
->vm_start
;
358 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
361 first
= ERR_PTR(err
);
370 /* Apply policy to a single VMA */
371 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
374 struct mempolicy
*old
= vma
->vm_policy
;
376 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
377 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
378 vma
->vm_ops
, vma
->vm_file
,
379 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
381 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
382 err
= vma
->vm_ops
->set_policy(vma
, new);
385 vma
->vm_policy
= new;
391 /* Step 2: apply policy to a range and do splits. */
392 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
393 unsigned long end
, struct mempolicy
*new)
395 struct vm_area_struct
*next
;
399 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
401 if (vma
->vm_start
< start
)
402 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
403 if (!err
&& vma
->vm_end
> end
)
404 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
406 err
= policy_vma(vma
, new);
414 * Update task->flags PF_MEMPOLICY bit: set iff non-default
415 * mempolicy. Allows more rapid checking of this (combined perhaps
416 * with other PF_* flag bits) on memory allocation hot code paths.
418 * If called from outside this file, the task 'p' should -only- be
419 * a newly forked child not yet visible on the task list, because
420 * manipulating the task flags of a visible task is not safe.
422 * The above limitation is why this routine has the funny name
423 * mpol_fix_fork_child_flag().
425 * It is also safe to call this with a task pointer of current,
426 * which the static wrapper mpol_set_task_struct_flag() does,
427 * for use within this file.
430 void mpol_fix_fork_child_flag(struct task_struct
*p
)
433 p
->flags
|= PF_MEMPOLICY
;
435 p
->flags
&= ~PF_MEMPOLICY
;
438 static void mpol_set_task_struct_flag(void)
440 mpol_fix_fork_child_flag(current
);
443 /* Set the process memory policy */
444 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
447 struct mempolicy
*new;
449 new = mpol_new(mode
, flags
, nodes
);
452 mpol_free(current
->mempolicy
);
453 current
->mempolicy
= new;
454 mpol_set_task_struct_flag();
455 if (new && new->policy
== MPOL_INTERLEAVE
&&
456 nodes_weight(new->v
.nodes
))
457 current
->il_next
= first_node(new->v
.nodes
);
461 /* Fill a zone bitmap for a policy */
462 static void get_zonemask(struct mempolicy
*p
, nodemask_t
*nodes
)
470 case MPOL_INTERLEAVE
:
474 /* or use current node instead of memory_map? */
475 if (p
->v
.preferred_node
< 0)
476 *nodes
= node_states
[N_HIGH_MEMORY
];
478 node_set(p
->v
.preferred_node
, *nodes
);
485 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
490 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
492 err
= page_to_nid(p
);
498 /* Retrieve NUMA policy */
499 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
500 unsigned long addr
, unsigned long flags
)
503 struct mm_struct
*mm
= current
->mm
;
504 struct vm_area_struct
*vma
= NULL
;
505 struct mempolicy
*pol
= current
->mempolicy
;
507 cpuset_update_task_memory_state();
509 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
512 if (flags
& MPOL_F_MEMS_ALLOWED
) {
513 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
515 *policy
= 0; /* just so it's initialized */
516 *nmask
= cpuset_current_mems_allowed
;
520 if (flags
& MPOL_F_ADDR
) {
521 down_read(&mm
->mmap_sem
);
522 vma
= find_vma_intersection(mm
, addr
, addr
+1);
524 up_read(&mm
->mmap_sem
);
527 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
528 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
530 pol
= vma
->vm_policy
;
535 pol
= &default_policy
;
537 if (flags
& MPOL_F_NODE
) {
538 if (flags
& MPOL_F_ADDR
) {
539 err
= lookup_node(mm
, addr
);
543 } else if (pol
== current
->mempolicy
&&
544 pol
->policy
== MPOL_INTERLEAVE
) {
545 *policy
= current
->il_next
;
551 *policy
= pol
->policy
| pol
->flags
;
554 up_read(¤t
->mm
->mmap_sem
);
560 get_zonemask(pol
, nmask
);
564 up_read(¤t
->mm
->mmap_sem
);
568 #ifdef CONFIG_MIGRATION
572 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
576 * Avoid migrating a page that is shared with others.
578 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1)
579 isolate_lru_page(page
, pagelist
);
582 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
584 return alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
588 * Migrate pages from one node to a target node.
589 * Returns error or the number of pages not migrated.
591 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
599 node_set(source
, nmask
);
601 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nmask
,
602 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
604 if (!list_empty(&pagelist
))
605 err
= migrate_pages(&pagelist
, new_node_page
, dest
);
611 * Move pages between the two nodesets so as to preserve the physical
612 * layout as much as possible.
614 * Returns the number of page that could not be moved.
616 int do_migrate_pages(struct mm_struct
*mm
,
617 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
624 down_read(&mm
->mmap_sem
);
626 err
= migrate_vmas(mm
, from_nodes
, to_nodes
, flags
);
631 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
632 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
633 * bit in 'tmp', and return that <source, dest> pair for migration.
634 * The pair of nodemasks 'to' and 'from' define the map.
636 * If no pair of bits is found that way, fallback to picking some
637 * pair of 'source' and 'dest' bits that are not the same. If the
638 * 'source' and 'dest' bits are the same, this represents a node
639 * that will be migrating to itself, so no pages need move.
641 * If no bits are left in 'tmp', or if all remaining bits left
642 * in 'tmp' correspond to the same bit in 'to', return false
643 * (nothing left to migrate).
645 * This lets us pick a pair of nodes to migrate between, such that
646 * if possible the dest node is not already occupied by some other
647 * source node, minimizing the risk of overloading the memory on a
648 * node that would happen if we migrated incoming memory to a node
649 * before migrating outgoing memory source that same node.
651 * A single scan of tmp is sufficient. As we go, we remember the
652 * most recent <s, d> pair that moved (s != d). If we find a pair
653 * that not only moved, but what's better, moved to an empty slot
654 * (d is not set in tmp), then we break out then, with that pair.
655 * Otherwise when we finish scannng from_tmp, we at least have the
656 * most recent <s, d> pair that moved. If we get all the way through
657 * the scan of tmp without finding any node that moved, much less
658 * moved to an empty node, then there is nothing left worth migrating.
662 while (!nodes_empty(tmp
)) {
667 for_each_node_mask(s
, tmp
) {
668 d
= node_remap(s
, *from_nodes
, *to_nodes
);
672 source
= s
; /* Node moved. Memorize */
675 /* dest not in remaining from nodes? */
676 if (!node_isset(dest
, tmp
))
682 node_clear(source
, tmp
);
683 err
= migrate_to_node(mm
, source
, dest
, flags
);
690 up_read(&mm
->mmap_sem
);
698 * Allocate a new page for page migration based on vma policy.
699 * Start assuming that page is mapped by vma pointed to by @private.
700 * Search forward from there, if not. N.B., this assumes that the
701 * list of pages handed to migrate_pages()--which is how we get here--
702 * is in virtual address order.
704 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
706 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
707 unsigned long uninitialized_var(address
);
710 address
= page_address_in_vma(page
, vma
);
711 if (address
!= -EFAULT
)
717 * if !vma, alloc_page_vma() will use task or system default policy
719 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
723 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
728 int do_migrate_pages(struct mm_struct
*mm
,
729 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
734 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
740 static long do_mbind(unsigned long start
, unsigned long len
,
741 unsigned short mode
, unsigned short mode_flags
,
742 nodemask_t
*nmask
, unsigned long flags
)
744 struct vm_area_struct
*vma
;
745 struct mm_struct
*mm
= current
->mm
;
746 struct mempolicy
*new;
751 if (flags
& ~(unsigned long)(MPOL_MF_STRICT
|
752 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
754 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
757 if (start
& ~PAGE_MASK
)
760 if (mode
== MPOL_DEFAULT
)
761 flags
&= ~MPOL_MF_STRICT
;
763 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
771 new = mpol_new(mode
, mode_flags
, nmask
);
776 * If we are using the default policy then operation
777 * on discontinuous address spaces is okay after all
780 flags
|= MPOL_MF_DISCONTIG_OK
;
782 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
783 start
, start
+ len
, mode
, mode_flags
,
784 nmask
? nodes_addr(*nmask
)[0] : -1);
786 down_write(&mm
->mmap_sem
);
787 vma
= check_range(mm
, start
, end
, nmask
,
788 flags
| MPOL_MF_INVERT
, &pagelist
);
794 err
= mbind_range(vma
, start
, end
, new);
796 if (!list_empty(&pagelist
))
797 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
800 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
804 up_write(&mm
->mmap_sem
);
810 * User space interface with variable sized bitmaps for nodelists.
813 /* Copy a node mask from user space. */
814 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
815 unsigned long maxnode
)
818 unsigned long nlongs
;
819 unsigned long endmask
;
823 if (maxnode
== 0 || !nmask
)
825 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
828 nlongs
= BITS_TO_LONGS(maxnode
);
829 if ((maxnode
% BITS_PER_LONG
) == 0)
832 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
834 /* When the user specified more nodes than supported just check
835 if the non supported part is all zero. */
836 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
837 if (nlongs
> PAGE_SIZE
/sizeof(long))
839 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
841 if (get_user(t
, nmask
+ k
))
843 if (k
== nlongs
- 1) {
849 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
853 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
855 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
859 /* Copy a kernel node mask to user space */
860 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
863 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
864 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
867 if (copy
> PAGE_SIZE
)
869 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
873 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
876 asmlinkage
long sys_mbind(unsigned long start
, unsigned long len
,
878 unsigned long __user
*nmask
, unsigned long maxnode
,
883 unsigned short mode_flags
;
885 mode_flags
= mode
& MPOL_MODE_FLAGS
;
886 mode
&= ~MPOL_MODE_FLAGS
;
887 if (mode
>= MPOL_MAX
)
889 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
890 (mode_flags
& MPOL_F_RELATIVE_NODES
))
892 err
= get_nodes(&nodes
, nmask
, maxnode
);
895 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
898 /* Set the process memory policy */
899 asmlinkage
long sys_set_mempolicy(int mode
, unsigned long __user
*nmask
,
900 unsigned long maxnode
)
904 unsigned short flags
;
906 flags
= mode
& MPOL_MODE_FLAGS
;
907 mode
&= ~MPOL_MODE_FLAGS
;
908 if ((unsigned int)mode
>= MPOL_MAX
)
910 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
912 err
= get_nodes(&nodes
, nmask
, maxnode
);
915 return do_set_mempolicy(mode
, flags
, &nodes
);
918 asmlinkage
long sys_migrate_pages(pid_t pid
, unsigned long maxnode
,
919 const unsigned long __user
*old_nodes
,
920 const unsigned long __user
*new_nodes
)
922 struct mm_struct
*mm
;
923 struct task_struct
*task
;
926 nodemask_t task_nodes
;
929 err
= get_nodes(&old
, old_nodes
, maxnode
);
933 err
= get_nodes(&new, new_nodes
, maxnode
);
937 /* Find the mm_struct */
938 read_lock(&tasklist_lock
);
939 task
= pid
? find_task_by_vpid(pid
) : current
;
941 read_unlock(&tasklist_lock
);
944 mm
= get_task_mm(task
);
945 read_unlock(&tasklist_lock
);
951 * Check if this process has the right to modify the specified
952 * process. The right exists if the process has administrative
953 * capabilities, superuser privileges or the same
954 * userid as the target process.
956 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
957 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
958 !capable(CAP_SYS_NICE
)) {
963 task_nodes
= cpuset_mems_allowed(task
);
964 /* Is the user allowed to access the target nodes? */
965 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
970 if (!nodes_subset(new, node_states
[N_HIGH_MEMORY
])) {
975 err
= security_task_movememory(task
);
979 err
= do_migrate_pages(mm
, &old
, &new,
980 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
987 /* Retrieve NUMA policy */
988 asmlinkage
long sys_get_mempolicy(int __user
*policy
,
989 unsigned long __user
*nmask
,
990 unsigned long maxnode
,
991 unsigned long addr
, unsigned long flags
)
994 int uninitialized_var(pval
);
997 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1000 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1005 if (policy
&& put_user(pval
, policy
))
1009 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1014 #ifdef CONFIG_COMPAT
1016 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1017 compat_ulong_t __user
*nmask
,
1018 compat_ulong_t maxnode
,
1019 compat_ulong_t addr
, compat_ulong_t flags
)
1022 unsigned long __user
*nm
= NULL
;
1023 unsigned long nr_bits
, alloc_size
;
1024 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1026 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1027 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1030 nm
= compat_alloc_user_space(alloc_size
);
1032 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1034 if (!err
&& nmask
) {
1035 err
= copy_from_user(bm
, nm
, alloc_size
);
1036 /* ensure entire bitmap is zeroed */
1037 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1038 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1044 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1045 compat_ulong_t maxnode
)
1048 unsigned long __user
*nm
= NULL
;
1049 unsigned long nr_bits
, alloc_size
;
1050 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1052 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1053 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1056 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1057 nm
= compat_alloc_user_space(alloc_size
);
1058 err
|= copy_to_user(nm
, bm
, alloc_size
);
1064 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1067 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1068 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1069 compat_ulong_t maxnode
, compat_ulong_t flags
)
1072 unsigned long __user
*nm
= NULL
;
1073 unsigned long nr_bits
, alloc_size
;
1076 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1077 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1080 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1081 nm
= compat_alloc_user_space(alloc_size
);
1082 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1088 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1094 * get_vma_policy(@task, @vma, @addr)
1095 * @task - task for fallback if vma policy == default
1096 * @vma - virtual memory area whose policy is sought
1097 * @addr - address in @vma for shared policy lookup
1099 * Returns effective policy for a VMA at specified address.
1100 * Falls back to @task or system default policy, as necessary.
1101 * Returned policy has extra reference count if shared, vma,
1102 * or some other task's policy [show_numa_maps() can pass
1103 * @task != current]. It is the caller's responsibility to
1104 * free the reference in these cases.
1106 static struct mempolicy
* get_vma_policy(struct task_struct
*task
,
1107 struct vm_area_struct
*vma
, unsigned long addr
)
1109 struct mempolicy
*pol
= task
->mempolicy
;
1113 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1114 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1115 shared_pol
= 1; /* if pol non-NULL, add ref below */
1116 } else if (vma
->vm_policy
&&
1117 vma
->vm_policy
->policy
!= MPOL_DEFAULT
)
1118 pol
= vma
->vm_policy
;
1121 pol
= &default_policy
;
1122 else if (!shared_pol
&& pol
!= current
->mempolicy
)
1123 mpol_get(pol
); /* vma or other task's policy */
1127 /* Return a nodemask representing a mempolicy */
1128 static nodemask_t
*nodemask_policy(gfp_t gfp
, struct mempolicy
*policy
)
1130 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1131 if (unlikely(policy
->policy
== MPOL_BIND
) &&
1132 gfp_zone(gfp
) >= policy_zone
&&
1133 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1134 return &policy
->v
.nodes
;
1139 /* Return a zonelist representing a mempolicy */
1140 static struct zonelist
*zonelist_policy(gfp_t gfp
, struct mempolicy
*policy
)
1144 switch (policy
->policy
) {
1145 case MPOL_PREFERRED
:
1146 nd
= policy
->v
.preferred_node
;
1148 nd
= numa_node_id();
1152 * Normally, MPOL_BIND allocations node-local are node-local
1153 * within the allowed nodemask. However, if __GFP_THISNODE is
1154 * set and the current node is part of the mask, we use the
1155 * the zonelist for the first node in the mask instead.
1157 nd
= numa_node_id();
1158 if (unlikely(gfp
& __GFP_THISNODE
) &&
1159 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1160 nd
= first_node(policy
->v
.nodes
);
1162 case MPOL_INTERLEAVE
: /* should not happen */
1164 nd
= numa_node_id();
1170 return node_zonelist(nd
, gfp
);
1173 /* Do dynamic interleaving for a process */
1174 static unsigned interleave_nodes(struct mempolicy
*policy
)
1177 struct task_struct
*me
= current
;
1180 next
= next_node(nid
, policy
->v
.nodes
);
1181 if (next
>= MAX_NUMNODES
)
1182 next
= first_node(policy
->v
.nodes
);
1183 if (next
< MAX_NUMNODES
)
1189 * Depending on the memory policy provide a node from which to allocate the
1192 unsigned slab_node(struct mempolicy
*policy
)
1194 unsigned short pol
= policy
? policy
->policy
: MPOL_DEFAULT
;
1197 case MPOL_INTERLEAVE
:
1198 return interleave_nodes(policy
);
1202 * Follow bind policy behavior and start allocation at the
1205 struct zonelist
*zonelist
;
1207 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1208 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1209 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1215 case MPOL_PREFERRED
:
1216 if (policy
->v
.preferred_node
>= 0)
1217 return policy
->v
.preferred_node
;
1221 return numa_node_id();
1225 /* Do static interleaving for a VMA with known offset. */
1226 static unsigned offset_il_node(struct mempolicy
*pol
,
1227 struct vm_area_struct
*vma
, unsigned long off
)
1229 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1235 return numa_node_id();
1236 target
= (unsigned int)off
% nnodes
;
1239 nid
= next_node(nid
, pol
->v
.nodes
);
1241 } while (c
<= target
);
1245 /* Determine a node number for interleave */
1246 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1247 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1253 * for small pages, there is no difference between
1254 * shift and PAGE_SHIFT, so the bit-shift is safe.
1255 * for huge pages, since vm_pgoff is in units of small
1256 * pages, we need to shift off the always 0 bits to get
1259 BUG_ON(shift
< PAGE_SHIFT
);
1260 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1261 off
+= (addr
- vma
->vm_start
) >> shift
;
1262 return offset_il_node(pol
, vma
, off
);
1264 return interleave_nodes(pol
);
1267 #ifdef CONFIG_HUGETLBFS
1269 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1270 * @vma = virtual memory area whose policy is sought
1271 * @addr = address in @vma for shared policy lookup and interleave policy
1272 * @gfp_flags = for requested zone
1273 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1274 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1276 * Returns a zonelist suitable for a huge page allocation.
1277 * If the effective policy is 'BIND, returns pointer to local node's zonelist,
1278 * and a pointer to the mempolicy's @nodemask for filtering the zonelist.
1279 * If it is also a policy for which get_vma_policy() returns an extra
1280 * reference, we must hold that reference until after the allocation.
1281 * In that case, return policy via @mpol so hugetlb allocation can drop
1282 * the reference. For non-'BIND referenced policies, we can/do drop the
1283 * reference here, so the caller doesn't need to know about the special case
1284 * for default and current task policy.
1286 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1287 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1288 nodemask_t
**nodemask
)
1290 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1291 struct zonelist
*zl
;
1293 *mpol
= NULL
; /* probably no unref needed */
1294 *nodemask
= NULL
; /* assume !MPOL_BIND */
1295 if (pol
->policy
== MPOL_BIND
) {
1296 *nodemask
= &pol
->v
.nodes
;
1297 } else if (pol
->policy
== MPOL_INTERLEAVE
) {
1300 nid
= interleave_nid(pol
, vma
, addr
, HPAGE_SHIFT
);
1301 if (unlikely(pol
!= &default_policy
&&
1302 pol
!= current
->mempolicy
))
1303 __mpol_free(pol
); /* finished with pol */
1304 return node_zonelist(nid
, gfp_flags
);
1307 zl
= zonelist_policy(GFP_HIGHUSER
, pol
);
1308 if (unlikely(pol
!= &default_policy
&& pol
!= current
->mempolicy
)) {
1309 if (pol
->policy
!= MPOL_BIND
)
1310 __mpol_free(pol
); /* finished with pol */
1312 *mpol
= pol
; /* unref needed after allocation */
1318 /* Allocate a page in interleaved policy.
1319 Own path because it needs to do special accounting. */
1320 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1323 struct zonelist
*zl
;
1326 zl
= node_zonelist(nid
, gfp
);
1327 page
= __alloc_pages(gfp
, order
, zl
);
1328 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1329 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1334 * alloc_page_vma - Allocate a page for a VMA.
1337 * %GFP_USER user allocation.
1338 * %GFP_KERNEL kernel allocations,
1339 * %GFP_HIGHMEM highmem/user allocations,
1340 * %GFP_FS allocation should not call back into a file system.
1341 * %GFP_ATOMIC don't sleep.
1343 * @vma: Pointer to VMA or NULL if not available.
1344 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1346 * This function allocates a page from the kernel page pool and applies
1347 * a NUMA policy associated with the VMA or the current process.
1348 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1349 * mm_struct of the VMA to prevent it from going away. Should be used for
1350 * all allocations for pages that will be mapped into
1351 * user space. Returns NULL when no page can be allocated.
1353 * Should be called with the mm_sem of the vma hold.
1356 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1358 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1359 struct zonelist
*zl
;
1361 cpuset_update_task_memory_state();
1363 if (unlikely(pol
->policy
== MPOL_INTERLEAVE
)) {
1366 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1367 if (unlikely(pol
!= &default_policy
&&
1368 pol
!= current
->mempolicy
))
1369 __mpol_free(pol
); /* finished with pol */
1370 return alloc_page_interleave(gfp
, 0, nid
);
1372 zl
= zonelist_policy(gfp
, pol
);
1373 if (pol
!= &default_policy
&& pol
!= current
->mempolicy
) {
1375 * slow path: ref counted policy -- shared or vma
1377 struct page
*page
= __alloc_pages_nodemask(gfp
, 0,
1378 zl
, nodemask_policy(gfp
, pol
));
1383 * fast path: default or task policy
1385 return __alloc_pages_nodemask(gfp
, 0, zl
, nodemask_policy(gfp
, pol
));
1389 * alloc_pages_current - Allocate pages.
1392 * %GFP_USER user allocation,
1393 * %GFP_KERNEL kernel allocation,
1394 * %GFP_HIGHMEM highmem allocation,
1395 * %GFP_FS don't call back into a file system.
1396 * %GFP_ATOMIC don't sleep.
1397 * @order: Power of two of allocation size in pages. 0 is a single page.
1399 * Allocate a page from the kernel page pool. When not in
1400 * interrupt context and apply the current process NUMA policy.
1401 * Returns NULL when no page can be allocated.
1403 * Don't call cpuset_update_task_memory_state() unless
1404 * 1) it's ok to take cpuset_sem (can WAIT), and
1405 * 2) allocating for current task (not interrupt).
1407 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1409 struct mempolicy
*pol
= current
->mempolicy
;
1411 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1412 cpuset_update_task_memory_state();
1413 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
1414 pol
= &default_policy
;
1415 if (pol
->policy
== MPOL_INTERLEAVE
)
1416 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1417 return __alloc_pages_nodemask(gfp
, order
,
1418 zonelist_policy(gfp
, pol
), nodemask_policy(gfp
, pol
));
1420 EXPORT_SYMBOL(alloc_pages_current
);
1423 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1424 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1425 * with the mems_allowed returned by cpuset_mems_allowed(). This
1426 * keeps mempolicies cpuset relative after its cpuset moves. See
1427 * further kernel/cpuset.c update_nodemask().
1430 /* Slow path of a mempolicy copy */
1431 struct mempolicy
*__mpol_copy(struct mempolicy
*old
)
1433 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1436 return ERR_PTR(-ENOMEM
);
1437 if (current_cpuset_is_being_rebound()) {
1438 nodemask_t mems
= cpuset_mems_allowed(current
);
1439 mpol_rebind_policy(old
, &mems
);
1442 atomic_set(&new->refcnt
, 1);
1446 static int mpol_match_intent(const struct mempolicy
*a
,
1447 const struct mempolicy
*b
)
1449 if (a
->flags
!= b
->flags
)
1451 if (!mpol_store_user_nodemask(a
))
1453 return nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
);
1456 /* Slow path of a mempolicy comparison */
1457 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1461 if (a
->policy
!= b
->policy
)
1463 if (a
->policy
!= MPOL_DEFAULT
&& !mpol_match_intent(a
, b
))
1465 switch (a
->policy
) {
1470 case MPOL_INTERLEAVE
:
1471 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1472 case MPOL_PREFERRED
:
1473 return a
->v
.preferred_node
== b
->v
.preferred_node
;
1480 /* Slow path of a mpol destructor. */
1481 void __mpol_free(struct mempolicy
*p
)
1483 if (!atomic_dec_and_test(&p
->refcnt
))
1485 p
->policy
= MPOL_DEFAULT
;
1486 kmem_cache_free(policy_cache
, p
);
1490 * Shared memory backing store policy support.
1492 * Remember policies even when nobody has shared memory mapped.
1493 * The policies are kept in Red-Black tree linked from the inode.
1494 * They are protected by the sp->lock spinlock, which should be held
1495 * for any accesses to the tree.
1498 /* lookup first element intersecting start-end */
1499 /* Caller holds sp->lock */
1500 static struct sp_node
*
1501 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1503 struct rb_node
*n
= sp
->root
.rb_node
;
1506 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1508 if (start
>= p
->end
)
1510 else if (end
<= p
->start
)
1518 struct sp_node
*w
= NULL
;
1519 struct rb_node
*prev
= rb_prev(n
);
1522 w
= rb_entry(prev
, struct sp_node
, nd
);
1523 if (w
->end
<= start
)
1527 return rb_entry(n
, struct sp_node
, nd
);
1530 /* Insert a new shared policy into the list. */
1531 /* Caller holds sp->lock */
1532 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1534 struct rb_node
**p
= &sp
->root
.rb_node
;
1535 struct rb_node
*parent
= NULL
;
1540 nd
= rb_entry(parent
, struct sp_node
, nd
);
1541 if (new->start
< nd
->start
)
1543 else if (new->end
> nd
->end
)
1544 p
= &(*p
)->rb_right
;
1548 rb_link_node(&new->nd
, parent
, p
);
1549 rb_insert_color(&new->nd
, &sp
->root
);
1550 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
1551 new->policy
? new->policy
->policy
: 0);
1554 /* Find shared policy intersecting idx */
1556 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1558 struct mempolicy
*pol
= NULL
;
1561 if (!sp
->root
.rb_node
)
1563 spin_lock(&sp
->lock
);
1564 sn
= sp_lookup(sp
, idx
, idx
+1);
1566 mpol_get(sn
->policy
);
1569 spin_unlock(&sp
->lock
);
1573 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1575 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
1576 rb_erase(&n
->nd
, &sp
->root
);
1577 mpol_free(n
->policy
);
1578 kmem_cache_free(sn_cache
, n
);
1581 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
1582 struct mempolicy
*pol
)
1584 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1595 /* Replace a policy range. */
1596 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1597 unsigned long end
, struct sp_node
*new)
1599 struct sp_node
*n
, *new2
= NULL
;
1602 spin_lock(&sp
->lock
);
1603 n
= sp_lookup(sp
, start
, end
);
1604 /* Take care of old policies in the same range. */
1605 while (n
&& n
->start
< end
) {
1606 struct rb_node
*next
= rb_next(&n
->nd
);
1607 if (n
->start
>= start
) {
1613 /* Old policy spanning whole new range. */
1616 spin_unlock(&sp
->lock
);
1617 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1623 sp_insert(sp
, new2
);
1631 n
= rb_entry(next
, struct sp_node
, nd
);
1635 spin_unlock(&sp
->lock
);
1637 mpol_free(new2
->policy
);
1638 kmem_cache_free(sn_cache
, new2
);
1643 void mpol_shared_policy_init(struct shared_policy
*info
, unsigned short policy
,
1644 unsigned short flags
, nodemask_t
*policy_nodes
)
1646 info
->root
= RB_ROOT
;
1647 spin_lock_init(&info
->lock
);
1649 if (policy
!= MPOL_DEFAULT
) {
1650 struct mempolicy
*newpol
;
1652 /* Falls back to MPOL_DEFAULT on any error */
1653 newpol
= mpol_new(policy
, flags
, policy_nodes
);
1654 if (!IS_ERR(newpol
)) {
1655 /* Create pseudo-vma that contains just the policy */
1656 struct vm_area_struct pvma
;
1658 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
1659 /* Policy covers entire file */
1660 pvma
.vm_end
= TASK_SIZE
;
1661 mpol_set_shared_policy(info
, &pvma
, newpol
);
1667 int mpol_set_shared_policy(struct shared_policy
*info
,
1668 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1671 struct sp_node
*new = NULL
;
1672 unsigned long sz
= vma_pages(vma
);
1674 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1676 sz
, npol
? npol
->policy
: -1,
1677 npol
? npol
->flags
: -1,
1678 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1681 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1685 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1687 kmem_cache_free(sn_cache
, new);
1691 /* Free a backing policy store on inode delete. */
1692 void mpol_free_shared_policy(struct shared_policy
*p
)
1695 struct rb_node
*next
;
1697 if (!p
->root
.rb_node
)
1699 spin_lock(&p
->lock
);
1700 next
= rb_first(&p
->root
);
1702 n
= rb_entry(next
, struct sp_node
, nd
);
1703 next
= rb_next(&n
->nd
);
1704 rb_erase(&n
->nd
, &p
->root
);
1705 mpol_free(n
->policy
);
1706 kmem_cache_free(sn_cache
, n
);
1708 spin_unlock(&p
->lock
);
1711 /* assumes fs == KERNEL_DS */
1712 void __init
numa_policy_init(void)
1714 nodemask_t interleave_nodes
;
1715 unsigned long largest
= 0;
1716 int nid
, prefer
= 0;
1718 policy_cache
= kmem_cache_create("numa_policy",
1719 sizeof(struct mempolicy
),
1720 0, SLAB_PANIC
, NULL
);
1722 sn_cache
= kmem_cache_create("shared_policy_node",
1723 sizeof(struct sp_node
),
1724 0, SLAB_PANIC
, NULL
);
1727 * Set interleaving policy for system init. Interleaving is only
1728 * enabled across suitably sized nodes (default is >= 16MB), or
1729 * fall back to the largest node if they're all smaller.
1731 nodes_clear(interleave_nodes
);
1732 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1733 unsigned long total_pages
= node_present_pages(nid
);
1735 /* Preserve the largest node */
1736 if (largest
< total_pages
) {
1737 largest
= total_pages
;
1741 /* Interleave this node? */
1742 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
1743 node_set(nid
, interleave_nodes
);
1746 /* All too small, use the largest */
1747 if (unlikely(nodes_empty(interleave_nodes
)))
1748 node_set(prefer
, interleave_nodes
);
1750 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
1751 printk("numa_policy_init: interleaving failed\n");
1754 /* Reset policy of current process to default */
1755 void numa_default_policy(void)
1757 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
1760 /* Migrate a policy to a different set of nodes */
1761 static void mpol_rebind_policy(struct mempolicy
*pol
,
1762 const nodemask_t
*newmask
)
1770 static_nodes
= pol
->flags
& MPOL_F_STATIC_NODES
;
1771 relative_nodes
= pol
->flags
& MPOL_F_RELATIVE_NODES
;
1772 if (!mpol_store_user_nodemask(pol
) &&
1773 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
1776 switch (pol
->policy
) {
1781 case MPOL_INTERLEAVE
:
1783 nodes_and(tmp
, pol
->w
.user_nodemask
, *newmask
);
1784 else if (relative_nodes
)
1785 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
,
1788 nodes_remap(tmp
, pol
->v
.nodes
,
1789 pol
->w
.cpuset_mems_allowed
, *newmask
);
1790 pol
->w
.cpuset_mems_allowed
= *newmask
;
1793 if (!node_isset(current
->il_next
, tmp
)) {
1794 current
->il_next
= next_node(current
->il_next
, tmp
);
1795 if (current
->il_next
>= MAX_NUMNODES
)
1796 current
->il_next
= first_node(tmp
);
1797 if (current
->il_next
>= MAX_NUMNODES
)
1798 current
->il_next
= numa_node_id();
1801 case MPOL_PREFERRED
:
1803 int node
= first_node(pol
->w
.user_nodemask
);
1805 if (node_isset(node
, *newmask
))
1806 pol
->v
.preferred_node
= node
;
1808 pol
->v
.preferred_node
= -1;
1809 } else if (relative_nodes
) {
1810 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
,
1812 pol
->v
.preferred_node
= first_node(tmp
);
1814 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
1815 pol
->w
.cpuset_mems_allowed
, *newmask
);
1816 pol
->w
.cpuset_mems_allowed
= *newmask
;
1826 * Wrapper for mpol_rebind_policy() that just requires task
1827 * pointer, and updates task mempolicy.
1830 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
1832 mpol_rebind_policy(tsk
->mempolicy
, new);
1836 * Rebind each vma in mm to new nodemask.
1838 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1841 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
1843 struct vm_area_struct
*vma
;
1845 down_write(&mm
->mmap_sem
);
1846 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
1847 mpol_rebind_policy(vma
->vm_policy
, new);
1848 up_write(&mm
->mmap_sem
);
1852 * Display pages allocated per node and memory policy via /proc.
1855 static const char * const policy_types
[] =
1856 { "default", "prefer", "bind", "interleave" };
1859 * Convert a mempolicy into a string.
1860 * Returns the number of characters in buffer (if positive)
1861 * or an error (negative)
1863 static inline int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
1868 unsigned short mode
= pol
? pol
->policy
: MPOL_DEFAULT
;
1869 unsigned short flags
= pol
? pol
->flags
: 0;
1876 case MPOL_PREFERRED
:
1878 node_set(pol
->v
.preferred_node
, nodes
);
1883 case MPOL_INTERLEAVE
:
1884 nodes
= pol
->v
.nodes
;
1892 l
= strlen(policy_types
[mode
]);
1893 if (buffer
+ maxlen
< p
+ l
+ 1)
1896 strcpy(p
, policy_types
[mode
]);
1902 if (buffer
+ maxlen
< p
+ 2)
1906 if (flags
& MPOL_F_STATIC_NODES
)
1907 p
+= sprintf(p
, "%sstatic", need_bar
++ ? "|" : "");
1908 if (flags
& MPOL_F_RELATIVE_NODES
)
1909 p
+= sprintf(p
, "%srelative", need_bar
++ ? "|" : "");
1912 if (!nodes_empty(nodes
)) {
1913 if (buffer
+ maxlen
< p
+ 2)
1916 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
1922 unsigned long pages
;
1924 unsigned long active
;
1925 unsigned long writeback
;
1926 unsigned long mapcount_max
;
1927 unsigned long dirty
;
1928 unsigned long swapcache
;
1929 unsigned long node
[MAX_NUMNODES
];
1932 static void gather_stats(struct page
*page
, void *private, int pte_dirty
)
1934 struct numa_maps
*md
= private;
1935 int count
= page_mapcount(page
);
1938 if (pte_dirty
|| PageDirty(page
))
1941 if (PageSwapCache(page
))
1944 if (PageActive(page
))
1947 if (PageWriteback(page
))
1953 if (count
> md
->mapcount_max
)
1954 md
->mapcount_max
= count
;
1956 md
->node
[page_to_nid(page
)]++;
1959 #ifdef CONFIG_HUGETLB_PAGE
1960 static void check_huge_range(struct vm_area_struct
*vma
,
1961 unsigned long start
, unsigned long end
,
1962 struct numa_maps
*md
)
1967 for (addr
= start
; addr
< end
; addr
+= HPAGE_SIZE
) {
1968 pte_t
*ptep
= huge_pte_offset(vma
->vm_mm
, addr
& HPAGE_MASK
);
1978 page
= pte_page(pte
);
1982 gather_stats(page
, md
, pte_dirty(*ptep
));
1986 static inline void check_huge_range(struct vm_area_struct
*vma
,
1987 unsigned long start
, unsigned long end
,
1988 struct numa_maps
*md
)
1993 int show_numa_map(struct seq_file
*m
, void *v
)
1995 struct proc_maps_private
*priv
= m
->private;
1996 struct vm_area_struct
*vma
= v
;
1997 struct numa_maps
*md
;
1998 struct file
*file
= vma
->vm_file
;
1999 struct mm_struct
*mm
= vma
->vm_mm
;
2000 struct mempolicy
*pol
;
2007 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
2011 pol
= get_vma_policy(priv
->task
, vma
, vma
->vm_start
);
2012 mpol_to_str(buffer
, sizeof(buffer
), pol
);
2014 * unref shared or other task's mempolicy
2016 if (pol
!= &default_policy
&& pol
!= current
->mempolicy
)
2019 seq_printf(m
, "%08lx %s", vma
->vm_start
, buffer
);
2022 seq_printf(m
, " file=");
2023 seq_path(m
, &file
->f_path
, "\n\t= ");
2024 } else if (vma
->vm_start
<= mm
->brk
&& vma
->vm_end
>= mm
->start_brk
) {
2025 seq_printf(m
, " heap");
2026 } else if (vma
->vm_start
<= mm
->start_stack
&&
2027 vma
->vm_end
>= mm
->start_stack
) {
2028 seq_printf(m
, " stack");
2031 if (is_vm_hugetlb_page(vma
)) {
2032 check_huge_range(vma
, vma
->vm_start
, vma
->vm_end
, md
);
2033 seq_printf(m
, " huge");
2035 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
2036 &node_states
[N_HIGH_MEMORY
], MPOL_MF_STATS
, md
);
2043 seq_printf(m
," anon=%lu",md
->anon
);
2046 seq_printf(m
," dirty=%lu",md
->dirty
);
2048 if (md
->pages
!= md
->anon
&& md
->pages
!= md
->dirty
)
2049 seq_printf(m
, " mapped=%lu", md
->pages
);
2051 if (md
->mapcount_max
> 1)
2052 seq_printf(m
, " mapmax=%lu", md
->mapcount_max
);
2055 seq_printf(m
," swapcache=%lu", md
->swapcache
);
2057 if (md
->active
< md
->pages
&& !is_vm_hugetlb_page(vma
))
2058 seq_printf(m
," active=%lu", md
->active
);
2061 seq_printf(m
," writeback=%lu", md
->writeback
);
2063 for_each_node_state(n
, N_HIGH_MEMORY
)
2065 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
2070 if (m
->count
< m
->size
)
2071 m
->version
= (vma
!= priv
->tail_vma
) ? vma
->vm_start
: 0;