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 /* Do sanity checking on a policy */
117 static int mpol_check_policy(int mode
, nodemask_t
*nodes
)
119 int was_empty
, is_empty
;
125 * "Contextualize" the in-coming nodemast for cpusets:
126 * Remember whether in-coming nodemask was empty, If not,
127 * restrict the nodes to the allowed nodes in the cpuset.
128 * This is guaranteed to be a subset of nodes with memory.
130 cpuset_update_task_memory_state();
131 is_empty
= was_empty
= nodes_empty(*nodes
);
133 nodes_and(*nodes
, *nodes
, cpuset_current_mems_allowed
);
134 is_empty
= nodes_empty(*nodes
); /* after "contextualization" */
140 * require caller to specify an empty nodemask
141 * before "contextualization"
147 case MPOL_INTERLEAVE
:
149 * require at least 1 valid node after "contextualization"
156 * Did caller specify invalid nodes?
157 * Don't silently accept this as "local allocation".
159 if (!was_empty
&& is_empty
)
166 /* Generate a custom zonelist for the BIND policy. */
167 static struct zonelist
*bind_zonelist(nodemask_t
*nodes
)
173 max
= 1 + MAX_NR_ZONES
* nodes_weight(*nodes
);
174 max
++; /* space for zlcache_ptr (see mmzone.h) */
175 zl
= kmalloc(sizeof(struct zone
*) * max
, GFP_KERNEL
);
177 return ERR_PTR(-ENOMEM
);
178 zl
->zlcache_ptr
= NULL
;
180 /* First put in the highest zones from all nodes, then all the next
181 lower zones etc. Avoid empty zones because the memory allocator
182 doesn't like them. If you implement node hot removal you
184 k
= MAX_NR_ZONES
- 1;
186 for_each_node_mask(nd
, *nodes
) {
187 struct zone
*z
= &NODE_DATA(nd
)->node_zones
[k
];
188 if (z
->present_pages
> 0)
189 zoneref_set_zone(z
, &zl
->_zonerefs
[num
++]);
197 return ERR_PTR(-EINVAL
);
199 zl
->_zonerefs
[num
].zone
= NULL
;
200 zl
->_zonerefs
[num
].zone_idx
= 0;
204 /* Create a new policy */
205 static struct mempolicy
*mpol_new(int mode
, nodemask_t
*nodes
)
207 struct mempolicy
*policy
;
209 pr_debug("setting mode %d nodes[0] %lx\n",
210 mode
, nodes
? nodes_addr(*nodes
)[0] : -1);
212 if (mode
== MPOL_DEFAULT
)
214 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
216 return ERR_PTR(-ENOMEM
);
217 atomic_set(&policy
->refcnt
, 1);
219 case MPOL_INTERLEAVE
:
220 policy
->v
.nodes
= *nodes
;
221 if (nodes_weight(policy
->v
.nodes
) == 0) {
222 kmem_cache_free(policy_cache
, policy
);
223 return ERR_PTR(-EINVAL
);
227 policy
->v
.preferred_node
= first_node(*nodes
);
228 if (policy
->v
.preferred_node
>= MAX_NUMNODES
)
229 policy
->v
.preferred_node
= -1;
232 policy
->v
.zonelist
= bind_zonelist(nodes
);
233 if (IS_ERR(policy
->v
.zonelist
)) {
234 void *error_code
= policy
->v
.zonelist
;
235 kmem_cache_free(policy_cache
, policy
);
240 policy
->policy
= mode
;
241 policy
->cpuset_mems_allowed
= cpuset_mems_allowed(current
);
245 static void gather_stats(struct page
*, void *, int pte_dirty
);
246 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
247 unsigned long flags
);
249 /* Scan through pages checking if pages follow certain conditions. */
250 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
251 unsigned long addr
, unsigned long end
,
252 const nodemask_t
*nodes
, unsigned long flags
,
259 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
264 if (!pte_present(*pte
))
266 page
= vm_normal_page(vma
, addr
, *pte
);
270 * The check for PageReserved here is important to avoid
271 * handling zero pages and other pages that may have been
272 * marked special by the system.
274 * If the PageReserved would not be checked here then f.e.
275 * the location of the zero page could have an influence
276 * on MPOL_MF_STRICT, zero pages would be counted for
277 * the per node stats, and there would be useless attempts
278 * to put zero pages on the migration list.
280 if (PageReserved(page
))
282 nid
= page_to_nid(page
);
283 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
286 if (flags
& MPOL_MF_STATS
)
287 gather_stats(page
, private, pte_dirty(*pte
));
288 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
289 migrate_page_add(page
, private, flags
);
292 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
293 pte_unmap_unlock(orig_pte
, ptl
);
297 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
298 unsigned long addr
, unsigned long end
,
299 const nodemask_t
*nodes
, unsigned long flags
,
305 pmd
= pmd_offset(pud
, addr
);
307 next
= pmd_addr_end(addr
, end
);
308 if (pmd_none_or_clear_bad(pmd
))
310 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
313 } while (pmd
++, addr
= next
, addr
!= end
);
317 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
318 unsigned long addr
, unsigned long end
,
319 const nodemask_t
*nodes
, unsigned long flags
,
325 pud
= pud_offset(pgd
, addr
);
327 next
= pud_addr_end(addr
, end
);
328 if (pud_none_or_clear_bad(pud
))
330 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
333 } while (pud
++, addr
= next
, addr
!= end
);
337 static inline int check_pgd_range(struct vm_area_struct
*vma
,
338 unsigned long addr
, unsigned long end
,
339 const nodemask_t
*nodes
, unsigned long flags
,
345 pgd
= pgd_offset(vma
->vm_mm
, addr
);
347 next
= pgd_addr_end(addr
, end
);
348 if (pgd_none_or_clear_bad(pgd
))
350 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
353 } while (pgd
++, addr
= next
, addr
!= end
);
358 * Check if all pages in a range are on a set of nodes.
359 * If pagelist != NULL then isolate pages from the LRU and
360 * put them on the pagelist.
362 static struct vm_area_struct
*
363 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
364 const nodemask_t
*nodes
, unsigned long flags
, void *private)
367 struct vm_area_struct
*first
, *vma
, *prev
;
369 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
371 err
= migrate_prep();
376 first
= find_vma(mm
, start
);
378 return ERR_PTR(-EFAULT
);
380 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
381 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
382 if (!vma
->vm_next
&& vma
->vm_end
< end
)
383 return ERR_PTR(-EFAULT
);
384 if (prev
&& prev
->vm_end
< vma
->vm_start
)
385 return ERR_PTR(-EFAULT
);
387 if (!is_vm_hugetlb_page(vma
) &&
388 ((flags
& MPOL_MF_STRICT
) ||
389 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
390 vma_migratable(vma
)))) {
391 unsigned long endvma
= vma
->vm_end
;
395 if (vma
->vm_start
> start
)
396 start
= vma
->vm_start
;
397 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
400 first
= ERR_PTR(err
);
409 /* Apply policy to a single VMA */
410 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
413 struct mempolicy
*old
= vma
->vm_policy
;
415 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
416 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
417 vma
->vm_ops
, vma
->vm_file
,
418 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
420 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
421 err
= vma
->vm_ops
->set_policy(vma
, new);
424 vma
->vm_policy
= new;
430 /* Step 2: apply policy to a range and do splits. */
431 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
432 unsigned long end
, struct mempolicy
*new)
434 struct vm_area_struct
*next
;
438 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
440 if (vma
->vm_start
< start
)
441 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
442 if (!err
&& vma
->vm_end
> end
)
443 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
445 err
= policy_vma(vma
, new);
453 * Update task->flags PF_MEMPOLICY bit: set iff non-default
454 * mempolicy. Allows more rapid checking of this (combined perhaps
455 * with other PF_* flag bits) on memory allocation hot code paths.
457 * If called from outside this file, the task 'p' should -only- be
458 * a newly forked child not yet visible on the task list, because
459 * manipulating the task flags of a visible task is not safe.
461 * The above limitation is why this routine has the funny name
462 * mpol_fix_fork_child_flag().
464 * It is also safe to call this with a task pointer of current,
465 * which the static wrapper mpol_set_task_struct_flag() does,
466 * for use within this file.
469 void mpol_fix_fork_child_flag(struct task_struct
*p
)
472 p
->flags
|= PF_MEMPOLICY
;
474 p
->flags
&= ~PF_MEMPOLICY
;
477 static void mpol_set_task_struct_flag(void)
479 mpol_fix_fork_child_flag(current
);
482 /* Set the process memory policy */
483 static long do_set_mempolicy(int mode
, nodemask_t
*nodes
)
485 struct mempolicy
*new;
487 if (mpol_check_policy(mode
, nodes
))
489 new = mpol_new(mode
, nodes
);
492 mpol_free(current
->mempolicy
);
493 current
->mempolicy
= new;
494 mpol_set_task_struct_flag();
495 if (new && new->policy
== MPOL_INTERLEAVE
)
496 current
->il_next
= first_node(new->v
.nodes
);
500 /* Fill a zone bitmap for a policy */
501 static void get_zonemask(struct mempolicy
*p
, nodemask_t
*nodes
)
508 for (i
= 0; p
->v
.zonelist
->_zonerefs
[i
].zone
; i
++) {
509 struct zoneref
*zref
;
510 zref
= &p
->v
.zonelist
->_zonerefs
[i
];
511 node_set(zonelist_node_idx(zref
), *nodes
);
516 case MPOL_INTERLEAVE
:
520 /* or use current node instead of memory_map? */
521 if (p
->v
.preferred_node
< 0)
522 *nodes
= node_states
[N_HIGH_MEMORY
];
524 node_set(p
->v
.preferred_node
, *nodes
);
531 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
536 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
538 err
= page_to_nid(p
);
544 /* Retrieve NUMA policy */
545 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
546 unsigned long addr
, unsigned long flags
)
549 struct mm_struct
*mm
= current
->mm
;
550 struct vm_area_struct
*vma
= NULL
;
551 struct mempolicy
*pol
= current
->mempolicy
;
553 cpuset_update_task_memory_state();
555 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
558 if (flags
& MPOL_F_MEMS_ALLOWED
) {
559 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
561 *policy
= 0; /* just so it's initialized */
562 *nmask
= cpuset_current_mems_allowed
;
566 if (flags
& MPOL_F_ADDR
) {
567 down_read(&mm
->mmap_sem
);
568 vma
= find_vma_intersection(mm
, addr
, addr
+1);
570 up_read(&mm
->mmap_sem
);
573 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
574 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
576 pol
= vma
->vm_policy
;
581 pol
= &default_policy
;
583 if (flags
& MPOL_F_NODE
) {
584 if (flags
& MPOL_F_ADDR
) {
585 err
= lookup_node(mm
, addr
);
589 } else if (pol
== current
->mempolicy
&&
590 pol
->policy
== MPOL_INTERLEAVE
) {
591 *policy
= current
->il_next
;
597 *policy
= pol
->policy
;
600 up_read(¤t
->mm
->mmap_sem
);
606 get_zonemask(pol
, nmask
);
610 up_read(¤t
->mm
->mmap_sem
);
614 #ifdef CONFIG_MIGRATION
618 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
622 * Avoid migrating a page that is shared with others.
624 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1)
625 isolate_lru_page(page
, pagelist
);
628 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
630 return alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
634 * Migrate pages from one node to a target node.
635 * Returns error or the number of pages not migrated.
637 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
645 node_set(source
, nmask
);
647 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nmask
,
648 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
650 if (!list_empty(&pagelist
))
651 err
= migrate_pages(&pagelist
, new_node_page
, dest
);
657 * Move pages between the two nodesets so as to preserve the physical
658 * layout as much as possible.
660 * Returns the number of page that could not be moved.
662 int do_migrate_pages(struct mm_struct
*mm
,
663 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
670 down_read(&mm
->mmap_sem
);
672 err
= migrate_vmas(mm
, from_nodes
, to_nodes
, flags
);
677 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
678 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
679 * bit in 'tmp', and return that <source, dest> pair for migration.
680 * The pair of nodemasks 'to' and 'from' define the map.
682 * If no pair of bits is found that way, fallback to picking some
683 * pair of 'source' and 'dest' bits that are not the same. If the
684 * 'source' and 'dest' bits are the same, this represents a node
685 * that will be migrating to itself, so no pages need move.
687 * If no bits are left in 'tmp', or if all remaining bits left
688 * in 'tmp' correspond to the same bit in 'to', return false
689 * (nothing left to migrate).
691 * This lets us pick a pair of nodes to migrate between, such that
692 * if possible the dest node is not already occupied by some other
693 * source node, minimizing the risk of overloading the memory on a
694 * node that would happen if we migrated incoming memory to a node
695 * before migrating outgoing memory source that same node.
697 * A single scan of tmp is sufficient. As we go, we remember the
698 * most recent <s, d> pair that moved (s != d). If we find a pair
699 * that not only moved, but what's better, moved to an empty slot
700 * (d is not set in tmp), then we break out then, with that pair.
701 * Otherwise when we finish scannng from_tmp, we at least have the
702 * most recent <s, d> pair that moved. If we get all the way through
703 * the scan of tmp without finding any node that moved, much less
704 * moved to an empty node, then there is nothing left worth migrating.
708 while (!nodes_empty(tmp
)) {
713 for_each_node_mask(s
, tmp
) {
714 d
= node_remap(s
, *from_nodes
, *to_nodes
);
718 source
= s
; /* Node moved. Memorize */
721 /* dest not in remaining from nodes? */
722 if (!node_isset(dest
, tmp
))
728 node_clear(source
, tmp
);
729 err
= migrate_to_node(mm
, source
, dest
, flags
);
736 up_read(&mm
->mmap_sem
);
744 * Allocate a new page for page migration based on vma policy.
745 * Start assuming that page is mapped by vma pointed to by @private.
746 * Search forward from there, if not. N.B., this assumes that the
747 * list of pages handed to migrate_pages()--which is how we get here--
748 * is in virtual address order.
750 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
752 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
753 unsigned long uninitialized_var(address
);
756 address
= page_address_in_vma(page
, vma
);
757 if (address
!= -EFAULT
)
763 * if !vma, alloc_page_vma() will use task or system default policy
765 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
769 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
774 int do_migrate_pages(struct mm_struct
*mm
,
775 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
780 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
786 static long do_mbind(unsigned long start
, unsigned long len
,
787 unsigned long mode
, nodemask_t
*nmask
,
790 struct vm_area_struct
*vma
;
791 struct mm_struct
*mm
= current
->mm
;
792 struct mempolicy
*new;
797 if ((flags
& ~(unsigned long)(MPOL_MF_STRICT
|
798 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
801 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
804 if (start
& ~PAGE_MASK
)
807 if (mode
== MPOL_DEFAULT
)
808 flags
&= ~MPOL_MF_STRICT
;
810 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
818 if (mpol_check_policy(mode
, nmask
))
821 new = mpol_new(mode
, nmask
);
826 * If we are using the default policy then operation
827 * on discontinuous address spaces is okay after all
830 flags
|= MPOL_MF_DISCONTIG_OK
;
832 pr_debug("mbind %lx-%lx mode:%ld nodes:%lx\n",start
,start
+len
,
833 mode
, nmask
? nodes_addr(*nmask
)[0] : -1);
835 down_write(&mm
->mmap_sem
);
836 vma
= check_range(mm
, start
, end
, nmask
,
837 flags
| MPOL_MF_INVERT
, &pagelist
);
843 err
= mbind_range(vma
, start
, end
, new);
845 if (!list_empty(&pagelist
))
846 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
849 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
853 up_write(&mm
->mmap_sem
);
859 * User space interface with variable sized bitmaps for nodelists.
862 /* Copy a node mask from user space. */
863 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
864 unsigned long maxnode
)
867 unsigned long nlongs
;
868 unsigned long endmask
;
872 if (maxnode
== 0 || !nmask
)
874 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
877 nlongs
= BITS_TO_LONGS(maxnode
);
878 if ((maxnode
% BITS_PER_LONG
) == 0)
881 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
883 /* When the user specified more nodes than supported just check
884 if the non supported part is all zero. */
885 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
886 if (nlongs
> PAGE_SIZE
/sizeof(long))
888 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
890 if (get_user(t
, nmask
+ k
))
892 if (k
== nlongs
- 1) {
898 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
902 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
904 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
908 /* Copy a kernel node mask to user space */
909 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
912 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
913 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
916 if (copy
> PAGE_SIZE
)
918 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
922 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
925 asmlinkage
long sys_mbind(unsigned long start
, unsigned long len
,
927 unsigned long __user
*nmask
, unsigned long maxnode
,
933 err
= get_nodes(&nodes
, nmask
, maxnode
);
936 return do_mbind(start
, len
, mode
, &nodes
, flags
);
939 /* Set the process memory policy */
940 asmlinkage
long sys_set_mempolicy(int mode
, unsigned long __user
*nmask
,
941 unsigned long maxnode
)
946 if (mode
< 0 || mode
> MPOL_MAX
)
948 err
= get_nodes(&nodes
, nmask
, maxnode
);
951 return do_set_mempolicy(mode
, &nodes
);
954 asmlinkage
long sys_migrate_pages(pid_t pid
, unsigned long maxnode
,
955 const unsigned long __user
*old_nodes
,
956 const unsigned long __user
*new_nodes
)
958 struct mm_struct
*mm
;
959 struct task_struct
*task
;
962 nodemask_t task_nodes
;
965 err
= get_nodes(&old
, old_nodes
, maxnode
);
969 err
= get_nodes(&new, new_nodes
, maxnode
);
973 /* Find the mm_struct */
974 read_lock(&tasklist_lock
);
975 task
= pid
? find_task_by_vpid(pid
) : current
;
977 read_unlock(&tasklist_lock
);
980 mm
= get_task_mm(task
);
981 read_unlock(&tasklist_lock
);
987 * Check if this process has the right to modify the specified
988 * process. The right exists if the process has administrative
989 * capabilities, superuser privileges or the same
990 * userid as the target process.
992 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
993 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
994 !capable(CAP_SYS_NICE
)) {
999 task_nodes
= cpuset_mems_allowed(task
);
1000 /* Is the user allowed to access the target nodes? */
1001 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1006 if (!nodes_subset(new, node_states
[N_HIGH_MEMORY
])) {
1011 err
= security_task_movememory(task
);
1015 err
= do_migrate_pages(mm
, &old
, &new,
1016 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1023 /* Retrieve NUMA policy */
1024 asmlinkage
long sys_get_mempolicy(int __user
*policy
,
1025 unsigned long __user
*nmask
,
1026 unsigned long maxnode
,
1027 unsigned long addr
, unsigned long flags
)
1030 int uninitialized_var(pval
);
1033 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1036 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1041 if (policy
&& put_user(pval
, policy
))
1045 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1050 #ifdef CONFIG_COMPAT
1052 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1053 compat_ulong_t __user
*nmask
,
1054 compat_ulong_t maxnode
,
1055 compat_ulong_t addr
, compat_ulong_t flags
)
1058 unsigned long __user
*nm
= NULL
;
1059 unsigned long nr_bits
, alloc_size
;
1060 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1062 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1063 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1066 nm
= compat_alloc_user_space(alloc_size
);
1068 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1070 if (!err
&& nmask
) {
1071 err
= copy_from_user(bm
, nm
, alloc_size
);
1072 /* ensure entire bitmap is zeroed */
1073 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1074 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1080 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1081 compat_ulong_t maxnode
)
1084 unsigned long __user
*nm
= NULL
;
1085 unsigned long nr_bits
, alloc_size
;
1086 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1088 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1089 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1092 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1093 nm
= compat_alloc_user_space(alloc_size
);
1094 err
|= copy_to_user(nm
, bm
, alloc_size
);
1100 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1103 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1104 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1105 compat_ulong_t maxnode
, compat_ulong_t flags
)
1108 unsigned long __user
*nm
= NULL
;
1109 unsigned long nr_bits
, alloc_size
;
1112 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1113 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1116 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1117 nm
= compat_alloc_user_space(alloc_size
);
1118 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1124 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1130 * get_vma_policy(@task, @vma, @addr)
1131 * @task - task for fallback if vma policy == default
1132 * @vma - virtual memory area whose policy is sought
1133 * @addr - address in @vma for shared policy lookup
1135 * Returns effective policy for a VMA at specified address.
1136 * Falls back to @task or system default policy, as necessary.
1137 * Returned policy has extra reference count if shared, vma,
1138 * or some other task's policy [show_numa_maps() can pass
1139 * @task != current]. It is the caller's responsibility to
1140 * free the reference in these cases.
1142 static struct mempolicy
* get_vma_policy(struct task_struct
*task
,
1143 struct vm_area_struct
*vma
, unsigned long addr
)
1145 struct mempolicy
*pol
= task
->mempolicy
;
1149 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1150 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1151 shared_pol
= 1; /* if pol non-NULL, add ref below */
1152 } else if (vma
->vm_policy
&&
1153 vma
->vm_policy
->policy
!= MPOL_DEFAULT
)
1154 pol
= vma
->vm_policy
;
1157 pol
= &default_policy
;
1158 else if (!shared_pol
&& pol
!= current
->mempolicy
)
1159 mpol_get(pol
); /* vma or other task's policy */
1163 /* Return a zonelist representing a mempolicy */
1164 static struct zonelist
*zonelist_policy(gfp_t gfp
, struct mempolicy
*policy
)
1168 switch (policy
->policy
) {
1169 case MPOL_PREFERRED
:
1170 nd
= policy
->v
.preferred_node
;
1172 nd
= numa_node_id();
1175 /* Lower zones don't get a policy applied */
1176 /* Careful: current->mems_allowed might have moved */
1177 if (gfp_zone(gfp
) >= policy_zone
)
1178 if (cpuset_zonelist_valid_mems_allowed(policy
->v
.zonelist
))
1179 return policy
->v
.zonelist
;
1181 case MPOL_INTERLEAVE
: /* should not happen */
1183 nd
= numa_node_id();
1189 return node_zonelist(nd
, gfp
);
1192 /* Do dynamic interleaving for a process */
1193 static unsigned interleave_nodes(struct mempolicy
*policy
)
1196 struct task_struct
*me
= current
;
1199 next
= next_node(nid
, policy
->v
.nodes
);
1200 if (next
>= MAX_NUMNODES
)
1201 next
= first_node(policy
->v
.nodes
);
1207 * Depending on the memory policy provide a node from which to allocate the
1210 unsigned slab_node(struct mempolicy
*policy
)
1212 int pol
= policy
? policy
->policy
: MPOL_DEFAULT
;
1215 case MPOL_INTERLEAVE
:
1216 return interleave_nodes(policy
);
1220 * Follow bind policy behavior and start allocation at the
1223 return zonelist_node_idx(policy
->v
.zonelist
->_zonerefs
);
1226 case MPOL_PREFERRED
:
1227 if (policy
->v
.preferred_node
>= 0)
1228 return policy
->v
.preferred_node
;
1232 return numa_node_id();
1236 /* Do static interleaving for a VMA with known offset. */
1237 static unsigned offset_il_node(struct mempolicy
*pol
,
1238 struct vm_area_struct
*vma
, unsigned long off
)
1240 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1241 unsigned target
= (unsigned)off
% nnodes
;
1247 nid
= next_node(nid
, pol
->v
.nodes
);
1249 } while (c
<= target
);
1253 /* Determine a node number for interleave */
1254 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1255 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1261 * for small pages, there is no difference between
1262 * shift and PAGE_SHIFT, so the bit-shift is safe.
1263 * for huge pages, since vm_pgoff is in units of small
1264 * pages, we need to shift off the always 0 bits to get
1267 BUG_ON(shift
< PAGE_SHIFT
);
1268 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1269 off
+= (addr
- vma
->vm_start
) >> shift
;
1270 return offset_il_node(pol
, vma
, off
);
1272 return interleave_nodes(pol
);
1275 #ifdef CONFIG_HUGETLBFS
1277 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1278 * @vma = virtual memory area whose policy is sought
1279 * @addr = address in @vma for shared policy lookup and interleave policy
1280 * @gfp_flags = for requested zone
1281 * @mpol = pointer to mempolicy pointer for reference counted 'BIND policy
1283 * Returns a zonelist suitable for a huge page allocation.
1284 * If the effective policy is 'BIND, returns pointer to policy's zonelist.
1285 * If it is also a policy for which get_vma_policy() returns an extra
1286 * reference, we must hold that reference until after allocation.
1287 * In that case, return policy via @mpol so hugetlb allocation can drop
1288 * the reference. For non-'BIND referenced policies, we can/do drop the
1289 * reference here, so the caller doesn't need to know about the special case
1290 * for default and current task policy.
1292 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1293 gfp_t gfp_flags
, struct mempolicy
**mpol
)
1295 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1296 struct zonelist
*zl
;
1298 *mpol
= NULL
; /* probably no unref needed */
1299 if (pol
->policy
== MPOL_INTERLEAVE
) {
1302 nid
= interleave_nid(pol
, vma
, addr
, HPAGE_SHIFT
);
1303 if (unlikely(pol
!= &default_policy
&&
1304 pol
!= current
->mempolicy
))
1305 __mpol_free(pol
); /* finished with pol */
1306 return node_zonelist(nid
, gfp_flags
);
1309 zl
= zonelist_policy(GFP_HIGHUSER
, pol
);
1310 if (unlikely(pol
!= &default_policy
&& pol
!= current
->mempolicy
)) {
1311 if (pol
->policy
!= MPOL_BIND
)
1312 __mpol_free(pol
); /* finished with pol */
1314 *mpol
= pol
; /* unref needed after allocation */
1320 /* Allocate a page in interleaved policy.
1321 Own path because it needs to do special accounting. */
1322 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1325 struct zonelist
*zl
;
1328 zl
= node_zonelist(nid
, gfp
);
1329 page
= __alloc_pages(gfp
, order
, zl
);
1330 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1331 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1336 * alloc_page_vma - Allocate a page for a VMA.
1339 * %GFP_USER user allocation.
1340 * %GFP_KERNEL kernel allocations,
1341 * %GFP_HIGHMEM highmem/user allocations,
1342 * %GFP_FS allocation should not call back into a file system.
1343 * %GFP_ATOMIC don't sleep.
1345 * @vma: Pointer to VMA or NULL if not available.
1346 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1348 * This function allocates a page from the kernel page pool and applies
1349 * a NUMA policy associated with the VMA or the current process.
1350 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1351 * mm_struct of the VMA to prevent it from going away. Should be used for
1352 * all allocations for pages that will be mapped into
1353 * user space. Returns NULL when no page can be allocated.
1355 * Should be called with the mm_sem of the vma hold.
1358 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1360 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1361 struct zonelist
*zl
;
1363 cpuset_update_task_memory_state();
1365 if (unlikely(pol
->policy
== MPOL_INTERLEAVE
)) {
1368 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1369 if (unlikely(pol
!= &default_policy
&&
1370 pol
!= current
->mempolicy
))
1371 __mpol_free(pol
); /* finished with pol */
1372 return alloc_page_interleave(gfp
, 0, nid
);
1374 zl
= zonelist_policy(gfp
, pol
);
1375 if (pol
!= &default_policy
&& pol
!= current
->mempolicy
) {
1377 * slow path: ref counted policy -- shared or vma
1379 struct page
*page
= __alloc_pages(gfp
, 0, zl
);
1384 * fast path: default or task policy
1386 return __alloc_pages(gfp
, 0, zl
);
1390 * alloc_pages_current - Allocate pages.
1393 * %GFP_USER user allocation,
1394 * %GFP_KERNEL kernel allocation,
1395 * %GFP_HIGHMEM highmem allocation,
1396 * %GFP_FS don't call back into a file system.
1397 * %GFP_ATOMIC don't sleep.
1398 * @order: Power of two of allocation size in pages. 0 is a single page.
1400 * Allocate a page from the kernel page pool. When not in
1401 * interrupt context and apply the current process NUMA policy.
1402 * Returns NULL when no page can be allocated.
1404 * Don't call cpuset_update_task_memory_state() unless
1405 * 1) it's ok to take cpuset_sem (can WAIT), and
1406 * 2) allocating for current task (not interrupt).
1408 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1410 struct mempolicy
*pol
= current
->mempolicy
;
1412 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1413 cpuset_update_task_memory_state();
1414 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
1415 pol
= &default_policy
;
1416 if (pol
->policy
== MPOL_INTERLEAVE
)
1417 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1418 return __alloc_pages(gfp
, order
, zonelist_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);
1443 if (new->policy
== MPOL_BIND
) {
1444 int sz
= ksize(old
->v
.zonelist
);
1445 new->v
.zonelist
= kmemdup(old
->v
.zonelist
, sz
, GFP_KERNEL
);
1446 if (!new->v
.zonelist
) {
1447 kmem_cache_free(policy_cache
, new);
1448 return ERR_PTR(-ENOMEM
);
1454 /* Slow path of a mempolicy comparison */
1455 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1459 if (a
->policy
!= b
->policy
)
1461 switch (a
->policy
) {
1464 case MPOL_INTERLEAVE
:
1465 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1466 case MPOL_PREFERRED
:
1467 return a
->v
.preferred_node
== b
->v
.preferred_node
;
1470 for (i
= 0; a
->v
.zonelist
->_zonerefs
[i
].zone
; i
++) {
1471 struct zone
*za
, *zb
;
1472 za
= zonelist_zone(&a
->v
.zonelist
->_zonerefs
[i
]);
1473 zb
= zonelist_zone(&b
->v
.zonelist
->_zonerefs
[i
]);
1477 return b
->v
.zonelist
->_zonerefs
[i
].zone
== NULL
;
1485 /* Slow path of a mpol destructor. */
1486 void __mpol_free(struct mempolicy
*p
)
1488 if (!atomic_dec_and_test(&p
->refcnt
))
1490 if (p
->policy
== MPOL_BIND
)
1491 kfree(p
->v
.zonelist
);
1492 p
->policy
= MPOL_DEFAULT
;
1493 kmem_cache_free(policy_cache
, p
);
1497 * Shared memory backing store policy support.
1499 * Remember policies even when nobody has shared memory mapped.
1500 * The policies are kept in Red-Black tree linked from the inode.
1501 * They are protected by the sp->lock spinlock, which should be held
1502 * for any accesses to the tree.
1505 /* lookup first element intersecting start-end */
1506 /* Caller holds sp->lock */
1507 static struct sp_node
*
1508 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1510 struct rb_node
*n
= sp
->root
.rb_node
;
1513 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1515 if (start
>= p
->end
)
1517 else if (end
<= p
->start
)
1525 struct sp_node
*w
= NULL
;
1526 struct rb_node
*prev
= rb_prev(n
);
1529 w
= rb_entry(prev
, struct sp_node
, nd
);
1530 if (w
->end
<= start
)
1534 return rb_entry(n
, struct sp_node
, nd
);
1537 /* Insert a new shared policy into the list. */
1538 /* Caller holds sp->lock */
1539 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1541 struct rb_node
**p
= &sp
->root
.rb_node
;
1542 struct rb_node
*parent
= NULL
;
1547 nd
= rb_entry(parent
, struct sp_node
, nd
);
1548 if (new->start
< nd
->start
)
1550 else if (new->end
> nd
->end
)
1551 p
= &(*p
)->rb_right
;
1555 rb_link_node(&new->nd
, parent
, p
);
1556 rb_insert_color(&new->nd
, &sp
->root
);
1557 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
1558 new->policy
? new->policy
->policy
: 0);
1561 /* Find shared policy intersecting idx */
1563 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1565 struct mempolicy
*pol
= NULL
;
1568 if (!sp
->root
.rb_node
)
1570 spin_lock(&sp
->lock
);
1571 sn
= sp_lookup(sp
, idx
, idx
+1);
1573 mpol_get(sn
->policy
);
1576 spin_unlock(&sp
->lock
);
1580 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1582 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
1583 rb_erase(&n
->nd
, &sp
->root
);
1584 mpol_free(n
->policy
);
1585 kmem_cache_free(sn_cache
, n
);
1588 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
1589 struct mempolicy
*pol
)
1591 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1602 /* Replace a policy range. */
1603 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1604 unsigned long end
, struct sp_node
*new)
1606 struct sp_node
*n
, *new2
= NULL
;
1609 spin_lock(&sp
->lock
);
1610 n
= sp_lookup(sp
, start
, end
);
1611 /* Take care of old policies in the same range. */
1612 while (n
&& n
->start
< end
) {
1613 struct rb_node
*next
= rb_next(&n
->nd
);
1614 if (n
->start
>= start
) {
1620 /* Old policy spanning whole new range. */
1623 spin_unlock(&sp
->lock
);
1624 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1630 sp_insert(sp
, new2
);
1638 n
= rb_entry(next
, struct sp_node
, nd
);
1642 spin_unlock(&sp
->lock
);
1644 mpol_free(new2
->policy
);
1645 kmem_cache_free(sn_cache
, new2
);
1650 void mpol_shared_policy_init(struct shared_policy
*info
, int policy
,
1651 nodemask_t
*policy_nodes
)
1653 info
->root
= RB_ROOT
;
1654 spin_lock_init(&info
->lock
);
1656 if (policy
!= MPOL_DEFAULT
) {
1657 struct mempolicy
*newpol
;
1659 /* Falls back to MPOL_DEFAULT on any error */
1660 newpol
= mpol_new(policy
, policy_nodes
);
1661 if (!IS_ERR(newpol
)) {
1662 /* Create pseudo-vma that contains just the policy */
1663 struct vm_area_struct pvma
;
1665 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
1666 /* Policy covers entire file */
1667 pvma
.vm_end
= TASK_SIZE
;
1668 mpol_set_shared_policy(info
, &pvma
, newpol
);
1674 int mpol_set_shared_policy(struct shared_policy
*info
,
1675 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1678 struct sp_node
*new = NULL
;
1679 unsigned long sz
= vma_pages(vma
);
1681 pr_debug("set_shared_policy %lx sz %lu %d %lx\n",
1683 sz
, npol
? npol
->policy
: -1,
1684 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1687 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1691 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1693 kmem_cache_free(sn_cache
, new);
1697 /* Free a backing policy store on inode delete. */
1698 void mpol_free_shared_policy(struct shared_policy
*p
)
1701 struct rb_node
*next
;
1703 if (!p
->root
.rb_node
)
1705 spin_lock(&p
->lock
);
1706 next
= rb_first(&p
->root
);
1708 n
= rb_entry(next
, struct sp_node
, nd
);
1709 next
= rb_next(&n
->nd
);
1710 rb_erase(&n
->nd
, &p
->root
);
1711 mpol_free(n
->policy
);
1712 kmem_cache_free(sn_cache
, n
);
1714 spin_unlock(&p
->lock
);
1717 /* assumes fs == KERNEL_DS */
1718 void __init
numa_policy_init(void)
1720 nodemask_t interleave_nodes
;
1721 unsigned long largest
= 0;
1722 int nid
, prefer
= 0;
1724 policy_cache
= kmem_cache_create("numa_policy",
1725 sizeof(struct mempolicy
),
1726 0, SLAB_PANIC
, NULL
);
1728 sn_cache
= kmem_cache_create("shared_policy_node",
1729 sizeof(struct sp_node
),
1730 0, SLAB_PANIC
, NULL
);
1733 * Set interleaving policy for system init. Interleaving is only
1734 * enabled across suitably sized nodes (default is >= 16MB), or
1735 * fall back to the largest node if they're all smaller.
1737 nodes_clear(interleave_nodes
);
1738 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1739 unsigned long total_pages
= node_present_pages(nid
);
1741 /* Preserve the largest node */
1742 if (largest
< total_pages
) {
1743 largest
= total_pages
;
1747 /* Interleave this node? */
1748 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
1749 node_set(nid
, interleave_nodes
);
1752 /* All too small, use the largest */
1753 if (unlikely(nodes_empty(interleave_nodes
)))
1754 node_set(prefer
, interleave_nodes
);
1756 if (do_set_mempolicy(MPOL_INTERLEAVE
, &interleave_nodes
))
1757 printk("numa_policy_init: interleaving failed\n");
1760 /* Reset policy of current process to default */
1761 void numa_default_policy(void)
1763 do_set_mempolicy(MPOL_DEFAULT
, NULL
);
1766 /* Migrate a policy to a different set of nodes */
1767 static void mpol_rebind_policy(struct mempolicy
*pol
,
1768 const nodemask_t
*newmask
)
1770 nodemask_t
*mpolmask
;
1775 mpolmask
= &pol
->cpuset_mems_allowed
;
1776 if (nodes_equal(*mpolmask
, *newmask
))
1779 switch (pol
->policy
) {
1782 case MPOL_INTERLEAVE
:
1783 nodes_remap(tmp
, pol
->v
.nodes
, *mpolmask
, *newmask
);
1785 *mpolmask
= *newmask
;
1786 current
->il_next
= node_remap(current
->il_next
,
1787 *mpolmask
, *newmask
);
1789 case MPOL_PREFERRED
:
1790 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
1791 *mpolmask
, *newmask
);
1792 *mpolmask
= *newmask
;
1797 struct zonelist
*zonelist
;
1800 for (z
= pol
->v
.zonelist
->_zonerefs
; z
->zone
; z
++)
1801 node_set(zonelist_node_idx(z
), nodes
);
1802 nodes_remap(tmp
, nodes
, *mpolmask
, *newmask
);
1805 zonelist
= bind_zonelist(&nodes
);
1807 /* If no mem, then zonelist is NULL and we keep old zonelist.
1808 * If that old zonelist has no remaining mems_allowed nodes,
1809 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1812 if (!IS_ERR(zonelist
)) {
1813 /* Good - got mem - substitute new zonelist */
1814 kfree(pol
->v
.zonelist
);
1815 pol
->v
.zonelist
= zonelist
;
1817 *mpolmask
= *newmask
;
1827 * Wrapper for mpol_rebind_policy() that just requires task
1828 * pointer, and updates task mempolicy.
1831 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
1833 mpol_rebind_policy(tsk
->mempolicy
, new);
1837 * Rebind each vma in mm to new nodemask.
1839 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1842 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
1844 struct vm_area_struct
*vma
;
1846 down_write(&mm
->mmap_sem
);
1847 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
1848 mpol_rebind_policy(vma
->vm_policy
, new);
1849 up_write(&mm
->mmap_sem
);
1853 * Display pages allocated per node and memory policy via /proc.
1856 static const char * const policy_types
[] =
1857 { "default", "prefer", "bind", "interleave" };
1860 * Convert a mempolicy into a string.
1861 * Returns the number of characters in buffer (if positive)
1862 * or an error (negative)
1864 static inline int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
1869 int mode
= pol
? pol
->policy
: MPOL_DEFAULT
;
1876 case MPOL_PREFERRED
:
1878 node_set(pol
->v
.preferred_node
, nodes
);
1882 get_zonemask(pol
, &nodes
);
1885 case MPOL_INTERLEAVE
:
1886 nodes
= pol
->v
.nodes
;
1894 l
= strlen(policy_types
[mode
]);
1895 if (buffer
+ maxlen
< p
+ l
+ 1)
1898 strcpy(p
, policy_types
[mode
]);
1901 if (!nodes_empty(nodes
)) {
1902 if (buffer
+ maxlen
< p
+ 2)
1905 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
1911 unsigned long pages
;
1913 unsigned long active
;
1914 unsigned long writeback
;
1915 unsigned long mapcount_max
;
1916 unsigned long dirty
;
1917 unsigned long swapcache
;
1918 unsigned long node
[MAX_NUMNODES
];
1921 static void gather_stats(struct page
*page
, void *private, int pte_dirty
)
1923 struct numa_maps
*md
= private;
1924 int count
= page_mapcount(page
);
1927 if (pte_dirty
|| PageDirty(page
))
1930 if (PageSwapCache(page
))
1933 if (PageActive(page
))
1936 if (PageWriteback(page
))
1942 if (count
> md
->mapcount_max
)
1943 md
->mapcount_max
= count
;
1945 md
->node
[page_to_nid(page
)]++;
1948 #ifdef CONFIG_HUGETLB_PAGE
1949 static void check_huge_range(struct vm_area_struct
*vma
,
1950 unsigned long start
, unsigned long end
,
1951 struct numa_maps
*md
)
1956 for (addr
= start
; addr
< end
; addr
+= HPAGE_SIZE
) {
1957 pte_t
*ptep
= huge_pte_offset(vma
->vm_mm
, addr
& HPAGE_MASK
);
1967 page
= pte_page(pte
);
1971 gather_stats(page
, md
, pte_dirty(*ptep
));
1975 static inline void check_huge_range(struct vm_area_struct
*vma
,
1976 unsigned long start
, unsigned long end
,
1977 struct numa_maps
*md
)
1982 int show_numa_map(struct seq_file
*m
, void *v
)
1984 struct proc_maps_private
*priv
= m
->private;
1985 struct vm_area_struct
*vma
= v
;
1986 struct numa_maps
*md
;
1987 struct file
*file
= vma
->vm_file
;
1988 struct mm_struct
*mm
= vma
->vm_mm
;
1989 struct mempolicy
*pol
;
1996 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
2000 pol
= get_vma_policy(priv
->task
, vma
, vma
->vm_start
);
2001 mpol_to_str(buffer
, sizeof(buffer
), pol
);
2003 * unref shared or other task's mempolicy
2005 if (pol
!= &default_policy
&& pol
!= current
->mempolicy
)
2008 seq_printf(m
, "%08lx %s", vma
->vm_start
, buffer
);
2011 seq_printf(m
, " file=");
2012 seq_path(m
, &file
->f_path
, "\n\t= ");
2013 } else if (vma
->vm_start
<= mm
->brk
&& vma
->vm_end
>= mm
->start_brk
) {
2014 seq_printf(m
, " heap");
2015 } else if (vma
->vm_start
<= mm
->start_stack
&&
2016 vma
->vm_end
>= mm
->start_stack
) {
2017 seq_printf(m
, " stack");
2020 if (is_vm_hugetlb_page(vma
)) {
2021 check_huge_range(vma
, vma
->vm_start
, vma
->vm_end
, md
);
2022 seq_printf(m
, " huge");
2024 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
2025 &node_states
[N_HIGH_MEMORY
], MPOL_MF_STATS
, md
);
2032 seq_printf(m
," anon=%lu",md
->anon
);
2035 seq_printf(m
," dirty=%lu",md
->dirty
);
2037 if (md
->pages
!= md
->anon
&& md
->pages
!= md
->dirty
)
2038 seq_printf(m
, " mapped=%lu", md
->pages
);
2040 if (md
->mapcount_max
> 1)
2041 seq_printf(m
, " mapmax=%lu", md
->mapcount_max
);
2044 seq_printf(m
," swapcache=%lu", md
->swapcache
);
2046 if (md
->active
< md
->pages
&& !is_vm_hugetlb_page(vma
))
2047 seq_printf(m
," active=%lu", md
->active
);
2050 seq_printf(m
," writeback=%lu", md
->writeback
);
2052 for_each_node_state(n
, N_HIGH_MEMORY
)
2054 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
2059 if (m
->count
< m
->size
)
2060 m
->version
= (vma
!= priv
->tail_vma
) ? vma
->vm_start
: 0;