2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/gfp.h>
77 #include <linux/slab.h>
78 #include <linux/string.h>
79 #include <linux/module.h>
80 #include <linux/nsproxy.h>
81 #include <linux/interrupt.h>
82 #include <linux/init.h>
83 #include <linux/compat.h>
84 #include <linux/swap.h>
85 #include <linux/seq_file.h>
86 #include <linux/proc_fs.h>
87 #include <linux/migrate.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.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;
109 * run-time system-wide default policy => local allocation
111 struct mempolicy default_policy
= {
112 .refcnt
= ATOMIC_INIT(1), /* never free it */
113 .mode
= MPOL_PREFERRED
,
114 .flags
= MPOL_F_LOCAL
,
117 static const struct mempolicy_operations
{
118 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
119 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
120 } mpol_ops
[MPOL_MAX
];
122 /* Check that the nodemask contains at least one populated zone */
123 static int is_valid_nodemask(const nodemask_t
*nodemask
)
127 /* Check that there is something useful in this mask */
130 for_each_node_mask(nd
, *nodemask
) {
133 for (k
= 0; k
<= policy_zone
; k
++) {
134 z
= &NODE_DATA(nd
)->node_zones
[k
];
135 if (z
->present_pages
> 0)
143 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
145 return pol
->flags
& (MPOL_F_STATIC_NODES
| MPOL_F_RELATIVE_NODES
);
148 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
149 const nodemask_t
*rel
)
152 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
153 nodes_onto(*ret
, tmp
, *rel
);
156 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
158 if (nodes_empty(*nodes
))
160 pol
->v
.nodes
= *nodes
;
164 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
167 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
168 else if (nodes_empty(*nodes
))
169 return -EINVAL
; /* no allowed nodes */
171 pol
->v
.preferred_node
= first_node(*nodes
);
175 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
177 if (!is_valid_nodemask(nodes
))
179 pol
->v
.nodes
= *nodes
;
183 /* Create a new policy */
184 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
187 struct mempolicy
*policy
;
188 nodemask_t cpuset_context_nmask
;
191 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
192 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : -1);
194 if (mode
== MPOL_DEFAULT
) {
195 if (nodes
&& !nodes_empty(*nodes
))
196 return ERR_PTR(-EINVAL
);
197 return NULL
; /* simply delete any existing policy */
202 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
203 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
204 * All other modes require a valid pointer to a non-empty nodemask.
206 if (mode
== MPOL_PREFERRED
) {
207 if (nodes_empty(*nodes
)) {
208 if (((flags
& MPOL_F_STATIC_NODES
) ||
209 (flags
& MPOL_F_RELATIVE_NODES
)))
210 return ERR_PTR(-EINVAL
);
211 nodes
= NULL
; /* flag local alloc */
213 } else if (nodes_empty(*nodes
))
214 return ERR_PTR(-EINVAL
);
215 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
217 return ERR_PTR(-ENOMEM
);
218 atomic_set(&policy
->refcnt
, 1);
220 policy
->flags
= flags
;
224 * cpuset related setup doesn't apply to local allocation
226 cpuset_update_task_memory_state();
227 if (flags
& MPOL_F_RELATIVE_NODES
)
228 mpol_relative_nodemask(&cpuset_context_nmask
, nodes
,
229 &cpuset_current_mems_allowed
);
231 nodes_and(cpuset_context_nmask
, *nodes
,
232 cpuset_current_mems_allowed
);
233 if (mpol_store_user_nodemask(policy
))
234 policy
->w
.user_nodemask
= *nodes
;
236 policy
->w
.cpuset_mems_allowed
=
237 cpuset_mems_allowed(current
);
240 ret
= mpol_ops
[mode
].create(policy
,
241 nodes
? &cpuset_context_nmask
: NULL
);
243 kmem_cache_free(policy_cache
, policy
);
249 /* Slow path of a mpol destructor. */
250 void __mpol_put(struct mempolicy
*p
)
252 if (!atomic_dec_and_test(&p
->refcnt
))
254 kmem_cache_free(policy_cache
, p
);
257 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
261 static void mpol_rebind_nodemask(struct mempolicy
*pol
,
262 const nodemask_t
*nodes
)
266 if (pol
->flags
& MPOL_F_STATIC_NODES
)
267 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
268 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
269 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
271 nodes_remap(tmp
, pol
->v
.nodes
, pol
->w
.cpuset_mems_allowed
,
273 pol
->w
.cpuset_mems_allowed
= *nodes
;
277 if (!node_isset(current
->il_next
, tmp
)) {
278 current
->il_next
= next_node(current
->il_next
, tmp
);
279 if (current
->il_next
>= MAX_NUMNODES
)
280 current
->il_next
= first_node(tmp
);
281 if (current
->il_next
>= MAX_NUMNODES
)
282 current
->il_next
= numa_node_id();
286 static void mpol_rebind_preferred(struct mempolicy
*pol
,
287 const nodemask_t
*nodes
)
291 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
292 int node
= first_node(pol
->w
.user_nodemask
);
294 if (node_isset(node
, *nodes
)) {
295 pol
->v
.preferred_node
= node
;
296 pol
->flags
&= ~MPOL_F_LOCAL
;
298 pol
->flags
|= MPOL_F_LOCAL
;
299 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
300 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
301 pol
->v
.preferred_node
= first_node(tmp
);
302 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
303 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
304 pol
->w
.cpuset_mems_allowed
,
306 pol
->w
.cpuset_mems_allowed
= *nodes
;
310 /* Migrate a policy to a different set of nodes */
311 static void mpol_rebind_policy(struct mempolicy
*pol
,
312 const nodemask_t
*newmask
)
316 if (!mpol_store_user_nodemask(pol
) &&
317 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
319 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
323 * Wrapper for mpol_rebind_policy() that just requires task
324 * pointer, and updates task mempolicy.
327 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
329 mpol_rebind_policy(tsk
->mempolicy
, new);
333 * Rebind each vma in mm to new nodemask.
335 * Call holding a reference to mm. Takes mm->mmap_sem during call.
338 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
340 struct vm_area_struct
*vma
;
342 down_write(&mm
->mmap_sem
);
343 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
344 mpol_rebind_policy(vma
->vm_policy
, new);
345 up_write(&mm
->mmap_sem
);
348 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
350 .rebind
= mpol_rebind_default
,
352 [MPOL_INTERLEAVE
] = {
353 .create
= mpol_new_interleave
,
354 .rebind
= mpol_rebind_nodemask
,
357 .create
= mpol_new_preferred
,
358 .rebind
= mpol_rebind_preferred
,
361 .create
= mpol_new_bind
,
362 .rebind
= mpol_rebind_nodemask
,
366 static void gather_stats(struct page
*, void *, int pte_dirty
);
367 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
368 unsigned long flags
);
370 /* Scan through pages checking if pages follow certain conditions. */
371 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
372 unsigned long addr
, unsigned long end
,
373 const nodemask_t
*nodes
, unsigned long flags
,
380 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
385 if (!pte_present(*pte
))
387 page
= vm_normal_page(vma
, addr
, *pte
);
391 * The check for PageReserved here is important to avoid
392 * handling zero pages and other pages that may have been
393 * marked special by the system.
395 * If the PageReserved would not be checked here then f.e.
396 * the location of the zero page could have an influence
397 * on MPOL_MF_STRICT, zero pages would be counted for
398 * the per node stats, and there would be useless attempts
399 * to put zero pages on the migration list.
401 if (PageReserved(page
))
403 nid
= page_to_nid(page
);
404 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
407 if (flags
& MPOL_MF_STATS
)
408 gather_stats(page
, private, pte_dirty(*pte
));
409 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
410 migrate_page_add(page
, private, flags
);
413 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
414 pte_unmap_unlock(orig_pte
, ptl
);
418 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
419 unsigned long addr
, unsigned long end
,
420 const nodemask_t
*nodes
, unsigned long flags
,
426 pmd
= pmd_offset(pud
, addr
);
428 next
= pmd_addr_end(addr
, end
);
429 if (pmd_none_or_clear_bad(pmd
))
431 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
434 } while (pmd
++, addr
= next
, addr
!= end
);
438 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
439 unsigned long addr
, unsigned long end
,
440 const nodemask_t
*nodes
, unsigned long flags
,
446 pud
= pud_offset(pgd
, addr
);
448 next
= pud_addr_end(addr
, end
);
449 if (pud_none_or_clear_bad(pud
))
451 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
454 } while (pud
++, addr
= next
, addr
!= end
);
458 static inline int check_pgd_range(struct vm_area_struct
*vma
,
459 unsigned long addr
, unsigned long end
,
460 const nodemask_t
*nodes
, unsigned long flags
,
466 pgd
= pgd_offset(vma
->vm_mm
, addr
);
468 next
= pgd_addr_end(addr
, end
);
469 if (pgd_none_or_clear_bad(pgd
))
471 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
474 } while (pgd
++, addr
= next
, addr
!= end
);
479 * Check if all pages in a range are on a set of nodes.
480 * If pagelist != NULL then isolate pages from the LRU and
481 * put them on the pagelist.
483 static struct vm_area_struct
*
484 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
485 const nodemask_t
*nodes
, unsigned long flags
, void *private)
488 struct vm_area_struct
*first
, *vma
, *prev
;
490 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
492 err
= migrate_prep();
497 first
= find_vma(mm
, start
);
499 return ERR_PTR(-EFAULT
);
501 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
502 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
503 if (!vma
->vm_next
&& vma
->vm_end
< end
)
504 return ERR_PTR(-EFAULT
);
505 if (prev
&& prev
->vm_end
< vma
->vm_start
)
506 return ERR_PTR(-EFAULT
);
508 if (!is_vm_hugetlb_page(vma
) &&
509 ((flags
& MPOL_MF_STRICT
) ||
510 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
511 vma_migratable(vma
)))) {
512 unsigned long endvma
= vma
->vm_end
;
516 if (vma
->vm_start
> start
)
517 start
= vma
->vm_start
;
518 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
521 first
= ERR_PTR(err
);
530 /* Apply policy to a single VMA */
531 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
534 struct mempolicy
*old
= vma
->vm_policy
;
536 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
537 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
538 vma
->vm_ops
, vma
->vm_file
,
539 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
541 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
542 err
= vma
->vm_ops
->set_policy(vma
, new);
545 vma
->vm_policy
= new;
551 /* Step 2: apply policy to a range and do splits. */
552 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
553 unsigned long end
, struct mempolicy
*new)
555 struct vm_area_struct
*next
;
559 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
561 if (vma
->vm_start
< start
)
562 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
563 if (!err
&& vma
->vm_end
> end
)
564 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
566 err
= policy_vma(vma
, new);
574 * Update task->flags PF_MEMPOLICY bit: set iff non-default
575 * mempolicy. Allows more rapid checking of this (combined perhaps
576 * with other PF_* flag bits) on memory allocation hot code paths.
578 * If called from outside this file, the task 'p' should -only- be
579 * a newly forked child not yet visible on the task list, because
580 * manipulating the task flags of a visible task is not safe.
582 * The above limitation is why this routine has the funny name
583 * mpol_fix_fork_child_flag().
585 * It is also safe to call this with a task pointer of current,
586 * which the static wrapper mpol_set_task_struct_flag() does,
587 * for use within this file.
590 void mpol_fix_fork_child_flag(struct task_struct
*p
)
593 p
->flags
|= PF_MEMPOLICY
;
595 p
->flags
&= ~PF_MEMPOLICY
;
598 static void mpol_set_task_struct_flag(void)
600 mpol_fix_fork_child_flag(current
);
603 /* Set the process memory policy */
604 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
607 struct mempolicy
*new;
608 struct mm_struct
*mm
= current
->mm
;
610 new = mpol_new(mode
, flags
, nodes
);
615 * prevent changing our mempolicy while show_numa_maps()
617 * Note: do_set_mempolicy() can be called at init time
621 down_write(&mm
->mmap_sem
);
622 mpol_put(current
->mempolicy
);
623 current
->mempolicy
= new;
624 mpol_set_task_struct_flag();
625 if (new && new->mode
== MPOL_INTERLEAVE
&&
626 nodes_weight(new->v
.nodes
))
627 current
->il_next
= first_node(new->v
.nodes
);
629 up_write(&mm
->mmap_sem
);
635 * Return nodemask for policy for get_mempolicy() query
637 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
640 if (p
== &default_policy
)
646 case MPOL_INTERLEAVE
:
650 if (!(p
->flags
& MPOL_F_LOCAL
))
651 node_set(p
->v
.preferred_node
, *nodes
);
652 /* else return empty node mask for local allocation */
659 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
664 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
666 err
= page_to_nid(p
);
672 /* Retrieve NUMA policy */
673 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
674 unsigned long addr
, unsigned long flags
)
677 struct mm_struct
*mm
= current
->mm
;
678 struct vm_area_struct
*vma
= NULL
;
679 struct mempolicy
*pol
= current
->mempolicy
;
681 cpuset_update_task_memory_state();
683 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
686 if (flags
& MPOL_F_MEMS_ALLOWED
) {
687 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
689 *policy
= 0; /* just so it's initialized */
690 *nmask
= cpuset_current_mems_allowed
;
694 if (flags
& MPOL_F_ADDR
) {
696 * Do NOT fall back to task policy if the
697 * vma/shared policy at addr is NULL. We
698 * want to return MPOL_DEFAULT in this case.
700 down_read(&mm
->mmap_sem
);
701 vma
= find_vma_intersection(mm
, addr
, addr
+1);
703 up_read(&mm
->mmap_sem
);
706 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
707 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
709 pol
= vma
->vm_policy
;
714 pol
= &default_policy
; /* indicates default behavior */
716 if (flags
& MPOL_F_NODE
) {
717 if (flags
& MPOL_F_ADDR
) {
718 err
= lookup_node(mm
, addr
);
722 } else if (pol
== current
->mempolicy
&&
723 pol
->mode
== MPOL_INTERLEAVE
) {
724 *policy
= current
->il_next
;
730 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
733 * Internal mempolicy flags must be masked off before exposing
734 * the policy to userspace.
736 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
740 up_read(¤t
->mm
->mmap_sem
);
746 get_policy_nodemask(pol
, nmask
);
751 up_read(¤t
->mm
->mmap_sem
);
755 #ifdef CONFIG_MIGRATION
759 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
763 * Avoid migrating a page that is shared with others.
765 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1)
766 isolate_lru_page(page
, pagelist
);
769 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
771 return alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
775 * Migrate pages from one node to a target node.
776 * Returns error or the number of pages not migrated.
778 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
786 node_set(source
, nmask
);
788 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nmask
,
789 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
791 if (!list_empty(&pagelist
))
792 err
= migrate_pages(&pagelist
, new_node_page
, dest
);
798 * Move pages between the two nodesets so as to preserve the physical
799 * layout as much as possible.
801 * Returns the number of page that could not be moved.
803 int do_migrate_pages(struct mm_struct
*mm
,
804 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
810 down_read(&mm
->mmap_sem
);
812 err
= migrate_vmas(mm
, from_nodes
, to_nodes
, flags
);
817 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
818 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
819 * bit in 'tmp', and return that <source, dest> pair for migration.
820 * The pair of nodemasks 'to' and 'from' define the map.
822 * If no pair of bits is found that way, fallback to picking some
823 * pair of 'source' and 'dest' bits that are not the same. If the
824 * 'source' and 'dest' bits are the same, this represents a node
825 * that will be migrating to itself, so no pages need move.
827 * If no bits are left in 'tmp', or if all remaining bits left
828 * in 'tmp' correspond to the same bit in 'to', return false
829 * (nothing left to migrate).
831 * This lets us pick a pair of nodes to migrate between, such that
832 * if possible the dest node is not already occupied by some other
833 * source node, minimizing the risk of overloading the memory on a
834 * node that would happen if we migrated incoming memory to a node
835 * before migrating outgoing memory source that same node.
837 * A single scan of tmp is sufficient. As we go, we remember the
838 * most recent <s, d> pair that moved (s != d). If we find a pair
839 * that not only moved, but what's better, moved to an empty slot
840 * (d is not set in tmp), then we break out then, with that pair.
841 * Otherwise when we finish scannng from_tmp, we at least have the
842 * most recent <s, d> pair that moved. If we get all the way through
843 * the scan of tmp without finding any node that moved, much less
844 * moved to an empty node, then there is nothing left worth migrating.
848 while (!nodes_empty(tmp
)) {
853 for_each_node_mask(s
, tmp
) {
854 d
= node_remap(s
, *from_nodes
, *to_nodes
);
858 source
= s
; /* Node moved. Memorize */
861 /* dest not in remaining from nodes? */
862 if (!node_isset(dest
, tmp
))
868 node_clear(source
, tmp
);
869 err
= migrate_to_node(mm
, source
, dest
, flags
);
876 up_read(&mm
->mmap_sem
);
884 * Allocate a new page for page migration based on vma policy.
885 * Start assuming that page is mapped by vma pointed to by @private.
886 * Search forward from there, if not. N.B., this assumes that the
887 * list of pages handed to migrate_pages()--which is how we get here--
888 * is in virtual address order.
890 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
892 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
893 unsigned long uninitialized_var(address
);
896 address
= page_address_in_vma(page
, vma
);
897 if (address
!= -EFAULT
)
903 * if !vma, alloc_page_vma() will use task or system default policy
905 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
909 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
914 int do_migrate_pages(struct mm_struct
*mm
,
915 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
920 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
926 static long do_mbind(unsigned long start
, unsigned long len
,
927 unsigned short mode
, unsigned short mode_flags
,
928 nodemask_t
*nmask
, unsigned long flags
)
930 struct vm_area_struct
*vma
;
931 struct mm_struct
*mm
= current
->mm
;
932 struct mempolicy
*new;
937 if (flags
& ~(unsigned long)(MPOL_MF_STRICT
|
938 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
940 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
943 if (start
& ~PAGE_MASK
)
946 if (mode
== MPOL_DEFAULT
)
947 flags
&= ~MPOL_MF_STRICT
;
949 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
957 new = mpol_new(mode
, mode_flags
, nmask
);
962 * If we are using the default policy then operation
963 * on discontinuous address spaces is okay after all
966 flags
|= MPOL_MF_DISCONTIG_OK
;
968 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
969 start
, start
+ len
, mode
, mode_flags
,
970 nmask
? nodes_addr(*nmask
)[0] : -1);
972 down_write(&mm
->mmap_sem
);
973 vma
= check_range(mm
, start
, end
, nmask
,
974 flags
| MPOL_MF_INVERT
, &pagelist
);
980 err
= mbind_range(vma
, start
, end
, new);
982 if (!list_empty(&pagelist
))
983 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
986 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
990 up_write(&mm
->mmap_sem
);
996 * User space interface with variable sized bitmaps for nodelists.
999 /* Copy a node mask from user space. */
1000 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1001 unsigned long maxnode
)
1004 unsigned long nlongs
;
1005 unsigned long endmask
;
1008 nodes_clear(*nodes
);
1009 if (maxnode
== 0 || !nmask
)
1011 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1014 nlongs
= BITS_TO_LONGS(maxnode
);
1015 if ((maxnode
% BITS_PER_LONG
) == 0)
1018 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1020 /* When the user specified more nodes than supported just check
1021 if the non supported part is all zero. */
1022 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1023 if (nlongs
> PAGE_SIZE
/sizeof(long))
1025 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1027 if (get_user(t
, nmask
+ k
))
1029 if (k
== nlongs
- 1) {
1035 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1039 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1041 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1045 /* Copy a kernel node mask to user space */
1046 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1049 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1050 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1052 if (copy
> nbytes
) {
1053 if (copy
> PAGE_SIZE
)
1055 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1059 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1062 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1063 unsigned long, mode
, unsigned long __user
*, nmask
,
1064 unsigned long, maxnode
, unsigned, flags
)
1068 unsigned short mode_flags
;
1070 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1071 mode
&= ~MPOL_MODE_FLAGS
;
1072 if (mode
>= MPOL_MAX
)
1074 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1075 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1077 err
= get_nodes(&nodes
, nmask
, maxnode
);
1080 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1083 /* Set the process memory policy */
1084 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, unsigned long __user
*, nmask
,
1085 unsigned long, maxnode
)
1089 unsigned short flags
;
1091 flags
= mode
& MPOL_MODE_FLAGS
;
1092 mode
&= ~MPOL_MODE_FLAGS
;
1093 if ((unsigned int)mode
>= MPOL_MAX
)
1095 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1097 err
= get_nodes(&nodes
, nmask
, maxnode
);
1100 return do_set_mempolicy(mode
, flags
, &nodes
);
1103 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1104 const unsigned long __user
*, old_nodes
,
1105 const unsigned long __user
*, new_nodes
)
1107 struct mm_struct
*mm
;
1108 struct task_struct
*task
;
1111 nodemask_t task_nodes
;
1114 err
= get_nodes(&old
, old_nodes
, maxnode
);
1118 err
= get_nodes(&new, new_nodes
, maxnode
);
1122 /* Find the mm_struct */
1123 read_lock(&tasklist_lock
);
1124 task
= pid
? find_task_by_vpid(pid
) : current
;
1126 read_unlock(&tasklist_lock
);
1129 mm
= get_task_mm(task
);
1130 read_unlock(&tasklist_lock
);
1136 * Check if this process has the right to modify the specified
1137 * process. The right exists if the process has administrative
1138 * capabilities, superuser privileges or the same
1139 * userid as the target process.
1141 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
1142 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
1143 !capable(CAP_SYS_NICE
)) {
1148 task_nodes
= cpuset_mems_allowed(task
);
1149 /* Is the user allowed to access the target nodes? */
1150 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1155 if (!nodes_subset(new, node_states
[N_HIGH_MEMORY
])) {
1160 err
= security_task_movememory(task
);
1164 err
= do_migrate_pages(mm
, &old
, &new,
1165 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1172 /* Retrieve NUMA policy */
1173 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1174 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1175 unsigned long, addr
, unsigned long, flags
)
1178 int uninitialized_var(pval
);
1181 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1184 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1189 if (policy
&& put_user(pval
, policy
))
1193 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1198 #ifdef CONFIG_COMPAT
1200 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1201 compat_ulong_t __user
*nmask
,
1202 compat_ulong_t maxnode
,
1203 compat_ulong_t addr
, compat_ulong_t flags
)
1206 unsigned long __user
*nm
= NULL
;
1207 unsigned long nr_bits
, alloc_size
;
1208 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1210 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1211 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1214 nm
= compat_alloc_user_space(alloc_size
);
1216 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1218 if (!err
&& nmask
) {
1219 err
= copy_from_user(bm
, nm
, alloc_size
);
1220 /* ensure entire bitmap is zeroed */
1221 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1222 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1228 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1229 compat_ulong_t maxnode
)
1232 unsigned long __user
*nm
= NULL
;
1233 unsigned long nr_bits
, alloc_size
;
1234 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1236 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1237 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1240 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1241 nm
= compat_alloc_user_space(alloc_size
);
1242 err
|= copy_to_user(nm
, bm
, alloc_size
);
1248 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1251 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1252 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1253 compat_ulong_t maxnode
, compat_ulong_t flags
)
1256 unsigned long __user
*nm
= NULL
;
1257 unsigned long nr_bits
, alloc_size
;
1260 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1261 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1264 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1265 nm
= compat_alloc_user_space(alloc_size
);
1266 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1272 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1278 * get_vma_policy(@task, @vma, @addr)
1279 * @task - task for fallback if vma policy == default
1280 * @vma - virtual memory area whose policy is sought
1281 * @addr - address in @vma for shared policy lookup
1283 * Returns effective policy for a VMA at specified address.
1284 * Falls back to @task or system default policy, as necessary.
1285 * Current or other task's task mempolicy and non-shared vma policies
1286 * are protected by the task's mmap_sem, which must be held for read by
1288 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1289 * count--added by the get_policy() vm_op, as appropriate--to protect against
1290 * freeing by another task. It is the caller's responsibility to free the
1291 * extra reference for shared policies.
1293 static struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1294 struct vm_area_struct
*vma
, unsigned long addr
)
1296 struct mempolicy
*pol
= task
->mempolicy
;
1299 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1300 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1304 } else if (vma
->vm_policy
)
1305 pol
= vma
->vm_policy
;
1308 pol
= &default_policy
;
1313 * Return a nodemask representing a mempolicy for filtering nodes for
1316 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1318 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1319 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1320 gfp_zone(gfp
) >= policy_zone
&&
1321 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1322 return &policy
->v
.nodes
;
1327 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1328 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
)
1330 int nd
= numa_node_id();
1332 switch (policy
->mode
) {
1333 case MPOL_PREFERRED
:
1334 if (!(policy
->flags
& MPOL_F_LOCAL
))
1335 nd
= policy
->v
.preferred_node
;
1339 * Normally, MPOL_BIND allocations are node-local within the
1340 * allowed nodemask. However, if __GFP_THISNODE is set and the
1341 * current node is part of the mask, we use the zonelist for
1342 * the first node in the mask instead.
1344 if (unlikely(gfp
& __GFP_THISNODE
) &&
1345 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1346 nd
= first_node(policy
->v
.nodes
);
1348 case MPOL_INTERLEAVE
: /* should not happen */
1353 return node_zonelist(nd
, gfp
);
1356 /* Do dynamic interleaving for a process */
1357 static unsigned interleave_nodes(struct mempolicy
*policy
)
1360 struct task_struct
*me
= current
;
1363 next
= next_node(nid
, policy
->v
.nodes
);
1364 if (next
>= MAX_NUMNODES
)
1365 next
= first_node(policy
->v
.nodes
);
1366 if (next
< MAX_NUMNODES
)
1372 * Depending on the memory policy provide a node from which to allocate the
1374 * @policy must be protected by freeing by the caller. If @policy is
1375 * the current task's mempolicy, this protection is implicit, as only the
1376 * task can change it's policy. The system default policy requires no
1379 unsigned slab_node(struct mempolicy
*policy
)
1381 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1382 return numa_node_id();
1384 switch (policy
->mode
) {
1385 case MPOL_PREFERRED
:
1387 * handled MPOL_F_LOCAL above
1389 return policy
->v
.preferred_node
;
1391 case MPOL_INTERLEAVE
:
1392 return interleave_nodes(policy
);
1396 * Follow bind policy behavior and start allocation at the
1399 struct zonelist
*zonelist
;
1401 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1402 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1403 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1414 /* Do static interleaving for a VMA with known offset. */
1415 static unsigned offset_il_node(struct mempolicy
*pol
,
1416 struct vm_area_struct
*vma
, unsigned long off
)
1418 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1424 return numa_node_id();
1425 target
= (unsigned int)off
% nnodes
;
1428 nid
= next_node(nid
, pol
->v
.nodes
);
1430 } while (c
<= target
);
1434 /* Determine a node number for interleave */
1435 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1436 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1442 * for small pages, there is no difference between
1443 * shift and PAGE_SHIFT, so the bit-shift is safe.
1444 * for huge pages, since vm_pgoff is in units of small
1445 * pages, we need to shift off the always 0 bits to get
1448 BUG_ON(shift
< PAGE_SHIFT
);
1449 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1450 off
+= (addr
- vma
->vm_start
) >> shift
;
1451 return offset_il_node(pol
, vma
, off
);
1453 return interleave_nodes(pol
);
1456 #ifdef CONFIG_HUGETLBFS
1458 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1459 * @vma = virtual memory area whose policy is sought
1460 * @addr = address in @vma for shared policy lookup and interleave policy
1461 * @gfp_flags = for requested zone
1462 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1463 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1465 * Returns a zonelist suitable for a huge page allocation and a pointer
1466 * to the struct mempolicy for conditional unref after allocation.
1467 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1468 * @nodemask for filtering the zonelist.
1470 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1471 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1472 nodemask_t
**nodemask
)
1474 struct zonelist
*zl
;
1476 *mpol
= get_vma_policy(current
, vma
, addr
);
1477 *nodemask
= NULL
; /* assume !MPOL_BIND */
1479 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1480 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1481 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1483 zl
= policy_zonelist(gfp_flags
, *mpol
);
1484 if ((*mpol
)->mode
== MPOL_BIND
)
1485 *nodemask
= &(*mpol
)->v
.nodes
;
1491 /* Allocate a page in interleaved policy.
1492 Own path because it needs to do special accounting. */
1493 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1496 struct zonelist
*zl
;
1499 zl
= node_zonelist(nid
, gfp
);
1500 page
= __alloc_pages(gfp
, order
, zl
);
1501 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1502 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1507 * alloc_page_vma - Allocate a page for a VMA.
1510 * %GFP_USER user allocation.
1511 * %GFP_KERNEL kernel allocations,
1512 * %GFP_HIGHMEM highmem/user allocations,
1513 * %GFP_FS allocation should not call back into a file system.
1514 * %GFP_ATOMIC don't sleep.
1516 * @vma: Pointer to VMA or NULL if not available.
1517 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1519 * This function allocates a page from the kernel page pool and applies
1520 * a NUMA policy associated with the VMA or the current process.
1521 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1522 * mm_struct of the VMA to prevent it from going away. Should be used for
1523 * all allocations for pages that will be mapped into
1524 * user space. Returns NULL when no page can be allocated.
1526 * Should be called with the mm_sem of the vma hold.
1529 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1531 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1532 struct zonelist
*zl
;
1534 cpuset_update_task_memory_state();
1536 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
1539 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1541 return alloc_page_interleave(gfp
, 0, nid
);
1543 zl
= policy_zonelist(gfp
, pol
);
1544 if (unlikely(mpol_needs_cond_ref(pol
))) {
1546 * slow path: ref counted shared policy
1548 struct page
*page
= __alloc_pages_nodemask(gfp
, 0,
1549 zl
, policy_nodemask(gfp
, pol
));
1554 * fast path: default or task policy
1556 return __alloc_pages_nodemask(gfp
, 0, zl
, policy_nodemask(gfp
, pol
));
1560 * alloc_pages_current - Allocate pages.
1563 * %GFP_USER user allocation,
1564 * %GFP_KERNEL kernel allocation,
1565 * %GFP_HIGHMEM highmem allocation,
1566 * %GFP_FS don't call back into a file system.
1567 * %GFP_ATOMIC don't sleep.
1568 * @order: Power of two of allocation size in pages. 0 is a single page.
1570 * Allocate a page from the kernel page pool. When not in
1571 * interrupt context and apply the current process NUMA policy.
1572 * Returns NULL when no page can be allocated.
1574 * Don't call cpuset_update_task_memory_state() unless
1575 * 1) it's ok to take cpuset_sem (can WAIT), and
1576 * 2) allocating for current task (not interrupt).
1578 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1580 struct mempolicy
*pol
= current
->mempolicy
;
1582 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1583 cpuset_update_task_memory_state();
1584 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
1585 pol
= &default_policy
;
1588 * No reference counting needed for current->mempolicy
1589 * nor system default_policy
1591 if (pol
->mode
== MPOL_INTERLEAVE
)
1592 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1593 return __alloc_pages_nodemask(gfp
, order
,
1594 policy_zonelist(gfp
, pol
), policy_nodemask(gfp
, pol
));
1596 EXPORT_SYMBOL(alloc_pages_current
);
1599 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1600 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1601 * with the mems_allowed returned by cpuset_mems_allowed(). This
1602 * keeps mempolicies cpuset relative after its cpuset moves. See
1603 * further kernel/cpuset.c update_nodemask().
1606 /* Slow path of a mempolicy duplicate */
1607 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
1609 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1612 return ERR_PTR(-ENOMEM
);
1613 if (current_cpuset_is_being_rebound()) {
1614 nodemask_t mems
= cpuset_mems_allowed(current
);
1615 mpol_rebind_policy(old
, &mems
);
1618 atomic_set(&new->refcnt
, 1);
1623 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1624 * eliminate the * MPOL_F_* flags that require conditional ref and
1625 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1626 * after return. Use the returned value.
1628 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1629 * policy lookup, even if the policy needs/has extra ref on lookup.
1630 * shmem_readahead needs this.
1632 struct mempolicy
*__mpol_cond_copy(struct mempolicy
*tompol
,
1633 struct mempolicy
*frompol
)
1635 if (!mpol_needs_cond_ref(frompol
))
1639 tompol
->flags
&= ~MPOL_F_SHARED
; /* copy doesn't need unref */
1640 __mpol_put(frompol
);
1644 static int mpol_match_intent(const struct mempolicy
*a
,
1645 const struct mempolicy
*b
)
1647 if (a
->flags
!= b
->flags
)
1649 if (!mpol_store_user_nodemask(a
))
1651 return nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
);
1654 /* Slow path of a mempolicy comparison */
1655 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1659 if (a
->mode
!= b
->mode
)
1661 if (a
->mode
!= MPOL_DEFAULT
&& !mpol_match_intent(a
, b
))
1666 case MPOL_INTERLEAVE
:
1667 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1668 case MPOL_PREFERRED
:
1669 return a
->v
.preferred_node
== b
->v
.preferred_node
&&
1670 a
->flags
== b
->flags
;
1678 * Shared memory backing store policy support.
1680 * Remember policies even when nobody has shared memory mapped.
1681 * The policies are kept in Red-Black tree linked from the inode.
1682 * They are protected by the sp->lock spinlock, which should be held
1683 * for any accesses to the tree.
1686 /* lookup first element intersecting start-end */
1687 /* Caller holds sp->lock */
1688 static struct sp_node
*
1689 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1691 struct rb_node
*n
= sp
->root
.rb_node
;
1694 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1696 if (start
>= p
->end
)
1698 else if (end
<= p
->start
)
1706 struct sp_node
*w
= NULL
;
1707 struct rb_node
*prev
= rb_prev(n
);
1710 w
= rb_entry(prev
, struct sp_node
, nd
);
1711 if (w
->end
<= start
)
1715 return rb_entry(n
, struct sp_node
, nd
);
1718 /* Insert a new shared policy into the list. */
1719 /* Caller holds sp->lock */
1720 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1722 struct rb_node
**p
= &sp
->root
.rb_node
;
1723 struct rb_node
*parent
= NULL
;
1728 nd
= rb_entry(parent
, struct sp_node
, nd
);
1729 if (new->start
< nd
->start
)
1731 else if (new->end
> nd
->end
)
1732 p
= &(*p
)->rb_right
;
1736 rb_link_node(&new->nd
, parent
, p
);
1737 rb_insert_color(&new->nd
, &sp
->root
);
1738 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
1739 new->policy
? new->policy
->mode
: 0);
1742 /* Find shared policy intersecting idx */
1744 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1746 struct mempolicy
*pol
= NULL
;
1749 if (!sp
->root
.rb_node
)
1751 spin_lock(&sp
->lock
);
1752 sn
= sp_lookup(sp
, idx
, idx
+1);
1754 mpol_get(sn
->policy
);
1757 spin_unlock(&sp
->lock
);
1761 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1763 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
1764 rb_erase(&n
->nd
, &sp
->root
);
1765 mpol_put(n
->policy
);
1766 kmem_cache_free(sn_cache
, n
);
1769 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
1770 struct mempolicy
*pol
)
1772 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1779 pol
->flags
|= MPOL_F_SHARED
; /* for unref */
1784 /* Replace a policy range. */
1785 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1786 unsigned long end
, struct sp_node
*new)
1788 struct sp_node
*n
, *new2
= NULL
;
1791 spin_lock(&sp
->lock
);
1792 n
= sp_lookup(sp
, start
, end
);
1793 /* Take care of old policies in the same range. */
1794 while (n
&& n
->start
< end
) {
1795 struct rb_node
*next
= rb_next(&n
->nd
);
1796 if (n
->start
>= start
) {
1802 /* Old policy spanning whole new range. */
1805 spin_unlock(&sp
->lock
);
1806 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1812 sp_insert(sp
, new2
);
1820 n
= rb_entry(next
, struct sp_node
, nd
);
1824 spin_unlock(&sp
->lock
);
1826 mpol_put(new2
->policy
);
1827 kmem_cache_free(sn_cache
, new2
);
1833 * mpol_shared_policy_init - initialize shared policy for inode
1834 * @sp: pointer to inode shared policy
1835 * @mpol: struct mempolicy to install
1837 * Install non-NULL @mpol in inode's shared policy rb-tree.
1838 * On entry, the current task has a reference on a non-NULL @mpol.
1839 * This must be released on exit.
1841 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
1843 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
1844 spin_lock_init(&sp
->lock
);
1847 struct vm_area_struct pvma
;
1848 struct mempolicy
*new;
1850 /* contextualize the tmpfs mount point mempolicy */
1851 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
1852 mpol_put(mpol
); /* drop our ref on sb mpol */
1854 return; /* no valid nodemask intersection */
1856 /* Create pseudo-vma that contains just the policy */
1857 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
1858 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
1859 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
1860 mpol_put(new); /* drop initial ref */
1864 int mpol_set_shared_policy(struct shared_policy
*info
,
1865 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1868 struct sp_node
*new = NULL
;
1869 unsigned long sz
= vma_pages(vma
);
1871 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1873 sz
, npol
? npol
->mode
: -1,
1874 npol
? npol
->flags
: -1,
1875 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1878 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1882 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1884 kmem_cache_free(sn_cache
, new);
1888 /* Free a backing policy store on inode delete. */
1889 void mpol_free_shared_policy(struct shared_policy
*p
)
1892 struct rb_node
*next
;
1894 if (!p
->root
.rb_node
)
1896 spin_lock(&p
->lock
);
1897 next
= rb_first(&p
->root
);
1899 n
= rb_entry(next
, struct sp_node
, nd
);
1900 next
= rb_next(&n
->nd
);
1901 rb_erase(&n
->nd
, &p
->root
);
1902 mpol_put(n
->policy
);
1903 kmem_cache_free(sn_cache
, n
);
1905 spin_unlock(&p
->lock
);
1908 /* assumes fs == KERNEL_DS */
1909 void __init
numa_policy_init(void)
1911 nodemask_t interleave_nodes
;
1912 unsigned long largest
= 0;
1913 int nid
, prefer
= 0;
1915 policy_cache
= kmem_cache_create("numa_policy",
1916 sizeof(struct mempolicy
),
1917 0, SLAB_PANIC
, NULL
);
1919 sn_cache
= kmem_cache_create("shared_policy_node",
1920 sizeof(struct sp_node
),
1921 0, SLAB_PANIC
, NULL
);
1924 * Set interleaving policy for system init. Interleaving is only
1925 * enabled across suitably sized nodes (default is >= 16MB), or
1926 * fall back to the largest node if they're all smaller.
1928 nodes_clear(interleave_nodes
);
1929 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1930 unsigned long total_pages
= node_present_pages(nid
);
1932 /* Preserve the largest node */
1933 if (largest
< total_pages
) {
1934 largest
= total_pages
;
1938 /* Interleave this node? */
1939 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
1940 node_set(nid
, interleave_nodes
);
1943 /* All too small, use the largest */
1944 if (unlikely(nodes_empty(interleave_nodes
)))
1945 node_set(prefer
, interleave_nodes
);
1947 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
1948 printk("numa_policy_init: interleaving failed\n");
1951 /* Reset policy of current process to default */
1952 void numa_default_policy(void)
1954 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
1958 * Parse and format mempolicy from/to strings
1962 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
1963 * Used only for mpol_parse_str() and mpol_to_str()
1965 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
1966 static const char * const policy_types
[] =
1967 { "default", "prefer", "bind", "interleave", "local" };
1972 * mpol_parse_str - parse string to mempolicy
1973 * @str: string containing mempolicy to parse
1974 * @mpol: pointer to struct mempolicy pointer, returned on success.
1975 * @no_context: flag whether to "contextualize" the mempolicy
1978 * <mode>[=<flags>][:<nodelist>]
1980 * if @no_context is true, save the input nodemask in w.user_nodemask in
1981 * the returned mempolicy. This will be used to "clone" the mempolicy in
1982 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
1983 * mount option. Note that if 'static' or 'relative' mode flags were
1984 * specified, the input nodemask will already have been saved. Saving
1985 * it again is redundant, but safe.
1987 * On success, returns 0, else 1
1989 int mpol_parse_str(char *str
, struct mempolicy
**mpol
, int no_context
)
1991 struct mempolicy
*new = NULL
;
1992 unsigned short uninitialized_var(mode
);
1993 unsigned short uninitialized_var(mode_flags
);
1995 char *nodelist
= strchr(str
, ':');
1996 char *flags
= strchr(str
, '=');
2001 /* NUL-terminate mode or flags string */
2003 if (nodelist_parse(nodelist
, nodes
))
2005 if (!nodes_subset(nodes
, node_states
[N_HIGH_MEMORY
]))
2011 *flags
++ = '\0'; /* terminate mode string */
2013 for (i
= 0; i
<= MPOL_LOCAL
; i
++) {
2014 if (!strcmp(str
, policy_types
[i
])) {
2023 case MPOL_PREFERRED
:
2025 * Insist on a nodelist of one node only
2028 char *rest
= nodelist
;
2029 while (isdigit(*rest
))
2035 case MPOL_INTERLEAVE
:
2037 * Default to online nodes with memory if no nodelist
2040 nodes
= node_states
[N_HIGH_MEMORY
];
2045 * Don't allow a nodelist; mpol_new() checks flags
2049 mode
= MPOL_PREFERRED
;
2053 * case MPOL_BIND: mpol_new() enforces non-empty nodemask.
2054 * case MPOL_DEFAULT: mpol_new() enforces empty nodemask, ignores flags.
2061 * Currently, we only support two mutually exclusive
2064 if (!strcmp(flags
, "static"))
2065 mode_flags
|= MPOL_F_STATIC_NODES
;
2066 else if (!strcmp(flags
, "relative"))
2067 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2072 new = mpol_new(mode
, mode_flags
, &nodes
);
2075 else if (no_context
)
2076 new->w
.user_nodemask
= nodes
; /* save for contextualization */
2079 /* Restore string for error message */
2088 #endif /* CONFIG_TMPFS */
2091 * mpol_to_str - format a mempolicy structure for printing
2092 * @buffer: to contain formatted mempolicy string
2093 * @maxlen: length of @buffer
2094 * @pol: pointer to mempolicy to be formatted
2095 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2097 * Convert a mempolicy into a string.
2098 * Returns the number of characters in buffer (if positive)
2099 * or an error (negative)
2101 int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
, int no_context
)
2106 unsigned short mode
;
2107 unsigned short flags
= pol
? pol
->flags
: 0;
2110 * Sanity check: room for longest mode, flag and some nodes
2112 VM_BUG_ON(maxlen
< strlen("interleave") + strlen("relative") + 16);
2114 if (!pol
|| pol
== &default_policy
)
2115 mode
= MPOL_DEFAULT
;
2124 case MPOL_PREFERRED
:
2126 if (flags
& MPOL_F_LOCAL
)
2127 mode
= MPOL_LOCAL
; /* pseudo-policy */
2129 node_set(pol
->v
.preferred_node
, nodes
);
2134 case MPOL_INTERLEAVE
:
2136 nodes
= pol
->w
.user_nodemask
;
2138 nodes
= pol
->v
.nodes
;
2145 l
= strlen(policy_types
[mode
]);
2146 if (buffer
+ maxlen
< p
+ l
+ 1)
2149 strcpy(p
, policy_types
[mode
]);
2152 if (flags
& MPOL_MODE_FLAGS
) {
2153 if (buffer
+ maxlen
< p
+ 2)
2158 * Currently, the only defined flags are mutually exclusive
2160 if (flags
& MPOL_F_STATIC_NODES
)
2161 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2162 else if (flags
& MPOL_F_RELATIVE_NODES
)
2163 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2166 if (!nodes_empty(nodes
)) {
2167 if (buffer
+ maxlen
< p
+ 2)
2170 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
2176 unsigned long pages
;
2178 unsigned long active
;
2179 unsigned long writeback
;
2180 unsigned long mapcount_max
;
2181 unsigned long dirty
;
2182 unsigned long swapcache
;
2183 unsigned long node
[MAX_NUMNODES
];
2186 static void gather_stats(struct page
*page
, void *private, int pte_dirty
)
2188 struct numa_maps
*md
= private;
2189 int count
= page_mapcount(page
);
2192 if (pte_dirty
|| PageDirty(page
))
2195 if (PageSwapCache(page
))
2198 if (PageActive(page
))
2201 if (PageWriteback(page
))
2207 if (count
> md
->mapcount_max
)
2208 md
->mapcount_max
= count
;
2210 md
->node
[page_to_nid(page
)]++;
2213 #ifdef CONFIG_HUGETLB_PAGE
2214 static void check_huge_range(struct vm_area_struct
*vma
,
2215 unsigned long start
, unsigned long end
,
2216 struct numa_maps
*md
)
2220 struct hstate
*h
= hstate_vma(vma
);
2221 unsigned long sz
= huge_page_size(h
);
2223 for (addr
= start
; addr
< end
; addr
+= sz
) {
2224 pte_t
*ptep
= huge_pte_offset(vma
->vm_mm
,
2225 addr
& huge_page_mask(h
));
2235 page
= pte_page(pte
);
2239 gather_stats(page
, md
, pte_dirty(*ptep
));
2243 static inline void check_huge_range(struct vm_area_struct
*vma
,
2244 unsigned long start
, unsigned long end
,
2245 struct numa_maps
*md
)
2251 * Display pages allocated per node and memory policy via /proc.
2253 int show_numa_map(struct seq_file
*m
, void *v
)
2255 struct proc_maps_private
*priv
= m
->private;
2256 struct vm_area_struct
*vma
= v
;
2257 struct numa_maps
*md
;
2258 struct file
*file
= vma
->vm_file
;
2259 struct mm_struct
*mm
= vma
->vm_mm
;
2260 struct mempolicy
*pol
;
2267 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
2271 pol
= get_vma_policy(priv
->task
, vma
, vma
->vm_start
);
2272 mpol_to_str(buffer
, sizeof(buffer
), pol
, 0);
2275 seq_printf(m
, "%08lx %s", vma
->vm_start
, buffer
);
2278 seq_printf(m
, " file=");
2279 seq_path(m
, &file
->f_path
, "\n\t= ");
2280 } else if (vma
->vm_start
<= mm
->brk
&& vma
->vm_end
>= mm
->start_brk
) {
2281 seq_printf(m
, " heap");
2282 } else if (vma
->vm_start
<= mm
->start_stack
&&
2283 vma
->vm_end
>= mm
->start_stack
) {
2284 seq_printf(m
, " stack");
2287 if (is_vm_hugetlb_page(vma
)) {
2288 check_huge_range(vma
, vma
->vm_start
, vma
->vm_end
, md
);
2289 seq_printf(m
, " huge");
2291 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
2292 &node_states
[N_HIGH_MEMORY
], MPOL_MF_STATS
, md
);
2299 seq_printf(m
," anon=%lu",md
->anon
);
2302 seq_printf(m
," dirty=%lu",md
->dirty
);
2304 if (md
->pages
!= md
->anon
&& md
->pages
!= md
->dirty
)
2305 seq_printf(m
, " mapped=%lu", md
->pages
);
2307 if (md
->mapcount_max
> 1)
2308 seq_printf(m
, " mapmax=%lu", md
->mapcount_max
);
2311 seq_printf(m
," swapcache=%lu", md
->swapcache
);
2313 if (md
->active
< md
->pages
&& !is_vm_hugetlb_page(vma
))
2314 seq_printf(m
," active=%lu", md
->active
);
2317 seq_printf(m
," writeback=%lu", md
->writeback
);
2319 for_each_node_state(n
, N_HIGH_MEMORY
)
2321 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
2326 if (m
->count
< m
->size
)
2327 m
->version
= (vma
!= priv
->tail_vma
) ? vma
->vm_start
: 0;