2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/gfp.h>
77 #include <linux/slab.h>
78 #include <linux/string.h>
79 #include <linux/module.h>
80 #include <linux/nsproxy.h>
81 #include <linux/interrupt.h>
82 #include <linux/init.h>
83 #include <linux/compat.h>
84 #include <linux/swap.h>
85 #include <linux/seq_file.h>
86 #include <linux/proc_fs.h>
87 #include <linux/migrate.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
92 #include <asm/tlbflush.h>
93 #include <asm/uaccess.h>
96 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
97 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
98 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
100 static struct kmem_cache
*policy_cache
;
101 static struct kmem_cache
*sn_cache
;
103 /* Highest zone. An specific allocation for a zone below that is not
105 enum zone_type policy_zone
= 0;
107 struct mempolicy default_policy
= {
108 .refcnt
= ATOMIC_INIT(1), /* never free it */
109 .mode
= MPOL_DEFAULT
,
112 static const struct mempolicy_operations
{
113 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
114 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
115 } mpol_ops
[MPOL_MAX
];
117 /* Check that the nodemask contains at least one populated zone */
118 static int is_valid_nodemask(const nodemask_t
*nodemask
)
122 /* Check that there is something useful in this mask */
125 for_each_node_mask(nd
, *nodemask
) {
128 for (k
= 0; k
<= policy_zone
; k
++) {
129 z
= &NODE_DATA(nd
)->node_zones
[k
];
130 if (z
->present_pages
> 0)
138 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
140 return pol
->flags
& (MPOL_F_STATIC_NODES
| MPOL_F_RELATIVE_NODES
);
143 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
144 const nodemask_t
*rel
)
147 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
148 nodes_onto(*ret
, tmp
, *rel
);
151 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
153 if (nodes_empty(*nodes
))
155 pol
->v
.nodes
= *nodes
;
159 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
162 pol
->v
.preferred_node
= -1; /* local allocation */
163 else if (nodes_empty(*nodes
))
164 return -EINVAL
; /* no allowed nodes */
166 pol
->v
.preferred_node
= first_node(*nodes
);
170 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
172 if (!is_valid_nodemask(nodes
))
174 pol
->v
.nodes
= *nodes
;
178 /* Create a new policy */
179 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
182 struct mempolicy
*policy
;
183 nodemask_t cpuset_context_nmask
;
186 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
187 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : -1);
189 if (mode
== MPOL_DEFAULT
) {
190 if (nodes
&& !nodes_empty(*nodes
))
191 return ERR_PTR(-EINVAL
);
197 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
198 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
199 * All other modes require a valid pointer to a non-empty nodemask.
201 if (mode
== MPOL_PREFERRED
) {
202 if (nodes_empty(*nodes
)) {
203 if (((flags
& MPOL_F_STATIC_NODES
) ||
204 (flags
& MPOL_F_RELATIVE_NODES
)))
205 return ERR_PTR(-EINVAL
);
206 nodes
= NULL
; /* flag local alloc */
208 } else if (nodes_empty(*nodes
))
209 return ERR_PTR(-EINVAL
);
210 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
212 return ERR_PTR(-ENOMEM
);
213 atomic_set(&policy
->refcnt
, 1);
215 policy
->flags
= flags
;
219 * cpuset related setup doesn't apply to local allocation
221 cpuset_update_task_memory_state();
222 if (flags
& MPOL_F_RELATIVE_NODES
)
223 mpol_relative_nodemask(&cpuset_context_nmask
, nodes
,
224 &cpuset_current_mems_allowed
);
226 nodes_and(cpuset_context_nmask
, *nodes
,
227 cpuset_current_mems_allowed
);
228 if (mpol_store_user_nodemask(policy
))
229 policy
->w
.user_nodemask
= *nodes
;
231 policy
->w
.cpuset_mems_allowed
=
232 cpuset_mems_allowed(current
);
235 ret
= mpol_ops
[mode
].create(policy
,
236 nodes
? &cpuset_context_nmask
: NULL
);
238 kmem_cache_free(policy_cache
, policy
);
244 /* Slow path of a mpol destructor. */
245 void __mpol_put(struct mempolicy
*p
)
247 if (!atomic_dec_and_test(&p
->refcnt
))
249 p
->mode
= MPOL_DEFAULT
;
250 kmem_cache_free(policy_cache
, p
);
253 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
257 static void mpol_rebind_nodemask(struct mempolicy
*pol
,
258 const nodemask_t
*nodes
)
262 if (pol
->flags
& MPOL_F_STATIC_NODES
)
263 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
264 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
265 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
267 nodes_remap(tmp
, pol
->v
.nodes
, pol
->w
.cpuset_mems_allowed
,
269 pol
->w
.cpuset_mems_allowed
= *nodes
;
273 if (!node_isset(current
->il_next
, tmp
)) {
274 current
->il_next
= next_node(current
->il_next
, tmp
);
275 if (current
->il_next
>= MAX_NUMNODES
)
276 current
->il_next
= first_node(tmp
);
277 if (current
->il_next
>= MAX_NUMNODES
)
278 current
->il_next
= numa_node_id();
282 static void mpol_rebind_preferred(struct mempolicy
*pol
,
283 const nodemask_t
*nodes
)
287 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
288 int node
= first_node(pol
->w
.user_nodemask
);
290 if (node_isset(node
, *nodes
))
291 pol
->v
.preferred_node
= node
;
293 pol
->v
.preferred_node
= -1;
294 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
295 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
296 pol
->v
.preferred_node
= first_node(tmp
);
297 } else if (pol
->v
.preferred_node
!= -1) {
298 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
299 pol
->w
.cpuset_mems_allowed
,
301 pol
->w
.cpuset_mems_allowed
= *nodes
;
305 /* Migrate a policy to a different set of nodes */
306 static void mpol_rebind_policy(struct mempolicy
*pol
,
307 const nodemask_t
*newmask
)
311 if (!mpol_store_user_nodemask(pol
) &&
312 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
314 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
318 * Wrapper for mpol_rebind_policy() that just requires task
319 * pointer, and updates task mempolicy.
322 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
324 mpol_rebind_policy(tsk
->mempolicy
, new);
328 * Rebind each vma in mm to new nodemask.
330 * Call holding a reference to mm. Takes mm->mmap_sem during call.
333 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
335 struct vm_area_struct
*vma
;
337 down_write(&mm
->mmap_sem
);
338 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
339 mpol_rebind_policy(vma
->vm_policy
, new);
340 up_write(&mm
->mmap_sem
);
343 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
345 .rebind
= mpol_rebind_default
,
347 [MPOL_INTERLEAVE
] = {
348 .create
= mpol_new_interleave
,
349 .rebind
= mpol_rebind_nodemask
,
352 .create
= mpol_new_preferred
,
353 .rebind
= mpol_rebind_preferred
,
356 .create
= mpol_new_bind
,
357 .rebind
= mpol_rebind_nodemask
,
361 static void gather_stats(struct page
*, void *, int pte_dirty
);
362 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
363 unsigned long flags
);
365 /* Scan through pages checking if pages follow certain conditions. */
366 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
367 unsigned long addr
, unsigned long end
,
368 const nodemask_t
*nodes
, unsigned long flags
,
375 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
380 if (!pte_present(*pte
))
382 page
= vm_normal_page(vma
, addr
, *pte
);
386 * The check for PageReserved here is important to avoid
387 * handling zero pages and other pages that may have been
388 * marked special by the system.
390 * If the PageReserved would not be checked here then f.e.
391 * the location of the zero page could have an influence
392 * on MPOL_MF_STRICT, zero pages would be counted for
393 * the per node stats, and there would be useless attempts
394 * to put zero pages on the migration list.
396 if (PageReserved(page
))
398 nid
= page_to_nid(page
);
399 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
402 if (flags
& MPOL_MF_STATS
)
403 gather_stats(page
, private, pte_dirty(*pte
));
404 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
405 migrate_page_add(page
, private, flags
);
408 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
409 pte_unmap_unlock(orig_pte
, ptl
);
413 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
414 unsigned long addr
, unsigned long end
,
415 const nodemask_t
*nodes
, unsigned long flags
,
421 pmd
= pmd_offset(pud
, addr
);
423 next
= pmd_addr_end(addr
, end
);
424 if (pmd_none_or_clear_bad(pmd
))
426 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
429 } while (pmd
++, addr
= next
, addr
!= end
);
433 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
434 unsigned long addr
, unsigned long end
,
435 const nodemask_t
*nodes
, unsigned long flags
,
441 pud
= pud_offset(pgd
, addr
);
443 next
= pud_addr_end(addr
, end
);
444 if (pud_none_or_clear_bad(pud
))
446 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
449 } while (pud
++, addr
= next
, addr
!= end
);
453 static inline int check_pgd_range(struct vm_area_struct
*vma
,
454 unsigned long addr
, unsigned long end
,
455 const nodemask_t
*nodes
, unsigned long flags
,
461 pgd
= pgd_offset(vma
->vm_mm
, addr
);
463 next
= pgd_addr_end(addr
, end
);
464 if (pgd_none_or_clear_bad(pgd
))
466 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
469 } while (pgd
++, addr
= next
, addr
!= end
);
474 * Check if all pages in a range are on a set of nodes.
475 * If pagelist != NULL then isolate pages from the LRU and
476 * put them on the pagelist.
478 static struct vm_area_struct
*
479 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
480 const nodemask_t
*nodes
, unsigned long flags
, void *private)
483 struct vm_area_struct
*first
, *vma
, *prev
;
485 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
487 err
= migrate_prep();
492 first
= find_vma(mm
, start
);
494 return ERR_PTR(-EFAULT
);
496 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
497 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
498 if (!vma
->vm_next
&& vma
->vm_end
< end
)
499 return ERR_PTR(-EFAULT
);
500 if (prev
&& prev
->vm_end
< vma
->vm_start
)
501 return ERR_PTR(-EFAULT
);
503 if (!is_vm_hugetlb_page(vma
) &&
504 ((flags
& MPOL_MF_STRICT
) ||
505 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
506 vma_migratable(vma
)))) {
507 unsigned long endvma
= vma
->vm_end
;
511 if (vma
->vm_start
> start
)
512 start
= vma
->vm_start
;
513 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
516 first
= ERR_PTR(err
);
525 /* Apply policy to a single VMA */
526 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
529 struct mempolicy
*old
= vma
->vm_policy
;
531 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
532 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
533 vma
->vm_ops
, vma
->vm_file
,
534 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
536 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
537 err
= vma
->vm_ops
->set_policy(vma
, new);
540 vma
->vm_policy
= new;
546 /* Step 2: apply policy to a range and do splits. */
547 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
548 unsigned long end
, struct mempolicy
*new)
550 struct vm_area_struct
*next
;
554 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
556 if (vma
->vm_start
< start
)
557 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
558 if (!err
&& vma
->vm_end
> end
)
559 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
561 err
= policy_vma(vma
, new);
569 * Update task->flags PF_MEMPOLICY bit: set iff non-default
570 * mempolicy. Allows more rapid checking of this (combined perhaps
571 * with other PF_* flag bits) on memory allocation hot code paths.
573 * If called from outside this file, the task 'p' should -only- be
574 * a newly forked child not yet visible on the task list, because
575 * manipulating the task flags of a visible task is not safe.
577 * The above limitation is why this routine has the funny name
578 * mpol_fix_fork_child_flag().
580 * It is also safe to call this with a task pointer of current,
581 * which the static wrapper mpol_set_task_struct_flag() does,
582 * for use within this file.
585 void mpol_fix_fork_child_flag(struct task_struct
*p
)
588 p
->flags
|= PF_MEMPOLICY
;
590 p
->flags
&= ~PF_MEMPOLICY
;
593 static void mpol_set_task_struct_flag(void)
595 mpol_fix_fork_child_flag(current
);
598 /* Set the process memory policy */
599 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
602 struct mempolicy
*new;
603 struct mm_struct
*mm
= current
->mm
;
605 new = mpol_new(mode
, flags
, nodes
);
610 * prevent changing our mempolicy while show_numa_maps()
612 * Note: do_set_mempolicy() can be called at init time
616 down_write(&mm
->mmap_sem
);
617 mpol_put(current
->mempolicy
);
618 current
->mempolicy
= new;
619 mpol_set_task_struct_flag();
620 if (new && new->mode
== MPOL_INTERLEAVE
&&
621 nodes_weight(new->v
.nodes
))
622 current
->il_next
= first_node(new->v
.nodes
);
624 up_write(&mm
->mmap_sem
);
629 /* Fill a zone bitmap for a policy */
630 static void get_zonemask(struct mempolicy
*p
, nodemask_t
*nodes
)
638 case MPOL_INTERLEAVE
:
642 /* or use current node instead of memory_map? */
643 if (p
->v
.preferred_node
< 0)
644 *nodes
= node_states
[N_HIGH_MEMORY
];
646 node_set(p
->v
.preferred_node
, *nodes
);
653 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
658 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
660 err
= page_to_nid(p
);
666 /* Retrieve NUMA policy */
667 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
668 unsigned long addr
, unsigned long flags
)
671 struct mm_struct
*mm
= current
->mm
;
672 struct vm_area_struct
*vma
= NULL
;
673 struct mempolicy
*pol
= current
->mempolicy
;
675 cpuset_update_task_memory_state();
677 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
680 if (flags
& MPOL_F_MEMS_ALLOWED
) {
681 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
683 *policy
= 0; /* just so it's initialized */
684 *nmask
= cpuset_current_mems_allowed
;
688 if (flags
& MPOL_F_ADDR
) {
689 down_read(&mm
->mmap_sem
);
690 vma
= find_vma_intersection(mm
, addr
, addr
+1);
692 up_read(&mm
->mmap_sem
);
695 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
696 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
698 pol
= vma
->vm_policy
;
703 pol
= &default_policy
;
705 if (flags
& MPOL_F_NODE
) {
706 if (flags
& MPOL_F_ADDR
) {
707 err
= lookup_node(mm
, addr
);
711 } else if (pol
== current
->mempolicy
&&
712 pol
->mode
== MPOL_INTERLEAVE
) {
713 *policy
= current
->il_next
;
719 *policy
= pol
->mode
| pol
->flags
;
722 up_read(¤t
->mm
->mmap_sem
);
728 get_zonemask(pol
, nmask
);
733 up_read(¤t
->mm
->mmap_sem
);
737 #ifdef CONFIG_MIGRATION
741 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
745 * Avoid migrating a page that is shared with others.
747 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1)
748 isolate_lru_page(page
, pagelist
);
751 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
753 return alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
757 * Migrate pages from one node to a target node.
758 * Returns error or the number of pages not migrated.
760 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
768 node_set(source
, nmask
);
770 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nmask
,
771 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
773 if (!list_empty(&pagelist
))
774 err
= migrate_pages(&pagelist
, new_node_page
, dest
);
780 * Move pages between the two nodesets so as to preserve the physical
781 * layout as much as possible.
783 * Returns the number of page that could not be moved.
785 int do_migrate_pages(struct mm_struct
*mm
,
786 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
793 down_read(&mm
->mmap_sem
);
795 err
= migrate_vmas(mm
, from_nodes
, to_nodes
, flags
);
800 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
801 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
802 * bit in 'tmp', and return that <source, dest> pair for migration.
803 * The pair of nodemasks 'to' and 'from' define the map.
805 * If no pair of bits is found that way, fallback to picking some
806 * pair of 'source' and 'dest' bits that are not the same. If the
807 * 'source' and 'dest' bits are the same, this represents a node
808 * that will be migrating to itself, so no pages need move.
810 * If no bits are left in 'tmp', or if all remaining bits left
811 * in 'tmp' correspond to the same bit in 'to', return false
812 * (nothing left to migrate).
814 * This lets us pick a pair of nodes to migrate between, such that
815 * if possible the dest node is not already occupied by some other
816 * source node, minimizing the risk of overloading the memory on a
817 * node that would happen if we migrated incoming memory to a node
818 * before migrating outgoing memory source that same node.
820 * A single scan of tmp is sufficient. As we go, we remember the
821 * most recent <s, d> pair that moved (s != d). If we find a pair
822 * that not only moved, but what's better, moved to an empty slot
823 * (d is not set in tmp), then we break out then, with that pair.
824 * Otherwise when we finish scannng from_tmp, we at least have the
825 * most recent <s, d> pair that moved. If we get all the way through
826 * the scan of tmp without finding any node that moved, much less
827 * moved to an empty node, then there is nothing left worth migrating.
831 while (!nodes_empty(tmp
)) {
836 for_each_node_mask(s
, tmp
) {
837 d
= node_remap(s
, *from_nodes
, *to_nodes
);
841 source
= s
; /* Node moved. Memorize */
844 /* dest not in remaining from nodes? */
845 if (!node_isset(dest
, tmp
))
851 node_clear(source
, tmp
);
852 err
= migrate_to_node(mm
, source
, dest
, flags
);
859 up_read(&mm
->mmap_sem
);
867 * Allocate a new page for page migration based on vma policy.
868 * Start assuming that page is mapped by vma pointed to by @private.
869 * Search forward from there, if not. N.B., this assumes that the
870 * list of pages handed to migrate_pages()--which is how we get here--
871 * is in virtual address order.
873 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
875 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
876 unsigned long uninitialized_var(address
);
879 address
= page_address_in_vma(page
, vma
);
880 if (address
!= -EFAULT
)
886 * if !vma, alloc_page_vma() will use task or system default policy
888 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
892 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
897 int do_migrate_pages(struct mm_struct
*mm
,
898 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
903 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
909 static long do_mbind(unsigned long start
, unsigned long len
,
910 unsigned short mode
, unsigned short mode_flags
,
911 nodemask_t
*nmask
, unsigned long flags
)
913 struct vm_area_struct
*vma
;
914 struct mm_struct
*mm
= current
->mm
;
915 struct mempolicy
*new;
920 if (flags
& ~(unsigned long)(MPOL_MF_STRICT
|
921 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
923 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
926 if (start
& ~PAGE_MASK
)
929 if (mode
== MPOL_DEFAULT
)
930 flags
&= ~MPOL_MF_STRICT
;
932 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
940 new = mpol_new(mode
, mode_flags
, nmask
);
945 * If we are using the default policy then operation
946 * on discontinuous address spaces is okay after all
949 flags
|= MPOL_MF_DISCONTIG_OK
;
951 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
952 start
, start
+ len
, mode
, mode_flags
,
953 nmask
? nodes_addr(*nmask
)[0] : -1);
955 down_write(&mm
->mmap_sem
);
956 vma
= check_range(mm
, start
, end
, nmask
,
957 flags
| MPOL_MF_INVERT
, &pagelist
);
963 err
= mbind_range(vma
, start
, end
, new);
965 if (!list_empty(&pagelist
))
966 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
969 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
973 up_write(&mm
->mmap_sem
);
979 * User space interface with variable sized bitmaps for nodelists.
982 /* Copy a node mask from user space. */
983 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
984 unsigned long maxnode
)
987 unsigned long nlongs
;
988 unsigned long endmask
;
992 if (maxnode
== 0 || !nmask
)
994 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
997 nlongs
= BITS_TO_LONGS(maxnode
);
998 if ((maxnode
% BITS_PER_LONG
) == 0)
1001 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1003 /* When the user specified more nodes than supported just check
1004 if the non supported part is all zero. */
1005 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1006 if (nlongs
> PAGE_SIZE
/sizeof(long))
1008 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1010 if (get_user(t
, nmask
+ k
))
1012 if (k
== nlongs
- 1) {
1018 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1022 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1024 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1028 /* Copy a kernel node mask to user space */
1029 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1032 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1033 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1035 if (copy
> nbytes
) {
1036 if (copy
> PAGE_SIZE
)
1038 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1042 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1045 asmlinkage
long sys_mbind(unsigned long start
, unsigned long len
,
1047 unsigned long __user
*nmask
, unsigned long maxnode
,
1052 unsigned short mode_flags
;
1054 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1055 mode
&= ~MPOL_MODE_FLAGS
;
1056 if (mode
>= MPOL_MAX
)
1058 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1059 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1061 err
= get_nodes(&nodes
, nmask
, maxnode
);
1064 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1067 /* Set the process memory policy */
1068 asmlinkage
long sys_set_mempolicy(int mode
, unsigned long __user
*nmask
,
1069 unsigned long maxnode
)
1073 unsigned short flags
;
1075 flags
= mode
& MPOL_MODE_FLAGS
;
1076 mode
&= ~MPOL_MODE_FLAGS
;
1077 if ((unsigned int)mode
>= MPOL_MAX
)
1079 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1081 err
= get_nodes(&nodes
, nmask
, maxnode
);
1084 return do_set_mempolicy(mode
, flags
, &nodes
);
1087 asmlinkage
long sys_migrate_pages(pid_t pid
, unsigned long maxnode
,
1088 const unsigned long __user
*old_nodes
,
1089 const unsigned long __user
*new_nodes
)
1091 struct mm_struct
*mm
;
1092 struct task_struct
*task
;
1095 nodemask_t task_nodes
;
1098 err
= get_nodes(&old
, old_nodes
, maxnode
);
1102 err
= get_nodes(&new, new_nodes
, maxnode
);
1106 /* Find the mm_struct */
1107 read_lock(&tasklist_lock
);
1108 task
= pid
? find_task_by_vpid(pid
) : current
;
1110 read_unlock(&tasklist_lock
);
1113 mm
= get_task_mm(task
);
1114 read_unlock(&tasklist_lock
);
1120 * Check if this process has the right to modify the specified
1121 * process. The right exists if the process has administrative
1122 * capabilities, superuser privileges or the same
1123 * userid as the target process.
1125 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
1126 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
1127 !capable(CAP_SYS_NICE
)) {
1132 task_nodes
= cpuset_mems_allowed(task
);
1133 /* Is the user allowed to access the target nodes? */
1134 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1139 if (!nodes_subset(new, node_states
[N_HIGH_MEMORY
])) {
1144 err
= security_task_movememory(task
);
1148 err
= do_migrate_pages(mm
, &old
, &new,
1149 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1156 /* Retrieve NUMA policy */
1157 asmlinkage
long sys_get_mempolicy(int __user
*policy
,
1158 unsigned long __user
*nmask
,
1159 unsigned long maxnode
,
1160 unsigned long addr
, unsigned long flags
)
1163 int uninitialized_var(pval
);
1166 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1169 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1174 if (policy
&& put_user(pval
, policy
))
1178 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1183 #ifdef CONFIG_COMPAT
1185 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1186 compat_ulong_t __user
*nmask
,
1187 compat_ulong_t maxnode
,
1188 compat_ulong_t addr
, compat_ulong_t flags
)
1191 unsigned long __user
*nm
= NULL
;
1192 unsigned long nr_bits
, alloc_size
;
1193 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1195 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1196 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1199 nm
= compat_alloc_user_space(alloc_size
);
1201 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1203 if (!err
&& nmask
) {
1204 err
= copy_from_user(bm
, nm
, alloc_size
);
1205 /* ensure entire bitmap is zeroed */
1206 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1207 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1213 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1214 compat_ulong_t maxnode
)
1217 unsigned long __user
*nm
= NULL
;
1218 unsigned long nr_bits
, alloc_size
;
1219 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1221 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1222 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1225 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1226 nm
= compat_alloc_user_space(alloc_size
);
1227 err
|= copy_to_user(nm
, bm
, alloc_size
);
1233 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1236 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1237 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1238 compat_ulong_t maxnode
, compat_ulong_t flags
)
1241 unsigned long __user
*nm
= NULL
;
1242 unsigned long nr_bits
, alloc_size
;
1245 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1246 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1249 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1250 nm
= compat_alloc_user_space(alloc_size
);
1251 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1257 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1263 * get_vma_policy(@task, @vma, @addr)
1264 * @task - task for fallback if vma policy == default
1265 * @vma - virtual memory area whose policy is sought
1266 * @addr - address in @vma for shared policy lookup
1268 * Returns effective policy for a VMA at specified address.
1269 * Falls back to @task or system default policy, as necessary.
1270 * Current or other task's task mempolicy and non-shared vma policies
1271 * are protected by the task's mmap_sem, which must be held for read by
1273 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1274 * count--added by the get_policy() vm_op, as appropriate--to protect against
1275 * freeing by another task. It is the caller's responsibility to free the
1276 * extra reference for shared policies.
1278 static struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1279 struct vm_area_struct
*vma
, unsigned long addr
)
1281 struct mempolicy
*pol
= task
->mempolicy
;
1284 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1285 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1289 } else if (vma
->vm_policy
&&
1290 vma
->vm_policy
->mode
!= MPOL_DEFAULT
)
1291 pol
= vma
->vm_policy
;
1294 pol
= &default_policy
;
1299 * Return a nodemask representing a mempolicy for filtering nodes for
1302 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1304 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1305 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1306 gfp_zone(gfp
) >= policy_zone
&&
1307 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1308 return &policy
->v
.nodes
;
1313 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1314 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
)
1318 switch (policy
->mode
) {
1319 case MPOL_PREFERRED
:
1320 nd
= policy
->v
.preferred_node
;
1322 nd
= numa_node_id();
1326 * Normally, MPOL_BIND allocations are node-local within the
1327 * allowed nodemask. However, if __GFP_THISNODE is set and the
1328 * current node is part of the mask, we use the zonelist for
1329 * the first node in the mask instead.
1331 nd
= numa_node_id();
1332 if (unlikely(gfp
& __GFP_THISNODE
) &&
1333 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1334 nd
= first_node(policy
->v
.nodes
);
1336 case MPOL_INTERLEAVE
: /* should not happen */
1338 nd
= numa_node_id();
1344 return node_zonelist(nd
, gfp
);
1347 /* Do dynamic interleaving for a process */
1348 static unsigned interleave_nodes(struct mempolicy
*policy
)
1351 struct task_struct
*me
= current
;
1354 next
= next_node(nid
, policy
->v
.nodes
);
1355 if (next
>= MAX_NUMNODES
)
1356 next
= first_node(policy
->v
.nodes
);
1357 if (next
< MAX_NUMNODES
)
1363 * Depending on the memory policy provide a node from which to allocate the
1365 * @policy must be protected by freeing by the caller. If @policy is
1366 * the current task's mempolicy, this protection is implicit, as only the
1367 * task can change it's policy. The system default policy requires no
1370 unsigned slab_node(struct mempolicy
*policy
)
1372 unsigned short pol
= policy
? policy
->mode
: MPOL_DEFAULT
;
1375 case MPOL_INTERLEAVE
:
1376 return interleave_nodes(policy
);
1380 * Follow bind policy behavior and start allocation at the
1383 struct zonelist
*zonelist
;
1385 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1386 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1387 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1393 case MPOL_PREFERRED
:
1394 if (policy
->v
.preferred_node
>= 0)
1395 return policy
->v
.preferred_node
;
1399 return numa_node_id();
1403 /* Do static interleaving for a VMA with known offset. */
1404 static unsigned offset_il_node(struct mempolicy
*pol
,
1405 struct vm_area_struct
*vma
, unsigned long off
)
1407 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1413 return numa_node_id();
1414 target
= (unsigned int)off
% nnodes
;
1417 nid
= next_node(nid
, pol
->v
.nodes
);
1419 } while (c
<= target
);
1423 /* Determine a node number for interleave */
1424 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1425 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1431 * for small pages, there is no difference between
1432 * shift and PAGE_SHIFT, so the bit-shift is safe.
1433 * for huge pages, since vm_pgoff is in units of small
1434 * pages, we need to shift off the always 0 bits to get
1437 BUG_ON(shift
< PAGE_SHIFT
);
1438 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1439 off
+= (addr
- vma
->vm_start
) >> shift
;
1440 return offset_il_node(pol
, vma
, off
);
1442 return interleave_nodes(pol
);
1445 #ifdef CONFIG_HUGETLBFS
1447 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1448 * @vma = virtual memory area whose policy is sought
1449 * @addr = address in @vma for shared policy lookup and interleave policy
1450 * @gfp_flags = for requested zone
1451 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1452 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1454 * Returns a zonelist suitable for a huge page allocation and a pointer
1455 * to the struct mempolicy for conditional unref after allocation.
1456 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1457 * @nodemask for filtering the zonelist.
1459 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1460 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1461 nodemask_t
**nodemask
)
1463 struct zonelist
*zl
;
1465 *mpol
= get_vma_policy(current
, vma
, addr
);
1466 *nodemask
= NULL
; /* assume !MPOL_BIND */
1468 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1469 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1470 HPAGE_SHIFT
), gfp_flags
);
1472 zl
= policy_zonelist(gfp_flags
, *mpol
);
1473 if ((*mpol
)->mode
== MPOL_BIND
)
1474 *nodemask
= &(*mpol
)->v
.nodes
;
1480 /* Allocate a page in interleaved policy.
1481 Own path because it needs to do special accounting. */
1482 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1485 struct zonelist
*zl
;
1488 zl
= node_zonelist(nid
, gfp
);
1489 page
= __alloc_pages(gfp
, order
, zl
);
1490 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1491 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1496 * alloc_page_vma - Allocate a page for a VMA.
1499 * %GFP_USER user allocation.
1500 * %GFP_KERNEL kernel allocations,
1501 * %GFP_HIGHMEM highmem/user allocations,
1502 * %GFP_FS allocation should not call back into a file system.
1503 * %GFP_ATOMIC don't sleep.
1505 * @vma: Pointer to VMA or NULL if not available.
1506 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1508 * This function allocates a page from the kernel page pool and applies
1509 * a NUMA policy associated with the VMA or the current process.
1510 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1511 * mm_struct of the VMA to prevent it from going away. Should be used for
1512 * all allocations for pages that will be mapped into
1513 * user space. Returns NULL when no page can be allocated.
1515 * Should be called with the mm_sem of the vma hold.
1518 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1520 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1521 struct zonelist
*zl
;
1523 cpuset_update_task_memory_state();
1525 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
1528 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1530 return alloc_page_interleave(gfp
, 0, nid
);
1532 zl
= policy_zonelist(gfp
, pol
);
1533 if (unlikely(mpol_needs_cond_ref(pol
))) {
1535 * slow path: ref counted shared policy
1537 struct page
*page
= __alloc_pages_nodemask(gfp
, 0,
1538 zl
, policy_nodemask(gfp
, pol
));
1543 * fast path: default or task policy
1545 return __alloc_pages_nodemask(gfp
, 0, zl
, policy_nodemask(gfp
, pol
));
1549 * alloc_pages_current - Allocate pages.
1552 * %GFP_USER user allocation,
1553 * %GFP_KERNEL kernel allocation,
1554 * %GFP_HIGHMEM highmem allocation,
1555 * %GFP_FS don't call back into a file system.
1556 * %GFP_ATOMIC don't sleep.
1557 * @order: Power of two of allocation size in pages. 0 is a single page.
1559 * Allocate a page from the kernel page pool. When not in
1560 * interrupt context and apply the current process NUMA policy.
1561 * Returns NULL when no page can be allocated.
1563 * Don't call cpuset_update_task_memory_state() unless
1564 * 1) it's ok to take cpuset_sem (can WAIT), and
1565 * 2) allocating for current task (not interrupt).
1567 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1569 struct mempolicy
*pol
= current
->mempolicy
;
1571 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1572 cpuset_update_task_memory_state();
1573 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
1574 pol
= &default_policy
;
1577 * No reference counting needed for current->mempolicy
1578 * nor system default_policy
1580 if (pol
->mode
== MPOL_INTERLEAVE
)
1581 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1582 return __alloc_pages_nodemask(gfp
, order
,
1583 policy_zonelist(gfp
, pol
), policy_nodemask(gfp
, pol
));
1585 EXPORT_SYMBOL(alloc_pages_current
);
1588 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1589 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1590 * with the mems_allowed returned by cpuset_mems_allowed(). This
1591 * keeps mempolicies cpuset relative after its cpuset moves. See
1592 * further kernel/cpuset.c update_nodemask().
1595 /* Slow path of a mempolicy duplicate */
1596 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
1598 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1601 return ERR_PTR(-ENOMEM
);
1602 if (current_cpuset_is_being_rebound()) {
1603 nodemask_t mems
= cpuset_mems_allowed(current
);
1604 mpol_rebind_policy(old
, &mems
);
1607 atomic_set(&new->refcnt
, 1);
1612 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1613 * eliminate the * MPOL_F_* flags that require conditional ref and
1614 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1615 * after return. Use the returned value.
1617 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1618 * policy lookup, even if the policy needs/has extra ref on lookup.
1619 * shmem_readahead needs this.
1621 struct mempolicy
*__mpol_cond_copy(struct mempolicy
*tompol
,
1622 struct mempolicy
*frompol
)
1624 if (!mpol_needs_cond_ref(frompol
))
1628 tompol
->flags
&= ~MPOL_F_SHARED
; /* copy doesn't need unref */
1629 __mpol_put(frompol
);
1633 static int mpol_match_intent(const struct mempolicy
*a
,
1634 const struct mempolicy
*b
)
1636 if (a
->flags
!= b
->flags
)
1638 if (!mpol_store_user_nodemask(a
))
1640 return nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
);
1643 /* Slow path of a mempolicy comparison */
1644 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1648 if (a
->mode
!= b
->mode
)
1650 if (a
->mode
!= MPOL_DEFAULT
&& !mpol_match_intent(a
, b
))
1657 case MPOL_INTERLEAVE
:
1658 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1659 case MPOL_PREFERRED
:
1660 return a
->v
.preferred_node
== b
->v
.preferred_node
;
1668 * Shared memory backing store policy support.
1670 * Remember policies even when nobody has shared memory mapped.
1671 * The policies are kept in Red-Black tree linked from the inode.
1672 * They are protected by the sp->lock spinlock, which should be held
1673 * for any accesses to the tree.
1676 /* lookup first element intersecting start-end */
1677 /* Caller holds sp->lock */
1678 static struct sp_node
*
1679 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1681 struct rb_node
*n
= sp
->root
.rb_node
;
1684 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1686 if (start
>= p
->end
)
1688 else if (end
<= p
->start
)
1696 struct sp_node
*w
= NULL
;
1697 struct rb_node
*prev
= rb_prev(n
);
1700 w
= rb_entry(prev
, struct sp_node
, nd
);
1701 if (w
->end
<= start
)
1705 return rb_entry(n
, struct sp_node
, nd
);
1708 /* Insert a new shared policy into the list. */
1709 /* Caller holds sp->lock */
1710 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1712 struct rb_node
**p
= &sp
->root
.rb_node
;
1713 struct rb_node
*parent
= NULL
;
1718 nd
= rb_entry(parent
, struct sp_node
, nd
);
1719 if (new->start
< nd
->start
)
1721 else if (new->end
> nd
->end
)
1722 p
= &(*p
)->rb_right
;
1726 rb_link_node(&new->nd
, parent
, p
);
1727 rb_insert_color(&new->nd
, &sp
->root
);
1728 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
1729 new->policy
? new->policy
->mode
: 0);
1732 /* Find shared policy intersecting idx */
1734 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1736 struct mempolicy
*pol
= NULL
;
1739 if (!sp
->root
.rb_node
)
1741 spin_lock(&sp
->lock
);
1742 sn
= sp_lookup(sp
, idx
, idx
+1);
1744 mpol_get(sn
->policy
);
1747 spin_unlock(&sp
->lock
);
1751 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1753 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
1754 rb_erase(&n
->nd
, &sp
->root
);
1755 mpol_put(n
->policy
);
1756 kmem_cache_free(sn_cache
, n
);
1759 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
1760 struct mempolicy
*pol
)
1762 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1769 pol
->flags
|= MPOL_F_SHARED
; /* for unref */
1774 /* Replace a policy range. */
1775 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1776 unsigned long end
, struct sp_node
*new)
1778 struct sp_node
*n
, *new2
= NULL
;
1781 spin_lock(&sp
->lock
);
1782 n
= sp_lookup(sp
, start
, end
);
1783 /* Take care of old policies in the same range. */
1784 while (n
&& n
->start
< end
) {
1785 struct rb_node
*next
= rb_next(&n
->nd
);
1786 if (n
->start
>= start
) {
1792 /* Old policy spanning whole new range. */
1795 spin_unlock(&sp
->lock
);
1796 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1802 sp_insert(sp
, new2
);
1810 n
= rb_entry(next
, struct sp_node
, nd
);
1814 spin_unlock(&sp
->lock
);
1816 mpol_put(new2
->policy
);
1817 kmem_cache_free(sn_cache
, new2
);
1822 void mpol_shared_policy_init(struct shared_policy
*info
, unsigned short policy
,
1823 unsigned short flags
, nodemask_t
*policy_nodes
)
1825 info
->root
= RB_ROOT
;
1826 spin_lock_init(&info
->lock
);
1828 if (policy
!= MPOL_DEFAULT
) {
1829 struct mempolicy
*newpol
;
1831 /* Falls back to MPOL_DEFAULT on any error */
1832 newpol
= mpol_new(policy
, flags
, policy_nodes
);
1833 if (!IS_ERR(newpol
)) {
1834 /* Create pseudo-vma that contains just the policy */
1835 struct vm_area_struct pvma
;
1837 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
1838 /* Policy covers entire file */
1839 pvma
.vm_end
= TASK_SIZE
;
1840 mpol_set_shared_policy(info
, &pvma
, newpol
);
1846 int mpol_set_shared_policy(struct shared_policy
*info
,
1847 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1850 struct sp_node
*new = NULL
;
1851 unsigned long sz
= vma_pages(vma
);
1853 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1855 sz
, npol
? npol
->mode
: -1,
1856 npol
? npol
->flags
: -1,
1857 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1860 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1864 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1866 kmem_cache_free(sn_cache
, new);
1870 /* Free a backing policy store on inode delete. */
1871 void mpol_free_shared_policy(struct shared_policy
*p
)
1874 struct rb_node
*next
;
1876 if (!p
->root
.rb_node
)
1878 spin_lock(&p
->lock
);
1879 next
= rb_first(&p
->root
);
1881 n
= rb_entry(next
, struct sp_node
, nd
);
1882 next
= rb_next(&n
->nd
);
1883 rb_erase(&n
->nd
, &p
->root
);
1884 mpol_put(n
->policy
);
1885 kmem_cache_free(sn_cache
, n
);
1887 spin_unlock(&p
->lock
);
1890 /* assumes fs == KERNEL_DS */
1891 void __init
numa_policy_init(void)
1893 nodemask_t interleave_nodes
;
1894 unsigned long largest
= 0;
1895 int nid
, prefer
= 0;
1897 policy_cache
= kmem_cache_create("numa_policy",
1898 sizeof(struct mempolicy
),
1899 0, SLAB_PANIC
, NULL
);
1901 sn_cache
= kmem_cache_create("shared_policy_node",
1902 sizeof(struct sp_node
),
1903 0, SLAB_PANIC
, NULL
);
1906 * Set interleaving policy for system init. Interleaving is only
1907 * enabled across suitably sized nodes (default is >= 16MB), or
1908 * fall back to the largest node if they're all smaller.
1910 nodes_clear(interleave_nodes
);
1911 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1912 unsigned long total_pages
= node_present_pages(nid
);
1914 /* Preserve the largest node */
1915 if (largest
< total_pages
) {
1916 largest
= total_pages
;
1920 /* Interleave this node? */
1921 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
1922 node_set(nid
, interleave_nodes
);
1925 /* All too small, use the largest */
1926 if (unlikely(nodes_empty(interleave_nodes
)))
1927 node_set(prefer
, interleave_nodes
);
1929 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
1930 printk("numa_policy_init: interleaving failed\n");
1933 /* Reset policy of current process to default */
1934 void numa_default_policy(void)
1936 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
1940 * Display pages allocated per node and memory policy via /proc.
1942 static const char * const policy_types
[] =
1943 { "default", "prefer", "bind", "interleave" };
1946 * Convert a mempolicy into a string.
1947 * Returns the number of characters in buffer (if positive)
1948 * or an error (negative)
1950 static inline int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
1955 unsigned short mode
= pol
? pol
->mode
: MPOL_DEFAULT
;
1956 unsigned short flags
= pol
? pol
->flags
: 0;
1963 case MPOL_PREFERRED
:
1965 node_set(pol
->v
.preferred_node
, nodes
);
1970 case MPOL_INTERLEAVE
:
1971 nodes
= pol
->v
.nodes
;
1979 l
= strlen(policy_types
[mode
]);
1980 if (buffer
+ maxlen
< p
+ l
+ 1)
1983 strcpy(p
, policy_types
[mode
]);
1989 if (buffer
+ maxlen
< p
+ 2)
1993 if (flags
& MPOL_F_STATIC_NODES
)
1994 p
+= sprintf(p
, "%sstatic", need_bar
++ ? "|" : "");
1995 if (flags
& MPOL_F_RELATIVE_NODES
)
1996 p
+= sprintf(p
, "%srelative", need_bar
++ ? "|" : "");
1999 if (!nodes_empty(nodes
)) {
2000 if (buffer
+ maxlen
< p
+ 2)
2003 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
2009 unsigned long pages
;
2011 unsigned long active
;
2012 unsigned long writeback
;
2013 unsigned long mapcount_max
;
2014 unsigned long dirty
;
2015 unsigned long swapcache
;
2016 unsigned long node
[MAX_NUMNODES
];
2019 static void gather_stats(struct page
*page
, void *private, int pte_dirty
)
2021 struct numa_maps
*md
= private;
2022 int count
= page_mapcount(page
);
2025 if (pte_dirty
|| PageDirty(page
))
2028 if (PageSwapCache(page
))
2031 if (PageActive(page
))
2034 if (PageWriteback(page
))
2040 if (count
> md
->mapcount_max
)
2041 md
->mapcount_max
= count
;
2043 md
->node
[page_to_nid(page
)]++;
2046 #ifdef CONFIG_HUGETLB_PAGE
2047 static void check_huge_range(struct vm_area_struct
*vma
,
2048 unsigned long start
, unsigned long end
,
2049 struct numa_maps
*md
)
2054 for (addr
= start
; addr
< end
; addr
+= HPAGE_SIZE
) {
2055 pte_t
*ptep
= huge_pte_offset(vma
->vm_mm
, addr
& HPAGE_MASK
);
2065 page
= pte_page(pte
);
2069 gather_stats(page
, md
, pte_dirty(*ptep
));
2073 static inline void check_huge_range(struct vm_area_struct
*vma
,
2074 unsigned long start
, unsigned long end
,
2075 struct numa_maps
*md
)
2080 int show_numa_map(struct seq_file
*m
, void *v
)
2082 struct proc_maps_private
*priv
= m
->private;
2083 struct vm_area_struct
*vma
= v
;
2084 struct numa_maps
*md
;
2085 struct file
*file
= vma
->vm_file
;
2086 struct mm_struct
*mm
= vma
->vm_mm
;
2087 struct mempolicy
*pol
;
2094 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
2098 pol
= get_vma_policy(priv
->task
, vma
, vma
->vm_start
);
2099 mpol_to_str(buffer
, sizeof(buffer
), pol
);
2102 seq_printf(m
, "%08lx %s", vma
->vm_start
, buffer
);
2105 seq_printf(m
, " file=");
2106 seq_path(m
, &file
->f_path
, "\n\t= ");
2107 } else if (vma
->vm_start
<= mm
->brk
&& vma
->vm_end
>= mm
->start_brk
) {
2108 seq_printf(m
, " heap");
2109 } else if (vma
->vm_start
<= mm
->start_stack
&&
2110 vma
->vm_end
>= mm
->start_stack
) {
2111 seq_printf(m
, " stack");
2114 if (is_vm_hugetlb_page(vma
)) {
2115 check_huge_range(vma
, vma
->vm_start
, vma
->vm_end
, md
);
2116 seq_printf(m
, " huge");
2118 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
2119 &node_states
[N_HIGH_MEMORY
], MPOL_MF_STATS
, md
);
2126 seq_printf(m
," anon=%lu",md
->anon
);
2129 seq_printf(m
," dirty=%lu",md
->dirty
);
2131 if (md
->pages
!= md
->anon
&& md
->pages
!= md
->dirty
)
2132 seq_printf(m
, " mapped=%lu", md
->pages
);
2134 if (md
->mapcount_max
> 1)
2135 seq_printf(m
, " mapmax=%lu", md
->mapcount_max
);
2138 seq_printf(m
," swapcache=%lu", md
->swapcache
);
2140 if (md
->active
< md
->pages
&& !is_vm_hugetlb_page(vma
))
2141 seq_printf(m
," active=%lu", md
->active
);
2144 seq_printf(m
," writeback=%lu", md
->writeback
);
2146 for_each_node_state(n
, N_HIGH_MEMORY
)
2148 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
2153 if (m
->count
< m
->size
)
2154 m
->version
= (vma
!= priv
->tail_vma
) ? vma
->vm_start
: 0;