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/ksm.h>
89 #include <linux/rmap.h>
90 #include <linux/security.h>
91 #include <linux/syscalls.h>
92 #include <linux/ctype.h>
93 #include <linux/mm_inline.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
101 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
103 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
105 static struct kmem_cache
*policy_cache
;
106 static struct kmem_cache
*sn_cache
;
108 /* Highest zone. An specific allocation for a zone below that is not
110 enum zone_type policy_zone
= 0;
113 * run-time system-wide default policy => local allocation
115 struct mempolicy default_policy
= {
116 .refcnt
= ATOMIC_INIT(1), /* never free it */
117 .mode
= MPOL_PREFERRED
,
118 .flags
= MPOL_F_LOCAL
,
121 static const struct mempolicy_operations
{
122 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
123 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
124 } mpol_ops
[MPOL_MAX
];
126 /* Check that the nodemask contains at least one populated zone */
127 static int is_valid_nodemask(const nodemask_t
*nodemask
)
131 /* Check that there is something useful in this mask */
134 for_each_node_mask(nd
, *nodemask
) {
137 for (k
= 0; k
<= policy_zone
; k
++) {
138 z
= &NODE_DATA(nd
)->node_zones
[k
];
139 if (z
->present_pages
> 0)
147 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
149 return pol
->flags
& (MPOL_F_STATIC_NODES
| MPOL_F_RELATIVE_NODES
);
152 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
153 const nodemask_t
*rel
)
156 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
157 nodes_onto(*ret
, tmp
, *rel
);
160 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
162 if (nodes_empty(*nodes
))
164 pol
->v
.nodes
= *nodes
;
168 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
171 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
172 else if (nodes_empty(*nodes
))
173 return -EINVAL
; /* no allowed nodes */
175 pol
->v
.preferred_node
= first_node(*nodes
);
179 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
181 if (!is_valid_nodemask(nodes
))
183 pol
->v
.nodes
= *nodes
;
188 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
189 * any, for the new policy. mpol_new() has already validated the nodes
190 * parameter with respect to the policy mode and flags. But, we need to
191 * handle an empty nodemask with MPOL_PREFERRED here.
193 * Must be called holding task's alloc_lock to protect task's mems_allowed
194 * and mempolicy. May also be called holding the mmap_semaphore for write.
196 static int mpol_set_nodemask(struct mempolicy
*pol
,
197 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
201 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
204 /* Check N_HIGH_MEMORY */
205 nodes_and(nsc
->mask1
,
206 cpuset_current_mems_allowed
, node_states
[N_HIGH_MEMORY
]);
209 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
210 nodes
= NULL
; /* explicit local allocation */
212 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
213 mpol_relative_nodemask(&nsc
->mask2
, nodes
,&nsc
->mask1
);
215 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
217 if (mpol_store_user_nodemask(pol
))
218 pol
->w
.user_nodemask
= *nodes
;
220 pol
->w
.cpuset_mems_allowed
=
221 cpuset_current_mems_allowed
;
225 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
227 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
232 * This function just creates a new policy, does some check and simple
233 * initialization. You must invoke mpol_set_nodemask() to set nodes.
235 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
238 struct mempolicy
*policy
;
240 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
241 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : -1);
243 if (mode
== MPOL_DEFAULT
) {
244 if (nodes
&& !nodes_empty(*nodes
))
245 return ERR_PTR(-EINVAL
);
246 return NULL
; /* simply delete any existing policy */
251 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
252 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
253 * All other modes require a valid pointer to a non-empty nodemask.
255 if (mode
== MPOL_PREFERRED
) {
256 if (nodes_empty(*nodes
)) {
257 if (((flags
& MPOL_F_STATIC_NODES
) ||
258 (flags
& MPOL_F_RELATIVE_NODES
)))
259 return ERR_PTR(-EINVAL
);
261 } else if (nodes_empty(*nodes
))
262 return ERR_PTR(-EINVAL
);
263 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
265 return ERR_PTR(-ENOMEM
);
266 atomic_set(&policy
->refcnt
, 1);
268 policy
->flags
= flags
;
273 /* Slow path of a mpol destructor. */
274 void __mpol_put(struct mempolicy
*p
)
276 if (!atomic_dec_and_test(&p
->refcnt
))
278 kmem_cache_free(policy_cache
, p
);
281 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
285 static void mpol_rebind_nodemask(struct mempolicy
*pol
,
286 const nodemask_t
*nodes
)
290 if (pol
->flags
& MPOL_F_STATIC_NODES
)
291 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
292 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
293 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
295 nodes_remap(tmp
, pol
->v
.nodes
, pol
->w
.cpuset_mems_allowed
,
297 pol
->w
.cpuset_mems_allowed
= *nodes
;
301 if (!node_isset(current
->il_next
, tmp
)) {
302 current
->il_next
= next_node(current
->il_next
, tmp
);
303 if (current
->il_next
>= MAX_NUMNODES
)
304 current
->il_next
= first_node(tmp
);
305 if (current
->il_next
>= MAX_NUMNODES
)
306 current
->il_next
= numa_node_id();
310 static void mpol_rebind_preferred(struct mempolicy
*pol
,
311 const nodemask_t
*nodes
)
315 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
316 int node
= first_node(pol
->w
.user_nodemask
);
318 if (node_isset(node
, *nodes
)) {
319 pol
->v
.preferred_node
= node
;
320 pol
->flags
&= ~MPOL_F_LOCAL
;
322 pol
->flags
|= MPOL_F_LOCAL
;
323 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
324 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
325 pol
->v
.preferred_node
= first_node(tmp
);
326 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
327 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
328 pol
->w
.cpuset_mems_allowed
,
330 pol
->w
.cpuset_mems_allowed
= *nodes
;
334 /* Migrate a policy to a different set of nodes */
335 static void mpol_rebind_policy(struct mempolicy
*pol
,
336 const nodemask_t
*newmask
)
340 if (!mpol_store_user_nodemask(pol
) &&
341 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
343 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
347 * Wrapper for mpol_rebind_policy() that just requires task
348 * pointer, and updates task mempolicy.
350 * Called with task's alloc_lock held.
353 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
355 mpol_rebind_policy(tsk
->mempolicy
, new);
359 * Rebind each vma in mm to new nodemask.
361 * Call holding a reference to mm. Takes mm->mmap_sem during call.
364 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
366 struct vm_area_struct
*vma
;
368 down_write(&mm
->mmap_sem
);
369 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
370 mpol_rebind_policy(vma
->vm_policy
, new);
371 up_write(&mm
->mmap_sem
);
374 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
376 .rebind
= mpol_rebind_default
,
378 [MPOL_INTERLEAVE
] = {
379 .create
= mpol_new_interleave
,
380 .rebind
= mpol_rebind_nodemask
,
383 .create
= mpol_new_preferred
,
384 .rebind
= mpol_rebind_preferred
,
387 .create
= mpol_new_bind
,
388 .rebind
= mpol_rebind_nodemask
,
392 static void gather_stats(struct page
*, void *, int pte_dirty
);
393 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
394 unsigned long flags
);
396 /* Scan through pages checking if pages follow certain conditions. */
397 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
398 unsigned long addr
, unsigned long end
,
399 const nodemask_t
*nodes
, unsigned long flags
,
406 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
411 if (!pte_present(*pte
))
413 page
= vm_normal_page(vma
, addr
, *pte
);
417 * vm_normal_page() filters out zero pages, but there might
418 * still be PageReserved pages to skip, perhaps in a VDSO.
419 * And we cannot move PageKsm pages sensibly or safely yet.
421 if (PageReserved(page
) || PageKsm(page
))
423 nid
= page_to_nid(page
);
424 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
427 if (flags
& MPOL_MF_STATS
)
428 gather_stats(page
, private, pte_dirty(*pte
));
429 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
430 migrate_page_add(page
, private, flags
);
433 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
434 pte_unmap_unlock(orig_pte
, ptl
);
438 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
439 unsigned long addr
, unsigned long end
,
440 const nodemask_t
*nodes
, unsigned long flags
,
446 pmd
= pmd_offset(pud
, addr
);
448 next
= pmd_addr_end(addr
, end
);
449 if (pmd_none_or_clear_bad(pmd
))
451 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
454 } while (pmd
++, addr
= next
, addr
!= end
);
458 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
459 unsigned long addr
, unsigned long end
,
460 const nodemask_t
*nodes
, unsigned long flags
,
466 pud
= pud_offset(pgd
, addr
);
468 next
= pud_addr_end(addr
, end
);
469 if (pud_none_or_clear_bad(pud
))
471 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
474 } while (pud
++, addr
= next
, addr
!= end
);
478 static inline int check_pgd_range(struct vm_area_struct
*vma
,
479 unsigned long addr
, unsigned long end
,
480 const nodemask_t
*nodes
, unsigned long flags
,
486 pgd
= pgd_offset(vma
->vm_mm
, addr
);
488 next
= pgd_addr_end(addr
, end
);
489 if (pgd_none_or_clear_bad(pgd
))
491 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
494 } while (pgd
++, addr
= next
, addr
!= end
);
499 * Check if all pages in a range are on a set of nodes.
500 * If pagelist != NULL then isolate pages from the LRU and
501 * put them on the pagelist.
503 static struct vm_area_struct
*
504 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
505 const nodemask_t
*nodes
, unsigned long flags
, void *private)
508 struct vm_area_struct
*first
, *vma
, *prev
;
511 first
= find_vma(mm
, start
);
513 return ERR_PTR(-EFAULT
);
515 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
516 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
517 if (!vma
->vm_next
&& vma
->vm_end
< end
)
518 return ERR_PTR(-EFAULT
);
519 if (prev
&& prev
->vm_end
< vma
->vm_start
)
520 return ERR_PTR(-EFAULT
);
522 if (!is_vm_hugetlb_page(vma
) &&
523 ((flags
& MPOL_MF_STRICT
) ||
524 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
525 vma_migratable(vma
)))) {
526 unsigned long endvma
= vma
->vm_end
;
530 if (vma
->vm_start
> start
)
531 start
= vma
->vm_start
;
532 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
535 first
= ERR_PTR(err
);
544 /* Apply policy to a single VMA */
545 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
548 struct mempolicy
*old
= vma
->vm_policy
;
550 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
551 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
552 vma
->vm_ops
, vma
->vm_file
,
553 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
555 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
556 err
= vma
->vm_ops
->set_policy(vma
, new);
559 vma
->vm_policy
= new;
565 /* Step 2: apply policy to a range and do splits. */
566 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
567 unsigned long end
, struct mempolicy
*new_pol
)
569 struct vm_area_struct
*next
;
570 struct vm_area_struct
*prev
;
571 struct vm_area_struct
*vma
;
574 unsigned long vmstart
;
577 vma
= find_vma_prev(mm
, start
, &prev
);
578 if (!vma
|| vma
->vm_start
> start
)
581 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
583 vmstart
= max(start
, vma
->vm_start
);
584 vmend
= min(end
, vma
->vm_end
);
586 pgoff
= vma
->vm_pgoff
+ ((start
- vma
->vm_start
) >> PAGE_SHIFT
);
587 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
588 vma
->anon_vma
, vma
->vm_file
, pgoff
, new_pol
);
594 if (vma
->vm_start
!= vmstart
) {
595 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
599 if (vma
->vm_end
!= vmend
) {
600 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
604 err
= policy_vma(vma
, new_pol
);
614 * Update task->flags PF_MEMPOLICY bit: set iff non-default
615 * mempolicy. Allows more rapid checking of this (combined perhaps
616 * with other PF_* flag bits) on memory allocation hot code paths.
618 * If called from outside this file, the task 'p' should -only- be
619 * a newly forked child not yet visible on the task list, because
620 * manipulating the task flags of a visible task is not safe.
622 * The above limitation is why this routine has the funny name
623 * mpol_fix_fork_child_flag().
625 * It is also safe to call this with a task pointer of current,
626 * which the static wrapper mpol_set_task_struct_flag() does,
627 * for use within this file.
630 void mpol_fix_fork_child_flag(struct task_struct
*p
)
633 p
->flags
|= PF_MEMPOLICY
;
635 p
->flags
&= ~PF_MEMPOLICY
;
638 static void mpol_set_task_struct_flag(void)
640 mpol_fix_fork_child_flag(current
);
643 /* Set the process memory policy */
644 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
647 struct mempolicy
*new, *old
;
648 struct mm_struct
*mm
= current
->mm
;
649 NODEMASK_SCRATCH(scratch
);
655 new = mpol_new(mode
, flags
, nodes
);
661 * prevent changing our mempolicy while show_numa_maps()
663 * Note: do_set_mempolicy() can be called at init time
667 down_write(&mm
->mmap_sem
);
669 ret
= mpol_set_nodemask(new, nodes
, scratch
);
671 task_unlock(current
);
673 up_write(&mm
->mmap_sem
);
677 old
= current
->mempolicy
;
678 current
->mempolicy
= new;
679 mpol_set_task_struct_flag();
680 if (new && new->mode
== MPOL_INTERLEAVE
&&
681 nodes_weight(new->v
.nodes
))
682 current
->il_next
= first_node(new->v
.nodes
);
683 task_unlock(current
);
685 up_write(&mm
->mmap_sem
);
690 NODEMASK_SCRATCH_FREE(scratch
);
695 * Return nodemask for policy for get_mempolicy() query
697 * Called with task's alloc_lock held
699 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
702 if (p
== &default_policy
)
708 case MPOL_INTERLEAVE
:
712 if (!(p
->flags
& MPOL_F_LOCAL
))
713 node_set(p
->v
.preferred_node
, *nodes
);
714 /* else return empty node mask for local allocation */
721 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
726 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
728 err
= page_to_nid(p
);
734 /* Retrieve NUMA policy */
735 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
736 unsigned long addr
, unsigned long flags
)
739 struct mm_struct
*mm
= current
->mm
;
740 struct vm_area_struct
*vma
= NULL
;
741 struct mempolicy
*pol
= current
->mempolicy
;
744 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
747 if (flags
& MPOL_F_MEMS_ALLOWED
) {
748 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
750 *policy
= 0; /* just so it's initialized */
752 *nmask
= cpuset_current_mems_allowed
;
753 task_unlock(current
);
757 if (flags
& MPOL_F_ADDR
) {
759 * Do NOT fall back to task policy if the
760 * vma/shared policy at addr is NULL. We
761 * want to return MPOL_DEFAULT in this case.
763 down_read(&mm
->mmap_sem
);
764 vma
= find_vma_intersection(mm
, addr
, addr
+1);
766 up_read(&mm
->mmap_sem
);
769 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
770 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
772 pol
= vma
->vm_policy
;
777 pol
= &default_policy
; /* indicates default behavior */
779 if (flags
& MPOL_F_NODE
) {
780 if (flags
& MPOL_F_ADDR
) {
781 err
= lookup_node(mm
, addr
);
785 } else if (pol
== current
->mempolicy
&&
786 pol
->mode
== MPOL_INTERLEAVE
) {
787 *policy
= current
->il_next
;
793 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
796 * Internal mempolicy flags must be masked off before exposing
797 * the policy to userspace.
799 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
803 up_read(¤t
->mm
->mmap_sem
);
809 if (mpol_store_user_nodemask(pol
)) {
810 *nmask
= pol
->w
.user_nodemask
;
813 get_policy_nodemask(pol
, nmask
);
814 task_unlock(current
);
821 up_read(¤t
->mm
->mmap_sem
);
825 #ifdef CONFIG_MIGRATION
829 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
833 * Avoid migrating a page that is shared with others.
835 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
836 if (!isolate_lru_page(page
)) {
837 list_add_tail(&page
->lru
, pagelist
);
838 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
839 page_is_file_cache(page
));
844 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
846 return alloc_pages_exact_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
850 * Migrate pages from one node to a target node.
851 * Returns error or the number of pages not migrated.
853 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
861 node_set(source
, nmask
);
863 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nmask
,
864 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
866 if (!list_empty(&pagelist
))
867 err
= migrate_pages(&pagelist
, new_node_page
, dest
, 0);
873 * Move pages between the two nodesets so as to preserve the physical
874 * layout as much as possible.
876 * Returns the number of page that could not be moved.
878 int do_migrate_pages(struct mm_struct
*mm
,
879 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
885 err
= migrate_prep();
889 down_read(&mm
->mmap_sem
);
891 err
= migrate_vmas(mm
, from_nodes
, to_nodes
, flags
);
896 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
897 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
898 * bit in 'tmp', and return that <source, dest> pair for migration.
899 * The pair of nodemasks 'to' and 'from' define the map.
901 * If no pair of bits is found that way, fallback to picking some
902 * pair of 'source' and 'dest' bits that are not the same. If the
903 * 'source' and 'dest' bits are the same, this represents a node
904 * that will be migrating to itself, so no pages need move.
906 * If no bits are left in 'tmp', or if all remaining bits left
907 * in 'tmp' correspond to the same bit in 'to', return false
908 * (nothing left to migrate).
910 * This lets us pick a pair of nodes to migrate between, such that
911 * if possible the dest node is not already occupied by some other
912 * source node, minimizing the risk of overloading the memory on a
913 * node that would happen if we migrated incoming memory to a node
914 * before migrating outgoing memory source that same node.
916 * A single scan of tmp is sufficient. As we go, we remember the
917 * most recent <s, d> pair that moved (s != d). If we find a pair
918 * that not only moved, but what's better, moved to an empty slot
919 * (d is not set in tmp), then we break out then, with that pair.
920 * Otherwise when we finish scannng from_tmp, we at least have the
921 * most recent <s, d> pair that moved. If we get all the way through
922 * the scan of tmp without finding any node that moved, much less
923 * moved to an empty node, then there is nothing left worth migrating.
927 while (!nodes_empty(tmp
)) {
932 for_each_node_mask(s
, tmp
) {
933 d
= node_remap(s
, *from_nodes
, *to_nodes
);
937 source
= s
; /* Node moved. Memorize */
940 /* dest not in remaining from nodes? */
941 if (!node_isset(dest
, tmp
))
947 node_clear(source
, tmp
);
948 err
= migrate_to_node(mm
, source
, dest
, flags
);
955 up_read(&mm
->mmap_sem
);
963 * Allocate a new page for page migration based on vma policy.
964 * Start assuming that page is mapped by vma pointed to by @private.
965 * Search forward from there, if not. N.B., this assumes that the
966 * list of pages handed to migrate_pages()--which is how we get here--
967 * is in virtual address order.
969 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
971 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
972 unsigned long uninitialized_var(address
);
975 address
= page_address_in_vma(page
, vma
);
976 if (address
!= -EFAULT
)
982 * if !vma, alloc_page_vma() will use task or system default policy
984 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
988 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
993 int do_migrate_pages(struct mm_struct
*mm
,
994 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
999 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
1005 static long do_mbind(unsigned long start
, unsigned long len
,
1006 unsigned short mode
, unsigned short mode_flags
,
1007 nodemask_t
*nmask
, unsigned long flags
)
1009 struct vm_area_struct
*vma
;
1010 struct mm_struct
*mm
= current
->mm
;
1011 struct mempolicy
*new;
1014 LIST_HEAD(pagelist
);
1016 if (flags
& ~(unsigned long)(MPOL_MF_STRICT
|
1017 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
1019 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1022 if (start
& ~PAGE_MASK
)
1025 if (mode
== MPOL_DEFAULT
)
1026 flags
&= ~MPOL_MF_STRICT
;
1028 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1036 new = mpol_new(mode
, mode_flags
, nmask
);
1038 return PTR_ERR(new);
1041 * If we are using the default policy then operation
1042 * on discontinuous address spaces is okay after all
1045 flags
|= MPOL_MF_DISCONTIG_OK
;
1047 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1048 start
, start
+ len
, mode
, mode_flags
,
1049 nmask
? nodes_addr(*nmask
)[0] : -1);
1051 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1053 err
= migrate_prep();
1058 NODEMASK_SCRATCH(scratch
);
1060 down_write(&mm
->mmap_sem
);
1062 err
= mpol_set_nodemask(new, nmask
, scratch
);
1063 task_unlock(current
);
1065 up_write(&mm
->mmap_sem
);
1068 NODEMASK_SCRATCH_FREE(scratch
);
1073 vma
= check_range(mm
, start
, end
, nmask
,
1074 flags
| MPOL_MF_INVERT
, &pagelist
);
1080 err
= mbind_range(mm
, start
, end
, new);
1082 if (!list_empty(&pagelist
))
1083 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
1084 (unsigned long)vma
, 0);
1086 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
1089 putback_lru_pages(&pagelist
);
1091 up_write(&mm
->mmap_sem
);
1098 * User space interface with variable sized bitmaps for nodelists.
1101 /* Copy a node mask from user space. */
1102 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1103 unsigned long maxnode
)
1106 unsigned long nlongs
;
1107 unsigned long endmask
;
1110 nodes_clear(*nodes
);
1111 if (maxnode
== 0 || !nmask
)
1113 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1116 nlongs
= BITS_TO_LONGS(maxnode
);
1117 if ((maxnode
% BITS_PER_LONG
) == 0)
1120 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1122 /* When the user specified more nodes than supported just check
1123 if the non supported part is all zero. */
1124 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1125 if (nlongs
> PAGE_SIZE
/sizeof(long))
1127 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1129 if (get_user(t
, nmask
+ k
))
1131 if (k
== nlongs
- 1) {
1137 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1141 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1143 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1147 /* Copy a kernel node mask to user space */
1148 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1151 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1152 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1154 if (copy
> nbytes
) {
1155 if (copy
> PAGE_SIZE
)
1157 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1161 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1164 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1165 unsigned long, mode
, unsigned long __user
*, nmask
,
1166 unsigned long, maxnode
, unsigned, flags
)
1170 unsigned short mode_flags
;
1172 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1173 mode
&= ~MPOL_MODE_FLAGS
;
1174 if (mode
>= MPOL_MAX
)
1176 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1177 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1179 err
= get_nodes(&nodes
, nmask
, maxnode
);
1182 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1185 /* Set the process memory policy */
1186 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, unsigned long __user
*, nmask
,
1187 unsigned long, maxnode
)
1191 unsigned short flags
;
1193 flags
= mode
& MPOL_MODE_FLAGS
;
1194 mode
&= ~MPOL_MODE_FLAGS
;
1195 if ((unsigned int)mode
>= MPOL_MAX
)
1197 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1199 err
= get_nodes(&nodes
, nmask
, maxnode
);
1202 return do_set_mempolicy(mode
, flags
, &nodes
);
1205 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1206 const unsigned long __user
*, old_nodes
,
1207 const unsigned long __user
*, new_nodes
)
1209 const struct cred
*cred
= current_cred(), *tcred
;
1210 struct mm_struct
*mm
;
1211 struct task_struct
*task
;
1214 nodemask_t task_nodes
;
1217 err
= get_nodes(&old
, old_nodes
, maxnode
);
1221 err
= get_nodes(&new, new_nodes
, maxnode
);
1225 /* Find the mm_struct */
1226 read_lock(&tasklist_lock
);
1227 task
= pid
? find_task_by_vpid(pid
) : current
;
1229 read_unlock(&tasklist_lock
);
1232 mm
= get_task_mm(task
);
1233 read_unlock(&tasklist_lock
);
1239 * Check if this process has the right to modify the specified
1240 * process. The right exists if the process has administrative
1241 * capabilities, superuser privileges or the same
1242 * userid as the target process.
1245 tcred
= __task_cred(task
);
1246 if (cred
->euid
!= tcred
->suid
&& cred
->euid
!= tcred
->uid
&&
1247 cred
->uid
!= tcred
->suid
&& cred
->uid
!= tcred
->uid
&&
1248 !capable(CAP_SYS_NICE
)) {
1255 task_nodes
= cpuset_mems_allowed(task
);
1256 /* Is the user allowed to access the target nodes? */
1257 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1262 if (!nodes_subset(new, node_states
[N_HIGH_MEMORY
])) {
1267 err
= security_task_movememory(task
);
1271 err
= do_migrate_pages(mm
, &old
, &new,
1272 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1279 /* Retrieve NUMA policy */
1280 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1281 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1282 unsigned long, addr
, unsigned long, flags
)
1285 int uninitialized_var(pval
);
1288 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1291 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1296 if (policy
&& put_user(pval
, policy
))
1300 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1305 #ifdef CONFIG_COMPAT
1307 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1308 compat_ulong_t __user
*nmask
,
1309 compat_ulong_t maxnode
,
1310 compat_ulong_t addr
, compat_ulong_t flags
)
1313 unsigned long __user
*nm
= NULL
;
1314 unsigned long nr_bits
, alloc_size
;
1315 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1317 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1318 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1321 nm
= compat_alloc_user_space(alloc_size
);
1323 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1325 if (!err
&& nmask
) {
1326 err
= copy_from_user(bm
, nm
, alloc_size
);
1327 /* ensure entire bitmap is zeroed */
1328 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1329 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1335 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1336 compat_ulong_t maxnode
)
1339 unsigned long __user
*nm
= NULL
;
1340 unsigned long nr_bits
, alloc_size
;
1341 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1343 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1344 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1347 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1348 nm
= compat_alloc_user_space(alloc_size
);
1349 err
|= copy_to_user(nm
, bm
, alloc_size
);
1355 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1358 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1359 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1360 compat_ulong_t maxnode
, compat_ulong_t flags
)
1363 unsigned long __user
*nm
= NULL
;
1364 unsigned long nr_bits
, alloc_size
;
1367 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1368 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1371 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1372 nm
= compat_alloc_user_space(alloc_size
);
1373 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1379 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1385 * get_vma_policy(@task, @vma, @addr)
1386 * @task - task for fallback if vma policy == default
1387 * @vma - virtual memory area whose policy is sought
1388 * @addr - address in @vma for shared policy lookup
1390 * Returns effective policy for a VMA at specified address.
1391 * Falls back to @task or system default policy, as necessary.
1392 * Current or other task's task mempolicy and non-shared vma policies
1393 * are protected by the task's mmap_sem, which must be held for read by
1395 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1396 * count--added by the get_policy() vm_op, as appropriate--to protect against
1397 * freeing by another task. It is the caller's responsibility to free the
1398 * extra reference for shared policies.
1400 static struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1401 struct vm_area_struct
*vma
, unsigned long addr
)
1403 struct mempolicy
*pol
= task
->mempolicy
;
1406 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1407 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1411 } else if (vma
->vm_policy
)
1412 pol
= vma
->vm_policy
;
1415 pol
= &default_policy
;
1420 * Return a nodemask representing a mempolicy for filtering nodes for
1423 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1425 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1426 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1427 gfp_zone(gfp
) >= policy_zone
&&
1428 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1429 return &policy
->v
.nodes
;
1434 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1435 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
)
1437 int nd
= numa_node_id();
1439 switch (policy
->mode
) {
1440 case MPOL_PREFERRED
:
1441 if (!(policy
->flags
& MPOL_F_LOCAL
))
1442 nd
= policy
->v
.preferred_node
;
1446 * Normally, MPOL_BIND allocations are node-local within the
1447 * allowed nodemask. However, if __GFP_THISNODE is set and the
1448 * current node is part of the mask, we use the zonelist for
1449 * the first node in the mask instead.
1451 if (unlikely(gfp
& __GFP_THISNODE
) &&
1452 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1453 nd
= first_node(policy
->v
.nodes
);
1455 case MPOL_INTERLEAVE
: /* should not happen */
1460 return node_zonelist(nd
, gfp
);
1463 /* Do dynamic interleaving for a process */
1464 static unsigned interleave_nodes(struct mempolicy
*policy
)
1467 struct task_struct
*me
= current
;
1470 next
= next_node(nid
, policy
->v
.nodes
);
1471 if (next
>= MAX_NUMNODES
)
1472 next
= first_node(policy
->v
.nodes
);
1473 if (next
< MAX_NUMNODES
)
1479 * Depending on the memory policy provide a node from which to allocate the
1481 * @policy must be protected by freeing by the caller. If @policy is
1482 * the current task's mempolicy, this protection is implicit, as only the
1483 * task can change it's policy. The system default policy requires no
1486 unsigned slab_node(struct mempolicy
*policy
)
1488 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1489 return numa_node_id();
1491 switch (policy
->mode
) {
1492 case MPOL_PREFERRED
:
1494 * handled MPOL_F_LOCAL above
1496 return policy
->v
.preferred_node
;
1498 case MPOL_INTERLEAVE
:
1499 return interleave_nodes(policy
);
1503 * Follow bind policy behavior and start allocation at the
1506 struct zonelist
*zonelist
;
1508 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1509 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1510 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1521 /* Do static interleaving for a VMA with known offset. */
1522 static unsigned offset_il_node(struct mempolicy
*pol
,
1523 struct vm_area_struct
*vma
, unsigned long off
)
1525 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1531 return numa_node_id();
1532 target
= (unsigned int)off
% nnodes
;
1535 nid
= next_node(nid
, pol
->v
.nodes
);
1537 } while (c
<= target
);
1541 /* Determine a node number for interleave */
1542 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1543 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1549 * for small pages, there is no difference between
1550 * shift and PAGE_SHIFT, so the bit-shift is safe.
1551 * for huge pages, since vm_pgoff is in units of small
1552 * pages, we need to shift off the always 0 bits to get
1555 BUG_ON(shift
< PAGE_SHIFT
);
1556 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1557 off
+= (addr
- vma
->vm_start
) >> shift
;
1558 return offset_il_node(pol
, vma
, off
);
1560 return interleave_nodes(pol
);
1563 #ifdef CONFIG_HUGETLBFS
1565 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1566 * @vma = virtual memory area whose policy is sought
1567 * @addr = address in @vma for shared policy lookup and interleave policy
1568 * @gfp_flags = for requested zone
1569 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1570 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1572 * Returns a zonelist suitable for a huge page allocation and a pointer
1573 * to the struct mempolicy for conditional unref after allocation.
1574 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1575 * @nodemask for filtering the zonelist.
1577 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1578 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1579 nodemask_t
**nodemask
)
1581 struct zonelist
*zl
;
1583 *mpol
= get_vma_policy(current
, vma
, addr
);
1584 *nodemask
= NULL
; /* assume !MPOL_BIND */
1586 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1587 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1588 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1590 zl
= policy_zonelist(gfp_flags
, *mpol
);
1591 if ((*mpol
)->mode
== MPOL_BIND
)
1592 *nodemask
= &(*mpol
)->v
.nodes
;
1598 * init_nodemask_of_mempolicy
1600 * If the current task's mempolicy is "default" [NULL], return 'false'
1601 * to indicate default policy. Otherwise, extract the policy nodemask
1602 * for 'bind' or 'interleave' policy into the argument nodemask, or
1603 * initialize the argument nodemask to contain the single node for
1604 * 'preferred' or 'local' policy and return 'true' to indicate presence
1605 * of non-default mempolicy.
1607 * We don't bother with reference counting the mempolicy [mpol_get/put]
1608 * because the current task is examining it's own mempolicy and a task's
1609 * mempolicy is only ever changed by the task itself.
1611 * N.B., it is the caller's responsibility to free a returned nodemask.
1613 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1615 struct mempolicy
*mempolicy
;
1618 if (!(mask
&& current
->mempolicy
))
1621 mempolicy
= current
->mempolicy
;
1622 switch (mempolicy
->mode
) {
1623 case MPOL_PREFERRED
:
1624 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1625 nid
= numa_node_id();
1627 nid
= mempolicy
->v
.preferred_node
;
1628 init_nodemask_of_node(mask
, nid
);
1633 case MPOL_INTERLEAVE
:
1634 *mask
= mempolicy
->v
.nodes
;
1645 /* Allocate a page in interleaved policy.
1646 Own path because it needs to do special accounting. */
1647 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1650 struct zonelist
*zl
;
1653 zl
= node_zonelist(nid
, gfp
);
1654 page
= __alloc_pages(gfp
, order
, zl
);
1655 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1656 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1661 * alloc_page_vma - Allocate a page for a VMA.
1664 * %GFP_USER user allocation.
1665 * %GFP_KERNEL kernel allocations,
1666 * %GFP_HIGHMEM highmem/user allocations,
1667 * %GFP_FS allocation should not call back into a file system.
1668 * %GFP_ATOMIC don't sleep.
1670 * @vma: Pointer to VMA or NULL if not available.
1671 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1673 * This function allocates a page from the kernel page pool and applies
1674 * a NUMA policy associated with the VMA or the current process.
1675 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1676 * mm_struct of the VMA to prevent it from going away. Should be used for
1677 * all allocations for pages that will be mapped into
1678 * user space. Returns NULL when no page can be allocated.
1680 * Should be called with the mm_sem of the vma hold.
1683 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1685 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1686 struct zonelist
*zl
;
1688 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
1691 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1693 return alloc_page_interleave(gfp
, 0, nid
);
1695 zl
= policy_zonelist(gfp
, pol
);
1696 if (unlikely(mpol_needs_cond_ref(pol
))) {
1698 * slow path: ref counted shared policy
1700 struct page
*page
= __alloc_pages_nodemask(gfp
, 0,
1701 zl
, policy_nodemask(gfp
, pol
));
1706 * fast path: default or task policy
1708 return __alloc_pages_nodemask(gfp
, 0, zl
, policy_nodemask(gfp
, pol
));
1712 * alloc_pages_current - Allocate pages.
1715 * %GFP_USER user allocation,
1716 * %GFP_KERNEL kernel allocation,
1717 * %GFP_HIGHMEM highmem allocation,
1718 * %GFP_FS don't call back into a file system.
1719 * %GFP_ATOMIC don't sleep.
1720 * @order: Power of two of allocation size in pages. 0 is a single page.
1722 * Allocate a page from the kernel page pool. When not in
1723 * interrupt context and apply the current process NUMA policy.
1724 * Returns NULL when no page can be allocated.
1726 * Don't call cpuset_update_task_memory_state() unless
1727 * 1) it's ok to take cpuset_sem (can WAIT), and
1728 * 2) allocating for current task (not interrupt).
1730 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1732 struct mempolicy
*pol
= current
->mempolicy
;
1734 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
1735 pol
= &default_policy
;
1738 * No reference counting needed for current->mempolicy
1739 * nor system default_policy
1741 if (pol
->mode
== MPOL_INTERLEAVE
)
1742 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1743 return __alloc_pages_nodemask(gfp
, order
,
1744 policy_zonelist(gfp
, pol
), policy_nodemask(gfp
, pol
));
1746 EXPORT_SYMBOL(alloc_pages_current
);
1749 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1750 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1751 * with the mems_allowed returned by cpuset_mems_allowed(). This
1752 * keeps mempolicies cpuset relative after its cpuset moves. See
1753 * further kernel/cpuset.c update_nodemask().
1756 /* Slow path of a mempolicy duplicate */
1757 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
1759 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1762 return ERR_PTR(-ENOMEM
);
1764 if (current_cpuset_is_being_rebound()) {
1765 nodemask_t mems
= cpuset_mems_allowed(current
);
1766 mpol_rebind_policy(old
, &mems
);
1770 atomic_set(&new->refcnt
, 1);
1775 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1776 * eliminate the * MPOL_F_* flags that require conditional ref and
1777 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1778 * after return. Use the returned value.
1780 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1781 * policy lookup, even if the policy needs/has extra ref on lookup.
1782 * shmem_readahead needs this.
1784 struct mempolicy
*__mpol_cond_copy(struct mempolicy
*tompol
,
1785 struct mempolicy
*frompol
)
1787 if (!mpol_needs_cond_ref(frompol
))
1791 tompol
->flags
&= ~MPOL_F_SHARED
; /* copy doesn't need unref */
1792 __mpol_put(frompol
);
1796 static int mpol_match_intent(const struct mempolicy
*a
,
1797 const struct mempolicy
*b
)
1799 if (a
->flags
!= b
->flags
)
1801 if (!mpol_store_user_nodemask(a
))
1803 return nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
);
1806 /* Slow path of a mempolicy comparison */
1807 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1811 if (a
->mode
!= b
->mode
)
1813 if (a
->mode
!= MPOL_DEFAULT
&& !mpol_match_intent(a
, b
))
1818 case MPOL_INTERLEAVE
:
1819 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1820 case MPOL_PREFERRED
:
1821 return a
->v
.preferred_node
== b
->v
.preferred_node
&&
1822 a
->flags
== b
->flags
;
1830 * Shared memory backing store policy support.
1832 * Remember policies even when nobody has shared memory mapped.
1833 * The policies are kept in Red-Black tree linked from the inode.
1834 * They are protected by the sp->lock spinlock, which should be held
1835 * for any accesses to the tree.
1838 /* lookup first element intersecting start-end */
1839 /* Caller holds sp->lock */
1840 static struct sp_node
*
1841 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1843 struct rb_node
*n
= sp
->root
.rb_node
;
1846 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1848 if (start
>= p
->end
)
1850 else if (end
<= p
->start
)
1858 struct sp_node
*w
= NULL
;
1859 struct rb_node
*prev
= rb_prev(n
);
1862 w
= rb_entry(prev
, struct sp_node
, nd
);
1863 if (w
->end
<= start
)
1867 return rb_entry(n
, struct sp_node
, nd
);
1870 /* Insert a new shared policy into the list. */
1871 /* Caller holds sp->lock */
1872 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1874 struct rb_node
**p
= &sp
->root
.rb_node
;
1875 struct rb_node
*parent
= NULL
;
1880 nd
= rb_entry(parent
, struct sp_node
, nd
);
1881 if (new->start
< nd
->start
)
1883 else if (new->end
> nd
->end
)
1884 p
= &(*p
)->rb_right
;
1888 rb_link_node(&new->nd
, parent
, p
);
1889 rb_insert_color(&new->nd
, &sp
->root
);
1890 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
1891 new->policy
? new->policy
->mode
: 0);
1894 /* Find shared policy intersecting idx */
1896 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1898 struct mempolicy
*pol
= NULL
;
1901 if (!sp
->root
.rb_node
)
1903 spin_lock(&sp
->lock
);
1904 sn
= sp_lookup(sp
, idx
, idx
+1);
1906 mpol_get(sn
->policy
);
1909 spin_unlock(&sp
->lock
);
1913 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1915 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
1916 rb_erase(&n
->nd
, &sp
->root
);
1917 mpol_put(n
->policy
);
1918 kmem_cache_free(sn_cache
, n
);
1921 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
1922 struct mempolicy
*pol
)
1924 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1931 pol
->flags
|= MPOL_F_SHARED
; /* for unref */
1936 /* Replace a policy range. */
1937 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1938 unsigned long end
, struct sp_node
*new)
1940 struct sp_node
*n
, *new2
= NULL
;
1943 spin_lock(&sp
->lock
);
1944 n
= sp_lookup(sp
, start
, end
);
1945 /* Take care of old policies in the same range. */
1946 while (n
&& n
->start
< end
) {
1947 struct rb_node
*next
= rb_next(&n
->nd
);
1948 if (n
->start
>= start
) {
1954 /* Old policy spanning whole new range. */
1957 spin_unlock(&sp
->lock
);
1958 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1964 sp_insert(sp
, new2
);
1972 n
= rb_entry(next
, struct sp_node
, nd
);
1976 spin_unlock(&sp
->lock
);
1978 mpol_put(new2
->policy
);
1979 kmem_cache_free(sn_cache
, new2
);
1985 * mpol_shared_policy_init - initialize shared policy for inode
1986 * @sp: pointer to inode shared policy
1987 * @mpol: struct mempolicy to install
1989 * Install non-NULL @mpol in inode's shared policy rb-tree.
1990 * On entry, the current task has a reference on a non-NULL @mpol.
1991 * This must be released on exit.
1992 * This is called at get_inode() calls and we can use GFP_KERNEL.
1994 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
1998 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
1999 spin_lock_init(&sp
->lock
);
2002 struct vm_area_struct pvma
;
2003 struct mempolicy
*new;
2004 NODEMASK_SCRATCH(scratch
);
2008 /* contextualize the tmpfs mount point mempolicy */
2009 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2011 mpol_put(mpol
); /* drop our ref on sb mpol */
2012 NODEMASK_SCRATCH_FREE(scratch
);
2013 return; /* no valid nodemask intersection */
2017 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2018 task_unlock(current
);
2019 mpol_put(mpol
); /* drop our ref on sb mpol */
2021 NODEMASK_SCRATCH_FREE(scratch
);
2026 /* Create pseudo-vma that contains just the policy */
2027 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2028 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2029 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2030 mpol_put(new); /* drop initial ref */
2031 NODEMASK_SCRATCH_FREE(scratch
);
2035 int mpol_set_shared_policy(struct shared_policy
*info
,
2036 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2039 struct sp_node
*new = NULL
;
2040 unsigned long sz
= vma_pages(vma
);
2042 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2044 sz
, npol
? npol
->mode
: -1,
2045 npol
? npol
->flags
: -1,
2046 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
2049 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2053 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2055 kmem_cache_free(sn_cache
, new);
2059 /* Free a backing policy store on inode delete. */
2060 void mpol_free_shared_policy(struct shared_policy
*p
)
2063 struct rb_node
*next
;
2065 if (!p
->root
.rb_node
)
2067 spin_lock(&p
->lock
);
2068 next
= rb_first(&p
->root
);
2070 n
= rb_entry(next
, struct sp_node
, nd
);
2071 next
= rb_next(&n
->nd
);
2072 rb_erase(&n
->nd
, &p
->root
);
2073 mpol_put(n
->policy
);
2074 kmem_cache_free(sn_cache
, n
);
2076 spin_unlock(&p
->lock
);
2079 /* assumes fs == KERNEL_DS */
2080 void __init
numa_policy_init(void)
2082 nodemask_t interleave_nodes
;
2083 unsigned long largest
= 0;
2084 int nid
, prefer
= 0;
2086 policy_cache
= kmem_cache_create("numa_policy",
2087 sizeof(struct mempolicy
),
2088 0, SLAB_PANIC
, NULL
);
2090 sn_cache
= kmem_cache_create("shared_policy_node",
2091 sizeof(struct sp_node
),
2092 0, SLAB_PANIC
, NULL
);
2095 * Set interleaving policy for system init. Interleaving is only
2096 * enabled across suitably sized nodes (default is >= 16MB), or
2097 * fall back to the largest node if they're all smaller.
2099 nodes_clear(interleave_nodes
);
2100 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2101 unsigned long total_pages
= node_present_pages(nid
);
2103 /* Preserve the largest node */
2104 if (largest
< total_pages
) {
2105 largest
= total_pages
;
2109 /* Interleave this node? */
2110 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2111 node_set(nid
, interleave_nodes
);
2114 /* All too small, use the largest */
2115 if (unlikely(nodes_empty(interleave_nodes
)))
2116 node_set(prefer
, interleave_nodes
);
2118 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2119 printk("numa_policy_init: interleaving failed\n");
2122 /* Reset policy of current process to default */
2123 void numa_default_policy(void)
2125 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2129 * Parse and format mempolicy from/to strings
2133 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2134 * Used only for mpol_parse_str() and mpol_to_str()
2136 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
2137 static const char * const policy_types
[] =
2138 { "default", "prefer", "bind", "interleave", "local" };
2143 * mpol_parse_str - parse string to mempolicy
2144 * @str: string containing mempolicy to parse
2145 * @mpol: pointer to struct mempolicy pointer, returned on success.
2146 * @no_context: flag whether to "contextualize" the mempolicy
2149 * <mode>[=<flags>][:<nodelist>]
2151 * if @no_context is true, save the input nodemask in w.user_nodemask in
2152 * the returned mempolicy. This will be used to "clone" the mempolicy in
2153 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2154 * mount option. Note that if 'static' or 'relative' mode flags were
2155 * specified, the input nodemask will already have been saved. Saving
2156 * it again is redundant, but safe.
2158 * On success, returns 0, else 1
2160 int mpol_parse_str(char *str
, struct mempolicy
**mpol
, int no_context
)
2162 struct mempolicy
*new = NULL
;
2163 unsigned short uninitialized_var(mode
);
2164 unsigned short uninitialized_var(mode_flags
);
2166 char *nodelist
= strchr(str
, ':');
2167 char *flags
= strchr(str
, '=');
2172 /* NUL-terminate mode or flags string */
2174 if (nodelist_parse(nodelist
, nodes
))
2176 if (!nodes_subset(nodes
, node_states
[N_HIGH_MEMORY
]))
2182 *flags
++ = '\0'; /* terminate mode string */
2184 for (i
= 0; i
<= MPOL_LOCAL
; i
++) {
2185 if (!strcmp(str
, policy_types
[i
])) {
2194 case MPOL_PREFERRED
:
2196 * Insist on a nodelist of one node only
2199 char *rest
= nodelist
;
2200 while (isdigit(*rest
))
2206 case MPOL_INTERLEAVE
:
2208 * Default to online nodes with memory if no nodelist
2211 nodes
= node_states
[N_HIGH_MEMORY
];
2215 * Don't allow a nodelist; mpol_new() checks flags
2219 mode
= MPOL_PREFERRED
;
2223 * Insist on a empty nodelist
2230 * Insist on a nodelist
2239 * Currently, we only support two mutually exclusive
2242 if (!strcmp(flags
, "static"))
2243 mode_flags
|= MPOL_F_STATIC_NODES
;
2244 else if (!strcmp(flags
, "relative"))
2245 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2250 new = mpol_new(mode
, mode_flags
, &nodes
);
2256 NODEMASK_SCRATCH(scratch
);
2259 ret
= mpol_set_nodemask(new, &nodes
, scratch
);
2260 task_unlock(current
);
2263 NODEMASK_SCRATCH_FREE(scratch
);
2271 /* save for contextualization */
2272 new->w
.user_nodemask
= nodes
;
2276 /* Restore string for error message */
2285 #endif /* CONFIG_TMPFS */
2288 * mpol_to_str - format a mempolicy structure for printing
2289 * @buffer: to contain formatted mempolicy string
2290 * @maxlen: length of @buffer
2291 * @pol: pointer to mempolicy to be formatted
2292 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2294 * Convert a mempolicy into a string.
2295 * Returns the number of characters in buffer (if positive)
2296 * or an error (negative)
2298 int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
, int no_context
)
2303 unsigned short mode
;
2304 unsigned short flags
= pol
? pol
->flags
: 0;
2307 * Sanity check: room for longest mode, flag and some nodes
2309 VM_BUG_ON(maxlen
< strlen("interleave") + strlen("relative") + 16);
2311 if (!pol
|| pol
== &default_policy
)
2312 mode
= MPOL_DEFAULT
;
2321 case MPOL_PREFERRED
:
2323 if (flags
& MPOL_F_LOCAL
)
2324 mode
= MPOL_LOCAL
; /* pseudo-policy */
2326 node_set(pol
->v
.preferred_node
, nodes
);
2331 case MPOL_INTERLEAVE
:
2333 nodes
= pol
->w
.user_nodemask
;
2335 nodes
= pol
->v
.nodes
;
2342 l
= strlen(policy_types
[mode
]);
2343 if (buffer
+ maxlen
< p
+ l
+ 1)
2346 strcpy(p
, policy_types
[mode
]);
2349 if (flags
& MPOL_MODE_FLAGS
) {
2350 if (buffer
+ maxlen
< p
+ 2)
2355 * Currently, the only defined flags are mutually exclusive
2357 if (flags
& MPOL_F_STATIC_NODES
)
2358 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2359 else if (flags
& MPOL_F_RELATIVE_NODES
)
2360 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2363 if (!nodes_empty(nodes
)) {
2364 if (buffer
+ maxlen
< p
+ 2)
2367 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
2373 unsigned long pages
;
2375 unsigned long active
;
2376 unsigned long writeback
;
2377 unsigned long mapcount_max
;
2378 unsigned long dirty
;
2379 unsigned long swapcache
;
2380 unsigned long node
[MAX_NUMNODES
];
2383 static void gather_stats(struct page
*page
, void *private, int pte_dirty
)
2385 struct numa_maps
*md
= private;
2386 int count
= page_mapcount(page
);
2389 if (pte_dirty
|| PageDirty(page
))
2392 if (PageSwapCache(page
))
2395 if (PageActive(page
) || PageUnevictable(page
))
2398 if (PageWriteback(page
))
2404 if (count
> md
->mapcount_max
)
2405 md
->mapcount_max
= count
;
2407 md
->node
[page_to_nid(page
)]++;
2410 #ifdef CONFIG_HUGETLB_PAGE
2411 static void check_huge_range(struct vm_area_struct
*vma
,
2412 unsigned long start
, unsigned long end
,
2413 struct numa_maps
*md
)
2417 struct hstate
*h
= hstate_vma(vma
);
2418 unsigned long sz
= huge_page_size(h
);
2420 for (addr
= start
; addr
< end
; addr
+= sz
) {
2421 pte_t
*ptep
= huge_pte_offset(vma
->vm_mm
,
2422 addr
& huge_page_mask(h
));
2432 page
= pte_page(pte
);
2436 gather_stats(page
, md
, pte_dirty(*ptep
));
2440 static inline void check_huge_range(struct vm_area_struct
*vma
,
2441 unsigned long start
, unsigned long end
,
2442 struct numa_maps
*md
)
2448 * Display pages allocated per node and memory policy via /proc.
2450 int show_numa_map(struct seq_file
*m
, void *v
)
2452 struct proc_maps_private
*priv
= m
->private;
2453 struct vm_area_struct
*vma
= v
;
2454 struct numa_maps
*md
;
2455 struct file
*file
= vma
->vm_file
;
2456 struct mm_struct
*mm
= vma
->vm_mm
;
2457 struct mempolicy
*pol
;
2464 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
2468 pol
= get_vma_policy(priv
->task
, vma
, vma
->vm_start
);
2469 mpol_to_str(buffer
, sizeof(buffer
), pol
, 0);
2472 seq_printf(m
, "%08lx %s", vma
->vm_start
, buffer
);
2475 seq_printf(m
, " file=");
2476 seq_path(m
, &file
->f_path
, "\n\t= ");
2477 } else if (vma
->vm_start
<= mm
->brk
&& vma
->vm_end
>= mm
->start_brk
) {
2478 seq_printf(m
, " heap");
2479 } else if (vma
->vm_start
<= mm
->start_stack
&&
2480 vma
->vm_end
>= mm
->start_stack
) {
2481 seq_printf(m
, " stack");
2484 if (is_vm_hugetlb_page(vma
)) {
2485 check_huge_range(vma
, vma
->vm_start
, vma
->vm_end
, md
);
2486 seq_printf(m
, " huge");
2488 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
2489 &node_states
[N_HIGH_MEMORY
], MPOL_MF_STATS
, md
);
2496 seq_printf(m
," anon=%lu",md
->anon
);
2499 seq_printf(m
," dirty=%lu",md
->dirty
);
2501 if (md
->pages
!= md
->anon
&& md
->pages
!= md
->dirty
)
2502 seq_printf(m
, " mapped=%lu", md
->pages
);
2504 if (md
->mapcount_max
> 1)
2505 seq_printf(m
, " mapmax=%lu", md
->mapcount_max
);
2508 seq_printf(m
," swapcache=%lu", md
->swapcache
);
2510 if (md
->active
< md
->pages
&& !is_vm_hugetlb_page(vma
))
2511 seq_printf(m
," active=%lu", md
->active
);
2514 seq_printf(m
," writeback=%lu", md
->writeback
);
2516 for_each_node_state(n
, N_HIGH_MEMORY
)
2518 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
2523 if (m
->count
< m
->size
)
2524 m
->version
= (vma
!= priv
->tail_vma
) ? vma
->vm_start
: 0;