2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66 could replace all the switch()es with a mempolicy_ops structure.
69 #include <linux/mempolicy.h>
71 #include <linux/highmem.h>
72 #include <linux/hugetlb.h>
73 #include <linux/kernel.h>
74 #include <linux/sched.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/gfp.h>
79 #include <linux/slab.h>
80 #include <linux/string.h>
81 #include <linux/module.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/mempolicy.h>
86 #include <linux/swap.h>
87 #include <linux/seq_file.h>
88 #include <linux/proc_fs.h>
90 #include <asm/tlbflush.h>
91 #include <asm/uaccess.h>
94 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
95 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
96 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
98 static kmem_cache_t
*policy_cache
;
99 static kmem_cache_t
*sn_cache
;
101 #define PDprintk(fmt...)
103 /* Highest zone. An specific allocation for a zone below that is not
105 int policy_zone
= ZONE_DMA
;
107 struct mempolicy default_policy
= {
108 .refcnt
= ATOMIC_INIT(1), /* never free it */
109 .policy
= MPOL_DEFAULT
,
112 /* Do sanity checking on a policy */
113 static int mpol_check_policy(int mode
, nodemask_t
*nodes
)
115 int empty
= nodes_empty(*nodes
);
123 case MPOL_INTERLEAVE
:
124 /* Preferred will only use the first bit, but allow
130 return nodes_subset(*nodes
, node_online_map
) ? 0 : -EINVAL
;
132 /* Generate a custom zonelist for the BIND policy. */
133 static struct zonelist
*bind_zonelist(nodemask_t
*nodes
)
138 max
= 1 + MAX_NR_ZONES
* nodes_weight(*nodes
);
139 zl
= kmalloc(sizeof(void *) * max
, GFP_KERNEL
);
143 for_each_node_mask(nd
, *nodes
)
144 zl
->zones
[num
++] = &NODE_DATA(nd
)->node_zones
[policy_zone
];
145 zl
->zones
[num
] = NULL
;
149 /* Create a new policy */
150 static struct mempolicy
*mpol_new(int mode
, nodemask_t
*nodes
)
152 struct mempolicy
*policy
;
154 PDprintk("setting mode %d nodes[0] %lx\n", mode
, nodes_addr(*nodes
)[0]);
155 if (mode
== MPOL_DEFAULT
)
157 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
159 return ERR_PTR(-ENOMEM
);
160 atomic_set(&policy
->refcnt
, 1);
162 case MPOL_INTERLEAVE
:
163 policy
->v
.nodes
= *nodes
;
164 if (nodes_weight(*nodes
) == 0) {
165 kmem_cache_free(policy_cache
, policy
);
166 return ERR_PTR(-EINVAL
);
170 policy
->v
.preferred_node
= first_node(*nodes
);
171 if (policy
->v
.preferred_node
>= MAX_NUMNODES
)
172 policy
->v
.preferred_node
= -1;
175 policy
->v
.zonelist
= bind_zonelist(nodes
);
176 if (policy
->v
.zonelist
== NULL
) {
177 kmem_cache_free(policy_cache
, policy
);
178 return ERR_PTR(-ENOMEM
);
182 policy
->policy
= mode
;
186 /* Check if we are the only process mapping the page in question */
187 static inline int single_mm_mapping(struct mm_struct
*mm
,
188 struct address_space
*mapping
)
190 struct vm_area_struct
*vma
;
191 struct prio_tree_iter iter
;
194 spin_lock(&mapping
->i_mmap_lock
);
195 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
196 if (mm
!= vma
->vm_mm
) {
200 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
201 if (mm
!= vma
->vm_mm
) {
206 spin_unlock(&mapping
->i_mmap_lock
);
211 * Add a page to be migrated to the pagelist
213 static void migrate_page_add(struct vm_area_struct
*vma
,
214 struct page
*page
, struct list_head
*pagelist
, unsigned long flags
)
217 * Avoid migrating a page that is shared by others and not writable.
219 if ((flags
& MPOL_MF_MOVE_ALL
) || !page
->mapping
|| PageAnon(page
) ||
220 mapping_writably_mapped(page
->mapping
) ||
221 single_mm_mapping(vma
->vm_mm
, page
->mapping
)) {
222 int rc
= isolate_lru_page(page
);
225 list_add(&page
->lru
, pagelist
);
227 * If the isolate attempt was not successful then we just
228 * encountered an unswappable page. Something must be wrong.
234 static void gather_stats(struct page
*, void *);
236 /* Scan through pages checking if pages follow certain conditions. */
237 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
238 unsigned long addr
, unsigned long end
,
239 const nodemask_t
*nodes
, unsigned long flags
,
246 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
251 if (!pte_present(*pte
))
253 page
= vm_normal_page(vma
, addr
, *pte
);
256 nid
= page_to_nid(page
);
257 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
260 if (flags
& MPOL_MF_STATS
)
261 gather_stats(page
, private);
262 else if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
264 migrate_page_add(vma
, page
, private, flags
);
269 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
270 pte_unmap_unlock(orig_pte
, ptl
);
274 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
275 unsigned long addr
, unsigned long end
,
276 const nodemask_t
*nodes
, unsigned long flags
,
282 pmd
= pmd_offset(pud
, addr
);
284 next
= pmd_addr_end(addr
, end
);
285 if (pmd_none_or_clear_bad(pmd
))
287 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
290 } while (pmd
++, addr
= next
, addr
!= end
);
294 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
295 unsigned long addr
, unsigned long end
,
296 const nodemask_t
*nodes
, unsigned long flags
,
302 pud
= pud_offset(pgd
, addr
);
304 next
= pud_addr_end(addr
, end
);
305 if (pud_none_or_clear_bad(pud
))
307 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
310 } while (pud
++, addr
= next
, addr
!= end
);
314 static inline int check_pgd_range(struct vm_area_struct
*vma
,
315 unsigned long addr
, unsigned long end
,
316 const nodemask_t
*nodes
, unsigned long flags
,
322 pgd
= pgd_offset(vma
->vm_mm
, addr
);
324 next
= pgd_addr_end(addr
, end
);
325 if (pgd_none_or_clear_bad(pgd
))
327 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
330 } while (pgd
++, addr
= next
, addr
!= end
);
334 /* Check if a vma is migratable */
335 static inline int vma_migratable(struct vm_area_struct
*vma
)
337 if (vma
->vm_flags
& (
338 VM_LOCKED
|VM_IO
|VM_HUGETLB
|VM_PFNMAP
))
344 * Check if all pages in a range are on a set of nodes.
345 * If pagelist != NULL then isolate pages from the LRU and
346 * put them on the pagelist.
348 static struct vm_area_struct
*
349 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
350 const nodemask_t
*nodes
, unsigned long flags
, void *private)
353 struct vm_area_struct
*first
, *vma
, *prev
;
355 first
= find_vma(mm
, start
);
357 return ERR_PTR(-EFAULT
);
359 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
360 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
361 if (!vma
->vm_next
&& vma
->vm_end
< end
)
362 return ERR_PTR(-EFAULT
);
363 if (prev
&& prev
->vm_end
< vma
->vm_start
)
364 return ERR_PTR(-EFAULT
);
366 if (!is_vm_hugetlb_page(vma
) &&
367 ((flags
& MPOL_MF_STRICT
) ||
368 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
369 vma_migratable(vma
)))) {
370 unsigned long endvma
= vma
->vm_end
;
374 if (vma
->vm_start
> start
)
375 start
= vma
->vm_start
;
376 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
379 first
= ERR_PTR(err
);
388 /* Apply policy to a single VMA */
389 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
392 struct mempolicy
*old
= vma
->vm_policy
;
394 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
395 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
396 vma
->vm_ops
, vma
->vm_file
,
397 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
399 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
400 err
= vma
->vm_ops
->set_policy(vma
, new);
403 vma
->vm_policy
= new;
409 /* Step 2: apply policy to a range and do splits. */
410 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
411 unsigned long end
, struct mempolicy
*new)
413 struct vm_area_struct
*next
;
417 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
419 if (vma
->vm_start
< start
)
420 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
421 if (!err
&& vma
->vm_end
> end
)
422 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
424 err
= policy_vma(vma
, new);
431 static int contextualize_policy(int mode
, nodemask_t
*nodes
)
436 /* Update current mems_allowed */
437 cpuset_update_current_mems_allowed();
438 /* Ignore nodes not set in current->mems_allowed */
439 cpuset_restrict_to_mems_allowed(nodes
->bits
);
440 return mpol_check_policy(mode
, nodes
);
443 static int swap_pages(struct list_head
*pagelist
)
449 n
= migrate_pages(pagelist
, NULL
, &moved
, &failed
);
450 putback_lru_pages(&failed
);
451 putback_lru_pages(&moved
);
456 long do_mbind(unsigned long start
, unsigned long len
,
457 unsigned long mode
, nodemask_t
*nmask
, unsigned long flags
)
459 struct vm_area_struct
*vma
;
460 struct mm_struct
*mm
= current
->mm
;
461 struct mempolicy
*new;
466 if ((flags
& ~(unsigned long)(MPOL_MF_STRICT
|
467 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
470 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_RESOURCE
))
473 if (start
& ~PAGE_MASK
)
476 if (mode
== MPOL_DEFAULT
)
477 flags
&= ~MPOL_MF_STRICT
;
479 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
487 if (mpol_check_policy(mode
, nmask
))
490 new = mpol_new(mode
, nmask
);
495 * If we are using the default policy then operation
496 * on discontinuous address spaces is okay after all
499 flags
|= MPOL_MF_DISCONTIG_OK
;
501 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start
,start
+len
,
502 mode
,nodes_addr(nodes
)[0]);
504 down_write(&mm
->mmap_sem
);
505 vma
= check_range(mm
, start
, end
, nmask
,
506 flags
| MPOL_MF_INVERT
, &pagelist
);
512 err
= mbind_range(vma
, start
, end
, new);
513 if (!list_empty(&pagelist
))
514 nr_failed
= swap_pages(&pagelist
);
516 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
519 if (!list_empty(&pagelist
))
520 putback_lru_pages(&pagelist
);
522 up_write(&mm
->mmap_sem
);
527 /* Set the process memory policy */
528 long do_set_mempolicy(int mode
, nodemask_t
*nodes
)
530 struct mempolicy
*new;
532 if (contextualize_policy(mode
, nodes
))
534 new = mpol_new(mode
, nodes
);
537 mpol_free(current
->mempolicy
);
538 current
->mempolicy
= new;
539 if (new && new->policy
== MPOL_INTERLEAVE
)
540 current
->il_next
= first_node(new->v
.nodes
);
544 /* Fill a zone bitmap for a policy */
545 static void get_zonemask(struct mempolicy
*p
, nodemask_t
*nodes
)
552 for (i
= 0; p
->v
.zonelist
->zones
[i
]; i
++)
553 node_set(p
->v
.zonelist
->zones
[i
]->zone_pgdat
->node_id
,
558 case MPOL_INTERLEAVE
:
562 /* or use current node instead of online map? */
563 if (p
->v
.preferred_node
< 0)
564 *nodes
= node_online_map
;
566 node_set(p
->v
.preferred_node
, *nodes
);
573 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
578 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
580 err
= page_to_nid(p
);
586 /* Retrieve NUMA policy */
587 long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
588 unsigned long addr
, unsigned long flags
)
591 struct mm_struct
*mm
= current
->mm
;
592 struct vm_area_struct
*vma
= NULL
;
593 struct mempolicy
*pol
= current
->mempolicy
;
595 cpuset_update_current_mems_allowed();
596 if (flags
& ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
))
598 if (flags
& MPOL_F_ADDR
) {
599 down_read(&mm
->mmap_sem
);
600 vma
= find_vma_intersection(mm
, addr
, addr
+1);
602 up_read(&mm
->mmap_sem
);
605 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
606 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
608 pol
= vma
->vm_policy
;
613 pol
= &default_policy
;
615 if (flags
& MPOL_F_NODE
) {
616 if (flags
& MPOL_F_ADDR
) {
617 err
= lookup_node(mm
, addr
);
621 } else if (pol
== current
->mempolicy
&&
622 pol
->policy
== MPOL_INTERLEAVE
) {
623 *policy
= current
->il_next
;
629 *policy
= pol
->policy
;
632 up_read(¤t
->mm
->mmap_sem
);
638 get_zonemask(pol
, nmask
);
642 up_read(¤t
->mm
->mmap_sem
);
647 * For now migrate_pages simply swaps out the pages from nodes that are in
648 * the source set but not in the target set. In the future, we would
649 * want a function that moves pages between the two nodesets in such
650 * a way as to preserve the physical layout as much as possible.
652 * Returns the number of page that could not be moved.
654 int do_migrate_pages(struct mm_struct
*mm
,
655 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
661 nodes_andnot(nodes
, *from_nodes
, *to_nodes
);
663 down_read(&mm
->mmap_sem
);
664 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nodes
,
665 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
667 if (!list_empty(&pagelist
)) {
668 count
= swap_pages(&pagelist
);
669 putback_lru_pages(&pagelist
);
672 up_read(&mm
->mmap_sem
);
677 * User space interface with variable sized bitmaps for nodelists.
680 /* Copy a node mask from user space. */
681 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
682 unsigned long maxnode
)
685 unsigned long nlongs
;
686 unsigned long endmask
;
690 if (maxnode
== 0 || !nmask
)
693 nlongs
= BITS_TO_LONGS(maxnode
);
694 if ((maxnode
% BITS_PER_LONG
) == 0)
697 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
699 /* When the user specified more nodes than supported just check
700 if the non supported part is all zero. */
701 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
702 if (nlongs
> PAGE_SIZE
/sizeof(long))
704 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
706 if (get_user(t
, nmask
+ k
))
708 if (k
== nlongs
- 1) {
714 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
718 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
720 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
724 /* Copy a kernel node mask to user space */
725 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
728 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
729 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
732 if (copy
> PAGE_SIZE
)
734 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
738 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
741 asmlinkage
long sys_mbind(unsigned long start
, unsigned long len
,
743 unsigned long __user
*nmask
, unsigned long maxnode
,
749 err
= get_nodes(&nodes
, nmask
, maxnode
);
752 return do_mbind(start
, len
, mode
, &nodes
, flags
);
755 /* Set the process memory policy */
756 asmlinkage
long sys_set_mempolicy(int mode
, unsigned long __user
*nmask
,
757 unsigned long maxnode
)
762 if (mode
< 0 || mode
> MPOL_MAX
)
764 err
= get_nodes(&nodes
, nmask
, maxnode
);
767 return do_set_mempolicy(mode
, &nodes
);
770 /* Macro needed until Paul implements this function in kernel/cpusets.c */
771 #define cpuset_mems_allowed(task) node_online_map
773 asmlinkage
long sys_migrate_pages(pid_t pid
, unsigned long maxnode
,
774 const unsigned long __user
*old_nodes
,
775 const unsigned long __user
*new_nodes
)
777 struct mm_struct
*mm
;
778 struct task_struct
*task
;
781 nodemask_t task_nodes
;
784 err
= get_nodes(&old
, old_nodes
, maxnode
);
788 err
= get_nodes(&new, new_nodes
, maxnode
);
792 /* Find the mm_struct */
793 read_lock(&tasklist_lock
);
794 task
= pid
? find_task_by_pid(pid
) : current
;
796 read_unlock(&tasklist_lock
);
799 mm
= get_task_mm(task
);
800 read_unlock(&tasklist_lock
);
806 * Check if this process has the right to modify the specified
807 * process. The right exists if the process has administrative
808 * capabilities, superuser priviledges or the same
809 * userid as the target process.
811 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
812 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
813 !capable(CAP_SYS_ADMIN
)) {
818 task_nodes
= cpuset_mems_allowed(task
);
819 /* Is the user allowed to access the target nodes? */
820 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_ADMIN
)) {
825 err
= do_migrate_pages(mm
, &old
, &new, MPOL_MF_MOVE
);
832 /* Retrieve NUMA policy */
833 asmlinkage
long sys_get_mempolicy(int __user
*policy
,
834 unsigned long __user
*nmask
,
835 unsigned long maxnode
,
836 unsigned long addr
, unsigned long flags
)
841 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
844 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
849 if (policy
&& put_user(pval
, policy
))
853 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
860 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
861 compat_ulong_t __user
*nmask
,
862 compat_ulong_t maxnode
,
863 compat_ulong_t addr
, compat_ulong_t flags
)
866 unsigned long __user
*nm
= NULL
;
867 unsigned long nr_bits
, alloc_size
;
868 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
870 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
871 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
874 nm
= compat_alloc_user_space(alloc_size
);
876 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
879 err
= copy_from_user(bm
, nm
, alloc_size
);
880 /* ensure entire bitmap is zeroed */
881 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
882 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
888 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
889 compat_ulong_t maxnode
)
892 unsigned long __user
*nm
= NULL
;
893 unsigned long nr_bits
, alloc_size
;
894 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
896 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
897 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
900 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
901 nm
= compat_alloc_user_space(alloc_size
);
902 err
|= copy_to_user(nm
, bm
, alloc_size
);
908 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
911 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
912 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
913 compat_ulong_t maxnode
, compat_ulong_t flags
)
916 unsigned long __user
*nm
= NULL
;
917 unsigned long nr_bits
, alloc_size
;
920 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
921 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
924 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
925 nm
= compat_alloc_user_space(alloc_size
);
926 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
932 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
937 /* Return effective policy for a VMA */
939 get_vma_policy(struct task_struct
*task
, struct vm_area_struct
*vma
, unsigned long addr
)
941 struct mempolicy
*pol
= task
->mempolicy
;
944 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
945 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
946 else if (vma
->vm_policy
&&
947 vma
->vm_policy
->policy
!= MPOL_DEFAULT
)
948 pol
= vma
->vm_policy
;
951 pol
= &default_policy
;
955 /* Return a zonelist representing a mempolicy */
956 static struct zonelist
*zonelist_policy(gfp_t gfp
, struct mempolicy
*policy
)
960 switch (policy
->policy
) {
962 nd
= policy
->v
.preferred_node
;
967 /* Lower zones don't get a policy applied */
968 /* Careful: current->mems_allowed might have moved */
969 if (gfp_zone(gfp
) >= policy_zone
)
970 if (cpuset_zonelist_valid_mems_allowed(policy
->v
.zonelist
))
971 return policy
->v
.zonelist
;
973 case MPOL_INTERLEAVE
: /* should not happen */
981 return NODE_DATA(nd
)->node_zonelists
+ gfp_zone(gfp
);
984 /* Do dynamic interleaving for a process */
985 static unsigned interleave_nodes(struct mempolicy
*policy
)
988 struct task_struct
*me
= current
;
991 next
= next_node(nid
, policy
->v
.nodes
);
992 if (next
>= MAX_NUMNODES
)
993 next
= first_node(policy
->v
.nodes
);
998 /* Do static interleaving for a VMA with known offset. */
999 static unsigned offset_il_node(struct mempolicy
*pol
,
1000 struct vm_area_struct
*vma
, unsigned long off
)
1002 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1003 unsigned target
= (unsigned)off
% nnodes
;
1009 nid
= next_node(nid
, pol
->v
.nodes
);
1011 } while (c
<= target
);
1015 /* Determine a node number for interleave */
1016 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1017 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1022 off
= vma
->vm_pgoff
;
1023 off
+= (addr
- vma
->vm_start
) >> shift
;
1024 return offset_il_node(pol
, vma
, off
);
1026 return interleave_nodes(pol
);
1029 /* Return a zonelist suitable for a huge page allocation. */
1030 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
)
1032 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1034 if (pol
->policy
== MPOL_INTERLEAVE
) {
1037 nid
= interleave_nid(pol
, vma
, addr
, HPAGE_SHIFT
);
1038 return NODE_DATA(nid
)->node_zonelists
+ gfp_zone(GFP_HIGHUSER
);
1040 return zonelist_policy(GFP_HIGHUSER
, pol
);
1043 /* Allocate a page in interleaved policy.
1044 Own path because it needs to do special accounting. */
1045 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1048 struct zonelist
*zl
;
1051 zl
= NODE_DATA(nid
)->node_zonelists
+ gfp_zone(gfp
);
1052 page
= __alloc_pages(gfp
, order
, zl
);
1053 if (page
&& page_zone(page
) == zl
->zones
[0]) {
1054 zone_pcp(zl
->zones
[0],get_cpu())->interleave_hit
++;
1061 * alloc_page_vma - Allocate a page for a VMA.
1064 * %GFP_USER user allocation.
1065 * %GFP_KERNEL kernel allocations,
1066 * %GFP_HIGHMEM highmem/user allocations,
1067 * %GFP_FS allocation should not call back into a file system.
1068 * %GFP_ATOMIC don't sleep.
1070 * @vma: Pointer to VMA or NULL if not available.
1071 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1073 * This function allocates a page from the kernel page pool and applies
1074 * a NUMA policy associated with the VMA or the current process.
1075 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1076 * mm_struct of the VMA to prevent it from going away. Should be used for
1077 * all allocations for pages that will be mapped into
1078 * user space. Returns NULL when no page can be allocated.
1080 * Should be called with the mm_sem of the vma hold.
1083 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1085 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1087 cpuset_update_current_mems_allowed();
1089 if (unlikely(pol
->policy
== MPOL_INTERLEAVE
)) {
1092 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1093 return alloc_page_interleave(gfp
, 0, nid
);
1095 return __alloc_pages(gfp
, 0, zonelist_policy(gfp
, pol
));
1099 * alloc_pages_current - Allocate pages.
1102 * %GFP_USER user allocation,
1103 * %GFP_KERNEL kernel allocation,
1104 * %GFP_HIGHMEM highmem allocation,
1105 * %GFP_FS don't call back into a file system.
1106 * %GFP_ATOMIC don't sleep.
1107 * @order: Power of two of allocation size in pages. 0 is a single page.
1109 * Allocate a page from the kernel page pool. When not in
1110 * interrupt context and apply the current process NUMA policy.
1111 * Returns NULL when no page can be allocated.
1113 * Don't call cpuset_update_current_mems_allowed() unless
1114 * 1) it's ok to take cpuset_sem (can WAIT), and
1115 * 2) allocating for current task (not interrupt).
1117 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1119 struct mempolicy
*pol
= current
->mempolicy
;
1121 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1122 cpuset_update_current_mems_allowed();
1123 if (!pol
|| in_interrupt())
1124 pol
= &default_policy
;
1125 if (pol
->policy
== MPOL_INTERLEAVE
)
1126 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1127 return __alloc_pages(gfp
, order
, zonelist_policy(gfp
, pol
));
1129 EXPORT_SYMBOL(alloc_pages_current
);
1131 /* Slow path of a mempolicy copy */
1132 struct mempolicy
*__mpol_copy(struct mempolicy
*old
)
1134 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1137 return ERR_PTR(-ENOMEM
);
1139 atomic_set(&new->refcnt
, 1);
1140 if (new->policy
== MPOL_BIND
) {
1141 int sz
= ksize(old
->v
.zonelist
);
1142 new->v
.zonelist
= kmalloc(sz
, SLAB_KERNEL
);
1143 if (!new->v
.zonelist
) {
1144 kmem_cache_free(policy_cache
, new);
1145 return ERR_PTR(-ENOMEM
);
1147 memcpy(new->v
.zonelist
, old
->v
.zonelist
, sz
);
1152 /* Slow path of a mempolicy comparison */
1153 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1157 if (a
->policy
!= b
->policy
)
1159 switch (a
->policy
) {
1162 case MPOL_INTERLEAVE
:
1163 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1164 case MPOL_PREFERRED
:
1165 return a
->v
.preferred_node
== b
->v
.preferred_node
;
1168 for (i
= 0; a
->v
.zonelist
->zones
[i
]; i
++)
1169 if (a
->v
.zonelist
->zones
[i
] != b
->v
.zonelist
->zones
[i
])
1171 return b
->v
.zonelist
->zones
[i
] == NULL
;
1179 /* Slow path of a mpol destructor. */
1180 void __mpol_free(struct mempolicy
*p
)
1182 if (!atomic_dec_and_test(&p
->refcnt
))
1184 if (p
->policy
== MPOL_BIND
)
1185 kfree(p
->v
.zonelist
);
1186 p
->policy
= MPOL_DEFAULT
;
1187 kmem_cache_free(policy_cache
, p
);
1191 * Shared memory backing store policy support.
1193 * Remember policies even when nobody has shared memory mapped.
1194 * The policies are kept in Red-Black tree linked from the inode.
1195 * They are protected by the sp->lock spinlock, which should be held
1196 * for any accesses to the tree.
1199 /* lookup first element intersecting start-end */
1200 /* Caller holds sp->lock */
1201 static struct sp_node
*
1202 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1204 struct rb_node
*n
= sp
->root
.rb_node
;
1207 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1209 if (start
>= p
->end
)
1211 else if (end
<= p
->start
)
1219 struct sp_node
*w
= NULL
;
1220 struct rb_node
*prev
= rb_prev(n
);
1223 w
= rb_entry(prev
, struct sp_node
, nd
);
1224 if (w
->end
<= start
)
1228 return rb_entry(n
, struct sp_node
, nd
);
1231 /* Insert a new shared policy into the list. */
1232 /* Caller holds sp->lock */
1233 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1235 struct rb_node
**p
= &sp
->root
.rb_node
;
1236 struct rb_node
*parent
= NULL
;
1241 nd
= rb_entry(parent
, struct sp_node
, nd
);
1242 if (new->start
< nd
->start
)
1244 else if (new->end
> nd
->end
)
1245 p
= &(*p
)->rb_right
;
1249 rb_link_node(&new->nd
, parent
, p
);
1250 rb_insert_color(&new->nd
, &sp
->root
);
1251 PDprintk("inserting %lx-%lx: %d\n", new->start
, new->end
,
1252 new->policy
? new->policy
->policy
: 0);
1255 /* Find shared policy intersecting idx */
1257 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1259 struct mempolicy
*pol
= NULL
;
1262 if (!sp
->root
.rb_node
)
1264 spin_lock(&sp
->lock
);
1265 sn
= sp_lookup(sp
, idx
, idx
+1);
1267 mpol_get(sn
->policy
);
1270 spin_unlock(&sp
->lock
);
1274 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1276 PDprintk("deleting %lx-l%x\n", n
->start
, n
->end
);
1277 rb_erase(&n
->nd
, &sp
->root
);
1278 mpol_free(n
->policy
);
1279 kmem_cache_free(sn_cache
, n
);
1283 sp_alloc(unsigned long start
, unsigned long end
, struct mempolicy
*pol
)
1285 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1296 /* Replace a policy range. */
1297 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1298 unsigned long end
, struct sp_node
*new)
1300 struct sp_node
*n
, *new2
= NULL
;
1303 spin_lock(&sp
->lock
);
1304 n
= sp_lookup(sp
, start
, end
);
1305 /* Take care of old policies in the same range. */
1306 while (n
&& n
->start
< end
) {
1307 struct rb_node
*next
= rb_next(&n
->nd
);
1308 if (n
->start
>= start
) {
1314 /* Old policy spanning whole new range. */
1317 spin_unlock(&sp
->lock
);
1318 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1324 sp_insert(sp
, new2
);
1332 n
= rb_entry(next
, struct sp_node
, nd
);
1336 spin_unlock(&sp
->lock
);
1338 mpol_free(new2
->policy
);
1339 kmem_cache_free(sn_cache
, new2
);
1344 int mpol_set_shared_policy(struct shared_policy
*info
,
1345 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1348 struct sp_node
*new = NULL
;
1349 unsigned long sz
= vma_pages(vma
);
1351 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1353 sz
, npol
? npol
->policy
: -1,
1354 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1357 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1361 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1363 kmem_cache_free(sn_cache
, new);
1367 /* Free a backing policy store on inode delete. */
1368 void mpol_free_shared_policy(struct shared_policy
*p
)
1371 struct rb_node
*next
;
1373 if (!p
->root
.rb_node
)
1375 spin_lock(&p
->lock
);
1376 next
= rb_first(&p
->root
);
1378 n
= rb_entry(next
, struct sp_node
, nd
);
1379 next
= rb_next(&n
->nd
);
1380 rb_erase(&n
->nd
, &p
->root
);
1381 mpol_free(n
->policy
);
1382 kmem_cache_free(sn_cache
, n
);
1384 spin_unlock(&p
->lock
);
1387 /* assumes fs == KERNEL_DS */
1388 void __init
numa_policy_init(void)
1390 policy_cache
= kmem_cache_create("numa_policy",
1391 sizeof(struct mempolicy
),
1392 0, SLAB_PANIC
, NULL
, NULL
);
1394 sn_cache
= kmem_cache_create("shared_policy_node",
1395 sizeof(struct sp_node
),
1396 0, SLAB_PANIC
, NULL
, NULL
);
1398 /* Set interleaving policy for system init. This way not all
1399 the data structures allocated at system boot end up in node zero. */
1401 if (do_set_mempolicy(MPOL_INTERLEAVE
, &node_online_map
))
1402 printk("numa_policy_init: interleaving failed\n");
1405 /* Reset policy of current process to default */
1406 void numa_default_policy(void)
1408 do_set_mempolicy(MPOL_DEFAULT
, NULL
);
1411 /* Migrate a policy to a different set of nodes */
1412 static void rebind_policy(struct mempolicy
*pol
, const nodemask_t
*old
,
1413 const nodemask_t
*new)
1420 switch (pol
->policy
) {
1423 case MPOL_INTERLEAVE
:
1424 nodes_remap(tmp
, pol
->v
.nodes
, *old
, *new);
1426 current
->il_next
= node_remap(current
->il_next
, *old
, *new);
1428 case MPOL_PREFERRED
:
1429 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
1435 struct zonelist
*zonelist
;
1438 for (z
= pol
->v
.zonelist
->zones
; *z
; z
++)
1439 node_set((*z
)->zone_pgdat
->node_id
, nodes
);
1440 nodes_remap(tmp
, nodes
, *old
, *new);
1443 zonelist
= bind_zonelist(&nodes
);
1445 /* If no mem, then zonelist is NULL and we keep old zonelist.
1446 * If that old zonelist has no remaining mems_allowed nodes,
1447 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1451 /* Good - got mem - substitute new zonelist */
1452 kfree(pol
->v
.zonelist
);
1453 pol
->v
.zonelist
= zonelist
;
1464 * Someone moved this task to different nodes. Fixup mempolicies.
1466 * TODO - fixup current->mm->vma and shmfs/tmpfs/hugetlbfs policies as well,
1467 * once we have a cpuset mechanism to mark which cpuset subtree is migrating.
1469 void numa_policy_rebind(const nodemask_t
*old
, const nodemask_t
*new)
1471 rebind_policy(current
->mempolicy
, old
, new);
1475 * Display pages allocated per node and memory policy via /proc.
1478 static const char *policy_types
[] = { "default", "prefer", "bind",
1482 * Convert a mempolicy into a string.
1483 * Returns the number of characters in buffer (if positive)
1484 * or an error (negative)
1486 static inline int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
1491 int mode
= pol
? pol
->policy
: MPOL_DEFAULT
;
1498 case MPOL_PREFERRED
:
1500 node_set(pol
->v
.preferred_node
, nodes
);
1504 get_zonemask(pol
, &nodes
);
1507 case MPOL_INTERLEAVE
:
1508 nodes
= pol
->v
.nodes
;
1516 l
= strlen(policy_types
[mode
]);
1517 if (buffer
+ maxlen
< p
+ l
+ 1)
1520 strcpy(p
, policy_types
[mode
]);
1523 if (!nodes_empty(nodes
)) {
1524 if (buffer
+ maxlen
< p
+ 2)
1527 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
1533 unsigned long pages
;
1535 unsigned long mapped
;
1536 unsigned long mapcount_max
;
1537 unsigned long node
[MAX_NUMNODES
];
1540 static void gather_stats(struct page
*page
, void *private)
1542 struct numa_maps
*md
= private;
1543 int count
= page_mapcount(page
);
1548 if (count
> md
->mapcount_max
)
1549 md
->mapcount_max
= count
;
1556 md
->node
[page_to_nid(page
)]++;
1560 int show_numa_map(struct seq_file
*m
, void *v
)
1562 struct task_struct
*task
= m
->private;
1563 struct vm_area_struct
*vma
= v
;
1564 struct numa_maps
*md
;
1571 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
1575 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
1576 &node_online_map
, MPOL_MF_STATS
, md
);
1579 mpol_to_str(buffer
, sizeof(buffer
),
1580 get_vma_policy(task
, vma
, vma
->vm_start
));
1582 seq_printf(m
, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1583 vma
->vm_start
, buffer
, md
->pages
,
1584 md
->mapped
, md
->mapcount_max
);
1587 seq_printf(m
," anon=%lu",md
->anon
);
1589 for_each_online_node(n
)
1591 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
1597 if (m
->count
< m
->size
)
1598 m
->version
= (vma
!= get_gate_vma(task
)) ? vma
->vm_start
: 0;