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
))
263 migrate_page_add(vma
, page
, private, flags
);
266 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
267 pte_unmap_unlock(orig_pte
, ptl
);
271 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
272 unsigned long addr
, unsigned long end
,
273 const nodemask_t
*nodes
, unsigned long flags
,
279 pmd
= pmd_offset(pud
, addr
);
281 next
= pmd_addr_end(addr
, end
);
282 if (pmd_none_or_clear_bad(pmd
))
284 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
287 } while (pmd
++, addr
= next
, addr
!= end
);
291 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
292 unsigned long addr
, unsigned long end
,
293 const nodemask_t
*nodes
, unsigned long flags
,
299 pud
= pud_offset(pgd
, addr
);
301 next
= pud_addr_end(addr
, end
);
302 if (pud_none_or_clear_bad(pud
))
304 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
307 } while (pud
++, addr
= next
, addr
!= end
);
311 static inline int check_pgd_range(struct vm_area_struct
*vma
,
312 unsigned long addr
, unsigned long end
,
313 const nodemask_t
*nodes
, unsigned long flags
,
319 pgd
= pgd_offset(vma
->vm_mm
, addr
);
321 next
= pgd_addr_end(addr
, end
);
322 if (pgd_none_or_clear_bad(pgd
))
324 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
327 } while (pgd
++, addr
= next
, addr
!= end
);
331 /* Check if a vma is migratable */
332 static inline int vma_migratable(struct vm_area_struct
*vma
)
334 if (vma
->vm_flags
& (
335 VM_LOCKED
|VM_IO
|VM_HUGETLB
|VM_PFNMAP
))
341 * Check if all pages in a range are on a set of nodes.
342 * If pagelist != NULL then isolate pages from the LRU and
343 * put them on the pagelist.
345 static struct vm_area_struct
*
346 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
347 const nodemask_t
*nodes
, unsigned long flags
, void *private)
350 struct vm_area_struct
*first
, *vma
, *prev
;
352 first
= find_vma(mm
, start
);
354 return ERR_PTR(-EFAULT
);
356 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
357 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
358 if (!vma
->vm_next
&& vma
->vm_end
< end
)
359 return ERR_PTR(-EFAULT
);
360 if (prev
&& prev
->vm_end
< vma
->vm_start
)
361 return ERR_PTR(-EFAULT
);
363 if (!is_vm_hugetlb_page(vma
) &&
364 ((flags
& MPOL_MF_STRICT
) ||
365 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
366 vma_migratable(vma
)))) {
367 unsigned long endvma
= vma
->vm_end
;
371 if (vma
->vm_start
> start
)
372 start
= vma
->vm_start
;
373 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
376 first
= ERR_PTR(err
);
385 /* Apply policy to a single VMA */
386 static int policy_vma(struct vm_area_struct
*vma
, struct mempolicy
*new)
389 struct mempolicy
*old
= vma
->vm_policy
;
391 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
392 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
393 vma
->vm_ops
, vma
->vm_file
,
394 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
396 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
)
397 err
= vma
->vm_ops
->set_policy(vma
, new);
400 vma
->vm_policy
= new;
406 /* Step 2: apply policy to a range and do splits. */
407 static int mbind_range(struct vm_area_struct
*vma
, unsigned long start
,
408 unsigned long end
, struct mempolicy
*new)
410 struct vm_area_struct
*next
;
414 for (; vma
&& vma
->vm_start
< end
; vma
= next
) {
416 if (vma
->vm_start
< start
)
417 err
= split_vma(vma
->vm_mm
, vma
, start
, 1);
418 if (!err
&& vma
->vm_end
> end
)
419 err
= split_vma(vma
->vm_mm
, vma
, end
, 0);
421 err
= policy_vma(vma
, new);
428 static int contextualize_policy(int mode
, nodemask_t
*nodes
)
433 /* Update current mems_allowed */
434 cpuset_update_current_mems_allowed();
435 /* Ignore nodes not set in current->mems_allowed */
436 cpuset_restrict_to_mems_allowed(nodes
->bits
);
437 return mpol_check_policy(mode
, nodes
);
440 static int swap_pages(struct list_head
*pagelist
)
446 n
= migrate_pages(pagelist
, NULL
, &moved
, &failed
);
447 putback_lru_pages(&failed
);
448 putback_lru_pages(&moved
);
453 long do_mbind(unsigned long start
, unsigned long len
,
454 unsigned long mode
, nodemask_t
*nmask
, unsigned long flags
)
456 struct vm_area_struct
*vma
;
457 struct mm_struct
*mm
= current
->mm
;
458 struct mempolicy
*new;
463 if ((flags
& ~(unsigned long)(MPOL_MF_STRICT
|
464 MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
467 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_RESOURCE
))
470 if (start
& ~PAGE_MASK
)
473 if (mode
== MPOL_DEFAULT
)
474 flags
&= ~MPOL_MF_STRICT
;
476 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
484 if (mpol_check_policy(mode
, nmask
))
487 new = mpol_new(mode
, nmask
);
492 * If we are using the default policy then operation
493 * on discontinuous address spaces is okay after all
496 flags
|= MPOL_MF_DISCONTIG_OK
;
498 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start
,start
+len
,
499 mode
,nodes_addr(nodes
)[0]);
501 down_write(&mm
->mmap_sem
);
502 vma
= check_range(mm
, start
, end
, nmask
,
503 flags
| MPOL_MF_INVERT
, &pagelist
);
509 err
= mbind_range(vma
, start
, end
, new);
510 if (!list_empty(&pagelist
))
511 nr_failed
= swap_pages(&pagelist
);
513 if (!err
&& nr_failed
&& (flags
& MPOL_MF_STRICT
))
516 if (!list_empty(&pagelist
))
517 putback_lru_pages(&pagelist
);
519 up_write(&mm
->mmap_sem
);
524 /* Set the process memory policy */
525 long do_set_mempolicy(int mode
, nodemask_t
*nodes
)
527 struct mempolicy
*new;
529 if (contextualize_policy(mode
, nodes
))
531 new = mpol_new(mode
, nodes
);
534 mpol_free(current
->mempolicy
);
535 current
->mempolicy
= new;
536 if (new && new->policy
== MPOL_INTERLEAVE
)
537 current
->il_next
= first_node(new->v
.nodes
);
541 /* Fill a zone bitmap for a policy */
542 static void get_zonemask(struct mempolicy
*p
, nodemask_t
*nodes
)
549 for (i
= 0; p
->v
.zonelist
->zones
[i
]; i
++)
550 node_set(p
->v
.zonelist
->zones
[i
]->zone_pgdat
->node_id
,
555 case MPOL_INTERLEAVE
:
559 /* or use current node instead of online map? */
560 if (p
->v
.preferred_node
< 0)
561 *nodes
= node_online_map
;
563 node_set(p
->v
.preferred_node
, *nodes
);
570 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
575 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
577 err
= page_to_nid(p
);
583 /* Retrieve NUMA policy */
584 long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
585 unsigned long addr
, unsigned long flags
)
588 struct mm_struct
*mm
= current
->mm
;
589 struct vm_area_struct
*vma
= NULL
;
590 struct mempolicy
*pol
= current
->mempolicy
;
592 cpuset_update_current_mems_allowed();
593 if (flags
& ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
))
595 if (flags
& MPOL_F_ADDR
) {
596 down_read(&mm
->mmap_sem
);
597 vma
= find_vma_intersection(mm
, addr
, addr
+1);
599 up_read(&mm
->mmap_sem
);
602 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
603 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
605 pol
= vma
->vm_policy
;
610 pol
= &default_policy
;
612 if (flags
& MPOL_F_NODE
) {
613 if (flags
& MPOL_F_ADDR
) {
614 err
= lookup_node(mm
, addr
);
618 } else if (pol
== current
->mempolicy
&&
619 pol
->policy
== MPOL_INTERLEAVE
) {
620 *policy
= current
->il_next
;
626 *policy
= pol
->policy
;
629 up_read(¤t
->mm
->mmap_sem
);
635 get_zonemask(pol
, nmask
);
639 up_read(¤t
->mm
->mmap_sem
);
644 * For now migrate_pages simply swaps out the pages from nodes that are in
645 * the source set but not in the target set. In the future, we would
646 * want a function that moves pages between the two nodesets in such
647 * a way as to preserve the physical layout as much as possible.
649 * Returns the number of page that could not be moved.
651 int do_migrate_pages(struct mm_struct
*mm
,
652 const nodemask_t
*from_nodes
, const nodemask_t
*to_nodes
, int flags
)
658 nodes_andnot(nodes
, *from_nodes
, *to_nodes
);
660 down_read(&mm
->mmap_sem
);
661 check_range(mm
, mm
->mmap
->vm_start
, TASK_SIZE
, &nodes
,
662 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
664 if (!list_empty(&pagelist
)) {
665 count
= swap_pages(&pagelist
);
666 putback_lru_pages(&pagelist
);
669 up_read(&mm
->mmap_sem
);
674 * User space interface with variable sized bitmaps for nodelists.
677 /* Copy a node mask from user space. */
678 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
679 unsigned long maxnode
)
682 unsigned long nlongs
;
683 unsigned long endmask
;
687 if (maxnode
== 0 || !nmask
)
690 nlongs
= BITS_TO_LONGS(maxnode
);
691 if ((maxnode
% BITS_PER_LONG
) == 0)
694 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
696 /* When the user specified more nodes than supported just check
697 if the non supported part is all zero. */
698 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
699 if (nlongs
> PAGE_SIZE
/sizeof(long))
701 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
703 if (get_user(t
, nmask
+ k
))
705 if (k
== nlongs
- 1) {
711 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
715 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
717 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
721 /* Copy a kernel node mask to user space */
722 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
725 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
726 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
729 if (copy
> PAGE_SIZE
)
731 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
735 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
738 asmlinkage
long sys_mbind(unsigned long start
, unsigned long len
,
740 unsigned long __user
*nmask
, unsigned long maxnode
,
746 err
= get_nodes(&nodes
, nmask
, maxnode
);
749 return do_mbind(start
, len
, mode
, &nodes
, flags
);
752 /* Set the process memory policy */
753 asmlinkage
long sys_set_mempolicy(int mode
, unsigned long __user
*nmask
,
754 unsigned long maxnode
)
759 if (mode
< 0 || mode
> MPOL_MAX
)
761 err
= get_nodes(&nodes
, nmask
, maxnode
);
764 return do_set_mempolicy(mode
, &nodes
);
767 /* Macro needed until Paul implements this function in kernel/cpusets.c */
768 #define cpuset_mems_allowed(task) node_online_map
770 asmlinkage
long sys_migrate_pages(pid_t pid
, unsigned long maxnode
,
771 const unsigned long __user
*old_nodes
,
772 const unsigned long __user
*new_nodes
)
774 struct mm_struct
*mm
;
775 struct task_struct
*task
;
778 nodemask_t task_nodes
;
781 err
= get_nodes(&old
, old_nodes
, maxnode
);
785 err
= get_nodes(&new, new_nodes
, maxnode
);
789 /* Find the mm_struct */
790 read_lock(&tasklist_lock
);
791 task
= pid
? find_task_by_pid(pid
) : current
;
793 read_unlock(&tasklist_lock
);
796 mm
= get_task_mm(task
);
797 read_unlock(&tasklist_lock
);
803 * Check if this process has the right to modify the specified
804 * process. The right exists if the process has administrative
805 * capabilities, superuser priviledges or the same
806 * userid as the target process.
808 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
809 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
810 !capable(CAP_SYS_ADMIN
)) {
815 task_nodes
= cpuset_mems_allowed(task
);
816 /* Is the user allowed to access the target nodes? */
817 if (!nodes_subset(new, task_nodes
) && !capable(CAP_SYS_ADMIN
)) {
822 err
= do_migrate_pages(mm
, &old
, &new, MPOL_MF_MOVE
);
829 /* Retrieve NUMA policy */
830 asmlinkage
long sys_get_mempolicy(int __user
*policy
,
831 unsigned long __user
*nmask
,
832 unsigned long maxnode
,
833 unsigned long addr
, unsigned long flags
)
838 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
841 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
846 if (policy
&& put_user(pval
, policy
))
850 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
857 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
858 compat_ulong_t __user
*nmask
,
859 compat_ulong_t maxnode
,
860 compat_ulong_t addr
, compat_ulong_t flags
)
863 unsigned long __user
*nm
= NULL
;
864 unsigned long nr_bits
, alloc_size
;
865 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
867 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
868 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
871 nm
= compat_alloc_user_space(alloc_size
);
873 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
876 err
= copy_from_user(bm
, nm
, alloc_size
);
877 /* ensure entire bitmap is zeroed */
878 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
879 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
885 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
886 compat_ulong_t maxnode
)
889 unsigned long __user
*nm
= NULL
;
890 unsigned long nr_bits
, alloc_size
;
891 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
893 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
894 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
897 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
898 nm
= compat_alloc_user_space(alloc_size
);
899 err
|= copy_to_user(nm
, bm
, alloc_size
);
905 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
908 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
909 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
910 compat_ulong_t maxnode
, compat_ulong_t flags
)
913 unsigned long __user
*nm
= NULL
;
914 unsigned long nr_bits
, alloc_size
;
917 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
918 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
921 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
922 nm
= compat_alloc_user_space(alloc_size
);
923 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
929 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
934 /* Return effective policy for a VMA */
936 get_vma_policy(struct task_struct
*task
, struct vm_area_struct
*vma
, unsigned long addr
)
938 struct mempolicy
*pol
= task
->mempolicy
;
941 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
942 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
943 else if (vma
->vm_policy
&&
944 vma
->vm_policy
->policy
!= MPOL_DEFAULT
)
945 pol
= vma
->vm_policy
;
948 pol
= &default_policy
;
952 /* Return a zonelist representing a mempolicy */
953 static struct zonelist
*zonelist_policy(gfp_t gfp
, struct mempolicy
*policy
)
957 switch (policy
->policy
) {
959 nd
= policy
->v
.preferred_node
;
964 /* Lower zones don't get a policy applied */
965 /* Careful: current->mems_allowed might have moved */
966 if (gfp_zone(gfp
) >= policy_zone
)
967 if (cpuset_zonelist_valid_mems_allowed(policy
->v
.zonelist
))
968 return policy
->v
.zonelist
;
970 case MPOL_INTERLEAVE
: /* should not happen */
978 return NODE_DATA(nd
)->node_zonelists
+ gfp_zone(gfp
);
981 /* Do dynamic interleaving for a process */
982 static unsigned interleave_nodes(struct mempolicy
*policy
)
985 struct task_struct
*me
= current
;
988 next
= next_node(nid
, policy
->v
.nodes
);
989 if (next
>= MAX_NUMNODES
)
990 next
= first_node(policy
->v
.nodes
);
995 /* Do static interleaving for a VMA with known offset. */
996 static unsigned offset_il_node(struct mempolicy
*pol
,
997 struct vm_area_struct
*vma
, unsigned long off
)
999 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1000 unsigned target
= (unsigned)off
% nnodes
;
1006 nid
= next_node(nid
, pol
->v
.nodes
);
1008 } while (c
<= target
);
1012 /* Determine a node number for interleave */
1013 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1014 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1019 off
= vma
->vm_pgoff
;
1020 off
+= (addr
- vma
->vm_start
) >> shift
;
1021 return offset_il_node(pol
, vma
, off
);
1023 return interleave_nodes(pol
);
1026 /* Return a zonelist suitable for a huge page allocation. */
1027 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
)
1029 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1031 if (pol
->policy
== MPOL_INTERLEAVE
) {
1034 nid
= interleave_nid(pol
, vma
, addr
, HPAGE_SHIFT
);
1035 return NODE_DATA(nid
)->node_zonelists
+ gfp_zone(GFP_HIGHUSER
);
1037 return zonelist_policy(GFP_HIGHUSER
, pol
);
1040 /* Allocate a page in interleaved policy.
1041 Own path because it needs to do special accounting. */
1042 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1045 struct zonelist
*zl
;
1048 zl
= NODE_DATA(nid
)->node_zonelists
+ gfp_zone(gfp
);
1049 page
= __alloc_pages(gfp
, order
, zl
);
1050 if (page
&& page_zone(page
) == zl
->zones
[0]) {
1051 zone_pcp(zl
->zones
[0],get_cpu())->interleave_hit
++;
1058 * alloc_page_vma - Allocate a page for a VMA.
1061 * %GFP_USER user allocation.
1062 * %GFP_KERNEL kernel allocations,
1063 * %GFP_HIGHMEM highmem/user allocations,
1064 * %GFP_FS allocation should not call back into a file system.
1065 * %GFP_ATOMIC don't sleep.
1067 * @vma: Pointer to VMA or NULL if not available.
1068 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1070 * This function allocates a page from the kernel page pool and applies
1071 * a NUMA policy associated with the VMA or the current process.
1072 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1073 * mm_struct of the VMA to prevent it from going away. Should be used for
1074 * all allocations for pages that will be mapped into
1075 * user space. Returns NULL when no page can be allocated.
1077 * Should be called with the mm_sem of the vma hold.
1080 alloc_page_vma(gfp_t gfp
, struct vm_area_struct
*vma
, unsigned long addr
)
1082 struct mempolicy
*pol
= get_vma_policy(current
, vma
, addr
);
1084 cpuset_update_current_mems_allowed();
1086 if (unlikely(pol
->policy
== MPOL_INTERLEAVE
)) {
1089 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
);
1090 return alloc_page_interleave(gfp
, 0, nid
);
1092 return __alloc_pages(gfp
, 0, zonelist_policy(gfp
, pol
));
1096 * alloc_pages_current - Allocate pages.
1099 * %GFP_USER user allocation,
1100 * %GFP_KERNEL kernel allocation,
1101 * %GFP_HIGHMEM highmem allocation,
1102 * %GFP_FS don't call back into a file system.
1103 * %GFP_ATOMIC don't sleep.
1104 * @order: Power of two of allocation size in pages. 0 is a single page.
1106 * Allocate a page from the kernel page pool. When not in
1107 * interrupt context and apply the current process NUMA policy.
1108 * Returns NULL when no page can be allocated.
1110 * Don't call cpuset_update_current_mems_allowed() unless
1111 * 1) it's ok to take cpuset_sem (can WAIT), and
1112 * 2) allocating for current task (not interrupt).
1114 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
1116 struct mempolicy
*pol
= current
->mempolicy
;
1118 if ((gfp
& __GFP_WAIT
) && !in_interrupt())
1119 cpuset_update_current_mems_allowed();
1120 if (!pol
|| in_interrupt())
1121 pol
= &default_policy
;
1122 if (pol
->policy
== MPOL_INTERLEAVE
)
1123 return alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
1124 return __alloc_pages(gfp
, order
, zonelist_policy(gfp
, pol
));
1126 EXPORT_SYMBOL(alloc_pages_current
);
1128 /* Slow path of a mempolicy copy */
1129 struct mempolicy
*__mpol_copy(struct mempolicy
*old
)
1131 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
1134 return ERR_PTR(-ENOMEM
);
1136 atomic_set(&new->refcnt
, 1);
1137 if (new->policy
== MPOL_BIND
) {
1138 int sz
= ksize(old
->v
.zonelist
);
1139 new->v
.zonelist
= kmalloc(sz
, SLAB_KERNEL
);
1140 if (!new->v
.zonelist
) {
1141 kmem_cache_free(policy_cache
, new);
1142 return ERR_PTR(-ENOMEM
);
1144 memcpy(new->v
.zonelist
, old
->v
.zonelist
, sz
);
1149 /* Slow path of a mempolicy comparison */
1150 int __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
1154 if (a
->policy
!= b
->policy
)
1156 switch (a
->policy
) {
1159 case MPOL_INTERLEAVE
:
1160 return nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
1161 case MPOL_PREFERRED
:
1162 return a
->v
.preferred_node
== b
->v
.preferred_node
;
1165 for (i
= 0; a
->v
.zonelist
->zones
[i
]; i
++)
1166 if (a
->v
.zonelist
->zones
[i
] != b
->v
.zonelist
->zones
[i
])
1168 return b
->v
.zonelist
->zones
[i
] == NULL
;
1176 /* Slow path of a mpol destructor. */
1177 void __mpol_free(struct mempolicy
*p
)
1179 if (!atomic_dec_and_test(&p
->refcnt
))
1181 if (p
->policy
== MPOL_BIND
)
1182 kfree(p
->v
.zonelist
);
1183 p
->policy
= MPOL_DEFAULT
;
1184 kmem_cache_free(policy_cache
, p
);
1188 * Shared memory backing store policy support.
1190 * Remember policies even when nobody has shared memory mapped.
1191 * The policies are kept in Red-Black tree linked from the inode.
1192 * They are protected by the sp->lock spinlock, which should be held
1193 * for any accesses to the tree.
1196 /* lookup first element intersecting start-end */
1197 /* Caller holds sp->lock */
1198 static struct sp_node
*
1199 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
1201 struct rb_node
*n
= sp
->root
.rb_node
;
1204 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
1206 if (start
>= p
->end
)
1208 else if (end
<= p
->start
)
1216 struct sp_node
*w
= NULL
;
1217 struct rb_node
*prev
= rb_prev(n
);
1220 w
= rb_entry(prev
, struct sp_node
, nd
);
1221 if (w
->end
<= start
)
1225 return rb_entry(n
, struct sp_node
, nd
);
1228 /* Insert a new shared policy into the list. */
1229 /* Caller holds sp->lock */
1230 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
1232 struct rb_node
**p
= &sp
->root
.rb_node
;
1233 struct rb_node
*parent
= NULL
;
1238 nd
= rb_entry(parent
, struct sp_node
, nd
);
1239 if (new->start
< nd
->start
)
1241 else if (new->end
> nd
->end
)
1242 p
= &(*p
)->rb_right
;
1246 rb_link_node(&new->nd
, parent
, p
);
1247 rb_insert_color(&new->nd
, &sp
->root
);
1248 PDprintk("inserting %lx-%lx: %d\n", new->start
, new->end
,
1249 new->policy
? new->policy
->policy
: 0);
1252 /* Find shared policy intersecting idx */
1254 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
1256 struct mempolicy
*pol
= NULL
;
1259 if (!sp
->root
.rb_node
)
1261 spin_lock(&sp
->lock
);
1262 sn
= sp_lookup(sp
, idx
, idx
+1);
1264 mpol_get(sn
->policy
);
1267 spin_unlock(&sp
->lock
);
1271 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
1273 PDprintk("deleting %lx-l%x\n", n
->start
, n
->end
);
1274 rb_erase(&n
->nd
, &sp
->root
);
1275 mpol_free(n
->policy
);
1276 kmem_cache_free(sn_cache
, n
);
1280 sp_alloc(unsigned long start
, unsigned long end
, struct mempolicy
*pol
)
1282 struct sp_node
*n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
1293 /* Replace a policy range. */
1294 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
1295 unsigned long end
, struct sp_node
*new)
1297 struct sp_node
*n
, *new2
= NULL
;
1300 spin_lock(&sp
->lock
);
1301 n
= sp_lookup(sp
, start
, end
);
1302 /* Take care of old policies in the same range. */
1303 while (n
&& n
->start
< end
) {
1304 struct rb_node
*next
= rb_next(&n
->nd
);
1305 if (n
->start
>= start
) {
1311 /* Old policy spanning whole new range. */
1314 spin_unlock(&sp
->lock
);
1315 new2
= sp_alloc(end
, n
->end
, n
->policy
);
1321 sp_insert(sp
, new2
);
1329 n
= rb_entry(next
, struct sp_node
, nd
);
1333 spin_unlock(&sp
->lock
);
1335 mpol_free(new2
->policy
);
1336 kmem_cache_free(sn_cache
, new2
);
1341 int mpol_set_shared_policy(struct shared_policy
*info
,
1342 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
1345 struct sp_node
*new = NULL
;
1346 unsigned long sz
= vma_pages(vma
);
1348 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1350 sz
, npol
? npol
->policy
: -1,
1351 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
1354 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
1358 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
1360 kmem_cache_free(sn_cache
, new);
1364 /* Free a backing policy store on inode delete. */
1365 void mpol_free_shared_policy(struct shared_policy
*p
)
1368 struct rb_node
*next
;
1370 if (!p
->root
.rb_node
)
1372 spin_lock(&p
->lock
);
1373 next
= rb_first(&p
->root
);
1375 n
= rb_entry(next
, struct sp_node
, nd
);
1376 next
= rb_next(&n
->nd
);
1377 rb_erase(&n
->nd
, &p
->root
);
1378 mpol_free(n
->policy
);
1379 kmem_cache_free(sn_cache
, n
);
1381 spin_unlock(&p
->lock
);
1384 /* assumes fs == KERNEL_DS */
1385 void __init
numa_policy_init(void)
1387 policy_cache
= kmem_cache_create("numa_policy",
1388 sizeof(struct mempolicy
),
1389 0, SLAB_PANIC
, NULL
, NULL
);
1391 sn_cache
= kmem_cache_create("shared_policy_node",
1392 sizeof(struct sp_node
),
1393 0, SLAB_PANIC
, NULL
, NULL
);
1395 /* Set interleaving policy for system init. This way not all
1396 the data structures allocated at system boot end up in node zero. */
1398 if (do_set_mempolicy(MPOL_INTERLEAVE
, &node_online_map
))
1399 printk("numa_policy_init: interleaving failed\n");
1402 /* Reset policy of current process to default */
1403 void numa_default_policy(void)
1405 do_set_mempolicy(MPOL_DEFAULT
, NULL
);
1408 /* Migrate a policy to a different set of nodes */
1409 static void rebind_policy(struct mempolicy
*pol
, const nodemask_t
*old
,
1410 const nodemask_t
*new)
1417 switch (pol
->policy
) {
1420 case MPOL_INTERLEAVE
:
1421 nodes_remap(tmp
, pol
->v
.nodes
, *old
, *new);
1423 current
->il_next
= node_remap(current
->il_next
, *old
, *new);
1425 case MPOL_PREFERRED
:
1426 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
1432 struct zonelist
*zonelist
;
1435 for (z
= pol
->v
.zonelist
->zones
; *z
; z
++)
1436 node_set((*z
)->zone_pgdat
->node_id
, nodes
);
1437 nodes_remap(tmp
, nodes
, *old
, *new);
1440 zonelist
= bind_zonelist(&nodes
);
1442 /* If no mem, then zonelist is NULL and we keep old zonelist.
1443 * If that old zonelist has no remaining mems_allowed nodes,
1444 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1448 /* Good - got mem - substitute new zonelist */
1449 kfree(pol
->v
.zonelist
);
1450 pol
->v
.zonelist
= zonelist
;
1461 * Someone moved this task to different nodes. Fixup mempolicies.
1463 * TODO - fixup current->mm->vma and shmfs/tmpfs/hugetlbfs policies as well,
1464 * once we have a cpuset mechanism to mark which cpuset subtree is migrating.
1466 void numa_policy_rebind(const nodemask_t
*old
, const nodemask_t
*new)
1468 rebind_policy(current
->mempolicy
, old
, new);
1472 * Display pages allocated per node and memory policy via /proc.
1475 static const char *policy_types
[] = { "default", "prefer", "bind",
1479 * Convert a mempolicy into a string.
1480 * Returns the number of characters in buffer (if positive)
1481 * or an error (negative)
1483 static inline int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
1488 int mode
= pol
? pol
->policy
: MPOL_DEFAULT
;
1495 case MPOL_PREFERRED
:
1497 node_set(pol
->v
.preferred_node
, nodes
);
1501 get_zonemask(pol
, &nodes
);
1504 case MPOL_INTERLEAVE
:
1505 nodes
= pol
->v
.nodes
;
1513 l
= strlen(policy_types
[mode
]);
1514 if (buffer
+ maxlen
< p
+ l
+ 1)
1517 strcpy(p
, policy_types
[mode
]);
1520 if (!nodes_empty(nodes
)) {
1521 if (buffer
+ maxlen
< p
+ 2)
1524 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);
1530 unsigned long pages
;
1532 unsigned long mapped
;
1533 unsigned long mapcount_max
;
1534 unsigned long node
[MAX_NUMNODES
];
1537 static void gather_stats(struct page
*page
, void *private)
1539 struct numa_maps
*md
= private;
1540 int count
= page_mapcount(page
);
1545 if (count
> md
->mapcount_max
)
1546 md
->mapcount_max
= count
;
1553 md
->node
[page_to_nid(page
)]++;
1557 int show_numa_map(struct seq_file
*m
, void *v
)
1559 struct task_struct
*task
= m
->private;
1560 struct vm_area_struct
*vma
= v
;
1561 struct numa_maps
*md
;
1568 md
= kzalloc(sizeof(struct numa_maps
), GFP_KERNEL
);
1572 check_pgd_range(vma
, vma
->vm_start
, vma
->vm_end
,
1573 &node_online_map
, MPOL_MF_STATS
, md
);
1576 mpol_to_str(buffer
, sizeof(buffer
),
1577 get_vma_policy(task
, vma
, vma
->vm_start
));
1579 seq_printf(m
, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1580 vma
->vm_start
, buffer
, md
->pages
,
1581 md
->mapped
, md
->mapcount_max
);
1584 seq_printf(m
," anon=%lu",md
->anon
);
1586 for_each_online_node(n
)
1588 seq_printf(m
, " N%d=%lu", n
, md
->node
[n
]);
1594 if (m
->count
< m
->size
)
1595 m
->version
= (vma
!= get_gate_vma(task
)) ? vma
->vm_start
: 0;
This page took 0.063617 seconds and 6 git commands to generate.