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 NUMA_NO_NODE 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/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
93 #include <linux/mmu_notifier.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
97 #include <linux/random.h>
102 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
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 static struct mempolicy default_policy
= {
116 .refcnt
= ATOMIC_INIT(1), /* never free it */
117 .mode
= MPOL_PREFERRED
,
118 .flags
= MPOL_F_LOCAL
,
121 static struct mempolicy preferred_node_policy
[MAX_NUMNODES
];
123 static struct mempolicy
*get_task_policy(struct task_struct
*p
)
125 struct mempolicy
*pol
= p
->mempolicy
;
129 node
= numa_node_id();
130 if (node
!= NUMA_NO_NODE
)
131 pol
= &preferred_node_policy
[node
];
133 /* preferred_node_policy is not initialised early in boot */
141 static const struct mempolicy_operations
{
142 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
144 * If read-side task has no lock to protect task->mempolicy, write-side
145 * task will rebind the task->mempolicy by two step. The first step is
146 * setting all the newly nodes, and the second step is cleaning all the
147 * disallowed nodes. In this way, we can avoid finding no node to alloc
149 * If we have a lock to protect task->mempolicy in read-side, we do
153 * MPOL_REBIND_ONCE - do rebind work at once
154 * MPOL_REBIND_STEP1 - set all the newly nodes
155 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
157 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
,
158 enum mpol_rebind_step step
);
159 } mpol_ops
[MPOL_MAX
];
161 /* Check that the nodemask contains at least one populated zone */
162 static int is_valid_nodemask(const nodemask_t
*nodemask
)
164 return nodes_intersects(*nodemask
, node_states
[N_MEMORY
]);
167 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
169 return pol
->flags
& MPOL_MODE_FLAGS
;
172 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
173 const nodemask_t
*rel
)
176 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
177 nodes_onto(*ret
, tmp
, *rel
);
180 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
182 if (nodes_empty(*nodes
))
184 pol
->v
.nodes
= *nodes
;
188 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
191 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
192 else if (nodes_empty(*nodes
))
193 return -EINVAL
; /* no allowed nodes */
195 pol
->v
.preferred_node
= first_node(*nodes
);
199 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
201 if (!is_valid_nodemask(nodes
))
203 pol
->v
.nodes
= *nodes
;
208 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
209 * any, for the new policy. mpol_new() has already validated the nodes
210 * parameter with respect to the policy mode and flags. But, we need to
211 * handle an empty nodemask with MPOL_PREFERRED here.
213 * Must be called holding task's alloc_lock to protect task's mems_allowed
214 * and mempolicy. May also be called holding the mmap_semaphore for write.
216 static int mpol_set_nodemask(struct mempolicy
*pol
,
217 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
221 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
225 nodes_and(nsc
->mask1
,
226 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
229 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
230 nodes
= NULL
; /* explicit local allocation */
232 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
233 mpol_relative_nodemask(&nsc
->mask2
, nodes
,&nsc
->mask1
);
235 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
237 if (mpol_store_user_nodemask(pol
))
238 pol
->w
.user_nodemask
= *nodes
;
240 pol
->w
.cpuset_mems_allowed
=
241 cpuset_current_mems_allowed
;
245 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
247 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
252 * This function just creates a new policy, does some check and simple
253 * initialization. You must invoke mpol_set_nodemask() to set nodes.
255 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
258 struct mempolicy
*policy
;
260 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
261 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
263 if (mode
== MPOL_DEFAULT
) {
264 if (nodes
&& !nodes_empty(*nodes
))
265 return ERR_PTR(-EINVAL
);
271 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
272 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
273 * All other modes require a valid pointer to a non-empty nodemask.
275 if (mode
== MPOL_PREFERRED
) {
276 if (nodes_empty(*nodes
)) {
277 if (((flags
& MPOL_F_STATIC_NODES
) ||
278 (flags
& MPOL_F_RELATIVE_NODES
)))
279 return ERR_PTR(-EINVAL
);
281 } else if (mode
== MPOL_LOCAL
) {
282 if (!nodes_empty(*nodes
))
283 return ERR_PTR(-EINVAL
);
284 mode
= MPOL_PREFERRED
;
285 } else if (nodes_empty(*nodes
))
286 return ERR_PTR(-EINVAL
);
287 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
289 return ERR_PTR(-ENOMEM
);
290 atomic_set(&policy
->refcnt
, 1);
292 policy
->flags
= flags
;
297 /* Slow path of a mpol destructor. */
298 void __mpol_put(struct mempolicy
*p
)
300 if (!atomic_dec_and_test(&p
->refcnt
))
302 kmem_cache_free(policy_cache
, p
);
305 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
,
306 enum mpol_rebind_step step
)
312 * MPOL_REBIND_ONCE - do rebind work at once
313 * MPOL_REBIND_STEP1 - set all the newly nodes
314 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
316 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
,
317 enum mpol_rebind_step step
)
321 if (pol
->flags
& MPOL_F_STATIC_NODES
)
322 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
323 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
324 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
327 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
330 if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP1
) {
331 nodes_remap(tmp
, pol
->v
.nodes
,
332 pol
->w
.cpuset_mems_allowed
, *nodes
);
333 pol
->w
.cpuset_mems_allowed
= step
? tmp
: *nodes
;
334 } else if (step
== MPOL_REBIND_STEP2
) {
335 tmp
= pol
->w
.cpuset_mems_allowed
;
336 pol
->w
.cpuset_mems_allowed
= *nodes
;
341 if (nodes_empty(tmp
))
344 if (step
== MPOL_REBIND_STEP1
)
345 nodes_or(pol
->v
.nodes
, pol
->v
.nodes
, tmp
);
346 else if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP2
)
351 if (!node_isset(current
->il_next
, tmp
)) {
352 current
->il_next
= next_node(current
->il_next
, tmp
);
353 if (current
->il_next
>= MAX_NUMNODES
)
354 current
->il_next
= first_node(tmp
);
355 if (current
->il_next
>= MAX_NUMNODES
)
356 current
->il_next
= numa_node_id();
360 static void mpol_rebind_preferred(struct mempolicy
*pol
,
361 const nodemask_t
*nodes
,
362 enum mpol_rebind_step step
)
366 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
367 int node
= first_node(pol
->w
.user_nodemask
);
369 if (node_isset(node
, *nodes
)) {
370 pol
->v
.preferred_node
= node
;
371 pol
->flags
&= ~MPOL_F_LOCAL
;
373 pol
->flags
|= MPOL_F_LOCAL
;
374 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
375 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
376 pol
->v
.preferred_node
= first_node(tmp
);
377 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
378 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
379 pol
->w
.cpuset_mems_allowed
,
381 pol
->w
.cpuset_mems_allowed
= *nodes
;
386 * mpol_rebind_policy - Migrate a policy to a different set of nodes
388 * If read-side task has no lock to protect task->mempolicy, write-side
389 * task will rebind the task->mempolicy by two step. The first step is
390 * setting all the newly nodes, and the second step is cleaning all the
391 * disallowed nodes. In this way, we can avoid finding no node to alloc
393 * If we have a lock to protect task->mempolicy in read-side, we do
397 * MPOL_REBIND_ONCE - do rebind work at once
398 * MPOL_REBIND_STEP1 - set all the newly nodes
399 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
401 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
,
402 enum mpol_rebind_step step
)
406 if (!mpol_store_user_nodemask(pol
) && step
== MPOL_REBIND_ONCE
&&
407 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
410 if (step
== MPOL_REBIND_STEP1
&& (pol
->flags
& MPOL_F_REBINDING
))
413 if (step
== MPOL_REBIND_STEP2
&& !(pol
->flags
& MPOL_F_REBINDING
))
416 if (step
== MPOL_REBIND_STEP1
)
417 pol
->flags
|= MPOL_F_REBINDING
;
418 else if (step
== MPOL_REBIND_STEP2
)
419 pol
->flags
&= ~MPOL_F_REBINDING
;
420 else if (step
>= MPOL_REBIND_NSTEP
)
423 mpol_ops
[pol
->mode
].rebind(pol
, newmask
, step
);
427 * Wrapper for mpol_rebind_policy() that just requires task
428 * pointer, and updates task mempolicy.
430 * Called with task's alloc_lock held.
433 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new,
434 enum mpol_rebind_step step
)
436 mpol_rebind_policy(tsk
->mempolicy
, new, step
);
440 * Rebind each vma in mm to new nodemask.
442 * Call holding a reference to mm. Takes mm->mmap_sem during call.
445 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
447 struct vm_area_struct
*vma
;
449 down_write(&mm
->mmap_sem
);
450 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
451 mpol_rebind_policy(vma
->vm_policy
, new, MPOL_REBIND_ONCE
);
452 up_write(&mm
->mmap_sem
);
455 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
457 .rebind
= mpol_rebind_default
,
459 [MPOL_INTERLEAVE
] = {
460 .create
= mpol_new_interleave
,
461 .rebind
= mpol_rebind_nodemask
,
464 .create
= mpol_new_preferred
,
465 .rebind
= mpol_rebind_preferred
,
468 .create
= mpol_new_bind
,
469 .rebind
= mpol_rebind_nodemask
,
473 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
474 unsigned long flags
);
476 /* Scan through pages checking if pages follow certain conditions. */
477 static int check_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
478 unsigned long addr
, unsigned long end
,
479 const nodemask_t
*nodes
, unsigned long flags
,
486 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
491 if (!pte_present(*pte
))
493 page
= vm_normal_page(vma
, addr
, *pte
);
497 * vm_normal_page() filters out zero pages, but there might
498 * still be PageReserved pages to skip, perhaps in a VDSO.
500 if (PageReserved(page
))
502 nid
= page_to_nid(page
);
503 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
506 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
507 migrate_page_add(page
, private, flags
);
510 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
511 pte_unmap_unlock(orig_pte
, ptl
);
515 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
516 unsigned long addr
, unsigned long end
,
517 const nodemask_t
*nodes
, unsigned long flags
,
523 pmd
= pmd_offset(pud
, addr
);
525 next
= pmd_addr_end(addr
, end
);
526 split_huge_page_pmd(vma
, addr
, pmd
);
527 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
529 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
532 } while (pmd
++, addr
= next
, addr
!= end
);
536 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
537 unsigned long addr
, unsigned long end
,
538 const nodemask_t
*nodes
, unsigned long flags
,
544 pud
= pud_offset(pgd
, addr
);
546 next
= pud_addr_end(addr
, end
);
547 if (pud_none_or_clear_bad(pud
))
549 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
552 } while (pud
++, addr
= next
, addr
!= end
);
556 static inline int check_pgd_range(struct vm_area_struct
*vma
,
557 unsigned long addr
, unsigned long end
,
558 const nodemask_t
*nodes
, unsigned long flags
,
564 pgd
= pgd_offset(vma
->vm_mm
, addr
);
566 next
= pgd_addr_end(addr
, end
);
567 if (pgd_none_or_clear_bad(pgd
))
569 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
572 } while (pgd
++, addr
= next
, addr
!= end
);
576 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
578 * This is used to mark a range of virtual addresses to be inaccessible.
579 * These are later cleared by a NUMA hinting fault. Depending on these
580 * faults, pages may be migrated for better NUMA placement.
582 * This is assuming that NUMA faults are handled using PROT_NONE. If
583 * an architecture makes a different choice, it will need further
584 * changes to the core.
586 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
587 unsigned long addr
, unsigned long end
)
590 BUILD_BUG_ON(_PAGE_NUMA
!= _PAGE_PROTNONE
);
592 nr_updated
= change_protection(vma
, addr
, end
, vma
->vm_page_prot
, 0, 1);
594 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
599 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
600 unsigned long addr
, unsigned long end
)
604 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
607 * Check if all pages in a range are on a set of nodes.
608 * If pagelist != NULL then isolate pages from the LRU and
609 * put them on the pagelist.
612 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
613 const nodemask_t
*nodes
, unsigned long flags
, void *private)
616 struct vm_area_struct
*vma
, *prev
;
618 vma
= find_vma(mm
, start
);
622 for (; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
623 unsigned long endvma
= vma
->vm_end
;
627 if (vma
->vm_start
> start
)
628 start
= vma
->vm_start
;
630 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
631 if (!vma
->vm_next
&& vma
->vm_end
< end
)
633 if (prev
&& prev
->vm_end
< vma
->vm_start
)
637 if (is_vm_hugetlb_page(vma
))
640 if (flags
& MPOL_MF_LAZY
) {
641 change_prot_numa(vma
, start
, endvma
);
645 if ((flags
& MPOL_MF_STRICT
) ||
646 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
647 vma_migratable(vma
))) {
649 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
661 * Apply policy to a single VMA
662 * This must be called with the mmap_sem held for writing.
664 static int vma_replace_policy(struct vm_area_struct
*vma
,
665 struct mempolicy
*pol
)
668 struct mempolicy
*old
;
669 struct mempolicy
*new;
671 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
672 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
673 vma
->vm_ops
, vma
->vm_file
,
674 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
680 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
681 err
= vma
->vm_ops
->set_policy(vma
, new);
686 old
= vma
->vm_policy
;
687 vma
->vm_policy
= new; /* protected by mmap_sem */
696 /* Step 2: apply policy to a range and do splits. */
697 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
698 unsigned long end
, struct mempolicy
*new_pol
)
700 struct vm_area_struct
*next
;
701 struct vm_area_struct
*prev
;
702 struct vm_area_struct
*vma
;
705 unsigned long vmstart
;
708 vma
= find_vma(mm
, start
);
709 if (!vma
|| vma
->vm_start
> start
)
713 if (start
> vma
->vm_start
)
716 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
718 vmstart
= max(start
, vma
->vm_start
);
719 vmend
= min(end
, vma
->vm_end
);
721 if (mpol_equal(vma_policy(vma
), new_pol
))
724 pgoff
= vma
->vm_pgoff
+
725 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
726 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
727 vma
->anon_vma
, vma
->vm_file
, pgoff
,
728 new_pol
, vma_get_anon_name(vma
));
732 if (mpol_equal(vma_policy(vma
), new_pol
))
734 /* vma_merge() joined vma && vma->next, case 8 */
737 if (vma
->vm_start
!= vmstart
) {
738 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
742 if (vma
->vm_end
!= vmend
) {
743 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
748 err
= vma_replace_policy(vma
, new_pol
);
758 * Update task->flags PF_MEMPOLICY bit: set iff non-default
759 * mempolicy. Allows more rapid checking of this (combined perhaps
760 * with other PF_* flag bits) on memory allocation hot code paths.
762 * If called from outside this file, the task 'p' should -only- be
763 * a newly forked child not yet visible on the task list, because
764 * manipulating the task flags of a visible task is not safe.
766 * The above limitation is why this routine has the funny name
767 * mpol_fix_fork_child_flag().
769 * It is also safe to call this with a task pointer of current,
770 * which the static wrapper mpol_set_task_struct_flag() does,
771 * for use within this file.
774 void mpol_fix_fork_child_flag(struct task_struct
*p
)
777 p
->flags
|= PF_MEMPOLICY
;
779 p
->flags
&= ~PF_MEMPOLICY
;
782 static void mpol_set_task_struct_flag(void)
784 mpol_fix_fork_child_flag(current
);
787 /* Set the process memory policy */
788 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
791 struct mempolicy
*new, *old
;
792 struct mm_struct
*mm
= current
->mm
;
793 NODEMASK_SCRATCH(scratch
);
799 new = mpol_new(mode
, flags
, nodes
);
805 * prevent changing our mempolicy while show_numa_maps()
807 * Note: do_set_mempolicy() can be called at init time
811 down_write(&mm
->mmap_sem
);
813 ret
= mpol_set_nodemask(new, nodes
, scratch
);
815 task_unlock(current
);
817 up_write(&mm
->mmap_sem
);
821 old
= current
->mempolicy
;
822 current
->mempolicy
= new;
823 mpol_set_task_struct_flag();
824 if (new && new->mode
== MPOL_INTERLEAVE
&&
825 nodes_weight(new->v
.nodes
))
826 current
->il_next
= first_node(new->v
.nodes
);
827 task_unlock(current
);
829 up_write(&mm
->mmap_sem
);
834 NODEMASK_SCRATCH_FREE(scratch
);
839 * Return nodemask for policy for get_mempolicy() query
841 * Called with task's alloc_lock held
843 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
846 if (p
== &default_policy
)
852 case MPOL_INTERLEAVE
:
856 if (!(p
->flags
& MPOL_F_LOCAL
))
857 node_set(p
->v
.preferred_node
, *nodes
);
858 /* else return empty node mask for local allocation */
865 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
870 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
872 err
= page_to_nid(p
);
878 /* Retrieve NUMA policy */
879 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
880 unsigned long addr
, unsigned long flags
)
883 struct mm_struct
*mm
= current
->mm
;
884 struct vm_area_struct
*vma
= NULL
;
885 struct mempolicy
*pol
= current
->mempolicy
;
888 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
891 if (flags
& MPOL_F_MEMS_ALLOWED
) {
892 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
894 *policy
= 0; /* just so it's initialized */
896 *nmask
= cpuset_current_mems_allowed
;
897 task_unlock(current
);
901 if (flags
& MPOL_F_ADDR
) {
903 * Do NOT fall back to task policy if the
904 * vma/shared policy at addr is NULL. We
905 * want to return MPOL_DEFAULT in this case.
907 down_read(&mm
->mmap_sem
);
908 vma
= find_vma_intersection(mm
, addr
, addr
+1);
910 up_read(&mm
->mmap_sem
);
913 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
914 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
916 pol
= vma
->vm_policy
;
921 pol
= &default_policy
; /* indicates default behavior */
923 if (flags
& MPOL_F_NODE
) {
924 if (flags
& MPOL_F_ADDR
) {
925 err
= lookup_node(mm
, addr
);
929 } else if (pol
== current
->mempolicy
&&
930 pol
->mode
== MPOL_INTERLEAVE
) {
931 *policy
= current
->il_next
;
937 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
940 * Internal mempolicy flags must be masked off before exposing
941 * the policy to userspace.
943 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
947 up_read(¤t
->mm
->mmap_sem
);
953 if (mpol_store_user_nodemask(pol
)) {
954 *nmask
= pol
->w
.user_nodemask
;
957 get_policy_nodemask(pol
, nmask
);
958 task_unlock(current
);
965 up_read(¤t
->mm
->mmap_sem
);
969 #ifdef CONFIG_MIGRATION
973 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
977 * Avoid migrating a page that is shared with others.
979 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
980 if (!isolate_lru_page(page
)) {
981 list_add_tail(&page
->lru
, pagelist
);
982 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
983 page_is_file_cache(page
));
988 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
990 return alloc_pages_exact_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
994 * Migrate pages from one node to a target node.
995 * Returns error or the number of pages not migrated.
997 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1001 LIST_HEAD(pagelist
);
1005 node_set(source
, nmask
);
1008 * This does not "check" the range but isolates all pages that
1009 * need migration. Between passing in the full user address
1010 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1012 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1013 check_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1014 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1016 if (!list_empty(&pagelist
)) {
1017 err
= migrate_pages(&pagelist
, new_node_page
, dest
,
1018 MIGRATE_SYNC
, MR_SYSCALL
);
1020 putback_lru_pages(&pagelist
);
1027 * Move pages between the two nodesets so as to preserve the physical
1028 * layout as much as possible.
1030 * Returns the number of page that could not be moved.
1032 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1033 const nodemask_t
*to
, int flags
)
1039 err
= migrate_prep();
1043 down_read(&mm
->mmap_sem
);
1045 err
= migrate_vmas(mm
, from
, to
, flags
);
1050 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1051 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1052 * bit in 'tmp', and return that <source, dest> pair for migration.
1053 * The pair of nodemasks 'to' and 'from' define the map.
1055 * If no pair of bits is found that way, fallback to picking some
1056 * pair of 'source' and 'dest' bits that are not the same. If the
1057 * 'source' and 'dest' bits are the same, this represents a node
1058 * that will be migrating to itself, so no pages need move.
1060 * If no bits are left in 'tmp', or if all remaining bits left
1061 * in 'tmp' correspond to the same bit in 'to', return false
1062 * (nothing left to migrate).
1064 * This lets us pick a pair of nodes to migrate between, such that
1065 * if possible the dest node is not already occupied by some other
1066 * source node, minimizing the risk of overloading the memory on a
1067 * node that would happen if we migrated incoming memory to a node
1068 * before migrating outgoing memory source that same node.
1070 * A single scan of tmp is sufficient. As we go, we remember the
1071 * most recent <s, d> pair that moved (s != d). If we find a pair
1072 * that not only moved, but what's better, moved to an empty slot
1073 * (d is not set in tmp), then we break out then, with that pair.
1074 * Otherwise when we finish scanning from_tmp, we at least have the
1075 * most recent <s, d> pair that moved. If we get all the way through
1076 * the scan of tmp without finding any node that moved, much less
1077 * moved to an empty node, then there is nothing left worth migrating.
1081 while (!nodes_empty(tmp
)) {
1086 for_each_node_mask(s
, tmp
) {
1089 * do_migrate_pages() tries to maintain the relative
1090 * node relationship of the pages established between
1091 * threads and memory areas.
1093 * However if the number of source nodes is not equal to
1094 * the number of destination nodes we can not preserve
1095 * this node relative relationship. In that case, skip
1096 * copying memory from a node that is in the destination
1099 * Example: [2,3,4] -> [3,4,5] moves everything.
1100 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1103 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1104 (node_isset(s
, *to
)))
1107 d
= node_remap(s
, *from
, *to
);
1111 source
= s
; /* Node moved. Memorize */
1114 /* dest not in remaining from nodes? */
1115 if (!node_isset(dest
, tmp
))
1121 node_clear(source
, tmp
);
1122 err
= migrate_to_node(mm
, source
, dest
, flags
);
1129 up_read(&mm
->mmap_sem
);
1137 * Allocate a new page for page migration based on vma policy.
1138 * Start by assuming the page is mapped by the same vma as contains @start.
1139 * Search forward from there, if not. N.B., this assumes that the
1140 * list of pages handed to migrate_pages()--which is how we get here--
1141 * is in virtual address order.
1143 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1145 struct vm_area_struct
*vma
;
1146 unsigned long uninitialized_var(address
);
1148 vma
= find_vma(current
->mm
, start
);
1150 address
= page_address_in_vma(page
, vma
);
1151 if (address
!= -EFAULT
)
1157 * if !vma, alloc_page_vma() will use task or system default policy
1159 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1163 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1164 unsigned long flags
)
1168 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1169 const nodemask_t
*to
, int flags
)
1174 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1180 static long do_mbind(unsigned long start
, unsigned long len
,
1181 unsigned short mode
, unsigned short mode_flags
,
1182 nodemask_t
*nmask
, unsigned long flags
)
1184 struct mm_struct
*mm
= current
->mm
;
1185 struct mempolicy
*new;
1188 LIST_HEAD(pagelist
);
1190 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1192 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1195 if (start
& ~PAGE_MASK
)
1198 if (mode
== MPOL_DEFAULT
)
1199 flags
&= ~MPOL_MF_STRICT
;
1201 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1209 new = mpol_new(mode
, mode_flags
, nmask
);
1211 return PTR_ERR(new);
1213 if (flags
& MPOL_MF_LAZY
)
1214 new->flags
|= MPOL_F_MOF
;
1217 * If we are using the default policy then operation
1218 * on discontinuous address spaces is okay after all
1221 flags
|= MPOL_MF_DISCONTIG_OK
;
1223 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1224 start
, start
+ len
, mode
, mode_flags
,
1225 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1227 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1229 err
= migrate_prep();
1234 NODEMASK_SCRATCH(scratch
);
1236 down_write(&mm
->mmap_sem
);
1238 err
= mpol_set_nodemask(new, nmask
, scratch
);
1239 task_unlock(current
);
1241 up_write(&mm
->mmap_sem
);
1244 NODEMASK_SCRATCH_FREE(scratch
);
1249 err
= check_range(mm
, start
, end
, nmask
,
1250 flags
| MPOL_MF_INVERT
, &pagelist
);
1252 err
= mbind_range(mm
, start
, end
, new);
1257 if (!list_empty(&pagelist
)) {
1258 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1259 nr_failed
= migrate_pages(&pagelist
, new_page
,
1260 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1262 putback_lru_pages(&pagelist
);
1265 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1268 putback_lru_pages(&pagelist
);
1270 up_write(&mm
->mmap_sem
);
1277 * User space interface with variable sized bitmaps for nodelists.
1280 /* Copy a node mask from user space. */
1281 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1282 unsigned long maxnode
)
1285 unsigned long nlongs
;
1286 unsigned long endmask
;
1289 nodes_clear(*nodes
);
1290 if (maxnode
== 0 || !nmask
)
1292 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1295 nlongs
= BITS_TO_LONGS(maxnode
);
1296 if ((maxnode
% BITS_PER_LONG
) == 0)
1299 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1301 /* When the user specified more nodes than supported just check
1302 if the non supported part is all zero. */
1303 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1304 if (nlongs
> PAGE_SIZE
/sizeof(long))
1306 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1308 if (get_user(t
, nmask
+ k
))
1310 if (k
== nlongs
- 1) {
1316 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1320 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1322 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1326 /* Copy a kernel node mask to user space */
1327 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1330 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1331 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1333 if (copy
> nbytes
) {
1334 if (copy
> PAGE_SIZE
)
1336 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1340 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1343 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1344 unsigned long, mode
, unsigned long __user
*, nmask
,
1345 unsigned long, maxnode
, unsigned, flags
)
1349 unsigned short mode_flags
;
1351 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1352 mode
&= ~MPOL_MODE_FLAGS
;
1353 if (mode
>= MPOL_MAX
)
1355 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1356 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1358 err
= get_nodes(&nodes
, nmask
, maxnode
);
1361 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1364 /* Set the process memory policy */
1365 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, unsigned long __user
*, nmask
,
1366 unsigned long, maxnode
)
1370 unsigned short flags
;
1372 flags
= mode
& MPOL_MODE_FLAGS
;
1373 mode
&= ~MPOL_MODE_FLAGS
;
1374 if ((unsigned int)mode
>= MPOL_MAX
)
1376 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1378 err
= get_nodes(&nodes
, nmask
, maxnode
);
1381 return do_set_mempolicy(mode
, flags
, &nodes
);
1384 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1385 const unsigned long __user
*, old_nodes
,
1386 const unsigned long __user
*, new_nodes
)
1388 const struct cred
*cred
= current_cred(), *tcred
;
1389 struct mm_struct
*mm
= NULL
;
1390 struct task_struct
*task
;
1391 nodemask_t task_nodes
;
1395 NODEMASK_SCRATCH(scratch
);
1400 old
= &scratch
->mask1
;
1401 new = &scratch
->mask2
;
1403 err
= get_nodes(old
, old_nodes
, maxnode
);
1407 err
= get_nodes(new, new_nodes
, maxnode
);
1411 /* Find the mm_struct */
1413 task
= pid
? find_task_by_vpid(pid
) : current
;
1419 get_task_struct(task
);
1424 * Check if this process has the right to modify the specified
1425 * process. The right exists if the process has administrative
1426 * capabilities, superuser privileges or the same
1427 * userid as the target process.
1429 tcred
= __task_cred(task
);
1430 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1431 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1432 !capable(CAP_SYS_NICE
)) {
1439 task_nodes
= cpuset_mems_allowed(task
);
1440 /* Is the user allowed to access the target nodes? */
1441 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1446 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1451 err
= security_task_movememory(task
);
1455 mm
= get_task_mm(task
);
1456 put_task_struct(task
);
1463 err
= do_migrate_pages(mm
, old
, new,
1464 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1468 NODEMASK_SCRATCH_FREE(scratch
);
1473 put_task_struct(task
);
1479 /* Retrieve NUMA policy */
1480 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1481 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1482 unsigned long, addr
, unsigned long, flags
)
1485 int uninitialized_var(pval
);
1488 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1491 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1496 if (policy
&& put_user(pval
, policy
))
1500 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1505 #ifdef CONFIG_COMPAT
1507 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1508 compat_ulong_t __user
*nmask
,
1509 compat_ulong_t maxnode
,
1510 compat_ulong_t addr
, compat_ulong_t flags
)
1513 unsigned long __user
*nm
= NULL
;
1514 unsigned long nr_bits
, alloc_size
;
1515 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1517 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1518 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1521 nm
= compat_alloc_user_space(alloc_size
);
1523 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1525 if (!err
&& nmask
) {
1526 unsigned long copy_size
;
1527 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1528 err
= copy_from_user(bm
, nm
, copy_size
);
1529 /* ensure entire bitmap is zeroed */
1530 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1531 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1537 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1538 compat_ulong_t maxnode
)
1540 unsigned long __user
*nm
= NULL
;
1541 unsigned long nr_bits
, alloc_size
;
1542 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1544 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1545 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1548 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1550 nm
= compat_alloc_user_space(alloc_size
);
1551 if (copy_to_user(nm
, bm
, alloc_size
))
1555 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1558 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1559 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1560 compat_ulong_t maxnode
, compat_ulong_t flags
)
1562 unsigned long __user
*nm
= NULL
;
1563 unsigned long nr_bits
, alloc_size
;
1566 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1567 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1570 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1572 nm
= compat_alloc_user_space(alloc_size
);
1573 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1577 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1583 * get_vma_policy(@task, @vma, @addr)
1584 * @task - task for fallback if vma policy == default
1585 * @vma - virtual memory area whose policy is sought
1586 * @addr - address in @vma for shared policy lookup
1588 * Returns effective policy for a VMA at specified address.
1589 * Falls back to @task or system default policy, as necessary.
1590 * Current or other task's task mempolicy and non-shared vma policies must be
1591 * protected by task_lock(task) by the caller.
1592 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1593 * count--added by the get_policy() vm_op, as appropriate--to protect against
1594 * freeing by another task. It is the caller's responsibility to free the
1595 * extra reference for shared policies.
1597 struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1598 struct vm_area_struct
*vma
, unsigned long addr
)
1600 struct mempolicy
*pol
= get_task_policy(task
);
1603 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1604 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1608 } else if (vma
->vm_policy
) {
1609 pol
= vma
->vm_policy
;
1612 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1613 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1614 * count on these policies which will be dropped by
1615 * mpol_cond_put() later
1617 if (mpol_needs_cond_ref(pol
))
1622 pol
= &default_policy
;
1626 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1628 enum zone_type dynamic_policy_zone
= policy_zone
;
1630 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1633 * if policy->v.nodes has movable memory only,
1634 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1636 * policy->v.nodes is intersect with node_states[N_MEMORY].
1637 * so if the following test faile, it implies
1638 * policy->v.nodes has movable memory only.
1640 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1641 dynamic_policy_zone
= ZONE_MOVABLE
;
1643 return zone
>= dynamic_policy_zone
;
1647 * Return a nodemask representing a mempolicy for filtering nodes for
1650 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1652 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1653 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1654 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1655 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1656 return &policy
->v
.nodes
;
1661 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1662 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1665 switch (policy
->mode
) {
1666 case MPOL_PREFERRED
:
1667 if (!(policy
->flags
& MPOL_F_LOCAL
))
1668 nd
= policy
->v
.preferred_node
;
1672 * Normally, MPOL_BIND allocations are node-local within the
1673 * allowed nodemask. However, if __GFP_THISNODE is set and the
1674 * current node isn't part of the mask, we use the zonelist for
1675 * the first node in the mask instead.
1677 if (unlikely(gfp
& __GFP_THISNODE
) &&
1678 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1679 nd
= first_node(policy
->v
.nodes
);
1684 return node_zonelist(nd
, gfp
);
1687 /* Do dynamic interleaving for a process */
1688 static unsigned interleave_nodes(struct mempolicy
*policy
)
1691 struct task_struct
*me
= current
;
1694 next
= next_node(nid
, policy
->v
.nodes
);
1695 if (next
>= MAX_NUMNODES
)
1696 next
= first_node(policy
->v
.nodes
);
1697 if (next
< MAX_NUMNODES
)
1703 * Depending on the memory policy provide a node from which to allocate the
1705 * @policy must be protected by freeing by the caller. If @policy is
1706 * the current task's mempolicy, this protection is implicit, as only the
1707 * task can change it's policy. The system default policy requires no
1710 unsigned slab_node(void)
1712 struct mempolicy
*policy
;
1715 return numa_node_id();
1717 policy
= current
->mempolicy
;
1718 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1719 return numa_node_id();
1721 switch (policy
->mode
) {
1722 case MPOL_PREFERRED
:
1724 * handled MPOL_F_LOCAL above
1726 return policy
->v
.preferred_node
;
1728 case MPOL_INTERLEAVE
:
1729 return interleave_nodes(policy
);
1733 * Follow bind policy behavior and start allocation at the
1736 struct zonelist
*zonelist
;
1738 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1739 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1740 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1743 return zone
? zone
->node
: numa_node_id();
1751 /* Do static interleaving for a VMA with known offset. */
1752 static unsigned offset_il_node(struct mempolicy
*pol
,
1753 struct vm_area_struct
*vma
, unsigned long off
)
1755 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1761 return numa_node_id();
1762 target
= (unsigned int)off
% nnodes
;
1765 nid
= next_node(nid
, pol
->v
.nodes
);
1767 } while (c
<= target
);
1771 /* Determine a node number for interleave */
1772 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1773 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1779 * for small pages, there is no difference between
1780 * shift and PAGE_SHIFT, so the bit-shift is safe.
1781 * for huge pages, since vm_pgoff is in units of small
1782 * pages, we need to shift off the always 0 bits to get
1785 BUG_ON(shift
< PAGE_SHIFT
);
1786 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1787 off
+= (addr
- vma
->vm_start
) >> shift
;
1788 return offset_il_node(pol
, vma
, off
);
1790 return interleave_nodes(pol
);
1794 * Return the bit number of a random bit set in the nodemask.
1795 * (returns -1 if nodemask is empty)
1797 int node_random(const nodemask_t
*maskp
)
1801 w
= nodes_weight(*maskp
);
1803 bit
= bitmap_ord_to_pos(maskp
->bits
,
1804 get_random_int() % w
, MAX_NUMNODES
);
1808 #ifdef CONFIG_HUGETLBFS
1810 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1811 * @vma = virtual memory area whose policy is sought
1812 * @addr = address in @vma for shared policy lookup and interleave policy
1813 * @gfp_flags = for requested zone
1814 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1815 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1817 * Returns a zonelist suitable for a huge page allocation and a pointer
1818 * to the struct mempolicy for conditional unref after allocation.
1819 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1820 * @nodemask for filtering the zonelist.
1822 * Must be protected by get_mems_allowed()
1824 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1825 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1826 nodemask_t
**nodemask
)
1828 struct zonelist
*zl
;
1830 *mpol
= get_vma_policy(current
, vma
, addr
);
1831 *nodemask
= NULL
; /* assume !MPOL_BIND */
1833 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1834 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1835 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1837 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1838 if ((*mpol
)->mode
== MPOL_BIND
)
1839 *nodemask
= &(*mpol
)->v
.nodes
;
1845 * init_nodemask_of_mempolicy
1847 * If the current task's mempolicy is "default" [NULL], return 'false'
1848 * to indicate default policy. Otherwise, extract the policy nodemask
1849 * for 'bind' or 'interleave' policy into the argument nodemask, or
1850 * initialize the argument nodemask to contain the single node for
1851 * 'preferred' or 'local' policy and return 'true' to indicate presence
1852 * of non-default mempolicy.
1854 * We don't bother with reference counting the mempolicy [mpol_get/put]
1855 * because the current task is examining it's own mempolicy and a task's
1856 * mempolicy is only ever changed by the task itself.
1858 * N.B., it is the caller's responsibility to free a returned nodemask.
1860 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1862 struct mempolicy
*mempolicy
;
1865 if (!(mask
&& current
->mempolicy
))
1869 mempolicy
= current
->mempolicy
;
1870 switch (mempolicy
->mode
) {
1871 case MPOL_PREFERRED
:
1872 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1873 nid
= numa_node_id();
1875 nid
= mempolicy
->v
.preferred_node
;
1876 init_nodemask_of_node(mask
, nid
);
1881 case MPOL_INTERLEAVE
:
1882 *mask
= mempolicy
->v
.nodes
;
1888 task_unlock(current
);
1895 * mempolicy_nodemask_intersects
1897 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1898 * policy. Otherwise, check for intersection between mask and the policy
1899 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1900 * policy, always return true since it may allocate elsewhere on fallback.
1902 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1904 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1905 const nodemask_t
*mask
)
1907 struct mempolicy
*mempolicy
;
1913 mempolicy
= tsk
->mempolicy
;
1917 switch (mempolicy
->mode
) {
1918 case MPOL_PREFERRED
:
1920 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1921 * allocate from, they may fallback to other nodes when oom.
1922 * Thus, it's possible for tsk to have allocated memory from
1927 case MPOL_INTERLEAVE
:
1928 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1938 /* Allocate a page in interleaved policy.
1939 Own path because it needs to do special accounting. */
1940 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1943 struct zonelist
*zl
;
1946 zl
= node_zonelist(nid
, gfp
);
1947 page
= __alloc_pages(gfp
, order
, zl
);
1948 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1949 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1954 * alloc_pages_vma - Allocate a page for a VMA.
1957 * %GFP_USER user allocation.
1958 * %GFP_KERNEL kernel allocations,
1959 * %GFP_HIGHMEM highmem/user allocations,
1960 * %GFP_FS allocation should not call back into a file system.
1961 * %GFP_ATOMIC don't sleep.
1963 * @order:Order of the GFP allocation.
1964 * @vma: Pointer to VMA or NULL if not available.
1965 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1967 * This function allocates a page from the kernel page pool and applies
1968 * a NUMA policy associated with the VMA or the current process.
1969 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1970 * mm_struct of the VMA to prevent it from going away. Should be used for
1971 * all allocations for pages that will be mapped into
1972 * user space. Returns NULL when no page can be allocated.
1974 * Should be called with the mm_sem of the vma hold.
1977 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1978 unsigned long addr
, int node
)
1980 struct mempolicy
*pol
;
1982 unsigned int cpuset_mems_cookie
;
1985 pol
= get_vma_policy(current
, vma
, addr
);
1986 cpuset_mems_cookie
= get_mems_allowed();
1988 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
1991 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
1993 page
= alloc_page_interleave(gfp
, order
, nid
);
1994 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
1999 page
= __alloc_pages_nodemask(gfp
, order
,
2000 policy_zonelist(gfp
, pol
, node
),
2001 policy_nodemask(gfp
, pol
));
2002 if (unlikely(mpol_needs_cond_ref(pol
)))
2004 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2010 * alloc_pages_current - Allocate pages.
2013 * %GFP_USER user allocation,
2014 * %GFP_KERNEL kernel allocation,
2015 * %GFP_HIGHMEM highmem allocation,
2016 * %GFP_FS don't call back into a file system.
2017 * %GFP_ATOMIC don't sleep.
2018 * @order: Power of two of allocation size in pages. 0 is a single page.
2020 * Allocate a page from the kernel page pool. When not in
2021 * interrupt context and apply the current process NUMA policy.
2022 * Returns NULL when no page can be allocated.
2024 * Don't call cpuset_update_task_memory_state() unless
2025 * 1) it's ok to take cpuset_sem (can WAIT), and
2026 * 2) allocating for current task (not interrupt).
2028 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2030 struct mempolicy
*pol
= get_task_policy(current
);
2032 unsigned int cpuset_mems_cookie
;
2034 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
2035 pol
= &default_policy
;
2038 cpuset_mems_cookie
= get_mems_allowed();
2041 * No reference counting needed for current->mempolicy
2042 * nor system default_policy
2044 if (pol
->mode
== MPOL_INTERLEAVE
)
2045 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2047 page
= __alloc_pages_nodemask(gfp
, order
,
2048 policy_zonelist(gfp
, pol
, numa_node_id()),
2049 policy_nodemask(gfp
, pol
));
2051 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2056 EXPORT_SYMBOL(alloc_pages_current
);
2059 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2060 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2061 * with the mems_allowed returned by cpuset_mems_allowed(). This
2062 * keeps mempolicies cpuset relative after its cpuset moves. See
2063 * further kernel/cpuset.c update_nodemask().
2065 * current's mempolicy may be rebinded by the other task(the task that changes
2066 * cpuset's mems), so we needn't do rebind work for current task.
2069 /* Slow path of a mempolicy duplicate */
2070 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2072 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2075 return ERR_PTR(-ENOMEM
);
2077 /* task's mempolicy is protected by alloc_lock */
2078 if (old
== current
->mempolicy
) {
2081 task_unlock(current
);
2085 if (current_cpuset_is_being_rebound()) {
2086 nodemask_t mems
= cpuset_mems_allowed(current
);
2087 if (new->flags
& MPOL_F_REBINDING
)
2088 mpol_rebind_policy(new, &mems
, MPOL_REBIND_STEP2
);
2090 mpol_rebind_policy(new, &mems
, MPOL_REBIND_ONCE
);
2092 atomic_set(&new->refcnt
, 1);
2096 /* Slow path of a mempolicy comparison */
2097 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2101 if (a
->mode
!= b
->mode
)
2103 if (a
->flags
!= b
->flags
)
2105 if (mpol_store_user_nodemask(a
))
2106 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2112 case MPOL_INTERLEAVE
:
2113 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2114 case MPOL_PREFERRED
:
2115 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2123 * Shared memory backing store policy support.
2125 * Remember policies even when nobody has shared memory mapped.
2126 * The policies are kept in Red-Black tree linked from the inode.
2127 * They are protected by the sp->lock spinlock, which should be held
2128 * for any accesses to the tree.
2131 /* lookup first element intersecting start-end */
2132 /* Caller holds sp->lock */
2133 static struct sp_node
*
2134 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2136 struct rb_node
*n
= sp
->root
.rb_node
;
2139 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2141 if (start
>= p
->end
)
2143 else if (end
<= p
->start
)
2151 struct sp_node
*w
= NULL
;
2152 struct rb_node
*prev
= rb_prev(n
);
2155 w
= rb_entry(prev
, struct sp_node
, nd
);
2156 if (w
->end
<= start
)
2160 return rb_entry(n
, struct sp_node
, nd
);
2163 /* Insert a new shared policy into the list. */
2164 /* Caller holds sp->lock */
2165 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2167 struct rb_node
**p
= &sp
->root
.rb_node
;
2168 struct rb_node
*parent
= NULL
;
2173 nd
= rb_entry(parent
, struct sp_node
, nd
);
2174 if (new->start
< nd
->start
)
2176 else if (new->end
> nd
->end
)
2177 p
= &(*p
)->rb_right
;
2181 rb_link_node(&new->nd
, parent
, p
);
2182 rb_insert_color(&new->nd
, &sp
->root
);
2183 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2184 new->policy
? new->policy
->mode
: 0);
2187 /* Find shared policy intersecting idx */
2189 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2191 struct mempolicy
*pol
= NULL
;
2194 if (!sp
->root
.rb_node
)
2196 spin_lock(&sp
->lock
);
2197 sn
= sp_lookup(sp
, idx
, idx
+1);
2199 mpol_get(sn
->policy
);
2202 spin_unlock(&sp
->lock
);
2206 static void sp_free(struct sp_node
*n
)
2208 mpol_put(n
->policy
);
2209 kmem_cache_free(sn_cache
, n
);
2213 * mpol_misplaced - check whether current page node is valid in policy
2215 * @page - page to be checked
2216 * @vma - vm area where page mapped
2217 * @addr - virtual address where page mapped
2219 * Lookup current policy node id for vma,addr and "compare to" page's
2223 * -1 - not misplaced, page is in the right node
2224 * node - node id where the page should be
2226 * Policy determination "mimics" alloc_page_vma().
2227 * Called from fault path where we know the vma and faulting address.
2229 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2231 struct mempolicy
*pol
;
2233 int curnid
= page_to_nid(page
);
2234 unsigned long pgoff
;
2240 pol
= get_vma_policy(current
, vma
, addr
);
2241 if (!(pol
->flags
& MPOL_F_MOF
))
2244 switch (pol
->mode
) {
2245 case MPOL_INTERLEAVE
:
2246 BUG_ON(addr
>= vma
->vm_end
);
2247 BUG_ON(addr
< vma
->vm_start
);
2249 pgoff
= vma
->vm_pgoff
;
2250 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2251 polnid
= offset_il_node(pol
, vma
, pgoff
);
2254 case MPOL_PREFERRED
:
2255 if (pol
->flags
& MPOL_F_LOCAL
)
2256 polnid
= numa_node_id();
2258 polnid
= pol
->v
.preferred_node
;
2263 * allows binding to multiple nodes.
2264 * use current page if in policy nodemask,
2265 * else select nearest allowed node, if any.
2266 * If no allowed nodes, use current [!misplaced].
2268 if (node_isset(curnid
, pol
->v
.nodes
))
2270 (void)first_zones_zonelist(
2271 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2272 gfp_zone(GFP_HIGHUSER
),
2273 &pol
->v
.nodes
, &zone
);
2274 polnid
= zone
->node
;
2281 /* Migrate the page towards the node whose CPU is referencing it */
2282 if (pol
->flags
& MPOL_F_MORON
) {
2285 polnid
= numa_node_id();
2288 * Multi-stage node selection is used in conjunction
2289 * with a periodic migration fault to build a temporal
2290 * task<->page relation. By using a two-stage filter we
2291 * remove short/unlikely relations.
2293 * Using P(p) ~ n_p / n_t as per frequentist
2294 * probability, we can equate a task's usage of a
2295 * particular page (n_p) per total usage of this
2296 * page (n_t) (in a given time-span) to a probability.
2298 * Our periodic faults will sample this probability and
2299 * getting the same result twice in a row, given these
2300 * samples are fully independent, is then given by
2301 * P(n)^2, provided our sample period is sufficiently
2302 * short compared to the usage pattern.
2304 * This quadric squishes small probabilities, making
2305 * it less likely we act on an unlikely task<->page
2308 last_nid
= page_nid_xchg_last(page
, polnid
);
2309 if (last_nid
!= polnid
)
2313 if (curnid
!= polnid
)
2321 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2323 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2324 rb_erase(&n
->nd
, &sp
->root
);
2328 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2329 unsigned long end
, struct mempolicy
*pol
)
2331 node
->start
= start
;
2336 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2337 struct mempolicy
*pol
)
2340 struct mempolicy
*newpol
;
2342 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2346 newpol
= mpol_dup(pol
);
2347 if (IS_ERR(newpol
)) {
2348 kmem_cache_free(sn_cache
, n
);
2351 newpol
->flags
|= MPOL_F_SHARED
;
2352 sp_node_init(n
, start
, end
, newpol
);
2357 /* Replace a policy range. */
2358 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2359 unsigned long end
, struct sp_node
*new)
2362 struct sp_node
*n_new
= NULL
;
2363 struct mempolicy
*mpol_new
= NULL
;
2367 spin_lock(&sp
->lock
);
2368 n
= sp_lookup(sp
, start
, end
);
2369 /* Take care of old policies in the same range. */
2370 while (n
&& n
->start
< end
) {
2371 struct rb_node
*next
= rb_next(&n
->nd
);
2372 if (n
->start
>= start
) {
2378 /* Old policy spanning whole new range. */
2383 *mpol_new
= *n
->policy
;
2384 atomic_set(&mpol_new
->refcnt
, 1);
2385 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2387 sp_insert(sp
, n_new
);
2396 n
= rb_entry(next
, struct sp_node
, nd
);
2400 spin_unlock(&sp
->lock
);
2407 kmem_cache_free(sn_cache
, n_new
);
2412 spin_unlock(&sp
->lock
);
2414 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2417 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2424 * mpol_shared_policy_init - initialize shared policy for inode
2425 * @sp: pointer to inode shared policy
2426 * @mpol: struct mempolicy to install
2428 * Install non-NULL @mpol in inode's shared policy rb-tree.
2429 * On entry, the current task has a reference on a non-NULL @mpol.
2430 * This must be released on exit.
2431 * This is called at get_inode() calls and we can use GFP_KERNEL.
2433 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2437 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2438 spin_lock_init(&sp
->lock
);
2441 struct vm_area_struct pvma
;
2442 struct mempolicy
*new;
2443 NODEMASK_SCRATCH(scratch
);
2447 /* contextualize the tmpfs mount point mempolicy */
2448 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2450 goto free_scratch
; /* no valid nodemask intersection */
2453 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2454 task_unlock(current
);
2458 /* Create pseudo-vma that contains just the policy */
2459 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2460 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2461 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2464 mpol_put(new); /* drop initial ref */
2466 NODEMASK_SCRATCH_FREE(scratch
);
2468 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2472 int mpol_set_shared_policy(struct shared_policy
*info
,
2473 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2476 struct sp_node
*new = NULL
;
2477 unsigned long sz
= vma_pages(vma
);
2479 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2481 sz
, npol
? npol
->mode
: -1,
2482 npol
? npol
->flags
: -1,
2483 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2486 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2490 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2496 /* Free a backing policy store on inode delete. */
2497 void mpol_free_shared_policy(struct shared_policy
*p
)
2500 struct rb_node
*next
;
2502 if (!p
->root
.rb_node
)
2504 spin_lock(&p
->lock
);
2505 next
= rb_first(&p
->root
);
2507 n
= rb_entry(next
, struct sp_node
, nd
);
2508 next
= rb_next(&n
->nd
);
2511 spin_unlock(&p
->lock
);
2514 #ifdef CONFIG_NUMA_BALANCING
2515 static bool __initdata numabalancing_override
;
2517 static void __init
check_numabalancing_enable(void)
2519 bool numabalancing_default
= false;
2521 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2522 numabalancing_default
= true;
2524 if (nr_node_ids
> 1 && !numabalancing_override
) {
2525 printk(KERN_INFO
"Enabling automatic NUMA balancing. "
2526 "Configure with numa_balancing= or sysctl");
2527 set_numabalancing_state(numabalancing_default
);
2531 static int __init
setup_numabalancing(char *str
)
2536 numabalancing_override
= true;
2538 if (!strcmp(str
, "enable")) {
2539 set_numabalancing_state(true);
2541 } else if (!strcmp(str
, "disable")) {
2542 set_numabalancing_state(false);
2547 printk(KERN_WARNING
"Unable to parse numa_balancing=\n");
2551 __setup("numa_balancing=", setup_numabalancing
);
2553 static inline void __init
check_numabalancing_enable(void)
2556 #endif /* CONFIG_NUMA_BALANCING */
2558 /* assumes fs == KERNEL_DS */
2559 void __init
numa_policy_init(void)
2561 nodemask_t interleave_nodes
;
2562 unsigned long largest
= 0;
2563 int nid
, prefer
= 0;
2565 policy_cache
= kmem_cache_create("numa_policy",
2566 sizeof(struct mempolicy
),
2567 0, SLAB_PANIC
, NULL
);
2569 sn_cache
= kmem_cache_create("shared_policy_node",
2570 sizeof(struct sp_node
),
2571 0, SLAB_PANIC
, NULL
);
2573 for_each_node(nid
) {
2574 preferred_node_policy
[nid
] = (struct mempolicy
) {
2575 .refcnt
= ATOMIC_INIT(1),
2576 .mode
= MPOL_PREFERRED
,
2577 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2578 .v
= { .preferred_node
= nid
, },
2583 * Set interleaving policy for system init. Interleaving is only
2584 * enabled across suitably sized nodes (default is >= 16MB), or
2585 * fall back to the largest node if they're all smaller.
2587 nodes_clear(interleave_nodes
);
2588 for_each_node_state(nid
, N_MEMORY
) {
2589 unsigned long total_pages
= node_present_pages(nid
);
2591 /* Preserve the largest node */
2592 if (largest
< total_pages
) {
2593 largest
= total_pages
;
2597 /* Interleave this node? */
2598 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2599 node_set(nid
, interleave_nodes
);
2602 /* All too small, use the largest */
2603 if (unlikely(nodes_empty(interleave_nodes
)))
2604 node_set(prefer
, interleave_nodes
);
2606 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2607 printk("numa_policy_init: interleaving failed\n");
2609 check_numabalancing_enable();
2612 /* Reset policy of current process to default */
2613 void numa_default_policy(void)
2615 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2619 * Parse and format mempolicy from/to strings
2623 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2625 static const char * const policy_modes
[] =
2627 [MPOL_DEFAULT
] = "default",
2628 [MPOL_PREFERRED
] = "prefer",
2629 [MPOL_BIND
] = "bind",
2630 [MPOL_INTERLEAVE
] = "interleave",
2631 [MPOL_LOCAL
] = "local",
2637 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2638 * @str: string containing mempolicy to parse
2639 * @mpol: pointer to struct mempolicy pointer, returned on success.
2642 * <mode>[=<flags>][:<nodelist>]
2644 * On success, returns 0, else 1
2646 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2648 struct mempolicy
*new = NULL
;
2649 unsigned short mode
;
2650 unsigned short mode_flags
;
2652 char *nodelist
= strchr(str
, ':');
2653 char *flags
= strchr(str
, '=');
2657 /* NUL-terminate mode or flags string */
2659 if (nodelist_parse(nodelist
, nodes
))
2661 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2667 *flags
++ = '\0'; /* terminate mode string */
2669 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2670 if (!strcmp(str
, policy_modes
[mode
])) {
2674 if (mode
>= MPOL_MAX
)
2678 case MPOL_PREFERRED
:
2680 * Insist on a nodelist of one node only
2683 char *rest
= nodelist
;
2684 while (isdigit(*rest
))
2690 case MPOL_INTERLEAVE
:
2692 * Default to online nodes with memory if no nodelist
2695 nodes
= node_states
[N_MEMORY
];
2699 * Don't allow a nodelist; mpol_new() checks flags
2703 mode
= MPOL_PREFERRED
;
2707 * Insist on a empty nodelist
2714 * Insist on a nodelist
2723 * Currently, we only support two mutually exclusive
2726 if (!strcmp(flags
, "static"))
2727 mode_flags
|= MPOL_F_STATIC_NODES
;
2728 else if (!strcmp(flags
, "relative"))
2729 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2734 new = mpol_new(mode
, mode_flags
, &nodes
);
2739 * Save nodes for mpol_to_str() to show the tmpfs mount options
2740 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2742 if (mode
!= MPOL_PREFERRED
)
2743 new->v
.nodes
= nodes
;
2745 new->v
.preferred_node
= first_node(nodes
);
2747 new->flags
|= MPOL_F_LOCAL
;
2750 * Save nodes for contextualization: this will be used to "clone"
2751 * the mempolicy in a specific context [cpuset] at a later time.
2753 new->w
.user_nodemask
= nodes
;
2758 /* Restore string for error message */
2767 #endif /* CONFIG_TMPFS */
2770 * mpol_to_str - format a mempolicy structure for printing
2771 * @buffer: to contain formatted mempolicy string
2772 * @maxlen: length of @buffer
2773 * @pol: pointer to mempolicy to be formatted
2775 * Convert a mempolicy into a string.
2776 * Returns the number of characters in buffer (if positive)
2777 * or an error (negative)
2779 int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2784 unsigned short mode
;
2785 unsigned short flags
= pol
? pol
->flags
: 0;
2788 * Sanity check: room for longest mode, flag and some nodes
2790 VM_BUG_ON(maxlen
< strlen("interleave") + strlen("relative") + 16);
2792 if (!pol
|| pol
== &default_policy
|| (pol
->flags
& MPOL_F_MORON
))
2793 mode
= MPOL_DEFAULT
;
2802 case MPOL_PREFERRED
:
2804 if (flags
& MPOL_F_LOCAL
)
2807 node_set(pol
->v
.preferred_node
, nodes
);
2812 case MPOL_INTERLEAVE
:
2813 nodes
= pol
->v
.nodes
;
2820 l
= strlen(policy_modes
[mode
]);
2821 if (buffer
+ maxlen
< p
+ l
+ 1)
2824 strcpy(p
, policy_modes
[mode
]);
2827 if (flags
& MPOL_MODE_FLAGS
) {
2828 if (buffer
+ maxlen
< p
+ 2)
2833 * Currently, the only defined flags are mutually exclusive
2835 if (flags
& MPOL_F_STATIC_NODES
)
2836 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2837 else if (flags
& MPOL_F_RELATIVE_NODES
)
2838 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2841 if (!nodes_empty(nodes
)) {
2842 if (buffer
+ maxlen
< p
+ 2)
2845 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);