2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #include <linux/mempolicy.h>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/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();
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] : -1);
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.
611 static struct vm_area_struct
*
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
*first
, *vma
, *prev
;
619 first
= find_vma(mm
, start
);
621 return ERR_PTR(-EFAULT
);
623 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
624 unsigned long endvma
= vma
->vm_end
;
628 if (vma
->vm_start
> start
)
629 start
= vma
->vm_start
;
631 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
632 if (!vma
->vm_next
&& vma
->vm_end
< end
)
633 return ERR_PTR(-EFAULT
);
634 if (prev
&& prev
->vm_end
< vma
->vm_start
)
635 return ERR_PTR(-EFAULT
);
638 if (is_vm_hugetlb_page(vma
))
641 if (flags
& MPOL_MF_LAZY
) {
642 change_prot_numa(vma
, start
, endvma
);
646 if ((flags
& MPOL_MF_STRICT
) ||
647 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
648 vma_migratable(vma
))) {
650 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
653 first
= ERR_PTR(err
);
664 * Apply policy to a single VMA
665 * This must be called with the mmap_sem held for writing.
667 static int vma_replace_policy(struct vm_area_struct
*vma
,
668 struct mempolicy
*pol
)
671 struct mempolicy
*old
;
672 struct mempolicy
*new;
674 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
675 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
676 vma
->vm_ops
, vma
->vm_file
,
677 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
683 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
684 err
= vma
->vm_ops
->set_policy(vma
, new);
689 old
= vma
->vm_policy
;
690 vma
->vm_policy
= new; /* protected by mmap_sem */
699 /* Step 2: apply policy to a range and do splits. */
700 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
701 unsigned long end
, struct mempolicy
*new_pol
)
703 struct vm_area_struct
*next
;
704 struct vm_area_struct
*prev
;
705 struct vm_area_struct
*vma
;
708 unsigned long vmstart
;
711 vma
= find_vma(mm
, start
);
712 if (!vma
|| vma
->vm_start
> start
)
716 if (start
> vma
->vm_start
)
719 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
721 vmstart
= max(start
, vma
->vm_start
);
722 vmend
= min(end
, vma
->vm_end
);
724 if (mpol_equal(vma_policy(vma
), new_pol
))
727 pgoff
= vma
->vm_pgoff
+
728 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
729 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
730 vma
->anon_vma
, vma
->vm_file
, pgoff
,
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);
747 err
= vma_replace_policy(vma
, new_pol
);
757 * Update task->flags PF_MEMPOLICY bit: set iff non-default
758 * mempolicy. Allows more rapid checking of this (combined perhaps
759 * with other PF_* flag bits) on memory allocation hot code paths.
761 * If called from outside this file, the task 'p' should -only- be
762 * a newly forked child not yet visible on the task list, because
763 * manipulating the task flags of a visible task is not safe.
765 * The above limitation is why this routine has the funny name
766 * mpol_fix_fork_child_flag().
768 * It is also safe to call this with a task pointer of current,
769 * which the static wrapper mpol_set_task_struct_flag() does,
770 * for use within this file.
773 void mpol_fix_fork_child_flag(struct task_struct
*p
)
776 p
->flags
|= PF_MEMPOLICY
;
778 p
->flags
&= ~PF_MEMPOLICY
;
781 static void mpol_set_task_struct_flag(void)
783 mpol_fix_fork_child_flag(current
);
786 /* Set the process memory policy */
787 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
790 struct mempolicy
*new, *old
;
791 struct mm_struct
*mm
= current
->mm
;
792 NODEMASK_SCRATCH(scratch
);
798 new = mpol_new(mode
, flags
, nodes
);
804 * prevent changing our mempolicy while show_numa_maps()
806 * Note: do_set_mempolicy() can be called at init time
810 down_write(&mm
->mmap_sem
);
812 ret
= mpol_set_nodemask(new, nodes
, scratch
);
814 task_unlock(current
);
816 up_write(&mm
->mmap_sem
);
820 old
= current
->mempolicy
;
821 current
->mempolicy
= new;
822 mpol_set_task_struct_flag();
823 if (new && new->mode
== MPOL_INTERLEAVE
&&
824 nodes_weight(new->v
.nodes
))
825 current
->il_next
= first_node(new->v
.nodes
);
826 task_unlock(current
);
828 up_write(&mm
->mmap_sem
);
833 NODEMASK_SCRATCH_FREE(scratch
);
838 * Return nodemask for policy for get_mempolicy() query
840 * Called with task's alloc_lock held
842 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
845 if (p
== &default_policy
)
851 case MPOL_INTERLEAVE
:
855 if (!(p
->flags
& MPOL_F_LOCAL
))
856 node_set(p
->v
.preferred_node
, *nodes
);
857 /* else return empty node mask for local allocation */
864 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
869 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
871 err
= page_to_nid(p
);
877 /* Retrieve NUMA policy */
878 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
879 unsigned long addr
, unsigned long flags
)
882 struct mm_struct
*mm
= current
->mm
;
883 struct vm_area_struct
*vma
= NULL
;
884 struct mempolicy
*pol
= current
->mempolicy
;
887 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
890 if (flags
& MPOL_F_MEMS_ALLOWED
) {
891 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
893 *policy
= 0; /* just so it's initialized */
895 *nmask
= cpuset_current_mems_allowed
;
896 task_unlock(current
);
900 if (flags
& MPOL_F_ADDR
) {
902 * Do NOT fall back to task policy if the
903 * vma/shared policy at addr is NULL. We
904 * want to return MPOL_DEFAULT in this case.
906 down_read(&mm
->mmap_sem
);
907 vma
= find_vma_intersection(mm
, addr
, addr
+1);
909 up_read(&mm
->mmap_sem
);
912 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
913 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
915 pol
= vma
->vm_policy
;
920 pol
= &default_policy
; /* indicates default behavior */
922 if (flags
& MPOL_F_NODE
) {
923 if (flags
& MPOL_F_ADDR
) {
924 err
= lookup_node(mm
, addr
);
928 } else if (pol
== current
->mempolicy
&&
929 pol
->mode
== MPOL_INTERLEAVE
) {
930 *policy
= current
->il_next
;
936 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
939 * Internal mempolicy flags must be masked off before exposing
940 * the policy to userspace.
942 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
946 up_read(¤t
->mm
->mmap_sem
);
952 if (mpol_store_user_nodemask(pol
)) {
953 *nmask
= pol
->w
.user_nodemask
;
956 get_policy_nodemask(pol
, nmask
);
957 task_unlock(current
);
964 up_read(¤t
->mm
->mmap_sem
);
968 #ifdef CONFIG_MIGRATION
972 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
976 * Avoid migrating a page that is shared with others.
978 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
979 if (!isolate_lru_page(page
)) {
980 list_add_tail(&page
->lru
, pagelist
);
981 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
982 page_is_file_cache(page
));
987 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
989 return alloc_pages_exact_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
993 * Migrate pages from one node to a target node.
994 * Returns error or the number of pages not migrated.
996 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1000 LIST_HEAD(pagelist
);
1004 node_set(source
, nmask
);
1007 * This does not "check" the range but isolates all pages that
1008 * need migration. Between passing in the full user address
1009 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1011 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1012 check_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1013 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1015 if (!list_empty(&pagelist
)) {
1016 err
= migrate_pages(&pagelist
, new_node_page
, dest
,
1017 false, MIGRATE_SYNC
,
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 assuming that page is mapped by vma pointed to by @private.
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_vma_page(struct page
*page
, unsigned long private, int **x
)
1145 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
1146 unsigned long uninitialized_var(address
);
1149 address
= page_address_in_vma(page
, vma
);
1150 if (address
!= -EFAULT
)
1156 * if !vma, alloc_page_vma() will use task or system default policy
1158 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1162 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1163 unsigned long flags
)
1167 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1168 const nodemask_t
*to
, int flags
)
1173 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
1179 static long do_mbind(unsigned long start
, unsigned long len
,
1180 unsigned short mode
, unsigned short mode_flags
,
1181 nodemask_t
*nmask
, unsigned long flags
)
1183 struct vm_area_struct
*vma
;
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] : -1);
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 vma
= check_range(mm
, start
, end
, nmask
,
1250 flags
| MPOL_MF_INVERT
, &pagelist
);
1252 err
= PTR_ERR(vma
); /* maybe ... */
1254 err
= mbind_range(mm
, start
, end
, new);
1259 if (!list_empty(&pagelist
)) {
1260 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1261 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
1263 false, MIGRATE_SYNC
,
1264 MR_MEMPOLICY_MBIND
);
1266 putback_lru_pages(&pagelist
);
1269 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1272 putback_lru_pages(&pagelist
);
1274 up_write(&mm
->mmap_sem
);
1281 * User space interface with variable sized bitmaps for nodelists.
1284 /* Copy a node mask from user space. */
1285 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1286 unsigned long maxnode
)
1289 unsigned long nlongs
;
1290 unsigned long endmask
;
1293 nodes_clear(*nodes
);
1294 if (maxnode
== 0 || !nmask
)
1296 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1299 nlongs
= BITS_TO_LONGS(maxnode
);
1300 if ((maxnode
% BITS_PER_LONG
) == 0)
1303 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1305 /* When the user specified more nodes than supported just check
1306 if the non supported part is all zero. */
1307 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1308 if (nlongs
> PAGE_SIZE
/sizeof(long))
1310 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1312 if (get_user(t
, nmask
+ k
))
1314 if (k
== nlongs
- 1) {
1320 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1324 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1326 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1330 /* Copy a kernel node mask to user space */
1331 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1334 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1335 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1337 if (copy
> nbytes
) {
1338 if (copy
> PAGE_SIZE
)
1340 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1344 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1347 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1348 unsigned long, mode
, unsigned long __user
*, nmask
,
1349 unsigned long, maxnode
, unsigned, flags
)
1353 unsigned short mode_flags
;
1355 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1356 mode
&= ~MPOL_MODE_FLAGS
;
1357 if (mode
>= MPOL_MAX
)
1359 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1360 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1362 err
= get_nodes(&nodes
, nmask
, maxnode
);
1365 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1368 /* Set the process memory policy */
1369 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, unsigned long __user
*, nmask
,
1370 unsigned long, maxnode
)
1374 unsigned short flags
;
1376 flags
= mode
& MPOL_MODE_FLAGS
;
1377 mode
&= ~MPOL_MODE_FLAGS
;
1378 if ((unsigned int)mode
>= MPOL_MAX
)
1380 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1382 err
= get_nodes(&nodes
, nmask
, maxnode
);
1385 return do_set_mempolicy(mode
, flags
, &nodes
);
1388 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1389 const unsigned long __user
*, old_nodes
,
1390 const unsigned long __user
*, new_nodes
)
1392 const struct cred
*cred
= current_cred(), *tcred
;
1393 struct mm_struct
*mm
= NULL
;
1394 struct task_struct
*task
;
1395 nodemask_t task_nodes
;
1399 NODEMASK_SCRATCH(scratch
);
1404 old
= &scratch
->mask1
;
1405 new = &scratch
->mask2
;
1407 err
= get_nodes(old
, old_nodes
, maxnode
);
1411 err
= get_nodes(new, new_nodes
, maxnode
);
1415 /* Find the mm_struct */
1417 task
= pid
? find_task_by_vpid(pid
) : current
;
1423 get_task_struct(task
);
1428 * Check if this process has the right to modify the specified
1429 * process. The right exists if the process has administrative
1430 * capabilities, superuser privileges or the same
1431 * userid as the target process.
1433 tcred
= __task_cred(task
);
1434 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1435 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1436 !capable(CAP_SYS_NICE
)) {
1443 task_nodes
= cpuset_mems_allowed(task
);
1444 /* Is the user allowed to access the target nodes? */
1445 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1450 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1455 err
= security_task_movememory(task
);
1459 mm
= get_task_mm(task
);
1460 put_task_struct(task
);
1467 err
= do_migrate_pages(mm
, old
, new,
1468 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1472 NODEMASK_SCRATCH_FREE(scratch
);
1477 put_task_struct(task
);
1483 /* Retrieve NUMA policy */
1484 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1485 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1486 unsigned long, addr
, unsigned long, flags
)
1489 int uninitialized_var(pval
);
1492 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1495 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1500 if (policy
&& put_user(pval
, policy
))
1504 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1509 #ifdef CONFIG_COMPAT
1511 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1512 compat_ulong_t __user
*nmask
,
1513 compat_ulong_t maxnode
,
1514 compat_ulong_t addr
, compat_ulong_t flags
)
1517 unsigned long __user
*nm
= NULL
;
1518 unsigned long nr_bits
, alloc_size
;
1519 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1521 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1522 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1525 nm
= compat_alloc_user_space(alloc_size
);
1527 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1529 if (!err
&& nmask
) {
1530 unsigned long copy_size
;
1531 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1532 err
= copy_from_user(bm
, nm
, copy_size
);
1533 /* ensure entire bitmap is zeroed */
1534 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1535 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1541 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1542 compat_ulong_t maxnode
)
1545 unsigned long __user
*nm
= NULL
;
1546 unsigned long nr_bits
, alloc_size
;
1547 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1549 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1550 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1553 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1554 nm
= compat_alloc_user_space(alloc_size
);
1555 err
|= copy_to_user(nm
, bm
, alloc_size
);
1561 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1564 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1565 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1566 compat_ulong_t maxnode
, compat_ulong_t flags
)
1569 unsigned long __user
*nm
= NULL
;
1570 unsigned long nr_bits
, alloc_size
;
1573 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1574 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1577 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1578 nm
= compat_alloc_user_space(alloc_size
);
1579 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1585 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1591 * get_vma_policy(@task, @vma, @addr)
1592 * @task - task for fallback if vma policy == default
1593 * @vma - virtual memory area whose policy is sought
1594 * @addr - address in @vma for shared policy lookup
1596 * Returns effective policy for a VMA at specified address.
1597 * Falls back to @task or system default policy, as necessary.
1598 * Current or other task's task mempolicy and non-shared vma policies must be
1599 * protected by task_lock(task) by the caller.
1600 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1601 * count--added by the get_policy() vm_op, as appropriate--to protect against
1602 * freeing by another task. It is the caller's responsibility to free the
1603 * extra reference for shared policies.
1605 struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1606 struct vm_area_struct
*vma
, unsigned long addr
)
1608 struct mempolicy
*pol
= get_task_policy(task
);
1611 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1612 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1616 } else if (vma
->vm_policy
) {
1617 pol
= vma
->vm_policy
;
1620 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1621 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1622 * count on these policies which will be dropped by
1623 * mpol_cond_put() later
1625 if (mpol_needs_cond_ref(pol
))
1630 pol
= &default_policy
;
1634 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1636 enum zone_type dynamic_policy_zone
= policy_zone
;
1638 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1641 * if policy->v.nodes has movable memory only,
1642 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1644 * policy->v.nodes is intersect with node_states[N_MEMORY].
1645 * so if the following test faile, it implies
1646 * policy->v.nodes has movable memory only.
1648 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1649 dynamic_policy_zone
= ZONE_MOVABLE
;
1651 return zone
>= dynamic_policy_zone
;
1655 * Return a nodemask representing a mempolicy for filtering nodes for
1658 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1660 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1661 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1662 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1663 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1664 return &policy
->v
.nodes
;
1669 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1670 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1673 switch (policy
->mode
) {
1674 case MPOL_PREFERRED
:
1675 if (!(policy
->flags
& MPOL_F_LOCAL
))
1676 nd
= policy
->v
.preferred_node
;
1680 * Normally, MPOL_BIND allocations are node-local within the
1681 * allowed nodemask. However, if __GFP_THISNODE is set and the
1682 * current node isn't part of the mask, we use the zonelist for
1683 * the first node in the mask instead.
1685 if (unlikely(gfp
& __GFP_THISNODE
) &&
1686 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1687 nd
= first_node(policy
->v
.nodes
);
1692 return node_zonelist(nd
, gfp
);
1695 /* Do dynamic interleaving for a process */
1696 static unsigned interleave_nodes(struct mempolicy
*policy
)
1699 struct task_struct
*me
= current
;
1702 next
= next_node(nid
, policy
->v
.nodes
);
1703 if (next
>= MAX_NUMNODES
)
1704 next
= first_node(policy
->v
.nodes
);
1705 if (next
< MAX_NUMNODES
)
1711 * Depending on the memory policy provide a node from which to allocate the
1713 * @policy must be protected by freeing by the caller. If @policy is
1714 * the current task's mempolicy, this protection is implicit, as only the
1715 * task can change it's policy. The system default policy requires no
1718 unsigned slab_node(void)
1720 struct mempolicy
*policy
;
1723 return numa_node_id();
1725 policy
= current
->mempolicy
;
1726 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1727 return numa_node_id();
1729 switch (policy
->mode
) {
1730 case MPOL_PREFERRED
:
1732 * handled MPOL_F_LOCAL above
1734 return policy
->v
.preferred_node
;
1736 case MPOL_INTERLEAVE
:
1737 return interleave_nodes(policy
);
1741 * Follow bind policy behavior and start allocation at the
1744 struct zonelist
*zonelist
;
1746 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1747 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1748 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1751 return zone
? zone
->node
: numa_node_id();
1759 /* Do static interleaving for a VMA with known offset. */
1760 static unsigned offset_il_node(struct mempolicy
*pol
,
1761 struct vm_area_struct
*vma
, unsigned long off
)
1763 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1769 return numa_node_id();
1770 target
= (unsigned int)off
% nnodes
;
1773 nid
= next_node(nid
, pol
->v
.nodes
);
1775 } while (c
<= target
);
1779 /* Determine a node number for interleave */
1780 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1781 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1787 * for small pages, there is no difference between
1788 * shift and PAGE_SHIFT, so the bit-shift is safe.
1789 * for huge pages, since vm_pgoff is in units of small
1790 * pages, we need to shift off the always 0 bits to get
1793 BUG_ON(shift
< PAGE_SHIFT
);
1794 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1795 off
+= (addr
- vma
->vm_start
) >> shift
;
1796 return offset_il_node(pol
, vma
, off
);
1798 return interleave_nodes(pol
);
1802 * Return the bit number of a random bit set in the nodemask.
1803 * (returns -1 if nodemask is empty)
1805 int node_random(const nodemask_t
*maskp
)
1809 w
= nodes_weight(*maskp
);
1811 bit
= bitmap_ord_to_pos(maskp
->bits
,
1812 get_random_int() % w
, MAX_NUMNODES
);
1816 #ifdef CONFIG_HUGETLBFS
1818 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1819 * @vma = virtual memory area whose policy is sought
1820 * @addr = address in @vma for shared policy lookup and interleave policy
1821 * @gfp_flags = for requested zone
1822 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1823 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1825 * Returns a zonelist suitable for a huge page allocation and a pointer
1826 * to the struct mempolicy for conditional unref after allocation.
1827 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1828 * @nodemask for filtering the zonelist.
1830 * Must be protected by get_mems_allowed()
1832 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1833 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1834 nodemask_t
**nodemask
)
1836 struct zonelist
*zl
;
1838 *mpol
= get_vma_policy(current
, vma
, addr
);
1839 *nodemask
= NULL
; /* assume !MPOL_BIND */
1841 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1842 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1843 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1845 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1846 if ((*mpol
)->mode
== MPOL_BIND
)
1847 *nodemask
= &(*mpol
)->v
.nodes
;
1853 * init_nodemask_of_mempolicy
1855 * If the current task's mempolicy is "default" [NULL], return 'false'
1856 * to indicate default policy. Otherwise, extract the policy nodemask
1857 * for 'bind' or 'interleave' policy into the argument nodemask, or
1858 * initialize the argument nodemask to contain the single node for
1859 * 'preferred' or 'local' policy and return 'true' to indicate presence
1860 * of non-default mempolicy.
1862 * We don't bother with reference counting the mempolicy [mpol_get/put]
1863 * because the current task is examining it's own mempolicy and a task's
1864 * mempolicy is only ever changed by the task itself.
1866 * N.B., it is the caller's responsibility to free a returned nodemask.
1868 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1870 struct mempolicy
*mempolicy
;
1873 if (!(mask
&& current
->mempolicy
))
1877 mempolicy
= current
->mempolicy
;
1878 switch (mempolicy
->mode
) {
1879 case MPOL_PREFERRED
:
1880 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1881 nid
= numa_node_id();
1883 nid
= mempolicy
->v
.preferred_node
;
1884 init_nodemask_of_node(mask
, nid
);
1889 case MPOL_INTERLEAVE
:
1890 *mask
= mempolicy
->v
.nodes
;
1896 task_unlock(current
);
1903 * mempolicy_nodemask_intersects
1905 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1906 * policy. Otherwise, check for intersection between mask and the policy
1907 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1908 * policy, always return true since it may allocate elsewhere on fallback.
1910 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1912 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1913 const nodemask_t
*mask
)
1915 struct mempolicy
*mempolicy
;
1921 mempolicy
= tsk
->mempolicy
;
1925 switch (mempolicy
->mode
) {
1926 case MPOL_PREFERRED
:
1928 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1929 * allocate from, they may fallback to other nodes when oom.
1930 * Thus, it's possible for tsk to have allocated memory from
1935 case MPOL_INTERLEAVE
:
1936 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1946 /* Allocate a page in interleaved policy.
1947 Own path because it needs to do special accounting. */
1948 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1951 struct zonelist
*zl
;
1954 zl
= node_zonelist(nid
, gfp
);
1955 page
= __alloc_pages(gfp
, order
, zl
);
1956 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1957 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1962 * alloc_pages_vma - Allocate a page for a VMA.
1965 * %GFP_USER user allocation.
1966 * %GFP_KERNEL kernel allocations,
1967 * %GFP_HIGHMEM highmem/user allocations,
1968 * %GFP_FS allocation should not call back into a file system.
1969 * %GFP_ATOMIC don't sleep.
1971 * @order:Order of the GFP allocation.
1972 * @vma: Pointer to VMA or NULL if not available.
1973 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1975 * This function allocates a page from the kernel page pool and applies
1976 * a NUMA policy associated with the VMA or the current process.
1977 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1978 * mm_struct of the VMA to prevent it from going away. Should be used for
1979 * all allocations for pages that will be mapped into
1980 * user space. Returns NULL when no page can be allocated.
1982 * Should be called with the mm_sem of the vma hold.
1985 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1986 unsigned long addr
, int node
)
1988 struct mempolicy
*pol
;
1990 unsigned int cpuset_mems_cookie
;
1993 pol
= get_vma_policy(current
, vma
, addr
);
1994 cpuset_mems_cookie
= get_mems_allowed();
1996 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
1999 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2001 page
= alloc_page_interleave(gfp
, order
, nid
);
2002 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2007 page
= __alloc_pages_nodemask(gfp
, order
,
2008 policy_zonelist(gfp
, pol
, node
),
2009 policy_nodemask(gfp
, pol
));
2010 if (unlikely(mpol_needs_cond_ref(pol
)))
2012 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2018 * alloc_pages_current - Allocate pages.
2021 * %GFP_USER user allocation,
2022 * %GFP_KERNEL kernel allocation,
2023 * %GFP_HIGHMEM highmem allocation,
2024 * %GFP_FS don't call back into a file system.
2025 * %GFP_ATOMIC don't sleep.
2026 * @order: Power of two of allocation size in pages. 0 is a single page.
2028 * Allocate a page from the kernel page pool. When not in
2029 * interrupt context and apply the current process NUMA policy.
2030 * Returns NULL when no page can be allocated.
2032 * Don't call cpuset_update_task_memory_state() unless
2033 * 1) it's ok to take cpuset_sem (can WAIT), and
2034 * 2) allocating for current task (not interrupt).
2036 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2038 struct mempolicy
*pol
= get_task_policy(current
);
2040 unsigned int cpuset_mems_cookie
;
2042 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
2043 pol
= &default_policy
;
2046 cpuset_mems_cookie
= get_mems_allowed();
2049 * No reference counting needed for current->mempolicy
2050 * nor system default_policy
2052 if (pol
->mode
== MPOL_INTERLEAVE
)
2053 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2055 page
= __alloc_pages_nodemask(gfp
, order
,
2056 policy_zonelist(gfp
, pol
, numa_node_id()),
2057 policy_nodemask(gfp
, pol
));
2059 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2064 EXPORT_SYMBOL(alloc_pages_current
);
2067 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2068 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2069 * with the mems_allowed returned by cpuset_mems_allowed(). This
2070 * keeps mempolicies cpuset relative after its cpuset moves. See
2071 * further kernel/cpuset.c update_nodemask().
2073 * current's mempolicy may be rebinded by the other task(the task that changes
2074 * cpuset's mems), so we needn't do rebind work for current task.
2077 /* Slow path of a mempolicy duplicate */
2078 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2080 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2083 return ERR_PTR(-ENOMEM
);
2085 /* task's mempolicy is protected by alloc_lock */
2086 if (old
== current
->mempolicy
) {
2089 task_unlock(current
);
2094 if (current_cpuset_is_being_rebound()) {
2095 nodemask_t mems
= cpuset_mems_allowed(current
);
2096 if (new->flags
& MPOL_F_REBINDING
)
2097 mpol_rebind_policy(new, &mems
, MPOL_REBIND_STEP2
);
2099 mpol_rebind_policy(new, &mems
, MPOL_REBIND_ONCE
);
2102 atomic_set(&new->refcnt
, 1);
2106 /* Slow path of a mempolicy comparison */
2107 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2111 if (a
->mode
!= b
->mode
)
2113 if (a
->flags
!= b
->flags
)
2115 if (mpol_store_user_nodemask(a
))
2116 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2122 case MPOL_INTERLEAVE
:
2123 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2124 case MPOL_PREFERRED
:
2125 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2133 * Shared memory backing store policy support.
2135 * Remember policies even when nobody has shared memory mapped.
2136 * The policies are kept in Red-Black tree linked from the inode.
2137 * They are protected by the sp->lock spinlock, which should be held
2138 * for any accesses to the tree.
2141 /* lookup first element intersecting start-end */
2142 /* Caller holds sp->lock */
2143 static struct sp_node
*
2144 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2146 struct rb_node
*n
= sp
->root
.rb_node
;
2149 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2151 if (start
>= p
->end
)
2153 else if (end
<= p
->start
)
2161 struct sp_node
*w
= NULL
;
2162 struct rb_node
*prev
= rb_prev(n
);
2165 w
= rb_entry(prev
, struct sp_node
, nd
);
2166 if (w
->end
<= start
)
2170 return rb_entry(n
, struct sp_node
, nd
);
2173 /* Insert a new shared policy into the list. */
2174 /* Caller holds sp->lock */
2175 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2177 struct rb_node
**p
= &sp
->root
.rb_node
;
2178 struct rb_node
*parent
= NULL
;
2183 nd
= rb_entry(parent
, struct sp_node
, nd
);
2184 if (new->start
< nd
->start
)
2186 else if (new->end
> nd
->end
)
2187 p
= &(*p
)->rb_right
;
2191 rb_link_node(&new->nd
, parent
, p
);
2192 rb_insert_color(&new->nd
, &sp
->root
);
2193 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2194 new->policy
? new->policy
->mode
: 0);
2197 /* Find shared policy intersecting idx */
2199 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2201 struct mempolicy
*pol
= NULL
;
2204 if (!sp
->root
.rb_node
)
2206 spin_lock(&sp
->lock
);
2207 sn
= sp_lookup(sp
, idx
, idx
+1);
2209 mpol_get(sn
->policy
);
2212 spin_unlock(&sp
->lock
);
2216 static void sp_free(struct sp_node
*n
)
2218 mpol_put(n
->policy
);
2219 kmem_cache_free(sn_cache
, n
);
2223 * mpol_misplaced - check whether current page node is valid in policy
2225 * @page - page to be checked
2226 * @vma - vm area where page mapped
2227 * @addr - virtual address where page mapped
2229 * Lookup current policy node id for vma,addr and "compare to" page's
2233 * -1 - not misplaced, page is in the right node
2234 * node - node id where the page should be
2236 * Policy determination "mimics" alloc_page_vma().
2237 * Called from fault path where we know the vma and faulting address.
2239 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2241 struct mempolicy
*pol
;
2243 int curnid
= page_to_nid(page
);
2244 unsigned long pgoff
;
2250 pol
= get_vma_policy(current
, vma
, addr
);
2251 if (!(pol
->flags
& MPOL_F_MOF
))
2254 switch (pol
->mode
) {
2255 case MPOL_INTERLEAVE
:
2256 BUG_ON(addr
>= vma
->vm_end
);
2257 BUG_ON(addr
< vma
->vm_start
);
2259 pgoff
= vma
->vm_pgoff
;
2260 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2261 polnid
= offset_il_node(pol
, vma
, pgoff
);
2264 case MPOL_PREFERRED
:
2265 if (pol
->flags
& MPOL_F_LOCAL
)
2266 polnid
= numa_node_id();
2268 polnid
= pol
->v
.preferred_node
;
2273 * allows binding to multiple nodes.
2274 * use current page if in policy nodemask,
2275 * else select nearest allowed node, if any.
2276 * If no allowed nodes, use current [!misplaced].
2278 if (node_isset(curnid
, pol
->v
.nodes
))
2280 (void)first_zones_zonelist(
2281 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2282 gfp_zone(GFP_HIGHUSER
),
2283 &pol
->v
.nodes
, &zone
);
2284 polnid
= zone
->node
;
2291 /* Migrate the page towards the node whose CPU is referencing it */
2292 if (pol
->flags
& MPOL_F_MORON
) {
2295 polnid
= numa_node_id();
2298 * Multi-stage node selection is used in conjunction
2299 * with a periodic migration fault to build a temporal
2300 * task<->page relation. By using a two-stage filter we
2301 * remove short/unlikely relations.
2303 * Using P(p) ~ n_p / n_t as per frequentist
2304 * probability, we can equate a task's usage of a
2305 * particular page (n_p) per total usage of this
2306 * page (n_t) (in a given time-span) to a probability.
2308 * Our periodic faults will sample this probability and
2309 * getting the same result twice in a row, given these
2310 * samples are fully independent, is then given by
2311 * P(n)^2, provided our sample period is sufficiently
2312 * short compared to the usage pattern.
2314 * This quadric squishes small probabilities, making
2315 * it less likely we act on an unlikely task<->page
2318 last_nid
= page_nid_xchg_last(page
, polnid
);
2319 if (last_nid
!= polnid
)
2323 if (curnid
!= polnid
)
2331 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2333 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2334 rb_erase(&n
->nd
, &sp
->root
);
2338 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2339 unsigned long end
, struct mempolicy
*pol
)
2341 node
->start
= start
;
2346 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2347 struct mempolicy
*pol
)
2350 struct mempolicy
*newpol
;
2352 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2356 newpol
= mpol_dup(pol
);
2357 if (IS_ERR(newpol
)) {
2358 kmem_cache_free(sn_cache
, n
);
2361 newpol
->flags
|= MPOL_F_SHARED
;
2362 sp_node_init(n
, start
, end
, newpol
);
2367 /* Replace a policy range. */
2368 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2369 unsigned long end
, struct sp_node
*new)
2372 struct sp_node
*n_new
= NULL
;
2373 struct mempolicy
*mpol_new
= NULL
;
2377 spin_lock(&sp
->lock
);
2378 n
= sp_lookup(sp
, start
, end
);
2379 /* Take care of old policies in the same range. */
2380 while (n
&& n
->start
< end
) {
2381 struct rb_node
*next
= rb_next(&n
->nd
);
2382 if (n
->start
>= start
) {
2388 /* Old policy spanning whole new range. */
2393 *mpol_new
= *n
->policy
;
2394 atomic_set(&mpol_new
->refcnt
, 1);
2395 sp_node_init(n_new
, n
->end
, end
, mpol_new
);
2396 sp_insert(sp
, n_new
);
2406 n
= rb_entry(next
, struct sp_node
, nd
);
2410 spin_unlock(&sp
->lock
);
2417 kmem_cache_free(sn_cache
, n_new
);
2422 spin_unlock(&sp
->lock
);
2424 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2427 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2434 * mpol_shared_policy_init - initialize shared policy for inode
2435 * @sp: pointer to inode shared policy
2436 * @mpol: struct mempolicy to install
2438 * Install non-NULL @mpol in inode's shared policy rb-tree.
2439 * On entry, the current task has a reference on a non-NULL @mpol.
2440 * This must be released on exit.
2441 * This is called at get_inode() calls and we can use GFP_KERNEL.
2443 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2447 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2448 spin_lock_init(&sp
->lock
);
2451 struct vm_area_struct pvma
;
2452 struct mempolicy
*new;
2453 NODEMASK_SCRATCH(scratch
);
2457 /* contextualize the tmpfs mount point mempolicy */
2458 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2460 goto free_scratch
; /* no valid nodemask intersection */
2463 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2464 task_unlock(current
);
2468 /* Create pseudo-vma that contains just the policy */
2469 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2470 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2471 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2474 mpol_put(new); /* drop initial ref */
2476 NODEMASK_SCRATCH_FREE(scratch
);
2478 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2482 int mpol_set_shared_policy(struct shared_policy
*info
,
2483 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2486 struct sp_node
*new = NULL
;
2487 unsigned long sz
= vma_pages(vma
);
2489 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2491 sz
, npol
? npol
->mode
: -1,
2492 npol
? npol
->flags
: -1,
2493 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
2496 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2500 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2506 /* Free a backing policy store on inode delete. */
2507 void mpol_free_shared_policy(struct shared_policy
*p
)
2510 struct rb_node
*next
;
2512 if (!p
->root
.rb_node
)
2514 spin_lock(&p
->lock
);
2515 next
= rb_first(&p
->root
);
2517 n
= rb_entry(next
, struct sp_node
, nd
);
2518 next
= rb_next(&n
->nd
);
2521 spin_unlock(&p
->lock
);
2524 #ifdef CONFIG_NUMA_BALANCING
2525 static bool __initdata numabalancing_override
;
2527 static void __init
check_numabalancing_enable(void)
2529 bool numabalancing_default
= false;
2531 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2532 numabalancing_default
= true;
2534 if (nr_node_ids
> 1 && !numabalancing_override
) {
2535 printk(KERN_INFO
"Enabling automatic NUMA balancing. "
2536 "Configure with numa_balancing= or sysctl");
2537 set_numabalancing_state(numabalancing_default
);
2541 static int __init
setup_numabalancing(char *str
)
2546 numabalancing_override
= true;
2548 if (!strcmp(str
, "enable")) {
2549 set_numabalancing_state(true);
2551 } else if (!strcmp(str
, "disable")) {
2552 set_numabalancing_state(false);
2557 printk(KERN_WARNING
"Unable to parse numa_balancing=\n");
2561 __setup("numa_balancing=", setup_numabalancing
);
2563 static inline void __init
check_numabalancing_enable(void)
2566 #endif /* CONFIG_NUMA_BALANCING */
2568 /* assumes fs == KERNEL_DS */
2569 void __init
numa_policy_init(void)
2571 nodemask_t interleave_nodes
;
2572 unsigned long largest
= 0;
2573 int nid
, prefer
= 0;
2575 policy_cache
= kmem_cache_create("numa_policy",
2576 sizeof(struct mempolicy
),
2577 0, SLAB_PANIC
, NULL
);
2579 sn_cache
= kmem_cache_create("shared_policy_node",
2580 sizeof(struct sp_node
),
2581 0, SLAB_PANIC
, NULL
);
2583 for_each_node(nid
) {
2584 preferred_node_policy
[nid
] = (struct mempolicy
) {
2585 .refcnt
= ATOMIC_INIT(1),
2586 .mode
= MPOL_PREFERRED
,
2587 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2588 .v
= { .preferred_node
= nid
, },
2593 * Set interleaving policy for system init. Interleaving is only
2594 * enabled across suitably sized nodes (default is >= 16MB), or
2595 * fall back to the largest node if they're all smaller.
2597 nodes_clear(interleave_nodes
);
2598 for_each_node_state(nid
, N_MEMORY
) {
2599 unsigned long total_pages
= node_present_pages(nid
);
2601 /* Preserve the largest node */
2602 if (largest
< total_pages
) {
2603 largest
= total_pages
;
2607 /* Interleave this node? */
2608 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2609 node_set(nid
, interleave_nodes
);
2612 /* All too small, use the largest */
2613 if (unlikely(nodes_empty(interleave_nodes
)))
2614 node_set(prefer
, interleave_nodes
);
2616 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2617 printk("numa_policy_init: interleaving failed\n");
2619 check_numabalancing_enable();
2622 /* Reset policy of current process to default */
2623 void numa_default_policy(void)
2625 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2629 * Parse and format mempolicy from/to strings
2633 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2635 static const char * const policy_modes
[] =
2637 [MPOL_DEFAULT
] = "default",
2638 [MPOL_PREFERRED
] = "prefer",
2639 [MPOL_BIND
] = "bind",
2640 [MPOL_INTERLEAVE
] = "interleave",
2641 [MPOL_LOCAL
] = "local",
2647 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2648 * @str: string containing mempolicy to parse
2649 * @mpol: pointer to struct mempolicy pointer, returned on success.
2652 * <mode>[=<flags>][:<nodelist>]
2654 * On success, returns 0, else 1
2656 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2658 struct mempolicy
*new = NULL
;
2659 unsigned short mode
;
2660 unsigned short mode_flags
;
2662 char *nodelist
= strchr(str
, ':');
2663 char *flags
= strchr(str
, '=');
2667 /* NUL-terminate mode or flags string */
2669 if (nodelist_parse(nodelist
, nodes
))
2671 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2677 *flags
++ = '\0'; /* terminate mode string */
2679 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2680 if (!strcmp(str
, policy_modes
[mode
])) {
2684 if (mode
>= MPOL_MAX
)
2688 case MPOL_PREFERRED
:
2690 * Insist on a nodelist of one node only
2693 char *rest
= nodelist
;
2694 while (isdigit(*rest
))
2700 case MPOL_INTERLEAVE
:
2702 * Default to online nodes with memory if no nodelist
2705 nodes
= node_states
[N_MEMORY
];
2709 * Don't allow a nodelist; mpol_new() checks flags
2713 mode
= MPOL_PREFERRED
;
2717 * Insist on a empty nodelist
2724 * Insist on a nodelist
2733 * Currently, we only support two mutually exclusive
2736 if (!strcmp(flags
, "static"))
2737 mode_flags
|= MPOL_F_STATIC_NODES
;
2738 else if (!strcmp(flags
, "relative"))
2739 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2744 new = mpol_new(mode
, mode_flags
, &nodes
);
2749 * Save nodes for mpol_to_str() to show the tmpfs mount options
2750 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2752 if (mode
!= MPOL_PREFERRED
)
2753 new->v
.nodes
= nodes
;
2755 new->v
.preferred_node
= first_node(nodes
);
2757 new->flags
|= MPOL_F_LOCAL
;
2760 * Save nodes for contextualization: this will be used to "clone"
2761 * the mempolicy in a specific context [cpuset] at a later time.
2763 new->w
.user_nodemask
= nodes
;
2768 /* Restore string for error message */
2777 #endif /* CONFIG_TMPFS */
2780 * mpol_to_str - format a mempolicy structure for printing
2781 * @buffer: to contain formatted mempolicy string
2782 * @maxlen: length of @buffer
2783 * @pol: pointer to mempolicy to be formatted
2785 * Convert a mempolicy into a string.
2786 * Returns the number of characters in buffer (if positive)
2787 * or an error (negative)
2789 int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2794 unsigned short mode
;
2795 unsigned short flags
= pol
? pol
->flags
: 0;
2798 * Sanity check: room for longest mode, flag and some nodes
2800 VM_BUG_ON(maxlen
< strlen("interleave") + strlen("relative") + 16);
2802 if (!pol
|| pol
== &default_policy
)
2803 mode
= MPOL_DEFAULT
;
2812 case MPOL_PREFERRED
:
2814 if (flags
& MPOL_F_LOCAL
)
2817 node_set(pol
->v
.preferred_node
, nodes
);
2822 case MPOL_INTERLEAVE
:
2823 nodes
= pol
->v
.nodes
;
2830 l
= strlen(policy_modes
[mode
]);
2831 if (buffer
+ maxlen
< p
+ l
+ 1)
2834 strcpy(p
, policy_modes
[mode
]);
2837 if (flags
& MPOL_MODE_FLAGS
) {
2838 if (buffer
+ maxlen
< p
+ 2)
2843 * Currently, the only defined flags are mutually exclusive
2845 if (flags
& MPOL_F_STATIC_NODES
)
2846 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2847 else if (flags
& MPOL_F_RELATIVE_NODES
)
2848 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2851 if (!nodes_empty(nodes
)) {
2852 if (buffer
+ maxlen
< p
+ 2)
2855 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);