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.
499 * And we cannot move PageKsm pages sensibly or safely yet.
501 if (PageReserved(page
) || PageKsm(page
))
503 nid
= page_to_nid(page
);
504 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
507 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
508 migrate_page_add(page
, private, flags
);
511 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
512 pte_unmap_unlock(orig_pte
, ptl
);
516 static inline int check_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
517 unsigned long addr
, unsigned long end
,
518 const nodemask_t
*nodes
, unsigned long flags
,
524 pmd
= pmd_offset(pud
, addr
);
526 next
= pmd_addr_end(addr
, end
);
527 split_huge_page_pmd(vma
, addr
, pmd
);
528 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
530 if (check_pte_range(vma
, pmd
, addr
, next
, nodes
,
533 } while (pmd
++, addr
= next
, addr
!= end
);
537 static inline int check_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
538 unsigned long addr
, unsigned long end
,
539 const nodemask_t
*nodes
, unsigned long flags
,
545 pud
= pud_offset(pgd
, addr
);
547 next
= pud_addr_end(addr
, end
);
548 if (pud_none_or_clear_bad(pud
))
550 if (check_pmd_range(vma
, pud
, addr
, next
, nodes
,
553 } while (pud
++, addr
= next
, addr
!= end
);
557 static inline int check_pgd_range(struct vm_area_struct
*vma
,
558 unsigned long addr
, unsigned long end
,
559 const nodemask_t
*nodes
, unsigned long flags
,
565 pgd
= pgd_offset(vma
->vm_mm
, addr
);
567 next
= pgd_addr_end(addr
, end
);
568 if (pgd_none_or_clear_bad(pgd
))
570 if (check_pud_range(vma
, pgd
, addr
, next
, nodes
,
573 } while (pgd
++, addr
= next
, addr
!= end
);
577 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
579 * This is used to mark a range of virtual addresses to be inaccessible.
580 * These are later cleared by a NUMA hinting fault. Depending on these
581 * faults, pages may be migrated for better NUMA placement.
583 * This is assuming that NUMA faults are handled using PROT_NONE. If
584 * an architecture makes a different choice, it will need further
585 * changes to the core.
587 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
588 unsigned long addr
, unsigned long end
)
591 BUILD_BUG_ON(_PAGE_NUMA
!= _PAGE_PROTNONE
);
593 nr_updated
= change_protection(vma
, addr
, end
, vma
->vm_page_prot
, 0, 1);
595 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
600 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
601 unsigned long addr
, unsigned long end
)
605 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
608 * Check if all pages in a range are on a set of nodes.
609 * If pagelist != NULL then isolate pages from the LRU and
610 * put them on the pagelist.
612 static struct vm_area_struct
*
613 check_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
614 const nodemask_t
*nodes
, unsigned long flags
, void *private)
617 struct vm_area_struct
*first
, *vma
, *prev
;
620 first
= find_vma(mm
, start
);
622 return ERR_PTR(-EFAULT
);
624 for (vma
= first
; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
625 unsigned long endvma
= vma
->vm_end
;
629 if (vma
->vm_start
> start
)
630 start
= vma
->vm_start
;
632 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
633 if (!vma
->vm_next
&& vma
->vm_end
< end
)
634 return ERR_PTR(-EFAULT
);
635 if (prev
&& prev
->vm_end
< vma
->vm_start
)
636 return ERR_PTR(-EFAULT
);
639 if (is_vm_hugetlb_page(vma
))
642 if (flags
& MPOL_MF_LAZY
) {
643 change_prot_numa(vma
, start
, endvma
);
647 if ((flags
& MPOL_MF_STRICT
) ||
648 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
649 vma_migratable(vma
))) {
651 err
= check_pgd_range(vma
, start
, endvma
, nodes
,
654 first
= ERR_PTR(err
);
665 * Apply policy to a single VMA
666 * This must be called with the mmap_sem held for writing.
668 static int vma_replace_policy(struct vm_area_struct
*vma
,
669 struct mempolicy
*pol
)
672 struct mempolicy
*old
;
673 struct mempolicy
*new;
675 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
676 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
677 vma
->vm_ops
, vma
->vm_file
,
678 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
684 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
685 err
= vma
->vm_ops
->set_policy(vma
, new);
690 old
= vma
->vm_policy
;
691 vma
->vm_policy
= new; /* protected by mmap_sem */
700 /* Step 2: apply policy to a range and do splits. */
701 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
702 unsigned long end
, struct mempolicy
*new_pol
)
704 struct vm_area_struct
*next
;
705 struct vm_area_struct
*prev
;
706 struct vm_area_struct
*vma
;
709 unsigned long vmstart
;
712 vma
= find_vma(mm
, start
);
713 if (!vma
|| vma
->vm_start
> start
)
717 if (start
> vma
->vm_start
)
720 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
722 vmstart
= max(start
, vma
->vm_start
);
723 vmend
= min(end
, vma
->vm_end
);
725 if (mpol_equal(vma_policy(vma
), new_pol
))
728 pgoff
= vma
->vm_pgoff
+
729 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
730 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
731 vma
->anon_vma
, vma
->vm_file
, pgoff
,
738 if (vma
->vm_start
!= vmstart
) {
739 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
743 if (vma
->vm_end
!= vmend
) {
744 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 false, MIGRATE_SYNC
,
1021 putback_lru_pages(&pagelist
);
1028 * Move pages between the two nodesets so as to preserve the physical
1029 * layout as much as possible.
1031 * Returns the number of page that could not be moved.
1033 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1034 const nodemask_t
*to
, int flags
)
1040 err
= migrate_prep();
1044 down_read(&mm
->mmap_sem
);
1046 err
= migrate_vmas(mm
, from
, to
, flags
);
1051 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1052 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1053 * bit in 'tmp', and return that <source, dest> pair for migration.
1054 * The pair of nodemasks 'to' and 'from' define the map.
1056 * If no pair of bits is found that way, fallback to picking some
1057 * pair of 'source' and 'dest' bits that are not the same. If the
1058 * 'source' and 'dest' bits are the same, this represents a node
1059 * that will be migrating to itself, so no pages need move.
1061 * If no bits are left in 'tmp', or if all remaining bits left
1062 * in 'tmp' correspond to the same bit in 'to', return false
1063 * (nothing left to migrate).
1065 * This lets us pick a pair of nodes to migrate between, such that
1066 * if possible the dest node is not already occupied by some other
1067 * source node, minimizing the risk of overloading the memory on a
1068 * node that would happen if we migrated incoming memory to a node
1069 * before migrating outgoing memory source that same node.
1071 * A single scan of tmp is sufficient. As we go, we remember the
1072 * most recent <s, d> pair that moved (s != d). If we find a pair
1073 * that not only moved, but what's better, moved to an empty slot
1074 * (d is not set in tmp), then we break out then, with that pair.
1075 * Otherwise when we finish scanning from_tmp, we at least have the
1076 * most recent <s, d> pair that moved. If we get all the way through
1077 * the scan of tmp without finding any node that moved, much less
1078 * moved to an empty node, then there is nothing left worth migrating.
1082 while (!nodes_empty(tmp
)) {
1087 for_each_node_mask(s
, tmp
) {
1090 * do_migrate_pages() tries to maintain the relative
1091 * node relationship of the pages established between
1092 * threads and memory areas.
1094 * However if the number of source nodes is not equal to
1095 * the number of destination nodes we can not preserve
1096 * this node relative relationship. In that case, skip
1097 * copying memory from a node that is in the destination
1100 * Example: [2,3,4] -> [3,4,5] moves everything.
1101 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1104 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1105 (node_isset(s
, *to
)))
1108 d
= node_remap(s
, *from
, *to
);
1112 source
= s
; /* Node moved. Memorize */
1115 /* dest not in remaining from nodes? */
1116 if (!node_isset(dest
, tmp
))
1122 node_clear(source
, tmp
);
1123 err
= migrate_to_node(mm
, source
, dest
, flags
);
1130 up_read(&mm
->mmap_sem
);
1138 * Allocate a new page for page migration based on vma policy.
1139 * Start assuming that page is mapped by vma pointed to by @private.
1140 * Search forward from there, if not. N.B., this assumes that the
1141 * list of pages handed to migrate_pages()--which is how we get here--
1142 * is in virtual address order.
1144 static struct page
*new_vma_page(struct page
*page
, unsigned long private, int **x
)
1146 struct vm_area_struct
*vma
= (struct vm_area_struct
*)private;
1147 unsigned long uninitialized_var(address
);
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_vma_page(struct page
*page
, unsigned long private, 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 vm_area_struct
*vma
;
1185 struct mm_struct
*mm
= current
->mm
;
1186 struct mempolicy
*new;
1189 LIST_HEAD(pagelist
);
1191 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1193 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1196 if (start
& ~PAGE_MASK
)
1199 if (mode
== MPOL_DEFAULT
)
1200 flags
&= ~MPOL_MF_STRICT
;
1202 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1210 new = mpol_new(mode
, mode_flags
, nmask
);
1212 return PTR_ERR(new);
1214 if (flags
& MPOL_MF_LAZY
)
1215 new->flags
|= MPOL_F_MOF
;
1218 * If we are using the default policy then operation
1219 * on discontinuous address spaces is okay after all
1222 flags
|= MPOL_MF_DISCONTIG_OK
;
1224 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1225 start
, start
+ len
, mode
, mode_flags
,
1226 nmask
? nodes_addr(*nmask
)[0] : -1);
1228 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1230 err
= migrate_prep();
1235 NODEMASK_SCRATCH(scratch
);
1237 down_write(&mm
->mmap_sem
);
1239 err
= mpol_set_nodemask(new, nmask
, scratch
);
1240 task_unlock(current
);
1242 up_write(&mm
->mmap_sem
);
1245 NODEMASK_SCRATCH_FREE(scratch
);
1250 vma
= check_range(mm
, start
, end
, nmask
,
1251 flags
| MPOL_MF_INVERT
, &pagelist
);
1253 err
= PTR_ERR(vma
); /* maybe ... */
1255 err
= mbind_range(mm
, start
, end
, new);
1260 if (!list_empty(&pagelist
)) {
1261 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1262 nr_failed
= migrate_pages(&pagelist
, new_vma_page
,
1264 false, MIGRATE_SYNC
,
1265 MR_MEMPOLICY_MBIND
);
1267 putback_lru_pages(&pagelist
);
1270 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1273 putback_lru_pages(&pagelist
);
1275 up_write(&mm
->mmap_sem
);
1282 * User space interface with variable sized bitmaps for nodelists.
1285 /* Copy a node mask from user space. */
1286 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1287 unsigned long maxnode
)
1290 unsigned long nlongs
;
1291 unsigned long endmask
;
1294 nodes_clear(*nodes
);
1295 if (maxnode
== 0 || !nmask
)
1297 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1300 nlongs
= BITS_TO_LONGS(maxnode
);
1301 if ((maxnode
% BITS_PER_LONG
) == 0)
1304 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1306 /* When the user specified more nodes than supported just check
1307 if the non supported part is all zero. */
1308 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1309 if (nlongs
> PAGE_SIZE
/sizeof(long))
1311 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1313 if (get_user(t
, nmask
+ k
))
1315 if (k
== nlongs
- 1) {
1321 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1325 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1327 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1331 /* Copy a kernel node mask to user space */
1332 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1335 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1336 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1338 if (copy
> nbytes
) {
1339 if (copy
> PAGE_SIZE
)
1341 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1345 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1348 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1349 unsigned long, mode
, unsigned long __user
*, nmask
,
1350 unsigned long, maxnode
, unsigned, flags
)
1354 unsigned short mode_flags
;
1356 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1357 mode
&= ~MPOL_MODE_FLAGS
;
1358 if (mode
>= MPOL_MAX
)
1360 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1361 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1363 err
= get_nodes(&nodes
, nmask
, maxnode
);
1366 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1369 /* Set the process memory policy */
1370 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, unsigned long __user
*, nmask
,
1371 unsigned long, maxnode
)
1375 unsigned short flags
;
1377 flags
= mode
& MPOL_MODE_FLAGS
;
1378 mode
&= ~MPOL_MODE_FLAGS
;
1379 if ((unsigned int)mode
>= MPOL_MAX
)
1381 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1383 err
= get_nodes(&nodes
, nmask
, maxnode
);
1386 return do_set_mempolicy(mode
, flags
, &nodes
);
1389 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1390 const unsigned long __user
*, old_nodes
,
1391 const unsigned long __user
*, new_nodes
)
1393 const struct cred
*cred
= current_cred(), *tcred
;
1394 struct mm_struct
*mm
= NULL
;
1395 struct task_struct
*task
;
1396 nodemask_t task_nodes
;
1400 NODEMASK_SCRATCH(scratch
);
1405 old
= &scratch
->mask1
;
1406 new = &scratch
->mask2
;
1408 err
= get_nodes(old
, old_nodes
, maxnode
);
1412 err
= get_nodes(new, new_nodes
, maxnode
);
1416 /* Find the mm_struct */
1418 task
= pid
? find_task_by_vpid(pid
) : current
;
1424 get_task_struct(task
);
1429 * Check if this process has the right to modify the specified
1430 * process. The right exists if the process has administrative
1431 * capabilities, superuser privileges or the same
1432 * userid as the target process.
1434 tcred
= __task_cred(task
);
1435 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1436 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1437 !capable(CAP_SYS_NICE
)) {
1444 task_nodes
= cpuset_mems_allowed(task
);
1445 /* Is the user allowed to access the target nodes? */
1446 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1451 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1456 err
= security_task_movememory(task
);
1460 mm
= get_task_mm(task
);
1461 put_task_struct(task
);
1468 err
= do_migrate_pages(mm
, old
, new,
1469 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1473 NODEMASK_SCRATCH_FREE(scratch
);
1478 put_task_struct(task
);
1484 /* Retrieve NUMA policy */
1485 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1486 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1487 unsigned long, addr
, unsigned long, flags
)
1490 int uninitialized_var(pval
);
1493 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1496 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1501 if (policy
&& put_user(pval
, policy
))
1505 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1510 #ifdef CONFIG_COMPAT
1512 asmlinkage
long compat_sys_get_mempolicy(int __user
*policy
,
1513 compat_ulong_t __user
*nmask
,
1514 compat_ulong_t maxnode
,
1515 compat_ulong_t addr
, compat_ulong_t flags
)
1518 unsigned long __user
*nm
= NULL
;
1519 unsigned long nr_bits
, alloc_size
;
1520 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1522 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1523 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1526 nm
= compat_alloc_user_space(alloc_size
);
1528 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1530 if (!err
&& nmask
) {
1531 unsigned long copy_size
;
1532 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1533 err
= copy_from_user(bm
, nm
, copy_size
);
1534 /* ensure entire bitmap is zeroed */
1535 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1536 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1542 asmlinkage
long compat_sys_set_mempolicy(int mode
, compat_ulong_t __user
*nmask
,
1543 compat_ulong_t maxnode
)
1546 unsigned long __user
*nm
= NULL
;
1547 unsigned long nr_bits
, alloc_size
;
1548 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1550 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1551 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1554 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1555 nm
= compat_alloc_user_space(alloc_size
);
1556 err
|= copy_to_user(nm
, bm
, alloc_size
);
1562 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1565 asmlinkage
long compat_sys_mbind(compat_ulong_t start
, compat_ulong_t len
,
1566 compat_ulong_t mode
, compat_ulong_t __user
*nmask
,
1567 compat_ulong_t maxnode
, compat_ulong_t flags
)
1570 unsigned long __user
*nm
= NULL
;
1571 unsigned long nr_bits
, alloc_size
;
1574 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1575 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1578 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1579 nm
= compat_alloc_user_space(alloc_size
);
1580 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1586 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1592 * get_vma_policy(@task, @vma, @addr)
1593 * @task - task for fallback if vma policy == default
1594 * @vma - virtual memory area whose policy is sought
1595 * @addr - address in @vma for shared policy lookup
1597 * Returns effective policy for a VMA at specified address.
1598 * Falls back to @task or system default policy, as necessary.
1599 * Current or other task's task mempolicy and non-shared vma policies must be
1600 * protected by task_lock(task) by the caller.
1601 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1602 * count--added by the get_policy() vm_op, as appropriate--to protect against
1603 * freeing by another task. It is the caller's responsibility to free the
1604 * extra reference for shared policies.
1606 struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1607 struct vm_area_struct
*vma
, unsigned long addr
)
1609 struct mempolicy
*pol
= get_task_policy(task
);
1612 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1613 struct mempolicy
*vpol
= vma
->vm_ops
->get_policy(vma
,
1617 } else if (vma
->vm_policy
) {
1618 pol
= vma
->vm_policy
;
1621 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1622 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1623 * count on these policies which will be dropped by
1624 * mpol_cond_put() later
1626 if (mpol_needs_cond_ref(pol
))
1631 pol
= &default_policy
;
1635 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1637 enum zone_type dynamic_policy_zone
= policy_zone
;
1639 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1642 * if policy->v.nodes has movable memory only,
1643 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1645 * policy->v.nodes is intersect with node_states[N_MEMORY].
1646 * so if the following test faile, it implies
1647 * policy->v.nodes has movable memory only.
1649 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1650 dynamic_policy_zone
= ZONE_MOVABLE
;
1652 return zone
>= dynamic_policy_zone
;
1656 * Return a nodemask representing a mempolicy for filtering nodes for
1659 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1661 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1662 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1663 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1664 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1665 return &policy
->v
.nodes
;
1670 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1671 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1674 switch (policy
->mode
) {
1675 case MPOL_PREFERRED
:
1676 if (!(policy
->flags
& MPOL_F_LOCAL
))
1677 nd
= policy
->v
.preferred_node
;
1681 * Normally, MPOL_BIND allocations are node-local within the
1682 * allowed nodemask. However, if __GFP_THISNODE is set and the
1683 * current node isn't part of the mask, we use the zonelist for
1684 * the first node in the mask instead.
1686 if (unlikely(gfp
& __GFP_THISNODE
) &&
1687 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1688 nd
= first_node(policy
->v
.nodes
);
1693 return node_zonelist(nd
, gfp
);
1696 /* Do dynamic interleaving for a process */
1697 static unsigned interleave_nodes(struct mempolicy
*policy
)
1700 struct task_struct
*me
= current
;
1703 next
= next_node(nid
, policy
->v
.nodes
);
1704 if (next
>= MAX_NUMNODES
)
1705 next
= first_node(policy
->v
.nodes
);
1706 if (next
< MAX_NUMNODES
)
1712 * Depending on the memory policy provide a node from which to allocate the
1714 * @policy must be protected by freeing by the caller. If @policy is
1715 * the current task's mempolicy, this protection is implicit, as only the
1716 * task can change it's policy. The system default policy requires no
1719 unsigned slab_node(void)
1721 struct mempolicy
*policy
;
1724 return numa_node_id();
1726 policy
= current
->mempolicy
;
1727 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1728 return numa_node_id();
1730 switch (policy
->mode
) {
1731 case MPOL_PREFERRED
:
1733 * handled MPOL_F_LOCAL above
1735 return policy
->v
.preferred_node
;
1737 case MPOL_INTERLEAVE
:
1738 return interleave_nodes(policy
);
1742 * Follow bind policy behavior and start allocation at the
1745 struct zonelist
*zonelist
;
1747 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1748 zonelist
= &NODE_DATA(numa_node_id())->node_zonelists
[0];
1749 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1752 return zone
? zone
->node
: numa_node_id();
1760 /* Do static interleaving for a VMA with known offset. */
1761 static unsigned offset_il_node(struct mempolicy
*pol
,
1762 struct vm_area_struct
*vma
, unsigned long off
)
1764 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1770 return numa_node_id();
1771 target
= (unsigned int)off
% nnodes
;
1774 nid
= next_node(nid
, pol
->v
.nodes
);
1776 } while (c
<= target
);
1780 /* Determine a node number for interleave */
1781 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1782 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1788 * for small pages, there is no difference between
1789 * shift and PAGE_SHIFT, so the bit-shift is safe.
1790 * for huge pages, since vm_pgoff is in units of small
1791 * pages, we need to shift off the always 0 bits to get
1794 BUG_ON(shift
< PAGE_SHIFT
);
1795 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1796 off
+= (addr
- vma
->vm_start
) >> shift
;
1797 return offset_il_node(pol
, vma
, off
);
1799 return interleave_nodes(pol
);
1803 * Return the bit number of a random bit set in the nodemask.
1804 * (returns -1 if nodemask is empty)
1806 int node_random(const nodemask_t
*maskp
)
1810 w
= nodes_weight(*maskp
);
1812 bit
= bitmap_ord_to_pos(maskp
->bits
,
1813 get_random_int() % w
, MAX_NUMNODES
);
1817 #ifdef CONFIG_HUGETLBFS
1819 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1820 * @vma = virtual memory area whose policy is sought
1821 * @addr = address in @vma for shared policy lookup and interleave policy
1822 * @gfp_flags = for requested zone
1823 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1824 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1826 * Returns a zonelist suitable for a huge page allocation and a pointer
1827 * to the struct mempolicy for conditional unref after allocation.
1828 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1829 * @nodemask for filtering the zonelist.
1831 * Must be protected by get_mems_allowed()
1833 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1834 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1835 nodemask_t
**nodemask
)
1837 struct zonelist
*zl
;
1839 *mpol
= get_vma_policy(current
, vma
, addr
);
1840 *nodemask
= NULL
; /* assume !MPOL_BIND */
1842 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1843 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1844 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1846 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1847 if ((*mpol
)->mode
== MPOL_BIND
)
1848 *nodemask
= &(*mpol
)->v
.nodes
;
1854 * init_nodemask_of_mempolicy
1856 * If the current task's mempolicy is "default" [NULL], return 'false'
1857 * to indicate default policy. Otherwise, extract the policy nodemask
1858 * for 'bind' or 'interleave' policy into the argument nodemask, or
1859 * initialize the argument nodemask to contain the single node for
1860 * 'preferred' or 'local' policy and return 'true' to indicate presence
1861 * of non-default mempolicy.
1863 * We don't bother with reference counting the mempolicy [mpol_get/put]
1864 * because the current task is examining it's own mempolicy and a task's
1865 * mempolicy is only ever changed by the task itself.
1867 * N.B., it is the caller's responsibility to free a returned nodemask.
1869 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1871 struct mempolicy
*mempolicy
;
1874 if (!(mask
&& current
->mempolicy
))
1878 mempolicy
= current
->mempolicy
;
1879 switch (mempolicy
->mode
) {
1880 case MPOL_PREFERRED
:
1881 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1882 nid
= numa_node_id();
1884 nid
= mempolicy
->v
.preferred_node
;
1885 init_nodemask_of_node(mask
, nid
);
1890 case MPOL_INTERLEAVE
:
1891 *mask
= mempolicy
->v
.nodes
;
1897 task_unlock(current
);
1904 * mempolicy_nodemask_intersects
1906 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1907 * policy. Otherwise, check for intersection between mask and the policy
1908 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1909 * policy, always return true since it may allocate elsewhere on fallback.
1911 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1913 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1914 const nodemask_t
*mask
)
1916 struct mempolicy
*mempolicy
;
1922 mempolicy
= tsk
->mempolicy
;
1926 switch (mempolicy
->mode
) {
1927 case MPOL_PREFERRED
:
1929 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1930 * allocate from, they may fallback to other nodes when oom.
1931 * Thus, it's possible for tsk to have allocated memory from
1936 case MPOL_INTERLEAVE
:
1937 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1947 /* Allocate a page in interleaved policy.
1948 Own path because it needs to do special accounting. */
1949 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1952 struct zonelist
*zl
;
1955 zl
= node_zonelist(nid
, gfp
);
1956 page
= __alloc_pages(gfp
, order
, zl
);
1957 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1958 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1963 * alloc_pages_vma - Allocate a page for a VMA.
1966 * %GFP_USER user allocation.
1967 * %GFP_KERNEL kernel allocations,
1968 * %GFP_HIGHMEM highmem/user allocations,
1969 * %GFP_FS allocation should not call back into a file system.
1970 * %GFP_ATOMIC don't sleep.
1972 * @order:Order of the GFP allocation.
1973 * @vma: Pointer to VMA or NULL if not available.
1974 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1976 * This function allocates a page from the kernel page pool and applies
1977 * a NUMA policy associated with the VMA or the current process.
1978 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1979 * mm_struct of the VMA to prevent it from going away. Should be used for
1980 * all allocations for pages that will be mapped into
1981 * user space. Returns NULL when no page can be allocated.
1983 * Should be called with the mm_sem of the vma hold.
1986 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1987 unsigned long addr
, int node
)
1989 struct mempolicy
*pol
;
1991 unsigned int cpuset_mems_cookie
;
1994 pol
= get_vma_policy(current
, vma
, addr
);
1995 cpuset_mems_cookie
= get_mems_allowed();
1997 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
2000 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2002 page
= alloc_page_interleave(gfp
, order
, nid
);
2003 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2008 page
= __alloc_pages_nodemask(gfp
, order
,
2009 policy_zonelist(gfp
, pol
, node
),
2010 policy_nodemask(gfp
, pol
));
2011 if (unlikely(mpol_needs_cond_ref(pol
)))
2013 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2019 * alloc_pages_current - Allocate pages.
2022 * %GFP_USER user allocation,
2023 * %GFP_KERNEL kernel allocation,
2024 * %GFP_HIGHMEM highmem allocation,
2025 * %GFP_FS don't call back into a file system.
2026 * %GFP_ATOMIC don't sleep.
2027 * @order: Power of two of allocation size in pages. 0 is a single page.
2029 * Allocate a page from the kernel page pool. When not in
2030 * interrupt context and apply the current process NUMA policy.
2031 * Returns NULL when no page can be allocated.
2033 * Don't call cpuset_update_task_memory_state() unless
2034 * 1) it's ok to take cpuset_sem (can WAIT), and
2035 * 2) allocating for current task (not interrupt).
2037 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2039 struct mempolicy
*pol
= get_task_policy(current
);
2041 unsigned int cpuset_mems_cookie
;
2043 if (!pol
|| in_interrupt() || (gfp
& __GFP_THISNODE
))
2044 pol
= &default_policy
;
2047 cpuset_mems_cookie
= get_mems_allowed();
2050 * No reference counting needed for current->mempolicy
2051 * nor system default_policy
2053 if (pol
->mode
== MPOL_INTERLEAVE
)
2054 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2056 page
= __alloc_pages_nodemask(gfp
, order
,
2057 policy_zonelist(gfp
, pol
, numa_node_id()),
2058 policy_nodemask(gfp
, pol
));
2060 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2065 EXPORT_SYMBOL(alloc_pages_current
);
2068 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2069 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2070 * with the mems_allowed returned by cpuset_mems_allowed(). This
2071 * keeps mempolicies cpuset relative after its cpuset moves. See
2072 * further kernel/cpuset.c update_nodemask().
2074 * current's mempolicy may be rebinded by the other task(the task that changes
2075 * cpuset's mems), so we needn't do rebind work for current task.
2078 /* Slow path of a mempolicy duplicate */
2079 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2081 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2084 return ERR_PTR(-ENOMEM
);
2086 /* task's mempolicy is protected by alloc_lock */
2087 if (old
== current
->mempolicy
) {
2090 task_unlock(current
);
2095 if (current_cpuset_is_being_rebound()) {
2096 nodemask_t mems
= cpuset_mems_allowed(current
);
2097 if (new->flags
& MPOL_F_REBINDING
)
2098 mpol_rebind_policy(new, &mems
, MPOL_REBIND_STEP2
);
2100 mpol_rebind_policy(new, &mems
, MPOL_REBIND_ONCE
);
2103 atomic_set(&new->refcnt
, 1);
2107 /* Slow path of a mempolicy comparison */
2108 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2112 if (a
->mode
!= b
->mode
)
2114 if (a
->flags
!= b
->flags
)
2116 if (mpol_store_user_nodemask(a
))
2117 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2123 case MPOL_INTERLEAVE
:
2124 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2125 case MPOL_PREFERRED
:
2126 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2134 * Shared memory backing store policy support.
2136 * Remember policies even when nobody has shared memory mapped.
2137 * The policies are kept in Red-Black tree linked from the inode.
2138 * They are protected by the sp->lock spinlock, which should be held
2139 * for any accesses to the tree.
2142 /* lookup first element intersecting start-end */
2143 /* Caller holds sp->lock */
2144 static struct sp_node
*
2145 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2147 struct rb_node
*n
= sp
->root
.rb_node
;
2150 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2152 if (start
>= p
->end
)
2154 else if (end
<= p
->start
)
2162 struct sp_node
*w
= NULL
;
2163 struct rb_node
*prev
= rb_prev(n
);
2166 w
= rb_entry(prev
, struct sp_node
, nd
);
2167 if (w
->end
<= start
)
2171 return rb_entry(n
, struct sp_node
, nd
);
2174 /* Insert a new shared policy into the list. */
2175 /* Caller holds sp->lock */
2176 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2178 struct rb_node
**p
= &sp
->root
.rb_node
;
2179 struct rb_node
*parent
= NULL
;
2184 nd
= rb_entry(parent
, struct sp_node
, nd
);
2185 if (new->start
< nd
->start
)
2187 else if (new->end
> nd
->end
)
2188 p
= &(*p
)->rb_right
;
2192 rb_link_node(&new->nd
, parent
, p
);
2193 rb_insert_color(&new->nd
, &sp
->root
);
2194 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2195 new->policy
? new->policy
->mode
: 0);
2198 /* Find shared policy intersecting idx */
2200 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2202 struct mempolicy
*pol
= NULL
;
2205 if (!sp
->root
.rb_node
)
2207 spin_lock(&sp
->lock
);
2208 sn
= sp_lookup(sp
, idx
, idx
+1);
2210 mpol_get(sn
->policy
);
2213 spin_unlock(&sp
->lock
);
2217 static void sp_free(struct sp_node
*n
)
2219 mpol_put(n
->policy
);
2220 kmem_cache_free(sn_cache
, n
);
2224 * mpol_misplaced - check whether current page node is valid in policy
2226 * @page - page to be checked
2227 * @vma - vm area where page mapped
2228 * @addr - virtual address where page mapped
2230 * Lookup current policy node id for vma,addr and "compare to" page's
2234 * -1 - not misplaced, page is in the right node
2235 * node - node id where the page should be
2237 * Policy determination "mimics" alloc_page_vma().
2238 * Called from fault path where we know the vma and faulting address.
2240 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2242 struct mempolicy
*pol
;
2244 int curnid
= page_to_nid(page
);
2245 unsigned long pgoff
;
2251 pol
= get_vma_policy(current
, vma
, addr
);
2252 if (!(pol
->flags
& MPOL_F_MOF
))
2255 switch (pol
->mode
) {
2256 case MPOL_INTERLEAVE
:
2257 BUG_ON(addr
>= vma
->vm_end
);
2258 BUG_ON(addr
< vma
->vm_start
);
2260 pgoff
= vma
->vm_pgoff
;
2261 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2262 polnid
= offset_il_node(pol
, vma
, pgoff
);
2265 case MPOL_PREFERRED
:
2266 if (pol
->flags
& MPOL_F_LOCAL
)
2267 polnid
= numa_node_id();
2269 polnid
= pol
->v
.preferred_node
;
2274 * allows binding to multiple nodes.
2275 * use current page if in policy nodemask,
2276 * else select nearest allowed node, if any.
2277 * If no allowed nodes, use current [!misplaced].
2279 if (node_isset(curnid
, pol
->v
.nodes
))
2281 (void)first_zones_zonelist(
2282 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2283 gfp_zone(GFP_HIGHUSER
),
2284 &pol
->v
.nodes
, &zone
);
2285 polnid
= zone
->node
;
2292 /* Migrate the page towards the node whose CPU is referencing it */
2293 if (pol
->flags
& MPOL_F_MORON
) {
2296 polnid
= numa_node_id();
2299 * Multi-stage node selection is used in conjunction
2300 * with a periodic migration fault to build a temporal
2301 * task<->page relation. By using a two-stage filter we
2302 * remove short/unlikely relations.
2304 * Using P(p) ~ n_p / n_t as per frequentist
2305 * probability, we can equate a task's usage of a
2306 * particular page (n_p) per total usage of this
2307 * page (n_t) (in a given time-span) to a probability.
2309 * Our periodic faults will sample this probability and
2310 * getting the same result twice in a row, given these
2311 * samples are fully independent, is then given by
2312 * P(n)^2, provided our sample period is sufficiently
2313 * short compared to the usage pattern.
2315 * This quadric squishes small probabilities, making
2316 * it less likely we act on an unlikely task<->page
2319 last_nid
= page_nid_xchg_last(page
, polnid
);
2320 if (last_nid
!= polnid
)
2324 if (curnid
!= polnid
)
2332 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2334 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2335 rb_erase(&n
->nd
, &sp
->root
);
2339 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2340 unsigned long end
, struct mempolicy
*pol
)
2342 node
->start
= start
;
2347 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2348 struct mempolicy
*pol
)
2351 struct mempolicy
*newpol
;
2353 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2357 newpol
= mpol_dup(pol
);
2358 if (IS_ERR(newpol
)) {
2359 kmem_cache_free(sn_cache
, n
);
2362 newpol
->flags
|= MPOL_F_SHARED
;
2363 sp_node_init(n
, start
, end
, newpol
);
2368 /* Replace a policy range. */
2369 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2370 unsigned long end
, struct sp_node
*new)
2373 struct sp_node
*n_new
= NULL
;
2374 struct mempolicy
*mpol_new
= NULL
;
2378 spin_lock(&sp
->lock
);
2379 n
= sp_lookup(sp
, start
, end
);
2380 /* Take care of old policies in the same range. */
2381 while (n
&& n
->start
< end
) {
2382 struct rb_node
*next
= rb_next(&n
->nd
);
2383 if (n
->start
>= start
) {
2389 /* Old policy spanning whole new range. */
2394 *mpol_new
= *n
->policy
;
2395 atomic_set(&mpol_new
->refcnt
, 1);
2396 sp_node_init(n_new
, n
->end
, end
, mpol_new
);
2397 sp_insert(sp
, n_new
);
2407 n
= rb_entry(next
, struct sp_node
, nd
);
2411 spin_unlock(&sp
->lock
);
2418 kmem_cache_free(sn_cache
, n_new
);
2423 spin_unlock(&sp
->lock
);
2425 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2428 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2435 * mpol_shared_policy_init - initialize shared policy for inode
2436 * @sp: pointer to inode shared policy
2437 * @mpol: struct mempolicy to install
2439 * Install non-NULL @mpol in inode's shared policy rb-tree.
2440 * On entry, the current task has a reference on a non-NULL @mpol.
2441 * This must be released on exit.
2442 * This is called at get_inode() calls and we can use GFP_KERNEL.
2444 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2448 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2449 spin_lock_init(&sp
->lock
);
2452 struct vm_area_struct pvma
;
2453 struct mempolicy
*new;
2454 NODEMASK_SCRATCH(scratch
);
2458 /* contextualize the tmpfs mount point mempolicy */
2459 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2461 goto free_scratch
; /* no valid nodemask intersection */
2464 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2465 task_unlock(current
);
2469 /* Create pseudo-vma that contains just the policy */
2470 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2471 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2472 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2475 mpol_put(new); /* drop initial ref */
2477 NODEMASK_SCRATCH_FREE(scratch
);
2479 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2483 int mpol_set_shared_policy(struct shared_policy
*info
,
2484 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2487 struct sp_node
*new = NULL
;
2488 unsigned long sz
= vma_pages(vma
);
2490 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2492 sz
, npol
? npol
->mode
: -1,
2493 npol
? npol
->flags
: -1,
2494 npol
? nodes_addr(npol
->v
.nodes
)[0] : -1);
2497 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2501 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2507 /* Free a backing policy store on inode delete. */
2508 void mpol_free_shared_policy(struct shared_policy
*p
)
2511 struct rb_node
*next
;
2513 if (!p
->root
.rb_node
)
2515 spin_lock(&p
->lock
);
2516 next
= rb_first(&p
->root
);
2518 n
= rb_entry(next
, struct sp_node
, nd
);
2519 next
= rb_next(&n
->nd
);
2522 spin_unlock(&p
->lock
);
2525 #ifdef CONFIG_NUMA_BALANCING
2526 static bool __initdata numabalancing_override
;
2528 static void __init
check_numabalancing_enable(void)
2530 bool numabalancing_default
= false;
2532 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2533 numabalancing_default
= true;
2535 if (nr_node_ids
> 1 && !numabalancing_override
) {
2536 printk(KERN_INFO
"Enabling automatic NUMA balancing. "
2537 "Configure with numa_balancing= or sysctl");
2538 set_numabalancing_state(numabalancing_default
);
2542 static int __init
setup_numabalancing(char *str
)
2547 numabalancing_override
= true;
2549 if (!strcmp(str
, "enable")) {
2550 set_numabalancing_state(true);
2552 } else if (!strcmp(str
, "disable")) {
2553 set_numabalancing_state(false);
2558 printk(KERN_WARNING
"Unable to parse numa_balancing=\n");
2562 __setup("numa_balancing=", setup_numabalancing
);
2564 static inline void __init
check_numabalancing_enable(void)
2567 #endif /* CONFIG_NUMA_BALANCING */
2569 /* assumes fs == KERNEL_DS */
2570 void __init
numa_policy_init(void)
2572 nodemask_t interleave_nodes
;
2573 unsigned long largest
= 0;
2574 int nid
, prefer
= 0;
2576 policy_cache
= kmem_cache_create("numa_policy",
2577 sizeof(struct mempolicy
),
2578 0, SLAB_PANIC
, NULL
);
2580 sn_cache
= kmem_cache_create("shared_policy_node",
2581 sizeof(struct sp_node
),
2582 0, SLAB_PANIC
, NULL
);
2584 for_each_node(nid
) {
2585 preferred_node_policy
[nid
] = (struct mempolicy
) {
2586 .refcnt
= ATOMIC_INIT(1),
2587 .mode
= MPOL_PREFERRED
,
2588 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2589 .v
= { .preferred_node
= nid
, },
2594 * Set interleaving policy for system init. Interleaving is only
2595 * enabled across suitably sized nodes (default is >= 16MB), or
2596 * fall back to the largest node if they're all smaller.
2598 nodes_clear(interleave_nodes
);
2599 for_each_node_state(nid
, N_MEMORY
) {
2600 unsigned long total_pages
= node_present_pages(nid
);
2602 /* Preserve the largest node */
2603 if (largest
< total_pages
) {
2604 largest
= total_pages
;
2608 /* Interleave this node? */
2609 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2610 node_set(nid
, interleave_nodes
);
2613 /* All too small, use the largest */
2614 if (unlikely(nodes_empty(interleave_nodes
)))
2615 node_set(prefer
, interleave_nodes
);
2617 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2618 printk("numa_policy_init: interleaving failed\n");
2620 check_numabalancing_enable();
2623 /* Reset policy of current process to default */
2624 void numa_default_policy(void)
2626 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2630 * Parse and format mempolicy from/to strings
2634 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2636 static const char * const policy_modes
[] =
2638 [MPOL_DEFAULT
] = "default",
2639 [MPOL_PREFERRED
] = "prefer",
2640 [MPOL_BIND
] = "bind",
2641 [MPOL_INTERLEAVE
] = "interleave",
2642 [MPOL_LOCAL
] = "local",
2648 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2649 * @str: string containing mempolicy to parse
2650 * @mpol: pointer to struct mempolicy pointer, returned on success.
2653 * <mode>[=<flags>][:<nodelist>]
2655 * On success, returns 0, else 1
2657 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2659 struct mempolicy
*new = NULL
;
2660 unsigned short mode
;
2661 unsigned short mode_flags
;
2663 char *nodelist
= strchr(str
, ':');
2664 char *flags
= strchr(str
, '=');
2668 /* NUL-terminate mode or flags string */
2670 if (nodelist_parse(nodelist
, nodes
))
2672 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2678 *flags
++ = '\0'; /* terminate mode string */
2680 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2681 if (!strcmp(str
, policy_modes
[mode
])) {
2685 if (mode
>= MPOL_MAX
)
2689 case MPOL_PREFERRED
:
2691 * Insist on a nodelist of one node only
2694 char *rest
= nodelist
;
2695 while (isdigit(*rest
))
2701 case MPOL_INTERLEAVE
:
2703 * Default to online nodes with memory if no nodelist
2706 nodes
= node_states
[N_MEMORY
];
2710 * Don't allow a nodelist; mpol_new() checks flags
2714 mode
= MPOL_PREFERRED
;
2718 * Insist on a empty nodelist
2725 * Insist on a nodelist
2734 * Currently, we only support two mutually exclusive
2737 if (!strcmp(flags
, "static"))
2738 mode_flags
|= MPOL_F_STATIC_NODES
;
2739 else if (!strcmp(flags
, "relative"))
2740 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2745 new = mpol_new(mode
, mode_flags
, &nodes
);
2750 * Save nodes for mpol_to_str() to show the tmpfs mount options
2751 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2753 if (mode
!= MPOL_PREFERRED
)
2754 new->v
.nodes
= nodes
;
2756 new->v
.preferred_node
= first_node(nodes
);
2758 new->flags
|= MPOL_F_LOCAL
;
2761 * Save nodes for contextualization: this will be used to "clone"
2762 * the mempolicy in a specific context [cpuset] at a later time.
2764 new->w
.user_nodemask
= nodes
;
2769 /* Restore string for error message */
2778 #endif /* CONFIG_TMPFS */
2781 * mpol_to_str - format a mempolicy structure for printing
2782 * @buffer: to contain formatted mempolicy string
2783 * @maxlen: length of @buffer
2784 * @pol: pointer to mempolicy to be formatted
2786 * Convert a mempolicy into a string.
2787 * Returns the number of characters in buffer (if positive)
2788 * or an error (negative)
2790 int mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2795 unsigned short mode
;
2796 unsigned short flags
= pol
? pol
->flags
: 0;
2799 * Sanity check: room for longest mode, flag and some nodes
2801 VM_BUG_ON(maxlen
< strlen("interleave") + strlen("relative") + 16);
2803 if (!pol
|| pol
== &default_policy
)
2804 mode
= MPOL_DEFAULT
;
2813 case MPOL_PREFERRED
:
2815 if (flags
& MPOL_F_LOCAL
)
2818 node_set(pol
->v
.preferred_node
, nodes
);
2823 case MPOL_INTERLEAVE
:
2824 nodes
= pol
->v
.nodes
;
2831 l
= strlen(policy_modes
[mode
]);
2832 if (buffer
+ maxlen
< p
+ l
+ 1)
2835 strcpy(p
, policy_modes
[mode
]);
2838 if (flags
& MPOL_MODE_FLAGS
) {
2839 if (buffer
+ maxlen
< p
+ 2)
2844 * Currently, the only defined flags are mutually exclusive
2846 if (flags
& MPOL_F_STATIC_NODES
)
2847 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2848 else if (flags
& MPOL_F_RELATIVE_NODES
)
2849 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2852 if (!nodes_empty(nodes
)) {
2853 if (buffer
+ maxlen
< p
+ 2)
2856 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);