2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
68 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/export.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/ksm.h>
90 #include <linux/rmap.h>
91 #include <linux/security.h>
92 #include <linux/syscalls.h>
93 #include <linux/ctype.h>
94 #include <linux/mm_inline.h>
95 #include <linux/mmu_notifier.h>
96 #include <linux/printk.h>
98 #include <asm/tlbflush.h>
99 #include <asm/uaccess.h>
100 #include <linux/random.h>
102 #include "internal.h"
105 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
106 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
108 static struct kmem_cache
*policy_cache
;
109 static struct kmem_cache
*sn_cache
;
111 /* Highest zone. An specific allocation for a zone below that is not
113 enum zone_type policy_zone
= 0;
116 * run-time system-wide default policy => local allocation
118 static struct mempolicy default_policy
= {
119 .refcnt
= ATOMIC_INIT(1), /* never free it */
120 .mode
= MPOL_PREFERRED
,
121 .flags
= MPOL_F_LOCAL
,
124 static struct mempolicy preferred_node_policy
[MAX_NUMNODES
];
126 struct mempolicy
*get_task_policy(struct task_struct
*p
)
128 struct mempolicy
*pol
= p
->mempolicy
;
134 node
= numa_node_id();
135 if (node
!= NUMA_NO_NODE
) {
136 pol
= &preferred_node_policy
[node
];
137 /* preferred_node_policy is not initialised early in boot */
142 return &default_policy
;
145 static const struct mempolicy_operations
{
146 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
153 * If we have a lock to protect task->mempolicy in read-side, we do
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
161 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
,
162 enum mpol_rebind_step step
);
163 } mpol_ops
[MPOL_MAX
];
165 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
167 return pol
->flags
& MPOL_MODE_FLAGS
;
170 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
171 const nodemask_t
*rel
)
174 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
175 nodes_onto(*ret
, tmp
, *rel
);
178 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
180 if (nodes_empty(*nodes
))
182 pol
->v
.nodes
= *nodes
;
186 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
189 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
190 else if (nodes_empty(*nodes
))
191 return -EINVAL
; /* no allowed nodes */
193 pol
->v
.preferred_node
= first_node(*nodes
);
197 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
199 if (nodes_empty(*nodes
))
201 pol
->v
.nodes
= *nodes
;
206 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
207 * any, for the new policy. mpol_new() has already validated the nodes
208 * parameter with respect to the policy mode and flags. But, we need to
209 * handle an empty nodemask with MPOL_PREFERRED here.
211 * Must be called holding task's alloc_lock to protect task's mems_allowed
212 * and mempolicy. May also be called holding the mmap_semaphore for write.
214 static int mpol_set_nodemask(struct mempolicy
*pol
,
215 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
219 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
223 nodes_and(nsc
->mask1
,
224 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
227 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
228 nodes
= NULL
; /* explicit local allocation */
230 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
231 mpol_relative_nodemask(&nsc
->mask2
, nodes
, &nsc
->mask1
);
233 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
235 if (mpol_store_user_nodemask(pol
))
236 pol
->w
.user_nodemask
= *nodes
;
238 pol
->w
.cpuset_mems_allowed
=
239 cpuset_current_mems_allowed
;
243 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
245 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
250 * This function just creates a new policy, does some check and simple
251 * initialization. You must invoke mpol_set_nodemask() to set nodes.
253 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
256 struct mempolicy
*policy
;
258 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
259 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
261 if (mode
== MPOL_DEFAULT
) {
262 if (nodes
&& !nodes_empty(*nodes
))
263 return ERR_PTR(-EINVAL
);
269 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
270 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
271 * All other modes require a valid pointer to a non-empty nodemask.
273 if (mode
== MPOL_PREFERRED
) {
274 if (nodes_empty(*nodes
)) {
275 if (((flags
& MPOL_F_STATIC_NODES
) ||
276 (flags
& MPOL_F_RELATIVE_NODES
)))
277 return ERR_PTR(-EINVAL
);
279 } else if (mode
== MPOL_LOCAL
) {
280 if (!nodes_empty(*nodes
))
281 return ERR_PTR(-EINVAL
);
282 mode
= MPOL_PREFERRED
;
283 } else if (nodes_empty(*nodes
))
284 return ERR_PTR(-EINVAL
);
285 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
287 return ERR_PTR(-ENOMEM
);
288 atomic_set(&policy
->refcnt
, 1);
290 policy
->flags
= flags
;
295 /* Slow path of a mpol destructor. */
296 void __mpol_put(struct mempolicy
*p
)
298 if (!atomic_dec_and_test(&p
->refcnt
))
300 kmem_cache_free(policy_cache
, p
);
303 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
,
304 enum mpol_rebind_step step
)
310 * MPOL_REBIND_ONCE - do rebind work at once
311 * MPOL_REBIND_STEP1 - set all the newly nodes
312 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
314 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
,
315 enum mpol_rebind_step step
)
319 if (pol
->flags
& MPOL_F_STATIC_NODES
)
320 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
321 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
322 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
325 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
328 if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP1
) {
329 nodes_remap(tmp
, pol
->v
.nodes
,
330 pol
->w
.cpuset_mems_allowed
, *nodes
);
331 pol
->w
.cpuset_mems_allowed
= step
? tmp
: *nodes
;
332 } else if (step
== MPOL_REBIND_STEP2
) {
333 tmp
= pol
->w
.cpuset_mems_allowed
;
334 pol
->w
.cpuset_mems_allowed
= *nodes
;
339 if (nodes_empty(tmp
))
342 if (step
== MPOL_REBIND_STEP1
)
343 nodes_or(pol
->v
.nodes
, pol
->v
.nodes
, tmp
);
344 else if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP2
)
349 if (!node_isset(current
->il_next
, tmp
)) {
350 current
->il_next
= next_node(current
->il_next
, tmp
);
351 if (current
->il_next
>= MAX_NUMNODES
)
352 current
->il_next
= first_node(tmp
);
353 if (current
->il_next
>= MAX_NUMNODES
)
354 current
->il_next
= numa_node_id();
358 static void mpol_rebind_preferred(struct mempolicy
*pol
,
359 const nodemask_t
*nodes
,
360 enum mpol_rebind_step step
)
364 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
365 int node
= first_node(pol
->w
.user_nodemask
);
367 if (node_isset(node
, *nodes
)) {
368 pol
->v
.preferred_node
= node
;
369 pol
->flags
&= ~MPOL_F_LOCAL
;
371 pol
->flags
|= MPOL_F_LOCAL
;
372 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
373 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
374 pol
->v
.preferred_node
= first_node(tmp
);
375 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
376 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
377 pol
->w
.cpuset_mems_allowed
,
379 pol
->w
.cpuset_mems_allowed
= *nodes
;
384 * mpol_rebind_policy - Migrate a policy to a different set of nodes
386 * If read-side task has no lock to protect task->mempolicy, write-side
387 * task will rebind the task->mempolicy by two step. The first step is
388 * setting all the newly nodes, and the second step is cleaning all the
389 * disallowed nodes. In this way, we can avoid finding no node to alloc
391 * If we have a lock to protect task->mempolicy in read-side, we do
395 * MPOL_REBIND_ONCE - do rebind work at once
396 * MPOL_REBIND_STEP1 - set all the newly nodes
397 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
399 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
,
400 enum mpol_rebind_step step
)
404 if (!mpol_store_user_nodemask(pol
) && step
== MPOL_REBIND_ONCE
&&
405 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
408 if (step
== MPOL_REBIND_STEP1
&& (pol
->flags
& MPOL_F_REBINDING
))
411 if (step
== MPOL_REBIND_STEP2
&& !(pol
->flags
& MPOL_F_REBINDING
))
414 if (step
== MPOL_REBIND_STEP1
)
415 pol
->flags
|= MPOL_F_REBINDING
;
416 else if (step
== MPOL_REBIND_STEP2
)
417 pol
->flags
&= ~MPOL_F_REBINDING
;
418 else if (step
>= MPOL_REBIND_NSTEP
)
421 mpol_ops
[pol
->mode
].rebind(pol
, newmask
, step
);
425 * Wrapper for mpol_rebind_policy() that just requires task
426 * pointer, and updates task mempolicy.
428 * Called with task's alloc_lock held.
431 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new,
432 enum mpol_rebind_step step
)
434 mpol_rebind_policy(tsk
->mempolicy
, new, step
);
438 * Rebind each vma in mm to new nodemask.
440 * Call holding a reference to mm. Takes mm->mmap_sem during call.
443 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
445 struct vm_area_struct
*vma
;
447 down_write(&mm
->mmap_sem
);
448 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
449 mpol_rebind_policy(vma
->vm_policy
, new, MPOL_REBIND_ONCE
);
450 up_write(&mm
->mmap_sem
);
453 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
455 .rebind
= mpol_rebind_default
,
457 [MPOL_INTERLEAVE
] = {
458 .create
= mpol_new_interleave
,
459 .rebind
= mpol_rebind_nodemask
,
462 .create
= mpol_new_preferred
,
463 .rebind
= mpol_rebind_preferred
,
466 .create
= mpol_new_bind
,
467 .rebind
= mpol_rebind_nodemask
,
471 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
472 unsigned long flags
);
475 struct list_head
*pagelist
;
478 struct vm_area_struct
*prev
;
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
485 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
486 unsigned long end
, struct mm_walk
*walk
)
488 struct vm_area_struct
*vma
= walk
->vma
;
490 struct queue_pages
*qp
= walk
->private;
491 unsigned long flags
= qp
->flags
;
496 split_huge_page_pmd(vma
, addr
, pmd
);
497 if (pmd_trans_unstable(pmd
))
500 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
501 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
502 if (!pte_present(*pte
))
504 page
= vm_normal_page(vma
, addr
, *pte
);
508 * vm_normal_page() filters out zero pages, but there might
509 * still be PageReserved pages to skip, perhaps in a VDSO.
511 if (PageReserved(page
))
513 nid
= page_to_nid(page
);
514 if (node_isset(nid
, *qp
->nmask
) == !!(flags
& MPOL_MF_INVERT
))
517 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
518 migrate_page_add(page
, qp
->pagelist
, flags
);
520 pte_unmap_unlock(pte
- 1, ptl
);
525 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
526 unsigned long addr
, unsigned long end
,
527 struct mm_walk
*walk
)
529 #ifdef CONFIG_HUGETLB_PAGE
530 struct queue_pages
*qp
= walk
->private;
531 unsigned long flags
= qp
->flags
;
537 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
538 entry
= huge_ptep_get(pte
);
539 if (!pte_present(entry
))
541 page
= pte_page(entry
);
542 nid
= page_to_nid(page
);
543 if (node_isset(nid
, *qp
->nmask
) == !!(flags
& MPOL_MF_INVERT
))
545 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
546 if (flags
& (MPOL_MF_MOVE_ALL
) ||
547 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
548 isolate_huge_page(page
, qp
->pagelist
);
557 #ifdef CONFIG_NUMA_BALANCING
559 * This is used to mark a range of virtual addresses to be inaccessible.
560 * These are later cleared by a NUMA hinting fault. Depending on these
561 * faults, pages may be migrated for better NUMA placement.
563 * This is assuming that NUMA faults are handled using PROT_NONE. If
564 * an architecture makes a different choice, it will need further
565 * changes to the core.
567 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
568 unsigned long addr
, unsigned long end
)
572 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, 0, 1);
574 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
579 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
580 unsigned long addr
, unsigned long end
)
584 #endif /* CONFIG_NUMA_BALANCING */
586 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
587 struct mm_walk
*walk
)
589 struct vm_area_struct
*vma
= walk
->vma
;
590 struct queue_pages
*qp
= walk
->private;
591 unsigned long endvma
= vma
->vm_end
;
592 unsigned long flags
= qp
->flags
;
594 if (vma
->vm_flags
& VM_PFNMAP
)
599 if (vma
->vm_start
> start
)
600 start
= vma
->vm_start
;
602 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
603 if (!vma
->vm_next
&& vma
->vm_end
< end
)
605 if (qp
->prev
&& qp
->prev
->vm_end
< vma
->vm_start
)
611 if (flags
& MPOL_MF_LAZY
) {
612 /* Similar to task_numa_work, skip inaccessible VMAs */
613 if (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
))
614 change_prot_numa(vma
, start
, endvma
);
618 if ((flags
& MPOL_MF_STRICT
) ||
619 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
620 vma_migratable(vma
)))
621 /* queue pages from current vma */
627 * Walk through page tables and collect pages to be migrated.
629 * If pages found in a given range are on a set of nodes (determined by
630 * @nodes and @flags,) it's isolated and queued to the pagelist which is
631 * passed via @private.)
634 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
635 nodemask_t
*nodes
, unsigned long flags
,
636 struct list_head
*pagelist
)
638 struct queue_pages qp
= {
639 .pagelist
= pagelist
,
644 struct mm_walk queue_pages_walk
= {
645 .hugetlb_entry
= queue_pages_hugetlb
,
646 .pmd_entry
= queue_pages_pte_range
,
647 .test_walk
= queue_pages_test_walk
,
652 return walk_page_range(start
, end
, &queue_pages_walk
);
656 * Apply policy to a single VMA
657 * This must be called with the mmap_sem held for writing.
659 static int vma_replace_policy(struct vm_area_struct
*vma
,
660 struct mempolicy
*pol
)
663 struct mempolicy
*old
;
664 struct mempolicy
*new;
666 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
667 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
668 vma
->vm_ops
, vma
->vm_file
,
669 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
675 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
676 err
= vma
->vm_ops
->set_policy(vma
, new);
681 old
= vma
->vm_policy
;
682 vma
->vm_policy
= new; /* protected by mmap_sem */
691 /* Step 2: apply policy to a range and do splits. */
692 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
693 unsigned long end
, struct mempolicy
*new_pol
)
695 struct vm_area_struct
*next
;
696 struct vm_area_struct
*prev
;
697 struct vm_area_struct
*vma
;
700 unsigned long vmstart
;
703 vma
= find_vma(mm
, start
);
704 if (!vma
|| vma
->vm_start
> start
)
708 if (start
> vma
->vm_start
)
711 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
713 vmstart
= max(start
, vma
->vm_start
);
714 vmend
= min(end
, vma
->vm_end
);
716 if (mpol_equal(vma_policy(vma
), new_pol
))
719 pgoff
= vma
->vm_pgoff
+
720 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
721 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
722 vma
->anon_vma
, vma
->vm_file
, pgoff
,
723 new_pol
, vma
->vm_userfaultfd_ctx
);
727 if (mpol_equal(vma_policy(vma
), new_pol
))
729 /* vma_merge() joined vma && vma->next, case 8 */
732 if (vma
->vm_start
!= vmstart
) {
733 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
737 if (vma
->vm_end
!= vmend
) {
738 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
743 err
= vma_replace_policy(vma
, new_pol
);
752 /* Set the process memory policy */
753 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
756 struct mempolicy
*new, *old
;
757 NODEMASK_SCRATCH(scratch
);
763 new = mpol_new(mode
, flags
, nodes
);
770 ret
= mpol_set_nodemask(new, nodes
, scratch
);
772 task_unlock(current
);
776 old
= current
->mempolicy
;
777 current
->mempolicy
= new;
778 if (new && new->mode
== MPOL_INTERLEAVE
&&
779 nodes_weight(new->v
.nodes
))
780 current
->il_next
= first_node(new->v
.nodes
);
781 task_unlock(current
);
785 NODEMASK_SCRATCH_FREE(scratch
);
790 * Return nodemask for policy for get_mempolicy() query
792 * Called with task's alloc_lock held
794 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
797 if (p
== &default_policy
)
803 case MPOL_INTERLEAVE
:
807 if (!(p
->flags
& MPOL_F_LOCAL
))
808 node_set(p
->v
.preferred_node
, *nodes
);
809 /* else return empty node mask for local allocation */
816 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
821 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
823 err
= page_to_nid(p
);
829 /* Retrieve NUMA policy */
830 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
831 unsigned long addr
, unsigned long flags
)
834 struct mm_struct
*mm
= current
->mm
;
835 struct vm_area_struct
*vma
= NULL
;
836 struct mempolicy
*pol
= current
->mempolicy
;
839 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
842 if (flags
& MPOL_F_MEMS_ALLOWED
) {
843 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
845 *policy
= 0; /* just so it's initialized */
847 *nmask
= cpuset_current_mems_allowed
;
848 task_unlock(current
);
852 if (flags
& MPOL_F_ADDR
) {
854 * Do NOT fall back to task policy if the
855 * vma/shared policy at addr is NULL. We
856 * want to return MPOL_DEFAULT in this case.
858 down_read(&mm
->mmap_sem
);
859 vma
= find_vma_intersection(mm
, addr
, addr
+1);
861 up_read(&mm
->mmap_sem
);
864 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
865 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
867 pol
= vma
->vm_policy
;
872 pol
= &default_policy
; /* indicates default behavior */
874 if (flags
& MPOL_F_NODE
) {
875 if (flags
& MPOL_F_ADDR
) {
876 err
= lookup_node(mm
, addr
);
880 } else if (pol
== current
->mempolicy
&&
881 pol
->mode
== MPOL_INTERLEAVE
) {
882 *policy
= current
->il_next
;
888 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
891 * Internal mempolicy flags must be masked off before exposing
892 * the policy to userspace.
894 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
899 if (mpol_store_user_nodemask(pol
)) {
900 *nmask
= pol
->w
.user_nodemask
;
903 get_policy_nodemask(pol
, nmask
);
904 task_unlock(current
);
911 up_read(¤t
->mm
->mmap_sem
);
915 #ifdef CONFIG_MIGRATION
919 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
923 * Avoid migrating a page that is shared with others.
925 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
926 if (!isolate_lru_page(page
)) {
927 list_add_tail(&page
->lru
, pagelist
);
928 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
929 page_is_file_cache(page
));
934 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
937 return alloc_huge_page_node(page_hstate(compound_head(page
)),
940 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
945 * Migrate pages from one node to a target node.
946 * Returns error or the number of pages not migrated.
948 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
956 node_set(source
, nmask
);
959 * This does not "check" the range but isolates all pages that
960 * need migration. Between passing in the full user address
961 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
963 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
964 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
965 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
967 if (!list_empty(&pagelist
)) {
968 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
969 MIGRATE_SYNC
, MR_SYSCALL
);
971 putback_movable_pages(&pagelist
);
978 * Move pages between the two nodesets so as to preserve the physical
979 * layout as much as possible.
981 * Returns the number of page that could not be moved.
983 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
984 const nodemask_t
*to
, int flags
)
990 err
= migrate_prep();
994 down_read(&mm
->mmap_sem
);
997 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
998 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
999 * bit in 'tmp', and return that <source, dest> pair for migration.
1000 * The pair of nodemasks 'to' and 'from' define the map.
1002 * If no pair of bits is found that way, fallback to picking some
1003 * pair of 'source' and 'dest' bits that are not the same. If the
1004 * 'source' and 'dest' bits are the same, this represents a node
1005 * that will be migrating to itself, so no pages need move.
1007 * If no bits are left in 'tmp', or if all remaining bits left
1008 * in 'tmp' correspond to the same bit in 'to', return false
1009 * (nothing left to migrate).
1011 * This lets us pick a pair of nodes to migrate between, such that
1012 * if possible the dest node is not already occupied by some other
1013 * source node, minimizing the risk of overloading the memory on a
1014 * node that would happen if we migrated incoming memory to a node
1015 * before migrating outgoing memory source that same node.
1017 * A single scan of tmp is sufficient. As we go, we remember the
1018 * most recent <s, d> pair that moved (s != d). If we find a pair
1019 * that not only moved, but what's better, moved to an empty slot
1020 * (d is not set in tmp), then we break out then, with that pair.
1021 * Otherwise when we finish scanning from_tmp, we at least have the
1022 * most recent <s, d> pair that moved. If we get all the way through
1023 * the scan of tmp without finding any node that moved, much less
1024 * moved to an empty node, then there is nothing left worth migrating.
1028 while (!nodes_empty(tmp
)) {
1030 int source
= NUMA_NO_NODE
;
1033 for_each_node_mask(s
, tmp
) {
1036 * do_migrate_pages() tries to maintain the relative
1037 * node relationship of the pages established between
1038 * threads and memory areas.
1040 * However if the number of source nodes is not equal to
1041 * the number of destination nodes we can not preserve
1042 * this node relative relationship. In that case, skip
1043 * copying memory from a node that is in the destination
1046 * Example: [2,3,4] -> [3,4,5] moves everything.
1047 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1050 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1051 (node_isset(s
, *to
)))
1054 d
= node_remap(s
, *from
, *to
);
1058 source
= s
; /* Node moved. Memorize */
1061 /* dest not in remaining from nodes? */
1062 if (!node_isset(dest
, tmp
))
1065 if (source
== NUMA_NO_NODE
)
1068 node_clear(source
, tmp
);
1069 err
= migrate_to_node(mm
, source
, dest
, flags
);
1075 up_read(&mm
->mmap_sem
);
1083 * Allocate a new page for page migration based on vma policy.
1084 * Start by assuming the page is mapped by the same vma as contains @start.
1085 * Search forward from there, if not. N.B., this assumes that the
1086 * list of pages handed to migrate_pages()--which is how we get here--
1087 * is in virtual address order.
1089 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1091 struct vm_area_struct
*vma
;
1092 unsigned long uninitialized_var(address
);
1094 vma
= find_vma(current
->mm
, start
);
1096 address
= page_address_in_vma(page
, vma
);
1097 if (address
!= -EFAULT
)
1102 if (PageHuge(page
)) {
1104 return alloc_huge_page_noerr(vma
, address
, 1);
1107 * if !vma, alloc_page_vma() will use task or system default policy
1109 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1113 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1114 unsigned long flags
)
1118 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1119 const nodemask_t
*to
, int flags
)
1124 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1130 static long do_mbind(unsigned long start
, unsigned long len
,
1131 unsigned short mode
, unsigned short mode_flags
,
1132 nodemask_t
*nmask
, unsigned long flags
)
1134 struct mm_struct
*mm
= current
->mm
;
1135 struct mempolicy
*new;
1138 LIST_HEAD(pagelist
);
1140 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1142 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1145 if (start
& ~PAGE_MASK
)
1148 if (mode
== MPOL_DEFAULT
)
1149 flags
&= ~MPOL_MF_STRICT
;
1151 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1159 new = mpol_new(mode
, mode_flags
, nmask
);
1161 return PTR_ERR(new);
1163 if (flags
& MPOL_MF_LAZY
)
1164 new->flags
|= MPOL_F_MOF
;
1167 * If we are using the default policy then operation
1168 * on discontinuous address spaces is okay after all
1171 flags
|= MPOL_MF_DISCONTIG_OK
;
1173 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1174 start
, start
+ len
, mode
, mode_flags
,
1175 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1177 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1179 err
= migrate_prep();
1184 NODEMASK_SCRATCH(scratch
);
1186 down_write(&mm
->mmap_sem
);
1188 err
= mpol_set_nodemask(new, nmask
, scratch
);
1189 task_unlock(current
);
1191 up_write(&mm
->mmap_sem
);
1194 NODEMASK_SCRATCH_FREE(scratch
);
1199 err
= queue_pages_range(mm
, start
, end
, nmask
,
1200 flags
| MPOL_MF_INVERT
, &pagelist
);
1202 err
= mbind_range(mm
, start
, end
, new);
1207 if (!list_empty(&pagelist
)) {
1208 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1209 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1210 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1212 putback_movable_pages(&pagelist
);
1215 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1218 putback_movable_pages(&pagelist
);
1220 up_write(&mm
->mmap_sem
);
1227 * User space interface with variable sized bitmaps for nodelists.
1230 /* Copy a node mask from user space. */
1231 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1232 unsigned long maxnode
)
1235 unsigned long nlongs
;
1236 unsigned long endmask
;
1239 nodes_clear(*nodes
);
1240 if (maxnode
== 0 || !nmask
)
1242 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1245 nlongs
= BITS_TO_LONGS(maxnode
);
1246 if ((maxnode
% BITS_PER_LONG
) == 0)
1249 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1251 /* When the user specified more nodes than supported just check
1252 if the non supported part is all zero. */
1253 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1254 if (nlongs
> PAGE_SIZE
/sizeof(long))
1256 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1258 if (get_user(t
, nmask
+ k
))
1260 if (k
== nlongs
- 1) {
1266 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1270 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1272 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1276 /* Copy a kernel node mask to user space */
1277 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1280 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1281 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1283 if (copy
> nbytes
) {
1284 if (copy
> PAGE_SIZE
)
1286 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1290 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1293 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1294 unsigned long, mode
, const unsigned long __user
*, nmask
,
1295 unsigned long, maxnode
, unsigned, flags
)
1299 unsigned short mode_flags
;
1301 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1302 mode
&= ~MPOL_MODE_FLAGS
;
1303 if (mode
>= MPOL_MAX
)
1305 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1306 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1308 err
= get_nodes(&nodes
, nmask
, maxnode
);
1311 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1314 /* Set the process memory policy */
1315 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1316 unsigned long, maxnode
)
1320 unsigned short flags
;
1322 flags
= mode
& MPOL_MODE_FLAGS
;
1323 mode
&= ~MPOL_MODE_FLAGS
;
1324 if ((unsigned int)mode
>= MPOL_MAX
)
1326 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1328 err
= get_nodes(&nodes
, nmask
, maxnode
);
1331 return do_set_mempolicy(mode
, flags
, &nodes
);
1334 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1335 const unsigned long __user
*, old_nodes
,
1336 const unsigned long __user
*, new_nodes
)
1338 const struct cred
*cred
= current_cred(), *tcred
;
1339 struct mm_struct
*mm
= NULL
;
1340 struct task_struct
*task
;
1341 nodemask_t task_nodes
;
1345 NODEMASK_SCRATCH(scratch
);
1350 old
= &scratch
->mask1
;
1351 new = &scratch
->mask2
;
1353 err
= get_nodes(old
, old_nodes
, maxnode
);
1357 err
= get_nodes(new, new_nodes
, maxnode
);
1361 /* Find the mm_struct */
1363 task
= pid
? find_task_by_vpid(pid
) : current
;
1369 get_task_struct(task
);
1374 * Check if this process has the right to modify the specified
1375 * process. The right exists if the process has administrative
1376 * capabilities, superuser privileges or the same
1377 * userid as the target process.
1379 tcred
= __task_cred(task
);
1380 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1381 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1382 !capable(CAP_SYS_NICE
)) {
1389 task_nodes
= cpuset_mems_allowed(task
);
1390 /* Is the user allowed to access the target nodes? */
1391 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1396 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1401 err
= security_task_movememory(task
);
1405 mm
= get_task_mm(task
);
1406 put_task_struct(task
);
1413 err
= do_migrate_pages(mm
, old
, new,
1414 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1418 NODEMASK_SCRATCH_FREE(scratch
);
1423 put_task_struct(task
);
1429 /* Retrieve NUMA policy */
1430 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1431 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1432 unsigned long, addr
, unsigned long, flags
)
1435 int uninitialized_var(pval
);
1438 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1441 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1446 if (policy
&& put_user(pval
, policy
))
1450 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1455 #ifdef CONFIG_COMPAT
1457 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1458 compat_ulong_t __user
*, nmask
,
1459 compat_ulong_t
, maxnode
,
1460 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1463 unsigned long __user
*nm
= NULL
;
1464 unsigned long nr_bits
, alloc_size
;
1465 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1467 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1468 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1471 nm
= compat_alloc_user_space(alloc_size
);
1473 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1475 if (!err
&& nmask
) {
1476 unsigned long copy_size
;
1477 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1478 err
= copy_from_user(bm
, nm
, copy_size
);
1479 /* ensure entire bitmap is zeroed */
1480 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1481 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1487 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1488 compat_ulong_t
, maxnode
)
1490 unsigned long __user
*nm
= NULL
;
1491 unsigned long nr_bits
, alloc_size
;
1492 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1494 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1495 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1498 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1500 nm
= compat_alloc_user_space(alloc_size
);
1501 if (copy_to_user(nm
, bm
, alloc_size
))
1505 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1508 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1509 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1510 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1512 unsigned long __user
*nm
= NULL
;
1513 unsigned long nr_bits
, alloc_size
;
1516 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1517 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1520 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1522 nm
= compat_alloc_user_space(alloc_size
);
1523 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1527 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1532 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1535 struct mempolicy
*pol
= NULL
;
1538 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1539 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1540 } else if (vma
->vm_policy
) {
1541 pol
= vma
->vm_policy
;
1544 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1545 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1546 * count on these policies which will be dropped by
1547 * mpol_cond_put() later
1549 if (mpol_needs_cond_ref(pol
))
1558 * get_vma_policy(@vma, @addr)
1559 * @vma: virtual memory area whose policy is sought
1560 * @addr: address in @vma for shared policy lookup
1562 * Returns effective policy for a VMA at specified address.
1563 * Falls back to current->mempolicy or system default policy, as necessary.
1564 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1565 * count--added by the get_policy() vm_op, as appropriate--to protect against
1566 * freeing by another task. It is the caller's responsibility to free the
1567 * extra reference for shared policies.
1569 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1572 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1575 pol
= get_task_policy(current
);
1580 bool vma_policy_mof(struct vm_area_struct
*vma
)
1582 struct mempolicy
*pol
;
1584 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1587 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1588 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1595 pol
= vma
->vm_policy
;
1597 pol
= get_task_policy(current
);
1599 return pol
->flags
& MPOL_F_MOF
;
1602 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1604 enum zone_type dynamic_policy_zone
= policy_zone
;
1606 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1609 * if policy->v.nodes has movable memory only,
1610 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1612 * policy->v.nodes is intersect with node_states[N_MEMORY].
1613 * so if the following test faile, it implies
1614 * policy->v.nodes has movable memory only.
1616 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1617 dynamic_policy_zone
= ZONE_MOVABLE
;
1619 return zone
>= dynamic_policy_zone
;
1623 * Return a nodemask representing a mempolicy for filtering nodes for
1626 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1628 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1629 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1630 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1631 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1632 return &policy
->v
.nodes
;
1637 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1638 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1641 switch (policy
->mode
) {
1642 case MPOL_PREFERRED
:
1643 if (!(policy
->flags
& MPOL_F_LOCAL
))
1644 nd
= policy
->v
.preferred_node
;
1648 * Normally, MPOL_BIND allocations are node-local within the
1649 * allowed nodemask. However, if __GFP_THISNODE is set and the
1650 * current node isn't part of the mask, we use the zonelist for
1651 * the first node in the mask instead.
1653 if (unlikely(gfp
& __GFP_THISNODE
) &&
1654 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1655 nd
= first_node(policy
->v
.nodes
);
1660 return node_zonelist(nd
, gfp
);
1663 /* Do dynamic interleaving for a process */
1664 static unsigned interleave_nodes(struct mempolicy
*policy
)
1667 struct task_struct
*me
= current
;
1670 next
= next_node(nid
, policy
->v
.nodes
);
1671 if (next
>= MAX_NUMNODES
)
1672 next
= first_node(policy
->v
.nodes
);
1673 if (next
< MAX_NUMNODES
)
1679 * Depending on the memory policy provide a node from which to allocate the
1682 unsigned int mempolicy_slab_node(void)
1684 struct mempolicy
*policy
;
1685 int node
= numa_mem_id();
1690 policy
= current
->mempolicy
;
1691 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1694 switch (policy
->mode
) {
1695 case MPOL_PREFERRED
:
1697 * handled MPOL_F_LOCAL above
1699 return policy
->v
.preferred_node
;
1701 case MPOL_INTERLEAVE
:
1702 return interleave_nodes(policy
);
1706 * Follow bind policy behavior and start allocation at the
1709 struct zonelist
*zonelist
;
1711 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1712 zonelist
= &NODE_DATA(node
)->node_zonelists
[0];
1713 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1716 return zone
? zone
->node
: node
;
1724 /* Do static interleaving for a VMA with known offset. */
1725 static unsigned offset_il_node(struct mempolicy
*pol
,
1726 struct vm_area_struct
*vma
, unsigned long off
)
1728 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1731 int nid
= NUMA_NO_NODE
;
1734 return numa_node_id();
1735 target
= (unsigned int)off
% nnodes
;
1738 nid
= next_node(nid
, pol
->v
.nodes
);
1740 } while (c
<= target
);
1744 /* Determine a node number for interleave */
1745 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1746 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1752 * for small pages, there is no difference between
1753 * shift and PAGE_SHIFT, so the bit-shift is safe.
1754 * for huge pages, since vm_pgoff is in units of small
1755 * pages, we need to shift off the always 0 bits to get
1758 BUG_ON(shift
< PAGE_SHIFT
);
1759 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1760 off
+= (addr
- vma
->vm_start
) >> shift
;
1761 return offset_il_node(pol
, vma
, off
);
1763 return interleave_nodes(pol
);
1767 * Return the bit number of a random bit set in the nodemask.
1768 * (returns NUMA_NO_NODE if nodemask is empty)
1770 int node_random(const nodemask_t
*maskp
)
1772 int w
, bit
= NUMA_NO_NODE
;
1774 w
= nodes_weight(*maskp
);
1776 bit
= bitmap_ord_to_pos(maskp
->bits
,
1777 get_random_int() % w
, MAX_NUMNODES
);
1781 #ifdef CONFIG_HUGETLBFS
1783 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1784 * @vma: virtual memory area whose policy is sought
1785 * @addr: address in @vma for shared policy lookup and interleave policy
1786 * @gfp_flags: for requested zone
1787 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1788 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1790 * Returns a zonelist suitable for a huge page allocation and a pointer
1791 * to the struct mempolicy for conditional unref after allocation.
1792 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1793 * @nodemask for filtering the zonelist.
1795 * Must be protected by read_mems_allowed_begin()
1797 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1798 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1799 nodemask_t
**nodemask
)
1801 struct zonelist
*zl
;
1803 *mpol
= get_vma_policy(vma
, addr
);
1804 *nodemask
= NULL
; /* assume !MPOL_BIND */
1806 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1807 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1808 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1810 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1811 if ((*mpol
)->mode
== MPOL_BIND
)
1812 *nodemask
= &(*mpol
)->v
.nodes
;
1818 * init_nodemask_of_mempolicy
1820 * If the current task's mempolicy is "default" [NULL], return 'false'
1821 * to indicate default policy. Otherwise, extract the policy nodemask
1822 * for 'bind' or 'interleave' policy into the argument nodemask, or
1823 * initialize the argument nodemask to contain the single node for
1824 * 'preferred' or 'local' policy and return 'true' to indicate presence
1825 * of non-default mempolicy.
1827 * We don't bother with reference counting the mempolicy [mpol_get/put]
1828 * because the current task is examining it's own mempolicy and a task's
1829 * mempolicy is only ever changed by the task itself.
1831 * N.B., it is the caller's responsibility to free a returned nodemask.
1833 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1835 struct mempolicy
*mempolicy
;
1838 if (!(mask
&& current
->mempolicy
))
1842 mempolicy
= current
->mempolicy
;
1843 switch (mempolicy
->mode
) {
1844 case MPOL_PREFERRED
:
1845 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1846 nid
= numa_node_id();
1848 nid
= mempolicy
->v
.preferred_node
;
1849 init_nodemask_of_node(mask
, nid
);
1854 case MPOL_INTERLEAVE
:
1855 *mask
= mempolicy
->v
.nodes
;
1861 task_unlock(current
);
1868 * mempolicy_nodemask_intersects
1870 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1871 * policy. Otherwise, check for intersection between mask and the policy
1872 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1873 * policy, always return true since it may allocate elsewhere on fallback.
1875 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1877 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1878 const nodemask_t
*mask
)
1880 struct mempolicy
*mempolicy
;
1886 mempolicy
= tsk
->mempolicy
;
1890 switch (mempolicy
->mode
) {
1891 case MPOL_PREFERRED
:
1893 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1894 * allocate from, they may fallback to other nodes when oom.
1895 * Thus, it's possible for tsk to have allocated memory from
1900 case MPOL_INTERLEAVE
:
1901 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1911 /* Allocate a page in interleaved policy.
1912 Own path because it needs to do special accounting. */
1913 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1916 struct zonelist
*zl
;
1919 zl
= node_zonelist(nid
, gfp
);
1920 page
= __alloc_pages(gfp
, order
, zl
);
1921 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1922 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1927 * alloc_pages_vma - Allocate a page for a VMA.
1930 * %GFP_USER user allocation.
1931 * %GFP_KERNEL kernel allocations,
1932 * %GFP_HIGHMEM highmem/user allocations,
1933 * %GFP_FS allocation should not call back into a file system.
1934 * %GFP_ATOMIC don't sleep.
1936 * @order:Order of the GFP allocation.
1937 * @vma: Pointer to VMA or NULL if not available.
1938 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1939 * @node: Which node to prefer for allocation (modulo policy).
1940 * @hugepage: for hugepages try only the preferred node if possible
1942 * This function allocates a page from the kernel page pool and applies
1943 * a NUMA policy associated with the VMA or the current process.
1944 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1945 * mm_struct of the VMA to prevent it from going away. Should be used for
1946 * all allocations for pages that will be mapped into user space. Returns
1947 * NULL when no page can be allocated.
1950 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1951 unsigned long addr
, int node
, bool hugepage
)
1953 struct mempolicy
*pol
;
1955 unsigned int cpuset_mems_cookie
;
1956 struct zonelist
*zl
;
1960 pol
= get_vma_policy(vma
, addr
);
1961 cpuset_mems_cookie
= read_mems_allowed_begin();
1963 if (pol
->mode
== MPOL_INTERLEAVE
) {
1966 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
1968 page
= alloc_page_interleave(gfp
, order
, nid
);
1972 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
1973 int hpage_node
= node
;
1976 * For hugepage allocation and non-interleave policy which
1977 * allows the current node (or other explicitly preferred
1978 * node) we only try to allocate from the current/preferred
1979 * node and don't fall back to other nodes, as the cost of
1980 * remote accesses would likely offset THP benefits.
1982 * If the policy is interleave, or does not allow the current
1983 * node in its nodemask, we allocate the standard way.
1985 if (pol
->mode
== MPOL_PREFERRED
&&
1986 !(pol
->flags
& MPOL_F_LOCAL
))
1987 hpage_node
= pol
->v
.preferred_node
;
1989 nmask
= policy_nodemask(gfp
, pol
);
1990 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
1992 page
= __alloc_pages_node(hpage_node
,
1993 gfp
| __GFP_THISNODE
, order
);
1998 nmask
= policy_nodemask(gfp
, pol
);
1999 zl
= policy_zonelist(gfp
, pol
, node
);
2000 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nmask
);
2003 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2009 * alloc_pages_current - Allocate pages.
2012 * %GFP_USER user allocation,
2013 * %GFP_KERNEL kernel allocation,
2014 * %GFP_HIGHMEM highmem allocation,
2015 * %GFP_FS don't call back into a file system.
2016 * %GFP_ATOMIC don't sleep.
2017 * @order: Power of two of allocation size in pages. 0 is a single page.
2019 * Allocate a page from the kernel page pool. When not in
2020 * interrupt context and apply the current process NUMA policy.
2021 * Returns NULL when no page can be allocated.
2023 * Don't call cpuset_update_task_memory_state() unless
2024 * 1) it's ok to take cpuset_sem (can WAIT), and
2025 * 2) allocating for current task (not interrupt).
2027 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2029 struct mempolicy
*pol
= &default_policy
;
2031 unsigned int cpuset_mems_cookie
;
2033 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2034 pol
= get_task_policy(current
);
2037 cpuset_mems_cookie
= read_mems_allowed_begin();
2040 * No reference counting needed for current->mempolicy
2041 * nor system default_policy
2043 if (pol
->mode
== MPOL_INTERLEAVE
)
2044 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2046 page
= __alloc_pages_nodemask(gfp
, order
,
2047 policy_zonelist(gfp
, pol
, numa_node_id()),
2048 policy_nodemask(gfp
, pol
));
2050 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2055 EXPORT_SYMBOL(alloc_pages_current
);
2057 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2059 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2062 return PTR_ERR(pol
);
2063 dst
->vm_policy
= pol
;
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
);
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
);
2101 atomic_set(&new->refcnt
, 1);
2105 /* Slow path of a mempolicy comparison */
2106 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2110 if (a
->mode
!= b
->mode
)
2112 if (a
->flags
!= b
->flags
)
2114 if (mpol_store_user_nodemask(a
))
2115 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2121 case MPOL_INTERLEAVE
:
2122 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2123 case MPOL_PREFERRED
:
2124 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2132 * Shared memory backing store policy support.
2134 * Remember policies even when nobody has shared memory mapped.
2135 * The policies are kept in Red-Black tree linked from the inode.
2136 * They are protected by the sp->lock spinlock, which should be held
2137 * for any accesses to the tree.
2140 /* lookup first element intersecting start-end */
2141 /* Caller holds sp->lock */
2142 static struct sp_node
*
2143 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2145 struct rb_node
*n
= sp
->root
.rb_node
;
2148 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2150 if (start
>= p
->end
)
2152 else if (end
<= p
->start
)
2160 struct sp_node
*w
= NULL
;
2161 struct rb_node
*prev
= rb_prev(n
);
2164 w
= rb_entry(prev
, struct sp_node
, nd
);
2165 if (w
->end
<= start
)
2169 return rb_entry(n
, struct sp_node
, nd
);
2172 /* Insert a new shared policy into the list. */
2173 /* Caller holds sp->lock */
2174 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2176 struct rb_node
**p
= &sp
->root
.rb_node
;
2177 struct rb_node
*parent
= NULL
;
2182 nd
= rb_entry(parent
, struct sp_node
, nd
);
2183 if (new->start
< nd
->start
)
2185 else if (new->end
> nd
->end
)
2186 p
= &(*p
)->rb_right
;
2190 rb_link_node(&new->nd
, parent
, p
);
2191 rb_insert_color(&new->nd
, &sp
->root
);
2192 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2193 new->policy
? new->policy
->mode
: 0);
2196 /* Find shared policy intersecting idx */
2198 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2200 struct mempolicy
*pol
= NULL
;
2203 if (!sp
->root
.rb_node
)
2205 spin_lock(&sp
->lock
);
2206 sn
= sp_lookup(sp
, idx
, idx
+1);
2208 mpol_get(sn
->policy
);
2211 spin_unlock(&sp
->lock
);
2215 static void sp_free(struct sp_node
*n
)
2217 mpol_put(n
->policy
);
2218 kmem_cache_free(sn_cache
, n
);
2222 * mpol_misplaced - check whether current page node is valid in policy
2224 * @page: page to be checked
2225 * @vma: vm area where page mapped
2226 * @addr: virtual address where page mapped
2228 * Lookup current policy node id for vma,addr and "compare to" page's
2232 * -1 - not misplaced, page is in the right node
2233 * node - node id where the page should be
2235 * Policy determination "mimics" alloc_page_vma().
2236 * Called from fault path where we know the vma and faulting address.
2238 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2240 struct mempolicy
*pol
;
2242 int curnid
= page_to_nid(page
);
2243 unsigned long pgoff
;
2244 int thiscpu
= raw_smp_processor_id();
2245 int thisnid
= cpu_to_node(thiscpu
);
2251 pol
= get_vma_policy(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 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2300 if (curnid
!= polnid
)
2308 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2310 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2311 rb_erase(&n
->nd
, &sp
->root
);
2315 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2316 unsigned long end
, struct mempolicy
*pol
)
2318 node
->start
= start
;
2323 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2324 struct mempolicy
*pol
)
2327 struct mempolicy
*newpol
;
2329 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2333 newpol
= mpol_dup(pol
);
2334 if (IS_ERR(newpol
)) {
2335 kmem_cache_free(sn_cache
, n
);
2338 newpol
->flags
|= MPOL_F_SHARED
;
2339 sp_node_init(n
, start
, end
, newpol
);
2344 /* Replace a policy range. */
2345 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2346 unsigned long end
, struct sp_node
*new)
2349 struct sp_node
*n_new
= NULL
;
2350 struct mempolicy
*mpol_new
= NULL
;
2354 spin_lock(&sp
->lock
);
2355 n
= sp_lookup(sp
, start
, end
);
2356 /* Take care of old policies in the same range. */
2357 while (n
&& n
->start
< end
) {
2358 struct rb_node
*next
= rb_next(&n
->nd
);
2359 if (n
->start
>= start
) {
2365 /* Old policy spanning whole new range. */
2370 *mpol_new
= *n
->policy
;
2371 atomic_set(&mpol_new
->refcnt
, 1);
2372 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2374 sp_insert(sp
, n_new
);
2383 n
= rb_entry(next
, struct sp_node
, nd
);
2387 spin_unlock(&sp
->lock
);
2394 kmem_cache_free(sn_cache
, n_new
);
2399 spin_unlock(&sp
->lock
);
2401 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2404 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2411 * mpol_shared_policy_init - initialize shared policy for inode
2412 * @sp: pointer to inode shared policy
2413 * @mpol: struct mempolicy to install
2415 * Install non-NULL @mpol in inode's shared policy rb-tree.
2416 * On entry, the current task has a reference on a non-NULL @mpol.
2417 * This must be released on exit.
2418 * This is called at get_inode() calls and we can use GFP_KERNEL.
2420 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2424 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2425 spin_lock_init(&sp
->lock
);
2428 struct vm_area_struct pvma
;
2429 struct mempolicy
*new;
2430 NODEMASK_SCRATCH(scratch
);
2434 /* contextualize the tmpfs mount point mempolicy */
2435 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2437 goto free_scratch
; /* no valid nodemask intersection */
2440 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2441 task_unlock(current
);
2445 /* Create pseudo-vma that contains just the policy */
2446 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2447 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2448 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2451 mpol_put(new); /* drop initial ref */
2453 NODEMASK_SCRATCH_FREE(scratch
);
2455 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2459 int mpol_set_shared_policy(struct shared_policy
*info
,
2460 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2463 struct sp_node
*new = NULL
;
2464 unsigned long sz
= vma_pages(vma
);
2466 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2468 sz
, npol
? npol
->mode
: -1,
2469 npol
? npol
->flags
: -1,
2470 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2473 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2477 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2483 /* Free a backing policy store on inode delete. */
2484 void mpol_free_shared_policy(struct shared_policy
*p
)
2487 struct rb_node
*next
;
2489 if (!p
->root
.rb_node
)
2491 spin_lock(&p
->lock
);
2492 next
= rb_first(&p
->root
);
2494 n
= rb_entry(next
, struct sp_node
, nd
);
2495 next
= rb_next(&n
->nd
);
2498 spin_unlock(&p
->lock
);
2501 #ifdef CONFIG_NUMA_BALANCING
2502 static int __initdata numabalancing_override
;
2504 static void __init
check_numabalancing_enable(void)
2506 bool numabalancing_default
= false;
2508 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2509 numabalancing_default
= true;
2511 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2512 if (numabalancing_override
)
2513 set_numabalancing_state(numabalancing_override
== 1);
2515 if (num_online_nodes() > 1 && !numabalancing_override
) {
2516 pr_info("%s automatic NUMA balancing. "
2517 "Configure with numa_balancing= or the "
2518 "kernel.numa_balancing sysctl",
2519 numabalancing_default
? "Enabling" : "Disabling");
2520 set_numabalancing_state(numabalancing_default
);
2524 static int __init
setup_numabalancing(char *str
)
2530 if (!strcmp(str
, "enable")) {
2531 numabalancing_override
= 1;
2533 } else if (!strcmp(str
, "disable")) {
2534 numabalancing_override
= -1;
2539 pr_warn("Unable to parse numa_balancing=\n");
2543 __setup("numa_balancing=", setup_numabalancing
);
2545 static inline void __init
check_numabalancing_enable(void)
2548 #endif /* CONFIG_NUMA_BALANCING */
2550 /* assumes fs == KERNEL_DS */
2551 void __init
numa_policy_init(void)
2553 nodemask_t interleave_nodes
;
2554 unsigned long largest
= 0;
2555 int nid
, prefer
= 0;
2557 policy_cache
= kmem_cache_create("numa_policy",
2558 sizeof(struct mempolicy
),
2559 0, SLAB_PANIC
, NULL
);
2561 sn_cache
= kmem_cache_create("shared_policy_node",
2562 sizeof(struct sp_node
),
2563 0, SLAB_PANIC
, NULL
);
2565 for_each_node(nid
) {
2566 preferred_node_policy
[nid
] = (struct mempolicy
) {
2567 .refcnt
= ATOMIC_INIT(1),
2568 .mode
= MPOL_PREFERRED
,
2569 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2570 .v
= { .preferred_node
= nid
, },
2575 * Set interleaving policy for system init. Interleaving is only
2576 * enabled across suitably sized nodes (default is >= 16MB), or
2577 * fall back to the largest node if they're all smaller.
2579 nodes_clear(interleave_nodes
);
2580 for_each_node_state(nid
, N_MEMORY
) {
2581 unsigned long total_pages
= node_present_pages(nid
);
2583 /* Preserve the largest node */
2584 if (largest
< total_pages
) {
2585 largest
= total_pages
;
2589 /* Interleave this node? */
2590 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2591 node_set(nid
, interleave_nodes
);
2594 /* All too small, use the largest */
2595 if (unlikely(nodes_empty(interleave_nodes
)))
2596 node_set(prefer
, interleave_nodes
);
2598 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2599 pr_err("%s: interleaving failed\n", __func__
);
2601 check_numabalancing_enable();
2604 /* Reset policy of current process to default */
2605 void numa_default_policy(void)
2607 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2611 * Parse and format mempolicy from/to strings
2615 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2617 static const char * const policy_modes
[] =
2619 [MPOL_DEFAULT
] = "default",
2620 [MPOL_PREFERRED
] = "prefer",
2621 [MPOL_BIND
] = "bind",
2622 [MPOL_INTERLEAVE
] = "interleave",
2623 [MPOL_LOCAL
] = "local",
2629 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2630 * @str: string containing mempolicy to parse
2631 * @mpol: pointer to struct mempolicy pointer, returned on success.
2634 * <mode>[=<flags>][:<nodelist>]
2636 * On success, returns 0, else 1
2638 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2640 struct mempolicy
*new = NULL
;
2641 unsigned short mode
;
2642 unsigned short mode_flags
;
2644 char *nodelist
= strchr(str
, ':');
2645 char *flags
= strchr(str
, '=');
2649 /* NUL-terminate mode or flags string */
2651 if (nodelist_parse(nodelist
, nodes
))
2653 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2659 *flags
++ = '\0'; /* terminate mode string */
2661 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2662 if (!strcmp(str
, policy_modes
[mode
])) {
2666 if (mode
>= MPOL_MAX
)
2670 case MPOL_PREFERRED
:
2672 * Insist on a nodelist of one node only
2675 char *rest
= nodelist
;
2676 while (isdigit(*rest
))
2682 case MPOL_INTERLEAVE
:
2684 * Default to online nodes with memory if no nodelist
2687 nodes
= node_states
[N_MEMORY
];
2691 * Don't allow a nodelist; mpol_new() checks flags
2695 mode
= MPOL_PREFERRED
;
2699 * Insist on a empty nodelist
2706 * Insist on a nodelist
2715 * Currently, we only support two mutually exclusive
2718 if (!strcmp(flags
, "static"))
2719 mode_flags
|= MPOL_F_STATIC_NODES
;
2720 else if (!strcmp(flags
, "relative"))
2721 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2726 new = mpol_new(mode
, mode_flags
, &nodes
);
2731 * Save nodes for mpol_to_str() to show the tmpfs mount options
2732 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2734 if (mode
!= MPOL_PREFERRED
)
2735 new->v
.nodes
= nodes
;
2737 new->v
.preferred_node
= first_node(nodes
);
2739 new->flags
|= MPOL_F_LOCAL
;
2742 * Save nodes for contextualization: this will be used to "clone"
2743 * the mempolicy in a specific context [cpuset] at a later time.
2745 new->w
.user_nodemask
= nodes
;
2750 /* Restore string for error message */
2759 #endif /* CONFIG_TMPFS */
2762 * mpol_to_str - format a mempolicy structure for printing
2763 * @buffer: to contain formatted mempolicy string
2764 * @maxlen: length of @buffer
2765 * @pol: pointer to mempolicy to be formatted
2767 * Convert @pol into a string. If @buffer is too short, truncate the string.
2768 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2769 * longest flag, "relative", and to display at least a few node ids.
2771 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2774 nodemask_t nodes
= NODE_MASK_NONE
;
2775 unsigned short mode
= MPOL_DEFAULT
;
2776 unsigned short flags
= 0;
2778 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2786 case MPOL_PREFERRED
:
2787 if (flags
& MPOL_F_LOCAL
)
2790 node_set(pol
->v
.preferred_node
, nodes
);
2793 case MPOL_INTERLEAVE
:
2794 nodes
= pol
->v
.nodes
;
2798 snprintf(p
, maxlen
, "unknown");
2802 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2804 if (flags
& MPOL_MODE_FLAGS
) {
2805 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2808 * Currently, the only defined flags are mutually exclusive
2810 if (flags
& MPOL_F_STATIC_NODES
)
2811 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2812 else if (flags
& MPOL_F_RELATIVE_NODES
)
2813 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2816 if (!nodes_empty(nodes
))
2817 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2818 nodemask_pr_args(&nodes
));