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
,
724 vma_get_anon_name(vma
));
728 if (mpol_equal(vma_policy(vma
), new_pol
))
730 /* vma_merge() joined vma && vma->next, case 8 */
733 if (vma
->vm_start
!= vmstart
) {
734 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
738 if (vma
->vm_end
!= vmend
) {
739 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
744 err
= vma_replace_policy(vma
, new_pol
);
753 /* Set the process memory policy */
754 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
757 struct mempolicy
*new, *old
;
758 NODEMASK_SCRATCH(scratch
);
764 new = mpol_new(mode
, flags
, nodes
);
771 ret
= mpol_set_nodemask(new, nodes
, scratch
);
773 task_unlock(current
);
777 old
= current
->mempolicy
;
778 current
->mempolicy
= new;
779 if (new && new->mode
== MPOL_INTERLEAVE
&&
780 nodes_weight(new->v
.nodes
))
781 current
->il_next
= first_node(new->v
.nodes
);
782 task_unlock(current
);
786 NODEMASK_SCRATCH_FREE(scratch
);
791 * Return nodemask for policy for get_mempolicy() query
793 * Called with task's alloc_lock held
795 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
798 if (p
== &default_policy
)
804 case MPOL_INTERLEAVE
:
808 if (!(p
->flags
& MPOL_F_LOCAL
))
809 node_set(p
->v
.preferred_node
, *nodes
);
810 /* else return empty node mask for local allocation */
817 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
822 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
824 err
= page_to_nid(p
);
830 /* Retrieve NUMA policy */
831 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
832 unsigned long addr
, unsigned long flags
)
835 struct mm_struct
*mm
= current
->mm
;
836 struct vm_area_struct
*vma
= NULL
;
837 struct mempolicy
*pol
= current
->mempolicy
;
840 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
843 if (flags
& MPOL_F_MEMS_ALLOWED
) {
844 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
846 *policy
= 0; /* just so it's initialized */
848 *nmask
= cpuset_current_mems_allowed
;
849 task_unlock(current
);
853 if (flags
& MPOL_F_ADDR
) {
855 * Do NOT fall back to task policy if the
856 * vma/shared policy at addr is NULL. We
857 * want to return MPOL_DEFAULT in this case.
859 down_read(&mm
->mmap_sem
);
860 vma
= find_vma_intersection(mm
, addr
, addr
+1);
862 up_read(&mm
->mmap_sem
);
865 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
866 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
868 pol
= vma
->vm_policy
;
873 pol
= &default_policy
; /* indicates default behavior */
875 if (flags
& MPOL_F_NODE
) {
876 if (flags
& MPOL_F_ADDR
) {
877 err
= lookup_node(mm
, addr
);
881 } else if (pol
== current
->mempolicy
&&
882 pol
->mode
== MPOL_INTERLEAVE
) {
883 *policy
= current
->il_next
;
889 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
892 * Internal mempolicy flags must be masked off before exposing
893 * the policy to userspace.
895 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
900 if (mpol_store_user_nodemask(pol
)) {
901 *nmask
= pol
->w
.user_nodemask
;
904 get_policy_nodemask(pol
, nmask
);
905 task_unlock(current
);
912 up_read(¤t
->mm
->mmap_sem
);
916 #ifdef CONFIG_MIGRATION
920 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
924 * Avoid migrating a page that is shared with others.
926 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
927 if (!isolate_lru_page(page
)) {
928 list_add_tail(&page
->lru
, pagelist
);
929 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
930 page_is_file_cache(page
));
935 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
938 return alloc_huge_page_node(page_hstate(compound_head(page
)),
941 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
946 * Migrate pages from one node to a target node.
947 * Returns error or the number of pages not migrated.
949 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
957 node_set(source
, nmask
);
960 * This does not "check" the range but isolates all pages that
961 * need migration. Between passing in the full user address
962 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
964 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
965 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
966 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
968 if (!list_empty(&pagelist
)) {
969 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
970 MIGRATE_SYNC
, MR_SYSCALL
);
972 putback_movable_pages(&pagelist
);
979 * Move pages between the two nodesets so as to preserve the physical
980 * layout as much as possible.
982 * Returns the number of page that could not be moved.
984 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
985 const nodemask_t
*to
, int flags
)
991 err
= migrate_prep();
995 down_read(&mm
->mmap_sem
);
998 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
999 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1000 * bit in 'tmp', and return that <source, dest> pair for migration.
1001 * The pair of nodemasks 'to' and 'from' define the map.
1003 * If no pair of bits is found that way, fallback to picking some
1004 * pair of 'source' and 'dest' bits that are not the same. If the
1005 * 'source' and 'dest' bits are the same, this represents a node
1006 * that will be migrating to itself, so no pages need move.
1008 * If no bits are left in 'tmp', or if all remaining bits left
1009 * in 'tmp' correspond to the same bit in 'to', return false
1010 * (nothing left to migrate).
1012 * This lets us pick a pair of nodes to migrate between, such that
1013 * if possible the dest node is not already occupied by some other
1014 * source node, minimizing the risk of overloading the memory on a
1015 * node that would happen if we migrated incoming memory to a node
1016 * before migrating outgoing memory source that same node.
1018 * A single scan of tmp is sufficient. As we go, we remember the
1019 * most recent <s, d> pair that moved (s != d). If we find a pair
1020 * that not only moved, but what's better, moved to an empty slot
1021 * (d is not set in tmp), then we break out then, with that pair.
1022 * Otherwise when we finish scanning from_tmp, we at least have the
1023 * most recent <s, d> pair that moved. If we get all the way through
1024 * the scan of tmp without finding any node that moved, much less
1025 * moved to an empty node, then there is nothing left worth migrating.
1029 while (!nodes_empty(tmp
)) {
1031 int source
= NUMA_NO_NODE
;
1034 for_each_node_mask(s
, tmp
) {
1037 * do_migrate_pages() tries to maintain the relative
1038 * node relationship of the pages established between
1039 * threads and memory areas.
1041 * However if the number of source nodes is not equal to
1042 * the number of destination nodes we can not preserve
1043 * this node relative relationship. In that case, skip
1044 * copying memory from a node that is in the destination
1047 * Example: [2,3,4] -> [3,4,5] moves everything.
1048 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1051 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1052 (node_isset(s
, *to
)))
1055 d
= node_remap(s
, *from
, *to
);
1059 source
= s
; /* Node moved. Memorize */
1062 /* dest not in remaining from nodes? */
1063 if (!node_isset(dest
, tmp
))
1066 if (source
== NUMA_NO_NODE
)
1069 node_clear(source
, tmp
);
1070 err
= migrate_to_node(mm
, source
, dest
, flags
);
1076 up_read(&mm
->mmap_sem
);
1084 * Allocate a new page for page migration based on vma policy.
1085 * Start by assuming the page is mapped by the same vma as contains @start.
1086 * Search forward from there, if not. N.B., this assumes that the
1087 * list of pages handed to migrate_pages()--which is how we get here--
1088 * is in virtual address order.
1090 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1092 struct vm_area_struct
*vma
;
1093 unsigned long uninitialized_var(address
);
1095 vma
= find_vma(current
->mm
, start
);
1097 address
= page_address_in_vma(page
, vma
);
1098 if (address
!= -EFAULT
)
1103 if (PageHuge(page
)) {
1105 return alloc_huge_page_noerr(vma
, address
, 1);
1108 * if !vma, alloc_page_vma() will use task or system default policy
1110 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1114 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1115 unsigned long flags
)
1119 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1120 const nodemask_t
*to
, int flags
)
1125 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1131 static long do_mbind(unsigned long start
, unsigned long len
,
1132 unsigned short mode
, unsigned short mode_flags
,
1133 nodemask_t
*nmask
, unsigned long flags
)
1135 struct mm_struct
*mm
= current
->mm
;
1136 struct mempolicy
*new;
1139 LIST_HEAD(pagelist
);
1141 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1143 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1146 if (start
& ~PAGE_MASK
)
1149 if (mode
== MPOL_DEFAULT
)
1150 flags
&= ~MPOL_MF_STRICT
;
1152 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1160 new = mpol_new(mode
, mode_flags
, nmask
);
1162 return PTR_ERR(new);
1164 if (flags
& MPOL_MF_LAZY
)
1165 new->flags
|= MPOL_F_MOF
;
1168 * If we are using the default policy then operation
1169 * on discontinuous address spaces is okay after all
1172 flags
|= MPOL_MF_DISCONTIG_OK
;
1174 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1175 start
, start
+ len
, mode
, mode_flags
,
1176 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1178 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1180 err
= migrate_prep();
1185 NODEMASK_SCRATCH(scratch
);
1187 down_write(&mm
->mmap_sem
);
1189 err
= mpol_set_nodemask(new, nmask
, scratch
);
1190 task_unlock(current
);
1192 up_write(&mm
->mmap_sem
);
1195 NODEMASK_SCRATCH_FREE(scratch
);
1200 err
= queue_pages_range(mm
, start
, end
, nmask
,
1201 flags
| MPOL_MF_INVERT
, &pagelist
);
1203 err
= mbind_range(mm
, start
, end
, new);
1208 if (!list_empty(&pagelist
)) {
1209 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1210 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1211 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1213 putback_movable_pages(&pagelist
);
1216 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1219 putback_movable_pages(&pagelist
);
1221 up_write(&mm
->mmap_sem
);
1228 * User space interface with variable sized bitmaps for nodelists.
1231 /* Copy a node mask from user space. */
1232 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1233 unsigned long maxnode
)
1236 unsigned long nlongs
;
1237 unsigned long endmask
;
1240 nodes_clear(*nodes
);
1241 if (maxnode
== 0 || !nmask
)
1243 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1246 nlongs
= BITS_TO_LONGS(maxnode
);
1247 if ((maxnode
% BITS_PER_LONG
) == 0)
1250 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1252 /* When the user specified more nodes than supported just check
1253 if the non supported part is all zero. */
1254 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1255 if (nlongs
> PAGE_SIZE
/sizeof(long))
1257 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1259 if (get_user(t
, nmask
+ k
))
1261 if (k
== nlongs
- 1) {
1267 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1271 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1273 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1277 /* Copy a kernel node mask to user space */
1278 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1281 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1282 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1284 if (copy
> nbytes
) {
1285 if (copy
> PAGE_SIZE
)
1287 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1291 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1294 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1295 unsigned long, mode
, const unsigned long __user
*, nmask
,
1296 unsigned long, maxnode
, unsigned, flags
)
1300 unsigned short mode_flags
;
1302 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1303 mode
&= ~MPOL_MODE_FLAGS
;
1304 if (mode
>= MPOL_MAX
)
1306 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1307 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1309 err
= get_nodes(&nodes
, nmask
, maxnode
);
1312 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1315 /* Set the process memory policy */
1316 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1317 unsigned long, maxnode
)
1321 unsigned short flags
;
1323 flags
= mode
& MPOL_MODE_FLAGS
;
1324 mode
&= ~MPOL_MODE_FLAGS
;
1325 if ((unsigned int)mode
>= MPOL_MAX
)
1327 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1329 err
= get_nodes(&nodes
, nmask
, maxnode
);
1332 return do_set_mempolicy(mode
, flags
, &nodes
);
1335 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1336 const unsigned long __user
*, old_nodes
,
1337 const unsigned long __user
*, new_nodes
)
1339 const struct cred
*cred
= current_cred(), *tcred
;
1340 struct mm_struct
*mm
= NULL
;
1341 struct task_struct
*task
;
1342 nodemask_t task_nodes
;
1346 NODEMASK_SCRATCH(scratch
);
1351 old
= &scratch
->mask1
;
1352 new = &scratch
->mask2
;
1354 err
= get_nodes(old
, old_nodes
, maxnode
);
1358 err
= get_nodes(new, new_nodes
, maxnode
);
1362 /* Find the mm_struct */
1364 task
= pid
? find_task_by_vpid(pid
) : current
;
1370 get_task_struct(task
);
1375 * Check if this process has the right to modify the specified
1376 * process. The right exists if the process has administrative
1377 * capabilities, superuser privileges or the same
1378 * userid as the target process.
1380 tcred
= __task_cred(task
);
1381 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1382 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1383 !capable(CAP_SYS_NICE
)) {
1390 task_nodes
= cpuset_mems_allowed(task
);
1391 /* Is the user allowed to access the target nodes? */
1392 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1397 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1402 err
= security_task_movememory(task
);
1406 mm
= get_task_mm(task
);
1407 put_task_struct(task
);
1414 err
= do_migrate_pages(mm
, old
, new,
1415 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1419 NODEMASK_SCRATCH_FREE(scratch
);
1424 put_task_struct(task
);
1430 /* Retrieve NUMA policy */
1431 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1432 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1433 unsigned long, addr
, unsigned long, flags
)
1436 int uninitialized_var(pval
);
1439 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1442 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1447 if (policy
&& put_user(pval
, policy
))
1451 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1456 #ifdef CONFIG_COMPAT
1458 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1459 compat_ulong_t __user
*, nmask
,
1460 compat_ulong_t
, maxnode
,
1461 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1464 unsigned long __user
*nm
= NULL
;
1465 unsigned long nr_bits
, alloc_size
;
1466 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1468 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1469 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1472 nm
= compat_alloc_user_space(alloc_size
);
1474 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1476 if (!err
&& nmask
) {
1477 unsigned long copy_size
;
1478 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1479 err
= copy_from_user(bm
, nm
, copy_size
);
1480 /* ensure entire bitmap is zeroed */
1481 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1482 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1488 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1489 compat_ulong_t
, maxnode
)
1491 unsigned long __user
*nm
= NULL
;
1492 unsigned long nr_bits
, alloc_size
;
1493 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1495 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1496 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1499 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1501 nm
= compat_alloc_user_space(alloc_size
);
1502 if (copy_to_user(nm
, bm
, alloc_size
))
1506 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1509 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1510 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1511 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1513 unsigned long __user
*nm
= NULL
;
1514 unsigned long nr_bits
, alloc_size
;
1517 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1518 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1521 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1523 nm
= compat_alloc_user_space(alloc_size
);
1524 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1528 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1533 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1536 struct mempolicy
*pol
= NULL
;
1539 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1540 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1541 } else if (vma
->vm_policy
) {
1542 pol
= vma
->vm_policy
;
1545 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1546 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1547 * count on these policies which will be dropped by
1548 * mpol_cond_put() later
1550 if (mpol_needs_cond_ref(pol
))
1559 * get_vma_policy(@vma, @addr)
1560 * @vma: virtual memory area whose policy is sought
1561 * @addr: address in @vma for shared policy lookup
1563 * Returns effective policy for a VMA at specified address.
1564 * Falls back to current->mempolicy or system default policy, as necessary.
1565 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1566 * count--added by the get_policy() vm_op, as appropriate--to protect against
1567 * freeing by another task. It is the caller's responsibility to free the
1568 * extra reference for shared policies.
1570 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1573 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1576 pol
= get_task_policy(current
);
1581 bool vma_policy_mof(struct vm_area_struct
*vma
)
1583 struct mempolicy
*pol
;
1585 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1588 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1589 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1596 pol
= vma
->vm_policy
;
1598 pol
= get_task_policy(current
);
1600 return pol
->flags
& MPOL_F_MOF
;
1603 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1605 enum zone_type dynamic_policy_zone
= policy_zone
;
1607 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1610 * if policy->v.nodes has movable memory only,
1611 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1613 * policy->v.nodes is intersect with node_states[N_MEMORY].
1614 * so if the following test faile, it implies
1615 * policy->v.nodes has movable memory only.
1617 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1618 dynamic_policy_zone
= ZONE_MOVABLE
;
1620 return zone
>= dynamic_policy_zone
;
1624 * Return a nodemask representing a mempolicy for filtering nodes for
1627 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1629 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1630 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1631 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1632 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1633 return &policy
->v
.nodes
;
1638 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1639 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1642 switch (policy
->mode
) {
1643 case MPOL_PREFERRED
:
1644 if (!(policy
->flags
& MPOL_F_LOCAL
))
1645 nd
= policy
->v
.preferred_node
;
1649 * Normally, MPOL_BIND allocations are node-local within the
1650 * allowed nodemask. However, if __GFP_THISNODE is set and the
1651 * current node isn't part of the mask, we use the zonelist for
1652 * the first node in the mask instead.
1654 if (unlikely(gfp
& __GFP_THISNODE
) &&
1655 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1656 nd
= first_node(policy
->v
.nodes
);
1661 return node_zonelist(nd
, gfp
);
1664 /* Do dynamic interleaving for a process */
1665 static unsigned interleave_nodes(struct mempolicy
*policy
)
1668 struct task_struct
*me
= current
;
1671 next
= next_node(nid
, policy
->v
.nodes
);
1672 if (next
>= MAX_NUMNODES
)
1673 next
= first_node(policy
->v
.nodes
);
1674 if (next
< MAX_NUMNODES
)
1680 * Depending on the memory policy provide a node from which to allocate the
1683 unsigned int mempolicy_slab_node(void)
1685 struct mempolicy
*policy
;
1686 int node
= numa_mem_id();
1691 policy
= current
->mempolicy
;
1692 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1695 switch (policy
->mode
) {
1696 case MPOL_PREFERRED
:
1698 * handled MPOL_F_LOCAL above
1700 return policy
->v
.preferred_node
;
1702 case MPOL_INTERLEAVE
:
1703 return interleave_nodes(policy
);
1707 * Follow bind policy behavior and start allocation at the
1710 struct zonelist
*zonelist
;
1712 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1713 zonelist
= &NODE_DATA(node
)->node_zonelists
[0];
1714 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1717 return zone
? zone
->node
: node
;
1725 /* Do static interleaving for a VMA with known offset. */
1726 static unsigned offset_il_node(struct mempolicy
*pol
,
1727 struct vm_area_struct
*vma
, unsigned long off
)
1729 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1732 int nid
= NUMA_NO_NODE
;
1735 return numa_node_id();
1736 target
= (unsigned int)off
% nnodes
;
1739 nid
= next_node(nid
, pol
->v
.nodes
);
1741 } while (c
<= target
);
1745 /* Determine a node number for interleave */
1746 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1747 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1753 * for small pages, there is no difference between
1754 * shift and PAGE_SHIFT, so the bit-shift is safe.
1755 * for huge pages, since vm_pgoff is in units of small
1756 * pages, we need to shift off the always 0 bits to get
1759 BUG_ON(shift
< PAGE_SHIFT
);
1760 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1761 off
+= (addr
- vma
->vm_start
) >> shift
;
1762 return offset_il_node(pol
, vma
, off
);
1764 return interleave_nodes(pol
);
1768 * Return the bit number of a random bit set in the nodemask.
1769 * (returns NUMA_NO_NODE if nodemask is empty)
1771 int node_random(const nodemask_t
*maskp
)
1773 int w
, bit
= NUMA_NO_NODE
;
1775 w
= nodes_weight(*maskp
);
1777 bit
= bitmap_ord_to_pos(maskp
->bits
,
1778 get_random_int() % w
, MAX_NUMNODES
);
1782 #ifdef CONFIG_HUGETLBFS
1784 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1785 * @vma: virtual memory area whose policy is sought
1786 * @addr: address in @vma for shared policy lookup and interleave policy
1787 * @gfp_flags: for requested zone
1788 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1789 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1791 * Returns a zonelist suitable for a huge page allocation and a pointer
1792 * to the struct mempolicy for conditional unref after allocation.
1793 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1794 * @nodemask for filtering the zonelist.
1796 * Must be protected by read_mems_allowed_begin()
1798 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1799 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1800 nodemask_t
**nodemask
)
1802 struct zonelist
*zl
;
1804 *mpol
= get_vma_policy(vma
, addr
);
1805 *nodemask
= NULL
; /* assume !MPOL_BIND */
1807 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1808 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1809 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1811 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1812 if ((*mpol
)->mode
== MPOL_BIND
)
1813 *nodemask
= &(*mpol
)->v
.nodes
;
1819 * init_nodemask_of_mempolicy
1821 * If the current task's mempolicy is "default" [NULL], return 'false'
1822 * to indicate default policy. Otherwise, extract the policy nodemask
1823 * for 'bind' or 'interleave' policy into the argument nodemask, or
1824 * initialize the argument nodemask to contain the single node for
1825 * 'preferred' or 'local' policy and return 'true' to indicate presence
1826 * of non-default mempolicy.
1828 * We don't bother with reference counting the mempolicy [mpol_get/put]
1829 * because the current task is examining it's own mempolicy and a task's
1830 * mempolicy is only ever changed by the task itself.
1832 * N.B., it is the caller's responsibility to free a returned nodemask.
1834 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1836 struct mempolicy
*mempolicy
;
1839 if (!(mask
&& current
->mempolicy
))
1843 mempolicy
= current
->mempolicy
;
1844 switch (mempolicy
->mode
) {
1845 case MPOL_PREFERRED
:
1846 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1847 nid
= numa_node_id();
1849 nid
= mempolicy
->v
.preferred_node
;
1850 init_nodemask_of_node(mask
, nid
);
1855 case MPOL_INTERLEAVE
:
1856 *mask
= mempolicy
->v
.nodes
;
1862 task_unlock(current
);
1869 * mempolicy_nodemask_intersects
1871 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1872 * policy. Otherwise, check for intersection between mask and the policy
1873 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1874 * policy, always return true since it may allocate elsewhere on fallback.
1876 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1878 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1879 const nodemask_t
*mask
)
1881 struct mempolicy
*mempolicy
;
1887 mempolicy
= tsk
->mempolicy
;
1891 switch (mempolicy
->mode
) {
1892 case MPOL_PREFERRED
:
1894 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1895 * allocate from, they may fallback to other nodes when oom.
1896 * Thus, it's possible for tsk to have allocated memory from
1901 case MPOL_INTERLEAVE
:
1902 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1912 /* Allocate a page in interleaved policy.
1913 Own path because it needs to do special accounting. */
1914 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1917 struct zonelist
*zl
;
1920 zl
= node_zonelist(nid
, gfp
);
1921 page
= __alloc_pages(gfp
, order
, zl
);
1922 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1923 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1928 * alloc_pages_vma - Allocate a page for a VMA.
1931 * %GFP_USER user allocation.
1932 * %GFP_KERNEL kernel allocations,
1933 * %GFP_HIGHMEM highmem/user allocations,
1934 * %GFP_FS allocation should not call back into a file system.
1935 * %GFP_ATOMIC don't sleep.
1937 * @order:Order of the GFP allocation.
1938 * @vma: Pointer to VMA or NULL if not available.
1939 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1940 * @node: Which node to prefer for allocation (modulo policy).
1941 * @hugepage: for hugepages try only the preferred node if possible
1943 * This function allocates a page from the kernel page pool and applies
1944 * a NUMA policy associated with the VMA or the current process.
1945 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1946 * mm_struct of the VMA to prevent it from going away. Should be used for
1947 * all allocations for pages that will be mapped into user space. Returns
1948 * NULL when no page can be allocated.
1951 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1952 unsigned long addr
, int node
, bool hugepage
)
1954 struct mempolicy
*pol
;
1956 unsigned int cpuset_mems_cookie
;
1957 struct zonelist
*zl
;
1961 pol
= get_vma_policy(vma
, addr
);
1962 cpuset_mems_cookie
= read_mems_allowed_begin();
1964 if (pol
->mode
== MPOL_INTERLEAVE
) {
1967 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
1969 page
= alloc_page_interleave(gfp
, order
, nid
);
1973 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
1974 int hpage_node
= node
;
1977 * For hugepage allocation and non-interleave policy which
1978 * allows the current node (or other explicitly preferred
1979 * node) we only try to allocate from the current/preferred
1980 * node and don't fall back to other nodes, as the cost of
1981 * remote accesses would likely offset THP benefits.
1983 * If the policy is interleave, or does not allow the current
1984 * node in its nodemask, we allocate the standard way.
1986 if (pol
->mode
== MPOL_PREFERRED
&&
1987 !(pol
->flags
& MPOL_F_LOCAL
))
1988 hpage_node
= pol
->v
.preferred_node
;
1990 nmask
= policy_nodemask(gfp
, pol
);
1991 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
1993 page
= __alloc_pages_node(hpage_node
,
1994 gfp
| __GFP_THISNODE
, order
);
1999 nmask
= policy_nodemask(gfp
, pol
);
2000 zl
= policy_zonelist(gfp
, pol
, node
);
2001 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nmask
);
2004 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2010 * alloc_pages_current - Allocate pages.
2013 * %GFP_USER user allocation,
2014 * %GFP_KERNEL kernel allocation,
2015 * %GFP_HIGHMEM highmem allocation,
2016 * %GFP_FS don't call back into a file system.
2017 * %GFP_ATOMIC don't sleep.
2018 * @order: Power of two of allocation size in pages. 0 is a single page.
2020 * Allocate a page from the kernel page pool. When not in
2021 * interrupt context and apply the current process NUMA policy.
2022 * Returns NULL when no page can be allocated.
2024 * Don't call cpuset_update_task_memory_state() unless
2025 * 1) it's ok to take cpuset_sem (can WAIT), and
2026 * 2) allocating for current task (not interrupt).
2028 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2030 struct mempolicy
*pol
= &default_policy
;
2032 unsigned int cpuset_mems_cookie
;
2034 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2035 pol
= get_task_policy(current
);
2038 cpuset_mems_cookie
= read_mems_allowed_begin();
2041 * No reference counting needed for current->mempolicy
2042 * nor system default_policy
2044 if (pol
->mode
== MPOL_INTERLEAVE
)
2045 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2047 page
= __alloc_pages_nodemask(gfp
, order
,
2048 policy_zonelist(gfp
, pol
, numa_node_id()),
2049 policy_nodemask(gfp
, pol
));
2051 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2056 EXPORT_SYMBOL(alloc_pages_current
);
2058 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2060 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2063 return PTR_ERR(pol
);
2064 dst
->vm_policy
= pol
;
2069 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2070 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2071 * with the mems_allowed returned by cpuset_mems_allowed(). This
2072 * keeps mempolicies cpuset relative after its cpuset moves. See
2073 * further kernel/cpuset.c update_nodemask().
2075 * current's mempolicy may be rebinded by the other task(the task that changes
2076 * cpuset's mems), so we needn't do rebind work for current task.
2079 /* Slow path of a mempolicy duplicate */
2080 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2082 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2085 return ERR_PTR(-ENOMEM
);
2087 /* task's mempolicy is protected by alloc_lock */
2088 if (old
== current
->mempolicy
) {
2091 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
);
2102 atomic_set(&new->refcnt
, 1);
2106 /* Slow path of a mempolicy comparison */
2107 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2111 if (a
->mode
!= b
->mode
)
2113 if (a
->flags
!= b
->flags
)
2115 if (mpol_store_user_nodemask(a
))
2116 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2122 case MPOL_INTERLEAVE
:
2123 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2124 case MPOL_PREFERRED
:
2125 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2133 * Shared memory backing store policy support.
2135 * Remember policies even when nobody has shared memory mapped.
2136 * The policies are kept in Red-Black tree linked from the inode.
2137 * They are protected by the sp->lock spinlock, which should be held
2138 * for any accesses to the tree.
2141 /* lookup first element intersecting start-end */
2142 /* Caller holds sp->lock */
2143 static struct sp_node
*
2144 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2146 struct rb_node
*n
= sp
->root
.rb_node
;
2149 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2151 if (start
>= p
->end
)
2153 else if (end
<= p
->start
)
2161 struct sp_node
*w
= NULL
;
2162 struct rb_node
*prev
= rb_prev(n
);
2165 w
= rb_entry(prev
, struct sp_node
, nd
);
2166 if (w
->end
<= start
)
2170 return rb_entry(n
, struct sp_node
, nd
);
2173 /* Insert a new shared policy into the list. */
2174 /* Caller holds sp->lock */
2175 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2177 struct rb_node
**p
= &sp
->root
.rb_node
;
2178 struct rb_node
*parent
= NULL
;
2183 nd
= rb_entry(parent
, struct sp_node
, nd
);
2184 if (new->start
< nd
->start
)
2186 else if (new->end
> nd
->end
)
2187 p
= &(*p
)->rb_right
;
2191 rb_link_node(&new->nd
, parent
, p
);
2192 rb_insert_color(&new->nd
, &sp
->root
);
2193 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2194 new->policy
? new->policy
->mode
: 0);
2197 /* Find shared policy intersecting idx */
2199 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2201 struct mempolicy
*pol
= NULL
;
2204 if (!sp
->root
.rb_node
)
2206 spin_lock(&sp
->lock
);
2207 sn
= sp_lookup(sp
, idx
, idx
+1);
2209 mpol_get(sn
->policy
);
2212 spin_unlock(&sp
->lock
);
2216 static void sp_free(struct sp_node
*n
)
2218 mpol_put(n
->policy
);
2219 kmem_cache_free(sn_cache
, n
);
2223 * mpol_misplaced - check whether current page node is valid in policy
2225 * @page: page to be checked
2226 * @vma: vm area where page mapped
2227 * @addr: virtual address where page mapped
2229 * Lookup current policy node id for vma,addr and "compare to" page's
2233 * -1 - not misplaced, page is in the right node
2234 * node - node id where the page should be
2236 * Policy determination "mimics" alloc_page_vma().
2237 * Called from fault path where we know the vma and faulting address.
2239 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2241 struct mempolicy
*pol
;
2243 int curnid
= page_to_nid(page
);
2244 unsigned long pgoff
;
2245 int thiscpu
= raw_smp_processor_id();
2246 int thisnid
= cpu_to_node(thiscpu
);
2252 pol
= get_vma_policy(vma
, addr
);
2253 if (!(pol
->flags
& MPOL_F_MOF
))
2256 switch (pol
->mode
) {
2257 case MPOL_INTERLEAVE
:
2258 BUG_ON(addr
>= vma
->vm_end
);
2259 BUG_ON(addr
< vma
->vm_start
);
2261 pgoff
= vma
->vm_pgoff
;
2262 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2263 polnid
= offset_il_node(pol
, vma
, pgoff
);
2266 case MPOL_PREFERRED
:
2267 if (pol
->flags
& MPOL_F_LOCAL
)
2268 polnid
= numa_node_id();
2270 polnid
= pol
->v
.preferred_node
;
2275 * allows binding to multiple nodes.
2276 * use current page if in policy nodemask,
2277 * else select nearest allowed node, if any.
2278 * If no allowed nodes, use current [!misplaced].
2280 if (node_isset(curnid
, pol
->v
.nodes
))
2282 (void)first_zones_zonelist(
2283 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2284 gfp_zone(GFP_HIGHUSER
),
2285 &pol
->v
.nodes
, &zone
);
2286 polnid
= zone
->node
;
2293 /* Migrate the page towards the node whose CPU is referencing it */
2294 if (pol
->flags
& MPOL_F_MORON
) {
2297 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2301 if (curnid
!= polnid
)
2309 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2311 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2312 rb_erase(&n
->nd
, &sp
->root
);
2316 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2317 unsigned long end
, struct mempolicy
*pol
)
2319 node
->start
= start
;
2324 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2325 struct mempolicy
*pol
)
2328 struct mempolicy
*newpol
;
2330 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2334 newpol
= mpol_dup(pol
);
2335 if (IS_ERR(newpol
)) {
2336 kmem_cache_free(sn_cache
, n
);
2339 newpol
->flags
|= MPOL_F_SHARED
;
2340 sp_node_init(n
, start
, end
, newpol
);
2345 /* Replace a policy range. */
2346 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2347 unsigned long end
, struct sp_node
*new)
2350 struct sp_node
*n_new
= NULL
;
2351 struct mempolicy
*mpol_new
= NULL
;
2355 spin_lock(&sp
->lock
);
2356 n
= sp_lookup(sp
, start
, end
);
2357 /* Take care of old policies in the same range. */
2358 while (n
&& n
->start
< end
) {
2359 struct rb_node
*next
= rb_next(&n
->nd
);
2360 if (n
->start
>= start
) {
2366 /* Old policy spanning whole new range. */
2371 *mpol_new
= *n
->policy
;
2372 atomic_set(&mpol_new
->refcnt
, 1);
2373 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2375 sp_insert(sp
, n_new
);
2384 n
= rb_entry(next
, struct sp_node
, nd
);
2388 spin_unlock(&sp
->lock
);
2395 kmem_cache_free(sn_cache
, n_new
);
2400 spin_unlock(&sp
->lock
);
2402 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2405 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2412 * mpol_shared_policy_init - initialize shared policy for inode
2413 * @sp: pointer to inode shared policy
2414 * @mpol: struct mempolicy to install
2416 * Install non-NULL @mpol in inode's shared policy rb-tree.
2417 * On entry, the current task has a reference on a non-NULL @mpol.
2418 * This must be released on exit.
2419 * This is called at get_inode() calls and we can use GFP_KERNEL.
2421 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2425 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2426 spin_lock_init(&sp
->lock
);
2429 struct vm_area_struct pvma
;
2430 struct mempolicy
*new;
2431 NODEMASK_SCRATCH(scratch
);
2435 /* contextualize the tmpfs mount point mempolicy */
2436 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2438 goto free_scratch
; /* no valid nodemask intersection */
2441 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2442 task_unlock(current
);
2446 /* Create pseudo-vma that contains just the policy */
2447 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2448 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2449 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2452 mpol_put(new); /* drop initial ref */
2454 NODEMASK_SCRATCH_FREE(scratch
);
2456 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2460 int mpol_set_shared_policy(struct shared_policy
*info
,
2461 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2464 struct sp_node
*new = NULL
;
2465 unsigned long sz
= vma_pages(vma
);
2467 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2469 sz
, npol
? npol
->mode
: -1,
2470 npol
? npol
->flags
: -1,
2471 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2474 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2478 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2484 /* Free a backing policy store on inode delete. */
2485 void mpol_free_shared_policy(struct shared_policy
*p
)
2488 struct rb_node
*next
;
2490 if (!p
->root
.rb_node
)
2492 spin_lock(&p
->lock
);
2493 next
= rb_first(&p
->root
);
2495 n
= rb_entry(next
, struct sp_node
, nd
);
2496 next
= rb_next(&n
->nd
);
2499 spin_unlock(&p
->lock
);
2502 #ifdef CONFIG_NUMA_BALANCING
2503 static int __initdata numabalancing_override
;
2505 static void __init
check_numabalancing_enable(void)
2507 bool numabalancing_default
= false;
2509 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2510 numabalancing_default
= true;
2512 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2513 if (numabalancing_override
)
2514 set_numabalancing_state(numabalancing_override
== 1);
2516 if (num_online_nodes() > 1 && !numabalancing_override
) {
2517 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2518 numabalancing_default
? "Enabling" : "Disabling");
2519 set_numabalancing_state(numabalancing_default
);
2523 static int __init
setup_numabalancing(char *str
)
2529 if (!strcmp(str
, "enable")) {
2530 numabalancing_override
= 1;
2532 } else if (!strcmp(str
, "disable")) {
2533 numabalancing_override
= -1;
2538 pr_warn("Unable to parse numa_balancing=\n");
2542 __setup("numa_balancing=", setup_numabalancing
);
2544 static inline void __init
check_numabalancing_enable(void)
2547 #endif /* CONFIG_NUMA_BALANCING */
2549 /* assumes fs == KERNEL_DS */
2550 void __init
numa_policy_init(void)
2552 nodemask_t interleave_nodes
;
2553 unsigned long largest
= 0;
2554 int nid
, prefer
= 0;
2556 policy_cache
= kmem_cache_create("numa_policy",
2557 sizeof(struct mempolicy
),
2558 0, SLAB_PANIC
, NULL
);
2560 sn_cache
= kmem_cache_create("shared_policy_node",
2561 sizeof(struct sp_node
),
2562 0, SLAB_PANIC
, NULL
);
2564 for_each_node(nid
) {
2565 preferred_node_policy
[nid
] = (struct mempolicy
) {
2566 .refcnt
= ATOMIC_INIT(1),
2567 .mode
= MPOL_PREFERRED
,
2568 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2569 .v
= { .preferred_node
= nid
, },
2574 * Set interleaving policy for system init. Interleaving is only
2575 * enabled across suitably sized nodes (default is >= 16MB), or
2576 * fall back to the largest node if they're all smaller.
2578 nodes_clear(interleave_nodes
);
2579 for_each_node_state(nid
, N_MEMORY
) {
2580 unsigned long total_pages
= node_present_pages(nid
);
2582 /* Preserve the largest node */
2583 if (largest
< total_pages
) {
2584 largest
= total_pages
;
2588 /* Interleave this node? */
2589 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2590 node_set(nid
, interleave_nodes
);
2593 /* All too small, use the largest */
2594 if (unlikely(nodes_empty(interleave_nodes
)))
2595 node_set(prefer
, interleave_nodes
);
2597 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2598 pr_err("%s: interleaving failed\n", __func__
);
2600 check_numabalancing_enable();
2603 /* Reset policy of current process to default */
2604 void numa_default_policy(void)
2606 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2610 * Parse and format mempolicy from/to strings
2614 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2616 static const char * const policy_modes
[] =
2618 [MPOL_DEFAULT
] = "default",
2619 [MPOL_PREFERRED
] = "prefer",
2620 [MPOL_BIND
] = "bind",
2621 [MPOL_INTERLEAVE
] = "interleave",
2622 [MPOL_LOCAL
] = "local",
2628 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2629 * @str: string containing mempolicy to parse
2630 * @mpol: pointer to struct mempolicy pointer, returned on success.
2633 * <mode>[=<flags>][:<nodelist>]
2635 * On success, returns 0, else 1
2637 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2639 struct mempolicy
*new = NULL
;
2640 unsigned short mode
;
2641 unsigned short mode_flags
;
2643 char *nodelist
= strchr(str
, ':');
2644 char *flags
= strchr(str
, '=');
2648 /* NUL-terminate mode or flags string */
2650 if (nodelist_parse(nodelist
, nodes
))
2652 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2658 *flags
++ = '\0'; /* terminate mode string */
2660 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2661 if (!strcmp(str
, policy_modes
[mode
])) {
2665 if (mode
>= MPOL_MAX
)
2669 case MPOL_PREFERRED
:
2671 * Insist on a nodelist of one node only
2674 char *rest
= nodelist
;
2675 while (isdigit(*rest
))
2681 case MPOL_INTERLEAVE
:
2683 * Default to online nodes with memory if no nodelist
2686 nodes
= node_states
[N_MEMORY
];
2690 * Don't allow a nodelist; mpol_new() checks flags
2694 mode
= MPOL_PREFERRED
;
2698 * Insist on a empty nodelist
2705 * Insist on a nodelist
2714 * Currently, we only support two mutually exclusive
2717 if (!strcmp(flags
, "static"))
2718 mode_flags
|= MPOL_F_STATIC_NODES
;
2719 else if (!strcmp(flags
, "relative"))
2720 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2725 new = mpol_new(mode
, mode_flags
, &nodes
);
2730 * Save nodes for mpol_to_str() to show the tmpfs mount options
2731 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2733 if (mode
!= MPOL_PREFERRED
)
2734 new->v
.nodes
= nodes
;
2736 new->v
.preferred_node
= first_node(nodes
);
2738 new->flags
|= MPOL_F_LOCAL
;
2741 * Save nodes for contextualization: this will be used to "clone"
2742 * the mempolicy in a specific context [cpuset] at a later time.
2744 new->w
.user_nodemask
= nodes
;
2749 /* Restore string for error message */
2758 #endif /* CONFIG_TMPFS */
2761 * mpol_to_str - format a mempolicy structure for printing
2762 * @buffer: to contain formatted mempolicy string
2763 * @maxlen: length of @buffer
2764 * @pol: pointer to mempolicy to be formatted
2766 * Convert @pol into a string. If @buffer is too short, truncate the string.
2767 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2768 * longest flag, "relative", and to display at least a few node ids.
2770 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2773 nodemask_t nodes
= NODE_MASK_NONE
;
2774 unsigned short mode
= MPOL_DEFAULT
;
2775 unsigned short flags
= 0;
2777 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2785 case MPOL_PREFERRED
:
2786 if (flags
& MPOL_F_LOCAL
)
2789 node_set(pol
->v
.preferred_node
, nodes
);
2792 case MPOL_INTERLEAVE
:
2793 nodes
= pol
->v
.nodes
;
2797 snprintf(p
, maxlen
, "unknown");
2801 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2803 if (flags
& MPOL_MODE_FLAGS
) {
2804 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2807 * Currently, the only defined flags are mutually exclusive
2809 if (flags
& MPOL_F_STATIC_NODES
)
2810 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2811 else if (flags
& MPOL_F_RELATIVE_NODES
)
2812 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2815 if (!nodes_empty(nodes
))
2816 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2817 nodemask_pr_args(&nodes
));