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>
101 #include "internal.h"
104 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
105 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
107 static struct kmem_cache
*policy_cache
;
108 static struct kmem_cache
*sn_cache
;
110 /* Highest zone. An specific allocation for a zone below that is not
112 enum zone_type policy_zone
= 0;
115 * run-time system-wide default policy => local allocation
117 static struct mempolicy default_policy
= {
118 .refcnt
= ATOMIC_INIT(1), /* never free it */
119 .mode
= MPOL_PREFERRED
,
120 .flags
= MPOL_F_LOCAL
,
123 static struct mempolicy preferred_node_policy
[MAX_NUMNODES
];
125 struct mempolicy
*get_task_policy(struct task_struct
*p
)
127 struct mempolicy
*pol
= p
->mempolicy
;
133 node
= numa_node_id();
134 if (node
!= NUMA_NO_NODE
) {
135 pol
= &preferred_node_policy
[node
];
136 /* preferred_node_policy is not initialised early in boot */
141 return &default_policy
;
144 static const struct mempolicy_operations
{
145 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * If we have a lock to protect task->mempolicy in read-side, we do
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
,
161 enum mpol_rebind_step step
);
162 } mpol_ops
[MPOL_MAX
];
164 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
166 return pol
->flags
& MPOL_MODE_FLAGS
;
169 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
170 const nodemask_t
*rel
)
173 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
174 nodes_onto(*ret
, tmp
, *rel
);
177 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
179 if (nodes_empty(*nodes
))
181 pol
->v
.nodes
= *nodes
;
185 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
188 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
189 else if (nodes_empty(*nodes
))
190 return -EINVAL
; /* no allowed nodes */
192 pol
->v
.preferred_node
= first_node(*nodes
);
196 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
198 if (nodes_empty(*nodes
))
200 pol
->v
.nodes
= *nodes
;
205 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
206 * any, for the new policy. mpol_new() has already validated the nodes
207 * parameter with respect to the policy mode and flags. But, we need to
208 * handle an empty nodemask with MPOL_PREFERRED here.
210 * Must be called holding task's alloc_lock to protect task's mems_allowed
211 * and mempolicy. May also be called holding the mmap_semaphore for write.
213 static int mpol_set_nodemask(struct mempolicy
*pol
,
214 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
218 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
222 nodes_and(nsc
->mask1
,
223 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
226 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
227 nodes
= NULL
; /* explicit local allocation */
229 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
230 mpol_relative_nodemask(&nsc
->mask2
, nodes
, &nsc
->mask1
);
232 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
234 if (mpol_store_user_nodemask(pol
))
235 pol
->w
.user_nodemask
= *nodes
;
237 pol
->w
.cpuset_mems_allowed
=
238 cpuset_current_mems_allowed
;
242 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
244 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
249 * This function just creates a new policy, does some check and simple
250 * initialization. You must invoke mpol_set_nodemask() to set nodes.
252 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
255 struct mempolicy
*policy
;
257 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
258 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
260 if (mode
== MPOL_DEFAULT
) {
261 if (nodes
&& !nodes_empty(*nodes
))
262 return ERR_PTR(-EINVAL
);
268 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
269 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
270 * All other modes require a valid pointer to a non-empty nodemask.
272 if (mode
== MPOL_PREFERRED
) {
273 if (nodes_empty(*nodes
)) {
274 if (((flags
& MPOL_F_STATIC_NODES
) ||
275 (flags
& MPOL_F_RELATIVE_NODES
)))
276 return ERR_PTR(-EINVAL
);
278 } else if (mode
== MPOL_LOCAL
) {
279 if (!nodes_empty(*nodes
))
280 return ERR_PTR(-EINVAL
);
281 mode
= MPOL_PREFERRED
;
282 } else if (nodes_empty(*nodes
))
283 return ERR_PTR(-EINVAL
);
284 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
286 return ERR_PTR(-ENOMEM
);
287 atomic_set(&policy
->refcnt
, 1);
289 policy
->flags
= flags
;
294 /* Slow path of a mpol destructor. */
295 void __mpol_put(struct mempolicy
*p
)
297 if (!atomic_dec_and_test(&p
->refcnt
))
299 kmem_cache_free(policy_cache
, p
);
302 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
,
303 enum mpol_rebind_step step
)
309 * MPOL_REBIND_ONCE - do rebind work at once
310 * MPOL_REBIND_STEP1 - set all the newly nodes
311 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
313 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
,
314 enum mpol_rebind_step step
)
318 if (pol
->flags
& MPOL_F_STATIC_NODES
)
319 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
320 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
321 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
324 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
327 if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP1
) {
328 nodes_remap(tmp
, pol
->v
.nodes
,
329 pol
->w
.cpuset_mems_allowed
, *nodes
);
330 pol
->w
.cpuset_mems_allowed
= step
? tmp
: *nodes
;
331 } else if (step
== MPOL_REBIND_STEP2
) {
332 tmp
= pol
->w
.cpuset_mems_allowed
;
333 pol
->w
.cpuset_mems_allowed
= *nodes
;
338 if (nodes_empty(tmp
))
341 if (step
== MPOL_REBIND_STEP1
)
342 nodes_or(pol
->v
.nodes
, pol
->v
.nodes
, tmp
);
343 else if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP2
)
348 if (!node_isset(current
->il_next
, tmp
)) {
349 current
->il_next
= next_node_in(current
->il_next
, tmp
);
350 if (current
->il_next
>= MAX_NUMNODES
)
351 current
->il_next
= numa_node_id();
355 static void mpol_rebind_preferred(struct mempolicy
*pol
,
356 const nodemask_t
*nodes
,
357 enum mpol_rebind_step step
)
361 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
362 int node
= first_node(pol
->w
.user_nodemask
);
364 if (node_isset(node
, *nodes
)) {
365 pol
->v
.preferred_node
= node
;
366 pol
->flags
&= ~MPOL_F_LOCAL
;
368 pol
->flags
|= MPOL_F_LOCAL
;
369 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
370 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
371 pol
->v
.preferred_node
= first_node(tmp
);
372 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
373 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
374 pol
->w
.cpuset_mems_allowed
,
376 pol
->w
.cpuset_mems_allowed
= *nodes
;
381 * mpol_rebind_policy - Migrate a policy to a different set of nodes
383 * If read-side task has no lock to protect task->mempolicy, write-side
384 * task will rebind the task->mempolicy by two step. The first step is
385 * setting all the newly nodes, and the second step is cleaning all the
386 * disallowed nodes. In this way, we can avoid finding no node to alloc
388 * If we have a lock to protect task->mempolicy in read-side, we do
392 * MPOL_REBIND_ONCE - do rebind work at once
393 * MPOL_REBIND_STEP1 - set all the newly nodes
394 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
396 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
,
397 enum mpol_rebind_step step
)
401 if (!mpol_store_user_nodemask(pol
) && step
== MPOL_REBIND_ONCE
&&
402 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
405 if (step
== MPOL_REBIND_STEP1
&& (pol
->flags
& MPOL_F_REBINDING
))
408 if (step
== MPOL_REBIND_STEP2
&& !(pol
->flags
& MPOL_F_REBINDING
))
411 if (step
== MPOL_REBIND_STEP1
)
412 pol
->flags
|= MPOL_F_REBINDING
;
413 else if (step
== MPOL_REBIND_STEP2
)
414 pol
->flags
&= ~MPOL_F_REBINDING
;
415 else if (step
>= MPOL_REBIND_NSTEP
)
418 mpol_ops
[pol
->mode
].rebind(pol
, newmask
, step
);
422 * Wrapper for mpol_rebind_policy() that just requires task
423 * pointer, and updates task mempolicy.
425 * Called with task's alloc_lock held.
428 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new,
429 enum mpol_rebind_step step
)
431 mpol_rebind_policy(tsk
->mempolicy
, new, step
);
435 * Rebind each vma in mm to new nodemask.
437 * Call holding a reference to mm. Takes mm->mmap_sem during call.
440 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
442 struct vm_area_struct
*vma
;
444 down_write(&mm
->mmap_sem
);
445 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
446 mpol_rebind_policy(vma
->vm_policy
, new, MPOL_REBIND_ONCE
);
447 up_write(&mm
->mmap_sem
);
450 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
452 .rebind
= mpol_rebind_default
,
454 [MPOL_INTERLEAVE
] = {
455 .create
= mpol_new_interleave
,
456 .rebind
= mpol_rebind_nodemask
,
459 .create
= mpol_new_preferred
,
460 .rebind
= mpol_rebind_preferred
,
463 .create
= mpol_new_bind
,
464 .rebind
= mpol_rebind_nodemask
,
468 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
469 unsigned long flags
);
472 struct list_head
*pagelist
;
475 struct vm_area_struct
*prev
;
479 * Scan through pages checking if pages follow certain conditions,
480 * and move them to the pagelist if they do.
482 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
483 unsigned long end
, struct mm_walk
*walk
)
485 struct vm_area_struct
*vma
= walk
->vma
;
487 struct queue_pages
*qp
= walk
->private;
488 unsigned long flags
= qp
->flags
;
493 if (pmd_trans_huge(*pmd
)) {
494 ptl
= pmd_lock(walk
->mm
, pmd
);
495 if (pmd_trans_huge(*pmd
)) {
496 page
= pmd_page(*pmd
);
497 if (is_huge_zero_page(page
)) {
499 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
504 ret
= split_huge_page(page
);
515 if (pmd_trans_unstable(pmd
))
518 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
519 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
520 if (!pte_present(*pte
))
522 page
= vm_normal_page(vma
, addr
, *pte
);
526 * vm_normal_page() filters out zero pages, but there might
527 * still be PageReserved pages to skip, perhaps in a VDSO.
529 if (PageReserved(page
))
531 nid
= page_to_nid(page
);
532 if (node_isset(nid
, *qp
->nmask
) == !!(flags
& MPOL_MF_INVERT
))
534 if (PageTransCompound(page
)) {
536 pte_unmap_unlock(pte
, ptl
);
538 ret
= split_huge_page(page
);
541 /* Failed to split -- skip. */
543 pte
= pte_offset_map_lock(walk
->mm
, pmd
,
550 migrate_page_add(page
, qp
->pagelist
, flags
);
552 pte_unmap_unlock(pte
- 1, ptl
);
557 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
558 unsigned long addr
, unsigned long end
,
559 struct mm_walk
*walk
)
561 #ifdef CONFIG_HUGETLB_PAGE
562 struct queue_pages
*qp
= walk
->private;
563 unsigned long flags
= qp
->flags
;
569 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
570 entry
= huge_ptep_get(pte
);
571 if (!pte_present(entry
))
573 page
= pte_page(entry
);
574 nid
= page_to_nid(page
);
575 if (node_isset(nid
, *qp
->nmask
) == !!(flags
& MPOL_MF_INVERT
))
577 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
578 if (flags
& (MPOL_MF_MOVE_ALL
) ||
579 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
580 isolate_huge_page(page
, qp
->pagelist
);
589 #ifdef CONFIG_NUMA_BALANCING
591 * This is used to mark a range of virtual addresses to be inaccessible.
592 * These are later cleared by a NUMA hinting fault. Depending on these
593 * faults, pages may be migrated for better NUMA placement.
595 * This is assuming that NUMA faults are handled using PROT_NONE. If
596 * an architecture makes a different choice, it will need further
597 * changes to the core.
599 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
600 unsigned long addr
, unsigned long end
)
604 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, 0, 1);
606 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
611 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
612 unsigned long addr
, unsigned long end
)
616 #endif /* CONFIG_NUMA_BALANCING */
618 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
619 struct mm_walk
*walk
)
621 struct vm_area_struct
*vma
= walk
->vma
;
622 struct queue_pages
*qp
= walk
->private;
623 unsigned long endvma
= vma
->vm_end
;
624 unsigned long flags
= qp
->flags
;
626 if (!vma_migratable(vma
))
631 if (vma
->vm_start
> start
)
632 start
= vma
->vm_start
;
634 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
635 if (!vma
->vm_next
&& vma
->vm_end
< end
)
637 if (qp
->prev
&& qp
->prev
->vm_end
< vma
->vm_start
)
643 if (flags
& MPOL_MF_LAZY
) {
644 /* Similar to task_numa_work, skip inaccessible VMAs */
645 if (!is_vm_hugetlb_page(vma
) &&
646 (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)) &&
647 !(vma
->vm_flags
& VM_MIXEDMAP
))
648 change_prot_numa(vma
, start
, endvma
);
652 /* queue pages from current vma */
653 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
659 * Walk through page tables and collect pages to be migrated.
661 * If pages found in a given range are on a set of nodes (determined by
662 * @nodes and @flags,) it's isolated and queued to the pagelist which is
663 * passed via @private.)
666 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
667 nodemask_t
*nodes
, unsigned long flags
,
668 struct list_head
*pagelist
)
670 struct queue_pages qp
= {
671 .pagelist
= pagelist
,
676 struct mm_walk queue_pages_walk
= {
677 .hugetlb_entry
= queue_pages_hugetlb
,
678 .pmd_entry
= queue_pages_pte_range
,
679 .test_walk
= queue_pages_test_walk
,
684 return walk_page_range(start
, end
, &queue_pages_walk
);
688 * Apply policy to a single VMA
689 * This must be called with the mmap_sem held for writing.
691 static int vma_replace_policy(struct vm_area_struct
*vma
,
692 struct mempolicy
*pol
)
695 struct mempolicy
*old
;
696 struct mempolicy
*new;
698 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
699 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
700 vma
->vm_ops
, vma
->vm_file
,
701 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
707 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
708 err
= vma
->vm_ops
->set_policy(vma
, new);
713 old
= vma
->vm_policy
;
714 vma
->vm_policy
= new; /* protected by mmap_sem */
723 /* Step 2: apply policy to a range and do splits. */
724 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
725 unsigned long end
, struct mempolicy
*new_pol
)
727 struct vm_area_struct
*next
;
728 struct vm_area_struct
*prev
;
729 struct vm_area_struct
*vma
;
732 unsigned long vmstart
;
735 vma
= find_vma(mm
, start
);
736 if (!vma
|| vma
->vm_start
> start
)
740 if (start
> vma
->vm_start
)
743 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
745 vmstart
= max(start
, vma
->vm_start
);
746 vmend
= min(end
, vma
->vm_end
);
748 if (mpol_equal(vma_policy(vma
), new_pol
))
751 pgoff
= vma
->vm_pgoff
+
752 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
753 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
754 vma
->anon_vma
, vma
->vm_file
, pgoff
,
755 new_pol
, vma
->vm_userfaultfd_ctx
);
759 if (mpol_equal(vma_policy(vma
), new_pol
))
761 /* vma_merge() joined vma && vma->next, case 8 */
764 if (vma
->vm_start
!= vmstart
) {
765 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
769 if (vma
->vm_end
!= vmend
) {
770 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
775 err
= vma_replace_policy(vma
, new_pol
);
784 /* Set the process memory policy */
785 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
788 struct mempolicy
*new, *old
;
789 NODEMASK_SCRATCH(scratch
);
795 new = mpol_new(mode
, flags
, nodes
);
802 ret
= mpol_set_nodemask(new, nodes
, scratch
);
804 task_unlock(current
);
808 old
= current
->mempolicy
;
809 current
->mempolicy
= new;
810 if (new && new->mode
== MPOL_INTERLEAVE
&&
811 nodes_weight(new->v
.nodes
))
812 current
->il_next
= first_node(new->v
.nodes
);
813 task_unlock(current
);
817 NODEMASK_SCRATCH_FREE(scratch
);
822 * Return nodemask for policy for get_mempolicy() query
824 * Called with task's alloc_lock held
826 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
829 if (p
== &default_policy
)
835 case MPOL_INTERLEAVE
:
839 if (!(p
->flags
& MPOL_F_LOCAL
))
840 node_set(p
->v
.preferred_node
, *nodes
);
841 /* else return empty node mask for local allocation */
848 static int lookup_node(unsigned long addr
)
853 err
= get_user_pages(addr
& PAGE_MASK
, 1, 0, &p
, NULL
);
855 err
= page_to_nid(p
);
861 /* Retrieve NUMA policy */
862 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
863 unsigned long addr
, unsigned long flags
)
866 struct mm_struct
*mm
= current
->mm
;
867 struct vm_area_struct
*vma
= NULL
;
868 struct mempolicy
*pol
= current
->mempolicy
;
871 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
874 if (flags
& MPOL_F_MEMS_ALLOWED
) {
875 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
877 *policy
= 0; /* just so it's initialized */
879 *nmask
= cpuset_current_mems_allowed
;
880 task_unlock(current
);
884 if (flags
& MPOL_F_ADDR
) {
886 * Do NOT fall back to task policy if the
887 * vma/shared policy at addr is NULL. We
888 * want to return MPOL_DEFAULT in this case.
890 down_read(&mm
->mmap_sem
);
891 vma
= find_vma_intersection(mm
, addr
, addr
+1);
893 up_read(&mm
->mmap_sem
);
896 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
897 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
899 pol
= vma
->vm_policy
;
904 pol
= &default_policy
; /* indicates default behavior */
906 if (flags
& MPOL_F_NODE
) {
907 if (flags
& MPOL_F_ADDR
) {
908 err
= lookup_node(addr
);
912 } else if (pol
== current
->mempolicy
&&
913 pol
->mode
== MPOL_INTERLEAVE
) {
914 *policy
= current
->il_next
;
920 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
923 * Internal mempolicy flags must be masked off before exposing
924 * the policy to userspace.
926 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
930 up_read(¤t
->mm
->mmap_sem
);
936 if (mpol_store_user_nodemask(pol
)) {
937 *nmask
= pol
->w
.user_nodemask
;
940 get_policy_nodemask(pol
, nmask
);
941 task_unlock(current
);
948 up_read(¤t
->mm
->mmap_sem
);
952 #ifdef CONFIG_MIGRATION
956 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
960 * Avoid migrating a page that is shared with others.
962 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
963 if (!isolate_lru_page(page
)) {
964 list_add_tail(&page
->lru
, pagelist
);
965 inc_node_page_state(page
, NR_ISOLATED_ANON
+
966 page_is_file_cache(page
));
971 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
974 return alloc_huge_page_node(page_hstate(compound_head(page
)),
977 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
982 * Migrate pages from one node to a target node.
983 * Returns error or the number of pages not migrated.
985 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
993 node_set(source
, nmask
);
996 * This does not "check" the range but isolates all pages that
997 * need migration. Between passing in the full user address
998 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1000 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1001 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1002 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1004 if (!list_empty(&pagelist
)) {
1005 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
1006 MIGRATE_SYNC
, MR_SYSCALL
);
1008 putback_movable_pages(&pagelist
);
1015 * Move pages between the two nodesets so as to preserve the physical
1016 * layout as much as possible.
1018 * Returns the number of page that could not be moved.
1020 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1021 const nodemask_t
*to
, int flags
)
1027 err
= migrate_prep();
1031 down_read(&mm
->mmap_sem
);
1034 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1035 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1036 * bit in 'tmp', and return that <source, dest> pair for migration.
1037 * The pair of nodemasks 'to' and 'from' define the map.
1039 * If no pair of bits is found that way, fallback to picking some
1040 * pair of 'source' and 'dest' bits that are not the same. If the
1041 * 'source' and 'dest' bits are the same, this represents a node
1042 * that will be migrating to itself, so no pages need move.
1044 * If no bits are left in 'tmp', or if all remaining bits left
1045 * in 'tmp' correspond to the same bit in 'to', return false
1046 * (nothing left to migrate).
1048 * This lets us pick a pair of nodes to migrate between, such that
1049 * if possible the dest node is not already occupied by some other
1050 * source node, minimizing the risk of overloading the memory on a
1051 * node that would happen if we migrated incoming memory to a node
1052 * before migrating outgoing memory source that same node.
1054 * A single scan of tmp is sufficient. As we go, we remember the
1055 * most recent <s, d> pair that moved (s != d). If we find a pair
1056 * that not only moved, but what's better, moved to an empty slot
1057 * (d is not set in tmp), then we break out then, with that pair.
1058 * Otherwise when we finish scanning from_tmp, we at least have the
1059 * most recent <s, d> pair that moved. If we get all the way through
1060 * the scan of tmp without finding any node that moved, much less
1061 * moved to an empty node, then there is nothing left worth migrating.
1065 while (!nodes_empty(tmp
)) {
1067 int source
= NUMA_NO_NODE
;
1070 for_each_node_mask(s
, tmp
) {
1073 * do_migrate_pages() tries to maintain the relative
1074 * node relationship of the pages established between
1075 * threads and memory areas.
1077 * However if the number of source nodes is not equal to
1078 * the number of destination nodes we can not preserve
1079 * this node relative relationship. In that case, skip
1080 * copying memory from a node that is in the destination
1083 * Example: [2,3,4] -> [3,4,5] moves everything.
1084 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1087 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1088 (node_isset(s
, *to
)))
1091 d
= node_remap(s
, *from
, *to
);
1095 source
= s
; /* Node moved. Memorize */
1098 /* dest not in remaining from nodes? */
1099 if (!node_isset(dest
, tmp
))
1102 if (source
== NUMA_NO_NODE
)
1105 node_clear(source
, tmp
);
1106 err
= migrate_to_node(mm
, source
, dest
, flags
);
1112 up_read(&mm
->mmap_sem
);
1120 * Allocate a new page for page migration based on vma policy.
1121 * Start by assuming the page is mapped by the same vma as contains @start.
1122 * Search forward from there, if not. N.B., this assumes that the
1123 * list of pages handed to migrate_pages()--which is how we get here--
1124 * is in virtual address order.
1126 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1128 struct vm_area_struct
*vma
;
1129 unsigned long uninitialized_var(address
);
1131 vma
= find_vma(current
->mm
, start
);
1133 address
= page_address_in_vma(page
, vma
);
1134 if (address
!= -EFAULT
)
1139 if (PageHuge(page
)) {
1141 return alloc_huge_page_noerr(vma
, address
, 1);
1144 * if !vma, alloc_page_vma() will use task or system default policy
1146 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1150 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1151 unsigned long flags
)
1155 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1156 const nodemask_t
*to
, int flags
)
1161 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1167 static long do_mbind(unsigned long start
, unsigned long len
,
1168 unsigned short mode
, unsigned short mode_flags
,
1169 nodemask_t
*nmask
, unsigned long flags
)
1171 struct mm_struct
*mm
= current
->mm
;
1172 struct mempolicy
*new;
1175 LIST_HEAD(pagelist
);
1177 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1179 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1182 if (start
& ~PAGE_MASK
)
1185 if (mode
== MPOL_DEFAULT
)
1186 flags
&= ~MPOL_MF_STRICT
;
1188 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1196 new = mpol_new(mode
, mode_flags
, nmask
);
1198 return PTR_ERR(new);
1200 if (flags
& MPOL_MF_LAZY
)
1201 new->flags
|= MPOL_F_MOF
;
1204 * If we are using the default policy then operation
1205 * on discontinuous address spaces is okay after all
1208 flags
|= MPOL_MF_DISCONTIG_OK
;
1210 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1211 start
, start
+ len
, mode
, mode_flags
,
1212 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1214 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1216 err
= migrate_prep();
1221 NODEMASK_SCRATCH(scratch
);
1223 down_write(&mm
->mmap_sem
);
1225 err
= mpol_set_nodemask(new, nmask
, scratch
);
1226 task_unlock(current
);
1228 up_write(&mm
->mmap_sem
);
1231 NODEMASK_SCRATCH_FREE(scratch
);
1236 err
= queue_pages_range(mm
, start
, end
, nmask
,
1237 flags
| MPOL_MF_INVERT
, &pagelist
);
1239 err
= mbind_range(mm
, start
, end
, new);
1244 if (!list_empty(&pagelist
)) {
1245 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1246 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1247 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1249 putback_movable_pages(&pagelist
);
1252 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1255 putback_movable_pages(&pagelist
);
1257 up_write(&mm
->mmap_sem
);
1264 * User space interface with variable sized bitmaps for nodelists.
1267 /* Copy a node mask from user space. */
1268 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1269 unsigned long maxnode
)
1272 unsigned long nlongs
;
1273 unsigned long endmask
;
1276 nodes_clear(*nodes
);
1277 if (maxnode
== 0 || !nmask
)
1279 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1282 nlongs
= BITS_TO_LONGS(maxnode
);
1283 if ((maxnode
% BITS_PER_LONG
) == 0)
1286 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1288 /* When the user specified more nodes than supported just check
1289 if the non supported part is all zero. */
1290 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1291 if (nlongs
> PAGE_SIZE
/sizeof(long))
1293 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1295 if (get_user(t
, nmask
+ k
))
1297 if (k
== nlongs
- 1) {
1303 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1307 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1309 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1313 /* Copy a kernel node mask to user space */
1314 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1317 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1318 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1320 if (copy
> nbytes
) {
1321 if (copy
> PAGE_SIZE
)
1323 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1327 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1330 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1331 unsigned long, mode
, const unsigned long __user
*, nmask
,
1332 unsigned long, maxnode
, unsigned, flags
)
1336 unsigned short mode_flags
;
1338 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1339 mode
&= ~MPOL_MODE_FLAGS
;
1340 if (mode
>= MPOL_MAX
)
1342 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1343 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1345 err
= get_nodes(&nodes
, nmask
, maxnode
);
1348 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1351 /* Set the process memory policy */
1352 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1353 unsigned long, maxnode
)
1357 unsigned short flags
;
1359 flags
= mode
& MPOL_MODE_FLAGS
;
1360 mode
&= ~MPOL_MODE_FLAGS
;
1361 if ((unsigned int)mode
>= MPOL_MAX
)
1363 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1365 err
= get_nodes(&nodes
, nmask
, maxnode
);
1368 return do_set_mempolicy(mode
, flags
, &nodes
);
1371 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1372 const unsigned long __user
*, old_nodes
,
1373 const unsigned long __user
*, new_nodes
)
1375 const struct cred
*cred
= current_cred(), *tcred
;
1376 struct mm_struct
*mm
= NULL
;
1377 struct task_struct
*task
;
1378 nodemask_t task_nodes
;
1382 NODEMASK_SCRATCH(scratch
);
1387 old
= &scratch
->mask1
;
1388 new = &scratch
->mask2
;
1390 err
= get_nodes(old
, old_nodes
, maxnode
);
1394 err
= get_nodes(new, new_nodes
, maxnode
);
1398 /* Find the mm_struct */
1400 task
= pid
? find_task_by_vpid(pid
) : current
;
1406 get_task_struct(task
);
1411 * Check if this process has the right to modify the specified
1412 * process. The right exists if the process has administrative
1413 * capabilities, superuser privileges or the same
1414 * userid as the target process.
1416 tcred
= __task_cred(task
);
1417 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1418 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1419 !capable(CAP_SYS_NICE
)) {
1426 task_nodes
= cpuset_mems_allowed(task
);
1427 /* Is the user allowed to access the target nodes? */
1428 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1433 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1438 err
= security_task_movememory(task
);
1442 mm
= get_task_mm(task
);
1443 put_task_struct(task
);
1450 err
= do_migrate_pages(mm
, old
, new,
1451 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1455 NODEMASK_SCRATCH_FREE(scratch
);
1460 put_task_struct(task
);
1466 /* Retrieve NUMA policy */
1467 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1468 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1469 unsigned long, addr
, unsigned long, flags
)
1472 int uninitialized_var(pval
);
1475 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1478 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1483 if (policy
&& put_user(pval
, policy
))
1487 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1492 #ifdef CONFIG_COMPAT
1494 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1495 compat_ulong_t __user
*, nmask
,
1496 compat_ulong_t
, maxnode
,
1497 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1500 unsigned long __user
*nm
= NULL
;
1501 unsigned long nr_bits
, alloc_size
;
1502 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1504 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1505 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1508 nm
= compat_alloc_user_space(alloc_size
);
1510 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1512 if (!err
&& nmask
) {
1513 unsigned long copy_size
;
1514 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1515 err
= copy_from_user(bm
, nm
, copy_size
);
1516 /* ensure entire bitmap is zeroed */
1517 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1518 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1524 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1525 compat_ulong_t
, maxnode
)
1528 unsigned long __user
*nm
= NULL
;
1529 unsigned long nr_bits
, alloc_size
;
1530 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1532 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1533 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1536 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1537 nm
= compat_alloc_user_space(alloc_size
);
1538 err
|= copy_to_user(nm
, bm
, alloc_size
);
1544 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1547 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1548 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1549 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1552 unsigned long __user
*nm
= NULL
;
1553 unsigned long nr_bits
, alloc_size
;
1556 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1557 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1560 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1561 nm
= compat_alloc_user_space(alloc_size
);
1562 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1568 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1573 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1576 struct mempolicy
*pol
= NULL
;
1579 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1580 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1581 } else if (vma
->vm_policy
) {
1582 pol
= vma
->vm_policy
;
1585 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1586 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1587 * count on these policies which will be dropped by
1588 * mpol_cond_put() later
1590 if (mpol_needs_cond_ref(pol
))
1599 * get_vma_policy(@vma, @addr)
1600 * @vma: virtual memory area whose policy is sought
1601 * @addr: address in @vma for shared policy lookup
1603 * Returns effective policy for a VMA at specified address.
1604 * Falls back to current->mempolicy or system default policy, as necessary.
1605 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1606 * count--added by the get_policy() vm_op, as appropriate--to protect against
1607 * freeing by another task. It is the caller's responsibility to free the
1608 * extra reference for shared policies.
1610 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1613 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1616 pol
= get_task_policy(current
);
1621 bool vma_policy_mof(struct vm_area_struct
*vma
)
1623 struct mempolicy
*pol
;
1625 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1628 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1629 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1636 pol
= vma
->vm_policy
;
1638 pol
= get_task_policy(current
);
1640 return pol
->flags
& MPOL_F_MOF
;
1643 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1645 enum zone_type dynamic_policy_zone
= policy_zone
;
1647 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1650 * if policy->v.nodes has movable memory only,
1651 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1653 * policy->v.nodes is intersect with node_states[N_MEMORY].
1654 * so if the following test faile, it implies
1655 * policy->v.nodes has movable memory only.
1657 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1658 dynamic_policy_zone
= ZONE_MOVABLE
;
1660 return zone
>= dynamic_policy_zone
;
1664 * Return a nodemask representing a mempolicy for filtering nodes for
1667 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1669 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1670 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1671 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1672 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1673 return &policy
->v
.nodes
;
1678 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1679 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1682 if (policy
->mode
== MPOL_PREFERRED
&& !(policy
->flags
& MPOL_F_LOCAL
))
1683 nd
= policy
->v
.preferred_node
;
1686 * __GFP_THISNODE shouldn't even be used with the bind policy
1687 * because we might easily break the expectation to stay on the
1688 * requested node and not break the policy.
1690 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1693 return node_zonelist(nd
, gfp
);
1696 /* Do dynamic interleaving for a process */
1697 static unsigned interleave_nodes(struct mempolicy
*policy
)
1700 struct task_struct
*me
= current
;
1703 next
= next_node_in(nid
, policy
->v
.nodes
);
1704 if (next
< MAX_NUMNODES
)
1710 * Depending on the memory policy provide a node from which to allocate the
1713 unsigned int mempolicy_slab_node(void)
1715 struct mempolicy
*policy
;
1716 int node
= numa_mem_id();
1721 policy
= current
->mempolicy
;
1722 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1725 switch (policy
->mode
) {
1726 case MPOL_PREFERRED
:
1728 * handled MPOL_F_LOCAL above
1730 return policy
->v
.preferred_node
;
1732 case MPOL_INTERLEAVE
:
1733 return interleave_nodes(policy
);
1739 * Follow bind policy behavior and start allocation at the
1742 struct zonelist
*zonelist
;
1743 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1744 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1745 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1747 return z
->zone
? z
->zone
->node
: node
;
1756 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1757 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1758 * number of present nodes.
1760 static unsigned offset_il_node(struct mempolicy
*pol
,
1761 struct vm_area_struct
*vma
, unsigned long n
)
1763 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1769 return numa_node_id();
1770 target
= (unsigned int)n
% nnodes
;
1771 nid
= first_node(pol
->v
.nodes
);
1772 for (i
= 0; i
< target
; i
++)
1773 nid
= next_node(nid
, pol
->v
.nodes
);
1777 /* Determine a node number for interleave */
1778 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1779 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1785 * for small pages, there is no difference between
1786 * shift and PAGE_SHIFT, so the bit-shift is safe.
1787 * for huge pages, since vm_pgoff is in units of small
1788 * pages, we need to shift off the always 0 bits to get
1791 BUG_ON(shift
< PAGE_SHIFT
);
1792 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1793 off
+= (addr
- vma
->vm_start
) >> shift
;
1794 return offset_il_node(pol
, vma
, off
);
1796 return interleave_nodes(pol
);
1799 #ifdef CONFIG_HUGETLBFS
1801 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1802 * @vma: virtual memory area whose policy is sought
1803 * @addr: address in @vma for shared policy lookup and interleave policy
1804 * @gfp_flags: for requested zone
1805 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1806 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1808 * Returns a zonelist suitable for a huge page allocation and a pointer
1809 * to the struct mempolicy for conditional unref after allocation.
1810 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1811 * @nodemask for filtering the zonelist.
1813 * Must be protected by read_mems_allowed_begin()
1815 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1816 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1817 nodemask_t
**nodemask
)
1819 struct zonelist
*zl
;
1821 *mpol
= get_vma_policy(vma
, addr
);
1822 *nodemask
= NULL
; /* assume !MPOL_BIND */
1824 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1825 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1826 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1828 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1829 if ((*mpol
)->mode
== MPOL_BIND
)
1830 *nodemask
= &(*mpol
)->v
.nodes
;
1836 * init_nodemask_of_mempolicy
1838 * If the current task's mempolicy is "default" [NULL], return 'false'
1839 * to indicate default policy. Otherwise, extract the policy nodemask
1840 * for 'bind' or 'interleave' policy into the argument nodemask, or
1841 * initialize the argument nodemask to contain the single node for
1842 * 'preferred' or 'local' policy and return 'true' to indicate presence
1843 * of non-default mempolicy.
1845 * We don't bother with reference counting the mempolicy [mpol_get/put]
1846 * because the current task is examining it's own mempolicy and a task's
1847 * mempolicy is only ever changed by the task itself.
1849 * N.B., it is the caller's responsibility to free a returned nodemask.
1851 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1853 struct mempolicy
*mempolicy
;
1856 if (!(mask
&& current
->mempolicy
))
1860 mempolicy
= current
->mempolicy
;
1861 switch (mempolicy
->mode
) {
1862 case MPOL_PREFERRED
:
1863 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1864 nid
= numa_node_id();
1866 nid
= mempolicy
->v
.preferred_node
;
1867 init_nodemask_of_node(mask
, nid
);
1872 case MPOL_INTERLEAVE
:
1873 *mask
= mempolicy
->v
.nodes
;
1879 task_unlock(current
);
1886 * mempolicy_nodemask_intersects
1888 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1889 * policy. Otherwise, check for intersection between mask and the policy
1890 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1891 * policy, always return true since it may allocate elsewhere on fallback.
1893 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1895 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1896 const nodemask_t
*mask
)
1898 struct mempolicy
*mempolicy
;
1904 mempolicy
= tsk
->mempolicy
;
1908 switch (mempolicy
->mode
) {
1909 case MPOL_PREFERRED
:
1911 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1912 * allocate from, they may fallback to other nodes when oom.
1913 * Thus, it's possible for tsk to have allocated memory from
1918 case MPOL_INTERLEAVE
:
1919 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1929 /* Allocate a page in interleaved policy.
1930 Own path because it needs to do special accounting. */
1931 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1934 struct zonelist
*zl
;
1937 zl
= node_zonelist(nid
, gfp
);
1938 page
= __alloc_pages(gfp
, order
, zl
);
1939 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1940 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1945 * alloc_pages_vma - Allocate a page for a VMA.
1948 * %GFP_USER user allocation.
1949 * %GFP_KERNEL kernel allocations,
1950 * %GFP_HIGHMEM highmem/user allocations,
1951 * %GFP_FS allocation should not call back into a file system.
1952 * %GFP_ATOMIC don't sleep.
1954 * @order:Order of the GFP allocation.
1955 * @vma: Pointer to VMA or NULL if not available.
1956 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1957 * @node: Which node to prefer for allocation (modulo policy).
1958 * @hugepage: for hugepages try only the preferred node if possible
1960 * This function allocates a page from the kernel page pool and applies
1961 * a NUMA policy associated with the VMA or the current process.
1962 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1963 * mm_struct of the VMA to prevent it from going away. Should be used for
1964 * all allocations for pages that will be mapped into user space. Returns
1965 * NULL when no page can be allocated.
1968 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1969 unsigned long addr
, int node
, bool hugepage
)
1971 struct mempolicy
*pol
;
1973 unsigned int cpuset_mems_cookie
;
1974 struct zonelist
*zl
;
1978 pol
= get_vma_policy(vma
, addr
);
1979 cpuset_mems_cookie
= read_mems_allowed_begin();
1981 if (pol
->mode
== MPOL_INTERLEAVE
) {
1984 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
1986 page
= alloc_page_interleave(gfp
, order
, nid
);
1990 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
1991 int hpage_node
= node
;
1994 * For hugepage allocation and non-interleave policy which
1995 * allows the current node (or other explicitly preferred
1996 * node) we only try to allocate from the current/preferred
1997 * node and don't fall back to other nodes, as the cost of
1998 * remote accesses would likely offset THP benefits.
2000 * If the policy is interleave, or does not allow the current
2001 * node in its nodemask, we allocate the standard way.
2003 if (pol
->mode
== MPOL_PREFERRED
&&
2004 !(pol
->flags
& MPOL_F_LOCAL
))
2005 hpage_node
= pol
->v
.preferred_node
;
2007 nmask
= policy_nodemask(gfp
, pol
);
2008 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2010 page
= __alloc_pages_node(hpage_node
,
2011 gfp
| __GFP_THISNODE
, order
);
2016 nmask
= policy_nodemask(gfp
, pol
);
2017 zl
= policy_zonelist(gfp
, pol
, node
);
2019 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nmask
);
2021 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2027 * alloc_pages_current - Allocate pages.
2030 * %GFP_USER user allocation,
2031 * %GFP_KERNEL kernel allocation,
2032 * %GFP_HIGHMEM highmem allocation,
2033 * %GFP_FS don't call back into a file system.
2034 * %GFP_ATOMIC don't sleep.
2035 * @order: Power of two of allocation size in pages. 0 is a single page.
2037 * Allocate a page from the kernel page pool. When not in
2038 * interrupt context and apply the current process NUMA policy.
2039 * Returns NULL when no page can be allocated.
2041 * Don't call cpuset_update_task_memory_state() unless
2042 * 1) it's ok to take cpuset_sem (can WAIT), and
2043 * 2) allocating for current task (not interrupt).
2045 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2047 struct mempolicy
*pol
= &default_policy
;
2049 unsigned int cpuset_mems_cookie
;
2051 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2052 pol
= get_task_policy(current
);
2055 cpuset_mems_cookie
= read_mems_allowed_begin();
2058 * No reference counting needed for current->mempolicy
2059 * nor system default_policy
2061 if (pol
->mode
== MPOL_INTERLEAVE
)
2062 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2064 page
= __alloc_pages_nodemask(gfp
, order
,
2065 policy_zonelist(gfp
, pol
, numa_node_id()),
2066 policy_nodemask(gfp
, pol
));
2068 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2073 EXPORT_SYMBOL(alloc_pages_current
);
2075 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2077 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2080 return PTR_ERR(pol
);
2081 dst
->vm_policy
= pol
;
2086 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2087 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2088 * with the mems_allowed returned by cpuset_mems_allowed(). This
2089 * keeps mempolicies cpuset relative after its cpuset moves. See
2090 * further kernel/cpuset.c update_nodemask().
2092 * current's mempolicy may be rebinded by the other task(the task that changes
2093 * cpuset's mems), so we needn't do rebind work for current task.
2096 /* Slow path of a mempolicy duplicate */
2097 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2099 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2102 return ERR_PTR(-ENOMEM
);
2104 /* task's mempolicy is protected by alloc_lock */
2105 if (old
== current
->mempolicy
) {
2108 task_unlock(current
);
2112 if (current_cpuset_is_being_rebound()) {
2113 nodemask_t mems
= cpuset_mems_allowed(current
);
2114 if (new->flags
& MPOL_F_REBINDING
)
2115 mpol_rebind_policy(new, &mems
, MPOL_REBIND_STEP2
);
2117 mpol_rebind_policy(new, &mems
, MPOL_REBIND_ONCE
);
2119 atomic_set(&new->refcnt
, 1);
2123 /* Slow path of a mempolicy comparison */
2124 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2128 if (a
->mode
!= b
->mode
)
2130 if (a
->flags
!= b
->flags
)
2132 if (mpol_store_user_nodemask(a
))
2133 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2139 case MPOL_INTERLEAVE
:
2140 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2141 case MPOL_PREFERRED
:
2142 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2150 * Shared memory backing store policy support.
2152 * Remember policies even when nobody has shared memory mapped.
2153 * The policies are kept in Red-Black tree linked from the inode.
2154 * They are protected by the sp->lock rwlock, which should be held
2155 * for any accesses to the tree.
2159 * lookup first element intersecting start-end. Caller holds sp->lock for
2160 * reading or for writing
2162 static struct sp_node
*
2163 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2165 struct rb_node
*n
= sp
->root
.rb_node
;
2168 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2170 if (start
>= p
->end
)
2172 else if (end
<= p
->start
)
2180 struct sp_node
*w
= NULL
;
2181 struct rb_node
*prev
= rb_prev(n
);
2184 w
= rb_entry(prev
, struct sp_node
, nd
);
2185 if (w
->end
<= start
)
2189 return rb_entry(n
, struct sp_node
, nd
);
2193 * Insert a new shared policy into the list. Caller holds sp->lock for
2196 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2198 struct rb_node
**p
= &sp
->root
.rb_node
;
2199 struct rb_node
*parent
= NULL
;
2204 nd
= rb_entry(parent
, struct sp_node
, nd
);
2205 if (new->start
< nd
->start
)
2207 else if (new->end
> nd
->end
)
2208 p
= &(*p
)->rb_right
;
2212 rb_link_node(&new->nd
, parent
, p
);
2213 rb_insert_color(&new->nd
, &sp
->root
);
2214 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2215 new->policy
? new->policy
->mode
: 0);
2218 /* Find shared policy intersecting idx */
2220 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2222 struct mempolicy
*pol
= NULL
;
2225 if (!sp
->root
.rb_node
)
2227 read_lock(&sp
->lock
);
2228 sn
= sp_lookup(sp
, idx
, idx
+1);
2230 mpol_get(sn
->policy
);
2233 read_unlock(&sp
->lock
);
2237 static void sp_free(struct sp_node
*n
)
2239 mpol_put(n
->policy
);
2240 kmem_cache_free(sn_cache
, n
);
2244 * mpol_misplaced - check whether current page node is valid in policy
2246 * @page: page to be checked
2247 * @vma: vm area where page mapped
2248 * @addr: virtual address where page mapped
2250 * Lookup current policy node id for vma,addr and "compare to" page's
2254 * -1 - not misplaced, page is in the right node
2255 * node - node id where the page should be
2257 * Policy determination "mimics" alloc_page_vma().
2258 * Called from fault path where we know the vma and faulting address.
2260 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2262 struct mempolicy
*pol
;
2264 int curnid
= page_to_nid(page
);
2265 unsigned long pgoff
;
2266 int thiscpu
= raw_smp_processor_id();
2267 int thisnid
= cpu_to_node(thiscpu
);
2273 pol
= get_vma_policy(vma
, addr
);
2274 if (!(pol
->flags
& MPOL_F_MOF
))
2277 switch (pol
->mode
) {
2278 case MPOL_INTERLEAVE
:
2279 BUG_ON(addr
>= vma
->vm_end
);
2280 BUG_ON(addr
< vma
->vm_start
);
2282 pgoff
= vma
->vm_pgoff
;
2283 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2284 polnid
= offset_il_node(pol
, vma
, pgoff
);
2287 case MPOL_PREFERRED
:
2288 if (pol
->flags
& MPOL_F_LOCAL
)
2289 polnid
= numa_node_id();
2291 polnid
= pol
->v
.preferred_node
;
2297 * allows binding to multiple nodes.
2298 * use current page if in policy nodemask,
2299 * else select nearest allowed node, if any.
2300 * If no allowed nodes, use current [!misplaced].
2302 if (node_isset(curnid
, pol
->v
.nodes
))
2304 z
= first_zones_zonelist(
2305 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2306 gfp_zone(GFP_HIGHUSER
),
2308 polnid
= z
->zone
->node
;
2315 /* Migrate the page towards the node whose CPU is referencing it */
2316 if (pol
->flags
& MPOL_F_MORON
) {
2319 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2323 if (curnid
!= polnid
)
2332 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2333 * dropped after task->mempolicy is set to NULL so that any allocation done as
2334 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2337 void mpol_put_task_policy(struct task_struct
*task
)
2339 struct mempolicy
*pol
;
2342 pol
= task
->mempolicy
;
2343 task
->mempolicy
= NULL
;
2348 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2350 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2351 rb_erase(&n
->nd
, &sp
->root
);
2355 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2356 unsigned long end
, struct mempolicy
*pol
)
2358 node
->start
= start
;
2363 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2364 struct mempolicy
*pol
)
2367 struct mempolicy
*newpol
;
2369 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2373 newpol
= mpol_dup(pol
);
2374 if (IS_ERR(newpol
)) {
2375 kmem_cache_free(sn_cache
, n
);
2378 newpol
->flags
|= MPOL_F_SHARED
;
2379 sp_node_init(n
, start
, end
, newpol
);
2384 /* Replace a policy range. */
2385 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2386 unsigned long end
, struct sp_node
*new)
2389 struct sp_node
*n_new
= NULL
;
2390 struct mempolicy
*mpol_new
= NULL
;
2394 write_lock(&sp
->lock
);
2395 n
= sp_lookup(sp
, start
, end
);
2396 /* Take care of old policies in the same range. */
2397 while (n
&& n
->start
< end
) {
2398 struct rb_node
*next
= rb_next(&n
->nd
);
2399 if (n
->start
>= start
) {
2405 /* Old policy spanning whole new range. */
2410 *mpol_new
= *n
->policy
;
2411 atomic_set(&mpol_new
->refcnt
, 1);
2412 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2414 sp_insert(sp
, n_new
);
2423 n
= rb_entry(next
, struct sp_node
, nd
);
2427 write_unlock(&sp
->lock
);
2434 kmem_cache_free(sn_cache
, n_new
);
2439 write_unlock(&sp
->lock
);
2441 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2444 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2451 * mpol_shared_policy_init - initialize shared policy for inode
2452 * @sp: pointer to inode shared policy
2453 * @mpol: struct mempolicy to install
2455 * Install non-NULL @mpol in inode's shared policy rb-tree.
2456 * On entry, the current task has a reference on a non-NULL @mpol.
2457 * This must be released on exit.
2458 * This is called at get_inode() calls and we can use GFP_KERNEL.
2460 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2464 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2465 rwlock_init(&sp
->lock
);
2468 struct vm_area_struct pvma
;
2469 struct mempolicy
*new;
2470 NODEMASK_SCRATCH(scratch
);
2474 /* contextualize the tmpfs mount point mempolicy */
2475 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2477 goto free_scratch
; /* no valid nodemask intersection */
2480 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2481 task_unlock(current
);
2485 /* Create pseudo-vma that contains just the policy */
2486 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2487 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2488 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2491 mpol_put(new); /* drop initial ref */
2493 NODEMASK_SCRATCH_FREE(scratch
);
2495 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2499 int mpol_set_shared_policy(struct shared_policy
*info
,
2500 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2503 struct sp_node
*new = NULL
;
2504 unsigned long sz
= vma_pages(vma
);
2506 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2508 sz
, npol
? npol
->mode
: -1,
2509 npol
? npol
->flags
: -1,
2510 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2513 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2517 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2523 /* Free a backing policy store on inode delete. */
2524 void mpol_free_shared_policy(struct shared_policy
*p
)
2527 struct rb_node
*next
;
2529 if (!p
->root
.rb_node
)
2531 write_lock(&p
->lock
);
2532 next
= rb_first(&p
->root
);
2534 n
= rb_entry(next
, struct sp_node
, nd
);
2535 next
= rb_next(&n
->nd
);
2538 write_unlock(&p
->lock
);
2541 #ifdef CONFIG_NUMA_BALANCING
2542 static int __initdata numabalancing_override
;
2544 static void __init
check_numabalancing_enable(void)
2546 bool numabalancing_default
= false;
2548 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2549 numabalancing_default
= true;
2551 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2552 if (numabalancing_override
)
2553 set_numabalancing_state(numabalancing_override
== 1);
2555 if (num_online_nodes() > 1 && !numabalancing_override
) {
2556 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2557 numabalancing_default
? "Enabling" : "Disabling");
2558 set_numabalancing_state(numabalancing_default
);
2562 static int __init
setup_numabalancing(char *str
)
2568 if (!strcmp(str
, "enable")) {
2569 numabalancing_override
= 1;
2571 } else if (!strcmp(str
, "disable")) {
2572 numabalancing_override
= -1;
2577 pr_warn("Unable to parse numa_balancing=\n");
2581 __setup("numa_balancing=", setup_numabalancing
);
2583 static inline void __init
check_numabalancing_enable(void)
2586 #endif /* CONFIG_NUMA_BALANCING */
2588 /* assumes fs == KERNEL_DS */
2589 void __init
numa_policy_init(void)
2591 nodemask_t interleave_nodes
;
2592 unsigned long largest
= 0;
2593 int nid
, prefer
= 0;
2595 policy_cache
= kmem_cache_create("numa_policy",
2596 sizeof(struct mempolicy
),
2597 0, SLAB_PANIC
, NULL
);
2599 sn_cache
= kmem_cache_create("shared_policy_node",
2600 sizeof(struct sp_node
),
2601 0, SLAB_PANIC
, NULL
);
2603 for_each_node(nid
) {
2604 preferred_node_policy
[nid
] = (struct mempolicy
) {
2605 .refcnt
= ATOMIC_INIT(1),
2606 .mode
= MPOL_PREFERRED
,
2607 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2608 .v
= { .preferred_node
= nid
, },
2613 * Set interleaving policy for system init. Interleaving is only
2614 * enabled across suitably sized nodes (default is >= 16MB), or
2615 * fall back to the largest node if they're all smaller.
2617 nodes_clear(interleave_nodes
);
2618 for_each_node_state(nid
, N_MEMORY
) {
2619 unsigned long total_pages
= node_present_pages(nid
);
2621 /* Preserve the largest node */
2622 if (largest
< total_pages
) {
2623 largest
= total_pages
;
2627 /* Interleave this node? */
2628 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2629 node_set(nid
, interleave_nodes
);
2632 /* All too small, use the largest */
2633 if (unlikely(nodes_empty(interleave_nodes
)))
2634 node_set(prefer
, interleave_nodes
);
2636 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2637 pr_err("%s: interleaving failed\n", __func__
);
2639 check_numabalancing_enable();
2642 /* Reset policy of current process to default */
2643 void numa_default_policy(void)
2645 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2649 * Parse and format mempolicy from/to strings
2653 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2655 static const char * const policy_modes
[] =
2657 [MPOL_DEFAULT
] = "default",
2658 [MPOL_PREFERRED
] = "prefer",
2659 [MPOL_BIND
] = "bind",
2660 [MPOL_INTERLEAVE
] = "interleave",
2661 [MPOL_LOCAL
] = "local",
2667 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2668 * @str: string containing mempolicy to parse
2669 * @mpol: pointer to struct mempolicy pointer, returned on success.
2672 * <mode>[=<flags>][:<nodelist>]
2674 * On success, returns 0, else 1
2676 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2678 struct mempolicy
*new = NULL
;
2679 unsigned short mode
;
2680 unsigned short mode_flags
;
2682 char *nodelist
= strchr(str
, ':');
2683 char *flags
= strchr(str
, '=');
2687 /* NUL-terminate mode or flags string */
2689 if (nodelist_parse(nodelist
, nodes
))
2691 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2697 *flags
++ = '\0'; /* terminate mode string */
2699 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2700 if (!strcmp(str
, policy_modes
[mode
])) {
2704 if (mode
>= MPOL_MAX
)
2708 case MPOL_PREFERRED
:
2710 * Insist on a nodelist of one node only
2713 char *rest
= nodelist
;
2714 while (isdigit(*rest
))
2720 case MPOL_INTERLEAVE
:
2722 * Default to online nodes with memory if no nodelist
2725 nodes
= node_states
[N_MEMORY
];
2729 * Don't allow a nodelist; mpol_new() checks flags
2733 mode
= MPOL_PREFERRED
;
2737 * Insist on a empty nodelist
2744 * Insist on a nodelist
2753 * Currently, we only support two mutually exclusive
2756 if (!strcmp(flags
, "static"))
2757 mode_flags
|= MPOL_F_STATIC_NODES
;
2758 else if (!strcmp(flags
, "relative"))
2759 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2764 new = mpol_new(mode
, mode_flags
, &nodes
);
2769 * Save nodes for mpol_to_str() to show the tmpfs mount options
2770 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2772 if (mode
!= MPOL_PREFERRED
)
2773 new->v
.nodes
= nodes
;
2775 new->v
.preferred_node
= first_node(nodes
);
2777 new->flags
|= MPOL_F_LOCAL
;
2780 * Save nodes for contextualization: this will be used to "clone"
2781 * the mempolicy in a specific context [cpuset] at a later time.
2783 new->w
.user_nodemask
= nodes
;
2788 /* Restore string for error message */
2797 #endif /* CONFIG_TMPFS */
2800 * mpol_to_str - format a mempolicy structure for printing
2801 * @buffer: to contain formatted mempolicy string
2802 * @maxlen: length of @buffer
2803 * @pol: pointer to mempolicy to be formatted
2805 * Convert @pol into a string. If @buffer is too short, truncate the string.
2806 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2807 * longest flag, "relative", and to display at least a few node ids.
2809 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2812 nodemask_t nodes
= NODE_MASK_NONE
;
2813 unsigned short mode
= MPOL_DEFAULT
;
2814 unsigned short flags
= 0;
2816 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2824 case MPOL_PREFERRED
:
2825 if (flags
& MPOL_F_LOCAL
)
2828 node_set(pol
->v
.preferred_node
, nodes
);
2831 case MPOL_INTERLEAVE
:
2832 nodes
= pol
->v
.nodes
;
2836 snprintf(p
, maxlen
, "unknown");
2840 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2842 if (flags
& MPOL_MODE_FLAGS
) {
2843 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2846 * Currently, the only defined flags are mutually exclusive
2848 if (flags
& MPOL_F_STATIC_NODES
)
2849 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2850 else if (flags
& MPOL_F_RELATIVE_NODES
)
2851 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2854 if (!nodes_empty(nodes
))
2855 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2856 nodemask_pr_args(&nodes
));