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 /* Check that the nodemask contains at least one populated zone */
166 static int is_valid_nodemask(const nodemask_t
*nodemask
)
168 return nodes_intersects(*nodemask
, node_states
[N_MEMORY
]);
171 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
173 return pol
->flags
& MPOL_MODE_FLAGS
;
176 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
177 const nodemask_t
*rel
)
180 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
181 nodes_onto(*ret
, tmp
, *rel
);
184 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
186 if (nodes_empty(*nodes
))
188 pol
->v
.nodes
= *nodes
;
192 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
195 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
196 else if (nodes_empty(*nodes
))
197 return -EINVAL
; /* no allowed nodes */
199 pol
->v
.preferred_node
= first_node(*nodes
);
203 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
205 if (!is_valid_nodemask(nodes
))
207 pol
->v
.nodes
= *nodes
;
212 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
213 * any, for the new policy. mpol_new() has already validated the nodes
214 * parameter with respect to the policy mode and flags. But, we need to
215 * handle an empty nodemask with MPOL_PREFERRED here.
217 * Must be called holding task's alloc_lock to protect task's mems_allowed
218 * and mempolicy. May also be called holding the mmap_semaphore for write.
220 static int mpol_set_nodemask(struct mempolicy
*pol
,
221 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
225 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
229 nodes_and(nsc
->mask1
,
230 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
233 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
234 nodes
= NULL
; /* explicit local allocation */
236 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
237 mpol_relative_nodemask(&nsc
->mask2
, nodes
,&nsc
->mask1
);
239 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
241 if (mpol_store_user_nodemask(pol
))
242 pol
->w
.user_nodemask
= *nodes
;
244 pol
->w
.cpuset_mems_allowed
=
245 cpuset_current_mems_allowed
;
249 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
251 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
256 * This function just creates a new policy, does some check and simple
257 * initialization. You must invoke mpol_set_nodemask() to set nodes.
259 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
262 struct mempolicy
*policy
;
264 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
265 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
267 if (mode
== MPOL_DEFAULT
) {
268 if (nodes
&& !nodes_empty(*nodes
))
269 return ERR_PTR(-EINVAL
);
275 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
276 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
277 * All other modes require a valid pointer to a non-empty nodemask.
279 if (mode
== MPOL_PREFERRED
) {
280 if (nodes_empty(*nodes
)) {
281 if (((flags
& MPOL_F_STATIC_NODES
) ||
282 (flags
& MPOL_F_RELATIVE_NODES
)))
283 return ERR_PTR(-EINVAL
);
285 } else if (mode
== MPOL_LOCAL
) {
286 if (!nodes_empty(*nodes
))
287 return ERR_PTR(-EINVAL
);
288 mode
= MPOL_PREFERRED
;
289 } else if (nodes_empty(*nodes
))
290 return ERR_PTR(-EINVAL
);
291 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
293 return ERR_PTR(-ENOMEM
);
294 atomic_set(&policy
->refcnt
, 1);
296 policy
->flags
= flags
;
301 /* Slow path of a mpol destructor. */
302 void __mpol_put(struct mempolicy
*p
)
304 if (!atomic_dec_and_test(&p
->refcnt
))
306 kmem_cache_free(policy_cache
, p
);
309 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
,
310 enum mpol_rebind_step step
)
316 * MPOL_REBIND_ONCE - do rebind work at once
317 * MPOL_REBIND_STEP1 - set all the newly nodes
318 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
320 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
,
321 enum mpol_rebind_step step
)
325 if (pol
->flags
& MPOL_F_STATIC_NODES
)
326 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
327 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
328 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
331 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
334 if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP1
) {
335 nodes_remap(tmp
, pol
->v
.nodes
,
336 pol
->w
.cpuset_mems_allowed
, *nodes
);
337 pol
->w
.cpuset_mems_allowed
= step
? tmp
: *nodes
;
338 } else if (step
== MPOL_REBIND_STEP2
) {
339 tmp
= pol
->w
.cpuset_mems_allowed
;
340 pol
->w
.cpuset_mems_allowed
= *nodes
;
345 if (nodes_empty(tmp
))
348 if (step
== MPOL_REBIND_STEP1
)
349 nodes_or(pol
->v
.nodes
, pol
->v
.nodes
, tmp
);
350 else if (step
== MPOL_REBIND_ONCE
|| step
== MPOL_REBIND_STEP2
)
355 if (!node_isset(current
->il_next
, tmp
)) {
356 current
->il_next
= next_node(current
->il_next
, tmp
);
357 if (current
->il_next
>= MAX_NUMNODES
)
358 current
->il_next
= first_node(tmp
);
359 if (current
->il_next
>= MAX_NUMNODES
)
360 current
->il_next
= numa_node_id();
364 static void mpol_rebind_preferred(struct mempolicy
*pol
,
365 const nodemask_t
*nodes
,
366 enum mpol_rebind_step step
)
370 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
371 int node
= first_node(pol
->w
.user_nodemask
);
373 if (node_isset(node
, *nodes
)) {
374 pol
->v
.preferred_node
= node
;
375 pol
->flags
&= ~MPOL_F_LOCAL
;
377 pol
->flags
|= MPOL_F_LOCAL
;
378 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
379 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
380 pol
->v
.preferred_node
= first_node(tmp
);
381 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
382 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
383 pol
->w
.cpuset_mems_allowed
,
385 pol
->w
.cpuset_mems_allowed
= *nodes
;
390 * mpol_rebind_policy - Migrate a policy to a different set of nodes
392 * If read-side task has no lock to protect task->mempolicy, write-side
393 * task will rebind the task->mempolicy by two step. The first step is
394 * setting all the newly nodes, and the second step is cleaning all the
395 * disallowed nodes. In this way, we can avoid finding no node to alloc
397 * If we have a lock to protect task->mempolicy in read-side, we do
401 * MPOL_REBIND_ONCE - do rebind work at once
402 * MPOL_REBIND_STEP1 - set all the newly nodes
403 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
405 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
,
406 enum mpol_rebind_step step
)
410 if (!mpol_store_user_nodemask(pol
) && step
== MPOL_REBIND_ONCE
&&
411 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
414 if (step
== MPOL_REBIND_STEP1
&& (pol
->flags
& MPOL_F_REBINDING
))
417 if (step
== MPOL_REBIND_STEP2
&& !(pol
->flags
& MPOL_F_REBINDING
))
420 if (step
== MPOL_REBIND_STEP1
)
421 pol
->flags
|= MPOL_F_REBINDING
;
422 else if (step
== MPOL_REBIND_STEP2
)
423 pol
->flags
&= ~MPOL_F_REBINDING
;
424 else if (step
>= MPOL_REBIND_NSTEP
)
427 mpol_ops
[pol
->mode
].rebind(pol
, newmask
, step
);
431 * Wrapper for mpol_rebind_policy() that just requires task
432 * pointer, and updates task mempolicy.
434 * Called with task's alloc_lock held.
437 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new,
438 enum mpol_rebind_step step
)
440 mpol_rebind_policy(tsk
->mempolicy
, new, step
);
444 * Rebind each vma in mm to new nodemask.
446 * Call holding a reference to mm. Takes mm->mmap_sem during call.
449 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
451 struct vm_area_struct
*vma
;
453 down_write(&mm
->mmap_sem
);
454 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
455 mpol_rebind_policy(vma
->vm_policy
, new, MPOL_REBIND_ONCE
);
456 up_write(&mm
->mmap_sem
);
459 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
461 .rebind
= mpol_rebind_default
,
463 [MPOL_INTERLEAVE
] = {
464 .create
= mpol_new_interleave
,
465 .rebind
= mpol_rebind_nodemask
,
468 .create
= mpol_new_preferred
,
469 .rebind
= mpol_rebind_preferred
,
472 .create
= mpol_new_bind
,
473 .rebind
= mpol_rebind_nodemask
,
477 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
478 unsigned long flags
);
481 * Scan through pages checking if pages follow certain conditions,
482 * and move them to the pagelist if they do.
484 static int queue_pages_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
485 unsigned long addr
, unsigned long end
,
486 const nodemask_t
*nodes
, unsigned long flags
,
493 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
498 if (!pte_present(*pte
))
500 page
= vm_normal_page(vma
, addr
, *pte
);
504 * vm_normal_page() filters out zero pages, but there might
505 * still be PageReserved pages to skip, perhaps in a VDSO.
507 if (PageReserved(page
))
509 nid
= page_to_nid(page
);
510 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
513 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
514 migrate_page_add(page
, private, flags
);
517 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
518 pte_unmap_unlock(orig_pte
, ptl
);
522 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct
*vma
,
523 pmd_t
*pmd
, const nodemask_t
*nodes
, unsigned long flags
,
526 #ifdef CONFIG_HUGETLB_PAGE
532 ptl
= huge_pte_lock(hstate_vma(vma
), vma
->vm_mm
, (pte_t
*)pmd
);
533 entry
= huge_ptep_get((pte_t
*)pmd
);
534 if (!pte_present(entry
))
536 page
= pte_page(entry
);
537 nid
= page_to_nid(page
);
538 if (node_isset(nid
, *nodes
) == !!(flags
& MPOL_MF_INVERT
))
540 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
541 if (flags
& (MPOL_MF_MOVE_ALL
) ||
542 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
543 isolate_huge_page(page
, private);
551 static inline int queue_pages_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
552 unsigned long addr
, unsigned long end
,
553 const nodemask_t
*nodes
, unsigned long flags
,
559 pmd
= pmd_offset(pud
, addr
);
561 next
= pmd_addr_end(addr
, end
);
562 if (!pmd_present(*pmd
))
564 if (pmd_huge(*pmd
) && is_vm_hugetlb_page(vma
)) {
565 queue_pages_hugetlb_pmd_range(vma
, pmd
, nodes
,
569 split_huge_page_pmd(vma
, addr
, pmd
);
570 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
572 if (queue_pages_pte_range(vma
, pmd
, addr
, next
, nodes
,
575 } while (pmd
++, addr
= next
, addr
!= end
);
579 static inline int queue_pages_pud_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
580 unsigned long addr
, unsigned long end
,
581 const nodemask_t
*nodes
, unsigned long flags
,
587 pud
= pud_offset(pgd
, addr
);
589 next
= pud_addr_end(addr
, end
);
590 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
))
592 if (pud_none_or_clear_bad(pud
))
594 if (queue_pages_pmd_range(vma
, pud
, addr
, next
, nodes
,
597 } while (pud
++, addr
= next
, addr
!= end
);
601 static inline int queue_pages_pgd_range(struct vm_area_struct
*vma
,
602 unsigned long addr
, unsigned long end
,
603 const nodemask_t
*nodes
, unsigned long flags
,
609 pgd
= pgd_offset(vma
->vm_mm
, addr
);
611 next
= pgd_addr_end(addr
, end
);
612 if (pgd_none_or_clear_bad(pgd
))
614 if (queue_pages_pud_range(vma
, pgd
, addr
, next
, nodes
,
617 } while (pgd
++, addr
= next
, addr
!= end
);
621 #ifdef CONFIG_NUMA_BALANCING
623 * This is used to mark a range of virtual addresses to be inaccessible.
624 * These are later cleared by a NUMA hinting fault. Depending on these
625 * faults, pages may be migrated for better NUMA placement.
627 * This is assuming that NUMA faults are handled using PROT_NONE. If
628 * an architecture makes a different choice, it will need further
629 * changes to the core.
631 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
632 unsigned long addr
, unsigned long end
)
636 nr_updated
= change_protection(vma
, addr
, end
, vma
->vm_page_prot
, 0, 1);
638 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
643 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
644 unsigned long addr
, unsigned long end
)
648 #endif /* CONFIG_NUMA_BALANCING */
651 * Walk through page tables and collect pages to be migrated.
653 * If pages found in a given range are on a set of nodes (determined by
654 * @nodes and @flags,) it's isolated and queued to the pagelist which is
655 * passed via @private.)
658 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
659 const nodemask_t
*nodes
, unsigned long flags
, void *private)
662 struct vm_area_struct
*vma
, *prev
;
664 vma
= find_vma(mm
, start
);
668 for (; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
) {
669 unsigned long endvma
= vma
->vm_end
;
673 if (vma
->vm_start
> start
)
674 start
= vma
->vm_start
;
676 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
677 if (!vma
->vm_next
&& vma
->vm_end
< end
)
679 if (prev
&& prev
->vm_end
< vma
->vm_start
)
683 if (flags
& MPOL_MF_LAZY
) {
684 change_prot_numa(vma
, start
, endvma
);
688 if ((flags
& MPOL_MF_STRICT
) ||
689 ((flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) &&
690 vma_migratable(vma
))) {
692 err
= queue_pages_pgd_range(vma
, start
, endvma
, nodes
,
704 * Apply policy to a single VMA
705 * This must be called with the mmap_sem held for writing.
707 static int vma_replace_policy(struct vm_area_struct
*vma
,
708 struct mempolicy
*pol
)
711 struct mempolicy
*old
;
712 struct mempolicy
*new;
714 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
715 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
716 vma
->vm_ops
, vma
->vm_file
,
717 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
723 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
724 err
= vma
->vm_ops
->set_policy(vma
, new);
729 old
= vma
->vm_policy
;
730 vma
->vm_policy
= new; /* protected by mmap_sem */
739 /* Step 2: apply policy to a range and do splits. */
740 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
741 unsigned long end
, struct mempolicy
*new_pol
)
743 struct vm_area_struct
*next
;
744 struct vm_area_struct
*prev
;
745 struct vm_area_struct
*vma
;
748 unsigned long vmstart
;
751 vma
= find_vma(mm
, start
);
752 if (!vma
|| vma
->vm_start
> start
)
756 if (start
> vma
->vm_start
)
759 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
761 vmstart
= max(start
, vma
->vm_start
);
762 vmend
= min(end
, vma
->vm_end
);
764 if (mpol_equal(vma_policy(vma
), new_pol
))
767 pgoff
= vma
->vm_pgoff
+
768 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
769 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
770 vma
->anon_vma
, vma
->vm_file
, pgoff
,
775 if (mpol_equal(vma_policy(vma
), new_pol
))
777 /* vma_merge() joined vma && vma->next, case 8 */
780 if (vma
->vm_start
!= vmstart
) {
781 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
785 if (vma
->vm_end
!= vmend
) {
786 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
791 err
= vma_replace_policy(vma
, new_pol
);
800 /* Set the process memory policy */
801 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
804 struct mempolicy
*new, *old
;
805 NODEMASK_SCRATCH(scratch
);
811 new = mpol_new(mode
, flags
, nodes
);
818 ret
= mpol_set_nodemask(new, nodes
, scratch
);
820 task_unlock(current
);
824 old
= current
->mempolicy
;
825 current
->mempolicy
= new;
826 if (new && new->mode
== MPOL_INTERLEAVE
&&
827 nodes_weight(new->v
.nodes
))
828 current
->il_next
= first_node(new->v
.nodes
);
829 task_unlock(current
);
833 NODEMASK_SCRATCH_FREE(scratch
);
838 * Return nodemask for policy for get_mempolicy() query
840 * Called with task's alloc_lock held
842 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
845 if (p
== &default_policy
)
851 case MPOL_INTERLEAVE
:
855 if (!(p
->flags
& MPOL_F_LOCAL
))
856 node_set(p
->v
.preferred_node
, *nodes
);
857 /* else return empty node mask for local allocation */
864 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
869 err
= get_user_pages(current
, mm
, addr
& PAGE_MASK
, 1, 0, 0, &p
, NULL
);
871 err
= page_to_nid(p
);
877 /* Retrieve NUMA policy */
878 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
879 unsigned long addr
, unsigned long flags
)
882 struct mm_struct
*mm
= current
->mm
;
883 struct vm_area_struct
*vma
= NULL
;
884 struct mempolicy
*pol
= current
->mempolicy
;
887 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
890 if (flags
& MPOL_F_MEMS_ALLOWED
) {
891 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
893 *policy
= 0; /* just so it's initialized */
895 *nmask
= cpuset_current_mems_allowed
;
896 task_unlock(current
);
900 if (flags
& MPOL_F_ADDR
) {
902 * Do NOT fall back to task policy if the
903 * vma/shared policy at addr is NULL. We
904 * want to return MPOL_DEFAULT in this case.
906 down_read(&mm
->mmap_sem
);
907 vma
= find_vma_intersection(mm
, addr
, addr
+1);
909 up_read(&mm
->mmap_sem
);
912 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
913 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
915 pol
= vma
->vm_policy
;
920 pol
= &default_policy
; /* indicates default behavior */
922 if (flags
& MPOL_F_NODE
) {
923 if (flags
& MPOL_F_ADDR
) {
924 err
= lookup_node(mm
, addr
);
928 } else if (pol
== current
->mempolicy
&&
929 pol
->mode
== MPOL_INTERLEAVE
) {
930 *policy
= current
->il_next
;
936 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
939 * Internal mempolicy flags must be masked off before exposing
940 * the policy to userspace.
942 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
946 up_read(¤t
->mm
->mmap_sem
);
952 if (mpol_store_user_nodemask(pol
)) {
953 *nmask
= pol
->w
.user_nodemask
;
956 get_policy_nodemask(pol
, nmask
);
957 task_unlock(current
);
964 up_read(¤t
->mm
->mmap_sem
);
968 #ifdef CONFIG_MIGRATION
972 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
976 * Avoid migrating a page that is shared with others.
978 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(page
) == 1) {
979 if (!isolate_lru_page(page
)) {
980 list_add_tail(&page
->lru
, pagelist
);
981 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
982 page_is_file_cache(page
));
987 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
990 return alloc_huge_page_node(page_hstate(compound_head(page
)),
993 return alloc_pages_exact_node(node
, GFP_HIGHUSER_MOVABLE
, 0);
997 * Migrate pages from one node to a target node.
998 * Returns error or the number of pages not migrated.
1000 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1004 LIST_HEAD(pagelist
);
1008 node_set(source
, nmask
);
1011 * This does not "check" the range but isolates all pages that
1012 * need migration. Between passing in the full user address
1013 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1015 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1016 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1017 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1019 if (!list_empty(&pagelist
)) {
1020 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
1021 MIGRATE_SYNC
, MR_SYSCALL
);
1023 putback_movable_pages(&pagelist
);
1030 * Move pages between the two nodesets so as to preserve the physical
1031 * layout as much as possible.
1033 * Returns the number of page that could not be moved.
1035 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1036 const nodemask_t
*to
, int flags
)
1042 err
= migrate_prep();
1046 down_read(&mm
->mmap_sem
);
1048 err
= migrate_vmas(mm
, from
, to
, flags
);
1053 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1054 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1055 * bit in 'tmp', and return that <source, dest> pair for migration.
1056 * The pair of nodemasks 'to' and 'from' define the map.
1058 * If no pair of bits is found that way, fallback to picking some
1059 * pair of 'source' and 'dest' bits that are not the same. If the
1060 * 'source' and 'dest' bits are the same, this represents a node
1061 * that will be migrating to itself, so no pages need move.
1063 * If no bits are left in 'tmp', or if all remaining bits left
1064 * in 'tmp' correspond to the same bit in 'to', return false
1065 * (nothing left to migrate).
1067 * This lets us pick a pair of nodes to migrate between, such that
1068 * if possible the dest node is not already occupied by some other
1069 * source node, minimizing the risk of overloading the memory on a
1070 * node that would happen if we migrated incoming memory to a node
1071 * before migrating outgoing memory source that same node.
1073 * A single scan of tmp is sufficient. As we go, we remember the
1074 * most recent <s, d> pair that moved (s != d). If we find a pair
1075 * that not only moved, but what's better, moved to an empty slot
1076 * (d is not set in tmp), then we break out then, with that pair.
1077 * Otherwise when we finish scanning from_tmp, we at least have the
1078 * most recent <s, d> pair that moved. If we get all the way through
1079 * the scan of tmp without finding any node that moved, much less
1080 * moved to an empty node, then there is nothing left worth migrating.
1084 while (!nodes_empty(tmp
)) {
1086 int source
= NUMA_NO_NODE
;
1089 for_each_node_mask(s
, tmp
) {
1092 * do_migrate_pages() tries to maintain the relative
1093 * node relationship of the pages established between
1094 * threads and memory areas.
1096 * However if the number of source nodes is not equal to
1097 * the number of destination nodes we can not preserve
1098 * this node relative relationship. In that case, skip
1099 * copying memory from a node that is in the destination
1102 * Example: [2,3,4] -> [3,4,5] moves everything.
1103 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1106 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1107 (node_isset(s
, *to
)))
1110 d
= node_remap(s
, *from
, *to
);
1114 source
= s
; /* Node moved. Memorize */
1117 /* dest not in remaining from nodes? */
1118 if (!node_isset(dest
, tmp
))
1121 if (source
== NUMA_NO_NODE
)
1124 node_clear(source
, tmp
);
1125 err
= migrate_to_node(mm
, source
, dest
, flags
);
1132 up_read(&mm
->mmap_sem
);
1140 * Allocate a new page for page migration based on vma policy.
1141 * Start by assuming the page is mapped by the same vma as contains @start.
1142 * Search forward from there, if not. N.B., this assumes that the
1143 * list of pages handed to migrate_pages()--which is how we get here--
1144 * is in virtual address order.
1146 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1148 struct vm_area_struct
*vma
;
1149 unsigned long uninitialized_var(address
);
1151 vma
= find_vma(current
->mm
, start
);
1153 address
= page_address_in_vma(page
, vma
);
1154 if (address
!= -EFAULT
)
1159 if (PageHuge(page
)) {
1161 return alloc_huge_page_noerr(vma
, address
, 1);
1164 * if !vma, alloc_page_vma() will use task or system default policy
1166 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1170 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1171 unsigned long flags
)
1175 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1176 const nodemask_t
*to
, int flags
)
1181 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1187 static long do_mbind(unsigned long start
, unsigned long len
,
1188 unsigned short mode
, unsigned short mode_flags
,
1189 nodemask_t
*nmask
, unsigned long flags
)
1191 struct mm_struct
*mm
= current
->mm
;
1192 struct mempolicy
*new;
1195 LIST_HEAD(pagelist
);
1197 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1199 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1202 if (start
& ~PAGE_MASK
)
1205 if (mode
== MPOL_DEFAULT
)
1206 flags
&= ~MPOL_MF_STRICT
;
1208 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1216 new = mpol_new(mode
, mode_flags
, nmask
);
1218 return PTR_ERR(new);
1220 if (flags
& MPOL_MF_LAZY
)
1221 new->flags
|= MPOL_F_MOF
;
1224 * If we are using the default policy then operation
1225 * on discontinuous address spaces is okay after all
1228 flags
|= MPOL_MF_DISCONTIG_OK
;
1230 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1231 start
, start
+ len
, mode
, mode_flags
,
1232 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1234 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1236 err
= migrate_prep();
1241 NODEMASK_SCRATCH(scratch
);
1243 down_write(&mm
->mmap_sem
);
1245 err
= mpol_set_nodemask(new, nmask
, scratch
);
1246 task_unlock(current
);
1248 up_write(&mm
->mmap_sem
);
1251 NODEMASK_SCRATCH_FREE(scratch
);
1256 err
= queue_pages_range(mm
, start
, end
, nmask
,
1257 flags
| MPOL_MF_INVERT
, &pagelist
);
1259 err
= mbind_range(mm
, start
, end
, new);
1264 if (!list_empty(&pagelist
)) {
1265 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1266 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1267 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1269 putback_movable_pages(&pagelist
);
1272 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1275 putback_movable_pages(&pagelist
);
1277 up_write(&mm
->mmap_sem
);
1284 * User space interface with variable sized bitmaps for nodelists.
1287 /* Copy a node mask from user space. */
1288 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1289 unsigned long maxnode
)
1292 unsigned long nlongs
;
1293 unsigned long endmask
;
1296 nodes_clear(*nodes
);
1297 if (maxnode
== 0 || !nmask
)
1299 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1302 nlongs
= BITS_TO_LONGS(maxnode
);
1303 if ((maxnode
% BITS_PER_LONG
) == 0)
1306 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1308 /* When the user specified more nodes than supported just check
1309 if the non supported part is all zero. */
1310 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1311 if (nlongs
> PAGE_SIZE
/sizeof(long))
1313 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1315 if (get_user(t
, nmask
+ k
))
1317 if (k
== nlongs
- 1) {
1323 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1327 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1329 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1333 /* Copy a kernel node mask to user space */
1334 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1337 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1338 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1340 if (copy
> nbytes
) {
1341 if (copy
> PAGE_SIZE
)
1343 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1347 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1350 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1351 unsigned long, mode
, const unsigned long __user
*, nmask
,
1352 unsigned long, maxnode
, unsigned, flags
)
1356 unsigned short mode_flags
;
1358 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1359 mode
&= ~MPOL_MODE_FLAGS
;
1360 if (mode
>= MPOL_MAX
)
1362 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1363 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1365 err
= get_nodes(&nodes
, nmask
, maxnode
);
1368 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1371 /* Set the process memory policy */
1372 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1373 unsigned long, maxnode
)
1377 unsigned short flags
;
1379 flags
= mode
& MPOL_MODE_FLAGS
;
1380 mode
&= ~MPOL_MODE_FLAGS
;
1381 if ((unsigned int)mode
>= MPOL_MAX
)
1383 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1385 err
= get_nodes(&nodes
, nmask
, maxnode
);
1388 return do_set_mempolicy(mode
, flags
, &nodes
);
1391 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1392 const unsigned long __user
*, old_nodes
,
1393 const unsigned long __user
*, new_nodes
)
1395 const struct cred
*cred
= current_cred(), *tcred
;
1396 struct mm_struct
*mm
= NULL
;
1397 struct task_struct
*task
;
1398 nodemask_t task_nodes
;
1402 NODEMASK_SCRATCH(scratch
);
1407 old
= &scratch
->mask1
;
1408 new = &scratch
->mask2
;
1410 err
= get_nodes(old
, old_nodes
, maxnode
);
1414 err
= get_nodes(new, new_nodes
, maxnode
);
1418 /* Find the mm_struct */
1420 task
= pid
? find_task_by_vpid(pid
) : current
;
1426 get_task_struct(task
);
1431 * Check if this process has the right to modify the specified
1432 * process. The right exists if the process has administrative
1433 * capabilities, superuser privileges or the same
1434 * userid as the target process.
1436 tcred
= __task_cred(task
);
1437 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1438 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1439 !capable(CAP_SYS_NICE
)) {
1446 task_nodes
= cpuset_mems_allowed(task
);
1447 /* Is the user allowed to access the target nodes? */
1448 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1453 if (!nodes_subset(*new, node_states
[N_MEMORY
])) {
1458 err
= security_task_movememory(task
);
1462 mm
= get_task_mm(task
);
1463 put_task_struct(task
);
1470 err
= do_migrate_pages(mm
, old
, new,
1471 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1475 NODEMASK_SCRATCH_FREE(scratch
);
1480 put_task_struct(task
);
1486 /* Retrieve NUMA policy */
1487 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1488 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1489 unsigned long, addr
, unsigned long, flags
)
1492 int uninitialized_var(pval
);
1495 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1498 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1503 if (policy
&& put_user(pval
, policy
))
1507 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1512 #ifdef CONFIG_COMPAT
1514 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1515 compat_ulong_t __user
*, nmask
,
1516 compat_ulong_t
, maxnode
,
1517 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1520 unsigned long __user
*nm
= NULL
;
1521 unsigned long nr_bits
, alloc_size
;
1522 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1524 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1525 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1528 nm
= compat_alloc_user_space(alloc_size
);
1530 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1532 if (!err
&& nmask
) {
1533 unsigned long copy_size
;
1534 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1535 err
= copy_from_user(bm
, nm
, copy_size
);
1536 /* ensure entire bitmap is zeroed */
1537 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1538 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1544 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1545 compat_ulong_t
, maxnode
)
1548 unsigned long __user
*nm
= NULL
;
1549 unsigned long nr_bits
, alloc_size
;
1550 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1552 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1553 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1556 err
= compat_get_bitmap(bm
, nmask
, nr_bits
);
1557 nm
= compat_alloc_user_space(alloc_size
);
1558 err
|= copy_to_user(nm
, bm
, alloc_size
);
1564 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1567 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1568 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1569 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1572 unsigned long __user
*nm
= NULL
;
1573 unsigned long nr_bits
, alloc_size
;
1576 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1577 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1580 err
= compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
);
1581 nm
= compat_alloc_user_space(alloc_size
);
1582 err
|= copy_to_user(nm
, nodes_addr(bm
), alloc_size
);
1588 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1593 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1596 struct mempolicy
*pol
= NULL
;
1599 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1600 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1601 } else if (vma
->vm_policy
) {
1602 pol
= vma
->vm_policy
;
1605 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1606 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1607 * count on these policies which will be dropped by
1608 * mpol_cond_put() later
1610 if (mpol_needs_cond_ref(pol
))
1619 * get_vma_policy(@task, @vma, @addr)
1620 * @task: task for fallback if vma policy == default
1621 * @vma: virtual memory area whose policy is sought
1622 * @addr: address in @vma for shared policy lookup
1624 * Returns effective policy for a VMA at specified address.
1625 * Falls back to @task or system default policy, as necessary.
1626 * Current or other task's task mempolicy and non-shared vma policies must be
1627 * protected by task_lock(task) by the caller.
1628 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1629 * count--added by the get_policy() vm_op, as appropriate--to protect against
1630 * freeing by another task. It is the caller's responsibility to free the
1631 * extra reference for shared policies.
1633 struct mempolicy
*get_vma_policy(struct task_struct
*task
,
1634 struct vm_area_struct
*vma
, unsigned long addr
)
1636 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1639 pol
= get_task_policy(task
);
1644 bool vma_policy_mof(struct vm_area_struct
*vma
)
1646 struct mempolicy
*pol
;
1648 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1651 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1652 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1659 pol
= vma
->vm_policy
;
1661 pol
= get_task_policy(current
);
1663 return pol
->flags
& MPOL_F_MOF
;
1666 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1668 enum zone_type dynamic_policy_zone
= policy_zone
;
1670 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1673 * if policy->v.nodes has movable memory only,
1674 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1676 * policy->v.nodes is intersect with node_states[N_MEMORY].
1677 * so if the following test faile, it implies
1678 * policy->v.nodes has movable memory only.
1680 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1681 dynamic_policy_zone
= ZONE_MOVABLE
;
1683 return zone
>= dynamic_policy_zone
;
1687 * Return a nodemask representing a mempolicy for filtering nodes for
1690 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1692 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1693 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1694 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1695 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1696 return &policy
->v
.nodes
;
1701 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1702 static struct zonelist
*policy_zonelist(gfp_t gfp
, struct mempolicy
*policy
,
1705 switch (policy
->mode
) {
1706 case MPOL_PREFERRED
:
1707 if (!(policy
->flags
& MPOL_F_LOCAL
))
1708 nd
= policy
->v
.preferred_node
;
1712 * Normally, MPOL_BIND allocations are node-local within the
1713 * allowed nodemask. However, if __GFP_THISNODE is set and the
1714 * current node isn't part of the mask, we use the zonelist for
1715 * the first node in the mask instead.
1717 if (unlikely(gfp
& __GFP_THISNODE
) &&
1718 unlikely(!node_isset(nd
, policy
->v
.nodes
)))
1719 nd
= first_node(policy
->v
.nodes
);
1724 return node_zonelist(nd
, gfp
);
1727 /* Do dynamic interleaving for a process */
1728 static unsigned interleave_nodes(struct mempolicy
*policy
)
1731 struct task_struct
*me
= current
;
1734 next
= next_node(nid
, policy
->v
.nodes
);
1735 if (next
>= MAX_NUMNODES
)
1736 next
= first_node(policy
->v
.nodes
);
1737 if (next
< MAX_NUMNODES
)
1743 * Depending on the memory policy provide a node from which to allocate the
1746 unsigned int mempolicy_slab_node(void)
1748 struct mempolicy
*policy
;
1749 int node
= numa_mem_id();
1754 policy
= current
->mempolicy
;
1755 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1758 switch (policy
->mode
) {
1759 case MPOL_PREFERRED
:
1761 * handled MPOL_F_LOCAL above
1763 return policy
->v
.preferred_node
;
1765 case MPOL_INTERLEAVE
:
1766 return interleave_nodes(policy
);
1770 * Follow bind policy behavior and start allocation at the
1773 struct zonelist
*zonelist
;
1775 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1776 zonelist
= &NODE_DATA(node
)->node_zonelists
[0];
1777 (void)first_zones_zonelist(zonelist
, highest_zoneidx
,
1780 return zone
? zone
->node
: node
;
1788 /* Do static interleaving for a VMA with known offset. */
1789 static unsigned offset_il_node(struct mempolicy
*pol
,
1790 struct vm_area_struct
*vma
, unsigned long off
)
1792 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1795 int nid
= NUMA_NO_NODE
;
1798 return numa_node_id();
1799 target
= (unsigned int)off
% nnodes
;
1802 nid
= next_node(nid
, pol
->v
.nodes
);
1804 } while (c
<= target
);
1808 /* Determine a node number for interleave */
1809 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1810 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1816 * for small pages, there is no difference between
1817 * shift and PAGE_SHIFT, so the bit-shift is safe.
1818 * for huge pages, since vm_pgoff is in units of small
1819 * pages, we need to shift off the always 0 bits to get
1822 BUG_ON(shift
< PAGE_SHIFT
);
1823 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1824 off
+= (addr
- vma
->vm_start
) >> shift
;
1825 return offset_il_node(pol
, vma
, off
);
1827 return interleave_nodes(pol
);
1831 * Return the bit number of a random bit set in the nodemask.
1832 * (returns NUMA_NO_NODE if nodemask is empty)
1834 int node_random(const nodemask_t
*maskp
)
1836 int w
, bit
= NUMA_NO_NODE
;
1838 w
= nodes_weight(*maskp
);
1840 bit
= bitmap_ord_to_pos(maskp
->bits
,
1841 get_random_int() % w
, MAX_NUMNODES
);
1845 #ifdef CONFIG_HUGETLBFS
1847 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1848 * @vma: virtual memory area whose policy is sought
1849 * @addr: address in @vma for shared policy lookup and interleave policy
1850 * @gfp_flags: for requested zone
1851 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1852 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1854 * Returns a zonelist suitable for a huge page allocation and a pointer
1855 * to the struct mempolicy for conditional unref after allocation.
1856 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1857 * @nodemask for filtering the zonelist.
1859 * Must be protected by read_mems_allowed_begin()
1861 struct zonelist
*huge_zonelist(struct vm_area_struct
*vma
, unsigned long addr
,
1862 gfp_t gfp_flags
, struct mempolicy
**mpol
,
1863 nodemask_t
**nodemask
)
1865 struct zonelist
*zl
;
1867 *mpol
= get_vma_policy(current
, vma
, addr
);
1868 *nodemask
= NULL
; /* assume !MPOL_BIND */
1870 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1871 zl
= node_zonelist(interleave_nid(*mpol
, vma
, addr
,
1872 huge_page_shift(hstate_vma(vma
))), gfp_flags
);
1874 zl
= policy_zonelist(gfp_flags
, *mpol
, numa_node_id());
1875 if ((*mpol
)->mode
== MPOL_BIND
)
1876 *nodemask
= &(*mpol
)->v
.nodes
;
1882 * init_nodemask_of_mempolicy
1884 * If the current task's mempolicy is "default" [NULL], return 'false'
1885 * to indicate default policy. Otherwise, extract the policy nodemask
1886 * for 'bind' or 'interleave' policy into the argument nodemask, or
1887 * initialize the argument nodemask to contain the single node for
1888 * 'preferred' or 'local' policy and return 'true' to indicate presence
1889 * of non-default mempolicy.
1891 * We don't bother with reference counting the mempolicy [mpol_get/put]
1892 * because the current task is examining it's own mempolicy and a task's
1893 * mempolicy is only ever changed by the task itself.
1895 * N.B., it is the caller's responsibility to free a returned nodemask.
1897 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1899 struct mempolicy
*mempolicy
;
1902 if (!(mask
&& current
->mempolicy
))
1906 mempolicy
= current
->mempolicy
;
1907 switch (mempolicy
->mode
) {
1908 case MPOL_PREFERRED
:
1909 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1910 nid
= numa_node_id();
1912 nid
= mempolicy
->v
.preferred_node
;
1913 init_nodemask_of_node(mask
, nid
);
1918 case MPOL_INTERLEAVE
:
1919 *mask
= mempolicy
->v
.nodes
;
1925 task_unlock(current
);
1932 * mempolicy_nodemask_intersects
1934 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1935 * policy. Otherwise, check for intersection between mask and the policy
1936 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1937 * policy, always return true since it may allocate elsewhere on fallback.
1939 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1941 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1942 const nodemask_t
*mask
)
1944 struct mempolicy
*mempolicy
;
1950 mempolicy
= tsk
->mempolicy
;
1954 switch (mempolicy
->mode
) {
1955 case MPOL_PREFERRED
:
1957 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1958 * allocate from, they may fallback to other nodes when oom.
1959 * Thus, it's possible for tsk to have allocated memory from
1964 case MPOL_INTERLEAVE
:
1965 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1975 /* Allocate a page in interleaved policy.
1976 Own path because it needs to do special accounting. */
1977 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1980 struct zonelist
*zl
;
1983 zl
= node_zonelist(nid
, gfp
);
1984 page
= __alloc_pages(gfp
, order
, zl
);
1985 if (page
&& page_zone(page
) == zonelist_zone(&zl
->_zonerefs
[0]))
1986 inc_zone_page_state(page
, NUMA_INTERLEAVE_HIT
);
1991 * alloc_pages_vma - Allocate a page for a VMA.
1994 * %GFP_USER user allocation.
1995 * %GFP_KERNEL kernel allocations,
1996 * %GFP_HIGHMEM highmem/user allocations,
1997 * %GFP_FS allocation should not call back into a file system.
1998 * %GFP_ATOMIC don't sleep.
2000 * @order:Order of the GFP allocation.
2001 * @vma: Pointer to VMA or NULL if not available.
2002 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2004 * This function allocates a page from the kernel page pool and applies
2005 * a NUMA policy associated with the VMA or the current process.
2006 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2007 * mm_struct of the VMA to prevent it from going away. Should be used for
2008 * all allocations for pages that will be mapped into
2009 * user space. Returns NULL when no page can be allocated.
2011 * Should be called with the mm_sem of the vma hold.
2014 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
2015 unsigned long addr
, int node
)
2017 struct mempolicy
*pol
;
2019 unsigned int cpuset_mems_cookie
;
2022 pol
= get_vma_policy(current
, vma
, addr
);
2023 cpuset_mems_cookie
= read_mems_allowed_begin();
2025 if (unlikely(pol
->mode
== MPOL_INTERLEAVE
)) {
2028 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2030 page
= alloc_page_interleave(gfp
, order
, nid
);
2031 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2036 page
= __alloc_pages_nodemask(gfp
, order
,
2037 policy_zonelist(gfp
, pol
, node
),
2038 policy_nodemask(gfp
, pol
));
2040 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2046 * alloc_pages_current - Allocate pages.
2049 * %GFP_USER user allocation,
2050 * %GFP_KERNEL kernel allocation,
2051 * %GFP_HIGHMEM highmem allocation,
2052 * %GFP_FS don't call back into a file system.
2053 * %GFP_ATOMIC don't sleep.
2054 * @order: Power of two of allocation size in pages. 0 is a single page.
2056 * Allocate a page from the kernel page pool. When not in
2057 * interrupt context and apply the current process NUMA policy.
2058 * Returns NULL when no page can be allocated.
2060 * Don't call cpuset_update_task_memory_state() unless
2061 * 1) it's ok to take cpuset_sem (can WAIT), and
2062 * 2) allocating for current task (not interrupt).
2064 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2066 struct mempolicy
*pol
= &default_policy
;
2068 unsigned int cpuset_mems_cookie
;
2070 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2071 pol
= get_task_policy(current
);
2074 cpuset_mems_cookie
= read_mems_allowed_begin();
2077 * No reference counting needed for current->mempolicy
2078 * nor system default_policy
2080 if (pol
->mode
== MPOL_INTERLEAVE
)
2081 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2083 page
= __alloc_pages_nodemask(gfp
, order
,
2084 policy_zonelist(gfp
, pol
, numa_node_id()),
2085 policy_nodemask(gfp
, pol
));
2087 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2092 EXPORT_SYMBOL(alloc_pages_current
);
2094 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2096 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2099 return PTR_ERR(pol
);
2100 dst
->vm_policy
= pol
;
2105 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2106 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2107 * with the mems_allowed returned by cpuset_mems_allowed(). This
2108 * keeps mempolicies cpuset relative after its cpuset moves. See
2109 * further kernel/cpuset.c update_nodemask().
2111 * current's mempolicy may be rebinded by the other task(the task that changes
2112 * cpuset's mems), so we needn't do rebind work for current task.
2115 /* Slow path of a mempolicy duplicate */
2116 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2118 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2121 return ERR_PTR(-ENOMEM
);
2123 /* task's mempolicy is protected by alloc_lock */
2124 if (old
== current
->mempolicy
) {
2127 task_unlock(current
);
2131 if (current_cpuset_is_being_rebound()) {
2132 nodemask_t mems
= cpuset_mems_allowed(current
);
2133 if (new->flags
& MPOL_F_REBINDING
)
2134 mpol_rebind_policy(new, &mems
, MPOL_REBIND_STEP2
);
2136 mpol_rebind_policy(new, &mems
, MPOL_REBIND_ONCE
);
2138 atomic_set(&new->refcnt
, 1);
2142 /* Slow path of a mempolicy comparison */
2143 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2147 if (a
->mode
!= b
->mode
)
2149 if (a
->flags
!= b
->flags
)
2151 if (mpol_store_user_nodemask(a
))
2152 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2158 case MPOL_INTERLEAVE
:
2159 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2160 case MPOL_PREFERRED
:
2161 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2169 * Shared memory backing store policy support.
2171 * Remember policies even when nobody has shared memory mapped.
2172 * The policies are kept in Red-Black tree linked from the inode.
2173 * They are protected by the sp->lock spinlock, which should be held
2174 * for any accesses to the tree.
2177 /* lookup first element intersecting start-end */
2178 /* Caller holds sp->lock */
2179 static struct sp_node
*
2180 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2182 struct rb_node
*n
= sp
->root
.rb_node
;
2185 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2187 if (start
>= p
->end
)
2189 else if (end
<= p
->start
)
2197 struct sp_node
*w
= NULL
;
2198 struct rb_node
*prev
= rb_prev(n
);
2201 w
= rb_entry(prev
, struct sp_node
, nd
);
2202 if (w
->end
<= start
)
2206 return rb_entry(n
, struct sp_node
, nd
);
2209 /* Insert a new shared policy into the list. */
2210 /* Caller holds sp->lock */
2211 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2213 struct rb_node
**p
= &sp
->root
.rb_node
;
2214 struct rb_node
*parent
= NULL
;
2219 nd
= rb_entry(parent
, struct sp_node
, nd
);
2220 if (new->start
< nd
->start
)
2222 else if (new->end
> nd
->end
)
2223 p
= &(*p
)->rb_right
;
2227 rb_link_node(&new->nd
, parent
, p
);
2228 rb_insert_color(&new->nd
, &sp
->root
);
2229 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2230 new->policy
? new->policy
->mode
: 0);
2233 /* Find shared policy intersecting idx */
2235 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2237 struct mempolicy
*pol
= NULL
;
2240 if (!sp
->root
.rb_node
)
2242 spin_lock(&sp
->lock
);
2243 sn
= sp_lookup(sp
, idx
, idx
+1);
2245 mpol_get(sn
->policy
);
2248 spin_unlock(&sp
->lock
);
2252 static void sp_free(struct sp_node
*n
)
2254 mpol_put(n
->policy
);
2255 kmem_cache_free(sn_cache
, n
);
2259 * mpol_misplaced - check whether current page node is valid in policy
2261 * @page: page to be checked
2262 * @vma: vm area where page mapped
2263 * @addr: virtual address where page mapped
2265 * Lookup current policy node id for vma,addr and "compare to" page's
2269 * -1 - not misplaced, page is in the right node
2270 * node - node id where the page should be
2272 * Policy determination "mimics" alloc_page_vma().
2273 * Called from fault path where we know the vma and faulting address.
2275 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2277 struct mempolicy
*pol
;
2279 int curnid
= page_to_nid(page
);
2280 unsigned long pgoff
;
2281 int thiscpu
= raw_smp_processor_id();
2282 int thisnid
= cpu_to_node(thiscpu
);
2288 pol
= get_vma_policy(current
, vma
, addr
);
2289 if (!(pol
->flags
& MPOL_F_MOF
))
2292 switch (pol
->mode
) {
2293 case MPOL_INTERLEAVE
:
2294 BUG_ON(addr
>= vma
->vm_end
);
2295 BUG_ON(addr
< vma
->vm_start
);
2297 pgoff
= vma
->vm_pgoff
;
2298 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2299 polnid
= offset_il_node(pol
, vma
, pgoff
);
2302 case MPOL_PREFERRED
:
2303 if (pol
->flags
& MPOL_F_LOCAL
)
2304 polnid
= numa_node_id();
2306 polnid
= pol
->v
.preferred_node
;
2311 * allows binding to multiple nodes.
2312 * use current page if in policy nodemask,
2313 * else select nearest allowed node, if any.
2314 * If no allowed nodes, use current [!misplaced].
2316 if (node_isset(curnid
, pol
->v
.nodes
))
2318 (void)first_zones_zonelist(
2319 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2320 gfp_zone(GFP_HIGHUSER
),
2321 &pol
->v
.nodes
, &zone
);
2322 polnid
= zone
->node
;
2329 /* Migrate the page towards the node whose CPU is referencing it */
2330 if (pol
->flags
& MPOL_F_MORON
) {
2333 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2337 if (curnid
!= polnid
)
2345 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2347 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2348 rb_erase(&n
->nd
, &sp
->root
);
2352 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2353 unsigned long end
, struct mempolicy
*pol
)
2355 node
->start
= start
;
2360 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2361 struct mempolicy
*pol
)
2364 struct mempolicy
*newpol
;
2366 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2370 newpol
= mpol_dup(pol
);
2371 if (IS_ERR(newpol
)) {
2372 kmem_cache_free(sn_cache
, n
);
2375 newpol
->flags
|= MPOL_F_SHARED
;
2376 sp_node_init(n
, start
, end
, newpol
);
2381 /* Replace a policy range. */
2382 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2383 unsigned long end
, struct sp_node
*new)
2386 struct sp_node
*n_new
= NULL
;
2387 struct mempolicy
*mpol_new
= NULL
;
2391 spin_lock(&sp
->lock
);
2392 n
= sp_lookup(sp
, start
, end
);
2393 /* Take care of old policies in the same range. */
2394 while (n
&& n
->start
< end
) {
2395 struct rb_node
*next
= rb_next(&n
->nd
);
2396 if (n
->start
>= start
) {
2402 /* Old policy spanning whole new range. */
2407 *mpol_new
= *n
->policy
;
2408 atomic_set(&mpol_new
->refcnt
, 1);
2409 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2411 sp_insert(sp
, n_new
);
2420 n
= rb_entry(next
, struct sp_node
, nd
);
2424 spin_unlock(&sp
->lock
);
2431 kmem_cache_free(sn_cache
, n_new
);
2436 spin_unlock(&sp
->lock
);
2438 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2441 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2448 * mpol_shared_policy_init - initialize shared policy for inode
2449 * @sp: pointer to inode shared policy
2450 * @mpol: struct mempolicy to install
2452 * Install non-NULL @mpol in inode's shared policy rb-tree.
2453 * On entry, the current task has a reference on a non-NULL @mpol.
2454 * This must be released on exit.
2455 * This is called at get_inode() calls and we can use GFP_KERNEL.
2457 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2461 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2462 spin_lock_init(&sp
->lock
);
2465 struct vm_area_struct pvma
;
2466 struct mempolicy
*new;
2467 NODEMASK_SCRATCH(scratch
);
2471 /* contextualize the tmpfs mount point mempolicy */
2472 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2474 goto free_scratch
; /* no valid nodemask intersection */
2477 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2478 task_unlock(current
);
2482 /* Create pseudo-vma that contains just the policy */
2483 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2484 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2485 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2488 mpol_put(new); /* drop initial ref */
2490 NODEMASK_SCRATCH_FREE(scratch
);
2492 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2496 int mpol_set_shared_policy(struct shared_policy
*info
,
2497 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2500 struct sp_node
*new = NULL
;
2501 unsigned long sz
= vma_pages(vma
);
2503 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2505 sz
, npol
? npol
->mode
: -1,
2506 npol
? npol
->flags
: -1,
2507 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2510 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2514 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2520 /* Free a backing policy store on inode delete. */
2521 void mpol_free_shared_policy(struct shared_policy
*p
)
2524 struct rb_node
*next
;
2526 if (!p
->root
.rb_node
)
2528 spin_lock(&p
->lock
);
2529 next
= rb_first(&p
->root
);
2531 n
= rb_entry(next
, struct sp_node
, nd
);
2532 next
= rb_next(&n
->nd
);
2535 spin_unlock(&p
->lock
);
2538 #ifdef CONFIG_NUMA_BALANCING
2539 static int __initdata numabalancing_override
;
2541 static void __init
check_numabalancing_enable(void)
2543 bool numabalancing_default
= false;
2545 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2546 numabalancing_default
= true;
2548 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2549 if (numabalancing_override
)
2550 set_numabalancing_state(numabalancing_override
== 1);
2552 if (nr_node_ids
> 1 && !numabalancing_override
) {
2553 pr_info("%s automatic NUMA balancing. "
2554 "Configure with numa_balancing= or the "
2555 "kernel.numa_balancing sysctl",
2556 numabalancing_default
? "Enabling" : "Disabling");
2557 set_numabalancing_state(numabalancing_default
);
2561 static int __init
setup_numabalancing(char *str
)
2567 if (!strcmp(str
, "enable")) {
2568 numabalancing_override
= 1;
2570 } else if (!strcmp(str
, "disable")) {
2571 numabalancing_override
= -1;
2576 pr_warn("Unable to parse numa_balancing=\n");
2580 __setup("numa_balancing=", setup_numabalancing
);
2582 static inline void __init
check_numabalancing_enable(void)
2585 #endif /* CONFIG_NUMA_BALANCING */
2587 /* assumes fs == KERNEL_DS */
2588 void __init
numa_policy_init(void)
2590 nodemask_t interleave_nodes
;
2591 unsigned long largest
= 0;
2592 int nid
, prefer
= 0;
2594 policy_cache
= kmem_cache_create("numa_policy",
2595 sizeof(struct mempolicy
),
2596 0, SLAB_PANIC
, NULL
);
2598 sn_cache
= kmem_cache_create("shared_policy_node",
2599 sizeof(struct sp_node
),
2600 0, SLAB_PANIC
, NULL
);
2602 for_each_node(nid
) {
2603 preferred_node_policy
[nid
] = (struct mempolicy
) {
2604 .refcnt
= ATOMIC_INIT(1),
2605 .mode
= MPOL_PREFERRED
,
2606 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2607 .v
= { .preferred_node
= nid
, },
2612 * Set interleaving policy for system init. Interleaving is only
2613 * enabled across suitably sized nodes (default is >= 16MB), or
2614 * fall back to the largest node if they're all smaller.
2616 nodes_clear(interleave_nodes
);
2617 for_each_node_state(nid
, N_MEMORY
) {
2618 unsigned long total_pages
= node_present_pages(nid
);
2620 /* Preserve the largest node */
2621 if (largest
< total_pages
) {
2622 largest
= total_pages
;
2626 /* Interleave this node? */
2627 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2628 node_set(nid
, interleave_nodes
);
2631 /* All too small, use the largest */
2632 if (unlikely(nodes_empty(interleave_nodes
)))
2633 node_set(prefer
, interleave_nodes
);
2635 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2636 pr_err("%s: interleaving failed\n", __func__
);
2638 check_numabalancing_enable();
2641 /* Reset policy of current process to default */
2642 void numa_default_policy(void)
2644 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2648 * Parse and format mempolicy from/to strings
2652 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2654 static const char * const policy_modes
[] =
2656 [MPOL_DEFAULT
] = "default",
2657 [MPOL_PREFERRED
] = "prefer",
2658 [MPOL_BIND
] = "bind",
2659 [MPOL_INTERLEAVE
] = "interleave",
2660 [MPOL_LOCAL
] = "local",
2666 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2667 * @str: string containing mempolicy to parse
2668 * @mpol: pointer to struct mempolicy pointer, returned on success.
2671 * <mode>[=<flags>][:<nodelist>]
2673 * On success, returns 0, else 1
2675 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2677 struct mempolicy
*new = NULL
;
2678 unsigned short mode
;
2679 unsigned short mode_flags
;
2681 char *nodelist
= strchr(str
, ':');
2682 char *flags
= strchr(str
, '=');
2686 /* NUL-terminate mode or flags string */
2688 if (nodelist_parse(nodelist
, nodes
))
2690 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2696 *flags
++ = '\0'; /* terminate mode string */
2698 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2699 if (!strcmp(str
, policy_modes
[mode
])) {
2703 if (mode
>= MPOL_MAX
)
2707 case MPOL_PREFERRED
:
2709 * Insist on a nodelist of one node only
2712 char *rest
= nodelist
;
2713 while (isdigit(*rest
))
2719 case MPOL_INTERLEAVE
:
2721 * Default to online nodes with memory if no nodelist
2724 nodes
= node_states
[N_MEMORY
];
2728 * Don't allow a nodelist; mpol_new() checks flags
2732 mode
= MPOL_PREFERRED
;
2736 * Insist on a empty nodelist
2743 * Insist on a nodelist
2752 * Currently, we only support two mutually exclusive
2755 if (!strcmp(flags
, "static"))
2756 mode_flags
|= MPOL_F_STATIC_NODES
;
2757 else if (!strcmp(flags
, "relative"))
2758 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2763 new = mpol_new(mode
, mode_flags
, &nodes
);
2768 * Save nodes for mpol_to_str() to show the tmpfs mount options
2769 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2771 if (mode
!= MPOL_PREFERRED
)
2772 new->v
.nodes
= nodes
;
2774 new->v
.preferred_node
= first_node(nodes
);
2776 new->flags
|= MPOL_F_LOCAL
;
2779 * Save nodes for contextualization: this will be used to "clone"
2780 * the mempolicy in a specific context [cpuset] at a later time.
2782 new->w
.user_nodemask
= nodes
;
2787 /* Restore string for error message */
2796 #endif /* CONFIG_TMPFS */
2799 * mpol_to_str - format a mempolicy structure for printing
2800 * @buffer: to contain formatted mempolicy string
2801 * @maxlen: length of @buffer
2802 * @pol: pointer to mempolicy to be formatted
2804 * Convert @pol into a string. If @buffer is too short, truncate the string.
2805 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2806 * longest flag, "relative", and to display at least a few node ids.
2808 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2811 nodemask_t nodes
= NODE_MASK_NONE
;
2812 unsigned short mode
= MPOL_DEFAULT
;
2813 unsigned short flags
= 0;
2815 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2823 case MPOL_PREFERRED
:
2824 if (flags
& MPOL_F_LOCAL
)
2827 node_set(pol
->v
.preferred_node
, nodes
);
2830 case MPOL_INTERLEAVE
:
2831 nodes
= pol
->v
.nodes
;
2835 snprintf(p
, maxlen
, "unknown");
2839 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2841 if (flags
& MPOL_MODE_FLAGS
) {
2842 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2845 * Currently, the only defined flags are mutually exclusive
2847 if (flags
& MPOL_F_STATIC_NODES
)
2848 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2849 else if (flags
& MPOL_F_RELATIVE_NODES
)
2850 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2853 if (!nodes_empty(nodes
)) {
2854 p
+= snprintf(p
, buffer
+ maxlen
- p
, ":");
2855 p
+= nodelist_scnprintf(p
, buffer
+ maxlen
- p
, nodes
);