1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
10 #include <linux/sched/mm.h>
11 #include <linux/sched/task.h>
12 #include <linux/hugetlb.h>
13 #include <linux/mman.h>
14 #include <linux/slab.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/swap.h>
17 #include <linux/vmalloc.h>
18 #include <linux/pagemap.h>
19 #include <linux/namei.h>
20 #include <linux/shmem_fs.h>
21 #include <linux/blkdev.h>
22 #include <linux/random.h>
23 #include <linux/writeback.h>
24 #include <linux/proc_fs.h>
25 #include <linux/seq_file.h>
26 #include <linux/init.h>
27 #include <linux/ksm.h>
28 #include <linux/rmap.h>
29 #include <linux/security.h>
30 #include <linux/backing-dev.h>
31 #include <linux/mutex.h>
32 #include <linux/capability.h>
33 #include <linux/syscalls.h>
34 #include <linux/memcontrol.h>
35 #include <linux/poll.h>
36 #include <linux/oom.h>
37 #include <linux/frontswap.h>
38 #include <linux/swapfile.h>
39 #include <linux/export.h>
40 #include <linux/swap_slots.h>
41 #include <linux/sort.h>
43 #include <asm/tlbflush.h>
44 #include <linux/swapops.h>
45 #include <linux/swap_cgroup.h>
47 static bool swap_count_continued(struct swap_info_struct
*, pgoff_t
,
49 static void free_swap_count_continuations(struct swap_info_struct
*);
50 static sector_t
map_swap_entry(swp_entry_t
, struct block_device
**);
52 DEFINE_SPINLOCK(swap_lock
);
53 static unsigned int nr_swapfiles
;
54 atomic_long_t nr_swap_pages
;
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
60 EXPORT_SYMBOL_GPL(nr_swap_pages
);
61 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62 long total_swap_pages
;
63 static int least_priority
= -1;
65 static const char Bad_file
[] = "Bad swap file entry ";
66 static const char Unused_file
[] = "Unused swap file entry ";
67 static const char Bad_offset
[] = "Bad swap offset entry ";
68 static const char Unused_offset
[] = "Unused swap offset entry ";
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
74 PLIST_HEAD(swap_active_head
);
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
88 static struct plist_head
*swap_avail_heads
;
89 static DEFINE_SPINLOCK(swap_avail_lock
);
91 struct swap_info_struct
*swap_info
[MAX_SWAPFILES
];
93 static DEFINE_MUTEX(swapon_mutex
);
95 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait
);
96 /* Activity counter to indicate that a swapon or swapoff has occurred */
97 static atomic_t proc_poll_event
= ATOMIC_INIT(0);
99 atomic_t nr_rotate_swap
= ATOMIC_INIT(0);
101 static struct swap_info_struct
*swap_type_to_swap_info(int type
)
103 if (type
>= READ_ONCE(nr_swapfiles
))
106 smp_rmb(); /* Pairs with smp_wmb in alloc_swap_info. */
107 return READ_ONCE(swap_info
[type
]);
110 static inline unsigned char swap_count(unsigned char ent
)
112 return ent
& ~SWAP_HAS_CACHE
; /* may include COUNT_CONTINUED flag */
115 /* Reclaim the swap entry anyway if possible */
116 #define TTRS_ANYWAY 0x1
118 * Reclaim the swap entry if there are no more mappings of the
121 #define TTRS_UNMAPPED 0x2
122 /* Reclaim the swap entry if swap is getting full*/
123 #define TTRS_FULL 0x4
125 /* returns 1 if swap entry is freed */
126 static int __try_to_reclaim_swap(struct swap_info_struct
*si
,
127 unsigned long offset
, unsigned long flags
)
129 swp_entry_t entry
= swp_entry(si
->type
, offset
);
133 page
= find_get_page(swap_address_space(entry
), offset
);
137 * When this function is called from scan_swap_map_slots() and it's
138 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
139 * here. We have to use trylock for avoiding deadlock. This is a special
140 * case and you should use try_to_free_swap() with explicit lock_page()
141 * in usual operations.
143 if (trylock_page(page
)) {
144 if ((flags
& TTRS_ANYWAY
) ||
145 ((flags
& TTRS_UNMAPPED
) && !page_mapped(page
)) ||
146 ((flags
& TTRS_FULL
) && mem_cgroup_swap_full(page
)))
147 ret
= try_to_free_swap(page
);
154 static inline struct swap_extent
*first_se(struct swap_info_struct
*sis
)
156 struct rb_node
*rb
= rb_first(&sis
->swap_extent_root
);
157 return rb_entry(rb
, struct swap_extent
, rb_node
);
160 static inline struct swap_extent
*next_se(struct swap_extent
*se
)
162 struct rb_node
*rb
= rb_next(&se
->rb_node
);
163 return rb
? rb_entry(rb
, struct swap_extent
, rb_node
) : NULL
;
167 * swapon tell device that all the old swap contents can be discarded,
168 * to allow the swap device to optimize its wear-levelling.
170 static int discard_swap(struct swap_info_struct
*si
)
172 struct swap_extent
*se
;
173 sector_t start_block
;
177 /* Do not discard the swap header page! */
179 start_block
= (se
->start_block
+ 1) << (PAGE_SHIFT
- 9);
180 nr_blocks
= ((sector_t
)se
->nr_pages
- 1) << (PAGE_SHIFT
- 9);
182 err
= blkdev_issue_discard(si
->bdev
, start_block
,
183 nr_blocks
, GFP_KERNEL
, 0);
189 for (se
= next_se(se
); se
; se
= next_se(se
)) {
190 start_block
= se
->start_block
<< (PAGE_SHIFT
- 9);
191 nr_blocks
= (sector_t
)se
->nr_pages
<< (PAGE_SHIFT
- 9);
193 err
= blkdev_issue_discard(si
->bdev
, start_block
,
194 nr_blocks
, GFP_KERNEL
, 0);
200 return err
; /* That will often be -EOPNOTSUPP */
203 static struct swap_extent
*
204 offset_to_swap_extent(struct swap_info_struct
*sis
, unsigned long offset
)
206 struct swap_extent
*se
;
209 rb
= sis
->swap_extent_root
.rb_node
;
211 se
= rb_entry(rb
, struct swap_extent
, rb_node
);
212 if (offset
< se
->start_page
)
214 else if (offset
>= se
->start_page
+ se
->nr_pages
)
219 /* It *must* be present */
223 sector_t
swap_page_sector(struct page
*page
)
225 struct swap_info_struct
*sis
= page_swap_info(page
);
226 struct swap_extent
*se
;
230 offset
= __page_file_index(page
);
231 se
= offset_to_swap_extent(sis
, offset
);
232 sector
= se
->start_block
+ (offset
- se
->start_page
);
233 return sector
<< (PAGE_SHIFT
- 9);
237 * swap allocation tell device that a cluster of swap can now be discarded,
238 * to allow the swap device to optimize its wear-levelling.
240 static void discard_swap_cluster(struct swap_info_struct
*si
,
241 pgoff_t start_page
, pgoff_t nr_pages
)
243 struct swap_extent
*se
= offset_to_swap_extent(si
, start_page
);
246 pgoff_t offset
= start_page
- se
->start_page
;
247 sector_t start_block
= se
->start_block
+ offset
;
248 sector_t nr_blocks
= se
->nr_pages
- offset
;
250 if (nr_blocks
> nr_pages
)
251 nr_blocks
= nr_pages
;
252 start_page
+= nr_blocks
;
253 nr_pages
-= nr_blocks
;
255 start_block
<<= PAGE_SHIFT
- 9;
256 nr_blocks
<<= PAGE_SHIFT
- 9;
257 if (blkdev_issue_discard(si
->bdev
, start_block
,
258 nr_blocks
, GFP_NOIO
, 0))
265 #ifdef CONFIG_THP_SWAP
266 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
268 #define swap_entry_size(size) (size)
270 #define SWAPFILE_CLUSTER 256
273 * Define swap_entry_size() as constant to let compiler to optimize
274 * out some code if !CONFIG_THP_SWAP
276 #define swap_entry_size(size) 1
278 #define LATENCY_LIMIT 256
280 static inline void cluster_set_flag(struct swap_cluster_info
*info
,
286 static inline unsigned int cluster_count(struct swap_cluster_info
*info
)
291 static inline void cluster_set_count(struct swap_cluster_info
*info
,
297 static inline void cluster_set_count_flag(struct swap_cluster_info
*info
,
298 unsigned int c
, unsigned int f
)
304 static inline unsigned int cluster_next(struct swap_cluster_info
*info
)
309 static inline void cluster_set_next(struct swap_cluster_info
*info
,
315 static inline void cluster_set_next_flag(struct swap_cluster_info
*info
,
316 unsigned int n
, unsigned int f
)
322 static inline bool cluster_is_free(struct swap_cluster_info
*info
)
324 return info
->flags
& CLUSTER_FLAG_FREE
;
327 static inline bool cluster_is_null(struct swap_cluster_info
*info
)
329 return info
->flags
& CLUSTER_FLAG_NEXT_NULL
;
332 static inline void cluster_set_null(struct swap_cluster_info
*info
)
334 info
->flags
= CLUSTER_FLAG_NEXT_NULL
;
338 static inline bool cluster_is_huge(struct swap_cluster_info
*info
)
340 if (IS_ENABLED(CONFIG_THP_SWAP
))
341 return info
->flags
& CLUSTER_FLAG_HUGE
;
345 static inline void cluster_clear_huge(struct swap_cluster_info
*info
)
347 info
->flags
&= ~CLUSTER_FLAG_HUGE
;
350 static inline struct swap_cluster_info
*lock_cluster(struct swap_info_struct
*si
,
351 unsigned long offset
)
353 struct swap_cluster_info
*ci
;
355 ci
= si
->cluster_info
;
357 ci
+= offset
/ SWAPFILE_CLUSTER
;
358 spin_lock(&ci
->lock
);
363 static inline void unlock_cluster(struct swap_cluster_info
*ci
)
366 spin_unlock(&ci
->lock
);
370 * Determine the locking method in use for this device. Return
371 * swap_cluster_info if SSD-style cluster-based locking is in place.
373 static inline struct swap_cluster_info
*lock_cluster_or_swap_info(
374 struct swap_info_struct
*si
, unsigned long offset
)
376 struct swap_cluster_info
*ci
;
378 /* Try to use fine-grained SSD-style locking if available: */
379 ci
= lock_cluster(si
, offset
);
380 /* Otherwise, fall back to traditional, coarse locking: */
382 spin_lock(&si
->lock
);
387 static inline void unlock_cluster_or_swap_info(struct swap_info_struct
*si
,
388 struct swap_cluster_info
*ci
)
393 spin_unlock(&si
->lock
);
396 static inline bool cluster_list_empty(struct swap_cluster_list
*list
)
398 return cluster_is_null(&list
->head
);
401 static inline unsigned int cluster_list_first(struct swap_cluster_list
*list
)
403 return cluster_next(&list
->head
);
406 static void cluster_list_init(struct swap_cluster_list
*list
)
408 cluster_set_null(&list
->head
);
409 cluster_set_null(&list
->tail
);
412 static void cluster_list_add_tail(struct swap_cluster_list
*list
,
413 struct swap_cluster_info
*ci
,
416 if (cluster_list_empty(list
)) {
417 cluster_set_next_flag(&list
->head
, idx
, 0);
418 cluster_set_next_flag(&list
->tail
, idx
, 0);
420 struct swap_cluster_info
*ci_tail
;
421 unsigned int tail
= cluster_next(&list
->tail
);
424 * Nested cluster lock, but both cluster locks are
425 * only acquired when we held swap_info_struct->lock
428 spin_lock_nested(&ci_tail
->lock
, SINGLE_DEPTH_NESTING
);
429 cluster_set_next(ci_tail
, idx
);
430 spin_unlock(&ci_tail
->lock
);
431 cluster_set_next_flag(&list
->tail
, idx
, 0);
435 static unsigned int cluster_list_del_first(struct swap_cluster_list
*list
,
436 struct swap_cluster_info
*ci
)
440 idx
= cluster_next(&list
->head
);
441 if (cluster_next(&list
->tail
) == idx
) {
442 cluster_set_null(&list
->head
);
443 cluster_set_null(&list
->tail
);
445 cluster_set_next_flag(&list
->head
,
446 cluster_next(&ci
[idx
]), 0);
451 /* Add a cluster to discard list and schedule it to do discard */
452 static void swap_cluster_schedule_discard(struct swap_info_struct
*si
,
456 * If scan_swap_map() can't find a free cluster, it will check
457 * si->swap_map directly. To make sure the discarding cluster isn't
458 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
459 * will be cleared after discard
461 memset(si
->swap_map
+ idx
* SWAPFILE_CLUSTER
,
462 SWAP_MAP_BAD
, SWAPFILE_CLUSTER
);
464 cluster_list_add_tail(&si
->discard_clusters
, si
->cluster_info
, idx
);
466 schedule_work(&si
->discard_work
);
469 static void __free_cluster(struct swap_info_struct
*si
, unsigned long idx
)
471 struct swap_cluster_info
*ci
= si
->cluster_info
;
473 cluster_set_flag(ci
+ idx
, CLUSTER_FLAG_FREE
);
474 cluster_list_add_tail(&si
->free_clusters
, ci
, idx
);
478 * Doing discard actually. After a cluster discard is finished, the cluster
479 * will be added to free cluster list. caller should hold si->lock.
481 static void swap_do_scheduled_discard(struct swap_info_struct
*si
)
483 struct swap_cluster_info
*info
, *ci
;
486 info
= si
->cluster_info
;
488 while (!cluster_list_empty(&si
->discard_clusters
)) {
489 idx
= cluster_list_del_first(&si
->discard_clusters
, info
);
490 spin_unlock(&si
->lock
);
492 discard_swap_cluster(si
, idx
* SWAPFILE_CLUSTER
,
495 spin_lock(&si
->lock
);
496 ci
= lock_cluster(si
, idx
* SWAPFILE_CLUSTER
);
497 __free_cluster(si
, idx
);
498 memset(si
->swap_map
+ idx
* SWAPFILE_CLUSTER
,
499 0, SWAPFILE_CLUSTER
);
504 static void swap_discard_work(struct work_struct
*work
)
506 struct swap_info_struct
*si
;
508 si
= container_of(work
, struct swap_info_struct
, discard_work
);
510 spin_lock(&si
->lock
);
511 swap_do_scheduled_discard(si
);
512 spin_unlock(&si
->lock
);
515 static void alloc_cluster(struct swap_info_struct
*si
, unsigned long idx
)
517 struct swap_cluster_info
*ci
= si
->cluster_info
;
519 VM_BUG_ON(cluster_list_first(&si
->free_clusters
) != idx
);
520 cluster_list_del_first(&si
->free_clusters
, ci
);
521 cluster_set_count_flag(ci
+ idx
, 0, 0);
524 static void free_cluster(struct swap_info_struct
*si
, unsigned long idx
)
526 struct swap_cluster_info
*ci
= si
->cluster_info
+ idx
;
528 VM_BUG_ON(cluster_count(ci
) != 0);
530 * If the swap is discardable, prepare discard the cluster
531 * instead of free it immediately. The cluster will be freed
534 if ((si
->flags
& (SWP_WRITEOK
| SWP_PAGE_DISCARD
)) ==
535 (SWP_WRITEOK
| SWP_PAGE_DISCARD
)) {
536 swap_cluster_schedule_discard(si
, idx
);
540 __free_cluster(si
, idx
);
544 * The cluster corresponding to page_nr will be used. The cluster will be
545 * removed from free cluster list and its usage counter will be increased.
547 static void inc_cluster_info_page(struct swap_info_struct
*p
,
548 struct swap_cluster_info
*cluster_info
, unsigned long page_nr
)
550 unsigned long idx
= page_nr
/ SWAPFILE_CLUSTER
;
554 if (cluster_is_free(&cluster_info
[idx
]))
555 alloc_cluster(p
, idx
);
557 VM_BUG_ON(cluster_count(&cluster_info
[idx
]) >= SWAPFILE_CLUSTER
);
558 cluster_set_count(&cluster_info
[idx
],
559 cluster_count(&cluster_info
[idx
]) + 1);
563 * The cluster corresponding to page_nr decreases one usage. If the usage
564 * counter becomes 0, which means no page in the cluster is in using, we can
565 * optionally discard the cluster and add it to free cluster list.
567 static void dec_cluster_info_page(struct swap_info_struct
*p
,
568 struct swap_cluster_info
*cluster_info
, unsigned long page_nr
)
570 unsigned long idx
= page_nr
/ SWAPFILE_CLUSTER
;
575 VM_BUG_ON(cluster_count(&cluster_info
[idx
]) == 0);
576 cluster_set_count(&cluster_info
[idx
],
577 cluster_count(&cluster_info
[idx
]) - 1);
579 if (cluster_count(&cluster_info
[idx
]) == 0)
580 free_cluster(p
, idx
);
584 * It's possible scan_swap_map() uses a free cluster in the middle of free
585 * cluster list. Avoiding such abuse to avoid list corruption.
588 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct
*si
,
589 unsigned long offset
)
591 struct percpu_cluster
*percpu_cluster
;
594 offset
/= SWAPFILE_CLUSTER
;
595 conflict
= !cluster_list_empty(&si
->free_clusters
) &&
596 offset
!= cluster_list_first(&si
->free_clusters
) &&
597 cluster_is_free(&si
->cluster_info
[offset
]);
602 percpu_cluster
= this_cpu_ptr(si
->percpu_cluster
);
603 cluster_set_null(&percpu_cluster
->index
);
608 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
609 * might involve allocating a new cluster for current CPU too.
611 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct
*si
,
612 unsigned long *offset
, unsigned long *scan_base
)
614 struct percpu_cluster
*cluster
;
615 struct swap_cluster_info
*ci
;
616 unsigned long tmp
, max
;
619 cluster
= this_cpu_ptr(si
->percpu_cluster
);
620 if (cluster_is_null(&cluster
->index
)) {
621 if (!cluster_list_empty(&si
->free_clusters
)) {
622 cluster
->index
= si
->free_clusters
.head
;
623 cluster
->next
= cluster_next(&cluster
->index
) *
625 } else if (!cluster_list_empty(&si
->discard_clusters
)) {
627 * we don't have free cluster but have some clusters in
628 * discarding, do discard now and reclaim them, then
629 * reread cluster_next_cpu since we dropped si->lock
631 swap_do_scheduled_discard(si
);
632 *scan_base
= this_cpu_read(*si
->cluster_next_cpu
);
633 *offset
= *scan_base
;
640 * Other CPUs can use our cluster if they can't find a free cluster,
641 * check if there is still free entry in the cluster
644 max
= min_t(unsigned long, si
->max
,
645 (cluster_next(&cluster
->index
) + 1) * SWAPFILE_CLUSTER
);
647 ci
= lock_cluster(si
, tmp
);
649 if (!si
->swap_map
[tmp
])
656 cluster_set_null(&cluster
->index
);
659 cluster
->next
= tmp
+ 1;
665 static void __del_from_avail_list(struct swap_info_struct
*p
)
670 plist_del(&p
->avail_lists
[nid
], &swap_avail_heads
[nid
]);
673 static void del_from_avail_list(struct swap_info_struct
*p
)
675 spin_lock(&swap_avail_lock
);
676 __del_from_avail_list(p
);
677 spin_unlock(&swap_avail_lock
);
680 static void swap_range_alloc(struct swap_info_struct
*si
, unsigned long offset
,
681 unsigned int nr_entries
)
683 unsigned int end
= offset
+ nr_entries
- 1;
685 if (offset
== si
->lowest_bit
)
686 si
->lowest_bit
+= nr_entries
;
687 if (end
== si
->highest_bit
)
688 WRITE_ONCE(si
->highest_bit
, si
->highest_bit
- nr_entries
);
689 si
->inuse_pages
+= nr_entries
;
690 if (si
->inuse_pages
== si
->pages
) {
691 si
->lowest_bit
= si
->max
;
693 del_from_avail_list(si
);
697 static void add_to_avail_list(struct swap_info_struct
*p
)
701 spin_lock(&swap_avail_lock
);
703 WARN_ON(!plist_node_empty(&p
->avail_lists
[nid
]));
704 plist_add(&p
->avail_lists
[nid
], &swap_avail_heads
[nid
]);
706 spin_unlock(&swap_avail_lock
);
709 static void swap_range_free(struct swap_info_struct
*si
, unsigned long offset
,
710 unsigned int nr_entries
)
712 unsigned long begin
= offset
;
713 unsigned long end
= offset
+ nr_entries
- 1;
714 void (*swap_slot_free_notify
)(struct block_device
*, unsigned long);
716 if (offset
< si
->lowest_bit
)
717 si
->lowest_bit
= offset
;
718 if (end
> si
->highest_bit
) {
719 bool was_full
= !si
->highest_bit
;
721 WRITE_ONCE(si
->highest_bit
, end
);
722 if (was_full
&& (si
->flags
& SWP_WRITEOK
))
723 add_to_avail_list(si
);
725 atomic_long_add(nr_entries
, &nr_swap_pages
);
726 si
->inuse_pages
-= nr_entries
;
727 if (si
->flags
& SWP_BLKDEV
)
728 swap_slot_free_notify
=
729 si
->bdev
->bd_disk
->fops
->swap_slot_free_notify
;
731 swap_slot_free_notify
= NULL
;
732 while (offset
<= end
) {
733 arch_swap_invalidate_page(si
->type
, offset
);
734 frontswap_invalidate_page(si
->type
, offset
);
735 if (swap_slot_free_notify
)
736 swap_slot_free_notify(si
->bdev
, offset
);
739 clear_shadow_from_swap_cache(si
->type
, begin
, end
);
742 static void set_cluster_next(struct swap_info_struct
*si
, unsigned long next
)
746 if (!(si
->flags
& SWP_SOLIDSTATE
)) {
747 si
->cluster_next
= next
;
751 prev
= this_cpu_read(*si
->cluster_next_cpu
);
753 * Cross the swap address space size aligned trunk, choose
754 * another trunk randomly to avoid lock contention on swap
755 * address space if possible.
757 if ((prev
>> SWAP_ADDRESS_SPACE_SHIFT
) !=
758 (next
>> SWAP_ADDRESS_SPACE_SHIFT
)) {
759 /* No free swap slots available */
760 if (si
->highest_bit
<= si
->lowest_bit
)
762 next
= si
->lowest_bit
+
763 prandom_u32_max(si
->highest_bit
- si
->lowest_bit
+ 1);
764 next
= ALIGN_DOWN(next
, SWAP_ADDRESS_SPACE_PAGES
);
765 next
= max_t(unsigned int, next
, si
->lowest_bit
);
767 this_cpu_write(*si
->cluster_next_cpu
, next
);
770 static int scan_swap_map_slots(struct swap_info_struct
*si
,
771 unsigned char usage
, int nr
,
774 struct swap_cluster_info
*ci
;
775 unsigned long offset
;
776 unsigned long scan_base
;
777 unsigned long last_in_cluster
= 0;
778 int latency_ration
= LATENCY_LIMIT
;
780 bool scanned_many
= false;
783 * We try to cluster swap pages by allocating them sequentially
784 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
785 * way, however, we resort to first-free allocation, starting
786 * a new cluster. This prevents us from scattering swap pages
787 * all over the entire swap partition, so that we reduce
788 * overall disk seek times between swap pages. -- sct
789 * But we do now try to find an empty cluster. -Andrea
790 * And we let swap pages go all over an SSD partition. Hugh
793 si
->flags
+= SWP_SCANNING
;
795 * Use percpu scan base for SSD to reduce lock contention on
796 * cluster and swap cache. For HDD, sequential access is more
799 if (si
->flags
& SWP_SOLIDSTATE
)
800 scan_base
= this_cpu_read(*si
->cluster_next_cpu
);
802 scan_base
= si
->cluster_next
;
806 if (si
->cluster_info
) {
807 if (!scan_swap_map_try_ssd_cluster(si
, &offset
, &scan_base
))
809 } else if (unlikely(!si
->cluster_nr
--)) {
810 if (si
->pages
- si
->inuse_pages
< SWAPFILE_CLUSTER
) {
811 si
->cluster_nr
= SWAPFILE_CLUSTER
- 1;
815 spin_unlock(&si
->lock
);
818 * If seek is expensive, start searching for new cluster from
819 * start of partition, to minimize the span of allocated swap.
820 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
821 * case, just handled by scan_swap_map_try_ssd_cluster() above.
823 scan_base
= offset
= si
->lowest_bit
;
824 last_in_cluster
= offset
+ SWAPFILE_CLUSTER
- 1;
826 /* Locate the first empty (unaligned) cluster */
827 for (; last_in_cluster
<= si
->highest_bit
; offset
++) {
828 if (si
->swap_map
[offset
])
829 last_in_cluster
= offset
+ SWAPFILE_CLUSTER
;
830 else if (offset
== last_in_cluster
) {
831 spin_lock(&si
->lock
);
832 offset
-= SWAPFILE_CLUSTER
- 1;
833 si
->cluster_next
= offset
;
834 si
->cluster_nr
= SWAPFILE_CLUSTER
- 1;
837 if (unlikely(--latency_ration
< 0)) {
839 latency_ration
= LATENCY_LIMIT
;
844 spin_lock(&si
->lock
);
845 si
->cluster_nr
= SWAPFILE_CLUSTER
- 1;
849 if (si
->cluster_info
) {
850 while (scan_swap_map_ssd_cluster_conflict(si
, offset
)) {
851 /* take a break if we already got some slots */
854 if (!scan_swap_map_try_ssd_cluster(si
, &offset
,
859 if (!(si
->flags
& SWP_WRITEOK
))
861 if (!si
->highest_bit
)
863 if (offset
> si
->highest_bit
)
864 scan_base
= offset
= si
->lowest_bit
;
866 ci
= lock_cluster(si
, offset
);
867 /* reuse swap entry of cache-only swap if not busy. */
868 if (vm_swap_full() && si
->swap_map
[offset
] == SWAP_HAS_CACHE
) {
871 spin_unlock(&si
->lock
);
872 swap_was_freed
= __try_to_reclaim_swap(si
, offset
, TTRS_ANYWAY
);
873 spin_lock(&si
->lock
);
874 /* entry was freed successfully, try to use this again */
877 goto scan
; /* check next one */
880 if (si
->swap_map
[offset
]) {
887 WRITE_ONCE(si
->swap_map
[offset
], usage
);
888 inc_cluster_info_page(si
, si
->cluster_info
, offset
);
891 swap_range_alloc(si
, offset
, 1);
892 slots
[n_ret
++] = swp_entry(si
->type
, offset
);
894 /* got enough slots or reach max slots? */
895 if ((n_ret
== nr
) || (offset
>= si
->highest_bit
))
898 /* search for next available slot */
900 /* time to take a break? */
901 if (unlikely(--latency_ration
< 0)) {
904 spin_unlock(&si
->lock
);
906 spin_lock(&si
->lock
);
907 latency_ration
= LATENCY_LIMIT
;
910 /* try to get more slots in cluster */
911 if (si
->cluster_info
) {
912 if (scan_swap_map_try_ssd_cluster(si
, &offset
, &scan_base
))
914 } else if (si
->cluster_nr
&& !si
->swap_map
[++offset
]) {
915 /* non-ssd case, still more slots in cluster? */
921 * Even if there's no free clusters available (fragmented),
922 * try to scan a little more quickly with lock held unless we
923 * have scanned too many slots already.
926 unsigned long scan_limit
;
928 if (offset
< scan_base
)
929 scan_limit
= scan_base
;
931 scan_limit
= si
->highest_bit
;
932 for (; offset
<= scan_limit
&& --latency_ration
> 0;
934 if (!si
->swap_map
[offset
])
940 set_cluster_next(si
, offset
+ 1);
941 si
->flags
-= SWP_SCANNING
;
945 spin_unlock(&si
->lock
);
946 while (++offset
<= READ_ONCE(si
->highest_bit
)) {
947 if (data_race(!si
->swap_map
[offset
])) {
948 spin_lock(&si
->lock
);
951 if (vm_swap_full() &&
952 READ_ONCE(si
->swap_map
[offset
]) == SWAP_HAS_CACHE
) {
953 spin_lock(&si
->lock
);
956 if (unlikely(--latency_ration
< 0)) {
958 latency_ration
= LATENCY_LIMIT
;
962 offset
= si
->lowest_bit
;
963 while (offset
< scan_base
) {
964 if (data_race(!si
->swap_map
[offset
])) {
965 spin_lock(&si
->lock
);
968 if (vm_swap_full() &&
969 READ_ONCE(si
->swap_map
[offset
]) == SWAP_HAS_CACHE
) {
970 spin_lock(&si
->lock
);
973 if (unlikely(--latency_ration
< 0)) {
975 latency_ration
= LATENCY_LIMIT
;
980 spin_lock(&si
->lock
);
983 si
->flags
-= SWP_SCANNING
;
987 static int swap_alloc_cluster(struct swap_info_struct
*si
, swp_entry_t
*slot
)
990 struct swap_cluster_info
*ci
;
991 unsigned long offset
;
994 * Should not even be attempting cluster allocations when huge
995 * page swap is disabled. Warn and fail the allocation.
997 if (!IS_ENABLED(CONFIG_THP_SWAP
)) {
1002 if (cluster_list_empty(&si
->free_clusters
))
1005 idx
= cluster_list_first(&si
->free_clusters
);
1006 offset
= idx
* SWAPFILE_CLUSTER
;
1007 ci
= lock_cluster(si
, offset
);
1008 alloc_cluster(si
, idx
);
1009 cluster_set_count_flag(ci
, SWAPFILE_CLUSTER
, CLUSTER_FLAG_HUGE
);
1011 memset(si
->swap_map
+ offset
, SWAP_HAS_CACHE
, SWAPFILE_CLUSTER
);
1013 swap_range_alloc(si
, offset
, SWAPFILE_CLUSTER
);
1014 *slot
= swp_entry(si
->type
, offset
);
1019 static void swap_free_cluster(struct swap_info_struct
*si
, unsigned long idx
)
1021 unsigned long offset
= idx
* SWAPFILE_CLUSTER
;
1022 struct swap_cluster_info
*ci
;
1024 ci
= lock_cluster(si
, offset
);
1025 memset(si
->swap_map
+ offset
, 0, SWAPFILE_CLUSTER
);
1026 cluster_set_count_flag(ci
, 0, 0);
1027 free_cluster(si
, idx
);
1029 swap_range_free(si
, offset
, SWAPFILE_CLUSTER
);
1032 static unsigned long scan_swap_map(struct swap_info_struct
*si
,
1033 unsigned char usage
)
1038 n_ret
= scan_swap_map_slots(si
, usage
, 1, &entry
);
1041 return swp_offset(entry
);
1047 int get_swap_pages(int n_goal
, swp_entry_t swp_entries
[], int entry_size
)
1049 unsigned long size
= swap_entry_size(entry_size
);
1050 struct swap_info_struct
*si
, *next
;
1055 /* Only single cluster request supported */
1056 WARN_ON_ONCE(n_goal
> 1 && size
== SWAPFILE_CLUSTER
);
1058 spin_lock(&swap_avail_lock
);
1060 avail_pgs
= atomic_long_read(&nr_swap_pages
) / size
;
1061 if (avail_pgs
<= 0) {
1062 spin_unlock(&swap_avail_lock
);
1066 n_goal
= min3((long)n_goal
, (long)SWAP_BATCH
, avail_pgs
);
1068 atomic_long_sub(n_goal
* size
, &nr_swap_pages
);
1071 node
= numa_node_id();
1072 plist_for_each_entry_safe(si
, next
, &swap_avail_heads
[node
], avail_lists
[node
]) {
1073 /* requeue si to after same-priority siblings */
1074 plist_requeue(&si
->avail_lists
[node
], &swap_avail_heads
[node
]);
1075 spin_unlock(&swap_avail_lock
);
1076 spin_lock(&si
->lock
);
1077 if (!si
->highest_bit
|| !(si
->flags
& SWP_WRITEOK
)) {
1078 spin_lock(&swap_avail_lock
);
1079 if (plist_node_empty(&si
->avail_lists
[node
])) {
1080 spin_unlock(&si
->lock
);
1083 WARN(!si
->highest_bit
,
1084 "swap_info %d in list but !highest_bit\n",
1086 WARN(!(si
->flags
& SWP_WRITEOK
),
1087 "swap_info %d in list but !SWP_WRITEOK\n",
1089 __del_from_avail_list(si
);
1090 spin_unlock(&si
->lock
);
1093 if (size
== SWAPFILE_CLUSTER
) {
1094 if (si
->flags
& SWP_BLKDEV
)
1095 n_ret
= swap_alloc_cluster(si
, swp_entries
);
1097 n_ret
= scan_swap_map_slots(si
, SWAP_HAS_CACHE
,
1098 n_goal
, swp_entries
);
1099 spin_unlock(&si
->lock
);
1100 if (n_ret
|| size
== SWAPFILE_CLUSTER
)
1102 pr_debug("scan_swap_map of si %d failed to find offset\n",
1105 spin_lock(&swap_avail_lock
);
1108 * if we got here, it's likely that si was almost full before,
1109 * and since scan_swap_map() can drop the si->lock, multiple
1110 * callers probably all tried to get a page from the same si
1111 * and it filled up before we could get one; or, the si filled
1112 * up between us dropping swap_avail_lock and taking si->lock.
1113 * Since we dropped the swap_avail_lock, the swap_avail_head
1114 * list may have been modified; so if next is still in the
1115 * swap_avail_head list then try it, otherwise start over
1116 * if we have not gotten any slots.
1118 if (plist_node_empty(&next
->avail_lists
[node
]))
1122 spin_unlock(&swap_avail_lock
);
1126 atomic_long_add((long)(n_goal
- n_ret
) * size
,
1132 /* The only caller of this function is now suspend routine */
1133 swp_entry_t
get_swap_page_of_type(int type
)
1135 struct swap_info_struct
*si
= swap_type_to_swap_info(type
);
1141 spin_lock(&si
->lock
);
1142 if (si
->flags
& SWP_WRITEOK
) {
1143 /* This is called for allocating swap entry, not cache */
1144 offset
= scan_swap_map(si
, 1);
1146 atomic_long_dec(&nr_swap_pages
);
1147 spin_unlock(&si
->lock
);
1148 return swp_entry(type
, offset
);
1151 spin_unlock(&si
->lock
);
1153 return (swp_entry_t
) {0};
1156 static struct swap_info_struct
*__swap_info_get(swp_entry_t entry
)
1158 struct swap_info_struct
*p
;
1159 unsigned long offset
;
1163 p
= swp_swap_info(entry
);
1166 if (data_race(!(p
->flags
& SWP_USED
)))
1168 offset
= swp_offset(entry
);
1169 if (offset
>= p
->max
)
1174 pr_err("swap_info_get: %s%08lx\n", Bad_offset
, entry
.val
);
1177 pr_err("swap_info_get: %s%08lx\n", Unused_file
, entry
.val
);
1180 pr_err("swap_info_get: %s%08lx\n", Bad_file
, entry
.val
);
1185 static struct swap_info_struct
*_swap_info_get(swp_entry_t entry
)
1187 struct swap_info_struct
*p
;
1189 p
= __swap_info_get(entry
);
1192 if (data_race(!p
->swap_map
[swp_offset(entry
)]))
1197 pr_err("swap_info_get: %s%08lx\n", Unused_offset
, entry
.val
);
1202 static struct swap_info_struct
*swap_info_get(swp_entry_t entry
)
1204 struct swap_info_struct
*p
;
1206 p
= _swap_info_get(entry
);
1208 spin_lock(&p
->lock
);
1212 static struct swap_info_struct
*swap_info_get_cont(swp_entry_t entry
,
1213 struct swap_info_struct
*q
)
1215 struct swap_info_struct
*p
;
1217 p
= _swap_info_get(entry
);
1221 spin_unlock(&q
->lock
);
1223 spin_lock(&p
->lock
);
1228 static unsigned char __swap_entry_free_locked(struct swap_info_struct
*p
,
1229 unsigned long offset
,
1230 unsigned char usage
)
1232 unsigned char count
;
1233 unsigned char has_cache
;
1235 count
= p
->swap_map
[offset
];
1237 has_cache
= count
& SWAP_HAS_CACHE
;
1238 count
&= ~SWAP_HAS_CACHE
;
1240 if (usage
== SWAP_HAS_CACHE
) {
1241 VM_BUG_ON(!has_cache
);
1243 } else if (count
== SWAP_MAP_SHMEM
) {
1245 * Or we could insist on shmem.c using a special
1246 * swap_shmem_free() and free_shmem_swap_and_cache()...
1249 } else if ((count
& ~COUNT_CONTINUED
) <= SWAP_MAP_MAX
) {
1250 if (count
== COUNT_CONTINUED
) {
1251 if (swap_count_continued(p
, offset
, count
))
1252 count
= SWAP_MAP_MAX
| COUNT_CONTINUED
;
1254 count
= SWAP_MAP_MAX
;
1259 usage
= count
| has_cache
;
1261 WRITE_ONCE(p
->swap_map
[offset
], usage
);
1263 WRITE_ONCE(p
->swap_map
[offset
], SWAP_HAS_CACHE
);
1269 * Check whether swap entry is valid in the swap device. If so,
1270 * return pointer to swap_info_struct, and keep the swap entry valid
1271 * via preventing the swap device from being swapoff, until
1272 * put_swap_device() is called. Otherwise return NULL.
1274 * The entirety of the RCU read critical section must come before the
1275 * return from or after the call to synchronize_rcu() in
1276 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1277 * true, the si->map, si->cluster_info, etc. must be valid in the
1280 * Notice that swapoff or swapoff+swapon can still happen before the
1281 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1282 * in put_swap_device() if there isn't any other way to prevent
1283 * swapoff, such as page lock, page table lock, etc. The caller must
1284 * be prepared for that. For example, the following situation is
1289 * ... swapoff+swapon
1290 * __read_swap_cache_async()
1291 * swapcache_prepare()
1292 * __swap_duplicate()
1294 * // verify PTE not changed
1296 * In __swap_duplicate(), the swap_map need to be checked before
1297 * changing partly because the specified swap entry may be for another
1298 * swap device which has been swapoff. And in do_swap_page(), after
1299 * the page is read from the swap device, the PTE is verified not
1300 * changed with the page table locked to check whether the swap device
1301 * has been swapoff or swapoff+swapon.
1303 struct swap_info_struct
*get_swap_device(swp_entry_t entry
)
1305 struct swap_info_struct
*si
;
1306 unsigned long offset
;
1310 si
= swp_swap_info(entry
);
1315 if (data_race(!(si
->flags
& SWP_VALID
)))
1317 offset
= swp_offset(entry
);
1318 if (offset
>= si
->max
)
1323 pr_err("%s: %s%08lx\n", __func__
, Bad_file
, entry
.val
);
1331 static unsigned char __swap_entry_free(struct swap_info_struct
*p
,
1334 struct swap_cluster_info
*ci
;
1335 unsigned long offset
= swp_offset(entry
);
1336 unsigned char usage
;
1338 ci
= lock_cluster_or_swap_info(p
, offset
);
1339 usage
= __swap_entry_free_locked(p
, offset
, 1);
1340 unlock_cluster_or_swap_info(p
, ci
);
1342 free_swap_slot(entry
);
1347 static void swap_entry_free(struct swap_info_struct
*p
, swp_entry_t entry
)
1349 struct swap_cluster_info
*ci
;
1350 unsigned long offset
= swp_offset(entry
);
1351 unsigned char count
;
1353 ci
= lock_cluster(p
, offset
);
1354 count
= p
->swap_map
[offset
];
1355 VM_BUG_ON(count
!= SWAP_HAS_CACHE
);
1356 p
->swap_map
[offset
] = 0;
1357 dec_cluster_info_page(p
, p
->cluster_info
, offset
);
1360 mem_cgroup_uncharge_swap(entry
, 1);
1361 swap_range_free(p
, offset
, 1);
1365 * Caller has made sure that the swap device corresponding to entry
1366 * is still around or has not been recycled.
1368 void swap_free(swp_entry_t entry
)
1370 struct swap_info_struct
*p
;
1372 p
= _swap_info_get(entry
);
1374 __swap_entry_free(p
, entry
);
1378 * Called after dropping swapcache to decrease refcnt to swap entries.
1380 void put_swap_page(struct page
*page
, swp_entry_t entry
)
1382 unsigned long offset
= swp_offset(entry
);
1383 unsigned long idx
= offset
/ SWAPFILE_CLUSTER
;
1384 struct swap_cluster_info
*ci
;
1385 struct swap_info_struct
*si
;
1387 unsigned int i
, free_entries
= 0;
1389 int size
= swap_entry_size(thp_nr_pages(page
));
1391 si
= _swap_info_get(entry
);
1395 ci
= lock_cluster_or_swap_info(si
, offset
);
1396 if (size
== SWAPFILE_CLUSTER
) {
1397 VM_BUG_ON(!cluster_is_huge(ci
));
1398 map
= si
->swap_map
+ offset
;
1399 for (i
= 0; i
< SWAPFILE_CLUSTER
; i
++) {
1401 VM_BUG_ON(!(val
& SWAP_HAS_CACHE
));
1402 if (val
== SWAP_HAS_CACHE
)
1405 cluster_clear_huge(ci
);
1406 if (free_entries
== SWAPFILE_CLUSTER
) {
1407 unlock_cluster_or_swap_info(si
, ci
);
1408 spin_lock(&si
->lock
);
1409 mem_cgroup_uncharge_swap(entry
, SWAPFILE_CLUSTER
);
1410 swap_free_cluster(si
, idx
);
1411 spin_unlock(&si
->lock
);
1415 for (i
= 0; i
< size
; i
++, entry
.val
++) {
1416 if (!__swap_entry_free_locked(si
, offset
+ i
, SWAP_HAS_CACHE
)) {
1417 unlock_cluster_or_swap_info(si
, ci
);
1418 free_swap_slot(entry
);
1421 lock_cluster_or_swap_info(si
, offset
);
1424 unlock_cluster_or_swap_info(si
, ci
);
1427 #ifdef CONFIG_THP_SWAP
1428 int split_swap_cluster(swp_entry_t entry
)
1430 struct swap_info_struct
*si
;
1431 struct swap_cluster_info
*ci
;
1432 unsigned long offset
= swp_offset(entry
);
1434 si
= _swap_info_get(entry
);
1437 ci
= lock_cluster(si
, offset
);
1438 cluster_clear_huge(ci
);
1444 static int swp_entry_cmp(const void *ent1
, const void *ent2
)
1446 const swp_entry_t
*e1
= ent1
, *e2
= ent2
;
1448 return (int)swp_type(*e1
) - (int)swp_type(*e2
);
1451 void swapcache_free_entries(swp_entry_t
*entries
, int n
)
1453 struct swap_info_struct
*p
, *prev
;
1463 * Sort swap entries by swap device, so each lock is only taken once.
1464 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1465 * so low that it isn't necessary to optimize further.
1467 if (nr_swapfiles
> 1)
1468 sort(entries
, n
, sizeof(entries
[0]), swp_entry_cmp
, NULL
);
1469 for (i
= 0; i
< n
; ++i
) {
1470 p
= swap_info_get_cont(entries
[i
], prev
);
1472 swap_entry_free(p
, entries
[i
]);
1476 spin_unlock(&p
->lock
);
1480 * How many references to page are currently swapped out?
1481 * This does not give an exact answer when swap count is continued,
1482 * but does include the high COUNT_CONTINUED flag to allow for that.
1484 int page_swapcount(struct page
*page
)
1487 struct swap_info_struct
*p
;
1488 struct swap_cluster_info
*ci
;
1490 unsigned long offset
;
1492 entry
.val
= page_private(page
);
1493 p
= _swap_info_get(entry
);
1495 offset
= swp_offset(entry
);
1496 ci
= lock_cluster_or_swap_info(p
, offset
);
1497 count
= swap_count(p
->swap_map
[offset
]);
1498 unlock_cluster_or_swap_info(p
, ci
);
1503 int __swap_count(swp_entry_t entry
)
1505 struct swap_info_struct
*si
;
1506 pgoff_t offset
= swp_offset(entry
);
1509 si
= get_swap_device(entry
);
1511 count
= swap_count(si
->swap_map
[offset
]);
1512 put_swap_device(si
);
1517 static int swap_swapcount(struct swap_info_struct
*si
, swp_entry_t entry
)
1520 pgoff_t offset
= swp_offset(entry
);
1521 struct swap_cluster_info
*ci
;
1523 ci
= lock_cluster_or_swap_info(si
, offset
);
1524 count
= swap_count(si
->swap_map
[offset
]);
1525 unlock_cluster_or_swap_info(si
, ci
);
1530 * How many references to @entry are currently swapped out?
1531 * This does not give an exact answer when swap count is continued,
1532 * but does include the high COUNT_CONTINUED flag to allow for that.
1534 int __swp_swapcount(swp_entry_t entry
)
1537 struct swap_info_struct
*si
;
1539 si
= get_swap_device(entry
);
1541 count
= swap_swapcount(si
, entry
);
1542 put_swap_device(si
);
1548 * How many references to @entry are currently swapped out?
1549 * This considers COUNT_CONTINUED so it returns exact answer.
1551 int swp_swapcount(swp_entry_t entry
)
1553 int count
, tmp_count
, n
;
1554 struct swap_info_struct
*p
;
1555 struct swap_cluster_info
*ci
;
1560 p
= _swap_info_get(entry
);
1564 offset
= swp_offset(entry
);
1566 ci
= lock_cluster_or_swap_info(p
, offset
);
1568 count
= swap_count(p
->swap_map
[offset
]);
1569 if (!(count
& COUNT_CONTINUED
))
1572 count
&= ~COUNT_CONTINUED
;
1573 n
= SWAP_MAP_MAX
+ 1;
1575 page
= vmalloc_to_page(p
->swap_map
+ offset
);
1576 offset
&= ~PAGE_MASK
;
1577 VM_BUG_ON(page_private(page
) != SWP_CONTINUED
);
1580 page
= list_next_entry(page
, lru
);
1581 map
= kmap_atomic(page
);
1582 tmp_count
= map
[offset
];
1585 count
+= (tmp_count
& ~COUNT_CONTINUED
) * n
;
1586 n
*= (SWAP_CONT_MAX
+ 1);
1587 } while (tmp_count
& COUNT_CONTINUED
);
1589 unlock_cluster_or_swap_info(p
, ci
);
1593 static bool swap_page_trans_huge_swapped(struct swap_info_struct
*si
,
1596 struct swap_cluster_info
*ci
;
1597 unsigned char *map
= si
->swap_map
;
1598 unsigned long roffset
= swp_offset(entry
);
1599 unsigned long offset
= round_down(roffset
, SWAPFILE_CLUSTER
);
1603 ci
= lock_cluster_or_swap_info(si
, offset
);
1604 if (!ci
|| !cluster_is_huge(ci
)) {
1605 if (swap_count(map
[roffset
]))
1609 for (i
= 0; i
< SWAPFILE_CLUSTER
; i
++) {
1610 if (swap_count(map
[offset
+ i
])) {
1616 unlock_cluster_or_swap_info(si
, ci
);
1620 static bool page_swapped(struct page
*page
)
1623 struct swap_info_struct
*si
;
1625 if (!IS_ENABLED(CONFIG_THP_SWAP
) || likely(!PageTransCompound(page
)))
1626 return page_swapcount(page
) != 0;
1628 page
= compound_head(page
);
1629 entry
.val
= page_private(page
);
1630 si
= _swap_info_get(entry
);
1632 return swap_page_trans_huge_swapped(si
, entry
);
1636 static int page_trans_huge_map_swapcount(struct page
*page
, int *total_mapcount
,
1637 int *total_swapcount
)
1639 int i
, map_swapcount
, _total_mapcount
, _total_swapcount
;
1640 unsigned long offset
= 0;
1641 struct swap_info_struct
*si
;
1642 struct swap_cluster_info
*ci
= NULL
;
1643 unsigned char *map
= NULL
;
1644 int mapcount
, swapcount
= 0;
1646 /* hugetlbfs shouldn't call it */
1647 VM_BUG_ON_PAGE(PageHuge(page
), page
);
1649 if (!IS_ENABLED(CONFIG_THP_SWAP
) || likely(!PageTransCompound(page
))) {
1650 mapcount
= page_trans_huge_mapcount(page
, total_mapcount
);
1651 if (PageSwapCache(page
))
1652 swapcount
= page_swapcount(page
);
1653 if (total_swapcount
)
1654 *total_swapcount
= swapcount
;
1655 return mapcount
+ swapcount
;
1658 page
= compound_head(page
);
1660 _total_mapcount
= _total_swapcount
= map_swapcount
= 0;
1661 if (PageSwapCache(page
)) {
1664 entry
.val
= page_private(page
);
1665 si
= _swap_info_get(entry
);
1668 offset
= swp_offset(entry
);
1672 ci
= lock_cluster(si
, offset
);
1673 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1674 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
1675 _total_mapcount
+= mapcount
;
1677 swapcount
= swap_count(map
[offset
+ i
]);
1678 _total_swapcount
+= swapcount
;
1680 map_swapcount
= max(map_swapcount
, mapcount
+ swapcount
);
1683 if (PageDoubleMap(page
)) {
1685 _total_mapcount
-= HPAGE_PMD_NR
;
1687 mapcount
= compound_mapcount(page
);
1688 map_swapcount
+= mapcount
;
1689 _total_mapcount
+= mapcount
;
1691 *total_mapcount
= _total_mapcount
;
1692 if (total_swapcount
)
1693 *total_swapcount
= _total_swapcount
;
1695 return map_swapcount
;
1699 * We can write to an anon page without COW if there are no other references
1700 * to it. And as a side-effect, free up its swap: because the old content
1701 * on disk will never be read, and seeking back there to write new content
1702 * later would only waste time away from clustering.
1704 * NOTE: total_map_swapcount should not be relied upon by the caller if
1705 * reuse_swap_page() returns false, but it may be always overwritten
1706 * (see the other implementation for CONFIG_SWAP=n).
1708 bool reuse_swap_page(struct page
*page
, int *total_map_swapcount
)
1710 int count
, total_mapcount
, total_swapcount
;
1712 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1713 if (unlikely(PageKsm(page
)))
1715 count
= page_trans_huge_map_swapcount(page
, &total_mapcount
,
1717 if (total_map_swapcount
)
1718 *total_map_swapcount
= total_mapcount
+ total_swapcount
;
1719 if (count
== 1 && PageSwapCache(page
) &&
1720 (likely(!PageTransCompound(page
)) ||
1721 /* The remaining swap count will be freed soon */
1722 total_swapcount
== page_swapcount(page
))) {
1723 if (!PageWriteback(page
)) {
1724 page
= compound_head(page
);
1725 delete_from_swap_cache(page
);
1729 struct swap_info_struct
*p
;
1731 entry
.val
= page_private(page
);
1732 p
= swap_info_get(entry
);
1733 if (p
->flags
& SWP_STABLE_WRITES
) {
1734 spin_unlock(&p
->lock
);
1737 spin_unlock(&p
->lock
);
1745 * If swap is getting full, or if there are no more mappings of this page,
1746 * then try_to_free_swap is called to free its swap space.
1748 int try_to_free_swap(struct page
*page
)
1750 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1752 if (!PageSwapCache(page
))
1754 if (PageWriteback(page
))
1756 if (page_swapped(page
))
1760 * Once hibernation has begun to create its image of memory,
1761 * there's a danger that one of the calls to try_to_free_swap()
1762 * - most probably a call from __try_to_reclaim_swap() while
1763 * hibernation is allocating its own swap pages for the image,
1764 * but conceivably even a call from memory reclaim - will free
1765 * the swap from a page which has already been recorded in the
1766 * image as a clean swapcache page, and then reuse its swap for
1767 * another page of the image. On waking from hibernation, the
1768 * original page might be freed under memory pressure, then
1769 * later read back in from swap, now with the wrong data.
1771 * Hibernation suspends storage while it is writing the image
1772 * to disk so check that here.
1774 if (pm_suspended_storage())
1777 page
= compound_head(page
);
1778 delete_from_swap_cache(page
);
1784 * Free the swap entry like above, but also try to
1785 * free the page cache entry if it is the last user.
1787 int free_swap_and_cache(swp_entry_t entry
)
1789 struct swap_info_struct
*p
;
1790 unsigned char count
;
1792 if (non_swap_entry(entry
))
1795 p
= _swap_info_get(entry
);
1797 count
= __swap_entry_free(p
, entry
);
1798 if (count
== SWAP_HAS_CACHE
&&
1799 !swap_page_trans_huge_swapped(p
, entry
))
1800 __try_to_reclaim_swap(p
, swp_offset(entry
),
1801 TTRS_UNMAPPED
| TTRS_FULL
);
1806 #ifdef CONFIG_HIBERNATION
1808 * Find the swap type that corresponds to given device (if any).
1810 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1811 * from 0, in which the swap header is expected to be located.
1813 * This is needed for the suspend to disk (aka swsusp).
1815 int swap_type_of(dev_t device
, sector_t offset
)
1822 spin_lock(&swap_lock
);
1823 for (type
= 0; type
< nr_swapfiles
; type
++) {
1824 struct swap_info_struct
*sis
= swap_info
[type
];
1826 if (!(sis
->flags
& SWP_WRITEOK
))
1829 if (device
== sis
->bdev
->bd_dev
) {
1830 struct swap_extent
*se
= first_se(sis
);
1832 if (se
->start_block
== offset
) {
1833 spin_unlock(&swap_lock
);
1838 spin_unlock(&swap_lock
);
1842 int find_first_swap(dev_t
*device
)
1846 spin_lock(&swap_lock
);
1847 for (type
= 0; type
< nr_swapfiles
; type
++) {
1848 struct swap_info_struct
*sis
= swap_info
[type
];
1850 if (!(sis
->flags
& SWP_WRITEOK
))
1852 *device
= sis
->bdev
->bd_dev
;
1853 spin_unlock(&swap_lock
);
1856 spin_unlock(&swap_lock
);
1861 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1862 * corresponding to given index in swap_info (swap type).
1864 sector_t
swapdev_block(int type
, pgoff_t offset
)
1866 struct block_device
*bdev
;
1867 struct swap_info_struct
*si
= swap_type_to_swap_info(type
);
1869 if (!si
|| !(si
->flags
& SWP_WRITEOK
))
1871 return map_swap_entry(swp_entry(type
, offset
), &bdev
);
1875 * Return either the total number of swap pages of given type, or the number
1876 * of free pages of that type (depending on @free)
1878 * This is needed for software suspend
1880 unsigned int count_swap_pages(int type
, int free
)
1884 spin_lock(&swap_lock
);
1885 if ((unsigned int)type
< nr_swapfiles
) {
1886 struct swap_info_struct
*sis
= swap_info
[type
];
1888 spin_lock(&sis
->lock
);
1889 if (sis
->flags
& SWP_WRITEOK
) {
1892 n
-= sis
->inuse_pages
;
1894 spin_unlock(&sis
->lock
);
1896 spin_unlock(&swap_lock
);
1899 #endif /* CONFIG_HIBERNATION */
1901 static inline int pte_same_as_swp(pte_t pte
, pte_t swp_pte
)
1903 return pte_same(pte_swp_clear_soft_dirty(pte
), swp_pte
);
1907 * No need to decide whether this PTE shares the swap entry with others,
1908 * just let do_wp_page work it out if a write is requested later - to
1909 * force COW, vm_page_prot omits write permission from any private vma.
1911 static int unuse_pte(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1912 unsigned long addr
, swp_entry_t entry
, struct page
*page
)
1914 struct page
*swapcache
;
1920 page
= ksm_might_need_to_copy(page
, vma
, addr
);
1921 if (unlikely(!page
))
1924 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
1925 if (unlikely(!pte_same_as_swp(*pte
, swp_entry_to_pte(entry
)))) {
1930 dec_mm_counter(vma
->vm_mm
, MM_SWAPENTS
);
1931 inc_mm_counter(vma
->vm_mm
, MM_ANONPAGES
);
1933 set_pte_at(vma
->vm_mm
, addr
, pte
,
1934 pte_mkold(mk_pte(page
, vma
->vm_page_prot
)));
1935 if (page
== swapcache
) {
1936 page_add_anon_rmap(page
, vma
, addr
, false);
1937 } else { /* ksm created a completely new copy */
1938 page_add_new_anon_rmap(page
, vma
, addr
, false);
1939 lru_cache_add_inactive_or_unevictable(page
, vma
);
1943 pte_unmap_unlock(pte
, ptl
);
1944 if (page
!= swapcache
) {
1951 static int unuse_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1952 unsigned long addr
, unsigned long end
,
1953 unsigned int type
, bool frontswap
,
1954 unsigned long *fs_pages_to_unuse
)
1959 struct swap_info_struct
*si
;
1960 unsigned long offset
;
1962 volatile unsigned char *swap_map
;
1964 si
= swap_info
[type
];
1965 pte
= pte_offset_map(pmd
, addr
);
1967 struct vm_fault vmf
;
1969 if (!is_swap_pte(*pte
))
1972 entry
= pte_to_swp_entry(*pte
);
1973 if (swp_type(entry
) != type
)
1976 offset
= swp_offset(entry
);
1977 if (frontswap
&& !frontswap_test(si
, offset
))
1981 swap_map
= &si
->swap_map
[offset
];
1982 page
= lookup_swap_cache(entry
, vma
, addr
);
1987 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
1991 if (*swap_map
== 0 || *swap_map
== SWAP_MAP_BAD
)
1997 wait_on_page_writeback(page
);
1998 ret
= unuse_pte(vma
, pmd
, addr
, entry
, page
);
2005 try_to_free_swap(page
);
2009 if (*fs_pages_to_unuse
&& !--(*fs_pages_to_unuse
)) {
2010 ret
= FRONTSWAP_PAGES_UNUSED
;
2014 pte
= pte_offset_map(pmd
, addr
);
2015 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2023 static inline int unuse_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
2024 unsigned long addr
, unsigned long end
,
2025 unsigned int type
, bool frontswap
,
2026 unsigned long *fs_pages_to_unuse
)
2032 pmd
= pmd_offset(pud
, addr
);
2035 next
= pmd_addr_end(addr
, end
);
2036 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
2038 ret
= unuse_pte_range(vma
, pmd
, addr
, next
, type
,
2039 frontswap
, fs_pages_to_unuse
);
2042 } while (pmd
++, addr
= next
, addr
!= end
);
2046 static inline int unuse_pud_range(struct vm_area_struct
*vma
, p4d_t
*p4d
,
2047 unsigned long addr
, unsigned long end
,
2048 unsigned int type
, bool frontswap
,
2049 unsigned long *fs_pages_to_unuse
)
2055 pud
= pud_offset(p4d
, addr
);
2057 next
= pud_addr_end(addr
, end
);
2058 if (pud_none_or_clear_bad(pud
))
2060 ret
= unuse_pmd_range(vma
, pud
, addr
, next
, type
,
2061 frontswap
, fs_pages_to_unuse
);
2064 } while (pud
++, addr
= next
, addr
!= end
);
2068 static inline int unuse_p4d_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
2069 unsigned long addr
, unsigned long end
,
2070 unsigned int type
, bool frontswap
,
2071 unsigned long *fs_pages_to_unuse
)
2077 p4d
= p4d_offset(pgd
, addr
);
2079 next
= p4d_addr_end(addr
, end
);
2080 if (p4d_none_or_clear_bad(p4d
))
2082 ret
= unuse_pud_range(vma
, p4d
, addr
, next
, type
,
2083 frontswap
, fs_pages_to_unuse
);
2086 } while (p4d
++, addr
= next
, addr
!= end
);
2090 static int unuse_vma(struct vm_area_struct
*vma
, unsigned int type
,
2091 bool frontswap
, unsigned long *fs_pages_to_unuse
)
2094 unsigned long addr
, end
, next
;
2097 addr
= vma
->vm_start
;
2100 pgd
= pgd_offset(vma
->vm_mm
, addr
);
2102 next
= pgd_addr_end(addr
, end
);
2103 if (pgd_none_or_clear_bad(pgd
))
2105 ret
= unuse_p4d_range(vma
, pgd
, addr
, next
, type
,
2106 frontswap
, fs_pages_to_unuse
);
2109 } while (pgd
++, addr
= next
, addr
!= end
);
2113 static int unuse_mm(struct mm_struct
*mm
, unsigned int type
,
2114 bool frontswap
, unsigned long *fs_pages_to_unuse
)
2116 struct vm_area_struct
*vma
;
2120 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
2121 if (vma
->anon_vma
) {
2122 ret
= unuse_vma(vma
, type
, frontswap
,
2129 mmap_read_unlock(mm
);
2134 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2135 * from current position to next entry still in use. Return 0
2136 * if there are no inuse entries after prev till end of the map.
2138 static unsigned int find_next_to_unuse(struct swap_info_struct
*si
,
2139 unsigned int prev
, bool frontswap
)
2142 unsigned char count
;
2145 * No need for swap_lock here: we're just looking
2146 * for whether an entry is in use, not modifying it; false
2147 * hits are okay, and sys_swapoff() has already prevented new
2148 * allocations from this area (while holding swap_lock).
2150 for (i
= prev
+ 1; i
< si
->max
; i
++) {
2151 count
= READ_ONCE(si
->swap_map
[i
]);
2152 if (count
&& swap_count(count
) != SWAP_MAP_BAD
)
2153 if (!frontswap
|| frontswap_test(si
, i
))
2155 if ((i
% LATENCY_LIMIT
) == 0)
2166 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2167 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2169 int try_to_unuse(unsigned int type
, bool frontswap
,
2170 unsigned long pages_to_unuse
)
2172 struct mm_struct
*prev_mm
;
2173 struct mm_struct
*mm
;
2174 struct list_head
*p
;
2176 struct swap_info_struct
*si
= swap_info
[type
];
2181 if (!READ_ONCE(si
->inuse_pages
))
2188 retval
= shmem_unuse(type
, frontswap
, &pages_to_unuse
);
2195 spin_lock(&mmlist_lock
);
2196 p
= &init_mm
.mmlist
;
2197 while (READ_ONCE(si
->inuse_pages
) &&
2198 !signal_pending(current
) &&
2199 (p
= p
->next
) != &init_mm
.mmlist
) {
2201 mm
= list_entry(p
, struct mm_struct
, mmlist
);
2202 if (!mmget_not_zero(mm
))
2204 spin_unlock(&mmlist_lock
);
2207 retval
= unuse_mm(mm
, type
, frontswap
, &pages_to_unuse
);
2215 * Make sure that we aren't completely killing
2216 * interactive performance.
2219 spin_lock(&mmlist_lock
);
2221 spin_unlock(&mmlist_lock
);
2226 while (READ_ONCE(si
->inuse_pages
) &&
2227 !signal_pending(current
) &&
2228 (i
= find_next_to_unuse(si
, i
, frontswap
)) != 0) {
2230 entry
= swp_entry(type
, i
);
2231 page
= find_get_page(swap_address_space(entry
), i
);
2236 * It is conceivable that a racing task removed this page from
2237 * swap cache just before we acquired the page lock. The page
2238 * might even be back in swap cache on another swap area. But
2239 * that is okay, try_to_free_swap() only removes stale pages.
2242 wait_on_page_writeback(page
);
2243 try_to_free_swap(page
);
2248 * For frontswap, we just need to unuse pages_to_unuse, if
2249 * it was specified. Need not check frontswap again here as
2250 * we already zeroed out pages_to_unuse if not frontswap.
2252 if (pages_to_unuse
&& --pages_to_unuse
== 0)
2257 * Lets check again to see if there are still swap entries in the map.
2258 * If yes, we would need to do retry the unuse logic again.
2259 * Under global memory pressure, swap entries can be reinserted back
2260 * into process space after the mmlist loop above passes over them.
2262 * Limit the number of retries? No: when mmget_not_zero() above fails,
2263 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2264 * at its own independent pace; and even shmem_writepage() could have
2265 * been preempted after get_swap_page(), temporarily hiding that swap.
2266 * It's easy and robust (though cpu-intensive) just to keep retrying.
2268 if (READ_ONCE(si
->inuse_pages
)) {
2269 if (!signal_pending(current
))
2274 return (retval
== FRONTSWAP_PAGES_UNUSED
) ? 0 : retval
;
2278 * After a successful try_to_unuse, if no swap is now in use, we know
2279 * we can empty the mmlist. swap_lock must be held on entry and exit.
2280 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2281 * added to the mmlist just after page_duplicate - before would be racy.
2283 static void drain_mmlist(void)
2285 struct list_head
*p
, *next
;
2288 for (type
= 0; type
< nr_swapfiles
; type
++)
2289 if (swap_info
[type
]->inuse_pages
)
2291 spin_lock(&mmlist_lock
);
2292 list_for_each_safe(p
, next
, &init_mm
.mmlist
)
2294 spin_unlock(&mmlist_lock
);
2298 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2299 * corresponds to page offset for the specified swap entry.
2300 * Note that the type of this function is sector_t, but it returns page offset
2301 * into the bdev, not sector offset.
2303 static sector_t
map_swap_entry(swp_entry_t entry
, struct block_device
**bdev
)
2305 struct swap_info_struct
*sis
;
2306 struct swap_extent
*se
;
2309 sis
= swp_swap_info(entry
);
2312 offset
= swp_offset(entry
);
2313 se
= offset_to_swap_extent(sis
, offset
);
2314 return se
->start_block
+ (offset
- se
->start_page
);
2318 * Returns the page offset into bdev for the specified page's swap entry.
2320 sector_t
map_swap_page(struct page
*page
, struct block_device
**bdev
)
2323 entry
.val
= page_private(page
);
2324 return map_swap_entry(entry
, bdev
);
2328 * Free all of a swapdev's extent information
2330 static void destroy_swap_extents(struct swap_info_struct
*sis
)
2332 while (!RB_EMPTY_ROOT(&sis
->swap_extent_root
)) {
2333 struct rb_node
*rb
= sis
->swap_extent_root
.rb_node
;
2334 struct swap_extent
*se
= rb_entry(rb
, struct swap_extent
, rb_node
);
2336 rb_erase(rb
, &sis
->swap_extent_root
);
2340 if (sis
->flags
& SWP_ACTIVATED
) {
2341 struct file
*swap_file
= sis
->swap_file
;
2342 struct address_space
*mapping
= swap_file
->f_mapping
;
2344 sis
->flags
&= ~SWP_ACTIVATED
;
2345 if (mapping
->a_ops
->swap_deactivate
)
2346 mapping
->a_ops
->swap_deactivate(swap_file
);
2351 * Add a block range (and the corresponding page range) into this swapdev's
2354 * This function rather assumes that it is called in ascending page order.
2357 add_swap_extent(struct swap_info_struct
*sis
, unsigned long start_page
,
2358 unsigned long nr_pages
, sector_t start_block
)
2360 struct rb_node
**link
= &sis
->swap_extent_root
.rb_node
, *parent
= NULL
;
2361 struct swap_extent
*se
;
2362 struct swap_extent
*new_se
;
2365 * place the new node at the right most since the
2366 * function is called in ascending page order.
2370 link
= &parent
->rb_right
;
2374 se
= rb_entry(parent
, struct swap_extent
, rb_node
);
2375 BUG_ON(se
->start_page
+ se
->nr_pages
!= start_page
);
2376 if (se
->start_block
+ se
->nr_pages
== start_block
) {
2378 se
->nr_pages
+= nr_pages
;
2383 /* No merge, insert a new extent. */
2384 new_se
= kmalloc(sizeof(*se
), GFP_KERNEL
);
2387 new_se
->start_page
= start_page
;
2388 new_se
->nr_pages
= nr_pages
;
2389 new_se
->start_block
= start_block
;
2391 rb_link_node(&new_se
->rb_node
, parent
, link
);
2392 rb_insert_color(&new_se
->rb_node
, &sis
->swap_extent_root
);
2395 EXPORT_SYMBOL_GPL(add_swap_extent
);
2398 * A `swap extent' is a simple thing which maps a contiguous range of pages
2399 * onto a contiguous range of disk blocks. An ordered list of swap extents
2400 * is built at swapon time and is then used at swap_writepage/swap_readpage
2401 * time for locating where on disk a page belongs.
2403 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2404 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2405 * swap files identically.
2407 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2408 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2409 * swapfiles are handled *identically* after swapon time.
2411 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2412 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2413 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2414 * requirements, they are simply tossed out - we will never use those blocks
2417 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2418 * prevents users from writing to the swap device, which will corrupt memory.
2420 * The amount of disk space which a single swap extent represents varies.
2421 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2422 * extents in the list. To avoid much list walking, we cache the previous
2423 * search location in `curr_swap_extent', and start new searches from there.
2424 * This is extremely effective. The average number of iterations in
2425 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2427 static int setup_swap_extents(struct swap_info_struct
*sis
, sector_t
*span
)
2429 struct file
*swap_file
= sis
->swap_file
;
2430 struct address_space
*mapping
= swap_file
->f_mapping
;
2431 struct inode
*inode
= mapping
->host
;
2434 if (S_ISBLK(inode
->i_mode
)) {
2435 ret
= add_swap_extent(sis
, 0, sis
->max
, 0);
2440 if (mapping
->a_ops
->swap_activate
) {
2441 ret
= mapping
->a_ops
->swap_activate(sis
, swap_file
, span
);
2443 sis
->flags
|= SWP_ACTIVATED
;
2445 sis
->flags
|= SWP_FS_OPS
;
2446 ret
= add_swap_extent(sis
, 0, sis
->max
, 0);
2452 return generic_swapfile_activate(sis
, swap_file
, span
);
2455 static int swap_node(struct swap_info_struct
*p
)
2457 struct block_device
*bdev
;
2462 bdev
= p
->swap_file
->f_inode
->i_sb
->s_bdev
;
2464 return bdev
? bdev
->bd_disk
->node_id
: NUMA_NO_NODE
;
2467 static void setup_swap_info(struct swap_info_struct
*p
, int prio
,
2468 unsigned char *swap_map
,
2469 struct swap_cluster_info
*cluster_info
)
2476 p
->prio
= --least_priority
;
2478 * the plist prio is negated because plist ordering is
2479 * low-to-high, while swap ordering is high-to-low
2481 p
->list
.prio
= -p
->prio
;
2484 p
->avail_lists
[i
].prio
= -p
->prio
;
2486 if (swap_node(p
) == i
)
2487 p
->avail_lists
[i
].prio
= 1;
2489 p
->avail_lists
[i
].prio
= -p
->prio
;
2492 p
->swap_map
= swap_map
;
2493 p
->cluster_info
= cluster_info
;
2496 static void _enable_swap_info(struct swap_info_struct
*p
)
2498 p
->flags
|= SWP_WRITEOK
| SWP_VALID
;
2499 atomic_long_add(p
->pages
, &nr_swap_pages
);
2500 total_swap_pages
+= p
->pages
;
2502 assert_spin_locked(&swap_lock
);
2504 * both lists are plists, and thus priority ordered.
2505 * swap_active_head needs to be priority ordered for swapoff(),
2506 * which on removal of any swap_info_struct with an auto-assigned
2507 * (i.e. negative) priority increments the auto-assigned priority
2508 * of any lower-priority swap_info_structs.
2509 * swap_avail_head needs to be priority ordered for get_swap_page(),
2510 * which allocates swap pages from the highest available priority
2513 plist_add(&p
->list
, &swap_active_head
);
2514 add_to_avail_list(p
);
2517 static void enable_swap_info(struct swap_info_struct
*p
, int prio
,
2518 unsigned char *swap_map
,
2519 struct swap_cluster_info
*cluster_info
,
2520 unsigned long *frontswap_map
)
2522 frontswap_init(p
->type
, frontswap_map
);
2523 spin_lock(&swap_lock
);
2524 spin_lock(&p
->lock
);
2525 setup_swap_info(p
, prio
, swap_map
, cluster_info
);
2526 spin_unlock(&p
->lock
);
2527 spin_unlock(&swap_lock
);
2529 * Guarantee swap_map, cluster_info, etc. fields are valid
2530 * between get/put_swap_device() if SWP_VALID bit is set
2533 spin_lock(&swap_lock
);
2534 spin_lock(&p
->lock
);
2535 _enable_swap_info(p
);
2536 spin_unlock(&p
->lock
);
2537 spin_unlock(&swap_lock
);
2540 static void reinsert_swap_info(struct swap_info_struct
*p
)
2542 spin_lock(&swap_lock
);
2543 spin_lock(&p
->lock
);
2544 setup_swap_info(p
, p
->prio
, p
->swap_map
, p
->cluster_info
);
2545 _enable_swap_info(p
);
2546 spin_unlock(&p
->lock
);
2547 spin_unlock(&swap_lock
);
2550 bool has_usable_swap(void)
2554 spin_lock(&swap_lock
);
2555 if (plist_head_empty(&swap_active_head
))
2557 spin_unlock(&swap_lock
);
2561 SYSCALL_DEFINE1(swapoff
, const char __user
*, specialfile
)
2563 struct swap_info_struct
*p
= NULL
;
2564 unsigned char *swap_map
;
2565 struct swap_cluster_info
*cluster_info
;
2566 unsigned long *frontswap_map
;
2567 struct file
*swap_file
, *victim
;
2568 struct address_space
*mapping
;
2569 struct inode
*inode
;
2570 struct filename
*pathname
;
2572 unsigned int old_block_size
;
2574 if (!capable(CAP_SYS_ADMIN
))
2577 BUG_ON(!current
->mm
);
2579 pathname
= getname(specialfile
);
2580 if (IS_ERR(pathname
))
2581 return PTR_ERR(pathname
);
2583 victim
= file_open_name(pathname
, O_RDWR
|O_LARGEFILE
, 0);
2584 err
= PTR_ERR(victim
);
2588 mapping
= victim
->f_mapping
;
2589 spin_lock(&swap_lock
);
2590 plist_for_each_entry(p
, &swap_active_head
, list
) {
2591 if (p
->flags
& SWP_WRITEOK
) {
2592 if (p
->swap_file
->f_mapping
== mapping
) {
2600 spin_unlock(&swap_lock
);
2603 if (!security_vm_enough_memory_mm(current
->mm
, p
->pages
))
2604 vm_unacct_memory(p
->pages
);
2607 spin_unlock(&swap_lock
);
2610 del_from_avail_list(p
);
2611 spin_lock(&p
->lock
);
2613 struct swap_info_struct
*si
= p
;
2616 plist_for_each_entry_continue(si
, &swap_active_head
, list
) {
2619 for_each_node(nid
) {
2620 if (si
->avail_lists
[nid
].prio
!= 1)
2621 si
->avail_lists
[nid
].prio
--;
2626 plist_del(&p
->list
, &swap_active_head
);
2627 atomic_long_sub(p
->pages
, &nr_swap_pages
);
2628 total_swap_pages
-= p
->pages
;
2629 p
->flags
&= ~SWP_WRITEOK
;
2630 spin_unlock(&p
->lock
);
2631 spin_unlock(&swap_lock
);
2633 disable_swap_slots_cache_lock();
2635 set_current_oom_origin();
2636 err
= try_to_unuse(p
->type
, false, 0); /* force unuse all pages */
2637 clear_current_oom_origin();
2640 /* re-insert swap space back into swap_list */
2641 reinsert_swap_info(p
);
2642 reenable_swap_slots_cache_unlock();
2646 reenable_swap_slots_cache_unlock();
2648 spin_lock(&swap_lock
);
2649 spin_lock(&p
->lock
);
2650 p
->flags
&= ~SWP_VALID
; /* mark swap device as invalid */
2651 spin_unlock(&p
->lock
);
2652 spin_unlock(&swap_lock
);
2654 * wait for swap operations protected by get/put_swap_device()
2659 flush_work(&p
->discard_work
);
2661 destroy_swap_extents(p
);
2662 if (p
->flags
& SWP_CONTINUED
)
2663 free_swap_count_continuations(p
);
2665 if (!p
->bdev
|| !blk_queue_nonrot(bdev_get_queue(p
->bdev
)))
2666 atomic_dec(&nr_rotate_swap
);
2668 mutex_lock(&swapon_mutex
);
2669 spin_lock(&swap_lock
);
2670 spin_lock(&p
->lock
);
2673 /* wait for anyone still in scan_swap_map */
2674 p
->highest_bit
= 0; /* cuts scans short */
2675 while (p
->flags
>= SWP_SCANNING
) {
2676 spin_unlock(&p
->lock
);
2677 spin_unlock(&swap_lock
);
2678 schedule_timeout_uninterruptible(1);
2679 spin_lock(&swap_lock
);
2680 spin_lock(&p
->lock
);
2683 swap_file
= p
->swap_file
;
2684 old_block_size
= p
->old_block_size
;
2685 p
->swap_file
= NULL
;
2687 swap_map
= p
->swap_map
;
2689 cluster_info
= p
->cluster_info
;
2690 p
->cluster_info
= NULL
;
2691 frontswap_map
= frontswap_map_get(p
);
2692 spin_unlock(&p
->lock
);
2693 spin_unlock(&swap_lock
);
2694 arch_swap_invalidate_area(p
->type
);
2695 frontswap_invalidate_area(p
->type
);
2696 frontswap_map_set(p
, NULL
);
2697 mutex_unlock(&swapon_mutex
);
2698 free_percpu(p
->percpu_cluster
);
2699 p
->percpu_cluster
= NULL
;
2700 free_percpu(p
->cluster_next_cpu
);
2701 p
->cluster_next_cpu
= NULL
;
2703 kvfree(cluster_info
);
2704 kvfree(frontswap_map
);
2705 /* Destroy swap account information */
2706 swap_cgroup_swapoff(p
->type
);
2707 exit_swap_address_space(p
->type
);
2709 inode
= mapping
->host
;
2710 if (S_ISBLK(inode
->i_mode
)) {
2711 struct block_device
*bdev
= I_BDEV(inode
);
2713 set_blocksize(bdev
, old_block_size
);
2714 blkdev_put(bdev
, FMODE_READ
| FMODE_WRITE
| FMODE_EXCL
);
2718 inode
->i_flags
&= ~S_SWAPFILE
;
2719 inode_unlock(inode
);
2720 filp_close(swap_file
, NULL
);
2723 * Clear the SWP_USED flag after all resources are freed so that swapon
2724 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2725 * not hold p->lock after we cleared its SWP_WRITEOK.
2727 spin_lock(&swap_lock
);
2729 spin_unlock(&swap_lock
);
2732 atomic_inc(&proc_poll_event
);
2733 wake_up_interruptible(&proc_poll_wait
);
2736 filp_close(victim
, NULL
);
2742 #ifdef CONFIG_PROC_FS
2743 static __poll_t
swaps_poll(struct file
*file
, poll_table
*wait
)
2745 struct seq_file
*seq
= file
->private_data
;
2747 poll_wait(file
, &proc_poll_wait
, wait
);
2749 if (seq
->poll_event
!= atomic_read(&proc_poll_event
)) {
2750 seq
->poll_event
= atomic_read(&proc_poll_event
);
2751 return EPOLLIN
| EPOLLRDNORM
| EPOLLERR
| EPOLLPRI
;
2754 return EPOLLIN
| EPOLLRDNORM
;
2758 static void *swap_start(struct seq_file
*swap
, loff_t
*pos
)
2760 struct swap_info_struct
*si
;
2764 mutex_lock(&swapon_mutex
);
2767 return SEQ_START_TOKEN
;
2769 for (type
= 0; (si
= swap_type_to_swap_info(type
)); type
++) {
2770 if (!(si
->flags
& SWP_USED
) || !si
->swap_map
)
2779 static void *swap_next(struct seq_file
*swap
, void *v
, loff_t
*pos
)
2781 struct swap_info_struct
*si
= v
;
2784 if (v
== SEQ_START_TOKEN
)
2787 type
= si
->type
+ 1;
2790 for (; (si
= swap_type_to_swap_info(type
)); type
++) {
2791 if (!(si
->flags
& SWP_USED
) || !si
->swap_map
)
2799 static void swap_stop(struct seq_file
*swap
, void *v
)
2801 mutex_unlock(&swapon_mutex
);
2804 static int swap_show(struct seq_file
*swap
, void *v
)
2806 struct swap_info_struct
*si
= v
;
2809 unsigned int bytes
, inuse
;
2811 if (si
== SEQ_START_TOKEN
) {
2812 seq_puts(swap
,"Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2816 bytes
= si
->pages
<< (PAGE_SHIFT
- 10);
2817 inuse
= si
->inuse_pages
<< (PAGE_SHIFT
- 10);
2819 file
= si
->swap_file
;
2820 len
= seq_file_path(swap
, file
, " \t\n\\");
2821 seq_printf(swap
, "%*s%s\t%u\t%s%u\t%s%d\n",
2822 len
< 40 ? 40 - len
: 1, " ",
2823 S_ISBLK(file_inode(file
)->i_mode
) ?
2824 "partition" : "file\t",
2825 bytes
, bytes
< 10000000 ? "\t" : "",
2826 inuse
, inuse
< 10000000 ? "\t" : "",
2831 static const struct seq_operations swaps_op
= {
2832 .start
= swap_start
,
2838 static int swaps_open(struct inode
*inode
, struct file
*file
)
2840 struct seq_file
*seq
;
2843 ret
= seq_open(file
, &swaps_op
);
2847 seq
= file
->private_data
;
2848 seq
->poll_event
= atomic_read(&proc_poll_event
);
2852 static const struct proc_ops swaps_proc_ops
= {
2853 .proc_flags
= PROC_ENTRY_PERMANENT
,
2854 .proc_open
= swaps_open
,
2855 .proc_read
= seq_read
,
2856 .proc_lseek
= seq_lseek
,
2857 .proc_release
= seq_release
,
2858 .proc_poll
= swaps_poll
,
2861 static int __init
procswaps_init(void)
2863 proc_create("swaps", 0, NULL
, &swaps_proc_ops
);
2866 __initcall(procswaps_init
);
2867 #endif /* CONFIG_PROC_FS */
2869 #ifdef MAX_SWAPFILES_CHECK
2870 static int __init
max_swapfiles_check(void)
2872 MAX_SWAPFILES_CHECK();
2875 late_initcall(max_swapfiles_check
);
2878 static struct swap_info_struct
*alloc_swap_info(void)
2880 struct swap_info_struct
*p
;
2881 struct swap_info_struct
*defer
= NULL
;
2885 p
= kvzalloc(struct_size(p
, avail_lists
, nr_node_ids
), GFP_KERNEL
);
2887 return ERR_PTR(-ENOMEM
);
2889 spin_lock(&swap_lock
);
2890 for (type
= 0; type
< nr_swapfiles
; type
++) {
2891 if (!(swap_info
[type
]->flags
& SWP_USED
))
2894 if (type
>= MAX_SWAPFILES
) {
2895 spin_unlock(&swap_lock
);
2897 return ERR_PTR(-EPERM
);
2899 if (type
>= nr_swapfiles
) {
2901 WRITE_ONCE(swap_info
[type
], p
);
2903 * Write swap_info[type] before nr_swapfiles, in case a
2904 * racing procfs swap_start() or swap_next() is reading them.
2905 * (We never shrink nr_swapfiles, we never free this entry.)
2908 WRITE_ONCE(nr_swapfiles
, nr_swapfiles
+ 1);
2911 p
= swap_info
[type
];
2913 * Do not memset this entry: a racing procfs swap_next()
2914 * would be relying on p->type to remain valid.
2917 p
->swap_extent_root
= RB_ROOT
;
2918 plist_node_init(&p
->list
, 0);
2920 plist_node_init(&p
->avail_lists
[i
], 0);
2921 p
->flags
= SWP_USED
;
2922 spin_unlock(&swap_lock
);
2924 spin_lock_init(&p
->lock
);
2925 spin_lock_init(&p
->cont_lock
);
2930 static int claim_swapfile(struct swap_info_struct
*p
, struct inode
*inode
)
2934 if (S_ISBLK(inode
->i_mode
)) {
2935 p
->bdev
= blkdev_get_by_dev(inode
->i_rdev
,
2936 FMODE_READ
| FMODE_WRITE
| FMODE_EXCL
, p
);
2937 if (IS_ERR(p
->bdev
)) {
2938 error
= PTR_ERR(p
->bdev
);
2942 p
->old_block_size
= block_size(p
->bdev
);
2943 error
= set_blocksize(p
->bdev
, PAGE_SIZE
);
2947 * Zoned block devices contain zones that have a sequential
2948 * write only restriction. Hence zoned block devices are not
2949 * suitable for swapping. Disallow them here.
2951 if (blk_queue_is_zoned(p
->bdev
->bd_disk
->queue
))
2953 p
->flags
|= SWP_BLKDEV
;
2954 } else if (S_ISREG(inode
->i_mode
)) {
2955 p
->bdev
= inode
->i_sb
->s_bdev
;
2963 * Find out how many pages are allowed for a single swap device. There
2964 * are two limiting factors:
2965 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2966 * 2) the number of bits in the swap pte, as defined by the different
2969 * In order to find the largest possible bit mask, a swap entry with
2970 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2971 * decoded to a swp_entry_t again, and finally the swap offset is
2974 * This will mask all the bits from the initial ~0UL mask that can't
2975 * be encoded in either the swp_entry_t or the architecture definition
2978 unsigned long generic_max_swapfile_size(void)
2980 return swp_offset(pte_to_swp_entry(
2981 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2984 /* Can be overridden by an architecture for additional checks. */
2985 __weak
unsigned long max_swapfile_size(void)
2987 return generic_max_swapfile_size();
2990 static unsigned long read_swap_header(struct swap_info_struct
*p
,
2991 union swap_header
*swap_header
,
2992 struct inode
*inode
)
2995 unsigned long maxpages
;
2996 unsigned long swapfilepages
;
2997 unsigned long last_page
;
2999 if (memcmp("SWAPSPACE2", swap_header
->magic
.magic
, 10)) {
3000 pr_err("Unable to find swap-space signature\n");
3004 /* swap partition endianess hack... */
3005 if (swab32(swap_header
->info
.version
) == 1) {
3006 swab32s(&swap_header
->info
.version
);
3007 swab32s(&swap_header
->info
.last_page
);
3008 swab32s(&swap_header
->info
.nr_badpages
);
3009 if (swap_header
->info
.nr_badpages
> MAX_SWAP_BADPAGES
)
3011 for (i
= 0; i
< swap_header
->info
.nr_badpages
; i
++)
3012 swab32s(&swap_header
->info
.badpages
[i
]);
3014 /* Check the swap header's sub-version */
3015 if (swap_header
->info
.version
!= 1) {
3016 pr_warn("Unable to handle swap header version %d\n",
3017 swap_header
->info
.version
);
3022 p
->cluster_next
= 1;
3025 maxpages
= max_swapfile_size();
3026 last_page
= swap_header
->info
.last_page
;
3028 pr_warn("Empty swap-file\n");
3031 if (last_page
> maxpages
) {
3032 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3033 maxpages
<< (PAGE_SHIFT
- 10),
3034 last_page
<< (PAGE_SHIFT
- 10));
3036 if (maxpages
> last_page
) {
3037 maxpages
= last_page
+ 1;
3038 /* p->max is an unsigned int: don't overflow it */
3039 if ((unsigned int)maxpages
== 0)
3040 maxpages
= UINT_MAX
;
3042 p
->highest_bit
= maxpages
- 1;
3046 swapfilepages
= i_size_read(inode
) >> PAGE_SHIFT
;
3047 if (swapfilepages
&& maxpages
> swapfilepages
) {
3048 pr_warn("Swap area shorter than signature indicates\n");
3051 if (swap_header
->info
.nr_badpages
&& S_ISREG(inode
->i_mode
))
3053 if (swap_header
->info
.nr_badpages
> MAX_SWAP_BADPAGES
)
3059 #define SWAP_CLUSTER_INFO_COLS \
3060 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3061 #define SWAP_CLUSTER_SPACE_COLS \
3062 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3063 #define SWAP_CLUSTER_COLS \
3064 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3066 static int setup_swap_map_and_extents(struct swap_info_struct
*p
,
3067 union swap_header
*swap_header
,
3068 unsigned char *swap_map
,
3069 struct swap_cluster_info
*cluster_info
,
3070 unsigned long maxpages
,
3074 unsigned int nr_good_pages
;
3076 unsigned long nr_clusters
= DIV_ROUND_UP(maxpages
, SWAPFILE_CLUSTER
);
3077 unsigned long col
= p
->cluster_next
/ SWAPFILE_CLUSTER
% SWAP_CLUSTER_COLS
;
3078 unsigned long i
, idx
;
3080 nr_good_pages
= maxpages
- 1; /* omit header page */
3082 cluster_list_init(&p
->free_clusters
);
3083 cluster_list_init(&p
->discard_clusters
);
3085 for (i
= 0; i
< swap_header
->info
.nr_badpages
; i
++) {
3086 unsigned int page_nr
= swap_header
->info
.badpages
[i
];
3087 if (page_nr
== 0 || page_nr
> swap_header
->info
.last_page
)
3089 if (page_nr
< maxpages
) {
3090 swap_map
[page_nr
] = SWAP_MAP_BAD
;
3093 * Haven't marked the cluster free yet, no list
3094 * operation involved
3096 inc_cluster_info_page(p
, cluster_info
, page_nr
);
3100 /* Haven't marked the cluster free yet, no list operation involved */
3101 for (i
= maxpages
; i
< round_up(maxpages
, SWAPFILE_CLUSTER
); i
++)
3102 inc_cluster_info_page(p
, cluster_info
, i
);
3104 if (nr_good_pages
) {
3105 swap_map
[0] = SWAP_MAP_BAD
;
3107 * Not mark the cluster free yet, no list
3108 * operation involved
3110 inc_cluster_info_page(p
, cluster_info
, 0);
3112 p
->pages
= nr_good_pages
;
3113 nr_extents
= setup_swap_extents(p
, span
);
3116 nr_good_pages
= p
->pages
;
3118 if (!nr_good_pages
) {
3119 pr_warn("Empty swap-file\n");
3128 * Reduce false cache line sharing between cluster_info and
3129 * sharing same address space.
3131 for (k
= 0; k
< SWAP_CLUSTER_COLS
; k
++) {
3132 j
= (k
+ col
) % SWAP_CLUSTER_COLS
;
3133 for (i
= 0; i
< DIV_ROUND_UP(nr_clusters
, SWAP_CLUSTER_COLS
); i
++) {
3134 idx
= i
* SWAP_CLUSTER_COLS
+ j
;
3135 if (idx
>= nr_clusters
)
3137 if (cluster_count(&cluster_info
[idx
]))
3139 cluster_set_flag(&cluster_info
[idx
], CLUSTER_FLAG_FREE
);
3140 cluster_list_add_tail(&p
->free_clusters
, cluster_info
,
3148 * Helper to sys_swapon determining if a given swap
3149 * backing device queue supports DISCARD operations.
3151 static bool swap_discardable(struct swap_info_struct
*si
)
3153 struct request_queue
*q
= bdev_get_queue(si
->bdev
);
3155 if (!q
|| !blk_queue_discard(q
))
3161 SYSCALL_DEFINE2(swapon
, const char __user
*, specialfile
, int, swap_flags
)
3163 struct swap_info_struct
*p
;
3164 struct filename
*name
;
3165 struct file
*swap_file
= NULL
;
3166 struct address_space
*mapping
;
3169 union swap_header
*swap_header
;
3172 unsigned long maxpages
;
3173 unsigned char *swap_map
= NULL
;
3174 struct swap_cluster_info
*cluster_info
= NULL
;
3175 unsigned long *frontswap_map
= NULL
;
3176 struct page
*page
= NULL
;
3177 struct inode
*inode
= NULL
;
3178 bool inced_nr_rotate_swap
= false;
3180 if (swap_flags
& ~SWAP_FLAGS_VALID
)
3183 if (!capable(CAP_SYS_ADMIN
))
3186 if (!swap_avail_heads
)
3189 p
= alloc_swap_info();
3193 INIT_WORK(&p
->discard_work
, swap_discard_work
);
3195 name
= getname(specialfile
);
3197 error
= PTR_ERR(name
);
3201 swap_file
= file_open_name(name
, O_RDWR
|O_LARGEFILE
, 0);
3202 if (IS_ERR(swap_file
)) {
3203 error
= PTR_ERR(swap_file
);
3208 p
->swap_file
= swap_file
;
3209 mapping
= swap_file
->f_mapping
;
3210 inode
= mapping
->host
;
3212 error
= claim_swapfile(p
, inode
);
3213 if (unlikely(error
))
3217 if (IS_SWAPFILE(inode
)) {
3219 goto bad_swap_unlock_inode
;
3223 * Read the swap header.
3225 if (!mapping
->a_ops
->readpage
) {
3227 goto bad_swap_unlock_inode
;
3229 page
= read_mapping_page(mapping
, 0, swap_file
);
3231 error
= PTR_ERR(page
);
3232 goto bad_swap_unlock_inode
;
3234 swap_header
= kmap(page
);
3236 maxpages
= read_swap_header(p
, swap_header
, inode
);
3237 if (unlikely(!maxpages
)) {
3239 goto bad_swap_unlock_inode
;
3242 /* OK, set up the swap map and apply the bad block list */
3243 swap_map
= vzalloc(maxpages
);
3246 goto bad_swap_unlock_inode
;
3249 if (p
->bdev
&& blk_queue_stable_writes(p
->bdev
->bd_disk
->queue
))
3250 p
->flags
|= SWP_STABLE_WRITES
;
3252 if (p
->bdev
&& p
->bdev
->bd_disk
->fops
->rw_page
)
3253 p
->flags
|= SWP_SYNCHRONOUS_IO
;
3255 if (p
->bdev
&& blk_queue_nonrot(bdev_get_queue(p
->bdev
))) {
3257 unsigned long ci
, nr_cluster
;
3259 p
->flags
|= SWP_SOLIDSTATE
;
3260 p
->cluster_next_cpu
= alloc_percpu(unsigned int);
3261 if (!p
->cluster_next_cpu
) {
3263 goto bad_swap_unlock_inode
;
3266 * select a random position to start with to help wear leveling
3269 for_each_possible_cpu(cpu
) {
3270 per_cpu(*p
->cluster_next_cpu
, cpu
) =
3271 1 + prandom_u32_max(p
->highest_bit
);
3273 nr_cluster
= DIV_ROUND_UP(maxpages
, SWAPFILE_CLUSTER
);
3275 cluster_info
= kvcalloc(nr_cluster
, sizeof(*cluster_info
),
3277 if (!cluster_info
) {
3279 goto bad_swap_unlock_inode
;
3282 for (ci
= 0; ci
< nr_cluster
; ci
++)
3283 spin_lock_init(&((cluster_info
+ ci
)->lock
));
3285 p
->percpu_cluster
= alloc_percpu(struct percpu_cluster
);
3286 if (!p
->percpu_cluster
) {
3288 goto bad_swap_unlock_inode
;
3290 for_each_possible_cpu(cpu
) {
3291 struct percpu_cluster
*cluster
;
3292 cluster
= per_cpu_ptr(p
->percpu_cluster
, cpu
);
3293 cluster_set_null(&cluster
->index
);
3296 atomic_inc(&nr_rotate_swap
);
3297 inced_nr_rotate_swap
= true;
3300 error
= swap_cgroup_swapon(p
->type
, maxpages
);
3302 goto bad_swap_unlock_inode
;
3304 nr_extents
= setup_swap_map_and_extents(p
, swap_header
, swap_map
,
3305 cluster_info
, maxpages
, &span
);
3306 if (unlikely(nr_extents
< 0)) {
3308 goto bad_swap_unlock_inode
;
3310 /* frontswap enabled? set up bit-per-page map for frontswap */
3311 if (IS_ENABLED(CONFIG_FRONTSWAP
))
3312 frontswap_map
= kvcalloc(BITS_TO_LONGS(maxpages
),
3316 if (p
->bdev
&&(swap_flags
& SWAP_FLAG_DISCARD
) && swap_discardable(p
)) {
3318 * When discard is enabled for swap with no particular
3319 * policy flagged, we set all swap discard flags here in
3320 * order to sustain backward compatibility with older
3321 * swapon(8) releases.
3323 p
->flags
|= (SWP_DISCARDABLE
| SWP_AREA_DISCARD
|
3327 * By flagging sys_swapon, a sysadmin can tell us to
3328 * either do single-time area discards only, or to just
3329 * perform discards for released swap page-clusters.
3330 * Now it's time to adjust the p->flags accordingly.
3332 if (swap_flags
& SWAP_FLAG_DISCARD_ONCE
)
3333 p
->flags
&= ~SWP_PAGE_DISCARD
;
3334 else if (swap_flags
& SWAP_FLAG_DISCARD_PAGES
)
3335 p
->flags
&= ~SWP_AREA_DISCARD
;
3337 /* issue a swapon-time discard if it's still required */
3338 if (p
->flags
& SWP_AREA_DISCARD
) {
3339 int err
= discard_swap(p
);
3341 pr_err("swapon: discard_swap(%p): %d\n",
3346 error
= init_swap_address_space(p
->type
, maxpages
);
3348 goto bad_swap_unlock_inode
;
3351 * Flush any pending IO and dirty mappings before we start using this
3354 inode
->i_flags
|= S_SWAPFILE
;
3355 error
= inode_drain_writes(inode
);
3357 inode
->i_flags
&= ~S_SWAPFILE
;
3358 goto free_swap_address_space
;
3361 mutex_lock(&swapon_mutex
);
3363 if (swap_flags
& SWAP_FLAG_PREFER
)
3365 (swap_flags
& SWAP_FLAG_PRIO_MASK
) >> SWAP_FLAG_PRIO_SHIFT
;
3366 enable_swap_info(p
, prio
, swap_map
, cluster_info
, frontswap_map
);
3368 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3369 p
->pages
<<(PAGE_SHIFT
-10), name
->name
, p
->prio
,
3370 nr_extents
, (unsigned long long)span
<<(PAGE_SHIFT
-10),
3371 (p
->flags
& SWP_SOLIDSTATE
) ? "SS" : "",
3372 (p
->flags
& SWP_DISCARDABLE
) ? "D" : "",
3373 (p
->flags
& SWP_AREA_DISCARD
) ? "s" : "",
3374 (p
->flags
& SWP_PAGE_DISCARD
) ? "c" : "",
3375 (frontswap_map
) ? "FS" : "");
3377 mutex_unlock(&swapon_mutex
);
3378 atomic_inc(&proc_poll_event
);
3379 wake_up_interruptible(&proc_poll_wait
);
3383 free_swap_address_space
:
3384 exit_swap_address_space(p
->type
);
3385 bad_swap_unlock_inode
:
3386 inode_unlock(inode
);
3388 free_percpu(p
->percpu_cluster
);
3389 p
->percpu_cluster
= NULL
;
3390 free_percpu(p
->cluster_next_cpu
);
3391 p
->cluster_next_cpu
= NULL
;
3392 if (inode
&& S_ISBLK(inode
->i_mode
) && p
->bdev
) {
3393 set_blocksize(p
->bdev
, p
->old_block_size
);
3394 blkdev_put(p
->bdev
, FMODE_READ
| FMODE_WRITE
| FMODE_EXCL
);
3397 destroy_swap_extents(p
);
3398 swap_cgroup_swapoff(p
->type
);
3399 spin_lock(&swap_lock
);
3400 p
->swap_file
= NULL
;
3402 spin_unlock(&swap_lock
);
3404 kvfree(cluster_info
);
3405 kvfree(frontswap_map
);
3406 if (inced_nr_rotate_swap
)
3407 atomic_dec(&nr_rotate_swap
);
3409 filp_close(swap_file
, NULL
);
3411 if (page
&& !IS_ERR(page
)) {
3418 inode_unlock(inode
);
3420 enable_swap_slots_cache();
3424 void si_swapinfo(struct sysinfo
*val
)
3427 unsigned long nr_to_be_unused
= 0;
3429 spin_lock(&swap_lock
);
3430 for (type
= 0; type
< nr_swapfiles
; type
++) {
3431 struct swap_info_struct
*si
= swap_info
[type
];
3433 if ((si
->flags
& SWP_USED
) && !(si
->flags
& SWP_WRITEOK
))
3434 nr_to_be_unused
+= si
->inuse_pages
;
3436 val
->freeswap
= atomic_long_read(&nr_swap_pages
) + nr_to_be_unused
;
3437 val
->totalswap
= total_swap_pages
+ nr_to_be_unused
;
3438 spin_unlock(&swap_lock
);
3442 * Verify that a swap entry is valid and increment its swap map count.
3444 * Returns error code in following case.
3446 * - swp_entry is invalid -> EINVAL
3447 * - swp_entry is migration entry -> EINVAL
3448 * - swap-cache reference is requested but there is already one. -> EEXIST
3449 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3450 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3452 static int __swap_duplicate(swp_entry_t entry
, unsigned char usage
)
3454 struct swap_info_struct
*p
;
3455 struct swap_cluster_info
*ci
;
3456 unsigned long offset
;
3457 unsigned char count
;
3458 unsigned char has_cache
;
3461 p
= get_swap_device(entry
);
3465 offset
= swp_offset(entry
);
3466 ci
= lock_cluster_or_swap_info(p
, offset
);
3468 count
= p
->swap_map
[offset
];
3471 * swapin_readahead() doesn't check if a swap entry is valid, so the
3472 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3474 if (unlikely(swap_count(count
) == SWAP_MAP_BAD
)) {
3479 has_cache
= count
& SWAP_HAS_CACHE
;
3480 count
&= ~SWAP_HAS_CACHE
;
3483 if (usage
== SWAP_HAS_CACHE
) {
3485 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3486 if (!has_cache
&& count
)
3487 has_cache
= SWAP_HAS_CACHE
;
3488 else if (has_cache
) /* someone else added cache */
3490 else /* no users remaining */
3493 } else if (count
|| has_cache
) {
3495 if ((count
& ~COUNT_CONTINUED
) < SWAP_MAP_MAX
)
3497 else if ((count
& ~COUNT_CONTINUED
) > SWAP_MAP_MAX
)
3499 else if (swap_count_continued(p
, offset
, count
))
3500 count
= COUNT_CONTINUED
;
3504 err
= -ENOENT
; /* unused swap entry */
3506 WRITE_ONCE(p
->swap_map
[offset
], count
| has_cache
);
3509 unlock_cluster_or_swap_info(p
, ci
);
3516 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3517 * (in which case its reference count is never incremented).
3519 void swap_shmem_alloc(swp_entry_t entry
)
3521 __swap_duplicate(entry
, SWAP_MAP_SHMEM
);
3525 * Increase reference count of swap entry by 1.
3526 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3527 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3528 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3529 * might occur if a page table entry has got corrupted.
3531 int swap_duplicate(swp_entry_t entry
)
3535 while (!err
&& __swap_duplicate(entry
, 1) == -ENOMEM
)
3536 err
= add_swap_count_continuation(entry
, GFP_ATOMIC
);
3541 * @entry: swap entry for which we allocate swap cache.
3543 * Called when allocating swap cache for existing swap entry,
3544 * This can return error codes. Returns 0 at success.
3545 * -EEXIST means there is a swap cache.
3546 * Note: return code is different from swap_duplicate().
3548 int swapcache_prepare(swp_entry_t entry
)
3550 return __swap_duplicate(entry
, SWAP_HAS_CACHE
);
3553 struct swap_info_struct
*swp_swap_info(swp_entry_t entry
)
3555 return swap_type_to_swap_info(swp_type(entry
));
3558 struct swap_info_struct
*page_swap_info(struct page
*page
)
3560 swp_entry_t entry
= { .val
= page_private(page
) };
3561 return swp_swap_info(entry
);
3565 * out-of-line __page_file_ methods to avoid include hell.
3567 struct address_space
*__page_file_mapping(struct page
*page
)
3569 return page_swap_info(page
)->swap_file
->f_mapping
;
3571 EXPORT_SYMBOL_GPL(__page_file_mapping
);
3573 pgoff_t
__page_file_index(struct page
*page
)
3575 swp_entry_t swap
= { .val
= page_private(page
) };
3576 return swp_offset(swap
);
3578 EXPORT_SYMBOL_GPL(__page_file_index
);
3581 * add_swap_count_continuation - called when a swap count is duplicated
3582 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3583 * page of the original vmalloc'ed swap_map, to hold the continuation count
3584 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3585 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3587 * These continuation pages are seldom referenced: the common paths all work
3588 * on the original swap_map, only referring to a continuation page when the
3589 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3591 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3592 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3593 * can be called after dropping locks.
3595 int add_swap_count_continuation(swp_entry_t entry
, gfp_t gfp_mask
)
3597 struct swap_info_struct
*si
;
3598 struct swap_cluster_info
*ci
;
3601 struct page
*list_page
;
3603 unsigned char count
;
3607 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3608 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3610 page
= alloc_page(gfp_mask
| __GFP_HIGHMEM
);
3612 si
= get_swap_device(entry
);
3615 * An acceptable race has occurred since the failing
3616 * __swap_duplicate(): the swap device may be swapoff
3620 spin_lock(&si
->lock
);
3622 offset
= swp_offset(entry
);
3624 ci
= lock_cluster(si
, offset
);
3626 count
= swap_count(si
->swap_map
[offset
]);
3628 if ((count
& ~COUNT_CONTINUED
) != SWAP_MAP_MAX
) {
3630 * The higher the swap count, the more likely it is that tasks
3631 * will race to add swap count continuation: we need to avoid
3632 * over-provisioning.
3643 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3644 * no architecture is using highmem pages for kernel page tables: so it
3645 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3647 head
= vmalloc_to_page(si
->swap_map
+ offset
);
3648 offset
&= ~PAGE_MASK
;
3650 spin_lock(&si
->cont_lock
);
3652 * Page allocation does not initialize the page's lru field,
3653 * but it does always reset its private field.
3655 if (!page_private(head
)) {
3656 BUG_ON(count
& COUNT_CONTINUED
);
3657 INIT_LIST_HEAD(&head
->lru
);
3658 set_page_private(head
, SWP_CONTINUED
);
3659 si
->flags
|= SWP_CONTINUED
;
3662 list_for_each_entry(list_page
, &head
->lru
, lru
) {
3666 * If the previous map said no continuation, but we've found
3667 * a continuation page, free our allocation and use this one.
3669 if (!(count
& COUNT_CONTINUED
))
3670 goto out_unlock_cont
;
3672 map
= kmap_atomic(list_page
) + offset
;
3677 * If this continuation count now has some space in it,
3678 * free our allocation and use this one.
3680 if ((count
& ~COUNT_CONTINUED
) != SWAP_CONT_MAX
)
3681 goto out_unlock_cont
;
3684 list_add_tail(&page
->lru
, &head
->lru
);
3685 page
= NULL
; /* now it's attached, don't free it */
3687 spin_unlock(&si
->cont_lock
);
3690 spin_unlock(&si
->lock
);
3691 put_swap_device(si
);
3699 * swap_count_continued - when the original swap_map count is incremented
3700 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3701 * into, carry if so, or else fail until a new continuation page is allocated;
3702 * when the original swap_map count is decremented from 0 with continuation,
3703 * borrow from the continuation and report whether it still holds more.
3704 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3707 static bool swap_count_continued(struct swap_info_struct
*si
,
3708 pgoff_t offset
, unsigned char count
)
3715 head
= vmalloc_to_page(si
->swap_map
+ offset
);
3716 if (page_private(head
) != SWP_CONTINUED
) {
3717 BUG_ON(count
& COUNT_CONTINUED
);
3718 return false; /* need to add count continuation */
3721 spin_lock(&si
->cont_lock
);
3722 offset
&= ~PAGE_MASK
;
3723 page
= list_next_entry(head
, lru
);
3724 map
= kmap_atomic(page
) + offset
;
3726 if (count
== SWAP_MAP_MAX
) /* initial increment from swap_map */
3727 goto init_map
; /* jump over SWAP_CONT_MAX checks */
3729 if (count
== (SWAP_MAP_MAX
| COUNT_CONTINUED
)) { /* incrementing */
3731 * Think of how you add 1 to 999
3733 while (*map
== (SWAP_CONT_MAX
| COUNT_CONTINUED
)) {
3735 page
= list_next_entry(page
, lru
);
3736 BUG_ON(page
== head
);
3737 map
= kmap_atomic(page
) + offset
;
3739 if (*map
== SWAP_CONT_MAX
) {
3741 page
= list_next_entry(page
, lru
);
3743 ret
= false; /* add count continuation */
3746 map
= kmap_atomic(page
) + offset
;
3747 init_map
: *map
= 0; /* we didn't zero the page */
3751 while ((page
= list_prev_entry(page
, lru
)) != head
) {
3752 map
= kmap_atomic(page
) + offset
;
3753 *map
= COUNT_CONTINUED
;
3756 ret
= true; /* incremented */
3758 } else { /* decrementing */
3760 * Think of how you subtract 1 from 1000
3762 BUG_ON(count
!= COUNT_CONTINUED
);
3763 while (*map
== COUNT_CONTINUED
) {
3765 page
= list_next_entry(page
, lru
);
3766 BUG_ON(page
== head
);
3767 map
= kmap_atomic(page
) + offset
;
3774 while ((page
= list_prev_entry(page
, lru
)) != head
) {
3775 map
= kmap_atomic(page
) + offset
;
3776 *map
= SWAP_CONT_MAX
| count
;
3777 count
= COUNT_CONTINUED
;
3780 ret
= count
== COUNT_CONTINUED
;
3783 spin_unlock(&si
->cont_lock
);
3788 * free_swap_count_continuations - swapoff free all the continuation pages
3789 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3791 static void free_swap_count_continuations(struct swap_info_struct
*si
)
3795 for (offset
= 0; offset
< si
->max
; offset
+= PAGE_SIZE
) {
3797 head
= vmalloc_to_page(si
->swap_map
+ offset
);
3798 if (page_private(head
)) {
3799 struct page
*page
, *next
;
3801 list_for_each_entry_safe(page
, next
, &head
->lru
, lru
) {
3802 list_del(&page
->lru
);
3809 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3810 void cgroup_throttle_swaprate(struct page
*page
, gfp_t gfp_mask
)
3812 struct swap_info_struct
*si
, *next
;
3813 int nid
= page_to_nid(page
);
3815 if (!(gfp_mask
& __GFP_IO
))
3818 if (!blk_cgroup_congested())
3822 * We've already scheduled a throttle, avoid taking the global swap
3825 if (current
->throttle_queue
)
3828 spin_lock(&swap_avail_lock
);
3829 plist_for_each_entry_safe(si
, next
, &swap_avail_heads
[nid
],
3832 blkcg_schedule_throttle(bdev_get_queue(si
->bdev
), true);
3836 spin_unlock(&swap_avail_lock
);
3840 static int __init
swapfile_init(void)
3844 swap_avail_heads
= kmalloc_array(nr_node_ids
, sizeof(struct plist_head
),
3846 if (!swap_avail_heads
) {
3847 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3852 plist_head_init(&swap_avail_heads
[nid
]);
3856 subsys_initcall(swapfile_init
);