4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
9 #include <linux/sched/mm.h>
10 #include <linux/sched/task.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mman.h>
13 #include <linux/slab.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/swap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/pagemap.h>
18 #include <linux/namei.h>
19 #include <linux/shmem_fs.h>
20 #include <linux/blkdev.h>
21 #include <linux/random.h>
22 #include <linux/writeback.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/init.h>
26 #include <linux/ksm.h>
27 #include <linux/rmap.h>
28 #include <linux/security.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mutex.h>
31 #include <linux/capability.h>
32 #include <linux/syscalls.h>
33 #include <linux/memcontrol.h>
34 #include <linux/poll.h>
35 #include <linux/oom.h>
36 #include <linux/frontswap.h>
37 #include <linux/swapfile.h>
38 #include <linux/export.h>
39 #include <linux/swap_slots.h>
40 #include <linux/sort.h>
42 #include <asm/pgtable.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 inline unsigned char swap_count(unsigned char ent
)
103 return ent
& ~SWAP_HAS_CACHE
; /* may include COUNT_CONTINUED flag */
106 /* Reclaim the swap entry anyway if possible */
107 #define TTRS_ANYWAY 0x1
109 * Reclaim the swap entry if there are no more mappings of the
112 #define TTRS_UNMAPPED 0x2
113 /* Reclaim the swap entry if swap is getting full*/
114 #define TTRS_FULL 0x4
116 /* returns 1 if swap entry is freed */
117 static int __try_to_reclaim_swap(struct swap_info_struct
*si
,
118 unsigned long offset
, unsigned long flags
)
120 swp_entry_t entry
= swp_entry(si
->type
, offset
);
124 page
= find_get_page(swap_address_space(entry
), offset
);
128 * When this function is called from scan_swap_map_slots() and it's
129 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
130 * here. We have to use trylock for avoiding deadlock. This is a special
131 * case and you should use try_to_free_swap() with explicit lock_page()
132 * in usual operations.
134 if (trylock_page(page
)) {
135 if ((flags
& TTRS_ANYWAY
) ||
136 ((flags
& TTRS_UNMAPPED
) && !page_mapped(page
)) ||
137 ((flags
& TTRS_FULL
) && mem_cgroup_swap_full(page
)))
138 ret
= try_to_free_swap(page
);
146 * swapon tell device that all the old swap contents can be discarded,
147 * to allow the swap device to optimize its wear-levelling.
149 static int discard_swap(struct swap_info_struct
*si
)
151 struct swap_extent
*se
;
152 sector_t start_block
;
156 /* Do not discard the swap header page! */
157 se
= &si
->first_swap_extent
;
158 start_block
= (se
->start_block
+ 1) << (PAGE_SHIFT
- 9);
159 nr_blocks
= ((sector_t
)se
->nr_pages
- 1) << (PAGE_SHIFT
- 9);
161 err
= blkdev_issue_discard(si
->bdev
, start_block
,
162 nr_blocks
, GFP_KERNEL
, 0);
168 list_for_each_entry(se
, &si
->first_swap_extent
.list
, list
) {
169 start_block
= se
->start_block
<< (PAGE_SHIFT
- 9);
170 nr_blocks
= (sector_t
)se
->nr_pages
<< (PAGE_SHIFT
- 9);
172 err
= blkdev_issue_discard(si
->bdev
, start_block
,
173 nr_blocks
, GFP_KERNEL
, 0);
179 return err
; /* That will often be -EOPNOTSUPP */
183 * swap allocation tell device that a cluster of swap can now be discarded,
184 * to allow the swap device to optimize its wear-levelling.
186 static void discard_swap_cluster(struct swap_info_struct
*si
,
187 pgoff_t start_page
, pgoff_t nr_pages
)
189 struct swap_extent
*se
= si
->curr_swap_extent
;
190 int found_extent
= 0;
193 if (se
->start_page
<= start_page
&&
194 start_page
< se
->start_page
+ se
->nr_pages
) {
195 pgoff_t offset
= start_page
- se
->start_page
;
196 sector_t start_block
= se
->start_block
+ offset
;
197 sector_t nr_blocks
= se
->nr_pages
- offset
;
199 if (nr_blocks
> nr_pages
)
200 nr_blocks
= nr_pages
;
201 start_page
+= nr_blocks
;
202 nr_pages
-= nr_blocks
;
205 si
->curr_swap_extent
= se
;
207 start_block
<<= PAGE_SHIFT
- 9;
208 nr_blocks
<<= PAGE_SHIFT
- 9;
209 if (blkdev_issue_discard(si
->bdev
, start_block
,
210 nr_blocks
, GFP_NOIO
, 0))
214 se
= list_next_entry(se
, list
);
218 #ifdef CONFIG_THP_SWAP
219 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
221 #define swap_entry_size(size) (size)
223 #define SWAPFILE_CLUSTER 256
226 * Define swap_entry_size() as constant to let compiler to optimize
227 * out some code if !CONFIG_THP_SWAP
229 #define swap_entry_size(size) 1
231 #define LATENCY_LIMIT 256
233 static inline void cluster_set_flag(struct swap_cluster_info
*info
,
239 static inline unsigned int cluster_count(struct swap_cluster_info
*info
)
244 static inline void cluster_set_count(struct swap_cluster_info
*info
,
250 static inline void cluster_set_count_flag(struct swap_cluster_info
*info
,
251 unsigned int c
, unsigned int f
)
257 static inline unsigned int cluster_next(struct swap_cluster_info
*info
)
262 static inline void cluster_set_next(struct swap_cluster_info
*info
,
268 static inline void cluster_set_next_flag(struct swap_cluster_info
*info
,
269 unsigned int n
, unsigned int f
)
275 static inline bool cluster_is_free(struct swap_cluster_info
*info
)
277 return info
->flags
& CLUSTER_FLAG_FREE
;
280 static inline bool cluster_is_null(struct swap_cluster_info
*info
)
282 return info
->flags
& CLUSTER_FLAG_NEXT_NULL
;
285 static inline void cluster_set_null(struct swap_cluster_info
*info
)
287 info
->flags
= CLUSTER_FLAG_NEXT_NULL
;
291 static inline bool cluster_is_huge(struct swap_cluster_info
*info
)
293 if (IS_ENABLED(CONFIG_THP_SWAP
))
294 return info
->flags
& CLUSTER_FLAG_HUGE
;
298 static inline void cluster_clear_huge(struct swap_cluster_info
*info
)
300 info
->flags
&= ~CLUSTER_FLAG_HUGE
;
303 static inline struct swap_cluster_info
*lock_cluster(struct swap_info_struct
*si
,
304 unsigned long offset
)
306 struct swap_cluster_info
*ci
;
308 ci
= si
->cluster_info
;
310 ci
+= offset
/ SWAPFILE_CLUSTER
;
311 spin_lock(&ci
->lock
);
316 static inline void unlock_cluster(struct swap_cluster_info
*ci
)
319 spin_unlock(&ci
->lock
);
323 * Determine the locking method in use for this device. Return
324 * swap_cluster_info if SSD-style cluster-based locking is in place.
326 static inline struct swap_cluster_info
*lock_cluster_or_swap_info(
327 struct swap_info_struct
*si
, unsigned long offset
)
329 struct swap_cluster_info
*ci
;
331 /* Try to use fine-grained SSD-style locking if available: */
332 ci
= lock_cluster(si
, offset
);
333 /* Otherwise, fall back to traditional, coarse locking: */
335 spin_lock(&si
->lock
);
340 static inline void unlock_cluster_or_swap_info(struct swap_info_struct
*si
,
341 struct swap_cluster_info
*ci
)
346 spin_unlock(&si
->lock
);
349 static inline bool cluster_list_empty(struct swap_cluster_list
*list
)
351 return cluster_is_null(&list
->head
);
354 static inline unsigned int cluster_list_first(struct swap_cluster_list
*list
)
356 return cluster_next(&list
->head
);
359 static void cluster_list_init(struct swap_cluster_list
*list
)
361 cluster_set_null(&list
->head
);
362 cluster_set_null(&list
->tail
);
365 static void cluster_list_add_tail(struct swap_cluster_list
*list
,
366 struct swap_cluster_info
*ci
,
369 if (cluster_list_empty(list
)) {
370 cluster_set_next_flag(&list
->head
, idx
, 0);
371 cluster_set_next_flag(&list
->tail
, idx
, 0);
373 struct swap_cluster_info
*ci_tail
;
374 unsigned int tail
= cluster_next(&list
->tail
);
377 * Nested cluster lock, but both cluster locks are
378 * only acquired when we held swap_info_struct->lock
381 spin_lock_nested(&ci_tail
->lock
, SINGLE_DEPTH_NESTING
);
382 cluster_set_next(ci_tail
, idx
);
383 spin_unlock(&ci_tail
->lock
);
384 cluster_set_next_flag(&list
->tail
, idx
, 0);
388 static unsigned int cluster_list_del_first(struct swap_cluster_list
*list
,
389 struct swap_cluster_info
*ci
)
393 idx
= cluster_next(&list
->head
);
394 if (cluster_next(&list
->tail
) == idx
) {
395 cluster_set_null(&list
->head
);
396 cluster_set_null(&list
->tail
);
398 cluster_set_next_flag(&list
->head
,
399 cluster_next(&ci
[idx
]), 0);
404 /* Add a cluster to discard list and schedule it to do discard */
405 static void swap_cluster_schedule_discard(struct swap_info_struct
*si
,
409 * If scan_swap_map() can't find a free cluster, it will check
410 * si->swap_map directly. To make sure the discarding cluster isn't
411 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
412 * will be cleared after discard
414 memset(si
->swap_map
+ idx
* SWAPFILE_CLUSTER
,
415 SWAP_MAP_BAD
, SWAPFILE_CLUSTER
);
417 cluster_list_add_tail(&si
->discard_clusters
, si
->cluster_info
, idx
);
419 schedule_work(&si
->discard_work
);
422 static void __free_cluster(struct swap_info_struct
*si
, unsigned long idx
)
424 struct swap_cluster_info
*ci
= si
->cluster_info
;
426 cluster_set_flag(ci
+ idx
, CLUSTER_FLAG_FREE
);
427 cluster_list_add_tail(&si
->free_clusters
, ci
, idx
);
431 * Doing discard actually. After a cluster discard is finished, the cluster
432 * will be added to free cluster list. caller should hold si->lock.
434 static void swap_do_scheduled_discard(struct swap_info_struct
*si
)
436 struct swap_cluster_info
*info
, *ci
;
439 info
= si
->cluster_info
;
441 while (!cluster_list_empty(&si
->discard_clusters
)) {
442 idx
= cluster_list_del_first(&si
->discard_clusters
, info
);
443 spin_unlock(&si
->lock
);
445 discard_swap_cluster(si
, idx
* SWAPFILE_CLUSTER
,
448 spin_lock(&si
->lock
);
449 ci
= lock_cluster(si
, idx
* SWAPFILE_CLUSTER
);
450 __free_cluster(si
, idx
);
451 memset(si
->swap_map
+ idx
* SWAPFILE_CLUSTER
,
452 0, SWAPFILE_CLUSTER
);
457 static void swap_discard_work(struct work_struct
*work
)
459 struct swap_info_struct
*si
;
461 si
= container_of(work
, struct swap_info_struct
, discard_work
);
463 spin_lock(&si
->lock
);
464 swap_do_scheduled_discard(si
);
465 spin_unlock(&si
->lock
);
468 static void alloc_cluster(struct swap_info_struct
*si
, unsigned long idx
)
470 struct swap_cluster_info
*ci
= si
->cluster_info
;
472 VM_BUG_ON(cluster_list_first(&si
->free_clusters
) != idx
);
473 cluster_list_del_first(&si
->free_clusters
, ci
);
474 cluster_set_count_flag(ci
+ idx
, 0, 0);
477 static void free_cluster(struct swap_info_struct
*si
, unsigned long idx
)
479 struct swap_cluster_info
*ci
= si
->cluster_info
+ idx
;
481 VM_BUG_ON(cluster_count(ci
) != 0);
483 * If the swap is discardable, prepare discard the cluster
484 * instead of free it immediately. The cluster will be freed
487 if ((si
->flags
& (SWP_WRITEOK
| SWP_PAGE_DISCARD
)) ==
488 (SWP_WRITEOK
| SWP_PAGE_DISCARD
)) {
489 swap_cluster_schedule_discard(si
, idx
);
493 __free_cluster(si
, idx
);
497 * The cluster corresponding to page_nr will be used. The cluster will be
498 * removed from free cluster list and its usage counter will be increased.
500 static void inc_cluster_info_page(struct swap_info_struct
*p
,
501 struct swap_cluster_info
*cluster_info
, unsigned long page_nr
)
503 unsigned long idx
= page_nr
/ SWAPFILE_CLUSTER
;
507 if (cluster_is_free(&cluster_info
[idx
]))
508 alloc_cluster(p
, idx
);
510 VM_BUG_ON(cluster_count(&cluster_info
[idx
]) >= SWAPFILE_CLUSTER
);
511 cluster_set_count(&cluster_info
[idx
],
512 cluster_count(&cluster_info
[idx
]) + 1);
516 * The cluster corresponding to page_nr decreases one usage. If the usage
517 * counter becomes 0, which means no page in the cluster is in using, we can
518 * optionally discard the cluster and add it to free cluster list.
520 static void dec_cluster_info_page(struct swap_info_struct
*p
,
521 struct swap_cluster_info
*cluster_info
, unsigned long page_nr
)
523 unsigned long idx
= page_nr
/ SWAPFILE_CLUSTER
;
528 VM_BUG_ON(cluster_count(&cluster_info
[idx
]) == 0);
529 cluster_set_count(&cluster_info
[idx
],
530 cluster_count(&cluster_info
[idx
]) - 1);
532 if (cluster_count(&cluster_info
[idx
]) == 0)
533 free_cluster(p
, idx
);
537 * It's possible scan_swap_map() uses a free cluster in the middle of free
538 * cluster list. Avoiding such abuse to avoid list corruption.
541 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct
*si
,
542 unsigned long offset
)
544 struct percpu_cluster
*percpu_cluster
;
547 offset
/= SWAPFILE_CLUSTER
;
548 conflict
= !cluster_list_empty(&si
->free_clusters
) &&
549 offset
!= cluster_list_first(&si
->free_clusters
) &&
550 cluster_is_free(&si
->cluster_info
[offset
]);
555 percpu_cluster
= this_cpu_ptr(si
->percpu_cluster
);
556 cluster_set_null(&percpu_cluster
->index
);
561 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
562 * might involve allocating a new cluster for current CPU too.
564 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct
*si
,
565 unsigned long *offset
, unsigned long *scan_base
)
567 struct percpu_cluster
*cluster
;
568 struct swap_cluster_info
*ci
;
570 unsigned long tmp
, max
;
573 cluster
= this_cpu_ptr(si
->percpu_cluster
);
574 if (cluster_is_null(&cluster
->index
)) {
575 if (!cluster_list_empty(&si
->free_clusters
)) {
576 cluster
->index
= si
->free_clusters
.head
;
577 cluster
->next
= cluster_next(&cluster
->index
) *
579 } else if (!cluster_list_empty(&si
->discard_clusters
)) {
581 * we don't have free cluster but have some clusters in
582 * discarding, do discard now and reclaim them
584 swap_do_scheduled_discard(si
);
585 *scan_base
= *offset
= si
->cluster_next
;
594 * Other CPUs can use our cluster if they can't find a free cluster,
595 * check if there is still free entry in the cluster
598 max
= min_t(unsigned long, si
->max
,
599 (cluster_next(&cluster
->index
) + 1) * SWAPFILE_CLUSTER
);
601 cluster_set_null(&cluster
->index
);
604 ci
= lock_cluster(si
, tmp
);
606 if (!si
->swap_map
[tmp
]) {
614 cluster_set_null(&cluster
->index
);
617 cluster
->next
= tmp
+ 1;
623 static void __del_from_avail_list(struct swap_info_struct
*p
)
628 plist_del(&p
->avail_lists
[nid
], &swap_avail_heads
[nid
]);
631 static void del_from_avail_list(struct swap_info_struct
*p
)
633 spin_lock(&swap_avail_lock
);
634 __del_from_avail_list(p
);
635 spin_unlock(&swap_avail_lock
);
638 static void swap_range_alloc(struct swap_info_struct
*si
, unsigned long offset
,
639 unsigned int nr_entries
)
641 unsigned int end
= offset
+ nr_entries
- 1;
643 if (offset
== si
->lowest_bit
)
644 si
->lowest_bit
+= nr_entries
;
645 if (end
== si
->highest_bit
)
646 si
->highest_bit
-= nr_entries
;
647 si
->inuse_pages
+= nr_entries
;
648 if (si
->inuse_pages
== si
->pages
) {
649 si
->lowest_bit
= si
->max
;
651 del_from_avail_list(si
);
655 static void add_to_avail_list(struct swap_info_struct
*p
)
659 spin_lock(&swap_avail_lock
);
661 WARN_ON(!plist_node_empty(&p
->avail_lists
[nid
]));
662 plist_add(&p
->avail_lists
[nid
], &swap_avail_heads
[nid
]);
664 spin_unlock(&swap_avail_lock
);
667 static void swap_range_free(struct swap_info_struct
*si
, unsigned long offset
,
668 unsigned int nr_entries
)
670 unsigned long end
= offset
+ nr_entries
- 1;
671 void (*swap_slot_free_notify
)(struct block_device
*, unsigned long);
673 if (offset
< si
->lowest_bit
)
674 si
->lowest_bit
= offset
;
675 if (end
> si
->highest_bit
) {
676 bool was_full
= !si
->highest_bit
;
678 si
->highest_bit
= end
;
679 if (was_full
&& (si
->flags
& SWP_WRITEOK
))
680 add_to_avail_list(si
);
682 atomic_long_add(nr_entries
, &nr_swap_pages
);
683 si
->inuse_pages
-= nr_entries
;
684 if (si
->flags
& SWP_BLKDEV
)
685 swap_slot_free_notify
=
686 si
->bdev
->bd_disk
->fops
->swap_slot_free_notify
;
688 swap_slot_free_notify
= NULL
;
689 while (offset
<= end
) {
690 frontswap_invalidate_page(si
->type
, offset
);
691 if (swap_slot_free_notify
)
692 swap_slot_free_notify(si
->bdev
, offset
);
697 static int scan_swap_map_slots(struct swap_info_struct
*si
,
698 unsigned char usage
, int nr
,
701 struct swap_cluster_info
*ci
;
702 unsigned long offset
;
703 unsigned long scan_base
;
704 unsigned long last_in_cluster
= 0;
705 int latency_ration
= LATENCY_LIMIT
;
712 * We try to cluster swap pages by allocating them sequentially
713 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
714 * way, however, we resort to first-free allocation, starting
715 * a new cluster. This prevents us from scattering swap pages
716 * all over the entire swap partition, so that we reduce
717 * overall disk seek times between swap pages. -- sct
718 * But we do now try to find an empty cluster. -Andrea
719 * And we let swap pages go all over an SSD partition. Hugh
722 si
->flags
+= SWP_SCANNING
;
723 scan_base
= offset
= si
->cluster_next
;
726 if (si
->cluster_info
) {
727 if (scan_swap_map_try_ssd_cluster(si
, &offset
, &scan_base
))
733 if (unlikely(!si
->cluster_nr
--)) {
734 if (si
->pages
- si
->inuse_pages
< SWAPFILE_CLUSTER
) {
735 si
->cluster_nr
= SWAPFILE_CLUSTER
- 1;
739 spin_unlock(&si
->lock
);
742 * If seek is expensive, start searching for new cluster from
743 * start of partition, to minimize the span of allocated swap.
744 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
745 * case, just handled by scan_swap_map_try_ssd_cluster() above.
747 scan_base
= offset
= si
->lowest_bit
;
748 last_in_cluster
= offset
+ SWAPFILE_CLUSTER
- 1;
750 /* Locate the first empty (unaligned) cluster */
751 for (; last_in_cluster
<= si
->highest_bit
; offset
++) {
752 if (si
->swap_map
[offset
])
753 last_in_cluster
= offset
+ SWAPFILE_CLUSTER
;
754 else if (offset
== last_in_cluster
) {
755 spin_lock(&si
->lock
);
756 offset
-= SWAPFILE_CLUSTER
- 1;
757 si
->cluster_next
= offset
;
758 si
->cluster_nr
= SWAPFILE_CLUSTER
- 1;
761 if (unlikely(--latency_ration
< 0)) {
763 latency_ration
= LATENCY_LIMIT
;
768 spin_lock(&si
->lock
);
769 si
->cluster_nr
= SWAPFILE_CLUSTER
- 1;
773 if (si
->cluster_info
) {
774 while (scan_swap_map_ssd_cluster_conflict(si
, offset
)) {
775 /* take a break if we already got some slots */
778 if (!scan_swap_map_try_ssd_cluster(si
, &offset
,
783 if (!(si
->flags
& SWP_WRITEOK
))
785 if (!si
->highest_bit
)
787 if (offset
> si
->highest_bit
)
788 scan_base
= offset
= si
->lowest_bit
;
790 ci
= lock_cluster(si
, offset
);
791 /* reuse swap entry of cache-only swap if not busy. */
792 if (vm_swap_full() && si
->swap_map
[offset
] == SWAP_HAS_CACHE
) {
795 spin_unlock(&si
->lock
);
796 swap_was_freed
= __try_to_reclaim_swap(si
, offset
, TTRS_ANYWAY
);
797 spin_lock(&si
->lock
);
798 /* entry was freed successfully, try to use this again */
801 goto scan
; /* check next one */
804 if (si
->swap_map
[offset
]) {
811 si
->swap_map
[offset
] = usage
;
812 inc_cluster_info_page(si
, si
->cluster_info
, offset
);
815 swap_range_alloc(si
, offset
, 1);
816 si
->cluster_next
= offset
+ 1;
817 slots
[n_ret
++] = swp_entry(si
->type
, offset
);
819 /* got enough slots or reach max slots? */
820 if ((n_ret
== nr
) || (offset
>= si
->highest_bit
))
823 /* search for next available slot */
825 /* time to take a break? */
826 if (unlikely(--latency_ration
< 0)) {
829 spin_unlock(&si
->lock
);
831 spin_lock(&si
->lock
);
832 latency_ration
= LATENCY_LIMIT
;
835 /* try to get more slots in cluster */
836 if (si
->cluster_info
) {
837 if (scan_swap_map_try_ssd_cluster(si
, &offset
, &scan_base
))
845 /* non-ssd case, still more slots in cluster? */
846 if (si
->cluster_nr
&& !si
->swap_map
[offset
]) {
852 si
->flags
-= SWP_SCANNING
;
856 spin_unlock(&si
->lock
);
857 while (++offset
<= si
->highest_bit
) {
858 if (!si
->swap_map
[offset
]) {
859 spin_lock(&si
->lock
);
862 if (vm_swap_full() && si
->swap_map
[offset
] == SWAP_HAS_CACHE
) {
863 spin_lock(&si
->lock
);
866 if (unlikely(--latency_ration
< 0)) {
868 latency_ration
= LATENCY_LIMIT
;
871 offset
= si
->lowest_bit
;
872 while (offset
< scan_base
) {
873 if (!si
->swap_map
[offset
]) {
874 spin_lock(&si
->lock
);
877 if (vm_swap_full() && si
->swap_map
[offset
] == SWAP_HAS_CACHE
) {
878 spin_lock(&si
->lock
);
881 if (unlikely(--latency_ration
< 0)) {
883 latency_ration
= LATENCY_LIMIT
;
887 spin_lock(&si
->lock
);
890 si
->flags
-= SWP_SCANNING
;
894 static int swap_alloc_cluster(struct swap_info_struct
*si
, swp_entry_t
*slot
)
897 struct swap_cluster_info
*ci
;
898 unsigned long offset
, i
;
902 * Should not even be attempting cluster allocations when huge
903 * page swap is disabled. Warn and fail the allocation.
905 if (!IS_ENABLED(CONFIG_THP_SWAP
)) {
910 if (cluster_list_empty(&si
->free_clusters
))
913 idx
= cluster_list_first(&si
->free_clusters
);
914 offset
= idx
* SWAPFILE_CLUSTER
;
915 ci
= lock_cluster(si
, offset
);
916 alloc_cluster(si
, idx
);
917 cluster_set_count_flag(ci
, SWAPFILE_CLUSTER
, CLUSTER_FLAG_HUGE
);
919 map
= si
->swap_map
+ offset
;
920 for (i
= 0; i
< SWAPFILE_CLUSTER
; i
++)
921 map
[i
] = SWAP_HAS_CACHE
;
923 swap_range_alloc(si
, offset
, SWAPFILE_CLUSTER
);
924 *slot
= swp_entry(si
->type
, offset
);
929 static void swap_free_cluster(struct swap_info_struct
*si
, unsigned long idx
)
931 unsigned long offset
= idx
* SWAPFILE_CLUSTER
;
932 struct swap_cluster_info
*ci
;
934 ci
= lock_cluster(si
, offset
);
935 memset(si
->swap_map
+ offset
, 0, SWAPFILE_CLUSTER
);
936 cluster_set_count_flag(ci
, 0, 0);
937 free_cluster(si
, idx
);
939 swap_range_free(si
, offset
, SWAPFILE_CLUSTER
);
942 static unsigned long scan_swap_map(struct swap_info_struct
*si
,
948 n_ret
= scan_swap_map_slots(si
, usage
, 1, &entry
);
951 return swp_offset(entry
);
957 int get_swap_pages(int n_goal
, swp_entry_t swp_entries
[], int entry_size
)
959 unsigned long size
= swap_entry_size(entry_size
);
960 struct swap_info_struct
*si
, *next
;
965 /* Only single cluster request supported */
966 WARN_ON_ONCE(n_goal
> 1 && size
== SWAPFILE_CLUSTER
);
968 avail_pgs
= atomic_long_read(&nr_swap_pages
) / size
;
972 if (n_goal
> SWAP_BATCH
)
975 if (n_goal
> avail_pgs
)
978 atomic_long_sub(n_goal
* size
, &nr_swap_pages
);
980 spin_lock(&swap_avail_lock
);
983 node
= numa_node_id();
984 plist_for_each_entry_safe(si
, next
, &swap_avail_heads
[node
], avail_lists
[node
]) {
985 /* requeue si to after same-priority siblings */
986 plist_requeue(&si
->avail_lists
[node
], &swap_avail_heads
[node
]);
987 spin_unlock(&swap_avail_lock
);
988 spin_lock(&si
->lock
);
989 if (!si
->highest_bit
|| !(si
->flags
& SWP_WRITEOK
)) {
990 spin_lock(&swap_avail_lock
);
991 if (plist_node_empty(&si
->avail_lists
[node
])) {
992 spin_unlock(&si
->lock
);
995 WARN(!si
->highest_bit
,
996 "swap_info %d in list but !highest_bit\n",
998 WARN(!(si
->flags
& SWP_WRITEOK
),
999 "swap_info %d in list but !SWP_WRITEOK\n",
1001 __del_from_avail_list(si
);
1002 spin_unlock(&si
->lock
);
1005 if (size
== SWAPFILE_CLUSTER
) {
1006 if (!(si
->flags
& SWP_FS
))
1007 n_ret
= swap_alloc_cluster(si
, swp_entries
);
1009 n_ret
= scan_swap_map_slots(si
, SWAP_HAS_CACHE
,
1010 n_goal
, swp_entries
);
1011 spin_unlock(&si
->lock
);
1012 if (n_ret
|| size
== SWAPFILE_CLUSTER
)
1014 pr_debug("scan_swap_map of si %d failed to find offset\n",
1017 spin_lock(&swap_avail_lock
);
1020 * if we got here, it's likely that si was almost full before,
1021 * and since scan_swap_map() can drop the si->lock, multiple
1022 * callers probably all tried to get a page from the same si
1023 * and it filled up before we could get one; or, the si filled
1024 * up between us dropping swap_avail_lock and taking si->lock.
1025 * Since we dropped the swap_avail_lock, the swap_avail_head
1026 * list may have been modified; so if next is still in the
1027 * swap_avail_head list then try it, otherwise start over
1028 * if we have not gotten any slots.
1030 if (plist_node_empty(&next
->avail_lists
[node
]))
1034 spin_unlock(&swap_avail_lock
);
1038 atomic_long_add((long)(n_goal
- n_ret
) * size
,
1044 /* The only caller of this function is now suspend routine */
1045 swp_entry_t
get_swap_page_of_type(int type
)
1047 struct swap_info_struct
*si
;
1050 si
= swap_info
[type
];
1051 spin_lock(&si
->lock
);
1052 if (si
&& (si
->flags
& SWP_WRITEOK
)) {
1053 atomic_long_dec(&nr_swap_pages
);
1054 /* This is called for allocating swap entry, not cache */
1055 offset
= scan_swap_map(si
, 1);
1057 spin_unlock(&si
->lock
);
1058 return swp_entry(type
, offset
);
1060 atomic_long_inc(&nr_swap_pages
);
1062 spin_unlock(&si
->lock
);
1063 return (swp_entry_t
) {0};
1066 static struct swap_info_struct
*__swap_info_get(swp_entry_t entry
)
1068 struct swap_info_struct
*p
;
1069 unsigned long offset
, type
;
1073 type
= swp_type(entry
);
1074 if (type
>= nr_swapfiles
)
1076 p
= swap_info
[type
];
1077 if (!(p
->flags
& SWP_USED
))
1079 offset
= swp_offset(entry
);
1080 if (offset
>= p
->max
)
1085 pr_err("swap_info_get: %s%08lx\n", Bad_offset
, entry
.val
);
1088 pr_err("swap_info_get: %s%08lx\n", Unused_file
, entry
.val
);
1091 pr_err("swap_info_get: %s%08lx\n", Bad_file
, entry
.val
);
1096 static struct swap_info_struct
*_swap_info_get(swp_entry_t entry
)
1098 struct swap_info_struct
*p
;
1100 p
= __swap_info_get(entry
);
1103 if (!p
->swap_map
[swp_offset(entry
)])
1108 pr_err("swap_info_get: %s%08lx\n", Unused_offset
, entry
.val
);
1114 static struct swap_info_struct
*swap_info_get(swp_entry_t entry
)
1116 struct swap_info_struct
*p
;
1118 p
= _swap_info_get(entry
);
1120 spin_lock(&p
->lock
);
1124 static struct swap_info_struct
*swap_info_get_cont(swp_entry_t entry
,
1125 struct swap_info_struct
*q
)
1127 struct swap_info_struct
*p
;
1129 p
= _swap_info_get(entry
);
1133 spin_unlock(&q
->lock
);
1135 spin_lock(&p
->lock
);
1140 static unsigned char __swap_entry_free_locked(struct swap_info_struct
*p
,
1141 unsigned long offset
,
1142 unsigned char usage
)
1144 unsigned char count
;
1145 unsigned char has_cache
;
1147 count
= p
->swap_map
[offset
];
1149 has_cache
= count
& SWAP_HAS_CACHE
;
1150 count
&= ~SWAP_HAS_CACHE
;
1152 if (usage
== SWAP_HAS_CACHE
) {
1153 VM_BUG_ON(!has_cache
);
1155 } else if (count
== SWAP_MAP_SHMEM
) {
1157 * Or we could insist on shmem.c using a special
1158 * swap_shmem_free() and free_shmem_swap_and_cache()...
1161 } else if ((count
& ~COUNT_CONTINUED
) <= SWAP_MAP_MAX
) {
1162 if (count
== COUNT_CONTINUED
) {
1163 if (swap_count_continued(p
, offset
, count
))
1164 count
= SWAP_MAP_MAX
| COUNT_CONTINUED
;
1166 count
= SWAP_MAP_MAX
;
1171 usage
= count
| has_cache
;
1172 p
->swap_map
[offset
] = usage
? : SWAP_HAS_CACHE
;
1177 static unsigned char __swap_entry_free(struct swap_info_struct
*p
,
1178 swp_entry_t entry
, unsigned char usage
)
1180 struct swap_cluster_info
*ci
;
1181 unsigned long offset
= swp_offset(entry
);
1183 ci
= lock_cluster_or_swap_info(p
, offset
);
1184 usage
= __swap_entry_free_locked(p
, offset
, usage
);
1185 unlock_cluster_or_swap_info(p
, ci
);
1187 free_swap_slot(entry
);
1192 static void swap_entry_free(struct swap_info_struct
*p
, swp_entry_t entry
)
1194 struct swap_cluster_info
*ci
;
1195 unsigned long offset
= swp_offset(entry
);
1196 unsigned char count
;
1198 ci
= lock_cluster(p
, offset
);
1199 count
= p
->swap_map
[offset
];
1200 VM_BUG_ON(count
!= SWAP_HAS_CACHE
);
1201 p
->swap_map
[offset
] = 0;
1202 dec_cluster_info_page(p
, p
->cluster_info
, offset
);
1205 mem_cgroup_uncharge_swap(entry
, 1);
1206 swap_range_free(p
, offset
, 1);
1210 * Caller has made sure that the swap device corresponding to entry
1211 * is still around or has not been recycled.
1213 void swap_free(swp_entry_t entry
)
1215 struct swap_info_struct
*p
;
1217 p
= _swap_info_get(entry
);
1219 __swap_entry_free(p
, entry
, 1);
1223 * Called after dropping swapcache to decrease refcnt to swap entries.
1225 void put_swap_page(struct page
*page
, swp_entry_t entry
)
1227 unsigned long offset
= swp_offset(entry
);
1228 unsigned long idx
= offset
/ SWAPFILE_CLUSTER
;
1229 struct swap_cluster_info
*ci
;
1230 struct swap_info_struct
*si
;
1232 unsigned int i
, free_entries
= 0;
1234 int size
= swap_entry_size(hpage_nr_pages(page
));
1236 si
= _swap_info_get(entry
);
1240 ci
= lock_cluster_or_swap_info(si
, offset
);
1241 if (size
== SWAPFILE_CLUSTER
) {
1242 VM_BUG_ON(!cluster_is_huge(ci
));
1243 map
= si
->swap_map
+ offset
;
1244 for (i
= 0; i
< SWAPFILE_CLUSTER
; i
++) {
1246 VM_BUG_ON(!(val
& SWAP_HAS_CACHE
));
1247 if (val
== SWAP_HAS_CACHE
)
1250 cluster_clear_huge(ci
);
1251 if (free_entries
== SWAPFILE_CLUSTER
) {
1252 unlock_cluster_or_swap_info(si
, ci
);
1253 spin_lock(&si
->lock
);
1254 mem_cgroup_uncharge_swap(entry
, SWAPFILE_CLUSTER
);
1255 swap_free_cluster(si
, idx
);
1256 spin_unlock(&si
->lock
);
1260 for (i
= 0; i
< size
; i
++, entry
.val
++) {
1261 if (!__swap_entry_free_locked(si
, offset
+ i
, SWAP_HAS_CACHE
)) {
1262 unlock_cluster_or_swap_info(si
, ci
);
1263 free_swap_slot(entry
);
1266 lock_cluster_or_swap_info(si
, offset
);
1269 unlock_cluster_or_swap_info(si
, ci
);
1272 #ifdef CONFIG_THP_SWAP
1273 int split_swap_cluster(swp_entry_t entry
)
1275 struct swap_info_struct
*si
;
1276 struct swap_cluster_info
*ci
;
1277 unsigned long offset
= swp_offset(entry
);
1279 si
= _swap_info_get(entry
);
1282 ci
= lock_cluster(si
, offset
);
1283 cluster_clear_huge(ci
);
1289 static int swp_entry_cmp(const void *ent1
, const void *ent2
)
1291 const swp_entry_t
*e1
= ent1
, *e2
= ent2
;
1293 return (int)swp_type(*e1
) - (int)swp_type(*e2
);
1296 void swapcache_free_entries(swp_entry_t
*entries
, int n
)
1298 struct swap_info_struct
*p
, *prev
;
1308 * Sort swap entries by swap device, so each lock is only taken once.
1309 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1310 * so low that it isn't necessary to optimize further.
1312 if (nr_swapfiles
> 1)
1313 sort(entries
, n
, sizeof(entries
[0]), swp_entry_cmp
, NULL
);
1314 for (i
= 0; i
< n
; ++i
) {
1315 p
= swap_info_get_cont(entries
[i
], prev
);
1317 swap_entry_free(p
, entries
[i
]);
1321 spin_unlock(&p
->lock
);
1325 * How many references to page are currently swapped out?
1326 * This does not give an exact answer when swap count is continued,
1327 * but does include the high COUNT_CONTINUED flag to allow for that.
1329 int page_swapcount(struct page
*page
)
1332 struct swap_info_struct
*p
;
1333 struct swap_cluster_info
*ci
;
1335 unsigned long offset
;
1337 entry
.val
= page_private(page
);
1338 p
= _swap_info_get(entry
);
1340 offset
= swp_offset(entry
);
1341 ci
= lock_cluster_or_swap_info(p
, offset
);
1342 count
= swap_count(p
->swap_map
[offset
]);
1343 unlock_cluster_or_swap_info(p
, ci
);
1348 int __swap_count(struct swap_info_struct
*si
, swp_entry_t entry
)
1350 pgoff_t offset
= swp_offset(entry
);
1352 return swap_count(si
->swap_map
[offset
]);
1355 static int swap_swapcount(struct swap_info_struct
*si
, swp_entry_t entry
)
1358 pgoff_t offset
= swp_offset(entry
);
1359 struct swap_cluster_info
*ci
;
1361 ci
= lock_cluster_or_swap_info(si
, offset
);
1362 count
= swap_count(si
->swap_map
[offset
]);
1363 unlock_cluster_or_swap_info(si
, ci
);
1368 * How many references to @entry are currently swapped out?
1369 * This does not give an exact answer when swap count is continued,
1370 * but does include the high COUNT_CONTINUED flag to allow for that.
1372 int __swp_swapcount(swp_entry_t entry
)
1375 struct swap_info_struct
*si
;
1377 si
= __swap_info_get(entry
);
1379 count
= swap_swapcount(si
, entry
);
1384 * How many references to @entry are currently swapped out?
1385 * This considers COUNT_CONTINUED so it returns exact answer.
1387 int swp_swapcount(swp_entry_t entry
)
1389 int count
, tmp_count
, n
;
1390 struct swap_info_struct
*p
;
1391 struct swap_cluster_info
*ci
;
1396 p
= _swap_info_get(entry
);
1400 offset
= swp_offset(entry
);
1402 ci
= lock_cluster_or_swap_info(p
, offset
);
1404 count
= swap_count(p
->swap_map
[offset
]);
1405 if (!(count
& COUNT_CONTINUED
))
1408 count
&= ~COUNT_CONTINUED
;
1409 n
= SWAP_MAP_MAX
+ 1;
1411 page
= vmalloc_to_page(p
->swap_map
+ offset
);
1412 offset
&= ~PAGE_MASK
;
1413 VM_BUG_ON(page_private(page
) != SWP_CONTINUED
);
1416 page
= list_next_entry(page
, lru
);
1417 map
= kmap_atomic(page
);
1418 tmp_count
= map
[offset
];
1421 count
+= (tmp_count
& ~COUNT_CONTINUED
) * n
;
1422 n
*= (SWAP_CONT_MAX
+ 1);
1423 } while (tmp_count
& COUNT_CONTINUED
);
1425 unlock_cluster_or_swap_info(p
, ci
);
1429 static bool swap_page_trans_huge_swapped(struct swap_info_struct
*si
,
1432 struct swap_cluster_info
*ci
;
1433 unsigned char *map
= si
->swap_map
;
1434 unsigned long roffset
= swp_offset(entry
);
1435 unsigned long offset
= round_down(roffset
, SWAPFILE_CLUSTER
);
1439 ci
= lock_cluster_or_swap_info(si
, offset
);
1440 if (!ci
|| !cluster_is_huge(ci
)) {
1441 if (swap_count(map
[roffset
]))
1445 for (i
= 0; i
< SWAPFILE_CLUSTER
; i
++) {
1446 if (swap_count(map
[offset
+ i
])) {
1452 unlock_cluster_or_swap_info(si
, ci
);
1456 static bool page_swapped(struct page
*page
)
1459 struct swap_info_struct
*si
;
1461 if (!IS_ENABLED(CONFIG_THP_SWAP
) || likely(!PageTransCompound(page
)))
1462 return page_swapcount(page
) != 0;
1464 page
= compound_head(page
);
1465 entry
.val
= page_private(page
);
1466 si
= _swap_info_get(entry
);
1468 return swap_page_trans_huge_swapped(si
, entry
);
1472 static int page_trans_huge_map_swapcount(struct page
*page
, int *total_mapcount
,
1473 int *total_swapcount
)
1475 int i
, map_swapcount
, _total_mapcount
, _total_swapcount
;
1476 unsigned long offset
= 0;
1477 struct swap_info_struct
*si
;
1478 struct swap_cluster_info
*ci
= NULL
;
1479 unsigned char *map
= NULL
;
1480 int mapcount
, swapcount
= 0;
1482 /* hugetlbfs shouldn't call it */
1483 VM_BUG_ON_PAGE(PageHuge(page
), page
);
1485 if (!IS_ENABLED(CONFIG_THP_SWAP
) || likely(!PageTransCompound(page
))) {
1486 mapcount
= page_trans_huge_mapcount(page
, total_mapcount
);
1487 if (PageSwapCache(page
))
1488 swapcount
= page_swapcount(page
);
1489 if (total_swapcount
)
1490 *total_swapcount
= swapcount
;
1491 return mapcount
+ swapcount
;
1494 page
= compound_head(page
);
1496 _total_mapcount
= _total_swapcount
= map_swapcount
= 0;
1497 if (PageSwapCache(page
)) {
1500 entry
.val
= page_private(page
);
1501 si
= _swap_info_get(entry
);
1504 offset
= swp_offset(entry
);
1508 ci
= lock_cluster(si
, offset
);
1509 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1510 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
1511 _total_mapcount
+= mapcount
;
1513 swapcount
= swap_count(map
[offset
+ i
]);
1514 _total_swapcount
+= swapcount
;
1516 map_swapcount
= max(map_swapcount
, mapcount
+ swapcount
);
1519 if (PageDoubleMap(page
)) {
1521 _total_mapcount
-= HPAGE_PMD_NR
;
1523 mapcount
= compound_mapcount(page
);
1524 map_swapcount
+= mapcount
;
1525 _total_mapcount
+= mapcount
;
1527 *total_mapcount
= _total_mapcount
;
1528 if (total_swapcount
)
1529 *total_swapcount
= _total_swapcount
;
1531 return map_swapcount
;
1535 * We can write to an anon page without COW if there are no other references
1536 * to it. And as a side-effect, free up its swap: because the old content
1537 * on disk will never be read, and seeking back there to write new content
1538 * later would only waste time away from clustering.
1540 * NOTE: total_map_swapcount should not be relied upon by the caller if
1541 * reuse_swap_page() returns false, but it may be always overwritten
1542 * (see the other implementation for CONFIG_SWAP=n).
1544 bool reuse_swap_page(struct page
*page
, int *total_map_swapcount
)
1546 int count
, total_mapcount
, total_swapcount
;
1548 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1549 if (unlikely(PageKsm(page
)))
1551 count
= page_trans_huge_map_swapcount(page
, &total_mapcount
,
1553 if (total_map_swapcount
)
1554 *total_map_swapcount
= total_mapcount
+ total_swapcount
;
1555 if (count
== 1 && PageSwapCache(page
) &&
1556 (likely(!PageTransCompound(page
)) ||
1557 /* The remaining swap count will be freed soon */
1558 total_swapcount
== page_swapcount(page
))) {
1559 if (!PageWriteback(page
)) {
1560 page
= compound_head(page
);
1561 delete_from_swap_cache(page
);
1565 struct swap_info_struct
*p
;
1567 entry
.val
= page_private(page
);
1568 p
= swap_info_get(entry
);
1569 if (p
->flags
& SWP_STABLE_WRITES
) {
1570 spin_unlock(&p
->lock
);
1573 spin_unlock(&p
->lock
);
1581 * If swap is getting full, or if there are no more mappings of this page,
1582 * then try_to_free_swap is called to free its swap space.
1584 int try_to_free_swap(struct page
*page
)
1586 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1588 if (!PageSwapCache(page
))
1590 if (PageWriteback(page
))
1592 if (page_swapped(page
))
1596 * Once hibernation has begun to create its image of memory,
1597 * there's a danger that one of the calls to try_to_free_swap()
1598 * - most probably a call from __try_to_reclaim_swap() while
1599 * hibernation is allocating its own swap pages for the image,
1600 * but conceivably even a call from memory reclaim - will free
1601 * the swap from a page which has already been recorded in the
1602 * image as a clean swapcache page, and then reuse its swap for
1603 * another page of the image. On waking from hibernation, the
1604 * original page might be freed under memory pressure, then
1605 * later read back in from swap, now with the wrong data.
1607 * Hibernation suspends storage while it is writing the image
1608 * to disk so check that here.
1610 if (pm_suspended_storage())
1613 page
= compound_head(page
);
1614 delete_from_swap_cache(page
);
1620 * Free the swap entry like above, but also try to
1621 * free the page cache entry if it is the last user.
1623 int free_swap_and_cache(swp_entry_t entry
)
1625 struct swap_info_struct
*p
;
1626 unsigned char count
;
1628 if (non_swap_entry(entry
))
1631 p
= _swap_info_get(entry
);
1633 count
= __swap_entry_free(p
, entry
, 1);
1634 if (count
== SWAP_HAS_CACHE
&&
1635 !swap_page_trans_huge_swapped(p
, entry
))
1636 __try_to_reclaim_swap(p
, swp_offset(entry
),
1637 TTRS_UNMAPPED
| TTRS_FULL
);
1642 #ifdef CONFIG_HIBERNATION
1644 * Find the swap type that corresponds to given device (if any).
1646 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1647 * from 0, in which the swap header is expected to be located.
1649 * This is needed for the suspend to disk (aka swsusp).
1651 int swap_type_of(dev_t device
, sector_t offset
, struct block_device
**bdev_p
)
1653 struct block_device
*bdev
= NULL
;
1657 bdev
= bdget(device
);
1659 spin_lock(&swap_lock
);
1660 for (type
= 0; type
< nr_swapfiles
; type
++) {
1661 struct swap_info_struct
*sis
= swap_info
[type
];
1663 if (!(sis
->flags
& SWP_WRITEOK
))
1668 *bdev_p
= bdgrab(sis
->bdev
);
1670 spin_unlock(&swap_lock
);
1673 if (bdev
== sis
->bdev
) {
1674 struct swap_extent
*se
= &sis
->first_swap_extent
;
1676 if (se
->start_block
== offset
) {
1678 *bdev_p
= bdgrab(sis
->bdev
);
1680 spin_unlock(&swap_lock
);
1686 spin_unlock(&swap_lock
);
1694 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1695 * corresponding to given index in swap_info (swap type).
1697 sector_t
swapdev_block(int type
, pgoff_t offset
)
1699 struct block_device
*bdev
;
1701 if ((unsigned int)type
>= nr_swapfiles
)
1703 if (!(swap_info
[type
]->flags
& SWP_WRITEOK
))
1705 return map_swap_entry(swp_entry(type
, offset
), &bdev
);
1709 * Return either the total number of swap pages of given type, or the number
1710 * of free pages of that type (depending on @free)
1712 * This is needed for software suspend
1714 unsigned int count_swap_pages(int type
, int free
)
1718 spin_lock(&swap_lock
);
1719 if ((unsigned int)type
< nr_swapfiles
) {
1720 struct swap_info_struct
*sis
= swap_info
[type
];
1722 spin_lock(&sis
->lock
);
1723 if (sis
->flags
& SWP_WRITEOK
) {
1726 n
-= sis
->inuse_pages
;
1728 spin_unlock(&sis
->lock
);
1730 spin_unlock(&swap_lock
);
1733 #endif /* CONFIG_HIBERNATION */
1735 static inline int pte_same_as_swp(pte_t pte
, pte_t swp_pte
)
1737 return pte_same(pte_swp_clear_soft_dirty(pte
), swp_pte
);
1741 * No need to decide whether this PTE shares the swap entry with others,
1742 * just let do_wp_page work it out if a write is requested later - to
1743 * force COW, vm_page_prot omits write permission from any private vma.
1745 static int unuse_pte(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1746 unsigned long addr
, swp_entry_t entry
, struct page
*page
)
1748 struct page
*swapcache
;
1749 struct mem_cgroup
*memcg
;
1755 page
= ksm_might_need_to_copy(page
, vma
, addr
);
1756 if (unlikely(!page
))
1759 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
,
1765 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
1766 if (unlikely(!pte_same_as_swp(*pte
, swp_entry_to_pte(entry
)))) {
1767 mem_cgroup_cancel_charge(page
, memcg
, false);
1772 dec_mm_counter(vma
->vm_mm
, MM_SWAPENTS
);
1773 inc_mm_counter(vma
->vm_mm
, MM_ANONPAGES
);
1775 set_pte_at(vma
->vm_mm
, addr
, pte
,
1776 pte_mkold(mk_pte(page
, vma
->vm_page_prot
)));
1777 if (page
== swapcache
) {
1778 page_add_anon_rmap(page
, vma
, addr
, false);
1779 mem_cgroup_commit_charge(page
, memcg
, true, false);
1780 } else { /* ksm created a completely new copy */
1781 page_add_new_anon_rmap(page
, vma
, addr
, false);
1782 mem_cgroup_commit_charge(page
, memcg
, false, false);
1783 lru_cache_add_active_or_unevictable(page
, vma
);
1787 * Move the page to the active list so it is not
1788 * immediately swapped out again after swapon.
1790 activate_page(page
);
1792 pte_unmap_unlock(pte
, ptl
);
1794 if (page
!= swapcache
) {
1801 static int unuse_pte_range(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1802 unsigned long addr
, unsigned long end
,
1803 swp_entry_t entry
, struct page
*page
)
1805 pte_t swp_pte
= swp_entry_to_pte(entry
);
1810 * We don't actually need pte lock while scanning for swp_pte: since
1811 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1812 * page table while we're scanning; though it could get zapped, and on
1813 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1814 * of unmatched parts which look like swp_pte, so unuse_pte must
1815 * recheck under pte lock. Scanning without pte lock lets it be
1816 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1818 pte
= pte_offset_map(pmd
, addr
);
1821 * swapoff spends a _lot_ of time in this loop!
1822 * Test inline before going to call unuse_pte.
1824 if (unlikely(pte_same_as_swp(*pte
, swp_pte
))) {
1826 ret
= unuse_pte(vma
, pmd
, addr
, entry
, page
);
1829 pte
= pte_offset_map(pmd
, addr
);
1831 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1837 static inline int unuse_pmd_range(struct vm_area_struct
*vma
, pud_t
*pud
,
1838 unsigned long addr
, unsigned long end
,
1839 swp_entry_t entry
, struct page
*page
)
1845 pmd
= pmd_offset(pud
, addr
);
1848 next
= pmd_addr_end(addr
, end
);
1849 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1851 ret
= unuse_pte_range(vma
, pmd
, addr
, next
, entry
, page
);
1854 } while (pmd
++, addr
= next
, addr
!= end
);
1858 static inline int unuse_pud_range(struct vm_area_struct
*vma
, p4d_t
*p4d
,
1859 unsigned long addr
, unsigned long end
,
1860 swp_entry_t entry
, struct page
*page
)
1866 pud
= pud_offset(p4d
, addr
);
1868 next
= pud_addr_end(addr
, end
);
1869 if (pud_none_or_clear_bad(pud
))
1871 ret
= unuse_pmd_range(vma
, pud
, addr
, next
, entry
, page
);
1874 } while (pud
++, addr
= next
, addr
!= end
);
1878 static inline int unuse_p4d_range(struct vm_area_struct
*vma
, pgd_t
*pgd
,
1879 unsigned long addr
, unsigned long end
,
1880 swp_entry_t entry
, struct page
*page
)
1886 p4d
= p4d_offset(pgd
, addr
);
1888 next
= p4d_addr_end(addr
, end
);
1889 if (p4d_none_or_clear_bad(p4d
))
1891 ret
= unuse_pud_range(vma
, p4d
, addr
, next
, entry
, page
);
1894 } while (p4d
++, addr
= next
, addr
!= end
);
1898 static int unuse_vma(struct vm_area_struct
*vma
,
1899 swp_entry_t entry
, struct page
*page
)
1902 unsigned long addr
, end
, next
;
1905 if (page_anon_vma(page
)) {
1906 addr
= page_address_in_vma(page
, vma
);
1907 if (addr
== -EFAULT
)
1910 end
= addr
+ PAGE_SIZE
;
1912 addr
= vma
->vm_start
;
1916 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1918 next
= pgd_addr_end(addr
, end
);
1919 if (pgd_none_or_clear_bad(pgd
))
1921 ret
= unuse_p4d_range(vma
, pgd
, addr
, next
, entry
, page
);
1924 } while (pgd
++, addr
= next
, addr
!= end
);
1928 static int unuse_mm(struct mm_struct
*mm
,
1929 swp_entry_t entry
, struct page
*page
)
1931 struct vm_area_struct
*vma
;
1934 if (!down_read_trylock(&mm
->mmap_sem
)) {
1936 * Activate page so shrink_inactive_list is unlikely to unmap
1937 * its ptes while lock is dropped, so swapoff can make progress.
1939 activate_page(page
);
1941 down_read(&mm
->mmap_sem
);
1944 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
1945 if (vma
->anon_vma
&& (ret
= unuse_vma(vma
, entry
, page
)))
1949 up_read(&mm
->mmap_sem
);
1950 return (ret
< 0)? ret
: 0;
1954 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1955 * from current position to next entry still in use.
1956 * Recycle to start on reaching the end, returning 0 when empty.
1958 static unsigned int find_next_to_unuse(struct swap_info_struct
*si
,
1959 unsigned int prev
, bool frontswap
)
1961 unsigned int max
= si
->max
;
1962 unsigned int i
= prev
;
1963 unsigned char count
;
1966 * No need for swap_lock here: we're just looking
1967 * for whether an entry is in use, not modifying it; false
1968 * hits are okay, and sys_swapoff() has already prevented new
1969 * allocations from this area (while holding swap_lock).
1978 * No entries in use at top of swap_map,
1979 * loop back to start and recheck there.
1985 count
= READ_ONCE(si
->swap_map
[i
]);
1986 if (count
&& swap_count(count
) != SWAP_MAP_BAD
)
1987 if (!frontswap
|| frontswap_test(si
, i
))
1989 if ((i
% LATENCY_LIMIT
) == 0)
1996 * We completely avoid races by reading each swap page in advance,
1997 * and then search for the process using it. All the necessary
1998 * page table adjustments can then be made atomically.
2000 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2001 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2003 int try_to_unuse(unsigned int type
, bool frontswap
,
2004 unsigned long pages_to_unuse
)
2006 struct swap_info_struct
*si
= swap_info
[type
];
2007 struct mm_struct
*start_mm
;
2008 volatile unsigned char *swap_map
; /* swap_map is accessed without
2009 * locking. Mark it as volatile
2010 * to prevent compiler doing
2013 unsigned char swcount
;
2020 * When searching mms for an entry, a good strategy is to
2021 * start at the first mm we freed the previous entry from
2022 * (though actually we don't notice whether we or coincidence
2023 * freed the entry). Initialize this start_mm with a hold.
2025 * A simpler strategy would be to start at the last mm we
2026 * freed the previous entry from; but that would take less
2027 * advantage of mmlist ordering, which clusters forked mms
2028 * together, child after parent. If we race with dup_mmap(), we
2029 * prefer to resolve parent before child, lest we miss entries
2030 * duplicated after we scanned child: using last mm would invert
2033 start_mm
= &init_mm
;
2037 * Keep on scanning until all entries have gone. Usually,
2038 * one pass through swap_map is enough, but not necessarily:
2039 * there are races when an instance of an entry might be missed.
2041 while ((i
= find_next_to_unuse(si
, i
, frontswap
)) != 0) {
2042 if (signal_pending(current
)) {
2048 * Get a page for the entry, using the existing swap
2049 * cache page if there is one. Otherwise, get a clean
2050 * page and read the swap into it.
2052 swap_map
= &si
->swap_map
[i
];
2053 entry
= swp_entry(type
, i
);
2054 page
= read_swap_cache_async(entry
,
2055 GFP_HIGHUSER_MOVABLE
, NULL
, 0, false);
2058 * Either swap_duplicate() failed because entry
2059 * has been freed independently, and will not be
2060 * reused since sys_swapoff() already disabled
2061 * allocation from here, or alloc_page() failed.
2063 swcount
= *swap_map
;
2065 * We don't hold lock here, so the swap entry could be
2066 * SWAP_MAP_BAD (when the cluster is discarding).
2067 * Instead of fail out, We can just skip the swap
2068 * entry because swapoff will wait for discarding
2071 if (!swcount
|| swcount
== SWAP_MAP_BAD
)
2078 * Don't hold on to start_mm if it looks like exiting.
2080 if (atomic_read(&start_mm
->mm_users
) == 1) {
2082 start_mm
= &init_mm
;
2087 * Wait for and lock page. When do_swap_page races with
2088 * try_to_unuse, do_swap_page can handle the fault much
2089 * faster than try_to_unuse can locate the entry. This
2090 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2091 * defer to do_swap_page in such a case - in some tests,
2092 * do_swap_page and try_to_unuse repeatedly compete.
2094 wait_on_page_locked(page
);
2095 wait_on_page_writeback(page
);
2097 wait_on_page_writeback(page
);
2100 * Remove all references to entry.
2102 swcount
= *swap_map
;
2103 if (swap_count(swcount
) == SWAP_MAP_SHMEM
) {
2104 retval
= shmem_unuse(entry
, page
);
2105 /* page has already been unlocked and released */
2110 if (swap_count(swcount
) && start_mm
!= &init_mm
)
2111 retval
= unuse_mm(start_mm
, entry
, page
);
2113 if (swap_count(*swap_map
)) {
2114 int set_start_mm
= (*swap_map
>= swcount
);
2115 struct list_head
*p
= &start_mm
->mmlist
;
2116 struct mm_struct
*new_start_mm
= start_mm
;
2117 struct mm_struct
*prev_mm
= start_mm
;
2118 struct mm_struct
*mm
;
2120 mmget(new_start_mm
);
2122 spin_lock(&mmlist_lock
);
2123 while (swap_count(*swap_map
) && !retval
&&
2124 (p
= p
->next
) != &start_mm
->mmlist
) {
2125 mm
= list_entry(p
, struct mm_struct
, mmlist
);
2126 if (!mmget_not_zero(mm
))
2128 spin_unlock(&mmlist_lock
);
2134 swcount
= *swap_map
;
2135 if (!swap_count(swcount
)) /* any usage ? */
2137 else if (mm
== &init_mm
)
2140 retval
= unuse_mm(mm
, entry
, page
);
2142 if (set_start_mm
&& *swap_map
< swcount
) {
2143 mmput(new_start_mm
);
2148 spin_lock(&mmlist_lock
);
2150 spin_unlock(&mmlist_lock
);
2153 start_mm
= new_start_mm
;
2162 * If a reference remains (rare), we would like to leave
2163 * the page in the swap cache; but try_to_unmap could
2164 * then re-duplicate the entry once we drop page lock,
2165 * so we might loop indefinitely; also, that page could
2166 * not be swapped out to other storage meanwhile. So:
2167 * delete from cache even if there's another reference,
2168 * after ensuring that the data has been saved to disk -
2169 * since if the reference remains (rarer), it will be
2170 * read from disk into another page. Splitting into two
2171 * pages would be incorrect if swap supported "shared
2172 * private" pages, but they are handled by tmpfs files.
2174 * Given how unuse_vma() targets one particular offset
2175 * in an anon_vma, once the anon_vma has been determined,
2176 * this splitting happens to be just what is needed to
2177 * handle where KSM pages have been swapped out: re-reading
2178 * is unnecessarily slow, but we can fix that later on.
2180 if (swap_count(*swap_map
) &&
2181 PageDirty(page
) && PageSwapCache(page
)) {
2182 struct writeback_control wbc
= {
2183 .sync_mode
= WB_SYNC_NONE
,
2186 swap_writepage(compound_head(page
), &wbc
);
2188 wait_on_page_writeback(page
);
2192 * It is conceivable that a racing task removed this page from
2193 * swap cache just before we acquired the page lock at the top,
2194 * or while we dropped it in unuse_mm(). The page might even
2195 * be back in swap cache on another swap area: that we must not
2196 * delete, since it may not have been written out to swap yet.
2198 if (PageSwapCache(page
) &&
2199 likely(page_private(page
) == entry
.val
) &&
2200 (!PageTransCompound(page
) ||
2201 !swap_page_trans_huge_swapped(si
, entry
)))
2202 delete_from_swap_cache(compound_head(page
));
2205 * So we could skip searching mms once swap count went
2206 * to 1, we did not mark any present ptes as dirty: must
2207 * mark page dirty so shrink_page_list will preserve it.
2214 * Make sure that we aren't completely killing
2215 * interactive performance.
2218 if (frontswap
&& pages_to_unuse
> 0) {
2219 if (!--pages_to_unuse
)
2229 * After a successful try_to_unuse, if no swap is now in use, we know
2230 * we can empty the mmlist. swap_lock must be held on entry and exit.
2231 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2232 * added to the mmlist just after page_duplicate - before would be racy.
2234 static void drain_mmlist(void)
2236 struct list_head
*p
, *next
;
2239 for (type
= 0; type
< nr_swapfiles
; type
++)
2240 if (swap_info
[type
]->inuse_pages
)
2242 spin_lock(&mmlist_lock
);
2243 list_for_each_safe(p
, next
, &init_mm
.mmlist
)
2245 spin_unlock(&mmlist_lock
);
2249 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2250 * corresponds to page offset for the specified swap entry.
2251 * Note that the type of this function is sector_t, but it returns page offset
2252 * into the bdev, not sector offset.
2254 static sector_t
map_swap_entry(swp_entry_t entry
, struct block_device
**bdev
)
2256 struct swap_info_struct
*sis
;
2257 struct swap_extent
*start_se
;
2258 struct swap_extent
*se
;
2261 sis
= swap_info
[swp_type(entry
)];
2264 offset
= swp_offset(entry
);
2265 start_se
= sis
->curr_swap_extent
;
2269 if (se
->start_page
<= offset
&&
2270 offset
< (se
->start_page
+ se
->nr_pages
)) {
2271 return se
->start_block
+ (offset
- se
->start_page
);
2273 se
= list_next_entry(se
, list
);
2274 sis
->curr_swap_extent
= se
;
2275 BUG_ON(se
== start_se
); /* It *must* be present */
2280 * Returns the page offset into bdev for the specified page's swap entry.
2282 sector_t
map_swap_page(struct page
*page
, struct block_device
**bdev
)
2285 entry
.val
= page_private(page
);
2286 return map_swap_entry(entry
, bdev
);
2290 * Free all of a swapdev's extent information
2292 static void destroy_swap_extents(struct swap_info_struct
*sis
)
2294 while (!list_empty(&sis
->first_swap_extent
.list
)) {
2295 struct swap_extent
*se
;
2297 se
= list_first_entry(&sis
->first_swap_extent
.list
,
2298 struct swap_extent
, list
);
2299 list_del(&se
->list
);
2303 if (sis
->flags
& SWP_ACTIVATED
) {
2304 struct file
*swap_file
= sis
->swap_file
;
2305 struct address_space
*mapping
= swap_file
->f_mapping
;
2307 sis
->flags
&= ~SWP_ACTIVATED
;
2308 if (mapping
->a_ops
->swap_deactivate
)
2309 mapping
->a_ops
->swap_deactivate(swap_file
);
2314 * Add a block range (and the corresponding page range) into this swapdev's
2315 * extent list. The extent list is kept sorted in page order.
2317 * This function rather assumes that it is called in ascending page order.
2320 add_swap_extent(struct swap_info_struct
*sis
, unsigned long start_page
,
2321 unsigned long nr_pages
, sector_t start_block
)
2323 struct swap_extent
*se
;
2324 struct swap_extent
*new_se
;
2325 struct list_head
*lh
;
2327 if (start_page
== 0) {
2328 se
= &sis
->first_swap_extent
;
2329 sis
->curr_swap_extent
= se
;
2331 se
->nr_pages
= nr_pages
;
2332 se
->start_block
= start_block
;
2335 lh
= sis
->first_swap_extent
.list
.prev
; /* Highest extent */
2336 se
= list_entry(lh
, struct swap_extent
, list
);
2337 BUG_ON(se
->start_page
+ se
->nr_pages
!= start_page
);
2338 if (se
->start_block
+ se
->nr_pages
== start_block
) {
2340 se
->nr_pages
+= nr_pages
;
2346 * No merge. Insert a new extent, preserving ordering.
2348 new_se
= kmalloc(sizeof(*se
), GFP_KERNEL
);
2351 new_se
->start_page
= start_page
;
2352 new_se
->nr_pages
= nr_pages
;
2353 new_se
->start_block
= start_block
;
2355 list_add_tail(&new_se
->list
, &sis
->first_swap_extent
.list
);
2358 EXPORT_SYMBOL_GPL(add_swap_extent
);
2361 * A `swap extent' is a simple thing which maps a contiguous range of pages
2362 * onto a contiguous range of disk blocks. An ordered list of swap extents
2363 * is built at swapon time and is then used at swap_writepage/swap_readpage
2364 * time for locating where on disk a page belongs.
2366 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2367 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2368 * swap files identically.
2370 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2371 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2372 * swapfiles are handled *identically* after swapon time.
2374 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2375 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2376 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2377 * requirements, they are simply tossed out - we will never use those blocks
2380 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2381 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2382 * which will scribble on the fs.
2384 * The amount of disk space which a single swap extent represents varies.
2385 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2386 * extents in the list. To avoid much list walking, we cache the previous
2387 * search location in `curr_swap_extent', and start new searches from there.
2388 * This is extremely effective. The average number of iterations in
2389 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2391 static int setup_swap_extents(struct swap_info_struct
*sis
, sector_t
*span
)
2393 struct file
*swap_file
= sis
->swap_file
;
2394 struct address_space
*mapping
= swap_file
->f_mapping
;
2395 struct inode
*inode
= mapping
->host
;
2398 if (S_ISBLK(inode
->i_mode
)) {
2399 ret
= add_swap_extent(sis
, 0, sis
->max
, 0);
2404 if (mapping
->a_ops
->swap_activate
) {
2405 ret
= mapping
->a_ops
->swap_activate(sis
, swap_file
, span
);
2407 sis
->flags
|= SWP_ACTIVATED
;
2409 sis
->flags
|= SWP_FS
;
2410 ret
= add_swap_extent(sis
, 0, sis
->max
, 0);
2416 return generic_swapfile_activate(sis
, swap_file
, span
);
2419 static int swap_node(struct swap_info_struct
*p
)
2421 struct block_device
*bdev
;
2426 bdev
= p
->swap_file
->f_inode
->i_sb
->s_bdev
;
2428 return bdev
? bdev
->bd_disk
->node_id
: NUMA_NO_NODE
;
2431 static void _enable_swap_info(struct swap_info_struct
*p
, int prio
,
2432 unsigned char *swap_map
,
2433 struct swap_cluster_info
*cluster_info
)
2440 p
->prio
= --least_priority
;
2442 * the plist prio is negated because plist ordering is
2443 * low-to-high, while swap ordering is high-to-low
2445 p
->list
.prio
= -p
->prio
;
2448 p
->avail_lists
[i
].prio
= -p
->prio
;
2450 if (swap_node(p
) == i
)
2451 p
->avail_lists
[i
].prio
= 1;
2453 p
->avail_lists
[i
].prio
= -p
->prio
;
2456 p
->swap_map
= swap_map
;
2457 p
->cluster_info
= cluster_info
;
2458 p
->flags
|= SWP_WRITEOK
;
2459 atomic_long_add(p
->pages
, &nr_swap_pages
);
2460 total_swap_pages
+= p
->pages
;
2462 assert_spin_locked(&swap_lock
);
2464 * both lists are plists, and thus priority ordered.
2465 * swap_active_head needs to be priority ordered for swapoff(),
2466 * which on removal of any swap_info_struct with an auto-assigned
2467 * (i.e. negative) priority increments the auto-assigned priority
2468 * of any lower-priority swap_info_structs.
2469 * swap_avail_head needs to be priority ordered for get_swap_page(),
2470 * which allocates swap pages from the highest available priority
2473 plist_add(&p
->list
, &swap_active_head
);
2474 add_to_avail_list(p
);
2477 static void enable_swap_info(struct swap_info_struct
*p
, int prio
,
2478 unsigned char *swap_map
,
2479 struct swap_cluster_info
*cluster_info
,
2480 unsigned long *frontswap_map
)
2482 frontswap_init(p
->type
, frontswap_map
);
2483 spin_lock(&swap_lock
);
2484 spin_lock(&p
->lock
);
2485 _enable_swap_info(p
, prio
, swap_map
, cluster_info
);
2486 spin_unlock(&p
->lock
);
2487 spin_unlock(&swap_lock
);
2490 static void reinsert_swap_info(struct swap_info_struct
*p
)
2492 spin_lock(&swap_lock
);
2493 spin_lock(&p
->lock
);
2494 _enable_swap_info(p
, p
->prio
, p
->swap_map
, p
->cluster_info
);
2495 spin_unlock(&p
->lock
);
2496 spin_unlock(&swap_lock
);
2499 bool has_usable_swap(void)
2503 spin_lock(&swap_lock
);
2504 if (plist_head_empty(&swap_active_head
))
2506 spin_unlock(&swap_lock
);
2510 SYSCALL_DEFINE1(swapoff
, const char __user
*, specialfile
)
2512 struct swap_info_struct
*p
= NULL
;
2513 unsigned char *swap_map
;
2514 struct swap_cluster_info
*cluster_info
;
2515 unsigned long *frontswap_map
;
2516 struct file
*swap_file
, *victim
;
2517 struct address_space
*mapping
;
2518 struct inode
*inode
;
2519 struct filename
*pathname
;
2521 unsigned int old_block_size
;
2523 if (!capable(CAP_SYS_ADMIN
))
2526 BUG_ON(!current
->mm
);
2528 pathname
= getname(specialfile
);
2529 if (IS_ERR(pathname
))
2530 return PTR_ERR(pathname
);
2532 victim
= file_open_name(pathname
, O_RDWR
|O_LARGEFILE
, 0);
2533 err
= PTR_ERR(victim
);
2537 mapping
= victim
->f_mapping
;
2538 spin_lock(&swap_lock
);
2539 plist_for_each_entry(p
, &swap_active_head
, list
) {
2540 if (p
->flags
& SWP_WRITEOK
) {
2541 if (p
->swap_file
->f_mapping
== mapping
) {
2549 spin_unlock(&swap_lock
);
2552 if (!security_vm_enough_memory_mm(current
->mm
, p
->pages
))
2553 vm_unacct_memory(p
->pages
);
2556 spin_unlock(&swap_lock
);
2559 del_from_avail_list(p
);
2560 spin_lock(&p
->lock
);
2562 struct swap_info_struct
*si
= p
;
2565 plist_for_each_entry_continue(si
, &swap_active_head
, list
) {
2568 for_each_node(nid
) {
2569 if (si
->avail_lists
[nid
].prio
!= 1)
2570 si
->avail_lists
[nid
].prio
--;
2575 plist_del(&p
->list
, &swap_active_head
);
2576 atomic_long_sub(p
->pages
, &nr_swap_pages
);
2577 total_swap_pages
-= p
->pages
;
2578 p
->flags
&= ~SWP_WRITEOK
;
2579 spin_unlock(&p
->lock
);
2580 spin_unlock(&swap_lock
);
2582 disable_swap_slots_cache_lock();
2584 set_current_oom_origin();
2585 err
= try_to_unuse(p
->type
, false, 0); /* force unuse all pages */
2586 clear_current_oom_origin();
2589 /* re-insert swap space back into swap_list */
2590 reinsert_swap_info(p
);
2591 reenable_swap_slots_cache_unlock();
2595 reenable_swap_slots_cache_unlock();
2597 flush_work(&p
->discard_work
);
2599 destroy_swap_extents(p
);
2600 if (p
->flags
& SWP_CONTINUED
)
2601 free_swap_count_continuations(p
);
2603 if (!p
->bdev
|| !blk_queue_nonrot(bdev_get_queue(p
->bdev
)))
2604 atomic_dec(&nr_rotate_swap
);
2606 mutex_lock(&swapon_mutex
);
2607 spin_lock(&swap_lock
);
2608 spin_lock(&p
->lock
);
2611 /* wait for anyone still in scan_swap_map */
2612 p
->highest_bit
= 0; /* cuts scans short */
2613 while (p
->flags
>= SWP_SCANNING
) {
2614 spin_unlock(&p
->lock
);
2615 spin_unlock(&swap_lock
);
2616 schedule_timeout_uninterruptible(1);
2617 spin_lock(&swap_lock
);
2618 spin_lock(&p
->lock
);
2621 swap_file
= p
->swap_file
;
2622 old_block_size
= p
->old_block_size
;
2623 p
->swap_file
= NULL
;
2625 swap_map
= p
->swap_map
;
2627 cluster_info
= p
->cluster_info
;
2628 p
->cluster_info
= NULL
;
2629 frontswap_map
= frontswap_map_get(p
);
2630 spin_unlock(&p
->lock
);
2631 spin_unlock(&swap_lock
);
2632 frontswap_invalidate_area(p
->type
);
2633 frontswap_map_set(p
, NULL
);
2634 mutex_unlock(&swapon_mutex
);
2635 free_percpu(p
->percpu_cluster
);
2636 p
->percpu_cluster
= NULL
;
2638 kvfree(cluster_info
);
2639 kvfree(frontswap_map
);
2640 /* Destroy swap account information */
2641 swap_cgroup_swapoff(p
->type
);
2642 exit_swap_address_space(p
->type
);
2644 inode
= mapping
->host
;
2645 if (S_ISBLK(inode
->i_mode
)) {
2646 struct block_device
*bdev
= I_BDEV(inode
);
2647 set_blocksize(bdev
, old_block_size
);
2648 blkdev_put(bdev
, FMODE_READ
| FMODE_WRITE
| FMODE_EXCL
);
2651 inode
->i_flags
&= ~S_SWAPFILE
;
2652 inode_unlock(inode
);
2654 filp_close(swap_file
, NULL
);
2657 * Clear the SWP_USED flag after all resources are freed so that swapon
2658 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2659 * not hold p->lock after we cleared its SWP_WRITEOK.
2661 spin_lock(&swap_lock
);
2663 spin_unlock(&swap_lock
);
2666 atomic_inc(&proc_poll_event
);
2667 wake_up_interruptible(&proc_poll_wait
);
2670 filp_close(victim
, NULL
);
2676 #ifdef CONFIG_PROC_FS
2677 static __poll_t
swaps_poll(struct file
*file
, poll_table
*wait
)
2679 struct seq_file
*seq
= file
->private_data
;
2681 poll_wait(file
, &proc_poll_wait
, wait
);
2683 if (seq
->poll_event
!= atomic_read(&proc_poll_event
)) {
2684 seq
->poll_event
= atomic_read(&proc_poll_event
);
2685 return EPOLLIN
| EPOLLRDNORM
| EPOLLERR
| EPOLLPRI
;
2688 return EPOLLIN
| EPOLLRDNORM
;
2692 static void *swap_start(struct seq_file
*swap
, loff_t
*pos
)
2694 struct swap_info_struct
*si
;
2698 mutex_lock(&swapon_mutex
);
2701 return SEQ_START_TOKEN
;
2703 for (type
= 0; type
< nr_swapfiles
; type
++) {
2704 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2705 si
= swap_info
[type
];
2706 if (!(si
->flags
& SWP_USED
) || !si
->swap_map
)
2715 static void *swap_next(struct seq_file
*swap
, void *v
, loff_t
*pos
)
2717 struct swap_info_struct
*si
= v
;
2720 if (v
== SEQ_START_TOKEN
)
2723 type
= si
->type
+ 1;
2725 for (; type
< nr_swapfiles
; type
++) {
2726 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2727 si
= swap_info
[type
];
2728 if (!(si
->flags
& SWP_USED
) || !si
->swap_map
)
2737 static void swap_stop(struct seq_file
*swap
, void *v
)
2739 mutex_unlock(&swapon_mutex
);
2742 static int swap_show(struct seq_file
*swap
, void *v
)
2744 struct swap_info_struct
*si
= v
;
2748 if (si
== SEQ_START_TOKEN
) {
2749 seq_puts(swap
,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2753 file
= si
->swap_file
;
2754 len
= seq_file_path(swap
, file
, " \t\n\\");
2755 seq_printf(swap
, "%*s%s\t%u\t%u\t%d\n",
2756 len
< 40 ? 40 - len
: 1, " ",
2757 S_ISBLK(file_inode(file
)->i_mode
) ?
2758 "partition" : "file\t",
2759 si
->pages
<< (PAGE_SHIFT
- 10),
2760 si
->inuse_pages
<< (PAGE_SHIFT
- 10),
2765 static const struct seq_operations swaps_op
= {
2766 .start
= swap_start
,
2772 static int swaps_open(struct inode
*inode
, struct file
*file
)
2774 struct seq_file
*seq
;
2777 ret
= seq_open(file
, &swaps_op
);
2781 seq
= file
->private_data
;
2782 seq
->poll_event
= atomic_read(&proc_poll_event
);
2786 static const struct file_operations proc_swaps_operations
= {
2789 .llseek
= seq_lseek
,
2790 .release
= seq_release
,
2794 static int __init
procswaps_init(void)
2796 proc_create("swaps", 0, NULL
, &proc_swaps_operations
);
2799 __initcall(procswaps_init
);
2800 #endif /* CONFIG_PROC_FS */
2802 #ifdef MAX_SWAPFILES_CHECK
2803 static int __init
max_swapfiles_check(void)
2805 MAX_SWAPFILES_CHECK();
2808 late_initcall(max_swapfiles_check
);
2811 static struct swap_info_struct
*alloc_swap_info(void)
2813 struct swap_info_struct
*p
;
2816 int size
= sizeof(*p
) + nr_node_ids
* sizeof(struct plist_node
);
2818 p
= kvzalloc(size
, GFP_KERNEL
);
2820 return ERR_PTR(-ENOMEM
);
2822 spin_lock(&swap_lock
);
2823 for (type
= 0; type
< nr_swapfiles
; type
++) {
2824 if (!(swap_info
[type
]->flags
& SWP_USED
))
2827 if (type
>= MAX_SWAPFILES
) {
2828 spin_unlock(&swap_lock
);
2830 return ERR_PTR(-EPERM
);
2832 if (type
>= nr_swapfiles
) {
2834 swap_info
[type
] = p
;
2836 * Write swap_info[type] before nr_swapfiles, in case a
2837 * racing procfs swap_start() or swap_next() is reading them.
2838 * (We never shrink nr_swapfiles, we never free this entry.)
2844 p
= swap_info
[type
];
2846 * Do not memset this entry: a racing procfs swap_next()
2847 * would be relying on p->type to remain valid.
2850 INIT_LIST_HEAD(&p
->first_swap_extent
.list
);
2851 plist_node_init(&p
->list
, 0);
2853 plist_node_init(&p
->avail_lists
[i
], 0);
2854 p
->flags
= SWP_USED
;
2855 spin_unlock(&swap_lock
);
2856 spin_lock_init(&p
->lock
);
2857 spin_lock_init(&p
->cont_lock
);
2862 static int claim_swapfile(struct swap_info_struct
*p
, struct inode
*inode
)
2866 if (S_ISBLK(inode
->i_mode
)) {
2867 p
->bdev
= bdgrab(I_BDEV(inode
));
2868 error
= blkdev_get(p
->bdev
,
2869 FMODE_READ
| FMODE_WRITE
| FMODE_EXCL
, p
);
2874 p
->old_block_size
= block_size(p
->bdev
);
2875 error
= set_blocksize(p
->bdev
, PAGE_SIZE
);
2878 p
->flags
|= SWP_BLKDEV
;
2879 } else if (S_ISREG(inode
->i_mode
)) {
2880 p
->bdev
= inode
->i_sb
->s_bdev
;
2882 if (IS_SWAPFILE(inode
))
2892 * Find out how many pages are allowed for a single swap device. There
2893 * are two limiting factors:
2894 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2895 * 2) the number of bits in the swap pte, as defined by the different
2898 * In order to find the largest possible bit mask, a swap entry with
2899 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2900 * decoded to a swp_entry_t again, and finally the swap offset is
2903 * This will mask all the bits from the initial ~0UL mask that can't
2904 * be encoded in either the swp_entry_t or the architecture definition
2907 unsigned long generic_max_swapfile_size(void)
2909 return swp_offset(pte_to_swp_entry(
2910 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2913 /* Can be overridden by an architecture for additional checks. */
2914 __weak
unsigned long max_swapfile_size(void)
2916 return generic_max_swapfile_size();
2919 static unsigned long read_swap_header(struct swap_info_struct
*p
,
2920 union swap_header
*swap_header
,
2921 struct inode
*inode
)
2924 unsigned long maxpages
;
2925 unsigned long swapfilepages
;
2926 unsigned long last_page
;
2928 if (memcmp("SWAPSPACE2", swap_header
->magic
.magic
, 10)) {
2929 pr_err("Unable to find swap-space signature\n");
2933 /* swap partition endianess hack... */
2934 if (swab32(swap_header
->info
.version
) == 1) {
2935 swab32s(&swap_header
->info
.version
);
2936 swab32s(&swap_header
->info
.last_page
);
2937 swab32s(&swap_header
->info
.nr_badpages
);
2938 if (swap_header
->info
.nr_badpages
> MAX_SWAP_BADPAGES
)
2940 for (i
= 0; i
< swap_header
->info
.nr_badpages
; i
++)
2941 swab32s(&swap_header
->info
.badpages
[i
]);
2943 /* Check the swap header's sub-version */
2944 if (swap_header
->info
.version
!= 1) {
2945 pr_warn("Unable to handle swap header version %d\n",
2946 swap_header
->info
.version
);
2951 p
->cluster_next
= 1;
2954 maxpages
= max_swapfile_size();
2955 last_page
= swap_header
->info
.last_page
;
2957 pr_warn("Empty swap-file\n");
2960 if (last_page
> maxpages
) {
2961 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2962 maxpages
<< (PAGE_SHIFT
- 10),
2963 last_page
<< (PAGE_SHIFT
- 10));
2965 if (maxpages
> last_page
) {
2966 maxpages
= last_page
+ 1;
2967 /* p->max is an unsigned int: don't overflow it */
2968 if ((unsigned int)maxpages
== 0)
2969 maxpages
= UINT_MAX
;
2971 p
->highest_bit
= maxpages
- 1;
2975 swapfilepages
= i_size_read(inode
) >> PAGE_SHIFT
;
2976 if (swapfilepages
&& maxpages
> swapfilepages
) {
2977 pr_warn("Swap area shorter than signature indicates\n");
2980 if (swap_header
->info
.nr_badpages
&& S_ISREG(inode
->i_mode
))
2982 if (swap_header
->info
.nr_badpages
> MAX_SWAP_BADPAGES
)
2988 #define SWAP_CLUSTER_INFO_COLS \
2989 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2990 #define SWAP_CLUSTER_SPACE_COLS \
2991 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2992 #define SWAP_CLUSTER_COLS \
2993 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2995 static int setup_swap_map_and_extents(struct swap_info_struct
*p
,
2996 union swap_header
*swap_header
,
2997 unsigned char *swap_map
,
2998 struct swap_cluster_info
*cluster_info
,
2999 unsigned long maxpages
,
3003 unsigned int nr_good_pages
;
3005 unsigned long nr_clusters
= DIV_ROUND_UP(maxpages
, SWAPFILE_CLUSTER
);
3006 unsigned long col
= p
->cluster_next
/ SWAPFILE_CLUSTER
% SWAP_CLUSTER_COLS
;
3007 unsigned long i
, idx
;
3009 nr_good_pages
= maxpages
- 1; /* omit header page */
3011 cluster_list_init(&p
->free_clusters
);
3012 cluster_list_init(&p
->discard_clusters
);
3014 for (i
= 0; i
< swap_header
->info
.nr_badpages
; i
++) {
3015 unsigned int page_nr
= swap_header
->info
.badpages
[i
];
3016 if (page_nr
== 0 || page_nr
> swap_header
->info
.last_page
)
3018 if (page_nr
< maxpages
) {
3019 swap_map
[page_nr
] = SWAP_MAP_BAD
;
3022 * Haven't marked the cluster free yet, no list
3023 * operation involved
3025 inc_cluster_info_page(p
, cluster_info
, page_nr
);
3029 /* Haven't marked the cluster free yet, no list operation involved */
3030 for (i
= maxpages
; i
< round_up(maxpages
, SWAPFILE_CLUSTER
); i
++)
3031 inc_cluster_info_page(p
, cluster_info
, i
);
3033 if (nr_good_pages
) {
3034 swap_map
[0] = SWAP_MAP_BAD
;
3036 * Not mark the cluster free yet, no list
3037 * operation involved
3039 inc_cluster_info_page(p
, cluster_info
, 0);
3041 p
->pages
= nr_good_pages
;
3042 nr_extents
= setup_swap_extents(p
, span
);
3045 nr_good_pages
= p
->pages
;
3047 if (!nr_good_pages
) {
3048 pr_warn("Empty swap-file\n");
3057 * Reduce false cache line sharing between cluster_info and
3058 * sharing same address space.
3060 for (k
= 0; k
< SWAP_CLUSTER_COLS
; k
++) {
3061 j
= (k
+ col
) % SWAP_CLUSTER_COLS
;
3062 for (i
= 0; i
< DIV_ROUND_UP(nr_clusters
, SWAP_CLUSTER_COLS
); i
++) {
3063 idx
= i
* SWAP_CLUSTER_COLS
+ j
;
3064 if (idx
>= nr_clusters
)
3066 if (cluster_count(&cluster_info
[idx
]))
3068 cluster_set_flag(&cluster_info
[idx
], CLUSTER_FLAG_FREE
);
3069 cluster_list_add_tail(&p
->free_clusters
, cluster_info
,
3077 * Helper to sys_swapon determining if a given swap
3078 * backing device queue supports DISCARD operations.
3080 static bool swap_discardable(struct swap_info_struct
*si
)
3082 struct request_queue
*q
= bdev_get_queue(si
->bdev
);
3084 if (!q
|| !blk_queue_discard(q
))
3090 SYSCALL_DEFINE2(swapon
, const char __user
*, specialfile
, int, swap_flags
)
3092 struct swap_info_struct
*p
;
3093 struct filename
*name
;
3094 struct file
*swap_file
= NULL
;
3095 struct address_space
*mapping
;
3098 union swap_header
*swap_header
;
3101 unsigned long maxpages
;
3102 unsigned char *swap_map
= NULL
;
3103 struct swap_cluster_info
*cluster_info
= NULL
;
3104 unsigned long *frontswap_map
= NULL
;
3105 struct page
*page
= NULL
;
3106 struct inode
*inode
= NULL
;
3107 bool inced_nr_rotate_swap
= false;
3109 if (swap_flags
& ~SWAP_FLAGS_VALID
)
3112 if (!capable(CAP_SYS_ADMIN
))
3115 if (!swap_avail_heads
)
3118 p
= alloc_swap_info();
3122 INIT_WORK(&p
->discard_work
, swap_discard_work
);
3124 name
= getname(specialfile
);
3126 error
= PTR_ERR(name
);
3130 swap_file
= file_open_name(name
, O_RDWR
|O_LARGEFILE
, 0);
3131 if (IS_ERR(swap_file
)) {
3132 error
= PTR_ERR(swap_file
);
3137 p
->swap_file
= swap_file
;
3138 mapping
= swap_file
->f_mapping
;
3139 inode
= mapping
->host
;
3141 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3142 error
= claim_swapfile(p
, inode
);
3143 if (unlikely(error
))
3147 * Read the swap header.
3149 if (!mapping
->a_ops
->readpage
) {
3153 page
= read_mapping_page(mapping
, 0, swap_file
);
3155 error
= PTR_ERR(page
);
3158 swap_header
= kmap(page
);
3160 maxpages
= read_swap_header(p
, swap_header
, inode
);
3161 if (unlikely(!maxpages
)) {
3166 /* OK, set up the swap map and apply the bad block list */
3167 swap_map
= vzalloc(maxpages
);
3173 if (bdi_cap_stable_pages_required(inode_to_bdi(inode
)))
3174 p
->flags
|= SWP_STABLE_WRITES
;
3176 if (bdi_cap_synchronous_io(inode_to_bdi(inode
)))
3177 p
->flags
|= SWP_SYNCHRONOUS_IO
;
3179 if (p
->bdev
&& blk_queue_nonrot(bdev_get_queue(p
->bdev
))) {
3181 unsigned long ci
, nr_cluster
;
3183 p
->flags
|= SWP_SOLIDSTATE
;
3185 * select a random position to start with to help wear leveling
3188 p
->cluster_next
= 1 + (prandom_u32() % p
->highest_bit
);
3189 nr_cluster
= DIV_ROUND_UP(maxpages
, SWAPFILE_CLUSTER
);
3191 cluster_info
= kvcalloc(nr_cluster
, sizeof(*cluster_info
),
3193 if (!cluster_info
) {
3198 for (ci
= 0; ci
< nr_cluster
; ci
++)
3199 spin_lock_init(&((cluster_info
+ ci
)->lock
));
3201 p
->percpu_cluster
= alloc_percpu(struct percpu_cluster
);
3202 if (!p
->percpu_cluster
) {
3206 for_each_possible_cpu(cpu
) {
3207 struct percpu_cluster
*cluster
;
3208 cluster
= per_cpu_ptr(p
->percpu_cluster
, cpu
);
3209 cluster_set_null(&cluster
->index
);
3212 atomic_inc(&nr_rotate_swap
);
3213 inced_nr_rotate_swap
= true;
3216 error
= swap_cgroup_swapon(p
->type
, maxpages
);
3220 nr_extents
= setup_swap_map_and_extents(p
, swap_header
, swap_map
,
3221 cluster_info
, maxpages
, &span
);
3222 if (unlikely(nr_extents
< 0)) {
3226 /* frontswap enabled? set up bit-per-page map for frontswap */
3227 if (IS_ENABLED(CONFIG_FRONTSWAP
))
3228 frontswap_map
= kvcalloc(BITS_TO_LONGS(maxpages
),
3232 if (p
->bdev
&&(swap_flags
& SWAP_FLAG_DISCARD
) && swap_discardable(p
)) {
3234 * When discard is enabled for swap with no particular
3235 * policy flagged, we set all swap discard flags here in
3236 * order to sustain backward compatibility with older
3237 * swapon(8) releases.
3239 p
->flags
|= (SWP_DISCARDABLE
| SWP_AREA_DISCARD
|
3243 * By flagging sys_swapon, a sysadmin can tell us to
3244 * either do single-time area discards only, or to just
3245 * perform discards for released swap page-clusters.
3246 * Now it's time to adjust the p->flags accordingly.
3248 if (swap_flags
& SWAP_FLAG_DISCARD_ONCE
)
3249 p
->flags
&= ~SWP_PAGE_DISCARD
;
3250 else if (swap_flags
& SWAP_FLAG_DISCARD_PAGES
)
3251 p
->flags
&= ~SWP_AREA_DISCARD
;
3253 /* issue a swapon-time discard if it's still required */
3254 if (p
->flags
& SWP_AREA_DISCARD
) {
3255 int err
= discard_swap(p
);
3257 pr_err("swapon: discard_swap(%p): %d\n",
3262 error
= init_swap_address_space(p
->type
, maxpages
);
3266 mutex_lock(&swapon_mutex
);
3268 if (swap_flags
& SWAP_FLAG_PREFER
)
3270 (swap_flags
& SWAP_FLAG_PRIO_MASK
) >> SWAP_FLAG_PRIO_SHIFT
;
3271 enable_swap_info(p
, prio
, swap_map
, cluster_info
, frontswap_map
);
3273 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3274 p
->pages
<<(PAGE_SHIFT
-10), name
->name
, p
->prio
,
3275 nr_extents
, (unsigned long long)span
<<(PAGE_SHIFT
-10),
3276 (p
->flags
& SWP_SOLIDSTATE
) ? "SS" : "",
3277 (p
->flags
& SWP_DISCARDABLE
) ? "D" : "",
3278 (p
->flags
& SWP_AREA_DISCARD
) ? "s" : "",
3279 (p
->flags
& SWP_PAGE_DISCARD
) ? "c" : "",
3280 (frontswap_map
) ? "FS" : "");
3282 mutex_unlock(&swapon_mutex
);
3283 atomic_inc(&proc_poll_event
);
3284 wake_up_interruptible(&proc_poll_wait
);
3286 if (S_ISREG(inode
->i_mode
))
3287 inode
->i_flags
|= S_SWAPFILE
;
3291 free_percpu(p
->percpu_cluster
);
3292 p
->percpu_cluster
= NULL
;
3293 if (inode
&& S_ISBLK(inode
->i_mode
) && p
->bdev
) {
3294 set_blocksize(p
->bdev
, p
->old_block_size
);
3295 blkdev_put(p
->bdev
, FMODE_READ
| FMODE_WRITE
| FMODE_EXCL
);
3297 destroy_swap_extents(p
);
3298 swap_cgroup_swapoff(p
->type
);
3299 spin_lock(&swap_lock
);
3300 p
->swap_file
= NULL
;
3302 spin_unlock(&swap_lock
);
3304 kvfree(cluster_info
);
3305 kvfree(frontswap_map
);
3306 if (inced_nr_rotate_swap
)
3307 atomic_dec(&nr_rotate_swap
);
3309 if (inode
&& S_ISREG(inode
->i_mode
)) {
3310 inode_unlock(inode
);
3313 filp_close(swap_file
, NULL
);
3316 if (page
&& !IS_ERR(page
)) {
3322 if (inode
&& S_ISREG(inode
->i_mode
))
3323 inode_unlock(inode
);
3325 enable_swap_slots_cache();
3329 void si_swapinfo(struct sysinfo
*val
)
3332 unsigned long nr_to_be_unused
= 0;
3334 spin_lock(&swap_lock
);
3335 for (type
= 0; type
< nr_swapfiles
; type
++) {
3336 struct swap_info_struct
*si
= swap_info
[type
];
3338 if ((si
->flags
& SWP_USED
) && !(si
->flags
& SWP_WRITEOK
))
3339 nr_to_be_unused
+= si
->inuse_pages
;
3341 val
->freeswap
= atomic_long_read(&nr_swap_pages
) + nr_to_be_unused
;
3342 val
->totalswap
= total_swap_pages
+ nr_to_be_unused
;
3343 spin_unlock(&swap_lock
);
3347 * Verify that a swap entry is valid and increment its swap map count.
3349 * Returns error code in following case.
3351 * - swp_entry is invalid -> EINVAL
3352 * - swp_entry is migration entry -> EINVAL
3353 * - swap-cache reference is requested but there is already one. -> EEXIST
3354 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3355 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3357 static int __swap_duplicate(swp_entry_t entry
, unsigned char usage
)
3359 struct swap_info_struct
*p
;
3360 struct swap_cluster_info
*ci
;
3361 unsigned long offset
, type
;
3362 unsigned char count
;
3363 unsigned char has_cache
;
3366 if (non_swap_entry(entry
))
3369 type
= swp_type(entry
);
3370 if (type
>= nr_swapfiles
)
3372 p
= swap_info
[type
];
3373 offset
= swp_offset(entry
);
3374 if (unlikely(offset
>= p
->max
))
3377 ci
= lock_cluster_or_swap_info(p
, offset
);
3379 count
= p
->swap_map
[offset
];
3382 * swapin_readahead() doesn't check if a swap entry is valid, so the
3383 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3385 if (unlikely(swap_count(count
) == SWAP_MAP_BAD
)) {
3390 has_cache
= count
& SWAP_HAS_CACHE
;
3391 count
&= ~SWAP_HAS_CACHE
;
3394 if (usage
== SWAP_HAS_CACHE
) {
3396 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3397 if (!has_cache
&& count
)
3398 has_cache
= SWAP_HAS_CACHE
;
3399 else if (has_cache
) /* someone else added cache */
3401 else /* no users remaining */
3404 } else if (count
|| has_cache
) {
3406 if ((count
& ~COUNT_CONTINUED
) < SWAP_MAP_MAX
)
3408 else if ((count
& ~COUNT_CONTINUED
) > SWAP_MAP_MAX
)
3410 else if (swap_count_continued(p
, offset
, count
))
3411 count
= COUNT_CONTINUED
;
3415 err
= -ENOENT
; /* unused swap entry */
3417 p
->swap_map
[offset
] = count
| has_cache
;
3420 unlock_cluster_or_swap_info(p
, ci
);
3425 pr_err("swap_dup: %s%08lx\n", Bad_file
, entry
.val
);
3430 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3431 * (in which case its reference count is never incremented).
3433 void swap_shmem_alloc(swp_entry_t entry
)
3435 __swap_duplicate(entry
, SWAP_MAP_SHMEM
);
3439 * Increase reference count of swap entry by 1.
3440 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3441 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3442 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3443 * might occur if a page table entry has got corrupted.
3445 int swap_duplicate(swp_entry_t entry
)
3449 while (!err
&& __swap_duplicate(entry
, 1) == -ENOMEM
)
3450 err
= add_swap_count_continuation(entry
, GFP_ATOMIC
);
3455 * @entry: swap entry for which we allocate swap cache.
3457 * Called when allocating swap cache for existing swap entry,
3458 * This can return error codes. Returns 0 at success.
3459 * -EBUSY means there is a swap cache.
3460 * Note: return code is different from swap_duplicate().
3462 int swapcache_prepare(swp_entry_t entry
)
3464 return __swap_duplicate(entry
, SWAP_HAS_CACHE
);
3467 struct swap_info_struct
*swp_swap_info(swp_entry_t entry
)
3469 return swap_info
[swp_type(entry
)];
3472 struct swap_info_struct
*page_swap_info(struct page
*page
)
3474 swp_entry_t entry
= { .val
= page_private(page
) };
3475 return swp_swap_info(entry
);
3479 * out-of-line __page_file_ methods to avoid include hell.
3481 struct address_space
*__page_file_mapping(struct page
*page
)
3483 return page_swap_info(page
)->swap_file
->f_mapping
;
3485 EXPORT_SYMBOL_GPL(__page_file_mapping
);
3487 pgoff_t
__page_file_index(struct page
*page
)
3489 swp_entry_t swap
= { .val
= page_private(page
) };
3490 return swp_offset(swap
);
3492 EXPORT_SYMBOL_GPL(__page_file_index
);
3495 * add_swap_count_continuation - called when a swap count is duplicated
3496 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3497 * page of the original vmalloc'ed swap_map, to hold the continuation count
3498 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3499 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3501 * These continuation pages are seldom referenced: the common paths all work
3502 * on the original swap_map, only referring to a continuation page when the
3503 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3505 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3506 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3507 * can be called after dropping locks.
3509 int add_swap_count_continuation(swp_entry_t entry
, gfp_t gfp_mask
)
3511 struct swap_info_struct
*si
;
3512 struct swap_cluster_info
*ci
;
3515 struct page
*list_page
;
3517 unsigned char count
;
3520 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3521 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3523 page
= alloc_page(gfp_mask
| __GFP_HIGHMEM
);
3525 si
= swap_info_get(entry
);
3528 * An acceptable race has occurred since the failing
3529 * __swap_duplicate(): the swap entry has been freed,
3530 * perhaps even the whole swap_map cleared for swapoff.
3535 offset
= swp_offset(entry
);
3537 ci
= lock_cluster(si
, offset
);
3539 count
= si
->swap_map
[offset
] & ~SWAP_HAS_CACHE
;
3541 if ((count
& ~COUNT_CONTINUED
) != SWAP_MAP_MAX
) {
3543 * The higher the swap count, the more likely it is that tasks
3544 * will race to add swap count continuation: we need to avoid
3545 * over-provisioning.
3552 spin_unlock(&si
->lock
);
3557 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3558 * no architecture is using highmem pages for kernel page tables: so it
3559 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3561 head
= vmalloc_to_page(si
->swap_map
+ offset
);
3562 offset
&= ~PAGE_MASK
;
3564 spin_lock(&si
->cont_lock
);
3566 * Page allocation does not initialize the page's lru field,
3567 * but it does always reset its private field.
3569 if (!page_private(head
)) {
3570 BUG_ON(count
& COUNT_CONTINUED
);
3571 INIT_LIST_HEAD(&head
->lru
);
3572 set_page_private(head
, SWP_CONTINUED
);
3573 si
->flags
|= SWP_CONTINUED
;
3576 list_for_each_entry(list_page
, &head
->lru
, lru
) {
3580 * If the previous map said no continuation, but we've found
3581 * a continuation page, free our allocation and use this one.
3583 if (!(count
& COUNT_CONTINUED
))
3584 goto out_unlock_cont
;
3586 map
= kmap_atomic(list_page
) + offset
;
3591 * If this continuation count now has some space in it,
3592 * free our allocation and use this one.
3594 if ((count
& ~COUNT_CONTINUED
) != SWAP_CONT_MAX
)
3595 goto out_unlock_cont
;
3598 list_add_tail(&page
->lru
, &head
->lru
);
3599 page
= NULL
; /* now it's attached, don't free it */
3601 spin_unlock(&si
->cont_lock
);
3604 spin_unlock(&si
->lock
);
3612 * swap_count_continued - when the original swap_map count is incremented
3613 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3614 * into, carry if so, or else fail until a new continuation page is allocated;
3615 * when the original swap_map count is decremented from 0 with continuation,
3616 * borrow from the continuation and report whether it still holds more.
3617 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3620 static bool swap_count_continued(struct swap_info_struct
*si
,
3621 pgoff_t offset
, unsigned char count
)
3628 head
= vmalloc_to_page(si
->swap_map
+ offset
);
3629 if (page_private(head
) != SWP_CONTINUED
) {
3630 BUG_ON(count
& COUNT_CONTINUED
);
3631 return false; /* need to add count continuation */
3634 spin_lock(&si
->cont_lock
);
3635 offset
&= ~PAGE_MASK
;
3636 page
= list_entry(head
->lru
.next
, struct page
, lru
);
3637 map
= kmap_atomic(page
) + offset
;
3639 if (count
== SWAP_MAP_MAX
) /* initial increment from swap_map */
3640 goto init_map
; /* jump over SWAP_CONT_MAX checks */
3642 if (count
== (SWAP_MAP_MAX
| COUNT_CONTINUED
)) { /* incrementing */
3644 * Think of how you add 1 to 999
3646 while (*map
== (SWAP_CONT_MAX
| COUNT_CONTINUED
)) {
3648 page
= list_entry(page
->lru
.next
, struct page
, lru
);
3649 BUG_ON(page
== head
);
3650 map
= kmap_atomic(page
) + offset
;
3652 if (*map
== SWAP_CONT_MAX
) {
3654 page
= list_entry(page
->lru
.next
, struct page
, lru
);
3656 ret
= false; /* add count continuation */
3659 map
= kmap_atomic(page
) + offset
;
3660 init_map
: *map
= 0; /* we didn't zero the page */
3664 page
= list_entry(page
->lru
.prev
, struct page
, lru
);
3665 while (page
!= head
) {
3666 map
= kmap_atomic(page
) + offset
;
3667 *map
= COUNT_CONTINUED
;
3669 page
= list_entry(page
->lru
.prev
, struct page
, lru
);
3671 ret
= true; /* incremented */
3673 } else { /* decrementing */
3675 * Think of how you subtract 1 from 1000
3677 BUG_ON(count
!= COUNT_CONTINUED
);
3678 while (*map
== COUNT_CONTINUED
) {
3680 page
= list_entry(page
->lru
.next
, struct page
, lru
);
3681 BUG_ON(page
== head
);
3682 map
= kmap_atomic(page
) + offset
;
3689 page
= list_entry(page
->lru
.prev
, struct page
, lru
);
3690 while (page
!= head
) {
3691 map
= kmap_atomic(page
) + offset
;
3692 *map
= SWAP_CONT_MAX
| count
;
3693 count
= COUNT_CONTINUED
;
3695 page
= list_entry(page
->lru
.prev
, struct page
, lru
);
3697 ret
= count
== COUNT_CONTINUED
;
3700 spin_unlock(&si
->cont_lock
);
3705 * free_swap_count_continuations - swapoff free all the continuation pages
3706 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3708 static void free_swap_count_continuations(struct swap_info_struct
*si
)
3712 for (offset
= 0; offset
< si
->max
; offset
+= PAGE_SIZE
) {
3714 head
= vmalloc_to_page(si
->swap_map
+ offset
);
3715 if (page_private(head
)) {
3716 struct page
*page
, *next
;
3718 list_for_each_entry_safe(page
, next
, &head
->lru
, lru
) {
3719 list_del(&page
->lru
);
3726 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3727 void mem_cgroup_throttle_swaprate(struct mem_cgroup
*memcg
, int node
,
3730 struct swap_info_struct
*si
, *next
;
3731 if (!(gfp_mask
& __GFP_IO
) || !memcg
)
3734 if (!blk_cgroup_congested())
3738 * We've already scheduled a throttle, avoid taking the global swap
3741 if (current
->throttle_queue
)
3744 spin_lock(&swap_avail_lock
);
3745 plist_for_each_entry_safe(si
, next
, &swap_avail_heads
[node
],
3746 avail_lists
[node
]) {
3748 blkcg_schedule_throttle(bdev_get_queue(si
->bdev
),
3753 spin_unlock(&swap_avail_lock
);
3757 static int __init
swapfile_init(void)
3761 swap_avail_heads
= kmalloc_array(nr_node_ids
, sizeof(struct plist_head
),
3763 if (!swap_avail_heads
) {
3764 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3769 plist_head_init(&swap_avail_heads
[nid
]);
3773 subsys_initcall(swapfile_init
);