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 /* returns 1 if swap entry is freed */
108 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
110 swp_entry_t entry = swp_entry(si->type, offset);
114 page = find_get_page(swap_address_space(entry), swp_offset(entry));
118 * This function is called from scan_swap_map() and it's called
119 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
120 * We have to use trylock for avoiding deadlock. This is a special
121 * case and you should use try_to_free_swap() with explicit lock_page()
122 * in usual operations.
124 if (trylock_page(page)) {
125 ret = try_to_free_swap(page);
133 * swapon tell device that all the old swap contents can be discarded,
134 * to allow the swap device to optimize its wear-levelling.
136 static int discard_swap(struct swap_info_struct *si)
138 struct swap_extent *se;
139 sector_t start_block;
143 /* Do not discard the swap header page! */
144 se = &si->first_swap_extent;
145 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
146 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
148 err = blkdev_issue_discard(si->bdev, start_block,
149 nr_blocks, GFP_KERNEL, 0);
155 list_for_each_entry(se, &si->first_swap_extent.list, list) {
156 start_block = se->start_block << (PAGE_SHIFT - 9);
157 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
159 err = blkdev_issue_discard(si->bdev, start_block,
160 nr_blocks, GFP_KERNEL, 0);
166 return err; /* That will often be -EOPNOTSUPP */
170 * swap allocation tell device that a cluster of swap can now be discarded,
171 * to allow the swap device to optimize its wear-levelling.
173 static void discard_swap_cluster(struct swap_info_struct *si,
174 pgoff_t start_page, pgoff_t nr_pages)
176 struct swap_extent *se = si->curr_swap_extent;
177 int found_extent = 0;
180 if (se->start_page <= start_page &&
181 start_page < se->start_page + se->nr_pages) {
182 pgoff_t offset = start_page - se->start_page;
183 sector_t start_block = se->start_block + offset;
184 sector_t nr_blocks = se->nr_pages - offset;
186 if (nr_blocks > nr_pages)
187 nr_blocks = nr_pages;
188 start_page += nr_blocks;
189 nr_pages -= nr_blocks;
192 si->curr_swap_extent = se;
194 start_block <<= PAGE_SHIFT - 9;
195 nr_blocks <<= PAGE_SHIFT - 9;
196 if (blkdev_issue_discard(si->bdev, start_block,
197 nr_blocks, GFP_NOIO, 0))
201 se = list_next_entry(se, list);
205 #ifdef CONFIG_THP_SWAP
206 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
208 #define SWAPFILE_CLUSTER 256
210 #define LATENCY_LIMIT 256
212 static inline void cluster_set_flag(struct swap_cluster_info *info,
218 static inline unsigned int cluster_count(struct swap_cluster_info *info)
223 static inline void cluster_set_count(struct swap_cluster_info *info,
229 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
230 unsigned int c, unsigned int f)
236 static inline unsigned int cluster_next(struct swap_cluster_info *info)
241 static inline void cluster_set_next(struct swap_cluster_info *info,
247 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
248 unsigned int n, unsigned int f)
254 static inline bool cluster_is_free(struct swap_cluster_info *info)
256 return info->flags & CLUSTER_FLAG_FREE;
259 static inline bool cluster_is_null(struct swap_cluster_info *info)
261 return info->flags & CLUSTER_FLAG_NEXT_NULL;
264 static inline void cluster_set_null(struct swap_cluster_info *info)
266 info->flags = CLUSTER_FLAG_NEXT_NULL;
270 static inline bool cluster_is_huge(struct swap_cluster_info *info)
272 if (IS_ENABLED(CONFIG_THP_SWAP))
273 return info->flags & CLUSTER_FLAG_HUGE;
277 static inline void cluster_clear_huge(struct swap_cluster_info *info)
279 info->flags &= ~CLUSTER_FLAG_HUGE;
282 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
283 unsigned long offset)
285 struct swap_cluster_info *ci;
287 ci = si->cluster_info;
289 ci += offset / SWAPFILE_CLUSTER;
290 spin_lock(&ci->lock);
295 static inline void unlock_cluster(struct swap_cluster_info *ci)
298 spin_unlock(&ci->lock);
302 * Determine the locking method in use for this device. Return
303 * swap_cluster_info if SSD-style cluster-based locking is in place.
305 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
306 struct swap_info_struct *si, unsigned long offset)
308 struct swap_cluster_info *ci;
310 /* Try to use fine-grained SSD-style locking if available: */
311 ci = lock_cluster(si, offset);
312 /* Otherwise, fall back to traditional, coarse locking: */
314 spin_lock(&si->lock);
319 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
320 struct swap_cluster_info *ci)
325 spin_unlock(&si->lock);
328 static inline bool cluster_list_empty(struct swap_cluster_list *list)
330 return cluster_is_null(&list->head);
333 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
335 return cluster_next(&list->head);
338 static void cluster_list_init(struct swap_cluster_list *list)
340 cluster_set_null(&list->head);
341 cluster_set_null(&list->tail);
344 static void cluster_list_add_tail(struct swap_cluster_list *list,
345 struct swap_cluster_info *ci,
348 if (cluster_list_empty(list)) {
349 cluster_set_next_flag(&list->head, idx, 0);
350 cluster_set_next_flag(&list->tail, idx, 0);
352 struct swap_cluster_info *ci_tail;
353 unsigned int tail = cluster_next(&list->tail);
356 * Nested cluster lock, but both cluster locks are
357 * only acquired when we held swap_info_struct->lock
360 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
361 cluster_set_next(ci_tail, idx);
362 spin_unlock(&ci_tail->lock);
363 cluster_set_next_flag(&list->tail, idx, 0);
367 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
368 struct swap_cluster_info *ci)
372 idx = cluster_next(&list->head);
373 if (cluster_next(&list->tail) == idx) {
374 cluster_set_null(&list->head);
375 cluster_set_null(&list->tail);
377 cluster_set_next_flag(&list->head,
378 cluster_next(&ci[idx]), 0);
383 /* Add a cluster to discard list and schedule it to do discard */
384 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
388 * If scan_swap_map() can't find a free cluster, it will check
389 * si->swap_map directly. To make sure the discarding cluster isn't
390 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
391 * will be cleared after discard
393 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
394 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
396 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
398 schedule_work(&si->discard_work);
401 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
403 struct swap_cluster_info *ci = si->cluster_info;
405 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
406 cluster_list_add_tail(&si->free_clusters, ci, idx);
410 * Doing discard actually. After a cluster discard is finished, the cluster
411 * will be added to free cluster list. caller should hold si->lock.
413 static void swap_do_scheduled_discard(struct swap_info_struct *si)
415 struct swap_cluster_info *info, *ci;
418 info = si->cluster_info;
420 while (!cluster_list_empty(&si->discard_clusters)) {
421 idx = cluster_list_del_first(&si->discard_clusters, info);
422 spin_unlock(&si->lock);
424 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
427 spin_lock(&si->lock);
428 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
429 __free_cluster(si, idx);
430 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
431 0, SWAPFILE_CLUSTER);
436 static void swap_discard_work(struct work_struct *work)
438 struct swap_info_struct *si;
440 si = container_of(work, struct swap_info_struct, discard_work);
442 spin_lock(&si->lock);
443 swap_do_scheduled_discard(si);
444 spin_unlock(&si->lock);
447 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
449 struct swap_cluster_info *ci = si->cluster_info;
451 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
452 cluster_list_del_first(&si->free_clusters, ci);
453 cluster_set_count_flag(ci + idx, 0, 0);
456 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
458 struct swap_cluster_info *ci = si->cluster_info + idx;
460 VM_BUG_ON(cluster_count(ci) != 0);
462 * If the swap is discardable, prepare discard the cluster
463 * instead of free it immediately. The cluster will be freed
466 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
467 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
468 swap_cluster_schedule_discard(si, idx);
472 __free_cluster(si, idx);
476 * The cluster corresponding to page_nr will be used. The cluster will be
477 * removed from free cluster list and its usage counter will be increased.
479 static void inc_cluster_info_page(struct swap_info_struct *p,
480 struct swap_cluster_info *cluster_info, unsigned long page_nr)
482 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
486 if (cluster_is_free(&cluster_info[idx]))
487 alloc_cluster(p, idx);
489 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
490 cluster_set_count(&cluster_info[idx],
491 cluster_count(&cluster_info[idx]) + 1);
495 * The cluster corresponding to page_nr decreases one usage. If the usage
496 * counter becomes 0, which means no page in the cluster is in using, we can
497 * optionally discard the cluster and add it to free cluster list.
499 static void dec_cluster_info_page(struct swap_info_struct *p,
500 struct swap_cluster_info *cluster_info, unsigned long page_nr)
502 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
507 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
508 cluster_set_count(&cluster_info[idx],
509 cluster_count(&cluster_info[idx]) - 1);
511 if (cluster_count(&cluster_info[idx]) == 0)
512 free_cluster(p, idx);
516 * It's possible scan_swap_map() uses a free cluster in the middle of free
517 * cluster list. Avoiding such abuse to avoid list corruption.
520 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
521 unsigned long offset)
523 struct percpu_cluster *percpu_cluster;
526 offset /= SWAPFILE_CLUSTER;
527 conflict = !cluster_list_empty(&si->free_clusters) &&
528 offset != cluster_list_first(&si->free_clusters) &&
529 cluster_is_free(&si->cluster_info[offset]);
534 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
535 cluster_set_null(&percpu_cluster->index);
540 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
541 * might involve allocating a new cluster for current CPU too.
543 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
544 unsigned long *offset, unsigned long *scan_base)
546 struct percpu_cluster *cluster;
547 struct swap_cluster_info *ci;
549 unsigned long tmp, max;
552 cluster = this_cpu_ptr(si->percpu_cluster);
553 if (cluster_is_null(&cluster->index)) {
554 if (!cluster_list_empty(&si->free_clusters)) {
555 cluster->index = si->free_clusters.head;
556 cluster->next = cluster_next(&cluster->index) *
558 } else if (!cluster_list_empty(&si->discard_clusters)) {
560 * we don't have free cluster but have some clusters in
561 * discarding, do discard now and reclaim them
563 swap_do_scheduled_discard(si);
564 *scan_base = *offset = si->cluster_next;
573 * Other CPUs can use our cluster if they can't find a free cluster,
574 * check if there is still free entry in the cluster
577 max = min_t(unsigned long, si->max,
578 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
580 cluster_set_null(&cluster->index);
583 ci = lock_cluster(si, tmp);
585 if (!si->swap_map[tmp]) {
593 cluster_set_null(&cluster->index);
596 cluster->next = tmp + 1;
602 static void __del_from_avail_list(struct swap_info_struct *p)
607 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
610 static void del_from_avail_list(struct swap_info_struct *p)
612 spin_lock(&swap_avail_lock);
613 __del_from_avail_list(p);
614 spin_unlock(&swap_avail_lock);
617 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
618 unsigned int nr_entries)
620 unsigned int end = offset + nr_entries - 1;
622 if (offset == si->lowest_bit)
623 si->lowest_bit += nr_entries;
624 if (end == si->highest_bit)
625 si->highest_bit -= nr_entries;
626 si->inuse_pages += nr_entries;
627 if (si->inuse_pages == si->pages) {
628 si->lowest_bit = si->max;
630 del_from_avail_list(si);
634 static void add_to_avail_list(struct swap_info_struct *p)
638 spin_lock(&swap_avail_lock);
640 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
641 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
643 spin_unlock(&swap_avail_lock);
646 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
647 unsigned int nr_entries)
649 unsigned long end = offset + nr_entries - 1;
650 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
652 if (offset < si->lowest_bit)
653 si->lowest_bit = offset;
654 if (end > si->highest_bit) {
655 bool was_full = !si->highest_bit;
657 si->highest_bit = end;
658 if (was_full && (si->flags & SWP_WRITEOK))
659 add_to_avail_list(si);
661 atomic_long_add(nr_entries, &nr_swap_pages);
662 si->inuse_pages -= nr_entries;
663 if (si->flags & SWP_BLKDEV)
664 swap_slot_free_notify =
665 si->bdev->bd_disk->fops->swap_slot_free_notify;
667 swap_slot_free_notify = NULL;
668 while (offset <= end) {
669 frontswap_invalidate_page(si->type, offset);
670 if (swap_slot_free_notify)
671 swap_slot_free_notify(si->bdev, offset);
676 static int scan_swap_map_slots(struct swap_info_struct *si,
677 unsigned char usage, int nr,
680 struct swap_cluster_info *ci;
681 unsigned long offset;
682 unsigned long scan_base;
683 unsigned long last_in_cluster = 0;
684 int latency_ration = LATENCY_LIMIT;
691 * We try to cluster swap pages by allocating them sequentially
692 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
693 * way, however, we resort to first-free allocation, starting
694 * a new cluster. This prevents us from scattering swap pages
695 * all over the entire swap partition, so that we reduce
696 * overall disk seek times between swap pages. -- sct
697 * But we do now try to find an empty cluster. -Andrea
698 * And we let swap pages go all over an SSD partition. Hugh
701 si->flags += SWP_SCANNING;
702 scan_base = offset = si->cluster_next;
705 if (si->cluster_info) {
706 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
712 if (unlikely(!si->cluster_nr--)) {
713 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
714 si->cluster_nr = SWAPFILE_CLUSTER - 1;
718 spin_unlock(&si->lock);
721 * If seek is expensive, start searching for new cluster from
722 * start of partition, to minimize the span of allocated swap.
723 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
724 * case, just handled by scan_swap_map_try_ssd_cluster() above.
726 scan_base = offset = si->lowest_bit;
727 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
729 /* Locate the first empty (unaligned) cluster */
730 for (; last_in_cluster <= si->highest_bit; offset++) {
731 if (si->swap_map[offset])
732 last_in_cluster = offset + SWAPFILE_CLUSTER;
733 else if (offset == last_in_cluster) {
734 spin_lock(&si->lock);
735 offset -= SWAPFILE_CLUSTER - 1;
736 si->cluster_next = offset;
737 si->cluster_nr = SWAPFILE_CLUSTER - 1;
740 if (unlikely(--latency_ration < 0)) {
742 latency_ration = LATENCY_LIMIT;
747 spin_lock(&si->lock);
748 si->cluster_nr = SWAPFILE_CLUSTER - 1;
752 if (si->cluster_info) {
753 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
754 /* take a break if we already got some slots */
757 if (!scan_swap_map_try_ssd_cluster(si, &offset,
762 if (!(si->flags & SWP_WRITEOK))
764 if (!si->highest_bit)
766 if (offset > si->highest_bit)
767 scan_base = offset = si->lowest_bit;
769 ci = lock_cluster(si, offset);
770 /* reuse swap entry of cache-only swap if not busy. */
771 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
774 spin_unlock(&si->lock);
775 swap_was_freed = __try_to_reclaim_swap(si, offset);
776 spin_lock(&si->lock);
777 /* entry was freed successfully, try to use this again */
780 goto scan; /* check next one */
783 if (si->swap_map[offset]) {
790 si->swap_map[offset] = usage;
791 inc_cluster_info_page(si, si->cluster_info, offset);
794 swap_range_alloc(si, offset, 1);
795 si->cluster_next = offset + 1;
796 slots[n_ret++] = swp_entry(si->type, offset);
798 /* got enough slots or reach max slots? */
799 if ((n_ret == nr) || (offset >= si->highest_bit))
802 /* search for next available slot */
804 /* time to take a break? */
805 if (unlikely(--latency_ration < 0)) {
808 spin_unlock(&si->lock);
810 spin_lock(&si->lock);
811 latency_ration = LATENCY_LIMIT;
814 /* try to get more slots in cluster */
815 if (si->cluster_info) {
816 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
824 /* non-ssd case, still more slots in cluster? */
825 if (si->cluster_nr && !si->swap_map[offset]) {
831 si->flags -= SWP_SCANNING;
835 spin_unlock(&si->lock);
836 while (++offset <= si->highest_bit) {
837 if (!si->swap_map[offset]) {
838 spin_lock(&si->lock);
841 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
842 spin_lock(&si->lock);
845 if (unlikely(--latency_ration < 0)) {
847 latency_ration = LATENCY_LIMIT;
850 offset = si->lowest_bit;
851 while (offset < scan_base) {
852 if (!si->swap_map[offset]) {
853 spin_lock(&si->lock);
856 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
857 spin_lock(&si->lock);
860 if (unlikely(--latency_ration < 0)) {
862 latency_ration = LATENCY_LIMIT;
866 spin_lock(&si->lock);
869 si->flags -= SWP_SCANNING;
873 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
876 struct swap_cluster_info *ci;
877 unsigned long offset, i;
881 * Should not even be attempting cluster allocations when huge
882 * page swap is disabled. Warn and fail the allocation.
884 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
889 if (cluster_list_empty(&si->free_clusters))
892 idx = cluster_list_first(&si->free_clusters);
893 offset = idx * SWAPFILE_CLUSTER;
894 ci = lock_cluster(si, offset);
895 alloc_cluster(si, idx);
896 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
898 map = si->swap_map + offset;
899 for (i = 0; i < SWAPFILE_CLUSTER; i++)
900 map[i] = SWAP_HAS_CACHE;
902 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
903 *slot = swp_entry(si->type, offset);
908 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
910 unsigned long offset = idx * SWAPFILE_CLUSTER;
911 struct swap_cluster_info *ci;
913 ci = lock_cluster(si, offset);
914 cluster_set_count_flag(ci, 0, 0);
915 free_cluster(si, idx);
917 swap_range_free(si, offset, SWAPFILE_CLUSTER);
920 static unsigned long scan_swap_map(struct swap_info_struct *si,
926 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
929 return swp_offset(entry);
935 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
937 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
938 struct swap_info_struct *si, *next;
943 /* Only single cluster request supported */
944 WARN_ON_ONCE(n_goal > 1 && cluster);
946 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
950 if (n_goal > SWAP_BATCH)
953 if (n_goal > avail_pgs)
956 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
958 spin_lock(&swap_avail_lock);
961 node = numa_node_id();
962 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
963 /* requeue si to after same-priority siblings */
964 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
965 spin_unlock(&swap_avail_lock);
966 spin_lock(&si->lock);
967 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
968 spin_lock(&swap_avail_lock);
969 if (plist_node_empty(&si->avail_lists[node])) {
970 spin_unlock(&si->lock);
973 WARN(!si->highest_bit,
974 "swap_info %d in list but !highest_bit\n",
976 WARN(!(si->flags & SWP_WRITEOK),
977 "swap_info %d in list but !SWP_WRITEOK\n",
979 __del_from_avail_list(si);
980 spin_unlock(&si->lock);
984 if (!(si->flags & SWP_FILE))
985 n_ret = swap_alloc_cluster(si, swp_entries);
987 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
988 n_goal, swp_entries);
989 spin_unlock(&si->lock);
990 if (n_ret || cluster)
992 pr_debug("scan_swap_map of si %d failed to find offset\n",
995 spin_lock(&swap_avail_lock);
998 * if we got here, it's likely that si was almost full before,
999 * and since scan_swap_map() can drop the si->lock, multiple
1000 * callers probably all tried to get a page from the same si
1001 * and it filled up before we could get one; or, the si filled
1002 * up between us dropping swap_avail_lock and taking si->lock.
1003 * Since we dropped the swap_avail_lock, the swap_avail_head
1004 * list may have been modified; so if next is still in the
1005 * swap_avail_head list then try it, otherwise start over
1006 * if we have not gotten any slots.
1008 if (plist_node_empty(&next->avail_lists[node]))
1012 spin_unlock(&swap_avail_lock);
1016 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1022 /* The only caller of this function is now suspend routine */
1023 swp_entry_t get_swap_page_of_type(int type)
1025 struct swap_info_struct *si;
1028 si = swap_info[type];
1029 spin_lock(&si->lock);
1030 if (si && (si->flags & SWP_WRITEOK)) {
1031 atomic_long_dec(&nr_swap_pages);
1032 /* This is called for allocating swap entry, not cache */
1033 offset = scan_swap_map(si, 1);
1035 spin_unlock(&si->lock);
1036 return swp_entry(type, offset);
1038 atomic_long_inc(&nr_swap_pages);
1040 spin_unlock(&si->lock);
1041 return (swp_entry_t) {0};
1044 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1046 struct swap_info_struct *p;
1047 unsigned long offset, type;
1051 type = swp_type(entry);
1052 if (type >= nr_swapfiles)
1054 p = swap_info[type];
1055 if (!(p->flags & SWP_USED))
1057 offset = swp_offset(entry);
1058 if (offset >= p->max)
1063 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1066 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1069 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1074 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1076 struct swap_info_struct *p;
1078 p = __swap_info_get(entry);
1081 if (!p->swap_map[swp_offset(entry)])
1086 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1092 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1094 struct swap_info_struct *p;
1096 p = _swap_info_get(entry);
1098 spin_lock(&p->lock);
1102 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1103 struct swap_info_struct *q)
1105 struct swap_info_struct *p;
1107 p = _swap_info_get(entry);
1111 spin_unlock(&q->lock);
1113 spin_lock(&p->lock);
1118 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1119 swp_entry_t entry, unsigned char usage)
1121 struct swap_cluster_info *ci;
1122 unsigned long offset = swp_offset(entry);
1123 unsigned char count;
1124 unsigned char has_cache;
1126 ci = lock_cluster_or_swap_info(p, offset);
1128 count = p->swap_map[offset];
1130 has_cache = count & SWAP_HAS_CACHE;
1131 count &= ~SWAP_HAS_CACHE;
1133 if (usage == SWAP_HAS_CACHE) {
1134 VM_BUG_ON(!has_cache);
1136 } else if (count == SWAP_MAP_SHMEM) {
1138 * Or we could insist on shmem.c using a special
1139 * swap_shmem_free() and free_shmem_swap_and_cache()...
1142 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1143 if (count == COUNT_CONTINUED) {
1144 if (swap_count_continued(p, offset, count))
1145 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1147 count = SWAP_MAP_MAX;
1152 usage = count | has_cache;
1153 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1155 unlock_cluster_or_swap_info(p, ci);
1160 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1162 struct swap_cluster_info *ci;
1163 unsigned long offset = swp_offset(entry);
1164 unsigned char count;
1166 ci = lock_cluster(p, offset);
1167 count = p->swap_map[offset];
1168 VM_BUG_ON(count != SWAP_HAS_CACHE);
1169 p->swap_map[offset] = 0;
1170 dec_cluster_info_page(p, p->cluster_info, offset);
1173 mem_cgroup_uncharge_swap(entry, 1);
1174 swap_range_free(p, offset, 1);
1178 * Caller has made sure that the swap device corresponding to entry
1179 * is still around or has not been recycled.
1181 void swap_free(swp_entry_t entry)
1183 struct swap_info_struct *p;
1185 p = _swap_info_get(entry);
1187 if (!__swap_entry_free(p, entry, 1))
1188 free_swap_slot(entry);
1193 * Called after dropping swapcache to decrease refcnt to swap entries.
1195 static void swapcache_free(swp_entry_t entry)
1197 struct swap_info_struct *p;
1199 p = _swap_info_get(entry);
1201 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1202 free_swap_slot(entry);
1206 static void swapcache_free_cluster(swp_entry_t entry)
1208 unsigned long offset = swp_offset(entry);
1209 unsigned long idx = offset / SWAPFILE_CLUSTER;
1210 struct swap_cluster_info *ci;
1211 struct swap_info_struct *si;
1213 unsigned int i, free_entries = 0;
1216 if (!IS_ENABLED(CONFIG_THP_SWAP))
1219 si = _swap_info_get(entry);
1223 ci = lock_cluster(si, offset);
1224 VM_BUG_ON(!cluster_is_huge(ci));
1225 map = si->swap_map + offset;
1226 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1228 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1229 if (val == SWAP_HAS_CACHE)
1232 if (!free_entries) {
1233 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1234 map[i] &= ~SWAP_HAS_CACHE;
1236 cluster_clear_huge(ci);
1238 if (free_entries == SWAPFILE_CLUSTER) {
1239 spin_lock(&si->lock);
1240 ci = lock_cluster(si, offset);
1241 memset(map, 0, SWAPFILE_CLUSTER);
1243 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1244 swap_free_cluster(si, idx);
1245 spin_unlock(&si->lock);
1246 } else if (free_entries) {
1247 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1248 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1249 free_swap_slot(entry);
1254 #ifdef CONFIG_THP_SWAP
1255 int split_swap_cluster(swp_entry_t entry)
1257 struct swap_info_struct *si;
1258 struct swap_cluster_info *ci;
1259 unsigned long offset = swp_offset(entry);
1261 si = _swap_info_get(entry);
1264 ci = lock_cluster(si, offset);
1265 cluster_clear_huge(ci);
1271 void put_swap_page(struct page *page, swp_entry_t entry)
1273 if (!PageTransHuge(page))
1274 swapcache_free(entry);
1276 swapcache_free_cluster(entry);
1279 static int swp_entry_cmp(const void *ent1, const void *ent2)
1281 const swp_entry_t *e1 = ent1, *e2 = ent2;
1283 return (int)swp_type(*e1) - (int)swp_type(*e2);
1286 void swapcache_free_entries(swp_entry_t *entries, int n)
1288 struct swap_info_struct *p, *prev;
1298 * Sort swap entries by swap device, so each lock is only taken once.
1299 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1300 * so low that it isn't necessary to optimize further.
1302 if (nr_swapfiles > 1)
1303 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1304 for (i = 0; i < n; ++i) {
1305 p = swap_info_get_cont(entries[i], prev);
1307 swap_entry_free(p, entries[i]);
1311 spin_unlock(&p->lock);
1315 * How many references to page are currently swapped out?
1316 * This does not give an exact answer when swap count is continued,
1317 * but does include the high COUNT_CONTINUED flag to allow for that.
1319 int page_swapcount(struct page *page)
1322 struct swap_info_struct *p;
1323 struct swap_cluster_info *ci;
1325 unsigned long offset;
1327 entry.val = page_private(page);
1328 p = _swap_info_get(entry);
1330 offset = swp_offset(entry);
1331 ci = lock_cluster_or_swap_info(p, offset);
1332 count = swap_count(p->swap_map[offset]);
1333 unlock_cluster_or_swap_info(p, ci);
1338 int __swap_count(struct swap_info_struct *si, swp_entry_t entry)
1340 pgoff_t offset = swp_offset(entry);
1342 return swap_count(si->swap_map[offset]);
1345 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1348 pgoff_t offset = swp_offset(entry);
1349 struct swap_cluster_info *ci;
1351 ci = lock_cluster_or_swap_info(si, offset);
1352 count = swap_count(si->swap_map[offset]);
1353 unlock_cluster_or_swap_info(si, ci);
1358 * How many references to @entry are currently swapped out?
1359 * This does not give an exact answer when swap count is continued,
1360 * but does include the high COUNT_CONTINUED flag to allow for that.
1362 int __swp_swapcount(swp_entry_t entry)
1365 struct swap_info_struct *si;
1367 si = __swap_info_get(entry);
1369 count = swap_swapcount(si, entry);
1374 * How many references to @entry are currently swapped out?
1375 * This considers COUNT_CONTINUED so it returns exact answer.
1377 int swp_swapcount(swp_entry_t entry)
1379 int count, tmp_count, n;
1380 struct swap_info_struct *p;
1381 struct swap_cluster_info *ci;
1386 p = _swap_info_get(entry);
1390 offset = swp_offset(entry);
1392 ci = lock_cluster_or_swap_info(p, offset);
1394 count = swap_count(p->swap_map[offset]);
1395 if (!(count & COUNT_CONTINUED))
1398 count &= ~COUNT_CONTINUED;
1399 n = SWAP_MAP_MAX + 1;
1401 page = vmalloc_to_page(p->swap_map + offset);
1402 offset &= ~PAGE_MASK;
1403 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1406 page = list_next_entry(page, lru);
1407 map = kmap_atomic(page);
1408 tmp_count = map[offset];
1411 count += (tmp_count & ~COUNT_CONTINUED) * n;
1412 n *= (SWAP_CONT_MAX + 1);
1413 } while (tmp_count & COUNT_CONTINUED);
1415 unlock_cluster_or_swap_info(p, ci);
1419 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1422 struct swap_cluster_info *ci;
1423 unsigned char *map = si->swap_map;
1424 unsigned long roffset = swp_offset(entry);
1425 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1429 ci = lock_cluster_or_swap_info(si, offset);
1430 if (!ci || !cluster_is_huge(ci)) {
1431 if (swap_count(map[roffset]))
1435 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1436 if (swap_count(map[offset + i])) {
1442 unlock_cluster_or_swap_info(si, ci);
1446 static bool page_swapped(struct page *page)
1449 struct swap_info_struct *si;
1451 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1452 return page_swapcount(page) != 0;
1454 page = compound_head(page);
1455 entry.val = page_private(page);
1456 si = _swap_info_get(entry);
1458 return swap_page_trans_huge_swapped(si, entry);
1462 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1463 int *total_swapcount)
1465 int i, map_swapcount, _total_mapcount, _total_swapcount;
1466 unsigned long offset = 0;
1467 struct swap_info_struct *si;
1468 struct swap_cluster_info *ci = NULL;
1469 unsigned char *map = NULL;
1470 int mapcount, swapcount = 0;
1472 /* hugetlbfs shouldn't call it */
1473 VM_BUG_ON_PAGE(PageHuge(page), page);
1475 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1476 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1477 if (PageSwapCache(page))
1478 swapcount = page_swapcount(page);
1479 if (total_swapcount)
1480 *total_swapcount = swapcount;
1481 return mapcount + swapcount;
1484 page = compound_head(page);
1486 _total_mapcount = _total_swapcount = map_swapcount = 0;
1487 if (PageSwapCache(page)) {
1490 entry.val = page_private(page);
1491 si = _swap_info_get(entry);
1494 offset = swp_offset(entry);
1498 ci = lock_cluster(si, offset);
1499 for (i = 0; i < HPAGE_PMD_NR; i++) {
1500 mapcount = atomic_read(&page[i]._mapcount) + 1;
1501 _total_mapcount += mapcount;
1503 swapcount = swap_count(map[offset + i]);
1504 _total_swapcount += swapcount;
1506 map_swapcount = max(map_swapcount, mapcount + swapcount);
1509 if (PageDoubleMap(page)) {
1511 _total_mapcount -= HPAGE_PMD_NR;
1513 mapcount = compound_mapcount(page);
1514 map_swapcount += mapcount;
1515 _total_mapcount += mapcount;
1517 *total_mapcount = _total_mapcount;
1518 if (total_swapcount)
1519 *total_swapcount = _total_swapcount;
1521 return map_swapcount;
1525 * We can write to an anon page without COW if there are no other references
1526 * to it. And as a side-effect, free up its swap: because the old content
1527 * on disk will never be read, and seeking back there to write new content
1528 * later would only waste time away from clustering.
1530 * NOTE: total_map_swapcount should not be relied upon by the caller if
1531 * reuse_swap_page() returns false, but it may be always overwritten
1532 * (see the other implementation for CONFIG_SWAP=n).
1534 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1536 int count, total_mapcount, total_swapcount;
1538 VM_BUG_ON_PAGE(!PageLocked(page), page);
1539 if (unlikely(PageKsm(page)))
1541 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1543 if (total_map_swapcount)
1544 *total_map_swapcount = total_mapcount + total_swapcount;
1545 if (count == 1 && PageSwapCache(page) &&
1546 (likely(!PageTransCompound(page)) ||
1547 /* The remaining swap count will be freed soon */
1548 total_swapcount == page_swapcount(page))) {
1549 if (!PageWriteback(page)) {
1550 page = compound_head(page);
1551 delete_from_swap_cache(page);
1555 struct swap_info_struct *p;
1557 entry.val = page_private(page);
1558 p = swap_info_get(entry);
1559 if (p->flags & SWP_STABLE_WRITES) {
1560 spin_unlock(&p->lock);
1563 spin_unlock(&p->lock);
1571 * If swap is getting full, or if there are no more mappings of this page,
1572 * then try_to_free_swap is called to free its swap space.
1574 int try_to_free_swap(struct page *page)
1576 VM_BUG_ON_PAGE(!PageLocked(page), page);
1578 if (!PageSwapCache(page))
1580 if (PageWriteback(page))
1582 if (page_swapped(page))
1586 * Once hibernation has begun to create its image of memory,
1587 * there's a danger that one of the calls to try_to_free_swap()
1588 * - most probably a call from __try_to_reclaim_swap() while
1589 * hibernation is allocating its own swap pages for the image,
1590 * but conceivably even a call from memory reclaim - will free
1591 * the swap from a page which has already been recorded in the
1592 * image as a clean swapcache page, and then reuse its swap for
1593 * another page of the image. On waking from hibernation, the
1594 * original page might be freed under memory pressure, then
1595 * later read back in from swap, now with the wrong data.
1597 * Hibernation suspends storage while it is writing the image
1598 * to disk so check that here.
1600 if (pm_suspended_storage())
1603 page = compound_head(page);
1604 delete_from_swap_cache(page);
1610 * Free the swap entry like above, but also try to
1611 * free the page cache entry if it is the last user.
1613 int free_swap_and_cache(swp_entry_t entry)
1615 struct swap_info_struct *p;
1616 struct page *page = NULL;
1617 unsigned char count;
1619 if (non_swap_entry(entry))
1622 p = _swap_info_get(entry);
1624 count = __swap_entry_free(p, entry, 1);
1625 if (count == SWAP_HAS_CACHE &&
1626 !swap_page_trans_huge_swapped(p, entry)) {
1627 page = find_get_page(swap_address_space(entry),
1629 if (page && !trylock_page(page)) {
1634 free_swap_slot(entry);
1638 * Not mapped elsewhere, or swap space full? Free it!
1639 * Also recheck PageSwapCache now page is locked (above).
1641 if (PageSwapCache(page) && !PageWriteback(page) &&
1642 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1643 !swap_page_trans_huge_swapped(p, entry)) {
1644 page = compound_head(page);
1645 delete_from_swap_cache(page);
1654 #ifdef CONFIG_HIBERNATION
1656 * Find the swap type that corresponds to given device (if any).
1658 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1659 * from 0, in which the swap header is expected to be located.
1661 * This is needed for the suspend to disk (aka swsusp).
1663 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1665 struct block_device *bdev = NULL;
1669 bdev = bdget(device);
1671 spin_lock(&swap_lock);
1672 for (type = 0; type < nr_swapfiles; type++) {
1673 struct swap_info_struct *sis = swap_info[type];
1675 if (!(sis->flags & SWP_WRITEOK))
1680 *bdev_p = bdgrab(sis->bdev);
1682 spin_unlock(&swap_lock);
1685 if (bdev == sis->bdev) {
1686 struct swap_extent *se = &sis->first_swap_extent;
1688 if (se->start_block == offset) {
1690 *bdev_p = bdgrab(sis->bdev);
1692 spin_unlock(&swap_lock);
1698 spin_unlock(&swap_lock);
1706 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1707 * corresponding to given index in swap_info (swap type).
1709 sector_t swapdev_block(int type, pgoff_t offset)
1711 struct block_device *bdev;
1713 if ((unsigned int)type >= nr_swapfiles)
1715 if (!(swap_info[type]->flags & SWP_WRITEOK))
1717 return map_swap_entry(swp_entry(type, offset), &bdev);
1721 * Return either the total number of swap pages of given type, or the number
1722 * of free pages of that type (depending on @free)
1724 * This is needed for software suspend
1726 unsigned int count_swap_pages(int type, int free)
1730 spin_lock(&swap_lock);
1731 if ((unsigned int)type < nr_swapfiles) {
1732 struct swap_info_struct *sis = swap_info[type];
1734 spin_lock(&sis->lock);
1735 if (sis->flags & SWP_WRITEOK) {
1738 n -= sis->inuse_pages;
1740 spin_unlock(&sis->lock);
1742 spin_unlock(&swap_lock);
1745 #endif /* CONFIG_HIBERNATION */
1747 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1749 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1753 * No need to decide whether this PTE shares the swap entry with others,
1754 * just let do_wp_page work it out if a write is requested later - to
1755 * force COW, vm_page_prot omits write permission from any private vma.
1757 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1758 unsigned long addr, swp_entry_t entry, struct page *page)
1760 struct page *swapcache;
1761 struct mem_cgroup *memcg;
1767 page = ksm_might_need_to_copy(page, vma, addr);
1768 if (unlikely(!page))
1771 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1777 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1778 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1779 mem_cgroup_cancel_charge(page, memcg, false);
1784 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1785 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1787 set_pte_at(vma->vm_mm, addr, pte,
1788 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1789 if (page == swapcache) {
1790 page_add_anon_rmap(page, vma, addr, false);
1791 mem_cgroup_commit_charge(page, memcg, true, false);
1792 } else { /* ksm created a completely new copy */
1793 page_add_new_anon_rmap(page, vma, addr, false);
1794 mem_cgroup_commit_charge(page, memcg, false, false);
1795 lru_cache_add_active_or_unevictable(page, vma);
1799 * Move the page to the active list so it is not
1800 * immediately swapped out again after swapon.
1802 activate_page(page);
1804 pte_unmap_unlock(pte, ptl);
1806 if (page != swapcache) {
1813 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1814 unsigned long addr, unsigned long end,
1815 swp_entry_t entry, struct page *page)
1817 pte_t swp_pte = swp_entry_to_pte(entry);
1822 * We don't actually need pte lock while scanning for swp_pte: since
1823 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1824 * page table while we're scanning; though it could get zapped, and on
1825 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1826 * of unmatched parts which look like swp_pte, so unuse_pte must
1827 * recheck under pte lock. Scanning without pte lock lets it be
1828 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1830 pte = pte_offset_map(pmd, addr);
1833 * swapoff spends a _lot_ of time in this loop!
1834 * Test inline before going to call unuse_pte.
1836 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1838 ret = unuse_pte(vma, pmd, addr, entry, page);
1841 pte = pte_offset_map(pmd, addr);
1843 } while (pte++, addr += PAGE_SIZE, addr != end);
1849 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1850 unsigned long addr, unsigned long end,
1851 swp_entry_t entry, struct page *page)
1857 pmd = pmd_offset(pud, addr);
1860 next = pmd_addr_end(addr, end);
1861 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1863 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1866 } while (pmd++, addr = next, addr != end);
1870 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1871 unsigned long addr, unsigned long end,
1872 swp_entry_t entry, struct page *page)
1878 pud = pud_offset(p4d, addr);
1880 next = pud_addr_end(addr, end);
1881 if (pud_none_or_clear_bad(pud))
1883 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1886 } while (pud++, addr = next, addr != end);
1890 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1891 unsigned long addr, unsigned long end,
1892 swp_entry_t entry, struct page *page)
1898 p4d = p4d_offset(pgd, addr);
1900 next = p4d_addr_end(addr, end);
1901 if (p4d_none_or_clear_bad(p4d))
1903 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1906 } while (p4d++, addr = next, addr != end);
1910 static int unuse_vma(struct vm_area_struct *vma,
1911 swp_entry_t entry, struct page *page)
1914 unsigned long addr, end, next;
1917 if (page_anon_vma(page)) {
1918 addr = page_address_in_vma(page, vma);
1919 if (addr == -EFAULT)
1922 end = addr + PAGE_SIZE;
1924 addr = vma->vm_start;
1928 pgd = pgd_offset(vma->vm_mm, addr);
1930 next = pgd_addr_end(addr, end);
1931 if (pgd_none_or_clear_bad(pgd))
1933 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1936 } while (pgd++, addr = next, addr != end);
1940 static int unuse_mm(struct mm_struct *mm,
1941 swp_entry_t entry, struct page *page)
1943 struct vm_area_struct *vma;
1946 if (!down_read_trylock(&mm->mmap_sem)) {
1948 * Activate page so shrink_inactive_list is unlikely to unmap
1949 * its ptes while lock is dropped, so swapoff can make progress.
1951 activate_page(page);
1953 down_read(&mm->mmap_sem);
1956 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1957 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1961 up_read(&mm->mmap_sem);
1962 return (ret < 0)? ret: 0;
1966 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1967 * from current position to next entry still in use.
1968 * Recycle to start on reaching the end, returning 0 when empty.
1970 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1971 unsigned int prev, bool frontswap)
1973 unsigned int max = si->max;
1974 unsigned int i = prev;
1975 unsigned char count;
1978 * No need for swap_lock here: we're just looking
1979 * for whether an entry is in use, not modifying it; false
1980 * hits are okay, and sys_swapoff() has already prevented new
1981 * allocations from this area (while holding swap_lock).
1990 * No entries in use at top of swap_map,
1991 * loop back to start and recheck there.
1997 count = READ_ONCE(si->swap_map[i]);
1998 if (count && swap_count(count) != SWAP_MAP_BAD)
1999 if (!frontswap || frontswap_test(si, i))
2001 if ((i % LATENCY_LIMIT) == 0)
2008 * We completely avoid races by reading each swap page in advance,
2009 * and then search for the process using it. All the necessary
2010 * page table adjustments can then be made atomically.
2012 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2013 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2015 int try_to_unuse(unsigned int type, bool frontswap,
2016 unsigned long pages_to_unuse)
2018 struct swap_info_struct *si = swap_info[type];
2019 struct mm_struct *start_mm;
2020 volatile unsigned char *swap_map; /* swap_map is accessed without
2021 * locking. Mark it as volatile
2022 * to prevent compiler doing
2025 unsigned char swcount;
2032 * When searching mms for an entry, a good strategy is to
2033 * start at the first mm we freed the previous entry from
2034 * (though actually we don't notice whether we or coincidence
2035 * freed the entry). Initialize this start_mm with a hold.
2037 * A simpler strategy would be to start at the last mm we
2038 * freed the previous entry from; but that would take less
2039 * advantage of mmlist ordering, which clusters forked mms
2040 * together, child after parent. If we race with dup_mmap(), we
2041 * prefer to resolve parent before child, lest we miss entries
2042 * duplicated after we scanned child: using last mm would invert
2045 start_mm = &init_mm;
2049 * Keep on scanning until all entries have gone. Usually,
2050 * one pass through swap_map is enough, but not necessarily:
2051 * there are races when an instance of an entry might be missed.
2053 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2054 if (signal_pending(current)) {
2060 * Get a page for the entry, using the existing swap
2061 * cache page if there is one. Otherwise, get a clean
2062 * page and read the swap into it.
2064 swap_map = &si->swap_map[i];
2065 entry = swp_entry(type, i);
2066 page = read_swap_cache_async(entry,
2067 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2070 * Either swap_duplicate() failed because entry
2071 * has been freed independently, and will not be
2072 * reused since sys_swapoff() already disabled
2073 * allocation from here, or alloc_page() failed.
2075 swcount = *swap_map;
2077 * We don't hold lock here, so the swap entry could be
2078 * SWAP_MAP_BAD (when the cluster is discarding).
2079 * Instead of fail out, We can just skip the swap
2080 * entry because swapoff will wait for discarding
2083 if (!swcount || swcount == SWAP_MAP_BAD)
2090 * Don't hold on to start_mm if it looks like exiting.
2092 if (atomic_read(&start_mm->mm_users) == 1) {
2094 start_mm = &init_mm;
2099 * Wait for and lock page. When do_swap_page races with
2100 * try_to_unuse, do_swap_page can handle the fault much
2101 * faster than try_to_unuse can locate the entry. This
2102 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2103 * defer to do_swap_page in such a case - in some tests,
2104 * do_swap_page and try_to_unuse repeatedly compete.
2106 wait_on_page_locked(page);
2107 wait_on_page_writeback(page);
2109 wait_on_page_writeback(page);
2112 * Remove all references to entry.
2114 swcount = *swap_map;
2115 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2116 retval = shmem_unuse(entry, page);
2117 /* page has already been unlocked and released */
2122 if (swap_count(swcount) && start_mm != &init_mm)
2123 retval = unuse_mm(start_mm, entry, page);
2125 if (swap_count(*swap_map)) {
2126 int set_start_mm = (*swap_map >= swcount);
2127 struct list_head *p = &start_mm->mmlist;
2128 struct mm_struct *new_start_mm = start_mm;
2129 struct mm_struct *prev_mm = start_mm;
2130 struct mm_struct *mm;
2132 mmget(new_start_mm);
2134 spin_lock(&mmlist_lock);
2135 while (swap_count(*swap_map) && !retval &&
2136 (p = p->next) != &start_mm->mmlist) {
2137 mm = list_entry(p, struct mm_struct, mmlist);
2138 if (!mmget_not_zero(mm))
2140 spin_unlock(&mmlist_lock);
2146 swcount = *swap_map;
2147 if (!swap_count(swcount)) /* any usage ? */
2149 else if (mm == &init_mm)
2152 retval = unuse_mm(mm, entry, page);
2154 if (set_start_mm && *swap_map < swcount) {
2155 mmput(new_start_mm);
2160 spin_lock(&mmlist_lock);
2162 spin_unlock(&mmlist_lock);
2165 start_mm = new_start_mm;
2174 * If a reference remains (rare), we would like to leave
2175 * the page in the swap cache; but try_to_unmap could
2176 * then re-duplicate the entry once we drop page lock,
2177 * so we might loop indefinitely; also, that page could
2178 * not be swapped out to other storage meanwhile. So:
2179 * delete from cache even if there's another reference,
2180 * after ensuring that the data has been saved to disk -
2181 * since if the reference remains (rarer), it will be
2182 * read from disk into another page. Splitting into two
2183 * pages would be incorrect if swap supported "shared
2184 * private" pages, but they are handled by tmpfs files.
2186 * Given how unuse_vma() targets one particular offset
2187 * in an anon_vma, once the anon_vma has been determined,
2188 * this splitting happens to be just what is needed to
2189 * handle where KSM pages have been swapped out: re-reading
2190 * is unnecessarily slow, but we can fix that later on.
2192 if (swap_count(*swap_map) &&
2193 PageDirty(page) && PageSwapCache(page)) {
2194 struct writeback_control wbc = {
2195 .sync_mode = WB_SYNC_NONE,
2198 swap_writepage(compound_head(page), &wbc);
2200 wait_on_page_writeback(page);
2204 * It is conceivable that a racing task removed this page from
2205 * swap cache just before we acquired the page lock at the top,
2206 * or while we dropped it in unuse_mm(). The page might even
2207 * be back in swap cache on another swap area: that we must not
2208 * delete, since it may not have been written out to swap yet.
2210 if (PageSwapCache(page) &&
2211 likely(page_private(page) == entry.val) &&
2212 !page_swapped(page))
2213 delete_from_swap_cache(compound_head(page));
2216 * So we could skip searching mms once swap count went
2217 * to 1, we did not mark any present ptes as dirty: must
2218 * mark page dirty so shrink_page_list will preserve it.
2225 * Make sure that we aren't completely killing
2226 * interactive performance.
2229 if (frontswap && pages_to_unuse > 0) {
2230 if (!--pages_to_unuse)
2240 * After a successful try_to_unuse, if no swap is now in use, we know
2241 * we can empty the mmlist. swap_lock must be held on entry and exit.
2242 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2243 * added to the mmlist just after page_duplicate - before would be racy.
2245 static void drain_mmlist(void)
2247 struct list_head *p, *next;
2250 for (type = 0; type < nr_swapfiles; type++)
2251 if (swap_info[type]->inuse_pages)
2253 spin_lock(&mmlist_lock);
2254 list_for_each_safe(p, next, &init_mm.mmlist)
2256 spin_unlock(&mmlist_lock);
2260 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2261 * corresponds to page offset for the specified swap entry.
2262 * Note that the type of this function is sector_t, but it returns page offset
2263 * into the bdev, not sector offset.
2265 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2267 struct swap_info_struct *sis;
2268 struct swap_extent *start_se;
2269 struct swap_extent *se;
2272 sis = swap_info[swp_type(entry)];
2275 offset = swp_offset(entry);
2276 start_se = sis->curr_swap_extent;
2280 if (se->start_page <= offset &&
2281 offset < (se->start_page + se->nr_pages)) {
2282 return se->start_block + (offset - se->start_page);
2284 se = list_next_entry(se, list);
2285 sis->curr_swap_extent = se;
2286 BUG_ON(se == start_se); /* It *must* be present */
2291 * Returns the page offset into bdev for the specified page's swap entry.
2293 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2296 entry.val = page_private(page);
2297 return map_swap_entry(entry, bdev);
2301 * Free all of a swapdev's extent information
2303 static void destroy_swap_extents(struct swap_info_struct *sis)
2305 while (!list_empty(&sis->first_swap_extent.list)) {
2306 struct swap_extent *se;
2308 se = list_first_entry(&sis->first_swap_extent.list,
2309 struct swap_extent, list);
2310 list_del(&se->list);
2314 if (sis->flags & SWP_FILE) {
2315 struct file *swap_file = sis->swap_file;
2316 struct address_space *mapping = swap_file->f_mapping;
2318 sis->flags &= ~SWP_FILE;
2319 mapping->a_ops->swap_deactivate(swap_file);
2324 * Add a block range (and the corresponding page range) into this swapdev's
2325 * extent list. The extent list is kept sorted in page order.
2327 * This function rather assumes that it is called in ascending page order.
2330 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2331 unsigned long nr_pages, sector_t start_block)
2333 struct swap_extent *se;
2334 struct swap_extent *new_se;
2335 struct list_head *lh;
2337 if (start_page == 0) {
2338 se = &sis->first_swap_extent;
2339 sis->curr_swap_extent = se;
2341 se->nr_pages = nr_pages;
2342 se->start_block = start_block;
2345 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2346 se = list_entry(lh, struct swap_extent, list);
2347 BUG_ON(se->start_page + se->nr_pages != start_page);
2348 if (se->start_block + se->nr_pages == start_block) {
2350 se->nr_pages += nr_pages;
2356 * No merge. Insert a new extent, preserving ordering.
2358 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2361 new_se->start_page = start_page;
2362 new_se->nr_pages = nr_pages;
2363 new_se->start_block = start_block;
2365 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2370 * A `swap extent' is a simple thing which maps a contiguous range of pages
2371 * onto a contiguous range of disk blocks. An ordered list of swap extents
2372 * is built at swapon time and is then used at swap_writepage/swap_readpage
2373 * time for locating where on disk a page belongs.
2375 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2376 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2377 * swap files identically.
2379 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2380 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2381 * swapfiles are handled *identically* after swapon time.
2383 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2384 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2385 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2386 * requirements, they are simply tossed out - we will never use those blocks
2389 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2390 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2391 * which will scribble on the fs.
2393 * The amount of disk space which a single swap extent represents varies.
2394 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2395 * extents in the list. To avoid much list walking, we cache the previous
2396 * search location in `curr_swap_extent', and start new searches from there.
2397 * This is extremely effective. The average number of iterations in
2398 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2400 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2402 struct file *swap_file = sis->swap_file;
2403 struct address_space *mapping = swap_file->f_mapping;
2404 struct inode *inode = mapping->host;
2407 if (S_ISBLK(inode->i_mode)) {
2408 ret = add_swap_extent(sis, 0, sis->max, 0);
2413 if (mapping->a_ops->swap_activate) {
2414 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2416 sis->flags |= SWP_FILE;
2417 ret = add_swap_extent(sis, 0, sis->max, 0);
2423 return generic_swapfile_activate(sis, swap_file, span);
2426 static int swap_node(struct swap_info_struct *p)
2428 struct block_device *bdev;
2433 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2435 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2438 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2439 unsigned char *swap_map,
2440 struct swap_cluster_info *cluster_info)
2447 p->prio = --least_priority;
2449 * the plist prio is negated because plist ordering is
2450 * low-to-high, while swap ordering is high-to-low
2452 p->list.prio = -p->prio;
2455 p->avail_lists[i].prio = -p->prio;
2457 if (swap_node(p) == i)
2458 p->avail_lists[i].prio = 1;
2460 p->avail_lists[i].prio = -p->prio;
2463 p->swap_map = swap_map;
2464 p->cluster_info = cluster_info;
2465 p->flags |= SWP_WRITEOK;
2466 atomic_long_add(p->pages, &nr_swap_pages);
2467 total_swap_pages += p->pages;
2469 assert_spin_locked(&swap_lock);
2471 * both lists are plists, and thus priority ordered.
2472 * swap_active_head needs to be priority ordered for swapoff(),
2473 * which on removal of any swap_info_struct with an auto-assigned
2474 * (i.e. negative) priority increments the auto-assigned priority
2475 * of any lower-priority swap_info_structs.
2476 * swap_avail_head needs to be priority ordered for get_swap_page(),
2477 * which allocates swap pages from the highest available priority
2480 plist_add(&p->list, &swap_active_head);
2481 add_to_avail_list(p);
2484 static void enable_swap_info(struct swap_info_struct *p, int prio,
2485 unsigned char *swap_map,
2486 struct swap_cluster_info *cluster_info,
2487 unsigned long *frontswap_map)
2489 frontswap_init(p->type, frontswap_map);
2490 spin_lock(&swap_lock);
2491 spin_lock(&p->lock);
2492 _enable_swap_info(p, prio, swap_map, cluster_info);
2493 spin_unlock(&p->lock);
2494 spin_unlock(&swap_lock);
2497 static void reinsert_swap_info(struct swap_info_struct *p)
2499 spin_lock(&swap_lock);
2500 spin_lock(&p->lock);
2501 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2502 spin_unlock(&p->lock);
2503 spin_unlock(&swap_lock);
2506 bool has_usable_swap(void)
2510 spin_lock(&swap_lock);
2511 if (plist_head_empty(&swap_active_head))
2513 spin_unlock(&swap_lock);
2517 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2519 struct swap_info_struct *p = NULL;
2520 unsigned char *swap_map;
2521 struct swap_cluster_info *cluster_info;
2522 unsigned long *frontswap_map;
2523 struct file *swap_file, *victim;
2524 struct address_space *mapping;
2525 struct inode *inode;
2526 struct filename *pathname;
2528 unsigned int old_block_size;
2530 if (!capable(CAP_SYS_ADMIN))
2533 BUG_ON(!current->mm);
2535 pathname = getname(specialfile);
2536 if (IS_ERR(pathname))
2537 return PTR_ERR(pathname);
2539 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2540 err = PTR_ERR(victim);
2544 mapping = victim->f_mapping;
2545 spin_lock(&swap_lock);
2546 plist_for_each_entry(p, &swap_active_head, list) {
2547 if (p->flags & SWP_WRITEOK) {
2548 if (p->swap_file->f_mapping == mapping) {
2556 spin_unlock(&swap_lock);
2559 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2560 vm_unacct_memory(p->pages);
2563 spin_unlock(&swap_lock);
2566 del_from_avail_list(p);
2567 spin_lock(&p->lock);
2569 struct swap_info_struct *si = p;
2572 plist_for_each_entry_continue(si, &swap_active_head, list) {
2575 for_each_node(nid) {
2576 if (si->avail_lists[nid].prio != 1)
2577 si->avail_lists[nid].prio--;
2582 plist_del(&p->list, &swap_active_head);
2583 atomic_long_sub(p->pages, &nr_swap_pages);
2584 total_swap_pages -= p->pages;
2585 p->flags &= ~SWP_WRITEOK;
2586 spin_unlock(&p->lock);
2587 spin_unlock(&swap_lock);
2589 disable_swap_slots_cache_lock();
2591 set_current_oom_origin();
2592 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2593 clear_current_oom_origin();
2596 /* re-insert swap space back into swap_list */
2597 reinsert_swap_info(p);
2598 reenable_swap_slots_cache_unlock();
2602 reenable_swap_slots_cache_unlock();
2604 flush_work(&p->discard_work);
2606 destroy_swap_extents(p);
2607 if (p->flags & SWP_CONTINUED)
2608 free_swap_count_continuations(p);
2610 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2611 atomic_dec(&nr_rotate_swap);
2613 mutex_lock(&swapon_mutex);
2614 spin_lock(&swap_lock);
2615 spin_lock(&p->lock);
2618 /* wait for anyone still in scan_swap_map */
2619 p->highest_bit = 0; /* cuts scans short */
2620 while (p->flags >= SWP_SCANNING) {
2621 spin_unlock(&p->lock);
2622 spin_unlock(&swap_lock);
2623 schedule_timeout_uninterruptible(1);
2624 spin_lock(&swap_lock);
2625 spin_lock(&p->lock);
2628 swap_file = p->swap_file;
2629 old_block_size = p->old_block_size;
2630 p->swap_file = NULL;
2632 swap_map = p->swap_map;
2634 cluster_info = p->cluster_info;
2635 p->cluster_info = NULL;
2636 frontswap_map = frontswap_map_get(p);
2637 spin_unlock(&p->lock);
2638 spin_unlock(&swap_lock);
2639 frontswap_invalidate_area(p->type);
2640 frontswap_map_set(p, NULL);
2641 mutex_unlock(&swapon_mutex);
2642 free_percpu(p->percpu_cluster);
2643 p->percpu_cluster = NULL;
2645 kvfree(cluster_info);
2646 kvfree(frontswap_map);
2647 /* Destroy swap account information */
2648 swap_cgroup_swapoff(p->type);
2649 exit_swap_address_space(p->type);
2651 inode = mapping->host;
2652 if (S_ISBLK(inode->i_mode)) {
2653 struct block_device *bdev = I_BDEV(inode);
2654 set_blocksize(bdev, old_block_size);
2655 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2658 inode->i_flags &= ~S_SWAPFILE;
2659 inode_unlock(inode);
2661 filp_close(swap_file, NULL);
2664 * Clear the SWP_USED flag after all resources are freed so that swapon
2665 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2666 * not hold p->lock after we cleared its SWP_WRITEOK.
2668 spin_lock(&swap_lock);
2670 spin_unlock(&swap_lock);
2673 atomic_inc(&proc_poll_event);
2674 wake_up_interruptible(&proc_poll_wait);
2677 filp_close(victim, NULL);
2683 #ifdef CONFIG_PROC_FS
2684 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2686 struct seq_file *seq = file->private_data;
2688 poll_wait(file, &proc_poll_wait, wait);
2690 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2691 seq->poll_event = atomic_read(&proc_poll_event);
2692 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2695 return EPOLLIN | EPOLLRDNORM;
2699 static void *swap_start(struct seq_file *swap, loff_t *pos)
2701 struct swap_info_struct *si;
2705 mutex_lock(&swapon_mutex);
2708 return SEQ_START_TOKEN;
2710 for (type = 0; type < nr_swapfiles; type++) {
2711 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2712 si = swap_info[type];
2713 if (!(si->flags & SWP_USED) || !si->swap_map)
2722 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2724 struct swap_info_struct *si = v;
2727 if (v == SEQ_START_TOKEN)
2730 type = si->type + 1;
2732 for (; type < nr_swapfiles; type++) {
2733 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2734 si = swap_info[type];
2735 if (!(si->flags & SWP_USED) || !si->swap_map)
2744 static void swap_stop(struct seq_file *swap, void *v)
2746 mutex_unlock(&swapon_mutex);
2749 static int swap_show(struct seq_file *swap, void *v)
2751 struct swap_info_struct *si = v;
2755 if (si == SEQ_START_TOKEN) {
2756 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2760 file = si->swap_file;
2761 len = seq_file_path(swap, file, " \t\n\\");
2762 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2763 len < 40 ? 40 - len : 1, " ",
2764 S_ISBLK(file_inode(file)->i_mode) ?
2765 "partition" : "file\t",
2766 si->pages << (PAGE_SHIFT - 10),
2767 si->inuse_pages << (PAGE_SHIFT - 10),
2772 static const struct seq_operations swaps_op = {
2773 .start = swap_start,
2779 static int swaps_open(struct inode *inode, struct file *file)
2781 struct seq_file *seq;
2784 ret = seq_open(file, &swaps_op);
2788 seq = file->private_data;
2789 seq->poll_event = atomic_read(&proc_poll_event);
2793 static const struct file_operations proc_swaps_operations = {
2796 .llseek = seq_lseek,
2797 .release = seq_release,
2801 static int __init procswaps_init(void)
2803 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2806 __initcall(procswaps_init);
2807 #endif /* CONFIG_PROC_FS */
2809 #ifdef MAX_SWAPFILES_CHECK
2810 static int __init max_swapfiles_check(void)
2812 MAX_SWAPFILES_CHECK();
2815 late_initcall(max_swapfiles_check);
2818 static struct swap_info_struct *alloc_swap_info(void)
2820 struct swap_info_struct *p;
2824 p = kzalloc(sizeof(*p), GFP_KERNEL);
2826 return ERR_PTR(-ENOMEM);
2828 spin_lock(&swap_lock);
2829 for (type = 0; type < nr_swapfiles; type++) {
2830 if (!(swap_info[type]->flags & SWP_USED))
2833 if (type >= MAX_SWAPFILES) {
2834 spin_unlock(&swap_lock);
2836 return ERR_PTR(-EPERM);
2838 if (type >= nr_swapfiles) {
2840 swap_info[type] = p;
2842 * Write swap_info[type] before nr_swapfiles, in case a
2843 * racing procfs swap_start() or swap_next() is reading them.
2844 * (We never shrink nr_swapfiles, we never free this entry.)
2850 p = swap_info[type];
2852 * Do not memset this entry: a racing procfs swap_next()
2853 * would be relying on p->type to remain valid.
2856 INIT_LIST_HEAD(&p->first_swap_extent.list);
2857 plist_node_init(&p->list, 0);
2859 plist_node_init(&p->avail_lists[i], 0);
2860 p->flags = SWP_USED;
2861 spin_unlock(&swap_lock);
2862 spin_lock_init(&p->lock);
2863 spin_lock_init(&p->cont_lock);
2868 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2872 if (S_ISBLK(inode->i_mode)) {
2873 p->bdev = bdgrab(I_BDEV(inode));
2874 error = blkdev_get(p->bdev,
2875 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2880 p->old_block_size = block_size(p->bdev);
2881 error = set_blocksize(p->bdev, PAGE_SIZE);
2884 p->flags |= SWP_BLKDEV;
2885 } else if (S_ISREG(inode->i_mode)) {
2886 p->bdev = inode->i_sb->s_bdev;
2888 if (IS_SWAPFILE(inode))
2898 * Find out how many pages are allowed for a single swap device. There
2899 * are two limiting factors:
2900 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2901 * 2) the number of bits in the swap pte, as defined by the different
2904 * In order to find the largest possible bit mask, a swap entry with
2905 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2906 * decoded to a swp_entry_t again, and finally the swap offset is
2909 * This will mask all the bits from the initial ~0UL mask that can't
2910 * be encoded in either the swp_entry_t or the architecture definition
2913 unsigned long generic_max_swapfile_size(void)
2915 return swp_offset(pte_to_swp_entry(
2916 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2919 /* Can be overridden by an architecture for additional checks. */
2920 __weak unsigned long max_swapfile_size(void)
2922 return generic_max_swapfile_size();
2925 static unsigned long read_swap_header(struct swap_info_struct *p,
2926 union swap_header *swap_header,
2927 struct inode *inode)
2930 unsigned long maxpages;
2931 unsigned long swapfilepages;
2932 unsigned long last_page;
2934 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2935 pr_err("Unable to find swap-space signature\n");
2939 /* swap partition endianess hack... */
2940 if (swab32(swap_header->info.version) == 1) {
2941 swab32s(&swap_header->info.version);
2942 swab32s(&swap_header->info.last_page);
2943 swab32s(&swap_header->info.nr_badpages);
2944 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2946 for (i = 0; i < swap_header->info.nr_badpages; i++)
2947 swab32s(&swap_header->info.badpages[i]);
2949 /* Check the swap header's sub-version */
2950 if (swap_header->info.version != 1) {
2951 pr_warn("Unable to handle swap header version %d\n",
2952 swap_header->info.version);
2957 p->cluster_next = 1;
2960 maxpages = max_swapfile_size();
2961 last_page = swap_header->info.last_page;
2963 pr_warn("Empty swap-file\n");
2966 if (last_page > maxpages) {
2967 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2968 maxpages << (PAGE_SHIFT - 10),
2969 last_page << (PAGE_SHIFT - 10));
2971 if (maxpages > last_page) {
2972 maxpages = last_page + 1;
2973 /* p->max is an unsigned int: don't overflow it */
2974 if ((unsigned int)maxpages == 0)
2975 maxpages = UINT_MAX;
2977 p->highest_bit = maxpages - 1;
2981 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2982 if (swapfilepages && maxpages > swapfilepages) {
2983 pr_warn("Swap area shorter than signature indicates\n");
2986 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2988 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2994 #define SWAP_CLUSTER_INFO_COLS \
2995 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2996 #define SWAP_CLUSTER_SPACE_COLS \
2997 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2998 #define SWAP_CLUSTER_COLS \
2999 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3001 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3002 union swap_header *swap_header,
3003 unsigned char *swap_map,
3004 struct swap_cluster_info *cluster_info,
3005 unsigned long maxpages,
3009 unsigned int nr_good_pages;
3011 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3012 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3013 unsigned long i, idx;
3015 nr_good_pages = maxpages - 1; /* omit header page */
3017 cluster_list_init(&p->free_clusters);
3018 cluster_list_init(&p->discard_clusters);
3020 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3021 unsigned int page_nr = swap_header->info.badpages[i];
3022 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3024 if (page_nr < maxpages) {
3025 swap_map[page_nr] = SWAP_MAP_BAD;
3028 * Haven't marked the cluster free yet, no list
3029 * operation involved
3031 inc_cluster_info_page(p, cluster_info, page_nr);
3035 /* Haven't marked the cluster free yet, no list operation involved */
3036 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3037 inc_cluster_info_page(p, cluster_info, i);
3039 if (nr_good_pages) {
3040 swap_map[0] = SWAP_MAP_BAD;
3042 * Not mark the cluster free yet, no list
3043 * operation involved
3045 inc_cluster_info_page(p, cluster_info, 0);
3047 p->pages = nr_good_pages;
3048 nr_extents = setup_swap_extents(p, span);
3051 nr_good_pages = p->pages;
3053 if (!nr_good_pages) {
3054 pr_warn("Empty swap-file\n");
3063 * Reduce false cache line sharing between cluster_info and
3064 * sharing same address space.
3066 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3067 j = (k + col) % SWAP_CLUSTER_COLS;
3068 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3069 idx = i * SWAP_CLUSTER_COLS + j;
3070 if (idx >= nr_clusters)
3072 if (cluster_count(&cluster_info[idx]))
3074 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3075 cluster_list_add_tail(&p->free_clusters, cluster_info,
3083 * Helper to sys_swapon determining if a given swap
3084 * backing device queue supports DISCARD operations.
3086 static bool swap_discardable(struct swap_info_struct *si)
3088 struct request_queue *q = bdev_get_queue(si->bdev);
3090 if (!q || !blk_queue_discard(q))
3096 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3098 struct swap_info_struct *p;
3099 struct filename *name;
3100 struct file *swap_file = NULL;
3101 struct address_space *mapping;
3104 union swap_header *swap_header;
3107 unsigned long maxpages;
3108 unsigned char *swap_map = NULL;
3109 struct swap_cluster_info *cluster_info = NULL;
3110 unsigned long *frontswap_map = NULL;
3111 struct page *page = NULL;
3112 struct inode *inode = NULL;
3113 bool inced_nr_rotate_swap = false;
3115 if (swap_flags & ~SWAP_FLAGS_VALID)
3118 if (!capable(CAP_SYS_ADMIN))
3121 if (!swap_avail_heads)
3124 p = alloc_swap_info();
3128 INIT_WORK(&p->discard_work, swap_discard_work);
3130 name = getname(specialfile);
3132 error = PTR_ERR(name);
3136 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3137 if (IS_ERR(swap_file)) {
3138 error = PTR_ERR(swap_file);
3143 p->swap_file = swap_file;
3144 mapping = swap_file->f_mapping;
3145 inode = mapping->host;
3147 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3148 error = claim_swapfile(p, inode);
3149 if (unlikely(error))
3153 * Read the swap header.
3155 if (!mapping->a_ops->readpage) {
3159 page = read_mapping_page(mapping, 0, swap_file);
3161 error = PTR_ERR(page);
3164 swap_header = kmap(page);
3166 maxpages = read_swap_header(p, swap_header, inode);
3167 if (unlikely(!maxpages)) {
3172 /* OK, set up the swap map and apply the bad block list */
3173 swap_map = vzalloc(maxpages);
3179 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3180 p->flags |= SWP_STABLE_WRITES;
3182 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3183 p->flags |= SWP_SYNCHRONOUS_IO;
3185 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3187 unsigned long ci, nr_cluster;
3189 p->flags |= SWP_SOLIDSTATE;
3191 * select a random position to start with to help wear leveling
3194 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3195 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3197 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3199 if (!cluster_info) {
3204 for (ci = 0; ci < nr_cluster; ci++)
3205 spin_lock_init(&((cluster_info + ci)->lock));
3207 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3208 if (!p->percpu_cluster) {
3212 for_each_possible_cpu(cpu) {
3213 struct percpu_cluster *cluster;
3214 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3215 cluster_set_null(&cluster->index);
3218 atomic_inc(&nr_rotate_swap);
3219 inced_nr_rotate_swap = true;
3222 error = swap_cgroup_swapon(p->type, maxpages);
3226 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3227 cluster_info, maxpages, &span);
3228 if (unlikely(nr_extents < 0)) {
3232 /* frontswap enabled? set up bit-per-page map for frontswap */
3233 if (IS_ENABLED(CONFIG_FRONTSWAP))
3234 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3238 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3240 * When discard is enabled for swap with no particular
3241 * policy flagged, we set all swap discard flags here in
3242 * order to sustain backward compatibility with older
3243 * swapon(8) releases.
3245 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3249 * By flagging sys_swapon, a sysadmin can tell us to
3250 * either do single-time area discards only, or to just
3251 * perform discards for released swap page-clusters.
3252 * Now it's time to adjust the p->flags accordingly.
3254 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3255 p->flags &= ~SWP_PAGE_DISCARD;
3256 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3257 p->flags &= ~SWP_AREA_DISCARD;
3259 /* issue a swapon-time discard if it's still required */
3260 if (p->flags & SWP_AREA_DISCARD) {
3261 int err = discard_swap(p);
3263 pr_err("swapon: discard_swap(%p): %d\n",
3268 error = init_swap_address_space(p->type, maxpages);
3272 mutex_lock(&swapon_mutex);
3274 if (swap_flags & SWAP_FLAG_PREFER)
3276 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3277 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3279 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3280 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3281 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3282 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3283 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3284 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3285 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3286 (frontswap_map) ? "FS" : "");
3288 mutex_unlock(&swapon_mutex);
3289 atomic_inc(&proc_poll_event);
3290 wake_up_interruptible(&proc_poll_wait);
3292 if (S_ISREG(inode->i_mode))
3293 inode->i_flags |= S_SWAPFILE;
3297 free_percpu(p->percpu_cluster);
3298 p->percpu_cluster = NULL;
3299 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3300 set_blocksize(p->bdev, p->old_block_size);
3301 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3303 destroy_swap_extents(p);
3304 swap_cgroup_swapoff(p->type);
3305 spin_lock(&swap_lock);
3306 p->swap_file = NULL;
3308 spin_unlock(&swap_lock);
3310 kvfree(cluster_info);
3311 kvfree(frontswap_map);
3312 if (inced_nr_rotate_swap)
3313 atomic_dec(&nr_rotate_swap);
3315 if (inode && S_ISREG(inode->i_mode)) {
3316 inode_unlock(inode);
3319 filp_close(swap_file, NULL);
3322 if (page && !IS_ERR(page)) {
3328 if (inode && S_ISREG(inode->i_mode))
3329 inode_unlock(inode);
3331 enable_swap_slots_cache();
3335 void si_swapinfo(struct sysinfo *val)
3338 unsigned long nr_to_be_unused = 0;
3340 spin_lock(&swap_lock);
3341 for (type = 0; type < nr_swapfiles; type++) {
3342 struct swap_info_struct *si = swap_info[type];
3344 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3345 nr_to_be_unused += si->inuse_pages;
3347 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3348 val->totalswap = total_swap_pages + nr_to_be_unused;
3349 spin_unlock(&swap_lock);
3353 * Verify that a swap entry is valid and increment its swap map count.
3355 * Returns error code in following case.
3357 * - swp_entry is invalid -> EINVAL
3358 * - swp_entry is migration entry -> EINVAL
3359 * - swap-cache reference is requested but there is already one. -> EEXIST
3360 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3361 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3363 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3365 struct swap_info_struct *p;
3366 struct swap_cluster_info *ci;
3367 unsigned long offset, type;
3368 unsigned char count;
3369 unsigned char has_cache;
3372 if (non_swap_entry(entry))
3375 type = swp_type(entry);
3376 if (type >= nr_swapfiles)
3378 p = swap_info[type];
3379 offset = swp_offset(entry);
3380 if (unlikely(offset >= p->max))
3383 ci = lock_cluster_or_swap_info(p, offset);
3385 count = p->swap_map[offset];
3388 * swapin_readahead() doesn't check if a swap entry is valid, so the
3389 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3391 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3396 has_cache = count & SWAP_HAS_CACHE;
3397 count &= ~SWAP_HAS_CACHE;
3400 if (usage == SWAP_HAS_CACHE) {
3402 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3403 if (!has_cache && count)
3404 has_cache = SWAP_HAS_CACHE;
3405 else if (has_cache) /* someone else added cache */
3407 else /* no users remaining */
3410 } else if (count || has_cache) {
3412 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3414 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3416 else if (swap_count_continued(p, offset, count))
3417 count = COUNT_CONTINUED;
3421 err = -ENOENT; /* unused swap entry */
3423 p->swap_map[offset] = count | has_cache;
3426 unlock_cluster_or_swap_info(p, ci);
3431 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3436 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3437 * (in which case its reference count is never incremented).
3439 void swap_shmem_alloc(swp_entry_t entry)
3441 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3445 * Increase reference count of swap entry by 1.
3446 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3447 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3448 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3449 * might occur if a page table entry has got corrupted.
3451 int swap_duplicate(swp_entry_t entry)
3455 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3456 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3461 * @entry: swap entry for which we allocate swap cache.
3463 * Called when allocating swap cache for existing swap entry,
3464 * This can return error codes. Returns 0 at success.
3465 * -EBUSY means there is a swap cache.
3466 * Note: return code is different from swap_duplicate().
3468 int swapcache_prepare(swp_entry_t entry)
3470 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3473 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3475 return swap_info[swp_type(entry)];
3478 struct swap_info_struct *page_swap_info(struct page *page)
3480 swp_entry_t entry = { .val = page_private(page) };
3481 return swp_swap_info(entry);
3485 * out-of-line __page_file_ methods to avoid include hell.
3487 struct address_space *__page_file_mapping(struct page *page)
3489 return page_swap_info(page)->swap_file->f_mapping;
3491 EXPORT_SYMBOL_GPL(__page_file_mapping);
3493 pgoff_t __page_file_index(struct page *page)
3495 swp_entry_t swap = { .val = page_private(page) };
3496 return swp_offset(swap);
3498 EXPORT_SYMBOL_GPL(__page_file_index);
3501 * add_swap_count_continuation - called when a swap count is duplicated
3502 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3503 * page of the original vmalloc'ed swap_map, to hold the continuation count
3504 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3505 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3507 * These continuation pages are seldom referenced: the common paths all work
3508 * on the original swap_map, only referring to a continuation page when the
3509 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3511 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3512 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3513 * can be called after dropping locks.
3515 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3517 struct swap_info_struct *si;
3518 struct swap_cluster_info *ci;
3521 struct page *list_page;
3523 unsigned char count;
3526 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3527 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3529 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3531 si = swap_info_get(entry);
3534 * An acceptable race has occurred since the failing
3535 * __swap_duplicate(): the swap entry has been freed,
3536 * perhaps even the whole swap_map cleared for swapoff.
3541 offset = swp_offset(entry);
3543 ci = lock_cluster(si, offset);
3545 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3547 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3549 * The higher the swap count, the more likely it is that tasks
3550 * will race to add swap count continuation: we need to avoid
3551 * over-provisioning.
3558 spin_unlock(&si->lock);
3563 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3564 * no architecture is using highmem pages for kernel page tables: so it
3565 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3567 head = vmalloc_to_page(si->swap_map + offset);
3568 offset &= ~PAGE_MASK;
3570 spin_lock(&si->cont_lock);
3572 * Page allocation does not initialize the page's lru field,
3573 * but it does always reset its private field.
3575 if (!page_private(head)) {
3576 BUG_ON(count & COUNT_CONTINUED);
3577 INIT_LIST_HEAD(&head->lru);
3578 set_page_private(head, SWP_CONTINUED);
3579 si->flags |= SWP_CONTINUED;
3582 list_for_each_entry(list_page, &head->lru, lru) {
3586 * If the previous map said no continuation, but we've found
3587 * a continuation page, free our allocation and use this one.
3589 if (!(count & COUNT_CONTINUED))
3590 goto out_unlock_cont;
3592 map = kmap_atomic(list_page) + offset;
3597 * If this continuation count now has some space in it,
3598 * free our allocation and use this one.
3600 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3601 goto out_unlock_cont;
3604 list_add_tail(&page->lru, &head->lru);
3605 page = NULL; /* now it's attached, don't free it */
3607 spin_unlock(&si->cont_lock);
3610 spin_unlock(&si->lock);
3618 * swap_count_continued - when the original swap_map count is incremented
3619 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3620 * into, carry if so, or else fail until a new continuation page is allocated;
3621 * when the original swap_map count is decremented from 0 with continuation,
3622 * borrow from the continuation and report whether it still holds more.
3623 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3626 static bool swap_count_continued(struct swap_info_struct *si,
3627 pgoff_t offset, unsigned char count)
3634 head = vmalloc_to_page(si->swap_map + offset);
3635 if (page_private(head) != SWP_CONTINUED) {
3636 BUG_ON(count & COUNT_CONTINUED);
3637 return false; /* need to add count continuation */
3640 spin_lock(&si->cont_lock);
3641 offset &= ~PAGE_MASK;
3642 page = list_entry(head->lru.next, struct page, lru);
3643 map = kmap_atomic(page) + offset;
3645 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3646 goto init_map; /* jump over SWAP_CONT_MAX checks */
3648 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3650 * Think of how you add 1 to 999
3652 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3654 page = list_entry(page->lru.next, struct page, lru);
3655 BUG_ON(page == head);
3656 map = kmap_atomic(page) + offset;
3658 if (*map == SWAP_CONT_MAX) {
3660 page = list_entry(page->lru.next, struct page, lru);
3662 ret = false; /* add count continuation */
3665 map = kmap_atomic(page) + offset;
3666 init_map: *map = 0; /* we didn't zero the page */
3670 page = list_entry(page->lru.prev, struct page, lru);
3671 while (page != head) {
3672 map = kmap_atomic(page) + offset;
3673 *map = COUNT_CONTINUED;
3675 page = list_entry(page->lru.prev, struct page, lru);
3677 ret = true; /* incremented */
3679 } else { /* decrementing */
3681 * Think of how you subtract 1 from 1000
3683 BUG_ON(count != COUNT_CONTINUED);
3684 while (*map == COUNT_CONTINUED) {
3686 page = list_entry(page->lru.next, struct page, lru);
3687 BUG_ON(page == head);
3688 map = kmap_atomic(page) + offset;
3695 page = list_entry(page->lru.prev, struct page, lru);
3696 while (page != head) {
3697 map = kmap_atomic(page) + offset;
3698 *map = SWAP_CONT_MAX | count;
3699 count = COUNT_CONTINUED;
3701 page = list_entry(page->lru.prev, struct page, lru);
3703 ret = count == COUNT_CONTINUED;
3706 spin_unlock(&si->cont_lock);
3711 * free_swap_count_continuations - swapoff free all the continuation pages
3712 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3714 static void free_swap_count_continuations(struct swap_info_struct *si)
3718 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3720 head = vmalloc_to_page(si->swap_map + offset);
3721 if (page_private(head)) {
3722 struct page *page, *next;
3724 list_for_each_entry_safe(page, next, &head->lru, lru) {
3725 list_del(&page->lru);
3732 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3733 void mem_cgroup_throttle_swaprate(struct mem_cgroup *memcg, int node,
3736 struct swap_info_struct *si, *next;
3737 if (!(gfp_mask & __GFP_IO) || !memcg)
3740 if (!blk_cgroup_congested())
3744 * We've already scheduled a throttle, avoid taking the global swap
3747 if (current->throttle_queue)
3750 spin_lock(&swap_avail_lock);
3751 plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
3752 avail_lists[node]) {
3754 blkcg_schedule_throttle(bdev_get_queue(si->bdev),
3759 spin_unlock(&swap_avail_lock);
3763 static int __init swapfile_init(void)
3767 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3769 if (!swap_avail_heads) {
3770 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3775 plist_head_init(&swap_avail_heads[nid]);
3779 subsys_initcall(swapfile_init);