1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
9 #include <linux/blkdev.h>
11 #include <linux/sched/mm.h>
12 #include <linux/sched/task.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mman.h>
15 #include <linux/slab.h>
16 #include <linux/kernel_stat.h>
17 #include <linux/swap.h>
18 #include <linux/vmalloc.h>
19 #include <linux/pagemap.h>
20 #include <linux/namei.h>
21 #include <linux/shmem_fs.h>
22 #include <linux/blk-cgroup.h>
23 #include <linux/random.h>
24 #include <linux/writeback.h>
25 #include <linux/proc_fs.h>
26 #include <linux/seq_file.h>
27 #include <linux/init.h>
28 #include <linux/ksm.h>
29 #include <linux/rmap.h>
30 #include <linux/security.h>
31 #include <linux/backing-dev.h>
32 #include <linux/mutex.h>
33 #include <linux/capability.h>
34 #include <linux/syscalls.h>
35 #include <linux/memcontrol.h>
36 #include <linux/poll.h>
37 #include <linux/oom.h>
38 #include <linux/frontswap.h>
39 #include <linux/swapfile.h>
40 #include <linux/export.h>
41 #include <linux/swap_slots.h>
42 #include <linux/sort.h>
43 #include <linux/completion.h>
45 #include <asm/tlbflush.h>
46 #include <linux/swapops.h>
47 #include <linux/swap_cgroup.h>
50 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
52 static void free_swap_count_continuations(struct swap_info_struct *);
54 static DEFINE_SPINLOCK(swap_lock);
55 static unsigned int nr_swapfiles;
56 atomic_long_t nr_swap_pages;
58 * Some modules use swappable objects and may try to swap them out under
59 * memory pressure (via the shrinker). Before doing so, they may wish to
60 * check to see if any swap space is available.
62 EXPORT_SYMBOL_GPL(nr_swap_pages);
63 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
64 long total_swap_pages;
65 static int least_priority = -1;
67 static const char Bad_file[] = "Bad swap file entry ";
68 static const char Unused_file[] = "Unused swap file entry ";
69 static const char Bad_offset[] = "Bad swap offset entry ";
70 static const char Unused_offset[] = "Unused swap offset entry ";
73 * all active swap_info_structs
74 * protected with swap_lock, and ordered by priority.
76 static PLIST_HEAD(swap_active_head);
79 * all available (active, not full) swap_info_structs
80 * protected with swap_avail_lock, ordered by priority.
81 * This is used by folio_alloc_swap() instead of swap_active_head
82 * because swap_active_head includes all swap_info_structs,
83 * but folio_alloc_swap() doesn't need to look at full ones.
84 * This uses its own lock instead of swap_lock because when a
85 * swap_info_struct changes between not-full/full, it needs to
86 * add/remove itself to/from this list, but the swap_info_struct->lock
87 * is held and the locking order requires swap_lock to be taken
88 * before any swap_info_struct->lock.
90 static struct plist_head *swap_avail_heads;
91 static DEFINE_SPINLOCK(swap_avail_lock);
93 struct swap_info_struct *swap_info[MAX_SWAPFILES];
95 static DEFINE_MUTEX(swapon_mutex);
97 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
98 /* Activity counter to indicate that a swapon or swapoff has occurred */
99 static atomic_t proc_poll_event = ATOMIC_INIT(0);
101 atomic_t nr_rotate_swap = ATOMIC_INIT(0);
103 static struct swap_info_struct *swap_type_to_swap_info(int type)
105 if (type >= MAX_SWAPFILES)
108 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
111 static inline unsigned char swap_count(unsigned char ent)
113 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
116 /* Reclaim the swap entry anyway if possible */
117 #define TTRS_ANYWAY 0x1
119 * Reclaim the swap entry if there are no more mappings of the
122 #define TTRS_UNMAPPED 0x2
123 /* Reclaim the swap entry if swap is getting full*/
124 #define TTRS_FULL 0x4
126 /* returns 1 if swap entry is freed */
127 static int __try_to_reclaim_swap(struct swap_info_struct *si,
128 unsigned long offset, unsigned long flags)
130 swp_entry_t entry = swp_entry(si->type, offset);
134 page = find_get_page(swap_address_space(entry), offset);
138 * When this function is called from scan_swap_map_slots() and it's
139 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
140 * here. We have to use trylock for avoiding deadlock. This is a special
141 * case and you should use try_to_free_swap() with explicit lock_page()
142 * in usual operations.
144 if (trylock_page(page)) {
145 if ((flags & TTRS_ANYWAY) ||
146 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
147 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
148 ret = try_to_free_swap(page);
155 static inline struct swap_extent *first_se(struct swap_info_struct *sis)
157 struct rb_node *rb = rb_first(&sis->swap_extent_root);
158 return rb_entry(rb, struct swap_extent, rb_node);
161 static inline struct swap_extent *next_se(struct swap_extent *se)
163 struct rb_node *rb = rb_next(&se->rb_node);
164 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
168 * swapon tell device that all the old swap contents can be discarded,
169 * to allow the swap device to optimize its wear-levelling.
171 static int discard_swap(struct swap_info_struct *si)
173 struct swap_extent *se;
174 sector_t start_block;
178 /* Do not discard the swap header page! */
180 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
181 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
183 err = blkdev_issue_discard(si->bdev, start_block,
184 nr_blocks, GFP_KERNEL);
190 for (se = next_se(se); se; se = next_se(se)) {
191 start_block = se->start_block << (PAGE_SHIFT - 9);
192 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
194 err = blkdev_issue_discard(si->bdev, start_block,
195 nr_blocks, GFP_KERNEL);
201 return err; /* That will often be -EOPNOTSUPP */
204 static struct swap_extent *
205 offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
207 struct swap_extent *se;
210 rb = sis->swap_extent_root.rb_node;
212 se = rb_entry(rb, struct swap_extent, rb_node);
213 if (offset < se->start_page)
215 else if (offset >= se->start_page + se->nr_pages)
220 /* It *must* be present */
224 sector_t swap_page_sector(struct page *page)
226 struct swap_info_struct *sis = page_swap_info(page);
227 struct swap_extent *se;
231 offset = __page_file_index(page);
232 se = offset_to_swap_extent(sis, offset);
233 sector = se->start_block + (offset - se->start_page);
234 return sector << (PAGE_SHIFT - 9);
238 * swap allocation tell device that a cluster of swap can now be discarded,
239 * to allow the swap device to optimize its wear-levelling.
241 static void discard_swap_cluster(struct swap_info_struct *si,
242 pgoff_t start_page, pgoff_t nr_pages)
244 struct swap_extent *se = offset_to_swap_extent(si, start_page);
247 pgoff_t offset = start_page - se->start_page;
248 sector_t start_block = se->start_block + offset;
249 sector_t nr_blocks = se->nr_pages - offset;
251 if (nr_blocks > nr_pages)
252 nr_blocks = nr_pages;
253 start_page += nr_blocks;
254 nr_pages -= nr_blocks;
256 start_block <<= PAGE_SHIFT - 9;
257 nr_blocks <<= PAGE_SHIFT - 9;
258 if (blkdev_issue_discard(si->bdev, start_block,
259 nr_blocks, GFP_NOIO))
266 #ifdef CONFIG_THP_SWAP
267 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
269 #define swap_entry_size(size) (size)
271 #define SWAPFILE_CLUSTER 256
274 * Define swap_entry_size() as constant to let compiler to optimize
275 * out some code if !CONFIG_THP_SWAP
277 #define swap_entry_size(size) 1
279 #define LATENCY_LIMIT 256
281 static inline void cluster_set_flag(struct swap_cluster_info *info,
287 static inline unsigned int cluster_count(struct swap_cluster_info *info)
292 static inline void cluster_set_count(struct swap_cluster_info *info,
298 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
299 unsigned int c, unsigned int f)
305 static inline unsigned int cluster_next(struct swap_cluster_info *info)
310 static inline void cluster_set_next(struct swap_cluster_info *info,
316 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
317 unsigned int n, unsigned int f)
323 static inline bool cluster_is_free(struct swap_cluster_info *info)
325 return info->flags & CLUSTER_FLAG_FREE;
328 static inline bool cluster_is_null(struct swap_cluster_info *info)
330 return info->flags & CLUSTER_FLAG_NEXT_NULL;
333 static inline void cluster_set_null(struct swap_cluster_info *info)
335 info->flags = CLUSTER_FLAG_NEXT_NULL;
339 static inline bool cluster_is_huge(struct swap_cluster_info *info)
341 if (IS_ENABLED(CONFIG_THP_SWAP))
342 return info->flags & CLUSTER_FLAG_HUGE;
346 static inline void cluster_clear_huge(struct swap_cluster_info *info)
348 info->flags &= ~CLUSTER_FLAG_HUGE;
351 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
352 unsigned long offset)
354 struct swap_cluster_info *ci;
356 ci = si->cluster_info;
358 ci += offset / SWAPFILE_CLUSTER;
359 spin_lock(&ci->lock);
364 static inline void unlock_cluster(struct swap_cluster_info *ci)
367 spin_unlock(&ci->lock);
371 * Determine the locking method in use for this device. Return
372 * swap_cluster_info if SSD-style cluster-based locking is in place.
374 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
375 struct swap_info_struct *si, unsigned long offset)
377 struct swap_cluster_info *ci;
379 /* Try to use fine-grained SSD-style locking if available: */
380 ci = lock_cluster(si, offset);
381 /* Otherwise, fall back to traditional, coarse locking: */
383 spin_lock(&si->lock);
388 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
389 struct swap_cluster_info *ci)
394 spin_unlock(&si->lock);
397 static inline bool cluster_list_empty(struct swap_cluster_list *list)
399 return cluster_is_null(&list->head);
402 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
404 return cluster_next(&list->head);
407 static void cluster_list_init(struct swap_cluster_list *list)
409 cluster_set_null(&list->head);
410 cluster_set_null(&list->tail);
413 static void cluster_list_add_tail(struct swap_cluster_list *list,
414 struct swap_cluster_info *ci,
417 if (cluster_list_empty(list)) {
418 cluster_set_next_flag(&list->head, idx, 0);
419 cluster_set_next_flag(&list->tail, idx, 0);
421 struct swap_cluster_info *ci_tail;
422 unsigned int tail = cluster_next(&list->tail);
425 * Nested cluster lock, but both cluster locks are
426 * only acquired when we held swap_info_struct->lock
429 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
430 cluster_set_next(ci_tail, idx);
431 spin_unlock(&ci_tail->lock);
432 cluster_set_next_flag(&list->tail, idx, 0);
436 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
437 struct swap_cluster_info *ci)
441 idx = cluster_next(&list->head);
442 if (cluster_next(&list->tail) == idx) {
443 cluster_set_null(&list->head);
444 cluster_set_null(&list->tail);
446 cluster_set_next_flag(&list->head,
447 cluster_next(&ci[idx]), 0);
452 /* Add a cluster to discard list and schedule it to do discard */
453 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
457 * If scan_swap_map_slots() can't find a free cluster, it will check
458 * si->swap_map directly. To make sure the discarding cluster isn't
459 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
460 * It will be cleared after discard
462 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
463 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
465 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
467 schedule_work(&si->discard_work);
470 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
472 struct swap_cluster_info *ci = si->cluster_info;
474 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
475 cluster_list_add_tail(&si->free_clusters, ci, idx);
479 * Doing discard actually. After a cluster discard is finished, the cluster
480 * will be added to free cluster list. caller should hold si->lock.
482 static void swap_do_scheduled_discard(struct swap_info_struct *si)
484 struct swap_cluster_info *info, *ci;
487 info = si->cluster_info;
489 while (!cluster_list_empty(&si->discard_clusters)) {
490 idx = cluster_list_del_first(&si->discard_clusters, info);
491 spin_unlock(&si->lock);
493 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
496 spin_lock(&si->lock);
497 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
498 __free_cluster(si, idx);
499 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
500 0, SWAPFILE_CLUSTER);
505 static void swap_discard_work(struct work_struct *work)
507 struct swap_info_struct *si;
509 si = container_of(work, struct swap_info_struct, discard_work);
511 spin_lock(&si->lock);
512 swap_do_scheduled_discard(si);
513 spin_unlock(&si->lock);
516 static void swap_users_ref_free(struct percpu_ref *ref)
518 struct swap_info_struct *si;
520 si = container_of(ref, struct swap_info_struct, users);
524 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
526 struct swap_cluster_info *ci = si->cluster_info;
528 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
529 cluster_list_del_first(&si->free_clusters, ci);
530 cluster_set_count_flag(ci + idx, 0, 0);
533 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
535 struct swap_cluster_info *ci = si->cluster_info + idx;
537 VM_BUG_ON(cluster_count(ci) != 0);
539 * If the swap is discardable, prepare discard the cluster
540 * instead of free it immediately. The cluster will be freed
543 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
544 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
545 swap_cluster_schedule_discard(si, idx);
549 __free_cluster(si, idx);
553 * The cluster corresponding to page_nr will be used. The cluster will be
554 * removed from free cluster list and its usage counter will be increased.
556 static void inc_cluster_info_page(struct swap_info_struct *p,
557 struct swap_cluster_info *cluster_info, unsigned long page_nr)
559 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
563 if (cluster_is_free(&cluster_info[idx]))
564 alloc_cluster(p, idx);
566 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
567 cluster_set_count(&cluster_info[idx],
568 cluster_count(&cluster_info[idx]) + 1);
572 * The cluster corresponding to page_nr decreases one usage. If the usage
573 * counter becomes 0, which means no page in the cluster is in using, we can
574 * optionally discard the cluster and add it to free cluster list.
576 static void dec_cluster_info_page(struct swap_info_struct *p,
577 struct swap_cluster_info *cluster_info, unsigned long page_nr)
579 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
584 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
585 cluster_set_count(&cluster_info[idx],
586 cluster_count(&cluster_info[idx]) - 1);
588 if (cluster_count(&cluster_info[idx]) == 0)
589 free_cluster(p, idx);
593 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free
594 * cluster list. Avoiding such abuse to avoid list corruption.
597 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
598 unsigned long offset)
600 struct percpu_cluster *percpu_cluster;
603 offset /= SWAPFILE_CLUSTER;
604 conflict = !cluster_list_empty(&si->free_clusters) &&
605 offset != cluster_list_first(&si->free_clusters) &&
606 cluster_is_free(&si->cluster_info[offset]);
611 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
612 cluster_set_null(&percpu_cluster->index);
617 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
618 * might involve allocating a new cluster for current CPU too.
620 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
621 unsigned long *offset, unsigned long *scan_base)
623 struct percpu_cluster *cluster;
624 struct swap_cluster_info *ci;
625 unsigned long tmp, max;
628 cluster = this_cpu_ptr(si->percpu_cluster);
629 if (cluster_is_null(&cluster->index)) {
630 if (!cluster_list_empty(&si->free_clusters)) {
631 cluster->index = si->free_clusters.head;
632 cluster->next = cluster_next(&cluster->index) *
634 } else if (!cluster_list_empty(&si->discard_clusters)) {
636 * we don't have free cluster but have some clusters in
637 * discarding, do discard now and reclaim them, then
638 * reread cluster_next_cpu since we dropped si->lock
640 swap_do_scheduled_discard(si);
641 *scan_base = this_cpu_read(*si->cluster_next_cpu);
642 *offset = *scan_base;
649 * Other CPUs can use our cluster if they can't find a free cluster,
650 * check if there is still free entry in the cluster
653 max = min_t(unsigned long, si->max,
654 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
656 ci = lock_cluster(si, tmp);
658 if (!si->swap_map[tmp])
665 cluster_set_null(&cluster->index);
668 cluster->next = tmp + 1;
674 static void __del_from_avail_list(struct swap_info_struct *p)
679 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
682 static void del_from_avail_list(struct swap_info_struct *p)
684 spin_lock(&swap_avail_lock);
685 __del_from_avail_list(p);
686 spin_unlock(&swap_avail_lock);
689 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
690 unsigned int nr_entries)
692 unsigned int end = offset + nr_entries - 1;
694 if (offset == si->lowest_bit)
695 si->lowest_bit += nr_entries;
696 if (end == si->highest_bit)
697 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
698 si->inuse_pages += nr_entries;
699 if (si->inuse_pages == si->pages) {
700 si->lowest_bit = si->max;
702 del_from_avail_list(si);
706 static void add_to_avail_list(struct swap_info_struct *p)
710 spin_lock(&swap_avail_lock);
712 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
713 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
715 spin_unlock(&swap_avail_lock);
718 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
719 unsigned int nr_entries)
721 unsigned long begin = offset;
722 unsigned long end = offset + nr_entries - 1;
723 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
725 if (offset < si->lowest_bit)
726 si->lowest_bit = offset;
727 if (end > si->highest_bit) {
728 bool was_full = !si->highest_bit;
730 WRITE_ONCE(si->highest_bit, end);
731 if (was_full && (si->flags & SWP_WRITEOK))
732 add_to_avail_list(si);
734 atomic_long_add(nr_entries, &nr_swap_pages);
735 si->inuse_pages -= nr_entries;
736 if (si->flags & SWP_BLKDEV)
737 swap_slot_free_notify =
738 si->bdev->bd_disk->fops->swap_slot_free_notify;
740 swap_slot_free_notify = NULL;
741 while (offset <= end) {
742 arch_swap_invalidate_page(si->type, offset);
743 frontswap_invalidate_page(si->type, offset);
744 if (swap_slot_free_notify)
745 swap_slot_free_notify(si->bdev, offset);
748 clear_shadow_from_swap_cache(si->type, begin, end);
751 static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
755 if (!(si->flags & SWP_SOLIDSTATE)) {
756 si->cluster_next = next;
760 prev = this_cpu_read(*si->cluster_next_cpu);
762 * Cross the swap address space size aligned trunk, choose
763 * another trunk randomly to avoid lock contention on swap
764 * address space if possible.
766 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
767 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
768 /* No free swap slots available */
769 if (si->highest_bit <= si->lowest_bit)
771 next = si->lowest_bit +
772 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
773 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
774 next = max_t(unsigned int, next, si->lowest_bit);
776 this_cpu_write(*si->cluster_next_cpu, next);
779 static bool swap_offset_available_and_locked(struct swap_info_struct *si,
780 unsigned long offset)
782 if (data_race(!si->swap_map[offset])) {
783 spin_lock(&si->lock);
787 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
788 spin_lock(&si->lock);
795 static int scan_swap_map_slots(struct swap_info_struct *si,
796 unsigned char usage, int nr,
799 struct swap_cluster_info *ci;
800 unsigned long offset;
801 unsigned long scan_base;
802 unsigned long last_in_cluster = 0;
803 int latency_ration = LATENCY_LIMIT;
805 bool scanned_many = false;
808 * We try to cluster swap pages by allocating them sequentially
809 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
810 * way, however, we resort to first-free allocation, starting
811 * a new cluster. This prevents us from scattering swap pages
812 * all over the entire swap partition, so that we reduce
813 * overall disk seek times between swap pages. -- sct
814 * But we do now try to find an empty cluster. -Andrea
815 * And we let swap pages go all over an SSD partition. Hugh
818 si->flags += SWP_SCANNING;
820 * Use percpu scan base for SSD to reduce lock contention on
821 * cluster and swap cache. For HDD, sequential access is more
824 if (si->flags & SWP_SOLIDSTATE)
825 scan_base = this_cpu_read(*si->cluster_next_cpu);
827 scan_base = si->cluster_next;
831 if (si->cluster_info) {
832 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
834 } else if (unlikely(!si->cluster_nr--)) {
835 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
836 si->cluster_nr = SWAPFILE_CLUSTER - 1;
840 spin_unlock(&si->lock);
843 * If seek is expensive, start searching for new cluster from
844 * start of partition, to minimize the span of allocated swap.
845 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
846 * case, just handled by scan_swap_map_try_ssd_cluster() above.
848 scan_base = offset = si->lowest_bit;
849 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
851 /* Locate the first empty (unaligned) cluster */
852 for (; last_in_cluster <= si->highest_bit; offset++) {
853 if (si->swap_map[offset])
854 last_in_cluster = offset + SWAPFILE_CLUSTER;
855 else if (offset == last_in_cluster) {
856 spin_lock(&si->lock);
857 offset -= SWAPFILE_CLUSTER - 1;
858 si->cluster_next = offset;
859 si->cluster_nr = SWAPFILE_CLUSTER - 1;
862 if (unlikely(--latency_ration < 0)) {
864 latency_ration = LATENCY_LIMIT;
869 spin_lock(&si->lock);
870 si->cluster_nr = SWAPFILE_CLUSTER - 1;
874 if (si->cluster_info) {
875 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
876 /* take a break if we already got some slots */
879 if (!scan_swap_map_try_ssd_cluster(si, &offset,
884 if (!(si->flags & SWP_WRITEOK))
886 if (!si->highest_bit)
888 if (offset > si->highest_bit)
889 scan_base = offset = si->lowest_bit;
891 ci = lock_cluster(si, offset);
892 /* reuse swap entry of cache-only swap if not busy. */
893 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
896 spin_unlock(&si->lock);
897 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
898 spin_lock(&si->lock);
899 /* entry was freed successfully, try to use this again */
902 goto scan; /* check next one */
905 if (si->swap_map[offset]) {
912 WRITE_ONCE(si->swap_map[offset], usage);
913 inc_cluster_info_page(si, si->cluster_info, offset);
916 swap_range_alloc(si, offset, 1);
917 slots[n_ret++] = swp_entry(si->type, offset);
919 /* got enough slots or reach max slots? */
920 if ((n_ret == nr) || (offset >= si->highest_bit))
923 /* search for next available slot */
925 /* time to take a break? */
926 if (unlikely(--latency_ration < 0)) {
929 spin_unlock(&si->lock);
931 spin_lock(&si->lock);
932 latency_ration = LATENCY_LIMIT;
935 /* try to get more slots in cluster */
936 if (si->cluster_info) {
937 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
939 } else if (si->cluster_nr && !si->swap_map[++offset]) {
940 /* non-ssd case, still more slots in cluster? */
946 * Even if there's no free clusters available (fragmented),
947 * try to scan a little more quickly with lock held unless we
948 * have scanned too many slots already.
951 unsigned long scan_limit;
953 if (offset < scan_base)
954 scan_limit = scan_base;
956 scan_limit = si->highest_bit;
957 for (; offset <= scan_limit && --latency_ration > 0;
959 if (!si->swap_map[offset])
965 set_cluster_next(si, offset + 1);
966 si->flags -= SWP_SCANNING;
970 spin_unlock(&si->lock);
971 while (++offset <= READ_ONCE(si->highest_bit)) {
972 if (swap_offset_available_and_locked(si, offset))
974 if (unlikely(--latency_ration < 0)) {
976 latency_ration = LATENCY_LIMIT;
980 offset = si->lowest_bit;
981 while (offset < scan_base) {
982 if (swap_offset_available_and_locked(si, offset))
984 if (unlikely(--latency_ration < 0)) {
986 latency_ration = LATENCY_LIMIT;
991 spin_lock(&si->lock);
994 si->flags -= SWP_SCANNING;
998 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
1001 struct swap_cluster_info *ci;
1002 unsigned long offset;
1005 * Should not even be attempting cluster allocations when huge
1006 * page swap is disabled. Warn and fail the allocation.
1008 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1013 if (cluster_list_empty(&si->free_clusters))
1016 idx = cluster_list_first(&si->free_clusters);
1017 offset = idx * SWAPFILE_CLUSTER;
1018 ci = lock_cluster(si, offset);
1019 alloc_cluster(si, idx);
1020 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1022 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1024 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1025 *slot = swp_entry(si->type, offset);
1030 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1032 unsigned long offset = idx * SWAPFILE_CLUSTER;
1033 struct swap_cluster_info *ci;
1035 ci = lock_cluster(si, offset);
1036 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1037 cluster_set_count_flag(ci, 0, 0);
1038 free_cluster(si, idx);
1040 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1043 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1045 unsigned long size = swap_entry_size(entry_size);
1046 struct swap_info_struct *si, *next;
1051 /* Only single cluster request supported */
1052 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1054 spin_lock(&swap_avail_lock);
1056 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1057 if (avail_pgs <= 0) {
1058 spin_unlock(&swap_avail_lock);
1062 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1064 atomic_long_sub(n_goal * size, &nr_swap_pages);
1067 node = numa_node_id();
1068 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1069 /* requeue si to after same-priority siblings */
1070 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1071 spin_unlock(&swap_avail_lock);
1072 spin_lock(&si->lock);
1073 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1074 spin_lock(&swap_avail_lock);
1075 if (plist_node_empty(&si->avail_lists[node])) {
1076 spin_unlock(&si->lock);
1079 WARN(!si->highest_bit,
1080 "swap_info %d in list but !highest_bit\n",
1082 WARN(!(si->flags & SWP_WRITEOK),
1083 "swap_info %d in list but !SWP_WRITEOK\n",
1085 __del_from_avail_list(si);
1086 spin_unlock(&si->lock);
1089 if (size == SWAPFILE_CLUSTER) {
1090 if (si->flags & SWP_BLKDEV)
1091 n_ret = swap_alloc_cluster(si, swp_entries);
1093 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1094 n_goal, swp_entries);
1095 spin_unlock(&si->lock);
1096 if (n_ret || size == SWAPFILE_CLUSTER)
1098 pr_debug("scan_swap_map of si %d failed to find offset\n",
1101 spin_lock(&swap_avail_lock);
1104 * if we got here, it's likely that si was almost full before,
1105 * and since scan_swap_map_slots() can drop the si->lock,
1106 * multiple callers probably all tried to get a page from the
1107 * same si and it filled up before we could get one; or, the si
1108 * filled up between us dropping swap_avail_lock and taking
1109 * si->lock. Since we dropped the swap_avail_lock, the
1110 * swap_avail_head list may have been modified; so if next is
1111 * still in the swap_avail_head list then try it, otherwise
1112 * start over if we have not gotten any slots.
1114 if (plist_node_empty(&next->avail_lists[node]))
1118 spin_unlock(&swap_avail_lock);
1122 atomic_long_add((long)(n_goal - n_ret) * size,
1128 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1130 struct swap_info_struct *p;
1131 unsigned long offset;
1135 p = swp_swap_info(entry);
1138 if (data_race(!(p->flags & SWP_USED)))
1140 offset = swp_offset(entry);
1141 if (offset >= p->max)
1143 if (data_race(!p->swap_map[swp_offset(entry)]))
1148 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1151 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1154 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1157 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1162 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1163 struct swap_info_struct *q)
1165 struct swap_info_struct *p;
1167 p = _swap_info_get(entry);
1171 spin_unlock(&q->lock);
1173 spin_lock(&p->lock);
1178 static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1179 unsigned long offset,
1180 unsigned char usage)
1182 unsigned char count;
1183 unsigned char has_cache;
1185 count = p->swap_map[offset];
1187 has_cache = count & SWAP_HAS_CACHE;
1188 count &= ~SWAP_HAS_CACHE;
1190 if (usage == SWAP_HAS_CACHE) {
1191 VM_BUG_ON(!has_cache);
1193 } else if (count == SWAP_MAP_SHMEM) {
1195 * Or we could insist on shmem.c using a special
1196 * swap_shmem_free() and free_shmem_swap_and_cache()...
1199 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1200 if (count == COUNT_CONTINUED) {
1201 if (swap_count_continued(p, offset, count))
1202 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1204 count = SWAP_MAP_MAX;
1209 usage = count | has_cache;
1211 WRITE_ONCE(p->swap_map[offset], usage);
1213 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1219 * Check whether swap entry is valid in the swap device. If so,
1220 * return pointer to swap_info_struct, and keep the swap entry valid
1221 * via preventing the swap device from being swapoff, until
1222 * put_swap_device() is called. Otherwise return NULL.
1224 * Notice that swapoff or swapoff+swapon can still happen before the
1225 * percpu_ref_tryget_live() in get_swap_device() or after the
1226 * percpu_ref_put() in put_swap_device() if there isn't any other way
1227 * to prevent swapoff, such as page lock, page table lock, etc. The
1228 * caller must be prepared for that. For example, the following
1229 * situation is possible.
1233 * ... swapoff+swapon
1234 * __read_swap_cache_async()
1235 * swapcache_prepare()
1236 * __swap_duplicate()
1238 * // verify PTE not changed
1240 * In __swap_duplicate(), the swap_map need to be checked before
1241 * changing partly because the specified swap entry may be for another
1242 * swap device which has been swapoff. And in do_swap_page(), after
1243 * the page is read from the swap device, the PTE is verified not
1244 * changed with the page table locked to check whether the swap device
1245 * has been swapoff or swapoff+swapon.
1247 struct swap_info_struct *get_swap_device(swp_entry_t entry)
1249 struct swap_info_struct *si;
1250 unsigned long offset;
1254 si = swp_swap_info(entry);
1257 if (!percpu_ref_tryget_live(&si->users))
1260 * Guarantee the si->users are checked before accessing other
1261 * fields of swap_info_struct.
1263 * Paired with the spin_unlock() after setup_swap_info() in
1264 * enable_swap_info().
1267 offset = swp_offset(entry);
1268 if (offset >= si->max)
1273 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1277 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1278 percpu_ref_put(&si->users);
1282 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1285 struct swap_cluster_info *ci;
1286 unsigned long offset = swp_offset(entry);
1287 unsigned char usage;
1289 ci = lock_cluster_or_swap_info(p, offset);
1290 usage = __swap_entry_free_locked(p, offset, 1);
1291 unlock_cluster_or_swap_info(p, ci);
1293 free_swap_slot(entry);
1298 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1300 struct swap_cluster_info *ci;
1301 unsigned long offset = swp_offset(entry);
1302 unsigned char count;
1304 ci = lock_cluster(p, offset);
1305 count = p->swap_map[offset];
1306 VM_BUG_ON(count != SWAP_HAS_CACHE);
1307 p->swap_map[offset] = 0;
1308 dec_cluster_info_page(p, p->cluster_info, offset);
1311 mem_cgroup_uncharge_swap(entry, 1);
1312 swap_range_free(p, offset, 1);
1316 * Caller has made sure that the swap device corresponding to entry
1317 * is still around or has not been recycled.
1319 void swap_free(swp_entry_t entry)
1321 struct swap_info_struct *p;
1323 p = _swap_info_get(entry);
1325 __swap_entry_free(p, entry);
1329 * Called after dropping swapcache to decrease refcnt to swap entries.
1331 void put_swap_page(struct page *page, swp_entry_t entry)
1333 unsigned long offset = swp_offset(entry);
1334 unsigned long idx = offset / SWAPFILE_CLUSTER;
1335 struct swap_cluster_info *ci;
1336 struct swap_info_struct *si;
1338 unsigned int i, free_entries = 0;
1340 int size = swap_entry_size(thp_nr_pages(page));
1342 si = _swap_info_get(entry);
1346 ci = lock_cluster_or_swap_info(si, offset);
1347 if (size == SWAPFILE_CLUSTER) {
1348 VM_BUG_ON(!cluster_is_huge(ci));
1349 map = si->swap_map + offset;
1350 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1352 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1353 if (val == SWAP_HAS_CACHE)
1356 cluster_clear_huge(ci);
1357 if (free_entries == SWAPFILE_CLUSTER) {
1358 unlock_cluster_or_swap_info(si, ci);
1359 spin_lock(&si->lock);
1360 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1361 swap_free_cluster(si, idx);
1362 spin_unlock(&si->lock);
1366 for (i = 0; i < size; i++, entry.val++) {
1367 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1368 unlock_cluster_or_swap_info(si, ci);
1369 free_swap_slot(entry);
1372 lock_cluster_or_swap_info(si, offset);
1375 unlock_cluster_or_swap_info(si, ci);
1378 #ifdef CONFIG_THP_SWAP
1379 int split_swap_cluster(swp_entry_t entry)
1381 struct swap_info_struct *si;
1382 struct swap_cluster_info *ci;
1383 unsigned long offset = swp_offset(entry);
1385 si = _swap_info_get(entry);
1388 ci = lock_cluster(si, offset);
1389 cluster_clear_huge(ci);
1395 static int swp_entry_cmp(const void *ent1, const void *ent2)
1397 const swp_entry_t *e1 = ent1, *e2 = ent2;
1399 return (int)swp_type(*e1) - (int)swp_type(*e2);
1402 void swapcache_free_entries(swp_entry_t *entries, int n)
1404 struct swap_info_struct *p, *prev;
1414 * Sort swap entries by swap device, so each lock is only taken once.
1415 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1416 * so low that it isn't necessary to optimize further.
1418 if (nr_swapfiles > 1)
1419 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1420 for (i = 0; i < n; ++i) {
1421 p = swap_info_get_cont(entries[i], prev);
1423 swap_entry_free(p, entries[i]);
1427 spin_unlock(&p->lock);
1431 * How many references to page are currently swapped out?
1432 * This does not give an exact answer when swap count is continued,
1433 * but does include the high COUNT_CONTINUED flag to allow for that.
1435 static int page_swapcount(struct page *page)
1438 struct swap_info_struct *p;
1439 struct swap_cluster_info *ci;
1441 unsigned long offset;
1443 entry.val = page_private(page);
1444 p = _swap_info_get(entry);
1446 offset = swp_offset(entry);
1447 ci = lock_cluster_or_swap_info(p, offset);
1448 count = swap_count(p->swap_map[offset]);
1449 unlock_cluster_or_swap_info(p, ci);
1454 int __swap_count(swp_entry_t entry)
1456 struct swap_info_struct *si;
1457 pgoff_t offset = swp_offset(entry);
1460 si = get_swap_device(entry);
1462 count = swap_count(si->swap_map[offset]);
1463 put_swap_device(si);
1468 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1471 pgoff_t offset = swp_offset(entry);
1472 struct swap_cluster_info *ci;
1474 ci = lock_cluster_or_swap_info(si, offset);
1475 count = swap_count(si->swap_map[offset]);
1476 unlock_cluster_or_swap_info(si, ci);
1481 * How many references to @entry are currently swapped out?
1482 * This does not give an exact answer when swap count is continued,
1483 * but does include the high COUNT_CONTINUED flag to allow for that.
1485 int __swp_swapcount(swp_entry_t entry)
1488 struct swap_info_struct *si;
1490 si = get_swap_device(entry);
1492 count = swap_swapcount(si, entry);
1493 put_swap_device(si);
1499 * How many references to @entry are currently swapped out?
1500 * This considers COUNT_CONTINUED so it returns exact answer.
1502 int swp_swapcount(swp_entry_t entry)
1504 int count, tmp_count, n;
1505 struct swap_info_struct *p;
1506 struct swap_cluster_info *ci;
1511 p = _swap_info_get(entry);
1515 offset = swp_offset(entry);
1517 ci = lock_cluster_or_swap_info(p, offset);
1519 count = swap_count(p->swap_map[offset]);
1520 if (!(count & COUNT_CONTINUED))
1523 count &= ~COUNT_CONTINUED;
1524 n = SWAP_MAP_MAX + 1;
1526 page = vmalloc_to_page(p->swap_map + offset);
1527 offset &= ~PAGE_MASK;
1528 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1531 page = list_next_entry(page, lru);
1532 map = kmap_atomic(page);
1533 tmp_count = map[offset];
1536 count += (tmp_count & ~COUNT_CONTINUED) * n;
1537 n *= (SWAP_CONT_MAX + 1);
1538 } while (tmp_count & COUNT_CONTINUED);
1540 unlock_cluster_or_swap_info(p, ci);
1544 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1547 struct swap_cluster_info *ci;
1548 unsigned char *map = si->swap_map;
1549 unsigned long roffset = swp_offset(entry);
1550 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1554 ci = lock_cluster_or_swap_info(si, offset);
1555 if (!ci || !cluster_is_huge(ci)) {
1556 if (swap_count(map[roffset]))
1560 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1561 if (swap_count(map[offset + i])) {
1567 unlock_cluster_or_swap_info(si, ci);
1571 static bool page_swapped(struct page *page)
1574 struct swap_info_struct *si;
1576 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1577 return page_swapcount(page) != 0;
1579 page = compound_head(page);
1580 entry.val = page_private(page);
1581 si = _swap_info_get(entry);
1583 return swap_page_trans_huge_swapped(si, entry);
1588 * If swap is getting full, or if there are no more mappings of this page,
1589 * then try_to_free_swap is called to free its swap space.
1591 int try_to_free_swap(struct page *page)
1593 VM_BUG_ON_PAGE(!PageLocked(page), page);
1595 if (!PageSwapCache(page))
1597 if (PageWriteback(page))
1599 if (page_swapped(page))
1603 * Once hibernation has begun to create its image of memory,
1604 * there's a danger that one of the calls to try_to_free_swap()
1605 * - most probably a call from __try_to_reclaim_swap() while
1606 * hibernation is allocating its own swap pages for the image,
1607 * but conceivably even a call from memory reclaim - will free
1608 * the swap from a page which has already been recorded in the
1609 * image as a clean swapcache page, and then reuse its swap for
1610 * another page of the image. On waking from hibernation, the
1611 * original page might be freed under memory pressure, then
1612 * later read back in from swap, now with the wrong data.
1614 * Hibernation suspends storage while it is writing the image
1615 * to disk so check that here.
1617 if (pm_suspended_storage())
1620 page = compound_head(page);
1621 delete_from_swap_cache(page);
1627 * Free the swap entry like above, but also try to
1628 * free the page cache entry if it is the last user.
1630 int free_swap_and_cache(swp_entry_t entry)
1632 struct swap_info_struct *p;
1633 unsigned char count;
1635 if (non_swap_entry(entry))
1638 p = _swap_info_get(entry);
1640 count = __swap_entry_free(p, entry);
1641 if (count == SWAP_HAS_CACHE &&
1642 !swap_page_trans_huge_swapped(p, entry))
1643 __try_to_reclaim_swap(p, swp_offset(entry),
1644 TTRS_UNMAPPED | TTRS_FULL);
1649 #ifdef CONFIG_HIBERNATION
1651 swp_entry_t get_swap_page_of_type(int type)
1653 struct swap_info_struct *si = swap_type_to_swap_info(type);
1654 swp_entry_t entry = {0};
1659 /* This is called for allocating swap entry, not cache */
1660 spin_lock(&si->lock);
1661 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1662 atomic_long_dec(&nr_swap_pages);
1663 spin_unlock(&si->lock);
1669 * Find the swap type that corresponds to given device (if any).
1671 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1672 * from 0, in which the swap header is expected to be located.
1674 * This is needed for the suspend to disk (aka swsusp).
1676 int swap_type_of(dev_t device, sector_t offset)
1683 spin_lock(&swap_lock);
1684 for (type = 0; type < nr_swapfiles; type++) {
1685 struct swap_info_struct *sis = swap_info[type];
1687 if (!(sis->flags & SWP_WRITEOK))
1690 if (device == sis->bdev->bd_dev) {
1691 struct swap_extent *se = first_se(sis);
1693 if (se->start_block == offset) {
1694 spin_unlock(&swap_lock);
1699 spin_unlock(&swap_lock);
1703 int find_first_swap(dev_t *device)
1707 spin_lock(&swap_lock);
1708 for (type = 0; type < nr_swapfiles; type++) {
1709 struct swap_info_struct *sis = swap_info[type];
1711 if (!(sis->flags & SWP_WRITEOK))
1713 *device = sis->bdev->bd_dev;
1714 spin_unlock(&swap_lock);
1717 spin_unlock(&swap_lock);
1722 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1723 * corresponding to given index in swap_info (swap type).
1725 sector_t swapdev_block(int type, pgoff_t offset)
1727 struct swap_info_struct *si = swap_type_to_swap_info(type);
1728 struct swap_extent *se;
1730 if (!si || !(si->flags & SWP_WRITEOK))
1732 se = offset_to_swap_extent(si, offset);
1733 return se->start_block + (offset - se->start_page);
1737 * Return either the total number of swap pages of given type, or the number
1738 * of free pages of that type (depending on @free)
1740 * This is needed for software suspend
1742 unsigned int count_swap_pages(int type, int free)
1746 spin_lock(&swap_lock);
1747 if ((unsigned int)type < nr_swapfiles) {
1748 struct swap_info_struct *sis = swap_info[type];
1750 spin_lock(&sis->lock);
1751 if (sis->flags & SWP_WRITEOK) {
1754 n -= sis->inuse_pages;
1756 spin_unlock(&sis->lock);
1758 spin_unlock(&swap_lock);
1761 #endif /* CONFIG_HIBERNATION */
1763 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1765 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1769 * No need to decide whether this PTE shares the swap entry with others,
1770 * just let do_wp_page work it out if a write is requested later - to
1771 * force COW, vm_page_prot omits write permission from any private vma.
1773 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1774 unsigned long addr, swp_entry_t entry, struct page *page)
1776 struct page *swapcache;
1778 pte_t *pte, new_pte;
1782 page = ksm_might_need_to_copy(page, vma, addr);
1783 if (unlikely(!page))
1786 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1787 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1792 if (unlikely(!PageUptodate(page))) {
1795 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1796 pteval = swp_entry_to_pte(make_swapin_error_entry(page));
1797 set_pte_at(vma->vm_mm, addr, pte, pteval);
1803 /* See do_swap_page() */
1804 BUG_ON(!PageAnon(page) && PageMappedToDisk(page));
1805 BUG_ON(PageAnon(page) && PageAnonExclusive(page));
1807 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1808 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1810 if (page == swapcache) {
1811 rmap_t rmap_flags = RMAP_NONE;
1814 * See do_swap_page(): PageWriteback() would be problematic.
1815 * However, we do a wait_on_page_writeback() just before this
1816 * call and have the page locked.
1818 VM_BUG_ON_PAGE(PageWriteback(page), page);
1819 if (pte_swp_exclusive(*pte))
1820 rmap_flags |= RMAP_EXCLUSIVE;
1822 page_add_anon_rmap(page, vma, addr, rmap_flags);
1823 } else { /* ksm created a completely new copy */
1824 page_add_new_anon_rmap(page, vma, addr);
1825 lru_cache_add_inactive_or_unevictable(page, vma);
1827 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
1828 if (pte_swp_soft_dirty(*pte))
1829 new_pte = pte_mksoft_dirty(new_pte);
1830 if (pte_swp_uffd_wp(*pte))
1831 new_pte = pte_mkuffd_wp(new_pte);
1832 set_pte_at(vma->vm_mm, addr, pte, new_pte);
1835 pte_unmap_unlock(pte, ptl);
1836 if (page != swapcache) {
1843 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1844 unsigned long addr, unsigned long end,
1850 struct swap_info_struct *si;
1851 unsigned long offset;
1853 volatile unsigned char *swap_map;
1855 si = swap_info[type];
1856 pte = pte_offset_map(pmd, addr);
1858 if (!is_swap_pte(*pte))
1861 entry = pte_to_swp_entry(*pte);
1862 if (swp_type(entry) != type)
1865 offset = swp_offset(entry);
1867 swap_map = &si->swap_map[offset];
1868 page = lookup_swap_cache(entry, vma, addr);
1870 struct vm_fault vmf = {
1873 .real_address = addr,
1877 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1881 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1887 wait_on_page_writeback(page);
1888 ret = unuse_pte(vma, pmd, addr, entry, page);
1895 try_to_free_swap(page);
1899 pte = pte_offset_map(pmd, addr);
1900 } while (pte++, addr += PAGE_SIZE, addr != end);
1908 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1909 unsigned long addr, unsigned long end,
1916 pmd = pmd_offset(pud, addr);
1919 next = pmd_addr_end(addr, end);
1920 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1922 ret = unuse_pte_range(vma, pmd, addr, next, type);
1925 } while (pmd++, addr = next, addr != end);
1929 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1930 unsigned long addr, unsigned long end,
1937 pud = pud_offset(p4d, addr);
1939 next = pud_addr_end(addr, end);
1940 if (pud_none_or_clear_bad(pud))
1942 ret = unuse_pmd_range(vma, pud, addr, next, type);
1945 } while (pud++, addr = next, addr != end);
1949 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1950 unsigned long addr, unsigned long end,
1957 p4d = p4d_offset(pgd, addr);
1959 next = p4d_addr_end(addr, end);
1960 if (p4d_none_or_clear_bad(p4d))
1962 ret = unuse_pud_range(vma, p4d, addr, next, type);
1965 } while (p4d++, addr = next, addr != end);
1969 static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
1972 unsigned long addr, end, next;
1975 addr = vma->vm_start;
1978 pgd = pgd_offset(vma->vm_mm, addr);
1980 next = pgd_addr_end(addr, end);
1981 if (pgd_none_or_clear_bad(pgd))
1983 ret = unuse_p4d_range(vma, pgd, addr, next, type);
1986 } while (pgd++, addr = next, addr != end);
1990 static int unuse_mm(struct mm_struct *mm, unsigned int type)
1992 struct vm_area_struct *vma;
1996 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1997 if (vma->anon_vma) {
1998 ret = unuse_vma(vma, type);
2004 mmap_read_unlock(mm);
2009 * Scan swap_map from current position to next entry still in use.
2010 * Return 0 if there are no inuse entries after prev till end of
2013 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2017 unsigned char count;
2020 * No need for swap_lock here: we're just looking
2021 * for whether an entry is in use, not modifying it; false
2022 * hits are okay, and sys_swapoff() has already prevented new
2023 * allocations from this area (while holding swap_lock).
2025 for (i = prev + 1; i < si->max; i++) {
2026 count = READ_ONCE(si->swap_map[i]);
2027 if (count && swap_count(count) != SWAP_MAP_BAD)
2029 if ((i % LATENCY_LIMIT) == 0)
2039 static int try_to_unuse(unsigned int type)
2041 struct mm_struct *prev_mm;
2042 struct mm_struct *mm;
2043 struct list_head *p;
2045 struct swap_info_struct *si = swap_info[type];
2050 if (!READ_ONCE(si->inuse_pages))
2054 retval = shmem_unuse(type);
2061 spin_lock(&mmlist_lock);
2062 p = &init_mm.mmlist;
2063 while (READ_ONCE(si->inuse_pages) &&
2064 !signal_pending(current) &&
2065 (p = p->next) != &init_mm.mmlist) {
2067 mm = list_entry(p, struct mm_struct, mmlist);
2068 if (!mmget_not_zero(mm))
2070 spin_unlock(&mmlist_lock);
2073 retval = unuse_mm(mm, type);
2080 * Make sure that we aren't completely killing
2081 * interactive performance.
2084 spin_lock(&mmlist_lock);
2086 spin_unlock(&mmlist_lock);
2091 while (READ_ONCE(si->inuse_pages) &&
2092 !signal_pending(current) &&
2093 (i = find_next_to_unuse(si, i)) != 0) {
2095 entry = swp_entry(type, i);
2096 page = find_get_page(swap_address_space(entry), i);
2101 * It is conceivable that a racing task removed this page from
2102 * swap cache just before we acquired the page lock. The page
2103 * might even be back in swap cache on another swap area. But
2104 * that is okay, try_to_free_swap() only removes stale pages.
2107 wait_on_page_writeback(page);
2108 try_to_free_swap(page);
2114 * Lets check again to see if there are still swap entries in the map.
2115 * If yes, we would need to do retry the unuse logic again.
2116 * Under global memory pressure, swap entries can be reinserted back
2117 * into process space after the mmlist loop above passes over them.
2119 * Limit the number of retries? No: when mmget_not_zero()
2120 * above fails, that mm is likely to be freeing swap from
2121 * exit_mmap(), which proceeds at its own independent pace;
2122 * and even shmem_writepage() could have been preempted after
2123 * folio_alloc_swap(), temporarily hiding that swap. It's easy
2124 * and robust (though cpu-intensive) just to keep retrying.
2126 if (READ_ONCE(si->inuse_pages)) {
2127 if (!signal_pending(current))
2136 * After a successful try_to_unuse, if no swap is now in use, we know
2137 * we can empty the mmlist. swap_lock must be held on entry and exit.
2138 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2139 * added to the mmlist just after page_duplicate - before would be racy.
2141 static void drain_mmlist(void)
2143 struct list_head *p, *next;
2146 for (type = 0; type < nr_swapfiles; type++)
2147 if (swap_info[type]->inuse_pages)
2149 spin_lock(&mmlist_lock);
2150 list_for_each_safe(p, next, &init_mm.mmlist)
2152 spin_unlock(&mmlist_lock);
2156 * Free all of a swapdev's extent information
2158 static void destroy_swap_extents(struct swap_info_struct *sis)
2160 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2161 struct rb_node *rb = sis->swap_extent_root.rb_node;
2162 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2164 rb_erase(rb, &sis->swap_extent_root);
2168 if (sis->flags & SWP_ACTIVATED) {
2169 struct file *swap_file = sis->swap_file;
2170 struct address_space *mapping = swap_file->f_mapping;
2172 sis->flags &= ~SWP_ACTIVATED;
2173 if (mapping->a_ops->swap_deactivate)
2174 mapping->a_ops->swap_deactivate(swap_file);
2179 * Add a block range (and the corresponding page range) into this swapdev's
2182 * This function rather assumes that it is called in ascending page order.
2185 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2186 unsigned long nr_pages, sector_t start_block)
2188 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2189 struct swap_extent *se;
2190 struct swap_extent *new_se;
2193 * place the new node at the right most since the
2194 * function is called in ascending page order.
2198 link = &parent->rb_right;
2202 se = rb_entry(parent, struct swap_extent, rb_node);
2203 BUG_ON(se->start_page + se->nr_pages != start_page);
2204 if (se->start_block + se->nr_pages == start_block) {
2206 se->nr_pages += nr_pages;
2211 /* No merge, insert a new extent. */
2212 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2215 new_se->start_page = start_page;
2216 new_se->nr_pages = nr_pages;
2217 new_se->start_block = start_block;
2219 rb_link_node(&new_se->rb_node, parent, link);
2220 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2223 EXPORT_SYMBOL_GPL(add_swap_extent);
2226 * A `swap extent' is a simple thing which maps a contiguous range of pages
2227 * onto a contiguous range of disk blocks. A rbtree of swap extents is
2228 * built at swapon time and is then used at swap_writepage/swap_readpage
2229 * time for locating where on disk a page belongs.
2231 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2232 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2233 * swap files identically.
2235 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2236 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2237 * swapfiles are handled *identically* after swapon time.
2239 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2240 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
2241 * blocks are found which do not fall within the PAGE_SIZE alignment
2242 * requirements, they are simply tossed out - we will never use those blocks
2245 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2246 * prevents users from writing to the swap device, which will corrupt memory.
2248 * The amount of disk space which a single swap extent represents varies.
2249 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2250 * extents in the rbtree. - akpm.
2252 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2254 struct file *swap_file = sis->swap_file;
2255 struct address_space *mapping = swap_file->f_mapping;
2256 struct inode *inode = mapping->host;
2259 if (S_ISBLK(inode->i_mode)) {
2260 ret = add_swap_extent(sis, 0, sis->max, 0);
2265 if (mapping->a_ops->swap_activate) {
2266 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2269 sis->flags |= SWP_ACTIVATED;
2270 if ((sis->flags & SWP_FS_OPS) &&
2271 sio_pool_init() != 0) {
2272 destroy_swap_extents(sis);
2278 return generic_swapfile_activate(sis, swap_file, span);
2281 static int swap_node(struct swap_info_struct *p)
2283 struct block_device *bdev;
2288 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2290 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2293 static void setup_swap_info(struct swap_info_struct *p, int prio,
2294 unsigned char *swap_map,
2295 struct swap_cluster_info *cluster_info)
2302 p->prio = --least_priority;
2304 * the plist prio is negated because plist ordering is
2305 * low-to-high, while swap ordering is high-to-low
2307 p->list.prio = -p->prio;
2310 p->avail_lists[i].prio = -p->prio;
2312 if (swap_node(p) == i)
2313 p->avail_lists[i].prio = 1;
2315 p->avail_lists[i].prio = -p->prio;
2318 p->swap_map = swap_map;
2319 p->cluster_info = cluster_info;
2322 static void _enable_swap_info(struct swap_info_struct *p)
2324 p->flags |= SWP_WRITEOK;
2325 atomic_long_add(p->pages, &nr_swap_pages);
2326 total_swap_pages += p->pages;
2328 assert_spin_locked(&swap_lock);
2330 * both lists are plists, and thus priority ordered.
2331 * swap_active_head needs to be priority ordered for swapoff(),
2332 * which on removal of any swap_info_struct with an auto-assigned
2333 * (i.e. negative) priority increments the auto-assigned priority
2334 * of any lower-priority swap_info_structs.
2335 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2336 * which allocates swap pages from the highest available priority
2339 plist_add(&p->list, &swap_active_head);
2340 add_to_avail_list(p);
2343 static void enable_swap_info(struct swap_info_struct *p, int prio,
2344 unsigned char *swap_map,
2345 struct swap_cluster_info *cluster_info,
2346 unsigned long *frontswap_map)
2348 if (IS_ENABLED(CONFIG_FRONTSWAP))
2349 frontswap_init(p->type, frontswap_map);
2350 spin_lock(&swap_lock);
2351 spin_lock(&p->lock);
2352 setup_swap_info(p, prio, swap_map, cluster_info);
2353 spin_unlock(&p->lock);
2354 spin_unlock(&swap_lock);
2356 * Finished initializing swap device, now it's safe to reference it.
2358 percpu_ref_resurrect(&p->users);
2359 spin_lock(&swap_lock);
2360 spin_lock(&p->lock);
2361 _enable_swap_info(p);
2362 spin_unlock(&p->lock);
2363 spin_unlock(&swap_lock);
2366 static void reinsert_swap_info(struct swap_info_struct *p)
2368 spin_lock(&swap_lock);
2369 spin_lock(&p->lock);
2370 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2371 _enable_swap_info(p);
2372 spin_unlock(&p->lock);
2373 spin_unlock(&swap_lock);
2376 bool has_usable_swap(void)
2380 spin_lock(&swap_lock);
2381 if (plist_head_empty(&swap_active_head))
2383 spin_unlock(&swap_lock);
2387 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2389 struct swap_info_struct *p = NULL;
2390 unsigned char *swap_map;
2391 struct swap_cluster_info *cluster_info;
2392 unsigned long *frontswap_map;
2393 struct file *swap_file, *victim;
2394 struct address_space *mapping;
2395 struct inode *inode;
2396 struct filename *pathname;
2398 unsigned int old_block_size;
2400 if (!capable(CAP_SYS_ADMIN))
2403 BUG_ON(!current->mm);
2405 pathname = getname(specialfile);
2406 if (IS_ERR(pathname))
2407 return PTR_ERR(pathname);
2409 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2410 err = PTR_ERR(victim);
2414 mapping = victim->f_mapping;
2415 spin_lock(&swap_lock);
2416 plist_for_each_entry(p, &swap_active_head, list) {
2417 if (p->flags & SWP_WRITEOK) {
2418 if (p->swap_file->f_mapping == mapping) {
2426 spin_unlock(&swap_lock);
2429 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2430 vm_unacct_memory(p->pages);
2433 spin_unlock(&swap_lock);
2436 del_from_avail_list(p);
2437 spin_lock(&p->lock);
2439 struct swap_info_struct *si = p;
2442 plist_for_each_entry_continue(si, &swap_active_head, list) {
2445 for_each_node(nid) {
2446 if (si->avail_lists[nid].prio != 1)
2447 si->avail_lists[nid].prio--;
2452 plist_del(&p->list, &swap_active_head);
2453 atomic_long_sub(p->pages, &nr_swap_pages);
2454 total_swap_pages -= p->pages;
2455 p->flags &= ~SWP_WRITEOK;
2456 spin_unlock(&p->lock);
2457 spin_unlock(&swap_lock);
2459 disable_swap_slots_cache_lock();
2461 set_current_oom_origin();
2462 err = try_to_unuse(p->type);
2463 clear_current_oom_origin();
2466 /* re-insert swap space back into swap_list */
2467 reinsert_swap_info(p);
2468 reenable_swap_slots_cache_unlock();
2472 reenable_swap_slots_cache_unlock();
2475 * Wait for swap operations protected by get/put_swap_device()
2478 * We need synchronize_rcu() here to protect the accessing to
2479 * the swap cache data structure.
2481 percpu_ref_kill(&p->users);
2483 wait_for_completion(&p->comp);
2485 flush_work(&p->discard_work);
2487 destroy_swap_extents(p);
2488 if (p->flags & SWP_CONTINUED)
2489 free_swap_count_continuations(p);
2491 if (!p->bdev || !bdev_nonrot(p->bdev))
2492 atomic_dec(&nr_rotate_swap);
2494 mutex_lock(&swapon_mutex);
2495 spin_lock(&swap_lock);
2496 spin_lock(&p->lock);
2499 /* wait for anyone still in scan_swap_map_slots */
2500 p->highest_bit = 0; /* cuts scans short */
2501 while (p->flags >= SWP_SCANNING) {
2502 spin_unlock(&p->lock);
2503 spin_unlock(&swap_lock);
2504 schedule_timeout_uninterruptible(1);
2505 spin_lock(&swap_lock);
2506 spin_lock(&p->lock);
2509 swap_file = p->swap_file;
2510 old_block_size = p->old_block_size;
2511 p->swap_file = NULL;
2513 swap_map = p->swap_map;
2515 cluster_info = p->cluster_info;
2516 p->cluster_info = NULL;
2517 frontswap_map = frontswap_map_get(p);
2518 spin_unlock(&p->lock);
2519 spin_unlock(&swap_lock);
2520 arch_swap_invalidate_area(p->type);
2521 frontswap_invalidate_area(p->type);
2522 frontswap_map_set(p, NULL);
2523 mutex_unlock(&swapon_mutex);
2524 free_percpu(p->percpu_cluster);
2525 p->percpu_cluster = NULL;
2526 free_percpu(p->cluster_next_cpu);
2527 p->cluster_next_cpu = NULL;
2529 kvfree(cluster_info);
2530 kvfree(frontswap_map);
2531 /* Destroy swap account information */
2532 swap_cgroup_swapoff(p->type);
2533 exit_swap_address_space(p->type);
2535 inode = mapping->host;
2536 if (S_ISBLK(inode->i_mode)) {
2537 struct block_device *bdev = I_BDEV(inode);
2539 set_blocksize(bdev, old_block_size);
2540 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2544 inode->i_flags &= ~S_SWAPFILE;
2545 inode_unlock(inode);
2546 filp_close(swap_file, NULL);
2549 * Clear the SWP_USED flag after all resources are freed so that swapon
2550 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2551 * not hold p->lock after we cleared its SWP_WRITEOK.
2553 spin_lock(&swap_lock);
2555 spin_unlock(&swap_lock);
2558 atomic_inc(&proc_poll_event);
2559 wake_up_interruptible(&proc_poll_wait);
2562 filp_close(victim, NULL);
2568 #ifdef CONFIG_PROC_FS
2569 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2571 struct seq_file *seq = file->private_data;
2573 poll_wait(file, &proc_poll_wait, wait);
2575 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2576 seq->poll_event = atomic_read(&proc_poll_event);
2577 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2580 return EPOLLIN | EPOLLRDNORM;
2584 static void *swap_start(struct seq_file *swap, loff_t *pos)
2586 struct swap_info_struct *si;
2590 mutex_lock(&swapon_mutex);
2593 return SEQ_START_TOKEN;
2595 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2596 if (!(si->flags & SWP_USED) || !si->swap_map)
2605 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2607 struct swap_info_struct *si = v;
2610 if (v == SEQ_START_TOKEN)
2613 type = si->type + 1;
2616 for (; (si = swap_type_to_swap_info(type)); type++) {
2617 if (!(si->flags & SWP_USED) || !si->swap_map)
2625 static void swap_stop(struct seq_file *swap, void *v)
2627 mutex_unlock(&swapon_mutex);
2630 static int swap_show(struct seq_file *swap, void *v)
2632 struct swap_info_struct *si = v;
2635 unsigned long bytes, inuse;
2637 if (si == SEQ_START_TOKEN) {
2638 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2642 bytes = si->pages << (PAGE_SHIFT - 10);
2643 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2645 file = si->swap_file;
2646 len = seq_file_path(swap, file, " \t\n\\");
2647 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
2648 len < 40 ? 40 - len : 1, " ",
2649 S_ISBLK(file_inode(file)->i_mode) ?
2650 "partition" : "file\t",
2651 bytes, bytes < 10000000 ? "\t" : "",
2652 inuse, inuse < 10000000 ? "\t" : "",
2657 static const struct seq_operations swaps_op = {
2658 .start = swap_start,
2664 static int swaps_open(struct inode *inode, struct file *file)
2666 struct seq_file *seq;
2669 ret = seq_open(file, &swaps_op);
2673 seq = file->private_data;
2674 seq->poll_event = atomic_read(&proc_poll_event);
2678 static const struct proc_ops swaps_proc_ops = {
2679 .proc_flags = PROC_ENTRY_PERMANENT,
2680 .proc_open = swaps_open,
2681 .proc_read = seq_read,
2682 .proc_lseek = seq_lseek,
2683 .proc_release = seq_release,
2684 .proc_poll = swaps_poll,
2687 static int __init procswaps_init(void)
2689 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2692 __initcall(procswaps_init);
2693 #endif /* CONFIG_PROC_FS */
2695 #ifdef MAX_SWAPFILES_CHECK
2696 static int __init max_swapfiles_check(void)
2698 MAX_SWAPFILES_CHECK();
2701 late_initcall(max_swapfiles_check);
2704 static struct swap_info_struct *alloc_swap_info(void)
2706 struct swap_info_struct *p;
2707 struct swap_info_struct *defer = NULL;
2711 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2713 return ERR_PTR(-ENOMEM);
2715 if (percpu_ref_init(&p->users, swap_users_ref_free,
2716 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2718 return ERR_PTR(-ENOMEM);
2721 spin_lock(&swap_lock);
2722 for (type = 0; type < nr_swapfiles; type++) {
2723 if (!(swap_info[type]->flags & SWP_USED))
2726 if (type >= MAX_SWAPFILES) {
2727 spin_unlock(&swap_lock);
2728 percpu_ref_exit(&p->users);
2730 return ERR_PTR(-EPERM);
2732 if (type >= nr_swapfiles) {
2735 * Publish the swap_info_struct after initializing it.
2736 * Note that kvzalloc() above zeroes all its fields.
2738 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2742 p = swap_info[type];
2744 * Do not memset this entry: a racing procfs swap_next()
2745 * would be relying on p->type to remain valid.
2748 p->swap_extent_root = RB_ROOT;
2749 plist_node_init(&p->list, 0);
2751 plist_node_init(&p->avail_lists[i], 0);
2752 p->flags = SWP_USED;
2753 spin_unlock(&swap_lock);
2755 percpu_ref_exit(&defer->users);
2758 spin_lock_init(&p->lock);
2759 spin_lock_init(&p->cont_lock);
2760 init_completion(&p->comp);
2765 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2769 if (S_ISBLK(inode->i_mode)) {
2770 p->bdev = blkdev_get_by_dev(inode->i_rdev,
2771 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2772 if (IS_ERR(p->bdev)) {
2773 error = PTR_ERR(p->bdev);
2777 p->old_block_size = block_size(p->bdev);
2778 error = set_blocksize(p->bdev, PAGE_SIZE);
2782 * Zoned block devices contain zones that have a sequential
2783 * write only restriction. Hence zoned block devices are not
2784 * suitable for swapping. Disallow them here.
2786 if (bdev_is_zoned(p->bdev))
2788 p->flags |= SWP_BLKDEV;
2789 } else if (S_ISREG(inode->i_mode)) {
2790 p->bdev = inode->i_sb->s_bdev;
2798 * Find out how many pages are allowed for a single swap device. There
2799 * are two limiting factors:
2800 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2801 * 2) the number of bits in the swap pte, as defined by the different
2804 * In order to find the largest possible bit mask, a swap entry with
2805 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2806 * decoded to a swp_entry_t again, and finally the swap offset is
2809 * This will mask all the bits from the initial ~0UL mask that can't
2810 * be encoded in either the swp_entry_t or the architecture definition
2813 unsigned long generic_max_swapfile_size(void)
2815 return swp_offset(pte_to_swp_entry(
2816 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2819 /* Can be overridden by an architecture for additional checks. */
2820 __weak unsigned long max_swapfile_size(void)
2822 return generic_max_swapfile_size();
2825 static unsigned long read_swap_header(struct swap_info_struct *p,
2826 union swap_header *swap_header,
2827 struct inode *inode)
2830 unsigned long maxpages;
2831 unsigned long swapfilepages;
2832 unsigned long last_page;
2834 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2835 pr_err("Unable to find swap-space signature\n");
2839 /* swap partition endianness hack... */
2840 if (swab32(swap_header->info.version) == 1) {
2841 swab32s(&swap_header->info.version);
2842 swab32s(&swap_header->info.last_page);
2843 swab32s(&swap_header->info.nr_badpages);
2844 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2846 for (i = 0; i < swap_header->info.nr_badpages; i++)
2847 swab32s(&swap_header->info.badpages[i]);
2849 /* Check the swap header's sub-version */
2850 if (swap_header->info.version != 1) {
2851 pr_warn("Unable to handle swap header version %d\n",
2852 swap_header->info.version);
2857 p->cluster_next = 1;
2860 maxpages = max_swapfile_size();
2861 last_page = swap_header->info.last_page;
2863 pr_warn("Empty swap-file\n");
2866 if (last_page > maxpages) {
2867 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2868 maxpages << (PAGE_SHIFT - 10),
2869 last_page << (PAGE_SHIFT - 10));
2871 if (maxpages > last_page) {
2872 maxpages = last_page + 1;
2873 /* p->max is an unsigned int: don't overflow it */
2874 if ((unsigned int)maxpages == 0)
2875 maxpages = UINT_MAX;
2877 p->highest_bit = maxpages - 1;
2881 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2882 if (swapfilepages && maxpages > swapfilepages) {
2883 pr_warn("Swap area shorter than signature indicates\n");
2886 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2888 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2894 #define SWAP_CLUSTER_INFO_COLS \
2895 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2896 #define SWAP_CLUSTER_SPACE_COLS \
2897 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2898 #define SWAP_CLUSTER_COLS \
2899 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2901 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2902 union swap_header *swap_header,
2903 unsigned char *swap_map,
2904 struct swap_cluster_info *cluster_info,
2905 unsigned long maxpages,
2909 unsigned int nr_good_pages;
2911 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2912 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2913 unsigned long i, idx;
2915 nr_good_pages = maxpages - 1; /* omit header page */
2917 cluster_list_init(&p->free_clusters);
2918 cluster_list_init(&p->discard_clusters);
2920 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2921 unsigned int page_nr = swap_header->info.badpages[i];
2922 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2924 if (page_nr < maxpages) {
2925 swap_map[page_nr] = SWAP_MAP_BAD;
2928 * Haven't marked the cluster free yet, no list
2929 * operation involved
2931 inc_cluster_info_page(p, cluster_info, page_nr);
2935 /* Haven't marked the cluster free yet, no list operation involved */
2936 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2937 inc_cluster_info_page(p, cluster_info, i);
2939 if (nr_good_pages) {
2940 swap_map[0] = SWAP_MAP_BAD;
2942 * Not mark the cluster free yet, no list
2943 * operation involved
2945 inc_cluster_info_page(p, cluster_info, 0);
2947 p->pages = nr_good_pages;
2948 nr_extents = setup_swap_extents(p, span);
2951 nr_good_pages = p->pages;
2953 if (!nr_good_pages) {
2954 pr_warn("Empty swap-file\n");
2963 * Reduce false cache line sharing between cluster_info and
2964 * sharing same address space.
2966 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2967 j = (k + col) % SWAP_CLUSTER_COLS;
2968 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2969 idx = i * SWAP_CLUSTER_COLS + j;
2970 if (idx >= nr_clusters)
2972 if (cluster_count(&cluster_info[idx]))
2974 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2975 cluster_list_add_tail(&p->free_clusters, cluster_info,
2982 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2984 struct swap_info_struct *p;
2985 struct filename *name;
2986 struct file *swap_file = NULL;
2987 struct address_space *mapping;
2988 struct dentry *dentry;
2991 union swap_header *swap_header;
2994 unsigned long maxpages;
2995 unsigned char *swap_map = NULL;
2996 struct swap_cluster_info *cluster_info = NULL;
2997 unsigned long *frontswap_map = NULL;
2998 struct page *page = NULL;
2999 struct inode *inode = NULL;
3000 bool inced_nr_rotate_swap = false;
3002 if (swap_flags & ~SWAP_FLAGS_VALID)
3005 if (!capable(CAP_SYS_ADMIN))
3008 if (!swap_avail_heads)
3011 p = alloc_swap_info();
3015 INIT_WORK(&p->discard_work, swap_discard_work);
3017 name = getname(specialfile);
3019 error = PTR_ERR(name);
3023 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3024 if (IS_ERR(swap_file)) {
3025 error = PTR_ERR(swap_file);
3030 p->swap_file = swap_file;
3031 mapping = swap_file->f_mapping;
3032 dentry = swap_file->f_path.dentry;
3033 inode = mapping->host;
3035 error = claim_swapfile(p, inode);
3036 if (unlikely(error))
3040 if (d_unlinked(dentry) || cant_mount(dentry)) {
3042 goto bad_swap_unlock_inode;
3044 if (IS_SWAPFILE(inode)) {
3046 goto bad_swap_unlock_inode;
3050 * Read the swap header.
3052 if (!mapping->a_ops->read_folio) {
3054 goto bad_swap_unlock_inode;
3056 page = read_mapping_page(mapping, 0, swap_file);
3058 error = PTR_ERR(page);
3059 goto bad_swap_unlock_inode;
3061 swap_header = kmap(page);
3063 maxpages = read_swap_header(p, swap_header, inode);
3064 if (unlikely(!maxpages)) {
3066 goto bad_swap_unlock_inode;
3069 /* OK, set up the swap map and apply the bad block list */
3070 swap_map = vzalloc(maxpages);
3073 goto bad_swap_unlock_inode;
3076 if (p->bdev && bdev_stable_writes(p->bdev))
3077 p->flags |= SWP_STABLE_WRITES;
3079 if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3080 p->flags |= SWP_SYNCHRONOUS_IO;
3082 if (p->bdev && bdev_nonrot(p->bdev)) {
3084 unsigned long ci, nr_cluster;
3086 p->flags |= SWP_SOLIDSTATE;
3087 p->cluster_next_cpu = alloc_percpu(unsigned int);
3088 if (!p->cluster_next_cpu) {
3090 goto bad_swap_unlock_inode;
3093 * select a random position to start with to help wear leveling
3096 for_each_possible_cpu(cpu) {
3097 per_cpu(*p->cluster_next_cpu, cpu) =
3098 1 + prandom_u32_max(p->highest_bit);
3100 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3102 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3104 if (!cluster_info) {
3106 goto bad_swap_unlock_inode;
3109 for (ci = 0; ci < nr_cluster; ci++)
3110 spin_lock_init(&((cluster_info + ci)->lock));
3112 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3113 if (!p->percpu_cluster) {
3115 goto bad_swap_unlock_inode;
3117 for_each_possible_cpu(cpu) {
3118 struct percpu_cluster *cluster;
3119 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3120 cluster_set_null(&cluster->index);
3123 atomic_inc(&nr_rotate_swap);
3124 inced_nr_rotate_swap = true;
3127 error = swap_cgroup_swapon(p->type, maxpages);
3129 goto bad_swap_unlock_inode;
3131 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3132 cluster_info, maxpages, &span);
3133 if (unlikely(nr_extents < 0)) {
3135 goto bad_swap_unlock_inode;
3137 /* frontswap enabled? set up bit-per-page map for frontswap */
3138 if (IS_ENABLED(CONFIG_FRONTSWAP))
3139 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3143 if ((swap_flags & SWAP_FLAG_DISCARD) &&
3144 p->bdev && bdev_max_discard_sectors(p->bdev)) {
3146 * When discard is enabled for swap with no particular
3147 * policy flagged, we set all swap discard flags here in
3148 * order to sustain backward compatibility with older
3149 * swapon(8) releases.
3151 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3155 * By flagging sys_swapon, a sysadmin can tell us to
3156 * either do single-time area discards only, or to just
3157 * perform discards for released swap page-clusters.
3158 * Now it's time to adjust the p->flags accordingly.
3160 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3161 p->flags &= ~SWP_PAGE_DISCARD;
3162 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3163 p->flags &= ~SWP_AREA_DISCARD;
3165 /* issue a swapon-time discard if it's still required */
3166 if (p->flags & SWP_AREA_DISCARD) {
3167 int err = discard_swap(p);
3169 pr_err("swapon: discard_swap(%p): %d\n",
3174 error = init_swap_address_space(p->type, maxpages);
3176 goto bad_swap_unlock_inode;
3179 * Flush any pending IO and dirty mappings before we start using this
3182 inode->i_flags |= S_SWAPFILE;
3183 error = inode_drain_writes(inode);
3185 inode->i_flags &= ~S_SWAPFILE;
3186 goto free_swap_address_space;
3189 mutex_lock(&swapon_mutex);
3191 if (swap_flags & SWAP_FLAG_PREFER)
3193 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3194 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3196 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3197 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3198 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3199 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3200 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3201 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3202 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3203 (frontswap_map) ? "FS" : "");
3205 mutex_unlock(&swapon_mutex);
3206 atomic_inc(&proc_poll_event);
3207 wake_up_interruptible(&proc_poll_wait);
3211 free_swap_address_space:
3212 exit_swap_address_space(p->type);
3213 bad_swap_unlock_inode:
3214 inode_unlock(inode);
3216 free_percpu(p->percpu_cluster);
3217 p->percpu_cluster = NULL;
3218 free_percpu(p->cluster_next_cpu);
3219 p->cluster_next_cpu = NULL;
3220 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3221 set_blocksize(p->bdev, p->old_block_size);
3222 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3225 destroy_swap_extents(p);
3226 swap_cgroup_swapoff(p->type);
3227 spin_lock(&swap_lock);
3228 p->swap_file = NULL;
3230 spin_unlock(&swap_lock);
3232 kvfree(cluster_info);
3233 kvfree(frontswap_map);
3234 if (inced_nr_rotate_swap)
3235 atomic_dec(&nr_rotate_swap);
3237 filp_close(swap_file, NULL);
3239 if (page && !IS_ERR(page)) {
3246 inode_unlock(inode);
3248 enable_swap_slots_cache();
3252 void si_swapinfo(struct sysinfo *val)
3255 unsigned long nr_to_be_unused = 0;
3257 spin_lock(&swap_lock);
3258 for (type = 0; type < nr_swapfiles; type++) {
3259 struct swap_info_struct *si = swap_info[type];
3261 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3262 nr_to_be_unused += si->inuse_pages;
3264 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3265 val->totalswap = total_swap_pages + nr_to_be_unused;
3266 spin_unlock(&swap_lock);
3270 * Verify that a swap entry is valid and increment its swap map count.
3272 * Returns error code in following case.
3274 * - swp_entry is invalid -> EINVAL
3275 * - swp_entry is migration entry -> EINVAL
3276 * - swap-cache reference is requested but there is already one. -> EEXIST
3277 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3278 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3280 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3282 struct swap_info_struct *p;
3283 struct swap_cluster_info *ci;
3284 unsigned long offset;
3285 unsigned char count;
3286 unsigned char has_cache;
3289 p = get_swap_device(entry);
3293 offset = swp_offset(entry);
3294 ci = lock_cluster_or_swap_info(p, offset);
3296 count = p->swap_map[offset];
3299 * swapin_readahead() doesn't check if a swap entry is valid, so the
3300 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3302 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3307 has_cache = count & SWAP_HAS_CACHE;
3308 count &= ~SWAP_HAS_CACHE;
3311 if (usage == SWAP_HAS_CACHE) {
3313 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3314 if (!has_cache && count)
3315 has_cache = SWAP_HAS_CACHE;
3316 else if (has_cache) /* someone else added cache */
3318 else /* no users remaining */
3321 } else if (count || has_cache) {
3323 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3325 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3327 else if (swap_count_continued(p, offset, count))
3328 count = COUNT_CONTINUED;
3332 err = -ENOENT; /* unused swap entry */
3334 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3337 unlock_cluster_or_swap_info(p, ci);
3343 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3344 * (in which case its reference count is never incremented).
3346 void swap_shmem_alloc(swp_entry_t entry)
3348 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3352 * Increase reference count of swap entry by 1.
3353 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3354 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3355 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3356 * might occur if a page table entry has got corrupted.
3358 int swap_duplicate(swp_entry_t entry)
3362 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3363 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3368 * @entry: swap entry for which we allocate swap cache.
3370 * Called when allocating swap cache for existing swap entry,
3371 * This can return error codes. Returns 0 at success.
3372 * -EEXIST means there is a swap cache.
3373 * Note: return code is different from swap_duplicate().
3375 int swapcache_prepare(swp_entry_t entry)
3377 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3380 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3382 return swap_type_to_swap_info(swp_type(entry));
3385 struct swap_info_struct *page_swap_info(struct page *page)
3387 swp_entry_t entry = { .val = page_private(page) };
3388 return swp_swap_info(entry);
3392 * out-of-line methods to avoid include hell.
3394 struct address_space *swapcache_mapping(struct folio *folio)
3396 return page_swap_info(&folio->page)->swap_file->f_mapping;
3398 EXPORT_SYMBOL_GPL(swapcache_mapping);
3400 pgoff_t __page_file_index(struct page *page)
3402 swp_entry_t swap = { .val = page_private(page) };
3403 return swp_offset(swap);
3405 EXPORT_SYMBOL_GPL(__page_file_index);
3408 * add_swap_count_continuation - called when a swap count is duplicated
3409 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3410 * page of the original vmalloc'ed swap_map, to hold the continuation count
3411 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3412 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3414 * These continuation pages are seldom referenced: the common paths all work
3415 * on the original swap_map, only referring to a continuation page when the
3416 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3418 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3419 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3420 * can be called after dropping locks.
3422 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3424 struct swap_info_struct *si;
3425 struct swap_cluster_info *ci;
3428 struct page *list_page;
3430 unsigned char count;
3434 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3435 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3437 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3439 si = get_swap_device(entry);
3442 * An acceptable race has occurred since the failing
3443 * __swap_duplicate(): the swap device may be swapoff
3447 spin_lock(&si->lock);
3449 offset = swp_offset(entry);
3451 ci = lock_cluster(si, offset);
3453 count = swap_count(si->swap_map[offset]);
3455 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3457 * The higher the swap count, the more likely it is that tasks
3458 * will race to add swap count continuation: we need to avoid
3459 * over-provisioning.
3470 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3471 * no architecture is using highmem pages for kernel page tables: so it
3472 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3474 head = vmalloc_to_page(si->swap_map + offset);
3475 offset &= ~PAGE_MASK;
3477 spin_lock(&si->cont_lock);
3479 * Page allocation does not initialize the page's lru field,
3480 * but it does always reset its private field.
3482 if (!page_private(head)) {
3483 BUG_ON(count & COUNT_CONTINUED);
3484 INIT_LIST_HEAD(&head->lru);
3485 set_page_private(head, SWP_CONTINUED);
3486 si->flags |= SWP_CONTINUED;
3489 list_for_each_entry(list_page, &head->lru, lru) {
3493 * If the previous map said no continuation, but we've found
3494 * a continuation page, free our allocation and use this one.
3496 if (!(count & COUNT_CONTINUED))
3497 goto out_unlock_cont;
3499 map = kmap_atomic(list_page) + offset;
3504 * If this continuation count now has some space in it,
3505 * free our allocation and use this one.
3507 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3508 goto out_unlock_cont;
3511 list_add_tail(&page->lru, &head->lru);
3512 page = NULL; /* now it's attached, don't free it */
3514 spin_unlock(&si->cont_lock);
3517 spin_unlock(&si->lock);
3518 put_swap_device(si);
3526 * swap_count_continued - when the original swap_map count is incremented
3527 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3528 * into, carry if so, or else fail until a new continuation page is allocated;
3529 * when the original swap_map count is decremented from 0 with continuation,
3530 * borrow from the continuation and report whether it still holds more.
3531 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3534 static bool swap_count_continued(struct swap_info_struct *si,
3535 pgoff_t offset, unsigned char count)
3542 head = vmalloc_to_page(si->swap_map + offset);
3543 if (page_private(head) != SWP_CONTINUED) {
3544 BUG_ON(count & COUNT_CONTINUED);
3545 return false; /* need to add count continuation */
3548 spin_lock(&si->cont_lock);
3549 offset &= ~PAGE_MASK;
3550 page = list_next_entry(head, lru);
3551 map = kmap_atomic(page) + offset;
3553 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3554 goto init_map; /* jump over SWAP_CONT_MAX checks */
3556 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3558 * Think of how you add 1 to 999
3560 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3562 page = list_next_entry(page, lru);
3563 BUG_ON(page == head);
3564 map = kmap_atomic(page) + offset;
3566 if (*map == SWAP_CONT_MAX) {
3568 page = list_next_entry(page, lru);
3570 ret = false; /* add count continuation */
3573 map = kmap_atomic(page) + offset;
3574 init_map: *map = 0; /* we didn't zero the page */
3578 while ((page = list_prev_entry(page, lru)) != head) {
3579 map = kmap_atomic(page) + offset;
3580 *map = COUNT_CONTINUED;
3583 ret = true; /* incremented */
3585 } else { /* decrementing */
3587 * Think of how you subtract 1 from 1000
3589 BUG_ON(count != COUNT_CONTINUED);
3590 while (*map == COUNT_CONTINUED) {
3592 page = list_next_entry(page, lru);
3593 BUG_ON(page == head);
3594 map = kmap_atomic(page) + offset;
3601 while ((page = list_prev_entry(page, lru)) != head) {
3602 map = kmap_atomic(page) + offset;
3603 *map = SWAP_CONT_MAX | count;
3604 count = COUNT_CONTINUED;
3607 ret = count == COUNT_CONTINUED;
3610 spin_unlock(&si->cont_lock);
3615 * free_swap_count_continuations - swapoff free all the continuation pages
3616 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3618 static void free_swap_count_continuations(struct swap_info_struct *si)
3622 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3624 head = vmalloc_to_page(si->swap_map + offset);
3625 if (page_private(head)) {
3626 struct page *page, *next;
3628 list_for_each_entry_safe(page, next, &head->lru, lru) {
3629 list_del(&page->lru);
3636 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3637 void __cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3639 struct swap_info_struct *si, *next;
3640 int nid = page_to_nid(page);
3642 if (!(gfp_mask & __GFP_IO))
3645 if (!blk_cgroup_congested())
3649 * We've already scheduled a throttle, avoid taking the global swap
3652 if (current->throttle_queue)
3655 spin_lock(&swap_avail_lock);
3656 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3659 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3663 spin_unlock(&swap_avail_lock);
3667 static int __init swapfile_init(void)
3671 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3673 if (!swap_avail_heads) {
3674 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3679 plist_head_init(&swap_avail_heads[nid]);
3683 subsys_initcall(swapfile_init);