1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
10 #include <linux/sched/mm.h>
11 #include <linux/sched/task.h>
12 #include <linux/hugetlb.h>
13 #include <linux/mman.h>
14 #include <linux/slab.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/swap.h>
17 #include <linux/vmalloc.h>
18 #include <linux/pagemap.h>
19 #include <linux/namei.h>
20 #include <linux/shmem_fs.h>
21 #include <linux/blkdev.h>
22 #include <linux/random.h>
23 #include <linux/writeback.h>
24 #include <linux/proc_fs.h>
25 #include <linux/seq_file.h>
26 #include <linux/init.h>
27 #include <linux/ksm.h>
28 #include <linux/rmap.h>
29 #include <linux/security.h>
30 #include <linux/backing-dev.h>
31 #include <linux/mutex.h>
32 #include <linux/capability.h>
33 #include <linux/syscalls.h>
34 #include <linux/memcontrol.h>
35 #include <linux/poll.h>
36 #include <linux/oom.h>
37 #include <linux/frontswap.h>
38 #include <linux/swapfile.h>
39 #include <linux/export.h>
40 #include <linux/swap_slots.h>
41 #include <linux/sort.h>
43 #include <asm/tlbflush.h>
44 #include <linux/swapops.h>
45 #include <linux/swap_cgroup.h>
47 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
49 static void free_swap_count_continuations(struct swap_info_struct *);
50 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
52 DEFINE_SPINLOCK(swap_lock);
53 static unsigned int nr_swapfiles;
54 atomic_long_t nr_swap_pages;
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
60 EXPORT_SYMBOL_GPL(nr_swap_pages);
61 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62 long total_swap_pages;
63 static int least_priority = -1;
65 static const char Bad_file[] = "Bad swap file entry ";
66 static const char Unused_file[] = "Unused swap file entry ";
67 static const char Bad_offset[] = "Bad swap offset entry ";
68 static const char Unused_offset[] = "Unused swap offset entry ";
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
74 PLIST_HEAD(swap_active_head);
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
88 static struct plist_head *swap_avail_heads;
89 static DEFINE_SPINLOCK(swap_avail_lock);
91 struct swap_info_struct *swap_info[MAX_SWAPFILES];
93 static DEFINE_MUTEX(swapon_mutex);
95 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96 /* Activity counter to indicate that a swapon or swapoff has occurred */
97 static atomic_t proc_poll_event = ATOMIC_INIT(0);
99 atomic_t nr_rotate_swap = ATOMIC_INIT(0);
101 static struct swap_info_struct *swap_type_to_swap_info(int type)
103 if (type >= READ_ONCE(nr_swapfiles))
106 smp_rmb(); /* Pairs with smp_wmb in alloc_swap_info. */
107 return READ_ONCE(swap_info[type]);
110 static inline unsigned char swap_count(unsigned char ent)
112 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
115 /* Reclaim the swap entry anyway if possible */
116 #define TTRS_ANYWAY 0x1
118 * Reclaim the swap entry if there are no more mappings of the
121 #define TTRS_UNMAPPED 0x2
122 /* Reclaim the swap entry if swap is getting full*/
123 #define TTRS_FULL 0x4
125 /* returns 1 if swap entry is freed */
126 static int __try_to_reclaim_swap(struct swap_info_struct *si,
127 unsigned long offset, unsigned long flags)
129 swp_entry_t entry = swp_entry(si->type, offset);
133 page = find_get_page(swap_address_space(entry), offset);
137 * When this function is called from scan_swap_map_slots() and it's
138 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
139 * here. We have to use trylock for avoiding deadlock. This is a special
140 * case and you should use try_to_free_swap() with explicit lock_page()
141 * in usual operations.
143 if (trylock_page(page)) {
144 if ((flags & TTRS_ANYWAY) ||
145 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
146 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
147 ret = try_to_free_swap(page);
154 static inline struct swap_extent *first_se(struct swap_info_struct *sis)
156 struct rb_node *rb = rb_first(&sis->swap_extent_root);
157 return rb_entry(rb, struct swap_extent, rb_node);
160 static inline struct swap_extent *next_se(struct swap_extent *se)
162 struct rb_node *rb = rb_next(&se->rb_node);
163 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
167 * swapon tell device that all the old swap contents can be discarded,
168 * to allow the swap device to optimize its wear-levelling.
170 static int discard_swap(struct swap_info_struct *si)
172 struct swap_extent *se;
173 sector_t start_block;
177 /* Do not discard the swap header page! */
179 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
180 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
182 err = blkdev_issue_discard(si->bdev, start_block,
183 nr_blocks, GFP_KERNEL, 0);
189 for (se = next_se(se); se; se = next_se(se)) {
190 start_block = se->start_block << (PAGE_SHIFT - 9);
191 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
193 err = blkdev_issue_discard(si->bdev, start_block,
194 nr_blocks, GFP_KERNEL, 0);
200 return err; /* That will often be -EOPNOTSUPP */
203 static struct swap_extent *
204 offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
206 struct swap_extent *se;
209 rb = sis->swap_extent_root.rb_node;
211 se = rb_entry(rb, struct swap_extent, rb_node);
212 if (offset < se->start_page)
214 else if (offset >= se->start_page + se->nr_pages)
219 /* It *must* be present */
223 sector_t swap_page_sector(struct page *page)
225 struct swap_info_struct *sis = page_swap_info(page);
226 struct swap_extent *se;
230 offset = __page_file_index(page);
231 se = offset_to_swap_extent(sis, offset);
232 sector = se->start_block + (offset - se->start_page);
233 return sector << (PAGE_SHIFT - 9);
237 * swap allocation tell device that a cluster of swap can now be discarded,
238 * to allow the swap device to optimize its wear-levelling.
240 static void discard_swap_cluster(struct swap_info_struct *si,
241 pgoff_t start_page, pgoff_t nr_pages)
243 struct swap_extent *se = offset_to_swap_extent(si, start_page);
246 pgoff_t offset = start_page - se->start_page;
247 sector_t start_block = se->start_block + offset;
248 sector_t nr_blocks = se->nr_pages - offset;
250 if (nr_blocks > nr_pages)
251 nr_blocks = nr_pages;
252 start_page += nr_blocks;
253 nr_pages -= nr_blocks;
255 start_block <<= PAGE_SHIFT - 9;
256 nr_blocks <<= PAGE_SHIFT - 9;
257 if (blkdev_issue_discard(si->bdev, start_block,
258 nr_blocks, GFP_NOIO, 0))
265 #ifdef CONFIG_THP_SWAP
266 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
268 #define swap_entry_size(size) (size)
270 #define SWAPFILE_CLUSTER 256
273 * Define swap_entry_size() as constant to let compiler to optimize
274 * out some code if !CONFIG_THP_SWAP
276 #define swap_entry_size(size) 1
278 #define LATENCY_LIMIT 256
280 static inline void cluster_set_flag(struct swap_cluster_info *info,
286 static inline unsigned int cluster_count(struct swap_cluster_info *info)
291 static inline void cluster_set_count(struct swap_cluster_info *info,
297 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
298 unsigned int c, unsigned int f)
304 static inline unsigned int cluster_next(struct swap_cluster_info *info)
309 static inline void cluster_set_next(struct swap_cluster_info *info,
315 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
316 unsigned int n, unsigned int f)
322 static inline bool cluster_is_free(struct swap_cluster_info *info)
324 return info->flags & CLUSTER_FLAG_FREE;
327 static inline bool cluster_is_null(struct swap_cluster_info *info)
329 return info->flags & CLUSTER_FLAG_NEXT_NULL;
332 static inline void cluster_set_null(struct swap_cluster_info *info)
334 info->flags = CLUSTER_FLAG_NEXT_NULL;
338 static inline bool cluster_is_huge(struct swap_cluster_info *info)
340 if (IS_ENABLED(CONFIG_THP_SWAP))
341 return info->flags & CLUSTER_FLAG_HUGE;
345 static inline void cluster_clear_huge(struct swap_cluster_info *info)
347 info->flags &= ~CLUSTER_FLAG_HUGE;
350 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
351 unsigned long offset)
353 struct swap_cluster_info *ci;
355 ci = si->cluster_info;
357 ci += offset / SWAPFILE_CLUSTER;
358 spin_lock(&ci->lock);
363 static inline void unlock_cluster(struct swap_cluster_info *ci)
366 spin_unlock(&ci->lock);
370 * Determine the locking method in use for this device. Return
371 * swap_cluster_info if SSD-style cluster-based locking is in place.
373 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
374 struct swap_info_struct *si, unsigned long offset)
376 struct swap_cluster_info *ci;
378 /* Try to use fine-grained SSD-style locking if available: */
379 ci = lock_cluster(si, offset);
380 /* Otherwise, fall back to traditional, coarse locking: */
382 spin_lock(&si->lock);
387 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
388 struct swap_cluster_info *ci)
393 spin_unlock(&si->lock);
396 static inline bool cluster_list_empty(struct swap_cluster_list *list)
398 return cluster_is_null(&list->head);
401 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
403 return cluster_next(&list->head);
406 static void cluster_list_init(struct swap_cluster_list *list)
408 cluster_set_null(&list->head);
409 cluster_set_null(&list->tail);
412 static void cluster_list_add_tail(struct swap_cluster_list *list,
413 struct swap_cluster_info *ci,
416 if (cluster_list_empty(list)) {
417 cluster_set_next_flag(&list->head, idx, 0);
418 cluster_set_next_flag(&list->tail, idx, 0);
420 struct swap_cluster_info *ci_tail;
421 unsigned int tail = cluster_next(&list->tail);
424 * Nested cluster lock, but both cluster locks are
425 * only acquired when we held swap_info_struct->lock
428 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
429 cluster_set_next(ci_tail, idx);
430 spin_unlock(&ci_tail->lock);
431 cluster_set_next_flag(&list->tail, idx, 0);
435 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
436 struct swap_cluster_info *ci)
440 idx = cluster_next(&list->head);
441 if (cluster_next(&list->tail) == idx) {
442 cluster_set_null(&list->head);
443 cluster_set_null(&list->tail);
445 cluster_set_next_flag(&list->head,
446 cluster_next(&ci[idx]), 0);
451 /* Add a cluster to discard list and schedule it to do discard */
452 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
456 * If scan_swap_map() can't find a free cluster, it will check
457 * si->swap_map directly. To make sure the discarding cluster isn't
458 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
459 * will be cleared after discard
461 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
462 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
464 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
466 schedule_work(&si->discard_work);
469 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
471 struct swap_cluster_info *ci = si->cluster_info;
473 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
474 cluster_list_add_tail(&si->free_clusters, ci, idx);
478 * Doing discard actually. After a cluster discard is finished, the cluster
479 * will be added to free cluster list. caller should hold si->lock.
481 static void swap_do_scheduled_discard(struct swap_info_struct *si)
483 struct swap_cluster_info *info, *ci;
486 info = si->cluster_info;
488 while (!cluster_list_empty(&si->discard_clusters)) {
489 idx = cluster_list_del_first(&si->discard_clusters, info);
490 spin_unlock(&si->lock);
492 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
495 spin_lock(&si->lock);
496 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
497 __free_cluster(si, idx);
498 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
499 0, SWAPFILE_CLUSTER);
504 static void swap_discard_work(struct work_struct *work)
506 struct swap_info_struct *si;
508 si = container_of(work, struct swap_info_struct, discard_work);
510 spin_lock(&si->lock);
511 swap_do_scheduled_discard(si);
512 spin_unlock(&si->lock);
515 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
517 struct swap_cluster_info *ci = si->cluster_info;
519 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
520 cluster_list_del_first(&si->free_clusters, ci);
521 cluster_set_count_flag(ci + idx, 0, 0);
524 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
526 struct swap_cluster_info *ci = si->cluster_info + idx;
528 VM_BUG_ON(cluster_count(ci) != 0);
530 * If the swap is discardable, prepare discard the cluster
531 * instead of free it immediately. The cluster will be freed
534 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
535 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
536 swap_cluster_schedule_discard(si, idx);
540 __free_cluster(si, idx);
544 * The cluster corresponding to page_nr will be used. The cluster will be
545 * removed from free cluster list and its usage counter will be increased.
547 static void inc_cluster_info_page(struct swap_info_struct *p,
548 struct swap_cluster_info *cluster_info, unsigned long page_nr)
550 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
554 if (cluster_is_free(&cluster_info[idx]))
555 alloc_cluster(p, idx);
557 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
558 cluster_set_count(&cluster_info[idx],
559 cluster_count(&cluster_info[idx]) + 1);
563 * The cluster corresponding to page_nr decreases one usage. If the usage
564 * counter becomes 0, which means no page in the cluster is in using, we can
565 * optionally discard the cluster and add it to free cluster list.
567 static void dec_cluster_info_page(struct swap_info_struct *p,
568 struct swap_cluster_info *cluster_info, unsigned long page_nr)
570 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
575 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
576 cluster_set_count(&cluster_info[idx],
577 cluster_count(&cluster_info[idx]) - 1);
579 if (cluster_count(&cluster_info[idx]) == 0)
580 free_cluster(p, idx);
584 * It's possible scan_swap_map() uses a free cluster in the middle of free
585 * cluster list. Avoiding such abuse to avoid list corruption.
588 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
589 unsigned long offset)
591 struct percpu_cluster *percpu_cluster;
594 offset /= SWAPFILE_CLUSTER;
595 conflict = !cluster_list_empty(&si->free_clusters) &&
596 offset != cluster_list_first(&si->free_clusters) &&
597 cluster_is_free(&si->cluster_info[offset]);
602 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
603 cluster_set_null(&percpu_cluster->index);
608 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
609 * might involve allocating a new cluster for current CPU too.
611 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
612 unsigned long *offset, unsigned long *scan_base)
614 struct percpu_cluster *cluster;
615 struct swap_cluster_info *ci;
616 unsigned long tmp, max;
619 cluster = this_cpu_ptr(si->percpu_cluster);
620 if (cluster_is_null(&cluster->index)) {
621 if (!cluster_list_empty(&si->free_clusters)) {
622 cluster->index = si->free_clusters.head;
623 cluster->next = cluster_next(&cluster->index) *
625 } else if (!cluster_list_empty(&si->discard_clusters)) {
627 * we don't have free cluster but have some clusters in
628 * discarding, do discard now and reclaim them, then
629 * reread cluster_next_cpu since we dropped si->lock
631 swap_do_scheduled_discard(si);
632 *scan_base = this_cpu_read(*si->cluster_next_cpu);
633 *offset = *scan_base;
640 * Other CPUs can use our cluster if they can't find a free cluster,
641 * check if there is still free entry in the cluster
644 max = min_t(unsigned long, si->max,
645 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
647 ci = lock_cluster(si, tmp);
649 if (!si->swap_map[tmp])
656 cluster_set_null(&cluster->index);
659 cluster->next = tmp + 1;
665 static void __del_from_avail_list(struct swap_info_struct *p)
669 assert_spin_locked(&p->lock);
671 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
674 static void del_from_avail_list(struct swap_info_struct *p)
676 spin_lock(&swap_avail_lock);
677 __del_from_avail_list(p);
678 spin_unlock(&swap_avail_lock);
681 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
682 unsigned int nr_entries)
684 unsigned int end = offset + nr_entries - 1;
686 if (offset == si->lowest_bit)
687 si->lowest_bit += nr_entries;
688 if (end == si->highest_bit)
689 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
690 si->inuse_pages += nr_entries;
691 if (si->inuse_pages == si->pages) {
692 si->lowest_bit = si->max;
694 del_from_avail_list(si);
698 static void add_to_avail_list(struct swap_info_struct *p)
702 spin_lock(&swap_avail_lock);
704 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
705 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
707 spin_unlock(&swap_avail_lock);
710 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
711 unsigned int nr_entries)
713 unsigned long begin = offset;
714 unsigned long end = offset + nr_entries - 1;
715 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
717 if (offset < si->lowest_bit)
718 si->lowest_bit = offset;
719 if (end > si->highest_bit) {
720 bool was_full = !si->highest_bit;
722 WRITE_ONCE(si->highest_bit, end);
723 if (was_full && (si->flags & SWP_WRITEOK))
724 add_to_avail_list(si);
726 atomic_long_add(nr_entries, &nr_swap_pages);
727 si->inuse_pages -= nr_entries;
728 if (si->flags & SWP_BLKDEV)
729 swap_slot_free_notify =
730 si->bdev->bd_disk->fops->swap_slot_free_notify;
732 swap_slot_free_notify = NULL;
733 while (offset <= end) {
734 arch_swap_invalidate_page(si->type, offset);
735 frontswap_invalidate_page(si->type, offset);
736 if (swap_slot_free_notify)
737 swap_slot_free_notify(si->bdev, offset);
740 clear_shadow_from_swap_cache(si->type, begin, end);
743 static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
747 if (!(si->flags & SWP_SOLIDSTATE)) {
748 si->cluster_next = next;
752 prev = this_cpu_read(*si->cluster_next_cpu);
754 * Cross the swap address space size aligned trunk, choose
755 * another trunk randomly to avoid lock contention on swap
756 * address space if possible.
758 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
759 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
760 /* No free swap slots available */
761 if (si->highest_bit <= si->lowest_bit)
763 next = si->lowest_bit +
764 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
765 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
766 next = max_t(unsigned int, next, si->lowest_bit);
768 this_cpu_write(*si->cluster_next_cpu, next);
771 static int scan_swap_map_slots(struct swap_info_struct *si,
772 unsigned char usage, int nr,
775 struct swap_cluster_info *ci;
776 unsigned long offset;
777 unsigned long scan_base;
778 unsigned long last_in_cluster = 0;
779 int latency_ration = LATENCY_LIMIT;
781 bool scanned_many = false;
784 * We try to cluster swap pages by allocating them sequentially
785 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
786 * way, however, we resort to first-free allocation, starting
787 * a new cluster. This prevents us from scattering swap pages
788 * all over the entire swap partition, so that we reduce
789 * overall disk seek times between swap pages. -- sct
790 * But we do now try to find an empty cluster. -Andrea
791 * And we let swap pages go all over an SSD partition. Hugh
794 si->flags += SWP_SCANNING;
796 * Use percpu scan base for SSD to reduce lock contention on
797 * cluster and swap cache. For HDD, sequential access is more
800 if (si->flags & SWP_SOLIDSTATE)
801 scan_base = this_cpu_read(*si->cluster_next_cpu);
803 scan_base = si->cluster_next;
807 if (si->cluster_info) {
808 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
810 } else if (unlikely(!si->cluster_nr--)) {
811 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
812 si->cluster_nr = SWAPFILE_CLUSTER - 1;
816 spin_unlock(&si->lock);
819 * If seek is expensive, start searching for new cluster from
820 * start of partition, to minimize the span of allocated swap.
821 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
822 * case, just handled by scan_swap_map_try_ssd_cluster() above.
824 scan_base = offset = si->lowest_bit;
825 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
827 /* Locate the first empty (unaligned) cluster */
828 for (; last_in_cluster <= si->highest_bit; offset++) {
829 if (si->swap_map[offset])
830 last_in_cluster = offset + SWAPFILE_CLUSTER;
831 else if (offset == last_in_cluster) {
832 spin_lock(&si->lock);
833 offset -= SWAPFILE_CLUSTER - 1;
834 si->cluster_next = offset;
835 si->cluster_nr = SWAPFILE_CLUSTER - 1;
838 if (unlikely(--latency_ration < 0)) {
840 latency_ration = LATENCY_LIMIT;
845 spin_lock(&si->lock);
846 si->cluster_nr = SWAPFILE_CLUSTER - 1;
850 if (si->cluster_info) {
851 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
852 /* take a break if we already got some slots */
855 if (!scan_swap_map_try_ssd_cluster(si, &offset,
860 if (!(si->flags & SWP_WRITEOK))
862 if (!si->highest_bit)
864 if (offset > si->highest_bit)
865 scan_base = offset = si->lowest_bit;
867 ci = lock_cluster(si, offset);
868 /* reuse swap entry of cache-only swap if not busy. */
869 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
872 spin_unlock(&si->lock);
873 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
874 spin_lock(&si->lock);
875 /* entry was freed successfully, try to use this again */
878 goto scan; /* check next one */
881 if (si->swap_map[offset]) {
888 WRITE_ONCE(si->swap_map[offset], usage);
889 inc_cluster_info_page(si, si->cluster_info, offset);
892 swap_range_alloc(si, offset, 1);
893 slots[n_ret++] = swp_entry(si->type, offset);
895 /* got enough slots or reach max slots? */
896 if ((n_ret == nr) || (offset >= si->highest_bit))
899 /* search for next available slot */
901 /* time to take a break? */
902 if (unlikely(--latency_ration < 0)) {
905 spin_unlock(&si->lock);
907 spin_lock(&si->lock);
908 latency_ration = LATENCY_LIMIT;
911 /* try to get more slots in cluster */
912 if (si->cluster_info) {
913 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
915 } else if (si->cluster_nr && !si->swap_map[++offset]) {
916 /* non-ssd case, still more slots in cluster? */
922 * Even if there's no free clusters available (fragmented),
923 * try to scan a little more quickly with lock held unless we
924 * have scanned too many slots already.
927 unsigned long scan_limit;
929 if (offset < scan_base)
930 scan_limit = scan_base;
932 scan_limit = si->highest_bit;
933 for (; offset <= scan_limit && --latency_ration > 0;
935 if (!si->swap_map[offset])
941 set_cluster_next(si, offset + 1);
942 si->flags -= SWP_SCANNING;
946 spin_unlock(&si->lock);
947 while (++offset <= READ_ONCE(si->highest_bit)) {
948 if (data_race(!si->swap_map[offset])) {
949 spin_lock(&si->lock);
952 if (vm_swap_full() &&
953 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
954 spin_lock(&si->lock);
957 if (unlikely(--latency_ration < 0)) {
959 latency_ration = LATENCY_LIMIT;
963 offset = si->lowest_bit;
964 while (offset < scan_base) {
965 if (data_race(!si->swap_map[offset])) {
966 spin_lock(&si->lock);
969 if (vm_swap_full() &&
970 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
971 spin_lock(&si->lock);
974 if (unlikely(--latency_ration < 0)) {
976 latency_ration = LATENCY_LIMIT;
981 spin_lock(&si->lock);
984 si->flags -= SWP_SCANNING;
988 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
991 struct swap_cluster_info *ci;
992 unsigned long offset, i;
996 * Should not even be attempting cluster allocations when huge
997 * page swap is disabled. Warn and fail the allocation.
999 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1004 if (cluster_list_empty(&si->free_clusters))
1007 idx = cluster_list_first(&si->free_clusters);
1008 offset = idx * SWAPFILE_CLUSTER;
1009 ci = lock_cluster(si, offset);
1010 alloc_cluster(si, idx);
1011 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1013 map = si->swap_map + offset;
1014 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1015 map[i] = SWAP_HAS_CACHE;
1017 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1018 *slot = swp_entry(si->type, offset);
1023 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1025 unsigned long offset = idx * SWAPFILE_CLUSTER;
1026 struct swap_cluster_info *ci;
1028 ci = lock_cluster(si, offset);
1029 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1030 cluster_set_count_flag(ci, 0, 0);
1031 free_cluster(si, idx);
1033 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1036 static unsigned long scan_swap_map(struct swap_info_struct *si,
1037 unsigned char usage)
1042 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
1045 return swp_offset(entry);
1051 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1053 unsigned long size = swap_entry_size(entry_size);
1054 struct swap_info_struct *si, *next;
1059 /* Only single cluster request supported */
1060 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1062 spin_lock(&swap_avail_lock);
1064 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1065 if (avail_pgs <= 0) {
1066 spin_unlock(&swap_avail_lock);
1070 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1072 atomic_long_sub(n_goal * size, &nr_swap_pages);
1075 node = numa_node_id();
1076 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1077 /* requeue si to after same-priority siblings */
1078 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1079 spin_unlock(&swap_avail_lock);
1080 spin_lock(&si->lock);
1081 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1082 spin_lock(&swap_avail_lock);
1083 if (plist_node_empty(&si->avail_lists[node])) {
1084 spin_unlock(&si->lock);
1087 WARN(!si->highest_bit,
1088 "swap_info %d in list but !highest_bit\n",
1090 WARN(!(si->flags & SWP_WRITEOK),
1091 "swap_info %d in list but !SWP_WRITEOK\n",
1093 __del_from_avail_list(si);
1094 spin_unlock(&si->lock);
1097 if (size == SWAPFILE_CLUSTER) {
1098 if (si->flags & SWP_BLKDEV)
1099 n_ret = swap_alloc_cluster(si, swp_entries);
1101 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1102 n_goal, swp_entries);
1103 spin_unlock(&si->lock);
1104 if (n_ret || size == SWAPFILE_CLUSTER)
1106 pr_debug("scan_swap_map of si %d failed to find offset\n",
1110 spin_lock(&swap_avail_lock);
1113 * if we got here, it's likely that si was almost full before,
1114 * and since scan_swap_map() can drop the si->lock, multiple
1115 * callers probably all tried to get a page from the same si
1116 * and it filled up before we could get one; or, the si filled
1117 * up between us dropping swap_avail_lock and taking si->lock.
1118 * Since we dropped the swap_avail_lock, the swap_avail_head
1119 * list may have been modified; so if next is still in the
1120 * swap_avail_head list then try it, otherwise start over
1121 * if we have not gotten any slots.
1123 if (plist_node_empty(&next->avail_lists[node]))
1127 spin_unlock(&swap_avail_lock);
1131 atomic_long_add((long)(n_goal - n_ret) * size,
1137 /* The only caller of this function is now suspend routine */
1138 swp_entry_t get_swap_page_of_type(int type)
1140 struct swap_info_struct *si = swap_type_to_swap_info(type);
1146 spin_lock(&si->lock);
1147 if (si->flags & SWP_WRITEOK) {
1148 /* This is called for allocating swap entry, not cache */
1149 offset = scan_swap_map(si, 1);
1151 atomic_long_dec(&nr_swap_pages);
1152 spin_unlock(&si->lock);
1153 return swp_entry(type, offset);
1156 spin_unlock(&si->lock);
1158 return (swp_entry_t) {0};
1161 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1163 struct swap_info_struct *p;
1164 unsigned long offset;
1168 p = swp_swap_info(entry);
1171 if (data_race(!(p->flags & SWP_USED)))
1173 offset = swp_offset(entry);
1174 if (offset >= p->max)
1179 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1182 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1185 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1190 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1192 struct swap_info_struct *p;
1194 p = __swap_info_get(entry);
1197 if (data_race(!p->swap_map[swp_offset(entry)]))
1202 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1207 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1209 struct swap_info_struct *p;
1211 p = _swap_info_get(entry);
1213 spin_lock(&p->lock);
1217 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1218 struct swap_info_struct *q)
1220 struct swap_info_struct *p;
1222 p = _swap_info_get(entry);
1226 spin_unlock(&q->lock);
1228 spin_lock(&p->lock);
1233 static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1234 unsigned long offset,
1235 unsigned char usage)
1237 unsigned char count;
1238 unsigned char has_cache;
1240 count = p->swap_map[offset];
1242 has_cache = count & SWAP_HAS_CACHE;
1243 count &= ~SWAP_HAS_CACHE;
1245 if (usage == SWAP_HAS_CACHE) {
1246 VM_BUG_ON(!has_cache);
1248 } else if (count == SWAP_MAP_SHMEM) {
1250 * Or we could insist on shmem.c using a special
1251 * swap_shmem_free() and free_shmem_swap_and_cache()...
1254 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1255 if (count == COUNT_CONTINUED) {
1256 if (swap_count_continued(p, offset, count))
1257 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1259 count = SWAP_MAP_MAX;
1264 usage = count | has_cache;
1266 WRITE_ONCE(p->swap_map[offset], usage);
1268 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1274 * Note that when only holding the PTL, swapoff might succeed immediately
1275 * after freeing a swap entry. Therefore, immediately after
1276 * __swap_entry_free(), the swap info might become stale and should not
1277 * be touched without a prior get_swap_device().
1279 * Check whether swap entry is valid in the swap device. If so,
1280 * return pointer to swap_info_struct, and keep the swap entry valid
1281 * via preventing the swap device from being swapoff, until
1282 * put_swap_device() is called. Otherwise return NULL.
1284 * The entirety of the RCU read critical section must come before the
1285 * return from or after the call to synchronize_rcu() in
1286 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1287 * true, the si->map, si->cluster_info, etc. must be valid in the
1290 * Notice that swapoff or swapoff+swapon can still happen before the
1291 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1292 * in put_swap_device() if there isn't any other way to prevent
1293 * swapoff, such as page lock, page table lock, etc. The caller must
1294 * be prepared for that. For example, the following situation is
1299 * ... swapoff+swapon
1300 * __read_swap_cache_async()
1301 * swapcache_prepare()
1302 * __swap_duplicate()
1304 * // verify PTE not changed
1306 * In __swap_duplicate(), the swap_map need to be checked before
1307 * changing partly because the specified swap entry may be for another
1308 * swap device which has been swapoff. And in do_swap_page(), after
1309 * the page is read from the swap device, the PTE is verified not
1310 * changed with the page table locked to check whether the swap device
1311 * has been swapoff or swapoff+swapon.
1313 struct swap_info_struct *get_swap_device(swp_entry_t entry)
1315 struct swap_info_struct *si;
1316 unsigned long offset;
1320 si = swp_swap_info(entry);
1325 if (data_race(!(si->flags & SWP_VALID)))
1327 offset = swp_offset(entry);
1328 if (offset >= si->max)
1333 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1341 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1344 struct swap_cluster_info *ci;
1345 unsigned long offset = swp_offset(entry);
1346 unsigned char usage;
1348 ci = lock_cluster_or_swap_info(p, offset);
1349 usage = __swap_entry_free_locked(p, offset, 1);
1350 unlock_cluster_or_swap_info(p, ci);
1352 free_swap_slot(entry);
1357 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1359 struct swap_cluster_info *ci;
1360 unsigned long offset = swp_offset(entry);
1361 unsigned char count;
1363 ci = lock_cluster(p, offset);
1364 count = p->swap_map[offset];
1365 VM_BUG_ON(count != SWAP_HAS_CACHE);
1366 p->swap_map[offset] = 0;
1367 dec_cluster_info_page(p, p->cluster_info, offset);
1370 mem_cgroup_uncharge_swap(entry, 1);
1371 swap_range_free(p, offset, 1);
1375 * Caller has made sure that the swap device corresponding to entry
1376 * is still around or has not been recycled.
1378 void swap_free(swp_entry_t entry)
1380 struct swap_info_struct *p;
1382 p = _swap_info_get(entry);
1384 __swap_entry_free(p, entry);
1388 * Called after dropping swapcache to decrease refcnt to swap entries.
1390 void put_swap_page(struct page *page, swp_entry_t entry)
1392 unsigned long offset = swp_offset(entry);
1393 unsigned long idx = offset / SWAPFILE_CLUSTER;
1394 struct swap_cluster_info *ci;
1395 struct swap_info_struct *si;
1397 unsigned int i, free_entries = 0;
1399 int size = swap_entry_size(thp_nr_pages(page));
1401 si = _swap_info_get(entry);
1405 ci = lock_cluster_or_swap_info(si, offset);
1406 if (size == SWAPFILE_CLUSTER) {
1407 VM_BUG_ON(!cluster_is_huge(ci));
1408 map = si->swap_map + offset;
1409 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1411 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1412 if (val == SWAP_HAS_CACHE)
1415 cluster_clear_huge(ci);
1416 if (free_entries == SWAPFILE_CLUSTER) {
1417 unlock_cluster_or_swap_info(si, ci);
1418 spin_lock(&si->lock);
1419 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1420 swap_free_cluster(si, idx);
1421 spin_unlock(&si->lock);
1425 for (i = 0; i < size; i++, entry.val++) {
1426 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1427 unlock_cluster_or_swap_info(si, ci);
1428 free_swap_slot(entry);
1431 lock_cluster_or_swap_info(si, offset);
1434 unlock_cluster_or_swap_info(si, ci);
1437 #ifdef CONFIG_THP_SWAP
1438 int split_swap_cluster(swp_entry_t entry)
1440 struct swap_info_struct *si;
1441 struct swap_cluster_info *ci;
1442 unsigned long offset = swp_offset(entry);
1444 si = _swap_info_get(entry);
1447 ci = lock_cluster(si, offset);
1448 cluster_clear_huge(ci);
1454 static int swp_entry_cmp(const void *ent1, const void *ent2)
1456 const swp_entry_t *e1 = ent1, *e2 = ent2;
1458 return (int)swp_type(*e1) - (int)swp_type(*e2);
1461 void swapcache_free_entries(swp_entry_t *entries, int n)
1463 struct swap_info_struct *p, *prev;
1473 * Sort swap entries by swap device, so each lock is only taken once.
1474 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1475 * so low that it isn't necessary to optimize further.
1477 if (nr_swapfiles > 1)
1478 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1479 for (i = 0; i < n; ++i) {
1480 p = swap_info_get_cont(entries[i], prev);
1482 swap_entry_free(p, entries[i]);
1486 spin_unlock(&p->lock);
1490 * How many references to page are currently swapped out?
1491 * This does not give an exact answer when swap count is continued,
1492 * but does include the high COUNT_CONTINUED flag to allow for that.
1494 int page_swapcount(struct page *page)
1497 struct swap_info_struct *p;
1498 struct swap_cluster_info *ci;
1500 unsigned long offset;
1502 entry.val = page_private(page);
1503 p = _swap_info_get(entry);
1505 offset = swp_offset(entry);
1506 ci = lock_cluster_or_swap_info(p, offset);
1507 count = swap_count(p->swap_map[offset]);
1508 unlock_cluster_or_swap_info(p, ci);
1513 int __swap_count(swp_entry_t entry)
1515 struct swap_info_struct *si;
1516 pgoff_t offset = swp_offset(entry);
1519 si = get_swap_device(entry);
1521 count = swap_count(si->swap_map[offset]);
1522 put_swap_device(si);
1527 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1530 pgoff_t offset = swp_offset(entry);
1531 struct swap_cluster_info *ci;
1533 ci = lock_cluster_or_swap_info(si, offset);
1534 count = swap_count(si->swap_map[offset]);
1535 unlock_cluster_or_swap_info(si, ci);
1540 * How many references to @entry are currently swapped out?
1541 * This does not give an exact answer when swap count is continued,
1542 * but does include the high COUNT_CONTINUED flag to allow for that.
1544 int __swp_swapcount(swp_entry_t entry)
1547 struct swap_info_struct *si;
1549 si = get_swap_device(entry);
1551 count = swap_swapcount(si, entry);
1552 put_swap_device(si);
1558 * How many references to @entry are currently swapped out?
1559 * This considers COUNT_CONTINUED so it returns exact answer.
1561 int swp_swapcount(swp_entry_t entry)
1563 int count, tmp_count, n;
1564 struct swap_info_struct *p;
1565 struct swap_cluster_info *ci;
1570 p = _swap_info_get(entry);
1574 offset = swp_offset(entry);
1576 ci = lock_cluster_or_swap_info(p, offset);
1578 count = swap_count(p->swap_map[offset]);
1579 if (!(count & COUNT_CONTINUED))
1582 count &= ~COUNT_CONTINUED;
1583 n = SWAP_MAP_MAX + 1;
1585 page = vmalloc_to_page(p->swap_map + offset);
1586 offset &= ~PAGE_MASK;
1587 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1590 page = list_next_entry(page, lru);
1591 map = kmap_atomic(page);
1592 tmp_count = map[offset];
1595 count += (tmp_count & ~COUNT_CONTINUED) * n;
1596 n *= (SWAP_CONT_MAX + 1);
1597 } while (tmp_count & COUNT_CONTINUED);
1599 unlock_cluster_or_swap_info(p, ci);
1603 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1606 struct swap_cluster_info *ci;
1607 unsigned char *map = si->swap_map;
1608 unsigned long roffset = swp_offset(entry);
1609 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1613 ci = lock_cluster_or_swap_info(si, offset);
1614 if (!ci || !cluster_is_huge(ci)) {
1615 if (swap_count(map[roffset]))
1619 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1620 if (swap_count(map[offset + i])) {
1626 unlock_cluster_or_swap_info(si, ci);
1630 static bool page_swapped(struct page *page)
1633 struct swap_info_struct *si;
1635 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1636 return page_swapcount(page) != 0;
1638 page = compound_head(page);
1639 entry.val = page_private(page);
1640 si = _swap_info_get(entry);
1642 return swap_page_trans_huge_swapped(si, entry);
1646 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1647 int *total_swapcount)
1649 int i, map_swapcount, _total_mapcount, _total_swapcount;
1650 unsigned long offset = 0;
1651 struct swap_info_struct *si;
1652 struct swap_cluster_info *ci = NULL;
1653 unsigned char *map = NULL;
1654 int mapcount, swapcount = 0;
1656 /* hugetlbfs shouldn't call it */
1657 VM_BUG_ON_PAGE(PageHuge(page), page);
1659 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1660 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1661 if (PageSwapCache(page))
1662 swapcount = page_swapcount(page);
1663 if (total_swapcount)
1664 *total_swapcount = swapcount;
1665 return mapcount + swapcount;
1668 page = compound_head(page);
1670 _total_mapcount = _total_swapcount = map_swapcount = 0;
1671 if (PageSwapCache(page)) {
1674 entry.val = page_private(page);
1675 si = _swap_info_get(entry);
1678 offset = swp_offset(entry);
1682 ci = lock_cluster(si, offset);
1683 for (i = 0; i < HPAGE_PMD_NR; i++) {
1684 mapcount = atomic_read(&page[i]._mapcount) + 1;
1685 _total_mapcount += mapcount;
1687 swapcount = swap_count(map[offset + i]);
1688 _total_swapcount += swapcount;
1690 map_swapcount = max(map_swapcount, mapcount + swapcount);
1693 if (PageDoubleMap(page)) {
1695 _total_mapcount -= HPAGE_PMD_NR;
1697 mapcount = compound_mapcount(page);
1698 map_swapcount += mapcount;
1699 _total_mapcount += mapcount;
1701 *total_mapcount = _total_mapcount;
1702 if (total_swapcount)
1703 *total_swapcount = _total_swapcount;
1705 return map_swapcount;
1709 * We can write to an anon page without COW if there are no other references
1710 * to it. And as a side-effect, free up its swap: because the old content
1711 * on disk will never be read, and seeking back there to write new content
1712 * later would only waste time away from clustering.
1714 * NOTE: total_map_swapcount should not be relied upon by the caller if
1715 * reuse_swap_page() returns false, but it may be always overwritten
1716 * (see the other implementation for CONFIG_SWAP=n).
1718 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1720 int count, total_mapcount, total_swapcount;
1722 VM_BUG_ON_PAGE(!PageLocked(page), page);
1723 if (unlikely(PageKsm(page)))
1725 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1727 if (total_map_swapcount)
1728 *total_map_swapcount = total_mapcount + total_swapcount;
1729 if (count == 1 && PageSwapCache(page) &&
1730 (likely(!PageTransCompound(page)) ||
1731 /* The remaining swap count will be freed soon */
1732 total_swapcount == page_swapcount(page))) {
1733 if (!PageWriteback(page)) {
1734 page = compound_head(page);
1735 delete_from_swap_cache(page);
1739 struct swap_info_struct *p;
1741 entry.val = page_private(page);
1742 p = swap_info_get(entry);
1743 if (p->flags & SWP_STABLE_WRITES) {
1744 spin_unlock(&p->lock);
1747 spin_unlock(&p->lock);
1755 * If swap is getting full, or if there are no more mappings of this page,
1756 * then try_to_free_swap is called to free its swap space.
1758 int try_to_free_swap(struct page *page)
1760 VM_BUG_ON_PAGE(!PageLocked(page), page);
1762 if (!PageSwapCache(page))
1764 if (PageWriteback(page))
1766 if (page_swapped(page))
1770 * Once hibernation has begun to create its image of memory,
1771 * there's a danger that one of the calls to try_to_free_swap()
1772 * - most probably a call from __try_to_reclaim_swap() while
1773 * hibernation is allocating its own swap pages for the image,
1774 * but conceivably even a call from memory reclaim - will free
1775 * the swap from a page which has already been recorded in the
1776 * image as a clean swapcache page, and then reuse its swap for
1777 * another page of the image. On waking from hibernation, the
1778 * original page might be freed under memory pressure, then
1779 * later read back in from swap, now with the wrong data.
1781 * Hibernation suspends storage while it is writing the image
1782 * to disk so check that here.
1784 if (pm_suspended_storage())
1787 page = compound_head(page);
1788 delete_from_swap_cache(page);
1794 * Free the swap entry like above, but also try to
1795 * free the page cache entry if it is the last user.
1797 int free_swap_and_cache(swp_entry_t entry)
1799 struct swap_info_struct *p;
1800 unsigned char count;
1802 if (non_swap_entry(entry))
1805 p = get_swap_device(entry);
1807 if (WARN_ON(data_race(!p->swap_map[swp_offset(entry)]))) {
1812 count = __swap_entry_free(p, entry);
1813 if (count == SWAP_HAS_CACHE &&
1814 !swap_page_trans_huge_swapped(p, entry))
1815 __try_to_reclaim_swap(p, swp_offset(entry),
1816 TTRS_UNMAPPED | TTRS_FULL);
1822 #ifdef CONFIG_HIBERNATION
1824 * Find the swap type that corresponds to given device (if any).
1826 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1827 * from 0, in which the swap header is expected to be located.
1829 * This is needed for the suspend to disk (aka swsusp).
1831 int swap_type_of(dev_t device, sector_t offset)
1838 spin_lock(&swap_lock);
1839 for (type = 0; type < nr_swapfiles; type++) {
1840 struct swap_info_struct *sis = swap_info[type];
1842 if (!(sis->flags & SWP_WRITEOK))
1845 if (device == sis->bdev->bd_dev) {
1846 struct swap_extent *se = first_se(sis);
1848 if (se->start_block == offset) {
1849 spin_unlock(&swap_lock);
1854 spin_unlock(&swap_lock);
1858 int find_first_swap(dev_t *device)
1862 spin_lock(&swap_lock);
1863 for (type = 0; type < nr_swapfiles; type++) {
1864 struct swap_info_struct *sis = swap_info[type];
1866 if (!(sis->flags & SWP_WRITEOK))
1868 *device = sis->bdev->bd_dev;
1869 spin_unlock(&swap_lock);
1872 spin_unlock(&swap_lock);
1877 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1878 * corresponding to given index in swap_info (swap type).
1880 sector_t swapdev_block(int type, pgoff_t offset)
1882 struct block_device *bdev;
1883 struct swap_info_struct *si = swap_type_to_swap_info(type);
1885 if (!si || !(si->flags & SWP_WRITEOK))
1887 return map_swap_entry(swp_entry(type, offset), &bdev);
1891 * Return either the total number of swap pages of given type, or the number
1892 * of free pages of that type (depending on @free)
1894 * This is needed for software suspend
1896 unsigned int count_swap_pages(int type, int free)
1900 spin_lock(&swap_lock);
1901 if ((unsigned int)type < nr_swapfiles) {
1902 struct swap_info_struct *sis = swap_info[type];
1904 spin_lock(&sis->lock);
1905 if (sis->flags & SWP_WRITEOK) {
1908 n -= sis->inuse_pages;
1910 spin_unlock(&sis->lock);
1912 spin_unlock(&swap_lock);
1915 #endif /* CONFIG_HIBERNATION */
1917 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1919 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1923 * No need to decide whether this PTE shares the swap entry with others,
1924 * just let do_wp_page work it out if a write is requested later - to
1925 * force COW, vm_page_prot omits write permission from any private vma.
1927 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1928 unsigned long addr, swp_entry_t entry, struct page *page)
1930 struct page *swapcache;
1936 page = ksm_might_need_to_copy(page, vma, addr);
1937 if (unlikely(!page))
1940 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1941 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1946 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1947 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1949 set_pte_at(vma->vm_mm, addr, pte,
1950 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1951 if (page == swapcache) {
1952 page_add_anon_rmap(page, vma, addr, false);
1953 } else { /* ksm created a completely new copy */
1954 page_add_new_anon_rmap(page, vma, addr, false);
1955 lru_cache_add_inactive_or_unevictable(page, vma);
1959 pte_unmap_unlock(pte, ptl);
1960 if (page != swapcache) {
1967 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1968 unsigned long addr, unsigned long end,
1969 unsigned int type, bool frontswap,
1970 unsigned long *fs_pages_to_unuse)
1975 struct swap_info_struct *si;
1976 unsigned long offset;
1978 volatile unsigned char *swap_map;
1980 si = swap_info[type];
1981 pte = pte_offset_map(pmd, addr);
1983 struct vm_fault vmf;
1985 if (!is_swap_pte(*pte))
1988 entry = pte_to_swp_entry(*pte);
1989 if (swp_type(entry) != type)
1992 offset = swp_offset(entry);
1993 if (frontswap && !frontswap_test(si, offset))
1997 swap_map = &si->swap_map[offset];
1998 page = lookup_swap_cache(entry, vma, addr);
2003 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2007 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
2013 wait_on_page_writeback(page);
2014 ret = unuse_pte(vma, pmd, addr, entry, page);
2021 try_to_free_swap(page);
2025 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
2026 ret = FRONTSWAP_PAGES_UNUSED;
2030 pte = pte_offset_map(pmd, addr);
2031 } while (pte++, addr += PAGE_SIZE, addr != end);
2039 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2040 unsigned long addr, unsigned long end,
2041 unsigned int type, bool frontswap,
2042 unsigned long *fs_pages_to_unuse)
2048 pmd = pmd_offset(pud, addr);
2051 next = pmd_addr_end(addr, end);
2052 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2054 ret = unuse_pte_range(vma, pmd, addr, next, type,
2055 frontswap, fs_pages_to_unuse);
2058 } while (pmd++, addr = next, addr != end);
2062 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2063 unsigned long addr, unsigned long end,
2064 unsigned int type, bool frontswap,
2065 unsigned long *fs_pages_to_unuse)
2071 pud = pud_offset(p4d, addr);
2073 next = pud_addr_end(addr, end);
2074 if (pud_none_or_clear_bad(pud))
2076 ret = unuse_pmd_range(vma, pud, addr, next, type,
2077 frontswap, fs_pages_to_unuse);
2080 } while (pud++, addr = next, addr != end);
2084 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2085 unsigned long addr, unsigned long end,
2086 unsigned int type, bool frontswap,
2087 unsigned long *fs_pages_to_unuse)
2093 p4d = p4d_offset(pgd, addr);
2095 next = p4d_addr_end(addr, end);
2096 if (p4d_none_or_clear_bad(p4d))
2098 ret = unuse_pud_range(vma, p4d, addr, next, type,
2099 frontswap, fs_pages_to_unuse);
2102 } while (p4d++, addr = next, addr != end);
2106 static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2107 bool frontswap, unsigned long *fs_pages_to_unuse)
2110 unsigned long addr, end, next;
2113 addr = vma->vm_start;
2116 pgd = pgd_offset(vma->vm_mm, addr);
2118 next = pgd_addr_end(addr, end);
2119 if (pgd_none_or_clear_bad(pgd))
2121 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2122 frontswap, fs_pages_to_unuse);
2125 } while (pgd++, addr = next, addr != end);
2129 static int unuse_mm(struct mm_struct *mm, unsigned int type,
2130 bool frontswap, unsigned long *fs_pages_to_unuse)
2132 struct vm_area_struct *vma;
2136 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2137 if (vma->anon_vma) {
2138 ret = unuse_vma(vma, type, frontswap,
2145 mmap_read_unlock(mm);
2150 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2151 * from current position to next entry still in use. Return 0
2152 * if there are no inuse entries after prev till end of the map.
2154 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2155 unsigned int prev, bool frontswap)
2158 unsigned char count;
2161 * No need for swap_lock here: we're just looking
2162 * for whether an entry is in use, not modifying it; false
2163 * hits are okay, and sys_swapoff() has already prevented new
2164 * allocations from this area (while holding swap_lock).
2166 for (i = prev + 1; i < si->max; i++) {
2167 count = READ_ONCE(si->swap_map[i]);
2168 if (count && swap_count(count) != SWAP_MAP_BAD)
2169 if (!frontswap || frontswap_test(si, i))
2171 if ((i % LATENCY_LIMIT) == 0)
2182 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2183 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2185 int try_to_unuse(unsigned int type, bool frontswap,
2186 unsigned long pages_to_unuse)
2188 struct mm_struct *prev_mm;
2189 struct mm_struct *mm;
2190 struct list_head *p;
2192 struct swap_info_struct *si = swap_info[type];
2197 if (!READ_ONCE(si->inuse_pages))
2204 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2211 spin_lock(&mmlist_lock);
2212 p = &init_mm.mmlist;
2213 while (READ_ONCE(si->inuse_pages) &&
2214 !signal_pending(current) &&
2215 (p = p->next) != &init_mm.mmlist) {
2217 mm = list_entry(p, struct mm_struct, mmlist);
2218 if (!mmget_not_zero(mm))
2220 spin_unlock(&mmlist_lock);
2223 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2231 * Make sure that we aren't completely killing
2232 * interactive performance.
2235 spin_lock(&mmlist_lock);
2237 spin_unlock(&mmlist_lock);
2242 while (READ_ONCE(si->inuse_pages) &&
2243 !signal_pending(current) &&
2244 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2246 entry = swp_entry(type, i);
2247 page = find_get_page(swap_address_space(entry), i);
2252 * It is conceivable that a racing task removed this page from
2253 * swap cache just before we acquired the page lock. The page
2254 * might even be back in swap cache on another swap area. But
2255 * that is okay, try_to_free_swap() only removes stale pages.
2258 wait_on_page_writeback(page);
2259 try_to_free_swap(page);
2264 * For frontswap, we just need to unuse pages_to_unuse, if
2265 * it was specified. Need not check frontswap again here as
2266 * we already zeroed out pages_to_unuse if not frontswap.
2268 if (pages_to_unuse && --pages_to_unuse == 0)
2273 * Lets check again to see if there are still swap entries in the map.
2274 * If yes, we would need to do retry the unuse logic again.
2275 * Under global memory pressure, swap entries can be reinserted back
2276 * into process space after the mmlist loop above passes over them.
2278 * Limit the number of retries? No: when mmget_not_zero() above fails,
2279 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2280 * at its own independent pace; and even shmem_writepage() could have
2281 * been preempted after get_swap_page(), temporarily hiding that swap.
2282 * It's easy and robust (though cpu-intensive) just to keep retrying.
2284 if (READ_ONCE(si->inuse_pages)) {
2285 if (!signal_pending(current))
2290 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2294 * After a successful try_to_unuse, if no swap is now in use, we know
2295 * we can empty the mmlist. swap_lock must be held on entry and exit.
2296 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2297 * added to the mmlist just after page_duplicate - before would be racy.
2299 static void drain_mmlist(void)
2301 struct list_head *p, *next;
2304 for (type = 0; type < nr_swapfiles; type++)
2305 if (swap_info[type]->inuse_pages)
2307 spin_lock(&mmlist_lock);
2308 list_for_each_safe(p, next, &init_mm.mmlist)
2310 spin_unlock(&mmlist_lock);
2314 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2315 * corresponds to page offset for the specified swap entry.
2316 * Note that the type of this function is sector_t, but it returns page offset
2317 * into the bdev, not sector offset.
2319 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2321 struct swap_info_struct *sis;
2322 struct swap_extent *se;
2325 sis = swp_swap_info(entry);
2328 offset = swp_offset(entry);
2329 se = offset_to_swap_extent(sis, offset);
2330 return se->start_block + (offset - se->start_page);
2334 * Returns the page offset into bdev for the specified page's swap entry.
2336 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2339 entry.val = page_private(page);
2340 return map_swap_entry(entry, bdev);
2344 * Free all of a swapdev's extent information
2346 static void destroy_swap_extents(struct swap_info_struct *sis)
2348 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2349 struct rb_node *rb = sis->swap_extent_root.rb_node;
2350 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2352 rb_erase(rb, &sis->swap_extent_root);
2356 if (sis->flags & SWP_ACTIVATED) {
2357 struct file *swap_file = sis->swap_file;
2358 struct address_space *mapping = swap_file->f_mapping;
2360 sis->flags &= ~SWP_ACTIVATED;
2361 if (mapping->a_ops->swap_deactivate)
2362 mapping->a_ops->swap_deactivate(swap_file);
2367 * Add a block range (and the corresponding page range) into this swapdev's
2370 * This function rather assumes that it is called in ascending page order.
2373 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2374 unsigned long nr_pages, sector_t start_block)
2376 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2377 struct swap_extent *se;
2378 struct swap_extent *new_se;
2381 * place the new node at the right most since the
2382 * function is called in ascending page order.
2386 link = &parent->rb_right;
2390 se = rb_entry(parent, struct swap_extent, rb_node);
2391 BUG_ON(se->start_page + se->nr_pages != start_page);
2392 if (se->start_block + se->nr_pages == start_block) {
2394 se->nr_pages += nr_pages;
2399 /* No merge, insert a new extent. */
2400 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2403 new_se->start_page = start_page;
2404 new_se->nr_pages = nr_pages;
2405 new_se->start_block = start_block;
2407 rb_link_node(&new_se->rb_node, parent, link);
2408 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2411 EXPORT_SYMBOL_GPL(add_swap_extent);
2414 * A `swap extent' is a simple thing which maps a contiguous range of pages
2415 * onto a contiguous range of disk blocks. An ordered list of swap extents
2416 * is built at swapon time and is then used at swap_writepage/swap_readpage
2417 * time for locating where on disk a page belongs.
2419 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2420 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2421 * swap files identically.
2423 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2424 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2425 * swapfiles are handled *identically* after swapon time.
2427 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2428 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2429 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2430 * requirements, they are simply tossed out - we will never use those blocks
2433 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2434 * prevents users from writing to the swap device, which will corrupt memory.
2436 * The amount of disk space which a single swap extent represents varies.
2437 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2438 * extents in the list. To avoid much list walking, we cache the previous
2439 * search location in `curr_swap_extent', and start new searches from there.
2440 * This is extremely effective. The average number of iterations in
2441 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2443 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2445 struct file *swap_file = sis->swap_file;
2446 struct address_space *mapping = swap_file->f_mapping;
2447 struct inode *inode = mapping->host;
2450 if (S_ISBLK(inode->i_mode)) {
2451 ret = add_swap_extent(sis, 0, sis->max, 0);
2456 if (mapping->a_ops->swap_activate) {
2457 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2459 sis->flags |= SWP_ACTIVATED;
2461 sis->flags |= SWP_FS_OPS;
2462 ret = add_swap_extent(sis, 0, sis->max, 0);
2468 return generic_swapfile_activate(sis, swap_file, span);
2471 static int swap_node(struct swap_info_struct *p)
2473 struct block_device *bdev;
2478 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2480 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2483 static void setup_swap_info(struct swap_info_struct *p, int prio,
2484 unsigned char *swap_map,
2485 struct swap_cluster_info *cluster_info)
2492 p->prio = --least_priority;
2494 * the plist prio is negated because plist ordering is
2495 * low-to-high, while swap ordering is high-to-low
2497 p->list.prio = -p->prio;
2500 p->avail_lists[i].prio = -p->prio;
2502 if (swap_node(p) == i)
2503 p->avail_lists[i].prio = 1;
2505 p->avail_lists[i].prio = -p->prio;
2508 p->swap_map = swap_map;
2509 p->cluster_info = cluster_info;
2512 static void _enable_swap_info(struct swap_info_struct *p)
2514 p->flags |= SWP_WRITEOK | SWP_VALID;
2515 atomic_long_add(p->pages, &nr_swap_pages);
2516 total_swap_pages += p->pages;
2518 assert_spin_locked(&swap_lock);
2520 * both lists are plists, and thus priority ordered.
2521 * swap_active_head needs to be priority ordered for swapoff(),
2522 * which on removal of any swap_info_struct with an auto-assigned
2523 * (i.e. negative) priority increments the auto-assigned priority
2524 * of any lower-priority swap_info_structs.
2525 * swap_avail_head needs to be priority ordered for get_swap_page(),
2526 * which allocates swap pages from the highest available priority
2529 plist_add(&p->list, &swap_active_head);
2530 add_to_avail_list(p);
2533 static void enable_swap_info(struct swap_info_struct *p, int prio,
2534 unsigned char *swap_map,
2535 struct swap_cluster_info *cluster_info,
2536 unsigned long *frontswap_map)
2538 frontswap_init(p->type, frontswap_map);
2539 spin_lock(&swap_lock);
2540 spin_lock(&p->lock);
2541 setup_swap_info(p, prio, swap_map, cluster_info);
2542 spin_unlock(&p->lock);
2543 spin_unlock(&swap_lock);
2545 * Guarantee swap_map, cluster_info, etc. fields are valid
2546 * between get/put_swap_device() if SWP_VALID bit is set
2549 spin_lock(&swap_lock);
2550 spin_lock(&p->lock);
2551 _enable_swap_info(p);
2552 spin_unlock(&p->lock);
2553 spin_unlock(&swap_lock);
2556 static void reinsert_swap_info(struct swap_info_struct *p)
2558 spin_lock(&swap_lock);
2559 spin_lock(&p->lock);
2560 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2561 _enable_swap_info(p);
2562 spin_unlock(&p->lock);
2563 spin_unlock(&swap_lock);
2566 bool has_usable_swap(void)
2570 spin_lock(&swap_lock);
2571 if (plist_head_empty(&swap_active_head))
2573 spin_unlock(&swap_lock);
2577 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2579 struct swap_info_struct *p = NULL;
2580 unsigned char *swap_map;
2581 struct swap_cluster_info *cluster_info;
2582 unsigned long *frontswap_map;
2583 struct file *swap_file, *victim;
2584 struct address_space *mapping;
2585 struct inode *inode;
2586 struct filename *pathname;
2588 unsigned int old_block_size;
2590 if (!capable(CAP_SYS_ADMIN))
2593 BUG_ON(!current->mm);
2595 pathname = getname(specialfile);
2596 if (IS_ERR(pathname))
2597 return PTR_ERR(pathname);
2599 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2600 err = PTR_ERR(victim);
2604 mapping = victim->f_mapping;
2605 spin_lock(&swap_lock);
2606 plist_for_each_entry(p, &swap_active_head, list) {
2607 if (p->flags & SWP_WRITEOK) {
2608 if (p->swap_file->f_mapping == mapping) {
2616 spin_unlock(&swap_lock);
2619 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2620 vm_unacct_memory(p->pages);
2623 spin_unlock(&swap_lock);
2626 spin_lock(&p->lock);
2627 del_from_avail_list(p);
2629 struct swap_info_struct *si = p;
2632 plist_for_each_entry_continue(si, &swap_active_head, list) {
2635 for_each_node(nid) {
2636 if (si->avail_lists[nid].prio != 1)
2637 si->avail_lists[nid].prio--;
2642 plist_del(&p->list, &swap_active_head);
2643 atomic_long_sub(p->pages, &nr_swap_pages);
2644 total_swap_pages -= p->pages;
2645 p->flags &= ~SWP_WRITEOK;
2646 spin_unlock(&p->lock);
2647 spin_unlock(&swap_lock);
2649 disable_swap_slots_cache_lock();
2651 set_current_oom_origin();
2652 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2653 clear_current_oom_origin();
2656 /* re-insert swap space back into swap_list */
2657 reinsert_swap_info(p);
2658 reenable_swap_slots_cache_unlock();
2662 reenable_swap_slots_cache_unlock();
2664 spin_lock(&swap_lock);
2665 spin_lock(&p->lock);
2666 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2667 spin_unlock(&p->lock);
2668 spin_unlock(&swap_lock);
2670 * wait for swap operations protected by get/put_swap_device()
2675 flush_work(&p->discard_work);
2677 destroy_swap_extents(p);
2678 if (p->flags & SWP_CONTINUED)
2679 free_swap_count_continuations(p);
2681 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2682 atomic_dec(&nr_rotate_swap);
2684 mutex_lock(&swapon_mutex);
2685 spin_lock(&swap_lock);
2686 spin_lock(&p->lock);
2689 /* wait for anyone still in scan_swap_map */
2690 p->highest_bit = 0; /* cuts scans short */
2691 while (p->flags >= SWP_SCANNING) {
2692 spin_unlock(&p->lock);
2693 spin_unlock(&swap_lock);
2694 schedule_timeout_uninterruptible(1);
2695 spin_lock(&swap_lock);
2696 spin_lock(&p->lock);
2699 swap_file = p->swap_file;
2700 old_block_size = p->old_block_size;
2701 p->swap_file = NULL;
2703 swap_map = p->swap_map;
2705 cluster_info = p->cluster_info;
2706 p->cluster_info = NULL;
2707 frontswap_map = frontswap_map_get(p);
2708 spin_unlock(&p->lock);
2709 spin_unlock(&swap_lock);
2710 arch_swap_invalidate_area(p->type);
2711 frontswap_invalidate_area(p->type);
2712 frontswap_map_set(p, NULL);
2713 mutex_unlock(&swapon_mutex);
2714 free_percpu(p->percpu_cluster);
2715 p->percpu_cluster = NULL;
2716 free_percpu(p->cluster_next_cpu);
2717 p->cluster_next_cpu = NULL;
2719 kvfree(cluster_info);
2720 kvfree(frontswap_map);
2721 /* Destroy swap account information */
2722 swap_cgroup_swapoff(p->type);
2723 exit_swap_address_space(p->type);
2725 inode = mapping->host;
2726 if (S_ISBLK(inode->i_mode)) {
2727 struct block_device *bdev = I_BDEV(inode);
2729 set_blocksize(bdev, old_block_size);
2730 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2734 inode->i_flags &= ~S_SWAPFILE;
2735 inode_unlock(inode);
2736 filp_close(swap_file, NULL);
2739 * Clear the SWP_USED flag after all resources are freed so that swapon
2740 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2741 * not hold p->lock after we cleared its SWP_WRITEOK.
2743 spin_lock(&swap_lock);
2745 spin_unlock(&swap_lock);
2748 atomic_inc(&proc_poll_event);
2749 wake_up_interruptible(&proc_poll_wait);
2752 filp_close(victim, NULL);
2758 #ifdef CONFIG_PROC_FS
2759 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2761 struct seq_file *seq = file->private_data;
2763 poll_wait(file, &proc_poll_wait, wait);
2765 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2766 seq->poll_event = atomic_read(&proc_poll_event);
2767 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2770 return EPOLLIN | EPOLLRDNORM;
2774 static void *swap_start(struct seq_file *swap, loff_t *pos)
2776 struct swap_info_struct *si;
2780 mutex_lock(&swapon_mutex);
2783 return SEQ_START_TOKEN;
2785 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2786 if (!(si->flags & SWP_USED) || !si->swap_map)
2795 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2797 struct swap_info_struct *si = v;
2800 if (v == SEQ_START_TOKEN)
2803 type = si->type + 1;
2806 for (; (si = swap_type_to_swap_info(type)); type++) {
2807 if (!(si->flags & SWP_USED) || !si->swap_map)
2815 static void swap_stop(struct seq_file *swap, void *v)
2817 mutex_unlock(&swapon_mutex);
2820 static int swap_show(struct seq_file *swap, void *v)
2822 struct swap_info_struct *si = v;
2825 unsigned int bytes, inuse;
2827 if (si == SEQ_START_TOKEN) {
2828 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2832 bytes = si->pages << (PAGE_SHIFT - 10);
2833 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2835 file = si->swap_file;
2836 len = seq_file_path(swap, file, " \t\n\\");
2837 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2838 len < 40 ? 40 - len : 1, " ",
2839 S_ISBLK(file_inode(file)->i_mode) ?
2840 "partition" : "file\t",
2841 bytes, bytes < 10000000 ? "\t" : "",
2842 inuse, inuse < 10000000 ? "\t" : "",
2847 static const struct seq_operations swaps_op = {
2848 .start = swap_start,
2854 static int swaps_open(struct inode *inode, struct file *file)
2856 struct seq_file *seq;
2859 ret = seq_open(file, &swaps_op);
2863 seq = file->private_data;
2864 seq->poll_event = atomic_read(&proc_poll_event);
2868 static const struct proc_ops swaps_proc_ops = {
2869 .proc_flags = PROC_ENTRY_PERMANENT,
2870 .proc_open = swaps_open,
2871 .proc_read = seq_read,
2872 .proc_lseek = seq_lseek,
2873 .proc_release = seq_release,
2874 .proc_poll = swaps_poll,
2877 static int __init procswaps_init(void)
2879 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2882 __initcall(procswaps_init);
2883 #endif /* CONFIG_PROC_FS */
2885 #ifdef MAX_SWAPFILES_CHECK
2886 static int __init max_swapfiles_check(void)
2888 MAX_SWAPFILES_CHECK();
2891 late_initcall(max_swapfiles_check);
2894 static struct swap_info_struct *alloc_swap_info(void)
2896 struct swap_info_struct *p;
2897 struct swap_info_struct *defer = NULL;
2901 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2903 return ERR_PTR(-ENOMEM);
2905 spin_lock(&swap_lock);
2906 for (type = 0; type < nr_swapfiles; type++) {
2907 if (!(swap_info[type]->flags & SWP_USED))
2910 if (type >= MAX_SWAPFILES) {
2911 spin_unlock(&swap_lock);
2913 return ERR_PTR(-EPERM);
2915 if (type >= nr_swapfiles) {
2917 WRITE_ONCE(swap_info[type], p);
2919 * Write swap_info[type] before nr_swapfiles, in case a
2920 * racing procfs swap_start() or swap_next() is reading them.
2921 * (We never shrink nr_swapfiles, we never free this entry.)
2924 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2927 p = swap_info[type];
2929 * Do not memset this entry: a racing procfs swap_next()
2930 * would be relying on p->type to remain valid.
2933 p->swap_extent_root = RB_ROOT;
2934 plist_node_init(&p->list, 0);
2936 plist_node_init(&p->avail_lists[i], 0);
2937 p->flags = SWP_USED;
2938 spin_unlock(&swap_lock);
2940 spin_lock_init(&p->lock);
2941 spin_lock_init(&p->cont_lock);
2946 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2950 if (S_ISBLK(inode->i_mode)) {
2951 p->bdev = blkdev_get_by_dev(inode->i_rdev,
2952 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2953 if (IS_ERR(p->bdev)) {
2954 error = PTR_ERR(p->bdev);
2958 p->old_block_size = block_size(p->bdev);
2959 error = set_blocksize(p->bdev, PAGE_SIZE);
2963 * Zoned block devices contain zones that have a sequential
2964 * write only restriction. Hence zoned block devices are not
2965 * suitable for swapping. Disallow them here.
2967 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2969 p->flags |= SWP_BLKDEV;
2970 } else if (S_ISREG(inode->i_mode)) {
2971 p->bdev = inode->i_sb->s_bdev;
2979 * Find out how many pages are allowed for a single swap device. There
2980 * are two limiting factors:
2981 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2982 * 2) the number of bits in the swap pte, as defined by the different
2985 * In order to find the largest possible bit mask, a swap entry with
2986 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2987 * decoded to a swp_entry_t again, and finally the swap offset is
2990 * This will mask all the bits from the initial ~0UL mask that can't
2991 * be encoded in either the swp_entry_t or the architecture definition
2994 unsigned long generic_max_swapfile_size(void)
2996 return swp_offset(pte_to_swp_entry(
2997 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
3000 /* Can be overridden by an architecture for additional checks. */
3001 __weak unsigned long max_swapfile_size(void)
3003 return generic_max_swapfile_size();
3006 static unsigned long read_swap_header(struct swap_info_struct *p,
3007 union swap_header *swap_header,
3008 struct inode *inode)
3011 unsigned long maxpages;
3012 unsigned long swapfilepages;
3013 unsigned long last_page;
3015 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
3016 pr_err("Unable to find swap-space signature\n");
3020 /* swap partition endianess hack... */
3021 if (swab32(swap_header->info.version) == 1) {
3022 swab32s(&swap_header->info.version);
3023 swab32s(&swap_header->info.last_page);
3024 swab32s(&swap_header->info.nr_badpages);
3025 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3027 for (i = 0; i < swap_header->info.nr_badpages; i++)
3028 swab32s(&swap_header->info.badpages[i]);
3030 /* Check the swap header's sub-version */
3031 if (swap_header->info.version != 1) {
3032 pr_warn("Unable to handle swap header version %d\n",
3033 swap_header->info.version);
3038 p->cluster_next = 1;
3041 maxpages = max_swapfile_size();
3042 last_page = swap_header->info.last_page;
3044 pr_warn("Empty swap-file\n");
3047 if (last_page > maxpages) {
3048 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3049 maxpages << (PAGE_SHIFT - 10),
3050 last_page << (PAGE_SHIFT - 10));
3052 if (maxpages > last_page) {
3053 maxpages = last_page + 1;
3054 /* p->max is an unsigned int: don't overflow it */
3055 if ((unsigned int)maxpages == 0)
3056 maxpages = UINT_MAX;
3058 p->highest_bit = maxpages - 1;
3062 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3063 if (swapfilepages && maxpages > swapfilepages) {
3064 pr_warn("Swap area shorter than signature indicates\n");
3067 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3069 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3075 #define SWAP_CLUSTER_INFO_COLS \
3076 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3077 #define SWAP_CLUSTER_SPACE_COLS \
3078 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3079 #define SWAP_CLUSTER_COLS \
3080 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3082 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3083 union swap_header *swap_header,
3084 unsigned char *swap_map,
3085 struct swap_cluster_info *cluster_info,
3086 unsigned long maxpages,
3090 unsigned int nr_good_pages;
3092 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3093 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3094 unsigned long i, idx;
3096 nr_good_pages = maxpages - 1; /* omit header page */
3098 cluster_list_init(&p->free_clusters);
3099 cluster_list_init(&p->discard_clusters);
3101 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3102 unsigned int page_nr = swap_header->info.badpages[i];
3103 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3105 if (page_nr < maxpages) {
3106 swap_map[page_nr] = SWAP_MAP_BAD;
3109 * Haven't marked the cluster free yet, no list
3110 * operation involved
3112 inc_cluster_info_page(p, cluster_info, page_nr);
3116 /* Haven't marked the cluster free yet, no list operation involved */
3117 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3118 inc_cluster_info_page(p, cluster_info, i);
3120 if (nr_good_pages) {
3121 swap_map[0] = SWAP_MAP_BAD;
3123 * Not mark the cluster free yet, no list
3124 * operation involved
3126 inc_cluster_info_page(p, cluster_info, 0);
3128 p->pages = nr_good_pages;
3129 nr_extents = setup_swap_extents(p, span);
3132 nr_good_pages = p->pages;
3134 if (!nr_good_pages) {
3135 pr_warn("Empty swap-file\n");
3144 * Reduce false cache line sharing between cluster_info and
3145 * sharing same address space.
3147 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3148 j = (k + col) % SWAP_CLUSTER_COLS;
3149 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3150 idx = i * SWAP_CLUSTER_COLS + j;
3151 if (idx >= nr_clusters)
3153 if (cluster_count(&cluster_info[idx]))
3155 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3156 cluster_list_add_tail(&p->free_clusters, cluster_info,
3164 * Helper to sys_swapon determining if a given swap
3165 * backing device queue supports DISCARD operations.
3167 static bool swap_discardable(struct swap_info_struct *si)
3169 struct request_queue *q = bdev_get_queue(si->bdev);
3171 if (!q || !blk_queue_discard(q))
3177 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3179 struct swap_info_struct *p;
3180 struct filename *name;
3181 struct file *swap_file = NULL;
3182 struct address_space *mapping;
3185 union swap_header *swap_header;
3188 unsigned long maxpages;
3189 unsigned char *swap_map = NULL;
3190 struct swap_cluster_info *cluster_info = NULL;
3191 unsigned long *frontswap_map = NULL;
3192 struct page *page = NULL;
3193 struct inode *inode = NULL;
3194 bool inced_nr_rotate_swap = false;
3196 if (swap_flags & ~SWAP_FLAGS_VALID)
3199 if (!capable(CAP_SYS_ADMIN))
3202 if (!swap_avail_heads)
3205 p = alloc_swap_info();
3209 INIT_WORK(&p->discard_work, swap_discard_work);
3211 name = getname(specialfile);
3213 error = PTR_ERR(name);
3217 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3218 if (IS_ERR(swap_file)) {
3219 error = PTR_ERR(swap_file);
3224 p->swap_file = swap_file;
3225 mapping = swap_file->f_mapping;
3226 inode = mapping->host;
3228 error = claim_swapfile(p, inode);
3229 if (unlikely(error))
3233 if (IS_SWAPFILE(inode)) {
3235 goto bad_swap_unlock_inode;
3239 * Read the swap header.
3241 if (!mapping->a_ops->readpage) {
3243 goto bad_swap_unlock_inode;
3245 page = read_mapping_page(mapping, 0, swap_file);
3247 error = PTR_ERR(page);
3248 goto bad_swap_unlock_inode;
3250 swap_header = kmap(page);
3252 maxpages = read_swap_header(p, swap_header, inode);
3253 if (unlikely(!maxpages)) {
3255 goto bad_swap_unlock_inode;
3258 /* OK, set up the swap map and apply the bad block list */
3259 swap_map = vzalloc(maxpages);
3262 goto bad_swap_unlock_inode;
3265 if (p->bdev && blk_queue_stable_writes(p->bdev->bd_disk->queue))
3266 p->flags |= SWP_STABLE_WRITES;
3268 if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3269 p->flags |= SWP_SYNCHRONOUS_IO;
3271 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3273 unsigned long ci, nr_cluster;
3275 p->flags |= SWP_SOLIDSTATE;
3276 p->cluster_next_cpu = alloc_percpu(unsigned int);
3277 if (!p->cluster_next_cpu) {
3279 goto bad_swap_unlock_inode;
3282 * select a random position to start with to help wear leveling
3285 for_each_possible_cpu(cpu) {
3286 per_cpu(*p->cluster_next_cpu, cpu) =
3287 1 + prandom_u32_max(p->highest_bit);
3289 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3291 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3293 if (!cluster_info) {
3295 goto bad_swap_unlock_inode;
3298 for (ci = 0; ci < nr_cluster; ci++)
3299 spin_lock_init(&((cluster_info + ci)->lock));
3301 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3302 if (!p->percpu_cluster) {
3304 goto bad_swap_unlock_inode;
3306 for_each_possible_cpu(cpu) {
3307 struct percpu_cluster *cluster;
3308 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3309 cluster_set_null(&cluster->index);
3312 atomic_inc(&nr_rotate_swap);
3313 inced_nr_rotate_swap = true;
3316 error = swap_cgroup_swapon(p->type, maxpages);
3318 goto bad_swap_unlock_inode;
3320 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3321 cluster_info, maxpages, &span);
3322 if (unlikely(nr_extents < 0)) {
3324 goto bad_swap_unlock_inode;
3326 /* frontswap enabled? set up bit-per-page map for frontswap */
3327 if (IS_ENABLED(CONFIG_FRONTSWAP))
3328 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3332 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3334 * When discard is enabled for swap with no particular
3335 * policy flagged, we set all swap discard flags here in
3336 * order to sustain backward compatibility with older
3337 * swapon(8) releases.
3339 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3343 * By flagging sys_swapon, a sysadmin can tell us to
3344 * either do single-time area discards only, or to just
3345 * perform discards for released swap page-clusters.
3346 * Now it's time to adjust the p->flags accordingly.
3348 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3349 p->flags &= ~SWP_PAGE_DISCARD;
3350 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3351 p->flags &= ~SWP_AREA_DISCARD;
3353 /* issue a swapon-time discard if it's still required */
3354 if (p->flags & SWP_AREA_DISCARD) {
3355 int err = discard_swap(p);
3357 pr_err("swapon: discard_swap(%p): %d\n",
3362 error = init_swap_address_space(p->type, maxpages);
3364 goto bad_swap_unlock_inode;
3367 * Flush any pending IO and dirty mappings before we start using this
3370 inode->i_flags |= S_SWAPFILE;
3371 error = inode_drain_writes(inode);
3373 inode->i_flags &= ~S_SWAPFILE;
3374 goto free_swap_address_space;
3377 mutex_lock(&swapon_mutex);
3379 if (swap_flags & SWAP_FLAG_PREFER)
3381 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3382 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3384 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3385 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3386 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3387 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3388 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3389 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3390 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3391 (frontswap_map) ? "FS" : "");
3393 mutex_unlock(&swapon_mutex);
3394 atomic_inc(&proc_poll_event);
3395 wake_up_interruptible(&proc_poll_wait);
3399 free_swap_address_space:
3400 exit_swap_address_space(p->type);
3401 bad_swap_unlock_inode:
3402 inode_unlock(inode);
3404 free_percpu(p->percpu_cluster);
3405 p->percpu_cluster = NULL;
3406 free_percpu(p->cluster_next_cpu);
3407 p->cluster_next_cpu = NULL;
3408 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3409 set_blocksize(p->bdev, p->old_block_size);
3410 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3413 destroy_swap_extents(p);
3414 swap_cgroup_swapoff(p->type);
3415 spin_lock(&swap_lock);
3416 p->swap_file = NULL;
3418 spin_unlock(&swap_lock);
3420 kvfree(cluster_info);
3421 kvfree(frontswap_map);
3422 if (inced_nr_rotate_swap)
3423 atomic_dec(&nr_rotate_swap);
3425 filp_close(swap_file, NULL);
3427 if (page && !IS_ERR(page)) {
3434 inode_unlock(inode);
3436 enable_swap_slots_cache();
3440 void si_swapinfo(struct sysinfo *val)
3443 unsigned long nr_to_be_unused = 0;
3445 spin_lock(&swap_lock);
3446 for (type = 0; type < nr_swapfiles; type++) {
3447 struct swap_info_struct *si = swap_info[type];
3449 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3450 nr_to_be_unused += si->inuse_pages;
3452 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3453 val->totalswap = total_swap_pages + nr_to_be_unused;
3454 spin_unlock(&swap_lock);
3458 * Verify that a swap entry is valid and increment its swap map count.
3460 * Returns error code in following case.
3462 * - swp_entry is invalid -> EINVAL
3463 * - swp_entry is migration entry -> EINVAL
3464 * - swap-cache reference is requested but there is already one. -> EEXIST
3465 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3466 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3468 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3470 struct swap_info_struct *p;
3471 struct swap_cluster_info *ci;
3472 unsigned long offset;
3473 unsigned char count;
3474 unsigned char has_cache;
3477 p = get_swap_device(entry);
3481 offset = swp_offset(entry);
3482 ci = lock_cluster_or_swap_info(p, offset);
3484 count = p->swap_map[offset];
3487 * swapin_readahead() doesn't check if a swap entry is valid, so the
3488 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3490 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3495 has_cache = count & SWAP_HAS_CACHE;
3496 count &= ~SWAP_HAS_CACHE;
3499 if (usage == SWAP_HAS_CACHE) {
3501 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3502 if (!has_cache && count)
3503 has_cache = SWAP_HAS_CACHE;
3504 else if (has_cache) /* someone else added cache */
3506 else /* no users remaining */
3509 } else if (count || has_cache) {
3511 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3513 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3515 else if (swap_count_continued(p, offset, count))
3516 count = COUNT_CONTINUED;
3520 err = -ENOENT; /* unused swap entry */
3522 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3525 unlock_cluster_or_swap_info(p, ci);
3533 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3534 * (in which case its reference count is never incremented).
3536 void swap_shmem_alloc(swp_entry_t entry)
3538 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3542 * Increase reference count of swap entry by 1.
3543 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3544 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3545 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3546 * might occur if a page table entry has got corrupted.
3548 int swap_duplicate(swp_entry_t entry)
3552 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3553 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3558 * @entry: swap entry for which we allocate swap cache.
3560 * Called when allocating swap cache for existing swap entry,
3561 * This can return error codes. Returns 0 at success.
3562 * -EEXIST means there is a swap cache.
3563 * Note: return code is different from swap_duplicate().
3565 int swapcache_prepare(swp_entry_t entry)
3567 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3570 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3572 return swap_type_to_swap_info(swp_type(entry));
3575 struct swap_info_struct *page_swap_info(struct page *page)
3577 swp_entry_t entry = { .val = page_private(page) };
3578 return swp_swap_info(entry);
3582 * out-of-line __page_file_ methods to avoid include hell.
3584 struct address_space *__page_file_mapping(struct page *page)
3586 return page_swap_info(page)->swap_file->f_mapping;
3588 EXPORT_SYMBOL_GPL(__page_file_mapping);
3590 pgoff_t __page_file_index(struct page *page)
3592 swp_entry_t swap = { .val = page_private(page) };
3593 return swp_offset(swap);
3595 EXPORT_SYMBOL_GPL(__page_file_index);
3598 * add_swap_count_continuation - called when a swap count is duplicated
3599 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3600 * page of the original vmalloc'ed swap_map, to hold the continuation count
3601 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3602 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3604 * These continuation pages are seldom referenced: the common paths all work
3605 * on the original swap_map, only referring to a continuation page when the
3606 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3608 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3609 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3610 * can be called after dropping locks.
3612 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3614 struct swap_info_struct *si;
3615 struct swap_cluster_info *ci;
3618 struct page *list_page;
3620 unsigned char count;
3624 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3625 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3627 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3629 si = get_swap_device(entry);
3632 * An acceptable race has occurred since the failing
3633 * __swap_duplicate(): the swap device may be swapoff
3637 spin_lock(&si->lock);
3639 offset = swp_offset(entry);
3641 ci = lock_cluster(si, offset);
3643 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3645 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3647 * The higher the swap count, the more likely it is that tasks
3648 * will race to add swap count continuation: we need to avoid
3649 * over-provisioning.
3660 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3661 * no architecture is using highmem pages for kernel page tables: so it
3662 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3664 head = vmalloc_to_page(si->swap_map + offset);
3665 offset &= ~PAGE_MASK;
3667 spin_lock(&si->cont_lock);
3669 * Page allocation does not initialize the page's lru field,
3670 * but it does always reset its private field.
3672 if (!page_private(head)) {
3673 BUG_ON(count & COUNT_CONTINUED);
3674 INIT_LIST_HEAD(&head->lru);
3675 set_page_private(head, SWP_CONTINUED);
3676 si->flags |= SWP_CONTINUED;
3679 list_for_each_entry(list_page, &head->lru, lru) {
3683 * If the previous map said no continuation, but we've found
3684 * a continuation page, free our allocation and use this one.
3686 if (!(count & COUNT_CONTINUED))
3687 goto out_unlock_cont;
3689 map = kmap_atomic(list_page) + offset;
3694 * If this continuation count now has some space in it,
3695 * free our allocation and use this one.
3697 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3698 goto out_unlock_cont;
3701 list_add_tail(&page->lru, &head->lru);
3702 page = NULL; /* now it's attached, don't free it */
3704 spin_unlock(&si->cont_lock);
3707 spin_unlock(&si->lock);
3708 put_swap_device(si);
3716 * swap_count_continued - when the original swap_map count is incremented
3717 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3718 * into, carry if so, or else fail until a new continuation page is allocated;
3719 * when the original swap_map count is decremented from 0 with continuation,
3720 * borrow from the continuation and report whether it still holds more.
3721 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3724 static bool swap_count_continued(struct swap_info_struct *si,
3725 pgoff_t offset, unsigned char count)
3732 head = vmalloc_to_page(si->swap_map + offset);
3733 if (page_private(head) != SWP_CONTINUED) {
3734 BUG_ON(count & COUNT_CONTINUED);
3735 return false; /* need to add count continuation */
3738 spin_lock(&si->cont_lock);
3739 offset &= ~PAGE_MASK;
3740 page = list_next_entry(head, lru);
3741 map = kmap_atomic(page) + offset;
3743 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3744 goto init_map; /* jump over SWAP_CONT_MAX checks */
3746 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3748 * Think of how you add 1 to 999
3750 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3752 page = list_next_entry(page, lru);
3753 BUG_ON(page == head);
3754 map = kmap_atomic(page) + offset;
3756 if (*map == SWAP_CONT_MAX) {
3758 page = list_next_entry(page, lru);
3760 ret = false; /* add count continuation */
3763 map = kmap_atomic(page) + offset;
3764 init_map: *map = 0; /* we didn't zero the page */
3768 while ((page = list_prev_entry(page, lru)) != head) {
3769 map = kmap_atomic(page) + offset;
3770 *map = COUNT_CONTINUED;
3773 ret = true; /* incremented */
3775 } else { /* decrementing */
3777 * Think of how you subtract 1 from 1000
3779 BUG_ON(count != COUNT_CONTINUED);
3780 while (*map == COUNT_CONTINUED) {
3782 page = list_next_entry(page, lru);
3783 BUG_ON(page == head);
3784 map = kmap_atomic(page) + offset;
3791 while ((page = list_prev_entry(page, lru)) != head) {
3792 map = kmap_atomic(page) + offset;
3793 *map = SWAP_CONT_MAX | count;
3794 count = COUNT_CONTINUED;
3797 ret = count == COUNT_CONTINUED;
3800 spin_unlock(&si->cont_lock);
3805 * free_swap_count_continuations - swapoff free all the continuation pages
3806 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3808 static void free_swap_count_continuations(struct swap_info_struct *si)
3812 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3814 head = vmalloc_to_page(si->swap_map + offset);
3815 if (page_private(head)) {
3816 struct page *page, *next;
3818 list_for_each_entry_safe(page, next, &head->lru, lru) {
3819 list_del(&page->lru);
3826 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3827 void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3829 struct swap_info_struct *si, *next;
3830 int nid = page_to_nid(page);
3832 if (!(gfp_mask & __GFP_IO))
3835 if (!blk_cgroup_congested())
3839 * We've already scheduled a throttle, avoid taking the global swap
3842 if (current->throttle_queue)
3845 spin_lock(&swap_avail_lock);
3846 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3849 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3853 spin_unlock(&swap_avail_lock);
3857 static int __init swapfile_init(void)
3861 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3863 if (!swap_avail_heads) {
3864 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3869 plist_head_init(&swap_avail_heads[nid]);
3873 subsys_initcall(swapfile_init);