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 */
224 * swap allocation tell device that a cluster of swap can now be discarded,
225 * to allow the swap device to optimize its wear-levelling.
227 static void discard_swap_cluster(struct swap_info_struct *si,
228 pgoff_t start_page, pgoff_t nr_pages)
230 struct swap_extent *se = offset_to_swap_extent(si, start_page);
233 pgoff_t offset = start_page - se->start_page;
234 sector_t start_block = se->start_block + offset;
235 sector_t nr_blocks = se->nr_pages - offset;
237 if (nr_blocks > nr_pages)
238 nr_blocks = nr_pages;
239 start_page += nr_blocks;
240 nr_pages -= nr_blocks;
242 start_block <<= PAGE_SHIFT - 9;
243 nr_blocks <<= PAGE_SHIFT - 9;
244 if (blkdev_issue_discard(si->bdev, start_block,
245 nr_blocks, GFP_NOIO, 0))
252 #ifdef CONFIG_THP_SWAP
253 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
255 #define swap_entry_size(size) (size)
257 #define SWAPFILE_CLUSTER 256
260 * Define swap_entry_size() as constant to let compiler to optimize
261 * out some code if !CONFIG_THP_SWAP
263 #define swap_entry_size(size) 1
265 #define LATENCY_LIMIT 256
267 static inline void cluster_set_flag(struct swap_cluster_info *info,
273 static inline unsigned int cluster_count(struct swap_cluster_info *info)
278 static inline void cluster_set_count(struct swap_cluster_info *info,
284 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
285 unsigned int c, unsigned int f)
291 static inline unsigned int cluster_next(struct swap_cluster_info *info)
296 static inline void cluster_set_next(struct swap_cluster_info *info,
302 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
303 unsigned int n, unsigned int f)
309 static inline bool cluster_is_free(struct swap_cluster_info *info)
311 return info->flags & CLUSTER_FLAG_FREE;
314 static inline bool cluster_is_null(struct swap_cluster_info *info)
316 return info->flags & CLUSTER_FLAG_NEXT_NULL;
319 static inline void cluster_set_null(struct swap_cluster_info *info)
321 info->flags = CLUSTER_FLAG_NEXT_NULL;
325 static inline bool cluster_is_huge(struct swap_cluster_info *info)
327 if (IS_ENABLED(CONFIG_THP_SWAP))
328 return info->flags & CLUSTER_FLAG_HUGE;
332 static inline void cluster_clear_huge(struct swap_cluster_info *info)
334 info->flags &= ~CLUSTER_FLAG_HUGE;
337 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
338 unsigned long offset)
340 struct swap_cluster_info *ci;
342 ci = si->cluster_info;
344 ci += offset / SWAPFILE_CLUSTER;
345 spin_lock(&ci->lock);
350 static inline void unlock_cluster(struct swap_cluster_info *ci)
353 spin_unlock(&ci->lock);
357 * Determine the locking method in use for this device. Return
358 * swap_cluster_info if SSD-style cluster-based locking is in place.
360 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
361 struct swap_info_struct *si, unsigned long offset)
363 struct swap_cluster_info *ci;
365 /* Try to use fine-grained SSD-style locking if available: */
366 ci = lock_cluster(si, offset);
367 /* Otherwise, fall back to traditional, coarse locking: */
369 spin_lock(&si->lock);
374 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
375 struct swap_cluster_info *ci)
380 spin_unlock(&si->lock);
383 static inline bool cluster_list_empty(struct swap_cluster_list *list)
385 return cluster_is_null(&list->head);
388 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
390 return cluster_next(&list->head);
393 static void cluster_list_init(struct swap_cluster_list *list)
395 cluster_set_null(&list->head);
396 cluster_set_null(&list->tail);
399 static void cluster_list_add_tail(struct swap_cluster_list *list,
400 struct swap_cluster_info *ci,
403 if (cluster_list_empty(list)) {
404 cluster_set_next_flag(&list->head, idx, 0);
405 cluster_set_next_flag(&list->tail, idx, 0);
407 struct swap_cluster_info *ci_tail;
408 unsigned int tail = cluster_next(&list->tail);
411 * Nested cluster lock, but both cluster locks are
412 * only acquired when we held swap_info_struct->lock
415 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
416 cluster_set_next(ci_tail, idx);
417 spin_unlock(&ci_tail->lock);
418 cluster_set_next_flag(&list->tail, idx, 0);
422 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
423 struct swap_cluster_info *ci)
427 idx = cluster_next(&list->head);
428 if (cluster_next(&list->tail) == idx) {
429 cluster_set_null(&list->head);
430 cluster_set_null(&list->tail);
432 cluster_set_next_flag(&list->head,
433 cluster_next(&ci[idx]), 0);
438 /* Add a cluster to discard list and schedule it to do discard */
439 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
443 * If scan_swap_map() can't find a free cluster, it will check
444 * si->swap_map directly. To make sure the discarding cluster isn't
445 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
446 * will be cleared after discard
448 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
449 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
451 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
453 schedule_work(&si->discard_work);
456 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
458 struct swap_cluster_info *ci = si->cluster_info;
460 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
461 cluster_list_add_tail(&si->free_clusters, ci, idx);
465 * Doing discard actually. After a cluster discard is finished, the cluster
466 * will be added to free cluster list. caller should hold si->lock.
468 static void swap_do_scheduled_discard(struct swap_info_struct *si)
470 struct swap_cluster_info *info, *ci;
473 info = si->cluster_info;
475 while (!cluster_list_empty(&si->discard_clusters)) {
476 idx = cluster_list_del_first(&si->discard_clusters, info);
477 spin_unlock(&si->lock);
479 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
482 spin_lock(&si->lock);
483 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
484 __free_cluster(si, idx);
485 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
486 0, SWAPFILE_CLUSTER);
491 static void swap_discard_work(struct work_struct *work)
493 struct swap_info_struct *si;
495 si = container_of(work, struct swap_info_struct, discard_work);
497 spin_lock(&si->lock);
498 swap_do_scheduled_discard(si);
499 spin_unlock(&si->lock);
502 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
504 struct swap_cluster_info *ci = si->cluster_info;
506 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
507 cluster_list_del_first(&si->free_clusters, ci);
508 cluster_set_count_flag(ci + idx, 0, 0);
511 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
513 struct swap_cluster_info *ci = si->cluster_info + idx;
515 VM_BUG_ON(cluster_count(ci) != 0);
517 * If the swap is discardable, prepare discard the cluster
518 * instead of free it immediately. The cluster will be freed
521 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
522 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
523 swap_cluster_schedule_discard(si, idx);
527 __free_cluster(si, idx);
531 * The cluster corresponding to page_nr will be used. The cluster will be
532 * removed from free cluster list and its usage counter will be increased.
534 static void inc_cluster_info_page(struct swap_info_struct *p,
535 struct swap_cluster_info *cluster_info, unsigned long page_nr)
537 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
541 if (cluster_is_free(&cluster_info[idx]))
542 alloc_cluster(p, idx);
544 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
545 cluster_set_count(&cluster_info[idx],
546 cluster_count(&cluster_info[idx]) + 1);
550 * The cluster corresponding to page_nr decreases one usage. If the usage
551 * counter becomes 0, which means no page in the cluster is in using, we can
552 * optionally discard the cluster and add it to free cluster list.
554 static void dec_cluster_info_page(struct swap_info_struct *p,
555 struct swap_cluster_info *cluster_info, unsigned long page_nr)
557 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
562 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
563 cluster_set_count(&cluster_info[idx],
564 cluster_count(&cluster_info[idx]) - 1);
566 if (cluster_count(&cluster_info[idx]) == 0)
567 free_cluster(p, idx);
571 * It's possible scan_swap_map() uses a free cluster in the middle of free
572 * cluster list. Avoiding such abuse to avoid list corruption.
575 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
576 unsigned long offset)
578 struct percpu_cluster *percpu_cluster;
581 offset /= SWAPFILE_CLUSTER;
582 conflict = !cluster_list_empty(&si->free_clusters) &&
583 offset != cluster_list_first(&si->free_clusters) &&
584 cluster_is_free(&si->cluster_info[offset]);
589 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
590 cluster_set_null(&percpu_cluster->index);
595 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
596 * might involve allocating a new cluster for current CPU too.
598 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
599 unsigned long *offset, unsigned long *scan_base)
601 struct percpu_cluster *cluster;
602 struct swap_cluster_info *ci;
603 unsigned long tmp, max;
606 cluster = this_cpu_ptr(si->percpu_cluster);
607 if (cluster_is_null(&cluster->index)) {
608 if (!cluster_list_empty(&si->free_clusters)) {
609 cluster->index = si->free_clusters.head;
610 cluster->next = cluster_next(&cluster->index) *
612 } else if (!cluster_list_empty(&si->discard_clusters)) {
614 * we don't have free cluster but have some clusters in
615 * discarding, do discard now and reclaim them, then
616 * reread cluster_next_cpu since we dropped si->lock
618 swap_do_scheduled_discard(si);
619 *scan_base = this_cpu_read(*si->cluster_next_cpu);
620 *offset = *scan_base;
627 * Other CPUs can use our cluster if they can't find a free cluster,
628 * check if there is still free entry in the cluster
631 max = min_t(unsigned long, si->max,
632 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
634 ci = lock_cluster(si, tmp);
636 if (!si->swap_map[tmp])
643 cluster_set_null(&cluster->index);
646 cluster->next = tmp + 1;
652 static void __del_from_avail_list(struct swap_info_struct *p)
657 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
660 static void del_from_avail_list(struct swap_info_struct *p)
662 spin_lock(&swap_avail_lock);
663 __del_from_avail_list(p);
664 spin_unlock(&swap_avail_lock);
667 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
668 unsigned int nr_entries)
670 unsigned int end = offset + nr_entries - 1;
672 if (offset == si->lowest_bit)
673 si->lowest_bit += nr_entries;
674 if (end == si->highest_bit)
675 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
676 si->inuse_pages += nr_entries;
677 if (si->inuse_pages == si->pages) {
678 si->lowest_bit = si->max;
680 del_from_avail_list(si);
684 static void add_to_avail_list(struct swap_info_struct *p)
688 spin_lock(&swap_avail_lock);
690 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
691 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
693 spin_unlock(&swap_avail_lock);
696 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
697 unsigned int nr_entries)
699 unsigned long begin = offset;
700 unsigned long end = offset + nr_entries - 1;
701 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
703 if (offset < si->lowest_bit)
704 si->lowest_bit = offset;
705 if (end > si->highest_bit) {
706 bool was_full = !si->highest_bit;
708 WRITE_ONCE(si->highest_bit, end);
709 if (was_full && (si->flags & SWP_WRITEOK))
710 add_to_avail_list(si);
712 atomic_long_add(nr_entries, &nr_swap_pages);
713 si->inuse_pages -= nr_entries;
714 if (si->flags & SWP_BLKDEV)
715 swap_slot_free_notify =
716 si->bdev->bd_disk->fops->swap_slot_free_notify;
718 swap_slot_free_notify = NULL;
719 while (offset <= end) {
720 arch_swap_invalidate_page(si->type, offset);
721 frontswap_invalidate_page(si->type, offset);
722 if (swap_slot_free_notify)
723 swap_slot_free_notify(si->bdev, offset);
726 clear_shadow_from_swap_cache(si->type, begin, end);
729 static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
733 if (!(si->flags & SWP_SOLIDSTATE)) {
734 si->cluster_next = next;
738 prev = this_cpu_read(*si->cluster_next_cpu);
740 * Cross the swap address space size aligned trunk, choose
741 * another trunk randomly to avoid lock contention on swap
742 * address space if possible.
744 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
745 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
746 /* No free swap slots available */
747 if (si->highest_bit <= si->lowest_bit)
749 next = si->lowest_bit +
750 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
751 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
752 next = max_t(unsigned int, next, si->lowest_bit);
754 this_cpu_write(*si->cluster_next_cpu, next);
757 static int scan_swap_map_slots(struct swap_info_struct *si,
758 unsigned char usage, int nr,
761 struct swap_cluster_info *ci;
762 unsigned long offset;
763 unsigned long scan_base;
764 unsigned long last_in_cluster = 0;
765 int latency_ration = LATENCY_LIMIT;
767 bool scanned_many = false;
770 * We try to cluster swap pages by allocating them sequentially
771 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
772 * way, however, we resort to first-free allocation, starting
773 * a new cluster. This prevents us from scattering swap pages
774 * all over the entire swap partition, so that we reduce
775 * overall disk seek times between swap pages. -- sct
776 * But we do now try to find an empty cluster. -Andrea
777 * And we let swap pages go all over an SSD partition. Hugh
780 si->flags += SWP_SCANNING;
782 * Use percpu scan base for SSD to reduce lock contention on
783 * cluster and swap cache. For HDD, sequential access is more
786 if (si->flags & SWP_SOLIDSTATE)
787 scan_base = this_cpu_read(*si->cluster_next_cpu);
789 scan_base = si->cluster_next;
793 if (si->cluster_info) {
794 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
796 } else if (unlikely(!si->cluster_nr--)) {
797 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
798 si->cluster_nr = SWAPFILE_CLUSTER - 1;
802 spin_unlock(&si->lock);
805 * If seek is expensive, start searching for new cluster from
806 * start of partition, to minimize the span of allocated swap.
807 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
808 * case, just handled by scan_swap_map_try_ssd_cluster() above.
810 scan_base = offset = si->lowest_bit;
811 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
813 /* Locate the first empty (unaligned) cluster */
814 for (; last_in_cluster <= si->highest_bit; offset++) {
815 if (si->swap_map[offset])
816 last_in_cluster = offset + SWAPFILE_CLUSTER;
817 else if (offset == last_in_cluster) {
818 spin_lock(&si->lock);
819 offset -= SWAPFILE_CLUSTER - 1;
820 si->cluster_next = offset;
821 si->cluster_nr = SWAPFILE_CLUSTER - 1;
824 if (unlikely(--latency_ration < 0)) {
826 latency_ration = LATENCY_LIMIT;
831 spin_lock(&si->lock);
832 si->cluster_nr = SWAPFILE_CLUSTER - 1;
836 if (si->cluster_info) {
837 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
838 /* take a break if we already got some slots */
841 if (!scan_swap_map_try_ssd_cluster(si, &offset,
846 if (!(si->flags & SWP_WRITEOK))
848 if (!si->highest_bit)
850 if (offset > si->highest_bit)
851 scan_base = offset = si->lowest_bit;
853 ci = lock_cluster(si, offset);
854 /* reuse swap entry of cache-only swap if not busy. */
855 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
858 spin_unlock(&si->lock);
859 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
860 spin_lock(&si->lock);
861 /* entry was freed successfully, try to use this again */
864 goto scan; /* check next one */
867 if (si->swap_map[offset]) {
874 WRITE_ONCE(si->swap_map[offset], usage);
875 inc_cluster_info_page(si, si->cluster_info, offset);
878 swap_range_alloc(si, offset, 1);
879 slots[n_ret++] = swp_entry(si->type, offset);
881 /* got enough slots or reach max slots? */
882 if ((n_ret == nr) || (offset >= si->highest_bit))
885 /* search for next available slot */
887 /* time to take a break? */
888 if (unlikely(--latency_ration < 0)) {
891 spin_unlock(&si->lock);
893 spin_lock(&si->lock);
894 latency_ration = LATENCY_LIMIT;
897 /* try to get more slots in cluster */
898 if (si->cluster_info) {
899 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
901 } else if (si->cluster_nr && !si->swap_map[++offset]) {
902 /* non-ssd case, still more slots in cluster? */
908 * Even if there's no free clusters available (fragmented),
909 * try to scan a little more quickly with lock held unless we
910 * have scanned too many slots already.
913 unsigned long scan_limit;
915 if (offset < scan_base)
916 scan_limit = scan_base;
918 scan_limit = si->highest_bit;
919 for (; offset <= scan_limit && --latency_ration > 0;
921 if (!si->swap_map[offset])
927 set_cluster_next(si, offset + 1);
928 si->flags -= SWP_SCANNING;
932 spin_unlock(&si->lock);
933 while (++offset <= READ_ONCE(si->highest_bit)) {
934 if (data_race(!si->swap_map[offset])) {
935 spin_lock(&si->lock);
938 if (vm_swap_full() &&
939 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
940 spin_lock(&si->lock);
943 if (unlikely(--latency_ration < 0)) {
945 latency_ration = LATENCY_LIMIT;
949 offset = si->lowest_bit;
950 while (offset < scan_base) {
951 if (data_race(!si->swap_map[offset])) {
952 spin_lock(&si->lock);
955 if (vm_swap_full() &&
956 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
957 spin_lock(&si->lock);
960 if (unlikely(--latency_ration < 0)) {
962 latency_ration = LATENCY_LIMIT;
967 spin_lock(&si->lock);
970 si->flags -= SWP_SCANNING;
974 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
977 struct swap_cluster_info *ci;
978 unsigned long offset;
981 * Should not even be attempting cluster allocations when huge
982 * page swap is disabled. Warn and fail the allocation.
984 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
989 if (cluster_list_empty(&si->free_clusters))
992 idx = cluster_list_first(&si->free_clusters);
993 offset = idx * SWAPFILE_CLUSTER;
994 ci = lock_cluster(si, offset);
995 alloc_cluster(si, idx);
996 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
998 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1000 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1001 *slot = swp_entry(si->type, offset);
1006 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1008 unsigned long offset = idx * SWAPFILE_CLUSTER;
1009 struct swap_cluster_info *ci;
1011 ci = lock_cluster(si, offset);
1012 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1013 cluster_set_count_flag(ci, 0, 0);
1014 free_cluster(si, idx);
1016 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1019 static unsigned long scan_swap_map(struct swap_info_struct *si,
1020 unsigned char usage)
1025 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
1028 return swp_offset(entry);
1034 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1036 unsigned long size = swap_entry_size(entry_size);
1037 struct swap_info_struct *si, *next;
1042 /* Only single cluster request supported */
1043 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1045 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1049 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1051 atomic_long_sub(n_goal * size, &nr_swap_pages);
1053 spin_lock(&swap_avail_lock);
1056 node = numa_node_id();
1057 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1058 /* requeue si to after same-priority siblings */
1059 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1060 spin_unlock(&swap_avail_lock);
1061 spin_lock(&si->lock);
1062 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1063 spin_lock(&swap_avail_lock);
1064 if (plist_node_empty(&si->avail_lists[node])) {
1065 spin_unlock(&si->lock);
1068 WARN(!si->highest_bit,
1069 "swap_info %d in list but !highest_bit\n",
1071 WARN(!(si->flags & SWP_WRITEOK),
1072 "swap_info %d in list but !SWP_WRITEOK\n",
1074 __del_from_avail_list(si);
1075 spin_unlock(&si->lock);
1078 if (size == SWAPFILE_CLUSTER) {
1079 if (si->flags & SWP_BLKDEV)
1080 n_ret = swap_alloc_cluster(si, swp_entries);
1082 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1083 n_goal, swp_entries);
1084 spin_unlock(&si->lock);
1085 if (n_ret || size == SWAPFILE_CLUSTER)
1087 pr_debug("scan_swap_map of si %d failed to find offset\n",
1090 spin_lock(&swap_avail_lock);
1093 * if we got here, it's likely that si was almost full before,
1094 * and since scan_swap_map() can drop the si->lock, multiple
1095 * callers probably all tried to get a page from the same si
1096 * and it filled up before we could get one; or, the si filled
1097 * up between us dropping swap_avail_lock and taking si->lock.
1098 * Since we dropped the swap_avail_lock, the swap_avail_head
1099 * list may have been modified; so if next is still in the
1100 * swap_avail_head list then try it, otherwise start over
1101 * if we have not gotten any slots.
1103 if (plist_node_empty(&next->avail_lists[node]))
1107 spin_unlock(&swap_avail_lock);
1111 atomic_long_add((long)(n_goal - n_ret) * size,
1117 /* The only caller of this function is now suspend routine */
1118 swp_entry_t get_swap_page_of_type(int type)
1120 struct swap_info_struct *si = swap_type_to_swap_info(type);
1126 spin_lock(&si->lock);
1127 if (si->flags & SWP_WRITEOK) {
1128 atomic_long_dec(&nr_swap_pages);
1129 /* This is called for allocating swap entry, not cache */
1130 offset = scan_swap_map(si, 1);
1132 spin_unlock(&si->lock);
1133 return swp_entry(type, offset);
1135 atomic_long_inc(&nr_swap_pages);
1137 spin_unlock(&si->lock);
1139 return (swp_entry_t) {0};
1142 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1144 struct swap_info_struct *p;
1145 unsigned long offset;
1149 p = swp_swap_info(entry);
1152 if (data_race(!(p->flags & SWP_USED)))
1154 offset = swp_offset(entry);
1155 if (offset >= p->max)
1160 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1163 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1166 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1171 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1173 struct swap_info_struct *p;
1175 p = __swap_info_get(entry);
1178 if (data_race(!p->swap_map[swp_offset(entry)]))
1183 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1188 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1190 struct swap_info_struct *p;
1192 p = _swap_info_get(entry);
1194 spin_lock(&p->lock);
1198 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1199 struct swap_info_struct *q)
1201 struct swap_info_struct *p;
1203 p = _swap_info_get(entry);
1207 spin_unlock(&q->lock);
1209 spin_lock(&p->lock);
1214 static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1215 unsigned long offset,
1216 unsigned char usage)
1218 unsigned char count;
1219 unsigned char has_cache;
1221 count = p->swap_map[offset];
1223 has_cache = count & SWAP_HAS_CACHE;
1224 count &= ~SWAP_HAS_CACHE;
1226 if (usage == SWAP_HAS_CACHE) {
1227 VM_BUG_ON(!has_cache);
1229 } else if (count == SWAP_MAP_SHMEM) {
1231 * Or we could insist on shmem.c using a special
1232 * swap_shmem_free() and free_shmem_swap_and_cache()...
1235 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1236 if (count == COUNT_CONTINUED) {
1237 if (swap_count_continued(p, offset, count))
1238 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1240 count = SWAP_MAP_MAX;
1245 usage = count | has_cache;
1247 WRITE_ONCE(p->swap_map[offset], usage);
1249 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1255 * Check whether swap entry is valid in the swap device. If so,
1256 * return pointer to swap_info_struct, and keep the swap entry valid
1257 * via preventing the swap device from being swapoff, until
1258 * put_swap_device() is called. Otherwise return NULL.
1260 * The entirety of the RCU read critical section must come before the
1261 * return from or after the call to synchronize_rcu() in
1262 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1263 * true, the si->map, si->cluster_info, etc. must be valid in the
1266 * Notice that swapoff or swapoff+swapon can still happen before the
1267 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1268 * in put_swap_device() if there isn't any other way to prevent
1269 * swapoff, such as page lock, page table lock, etc. The caller must
1270 * be prepared for that. For example, the following situation is
1275 * ... swapoff+swapon
1276 * __read_swap_cache_async()
1277 * swapcache_prepare()
1278 * __swap_duplicate()
1280 * // verify PTE not changed
1282 * In __swap_duplicate(), the swap_map need to be checked before
1283 * changing partly because the specified swap entry may be for another
1284 * swap device which has been swapoff. And in do_swap_page(), after
1285 * the page is read from the swap device, the PTE is verified not
1286 * changed with the page table locked to check whether the swap device
1287 * has been swapoff or swapoff+swapon.
1289 struct swap_info_struct *get_swap_device(swp_entry_t entry)
1291 struct swap_info_struct *si;
1292 unsigned long offset;
1296 si = swp_swap_info(entry);
1301 if (data_race(!(si->flags & SWP_VALID)))
1303 offset = swp_offset(entry);
1304 if (offset >= si->max)
1309 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1317 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1320 struct swap_cluster_info *ci;
1321 unsigned long offset = swp_offset(entry);
1322 unsigned char usage;
1324 ci = lock_cluster_or_swap_info(p, offset);
1325 usage = __swap_entry_free_locked(p, offset, 1);
1326 unlock_cluster_or_swap_info(p, ci);
1328 free_swap_slot(entry);
1333 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1335 struct swap_cluster_info *ci;
1336 unsigned long offset = swp_offset(entry);
1337 unsigned char count;
1339 ci = lock_cluster(p, offset);
1340 count = p->swap_map[offset];
1341 VM_BUG_ON(count != SWAP_HAS_CACHE);
1342 p->swap_map[offset] = 0;
1343 dec_cluster_info_page(p, p->cluster_info, offset);
1346 mem_cgroup_uncharge_swap(entry, 1);
1347 swap_range_free(p, offset, 1);
1351 * Caller has made sure that the swap device corresponding to entry
1352 * is still around or has not been recycled.
1354 void swap_free(swp_entry_t entry)
1356 struct swap_info_struct *p;
1358 p = _swap_info_get(entry);
1360 __swap_entry_free(p, entry);
1364 * Called after dropping swapcache to decrease refcnt to swap entries.
1366 void put_swap_page(struct page *page, swp_entry_t entry)
1368 unsigned long offset = swp_offset(entry);
1369 unsigned long idx = offset / SWAPFILE_CLUSTER;
1370 struct swap_cluster_info *ci;
1371 struct swap_info_struct *si;
1373 unsigned int i, free_entries = 0;
1375 int size = swap_entry_size(thp_nr_pages(page));
1377 si = _swap_info_get(entry);
1381 ci = lock_cluster_or_swap_info(si, offset);
1382 if (size == SWAPFILE_CLUSTER) {
1383 VM_BUG_ON(!cluster_is_huge(ci));
1384 map = si->swap_map + offset;
1385 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1387 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1388 if (val == SWAP_HAS_CACHE)
1391 cluster_clear_huge(ci);
1392 if (free_entries == SWAPFILE_CLUSTER) {
1393 unlock_cluster_or_swap_info(si, ci);
1394 spin_lock(&si->lock);
1395 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1396 swap_free_cluster(si, idx);
1397 spin_unlock(&si->lock);
1401 for (i = 0; i < size; i++, entry.val++) {
1402 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1403 unlock_cluster_or_swap_info(si, ci);
1404 free_swap_slot(entry);
1407 lock_cluster_or_swap_info(si, offset);
1410 unlock_cluster_or_swap_info(si, ci);
1413 #ifdef CONFIG_THP_SWAP
1414 int split_swap_cluster(swp_entry_t entry)
1416 struct swap_info_struct *si;
1417 struct swap_cluster_info *ci;
1418 unsigned long offset = swp_offset(entry);
1420 si = _swap_info_get(entry);
1423 ci = lock_cluster(si, offset);
1424 cluster_clear_huge(ci);
1430 static int swp_entry_cmp(const void *ent1, const void *ent2)
1432 const swp_entry_t *e1 = ent1, *e2 = ent2;
1434 return (int)swp_type(*e1) - (int)swp_type(*e2);
1437 void swapcache_free_entries(swp_entry_t *entries, int n)
1439 struct swap_info_struct *p, *prev;
1449 * Sort swap entries by swap device, so each lock is only taken once.
1450 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1451 * so low that it isn't necessary to optimize further.
1453 if (nr_swapfiles > 1)
1454 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1455 for (i = 0; i < n; ++i) {
1456 p = swap_info_get_cont(entries[i], prev);
1458 swap_entry_free(p, entries[i]);
1462 spin_unlock(&p->lock);
1466 * How many references to page are currently swapped out?
1467 * This does not give an exact answer when swap count is continued,
1468 * but does include the high COUNT_CONTINUED flag to allow for that.
1470 int page_swapcount(struct page *page)
1473 struct swap_info_struct *p;
1474 struct swap_cluster_info *ci;
1476 unsigned long offset;
1478 entry.val = page_private(page);
1479 p = _swap_info_get(entry);
1481 offset = swp_offset(entry);
1482 ci = lock_cluster_or_swap_info(p, offset);
1483 count = swap_count(p->swap_map[offset]);
1484 unlock_cluster_or_swap_info(p, ci);
1489 int __swap_count(swp_entry_t entry)
1491 struct swap_info_struct *si;
1492 pgoff_t offset = swp_offset(entry);
1495 si = get_swap_device(entry);
1497 count = swap_count(si->swap_map[offset]);
1498 put_swap_device(si);
1503 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1506 pgoff_t offset = swp_offset(entry);
1507 struct swap_cluster_info *ci;
1509 ci = lock_cluster_or_swap_info(si, offset);
1510 count = swap_count(si->swap_map[offset]);
1511 unlock_cluster_or_swap_info(si, ci);
1516 * How many references to @entry are currently swapped out?
1517 * This does not give an exact answer when swap count is continued,
1518 * but does include the high COUNT_CONTINUED flag to allow for that.
1520 int __swp_swapcount(swp_entry_t entry)
1523 struct swap_info_struct *si;
1525 si = get_swap_device(entry);
1527 count = swap_swapcount(si, entry);
1528 put_swap_device(si);
1534 * How many references to @entry are currently swapped out?
1535 * This considers COUNT_CONTINUED so it returns exact answer.
1537 int swp_swapcount(swp_entry_t entry)
1539 int count, tmp_count, n;
1540 struct swap_info_struct *p;
1541 struct swap_cluster_info *ci;
1546 p = _swap_info_get(entry);
1550 offset = swp_offset(entry);
1552 ci = lock_cluster_or_swap_info(p, offset);
1554 count = swap_count(p->swap_map[offset]);
1555 if (!(count & COUNT_CONTINUED))
1558 count &= ~COUNT_CONTINUED;
1559 n = SWAP_MAP_MAX + 1;
1561 page = vmalloc_to_page(p->swap_map + offset);
1562 offset &= ~PAGE_MASK;
1563 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1566 page = list_next_entry(page, lru);
1567 map = kmap_atomic(page);
1568 tmp_count = map[offset];
1571 count += (tmp_count & ~COUNT_CONTINUED) * n;
1572 n *= (SWAP_CONT_MAX + 1);
1573 } while (tmp_count & COUNT_CONTINUED);
1575 unlock_cluster_or_swap_info(p, ci);
1579 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1582 struct swap_cluster_info *ci;
1583 unsigned char *map = si->swap_map;
1584 unsigned long roffset = swp_offset(entry);
1585 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1589 ci = lock_cluster_or_swap_info(si, offset);
1590 if (!ci || !cluster_is_huge(ci)) {
1591 if (swap_count(map[roffset]))
1595 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1596 if (swap_count(map[offset + i])) {
1602 unlock_cluster_or_swap_info(si, ci);
1606 static bool page_swapped(struct page *page)
1609 struct swap_info_struct *si;
1611 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1612 return page_swapcount(page) != 0;
1614 page = compound_head(page);
1615 entry.val = page_private(page);
1616 si = _swap_info_get(entry);
1618 return swap_page_trans_huge_swapped(si, entry);
1622 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1623 int *total_swapcount)
1625 int i, map_swapcount, _total_mapcount, _total_swapcount;
1626 unsigned long offset = 0;
1627 struct swap_info_struct *si;
1628 struct swap_cluster_info *ci = NULL;
1629 unsigned char *map = NULL;
1630 int mapcount, swapcount = 0;
1632 /* hugetlbfs shouldn't call it */
1633 VM_BUG_ON_PAGE(PageHuge(page), page);
1635 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1636 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1637 if (PageSwapCache(page))
1638 swapcount = page_swapcount(page);
1639 if (total_swapcount)
1640 *total_swapcount = swapcount;
1641 return mapcount + swapcount;
1644 page = compound_head(page);
1646 _total_mapcount = _total_swapcount = map_swapcount = 0;
1647 if (PageSwapCache(page)) {
1650 entry.val = page_private(page);
1651 si = _swap_info_get(entry);
1654 offset = swp_offset(entry);
1658 ci = lock_cluster(si, offset);
1659 for (i = 0; i < HPAGE_PMD_NR; i++) {
1660 mapcount = atomic_read(&page[i]._mapcount) + 1;
1661 _total_mapcount += mapcount;
1663 swapcount = swap_count(map[offset + i]);
1664 _total_swapcount += swapcount;
1666 map_swapcount = max(map_swapcount, mapcount + swapcount);
1669 if (PageDoubleMap(page)) {
1671 _total_mapcount -= HPAGE_PMD_NR;
1673 mapcount = compound_mapcount(page);
1674 map_swapcount += mapcount;
1675 _total_mapcount += mapcount;
1677 *total_mapcount = _total_mapcount;
1678 if (total_swapcount)
1679 *total_swapcount = _total_swapcount;
1681 return map_swapcount;
1685 * We can write to an anon page without COW if there are no other references
1686 * to it. And as a side-effect, free up its swap: because the old content
1687 * on disk will never be read, and seeking back there to write new content
1688 * later would only waste time away from clustering.
1690 * NOTE: total_map_swapcount should not be relied upon by the caller if
1691 * reuse_swap_page() returns false, but it may be always overwritten
1692 * (see the other implementation for CONFIG_SWAP=n).
1694 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1696 int count, total_mapcount, total_swapcount;
1698 VM_BUG_ON_PAGE(!PageLocked(page), page);
1699 if (unlikely(PageKsm(page)))
1701 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1703 if (total_map_swapcount)
1704 *total_map_swapcount = total_mapcount + total_swapcount;
1705 if (count == 1 && PageSwapCache(page) &&
1706 (likely(!PageTransCompound(page)) ||
1707 /* The remaining swap count will be freed soon */
1708 total_swapcount == page_swapcount(page))) {
1709 if (!PageWriteback(page)) {
1710 page = compound_head(page);
1711 delete_from_swap_cache(page);
1715 struct swap_info_struct *p;
1717 entry.val = page_private(page);
1718 p = swap_info_get(entry);
1719 if (p->flags & SWP_STABLE_WRITES) {
1720 spin_unlock(&p->lock);
1723 spin_unlock(&p->lock);
1731 * If swap is getting full, or if there are no more mappings of this page,
1732 * then try_to_free_swap is called to free its swap space.
1734 int try_to_free_swap(struct page *page)
1736 VM_BUG_ON_PAGE(!PageLocked(page), page);
1738 if (!PageSwapCache(page))
1740 if (PageWriteback(page))
1742 if (page_swapped(page))
1746 * Once hibernation has begun to create its image of memory,
1747 * there's a danger that one of the calls to try_to_free_swap()
1748 * - most probably a call from __try_to_reclaim_swap() while
1749 * hibernation is allocating its own swap pages for the image,
1750 * but conceivably even a call from memory reclaim - will free
1751 * the swap from a page which has already been recorded in the
1752 * image as a clean swapcache page, and then reuse its swap for
1753 * another page of the image. On waking from hibernation, the
1754 * original page might be freed under memory pressure, then
1755 * later read back in from swap, now with the wrong data.
1757 * Hibernation suspends storage while it is writing the image
1758 * to disk so check that here.
1760 if (pm_suspended_storage())
1763 page = compound_head(page);
1764 delete_from_swap_cache(page);
1770 * Free the swap entry like above, but also try to
1771 * free the page cache entry if it is the last user.
1773 int free_swap_and_cache(swp_entry_t entry)
1775 struct swap_info_struct *p;
1776 unsigned char count;
1778 if (non_swap_entry(entry))
1781 p = _swap_info_get(entry);
1783 count = __swap_entry_free(p, entry);
1784 if (count == SWAP_HAS_CACHE &&
1785 !swap_page_trans_huge_swapped(p, entry))
1786 __try_to_reclaim_swap(p, swp_offset(entry),
1787 TTRS_UNMAPPED | TTRS_FULL);
1792 #ifdef CONFIG_HIBERNATION
1794 * Find the swap type that corresponds to given device (if any).
1796 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1797 * from 0, in which the swap header is expected to be located.
1799 * This is needed for the suspend to disk (aka swsusp).
1801 int swap_type_of(dev_t device, sector_t offset)
1808 spin_lock(&swap_lock);
1809 for (type = 0; type < nr_swapfiles; type++) {
1810 struct swap_info_struct *sis = swap_info[type];
1812 if (!(sis->flags & SWP_WRITEOK))
1815 if (device == sis->bdev->bd_dev) {
1816 struct swap_extent *se = first_se(sis);
1818 if (se->start_block == offset) {
1819 spin_unlock(&swap_lock);
1824 spin_unlock(&swap_lock);
1828 int find_first_swap(dev_t *device)
1832 spin_lock(&swap_lock);
1833 for (type = 0; type < nr_swapfiles; type++) {
1834 struct swap_info_struct *sis = swap_info[type];
1836 if (!(sis->flags & SWP_WRITEOK))
1838 *device = sis->bdev->bd_dev;
1839 spin_unlock(&swap_lock);
1842 spin_unlock(&swap_lock);
1847 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1848 * corresponding to given index in swap_info (swap type).
1850 sector_t swapdev_block(int type, pgoff_t offset)
1852 struct block_device *bdev;
1853 struct swap_info_struct *si = swap_type_to_swap_info(type);
1855 if (!si || !(si->flags & SWP_WRITEOK))
1857 return map_swap_entry(swp_entry(type, offset), &bdev);
1861 * Return either the total number of swap pages of given type, or the number
1862 * of free pages of that type (depending on @free)
1864 * This is needed for software suspend
1866 unsigned int count_swap_pages(int type, int free)
1870 spin_lock(&swap_lock);
1871 if ((unsigned int)type < nr_swapfiles) {
1872 struct swap_info_struct *sis = swap_info[type];
1874 spin_lock(&sis->lock);
1875 if (sis->flags & SWP_WRITEOK) {
1878 n -= sis->inuse_pages;
1880 spin_unlock(&sis->lock);
1882 spin_unlock(&swap_lock);
1885 #endif /* CONFIG_HIBERNATION */
1887 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1889 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1893 * No need to decide whether this PTE shares the swap entry with others,
1894 * just let do_wp_page work it out if a write is requested later - to
1895 * force COW, vm_page_prot omits write permission from any private vma.
1897 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1898 unsigned long addr, swp_entry_t entry, struct page *page)
1900 struct page *swapcache;
1906 page = ksm_might_need_to_copy(page, vma, addr);
1907 if (unlikely(!page))
1910 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1911 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1916 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1917 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1919 set_pte_at(vma->vm_mm, addr, pte,
1920 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1921 if (page == swapcache) {
1922 page_add_anon_rmap(page, vma, addr, false);
1923 } else { /* ksm created a completely new copy */
1924 page_add_new_anon_rmap(page, vma, addr, false);
1925 lru_cache_add_inactive_or_unevictable(page, vma);
1929 pte_unmap_unlock(pte, ptl);
1930 if (page != swapcache) {
1937 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1938 unsigned long addr, unsigned long end,
1939 unsigned int type, bool frontswap,
1940 unsigned long *fs_pages_to_unuse)
1945 struct swap_info_struct *si;
1946 unsigned long offset;
1948 volatile unsigned char *swap_map;
1950 si = swap_info[type];
1951 pte = pte_offset_map(pmd, addr);
1953 struct vm_fault vmf;
1955 if (!is_swap_pte(*pte))
1958 entry = pte_to_swp_entry(*pte);
1959 if (swp_type(entry) != type)
1962 offset = swp_offset(entry);
1963 if (frontswap && !frontswap_test(si, offset))
1967 swap_map = &si->swap_map[offset];
1968 page = lookup_swap_cache(entry, vma, addr);
1973 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1977 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1983 wait_on_page_writeback(page);
1984 ret = unuse_pte(vma, pmd, addr, entry, page);
1991 try_to_free_swap(page);
1995 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
1996 ret = FRONTSWAP_PAGES_UNUSED;
2000 pte = pte_offset_map(pmd, addr);
2001 } while (pte++, addr += PAGE_SIZE, addr != end);
2009 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2010 unsigned long addr, unsigned long end,
2011 unsigned int type, bool frontswap,
2012 unsigned long *fs_pages_to_unuse)
2018 pmd = pmd_offset(pud, addr);
2021 next = pmd_addr_end(addr, end);
2022 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2024 ret = unuse_pte_range(vma, pmd, addr, next, type,
2025 frontswap, fs_pages_to_unuse);
2028 } while (pmd++, addr = next, addr != end);
2032 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2033 unsigned long addr, unsigned long end,
2034 unsigned int type, bool frontswap,
2035 unsigned long *fs_pages_to_unuse)
2041 pud = pud_offset(p4d, addr);
2043 next = pud_addr_end(addr, end);
2044 if (pud_none_or_clear_bad(pud))
2046 ret = unuse_pmd_range(vma, pud, addr, next, type,
2047 frontswap, fs_pages_to_unuse);
2050 } while (pud++, addr = next, addr != end);
2054 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2055 unsigned long addr, unsigned long end,
2056 unsigned int type, bool frontswap,
2057 unsigned long *fs_pages_to_unuse)
2063 p4d = p4d_offset(pgd, addr);
2065 next = p4d_addr_end(addr, end);
2066 if (p4d_none_or_clear_bad(p4d))
2068 ret = unuse_pud_range(vma, p4d, addr, next, type,
2069 frontswap, fs_pages_to_unuse);
2072 } while (p4d++, addr = next, addr != end);
2076 static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2077 bool frontswap, unsigned long *fs_pages_to_unuse)
2080 unsigned long addr, end, next;
2083 addr = vma->vm_start;
2086 pgd = pgd_offset(vma->vm_mm, addr);
2088 next = pgd_addr_end(addr, end);
2089 if (pgd_none_or_clear_bad(pgd))
2091 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2092 frontswap, fs_pages_to_unuse);
2095 } while (pgd++, addr = next, addr != end);
2099 static int unuse_mm(struct mm_struct *mm, unsigned int type,
2100 bool frontswap, unsigned long *fs_pages_to_unuse)
2102 struct vm_area_struct *vma;
2106 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2107 if (vma->anon_vma) {
2108 ret = unuse_vma(vma, type, frontswap,
2115 mmap_read_unlock(mm);
2120 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2121 * from current position to next entry still in use. Return 0
2122 * if there are no inuse entries after prev till end of the map.
2124 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2125 unsigned int prev, bool frontswap)
2128 unsigned char count;
2131 * No need for swap_lock here: we're just looking
2132 * for whether an entry is in use, not modifying it; false
2133 * hits are okay, and sys_swapoff() has already prevented new
2134 * allocations from this area (while holding swap_lock).
2136 for (i = prev + 1; i < si->max; i++) {
2137 count = READ_ONCE(si->swap_map[i]);
2138 if (count && swap_count(count) != SWAP_MAP_BAD)
2139 if (!frontswap || frontswap_test(si, i))
2141 if ((i % LATENCY_LIMIT) == 0)
2152 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2153 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2155 int try_to_unuse(unsigned int type, bool frontswap,
2156 unsigned long pages_to_unuse)
2158 struct mm_struct *prev_mm;
2159 struct mm_struct *mm;
2160 struct list_head *p;
2162 struct swap_info_struct *si = swap_info[type];
2167 if (!READ_ONCE(si->inuse_pages))
2174 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2181 spin_lock(&mmlist_lock);
2182 p = &init_mm.mmlist;
2183 while (READ_ONCE(si->inuse_pages) &&
2184 !signal_pending(current) &&
2185 (p = p->next) != &init_mm.mmlist) {
2187 mm = list_entry(p, struct mm_struct, mmlist);
2188 if (!mmget_not_zero(mm))
2190 spin_unlock(&mmlist_lock);
2193 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2201 * Make sure that we aren't completely killing
2202 * interactive performance.
2205 spin_lock(&mmlist_lock);
2207 spin_unlock(&mmlist_lock);
2212 while (READ_ONCE(si->inuse_pages) &&
2213 !signal_pending(current) &&
2214 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2216 entry = swp_entry(type, i);
2217 page = find_get_page(swap_address_space(entry), i);
2222 * It is conceivable that a racing task removed this page from
2223 * swap cache just before we acquired the page lock. The page
2224 * might even be back in swap cache on another swap area. But
2225 * that is okay, try_to_free_swap() only removes stale pages.
2228 wait_on_page_writeback(page);
2229 try_to_free_swap(page);
2234 * For frontswap, we just need to unuse pages_to_unuse, if
2235 * it was specified. Need not check frontswap again here as
2236 * we already zeroed out pages_to_unuse if not frontswap.
2238 if (pages_to_unuse && --pages_to_unuse == 0)
2243 * Lets check again to see if there are still swap entries in the map.
2244 * If yes, we would need to do retry the unuse logic again.
2245 * Under global memory pressure, swap entries can be reinserted back
2246 * into process space after the mmlist loop above passes over them.
2248 * Limit the number of retries? No: when mmget_not_zero() above fails,
2249 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2250 * at its own independent pace; and even shmem_writepage() could have
2251 * been preempted after get_swap_page(), temporarily hiding that swap.
2252 * It's easy and robust (though cpu-intensive) just to keep retrying.
2254 if (READ_ONCE(si->inuse_pages)) {
2255 if (!signal_pending(current))
2260 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2264 * After a successful try_to_unuse, if no swap is now in use, we know
2265 * we can empty the mmlist. swap_lock must be held on entry and exit.
2266 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2267 * added to the mmlist just after page_duplicate - before would be racy.
2269 static void drain_mmlist(void)
2271 struct list_head *p, *next;
2274 for (type = 0; type < nr_swapfiles; type++)
2275 if (swap_info[type]->inuse_pages)
2277 spin_lock(&mmlist_lock);
2278 list_for_each_safe(p, next, &init_mm.mmlist)
2280 spin_unlock(&mmlist_lock);
2284 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2285 * corresponds to page offset for the specified swap entry.
2286 * Note that the type of this function is sector_t, but it returns page offset
2287 * into the bdev, not sector offset.
2289 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2291 struct swap_info_struct *sis;
2292 struct swap_extent *se;
2295 sis = swp_swap_info(entry);
2298 offset = swp_offset(entry);
2299 se = offset_to_swap_extent(sis, offset);
2300 return se->start_block + (offset - se->start_page);
2304 * Returns the page offset into bdev for the specified page's swap entry.
2306 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2309 entry.val = page_private(page);
2310 return map_swap_entry(entry, bdev);
2314 * Free all of a swapdev's extent information
2316 static void destroy_swap_extents(struct swap_info_struct *sis)
2318 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2319 struct rb_node *rb = sis->swap_extent_root.rb_node;
2320 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2322 rb_erase(rb, &sis->swap_extent_root);
2326 if (sis->flags & SWP_ACTIVATED) {
2327 struct file *swap_file = sis->swap_file;
2328 struct address_space *mapping = swap_file->f_mapping;
2330 sis->flags &= ~SWP_ACTIVATED;
2331 if (mapping->a_ops->swap_deactivate)
2332 mapping->a_ops->swap_deactivate(swap_file);
2337 * Add a block range (and the corresponding page range) into this swapdev's
2340 * This function rather assumes that it is called in ascending page order.
2343 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2344 unsigned long nr_pages, sector_t start_block)
2346 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2347 struct swap_extent *se;
2348 struct swap_extent *new_se;
2351 * place the new node at the right most since the
2352 * function is called in ascending page order.
2356 link = &parent->rb_right;
2360 se = rb_entry(parent, struct swap_extent, rb_node);
2361 BUG_ON(se->start_page + se->nr_pages != start_page);
2362 if (se->start_block + se->nr_pages == start_block) {
2364 se->nr_pages += nr_pages;
2369 /* No merge, insert a new extent. */
2370 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2373 new_se->start_page = start_page;
2374 new_se->nr_pages = nr_pages;
2375 new_se->start_block = start_block;
2377 rb_link_node(&new_se->rb_node, parent, link);
2378 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2381 EXPORT_SYMBOL_GPL(add_swap_extent);
2384 * A `swap extent' is a simple thing which maps a contiguous range of pages
2385 * onto a contiguous range of disk blocks. An ordered list of swap extents
2386 * is built at swapon time and is then used at swap_writepage/swap_readpage
2387 * time for locating where on disk a page belongs.
2389 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2390 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2391 * swap files identically.
2393 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2394 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2395 * swapfiles are handled *identically* after swapon time.
2397 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2398 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2399 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2400 * requirements, they are simply tossed out - we will never use those blocks
2403 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2404 * prevents users from writing to the swap device, which will corrupt memory.
2406 * The amount of disk space which a single swap extent represents varies.
2407 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2408 * extents in the list. To avoid much list walking, we cache the previous
2409 * search location in `curr_swap_extent', and start new searches from there.
2410 * This is extremely effective. The average number of iterations in
2411 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2413 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2415 struct file *swap_file = sis->swap_file;
2416 struct address_space *mapping = swap_file->f_mapping;
2417 struct inode *inode = mapping->host;
2420 if (S_ISBLK(inode->i_mode)) {
2421 ret = add_swap_extent(sis, 0, sis->max, 0);
2426 if (mapping->a_ops->swap_activate) {
2427 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2429 sis->flags |= SWP_ACTIVATED;
2431 sis->flags |= SWP_FS_OPS;
2432 ret = add_swap_extent(sis, 0, sis->max, 0);
2438 return generic_swapfile_activate(sis, swap_file, span);
2441 static int swap_node(struct swap_info_struct *p)
2443 struct block_device *bdev;
2448 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2450 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2453 static void setup_swap_info(struct swap_info_struct *p, int prio,
2454 unsigned char *swap_map,
2455 struct swap_cluster_info *cluster_info)
2462 p->prio = --least_priority;
2464 * the plist prio is negated because plist ordering is
2465 * low-to-high, while swap ordering is high-to-low
2467 p->list.prio = -p->prio;
2470 p->avail_lists[i].prio = -p->prio;
2472 if (swap_node(p) == i)
2473 p->avail_lists[i].prio = 1;
2475 p->avail_lists[i].prio = -p->prio;
2478 p->swap_map = swap_map;
2479 p->cluster_info = cluster_info;
2482 static void _enable_swap_info(struct swap_info_struct *p)
2484 p->flags |= SWP_WRITEOK | SWP_VALID;
2485 atomic_long_add(p->pages, &nr_swap_pages);
2486 total_swap_pages += p->pages;
2488 assert_spin_locked(&swap_lock);
2490 * both lists are plists, and thus priority ordered.
2491 * swap_active_head needs to be priority ordered for swapoff(),
2492 * which on removal of any swap_info_struct with an auto-assigned
2493 * (i.e. negative) priority increments the auto-assigned priority
2494 * of any lower-priority swap_info_structs.
2495 * swap_avail_head needs to be priority ordered for get_swap_page(),
2496 * which allocates swap pages from the highest available priority
2499 plist_add(&p->list, &swap_active_head);
2500 add_to_avail_list(p);
2503 static void enable_swap_info(struct swap_info_struct *p, int prio,
2504 unsigned char *swap_map,
2505 struct swap_cluster_info *cluster_info,
2506 unsigned long *frontswap_map)
2508 frontswap_init(p->type, frontswap_map);
2509 spin_lock(&swap_lock);
2510 spin_lock(&p->lock);
2511 setup_swap_info(p, prio, swap_map, cluster_info);
2512 spin_unlock(&p->lock);
2513 spin_unlock(&swap_lock);
2515 * Guarantee swap_map, cluster_info, etc. fields are valid
2516 * between get/put_swap_device() if SWP_VALID bit is set
2519 spin_lock(&swap_lock);
2520 spin_lock(&p->lock);
2521 _enable_swap_info(p);
2522 spin_unlock(&p->lock);
2523 spin_unlock(&swap_lock);
2526 static void reinsert_swap_info(struct swap_info_struct *p)
2528 spin_lock(&swap_lock);
2529 spin_lock(&p->lock);
2530 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2531 _enable_swap_info(p);
2532 spin_unlock(&p->lock);
2533 spin_unlock(&swap_lock);
2536 bool has_usable_swap(void)
2540 spin_lock(&swap_lock);
2541 if (plist_head_empty(&swap_active_head))
2543 spin_unlock(&swap_lock);
2547 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2549 struct swap_info_struct *p = NULL;
2550 unsigned char *swap_map;
2551 struct swap_cluster_info *cluster_info;
2552 unsigned long *frontswap_map;
2553 struct file *swap_file, *victim;
2554 struct address_space *mapping;
2555 struct inode *inode;
2556 struct filename *pathname;
2558 unsigned int old_block_size;
2560 if (!capable(CAP_SYS_ADMIN))
2563 BUG_ON(!current->mm);
2565 pathname = getname(specialfile);
2566 if (IS_ERR(pathname))
2567 return PTR_ERR(pathname);
2569 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2570 err = PTR_ERR(victim);
2574 mapping = victim->f_mapping;
2575 spin_lock(&swap_lock);
2576 plist_for_each_entry(p, &swap_active_head, list) {
2577 if (p->flags & SWP_WRITEOK) {
2578 if (p->swap_file->f_mapping == mapping) {
2586 spin_unlock(&swap_lock);
2589 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2590 vm_unacct_memory(p->pages);
2593 spin_unlock(&swap_lock);
2596 del_from_avail_list(p);
2597 spin_lock(&p->lock);
2599 struct swap_info_struct *si = p;
2602 plist_for_each_entry_continue(si, &swap_active_head, list) {
2605 for_each_node(nid) {
2606 if (si->avail_lists[nid].prio != 1)
2607 si->avail_lists[nid].prio--;
2612 plist_del(&p->list, &swap_active_head);
2613 atomic_long_sub(p->pages, &nr_swap_pages);
2614 total_swap_pages -= p->pages;
2615 p->flags &= ~SWP_WRITEOK;
2616 spin_unlock(&p->lock);
2617 spin_unlock(&swap_lock);
2619 disable_swap_slots_cache_lock();
2621 set_current_oom_origin();
2622 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2623 clear_current_oom_origin();
2626 /* re-insert swap space back into swap_list */
2627 reinsert_swap_info(p);
2628 reenable_swap_slots_cache_unlock();
2632 reenable_swap_slots_cache_unlock();
2634 spin_lock(&swap_lock);
2635 spin_lock(&p->lock);
2636 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2637 spin_unlock(&p->lock);
2638 spin_unlock(&swap_lock);
2640 * wait for swap operations protected by get/put_swap_device()
2645 flush_work(&p->discard_work);
2647 destroy_swap_extents(p);
2648 if (p->flags & SWP_CONTINUED)
2649 free_swap_count_continuations(p);
2651 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2652 atomic_dec(&nr_rotate_swap);
2654 mutex_lock(&swapon_mutex);
2655 spin_lock(&swap_lock);
2656 spin_lock(&p->lock);
2659 /* wait for anyone still in scan_swap_map */
2660 p->highest_bit = 0; /* cuts scans short */
2661 while (p->flags >= SWP_SCANNING) {
2662 spin_unlock(&p->lock);
2663 spin_unlock(&swap_lock);
2664 schedule_timeout_uninterruptible(1);
2665 spin_lock(&swap_lock);
2666 spin_lock(&p->lock);
2669 swap_file = p->swap_file;
2670 old_block_size = p->old_block_size;
2671 p->swap_file = NULL;
2673 swap_map = p->swap_map;
2675 cluster_info = p->cluster_info;
2676 p->cluster_info = NULL;
2677 frontswap_map = frontswap_map_get(p);
2678 spin_unlock(&p->lock);
2679 spin_unlock(&swap_lock);
2680 arch_swap_invalidate_area(p->type);
2681 frontswap_invalidate_area(p->type);
2682 frontswap_map_set(p, NULL);
2683 mutex_unlock(&swapon_mutex);
2684 free_percpu(p->percpu_cluster);
2685 p->percpu_cluster = NULL;
2686 free_percpu(p->cluster_next_cpu);
2687 p->cluster_next_cpu = NULL;
2689 kvfree(cluster_info);
2690 kvfree(frontswap_map);
2691 /* Destroy swap account information */
2692 swap_cgroup_swapoff(p->type);
2693 exit_swap_address_space(p->type);
2695 inode = mapping->host;
2696 if (S_ISBLK(inode->i_mode)) {
2697 struct block_device *bdev = I_BDEV(inode);
2699 set_blocksize(bdev, old_block_size);
2700 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2704 inode->i_flags &= ~S_SWAPFILE;
2705 inode_unlock(inode);
2706 filp_close(swap_file, NULL);
2709 * Clear the SWP_USED flag after all resources are freed so that swapon
2710 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2711 * not hold p->lock after we cleared its SWP_WRITEOK.
2713 spin_lock(&swap_lock);
2715 spin_unlock(&swap_lock);
2718 atomic_inc(&proc_poll_event);
2719 wake_up_interruptible(&proc_poll_wait);
2722 filp_close(victim, NULL);
2728 #ifdef CONFIG_PROC_FS
2729 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2731 struct seq_file *seq = file->private_data;
2733 poll_wait(file, &proc_poll_wait, wait);
2735 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2736 seq->poll_event = atomic_read(&proc_poll_event);
2737 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2740 return EPOLLIN | EPOLLRDNORM;
2744 static void *swap_start(struct seq_file *swap, loff_t *pos)
2746 struct swap_info_struct *si;
2750 mutex_lock(&swapon_mutex);
2753 return SEQ_START_TOKEN;
2755 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2756 if (!(si->flags & SWP_USED) || !si->swap_map)
2765 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2767 struct swap_info_struct *si = v;
2770 if (v == SEQ_START_TOKEN)
2773 type = si->type + 1;
2776 for (; (si = swap_type_to_swap_info(type)); type++) {
2777 if (!(si->flags & SWP_USED) || !si->swap_map)
2785 static void swap_stop(struct seq_file *swap, void *v)
2787 mutex_unlock(&swapon_mutex);
2790 static int swap_show(struct seq_file *swap, void *v)
2792 struct swap_info_struct *si = v;
2795 unsigned int bytes, inuse;
2797 if (si == SEQ_START_TOKEN) {
2798 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2802 bytes = si->pages << (PAGE_SHIFT - 10);
2803 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2805 file = si->swap_file;
2806 len = seq_file_path(swap, file, " \t\n\\");
2807 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2808 len < 40 ? 40 - len : 1, " ",
2809 S_ISBLK(file_inode(file)->i_mode) ?
2810 "partition" : "file\t",
2811 bytes, bytes < 10000000 ? "\t" : "",
2812 inuse, inuse < 10000000 ? "\t" : "",
2817 static const struct seq_operations swaps_op = {
2818 .start = swap_start,
2824 static int swaps_open(struct inode *inode, struct file *file)
2826 struct seq_file *seq;
2829 ret = seq_open(file, &swaps_op);
2833 seq = file->private_data;
2834 seq->poll_event = atomic_read(&proc_poll_event);
2838 static const struct proc_ops swaps_proc_ops = {
2839 .proc_flags = PROC_ENTRY_PERMANENT,
2840 .proc_open = swaps_open,
2841 .proc_read = seq_read,
2842 .proc_lseek = seq_lseek,
2843 .proc_release = seq_release,
2844 .proc_poll = swaps_poll,
2847 static int __init procswaps_init(void)
2849 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2852 __initcall(procswaps_init);
2853 #endif /* CONFIG_PROC_FS */
2855 #ifdef MAX_SWAPFILES_CHECK
2856 static int __init max_swapfiles_check(void)
2858 MAX_SWAPFILES_CHECK();
2861 late_initcall(max_swapfiles_check);
2864 static struct swap_info_struct *alloc_swap_info(void)
2866 struct swap_info_struct *p;
2867 struct swap_info_struct *defer = NULL;
2871 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2873 return ERR_PTR(-ENOMEM);
2875 spin_lock(&swap_lock);
2876 for (type = 0; type < nr_swapfiles; type++) {
2877 if (!(swap_info[type]->flags & SWP_USED))
2880 if (type >= MAX_SWAPFILES) {
2881 spin_unlock(&swap_lock);
2883 return ERR_PTR(-EPERM);
2885 if (type >= nr_swapfiles) {
2887 WRITE_ONCE(swap_info[type], p);
2889 * Write swap_info[type] before nr_swapfiles, in case a
2890 * racing procfs swap_start() or swap_next() is reading them.
2891 * (We never shrink nr_swapfiles, we never free this entry.)
2894 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2897 p = swap_info[type];
2899 * Do not memset this entry: a racing procfs swap_next()
2900 * would be relying on p->type to remain valid.
2903 p->swap_extent_root = RB_ROOT;
2904 plist_node_init(&p->list, 0);
2906 plist_node_init(&p->avail_lists[i], 0);
2907 p->flags = SWP_USED;
2908 spin_unlock(&swap_lock);
2910 spin_lock_init(&p->lock);
2911 spin_lock_init(&p->cont_lock);
2916 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2920 if (S_ISBLK(inode->i_mode)) {
2921 p->bdev = blkdev_get_by_dev(inode->i_rdev,
2922 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2923 if (IS_ERR(p->bdev)) {
2924 error = PTR_ERR(p->bdev);
2928 p->old_block_size = block_size(p->bdev);
2929 error = set_blocksize(p->bdev, PAGE_SIZE);
2933 * Zoned block devices contain zones that have a sequential
2934 * write only restriction. Hence zoned block devices are not
2935 * suitable for swapping. Disallow them here.
2937 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2939 p->flags |= SWP_BLKDEV;
2940 } else if (S_ISREG(inode->i_mode)) {
2941 p->bdev = inode->i_sb->s_bdev;
2949 * Find out how many pages are allowed for a single swap device. There
2950 * are two limiting factors:
2951 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2952 * 2) the number of bits in the swap pte, as defined by the different
2955 * In order to find the largest possible bit mask, a swap entry with
2956 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2957 * decoded to a swp_entry_t again, and finally the swap offset is
2960 * This will mask all the bits from the initial ~0UL mask that can't
2961 * be encoded in either the swp_entry_t or the architecture definition
2964 unsigned long generic_max_swapfile_size(void)
2966 return swp_offset(pte_to_swp_entry(
2967 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2970 /* Can be overridden by an architecture for additional checks. */
2971 __weak unsigned long max_swapfile_size(void)
2973 return generic_max_swapfile_size();
2976 static unsigned long read_swap_header(struct swap_info_struct *p,
2977 union swap_header *swap_header,
2978 struct inode *inode)
2981 unsigned long maxpages;
2982 unsigned long swapfilepages;
2983 unsigned long last_page;
2985 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2986 pr_err("Unable to find swap-space signature\n");
2990 /* swap partition endianess hack... */
2991 if (swab32(swap_header->info.version) == 1) {
2992 swab32s(&swap_header->info.version);
2993 swab32s(&swap_header->info.last_page);
2994 swab32s(&swap_header->info.nr_badpages);
2995 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2997 for (i = 0; i < swap_header->info.nr_badpages; i++)
2998 swab32s(&swap_header->info.badpages[i]);
3000 /* Check the swap header's sub-version */
3001 if (swap_header->info.version != 1) {
3002 pr_warn("Unable to handle swap header version %d\n",
3003 swap_header->info.version);
3008 p->cluster_next = 1;
3011 maxpages = max_swapfile_size();
3012 last_page = swap_header->info.last_page;
3014 pr_warn("Empty swap-file\n");
3017 if (last_page > maxpages) {
3018 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3019 maxpages << (PAGE_SHIFT - 10),
3020 last_page << (PAGE_SHIFT - 10));
3022 if (maxpages > last_page) {
3023 maxpages = last_page + 1;
3024 /* p->max is an unsigned int: don't overflow it */
3025 if ((unsigned int)maxpages == 0)
3026 maxpages = UINT_MAX;
3028 p->highest_bit = maxpages - 1;
3032 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3033 if (swapfilepages && maxpages > swapfilepages) {
3034 pr_warn("Swap area shorter than signature indicates\n");
3037 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3039 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3045 #define SWAP_CLUSTER_INFO_COLS \
3046 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3047 #define SWAP_CLUSTER_SPACE_COLS \
3048 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3049 #define SWAP_CLUSTER_COLS \
3050 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3052 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3053 union swap_header *swap_header,
3054 unsigned char *swap_map,
3055 struct swap_cluster_info *cluster_info,
3056 unsigned long maxpages,
3060 unsigned int nr_good_pages;
3062 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3063 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3064 unsigned long i, idx;
3066 nr_good_pages = maxpages - 1; /* omit header page */
3068 cluster_list_init(&p->free_clusters);
3069 cluster_list_init(&p->discard_clusters);
3071 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3072 unsigned int page_nr = swap_header->info.badpages[i];
3073 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3075 if (page_nr < maxpages) {
3076 swap_map[page_nr] = SWAP_MAP_BAD;
3079 * Haven't marked the cluster free yet, no list
3080 * operation involved
3082 inc_cluster_info_page(p, cluster_info, page_nr);
3086 /* Haven't marked the cluster free yet, no list operation involved */
3087 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3088 inc_cluster_info_page(p, cluster_info, i);
3090 if (nr_good_pages) {
3091 swap_map[0] = SWAP_MAP_BAD;
3093 * Not mark the cluster free yet, no list
3094 * operation involved
3096 inc_cluster_info_page(p, cluster_info, 0);
3098 p->pages = nr_good_pages;
3099 nr_extents = setup_swap_extents(p, span);
3102 nr_good_pages = p->pages;
3104 if (!nr_good_pages) {
3105 pr_warn("Empty swap-file\n");
3114 * Reduce false cache line sharing between cluster_info and
3115 * sharing same address space.
3117 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3118 j = (k + col) % SWAP_CLUSTER_COLS;
3119 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3120 idx = i * SWAP_CLUSTER_COLS + j;
3121 if (idx >= nr_clusters)
3123 if (cluster_count(&cluster_info[idx]))
3125 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3126 cluster_list_add_tail(&p->free_clusters, cluster_info,
3134 * Helper to sys_swapon determining if a given swap
3135 * backing device queue supports DISCARD operations.
3137 static bool swap_discardable(struct swap_info_struct *si)
3139 struct request_queue *q = bdev_get_queue(si->bdev);
3141 if (!q || !blk_queue_discard(q))
3147 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3149 struct swap_info_struct *p;
3150 struct filename *name;
3151 struct file *swap_file = NULL;
3152 struct address_space *mapping;
3155 union swap_header *swap_header;
3158 unsigned long maxpages;
3159 unsigned char *swap_map = NULL;
3160 struct swap_cluster_info *cluster_info = NULL;
3161 unsigned long *frontswap_map = NULL;
3162 struct page *page = NULL;
3163 struct inode *inode = NULL;
3164 bool inced_nr_rotate_swap = false;
3166 if (swap_flags & ~SWAP_FLAGS_VALID)
3169 if (!capable(CAP_SYS_ADMIN))
3172 if (!swap_avail_heads)
3175 p = alloc_swap_info();
3179 INIT_WORK(&p->discard_work, swap_discard_work);
3181 name = getname(specialfile);
3183 error = PTR_ERR(name);
3187 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3188 if (IS_ERR(swap_file)) {
3189 error = PTR_ERR(swap_file);
3194 p->swap_file = swap_file;
3195 mapping = swap_file->f_mapping;
3196 inode = mapping->host;
3198 error = claim_swapfile(p, inode);
3199 if (unlikely(error))
3203 if (IS_SWAPFILE(inode)) {
3205 goto bad_swap_unlock_inode;
3209 * Read the swap header.
3211 if (!mapping->a_ops->readpage) {
3213 goto bad_swap_unlock_inode;
3215 page = read_mapping_page(mapping, 0, swap_file);
3217 error = PTR_ERR(page);
3218 goto bad_swap_unlock_inode;
3220 swap_header = kmap(page);
3222 maxpages = read_swap_header(p, swap_header, inode);
3223 if (unlikely(!maxpages)) {
3225 goto bad_swap_unlock_inode;
3228 /* OK, set up the swap map and apply the bad block list */
3229 swap_map = vzalloc(maxpages);
3232 goto bad_swap_unlock_inode;
3235 if (p->bdev && blk_queue_stable_writes(p->bdev->bd_disk->queue))
3236 p->flags |= SWP_STABLE_WRITES;
3238 if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3239 p->flags |= SWP_SYNCHRONOUS_IO;
3241 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3243 unsigned long ci, nr_cluster;
3245 p->flags |= SWP_SOLIDSTATE;
3246 p->cluster_next_cpu = alloc_percpu(unsigned int);
3247 if (!p->cluster_next_cpu) {
3249 goto bad_swap_unlock_inode;
3252 * select a random position to start with to help wear leveling
3255 for_each_possible_cpu(cpu) {
3256 per_cpu(*p->cluster_next_cpu, cpu) =
3257 1 + prandom_u32_max(p->highest_bit);
3259 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3261 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3263 if (!cluster_info) {
3265 goto bad_swap_unlock_inode;
3268 for (ci = 0; ci < nr_cluster; ci++)
3269 spin_lock_init(&((cluster_info + ci)->lock));
3271 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3272 if (!p->percpu_cluster) {
3274 goto bad_swap_unlock_inode;
3276 for_each_possible_cpu(cpu) {
3277 struct percpu_cluster *cluster;
3278 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3279 cluster_set_null(&cluster->index);
3282 atomic_inc(&nr_rotate_swap);
3283 inced_nr_rotate_swap = true;
3286 error = swap_cgroup_swapon(p->type, maxpages);
3288 goto bad_swap_unlock_inode;
3290 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3291 cluster_info, maxpages, &span);
3292 if (unlikely(nr_extents < 0)) {
3294 goto bad_swap_unlock_inode;
3296 /* frontswap enabled? set up bit-per-page map for frontswap */
3297 if (IS_ENABLED(CONFIG_FRONTSWAP))
3298 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3302 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3304 * When discard is enabled for swap with no particular
3305 * policy flagged, we set all swap discard flags here in
3306 * order to sustain backward compatibility with older
3307 * swapon(8) releases.
3309 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3313 * By flagging sys_swapon, a sysadmin can tell us to
3314 * either do single-time area discards only, or to just
3315 * perform discards for released swap page-clusters.
3316 * Now it's time to adjust the p->flags accordingly.
3318 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3319 p->flags &= ~SWP_PAGE_DISCARD;
3320 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3321 p->flags &= ~SWP_AREA_DISCARD;
3323 /* issue a swapon-time discard if it's still required */
3324 if (p->flags & SWP_AREA_DISCARD) {
3325 int err = discard_swap(p);
3327 pr_err("swapon: discard_swap(%p): %d\n",
3332 error = init_swap_address_space(p->type, maxpages);
3334 goto bad_swap_unlock_inode;
3337 * Flush any pending IO and dirty mappings before we start using this
3340 inode->i_flags |= S_SWAPFILE;
3341 error = inode_drain_writes(inode);
3343 inode->i_flags &= ~S_SWAPFILE;
3344 goto free_swap_address_space;
3347 mutex_lock(&swapon_mutex);
3349 if (swap_flags & SWAP_FLAG_PREFER)
3351 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3352 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3354 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3355 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3356 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3357 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3358 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3359 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3360 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3361 (frontswap_map) ? "FS" : "");
3363 mutex_unlock(&swapon_mutex);
3364 atomic_inc(&proc_poll_event);
3365 wake_up_interruptible(&proc_poll_wait);
3369 free_swap_address_space:
3370 exit_swap_address_space(p->type);
3371 bad_swap_unlock_inode:
3372 inode_unlock(inode);
3374 free_percpu(p->percpu_cluster);
3375 p->percpu_cluster = NULL;
3376 free_percpu(p->cluster_next_cpu);
3377 p->cluster_next_cpu = NULL;
3378 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3379 set_blocksize(p->bdev, p->old_block_size);
3380 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3383 destroy_swap_extents(p);
3384 swap_cgroup_swapoff(p->type);
3385 spin_lock(&swap_lock);
3386 p->swap_file = NULL;
3388 spin_unlock(&swap_lock);
3390 kvfree(cluster_info);
3391 kvfree(frontswap_map);
3392 if (inced_nr_rotate_swap)
3393 atomic_dec(&nr_rotate_swap);
3395 filp_close(swap_file, NULL);
3397 if (page && !IS_ERR(page)) {
3404 inode_unlock(inode);
3406 enable_swap_slots_cache();
3410 void si_swapinfo(struct sysinfo *val)
3413 unsigned long nr_to_be_unused = 0;
3415 spin_lock(&swap_lock);
3416 for (type = 0; type < nr_swapfiles; type++) {
3417 struct swap_info_struct *si = swap_info[type];
3419 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3420 nr_to_be_unused += si->inuse_pages;
3422 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3423 val->totalswap = total_swap_pages + nr_to_be_unused;
3424 spin_unlock(&swap_lock);
3428 * Verify that a swap entry is valid and increment its swap map count.
3430 * Returns error code in following case.
3432 * - swp_entry is invalid -> EINVAL
3433 * - swp_entry is migration entry -> EINVAL
3434 * - swap-cache reference is requested but there is already one. -> EEXIST
3435 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3436 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3438 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3440 struct swap_info_struct *p;
3441 struct swap_cluster_info *ci;
3442 unsigned long offset;
3443 unsigned char count;
3444 unsigned char has_cache;
3447 p = get_swap_device(entry);
3451 offset = swp_offset(entry);
3452 ci = lock_cluster_or_swap_info(p, offset);
3454 count = p->swap_map[offset];
3457 * swapin_readahead() doesn't check if a swap entry is valid, so the
3458 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3460 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3465 has_cache = count & SWAP_HAS_CACHE;
3466 count &= ~SWAP_HAS_CACHE;
3469 if (usage == SWAP_HAS_CACHE) {
3471 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3472 if (!has_cache && count)
3473 has_cache = SWAP_HAS_CACHE;
3474 else if (has_cache) /* someone else added cache */
3476 else /* no users remaining */
3479 } else if (count || has_cache) {
3481 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3483 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3485 else if (swap_count_continued(p, offset, count))
3486 count = COUNT_CONTINUED;
3490 err = -ENOENT; /* unused swap entry */
3492 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3495 unlock_cluster_or_swap_info(p, ci);
3502 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3503 * (in which case its reference count is never incremented).
3505 void swap_shmem_alloc(swp_entry_t entry)
3507 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3511 * Increase reference count of swap entry by 1.
3512 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3513 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3514 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3515 * might occur if a page table entry has got corrupted.
3517 int swap_duplicate(swp_entry_t entry)
3521 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3522 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3527 * @entry: swap entry for which we allocate swap cache.
3529 * Called when allocating swap cache for existing swap entry,
3530 * This can return error codes. Returns 0 at success.
3531 * -EEXIST means there is a swap cache.
3532 * Note: return code is different from swap_duplicate().
3534 int swapcache_prepare(swp_entry_t entry)
3536 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3539 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3541 return swap_type_to_swap_info(swp_type(entry));
3544 struct swap_info_struct *page_swap_info(struct page *page)
3546 swp_entry_t entry = { .val = page_private(page) };
3547 return swp_swap_info(entry);
3551 * out-of-line __page_file_ methods to avoid include hell.
3553 struct address_space *__page_file_mapping(struct page *page)
3555 return page_swap_info(page)->swap_file->f_mapping;
3557 EXPORT_SYMBOL_GPL(__page_file_mapping);
3559 pgoff_t __page_file_index(struct page *page)
3561 swp_entry_t swap = { .val = page_private(page) };
3562 return swp_offset(swap);
3564 EXPORT_SYMBOL_GPL(__page_file_index);
3567 * add_swap_count_continuation - called when a swap count is duplicated
3568 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3569 * page of the original vmalloc'ed swap_map, to hold the continuation count
3570 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3571 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3573 * These continuation pages are seldom referenced: the common paths all work
3574 * on the original swap_map, only referring to a continuation page when the
3575 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3577 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3578 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3579 * can be called after dropping locks.
3581 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3583 struct swap_info_struct *si;
3584 struct swap_cluster_info *ci;
3587 struct page *list_page;
3589 unsigned char count;
3593 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3594 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3596 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3598 si = get_swap_device(entry);
3601 * An acceptable race has occurred since the failing
3602 * __swap_duplicate(): the swap device may be swapoff
3606 spin_lock(&si->lock);
3608 offset = swp_offset(entry);
3610 ci = lock_cluster(si, offset);
3612 count = swap_count(si->swap_map[offset]);
3614 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3616 * The higher the swap count, the more likely it is that tasks
3617 * will race to add swap count continuation: we need to avoid
3618 * over-provisioning.
3629 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3630 * no architecture is using highmem pages for kernel page tables: so it
3631 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3633 head = vmalloc_to_page(si->swap_map + offset);
3634 offset &= ~PAGE_MASK;
3636 spin_lock(&si->cont_lock);
3638 * Page allocation does not initialize the page's lru field,
3639 * but it does always reset its private field.
3641 if (!page_private(head)) {
3642 BUG_ON(count & COUNT_CONTINUED);
3643 INIT_LIST_HEAD(&head->lru);
3644 set_page_private(head, SWP_CONTINUED);
3645 si->flags |= SWP_CONTINUED;
3648 list_for_each_entry(list_page, &head->lru, lru) {
3652 * If the previous map said no continuation, but we've found
3653 * a continuation page, free our allocation and use this one.
3655 if (!(count & COUNT_CONTINUED))
3656 goto out_unlock_cont;
3658 map = kmap_atomic(list_page) + offset;
3663 * If this continuation count now has some space in it,
3664 * free our allocation and use this one.
3666 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3667 goto out_unlock_cont;
3670 list_add_tail(&page->lru, &head->lru);
3671 page = NULL; /* now it's attached, don't free it */
3673 spin_unlock(&si->cont_lock);
3676 spin_unlock(&si->lock);
3677 put_swap_device(si);
3685 * swap_count_continued - when the original swap_map count is incremented
3686 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3687 * into, carry if so, or else fail until a new continuation page is allocated;
3688 * when the original swap_map count is decremented from 0 with continuation,
3689 * borrow from the continuation and report whether it still holds more.
3690 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3693 static bool swap_count_continued(struct swap_info_struct *si,
3694 pgoff_t offset, unsigned char count)
3701 head = vmalloc_to_page(si->swap_map + offset);
3702 if (page_private(head) != SWP_CONTINUED) {
3703 BUG_ON(count & COUNT_CONTINUED);
3704 return false; /* need to add count continuation */
3707 spin_lock(&si->cont_lock);
3708 offset &= ~PAGE_MASK;
3709 page = list_next_entry(head, lru);
3710 map = kmap_atomic(page) + offset;
3712 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3713 goto init_map; /* jump over SWAP_CONT_MAX checks */
3715 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3717 * Think of how you add 1 to 999
3719 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3721 page = list_next_entry(page, lru);
3722 BUG_ON(page == head);
3723 map = kmap_atomic(page) + offset;
3725 if (*map == SWAP_CONT_MAX) {
3727 page = list_next_entry(page, lru);
3729 ret = false; /* add count continuation */
3732 map = kmap_atomic(page) + offset;
3733 init_map: *map = 0; /* we didn't zero the page */
3737 while ((page = list_prev_entry(page, lru)) != head) {
3738 map = kmap_atomic(page) + offset;
3739 *map = COUNT_CONTINUED;
3742 ret = true; /* incremented */
3744 } else { /* decrementing */
3746 * Think of how you subtract 1 from 1000
3748 BUG_ON(count != COUNT_CONTINUED);
3749 while (*map == COUNT_CONTINUED) {
3751 page = list_next_entry(page, lru);
3752 BUG_ON(page == head);
3753 map = kmap_atomic(page) + offset;
3760 while ((page = list_prev_entry(page, lru)) != head) {
3761 map = kmap_atomic(page) + offset;
3762 *map = SWAP_CONT_MAX | count;
3763 count = COUNT_CONTINUED;
3766 ret = count == COUNT_CONTINUED;
3769 spin_unlock(&si->cont_lock);
3774 * free_swap_count_continuations - swapoff free all the continuation pages
3775 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3777 static void free_swap_count_continuations(struct swap_info_struct *si)
3781 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3783 head = vmalloc_to_page(si->swap_map + offset);
3784 if (page_private(head)) {
3785 struct page *page, *next;
3787 list_for_each_entry_safe(page, next, &head->lru, lru) {
3788 list_del(&page->lru);
3795 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3796 void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3798 struct swap_info_struct *si, *next;
3799 int nid = page_to_nid(page);
3801 if (!(gfp_mask & __GFP_IO))
3804 if (!blk_cgroup_congested())
3808 * We've already scheduled a throttle, avoid taking the global swap
3811 if (current->throttle_queue)
3814 spin_lock(&swap_avail_lock);
3815 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3818 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3822 spin_unlock(&swap_avail_lock);
3826 static int __init swapfile_init(void)
3830 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3832 if (!swap_avail_heads) {
3833 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3838 plist_head_init(&swap_avail_heads[nid]);
3842 subsys_initcall(swapfile_init);