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, i;
982 * Should not even be attempting cluster allocations when huge
983 * page swap is disabled. Warn and fail the allocation.
985 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
990 if (cluster_list_empty(&si->free_clusters))
993 idx = cluster_list_first(&si->free_clusters);
994 offset = idx * SWAPFILE_CLUSTER;
995 ci = lock_cluster(si, offset);
996 alloc_cluster(si, idx);
997 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
999 map = si->swap_map + offset;
1000 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1001 map[i] = SWAP_HAS_CACHE;
1003 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1004 *slot = swp_entry(si->type, offset);
1009 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1011 unsigned long offset = idx * SWAPFILE_CLUSTER;
1012 struct swap_cluster_info *ci;
1014 ci = lock_cluster(si, offset);
1015 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1016 cluster_set_count_flag(ci, 0, 0);
1017 free_cluster(si, idx);
1019 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1022 static unsigned long scan_swap_map(struct swap_info_struct *si,
1023 unsigned char usage)
1028 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
1031 return swp_offset(entry);
1037 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1039 unsigned long size = swap_entry_size(entry_size);
1040 struct swap_info_struct *si, *next;
1045 /* Only single cluster request supported */
1046 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1048 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1052 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1054 atomic_long_sub(n_goal * size, &nr_swap_pages);
1056 spin_lock(&swap_avail_lock);
1059 node = numa_node_id();
1060 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1061 /* requeue si to after same-priority siblings */
1062 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1063 spin_unlock(&swap_avail_lock);
1064 spin_lock(&si->lock);
1065 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1066 spin_lock(&swap_avail_lock);
1067 if (plist_node_empty(&si->avail_lists[node])) {
1068 spin_unlock(&si->lock);
1071 WARN(!si->highest_bit,
1072 "swap_info %d in list but !highest_bit\n",
1074 WARN(!(si->flags & SWP_WRITEOK),
1075 "swap_info %d in list but !SWP_WRITEOK\n",
1077 __del_from_avail_list(si);
1078 spin_unlock(&si->lock);
1081 if (size == SWAPFILE_CLUSTER) {
1082 if (si->flags & SWP_BLKDEV)
1083 n_ret = swap_alloc_cluster(si, swp_entries);
1085 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1086 n_goal, swp_entries);
1087 spin_unlock(&si->lock);
1088 if (n_ret || size == SWAPFILE_CLUSTER)
1090 pr_debug("scan_swap_map of si %d failed to find offset\n",
1093 spin_lock(&swap_avail_lock);
1096 * if we got here, it's likely that si was almost full before,
1097 * and since scan_swap_map() can drop the si->lock, multiple
1098 * callers probably all tried to get a page from the same si
1099 * and it filled up before we could get one; or, the si filled
1100 * up between us dropping swap_avail_lock and taking si->lock.
1101 * Since we dropped the swap_avail_lock, the swap_avail_head
1102 * list may have been modified; so if next is still in the
1103 * swap_avail_head list then try it, otherwise start over
1104 * if we have not gotten any slots.
1106 if (plist_node_empty(&next->avail_lists[node]))
1110 spin_unlock(&swap_avail_lock);
1114 atomic_long_add((long)(n_goal - n_ret) * size,
1120 /* The only caller of this function is now suspend routine */
1121 swp_entry_t get_swap_page_of_type(int type)
1123 struct swap_info_struct *si = swap_type_to_swap_info(type);
1129 spin_lock(&si->lock);
1130 if (si->flags & SWP_WRITEOK) {
1131 atomic_long_dec(&nr_swap_pages);
1132 /* This is called for allocating swap entry, not cache */
1133 offset = scan_swap_map(si, 1);
1135 spin_unlock(&si->lock);
1136 return swp_entry(type, offset);
1138 atomic_long_inc(&nr_swap_pages);
1140 spin_unlock(&si->lock);
1142 return (swp_entry_t) {0};
1145 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1147 struct swap_info_struct *p;
1148 unsigned long offset;
1152 p = swp_swap_info(entry);
1155 if (data_race(!(p->flags & SWP_USED)))
1157 offset = swp_offset(entry);
1158 if (offset >= p->max)
1163 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1166 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1169 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1174 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1176 struct swap_info_struct *p;
1178 p = __swap_info_get(entry);
1181 if (data_race(!p->swap_map[swp_offset(entry)]))
1186 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1192 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1194 struct swap_info_struct *p;
1196 p = _swap_info_get(entry);
1198 spin_lock(&p->lock);
1202 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1203 struct swap_info_struct *q)
1205 struct swap_info_struct *p;
1207 p = _swap_info_get(entry);
1211 spin_unlock(&q->lock);
1213 spin_lock(&p->lock);
1218 static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1219 unsigned long offset,
1220 unsigned char usage)
1222 unsigned char count;
1223 unsigned char has_cache;
1225 count = p->swap_map[offset];
1227 has_cache = count & SWAP_HAS_CACHE;
1228 count &= ~SWAP_HAS_CACHE;
1230 if (usage == SWAP_HAS_CACHE) {
1231 VM_BUG_ON(!has_cache);
1233 } else if (count == SWAP_MAP_SHMEM) {
1235 * Or we could insist on shmem.c using a special
1236 * swap_shmem_free() and free_shmem_swap_and_cache()...
1239 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1240 if (count == COUNT_CONTINUED) {
1241 if (swap_count_continued(p, offset, count))
1242 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1244 count = SWAP_MAP_MAX;
1249 usage = count | has_cache;
1251 WRITE_ONCE(p->swap_map[offset], usage);
1253 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1259 * Check whether swap entry is valid in the swap device. If so,
1260 * return pointer to swap_info_struct, and keep the swap entry valid
1261 * via preventing the swap device from being swapoff, until
1262 * put_swap_device() is called. Otherwise return NULL.
1264 * The entirety of the RCU read critical section must come before the
1265 * return from or after the call to synchronize_rcu() in
1266 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1267 * true, the si->map, si->cluster_info, etc. must be valid in the
1270 * Notice that swapoff or swapoff+swapon can still happen before the
1271 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1272 * in put_swap_device() if there isn't any other way to prevent
1273 * swapoff, such as page lock, page table lock, etc. The caller must
1274 * be prepared for that. For example, the following situation is
1279 * ... swapoff+swapon
1280 * __read_swap_cache_async()
1281 * swapcache_prepare()
1282 * __swap_duplicate()
1284 * // verify PTE not changed
1286 * In __swap_duplicate(), the swap_map need to be checked before
1287 * changing partly because the specified swap entry may be for another
1288 * swap device which has been swapoff. And in do_swap_page(), after
1289 * the page is read from the swap device, the PTE is verified not
1290 * changed with the page table locked to check whether the swap device
1291 * has been swapoff or swapoff+swapon.
1293 struct swap_info_struct *get_swap_device(swp_entry_t entry)
1295 struct swap_info_struct *si;
1296 unsigned long offset;
1300 si = swp_swap_info(entry);
1305 if (data_race(!(si->flags & SWP_VALID)))
1307 offset = swp_offset(entry);
1308 if (offset >= si->max)
1313 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1321 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1324 struct swap_cluster_info *ci;
1325 unsigned long offset = swp_offset(entry);
1326 unsigned char usage;
1328 ci = lock_cluster_or_swap_info(p, offset);
1329 usage = __swap_entry_free_locked(p, offset, 1);
1330 unlock_cluster_or_swap_info(p, ci);
1332 free_swap_slot(entry);
1337 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1339 struct swap_cluster_info *ci;
1340 unsigned long offset = swp_offset(entry);
1341 unsigned char count;
1343 ci = lock_cluster(p, offset);
1344 count = p->swap_map[offset];
1345 VM_BUG_ON(count != SWAP_HAS_CACHE);
1346 p->swap_map[offset] = 0;
1347 dec_cluster_info_page(p, p->cluster_info, offset);
1350 mem_cgroup_uncharge_swap(entry, 1);
1351 swap_range_free(p, offset, 1);
1355 * Caller has made sure that the swap device corresponding to entry
1356 * is still around or has not been recycled.
1358 void swap_free(swp_entry_t entry)
1360 struct swap_info_struct *p;
1362 p = _swap_info_get(entry);
1364 __swap_entry_free(p, entry);
1368 * Called after dropping swapcache to decrease refcnt to swap entries.
1370 void put_swap_page(struct page *page, swp_entry_t entry)
1372 unsigned long offset = swp_offset(entry);
1373 unsigned long idx = offset / SWAPFILE_CLUSTER;
1374 struct swap_cluster_info *ci;
1375 struct swap_info_struct *si;
1377 unsigned int i, free_entries = 0;
1379 int size = swap_entry_size(thp_nr_pages(page));
1381 si = _swap_info_get(entry);
1385 ci = lock_cluster_or_swap_info(si, offset);
1386 if (size == SWAPFILE_CLUSTER) {
1387 VM_BUG_ON(!cluster_is_huge(ci));
1388 map = si->swap_map + offset;
1389 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1391 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1392 if (val == SWAP_HAS_CACHE)
1395 cluster_clear_huge(ci);
1396 if (free_entries == SWAPFILE_CLUSTER) {
1397 unlock_cluster_or_swap_info(si, ci);
1398 spin_lock(&si->lock);
1399 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1400 swap_free_cluster(si, idx);
1401 spin_unlock(&si->lock);
1405 for (i = 0; i < size; i++, entry.val++) {
1406 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1407 unlock_cluster_or_swap_info(si, ci);
1408 free_swap_slot(entry);
1411 lock_cluster_or_swap_info(si, offset);
1414 unlock_cluster_or_swap_info(si, ci);
1417 #ifdef CONFIG_THP_SWAP
1418 int split_swap_cluster(swp_entry_t entry)
1420 struct swap_info_struct *si;
1421 struct swap_cluster_info *ci;
1422 unsigned long offset = swp_offset(entry);
1424 si = _swap_info_get(entry);
1427 ci = lock_cluster(si, offset);
1428 cluster_clear_huge(ci);
1434 static int swp_entry_cmp(const void *ent1, const void *ent2)
1436 const swp_entry_t *e1 = ent1, *e2 = ent2;
1438 return (int)swp_type(*e1) - (int)swp_type(*e2);
1441 void swapcache_free_entries(swp_entry_t *entries, int n)
1443 struct swap_info_struct *p, *prev;
1453 * Sort swap entries by swap device, so each lock is only taken once.
1454 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1455 * so low that it isn't necessary to optimize further.
1457 if (nr_swapfiles > 1)
1458 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1459 for (i = 0; i < n; ++i) {
1460 p = swap_info_get_cont(entries[i], prev);
1462 swap_entry_free(p, entries[i]);
1466 spin_unlock(&p->lock);
1470 * How many references to page are currently swapped out?
1471 * This does not give an exact answer when swap count is continued,
1472 * but does include the high COUNT_CONTINUED flag to allow for that.
1474 int page_swapcount(struct page *page)
1477 struct swap_info_struct *p;
1478 struct swap_cluster_info *ci;
1480 unsigned long offset;
1482 entry.val = page_private(page);
1483 p = _swap_info_get(entry);
1485 offset = swp_offset(entry);
1486 ci = lock_cluster_or_swap_info(p, offset);
1487 count = swap_count(p->swap_map[offset]);
1488 unlock_cluster_or_swap_info(p, ci);
1493 int __swap_count(swp_entry_t entry)
1495 struct swap_info_struct *si;
1496 pgoff_t offset = swp_offset(entry);
1499 si = get_swap_device(entry);
1501 count = swap_count(si->swap_map[offset]);
1502 put_swap_device(si);
1507 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1510 pgoff_t offset = swp_offset(entry);
1511 struct swap_cluster_info *ci;
1513 ci = lock_cluster_or_swap_info(si, offset);
1514 count = swap_count(si->swap_map[offset]);
1515 unlock_cluster_or_swap_info(si, ci);
1520 * How many references to @entry are currently swapped out?
1521 * This does not give an exact answer when swap count is continued,
1522 * but does include the high COUNT_CONTINUED flag to allow for that.
1524 int __swp_swapcount(swp_entry_t entry)
1527 struct swap_info_struct *si;
1529 si = get_swap_device(entry);
1531 count = swap_swapcount(si, entry);
1532 put_swap_device(si);
1538 * How many references to @entry are currently swapped out?
1539 * This considers COUNT_CONTINUED so it returns exact answer.
1541 int swp_swapcount(swp_entry_t entry)
1543 int count, tmp_count, n;
1544 struct swap_info_struct *p;
1545 struct swap_cluster_info *ci;
1550 p = _swap_info_get(entry);
1554 offset = swp_offset(entry);
1556 ci = lock_cluster_or_swap_info(p, offset);
1558 count = swap_count(p->swap_map[offset]);
1559 if (!(count & COUNT_CONTINUED))
1562 count &= ~COUNT_CONTINUED;
1563 n = SWAP_MAP_MAX + 1;
1565 page = vmalloc_to_page(p->swap_map + offset);
1566 offset &= ~PAGE_MASK;
1567 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1570 page = list_next_entry(page, lru);
1571 map = kmap_atomic(page);
1572 tmp_count = map[offset];
1575 count += (tmp_count & ~COUNT_CONTINUED) * n;
1576 n *= (SWAP_CONT_MAX + 1);
1577 } while (tmp_count & COUNT_CONTINUED);
1579 unlock_cluster_or_swap_info(p, ci);
1583 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1586 struct swap_cluster_info *ci;
1587 unsigned char *map = si->swap_map;
1588 unsigned long roffset = swp_offset(entry);
1589 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1593 ci = lock_cluster_or_swap_info(si, offset);
1594 if (!ci || !cluster_is_huge(ci)) {
1595 if (swap_count(map[roffset]))
1599 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1600 if (swap_count(map[offset + i])) {
1606 unlock_cluster_or_swap_info(si, ci);
1610 static bool page_swapped(struct page *page)
1613 struct swap_info_struct *si;
1615 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1616 return page_swapcount(page) != 0;
1618 page = compound_head(page);
1619 entry.val = page_private(page);
1620 si = _swap_info_get(entry);
1622 return swap_page_trans_huge_swapped(si, entry);
1626 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1627 int *total_swapcount)
1629 int i, map_swapcount, _total_mapcount, _total_swapcount;
1630 unsigned long offset = 0;
1631 struct swap_info_struct *si;
1632 struct swap_cluster_info *ci = NULL;
1633 unsigned char *map = NULL;
1634 int mapcount, swapcount = 0;
1636 /* hugetlbfs shouldn't call it */
1637 VM_BUG_ON_PAGE(PageHuge(page), page);
1639 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1640 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1641 if (PageSwapCache(page))
1642 swapcount = page_swapcount(page);
1643 if (total_swapcount)
1644 *total_swapcount = swapcount;
1645 return mapcount + swapcount;
1648 page = compound_head(page);
1650 _total_mapcount = _total_swapcount = map_swapcount = 0;
1651 if (PageSwapCache(page)) {
1654 entry.val = page_private(page);
1655 si = _swap_info_get(entry);
1658 offset = swp_offset(entry);
1662 ci = lock_cluster(si, offset);
1663 for (i = 0; i < HPAGE_PMD_NR; i++) {
1664 mapcount = atomic_read(&page[i]._mapcount) + 1;
1665 _total_mapcount += mapcount;
1667 swapcount = swap_count(map[offset + i]);
1668 _total_swapcount += swapcount;
1670 map_swapcount = max(map_swapcount, mapcount + swapcount);
1673 if (PageDoubleMap(page)) {
1675 _total_mapcount -= HPAGE_PMD_NR;
1677 mapcount = compound_mapcount(page);
1678 map_swapcount += mapcount;
1679 _total_mapcount += mapcount;
1681 *total_mapcount = _total_mapcount;
1682 if (total_swapcount)
1683 *total_swapcount = _total_swapcount;
1685 return map_swapcount;
1689 * We can write to an anon page without COW if there are no other references
1690 * to it. And as a side-effect, free up its swap: because the old content
1691 * on disk will never be read, and seeking back there to write new content
1692 * later would only waste time away from clustering.
1694 * NOTE: total_map_swapcount should not be relied upon by the caller if
1695 * reuse_swap_page() returns false, but it may be always overwritten
1696 * (see the other implementation for CONFIG_SWAP=n).
1698 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1700 int count, total_mapcount, total_swapcount;
1702 VM_BUG_ON_PAGE(!PageLocked(page), page);
1703 if (unlikely(PageKsm(page)))
1705 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1707 if (total_map_swapcount)
1708 *total_map_swapcount = total_mapcount + total_swapcount;
1709 if (count == 1 && PageSwapCache(page) &&
1710 (likely(!PageTransCompound(page)) ||
1711 /* The remaining swap count will be freed soon */
1712 total_swapcount == page_swapcount(page))) {
1713 if (!PageWriteback(page)) {
1714 page = compound_head(page);
1715 delete_from_swap_cache(page);
1719 struct swap_info_struct *p;
1721 entry.val = page_private(page);
1722 p = swap_info_get(entry);
1723 if (p->flags & SWP_STABLE_WRITES) {
1724 spin_unlock(&p->lock);
1727 spin_unlock(&p->lock);
1735 * If swap is getting full, or if there are no more mappings of this page,
1736 * then try_to_free_swap is called to free its swap space.
1738 int try_to_free_swap(struct page *page)
1740 VM_BUG_ON_PAGE(!PageLocked(page), page);
1742 if (!PageSwapCache(page))
1744 if (PageWriteback(page))
1746 if (page_swapped(page))
1750 * Once hibernation has begun to create its image of memory,
1751 * there's a danger that one of the calls to try_to_free_swap()
1752 * - most probably a call from __try_to_reclaim_swap() while
1753 * hibernation is allocating its own swap pages for the image,
1754 * but conceivably even a call from memory reclaim - will free
1755 * the swap from a page which has already been recorded in the
1756 * image as a clean swapcache page, and then reuse its swap for
1757 * another page of the image. On waking from hibernation, the
1758 * original page might be freed under memory pressure, then
1759 * later read back in from swap, now with the wrong data.
1761 * Hibernation suspends storage while it is writing the image
1762 * to disk so check that here.
1764 if (pm_suspended_storage())
1767 page = compound_head(page);
1768 delete_from_swap_cache(page);
1774 * Free the swap entry like above, but also try to
1775 * free the page cache entry if it is the last user.
1777 int free_swap_and_cache(swp_entry_t entry)
1779 struct swap_info_struct *p;
1780 unsigned char count;
1782 if (non_swap_entry(entry))
1785 p = _swap_info_get(entry);
1787 count = __swap_entry_free(p, entry);
1788 if (count == SWAP_HAS_CACHE &&
1789 !swap_page_trans_huge_swapped(p, entry))
1790 __try_to_reclaim_swap(p, swp_offset(entry),
1791 TTRS_UNMAPPED | TTRS_FULL);
1796 #ifdef CONFIG_HIBERNATION
1798 * Find the swap type that corresponds to given device (if any).
1800 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1801 * from 0, in which the swap header is expected to be located.
1803 * This is needed for the suspend to disk (aka swsusp).
1805 int swap_type_of(dev_t device, sector_t offset)
1812 spin_lock(&swap_lock);
1813 for (type = 0; type < nr_swapfiles; type++) {
1814 struct swap_info_struct *sis = swap_info[type];
1816 if (!(sis->flags & SWP_WRITEOK))
1819 if (device == sis->bdev->bd_dev) {
1820 struct swap_extent *se = first_se(sis);
1822 if (se->start_block == offset) {
1823 spin_unlock(&swap_lock);
1828 spin_unlock(&swap_lock);
1832 int find_first_swap(dev_t *device)
1836 spin_lock(&swap_lock);
1837 for (type = 0; type < nr_swapfiles; type++) {
1838 struct swap_info_struct *sis = swap_info[type];
1840 if (!(sis->flags & SWP_WRITEOK))
1842 *device = sis->bdev->bd_dev;
1843 spin_unlock(&swap_lock);
1846 spin_unlock(&swap_lock);
1851 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1852 * corresponding to given index in swap_info (swap type).
1854 sector_t swapdev_block(int type, pgoff_t offset)
1856 struct block_device *bdev;
1857 struct swap_info_struct *si = swap_type_to_swap_info(type);
1859 if (!si || !(si->flags & SWP_WRITEOK))
1861 return map_swap_entry(swp_entry(type, offset), &bdev);
1865 * Return either the total number of swap pages of given type, or the number
1866 * of free pages of that type (depending on @free)
1868 * This is needed for software suspend
1870 unsigned int count_swap_pages(int type, int free)
1874 spin_lock(&swap_lock);
1875 if ((unsigned int)type < nr_swapfiles) {
1876 struct swap_info_struct *sis = swap_info[type];
1878 spin_lock(&sis->lock);
1879 if (sis->flags & SWP_WRITEOK) {
1882 n -= sis->inuse_pages;
1884 spin_unlock(&sis->lock);
1886 spin_unlock(&swap_lock);
1889 #endif /* CONFIG_HIBERNATION */
1891 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1893 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1897 * No need to decide whether this PTE shares the swap entry with others,
1898 * just let do_wp_page work it out if a write is requested later - to
1899 * force COW, vm_page_prot omits write permission from any private vma.
1901 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1902 unsigned long addr, swp_entry_t entry, struct page *page)
1904 struct page *swapcache;
1910 page = ksm_might_need_to_copy(page, vma, addr);
1911 if (unlikely(!page))
1914 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1915 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1920 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1921 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1923 set_pte_at(vma->vm_mm, addr, pte,
1924 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1925 if (page == swapcache) {
1926 page_add_anon_rmap(page, vma, addr, false);
1927 } else { /* ksm created a completely new copy */
1928 page_add_new_anon_rmap(page, vma, addr, false);
1929 lru_cache_add_inactive_or_unevictable(page, vma);
1933 * Move the page to the active list so it is not
1934 * immediately swapped out again after swapon.
1936 activate_page(page);
1938 pte_unmap_unlock(pte, ptl);
1939 if (page != swapcache) {
1946 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1947 unsigned long addr, unsigned long end,
1948 unsigned int type, bool frontswap,
1949 unsigned long *fs_pages_to_unuse)
1954 struct swap_info_struct *si;
1955 unsigned long offset;
1957 volatile unsigned char *swap_map;
1959 si = swap_info[type];
1960 pte = pte_offset_map(pmd, addr);
1962 struct vm_fault vmf;
1964 if (!is_swap_pte(*pte))
1967 entry = pte_to_swp_entry(*pte);
1968 if (swp_type(entry) != type)
1971 offset = swp_offset(entry);
1972 if (frontswap && !frontswap_test(si, offset))
1976 swap_map = &si->swap_map[offset];
1977 page = lookup_swap_cache(entry, vma, addr);
1982 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1986 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1992 wait_on_page_writeback(page);
1993 ret = unuse_pte(vma, pmd, addr, entry, page);
2000 try_to_free_swap(page);
2004 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
2005 ret = FRONTSWAP_PAGES_UNUSED;
2009 pte = pte_offset_map(pmd, addr);
2010 } while (pte++, addr += PAGE_SIZE, addr != end);
2018 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2019 unsigned long addr, unsigned long end,
2020 unsigned int type, bool frontswap,
2021 unsigned long *fs_pages_to_unuse)
2027 pmd = pmd_offset(pud, addr);
2030 next = pmd_addr_end(addr, end);
2031 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2033 ret = unuse_pte_range(vma, pmd, addr, next, type,
2034 frontswap, fs_pages_to_unuse);
2037 } while (pmd++, addr = next, addr != end);
2041 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2042 unsigned long addr, unsigned long end,
2043 unsigned int type, bool frontswap,
2044 unsigned long *fs_pages_to_unuse)
2050 pud = pud_offset(p4d, addr);
2052 next = pud_addr_end(addr, end);
2053 if (pud_none_or_clear_bad(pud))
2055 ret = unuse_pmd_range(vma, pud, addr, next, type,
2056 frontswap, fs_pages_to_unuse);
2059 } while (pud++, addr = next, addr != end);
2063 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2064 unsigned long addr, unsigned long end,
2065 unsigned int type, bool frontswap,
2066 unsigned long *fs_pages_to_unuse)
2072 p4d = p4d_offset(pgd, addr);
2074 next = p4d_addr_end(addr, end);
2075 if (p4d_none_or_clear_bad(p4d))
2077 ret = unuse_pud_range(vma, p4d, addr, next, type,
2078 frontswap, fs_pages_to_unuse);
2081 } while (p4d++, addr = next, addr != end);
2085 static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2086 bool frontswap, unsigned long *fs_pages_to_unuse)
2089 unsigned long addr, end, next;
2092 addr = vma->vm_start;
2095 pgd = pgd_offset(vma->vm_mm, addr);
2097 next = pgd_addr_end(addr, end);
2098 if (pgd_none_or_clear_bad(pgd))
2100 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2101 frontswap, fs_pages_to_unuse);
2104 } while (pgd++, addr = next, addr != end);
2108 static int unuse_mm(struct mm_struct *mm, unsigned int type,
2109 bool frontswap, unsigned long *fs_pages_to_unuse)
2111 struct vm_area_struct *vma;
2115 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2116 if (vma->anon_vma) {
2117 ret = unuse_vma(vma, type, frontswap,
2124 mmap_read_unlock(mm);
2129 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2130 * from current position to next entry still in use. Return 0
2131 * if there are no inuse entries after prev till end of the map.
2133 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2134 unsigned int prev, bool frontswap)
2137 unsigned char count;
2140 * No need for swap_lock here: we're just looking
2141 * for whether an entry is in use, not modifying it; false
2142 * hits are okay, and sys_swapoff() has already prevented new
2143 * allocations from this area (while holding swap_lock).
2145 for (i = prev + 1; i < si->max; i++) {
2146 count = READ_ONCE(si->swap_map[i]);
2147 if (count && swap_count(count) != SWAP_MAP_BAD)
2148 if (!frontswap || frontswap_test(si, i))
2150 if ((i % LATENCY_LIMIT) == 0)
2161 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2162 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2164 int try_to_unuse(unsigned int type, bool frontswap,
2165 unsigned long pages_to_unuse)
2167 struct mm_struct *prev_mm;
2168 struct mm_struct *mm;
2169 struct list_head *p;
2171 struct swap_info_struct *si = swap_info[type];
2176 if (!READ_ONCE(si->inuse_pages))
2183 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2190 spin_lock(&mmlist_lock);
2191 p = &init_mm.mmlist;
2192 while (READ_ONCE(si->inuse_pages) &&
2193 !signal_pending(current) &&
2194 (p = p->next) != &init_mm.mmlist) {
2196 mm = list_entry(p, struct mm_struct, mmlist);
2197 if (!mmget_not_zero(mm))
2199 spin_unlock(&mmlist_lock);
2202 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2210 * Make sure that we aren't completely killing
2211 * interactive performance.
2214 spin_lock(&mmlist_lock);
2216 spin_unlock(&mmlist_lock);
2221 while (READ_ONCE(si->inuse_pages) &&
2222 !signal_pending(current) &&
2223 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2225 entry = swp_entry(type, i);
2226 page = find_get_page(swap_address_space(entry), i);
2231 * It is conceivable that a racing task removed this page from
2232 * swap cache just before we acquired the page lock. The page
2233 * might even be back in swap cache on another swap area. But
2234 * that is okay, try_to_free_swap() only removes stale pages.
2237 wait_on_page_writeback(page);
2238 try_to_free_swap(page);
2243 * For frontswap, we just need to unuse pages_to_unuse, if
2244 * it was specified. Need not check frontswap again here as
2245 * we already zeroed out pages_to_unuse if not frontswap.
2247 if (pages_to_unuse && --pages_to_unuse == 0)
2252 * Lets check again to see if there are still swap entries in the map.
2253 * If yes, we would need to do retry the unuse logic again.
2254 * Under global memory pressure, swap entries can be reinserted back
2255 * into process space after the mmlist loop above passes over them.
2257 * Limit the number of retries? No: when mmget_not_zero() above fails,
2258 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2259 * at its own independent pace; and even shmem_writepage() could have
2260 * been preempted after get_swap_page(), temporarily hiding that swap.
2261 * It's easy and robust (though cpu-intensive) just to keep retrying.
2263 if (READ_ONCE(si->inuse_pages)) {
2264 if (!signal_pending(current))
2269 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2273 * After a successful try_to_unuse, if no swap is now in use, we know
2274 * we can empty the mmlist. swap_lock must be held on entry and exit.
2275 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2276 * added to the mmlist just after page_duplicate - before would be racy.
2278 static void drain_mmlist(void)
2280 struct list_head *p, *next;
2283 for (type = 0; type < nr_swapfiles; type++)
2284 if (swap_info[type]->inuse_pages)
2286 spin_lock(&mmlist_lock);
2287 list_for_each_safe(p, next, &init_mm.mmlist)
2289 spin_unlock(&mmlist_lock);
2293 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2294 * corresponds to page offset for the specified swap entry.
2295 * Note that the type of this function is sector_t, but it returns page offset
2296 * into the bdev, not sector offset.
2298 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2300 struct swap_info_struct *sis;
2301 struct swap_extent *se;
2304 sis = swp_swap_info(entry);
2307 offset = swp_offset(entry);
2308 se = offset_to_swap_extent(sis, offset);
2309 return se->start_block + (offset - se->start_page);
2313 * Returns the page offset into bdev for the specified page's swap entry.
2315 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2318 entry.val = page_private(page);
2319 return map_swap_entry(entry, bdev);
2323 * Free all of a swapdev's extent information
2325 static void destroy_swap_extents(struct swap_info_struct *sis)
2327 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2328 struct rb_node *rb = sis->swap_extent_root.rb_node;
2329 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2331 rb_erase(rb, &sis->swap_extent_root);
2335 if (sis->flags & SWP_ACTIVATED) {
2336 struct file *swap_file = sis->swap_file;
2337 struct address_space *mapping = swap_file->f_mapping;
2339 sis->flags &= ~SWP_ACTIVATED;
2340 if (mapping->a_ops->swap_deactivate)
2341 mapping->a_ops->swap_deactivate(swap_file);
2346 * Add a block range (and the corresponding page range) into this swapdev's
2349 * This function rather assumes that it is called in ascending page order.
2352 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2353 unsigned long nr_pages, sector_t start_block)
2355 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2356 struct swap_extent *se;
2357 struct swap_extent *new_se;
2360 * place the new node at the right most since the
2361 * function is called in ascending page order.
2365 link = &parent->rb_right;
2369 se = rb_entry(parent, struct swap_extent, rb_node);
2370 BUG_ON(se->start_page + se->nr_pages != start_page);
2371 if (se->start_block + se->nr_pages == start_block) {
2373 se->nr_pages += nr_pages;
2378 /* No merge, insert a new extent. */
2379 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2382 new_se->start_page = start_page;
2383 new_se->nr_pages = nr_pages;
2384 new_se->start_block = start_block;
2386 rb_link_node(&new_se->rb_node, parent, link);
2387 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2390 EXPORT_SYMBOL_GPL(add_swap_extent);
2393 * A `swap extent' is a simple thing which maps a contiguous range of pages
2394 * onto a contiguous range of disk blocks. An ordered list of swap extents
2395 * is built at swapon time and is then used at swap_writepage/swap_readpage
2396 * time for locating where on disk a page belongs.
2398 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2399 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2400 * swap files identically.
2402 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2403 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2404 * swapfiles are handled *identically* after swapon time.
2406 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2407 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2408 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2409 * requirements, they are simply tossed out - we will never use those blocks
2412 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2413 * prevents users from writing to the swap device, which will corrupt memory.
2415 * The amount of disk space which a single swap extent represents varies.
2416 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2417 * extents in the list. To avoid much list walking, we cache the previous
2418 * search location in `curr_swap_extent', and start new searches from there.
2419 * This is extremely effective. The average number of iterations in
2420 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2422 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2424 struct file *swap_file = sis->swap_file;
2425 struct address_space *mapping = swap_file->f_mapping;
2426 struct inode *inode = mapping->host;
2429 if (S_ISBLK(inode->i_mode)) {
2430 ret = add_swap_extent(sis, 0, sis->max, 0);
2435 if (mapping->a_ops->swap_activate) {
2436 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2438 sis->flags |= SWP_ACTIVATED;
2440 sis->flags |= SWP_FS;
2441 ret = add_swap_extent(sis, 0, sis->max, 0);
2447 return generic_swapfile_activate(sis, swap_file, span);
2450 static int swap_node(struct swap_info_struct *p)
2452 struct block_device *bdev;
2457 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2459 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2462 static void setup_swap_info(struct swap_info_struct *p, int prio,
2463 unsigned char *swap_map,
2464 struct swap_cluster_info *cluster_info)
2471 p->prio = --least_priority;
2473 * the plist prio is negated because plist ordering is
2474 * low-to-high, while swap ordering is high-to-low
2476 p->list.prio = -p->prio;
2479 p->avail_lists[i].prio = -p->prio;
2481 if (swap_node(p) == i)
2482 p->avail_lists[i].prio = 1;
2484 p->avail_lists[i].prio = -p->prio;
2487 p->swap_map = swap_map;
2488 p->cluster_info = cluster_info;
2491 static void _enable_swap_info(struct swap_info_struct *p)
2493 p->flags |= SWP_WRITEOK | SWP_VALID;
2494 atomic_long_add(p->pages, &nr_swap_pages);
2495 total_swap_pages += p->pages;
2497 assert_spin_locked(&swap_lock);
2499 * both lists are plists, and thus priority ordered.
2500 * swap_active_head needs to be priority ordered for swapoff(),
2501 * which on removal of any swap_info_struct with an auto-assigned
2502 * (i.e. negative) priority increments the auto-assigned priority
2503 * of any lower-priority swap_info_structs.
2504 * swap_avail_head needs to be priority ordered for get_swap_page(),
2505 * which allocates swap pages from the highest available priority
2508 plist_add(&p->list, &swap_active_head);
2509 add_to_avail_list(p);
2512 static void enable_swap_info(struct swap_info_struct *p, int prio,
2513 unsigned char *swap_map,
2514 struct swap_cluster_info *cluster_info,
2515 unsigned long *frontswap_map)
2517 frontswap_init(p->type, frontswap_map);
2518 spin_lock(&swap_lock);
2519 spin_lock(&p->lock);
2520 setup_swap_info(p, prio, swap_map, cluster_info);
2521 spin_unlock(&p->lock);
2522 spin_unlock(&swap_lock);
2524 * Guarantee swap_map, cluster_info, etc. fields are valid
2525 * between get/put_swap_device() if SWP_VALID bit is set
2528 spin_lock(&swap_lock);
2529 spin_lock(&p->lock);
2530 _enable_swap_info(p);
2531 spin_unlock(&p->lock);
2532 spin_unlock(&swap_lock);
2535 static void reinsert_swap_info(struct swap_info_struct *p)
2537 spin_lock(&swap_lock);
2538 spin_lock(&p->lock);
2539 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2540 _enable_swap_info(p);
2541 spin_unlock(&p->lock);
2542 spin_unlock(&swap_lock);
2545 bool has_usable_swap(void)
2549 spin_lock(&swap_lock);
2550 if (plist_head_empty(&swap_active_head))
2552 spin_unlock(&swap_lock);
2556 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2558 struct swap_info_struct *p = NULL;
2559 unsigned char *swap_map;
2560 struct swap_cluster_info *cluster_info;
2561 unsigned long *frontswap_map;
2562 struct file *swap_file, *victim;
2563 struct address_space *mapping;
2564 struct inode *inode;
2565 struct filename *pathname;
2567 unsigned int old_block_size;
2569 if (!capable(CAP_SYS_ADMIN))
2572 BUG_ON(!current->mm);
2574 pathname = getname(specialfile);
2575 if (IS_ERR(pathname))
2576 return PTR_ERR(pathname);
2578 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2579 err = PTR_ERR(victim);
2583 mapping = victim->f_mapping;
2584 spin_lock(&swap_lock);
2585 plist_for_each_entry(p, &swap_active_head, list) {
2586 if (p->flags & SWP_WRITEOK) {
2587 if (p->swap_file->f_mapping == mapping) {
2595 spin_unlock(&swap_lock);
2598 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2599 vm_unacct_memory(p->pages);
2602 spin_unlock(&swap_lock);
2605 del_from_avail_list(p);
2606 spin_lock(&p->lock);
2608 struct swap_info_struct *si = p;
2611 plist_for_each_entry_continue(si, &swap_active_head, list) {
2614 for_each_node(nid) {
2615 if (si->avail_lists[nid].prio != 1)
2616 si->avail_lists[nid].prio--;
2621 plist_del(&p->list, &swap_active_head);
2622 atomic_long_sub(p->pages, &nr_swap_pages);
2623 total_swap_pages -= p->pages;
2624 p->flags &= ~SWP_WRITEOK;
2625 spin_unlock(&p->lock);
2626 spin_unlock(&swap_lock);
2628 disable_swap_slots_cache_lock();
2630 set_current_oom_origin();
2631 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2632 clear_current_oom_origin();
2635 /* re-insert swap space back into swap_list */
2636 reinsert_swap_info(p);
2637 reenable_swap_slots_cache_unlock();
2641 reenable_swap_slots_cache_unlock();
2643 spin_lock(&swap_lock);
2644 spin_lock(&p->lock);
2645 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2646 spin_unlock(&p->lock);
2647 spin_unlock(&swap_lock);
2649 * wait for swap operations protected by get/put_swap_device()
2654 flush_work(&p->discard_work);
2656 destroy_swap_extents(p);
2657 if (p->flags & SWP_CONTINUED)
2658 free_swap_count_continuations(p);
2660 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2661 atomic_dec(&nr_rotate_swap);
2663 mutex_lock(&swapon_mutex);
2664 spin_lock(&swap_lock);
2665 spin_lock(&p->lock);
2668 /* wait for anyone still in scan_swap_map */
2669 p->highest_bit = 0; /* cuts scans short */
2670 while (p->flags >= SWP_SCANNING) {
2671 spin_unlock(&p->lock);
2672 spin_unlock(&swap_lock);
2673 schedule_timeout_uninterruptible(1);
2674 spin_lock(&swap_lock);
2675 spin_lock(&p->lock);
2678 swap_file = p->swap_file;
2679 old_block_size = p->old_block_size;
2680 p->swap_file = NULL;
2682 swap_map = p->swap_map;
2684 cluster_info = p->cluster_info;
2685 p->cluster_info = NULL;
2686 frontswap_map = frontswap_map_get(p);
2687 spin_unlock(&p->lock);
2688 spin_unlock(&swap_lock);
2689 arch_swap_invalidate_area(p->type);
2690 frontswap_invalidate_area(p->type);
2691 frontswap_map_set(p, NULL);
2692 mutex_unlock(&swapon_mutex);
2693 free_percpu(p->percpu_cluster);
2694 p->percpu_cluster = NULL;
2695 free_percpu(p->cluster_next_cpu);
2696 p->cluster_next_cpu = NULL;
2698 kvfree(cluster_info);
2699 kvfree(frontswap_map);
2700 /* Destroy swap account information */
2701 swap_cgroup_swapoff(p->type);
2702 exit_swap_address_space(p->type);
2704 inode = mapping->host;
2705 if (S_ISBLK(inode->i_mode)) {
2706 struct block_device *bdev = I_BDEV(inode);
2708 set_blocksize(bdev, old_block_size);
2709 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2713 inode->i_flags &= ~S_SWAPFILE;
2714 inode_unlock(inode);
2715 filp_close(swap_file, NULL);
2718 * Clear the SWP_USED flag after all resources are freed so that swapon
2719 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2720 * not hold p->lock after we cleared its SWP_WRITEOK.
2722 spin_lock(&swap_lock);
2724 spin_unlock(&swap_lock);
2727 atomic_inc(&proc_poll_event);
2728 wake_up_interruptible(&proc_poll_wait);
2731 filp_close(victim, NULL);
2737 #ifdef CONFIG_PROC_FS
2738 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2740 struct seq_file *seq = file->private_data;
2742 poll_wait(file, &proc_poll_wait, wait);
2744 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2745 seq->poll_event = atomic_read(&proc_poll_event);
2746 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2749 return EPOLLIN | EPOLLRDNORM;
2753 static void *swap_start(struct seq_file *swap, loff_t *pos)
2755 struct swap_info_struct *si;
2759 mutex_lock(&swapon_mutex);
2762 return SEQ_START_TOKEN;
2764 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2765 if (!(si->flags & SWP_USED) || !si->swap_map)
2774 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2776 struct swap_info_struct *si = v;
2779 if (v == SEQ_START_TOKEN)
2782 type = si->type + 1;
2785 for (; (si = swap_type_to_swap_info(type)); type++) {
2786 if (!(si->flags & SWP_USED) || !si->swap_map)
2794 static void swap_stop(struct seq_file *swap, void *v)
2796 mutex_unlock(&swapon_mutex);
2799 static int swap_show(struct seq_file *swap, void *v)
2801 struct swap_info_struct *si = v;
2804 unsigned int bytes, inuse;
2806 if (si == SEQ_START_TOKEN) {
2807 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2811 bytes = si->pages << (PAGE_SHIFT - 10);
2812 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2814 file = si->swap_file;
2815 len = seq_file_path(swap, file, " \t\n\\");
2816 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2817 len < 40 ? 40 - len : 1, " ",
2818 S_ISBLK(file_inode(file)->i_mode) ?
2819 "partition" : "file\t",
2820 bytes, bytes < 10000000 ? "\t" : "",
2821 inuse, inuse < 10000000 ? "\t" : "",
2826 static const struct seq_operations swaps_op = {
2827 .start = swap_start,
2833 static int swaps_open(struct inode *inode, struct file *file)
2835 struct seq_file *seq;
2838 ret = seq_open(file, &swaps_op);
2842 seq = file->private_data;
2843 seq->poll_event = atomic_read(&proc_poll_event);
2847 static const struct proc_ops swaps_proc_ops = {
2848 .proc_flags = PROC_ENTRY_PERMANENT,
2849 .proc_open = swaps_open,
2850 .proc_read = seq_read,
2851 .proc_lseek = seq_lseek,
2852 .proc_release = seq_release,
2853 .proc_poll = swaps_poll,
2856 static int __init procswaps_init(void)
2858 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2861 __initcall(procswaps_init);
2862 #endif /* CONFIG_PROC_FS */
2864 #ifdef MAX_SWAPFILES_CHECK
2865 static int __init max_swapfiles_check(void)
2867 MAX_SWAPFILES_CHECK();
2870 late_initcall(max_swapfiles_check);
2873 static struct swap_info_struct *alloc_swap_info(void)
2875 struct swap_info_struct *p;
2879 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2881 return ERR_PTR(-ENOMEM);
2883 spin_lock(&swap_lock);
2884 for (type = 0; type < nr_swapfiles; type++) {
2885 if (!(swap_info[type]->flags & SWP_USED))
2888 if (type >= MAX_SWAPFILES) {
2889 spin_unlock(&swap_lock);
2891 return ERR_PTR(-EPERM);
2893 if (type >= nr_swapfiles) {
2895 WRITE_ONCE(swap_info[type], p);
2897 * Write swap_info[type] before nr_swapfiles, in case a
2898 * racing procfs swap_start() or swap_next() is reading them.
2899 * (We never shrink nr_swapfiles, we never free this entry.)
2902 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2905 p = swap_info[type];
2907 * Do not memset this entry: a racing procfs swap_next()
2908 * would be relying on p->type to remain valid.
2911 p->swap_extent_root = RB_ROOT;
2912 plist_node_init(&p->list, 0);
2914 plist_node_init(&p->avail_lists[i], 0);
2915 p->flags = SWP_USED;
2916 spin_unlock(&swap_lock);
2917 spin_lock_init(&p->lock);
2918 spin_lock_init(&p->cont_lock);
2923 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2927 if (S_ISBLK(inode->i_mode)) {
2928 p->bdev = blkdev_get_by_dev(inode->i_rdev,
2929 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2930 if (IS_ERR(p->bdev)) {
2931 error = PTR_ERR(p->bdev);
2935 p->old_block_size = block_size(p->bdev);
2936 error = set_blocksize(p->bdev, PAGE_SIZE);
2940 * Zoned block devices contain zones that have a sequential
2941 * write only restriction. Hence zoned block devices are not
2942 * suitable for swapping. Disallow them here.
2944 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2946 p->flags |= SWP_BLKDEV;
2947 } else if (S_ISREG(inode->i_mode)) {
2948 p->bdev = inode->i_sb->s_bdev;
2956 * Find out how many pages are allowed for a single swap device. There
2957 * are two limiting factors:
2958 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2959 * 2) the number of bits in the swap pte, as defined by the different
2962 * In order to find the largest possible bit mask, a swap entry with
2963 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2964 * decoded to a swp_entry_t again, and finally the swap offset is
2967 * This will mask all the bits from the initial ~0UL mask that can't
2968 * be encoded in either the swp_entry_t or the architecture definition
2971 unsigned long generic_max_swapfile_size(void)
2973 return swp_offset(pte_to_swp_entry(
2974 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2977 /* Can be overridden by an architecture for additional checks. */
2978 __weak unsigned long max_swapfile_size(void)
2980 return generic_max_swapfile_size();
2983 static unsigned long read_swap_header(struct swap_info_struct *p,
2984 union swap_header *swap_header,
2985 struct inode *inode)
2988 unsigned long maxpages;
2989 unsigned long swapfilepages;
2990 unsigned long last_page;
2992 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2993 pr_err("Unable to find swap-space signature\n");
2997 /* swap partition endianess hack... */
2998 if (swab32(swap_header->info.version) == 1) {
2999 swab32s(&swap_header->info.version);
3000 swab32s(&swap_header->info.last_page);
3001 swab32s(&swap_header->info.nr_badpages);
3002 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3004 for (i = 0; i < swap_header->info.nr_badpages; i++)
3005 swab32s(&swap_header->info.badpages[i]);
3007 /* Check the swap header's sub-version */
3008 if (swap_header->info.version != 1) {
3009 pr_warn("Unable to handle swap header version %d\n",
3010 swap_header->info.version);
3015 p->cluster_next = 1;
3018 maxpages = max_swapfile_size();
3019 last_page = swap_header->info.last_page;
3021 pr_warn("Empty swap-file\n");
3024 if (last_page > maxpages) {
3025 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3026 maxpages << (PAGE_SHIFT - 10),
3027 last_page << (PAGE_SHIFT - 10));
3029 if (maxpages > last_page) {
3030 maxpages = last_page + 1;
3031 /* p->max is an unsigned int: don't overflow it */
3032 if ((unsigned int)maxpages == 0)
3033 maxpages = UINT_MAX;
3035 p->highest_bit = maxpages - 1;
3039 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3040 if (swapfilepages && maxpages > swapfilepages) {
3041 pr_warn("Swap area shorter than signature indicates\n");
3044 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3046 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3052 #define SWAP_CLUSTER_INFO_COLS \
3053 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3054 #define SWAP_CLUSTER_SPACE_COLS \
3055 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3056 #define SWAP_CLUSTER_COLS \
3057 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3059 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3060 union swap_header *swap_header,
3061 unsigned char *swap_map,
3062 struct swap_cluster_info *cluster_info,
3063 unsigned long maxpages,
3067 unsigned int nr_good_pages;
3069 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3070 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3071 unsigned long i, idx;
3073 nr_good_pages = maxpages - 1; /* omit header page */
3075 cluster_list_init(&p->free_clusters);
3076 cluster_list_init(&p->discard_clusters);
3078 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3079 unsigned int page_nr = swap_header->info.badpages[i];
3080 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3082 if (page_nr < maxpages) {
3083 swap_map[page_nr] = SWAP_MAP_BAD;
3086 * Haven't marked the cluster free yet, no list
3087 * operation involved
3089 inc_cluster_info_page(p, cluster_info, page_nr);
3093 /* Haven't marked the cluster free yet, no list operation involved */
3094 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3095 inc_cluster_info_page(p, cluster_info, i);
3097 if (nr_good_pages) {
3098 swap_map[0] = SWAP_MAP_BAD;
3100 * Not mark the cluster free yet, no list
3101 * operation involved
3103 inc_cluster_info_page(p, cluster_info, 0);
3105 p->pages = nr_good_pages;
3106 nr_extents = setup_swap_extents(p, span);
3109 nr_good_pages = p->pages;
3111 if (!nr_good_pages) {
3112 pr_warn("Empty swap-file\n");
3121 * Reduce false cache line sharing between cluster_info and
3122 * sharing same address space.
3124 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3125 j = (k + col) % SWAP_CLUSTER_COLS;
3126 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3127 idx = i * SWAP_CLUSTER_COLS + j;
3128 if (idx >= nr_clusters)
3130 if (cluster_count(&cluster_info[idx]))
3132 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3133 cluster_list_add_tail(&p->free_clusters, cluster_info,
3141 * Helper to sys_swapon determining if a given swap
3142 * backing device queue supports DISCARD operations.
3144 static bool swap_discardable(struct swap_info_struct *si)
3146 struct request_queue *q = bdev_get_queue(si->bdev);
3148 if (!q || !blk_queue_discard(q))
3154 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3156 struct swap_info_struct *p;
3157 struct filename *name;
3158 struct file *swap_file = NULL;
3159 struct address_space *mapping;
3162 union swap_header *swap_header;
3165 unsigned long maxpages;
3166 unsigned char *swap_map = NULL;
3167 struct swap_cluster_info *cluster_info = NULL;
3168 unsigned long *frontswap_map = NULL;
3169 struct page *page = NULL;
3170 struct inode *inode = NULL;
3171 bool inced_nr_rotate_swap = false;
3173 if (swap_flags & ~SWAP_FLAGS_VALID)
3176 if (!capable(CAP_SYS_ADMIN))
3179 if (!swap_avail_heads)
3182 p = alloc_swap_info();
3186 INIT_WORK(&p->discard_work, swap_discard_work);
3188 name = getname(specialfile);
3190 error = PTR_ERR(name);
3194 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3195 if (IS_ERR(swap_file)) {
3196 error = PTR_ERR(swap_file);
3201 p->swap_file = swap_file;
3202 mapping = swap_file->f_mapping;
3203 inode = mapping->host;
3205 error = claim_swapfile(p, inode);
3206 if (unlikely(error))
3210 if (IS_SWAPFILE(inode)) {
3212 goto bad_swap_unlock_inode;
3216 * Read the swap header.
3218 if (!mapping->a_ops->readpage) {
3220 goto bad_swap_unlock_inode;
3222 page = read_mapping_page(mapping, 0, swap_file);
3224 error = PTR_ERR(page);
3225 goto bad_swap_unlock_inode;
3227 swap_header = kmap(page);
3229 maxpages = read_swap_header(p, swap_header, inode);
3230 if (unlikely(!maxpages)) {
3232 goto bad_swap_unlock_inode;
3235 /* OK, set up the swap map and apply the bad block list */
3236 swap_map = vzalloc(maxpages);
3239 goto bad_swap_unlock_inode;
3242 if (p->bdev && blk_queue_stable_writes(p->bdev->bd_disk->queue))
3243 p->flags |= SWP_STABLE_WRITES;
3245 if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3246 p->flags |= SWP_SYNCHRONOUS_IO;
3248 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3250 unsigned long ci, nr_cluster;
3252 p->flags |= SWP_SOLIDSTATE;
3253 p->cluster_next_cpu = alloc_percpu(unsigned int);
3254 if (!p->cluster_next_cpu) {
3256 goto bad_swap_unlock_inode;
3259 * select a random position to start with to help wear leveling
3262 for_each_possible_cpu(cpu) {
3263 per_cpu(*p->cluster_next_cpu, cpu) =
3264 1 + prandom_u32_max(p->highest_bit);
3266 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3268 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3270 if (!cluster_info) {
3272 goto bad_swap_unlock_inode;
3275 for (ci = 0; ci < nr_cluster; ci++)
3276 spin_lock_init(&((cluster_info + ci)->lock));
3278 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3279 if (!p->percpu_cluster) {
3281 goto bad_swap_unlock_inode;
3283 for_each_possible_cpu(cpu) {
3284 struct percpu_cluster *cluster;
3285 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3286 cluster_set_null(&cluster->index);
3289 atomic_inc(&nr_rotate_swap);
3290 inced_nr_rotate_swap = true;
3293 error = swap_cgroup_swapon(p->type, maxpages);
3295 goto bad_swap_unlock_inode;
3297 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3298 cluster_info, maxpages, &span);
3299 if (unlikely(nr_extents < 0)) {
3301 goto bad_swap_unlock_inode;
3303 /* frontswap enabled? set up bit-per-page map for frontswap */
3304 if (IS_ENABLED(CONFIG_FRONTSWAP))
3305 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3309 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3311 * When discard is enabled for swap with no particular
3312 * policy flagged, we set all swap discard flags here in
3313 * order to sustain backward compatibility with older
3314 * swapon(8) releases.
3316 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3320 * By flagging sys_swapon, a sysadmin can tell us to
3321 * either do single-time area discards only, or to just
3322 * perform discards for released swap page-clusters.
3323 * Now it's time to adjust the p->flags accordingly.
3325 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3326 p->flags &= ~SWP_PAGE_DISCARD;
3327 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3328 p->flags &= ~SWP_AREA_DISCARD;
3330 /* issue a swapon-time discard if it's still required */
3331 if (p->flags & SWP_AREA_DISCARD) {
3332 int err = discard_swap(p);
3334 pr_err("swapon: discard_swap(%p): %d\n",
3339 error = init_swap_address_space(p->type, maxpages);
3341 goto bad_swap_unlock_inode;
3344 * Flush any pending IO and dirty mappings before we start using this
3347 inode->i_flags |= S_SWAPFILE;
3348 error = inode_drain_writes(inode);
3350 inode->i_flags &= ~S_SWAPFILE;
3351 goto bad_swap_unlock_inode;
3354 mutex_lock(&swapon_mutex);
3356 if (swap_flags & SWAP_FLAG_PREFER)
3358 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3359 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3361 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3362 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3363 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3364 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3365 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3366 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3367 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3368 (frontswap_map) ? "FS" : "");
3370 mutex_unlock(&swapon_mutex);
3371 atomic_inc(&proc_poll_event);
3372 wake_up_interruptible(&proc_poll_wait);
3376 bad_swap_unlock_inode:
3377 inode_unlock(inode);
3379 free_percpu(p->percpu_cluster);
3380 p->percpu_cluster = NULL;
3381 free_percpu(p->cluster_next_cpu);
3382 p->cluster_next_cpu = NULL;
3383 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3384 set_blocksize(p->bdev, p->old_block_size);
3385 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3388 destroy_swap_extents(p);
3389 swap_cgroup_swapoff(p->type);
3390 spin_lock(&swap_lock);
3391 p->swap_file = NULL;
3393 spin_unlock(&swap_lock);
3395 kvfree(cluster_info);
3396 kvfree(frontswap_map);
3397 if (inced_nr_rotate_swap)
3398 atomic_dec(&nr_rotate_swap);
3400 filp_close(swap_file, NULL);
3402 if (page && !IS_ERR(page)) {
3409 inode_unlock(inode);
3411 enable_swap_slots_cache();
3415 void si_swapinfo(struct sysinfo *val)
3418 unsigned long nr_to_be_unused = 0;
3420 spin_lock(&swap_lock);
3421 for (type = 0; type < nr_swapfiles; type++) {
3422 struct swap_info_struct *si = swap_info[type];
3424 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3425 nr_to_be_unused += si->inuse_pages;
3427 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3428 val->totalswap = total_swap_pages + nr_to_be_unused;
3429 spin_unlock(&swap_lock);
3433 * Verify that a swap entry is valid and increment its swap map count.
3435 * Returns error code in following case.
3437 * - swp_entry is invalid -> EINVAL
3438 * - swp_entry is migration entry -> EINVAL
3439 * - swap-cache reference is requested but there is already one. -> EEXIST
3440 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3441 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3443 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3445 struct swap_info_struct *p;
3446 struct swap_cluster_info *ci;
3447 unsigned long offset;
3448 unsigned char count;
3449 unsigned char has_cache;
3452 p = get_swap_device(entry);
3456 offset = swp_offset(entry);
3457 ci = lock_cluster_or_swap_info(p, offset);
3459 count = p->swap_map[offset];
3462 * swapin_readahead() doesn't check if a swap entry is valid, so the
3463 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3465 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3470 has_cache = count & SWAP_HAS_CACHE;
3471 count &= ~SWAP_HAS_CACHE;
3474 if (usage == SWAP_HAS_CACHE) {
3476 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3477 if (!has_cache && count)
3478 has_cache = SWAP_HAS_CACHE;
3479 else if (has_cache) /* someone else added cache */
3481 else /* no users remaining */
3484 } else if (count || has_cache) {
3486 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3488 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3490 else if (swap_count_continued(p, offset, count))
3491 count = COUNT_CONTINUED;
3495 err = -ENOENT; /* unused swap entry */
3497 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3500 unlock_cluster_or_swap_info(p, ci);
3508 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3509 * (in which case its reference count is never incremented).
3511 void swap_shmem_alloc(swp_entry_t entry)
3513 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3517 * Increase reference count of swap entry by 1.
3518 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3519 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3520 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3521 * might occur if a page table entry has got corrupted.
3523 int swap_duplicate(swp_entry_t entry)
3527 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3528 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3533 * @entry: swap entry for which we allocate swap cache.
3535 * Called when allocating swap cache for existing swap entry,
3536 * This can return error codes. Returns 0 at success.
3537 * -EEXIST means there is a swap cache.
3538 * Note: return code is different from swap_duplicate().
3540 int swapcache_prepare(swp_entry_t entry)
3542 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3545 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3547 return swap_type_to_swap_info(swp_type(entry));
3550 struct swap_info_struct *page_swap_info(struct page *page)
3552 swp_entry_t entry = { .val = page_private(page) };
3553 return swp_swap_info(entry);
3557 * out-of-line __page_file_ methods to avoid include hell.
3559 struct address_space *__page_file_mapping(struct page *page)
3561 return page_swap_info(page)->swap_file->f_mapping;
3563 EXPORT_SYMBOL_GPL(__page_file_mapping);
3565 pgoff_t __page_file_index(struct page *page)
3567 swp_entry_t swap = { .val = page_private(page) };
3568 return swp_offset(swap);
3570 EXPORT_SYMBOL_GPL(__page_file_index);
3573 * add_swap_count_continuation - called when a swap count is duplicated
3574 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3575 * page of the original vmalloc'ed swap_map, to hold the continuation count
3576 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3577 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3579 * These continuation pages are seldom referenced: the common paths all work
3580 * on the original swap_map, only referring to a continuation page when the
3581 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3583 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3584 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3585 * can be called after dropping locks.
3587 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3589 struct swap_info_struct *si;
3590 struct swap_cluster_info *ci;
3593 struct page *list_page;
3595 unsigned char count;
3599 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3600 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3602 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3604 si = get_swap_device(entry);
3607 * An acceptable race has occurred since the failing
3608 * __swap_duplicate(): the swap device may be swapoff
3612 spin_lock(&si->lock);
3614 offset = swp_offset(entry);
3616 ci = lock_cluster(si, offset);
3618 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3620 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3622 * The higher the swap count, the more likely it is that tasks
3623 * will race to add swap count continuation: we need to avoid
3624 * over-provisioning.
3635 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3636 * no architecture is using highmem pages for kernel page tables: so it
3637 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3639 head = vmalloc_to_page(si->swap_map + offset);
3640 offset &= ~PAGE_MASK;
3642 spin_lock(&si->cont_lock);
3644 * Page allocation does not initialize the page's lru field,
3645 * but it does always reset its private field.
3647 if (!page_private(head)) {
3648 BUG_ON(count & COUNT_CONTINUED);
3649 INIT_LIST_HEAD(&head->lru);
3650 set_page_private(head, SWP_CONTINUED);
3651 si->flags |= SWP_CONTINUED;
3654 list_for_each_entry(list_page, &head->lru, lru) {
3658 * If the previous map said no continuation, but we've found
3659 * a continuation page, free our allocation and use this one.
3661 if (!(count & COUNT_CONTINUED))
3662 goto out_unlock_cont;
3664 map = kmap_atomic(list_page) + offset;
3669 * If this continuation count now has some space in it,
3670 * free our allocation and use this one.
3672 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3673 goto out_unlock_cont;
3676 list_add_tail(&page->lru, &head->lru);
3677 page = NULL; /* now it's attached, don't free it */
3679 spin_unlock(&si->cont_lock);
3682 spin_unlock(&si->lock);
3683 put_swap_device(si);
3691 * swap_count_continued - when the original swap_map count is incremented
3692 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3693 * into, carry if so, or else fail until a new continuation page is allocated;
3694 * when the original swap_map count is decremented from 0 with continuation,
3695 * borrow from the continuation and report whether it still holds more.
3696 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3699 static bool swap_count_continued(struct swap_info_struct *si,
3700 pgoff_t offset, unsigned char count)
3707 head = vmalloc_to_page(si->swap_map + offset);
3708 if (page_private(head) != SWP_CONTINUED) {
3709 BUG_ON(count & COUNT_CONTINUED);
3710 return false; /* need to add count continuation */
3713 spin_lock(&si->cont_lock);
3714 offset &= ~PAGE_MASK;
3715 page = list_next_entry(head, lru);
3716 map = kmap_atomic(page) + offset;
3718 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3719 goto init_map; /* jump over SWAP_CONT_MAX checks */
3721 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3723 * Think of how you add 1 to 999
3725 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3727 page = list_next_entry(page, lru);
3728 BUG_ON(page == head);
3729 map = kmap_atomic(page) + offset;
3731 if (*map == SWAP_CONT_MAX) {
3733 page = list_next_entry(page, lru);
3735 ret = false; /* add count continuation */
3738 map = kmap_atomic(page) + offset;
3739 init_map: *map = 0; /* we didn't zero the page */
3743 while ((page = list_prev_entry(page, lru)) != head) {
3744 map = kmap_atomic(page) + offset;
3745 *map = COUNT_CONTINUED;
3748 ret = true; /* incremented */
3750 } else { /* decrementing */
3752 * Think of how you subtract 1 from 1000
3754 BUG_ON(count != COUNT_CONTINUED);
3755 while (*map == COUNT_CONTINUED) {
3757 page = list_next_entry(page, lru);
3758 BUG_ON(page == head);
3759 map = kmap_atomic(page) + offset;
3766 while ((page = list_prev_entry(page, lru)) != head) {
3767 map = kmap_atomic(page) + offset;
3768 *map = SWAP_CONT_MAX | count;
3769 count = COUNT_CONTINUED;
3772 ret = count == COUNT_CONTINUED;
3775 spin_unlock(&si->cont_lock);
3780 * free_swap_count_continuations - swapoff free all the continuation pages
3781 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3783 static void free_swap_count_continuations(struct swap_info_struct *si)
3787 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3789 head = vmalloc_to_page(si->swap_map + offset);
3790 if (page_private(head)) {
3791 struct page *page, *next;
3793 list_for_each_entry_safe(page, next, &head->lru, lru) {
3794 list_del(&page->lru);
3801 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3802 void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3804 struct swap_info_struct *si, *next;
3805 int nid = page_to_nid(page);
3807 if (!(gfp_mask & __GFP_IO))
3810 if (!blk_cgroup_congested())
3814 * We've already scheduled a throttle, avoid taking the global swap
3817 if (current->throttle_queue)
3820 spin_lock(&swap_avail_lock);
3821 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3824 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3828 spin_unlock(&swap_avail_lock);
3832 static int __init swapfile_init(void)
3836 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3838 if (!swap_avail_heads) {
3839 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3844 plist_head_init(&swap_avail_heads[nid]);
3848 subsys_initcall(swapfile_init);