4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
9 #include <linux/sched/mm.h>
10 #include <linux/sched/task.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mman.h>
13 #include <linux/slab.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/swap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/pagemap.h>
18 #include <linux/namei.h>
19 #include <linux/shmem_fs.h>
20 #include <linux/blkdev.h>
21 #include <linux/random.h>
22 #include <linux/writeback.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/init.h>
26 #include <linux/ksm.h>
27 #include <linux/rmap.h>
28 #include <linux/security.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mutex.h>
31 #include <linux/capability.h>
32 #include <linux/syscalls.h>
33 #include <linux/memcontrol.h>
34 #include <linux/poll.h>
35 #include <linux/oom.h>
36 #include <linux/frontswap.h>
37 #include <linux/swapfile.h>
38 #include <linux/export.h>
39 #include <linux/swap_slots.h>
40 #include <linux/sort.h>
42 #include <asm/pgtable.h>
43 #include <asm/tlbflush.h>
44 #include <linux/swapops.h>
45 #include <linux/swap_cgroup.h>
47 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
49 static void free_swap_count_continuations(struct swap_info_struct *);
50 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
52 DEFINE_SPINLOCK(swap_lock);
53 static unsigned int nr_swapfiles;
54 atomic_long_t nr_swap_pages;
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
60 EXPORT_SYMBOL_GPL(nr_swap_pages);
61 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62 long total_swap_pages;
63 static int least_priority = -1;
65 static const char Bad_file[] = "Bad swap file entry ";
66 static const char Unused_file[] = "Unused swap file entry ";
67 static const char Bad_offset[] = "Bad swap offset entry ";
68 static const char Unused_offset[] = "Unused swap offset entry ";
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
74 PLIST_HEAD(swap_active_head);
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
88 static struct plist_head *swap_avail_heads;
89 static DEFINE_SPINLOCK(swap_avail_lock);
91 struct swap_info_struct *swap_info[MAX_SWAPFILES];
93 static DEFINE_MUTEX(swapon_mutex);
95 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96 /* Activity counter to indicate that a swapon or swapoff has occurred */
97 static atomic_t proc_poll_event = ATOMIC_INIT(0);
99 atomic_t nr_rotate_swap = ATOMIC_INIT(0);
101 static inline unsigned char swap_count(unsigned char ent)
103 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
106 /* returns 1 if swap entry is freed */
108 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
110 swp_entry_t entry = swp_entry(si->type, offset);
114 page = find_get_page(swap_address_space(entry), swp_offset(entry));
118 * This function is called from scan_swap_map() and it's called
119 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
120 * We have to use trylock for avoiding deadlock. This is a special
121 * case and you should use try_to_free_swap() with explicit lock_page()
122 * in usual operations.
124 if (trylock_page(page)) {
125 ret = try_to_free_swap(page);
133 * swapon tell device that all the old swap contents can be discarded,
134 * to allow the swap device to optimize its wear-levelling.
136 static int discard_swap(struct swap_info_struct *si)
138 struct swap_extent *se;
139 sector_t start_block;
143 /* Do not discard the swap header page! */
144 se = &si->first_swap_extent;
145 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
146 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
148 err = blkdev_issue_discard(si->bdev, start_block,
149 nr_blocks, GFP_KERNEL, 0);
155 list_for_each_entry(se, &si->first_swap_extent.list, list) {
156 start_block = se->start_block << (PAGE_SHIFT - 9);
157 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
159 err = blkdev_issue_discard(si->bdev, start_block,
160 nr_blocks, GFP_KERNEL, 0);
166 return err; /* That will often be -EOPNOTSUPP */
170 * swap allocation tell device that a cluster of swap can now be discarded,
171 * to allow the swap device to optimize its wear-levelling.
173 static void discard_swap_cluster(struct swap_info_struct *si,
174 pgoff_t start_page, pgoff_t nr_pages)
176 struct swap_extent *se = si->curr_swap_extent;
177 int found_extent = 0;
180 if (se->start_page <= start_page &&
181 start_page < se->start_page + se->nr_pages) {
182 pgoff_t offset = start_page - se->start_page;
183 sector_t start_block = se->start_block + offset;
184 sector_t nr_blocks = se->nr_pages - offset;
186 if (nr_blocks > nr_pages)
187 nr_blocks = nr_pages;
188 start_page += nr_blocks;
189 nr_pages -= nr_blocks;
192 si->curr_swap_extent = se;
194 start_block <<= PAGE_SHIFT - 9;
195 nr_blocks <<= PAGE_SHIFT - 9;
196 if (blkdev_issue_discard(si->bdev, start_block,
197 nr_blocks, GFP_NOIO, 0))
201 se = list_next_entry(se, list);
205 #ifdef CONFIG_THP_SWAP
206 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
208 #define SWAPFILE_CLUSTER 256
210 #define LATENCY_LIMIT 256
212 static inline void cluster_set_flag(struct swap_cluster_info *info,
218 static inline unsigned int cluster_count(struct swap_cluster_info *info)
223 static inline void cluster_set_count(struct swap_cluster_info *info,
229 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
230 unsigned int c, unsigned int f)
236 static inline unsigned int cluster_next(struct swap_cluster_info *info)
241 static inline void cluster_set_next(struct swap_cluster_info *info,
247 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
248 unsigned int n, unsigned int f)
254 static inline bool cluster_is_free(struct swap_cluster_info *info)
256 return info->flags & CLUSTER_FLAG_FREE;
259 static inline bool cluster_is_null(struct swap_cluster_info *info)
261 return info->flags & CLUSTER_FLAG_NEXT_NULL;
264 static inline void cluster_set_null(struct swap_cluster_info *info)
266 info->flags = CLUSTER_FLAG_NEXT_NULL;
270 static inline bool cluster_is_huge(struct swap_cluster_info *info)
272 return info->flags & CLUSTER_FLAG_HUGE;
275 static inline void cluster_clear_huge(struct swap_cluster_info *info)
277 info->flags &= ~CLUSTER_FLAG_HUGE;
280 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
281 unsigned long offset)
283 struct swap_cluster_info *ci;
285 ci = si->cluster_info;
287 ci += offset / SWAPFILE_CLUSTER;
288 spin_lock(&ci->lock);
293 static inline void unlock_cluster(struct swap_cluster_info *ci)
296 spin_unlock(&ci->lock);
300 * Determine the locking method in use for this device. Return
301 * swap_cluster_info if SSD-style cluster-based locking is in place.
303 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
304 struct swap_info_struct *si, unsigned long offset)
306 struct swap_cluster_info *ci;
308 /* Try to use fine-grained SSD-style locking if available: */
309 ci = lock_cluster(si, offset);
310 /* Otherwise, fall back to traditional, coarse locking: */
312 spin_lock(&si->lock);
317 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
318 struct swap_cluster_info *ci)
323 spin_unlock(&si->lock);
326 static inline bool cluster_list_empty(struct swap_cluster_list *list)
328 return cluster_is_null(&list->head);
331 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
333 return cluster_next(&list->head);
336 static void cluster_list_init(struct swap_cluster_list *list)
338 cluster_set_null(&list->head);
339 cluster_set_null(&list->tail);
342 static void cluster_list_add_tail(struct swap_cluster_list *list,
343 struct swap_cluster_info *ci,
346 if (cluster_list_empty(list)) {
347 cluster_set_next_flag(&list->head, idx, 0);
348 cluster_set_next_flag(&list->tail, idx, 0);
350 struct swap_cluster_info *ci_tail;
351 unsigned int tail = cluster_next(&list->tail);
354 * Nested cluster lock, but both cluster locks are
355 * only acquired when we held swap_info_struct->lock
358 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
359 cluster_set_next(ci_tail, idx);
360 spin_unlock(&ci_tail->lock);
361 cluster_set_next_flag(&list->tail, idx, 0);
365 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
366 struct swap_cluster_info *ci)
370 idx = cluster_next(&list->head);
371 if (cluster_next(&list->tail) == idx) {
372 cluster_set_null(&list->head);
373 cluster_set_null(&list->tail);
375 cluster_set_next_flag(&list->head,
376 cluster_next(&ci[idx]), 0);
381 /* Add a cluster to discard list and schedule it to do discard */
382 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
386 * If scan_swap_map() can't find a free cluster, it will check
387 * si->swap_map directly. To make sure the discarding cluster isn't
388 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
389 * will be cleared after discard
391 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
392 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
394 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
396 schedule_work(&si->discard_work);
399 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
401 struct swap_cluster_info *ci = si->cluster_info;
403 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
404 cluster_list_add_tail(&si->free_clusters, ci, idx);
408 * Doing discard actually. After a cluster discard is finished, the cluster
409 * will be added to free cluster list. caller should hold si->lock.
411 static void swap_do_scheduled_discard(struct swap_info_struct *si)
413 struct swap_cluster_info *info, *ci;
416 info = si->cluster_info;
418 while (!cluster_list_empty(&si->discard_clusters)) {
419 idx = cluster_list_del_first(&si->discard_clusters, info);
420 spin_unlock(&si->lock);
422 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
425 spin_lock(&si->lock);
426 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
427 __free_cluster(si, idx);
428 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
429 0, SWAPFILE_CLUSTER);
434 static void swap_discard_work(struct work_struct *work)
436 struct swap_info_struct *si;
438 si = container_of(work, struct swap_info_struct, discard_work);
440 spin_lock(&si->lock);
441 swap_do_scheduled_discard(si);
442 spin_unlock(&si->lock);
445 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
447 struct swap_cluster_info *ci = si->cluster_info;
449 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
450 cluster_list_del_first(&si->free_clusters, ci);
451 cluster_set_count_flag(ci + idx, 0, 0);
454 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
456 struct swap_cluster_info *ci = si->cluster_info + idx;
458 VM_BUG_ON(cluster_count(ci) != 0);
460 * If the swap is discardable, prepare discard the cluster
461 * instead of free it immediately. The cluster will be freed
464 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
465 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
466 swap_cluster_schedule_discard(si, idx);
470 __free_cluster(si, idx);
474 * The cluster corresponding to page_nr will be used. The cluster will be
475 * removed from free cluster list and its usage counter will be increased.
477 static void inc_cluster_info_page(struct swap_info_struct *p,
478 struct swap_cluster_info *cluster_info, unsigned long page_nr)
480 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
484 if (cluster_is_free(&cluster_info[idx]))
485 alloc_cluster(p, idx);
487 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
488 cluster_set_count(&cluster_info[idx],
489 cluster_count(&cluster_info[idx]) + 1);
493 * The cluster corresponding to page_nr decreases one usage. If the usage
494 * counter becomes 0, which means no page in the cluster is in using, we can
495 * optionally discard the cluster and add it to free cluster list.
497 static void dec_cluster_info_page(struct swap_info_struct *p,
498 struct swap_cluster_info *cluster_info, unsigned long page_nr)
500 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
505 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
506 cluster_set_count(&cluster_info[idx],
507 cluster_count(&cluster_info[idx]) - 1);
509 if (cluster_count(&cluster_info[idx]) == 0)
510 free_cluster(p, idx);
514 * It's possible scan_swap_map() uses a free cluster in the middle of free
515 * cluster list. Avoiding such abuse to avoid list corruption.
518 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
519 unsigned long offset)
521 struct percpu_cluster *percpu_cluster;
524 offset /= SWAPFILE_CLUSTER;
525 conflict = !cluster_list_empty(&si->free_clusters) &&
526 offset != cluster_list_first(&si->free_clusters) &&
527 cluster_is_free(&si->cluster_info[offset]);
532 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
533 cluster_set_null(&percpu_cluster->index);
538 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
539 * might involve allocating a new cluster for current CPU too.
541 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
542 unsigned long *offset, unsigned long *scan_base)
544 struct percpu_cluster *cluster;
545 struct swap_cluster_info *ci;
547 unsigned long tmp, max;
550 cluster = this_cpu_ptr(si->percpu_cluster);
551 if (cluster_is_null(&cluster->index)) {
552 if (!cluster_list_empty(&si->free_clusters)) {
553 cluster->index = si->free_clusters.head;
554 cluster->next = cluster_next(&cluster->index) *
556 } else if (!cluster_list_empty(&si->discard_clusters)) {
558 * we don't have free cluster but have some clusters in
559 * discarding, do discard now and reclaim them
561 swap_do_scheduled_discard(si);
562 *scan_base = *offset = si->cluster_next;
571 * Other CPUs can use our cluster if they can't find a free cluster,
572 * check if there is still free entry in the cluster
575 max = min_t(unsigned long, si->max,
576 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
578 cluster_set_null(&cluster->index);
581 ci = lock_cluster(si, tmp);
583 if (!si->swap_map[tmp]) {
591 cluster_set_null(&cluster->index);
594 cluster->next = tmp + 1;
600 static void __del_from_avail_list(struct swap_info_struct *p)
605 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
608 static void del_from_avail_list(struct swap_info_struct *p)
610 spin_lock(&swap_avail_lock);
611 __del_from_avail_list(p);
612 spin_unlock(&swap_avail_lock);
615 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
616 unsigned int nr_entries)
618 unsigned int end = offset + nr_entries - 1;
620 if (offset == si->lowest_bit)
621 si->lowest_bit += nr_entries;
622 if (end == si->highest_bit)
623 si->highest_bit -= nr_entries;
624 si->inuse_pages += nr_entries;
625 if (si->inuse_pages == si->pages) {
626 si->lowest_bit = si->max;
628 del_from_avail_list(si);
632 static void add_to_avail_list(struct swap_info_struct *p)
636 spin_lock(&swap_avail_lock);
638 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
639 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
641 spin_unlock(&swap_avail_lock);
644 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
645 unsigned int nr_entries)
647 unsigned long end = offset + nr_entries - 1;
648 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
650 if (offset < si->lowest_bit)
651 si->lowest_bit = offset;
652 if (end > si->highest_bit) {
653 bool was_full = !si->highest_bit;
655 si->highest_bit = end;
656 if (was_full && (si->flags & SWP_WRITEOK))
657 add_to_avail_list(si);
659 atomic_long_add(nr_entries, &nr_swap_pages);
660 si->inuse_pages -= nr_entries;
661 if (si->flags & SWP_BLKDEV)
662 swap_slot_free_notify =
663 si->bdev->bd_disk->fops->swap_slot_free_notify;
665 swap_slot_free_notify = NULL;
666 while (offset <= end) {
667 frontswap_invalidate_page(si->type, offset);
668 if (swap_slot_free_notify)
669 swap_slot_free_notify(si->bdev, offset);
674 static int scan_swap_map_slots(struct swap_info_struct *si,
675 unsigned char usage, int nr,
678 struct swap_cluster_info *ci;
679 unsigned long offset;
680 unsigned long scan_base;
681 unsigned long last_in_cluster = 0;
682 int latency_ration = LATENCY_LIMIT;
689 * We try to cluster swap pages by allocating them sequentially
690 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
691 * way, however, we resort to first-free allocation, starting
692 * a new cluster. This prevents us from scattering swap pages
693 * all over the entire swap partition, so that we reduce
694 * overall disk seek times between swap pages. -- sct
695 * But we do now try to find an empty cluster. -Andrea
696 * And we let swap pages go all over an SSD partition. Hugh
699 si->flags += SWP_SCANNING;
700 scan_base = offset = si->cluster_next;
703 if (si->cluster_info) {
704 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
710 if (unlikely(!si->cluster_nr--)) {
711 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
712 si->cluster_nr = SWAPFILE_CLUSTER - 1;
716 spin_unlock(&si->lock);
719 * If seek is expensive, start searching for new cluster from
720 * start of partition, to minimize the span of allocated swap.
721 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
722 * case, just handled by scan_swap_map_try_ssd_cluster() above.
724 scan_base = offset = si->lowest_bit;
725 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
727 /* Locate the first empty (unaligned) cluster */
728 for (; last_in_cluster <= si->highest_bit; offset++) {
729 if (si->swap_map[offset])
730 last_in_cluster = offset + SWAPFILE_CLUSTER;
731 else if (offset == last_in_cluster) {
732 spin_lock(&si->lock);
733 offset -= SWAPFILE_CLUSTER - 1;
734 si->cluster_next = offset;
735 si->cluster_nr = SWAPFILE_CLUSTER - 1;
738 if (unlikely(--latency_ration < 0)) {
740 latency_ration = LATENCY_LIMIT;
745 spin_lock(&si->lock);
746 si->cluster_nr = SWAPFILE_CLUSTER - 1;
750 if (si->cluster_info) {
751 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
752 /* take a break if we already got some slots */
755 if (!scan_swap_map_try_ssd_cluster(si, &offset,
760 if (!(si->flags & SWP_WRITEOK))
762 if (!si->highest_bit)
764 if (offset > si->highest_bit)
765 scan_base = offset = si->lowest_bit;
767 ci = lock_cluster(si, offset);
768 /* reuse swap entry of cache-only swap if not busy. */
769 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
772 spin_unlock(&si->lock);
773 swap_was_freed = __try_to_reclaim_swap(si, offset);
774 spin_lock(&si->lock);
775 /* entry was freed successfully, try to use this again */
778 goto scan; /* check next one */
781 if (si->swap_map[offset]) {
788 si->swap_map[offset] = usage;
789 inc_cluster_info_page(si, si->cluster_info, offset);
792 swap_range_alloc(si, offset, 1);
793 si->cluster_next = offset + 1;
794 slots[n_ret++] = swp_entry(si->type, offset);
796 /* got enough slots or reach max slots? */
797 if ((n_ret == nr) || (offset >= si->highest_bit))
800 /* search for next available slot */
802 /* time to take a break? */
803 if (unlikely(--latency_ration < 0)) {
806 spin_unlock(&si->lock);
808 spin_lock(&si->lock);
809 latency_ration = LATENCY_LIMIT;
812 /* try to get more slots in cluster */
813 if (si->cluster_info) {
814 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
822 /* non-ssd case, still more slots in cluster? */
823 if (si->cluster_nr && !si->swap_map[offset]) {
829 si->flags -= SWP_SCANNING;
833 spin_unlock(&si->lock);
834 while (++offset <= si->highest_bit) {
835 if (!si->swap_map[offset]) {
836 spin_lock(&si->lock);
839 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
840 spin_lock(&si->lock);
843 if (unlikely(--latency_ration < 0)) {
845 latency_ration = LATENCY_LIMIT;
848 offset = si->lowest_bit;
849 while (offset < scan_base) {
850 if (!si->swap_map[offset]) {
851 spin_lock(&si->lock);
854 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
855 spin_lock(&si->lock);
858 if (unlikely(--latency_ration < 0)) {
860 latency_ration = LATENCY_LIMIT;
864 spin_lock(&si->lock);
867 si->flags -= SWP_SCANNING;
871 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
874 struct swap_cluster_info *ci;
875 unsigned long offset, i;
879 * Should not even be attempting cluster allocations when huge
880 * page swap is disabled. Warn and fail the allocation.
882 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
887 if (cluster_list_empty(&si->free_clusters))
890 idx = cluster_list_first(&si->free_clusters);
891 offset = idx * SWAPFILE_CLUSTER;
892 ci = lock_cluster(si, offset);
893 alloc_cluster(si, idx);
894 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
896 map = si->swap_map + offset;
897 for (i = 0; i < SWAPFILE_CLUSTER; i++)
898 map[i] = SWAP_HAS_CACHE;
900 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
901 *slot = swp_entry(si->type, offset);
906 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
908 unsigned long offset = idx * SWAPFILE_CLUSTER;
909 struct swap_cluster_info *ci;
911 ci = lock_cluster(si, offset);
912 cluster_set_count_flag(ci, 0, 0);
913 free_cluster(si, idx);
915 swap_range_free(si, offset, SWAPFILE_CLUSTER);
918 static unsigned long scan_swap_map(struct swap_info_struct *si,
924 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
927 return swp_offset(entry);
933 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
935 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
936 struct swap_info_struct *si, *next;
941 /* Only single cluster request supported */
942 WARN_ON_ONCE(n_goal > 1 && cluster);
944 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
948 if (n_goal > SWAP_BATCH)
951 if (n_goal > avail_pgs)
954 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
956 spin_lock(&swap_avail_lock);
959 node = numa_node_id();
960 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
961 /* requeue si to after same-priority siblings */
962 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
963 spin_unlock(&swap_avail_lock);
964 spin_lock(&si->lock);
965 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
966 spin_lock(&swap_avail_lock);
967 if (plist_node_empty(&si->avail_lists[node])) {
968 spin_unlock(&si->lock);
971 WARN(!si->highest_bit,
972 "swap_info %d in list but !highest_bit\n",
974 WARN(!(si->flags & SWP_WRITEOK),
975 "swap_info %d in list but !SWP_WRITEOK\n",
977 __del_from_avail_list(si);
978 spin_unlock(&si->lock);
982 if (!(si->flags & SWP_FILE))
983 n_ret = swap_alloc_cluster(si, swp_entries);
985 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
986 n_goal, swp_entries);
987 spin_unlock(&si->lock);
988 if (n_ret || cluster)
990 pr_debug("scan_swap_map of si %d failed to find offset\n",
993 spin_lock(&swap_avail_lock);
996 * if we got here, it's likely that si was almost full before,
997 * and since scan_swap_map() can drop the si->lock, multiple
998 * callers probably all tried to get a page from the same si
999 * and it filled up before we could get one; or, the si filled
1000 * up between us dropping swap_avail_lock and taking si->lock.
1001 * Since we dropped the swap_avail_lock, the swap_avail_head
1002 * list may have been modified; so if next is still in the
1003 * swap_avail_head list then try it, otherwise start over
1004 * if we have not gotten any slots.
1006 if (plist_node_empty(&next->avail_lists[node]))
1010 spin_unlock(&swap_avail_lock);
1014 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1020 /* The only caller of this function is now suspend routine */
1021 swp_entry_t get_swap_page_of_type(int type)
1023 struct swap_info_struct *si;
1026 si = swap_info[type];
1027 spin_lock(&si->lock);
1028 if (si && (si->flags & SWP_WRITEOK)) {
1029 atomic_long_dec(&nr_swap_pages);
1030 /* This is called for allocating swap entry, not cache */
1031 offset = scan_swap_map(si, 1);
1033 spin_unlock(&si->lock);
1034 return swp_entry(type, offset);
1036 atomic_long_inc(&nr_swap_pages);
1038 spin_unlock(&si->lock);
1039 return (swp_entry_t) {0};
1042 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1044 struct swap_info_struct *p;
1045 unsigned long offset, type;
1049 type = swp_type(entry);
1050 if (type >= nr_swapfiles)
1052 p = swap_info[type];
1053 if (!(p->flags & SWP_USED))
1055 offset = swp_offset(entry);
1056 if (offset >= p->max)
1061 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1064 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1067 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1072 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1074 struct swap_info_struct *p;
1076 p = __swap_info_get(entry);
1079 if (!p->swap_map[swp_offset(entry)])
1084 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1090 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1092 struct swap_info_struct *p;
1094 p = _swap_info_get(entry);
1096 spin_lock(&p->lock);
1100 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1101 struct swap_info_struct *q)
1103 struct swap_info_struct *p;
1105 p = _swap_info_get(entry);
1109 spin_unlock(&q->lock);
1111 spin_lock(&p->lock);
1116 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1117 swp_entry_t entry, unsigned char usage)
1119 struct swap_cluster_info *ci;
1120 unsigned long offset = swp_offset(entry);
1121 unsigned char count;
1122 unsigned char has_cache;
1124 ci = lock_cluster_or_swap_info(p, offset);
1126 count = p->swap_map[offset];
1128 has_cache = count & SWAP_HAS_CACHE;
1129 count &= ~SWAP_HAS_CACHE;
1131 if (usage == SWAP_HAS_CACHE) {
1132 VM_BUG_ON(!has_cache);
1134 } else if (count == SWAP_MAP_SHMEM) {
1136 * Or we could insist on shmem.c using a special
1137 * swap_shmem_free() and free_shmem_swap_and_cache()...
1140 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1141 if (count == COUNT_CONTINUED) {
1142 if (swap_count_continued(p, offset, count))
1143 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1145 count = SWAP_MAP_MAX;
1150 usage = count | has_cache;
1151 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1153 unlock_cluster_or_swap_info(p, ci);
1158 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1160 struct swap_cluster_info *ci;
1161 unsigned long offset = swp_offset(entry);
1162 unsigned char count;
1164 ci = lock_cluster(p, offset);
1165 count = p->swap_map[offset];
1166 VM_BUG_ON(count != SWAP_HAS_CACHE);
1167 p->swap_map[offset] = 0;
1168 dec_cluster_info_page(p, p->cluster_info, offset);
1171 mem_cgroup_uncharge_swap(entry, 1);
1172 swap_range_free(p, offset, 1);
1176 * Caller has made sure that the swap device corresponding to entry
1177 * is still around or has not been recycled.
1179 void swap_free(swp_entry_t entry)
1181 struct swap_info_struct *p;
1183 p = _swap_info_get(entry);
1185 if (!__swap_entry_free(p, entry, 1))
1186 free_swap_slot(entry);
1191 * Called after dropping swapcache to decrease refcnt to swap entries.
1193 static void swapcache_free(swp_entry_t entry)
1195 struct swap_info_struct *p;
1197 p = _swap_info_get(entry);
1199 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1200 free_swap_slot(entry);
1204 static void swapcache_free_cluster(swp_entry_t entry)
1206 unsigned long offset = swp_offset(entry);
1207 unsigned long idx = offset / SWAPFILE_CLUSTER;
1208 struct swap_cluster_info *ci;
1209 struct swap_info_struct *si;
1211 unsigned int i, free_entries = 0;
1214 if (!IS_ENABLED(CONFIG_THP_SWAP))
1217 si = _swap_info_get(entry);
1221 ci = lock_cluster(si, offset);
1222 VM_BUG_ON(!cluster_is_huge(ci));
1223 map = si->swap_map + offset;
1224 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1226 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1227 if (val == SWAP_HAS_CACHE)
1230 if (!free_entries) {
1231 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1232 map[i] &= ~SWAP_HAS_CACHE;
1234 cluster_clear_huge(ci);
1236 if (free_entries == SWAPFILE_CLUSTER) {
1237 spin_lock(&si->lock);
1238 ci = lock_cluster(si, offset);
1239 memset(map, 0, SWAPFILE_CLUSTER);
1241 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1242 swap_free_cluster(si, idx);
1243 spin_unlock(&si->lock);
1244 } else if (free_entries) {
1245 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1246 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1247 free_swap_slot(entry);
1252 #ifdef CONFIG_THP_SWAP
1253 int split_swap_cluster(swp_entry_t entry)
1255 struct swap_info_struct *si;
1256 struct swap_cluster_info *ci;
1257 unsigned long offset = swp_offset(entry);
1259 si = _swap_info_get(entry);
1262 ci = lock_cluster(si, offset);
1263 cluster_clear_huge(ci);
1269 void put_swap_page(struct page *page, swp_entry_t entry)
1271 if (!PageTransHuge(page))
1272 swapcache_free(entry);
1274 swapcache_free_cluster(entry);
1277 static int swp_entry_cmp(const void *ent1, const void *ent2)
1279 const swp_entry_t *e1 = ent1, *e2 = ent2;
1281 return (int)swp_type(*e1) - (int)swp_type(*e2);
1284 void swapcache_free_entries(swp_entry_t *entries, int n)
1286 struct swap_info_struct *p, *prev;
1296 * Sort swap entries by swap device, so each lock is only taken once.
1297 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1298 * so low that it isn't necessary to optimize further.
1300 if (nr_swapfiles > 1)
1301 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1302 for (i = 0; i < n; ++i) {
1303 p = swap_info_get_cont(entries[i], prev);
1305 swap_entry_free(p, entries[i]);
1309 spin_unlock(&p->lock);
1313 * How many references to page are currently swapped out?
1314 * This does not give an exact answer when swap count is continued,
1315 * but does include the high COUNT_CONTINUED flag to allow for that.
1317 int page_swapcount(struct page *page)
1320 struct swap_info_struct *p;
1321 struct swap_cluster_info *ci;
1323 unsigned long offset;
1325 entry.val = page_private(page);
1326 p = _swap_info_get(entry);
1328 offset = swp_offset(entry);
1329 ci = lock_cluster_or_swap_info(p, offset);
1330 count = swap_count(p->swap_map[offset]);
1331 unlock_cluster_or_swap_info(p, ci);
1336 int __swap_count(struct swap_info_struct *si, swp_entry_t entry)
1338 pgoff_t offset = swp_offset(entry);
1340 return swap_count(si->swap_map[offset]);
1343 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1346 pgoff_t offset = swp_offset(entry);
1347 struct swap_cluster_info *ci;
1349 ci = lock_cluster_or_swap_info(si, offset);
1350 count = swap_count(si->swap_map[offset]);
1351 unlock_cluster_or_swap_info(si, ci);
1356 * How many references to @entry are currently swapped out?
1357 * This does not give an exact answer when swap count is continued,
1358 * but does include the high COUNT_CONTINUED flag to allow for that.
1360 int __swp_swapcount(swp_entry_t entry)
1363 struct swap_info_struct *si;
1365 si = __swap_info_get(entry);
1367 count = swap_swapcount(si, entry);
1372 * How many references to @entry are currently swapped out?
1373 * This considers COUNT_CONTINUED so it returns exact answer.
1375 int swp_swapcount(swp_entry_t entry)
1377 int count, tmp_count, n;
1378 struct swap_info_struct *p;
1379 struct swap_cluster_info *ci;
1384 p = _swap_info_get(entry);
1388 offset = swp_offset(entry);
1390 ci = lock_cluster_or_swap_info(p, offset);
1392 count = swap_count(p->swap_map[offset]);
1393 if (!(count & COUNT_CONTINUED))
1396 count &= ~COUNT_CONTINUED;
1397 n = SWAP_MAP_MAX + 1;
1399 page = vmalloc_to_page(p->swap_map + offset);
1400 offset &= ~PAGE_MASK;
1401 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1404 page = list_next_entry(page, lru);
1405 map = kmap_atomic(page);
1406 tmp_count = map[offset];
1409 count += (tmp_count & ~COUNT_CONTINUED) * n;
1410 n *= (SWAP_CONT_MAX + 1);
1411 } while (tmp_count & COUNT_CONTINUED);
1413 unlock_cluster_or_swap_info(p, ci);
1417 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1420 struct swap_cluster_info *ci;
1421 unsigned char *map = si->swap_map;
1422 unsigned long roffset = swp_offset(entry);
1423 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1427 if (!IS_ENABLED(CONFIG_THP_SWAP))
1428 return swap_swapcount(si, entry) != 0;
1430 ci = lock_cluster_or_swap_info(si, offset);
1431 if (!ci || !cluster_is_huge(ci)) {
1432 if (map[roffset] != SWAP_HAS_CACHE)
1436 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1437 if (map[offset + i] != SWAP_HAS_CACHE) {
1443 unlock_cluster_or_swap_info(si, ci);
1447 static bool page_swapped(struct page *page)
1450 struct swap_info_struct *si;
1452 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1453 return page_swapcount(page) != 0;
1455 page = compound_head(page);
1456 entry.val = page_private(page);
1457 si = _swap_info_get(entry);
1459 return swap_page_trans_huge_swapped(si, entry);
1463 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1464 int *total_swapcount)
1466 int i, map_swapcount, _total_mapcount, _total_swapcount;
1467 unsigned long offset = 0;
1468 struct swap_info_struct *si;
1469 struct swap_cluster_info *ci = NULL;
1470 unsigned char *map = NULL;
1471 int mapcount, swapcount = 0;
1473 /* hugetlbfs shouldn't call it */
1474 VM_BUG_ON_PAGE(PageHuge(page), page);
1476 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1477 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1478 if (PageSwapCache(page))
1479 swapcount = page_swapcount(page);
1480 if (total_swapcount)
1481 *total_swapcount = swapcount;
1482 return mapcount + swapcount;
1485 page = compound_head(page);
1487 _total_mapcount = _total_swapcount = map_swapcount = 0;
1488 if (PageSwapCache(page)) {
1491 entry.val = page_private(page);
1492 si = _swap_info_get(entry);
1495 offset = swp_offset(entry);
1499 ci = lock_cluster(si, offset);
1500 for (i = 0; i < HPAGE_PMD_NR; i++) {
1501 mapcount = atomic_read(&page[i]._mapcount) + 1;
1502 _total_mapcount += mapcount;
1504 swapcount = swap_count(map[offset + i]);
1505 _total_swapcount += swapcount;
1507 map_swapcount = max(map_swapcount, mapcount + swapcount);
1510 if (PageDoubleMap(page)) {
1512 _total_mapcount -= HPAGE_PMD_NR;
1514 mapcount = compound_mapcount(page);
1515 map_swapcount += mapcount;
1516 _total_mapcount += mapcount;
1518 *total_mapcount = _total_mapcount;
1519 if (total_swapcount)
1520 *total_swapcount = _total_swapcount;
1522 return map_swapcount;
1526 * We can write to an anon page without COW if there are no other references
1527 * to it. And as a side-effect, free up its swap: because the old content
1528 * on disk will never be read, and seeking back there to write new content
1529 * later would only waste time away from clustering.
1531 * NOTE: total_map_swapcount should not be relied upon by the caller if
1532 * reuse_swap_page() returns false, but it may be always overwritten
1533 * (see the other implementation for CONFIG_SWAP=n).
1535 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1537 int count, total_mapcount, total_swapcount;
1539 VM_BUG_ON_PAGE(!PageLocked(page), page);
1540 if (unlikely(PageKsm(page)))
1542 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1544 if (total_map_swapcount)
1545 *total_map_swapcount = total_mapcount + total_swapcount;
1546 if (count == 1 && PageSwapCache(page) &&
1547 (likely(!PageTransCompound(page)) ||
1548 /* The remaining swap count will be freed soon */
1549 total_swapcount == page_swapcount(page))) {
1550 if (!PageWriteback(page)) {
1551 page = compound_head(page);
1552 delete_from_swap_cache(page);
1556 struct swap_info_struct *p;
1558 entry.val = page_private(page);
1559 p = swap_info_get(entry);
1560 if (p->flags & SWP_STABLE_WRITES) {
1561 spin_unlock(&p->lock);
1564 spin_unlock(&p->lock);
1572 * If swap is getting full, or if there are no more mappings of this page,
1573 * then try_to_free_swap is called to free its swap space.
1575 int try_to_free_swap(struct page *page)
1577 VM_BUG_ON_PAGE(!PageLocked(page), page);
1579 if (!PageSwapCache(page))
1581 if (PageWriteback(page))
1583 if (page_swapped(page))
1587 * Once hibernation has begun to create its image of memory,
1588 * there's a danger that one of the calls to try_to_free_swap()
1589 * - most probably a call from __try_to_reclaim_swap() while
1590 * hibernation is allocating its own swap pages for the image,
1591 * but conceivably even a call from memory reclaim - will free
1592 * the swap from a page which has already been recorded in the
1593 * image as a clean swapcache page, and then reuse its swap for
1594 * another page of the image. On waking from hibernation, the
1595 * original page might be freed under memory pressure, then
1596 * later read back in from swap, now with the wrong data.
1598 * Hibernation suspends storage while it is writing the image
1599 * to disk so check that here.
1601 if (pm_suspended_storage())
1604 page = compound_head(page);
1605 delete_from_swap_cache(page);
1611 * Free the swap entry like above, but also try to
1612 * free the page cache entry if it is the last user.
1614 int free_swap_and_cache(swp_entry_t entry)
1616 struct swap_info_struct *p;
1617 struct page *page = NULL;
1618 unsigned char count;
1620 if (non_swap_entry(entry))
1623 p = _swap_info_get(entry);
1625 count = __swap_entry_free(p, entry, 1);
1626 if (count == SWAP_HAS_CACHE &&
1627 !swap_page_trans_huge_swapped(p, entry)) {
1628 page = find_get_page(swap_address_space(entry),
1630 if (page && !trylock_page(page)) {
1635 free_swap_slot(entry);
1639 * Not mapped elsewhere, or swap space full? Free it!
1640 * Also recheck PageSwapCache now page is locked (above).
1642 if (PageSwapCache(page) && !PageWriteback(page) &&
1643 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1644 !swap_page_trans_huge_swapped(p, entry)) {
1645 page = compound_head(page);
1646 delete_from_swap_cache(page);
1655 #ifdef CONFIG_HIBERNATION
1657 * Find the swap type that corresponds to given device (if any).
1659 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1660 * from 0, in which the swap header is expected to be located.
1662 * This is needed for the suspend to disk (aka swsusp).
1664 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1666 struct block_device *bdev = NULL;
1670 bdev = bdget(device);
1672 spin_lock(&swap_lock);
1673 for (type = 0; type < nr_swapfiles; type++) {
1674 struct swap_info_struct *sis = swap_info[type];
1676 if (!(sis->flags & SWP_WRITEOK))
1681 *bdev_p = bdgrab(sis->bdev);
1683 spin_unlock(&swap_lock);
1686 if (bdev == sis->bdev) {
1687 struct swap_extent *se = &sis->first_swap_extent;
1689 if (se->start_block == offset) {
1691 *bdev_p = bdgrab(sis->bdev);
1693 spin_unlock(&swap_lock);
1699 spin_unlock(&swap_lock);
1707 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1708 * corresponding to given index in swap_info (swap type).
1710 sector_t swapdev_block(int type, pgoff_t offset)
1712 struct block_device *bdev;
1714 if ((unsigned int)type >= nr_swapfiles)
1716 if (!(swap_info[type]->flags & SWP_WRITEOK))
1718 return map_swap_entry(swp_entry(type, offset), &bdev);
1722 * Return either the total number of swap pages of given type, or the number
1723 * of free pages of that type (depending on @free)
1725 * This is needed for software suspend
1727 unsigned int count_swap_pages(int type, int free)
1731 spin_lock(&swap_lock);
1732 if ((unsigned int)type < nr_swapfiles) {
1733 struct swap_info_struct *sis = swap_info[type];
1735 spin_lock(&sis->lock);
1736 if (sis->flags & SWP_WRITEOK) {
1739 n -= sis->inuse_pages;
1741 spin_unlock(&sis->lock);
1743 spin_unlock(&swap_lock);
1746 #endif /* CONFIG_HIBERNATION */
1748 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1750 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1754 * No need to decide whether this PTE shares the swap entry with others,
1755 * just let do_wp_page work it out if a write is requested later - to
1756 * force COW, vm_page_prot omits write permission from any private vma.
1758 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1759 unsigned long addr, swp_entry_t entry, struct page *page)
1761 struct page *swapcache;
1762 struct mem_cgroup *memcg;
1768 page = ksm_might_need_to_copy(page, vma, addr);
1769 if (unlikely(!page))
1772 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1778 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1779 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1780 mem_cgroup_cancel_charge(page, memcg, false);
1785 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1786 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1788 set_pte_at(vma->vm_mm, addr, pte,
1789 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1790 if (page == swapcache) {
1791 page_add_anon_rmap(page, vma, addr, false);
1792 mem_cgroup_commit_charge(page, memcg, true, false);
1793 } else { /* ksm created a completely new copy */
1794 page_add_new_anon_rmap(page, vma, addr, false);
1795 mem_cgroup_commit_charge(page, memcg, false, false);
1796 lru_cache_add_active_or_unevictable(page, vma);
1800 * Move the page to the active list so it is not
1801 * immediately swapped out again after swapon.
1803 activate_page(page);
1805 pte_unmap_unlock(pte, ptl);
1807 if (page != swapcache) {
1814 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1815 unsigned long addr, unsigned long end,
1816 swp_entry_t entry, struct page *page)
1818 pte_t swp_pte = swp_entry_to_pte(entry);
1823 * We don't actually need pte lock while scanning for swp_pte: since
1824 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1825 * page table while we're scanning; though it could get zapped, and on
1826 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1827 * of unmatched parts which look like swp_pte, so unuse_pte must
1828 * recheck under pte lock. Scanning without pte lock lets it be
1829 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1831 pte = pte_offset_map(pmd, addr);
1834 * swapoff spends a _lot_ of time in this loop!
1835 * Test inline before going to call unuse_pte.
1837 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1839 ret = unuse_pte(vma, pmd, addr, entry, page);
1842 pte = pte_offset_map(pmd, addr);
1844 } while (pte++, addr += PAGE_SIZE, addr != end);
1850 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1851 unsigned long addr, unsigned long end,
1852 swp_entry_t entry, struct page *page)
1858 pmd = pmd_offset(pud, addr);
1861 next = pmd_addr_end(addr, end);
1862 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1864 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1867 } while (pmd++, addr = next, addr != end);
1871 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1872 unsigned long addr, unsigned long end,
1873 swp_entry_t entry, struct page *page)
1879 pud = pud_offset(p4d, addr);
1881 next = pud_addr_end(addr, end);
1882 if (pud_none_or_clear_bad(pud))
1884 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1887 } while (pud++, addr = next, addr != end);
1891 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1892 unsigned long addr, unsigned long end,
1893 swp_entry_t entry, struct page *page)
1899 p4d = p4d_offset(pgd, addr);
1901 next = p4d_addr_end(addr, end);
1902 if (p4d_none_or_clear_bad(p4d))
1904 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1907 } while (p4d++, addr = next, addr != end);
1911 static int unuse_vma(struct vm_area_struct *vma,
1912 swp_entry_t entry, struct page *page)
1915 unsigned long addr, end, next;
1918 if (page_anon_vma(page)) {
1919 addr = page_address_in_vma(page, vma);
1920 if (addr == -EFAULT)
1923 end = addr + PAGE_SIZE;
1925 addr = vma->vm_start;
1929 pgd = pgd_offset(vma->vm_mm, addr);
1931 next = pgd_addr_end(addr, end);
1932 if (pgd_none_or_clear_bad(pgd))
1934 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1937 } while (pgd++, addr = next, addr != end);
1941 static int unuse_mm(struct mm_struct *mm,
1942 swp_entry_t entry, struct page *page)
1944 struct vm_area_struct *vma;
1947 if (!down_read_trylock(&mm->mmap_sem)) {
1949 * Activate page so shrink_inactive_list is unlikely to unmap
1950 * its ptes while lock is dropped, so swapoff can make progress.
1952 activate_page(page);
1954 down_read(&mm->mmap_sem);
1957 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1958 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1962 up_read(&mm->mmap_sem);
1963 return (ret < 0)? ret: 0;
1967 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1968 * from current position to next entry still in use.
1969 * Recycle to start on reaching the end, returning 0 when empty.
1971 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1972 unsigned int prev, bool frontswap)
1974 unsigned int max = si->max;
1975 unsigned int i = prev;
1976 unsigned char count;
1979 * No need for swap_lock here: we're just looking
1980 * for whether an entry is in use, not modifying it; false
1981 * hits are okay, and sys_swapoff() has already prevented new
1982 * allocations from this area (while holding swap_lock).
1991 * No entries in use at top of swap_map,
1992 * loop back to start and recheck there.
1998 count = READ_ONCE(si->swap_map[i]);
1999 if (count && swap_count(count) != SWAP_MAP_BAD)
2000 if (!frontswap || frontswap_test(si, i))
2002 if ((i % LATENCY_LIMIT) == 0)
2009 * We completely avoid races by reading each swap page in advance,
2010 * and then search for the process using it. All the necessary
2011 * page table adjustments can then be made atomically.
2013 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2014 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2016 int try_to_unuse(unsigned int type, bool frontswap,
2017 unsigned long pages_to_unuse)
2019 struct swap_info_struct *si = swap_info[type];
2020 struct mm_struct *start_mm;
2021 volatile unsigned char *swap_map; /* swap_map is accessed without
2022 * locking. Mark it as volatile
2023 * to prevent compiler doing
2026 unsigned char swcount;
2033 * When searching mms for an entry, a good strategy is to
2034 * start at the first mm we freed the previous entry from
2035 * (though actually we don't notice whether we or coincidence
2036 * freed the entry). Initialize this start_mm with a hold.
2038 * A simpler strategy would be to start at the last mm we
2039 * freed the previous entry from; but that would take less
2040 * advantage of mmlist ordering, which clusters forked mms
2041 * together, child after parent. If we race with dup_mmap(), we
2042 * prefer to resolve parent before child, lest we miss entries
2043 * duplicated after we scanned child: using last mm would invert
2046 start_mm = &init_mm;
2050 * Keep on scanning until all entries have gone. Usually,
2051 * one pass through swap_map is enough, but not necessarily:
2052 * there are races when an instance of an entry might be missed.
2054 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2055 if (signal_pending(current)) {
2061 * Get a page for the entry, using the existing swap
2062 * cache page if there is one. Otherwise, get a clean
2063 * page and read the swap into it.
2065 swap_map = &si->swap_map[i];
2066 entry = swp_entry(type, i);
2067 page = read_swap_cache_async(entry,
2068 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2071 * Either swap_duplicate() failed because entry
2072 * has been freed independently, and will not be
2073 * reused since sys_swapoff() already disabled
2074 * allocation from here, or alloc_page() failed.
2076 swcount = *swap_map;
2078 * We don't hold lock here, so the swap entry could be
2079 * SWAP_MAP_BAD (when the cluster is discarding).
2080 * Instead of fail out, We can just skip the swap
2081 * entry because swapoff will wait for discarding
2084 if (!swcount || swcount == SWAP_MAP_BAD)
2091 * Don't hold on to start_mm if it looks like exiting.
2093 if (atomic_read(&start_mm->mm_users) == 1) {
2095 start_mm = &init_mm;
2100 * Wait for and lock page. When do_swap_page races with
2101 * try_to_unuse, do_swap_page can handle the fault much
2102 * faster than try_to_unuse can locate the entry. This
2103 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2104 * defer to do_swap_page in such a case - in some tests,
2105 * do_swap_page and try_to_unuse repeatedly compete.
2107 wait_on_page_locked(page);
2108 wait_on_page_writeback(page);
2110 wait_on_page_writeback(page);
2113 * Remove all references to entry.
2115 swcount = *swap_map;
2116 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2117 retval = shmem_unuse(entry, page);
2118 /* page has already been unlocked and released */
2123 if (swap_count(swcount) && start_mm != &init_mm)
2124 retval = unuse_mm(start_mm, entry, page);
2126 if (swap_count(*swap_map)) {
2127 int set_start_mm = (*swap_map >= swcount);
2128 struct list_head *p = &start_mm->mmlist;
2129 struct mm_struct *new_start_mm = start_mm;
2130 struct mm_struct *prev_mm = start_mm;
2131 struct mm_struct *mm;
2133 mmget(new_start_mm);
2135 spin_lock(&mmlist_lock);
2136 while (swap_count(*swap_map) && !retval &&
2137 (p = p->next) != &start_mm->mmlist) {
2138 mm = list_entry(p, struct mm_struct, mmlist);
2139 if (!mmget_not_zero(mm))
2141 spin_unlock(&mmlist_lock);
2147 swcount = *swap_map;
2148 if (!swap_count(swcount)) /* any usage ? */
2150 else if (mm == &init_mm)
2153 retval = unuse_mm(mm, entry, page);
2155 if (set_start_mm && *swap_map < swcount) {
2156 mmput(new_start_mm);
2161 spin_lock(&mmlist_lock);
2163 spin_unlock(&mmlist_lock);
2166 start_mm = new_start_mm;
2175 * If a reference remains (rare), we would like to leave
2176 * the page in the swap cache; but try_to_unmap could
2177 * then re-duplicate the entry once we drop page lock,
2178 * so we might loop indefinitely; also, that page could
2179 * not be swapped out to other storage meanwhile. So:
2180 * delete from cache even if there's another reference,
2181 * after ensuring that the data has been saved to disk -
2182 * since if the reference remains (rarer), it will be
2183 * read from disk into another page. Splitting into two
2184 * pages would be incorrect if swap supported "shared
2185 * private" pages, but they are handled by tmpfs files.
2187 * Given how unuse_vma() targets one particular offset
2188 * in an anon_vma, once the anon_vma has been determined,
2189 * this splitting happens to be just what is needed to
2190 * handle where KSM pages have been swapped out: re-reading
2191 * is unnecessarily slow, but we can fix that later on.
2193 if (swap_count(*swap_map) &&
2194 PageDirty(page) && PageSwapCache(page)) {
2195 struct writeback_control wbc = {
2196 .sync_mode = WB_SYNC_NONE,
2199 swap_writepage(compound_head(page), &wbc);
2201 wait_on_page_writeback(page);
2205 * It is conceivable that a racing task removed this page from
2206 * swap cache just before we acquired the page lock at the top,
2207 * or while we dropped it in unuse_mm(). The page might even
2208 * be back in swap cache on another swap area: that we must not
2209 * delete, since it may not have been written out to swap yet.
2211 if (PageSwapCache(page) &&
2212 likely(page_private(page) == entry.val) &&
2213 !page_swapped(page))
2214 delete_from_swap_cache(compound_head(page));
2217 * So we could skip searching mms once swap count went
2218 * to 1, we did not mark any present ptes as dirty: must
2219 * mark page dirty so shrink_page_list will preserve it.
2226 * Make sure that we aren't completely killing
2227 * interactive performance.
2230 if (frontswap && pages_to_unuse > 0) {
2231 if (!--pages_to_unuse)
2241 * After a successful try_to_unuse, if no swap is now in use, we know
2242 * we can empty the mmlist. swap_lock must be held on entry and exit.
2243 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2244 * added to the mmlist just after page_duplicate - before would be racy.
2246 static void drain_mmlist(void)
2248 struct list_head *p, *next;
2251 for (type = 0; type < nr_swapfiles; type++)
2252 if (swap_info[type]->inuse_pages)
2254 spin_lock(&mmlist_lock);
2255 list_for_each_safe(p, next, &init_mm.mmlist)
2257 spin_unlock(&mmlist_lock);
2261 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2262 * corresponds to page offset for the specified swap entry.
2263 * Note that the type of this function is sector_t, but it returns page offset
2264 * into the bdev, not sector offset.
2266 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2268 struct swap_info_struct *sis;
2269 struct swap_extent *start_se;
2270 struct swap_extent *se;
2273 sis = swap_info[swp_type(entry)];
2276 offset = swp_offset(entry);
2277 start_se = sis->curr_swap_extent;
2281 if (se->start_page <= offset &&
2282 offset < (se->start_page + se->nr_pages)) {
2283 return se->start_block + (offset - se->start_page);
2285 se = list_next_entry(se, list);
2286 sis->curr_swap_extent = se;
2287 BUG_ON(se == start_se); /* It *must* be present */
2292 * Returns the page offset into bdev for the specified page's swap entry.
2294 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2297 entry.val = page_private(page);
2298 return map_swap_entry(entry, bdev);
2302 * Free all of a swapdev's extent information
2304 static void destroy_swap_extents(struct swap_info_struct *sis)
2306 while (!list_empty(&sis->first_swap_extent.list)) {
2307 struct swap_extent *se;
2309 se = list_first_entry(&sis->first_swap_extent.list,
2310 struct swap_extent, list);
2311 list_del(&se->list);
2315 if (sis->flags & SWP_FILE) {
2316 struct file *swap_file = sis->swap_file;
2317 struct address_space *mapping = swap_file->f_mapping;
2319 sis->flags &= ~SWP_FILE;
2320 mapping->a_ops->swap_deactivate(swap_file);
2325 * Add a block range (and the corresponding page range) into this swapdev's
2326 * extent list. The extent list is kept sorted in page order.
2328 * This function rather assumes that it is called in ascending page order.
2331 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2332 unsigned long nr_pages, sector_t start_block)
2334 struct swap_extent *se;
2335 struct swap_extent *new_se;
2336 struct list_head *lh;
2338 if (start_page == 0) {
2339 se = &sis->first_swap_extent;
2340 sis->curr_swap_extent = se;
2342 se->nr_pages = nr_pages;
2343 se->start_block = start_block;
2346 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2347 se = list_entry(lh, struct swap_extent, list);
2348 BUG_ON(se->start_page + se->nr_pages != start_page);
2349 if (se->start_block + se->nr_pages == start_block) {
2351 se->nr_pages += nr_pages;
2357 * No merge. Insert a new extent, preserving ordering.
2359 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2362 new_se->start_page = start_page;
2363 new_se->nr_pages = nr_pages;
2364 new_se->start_block = start_block;
2366 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2371 * A `swap extent' is a simple thing which maps a contiguous range of pages
2372 * onto a contiguous range of disk blocks. An ordered list of swap extents
2373 * is built at swapon time and is then used at swap_writepage/swap_readpage
2374 * time for locating where on disk a page belongs.
2376 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2377 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2378 * swap files identically.
2380 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2381 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2382 * swapfiles are handled *identically* after swapon time.
2384 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2385 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2386 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2387 * requirements, they are simply tossed out - we will never use those blocks
2390 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2391 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2392 * which will scribble on the fs.
2394 * The amount of disk space which a single swap extent represents varies.
2395 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2396 * extents in the list. To avoid much list walking, we cache the previous
2397 * search location in `curr_swap_extent', and start new searches from there.
2398 * This is extremely effective. The average number of iterations in
2399 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2401 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2403 struct file *swap_file = sis->swap_file;
2404 struct address_space *mapping = swap_file->f_mapping;
2405 struct inode *inode = mapping->host;
2408 if (S_ISBLK(inode->i_mode)) {
2409 ret = add_swap_extent(sis, 0, sis->max, 0);
2414 if (mapping->a_ops->swap_activate) {
2415 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2417 sis->flags |= SWP_FILE;
2418 ret = add_swap_extent(sis, 0, sis->max, 0);
2424 return generic_swapfile_activate(sis, swap_file, span);
2427 static int swap_node(struct swap_info_struct *p)
2429 struct block_device *bdev;
2434 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2436 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2439 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2440 unsigned char *swap_map,
2441 struct swap_cluster_info *cluster_info)
2448 p->prio = --least_priority;
2450 * the plist prio is negated because plist ordering is
2451 * low-to-high, while swap ordering is high-to-low
2453 p->list.prio = -p->prio;
2456 p->avail_lists[i].prio = -p->prio;
2458 if (swap_node(p) == i)
2459 p->avail_lists[i].prio = 1;
2461 p->avail_lists[i].prio = -p->prio;
2464 p->swap_map = swap_map;
2465 p->cluster_info = cluster_info;
2466 p->flags |= SWP_WRITEOK;
2467 atomic_long_add(p->pages, &nr_swap_pages);
2468 total_swap_pages += p->pages;
2470 assert_spin_locked(&swap_lock);
2472 * both lists are plists, and thus priority ordered.
2473 * swap_active_head needs to be priority ordered for swapoff(),
2474 * which on removal of any swap_info_struct with an auto-assigned
2475 * (i.e. negative) priority increments the auto-assigned priority
2476 * of any lower-priority swap_info_structs.
2477 * swap_avail_head needs to be priority ordered for get_swap_page(),
2478 * which allocates swap pages from the highest available priority
2481 plist_add(&p->list, &swap_active_head);
2482 add_to_avail_list(p);
2485 static void enable_swap_info(struct swap_info_struct *p, int prio,
2486 unsigned char *swap_map,
2487 struct swap_cluster_info *cluster_info,
2488 unsigned long *frontswap_map)
2490 frontswap_init(p->type, frontswap_map);
2491 spin_lock(&swap_lock);
2492 spin_lock(&p->lock);
2493 _enable_swap_info(p, prio, swap_map, cluster_info);
2494 spin_unlock(&p->lock);
2495 spin_unlock(&swap_lock);
2498 static void reinsert_swap_info(struct swap_info_struct *p)
2500 spin_lock(&swap_lock);
2501 spin_lock(&p->lock);
2502 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2503 spin_unlock(&p->lock);
2504 spin_unlock(&swap_lock);
2507 bool has_usable_swap(void)
2511 spin_lock(&swap_lock);
2512 if (plist_head_empty(&swap_active_head))
2514 spin_unlock(&swap_lock);
2518 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2520 struct swap_info_struct *p = NULL;
2521 unsigned char *swap_map;
2522 struct swap_cluster_info *cluster_info;
2523 unsigned long *frontswap_map;
2524 struct file *swap_file, *victim;
2525 struct address_space *mapping;
2526 struct inode *inode;
2527 struct filename *pathname;
2529 unsigned int old_block_size;
2531 if (!capable(CAP_SYS_ADMIN))
2534 BUG_ON(!current->mm);
2536 pathname = getname(specialfile);
2537 if (IS_ERR(pathname))
2538 return PTR_ERR(pathname);
2540 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2541 err = PTR_ERR(victim);
2545 mapping = victim->f_mapping;
2546 spin_lock(&swap_lock);
2547 plist_for_each_entry(p, &swap_active_head, list) {
2548 if (p->flags & SWP_WRITEOK) {
2549 if (p->swap_file->f_mapping == mapping) {
2557 spin_unlock(&swap_lock);
2560 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2561 vm_unacct_memory(p->pages);
2564 spin_unlock(&swap_lock);
2567 del_from_avail_list(p);
2568 spin_lock(&p->lock);
2570 struct swap_info_struct *si = p;
2573 plist_for_each_entry_continue(si, &swap_active_head, list) {
2576 for_each_node(nid) {
2577 if (si->avail_lists[nid].prio != 1)
2578 si->avail_lists[nid].prio--;
2583 plist_del(&p->list, &swap_active_head);
2584 atomic_long_sub(p->pages, &nr_swap_pages);
2585 total_swap_pages -= p->pages;
2586 p->flags &= ~SWP_WRITEOK;
2587 spin_unlock(&p->lock);
2588 spin_unlock(&swap_lock);
2590 disable_swap_slots_cache_lock();
2592 set_current_oom_origin();
2593 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2594 clear_current_oom_origin();
2597 /* re-insert swap space back into swap_list */
2598 reinsert_swap_info(p);
2599 reenable_swap_slots_cache_unlock();
2603 reenable_swap_slots_cache_unlock();
2605 flush_work(&p->discard_work);
2607 destroy_swap_extents(p);
2608 if (p->flags & SWP_CONTINUED)
2609 free_swap_count_continuations(p);
2611 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2612 atomic_dec(&nr_rotate_swap);
2614 mutex_lock(&swapon_mutex);
2615 spin_lock(&swap_lock);
2616 spin_lock(&p->lock);
2619 /* wait for anyone still in scan_swap_map */
2620 p->highest_bit = 0; /* cuts scans short */
2621 while (p->flags >= SWP_SCANNING) {
2622 spin_unlock(&p->lock);
2623 spin_unlock(&swap_lock);
2624 schedule_timeout_uninterruptible(1);
2625 spin_lock(&swap_lock);
2626 spin_lock(&p->lock);
2629 swap_file = p->swap_file;
2630 old_block_size = p->old_block_size;
2631 p->swap_file = NULL;
2633 swap_map = p->swap_map;
2635 cluster_info = p->cluster_info;
2636 p->cluster_info = NULL;
2637 frontswap_map = frontswap_map_get(p);
2638 spin_unlock(&p->lock);
2639 spin_unlock(&swap_lock);
2640 frontswap_invalidate_area(p->type);
2641 frontswap_map_set(p, NULL);
2642 mutex_unlock(&swapon_mutex);
2643 free_percpu(p->percpu_cluster);
2644 p->percpu_cluster = NULL;
2646 kvfree(cluster_info);
2647 kvfree(frontswap_map);
2648 /* Destroy swap account information */
2649 swap_cgroup_swapoff(p->type);
2650 exit_swap_address_space(p->type);
2652 inode = mapping->host;
2653 if (S_ISBLK(inode->i_mode)) {
2654 struct block_device *bdev = I_BDEV(inode);
2655 set_blocksize(bdev, old_block_size);
2656 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2659 inode->i_flags &= ~S_SWAPFILE;
2660 inode_unlock(inode);
2662 filp_close(swap_file, NULL);
2665 * Clear the SWP_USED flag after all resources are freed so that swapon
2666 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2667 * not hold p->lock after we cleared its SWP_WRITEOK.
2669 spin_lock(&swap_lock);
2671 spin_unlock(&swap_lock);
2674 atomic_inc(&proc_poll_event);
2675 wake_up_interruptible(&proc_poll_wait);
2678 filp_close(victim, NULL);
2684 #ifdef CONFIG_PROC_FS
2685 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2687 struct seq_file *seq = file->private_data;
2689 poll_wait(file, &proc_poll_wait, wait);
2691 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2692 seq->poll_event = atomic_read(&proc_poll_event);
2693 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2696 return EPOLLIN | EPOLLRDNORM;
2700 static void *swap_start(struct seq_file *swap, loff_t *pos)
2702 struct swap_info_struct *si;
2706 mutex_lock(&swapon_mutex);
2709 return SEQ_START_TOKEN;
2711 for (type = 0; type < nr_swapfiles; type++) {
2712 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2713 si = swap_info[type];
2714 if (!(si->flags & SWP_USED) || !si->swap_map)
2723 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2725 struct swap_info_struct *si = v;
2728 if (v == SEQ_START_TOKEN)
2731 type = si->type + 1;
2733 for (; type < nr_swapfiles; type++) {
2734 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2735 si = swap_info[type];
2736 if (!(si->flags & SWP_USED) || !si->swap_map)
2745 static void swap_stop(struct seq_file *swap, void *v)
2747 mutex_unlock(&swapon_mutex);
2750 static int swap_show(struct seq_file *swap, void *v)
2752 struct swap_info_struct *si = v;
2756 if (si == SEQ_START_TOKEN) {
2757 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2761 file = si->swap_file;
2762 len = seq_file_path(swap, file, " \t\n\\");
2763 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2764 len < 40 ? 40 - len : 1, " ",
2765 S_ISBLK(file_inode(file)->i_mode) ?
2766 "partition" : "file\t",
2767 si->pages << (PAGE_SHIFT - 10),
2768 si->inuse_pages << (PAGE_SHIFT - 10),
2773 static const struct seq_operations swaps_op = {
2774 .start = swap_start,
2780 static int swaps_open(struct inode *inode, struct file *file)
2782 struct seq_file *seq;
2785 ret = seq_open(file, &swaps_op);
2789 seq = file->private_data;
2790 seq->poll_event = atomic_read(&proc_poll_event);
2794 static const struct file_operations proc_swaps_operations = {
2797 .llseek = seq_lseek,
2798 .release = seq_release,
2802 static int __init procswaps_init(void)
2804 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2807 __initcall(procswaps_init);
2808 #endif /* CONFIG_PROC_FS */
2810 #ifdef MAX_SWAPFILES_CHECK
2811 static int __init max_swapfiles_check(void)
2813 MAX_SWAPFILES_CHECK();
2816 late_initcall(max_swapfiles_check);
2819 static struct swap_info_struct *alloc_swap_info(void)
2821 struct swap_info_struct *p;
2825 p = kzalloc(sizeof(*p), GFP_KERNEL);
2827 return ERR_PTR(-ENOMEM);
2829 spin_lock(&swap_lock);
2830 for (type = 0; type < nr_swapfiles; type++) {
2831 if (!(swap_info[type]->flags & SWP_USED))
2834 if (type >= MAX_SWAPFILES) {
2835 spin_unlock(&swap_lock);
2837 return ERR_PTR(-EPERM);
2839 if (type >= nr_swapfiles) {
2841 swap_info[type] = p;
2843 * Write swap_info[type] before nr_swapfiles, in case a
2844 * racing procfs swap_start() or swap_next() is reading them.
2845 * (We never shrink nr_swapfiles, we never free this entry.)
2851 p = swap_info[type];
2853 * Do not memset this entry: a racing procfs swap_next()
2854 * would be relying on p->type to remain valid.
2857 INIT_LIST_HEAD(&p->first_swap_extent.list);
2858 plist_node_init(&p->list, 0);
2860 plist_node_init(&p->avail_lists[i], 0);
2861 p->flags = SWP_USED;
2862 spin_unlock(&swap_lock);
2863 spin_lock_init(&p->lock);
2864 spin_lock_init(&p->cont_lock);
2869 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2873 if (S_ISBLK(inode->i_mode)) {
2874 p->bdev = bdgrab(I_BDEV(inode));
2875 error = blkdev_get(p->bdev,
2876 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2881 p->old_block_size = block_size(p->bdev);
2882 error = set_blocksize(p->bdev, PAGE_SIZE);
2885 p->flags |= SWP_BLKDEV;
2886 } else if (S_ISREG(inode->i_mode)) {
2887 p->bdev = inode->i_sb->s_bdev;
2889 if (IS_SWAPFILE(inode))
2899 * Find out how many pages are allowed for a single swap device. There
2900 * are two limiting factors:
2901 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2902 * 2) the number of bits in the swap pte, as defined by the different
2905 * In order to find the largest possible bit mask, a swap entry with
2906 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2907 * decoded to a swp_entry_t again, and finally the swap offset is
2910 * This will mask all the bits from the initial ~0UL mask that can't
2911 * be encoded in either the swp_entry_t or the architecture definition
2914 unsigned long generic_max_swapfile_size(void)
2916 return swp_offset(pte_to_swp_entry(
2917 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2920 /* Can be overridden by an architecture for additional checks. */
2921 __weak unsigned long max_swapfile_size(void)
2923 return generic_max_swapfile_size();
2926 static unsigned long read_swap_header(struct swap_info_struct *p,
2927 union swap_header *swap_header,
2928 struct inode *inode)
2931 unsigned long maxpages;
2932 unsigned long swapfilepages;
2933 unsigned long last_page;
2935 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2936 pr_err("Unable to find swap-space signature\n");
2940 /* swap partition endianess hack... */
2941 if (swab32(swap_header->info.version) == 1) {
2942 swab32s(&swap_header->info.version);
2943 swab32s(&swap_header->info.last_page);
2944 swab32s(&swap_header->info.nr_badpages);
2945 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2947 for (i = 0; i < swap_header->info.nr_badpages; i++)
2948 swab32s(&swap_header->info.badpages[i]);
2950 /* Check the swap header's sub-version */
2951 if (swap_header->info.version != 1) {
2952 pr_warn("Unable to handle swap header version %d\n",
2953 swap_header->info.version);
2958 p->cluster_next = 1;
2961 maxpages = max_swapfile_size();
2962 last_page = swap_header->info.last_page;
2964 pr_warn("Empty swap-file\n");
2967 if (last_page > maxpages) {
2968 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2969 maxpages << (PAGE_SHIFT - 10),
2970 last_page << (PAGE_SHIFT - 10));
2972 if (maxpages > last_page) {
2973 maxpages = last_page + 1;
2974 /* p->max is an unsigned int: don't overflow it */
2975 if ((unsigned int)maxpages == 0)
2976 maxpages = UINT_MAX;
2978 p->highest_bit = maxpages - 1;
2982 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2983 if (swapfilepages && maxpages > swapfilepages) {
2984 pr_warn("Swap area shorter than signature indicates\n");
2987 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2989 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2995 #define SWAP_CLUSTER_INFO_COLS \
2996 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2997 #define SWAP_CLUSTER_SPACE_COLS \
2998 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2999 #define SWAP_CLUSTER_COLS \
3000 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3002 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3003 union swap_header *swap_header,
3004 unsigned char *swap_map,
3005 struct swap_cluster_info *cluster_info,
3006 unsigned long maxpages,
3010 unsigned int nr_good_pages;
3012 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3013 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3014 unsigned long i, idx;
3016 nr_good_pages = maxpages - 1; /* omit header page */
3018 cluster_list_init(&p->free_clusters);
3019 cluster_list_init(&p->discard_clusters);
3021 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3022 unsigned int page_nr = swap_header->info.badpages[i];
3023 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3025 if (page_nr < maxpages) {
3026 swap_map[page_nr] = SWAP_MAP_BAD;
3029 * Haven't marked the cluster free yet, no list
3030 * operation involved
3032 inc_cluster_info_page(p, cluster_info, page_nr);
3036 /* Haven't marked the cluster free yet, no list operation involved */
3037 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3038 inc_cluster_info_page(p, cluster_info, i);
3040 if (nr_good_pages) {
3041 swap_map[0] = SWAP_MAP_BAD;
3043 * Not mark the cluster free yet, no list
3044 * operation involved
3046 inc_cluster_info_page(p, cluster_info, 0);
3048 p->pages = nr_good_pages;
3049 nr_extents = setup_swap_extents(p, span);
3052 nr_good_pages = p->pages;
3054 if (!nr_good_pages) {
3055 pr_warn("Empty swap-file\n");
3064 * Reduce false cache line sharing between cluster_info and
3065 * sharing same address space.
3067 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3068 j = (k + col) % SWAP_CLUSTER_COLS;
3069 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3070 idx = i * SWAP_CLUSTER_COLS + j;
3071 if (idx >= nr_clusters)
3073 if (cluster_count(&cluster_info[idx]))
3075 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3076 cluster_list_add_tail(&p->free_clusters, cluster_info,
3084 * Helper to sys_swapon determining if a given swap
3085 * backing device queue supports DISCARD operations.
3087 static bool swap_discardable(struct swap_info_struct *si)
3089 struct request_queue *q = bdev_get_queue(si->bdev);
3091 if (!q || !blk_queue_discard(q))
3097 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3099 struct swap_info_struct *p;
3100 struct filename *name;
3101 struct file *swap_file = NULL;
3102 struct address_space *mapping;
3105 union swap_header *swap_header;
3108 unsigned long maxpages;
3109 unsigned char *swap_map = NULL;
3110 struct swap_cluster_info *cluster_info = NULL;
3111 unsigned long *frontswap_map = NULL;
3112 struct page *page = NULL;
3113 struct inode *inode = NULL;
3114 bool inced_nr_rotate_swap = false;
3116 if (swap_flags & ~SWAP_FLAGS_VALID)
3119 if (!capable(CAP_SYS_ADMIN))
3122 if (!swap_avail_heads)
3125 p = alloc_swap_info();
3129 INIT_WORK(&p->discard_work, swap_discard_work);
3131 name = getname(specialfile);
3133 error = PTR_ERR(name);
3137 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3138 if (IS_ERR(swap_file)) {
3139 error = PTR_ERR(swap_file);
3144 p->swap_file = swap_file;
3145 mapping = swap_file->f_mapping;
3146 inode = mapping->host;
3148 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3149 error = claim_swapfile(p, inode);
3150 if (unlikely(error))
3154 * Read the swap header.
3156 if (!mapping->a_ops->readpage) {
3160 page = read_mapping_page(mapping, 0, swap_file);
3162 error = PTR_ERR(page);
3165 swap_header = kmap(page);
3167 maxpages = read_swap_header(p, swap_header, inode);
3168 if (unlikely(!maxpages)) {
3173 /* OK, set up the swap map and apply the bad block list */
3174 swap_map = vzalloc(maxpages);
3180 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3181 p->flags |= SWP_STABLE_WRITES;
3183 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3184 p->flags |= SWP_SYNCHRONOUS_IO;
3186 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3188 unsigned long ci, nr_cluster;
3190 p->flags |= SWP_SOLIDSTATE;
3192 * select a random position to start with to help wear leveling
3195 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3196 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3198 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3200 if (!cluster_info) {
3205 for (ci = 0; ci < nr_cluster; ci++)
3206 spin_lock_init(&((cluster_info + ci)->lock));
3208 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3209 if (!p->percpu_cluster) {
3213 for_each_possible_cpu(cpu) {
3214 struct percpu_cluster *cluster;
3215 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3216 cluster_set_null(&cluster->index);
3219 atomic_inc(&nr_rotate_swap);
3220 inced_nr_rotate_swap = true;
3223 error = swap_cgroup_swapon(p->type, maxpages);
3227 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3228 cluster_info, maxpages, &span);
3229 if (unlikely(nr_extents < 0)) {
3233 /* frontswap enabled? set up bit-per-page map for frontswap */
3234 if (IS_ENABLED(CONFIG_FRONTSWAP))
3235 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3239 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3241 * When discard is enabled for swap with no particular
3242 * policy flagged, we set all swap discard flags here in
3243 * order to sustain backward compatibility with older
3244 * swapon(8) releases.
3246 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3250 * By flagging sys_swapon, a sysadmin can tell us to
3251 * either do single-time area discards only, or to just
3252 * perform discards for released swap page-clusters.
3253 * Now it's time to adjust the p->flags accordingly.
3255 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3256 p->flags &= ~SWP_PAGE_DISCARD;
3257 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3258 p->flags &= ~SWP_AREA_DISCARD;
3260 /* issue a swapon-time discard if it's still required */
3261 if (p->flags & SWP_AREA_DISCARD) {
3262 int err = discard_swap(p);
3264 pr_err("swapon: discard_swap(%p): %d\n",
3269 error = init_swap_address_space(p->type, maxpages);
3273 mutex_lock(&swapon_mutex);
3275 if (swap_flags & SWAP_FLAG_PREFER)
3277 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3278 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3280 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3281 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3282 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3283 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3284 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3285 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3286 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3287 (frontswap_map) ? "FS" : "");
3289 mutex_unlock(&swapon_mutex);
3290 atomic_inc(&proc_poll_event);
3291 wake_up_interruptible(&proc_poll_wait);
3293 if (S_ISREG(inode->i_mode))
3294 inode->i_flags |= S_SWAPFILE;
3298 free_percpu(p->percpu_cluster);
3299 p->percpu_cluster = NULL;
3300 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3301 set_blocksize(p->bdev, p->old_block_size);
3302 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3304 destroy_swap_extents(p);
3305 swap_cgroup_swapoff(p->type);
3306 spin_lock(&swap_lock);
3307 p->swap_file = NULL;
3309 spin_unlock(&swap_lock);
3311 kvfree(cluster_info);
3312 kvfree(frontswap_map);
3313 if (inced_nr_rotate_swap)
3314 atomic_dec(&nr_rotate_swap);
3316 if (inode && S_ISREG(inode->i_mode)) {
3317 inode_unlock(inode);
3320 filp_close(swap_file, NULL);
3323 if (page && !IS_ERR(page)) {
3329 if (inode && S_ISREG(inode->i_mode))
3330 inode_unlock(inode);
3332 enable_swap_slots_cache();
3336 void si_swapinfo(struct sysinfo *val)
3339 unsigned long nr_to_be_unused = 0;
3341 spin_lock(&swap_lock);
3342 for (type = 0; type < nr_swapfiles; type++) {
3343 struct swap_info_struct *si = swap_info[type];
3345 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3346 nr_to_be_unused += si->inuse_pages;
3348 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3349 val->totalswap = total_swap_pages + nr_to_be_unused;
3350 spin_unlock(&swap_lock);
3354 * Verify that a swap entry is valid and increment its swap map count.
3356 * Returns error code in following case.
3358 * - swp_entry is invalid -> EINVAL
3359 * - swp_entry is migration entry -> EINVAL
3360 * - swap-cache reference is requested but there is already one. -> EEXIST
3361 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3362 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3364 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3366 struct swap_info_struct *p;
3367 struct swap_cluster_info *ci;
3368 unsigned long offset, type;
3369 unsigned char count;
3370 unsigned char has_cache;
3373 if (non_swap_entry(entry))
3376 type = swp_type(entry);
3377 if (type >= nr_swapfiles)
3379 p = swap_info[type];
3380 offset = swp_offset(entry);
3381 if (unlikely(offset >= p->max))
3384 ci = lock_cluster_or_swap_info(p, offset);
3386 count = p->swap_map[offset];
3389 * swapin_readahead() doesn't check if a swap entry is valid, so the
3390 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3392 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3397 has_cache = count & SWAP_HAS_CACHE;
3398 count &= ~SWAP_HAS_CACHE;
3401 if (usage == SWAP_HAS_CACHE) {
3403 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3404 if (!has_cache && count)
3405 has_cache = SWAP_HAS_CACHE;
3406 else if (has_cache) /* someone else added cache */
3408 else /* no users remaining */
3411 } else if (count || has_cache) {
3413 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3415 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3417 else if (swap_count_continued(p, offset, count))
3418 count = COUNT_CONTINUED;
3422 err = -ENOENT; /* unused swap entry */
3424 p->swap_map[offset] = count | has_cache;
3427 unlock_cluster_or_swap_info(p, ci);
3432 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3437 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3438 * (in which case its reference count is never incremented).
3440 void swap_shmem_alloc(swp_entry_t entry)
3442 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3446 * Increase reference count of swap entry by 1.
3447 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3448 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3449 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3450 * might occur if a page table entry has got corrupted.
3452 int swap_duplicate(swp_entry_t entry)
3456 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3457 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3462 * @entry: swap entry for which we allocate swap cache.
3464 * Called when allocating swap cache for existing swap entry,
3465 * This can return error codes. Returns 0 at success.
3466 * -EBUSY means there is a swap cache.
3467 * Note: return code is different from swap_duplicate().
3469 int swapcache_prepare(swp_entry_t entry)
3471 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3474 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3476 return swap_info[swp_type(entry)];
3479 struct swap_info_struct *page_swap_info(struct page *page)
3481 swp_entry_t entry = { .val = page_private(page) };
3482 return swp_swap_info(entry);
3486 * out-of-line __page_file_ methods to avoid include hell.
3488 struct address_space *__page_file_mapping(struct page *page)
3490 return page_swap_info(page)->swap_file->f_mapping;
3492 EXPORT_SYMBOL_GPL(__page_file_mapping);
3494 pgoff_t __page_file_index(struct page *page)
3496 swp_entry_t swap = { .val = page_private(page) };
3497 return swp_offset(swap);
3499 EXPORT_SYMBOL_GPL(__page_file_index);
3502 * add_swap_count_continuation - called when a swap count is duplicated
3503 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3504 * page of the original vmalloc'ed swap_map, to hold the continuation count
3505 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3506 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3508 * These continuation pages are seldom referenced: the common paths all work
3509 * on the original swap_map, only referring to a continuation page when the
3510 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3512 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3513 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3514 * can be called after dropping locks.
3516 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3518 struct swap_info_struct *si;
3519 struct swap_cluster_info *ci;
3522 struct page *list_page;
3524 unsigned char count;
3527 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3528 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3530 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3532 si = swap_info_get(entry);
3535 * An acceptable race has occurred since the failing
3536 * __swap_duplicate(): the swap entry has been freed,
3537 * perhaps even the whole swap_map cleared for swapoff.
3542 offset = swp_offset(entry);
3544 ci = lock_cluster(si, offset);
3546 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3548 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3550 * The higher the swap count, the more likely it is that tasks
3551 * will race to add swap count continuation: we need to avoid
3552 * over-provisioning.
3559 spin_unlock(&si->lock);
3564 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3565 * no architecture is using highmem pages for kernel page tables: so it
3566 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3568 head = vmalloc_to_page(si->swap_map + offset);
3569 offset &= ~PAGE_MASK;
3571 spin_lock(&si->cont_lock);
3573 * Page allocation does not initialize the page's lru field,
3574 * but it does always reset its private field.
3576 if (!page_private(head)) {
3577 BUG_ON(count & COUNT_CONTINUED);
3578 INIT_LIST_HEAD(&head->lru);
3579 set_page_private(head, SWP_CONTINUED);
3580 si->flags |= SWP_CONTINUED;
3583 list_for_each_entry(list_page, &head->lru, lru) {
3587 * If the previous map said no continuation, but we've found
3588 * a continuation page, free our allocation and use this one.
3590 if (!(count & COUNT_CONTINUED))
3591 goto out_unlock_cont;
3593 map = kmap_atomic(list_page) + offset;
3598 * If this continuation count now has some space in it,
3599 * free our allocation and use this one.
3601 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3602 goto out_unlock_cont;
3605 list_add_tail(&page->lru, &head->lru);
3606 page = NULL; /* now it's attached, don't free it */
3608 spin_unlock(&si->cont_lock);
3611 spin_unlock(&si->lock);
3619 * swap_count_continued - when the original swap_map count is incremented
3620 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3621 * into, carry if so, or else fail until a new continuation page is allocated;
3622 * when the original swap_map count is decremented from 0 with continuation,
3623 * borrow from the continuation and report whether it still holds more.
3624 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3627 static bool swap_count_continued(struct swap_info_struct *si,
3628 pgoff_t offset, unsigned char count)
3635 head = vmalloc_to_page(si->swap_map + offset);
3636 if (page_private(head) != SWP_CONTINUED) {
3637 BUG_ON(count & COUNT_CONTINUED);
3638 return false; /* need to add count continuation */
3641 spin_lock(&si->cont_lock);
3642 offset &= ~PAGE_MASK;
3643 page = list_entry(head->lru.next, struct page, lru);
3644 map = kmap_atomic(page) + offset;
3646 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3647 goto init_map; /* jump over SWAP_CONT_MAX checks */
3649 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3651 * Think of how you add 1 to 999
3653 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3655 page = list_entry(page->lru.next, struct page, lru);
3656 BUG_ON(page == head);
3657 map = kmap_atomic(page) + offset;
3659 if (*map == SWAP_CONT_MAX) {
3661 page = list_entry(page->lru.next, struct page, lru);
3663 ret = false; /* add count continuation */
3666 map = kmap_atomic(page) + offset;
3667 init_map: *map = 0; /* we didn't zero the page */
3671 page = list_entry(page->lru.prev, struct page, lru);
3672 while (page != head) {
3673 map = kmap_atomic(page) + offset;
3674 *map = COUNT_CONTINUED;
3676 page = list_entry(page->lru.prev, struct page, lru);
3678 ret = true; /* incremented */
3680 } else { /* decrementing */
3682 * Think of how you subtract 1 from 1000
3684 BUG_ON(count != COUNT_CONTINUED);
3685 while (*map == COUNT_CONTINUED) {
3687 page = list_entry(page->lru.next, struct page, lru);
3688 BUG_ON(page == head);
3689 map = kmap_atomic(page) + offset;
3696 page = list_entry(page->lru.prev, struct page, lru);
3697 while (page != head) {
3698 map = kmap_atomic(page) + offset;
3699 *map = SWAP_CONT_MAX | count;
3700 count = COUNT_CONTINUED;
3702 page = list_entry(page->lru.prev, struct page, lru);
3704 ret = count == COUNT_CONTINUED;
3707 spin_unlock(&si->cont_lock);
3712 * free_swap_count_continuations - swapoff free all the continuation pages
3713 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3715 static void free_swap_count_continuations(struct swap_info_struct *si)
3719 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3721 head = vmalloc_to_page(si->swap_map + offset);
3722 if (page_private(head)) {
3723 struct page *page, *next;
3725 list_for_each_entry_safe(page, next, &head->lru, lru) {
3726 list_del(&page->lru);
3733 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3734 void mem_cgroup_throttle_swaprate(struct mem_cgroup *memcg, int node,
3737 struct swap_info_struct *si, *next;
3738 if (!(gfp_mask & __GFP_IO) || !memcg)
3741 if (!blk_cgroup_congested())
3745 * We've already scheduled a throttle, avoid taking the global swap
3748 if (current->throttle_queue)
3751 spin_lock(&swap_avail_lock);
3752 plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
3753 avail_lists[node]) {
3755 blkcg_schedule_throttle(bdev_get_queue(si->bdev),
3760 spin_unlock(&swap_avail_lock);
3764 static int __init swapfile_init(void)
3768 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3770 if (!swap_avail_heads) {
3771 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3776 plist_head_init(&swap_avail_heads[nid]);
3780 subsys_initcall(swapfile_init);