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 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 SWAP_HAS_CONT 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);
299 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
300 struct swap_info_struct *si,
301 unsigned long offset)
303 struct swap_cluster_info *ci;
305 ci = lock_cluster(si, offset);
307 spin_lock(&si->lock);
312 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
313 struct swap_cluster_info *ci)
318 spin_unlock(&si->lock);
321 static inline bool cluster_list_empty(struct swap_cluster_list *list)
323 return cluster_is_null(&list->head);
326 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
328 return cluster_next(&list->head);
331 static void cluster_list_init(struct swap_cluster_list *list)
333 cluster_set_null(&list->head);
334 cluster_set_null(&list->tail);
337 static void cluster_list_add_tail(struct swap_cluster_list *list,
338 struct swap_cluster_info *ci,
341 if (cluster_list_empty(list)) {
342 cluster_set_next_flag(&list->head, idx, 0);
343 cluster_set_next_flag(&list->tail, idx, 0);
345 struct swap_cluster_info *ci_tail;
346 unsigned int tail = cluster_next(&list->tail);
349 * Nested cluster lock, but both cluster locks are
350 * only acquired when we held swap_info_struct->lock
353 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
354 cluster_set_next(ci_tail, idx);
355 spin_unlock(&ci_tail->lock);
356 cluster_set_next_flag(&list->tail, idx, 0);
360 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
361 struct swap_cluster_info *ci)
365 idx = cluster_next(&list->head);
366 if (cluster_next(&list->tail) == idx) {
367 cluster_set_null(&list->head);
368 cluster_set_null(&list->tail);
370 cluster_set_next_flag(&list->head,
371 cluster_next(&ci[idx]), 0);
376 /* Add a cluster to discard list and schedule it to do discard */
377 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
381 * If scan_swap_map() can't find a free cluster, it will check
382 * si->swap_map directly. To make sure the discarding cluster isn't
383 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
384 * will be cleared after discard
386 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
387 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
389 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
391 schedule_work(&si->discard_work);
394 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
396 struct swap_cluster_info *ci = si->cluster_info;
398 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
399 cluster_list_add_tail(&si->free_clusters, ci, idx);
403 * Doing discard actually. After a cluster discard is finished, the cluster
404 * will be added to free cluster list. caller should hold si->lock.
406 static void swap_do_scheduled_discard(struct swap_info_struct *si)
408 struct swap_cluster_info *info, *ci;
411 info = si->cluster_info;
413 while (!cluster_list_empty(&si->discard_clusters)) {
414 idx = cluster_list_del_first(&si->discard_clusters, info);
415 spin_unlock(&si->lock);
417 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
420 spin_lock(&si->lock);
421 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
422 __free_cluster(si, idx);
423 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
424 0, SWAPFILE_CLUSTER);
429 static void swap_discard_work(struct work_struct *work)
431 struct swap_info_struct *si;
433 si = container_of(work, struct swap_info_struct, discard_work);
435 spin_lock(&si->lock);
436 swap_do_scheduled_discard(si);
437 spin_unlock(&si->lock);
440 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
442 struct swap_cluster_info *ci = si->cluster_info;
444 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
445 cluster_list_del_first(&si->free_clusters, ci);
446 cluster_set_count_flag(ci + idx, 0, 0);
449 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
451 struct swap_cluster_info *ci = si->cluster_info + idx;
453 VM_BUG_ON(cluster_count(ci) != 0);
455 * If the swap is discardable, prepare discard the cluster
456 * instead of free it immediately. The cluster will be freed
459 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
460 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
461 swap_cluster_schedule_discard(si, idx);
465 __free_cluster(si, idx);
469 * The cluster corresponding to page_nr will be used. The cluster will be
470 * removed from free cluster list and its usage counter will be increased.
472 static void inc_cluster_info_page(struct swap_info_struct *p,
473 struct swap_cluster_info *cluster_info, unsigned long page_nr)
475 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
479 if (cluster_is_free(&cluster_info[idx]))
480 alloc_cluster(p, idx);
482 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
483 cluster_set_count(&cluster_info[idx],
484 cluster_count(&cluster_info[idx]) + 1);
488 * The cluster corresponding to page_nr decreases one usage. If the usage
489 * counter becomes 0, which means no page in the cluster is in using, we can
490 * optionally discard the cluster and add it to free cluster list.
492 static void dec_cluster_info_page(struct swap_info_struct *p,
493 struct swap_cluster_info *cluster_info, unsigned long page_nr)
495 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
500 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
501 cluster_set_count(&cluster_info[idx],
502 cluster_count(&cluster_info[idx]) - 1);
504 if (cluster_count(&cluster_info[idx]) == 0)
505 free_cluster(p, idx);
509 * It's possible scan_swap_map() uses a free cluster in the middle of free
510 * cluster list. Avoiding such abuse to avoid list corruption.
513 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
514 unsigned long offset)
516 struct percpu_cluster *percpu_cluster;
519 offset /= SWAPFILE_CLUSTER;
520 conflict = !cluster_list_empty(&si->free_clusters) &&
521 offset != cluster_list_first(&si->free_clusters) &&
522 cluster_is_free(&si->cluster_info[offset]);
527 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
528 cluster_set_null(&percpu_cluster->index);
533 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
534 * might involve allocating a new cluster for current CPU too.
536 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
537 unsigned long *offset, unsigned long *scan_base)
539 struct percpu_cluster *cluster;
540 struct swap_cluster_info *ci;
542 unsigned long tmp, max;
545 cluster = this_cpu_ptr(si->percpu_cluster);
546 if (cluster_is_null(&cluster->index)) {
547 if (!cluster_list_empty(&si->free_clusters)) {
548 cluster->index = si->free_clusters.head;
549 cluster->next = cluster_next(&cluster->index) *
551 } else if (!cluster_list_empty(&si->discard_clusters)) {
553 * we don't have free cluster but have some clusters in
554 * discarding, do discard now and reclaim them
556 swap_do_scheduled_discard(si);
557 *scan_base = *offset = si->cluster_next;
566 * Other CPUs can use our cluster if they can't find a free cluster,
567 * check if there is still free entry in the cluster
570 max = min_t(unsigned long, si->max,
571 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
573 cluster_set_null(&cluster->index);
576 ci = lock_cluster(si, tmp);
578 if (!si->swap_map[tmp]) {
586 cluster_set_null(&cluster->index);
589 cluster->next = tmp + 1;
595 static void __del_from_avail_list(struct swap_info_struct *p)
599 assert_spin_locked(&p->lock);
601 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
604 static void del_from_avail_list(struct swap_info_struct *p)
606 spin_lock(&swap_avail_lock);
607 __del_from_avail_list(p);
608 spin_unlock(&swap_avail_lock);
611 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
612 unsigned int nr_entries)
614 unsigned int end = offset + nr_entries - 1;
616 if (offset == si->lowest_bit)
617 si->lowest_bit += nr_entries;
618 if (end == si->highest_bit)
619 si->highest_bit -= nr_entries;
620 si->inuse_pages += nr_entries;
621 if (si->inuse_pages == si->pages) {
622 si->lowest_bit = si->max;
624 del_from_avail_list(si);
628 static void add_to_avail_list(struct swap_info_struct *p)
632 spin_lock(&swap_avail_lock);
634 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
635 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
637 spin_unlock(&swap_avail_lock);
640 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
641 unsigned int nr_entries)
643 unsigned long end = offset + nr_entries - 1;
644 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
646 if (offset < si->lowest_bit)
647 si->lowest_bit = offset;
648 if (end > si->highest_bit) {
649 bool was_full = !si->highest_bit;
651 si->highest_bit = end;
652 if (was_full && (si->flags & SWP_WRITEOK))
653 add_to_avail_list(si);
655 atomic_long_add(nr_entries, &nr_swap_pages);
656 si->inuse_pages -= nr_entries;
657 if (si->flags & SWP_BLKDEV)
658 swap_slot_free_notify =
659 si->bdev->bd_disk->fops->swap_slot_free_notify;
661 swap_slot_free_notify = NULL;
662 while (offset <= end) {
663 frontswap_invalidate_page(si->type, offset);
664 if (swap_slot_free_notify)
665 swap_slot_free_notify(si->bdev, offset);
670 static int scan_swap_map_slots(struct swap_info_struct *si,
671 unsigned char usage, int nr,
674 struct swap_cluster_info *ci;
675 unsigned long offset;
676 unsigned long scan_base;
677 unsigned long last_in_cluster = 0;
678 int latency_ration = LATENCY_LIMIT;
685 * We try to cluster swap pages by allocating them sequentially
686 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
687 * way, however, we resort to first-free allocation, starting
688 * a new cluster. This prevents us from scattering swap pages
689 * all over the entire swap partition, so that we reduce
690 * overall disk seek times between swap pages. -- sct
691 * But we do now try to find an empty cluster. -Andrea
692 * And we let swap pages go all over an SSD partition. Hugh
695 si->flags += SWP_SCANNING;
696 scan_base = offset = si->cluster_next;
699 if (si->cluster_info) {
700 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
706 if (unlikely(!si->cluster_nr--)) {
707 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
708 si->cluster_nr = SWAPFILE_CLUSTER - 1;
712 spin_unlock(&si->lock);
715 * If seek is expensive, start searching for new cluster from
716 * start of partition, to minimize the span of allocated swap.
717 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
718 * case, just handled by scan_swap_map_try_ssd_cluster() above.
720 scan_base = offset = si->lowest_bit;
721 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
723 /* Locate the first empty (unaligned) cluster */
724 for (; last_in_cluster <= si->highest_bit; offset++) {
725 if (si->swap_map[offset])
726 last_in_cluster = offset + SWAPFILE_CLUSTER;
727 else if (offset == last_in_cluster) {
728 spin_lock(&si->lock);
729 offset -= SWAPFILE_CLUSTER - 1;
730 si->cluster_next = offset;
731 si->cluster_nr = SWAPFILE_CLUSTER - 1;
734 if (unlikely(--latency_ration < 0)) {
736 latency_ration = LATENCY_LIMIT;
741 spin_lock(&si->lock);
742 si->cluster_nr = SWAPFILE_CLUSTER - 1;
746 if (si->cluster_info) {
747 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
748 /* take a break if we already got some slots */
751 if (!scan_swap_map_try_ssd_cluster(si, &offset,
756 if (!(si->flags & SWP_WRITEOK))
758 if (!si->highest_bit)
760 if (offset > si->highest_bit)
761 scan_base = offset = si->lowest_bit;
763 ci = lock_cluster(si, offset);
764 /* reuse swap entry of cache-only swap if not busy. */
765 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
768 spin_unlock(&si->lock);
769 swap_was_freed = __try_to_reclaim_swap(si, offset);
770 spin_lock(&si->lock);
771 /* entry was freed successfully, try to use this again */
774 goto scan; /* check next one */
777 if (si->swap_map[offset]) {
784 si->swap_map[offset] = usage;
785 inc_cluster_info_page(si, si->cluster_info, offset);
788 swap_range_alloc(si, offset, 1);
789 si->cluster_next = offset + 1;
790 slots[n_ret++] = swp_entry(si->type, offset);
792 /* got enough slots or reach max slots? */
793 if ((n_ret == nr) || (offset >= si->highest_bit))
796 /* search for next available slot */
798 /* time to take a break? */
799 if (unlikely(--latency_ration < 0)) {
802 spin_unlock(&si->lock);
804 spin_lock(&si->lock);
805 latency_ration = LATENCY_LIMIT;
808 /* try to get more slots in cluster */
809 if (si->cluster_info) {
810 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
818 /* non-ssd case, still more slots in cluster? */
819 if (si->cluster_nr && !si->swap_map[offset]) {
825 si->flags -= SWP_SCANNING;
829 spin_unlock(&si->lock);
830 while (++offset <= si->highest_bit) {
831 if (!si->swap_map[offset]) {
832 spin_lock(&si->lock);
835 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
836 spin_lock(&si->lock);
839 if (unlikely(--latency_ration < 0)) {
841 latency_ration = LATENCY_LIMIT;
844 offset = si->lowest_bit;
845 while (offset < scan_base) {
846 if (!si->swap_map[offset]) {
847 spin_lock(&si->lock);
850 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
851 spin_lock(&si->lock);
854 if (unlikely(--latency_ration < 0)) {
856 latency_ration = LATENCY_LIMIT;
860 spin_lock(&si->lock);
863 si->flags -= SWP_SCANNING;
867 #ifdef CONFIG_THP_SWAP
868 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
871 struct swap_cluster_info *ci;
872 unsigned long offset, i;
875 if (cluster_list_empty(&si->free_clusters))
878 idx = cluster_list_first(&si->free_clusters);
879 offset = idx * SWAPFILE_CLUSTER;
880 ci = lock_cluster(si, offset);
881 alloc_cluster(si, idx);
882 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
884 map = si->swap_map + offset;
885 for (i = 0; i < SWAPFILE_CLUSTER; i++)
886 map[i] = SWAP_HAS_CACHE;
888 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
889 *slot = swp_entry(si->type, offset);
894 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
896 unsigned long offset = idx * SWAPFILE_CLUSTER;
897 struct swap_cluster_info *ci;
899 ci = lock_cluster(si, offset);
900 cluster_set_count_flag(ci, 0, 0);
901 free_cluster(si, idx);
903 swap_range_free(si, offset, SWAPFILE_CLUSTER);
906 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
911 #endif /* CONFIG_THP_SWAP */
913 static unsigned long scan_swap_map(struct swap_info_struct *si,
919 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
922 return swp_offset(entry);
928 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
930 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
931 struct swap_info_struct *si, *next;
936 /* Only single cluster request supported */
937 WARN_ON_ONCE(n_goal > 1 && cluster);
939 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
943 if (n_goal > SWAP_BATCH)
946 if (n_goal > avail_pgs)
949 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
951 spin_lock(&swap_avail_lock);
954 node = numa_node_id();
955 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
956 /* requeue si to after same-priority siblings */
957 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
958 spin_unlock(&swap_avail_lock);
959 spin_lock(&si->lock);
960 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
961 spin_lock(&swap_avail_lock);
962 if (plist_node_empty(&si->avail_lists[node])) {
963 spin_unlock(&si->lock);
966 WARN(!si->highest_bit,
967 "swap_info %d in list but !highest_bit\n",
969 WARN(!(si->flags & SWP_WRITEOK),
970 "swap_info %d in list but !SWP_WRITEOK\n",
972 __del_from_avail_list(si);
973 spin_unlock(&si->lock);
977 if (si->flags & SWP_BLKDEV)
978 n_ret = swap_alloc_cluster(si, swp_entries);
980 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
981 n_goal, swp_entries);
982 spin_unlock(&si->lock);
983 if (n_ret || cluster)
985 pr_debug("scan_swap_map of si %d failed to find offset\n",
989 spin_lock(&swap_avail_lock);
992 * if we got here, it's likely that si was almost full before,
993 * and since scan_swap_map() can drop the si->lock, multiple
994 * callers probably all tried to get a page from the same si
995 * and it filled up before we could get one; or, the si filled
996 * up between us dropping swap_avail_lock and taking si->lock.
997 * Since we dropped the swap_avail_lock, the swap_avail_head
998 * list may have been modified; so if next is still in the
999 * swap_avail_head list then try it, otherwise start over
1000 * if we have not gotten any slots.
1002 if (plist_node_empty(&next->avail_lists[node]))
1006 spin_unlock(&swap_avail_lock);
1010 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1016 /* The only caller of this function is now suspend routine */
1017 swp_entry_t get_swap_page_of_type(int type)
1019 struct swap_info_struct *si;
1022 si = swap_info[type];
1023 spin_lock(&si->lock);
1024 if (si && (si->flags & SWP_WRITEOK)) {
1025 atomic_long_dec(&nr_swap_pages);
1026 /* This is called for allocating swap entry, not cache */
1027 offset = scan_swap_map(si, 1);
1029 spin_unlock(&si->lock);
1030 return swp_entry(type, offset);
1032 atomic_long_inc(&nr_swap_pages);
1034 spin_unlock(&si->lock);
1035 return (swp_entry_t) {0};
1038 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1040 struct swap_info_struct *p;
1041 unsigned long offset, type;
1045 type = swp_type(entry);
1046 if (type >= nr_swapfiles)
1048 p = swap_info[type];
1049 if (!(p->flags & SWP_USED))
1051 offset = swp_offset(entry);
1052 if (offset >= p->max)
1057 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1060 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1063 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1068 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1070 struct swap_info_struct *p;
1072 p = __swap_info_get(entry);
1075 if (!p->swap_map[swp_offset(entry)])
1080 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1086 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1088 struct swap_info_struct *p;
1090 p = _swap_info_get(entry);
1092 spin_lock(&p->lock);
1096 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1097 struct swap_info_struct *q)
1099 struct swap_info_struct *p;
1101 p = _swap_info_get(entry);
1105 spin_unlock(&q->lock);
1107 spin_lock(&p->lock);
1112 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1113 swp_entry_t entry, unsigned char usage)
1115 struct swap_cluster_info *ci;
1116 unsigned long offset = swp_offset(entry);
1117 unsigned char count;
1118 unsigned char has_cache;
1120 ci = lock_cluster_or_swap_info(p, offset);
1122 count = p->swap_map[offset];
1124 has_cache = count & SWAP_HAS_CACHE;
1125 count &= ~SWAP_HAS_CACHE;
1127 if (usage == SWAP_HAS_CACHE) {
1128 VM_BUG_ON(!has_cache);
1130 } else if (count == SWAP_MAP_SHMEM) {
1132 * Or we could insist on shmem.c using a special
1133 * swap_shmem_free() and free_shmem_swap_and_cache()...
1136 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1137 if (count == COUNT_CONTINUED) {
1138 if (swap_count_continued(p, offset, count))
1139 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1141 count = SWAP_MAP_MAX;
1146 usage = count | has_cache;
1147 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1149 unlock_cluster_or_swap_info(p, ci);
1154 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1156 struct swap_cluster_info *ci;
1157 unsigned long offset = swp_offset(entry);
1158 unsigned char count;
1160 ci = lock_cluster(p, offset);
1161 count = p->swap_map[offset];
1162 VM_BUG_ON(count != SWAP_HAS_CACHE);
1163 p->swap_map[offset] = 0;
1164 dec_cluster_info_page(p, p->cluster_info, offset);
1167 mem_cgroup_uncharge_swap(entry, 1);
1168 swap_range_free(p, offset, 1);
1172 * Caller has made sure that the swap device corresponding to entry
1173 * is still around or has not been recycled.
1175 void swap_free(swp_entry_t entry)
1177 struct swap_info_struct *p;
1179 p = _swap_info_get(entry);
1181 if (!__swap_entry_free(p, entry, 1))
1182 free_swap_slot(entry);
1187 * Called after dropping swapcache to decrease refcnt to swap entries.
1189 static void swapcache_free(swp_entry_t entry)
1191 struct swap_info_struct *p;
1193 p = _swap_info_get(entry);
1195 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1196 free_swap_slot(entry);
1200 #ifdef CONFIG_THP_SWAP
1201 static void swapcache_free_cluster(swp_entry_t entry)
1203 unsigned long offset = swp_offset(entry);
1204 unsigned long idx = offset / SWAPFILE_CLUSTER;
1205 struct swap_cluster_info *ci;
1206 struct swap_info_struct *si;
1208 unsigned int i, free_entries = 0;
1211 si = _swap_info_get(entry);
1215 ci = lock_cluster(si, offset);
1216 VM_BUG_ON(!cluster_is_huge(ci));
1217 map = si->swap_map + offset;
1218 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1220 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1221 if (val == SWAP_HAS_CACHE)
1224 if (!free_entries) {
1225 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1226 map[i] &= ~SWAP_HAS_CACHE;
1228 cluster_clear_huge(ci);
1230 if (free_entries == SWAPFILE_CLUSTER) {
1231 spin_lock(&si->lock);
1232 ci = lock_cluster(si, offset);
1233 memset(map, 0, SWAPFILE_CLUSTER);
1235 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1236 swap_free_cluster(si, idx);
1237 spin_unlock(&si->lock);
1238 } else if (free_entries) {
1239 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1240 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1241 free_swap_slot(entry);
1246 int split_swap_cluster(swp_entry_t entry)
1248 struct swap_info_struct *si;
1249 struct swap_cluster_info *ci;
1250 unsigned long offset = swp_offset(entry);
1252 si = _swap_info_get(entry);
1255 ci = lock_cluster(si, offset);
1256 cluster_clear_huge(ci);
1261 static inline void swapcache_free_cluster(swp_entry_t entry)
1264 #endif /* CONFIG_THP_SWAP */
1266 void put_swap_page(struct page *page, swp_entry_t entry)
1268 if (!PageTransHuge(page))
1269 swapcache_free(entry);
1271 swapcache_free_cluster(entry);
1274 static int swp_entry_cmp(const void *ent1, const void *ent2)
1276 const swp_entry_t *e1 = ent1, *e2 = ent2;
1278 return (int)swp_type(*e1) - (int)swp_type(*e2);
1281 void swapcache_free_entries(swp_entry_t *entries, int n)
1283 struct swap_info_struct *p, *prev;
1293 * Sort swap entries by swap device, so each lock is only taken once.
1294 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1295 * so low that it isn't necessary to optimize further.
1297 if (nr_swapfiles > 1)
1298 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1299 for (i = 0; i < n; ++i) {
1300 p = swap_info_get_cont(entries[i], prev);
1302 swap_entry_free(p, entries[i]);
1306 spin_unlock(&p->lock);
1310 * How many references to page are currently swapped out?
1311 * This does not give an exact answer when swap count is continued,
1312 * but does include the high COUNT_CONTINUED flag to allow for that.
1314 int page_swapcount(struct page *page)
1317 struct swap_info_struct *p;
1318 struct swap_cluster_info *ci;
1320 unsigned long offset;
1322 entry.val = page_private(page);
1323 p = _swap_info_get(entry);
1325 offset = swp_offset(entry);
1326 ci = lock_cluster_or_swap_info(p, offset);
1327 count = swap_count(p->swap_map[offset]);
1328 unlock_cluster_or_swap_info(p, ci);
1333 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1336 pgoff_t offset = swp_offset(entry);
1337 struct swap_cluster_info *ci;
1339 ci = lock_cluster_or_swap_info(si, offset);
1340 count = swap_count(si->swap_map[offset]);
1341 unlock_cluster_or_swap_info(si, ci);
1346 * How many references to @entry are currently swapped out?
1347 * This does not give an exact answer when swap count is continued,
1348 * but does include the high COUNT_CONTINUED flag to allow for that.
1350 int __swp_swapcount(swp_entry_t entry)
1353 struct swap_info_struct *si;
1355 si = __swap_info_get(entry);
1357 count = swap_swapcount(si, entry);
1362 * How many references to @entry are currently swapped out?
1363 * This considers COUNT_CONTINUED so it returns exact answer.
1365 int swp_swapcount(swp_entry_t entry)
1367 int count, tmp_count, n;
1368 struct swap_info_struct *p;
1369 struct swap_cluster_info *ci;
1374 p = _swap_info_get(entry);
1378 offset = swp_offset(entry);
1380 ci = lock_cluster_or_swap_info(p, offset);
1382 count = swap_count(p->swap_map[offset]);
1383 if (!(count & COUNT_CONTINUED))
1386 count &= ~COUNT_CONTINUED;
1387 n = SWAP_MAP_MAX + 1;
1389 page = vmalloc_to_page(p->swap_map + offset);
1390 offset &= ~PAGE_MASK;
1391 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1394 page = list_next_entry(page, lru);
1395 map = kmap_atomic(page);
1396 tmp_count = map[offset];
1399 count += (tmp_count & ~COUNT_CONTINUED) * n;
1400 n *= (SWAP_CONT_MAX + 1);
1401 } while (tmp_count & COUNT_CONTINUED);
1403 unlock_cluster_or_swap_info(p, ci);
1407 #ifdef CONFIG_THP_SWAP
1408 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1411 struct swap_cluster_info *ci;
1412 unsigned char *map = si->swap_map;
1413 unsigned long roffset = swp_offset(entry);
1414 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1418 ci = lock_cluster_or_swap_info(si, offset);
1419 if (!ci || !cluster_is_huge(ci)) {
1420 if (map[roffset] != SWAP_HAS_CACHE)
1424 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1425 if (map[offset + i] != SWAP_HAS_CACHE) {
1431 unlock_cluster_or_swap_info(si, ci);
1435 static bool page_swapped(struct page *page)
1438 struct swap_info_struct *si;
1440 if (likely(!PageTransCompound(page)))
1441 return page_swapcount(page) != 0;
1443 page = compound_head(page);
1444 entry.val = page_private(page);
1445 si = _swap_info_get(entry);
1447 return swap_page_trans_huge_swapped(si, entry);
1451 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1452 int *total_swapcount)
1454 int i, map_swapcount, _total_mapcount, _total_swapcount;
1455 unsigned long offset = 0;
1456 struct swap_info_struct *si;
1457 struct swap_cluster_info *ci = NULL;
1458 unsigned char *map = NULL;
1459 int mapcount, swapcount = 0;
1461 /* hugetlbfs shouldn't call it */
1462 VM_BUG_ON_PAGE(PageHuge(page), page);
1464 if (likely(!PageTransCompound(page))) {
1465 mapcount = atomic_read(&page->_mapcount) + 1;
1467 *total_mapcount = mapcount;
1468 if (PageSwapCache(page))
1469 swapcount = page_swapcount(page);
1470 if (total_swapcount)
1471 *total_swapcount = swapcount;
1472 return mapcount + swapcount;
1475 page = compound_head(page);
1477 _total_mapcount = _total_swapcount = map_swapcount = 0;
1478 if (PageSwapCache(page)) {
1481 entry.val = page_private(page);
1482 si = _swap_info_get(entry);
1485 offset = swp_offset(entry);
1489 ci = lock_cluster(si, offset);
1490 for (i = 0; i < HPAGE_PMD_NR; i++) {
1491 mapcount = atomic_read(&page[i]._mapcount) + 1;
1492 _total_mapcount += mapcount;
1494 swapcount = swap_count(map[offset + i]);
1495 _total_swapcount += swapcount;
1497 map_swapcount = max(map_swapcount, mapcount + swapcount);
1500 if (PageDoubleMap(page)) {
1502 _total_mapcount -= HPAGE_PMD_NR;
1504 mapcount = compound_mapcount(page);
1505 map_swapcount += mapcount;
1506 _total_mapcount += mapcount;
1508 *total_mapcount = _total_mapcount;
1509 if (total_swapcount)
1510 *total_swapcount = _total_swapcount;
1512 return map_swapcount;
1515 #define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1516 #define page_swapped(page) (page_swapcount(page) != 0)
1518 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1519 int *total_swapcount)
1521 int mapcount, swapcount = 0;
1523 /* hugetlbfs shouldn't call it */
1524 VM_BUG_ON_PAGE(PageHuge(page), page);
1526 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1527 if (PageSwapCache(page))
1528 swapcount = page_swapcount(page);
1529 if (total_swapcount)
1530 *total_swapcount = swapcount;
1531 return mapcount + swapcount;
1536 * We can write to an anon page without COW if there are no other references
1537 * to it. And as a side-effect, free up its swap: because the old content
1538 * on disk will never be read, and seeking back there to write new content
1539 * later would only waste time away from clustering.
1541 * NOTE: total_map_swapcount should not be relied upon by the caller if
1542 * reuse_swap_page() returns false, but it may be always overwritten
1543 * (see the other implementation for CONFIG_SWAP=n).
1545 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1547 int count, total_mapcount, total_swapcount;
1549 VM_BUG_ON_PAGE(!PageLocked(page), page);
1550 if (unlikely(PageKsm(page)))
1552 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1554 if (total_map_swapcount)
1555 *total_map_swapcount = total_mapcount + total_swapcount;
1556 if (count == 1 && PageSwapCache(page) &&
1557 (likely(!PageTransCompound(page)) ||
1558 /* The remaining swap count will be freed soon */
1559 total_swapcount == page_swapcount(page))) {
1560 if (!PageWriteback(page)) {
1561 page = compound_head(page);
1562 delete_from_swap_cache(page);
1566 struct swap_info_struct *p;
1568 entry.val = page_private(page);
1569 p = swap_info_get(entry);
1570 if (p->flags & SWP_STABLE_WRITES) {
1571 spin_unlock(&p->lock);
1574 spin_unlock(&p->lock);
1582 * If swap is getting full, or if there are no more mappings of this page,
1583 * then try_to_free_swap is called to free its swap space.
1585 int try_to_free_swap(struct page *page)
1587 VM_BUG_ON_PAGE(!PageLocked(page), page);
1589 if (!PageSwapCache(page))
1591 if (PageWriteback(page))
1593 if (page_swapped(page))
1597 * Once hibernation has begun to create its image of memory,
1598 * there's a danger that one of the calls to try_to_free_swap()
1599 * - most probably a call from __try_to_reclaim_swap() while
1600 * hibernation is allocating its own swap pages for the image,
1601 * but conceivably even a call from memory reclaim - will free
1602 * the swap from a page which has already been recorded in the
1603 * image as a clean swapcache page, and then reuse its swap for
1604 * another page of the image. On waking from hibernation, the
1605 * original page might be freed under memory pressure, then
1606 * later read back in from swap, now with the wrong data.
1608 * Hibernation suspends storage while it is writing the image
1609 * to disk so check that here.
1611 if (pm_suspended_storage())
1614 page = compound_head(page);
1615 delete_from_swap_cache(page);
1621 * Free the swap entry like above, but also try to
1622 * free the page cache entry if it is the last user.
1624 int free_swap_and_cache(swp_entry_t entry)
1626 struct swap_info_struct *p;
1627 struct page *page = NULL;
1628 unsigned char count;
1630 if (non_swap_entry(entry))
1633 p = _swap_info_get(entry);
1635 count = __swap_entry_free(p, entry, 1);
1636 if (count == SWAP_HAS_CACHE &&
1637 !swap_page_trans_huge_swapped(p, entry)) {
1638 page = find_get_page(swap_address_space(entry),
1640 if (page && !trylock_page(page)) {
1645 free_swap_slot(entry);
1649 * Not mapped elsewhere, or swap space full? Free it!
1650 * Also recheck PageSwapCache now page is locked (above).
1652 if (PageSwapCache(page) && !PageWriteback(page) &&
1653 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1654 !swap_page_trans_huge_swapped(p, entry)) {
1655 page = compound_head(page);
1656 delete_from_swap_cache(page);
1665 #ifdef CONFIG_HIBERNATION
1667 * Find the swap type that corresponds to given device (if any).
1669 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1670 * from 0, in which the swap header is expected to be located.
1672 * This is needed for the suspend to disk (aka swsusp).
1674 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1676 struct block_device *bdev = NULL;
1680 bdev = bdget(device);
1682 spin_lock(&swap_lock);
1683 for (type = 0; type < nr_swapfiles; type++) {
1684 struct swap_info_struct *sis = swap_info[type];
1686 if (!(sis->flags & SWP_WRITEOK))
1691 *bdev_p = bdgrab(sis->bdev);
1693 spin_unlock(&swap_lock);
1696 if (bdev == sis->bdev) {
1697 struct swap_extent *se = &sis->first_swap_extent;
1699 if (se->start_block == offset) {
1701 *bdev_p = bdgrab(sis->bdev);
1703 spin_unlock(&swap_lock);
1709 spin_unlock(&swap_lock);
1717 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1718 * corresponding to given index in swap_info (swap type).
1720 sector_t swapdev_block(int type, pgoff_t offset)
1722 struct block_device *bdev;
1724 if ((unsigned int)type >= nr_swapfiles)
1726 if (!(swap_info[type]->flags & SWP_WRITEOK))
1728 return map_swap_entry(swp_entry(type, offset), &bdev);
1732 * Return either the total number of swap pages of given type, or the number
1733 * of free pages of that type (depending on @free)
1735 * This is needed for software suspend
1737 unsigned int count_swap_pages(int type, int free)
1741 spin_lock(&swap_lock);
1742 if ((unsigned int)type < nr_swapfiles) {
1743 struct swap_info_struct *sis = swap_info[type];
1745 spin_lock(&sis->lock);
1746 if (sis->flags & SWP_WRITEOK) {
1749 n -= sis->inuse_pages;
1751 spin_unlock(&sis->lock);
1753 spin_unlock(&swap_lock);
1756 #endif /* CONFIG_HIBERNATION */
1758 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1760 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1764 * No need to decide whether this PTE shares the swap entry with others,
1765 * just let do_wp_page work it out if a write is requested later - to
1766 * force COW, vm_page_prot omits write permission from any private vma.
1768 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1769 unsigned long addr, swp_entry_t entry, struct page *page)
1771 struct page *swapcache;
1772 struct mem_cgroup *memcg;
1778 page = ksm_might_need_to_copy(page, vma, addr);
1779 if (unlikely(!page))
1782 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1788 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1789 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1790 mem_cgroup_cancel_charge(page, memcg, false);
1795 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1796 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1798 set_pte_at(vma->vm_mm, addr, pte,
1799 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1800 if (page == swapcache) {
1801 page_add_anon_rmap(page, vma, addr, false);
1802 mem_cgroup_commit_charge(page, memcg, true, false);
1803 } else { /* ksm created a completely new copy */
1804 page_add_new_anon_rmap(page, vma, addr, false);
1805 mem_cgroup_commit_charge(page, memcg, false, false);
1806 lru_cache_add_active_or_unevictable(page, vma);
1810 * Move the page to the active list so it is not
1811 * immediately swapped out again after swapon.
1813 activate_page(page);
1815 pte_unmap_unlock(pte, ptl);
1817 if (page != swapcache) {
1824 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1825 unsigned long addr, unsigned long end,
1826 swp_entry_t entry, struct page *page)
1828 pte_t swp_pte = swp_entry_to_pte(entry);
1833 * We don't actually need pte lock while scanning for swp_pte: since
1834 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1835 * page table while we're scanning; though it could get zapped, and on
1836 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1837 * of unmatched parts which look like swp_pte, so unuse_pte must
1838 * recheck under pte lock. Scanning without pte lock lets it be
1839 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1841 pte = pte_offset_map(pmd, addr);
1844 * swapoff spends a _lot_ of time in this loop!
1845 * Test inline before going to call unuse_pte.
1847 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1849 ret = unuse_pte(vma, pmd, addr, entry, page);
1852 pte = pte_offset_map(pmd, addr);
1854 } while (pte++, addr += PAGE_SIZE, addr != end);
1860 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1861 unsigned long addr, unsigned long end,
1862 swp_entry_t entry, struct page *page)
1868 pmd = pmd_offset(pud, addr);
1871 next = pmd_addr_end(addr, end);
1872 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1874 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1877 } while (pmd++, addr = next, addr != end);
1881 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1882 unsigned long addr, unsigned long end,
1883 swp_entry_t entry, struct page *page)
1889 pud = pud_offset(p4d, addr);
1891 next = pud_addr_end(addr, end);
1892 if (pud_none_or_clear_bad(pud))
1894 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1897 } while (pud++, addr = next, addr != end);
1901 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1902 unsigned long addr, unsigned long end,
1903 swp_entry_t entry, struct page *page)
1909 p4d = p4d_offset(pgd, addr);
1911 next = p4d_addr_end(addr, end);
1912 if (p4d_none_or_clear_bad(p4d))
1914 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1917 } while (p4d++, addr = next, addr != end);
1921 static int unuse_vma(struct vm_area_struct *vma,
1922 swp_entry_t entry, struct page *page)
1925 unsigned long addr, end, next;
1928 if (page_anon_vma(page)) {
1929 addr = page_address_in_vma(page, vma);
1930 if (addr == -EFAULT)
1933 end = addr + PAGE_SIZE;
1935 addr = vma->vm_start;
1939 pgd = pgd_offset(vma->vm_mm, addr);
1941 next = pgd_addr_end(addr, end);
1942 if (pgd_none_or_clear_bad(pgd))
1944 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1947 } while (pgd++, addr = next, addr != end);
1951 static int unuse_mm(struct mm_struct *mm,
1952 swp_entry_t entry, struct page *page)
1954 struct vm_area_struct *vma;
1957 if (!down_read_trylock(&mm->mmap_sem)) {
1959 * Activate page so shrink_inactive_list is unlikely to unmap
1960 * its ptes while lock is dropped, so swapoff can make progress.
1962 activate_page(page);
1964 down_read(&mm->mmap_sem);
1967 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1968 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1972 up_read(&mm->mmap_sem);
1973 return (ret < 0)? ret: 0;
1977 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1978 * from current position to next entry still in use.
1979 * Recycle to start on reaching the end, returning 0 when empty.
1981 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1982 unsigned int prev, bool frontswap)
1984 unsigned int max = si->max;
1985 unsigned int i = prev;
1986 unsigned char count;
1989 * No need for swap_lock here: we're just looking
1990 * for whether an entry is in use, not modifying it; false
1991 * hits are okay, and sys_swapoff() has already prevented new
1992 * allocations from this area (while holding swap_lock).
2001 * No entries in use at top of swap_map,
2002 * loop back to start and recheck there.
2008 count = READ_ONCE(si->swap_map[i]);
2009 if (count && swap_count(count) != SWAP_MAP_BAD)
2010 if (!frontswap || frontswap_test(si, i))
2012 if ((i % LATENCY_LIMIT) == 0)
2019 * We completely avoid races by reading each swap page in advance,
2020 * and then search for the process using it. All the necessary
2021 * page table adjustments can then be made atomically.
2023 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2024 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2026 int try_to_unuse(unsigned int type, bool frontswap,
2027 unsigned long pages_to_unuse)
2029 struct swap_info_struct *si = swap_info[type];
2030 struct mm_struct *start_mm;
2031 volatile unsigned char *swap_map; /* swap_map is accessed without
2032 * locking. Mark it as volatile
2033 * to prevent compiler doing
2036 unsigned char swcount;
2043 * When searching mms for an entry, a good strategy is to
2044 * start at the first mm we freed the previous entry from
2045 * (though actually we don't notice whether we or coincidence
2046 * freed the entry). Initialize this start_mm with a hold.
2048 * A simpler strategy would be to start at the last mm we
2049 * freed the previous entry from; but that would take less
2050 * advantage of mmlist ordering, which clusters forked mms
2051 * together, child after parent. If we race with dup_mmap(), we
2052 * prefer to resolve parent before child, lest we miss entries
2053 * duplicated after we scanned child: using last mm would invert
2056 start_mm = &init_mm;
2060 * Keep on scanning until all entries have gone. Usually,
2061 * one pass through swap_map is enough, but not necessarily:
2062 * there are races when an instance of an entry might be missed.
2064 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2065 if (signal_pending(current)) {
2071 * Get a page for the entry, using the existing swap
2072 * cache page if there is one. Otherwise, get a clean
2073 * page and read the swap into it.
2075 swap_map = &si->swap_map[i];
2076 entry = swp_entry(type, i);
2077 page = read_swap_cache_async(entry,
2078 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2081 * Either swap_duplicate() failed because entry
2082 * has been freed independently, and will not be
2083 * reused since sys_swapoff() already disabled
2084 * allocation from here, or alloc_page() failed.
2086 swcount = *swap_map;
2088 * We don't hold lock here, so the swap entry could be
2089 * SWAP_MAP_BAD (when the cluster is discarding).
2090 * Instead of fail out, We can just skip the swap
2091 * entry because swapoff will wait for discarding
2094 if (!swcount || swcount == SWAP_MAP_BAD)
2101 * Don't hold on to start_mm if it looks like exiting.
2103 if (atomic_read(&start_mm->mm_users) == 1) {
2105 start_mm = &init_mm;
2110 * Wait for and lock page. When do_swap_page races with
2111 * try_to_unuse, do_swap_page can handle the fault much
2112 * faster than try_to_unuse can locate the entry. This
2113 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2114 * defer to do_swap_page in such a case - in some tests,
2115 * do_swap_page and try_to_unuse repeatedly compete.
2117 wait_on_page_locked(page);
2118 wait_on_page_writeback(page);
2120 wait_on_page_writeback(page);
2123 * Remove all references to entry.
2125 swcount = *swap_map;
2126 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2127 retval = shmem_unuse(entry, page);
2128 /* page has already been unlocked and released */
2133 if (swap_count(swcount) && start_mm != &init_mm)
2134 retval = unuse_mm(start_mm, entry, page);
2136 if (swap_count(*swap_map)) {
2137 int set_start_mm = (*swap_map >= swcount);
2138 struct list_head *p = &start_mm->mmlist;
2139 struct mm_struct *new_start_mm = start_mm;
2140 struct mm_struct *prev_mm = start_mm;
2141 struct mm_struct *mm;
2143 mmget(new_start_mm);
2145 spin_lock(&mmlist_lock);
2146 while (swap_count(*swap_map) && !retval &&
2147 (p = p->next) != &start_mm->mmlist) {
2148 mm = list_entry(p, struct mm_struct, mmlist);
2149 if (!mmget_not_zero(mm))
2151 spin_unlock(&mmlist_lock);
2157 swcount = *swap_map;
2158 if (!swap_count(swcount)) /* any usage ? */
2160 else if (mm == &init_mm)
2163 retval = unuse_mm(mm, entry, page);
2165 if (set_start_mm && *swap_map < swcount) {
2166 mmput(new_start_mm);
2171 spin_lock(&mmlist_lock);
2173 spin_unlock(&mmlist_lock);
2176 start_mm = new_start_mm;
2185 * If a reference remains (rare), we would like to leave
2186 * the page in the swap cache; but try_to_unmap could
2187 * then re-duplicate the entry once we drop page lock,
2188 * so we might loop indefinitely; also, that page could
2189 * not be swapped out to other storage meanwhile. So:
2190 * delete from cache even if there's another reference,
2191 * after ensuring that the data has been saved to disk -
2192 * since if the reference remains (rarer), it will be
2193 * read from disk into another page. Splitting into two
2194 * pages would be incorrect if swap supported "shared
2195 * private" pages, but they are handled by tmpfs files.
2197 * Given how unuse_vma() targets one particular offset
2198 * in an anon_vma, once the anon_vma has been determined,
2199 * this splitting happens to be just what is needed to
2200 * handle where KSM pages have been swapped out: re-reading
2201 * is unnecessarily slow, but we can fix that later on.
2203 if (swap_count(*swap_map) &&
2204 PageDirty(page) && PageSwapCache(page)) {
2205 struct writeback_control wbc = {
2206 .sync_mode = WB_SYNC_NONE,
2209 swap_writepage(compound_head(page), &wbc);
2211 wait_on_page_writeback(page);
2215 * It is conceivable that a racing task removed this page from
2216 * swap cache just before we acquired the page lock at the top,
2217 * or while we dropped it in unuse_mm(). The page might even
2218 * be back in swap cache on another swap area: that we must not
2219 * delete, since it may not have been written out to swap yet.
2221 if (PageSwapCache(page) &&
2222 likely(page_private(page) == entry.val) &&
2223 (!PageTransCompound(page) ||
2224 !swap_page_trans_huge_swapped(si, entry)))
2225 delete_from_swap_cache(compound_head(page));
2228 * So we could skip searching mms once swap count went
2229 * to 1, we did not mark any present ptes as dirty: must
2230 * mark page dirty so shrink_page_list will preserve it.
2237 * Make sure that we aren't completely killing
2238 * interactive performance.
2241 if (frontswap && pages_to_unuse > 0) {
2242 if (!--pages_to_unuse)
2252 * After a successful try_to_unuse, if no swap is now in use, we know
2253 * we can empty the mmlist. swap_lock must be held on entry and exit.
2254 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2255 * added to the mmlist just after page_duplicate - before would be racy.
2257 static void drain_mmlist(void)
2259 struct list_head *p, *next;
2262 for (type = 0; type < nr_swapfiles; type++)
2263 if (swap_info[type]->inuse_pages)
2265 spin_lock(&mmlist_lock);
2266 list_for_each_safe(p, next, &init_mm.mmlist)
2268 spin_unlock(&mmlist_lock);
2272 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2273 * corresponds to page offset for the specified swap entry.
2274 * Note that the type of this function is sector_t, but it returns page offset
2275 * into the bdev, not sector offset.
2277 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2279 struct swap_info_struct *sis;
2280 struct swap_extent *start_se;
2281 struct swap_extent *se;
2284 sis = swap_info[swp_type(entry)];
2287 offset = swp_offset(entry);
2288 start_se = sis->curr_swap_extent;
2292 if (se->start_page <= offset &&
2293 offset < (se->start_page + se->nr_pages)) {
2294 return se->start_block + (offset - se->start_page);
2296 se = list_next_entry(se, list);
2297 sis->curr_swap_extent = se;
2298 BUG_ON(se == start_se); /* It *must* be present */
2303 * Returns the page offset into bdev for the specified page's swap entry.
2305 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2308 entry.val = page_private(page);
2309 return map_swap_entry(entry, bdev) << (PAGE_SHIFT - 9);
2313 * Free all of a swapdev's extent information
2315 static void destroy_swap_extents(struct swap_info_struct *sis)
2317 while (!list_empty(&sis->first_swap_extent.list)) {
2318 struct swap_extent *se;
2320 se = list_first_entry(&sis->first_swap_extent.list,
2321 struct swap_extent, list);
2322 list_del(&se->list);
2326 if (sis->flags & SWP_FILE) {
2327 struct file *swap_file = sis->swap_file;
2328 struct address_space *mapping = swap_file->f_mapping;
2330 sis->flags &= ~SWP_FILE;
2331 mapping->a_ops->swap_deactivate(swap_file);
2336 * Add a block range (and the corresponding page range) into this swapdev's
2337 * extent list. The extent list is kept sorted in page order.
2339 * This function rather assumes that it is called in ascending page order.
2342 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2343 unsigned long nr_pages, sector_t start_block)
2345 struct swap_extent *se;
2346 struct swap_extent *new_se;
2347 struct list_head *lh;
2349 if (start_page == 0) {
2350 se = &sis->first_swap_extent;
2351 sis->curr_swap_extent = se;
2353 se->nr_pages = nr_pages;
2354 se->start_block = start_block;
2357 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2358 se = list_entry(lh, struct swap_extent, list);
2359 BUG_ON(se->start_page + se->nr_pages != start_page);
2360 if (se->start_block + se->nr_pages == start_block) {
2362 se->nr_pages += nr_pages;
2368 * No merge. Insert a new extent, preserving ordering.
2370 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2373 new_se->start_page = start_page;
2374 new_se->nr_pages = nr_pages;
2375 new_se->start_block = start_block;
2377 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2382 * A `swap extent' is a simple thing which maps a contiguous range of pages
2383 * onto a contiguous range of disk blocks. An ordered list of swap extents
2384 * is built at swapon time and is then used at swap_writepage/swap_readpage
2385 * time for locating where on disk a page belongs.
2387 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2388 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2389 * swap files identically.
2391 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2392 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2393 * swapfiles are handled *identically* after swapon time.
2395 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2396 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2397 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2398 * requirements, they are simply tossed out - we will never use those blocks
2401 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2402 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2403 * which will scribble on the fs.
2405 * The amount of disk space which a single swap extent represents varies.
2406 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2407 * extents in the list. To avoid much list walking, we cache the previous
2408 * search location in `curr_swap_extent', and start new searches from there.
2409 * This is extremely effective. The average number of iterations in
2410 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2412 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2414 struct file *swap_file = sis->swap_file;
2415 struct address_space *mapping = swap_file->f_mapping;
2416 struct inode *inode = mapping->host;
2419 if (S_ISBLK(inode->i_mode)) {
2420 ret = add_swap_extent(sis, 0, sis->max, 0);
2425 if (mapping->a_ops->swap_activate) {
2426 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2428 sis->flags |= SWP_FILE;
2429 ret = add_swap_extent(sis, 0, sis->max, 0);
2435 return generic_swapfile_activate(sis, swap_file, span);
2438 static int swap_node(struct swap_info_struct *p)
2440 struct block_device *bdev;
2445 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2447 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2450 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2451 unsigned char *swap_map,
2452 struct swap_cluster_info *cluster_info)
2459 p->prio = --least_priority;
2461 * the plist prio is negated because plist ordering is
2462 * low-to-high, while swap ordering is high-to-low
2464 p->list.prio = -p->prio;
2467 p->avail_lists[i].prio = -p->prio;
2469 if (swap_node(p) == i)
2470 p->avail_lists[i].prio = 1;
2472 p->avail_lists[i].prio = -p->prio;
2475 p->swap_map = swap_map;
2476 p->cluster_info = cluster_info;
2477 p->flags |= SWP_WRITEOK;
2478 atomic_long_add(p->pages, &nr_swap_pages);
2479 total_swap_pages += p->pages;
2481 assert_spin_locked(&swap_lock);
2483 * both lists are plists, and thus priority ordered.
2484 * swap_active_head needs to be priority ordered for swapoff(),
2485 * which on removal of any swap_info_struct with an auto-assigned
2486 * (i.e. negative) priority increments the auto-assigned priority
2487 * of any lower-priority swap_info_structs.
2488 * swap_avail_head needs to be priority ordered for get_swap_page(),
2489 * which allocates swap pages from the highest available priority
2492 plist_add(&p->list, &swap_active_head);
2493 add_to_avail_list(p);
2496 static void enable_swap_info(struct swap_info_struct *p, int prio,
2497 unsigned char *swap_map,
2498 struct swap_cluster_info *cluster_info,
2499 unsigned long *frontswap_map)
2501 frontswap_init(p->type, frontswap_map);
2502 spin_lock(&swap_lock);
2503 spin_lock(&p->lock);
2504 _enable_swap_info(p, prio, swap_map, cluster_info);
2505 spin_unlock(&p->lock);
2506 spin_unlock(&swap_lock);
2509 static void reinsert_swap_info(struct swap_info_struct *p)
2511 spin_lock(&swap_lock);
2512 spin_lock(&p->lock);
2513 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2514 spin_unlock(&p->lock);
2515 spin_unlock(&swap_lock);
2518 bool has_usable_swap(void)
2522 spin_lock(&swap_lock);
2523 if (plist_head_empty(&swap_active_head))
2525 spin_unlock(&swap_lock);
2529 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2531 struct swap_info_struct *p = NULL;
2532 unsigned char *swap_map;
2533 struct swap_cluster_info *cluster_info;
2534 unsigned long *frontswap_map;
2535 struct file *swap_file, *victim;
2536 struct address_space *mapping;
2537 struct inode *inode;
2538 struct filename *pathname;
2540 unsigned int old_block_size;
2542 if (!capable(CAP_SYS_ADMIN))
2545 BUG_ON(!current->mm);
2547 pathname = getname(specialfile);
2548 if (IS_ERR(pathname))
2549 return PTR_ERR(pathname);
2551 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2552 err = PTR_ERR(victim);
2556 mapping = victim->f_mapping;
2557 spin_lock(&swap_lock);
2558 plist_for_each_entry(p, &swap_active_head, list) {
2559 if (p->flags & SWP_WRITEOK) {
2560 if (p->swap_file->f_mapping == mapping) {
2568 spin_unlock(&swap_lock);
2571 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2572 vm_unacct_memory(p->pages);
2575 spin_unlock(&swap_lock);
2578 spin_lock(&p->lock);
2579 del_from_avail_list(p);
2581 struct swap_info_struct *si = p;
2584 plist_for_each_entry_continue(si, &swap_active_head, list) {
2587 for_each_node(nid) {
2588 if (si->avail_lists[nid].prio != 1)
2589 si->avail_lists[nid].prio--;
2594 plist_del(&p->list, &swap_active_head);
2595 atomic_long_sub(p->pages, &nr_swap_pages);
2596 total_swap_pages -= p->pages;
2597 p->flags &= ~SWP_WRITEOK;
2598 spin_unlock(&p->lock);
2599 spin_unlock(&swap_lock);
2601 disable_swap_slots_cache_lock();
2603 set_current_oom_origin();
2604 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2605 clear_current_oom_origin();
2608 /* re-insert swap space back into swap_list */
2609 reinsert_swap_info(p);
2610 reenable_swap_slots_cache_unlock();
2614 reenable_swap_slots_cache_unlock();
2616 flush_work(&p->discard_work);
2618 destroy_swap_extents(p);
2619 if (p->flags & SWP_CONTINUED)
2620 free_swap_count_continuations(p);
2622 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2623 atomic_dec(&nr_rotate_swap);
2625 mutex_lock(&swapon_mutex);
2626 spin_lock(&swap_lock);
2627 spin_lock(&p->lock);
2630 /* wait for anyone still in scan_swap_map */
2631 p->highest_bit = 0; /* cuts scans short */
2632 while (p->flags >= SWP_SCANNING) {
2633 spin_unlock(&p->lock);
2634 spin_unlock(&swap_lock);
2635 schedule_timeout_uninterruptible(1);
2636 spin_lock(&swap_lock);
2637 spin_lock(&p->lock);
2640 swap_file = p->swap_file;
2641 old_block_size = p->old_block_size;
2642 p->swap_file = NULL;
2644 swap_map = p->swap_map;
2646 cluster_info = p->cluster_info;
2647 p->cluster_info = NULL;
2648 frontswap_map = frontswap_map_get(p);
2649 spin_unlock(&p->lock);
2650 spin_unlock(&swap_lock);
2651 frontswap_invalidate_area(p->type);
2652 frontswap_map_set(p, NULL);
2653 mutex_unlock(&swapon_mutex);
2654 free_percpu(p->percpu_cluster);
2655 p->percpu_cluster = NULL;
2657 kvfree(cluster_info);
2658 kvfree(frontswap_map);
2659 /* Destroy swap account information */
2660 swap_cgroup_swapoff(p->type);
2661 exit_swap_address_space(p->type);
2663 inode = mapping->host;
2664 if (S_ISBLK(inode->i_mode)) {
2665 struct block_device *bdev = I_BDEV(inode);
2666 set_blocksize(bdev, old_block_size);
2667 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2670 inode->i_flags &= ~S_SWAPFILE;
2671 inode_unlock(inode);
2673 filp_close(swap_file, NULL);
2676 * Clear the SWP_USED flag after all resources are freed so that swapon
2677 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2678 * not hold p->lock after we cleared its SWP_WRITEOK.
2680 spin_lock(&swap_lock);
2682 spin_unlock(&swap_lock);
2685 atomic_inc(&proc_poll_event);
2686 wake_up_interruptible(&proc_poll_wait);
2689 filp_close(victim, NULL);
2695 #ifdef CONFIG_PROC_FS
2696 static unsigned swaps_poll(struct file *file, poll_table *wait)
2698 struct seq_file *seq = file->private_data;
2700 poll_wait(file, &proc_poll_wait, wait);
2702 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2703 seq->poll_event = atomic_read(&proc_poll_event);
2704 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2707 return POLLIN | POLLRDNORM;
2711 static void *swap_start(struct seq_file *swap, loff_t *pos)
2713 struct swap_info_struct *si;
2717 mutex_lock(&swapon_mutex);
2720 return SEQ_START_TOKEN;
2722 for (type = 0; type < nr_swapfiles; type++) {
2723 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2724 si = swap_info[type];
2725 if (!(si->flags & SWP_USED) || !si->swap_map)
2734 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2736 struct swap_info_struct *si = v;
2739 if (v == SEQ_START_TOKEN)
2742 type = si->type + 1;
2744 for (; type < nr_swapfiles; type++) {
2745 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2746 si = swap_info[type];
2747 if (!(si->flags & SWP_USED) || !si->swap_map)
2756 static void swap_stop(struct seq_file *swap, void *v)
2758 mutex_unlock(&swapon_mutex);
2761 static int swap_show(struct seq_file *swap, void *v)
2763 struct swap_info_struct *si = v;
2767 if (si == SEQ_START_TOKEN) {
2768 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2772 file = si->swap_file;
2773 len = seq_file_path(swap, file, " \t\n\\");
2774 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2775 len < 40 ? 40 - len : 1, " ",
2776 S_ISBLK(file_inode(file)->i_mode) ?
2777 "partition" : "file\t",
2778 si->pages << (PAGE_SHIFT - 10),
2779 si->inuse_pages << (PAGE_SHIFT - 10),
2784 static const struct seq_operations swaps_op = {
2785 .start = swap_start,
2791 static int swaps_open(struct inode *inode, struct file *file)
2793 struct seq_file *seq;
2796 ret = seq_open(file, &swaps_op);
2800 seq = file->private_data;
2801 seq->poll_event = atomic_read(&proc_poll_event);
2805 static const struct file_operations proc_swaps_operations = {
2808 .llseek = seq_lseek,
2809 .release = seq_release,
2813 static int __init procswaps_init(void)
2815 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2818 __initcall(procswaps_init);
2819 #endif /* CONFIG_PROC_FS */
2821 #ifdef MAX_SWAPFILES_CHECK
2822 static int __init max_swapfiles_check(void)
2824 MAX_SWAPFILES_CHECK();
2827 late_initcall(max_swapfiles_check);
2830 static struct swap_info_struct *alloc_swap_info(void)
2832 struct swap_info_struct *p;
2833 struct swap_info_struct *defer = NULL;
2836 int size = sizeof(*p) + nr_node_ids * sizeof(struct plist_node);
2838 p = kvzalloc(size, GFP_KERNEL);
2840 return ERR_PTR(-ENOMEM);
2842 spin_lock(&swap_lock);
2843 for (type = 0; type < nr_swapfiles; type++) {
2844 if (!(swap_info[type]->flags & SWP_USED))
2847 if (type >= MAX_SWAPFILES) {
2848 spin_unlock(&swap_lock);
2850 return ERR_PTR(-EPERM);
2852 if (type >= nr_swapfiles) {
2854 swap_info[type] = p;
2856 * Write swap_info[type] before nr_swapfiles, in case a
2857 * racing procfs swap_start() or swap_next() is reading them.
2858 * (We never shrink nr_swapfiles, we never free this entry.)
2864 p = swap_info[type];
2866 * Do not memset this entry: a racing procfs swap_next()
2867 * would be relying on p->type to remain valid.
2870 INIT_LIST_HEAD(&p->first_swap_extent.list);
2871 plist_node_init(&p->list, 0);
2873 plist_node_init(&p->avail_lists[i], 0);
2874 p->flags = SWP_USED;
2875 spin_unlock(&swap_lock);
2877 spin_lock_init(&p->lock);
2878 spin_lock_init(&p->cont_lock);
2883 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2887 if (S_ISBLK(inode->i_mode)) {
2888 p->bdev = bdgrab(I_BDEV(inode));
2889 error = blkdev_get(p->bdev,
2890 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2895 p->old_block_size = block_size(p->bdev);
2896 error = set_blocksize(p->bdev, PAGE_SIZE);
2899 p->flags |= SWP_BLKDEV;
2900 } else if (S_ISREG(inode->i_mode)) {
2901 p->bdev = inode->i_sb->s_bdev;
2903 if (IS_SWAPFILE(inode))
2913 * Find out how many pages are allowed for a single swap device. There
2914 * are two limiting factors:
2915 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2916 * 2) the number of bits in the swap pte, as defined by the different
2919 * In order to find the largest possible bit mask, a swap entry with
2920 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2921 * decoded to a swp_entry_t again, and finally the swap offset is
2924 * This will mask all the bits from the initial ~0UL mask that can't
2925 * be encoded in either the swp_entry_t or the architecture definition
2928 unsigned long generic_max_swapfile_size(void)
2930 return swp_offset(pte_to_swp_entry(
2931 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2934 /* Can be overridden by an architecture for additional checks. */
2935 __weak unsigned long max_swapfile_size(void)
2937 return generic_max_swapfile_size();
2940 static unsigned long read_swap_header(struct swap_info_struct *p,
2941 union swap_header *swap_header,
2942 struct inode *inode)
2945 unsigned long maxpages;
2946 unsigned long swapfilepages;
2947 unsigned long last_page;
2949 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2950 pr_err("Unable to find swap-space signature\n");
2954 /* swap partition endianess hack... */
2955 if (swab32(swap_header->info.version) == 1) {
2956 swab32s(&swap_header->info.version);
2957 swab32s(&swap_header->info.last_page);
2958 swab32s(&swap_header->info.nr_badpages);
2959 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2961 for (i = 0; i < swap_header->info.nr_badpages; i++)
2962 swab32s(&swap_header->info.badpages[i]);
2964 /* Check the swap header's sub-version */
2965 if (swap_header->info.version != 1) {
2966 pr_warn("Unable to handle swap header version %d\n",
2967 swap_header->info.version);
2972 p->cluster_next = 1;
2975 maxpages = max_swapfile_size();
2976 last_page = swap_header->info.last_page;
2978 pr_warn("Empty swap-file\n");
2981 if (last_page > maxpages) {
2982 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2983 maxpages << (PAGE_SHIFT - 10),
2984 last_page << (PAGE_SHIFT - 10));
2986 if (maxpages > last_page) {
2987 maxpages = last_page + 1;
2988 /* p->max is an unsigned int: don't overflow it */
2989 if ((unsigned int)maxpages == 0)
2990 maxpages = UINT_MAX;
2992 p->highest_bit = maxpages - 1;
2996 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2997 if (swapfilepages && maxpages > swapfilepages) {
2998 pr_warn("Swap area shorter than signature indicates\n");
3001 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3003 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3009 #define SWAP_CLUSTER_INFO_COLS \
3010 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3011 #define SWAP_CLUSTER_SPACE_COLS \
3012 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3013 #define SWAP_CLUSTER_COLS \
3014 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3016 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3017 union swap_header *swap_header,
3018 unsigned char *swap_map,
3019 struct swap_cluster_info *cluster_info,
3020 unsigned long maxpages,
3024 unsigned int nr_good_pages;
3026 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3027 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3028 unsigned long i, idx;
3030 nr_good_pages = maxpages - 1; /* omit header page */
3032 cluster_list_init(&p->free_clusters);
3033 cluster_list_init(&p->discard_clusters);
3035 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3036 unsigned int page_nr = swap_header->info.badpages[i];
3037 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3039 if (page_nr < maxpages) {
3040 swap_map[page_nr] = SWAP_MAP_BAD;
3043 * Haven't marked the cluster free yet, no list
3044 * operation involved
3046 inc_cluster_info_page(p, cluster_info, page_nr);
3050 /* Haven't marked the cluster free yet, no list operation involved */
3051 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3052 inc_cluster_info_page(p, cluster_info, i);
3054 if (nr_good_pages) {
3055 swap_map[0] = SWAP_MAP_BAD;
3057 * Not mark the cluster free yet, no list
3058 * operation involved
3060 inc_cluster_info_page(p, cluster_info, 0);
3062 p->pages = nr_good_pages;
3063 nr_extents = setup_swap_extents(p, span);
3066 nr_good_pages = p->pages;
3068 if (!nr_good_pages) {
3069 pr_warn("Empty swap-file\n");
3078 * Reduce false cache line sharing between cluster_info and
3079 * sharing same address space.
3081 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3082 j = (k + col) % SWAP_CLUSTER_COLS;
3083 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3084 idx = i * SWAP_CLUSTER_COLS + j;
3085 if (idx >= nr_clusters)
3087 if (cluster_count(&cluster_info[idx]))
3089 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3090 cluster_list_add_tail(&p->free_clusters, cluster_info,
3098 * Helper to sys_swapon determining if a given swap
3099 * backing device queue supports DISCARD operations.
3101 static bool swap_discardable(struct swap_info_struct *si)
3103 struct request_queue *q = bdev_get_queue(si->bdev);
3105 if (!q || !blk_queue_discard(q))
3111 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3113 struct swap_info_struct *p;
3114 struct filename *name;
3115 struct file *swap_file = NULL;
3116 struct address_space *mapping;
3119 union swap_header *swap_header;
3122 unsigned long maxpages;
3123 unsigned char *swap_map = NULL;
3124 struct swap_cluster_info *cluster_info = NULL;
3125 unsigned long *frontswap_map = NULL;
3126 struct page *page = NULL;
3127 struct inode *inode = NULL;
3129 if (swap_flags & ~SWAP_FLAGS_VALID)
3132 if (!capable(CAP_SYS_ADMIN))
3135 if (!swap_avail_heads)
3138 p = alloc_swap_info();
3142 INIT_WORK(&p->discard_work, swap_discard_work);
3144 name = getname(specialfile);
3146 error = PTR_ERR(name);
3150 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3151 if (IS_ERR(swap_file)) {
3152 error = PTR_ERR(swap_file);
3157 p->swap_file = swap_file;
3158 mapping = swap_file->f_mapping;
3159 inode = mapping->host;
3161 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3162 error = claim_swapfile(p, inode);
3163 if (unlikely(error))
3167 * Read the swap header.
3169 if (!mapping->a_ops->readpage) {
3173 page = read_mapping_page(mapping, 0, swap_file);
3175 error = PTR_ERR(page);
3178 swap_header = kmap(page);
3180 maxpages = read_swap_header(p, swap_header, inode);
3181 if (unlikely(!maxpages)) {
3186 /* OK, set up the swap map and apply the bad block list */
3187 swap_map = vzalloc(maxpages);
3193 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3194 p->flags |= SWP_STABLE_WRITES;
3196 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3198 unsigned long ci, nr_cluster;
3200 p->flags |= SWP_SOLIDSTATE;
3202 * select a random position to start with to help wear leveling
3205 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3206 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3208 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3210 if (!cluster_info) {
3215 for (ci = 0; ci < nr_cluster; ci++)
3216 spin_lock_init(&((cluster_info + ci)->lock));
3218 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3219 if (!p->percpu_cluster) {
3223 for_each_possible_cpu(cpu) {
3224 struct percpu_cluster *cluster;
3225 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3226 cluster_set_null(&cluster->index);
3229 atomic_inc(&nr_rotate_swap);
3231 error = swap_cgroup_swapon(p->type, maxpages);
3235 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3236 cluster_info, maxpages, &span);
3237 if (unlikely(nr_extents < 0)) {
3241 /* frontswap enabled? set up bit-per-page map for frontswap */
3242 if (IS_ENABLED(CONFIG_FRONTSWAP))
3243 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3246 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3248 * When discard is enabled for swap with no particular
3249 * policy flagged, we set all swap discard flags here in
3250 * order to sustain backward compatibility with older
3251 * swapon(8) releases.
3253 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3257 * By flagging sys_swapon, a sysadmin can tell us to
3258 * either do single-time area discards only, or to just
3259 * perform discards for released swap page-clusters.
3260 * Now it's time to adjust the p->flags accordingly.
3262 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3263 p->flags &= ~SWP_PAGE_DISCARD;
3264 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3265 p->flags &= ~SWP_AREA_DISCARD;
3267 /* issue a swapon-time discard if it's still required */
3268 if (p->flags & SWP_AREA_DISCARD) {
3269 int err = discard_swap(p);
3271 pr_err("swapon: discard_swap(%p): %d\n",
3276 error = init_swap_address_space(p->type, maxpages);
3280 mutex_lock(&swapon_mutex);
3282 if (swap_flags & SWAP_FLAG_PREFER)
3284 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3285 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3287 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3288 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3289 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3290 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3291 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3292 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3293 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3294 (frontswap_map) ? "FS" : "");
3296 mutex_unlock(&swapon_mutex);
3297 atomic_inc(&proc_poll_event);
3298 wake_up_interruptible(&proc_poll_wait);
3300 if (S_ISREG(inode->i_mode))
3301 inode->i_flags |= S_SWAPFILE;
3305 free_percpu(p->percpu_cluster);
3306 p->percpu_cluster = NULL;
3307 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3308 set_blocksize(p->bdev, p->old_block_size);
3309 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3311 destroy_swap_extents(p);
3312 swap_cgroup_swapoff(p->type);
3313 spin_lock(&swap_lock);
3314 p->swap_file = NULL;
3316 spin_unlock(&swap_lock);
3318 kvfree(cluster_info);
3319 kvfree(frontswap_map);
3321 if (inode && S_ISREG(inode->i_mode)) {
3322 inode_unlock(inode);
3325 filp_close(swap_file, NULL);
3328 if (page && !IS_ERR(page)) {
3334 if (inode && S_ISREG(inode->i_mode))
3335 inode_unlock(inode);
3337 enable_swap_slots_cache();
3341 void si_swapinfo(struct sysinfo *val)
3344 unsigned long nr_to_be_unused = 0;
3346 spin_lock(&swap_lock);
3347 for (type = 0; type < nr_swapfiles; type++) {
3348 struct swap_info_struct *si = swap_info[type];
3350 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3351 nr_to_be_unused += si->inuse_pages;
3353 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3354 val->totalswap = total_swap_pages + nr_to_be_unused;
3355 spin_unlock(&swap_lock);
3359 * Verify that a swap entry is valid and increment its swap map count.
3361 * Returns error code in following case.
3363 * - swp_entry is invalid -> EINVAL
3364 * - swp_entry is migration entry -> EINVAL
3365 * - swap-cache reference is requested but there is already one. -> EEXIST
3366 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3367 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3369 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3371 struct swap_info_struct *p;
3372 struct swap_cluster_info *ci;
3373 unsigned long offset, type;
3374 unsigned char count;
3375 unsigned char has_cache;
3378 if (non_swap_entry(entry))
3381 type = swp_type(entry);
3382 if (type >= nr_swapfiles)
3384 p = swap_info[type];
3385 offset = swp_offset(entry);
3386 if (unlikely(offset >= p->max))
3389 ci = lock_cluster_or_swap_info(p, offset);
3391 count = p->swap_map[offset];
3394 * swapin_readahead() doesn't check if a swap entry is valid, so the
3395 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3397 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3402 has_cache = count & SWAP_HAS_CACHE;
3403 count &= ~SWAP_HAS_CACHE;
3406 if (usage == SWAP_HAS_CACHE) {
3408 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3409 if (!has_cache && count)
3410 has_cache = SWAP_HAS_CACHE;
3411 else if (has_cache) /* someone else added cache */
3413 else /* no users remaining */
3416 } else if (count || has_cache) {
3418 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3420 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3422 else if (swap_count_continued(p, offset, count))
3423 count = COUNT_CONTINUED;
3427 err = -ENOENT; /* unused swap entry */
3429 p->swap_map[offset] = count | has_cache;
3432 unlock_cluster_or_swap_info(p, ci);
3437 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3442 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3443 * (in which case its reference count is never incremented).
3445 void swap_shmem_alloc(swp_entry_t entry)
3447 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3451 * Increase reference count of swap entry by 1.
3452 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3453 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3454 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3455 * might occur if a page table entry has got corrupted.
3457 int swap_duplicate(swp_entry_t entry)
3461 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3462 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3467 * @entry: swap entry for which we allocate swap cache.
3469 * Called when allocating swap cache for existing swap entry,
3470 * This can return error codes. Returns 0 at success.
3471 * -EBUSY means there is a swap cache.
3472 * Note: return code is different from swap_duplicate().
3474 int swapcache_prepare(swp_entry_t entry)
3476 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3479 struct swap_info_struct *page_swap_info(struct page *page)
3481 swp_entry_t swap = { .val = page_private(page) };
3482 return swap_info[swp_type(swap)];
3486 * out-of-line __page_file_ methods to avoid include hell.
3488 struct address_space *__page_file_mapping(struct page *page)
3490 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3491 return page_swap_info(page)->swap_file->f_mapping;
3493 EXPORT_SYMBOL_GPL(__page_file_mapping);
3495 pgoff_t __page_file_index(struct page *page)
3497 swp_entry_t swap = { .val = page_private(page) };
3498 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3499 return swp_offset(swap);
3501 EXPORT_SYMBOL_GPL(__page_file_index);
3504 * add_swap_count_continuation - called when a swap count is duplicated
3505 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3506 * page of the original vmalloc'ed swap_map, to hold the continuation count
3507 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3508 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3510 * These continuation pages are seldom referenced: the common paths all work
3511 * on the original swap_map, only referring to a continuation page when the
3512 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3514 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3515 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3516 * can be called after dropping locks.
3518 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3520 struct swap_info_struct *si;
3521 struct swap_cluster_info *ci;
3524 struct page *list_page;
3526 unsigned char count;
3529 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3530 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3532 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3534 si = swap_info_get(entry);
3537 * An acceptable race has occurred since the failing
3538 * __swap_duplicate(): the swap entry has been freed,
3539 * perhaps even the whole swap_map cleared for swapoff.
3544 offset = swp_offset(entry);
3546 ci = lock_cluster(si, offset);
3548 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3550 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3552 * The higher the swap count, the more likely it is that tasks
3553 * will race to add swap count continuation: we need to avoid
3554 * over-provisioning.
3561 spin_unlock(&si->lock);
3566 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3567 * no architecture is using highmem pages for kernel page tables: so it
3568 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3570 head = vmalloc_to_page(si->swap_map + offset);
3571 offset &= ~PAGE_MASK;
3573 spin_lock(&si->cont_lock);
3575 * Page allocation does not initialize the page's lru field,
3576 * but it does always reset its private field.
3578 if (!page_private(head)) {
3579 BUG_ON(count & COUNT_CONTINUED);
3580 INIT_LIST_HEAD(&head->lru);
3581 set_page_private(head, SWP_CONTINUED);
3582 si->flags |= SWP_CONTINUED;
3585 list_for_each_entry(list_page, &head->lru, lru) {
3589 * If the previous map said no continuation, but we've found
3590 * a continuation page, free our allocation and use this one.
3592 if (!(count & COUNT_CONTINUED))
3593 goto out_unlock_cont;
3595 map = kmap_atomic(list_page) + offset;
3600 * If this continuation count now has some space in it,
3601 * free our allocation and use this one.
3603 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3604 goto out_unlock_cont;
3607 list_add_tail(&page->lru, &head->lru);
3608 page = NULL; /* now it's attached, don't free it */
3610 spin_unlock(&si->cont_lock);
3613 spin_unlock(&si->lock);
3621 * swap_count_continued - when the original swap_map count is incremented
3622 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3623 * into, carry if so, or else fail until a new continuation page is allocated;
3624 * when the original swap_map count is decremented from 0 with continuation,
3625 * borrow from the continuation and report whether it still holds more.
3626 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3629 static bool swap_count_continued(struct swap_info_struct *si,
3630 pgoff_t offset, unsigned char count)
3637 head = vmalloc_to_page(si->swap_map + offset);
3638 if (page_private(head) != SWP_CONTINUED) {
3639 BUG_ON(count & COUNT_CONTINUED);
3640 return false; /* need to add count continuation */
3643 spin_lock(&si->cont_lock);
3644 offset &= ~PAGE_MASK;
3645 page = list_entry(head->lru.next, struct page, lru);
3646 map = kmap_atomic(page) + offset;
3648 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3649 goto init_map; /* jump over SWAP_CONT_MAX checks */
3651 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3653 * Think of how you add 1 to 999
3655 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3657 page = list_entry(page->lru.next, struct page, lru);
3658 BUG_ON(page == head);
3659 map = kmap_atomic(page) + offset;
3661 if (*map == SWAP_CONT_MAX) {
3663 page = list_entry(page->lru.next, struct page, lru);
3665 ret = false; /* add count continuation */
3668 map = kmap_atomic(page) + offset;
3669 init_map: *map = 0; /* we didn't zero the page */
3673 page = list_entry(page->lru.prev, struct page, lru);
3674 while (page != head) {
3675 map = kmap_atomic(page) + offset;
3676 *map = COUNT_CONTINUED;
3678 page = list_entry(page->lru.prev, struct page, lru);
3680 ret = true; /* incremented */
3682 } else { /* decrementing */
3684 * Think of how you subtract 1 from 1000
3686 BUG_ON(count != COUNT_CONTINUED);
3687 while (*map == COUNT_CONTINUED) {
3689 page = list_entry(page->lru.next, struct page, lru);
3690 BUG_ON(page == head);
3691 map = kmap_atomic(page) + offset;
3698 page = list_entry(page->lru.prev, struct page, lru);
3699 while (page != head) {
3700 map = kmap_atomic(page) + offset;
3701 *map = SWAP_CONT_MAX | count;
3702 count = COUNT_CONTINUED;
3704 page = list_entry(page->lru.prev, struct page, lru);
3706 ret = count == COUNT_CONTINUED;
3709 spin_unlock(&si->cont_lock);
3714 * free_swap_count_continuations - swapoff free all the continuation pages
3715 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3717 static void free_swap_count_continuations(struct swap_info_struct *si)
3721 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3723 head = vmalloc_to_page(si->swap_map + offset);
3724 if (page_private(head)) {
3725 struct page *page, *next;
3727 list_for_each_entry_safe(page, next, &head->lru, lru) {
3728 list_del(&page->lru);
3735 static int __init swapfile_init(void)
3739 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3741 if (!swap_avail_heads) {
3742 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3747 plist_head_init(&swap_avail_heads[nid]);
3751 subsys_initcall(swapfile_init);