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>
41 #include <asm/pgtable.h>
42 #include <asm/tlbflush.h>
43 #include <linux/swapops.h>
44 #include <linux/swap_cgroup.h>
46 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
48 static void free_swap_count_continuations(struct swap_info_struct *);
49 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
51 DEFINE_SPINLOCK(swap_lock);
52 static unsigned int nr_swapfiles;
53 atomic_long_t nr_swap_pages;
55 * Some modules use swappable objects and may try to swap them out under
56 * memory pressure (via the shrinker). Before doing so, they may wish to
57 * check to see if any swap space is available.
59 EXPORT_SYMBOL_GPL(nr_swap_pages);
60 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
61 long total_swap_pages;
62 static int least_priority;
64 static const char Bad_file[] = "Bad swap file entry ";
65 static const char Unused_file[] = "Unused swap file entry ";
66 static const char Bad_offset[] = "Bad swap offset entry ";
67 static const char Unused_offset[] = "Unused swap offset entry ";
70 * all active swap_info_structs
71 * protected with swap_lock, and ordered by priority.
73 PLIST_HEAD(swap_active_head);
76 * all available (active, not full) swap_info_structs
77 * protected with swap_avail_lock, ordered by priority.
78 * This is used by get_swap_page() instead of swap_active_head
79 * because swap_active_head includes all swap_info_structs,
80 * but get_swap_page() doesn't need to look at full ones.
81 * This uses its own lock instead of swap_lock because when a
82 * swap_info_struct changes between not-full/full, it needs to
83 * add/remove itself to/from this list, but the swap_info_struct->lock
84 * is held and the locking order requires swap_lock to be taken
85 * before any swap_info_struct->lock.
87 static PLIST_HEAD(swap_avail_head);
88 static DEFINE_SPINLOCK(swap_avail_lock);
90 struct swap_info_struct *swap_info[MAX_SWAPFILES];
92 static DEFINE_MUTEX(swapon_mutex);
94 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
95 /* Activity counter to indicate that a swapon or swapoff has occurred */
96 static atomic_t proc_poll_event = ATOMIC_INIT(0);
98 static inline unsigned char swap_count(unsigned char ent)
100 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
103 /* returns 1 if swap entry is freed */
105 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
107 swp_entry_t entry = swp_entry(si->type, offset);
111 page = find_get_page(swap_address_space(entry), swp_offset(entry));
115 * This function is called from scan_swap_map() and it's called
116 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
117 * We have to use trylock for avoiding deadlock. This is a special
118 * case and you should use try_to_free_swap() with explicit lock_page()
119 * in usual operations.
121 if (trylock_page(page)) {
122 ret = try_to_free_swap(page);
130 * swapon tell device that all the old swap contents can be discarded,
131 * to allow the swap device to optimize its wear-levelling.
133 static int discard_swap(struct swap_info_struct *si)
135 struct swap_extent *se;
136 sector_t start_block;
140 /* Do not discard the swap header page! */
141 se = &si->first_swap_extent;
142 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
143 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
145 err = blkdev_issue_discard(si->bdev, start_block,
146 nr_blocks, GFP_KERNEL, 0);
152 list_for_each_entry(se, &si->first_swap_extent.list, list) {
153 start_block = se->start_block << (PAGE_SHIFT - 9);
154 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
156 err = blkdev_issue_discard(si->bdev, start_block,
157 nr_blocks, GFP_KERNEL, 0);
163 return err; /* That will often be -EOPNOTSUPP */
167 * swap allocation tell device that a cluster of swap can now be discarded,
168 * to allow the swap device to optimize its wear-levelling.
170 static void discard_swap_cluster(struct swap_info_struct *si,
171 pgoff_t start_page, pgoff_t nr_pages)
173 struct swap_extent *se = si->curr_swap_extent;
174 int found_extent = 0;
177 if (se->start_page <= start_page &&
178 start_page < se->start_page + se->nr_pages) {
179 pgoff_t offset = start_page - se->start_page;
180 sector_t start_block = se->start_block + offset;
181 sector_t nr_blocks = se->nr_pages - offset;
183 if (nr_blocks > nr_pages)
184 nr_blocks = nr_pages;
185 start_page += nr_blocks;
186 nr_pages -= nr_blocks;
189 si->curr_swap_extent = se;
191 start_block <<= PAGE_SHIFT - 9;
192 nr_blocks <<= PAGE_SHIFT - 9;
193 if (blkdev_issue_discard(si->bdev, start_block,
194 nr_blocks, GFP_NOIO, 0))
198 se = list_next_entry(se, list);
202 #ifdef CONFIG_THP_SWAP
203 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
205 #define SWAPFILE_CLUSTER 256
207 #define LATENCY_LIMIT 256
209 static inline void cluster_set_flag(struct swap_cluster_info *info,
215 static inline unsigned int cluster_count(struct swap_cluster_info *info)
220 static inline void cluster_set_count(struct swap_cluster_info *info,
226 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
227 unsigned int c, unsigned int f)
233 static inline unsigned int cluster_next(struct swap_cluster_info *info)
238 static inline void cluster_set_next(struct swap_cluster_info *info,
244 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
245 unsigned int n, unsigned int f)
251 static inline bool cluster_is_free(struct swap_cluster_info *info)
253 return info->flags & CLUSTER_FLAG_FREE;
256 static inline bool cluster_is_null(struct swap_cluster_info *info)
258 return info->flags & CLUSTER_FLAG_NEXT_NULL;
261 static inline void cluster_set_null(struct swap_cluster_info *info)
263 info->flags = CLUSTER_FLAG_NEXT_NULL;
267 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
268 unsigned long offset)
270 struct swap_cluster_info *ci;
272 ci = si->cluster_info;
274 ci += offset / SWAPFILE_CLUSTER;
275 spin_lock(&ci->lock);
280 static inline void unlock_cluster(struct swap_cluster_info *ci)
283 spin_unlock(&ci->lock);
286 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
287 struct swap_info_struct *si,
288 unsigned long offset)
290 struct swap_cluster_info *ci;
292 ci = lock_cluster(si, offset);
294 spin_lock(&si->lock);
299 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
300 struct swap_cluster_info *ci)
305 spin_unlock(&si->lock);
308 static inline bool cluster_list_empty(struct swap_cluster_list *list)
310 return cluster_is_null(&list->head);
313 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
315 return cluster_next(&list->head);
318 static void cluster_list_init(struct swap_cluster_list *list)
320 cluster_set_null(&list->head);
321 cluster_set_null(&list->tail);
324 static void cluster_list_add_tail(struct swap_cluster_list *list,
325 struct swap_cluster_info *ci,
328 if (cluster_list_empty(list)) {
329 cluster_set_next_flag(&list->head, idx, 0);
330 cluster_set_next_flag(&list->tail, idx, 0);
332 struct swap_cluster_info *ci_tail;
333 unsigned int tail = cluster_next(&list->tail);
336 * Nested cluster lock, but both cluster locks are
337 * only acquired when we held swap_info_struct->lock
340 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
341 cluster_set_next(ci_tail, idx);
342 spin_unlock(&ci_tail->lock);
343 cluster_set_next_flag(&list->tail, idx, 0);
347 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
348 struct swap_cluster_info *ci)
352 idx = cluster_next(&list->head);
353 if (cluster_next(&list->tail) == idx) {
354 cluster_set_null(&list->head);
355 cluster_set_null(&list->tail);
357 cluster_set_next_flag(&list->head,
358 cluster_next(&ci[idx]), 0);
363 /* Add a cluster to discard list and schedule it to do discard */
364 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
368 * If scan_swap_map() can't find a free cluster, it will check
369 * si->swap_map directly. To make sure the discarding cluster isn't
370 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
371 * will be cleared after discard
373 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
374 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
376 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
378 schedule_work(&si->discard_work);
381 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
383 struct swap_cluster_info *ci = si->cluster_info;
385 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
386 cluster_list_add_tail(&si->free_clusters, ci, idx);
390 * Doing discard actually. After a cluster discard is finished, the cluster
391 * will be added to free cluster list. caller should hold si->lock.
393 static void swap_do_scheduled_discard(struct swap_info_struct *si)
395 struct swap_cluster_info *info, *ci;
398 info = si->cluster_info;
400 while (!cluster_list_empty(&si->discard_clusters)) {
401 idx = cluster_list_del_first(&si->discard_clusters, info);
402 spin_unlock(&si->lock);
404 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
407 spin_lock(&si->lock);
408 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
409 __free_cluster(si, idx);
410 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
411 0, SWAPFILE_CLUSTER);
416 static void swap_discard_work(struct work_struct *work)
418 struct swap_info_struct *si;
420 si = container_of(work, struct swap_info_struct, discard_work);
422 spin_lock(&si->lock);
423 swap_do_scheduled_discard(si);
424 spin_unlock(&si->lock);
427 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
429 struct swap_cluster_info *ci = si->cluster_info;
431 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
432 cluster_list_del_first(&si->free_clusters, ci);
433 cluster_set_count_flag(ci + idx, 0, 0);
436 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
438 struct swap_cluster_info *ci = si->cluster_info + idx;
440 VM_BUG_ON(cluster_count(ci) != 0);
442 * If the swap is discardable, prepare discard the cluster
443 * instead of free it immediately. The cluster will be freed
446 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
447 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
448 swap_cluster_schedule_discard(si, idx);
452 __free_cluster(si, idx);
456 * The cluster corresponding to page_nr will be used. The cluster will be
457 * removed from free cluster list and its usage counter will be increased.
459 static void inc_cluster_info_page(struct swap_info_struct *p,
460 struct swap_cluster_info *cluster_info, unsigned long page_nr)
462 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
466 if (cluster_is_free(&cluster_info[idx]))
467 alloc_cluster(p, idx);
469 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
470 cluster_set_count(&cluster_info[idx],
471 cluster_count(&cluster_info[idx]) + 1);
475 * The cluster corresponding to page_nr decreases one usage. If the usage
476 * counter becomes 0, which means no page in the cluster is in using, we can
477 * optionally discard the cluster and add it to free cluster list.
479 static void dec_cluster_info_page(struct swap_info_struct *p,
480 struct swap_cluster_info *cluster_info, unsigned long page_nr)
482 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
487 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
488 cluster_set_count(&cluster_info[idx],
489 cluster_count(&cluster_info[idx]) - 1);
491 if (cluster_count(&cluster_info[idx]) == 0)
492 free_cluster(p, idx);
496 * It's possible scan_swap_map() uses a free cluster in the middle of free
497 * cluster list. Avoiding such abuse to avoid list corruption.
500 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
501 unsigned long offset)
503 struct percpu_cluster *percpu_cluster;
506 offset /= SWAPFILE_CLUSTER;
507 conflict = !cluster_list_empty(&si->free_clusters) &&
508 offset != cluster_list_first(&si->free_clusters) &&
509 cluster_is_free(&si->cluster_info[offset]);
514 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
515 cluster_set_null(&percpu_cluster->index);
520 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
521 * might involve allocating a new cluster for current CPU too.
523 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
524 unsigned long *offset, unsigned long *scan_base)
526 struct percpu_cluster *cluster;
527 struct swap_cluster_info *ci;
529 unsigned long tmp, max;
532 cluster = this_cpu_ptr(si->percpu_cluster);
533 if (cluster_is_null(&cluster->index)) {
534 if (!cluster_list_empty(&si->free_clusters)) {
535 cluster->index = si->free_clusters.head;
536 cluster->next = cluster_next(&cluster->index) *
538 } else if (!cluster_list_empty(&si->discard_clusters)) {
540 * we don't have free cluster but have some clusters in
541 * discarding, do discard now and reclaim them
543 swap_do_scheduled_discard(si);
544 *scan_base = *offset = si->cluster_next;
553 * Other CPUs can use our cluster if they can't find a free cluster,
554 * check if there is still free entry in the cluster
557 max = min_t(unsigned long, si->max,
558 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
560 cluster_set_null(&cluster->index);
563 ci = lock_cluster(si, tmp);
565 if (!si->swap_map[tmp]) {
573 cluster_set_null(&cluster->index);
576 cluster->next = tmp + 1;
582 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
583 unsigned int nr_entries)
585 unsigned int end = offset + nr_entries - 1;
587 if (offset == si->lowest_bit)
588 si->lowest_bit += nr_entries;
589 if (end == si->highest_bit)
590 si->highest_bit -= nr_entries;
591 si->inuse_pages += nr_entries;
592 if (si->inuse_pages == si->pages) {
593 si->lowest_bit = si->max;
595 spin_lock(&swap_avail_lock);
596 plist_del(&si->avail_list, &swap_avail_head);
597 spin_unlock(&swap_avail_lock);
601 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
602 unsigned int nr_entries)
604 unsigned long end = offset + nr_entries - 1;
605 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
607 if (offset < si->lowest_bit)
608 si->lowest_bit = offset;
609 if (end > si->highest_bit) {
610 bool was_full = !si->highest_bit;
612 si->highest_bit = end;
613 if (was_full && (si->flags & SWP_WRITEOK)) {
614 spin_lock(&swap_avail_lock);
615 WARN_ON(!plist_node_empty(&si->avail_list));
616 if (plist_node_empty(&si->avail_list))
617 plist_add(&si->avail_list, &swap_avail_head);
618 spin_unlock(&swap_avail_lock);
621 atomic_long_add(nr_entries, &nr_swap_pages);
622 si->inuse_pages -= nr_entries;
623 if (si->flags & SWP_BLKDEV)
624 swap_slot_free_notify =
625 si->bdev->bd_disk->fops->swap_slot_free_notify;
627 swap_slot_free_notify = NULL;
628 while (offset <= end) {
629 frontswap_invalidate_page(si->type, offset);
630 if (swap_slot_free_notify)
631 swap_slot_free_notify(si->bdev, offset);
636 static int scan_swap_map_slots(struct swap_info_struct *si,
637 unsigned char usage, int nr,
640 struct swap_cluster_info *ci;
641 unsigned long offset;
642 unsigned long scan_base;
643 unsigned long last_in_cluster = 0;
644 int latency_ration = LATENCY_LIMIT;
651 * We try to cluster swap pages by allocating them sequentially
652 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
653 * way, however, we resort to first-free allocation, starting
654 * a new cluster. This prevents us from scattering swap pages
655 * all over the entire swap partition, so that we reduce
656 * overall disk seek times between swap pages. -- sct
657 * But we do now try to find an empty cluster. -Andrea
658 * And we let swap pages go all over an SSD partition. Hugh
661 si->flags += SWP_SCANNING;
662 scan_base = offset = si->cluster_next;
665 if (si->cluster_info) {
666 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
672 if (unlikely(!si->cluster_nr--)) {
673 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
674 si->cluster_nr = SWAPFILE_CLUSTER - 1;
678 spin_unlock(&si->lock);
681 * If seek is expensive, start searching for new cluster from
682 * start of partition, to minimize the span of allocated swap.
683 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
684 * case, just handled by scan_swap_map_try_ssd_cluster() above.
686 scan_base = offset = si->lowest_bit;
687 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
689 /* Locate the first empty (unaligned) cluster */
690 for (; last_in_cluster <= si->highest_bit; offset++) {
691 if (si->swap_map[offset])
692 last_in_cluster = offset + SWAPFILE_CLUSTER;
693 else if (offset == last_in_cluster) {
694 spin_lock(&si->lock);
695 offset -= SWAPFILE_CLUSTER - 1;
696 si->cluster_next = offset;
697 si->cluster_nr = SWAPFILE_CLUSTER - 1;
700 if (unlikely(--latency_ration < 0)) {
702 latency_ration = LATENCY_LIMIT;
707 spin_lock(&si->lock);
708 si->cluster_nr = SWAPFILE_CLUSTER - 1;
712 if (si->cluster_info) {
713 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
714 /* take a break if we already got some slots */
717 if (!scan_swap_map_try_ssd_cluster(si, &offset,
722 if (!(si->flags & SWP_WRITEOK))
724 if (!si->highest_bit)
726 if (offset > si->highest_bit)
727 scan_base = offset = si->lowest_bit;
729 ci = lock_cluster(si, offset);
730 /* reuse swap entry of cache-only swap if not busy. */
731 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
734 spin_unlock(&si->lock);
735 swap_was_freed = __try_to_reclaim_swap(si, offset);
736 spin_lock(&si->lock);
737 /* entry was freed successfully, try to use this again */
740 goto scan; /* check next one */
743 if (si->swap_map[offset]) {
750 si->swap_map[offset] = usage;
751 inc_cluster_info_page(si, si->cluster_info, offset);
754 swap_range_alloc(si, offset, 1);
755 si->cluster_next = offset + 1;
756 slots[n_ret++] = swp_entry(si->type, offset);
758 /* got enough slots or reach max slots? */
759 if ((n_ret == nr) || (offset >= si->highest_bit))
762 /* search for next available slot */
764 /* time to take a break? */
765 if (unlikely(--latency_ration < 0)) {
768 spin_unlock(&si->lock);
770 spin_lock(&si->lock);
771 latency_ration = LATENCY_LIMIT;
774 /* try to get more slots in cluster */
775 if (si->cluster_info) {
776 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
784 /* non-ssd case, still more slots in cluster? */
785 if (si->cluster_nr && !si->swap_map[offset]) {
791 si->flags -= SWP_SCANNING;
795 spin_unlock(&si->lock);
796 while (++offset <= si->highest_bit) {
797 if (!si->swap_map[offset]) {
798 spin_lock(&si->lock);
801 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
802 spin_lock(&si->lock);
805 if (unlikely(--latency_ration < 0)) {
807 latency_ration = LATENCY_LIMIT;
810 offset = si->lowest_bit;
811 while (offset < scan_base) {
812 if (!si->swap_map[offset]) {
813 spin_lock(&si->lock);
816 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
817 spin_lock(&si->lock);
820 if (unlikely(--latency_ration < 0)) {
822 latency_ration = LATENCY_LIMIT;
826 spin_lock(&si->lock);
829 si->flags -= SWP_SCANNING;
833 #ifdef CONFIG_THP_SWAP
834 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
837 struct swap_cluster_info *ci;
838 unsigned long offset, i;
841 if (cluster_list_empty(&si->free_clusters))
844 idx = cluster_list_first(&si->free_clusters);
845 offset = idx * SWAPFILE_CLUSTER;
846 ci = lock_cluster(si, offset);
847 alloc_cluster(si, idx);
848 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, 0);
850 map = si->swap_map + offset;
851 for (i = 0; i < SWAPFILE_CLUSTER; i++)
852 map[i] = SWAP_HAS_CACHE;
854 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
855 *slot = swp_entry(si->type, offset);
860 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
862 unsigned long offset = idx * SWAPFILE_CLUSTER;
863 struct swap_cluster_info *ci;
865 ci = lock_cluster(si, offset);
866 cluster_set_count_flag(ci, 0, 0);
867 free_cluster(si, idx);
869 swap_range_free(si, offset, SWAPFILE_CLUSTER);
872 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
877 #endif /* CONFIG_THP_SWAP */
879 static unsigned long scan_swap_map(struct swap_info_struct *si,
885 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
888 return swp_offset(entry);
894 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
896 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
897 struct swap_info_struct *si, *next;
901 /* Only single cluster request supported */
902 WARN_ON_ONCE(n_goal > 1 && cluster);
904 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
908 if (n_goal > SWAP_BATCH)
911 if (n_goal > avail_pgs)
914 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
916 spin_lock(&swap_avail_lock);
919 plist_for_each_entry_safe(si, next, &swap_avail_head, avail_list) {
920 /* requeue si to after same-priority siblings */
921 plist_requeue(&si->avail_list, &swap_avail_head);
922 spin_unlock(&swap_avail_lock);
923 spin_lock(&si->lock);
924 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
925 spin_lock(&swap_avail_lock);
926 if (plist_node_empty(&si->avail_list)) {
927 spin_unlock(&si->lock);
930 WARN(!si->highest_bit,
931 "swap_info %d in list but !highest_bit\n",
933 WARN(!(si->flags & SWP_WRITEOK),
934 "swap_info %d in list but !SWP_WRITEOK\n",
936 plist_del(&si->avail_list, &swap_avail_head);
937 spin_unlock(&si->lock);
941 n_ret = swap_alloc_cluster(si, swp_entries);
943 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
944 n_goal, swp_entries);
945 spin_unlock(&si->lock);
946 if (n_ret || cluster)
948 pr_debug("scan_swap_map of si %d failed to find offset\n",
951 spin_lock(&swap_avail_lock);
954 * if we got here, it's likely that si was almost full before,
955 * and since scan_swap_map() can drop the si->lock, multiple
956 * callers probably all tried to get a page from the same si
957 * and it filled up before we could get one; or, the si filled
958 * up between us dropping swap_avail_lock and taking si->lock.
959 * Since we dropped the swap_avail_lock, the swap_avail_head
960 * list may have been modified; so if next is still in the
961 * swap_avail_head list then try it, otherwise start over
962 * if we have not gotten any slots.
964 if (plist_node_empty(&next->avail_list))
968 spin_unlock(&swap_avail_lock);
972 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
978 /* The only caller of this function is now suspend routine */
979 swp_entry_t get_swap_page_of_type(int type)
981 struct swap_info_struct *si;
984 si = swap_info[type];
985 spin_lock(&si->lock);
986 if (si && (si->flags & SWP_WRITEOK)) {
987 atomic_long_dec(&nr_swap_pages);
988 /* This is called for allocating swap entry, not cache */
989 offset = scan_swap_map(si, 1);
991 spin_unlock(&si->lock);
992 return swp_entry(type, offset);
994 atomic_long_inc(&nr_swap_pages);
996 spin_unlock(&si->lock);
997 return (swp_entry_t) {0};
1000 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1002 struct swap_info_struct *p;
1003 unsigned long offset, type;
1007 type = swp_type(entry);
1008 if (type >= nr_swapfiles)
1010 p = swap_info[type];
1011 if (!(p->flags & SWP_USED))
1013 offset = swp_offset(entry);
1014 if (offset >= p->max)
1019 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1022 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1025 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1030 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1032 struct swap_info_struct *p;
1034 p = __swap_info_get(entry);
1037 if (!p->swap_map[swp_offset(entry)])
1042 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1048 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1050 struct swap_info_struct *p;
1052 p = _swap_info_get(entry);
1054 spin_lock(&p->lock);
1058 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1059 struct swap_info_struct *q)
1061 struct swap_info_struct *p;
1063 p = _swap_info_get(entry);
1067 spin_unlock(&q->lock);
1069 spin_lock(&p->lock);
1074 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1075 swp_entry_t entry, unsigned char usage)
1077 struct swap_cluster_info *ci;
1078 unsigned long offset = swp_offset(entry);
1079 unsigned char count;
1080 unsigned char has_cache;
1082 ci = lock_cluster_or_swap_info(p, offset);
1084 count = p->swap_map[offset];
1086 has_cache = count & SWAP_HAS_CACHE;
1087 count &= ~SWAP_HAS_CACHE;
1089 if (usage == SWAP_HAS_CACHE) {
1090 VM_BUG_ON(!has_cache);
1092 } else if (count == SWAP_MAP_SHMEM) {
1094 * Or we could insist on shmem.c using a special
1095 * swap_shmem_free() and free_shmem_swap_and_cache()...
1098 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1099 if (count == COUNT_CONTINUED) {
1100 if (swap_count_continued(p, offset, count))
1101 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1103 count = SWAP_MAP_MAX;
1108 usage = count | has_cache;
1109 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1111 unlock_cluster_or_swap_info(p, ci);
1116 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1118 struct swap_cluster_info *ci;
1119 unsigned long offset = swp_offset(entry);
1120 unsigned char count;
1122 ci = lock_cluster(p, offset);
1123 count = p->swap_map[offset];
1124 VM_BUG_ON(count != SWAP_HAS_CACHE);
1125 p->swap_map[offset] = 0;
1126 dec_cluster_info_page(p, p->cluster_info, offset);
1129 mem_cgroup_uncharge_swap(entry, 1);
1130 swap_range_free(p, offset, 1);
1134 * Caller has made sure that the swap device corresponding to entry
1135 * is still around or has not been recycled.
1137 void swap_free(swp_entry_t entry)
1139 struct swap_info_struct *p;
1141 p = _swap_info_get(entry);
1143 if (!__swap_entry_free(p, entry, 1))
1144 free_swap_slot(entry);
1149 * Called after dropping swapcache to decrease refcnt to swap entries.
1151 static void swapcache_free(swp_entry_t entry)
1153 struct swap_info_struct *p;
1155 p = _swap_info_get(entry);
1157 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1158 free_swap_slot(entry);
1162 #ifdef CONFIG_THP_SWAP
1163 static void swapcache_free_cluster(swp_entry_t entry)
1165 unsigned long offset = swp_offset(entry);
1166 unsigned long idx = offset / SWAPFILE_CLUSTER;
1167 struct swap_cluster_info *ci;
1168 struct swap_info_struct *si;
1172 si = swap_info_get(entry);
1176 ci = lock_cluster(si, offset);
1177 map = si->swap_map + offset;
1178 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1179 VM_BUG_ON(map[i] != SWAP_HAS_CACHE);
1183 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1184 swap_free_cluster(si, idx);
1185 spin_unlock(&si->lock);
1188 static inline void swapcache_free_cluster(swp_entry_t entry)
1191 #endif /* CONFIG_THP_SWAP */
1193 void put_swap_page(struct page *page, swp_entry_t entry)
1195 if (!PageTransHuge(page))
1196 swapcache_free(entry);
1198 swapcache_free_cluster(entry);
1201 void swapcache_free_entries(swp_entry_t *entries, int n)
1203 struct swap_info_struct *p, *prev;
1211 for (i = 0; i < n; ++i) {
1212 p = swap_info_get_cont(entries[i], prev);
1214 swap_entry_free(p, entries[i]);
1218 spin_unlock(&p->lock);
1222 * How many references to page are currently swapped out?
1223 * This does not give an exact answer when swap count is continued,
1224 * but does include the high COUNT_CONTINUED flag to allow for that.
1226 int page_swapcount(struct page *page)
1229 struct swap_info_struct *p;
1230 struct swap_cluster_info *ci;
1232 unsigned long offset;
1234 entry.val = page_private(page);
1235 p = _swap_info_get(entry);
1237 offset = swp_offset(entry);
1238 ci = lock_cluster_or_swap_info(p, offset);
1239 count = swap_count(p->swap_map[offset]);
1240 unlock_cluster_or_swap_info(p, ci);
1245 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1248 pgoff_t offset = swp_offset(entry);
1249 struct swap_cluster_info *ci;
1251 ci = lock_cluster_or_swap_info(si, offset);
1252 count = swap_count(si->swap_map[offset]);
1253 unlock_cluster_or_swap_info(si, ci);
1258 * How many references to @entry are currently swapped out?
1259 * This does not give an exact answer when swap count is continued,
1260 * but does include the high COUNT_CONTINUED flag to allow for that.
1262 int __swp_swapcount(swp_entry_t entry)
1265 struct swap_info_struct *si;
1267 si = __swap_info_get(entry);
1269 count = swap_swapcount(si, entry);
1274 * How many references to @entry are currently swapped out?
1275 * This considers COUNT_CONTINUED so it returns exact answer.
1277 int swp_swapcount(swp_entry_t entry)
1279 int count, tmp_count, n;
1280 struct swap_info_struct *p;
1281 struct swap_cluster_info *ci;
1286 p = _swap_info_get(entry);
1290 offset = swp_offset(entry);
1292 ci = lock_cluster_or_swap_info(p, offset);
1294 count = swap_count(p->swap_map[offset]);
1295 if (!(count & COUNT_CONTINUED))
1298 count &= ~COUNT_CONTINUED;
1299 n = SWAP_MAP_MAX + 1;
1301 page = vmalloc_to_page(p->swap_map + offset);
1302 offset &= ~PAGE_MASK;
1303 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1306 page = list_next_entry(page, lru);
1307 map = kmap_atomic(page);
1308 tmp_count = map[offset];
1311 count += (tmp_count & ~COUNT_CONTINUED) * n;
1312 n *= (SWAP_CONT_MAX + 1);
1313 } while (tmp_count & COUNT_CONTINUED);
1315 unlock_cluster_or_swap_info(p, ci);
1320 * We can write to an anon page without COW if there are no other references
1321 * to it. And as a side-effect, free up its swap: because the old content
1322 * on disk will never be read, and seeking back there to write new content
1323 * later would only waste time away from clustering.
1325 * NOTE: total_mapcount should not be relied upon by the caller if
1326 * reuse_swap_page() returns false, but it may be always overwritten
1327 * (see the other implementation for CONFIG_SWAP=n).
1329 bool reuse_swap_page(struct page *page, int *total_mapcount)
1333 VM_BUG_ON_PAGE(!PageLocked(page), page);
1334 if (unlikely(PageKsm(page)))
1336 count = page_trans_huge_mapcount(page, total_mapcount);
1337 if (count <= 1 && PageSwapCache(page)) {
1338 count += page_swapcount(page);
1341 if (!PageWriteback(page)) {
1342 delete_from_swap_cache(page);
1346 struct swap_info_struct *p;
1348 entry.val = page_private(page);
1349 p = swap_info_get(entry);
1350 if (p->flags & SWP_STABLE_WRITES) {
1351 spin_unlock(&p->lock);
1354 spin_unlock(&p->lock);
1362 * If swap is getting full, or if there are no more mappings of this page,
1363 * then try_to_free_swap is called to free its swap space.
1365 int try_to_free_swap(struct page *page)
1367 VM_BUG_ON_PAGE(!PageLocked(page), page);
1369 if (!PageSwapCache(page))
1371 if (PageWriteback(page))
1373 if (page_swapcount(page))
1377 * Once hibernation has begun to create its image of memory,
1378 * there's a danger that one of the calls to try_to_free_swap()
1379 * - most probably a call from __try_to_reclaim_swap() while
1380 * hibernation is allocating its own swap pages for the image,
1381 * but conceivably even a call from memory reclaim - will free
1382 * the swap from a page which has already been recorded in the
1383 * image as a clean swapcache page, and then reuse its swap for
1384 * another page of the image. On waking from hibernation, the
1385 * original page might be freed under memory pressure, then
1386 * later read back in from swap, now with the wrong data.
1388 * Hibernation suspends storage while it is writing the image
1389 * to disk so check that here.
1391 if (pm_suspended_storage())
1394 delete_from_swap_cache(page);
1400 * Free the swap entry like above, but also try to
1401 * free the page cache entry if it is the last user.
1403 int free_swap_and_cache(swp_entry_t entry)
1405 struct swap_info_struct *p;
1406 struct page *page = NULL;
1407 unsigned char count;
1409 if (non_swap_entry(entry))
1412 p = _swap_info_get(entry);
1414 count = __swap_entry_free(p, entry, 1);
1415 if (count == SWAP_HAS_CACHE) {
1416 page = find_get_page(swap_address_space(entry),
1418 if (page && !trylock_page(page)) {
1423 free_swap_slot(entry);
1427 * Not mapped elsewhere, or swap space full? Free it!
1428 * Also recheck PageSwapCache now page is locked (above).
1430 if (PageSwapCache(page) && !PageWriteback(page) &&
1431 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1432 !swap_swapcount(p, entry)) {
1433 delete_from_swap_cache(page);
1442 #ifdef CONFIG_HIBERNATION
1444 * Find the swap type that corresponds to given device (if any).
1446 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1447 * from 0, in which the swap header is expected to be located.
1449 * This is needed for the suspend to disk (aka swsusp).
1451 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1453 struct block_device *bdev = NULL;
1457 bdev = bdget(device);
1459 spin_lock(&swap_lock);
1460 for (type = 0; type < nr_swapfiles; type++) {
1461 struct swap_info_struct *sis = swap_info[type];
1463 if (!(sis->flags & SWP_WRITEOK))
1468 *bdev_p = bdgrab(sis->bdev);
1470 spin_unlock(&swap_lock);
1473 if (bdev == sis->bdev) {
1474 struct swap_extent *se = &sis->first_swap_extent;
1476 if (se->start_block == offset) {
1478 *bdev_p = bdgrab(sis->bdev);
1480 spin_unlock(&swap_lock);
1486 spin_unlock(&swap_lock);
1494 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1495 * corresponding to given index in swap_info (swap type).
1497 sector_t swapdev_block(int type, pgoff_t offset)
1499 struct block_device *bdev;
1501 if ((unsigned int)type >= nr_swapfiles)
1503 if (!(swap_info[type]->flags & SWP_WRITEOK))
1505 return map_swap_entry(swp_entry(type, offset), &bdev);
1509 * Return either the total number of swap pages of given type, or the number
1510 * of free pages of that type (depending on @free)
1512 * This is needed for software suspend
1514 unsigned int count_swap_pages(int type, int free)
1518 spin_lock(&swap_lock);
1519 if ((unsigned int)type < nr_swapfiles) {
1520 struct swap_info_struct *sis = swap_info[type];
1522 spin_lock(&sis->lock);
1523 if (sis->flags & SWP_WRITEOK) {
1526 n -= sis->inuse_pages;
1528 spin_unlock(&sis->lock);
1530 spin_unlock(&swap_lock);
1533 #endif /* CONFIG_HIBERNATION */
1535 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1537 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1541 * No need to decide whether this PTE shares the swap entry with others,
1542 * just let do_wp_page work it out if a write is requested later - to
1543 * force COW, vm_page_prot omits write permission from any private vma.
1545 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1546 unsigned long addr, swp_entry_t entry, struct page *page)
1548 struct page *swapcache;
1549 struct mem_cgroup *memcg;
1555 page = ksm_might_need_to_copy(page, vma, addr);
1556 if (unlikely(!page))
1559 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1565 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1566 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1567 mem_cgroup_cancel_charge(page, memcg, false);
1572 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1573 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1575 set_pte_at(vma->vm_mm, addr, pte,
1576 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1577 if (page == swapcache) {
1578 page_add_anon_rmap(page, vma, addr, false);
1579 mem_cgroup_commit_charge(page, memcg, true, false);
1580 } else { /* ksm created a completely new copy */
1581 page_add_new_anon_rmap(page, vma, addr, false);
1582 mem_cgroup_commit_charge(page, memcg, false, false);
1583 lru_cache_add_active_or_unevictable(page, vma);
1587 * Move the page to the active list so it is not
1588 * immediately swapped out again after swapon.
1590 activate_page(page);
1592 pte_unmap_unlock(pte, ptl);
1594 if (page != swapcache) {
1601 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1602 unsigned long addr, unsigned long end,
1603 swp_entry_t entry, struct page *page)
1605 pte_t swp_pte = swp_entry_to_pte(entry);
1610 * We don't actually need pte lock while scanning for swp_pte: since
1611 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1612 * page table while we're scanning; though it could get zapped, and on
1613 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1614 * of unmatched parts which look like swp_pte, so unuse_pte must
1615 * recheck under pte lock. Scanning without pte lock lets it be
1616 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1618 pte = pte_offset_map(pmd, addr);
1621 * swapoff spends a _lot_ of time in this loop!
1622 * Test inline before going to call unuse_pte.
1624 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1626 ret = unuse_pte(vma, pmd, addr, entry, page);
1629 pte = pte_offset_map(pmd, addr);
1631 } while (pte++, addr += PAGE_SIZE, addr != end);
1637 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1638 unsigned long addr, unsigned long end,
1639 swp_entry_t entry, struct page *page)
1645 pmd = pmd_offset(pud, addr);
1648 next = pmd_addr_end(addr, end);
1649 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1651 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1654 } while (pmd++, addr = next, addr != end);
1658 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1659 unsigned long addr, unsigned long end,
1660 swp_entry_t entry, struct page *page)
1666 pud = pud_offset(p4d, addr);
1668 next = pud_addr_end(addr, end);
1669 if (pud_none_or_clear_bad(pud))
1671 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1674 } while (pud++, addr = next, addr != end);
1678 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1679 unsigned long addr, unsigned long end,
1680 swp_entry_t entry, struct page *page)
1686 p4d = p4d_offset(pgd, addr);
1688 next = p4d_addr_end(addr, end);
1689 if (p4d_none_or_clear_bad(p4d))
1691 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1694 } while (p4d++, addr = next, addr != end);
1698 static int unuse_vma(struct vm_area_struct *vma,
1699 swp_entry_t entry, struct page *page)
1702 unsigned long addr, end, next;
1705 if (page_anon_vma(page)) {
1706 addr = page_address_in_vma(page, vma);
1707 if (addr == -EFAULT)
1710 end = addr + PAGE_SIZE;
1712 addr = vma->vm_start;
1716 pgd = pgd_offset(vma->vm_mm, addr);
1718 next = pgd_addr_end(addr, end);
1719 if (pgd_none_or_clear_bad(pgd))
1721 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1724 } while (pgd++, addr = next, addr != end);
1728 static int unuse_mm(struct mm_struct *mm,
1729 swp_entry_t entry, struct page *page)
1731 struct vm_area_struct *vma;
1734 if (!down_read_trylock(&mm->mmap_sem)) {
1736 * Activate page so shrink_inactive_list is unlikely to unmap
1737 * its ptes while lock is dropped, so swapoff can make progress.
1739 activate_page(page);
1741 down_read(&mm->mmap_sem);
1744 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1745 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1749 up_read(&mm->mmap_sem);
1750 return (ret < 0)? ret: 0;
1754 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1755 * from current position to next entry still in use.
1756 * Recycle to start on reaching the end, returning 0 when empty.
1758 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1759 unsigned int prev, bool frontswap)
1761 unsigned int max = si->max;
1762 unsigned int i = prev;
1763 unsigned char count;
1766 * No need for swap_lock here: we're just looking
1767 * for whether an entry is in use, not modifying it; false
1768 * hits are okay, and sys_swapoff() has already prevented new
1769 * allocations from this area (while holding swap_lock).
1778 * No entries in use at top of swap_map,
1779 * loop back to start and recheck there.
1785 count = READ_ONCE(si->swap_map[i]);
1786 if (count && swap_count(count) != SWAP_MAP_BAD)
1787 if (!frontswap || frontswap_test(si, i))
1789 if ((i % LATENCY_LIMIT) == 0)
1796 * We completely avoid races by reading each swap page in advance,
1797 * and then search for the process using it. All the necessary
1798 * page table adjustments can then be made atomically.
1800 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1801 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1803 int try_to_unuse(unsigned int type, bool frontswap,
1804 unsigned long pages_to_unuse)
1806 struct swap_info_struct *si = swap_info[type];
1807 struct mm_struct *start_mm;
1808 volatile unsigned char *swap_map; /* swap_map is accessed without
1809 * locking. Mark it as volatile
1810 * to prevent compiler doing
1813 unsigned char swcount;
1820 * When searching mms for an entry, a good strategy is to
1821 * start at the first mm we freed the previous entry from
1822 * (though actually we don't notice whether we or coincidence
1823 * freed the entry). Initialize this start_mm with a hold.
1825 * A simpler strategy would be to start at the last mm we
1826 * freed the previous entry from; but that would take less
1827 * advantage of mmlist ordering, which clusters forked mms
1828 * together, child after parent. If we race with dup_mmap(), we
1829 * prefer to resolve parent before child, lest we miss entries
1830 * duplicated after we scanned child: using last mm would invert
1833 start_mm = &init_mm;
1837 * Keep on scanning until all entries have gone. Usually,
1838 * one pass through swap_map is enough, but not necessarily:
1839 * there are races when an instance of an entry might be missed.
1841 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1842 if (signal_pending(current)) {
1848 * Get a page for the entry, using the existing swap
1849 * cache page if there is one. Otherwise, get a clean
1850 * page and read the swap into it.
1852 swap_map = &si->swap_map[i];
1853 entry = swp_entry(type, i);
1854 page = read_swap_cache_async(entry,
1855 GFP_HIGHUSER_MOVABLE, NULL, 0);
1858 * Either swap_duplicate() failed because entry
1859 * has been freed independently, and will not be
1860 * reused since sys_swapoff() already disabled
1861 * allocation from here, or alloc_page() failed.
1863 swcount = *swap_map;
1865 * We don't hold lock here, so the swap entry could be
1866 * SWAP_MAP_BAD (when the cluster is discarding).
1867 * Instead of fail out, We can just skip the swap
1868 * entry because swapoff will wait for discarding
1871 if (!swcount || swcount == SWAP_MAP_BAD)
1878 * Don't hold on to start_mm if it looks like exiting.
1880 if (atomic_read(&start_mm->mm_users) == 1) {
1882 start_mm = &init_mm;
1887 * Wait for and lock page. When do_swap_page races with
1888 * try_to_unuse, do_swap_page can handle the fault much
1889 * faster than try_to_unuse can locate the entry. This
1890 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1891 * defer to do_swap_page in such a case - in some tests,
1892 * do_swap_page and try_to_unuse repeatedly compete.
1894 wait_on_page_locked(page);
1895 wait_on_page_writeback(page);
1897 wait_on_page_writeback(page);
1900 * Remove all references to entry.
1902 swcount = *swap_map;
1903 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1904 retval = shmem_unuse(entry, page);
1905 /* page has already been unlocked and released */
1910 if (swap_count(swcount) && start_mm != &init_mm)
1911 retval = unuse_mm(start_mm, entry, page);
1913 if (swap_count(*swap_map)) {
1914 int set_start_mm = (*swap_map >= swcount);
1915 struct list_head *p = &start_mm->mmlist;
1916 struct mm_struct *new_start_mm = start_mm;
1917 struct mm_struct *prev_mm = start_mm;
1918 struct mm_struct *mm;
1920 mmget(new_start_mm);
1922 spin_lock(&mmlist_lock);
1923 while (swap_count(*swap_map) && !retval &&
1924 (p = p->next) != &start_mm->mmlist) {
1925 mm = list_entry(p, struct mm_struct, mmlist);
1926 if (!mmget_not_zero(mm))
1928 spin_unlock(&mmlist_lock);
1934 swcount = *swap_map;
1935 if (!swap_count(swcount)) /* any usage ? */
1937 else if (mm == &init_mm)
1940 retval = unuse_mm(mm, entry, page);
1942 if (set_start_mm && *swap_map < swcount) {
1943 mmput(new_start_mm);
1948 spin_lock(&mmlist_lock);
1950 spin_unlock(&mmlist_lock);
1953 start_mm = new_start_mm;
1962 * If a reference remains (rare), we would like to leave
1963 * the page in the swap cache; but try_to_unmap could
1964 * then re-duplicate the entry once we drop page lock,
1965 * so we might loop indefinitely; also, that page could
1966 * not be swapped out to other storage meanwhile. So:
1967 * delete from cache even if there's another reference,
1968 * after ensuring that the data has been saved to disk -
1969 * since if the reference remains (rarer), it will be
1970 * read from disk into another page. Splitting into two
1971 * pages would be incorrect if swap supported "shared
1972 * private" pages, but they are handled by tmpfs files.
1974 * Given how unuse_vma() targets one particular offset
1975 * in an anon_vma, once the anon_vma has been determined,
1976 * this splitting happens to be just what is needed to
1977 * handle where KSM pages have been swapped out: re-reading
1978 * is unnecessarily slow, but we can fix that later on.
1980 if (swap_count(*swap_map) &&
1981 PageDirty(page) && PageSwapCache(page)) {
1982 struct writeback_control wbc = {
1983 .sync_mode = WB_SYNC_NONE,
1986 swap_writepage(page, &wbc);
1988 wait_on_page_writeback(page);
1992 * It is conceivable that a racing task removed this page from
1993 * swap cache just before we acquired the page lock at the top,
1994 * or while we dropped it in unuse_mm(). The page might even
1995 * be back in swap cache on another swap area: that we must not
1996 * delete, since it may not have been written out to swap yet.
1998 if (PageSwapCache(page) &&
1999 likely(page_private(page) == entry.val))
2000 delete_from_swap_cache(page);
2003 * So we could skip searching mms once swap count went
2004 * to 1, we did not mark any present ptes as dirty: must
2005 * mark page dirty so shrink_page_list will preserve it.
2012 * Make sure that we aren't completely killing
2013 * interactive performance.
2016 if (frontswap && pages_to_unuse > 0) {
2017 if (!--pages_to_unuse)
2027 * After a successful try_to_unuse, if no swap is now in use, we know
2028 * we can empty the mmlist. swap_lock must be held on entry and exit.
2029 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2030 * added to the mmlist just after page_duplicate - before would be racy.
2032 static void drain_mmlist(void)
2034 struct list_head *p, *next;
2037 for (type = 0; type < nr_swapfiles; type++)
2038 if (swap_info[type]->inuse_pages)
2040 spin_lock(&mmlist_lock);
2041 list_for_each_safe(p, next, &init_mm.mmlist)
2043 spin_unlock(&mmlist_lock);
2047 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2048 * corresponds to page offset for the specified swap entry.
2049 * Note that the type of this function is sector_t, but it returns page offset
2050 * into the bdev, not sector offset.
2052 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2054 struct swap_info_struct *sis;
2055 struct swap_extent *start_se;
2056 struct swap_extent *se;
2059 sis = swap_info[swp_type(entry)];
2062 offset = swp_offset(entry);
2063 start_se = sis->curr_swap_extent;
2067 if (se->start_page <= offset &&
2068 offset < (se->start_page + se->nr_pages)) {
2069 return se->start_block + (offset - se->start_page);
2071 se = list_next_entry(se, list);
2072 sis->curr_swap_extent = se;
2073 BUG_ON(se == start_se); /* It *must* be present */
2078 * Returns the page offset into bdev for the specified page's swap entry.
2080 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2083 entry.val = page_private(page);
2084 return map_swap_entry(entry, bdev);
2088 * Free all of a swapdev's extent information
2090 static void destroy_swap_extents(struct swap_info_struct *sis)
2092 while (!list_empty(&sis->first_swap_extent.list)) {
2093 struct swap_extent *se;
2095 se = list_first_entry(&sis->first_swap_extent.list,
2096 struct swap_extent, list);
2097 list_del(&se->list);
2101 if (sis->flags & SWP_FILE) {
2102 struct file *swap_file = sis->swap_file;
2103 struct address_space *mapping = swap_file->f_mapping;
2105 sis->flags &= ~SWP_FILE;
2106 mapping->a_ops->swap_deactivate(swap_file);
2111 * Add a block range (and the corresponding page range) into this swapdev's
2112 * extent list. The extent list is kept sorted in page order.
2114 * This function rather assumes that it is called in ascending page order.
2117 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2118 unsigned long nr_pages, sector_t start_block)
2120 struct swap_extent *se;
2121 struct swap_extent *new_se;
2122 struct list_head *lh;
2124 if (start_page == 0) {
2125 se = &sis->first_swap_extent;
2126 sis->curr_swap_extent = se;
2128 se->nr_pages = nr_pages;
2129 se->start_block = start_block;
2132 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2133 se = list_entry(lh, struct swap_extent, list);
2134 BUG_ON(se->start_page + se->nr_pages != start_page);
2135 if (se->start_block + se->nr_pages == start_block) {
2137 se->nr_pages += nr_pages;
2143 * No merge. Insert a new extent, preserving ordering.
2145 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2148 new_se->start_page = start_page;
2149 new_se->nr_pages = nr_pages;
2150 new_se->start_block = start_block;
2152 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2157 * A `swap extent' is a simple thing which maps a contiguous range of pages
2158 * onto a contiguous range of disk blocks. An ordered list of swap extents
2159 * is built at swapon time and is then used at swap_writepage/swap_readpage
2160 * time for locating where on disk a page belongs.
2162 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2163 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2164 * swap files identically.
2166 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2167 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2168 * swapfiles are handled *identically* after swapon time.
2170 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2171 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2172 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2173 * requirements, they are simply tossed out - we will never use those blocks
2176 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2177 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2178 * which will scribble on the fs.
2180 * The amount of disk space which a single swap extent represents varies.
2181 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2182 * extents in the list. To avoid much list walking, we cache the previous
2183 * search location in `curr_swap_extent', and start new searches from there.
2184 * This is extremely effective. The average number of iterations in
2185 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2187 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2189 struct file *swap_file = sis->swap_file;
2190 struct address_space *mapping = swap_file->f_mapping;
2191 struct inode *inode = mapping->host;
2194 if (S_ISBLK(inode->i_mode)) {
2195 ret = add_swap_extent(sis, 0, sis->max, 0);
2200 if (mapping->a_ops->swap_activate) {
2201 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2203 sis->flags |= SWP_FILE;
2204 ret = add_swap_extent(sis, 0, sis->max, 0);
2210 return generic_swapfile_activate(sis, swap_file, span);
2213 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2214 unsigned char *swap_map,
2215 struct swap_cluster_info *cluster_info)
2220 p->prio = --least_priority;
2222 * the plist prio is negated because plist ordering is
2223 * low-to-high, while swap ordering is high-to-low
2225 p->list.prio = -p->prio;
2226 p->avail_list.prio = -p->prio;
2227 p->swap_map = swap_map;
2228 p->cluster_info = cluster_info;
2229 p->flags |= SWP_WRITEOK;
2230 atomic_long_add(p->pages, &nr_swap_pages);
2231 total_swap_pages += p->pages;
2233 assert_spin_locked(&swap_lock);
2235 * both lists are plists, and thus priority ordered.
2236 * swap_active_head needs to be priority ordered for swapoff(),
2237 * which on removal of any swap_info_struct with an auto-assigned
2238 * (i.e. negative) priority increments the auto-assigned priority
2239 * of any lower-priority swap_info_structs.
2240 * swap_avail_head needs to be priority ordered for get_swap_page(),
2241 * which allocates swap pages from the highest available priority
2244 plist_add(&p->list, &swap_active_head);
2245 spin_lock(&swap_avail_lock);
2246 plist_add(&p->avail_list, &swap_avail_head);
2247 spin_unlock(&swap_avail_lock);
2250 static void enable_swap_info(struct swap_info_struct *p, int prio,
2251 unsigned char *swap_map,
2252 struct swap_cluster_info *cluster_info,
2253 unsigned long *frontswap_map)
2255 frontswap_init(p->type, frontswap_map);
2256 spin_lock(&swap_lock);
2257 spin_lock(&p->lock);
2258 _enable_swap_info(p, prio, swap_map, cluster_info);
2259 spin_unlock(&p->lock);
2260 spin_unlock(&swap_lock);
2263 static void reinsert_swap_info(struct swap_info_struct *p)
2265 spin_lock(&swap_lock);
2266 spin_lock(&p->lock);
2267 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2268 spin_unlock(&p->lock);
2269 spin_unlock(&swap_lock);
2272 bool has_usable_swap(void)
2276 spin_lock(&swap_lock);
2277 if (plist_head_empty(&swap_active_head))
2279 spin_unlock(&swap_lock);
2283 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2285 struct swap_info_struct *p = NULL;
2286 unsigned char *swap_map;
2287 struct swap_cluster_info *cluster_info;
2288 unsigned long *frontswap_map;
2289 struct file *swap_file, *victim;
2290 struct address_space *mapping;
2291 struct inode *inode;
2292 struct filename *pathname;
2294 unsigned int old_block_size;
2296 if (!capable(CAP_SYS_ADMIN))
2299 BUG_ON(!current->mm);
2301 pathname = getname(specialfile);
2302 if (IS_ERR(pathname))
2303 return PTR_ERR(pathname);
2305 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2306 err = PTR_ERR(victim);
2310 mapping = victim->f_mapping;
2311 spin_lock(&swap_lock);
2312 plist_for_each_entry(p, &swap_active_head, list) {
2313 if (p->flags & SWP_WRITEOK) {
2314 if (p->swap_file->f_mapping == mapping) {
2322 spin_unlock(&swap_lock);
2325 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2326 vm_unacct_memory(p->pages);
2329 spin_unlock(&swap_lock);
2332 spin_lock(&swap_avail_lock);
2333 plist_del(&p->avail_list, &swap_avail_head);
2334 spin_unlock(&swap_avail_lock);
2335 spin_lock(&p->lock);
2337 struct swap_info_struct *si = p;
2339 plist_for_each_entry_continue(si, &swap_active_head, list) {
2342 si->avail_list.prio--;
2346 plist_del(&p->list, &swap_active_head);
2347 atomic_long_sub(p->pages, &nr_swap_pages);
2348 total_swap_pages -= p->pages;
2349 p->flags &= ~SWP_WRITEOK;
2350 spin_unlock(&p->lock);
2351 spin_unlock(&swap_lock);
2353 disable_swap_slots_cache_lock();
2355 set_current_oom_origin();
2356 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2357 clear_current_oom_origin();
2360 /* re-insert swap space back into swap_list */
2361 reinsert_swap_info(p);
2362 reenable_swap_slots_cache_unlock();
2366 reenable_swap_slots_cache_unlock();
2368 flush_work(&p->discard_work);
2370 destroy_swap_extents(p);
2371 if (p->flags & SWP_CONTINUED)
2372 free_swap_count_continuations(p);
2374 mutex_lock(&swapon_mutex);
2375 spin_lock(&swap_lock);
2376 spin_lock(&p->lock);
2379 /* wait for anyone still in scan_swap_map */
2380 p->highest_bit = 0; /* cuts scans short */
2381 while (p->flags >= SWP_SCANNING) {
2382 spin_unlock(&p->lock);
2383 spin_unlock(&swap_lock);
2384 schedule_timeout_uninterruptible(1);
2385 spin_lock(&swap_lock);
2386 spin_lock(&p->lock);
2389 swap_file = p->swap_file;
2390 old_block_size = p->old_block_size;
2391 p->swap_file = NULL;
2393 swap_map = p->swap_map;
2395 cluster_info = p->cluster_info;
2396 p->cluster_info = NULL;
2397 frontswap_map = frontswap_map_get(p);
2398 spin_unlock(&p->lock);
2399 spin_unlock(&swap_lock);
2400 frontswap_invalidate_area(p->type);
2401 frontswap_map_set(p, NULL);
2402 mutex_unlock(&swapon_mutex);
2403 free_percpu(p->percpu_cluster);
2404 p->percpu_cluster = NULL;
2406 kvfree(cluster_info);
2407 kvfree(frontswap_map);
2408 /* Destroy swap account information */
2409 swap_cgroup_swapoff(p->type);
2410 exit_swap_address_space(p->type);
2412 inode = mapping->host;
2413 if (S_ISBLK(inode->i_mode)) {
2414 struct block_device *bdev = I_BDEV(inode);
2415 set_blocksize(bdev, old_block_size);
2416 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2419 inode->i_flags &= ~S_SWAPFILE;
2420 inode_unlock(inode);
2422 filp_close(swap_file, NULL);
2425 * Clear the SWP_USED flag after all resources are freed so that swapon
2426 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2427 * not hold p->lock after we cleared its SWP_WRITEOK.
2429 spin_lock(&swap_lock);
2431 spin_unlock(&swap_lock);
2434 atomic_inc(&proc_poll_event);
2435 wake_up_interruptible(&proc_poll_wait);
2438 filp_close(victim, NULL);
2444 #ifdef CONFIG_PROC_FS
2445 static unsigned swaps_poll(struct file *file, poll_table *wait)
2447 struct seq_file *seq = file->private_data;
2449 poll_wait(file, &proc_poll_wait, wait);
2451 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2452 seq->poll_event = atomic_read(&proc_poll_event);
2453 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2456 return POLLIN | POLLRDNORM;
2460 static void *swap_start(struct seq_file *swap, loff_t *pos)
2462 struct swap_info_struct *si;
2466 mutex_lock(&swapon_mutex);
2469 return SEQ_START_TOKEN;
2471 for (type = 0; type < nr_swapfiles; type++) {
2472 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2473 si = swap_info[type];
2474 if (!(si->flags & SWP_USED) || !si->swap_map)
2483 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2485 struct swap_info_struct *si = v;
2488 if (v == SEQ_START_TOKEN)
2491 type = si->type + 1;
2493 for (; type < nr_swapfiles; type++) {
2494 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2495 si = swap_info[type];
2496 if (!(si->flags & SWP_USED) || !si->swap_map)
2505 static void swap_stop(struct seq_file *swap, void *v)
2507 mutex_unlock(&swapon_mutex);
2510 static int swap_show(struct seq_file *swap, void *v)
2512 struct swap_info_struct *si = v;
2516 if (si == SEQ_START_TOKEN) {
2517 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2521 file = si->swap_file;
2522 len = seq_file_path(swap, file, " \t\n\\");
2523 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2524 len < 40 ? 40 - len : 1, " ",
2525 S_ISBLK(file_inode(file)->i_mode) ?
2526 "partition" : "file\t",
2527 si->pages << (PAGE_SHIFT - 10),
2528 si->inuse_pages << (PAGE_SHIFT - 10),
2533 static const struct seq_operations swaps_op = {
2534 .start = swap_start,
2540 static int swaps_open(struct inode *inode, struct file *file)
2542 struct seq_file *seq;
2545 ret = seq_open(file, &swaps_op);
2549 seq = file->private_data;
2550 seq->poll_event = atomic_read(&proc_poll_event);
2554 static const struct file_operations proc_swaps_operations = {
2557 .llseek = seq_lseek,
2558 .release = seq_release,
2562 static int __init procswaps_init(void)
2564 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2567 __initcall(procswaps_init);
2568 #endif /* CONFIG_PROC_FS */
2570 #ifdef MAX_SWAPFILES_CHECK
2571 static int __init max_swapfiles_check(void)
2573 MAX_SWAPFILES_CHECK();
2576 late_initcall(max_swapfiles_check);
2579 static struct swap_info_struct *alloc_swap_info(void)
2581 struct swap_info_struct *p;
2584 p = kzalloc(sizeof(*p), GFP_KERNEL);
2586 return ERR_PTR(-ENOMEM);
2588 spin_lock(&swap_lock);
2589 for (type = 0; type < nr_swapfiles; type++) {
2590 if (!(swap_info[type]->flags & SWP_USED))
2593 if (type >= MAX_SWAPFILES) {
2594 spin_unlock(&swap_lock);
2596 return ERR_PTR(-EPERM);
2598 if (type >= nr_swapfiles) {
2600 swap_info[type] = p;
2602 * Write swap_info[type] before nr_swapfiles, in case a
2603 * racing procfs swap_start() or swap_next() is reading them.
2604 * (We never shrink nr_swapfiles, we never free this entry.)
2610 p = swap_info[type];
2612 * Do not memset this entry: a racing procfs swap_next()
2613 * would be relying on p->type to remain valid.
2616 INIT_LIST_HEAD(&p->first_swap_extent.list);
2617 plist_node_init(&p->list, 0);
2618 plist_node_init(&p->avail_list, 0);
2619 p->flags = SWP_USED;
2620 spin_unlock(&swap_lock);
2621 spin_lock_init(&p->lock);
2626 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2630 if (S_ISBLK(inode->i_mode)) {
2631 p->bdev = bdgrab(I_BDEV(inode));
2632 error = blkdev_get(p->bdev,
2633 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2638 p->old_block_size = block_size(p->bdev);
2639 error = set_blocksize(p->bdev, PAGE_SIZE);
2642 p->flags |= SWP_BLKDEV;
2643 } else if (S_ISREG(inode->i_mode)) {
2644 p->bdev = inode->i_sb->s_bdev;
2646 if (IS_SWAPFILE(inode))
2654 static unsigned long read_swap_header(struct swap_info_struct *p,
2655 union swap_header *swap_header,
2656 struct inode *inode)
2659 unsigned long maxpages;
2660 unsigned long swapfilepages;
2661 unsigned long last_page;
2663 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2664 pr_err("Unable to find swap-space signature\n");
2668 /* swap partition endianess hack... */
2669 if (swab32(swap_header->info.version) == 1) {
2670 swab32s(&swap_header->info.version);
2671 swab32s(&swap_header->info.last_page);
2672 swab32s(&swap_header->info.nr_badpages);
2673 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2675 for (i = 0; i < swap_header->info.nr_badpages; i++)
2676 swab32s(&swap_header->info.badpages[i]);
2678 /* Check the swap header's sub-version */
2679 if (swap_header->info.version != 1) {
2680 pr_warn("Unable to handle swap header version %d\n",
2681 swap_header->info.version);
2686 p->cluster_next = 1;
2690 * Find out how many pages are allowed for a single swap
2691 * device. There are two limiting factors: 1) the number
2692 * of bits for the swap offset in the swp_entry_t type, and
2693 * 2) the number of bits in the swap pte as defined by the
2694 * different architectures. In order to find the
2695 * largest possible bit mask, a swap entry with swap type 0
2696 * and swap offset ~0UL is created, encoded to a swap pte,
2697 * decoded to a swp_entry_t again, and finally the swap
2698 * offset is extracted. This will mask all the bits from
2699 * the initial ~0UL mask that can't be encoded in either
2700 * the swp_entry_t or the architecture definition of a
2703 maxpages = swp_offset(pte_to_swp_entry(
2704 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2705 last_page = swap_header->info.last_page;
2706 if (last_page > maxpages) {
2707 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2708 maxpages << (PAGE_SHIFT - 10),
2709 last_page << (PAGE_SHIFT - 10));
2711 if (maxpages > last_page) {
2712 maxpages = last_page + 1;
2713 /* p->max is an unsigned int: don't overflow it */
2714 if ((unsigned int)maxpages == 0)
2715 maxpages = UINT_MAX;
2717 p->highest_bit = maxpages - 1;
2721 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2722 if (swapfilepages && maxpages > swapfilepages) {
2723 pr_warn("Swap area shorter than signature indicates\n");
2726 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2728 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2734 #define SWAP_CLUSTER_INFO_COLS \
2735 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2736 #define SWAP_CLUSTER_SPACE_COLS \
2737 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2738 #define SWAP_CLUSTER_COLS \
2739 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2741 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2742 union swap_header *swap_header,
2743 unsigned char *swap_map,
2744 struct swap_cluster_info *cluster_info,
2745 unsigned long maxpages,
2749 unsigned int nr_good_pages;
2751 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2752 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2753 unsigned long i, idx;
2755 nr_good_pages = maxpages - 1; /* omit header page */
2757 cluster_list_init(&p->free_clusters);
2758 cluster_list_init(&p->discard_clusters);
2760 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2761 unsigned int page_nr = swap_header->info.badpages[i];
2762 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2764 if (page_nr < maxpages) {
2765 swap_map[page_nr] = SWAP_MAP_BAD;
2768 * Haven't marked the cluster free yet, no list
2769 * operation involved
2771 inc_cluster_info_page(p, cluster_info, page_nr);
2775 /* Haven't marked the cluster free yet, no list operation involved */
2776 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2777 inc_cluster_info_page(p, cluster_info, i);
2779 if (nr_good_pages) {
2780 swap_map[0] = SWAP_MAP_BAD;
2782 * Not mark the cluster free yet, no list
2783 * operation involved
2785 inc_cluster_info_page(p, cluster_info, 0);
2787 p->pages = nr_good_pages;
2788 nr_extents = setup_swap_extents(p, span);
2791 nr_good_pages = p->pages;
2793 if (!nr_good_pages) {
2794 pr_warn("Empty swap-file\n");
2803 * Reduce false cache line sharing between cluster_info and
2804 * sharing same address space.
2806 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2807 j = (k + col) % SWAP_CLUSTER_COLS;
2808 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2809 idx = i * SWAP_CLUSTER_COLS + j;
2810 if (idx >= nr_clusters)
2812 if (cluster_count(&cluster_info[idx]))
2814 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2815 cluster_list_add_tail(&p->free_clusters, cluster_info,
2823 * Helper to sys_swapon determining if a given swap
2824 * backing device queue supports DISCARD operations.
2826 static bool swap_discardable(struct swap_info_struct *si)
2828 struct request_queue *q = bdev_get_queue(si->bdev);
2830 if (!q || !blk_queue_discard(q))
2836 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2838 struct swap_info_struct *p;
2839 struct filename *name;
2840 struct file *swap_file = NULL;
2841 struct address_space *mapping;
2844 union swap_header *swap_header;
2847 unsigned long maxpages;
2848 unsigned char *swap_map = NULL;
2849 struct swap_cluster_info *cluster_info = NULL;
2850 unsigned long *frontswap_map = NULL;
2851 struct page *page = NULL;
2852 struct inode *inode = NULL;
2854 if (swap_flags & ~SWAP_FLAGS_VALID)
2857 if (!capable(CAP_SYS_ADMIN))
2860 p = alloc_swap_info();
2864 INIT_WORK(&p->discard_work, swap_discard_work);
2866 name = getname(specialfile);
2868 error = PTR_ERR(name);
2872 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2873 if (IS_ERR(swap_file)) {
2874 error = PTR_ERR(swap_file);
2879 p->swap_file = swap_file;
2880 mapping = swap_file->f_mapping;
2881 inode = mapping->host;
2883 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
2884 error = claim_swapfile(p, inode);
2885 if (unlikely(error))
2889 * Read the swap header.
2891 if (!mapping->a_ops->readpage) {
2895 page = read_mapping_page(mapping, 0, swap_file);
2897 error = PTR_ERR(page);
2900 swap_header = kmap(page);
2902 maxpages = read_swap_header(p, swap_header, inode);
2903 if (unlikely(!maxpages)) {
2908 /* OK, set up the swap map and apply the bad block list */
2909 swap_map = vzalloc(maxpages);
2915 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
2916 p->flags |= SWP_STABLE_WRITES;
2918 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2920 unsigned long ci, nr_cluster;
2922 p->flags |= SWP_SOLIDSTATE;
2924 * select a random position to start with to help wear leveling
2927 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2928 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2930 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
2932 if (!cluster_info) {
2937 for (ci = 0; ci < nr_cluster; ci++)
2938 spin_lock_init(&((cluster_info + ci)->lock));
2940 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
2941 if (!p->percpu_cluster) {
2945 for_each_possible_cpu(cpu) {
2946 struct percpu_cluster *cluster;
2947 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
2948 cluster_set_null(&cluster->index);
2952 error = swap_cgroup_swapon(p->type, maxpages);
2956 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2957 cluster_info, maxpages, &span);
2958 if (unlikely(nr_extents < 0)) {
2962 /* frontswap enabled? set up bit-per-page map for frontswap */
2963 if (IS_ENABLED(CONFIG_FRONTSWAP))
2964 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
2967 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2969 * When discard is enabled for swap with no particular
2970 * policy flagged, we set all swap discard flags here in
2971 * order to sustain backward compatibility with older
2972 * swapon(8) releases.
2974 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2978 * By flagging sys_swapon, a sysadmin can tell us to
2979 * either do single-time area discards only, or to just
2980 * perform discards for released swap page-clusters.
2981 * Now it's time to adjust the p->flags accordingly.
2983 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2984 p->flags &= ~SWP_PAGE_DISCARD;
2985 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2986 p->flags &= ~SWP_AREA_DISCARD;
2988 /* issue a swapon-time discard if it's still required */
2989 if (p->flags & SWP_AREA_DISCARD) {
2990 int err = discard_swap(p);
2992 pr_err("swapon: discard_swap(%p): %d\n",
2997 error = init_swap_address_space(p->type, maxpages);
3001 mutex_lock(&swapon_mutex);
3003 if (swap_flags & SWAP_FLAG_PREFER)
3005 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3006 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3008 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3009 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3010 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3011 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3012 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3013 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3014 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3015 (frontswap_map) ? "FS" : "");
3017 mutex_unlock(&swapon_mutex);
3018 atomic_inc(&proc_poll_event);
3019 wake_up_interruptible(&proc_poll_wait);
3021 if (S_ISREG(inode->i_mode))
3022 inode->i_flags |= S_SWAPFILE;
3026 free_percpu(p->percpu_cluster);
3027 p->percpu_cluster = NULL;
3028 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3029 set_blocksize(p->bdev, p->old_block_size);
3030 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3032 destroy_swap_extents(p);
3033 swap_cgroup_swapoff(p->type);
3034 spin_lock(&swap_lock);
3035 p->swap_file = NULL;
3037 spin_unlock(&swap_lock);
3039 vfree(cluster_info);
3041 if (inode && S_ISREG(inode->i_mode)) {
3042 inode_unlock(inode);
3045 filp_close(swap_file, NULL);
3048 if (page && !IS_ERR(page)) {
3054 if (inode && S_ISREG(inode->i_mode))
3055 inode_unlock(inode);
3057 enable_swap_slots_cache();
3061 void si_swapinfo(struct sysinfo *val)
3064 unsigned long nr_to_be_unused = 0;
3066 spin_lock(&swap_lock);
3067 for (type = 0; type < nr_swapfiles; type++) {
3068 struct swap_info_struct *si = swap_info[type];
3070 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3071 nr_to_be_unused += si->inuse_pages;
3073 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3074 val->totalswap = total_swap_pages + nr_to_be_unused;
3075 spin_unlock(&swap_lock);
3079 * Verify that a swap entry is valid and increment its swap map count.
3081 * Returns error code in following case.
3083 * - swp_entry is invalid -> EINVAL
3084 * - swp_entry is migration entry -> EINVAL
3085 * - swap-cache reference is requested but there is already one. -> EEXIST
3086 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3087 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3089 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3091 struct swap_info_struct *p;
3092 struct swap_cluster_info *ci;
3093 unsigned long offset, type;
3094 unsigned char count;
3095 unsigned char has_cache;
3098 if (non_swap_entry(entry))
3101 type = swp_type(entry);
3102 if (type >= nr_swapfiles)
3104 p = swap_info[type];
3105 offset = swp_offset(entry);
3106 if (unlikely(offset >= p->max))
3109 ci = lock_cluster_or_swap_info(p, offset);
3111 count = p->swap_map[offset];
3114 * swapin_readahead() doesn't check if a swap entry is valid, so the
3115 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3117 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3122 has_cache = count & SWAP_HAS_CACHE;
3123 count &= ~SWAP_HAS_CACHE;
3126 if (usage == SWAP_HAS_CACHE) {
3128 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3129 if (!has_cache && count)
3130 has_cache = SWAP_HAS_CACHE;
3131 else if (has_cache) /* someone else added cache */
3133 else /* no users remaining */
3136 } else if (count || has_cache) {
3138 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3140 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3142 else if (swap_count_continued(p, offset, count))
3143 count = COUNT_CONTINUED;
3147 err = -ENOENT; /* unused swap entry */
3149 p->swap_map[offset] = count | has_cache;
3152 unlock_cluster_or_swap_info(p, ci);
3157 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3162 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3163 * (in which case its reference count is never incremented).
3165 void swap_shmem_alloc(swp_entry_t entry)
3167 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3171 * Increase reference count of swap entry by 1.
3172 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3173 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3174 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3175 * might occur if a page table entry has got corrupted.
3177 int swap_duplicate(swp_entry_t entry)
3181 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3182 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3187 * @entry: swap entry for which we allocate swap cache.
3189 * Called when allocating swap cache for existing swap entry,
3190 * This can return error codes. Returns 0 at success.
3191 * -EBUSY means there is a swap cache.
3192 * Note: return code is different from swap_duplicate().
3194 int swapcache_prepare(swp_entry_t entry)
3196 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3199 struct swap_info_struct *page_swap_info(struct page *page)
3201 swp_entry_t swap = { .val = page_private(page) };
3202 return swap_info[swp_type(swap)];
3206 * out-of-line __page_file_ methods to avoid include hell.
3208 struct address_space *__page_file_mapping(struct page *page)
3210 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3211 return page_swap_info(page)->swap_file->f_mapping;
3213 EXPORT_SYMBOL_GPL(__page_file_mapping);
3215 pgoff_t __page_file_index(struct page *page)
3217 swp_entry_t swap = { .val = page_private(page) };
3218 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3219 return swp_offset(swap);
3221 EXPORT_SYMBOL_GPL(__page_file_index);
3224 * add_swap_count_continuation - called when a swap count is duplicated
3225 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3226 * page of the original vmalloc'ed swap_map, to hold the continuation count
3227 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3228 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3230 * These continuation pages are seldom referenced: the common paths all work
3231 * on the original swap_map, only referring to a continuation page when the
3232 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3234 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3235 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3236 * can be called after dropping locks.
3238 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3240 struct swap_info_struct *si;
3241 struct swap_cluster_info *ci;
3244 struct page *list_page;
3246 unsigned char count;
3249 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3250 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3252 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3254 si = swap_info_get(entry);
3257 * An acceptable race has occurred since the failing
3258 * __swap_duplicate(): the swap entry has been freed,
3259 * perhaps even the whole swap_map cleared for swapoff.
3264 offset = swp_offset(entry);
3266 ci = lock_cluster(si, offset);
3268 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3270 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3272 * The higher the swap count, the more likely it is that tasks
3273 * will race to add swap count continuation: we need to avoid
3274 * over-provisioning.
3281 spin_unlock(&si->lock);
3286 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3287 * no architecture is using highmem pages for kernel page tables: so it
3288 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3290 head = vmalloc_to_page(si->swap_map + offset);
3291 offset &= ~PAGE_MASK;
3294 * Page allocation does not initialize the page's lru field,
3295 * but it does always reset its private field.
3297 if (!page_private(head)) {
3298 BUG_ON(count & COUNT_CONTINUED);
3299 INIT_LIST_HEAD(&head->lru);
3300 set_page_private(head, SWP_CONTINUED);
3301 si->flags |= SWP_CONTINUED;
3304 list_for_each_entry(list_page, &head->lru, lru) {
3308 * If the previous map said no continuation, but we've found
3309 * a continuation page, free our allocation and use this one.
3311 if (!(count & COUNT_CONTINUED))
3314 map = kmap_atomic(list_page) + offset;
3319 * If this continuation count now has some space in it,
3320 * free our allocation and use this one.
3322 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3326 list_add_tail(&page->lru, &head->lru);
3327 page = NULL; /* now it's attached, don't free it */
3330 spin_unlock(&si->lock);
3338 * swap_count_continued - when the original swap_map count is incremented
3339 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3340 * into, carry if so, or else fail until a new continuation page is allocated;
3341 * when the original swap_map count is decremented from 0 with continuation,
3342 * borrow from the continuation and report whether it still holds more.
3343 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3346 static bool swap_count_continued(struct swap_info_struct *si,
3347 pgoff_t offset, unsigned char count)
3353 head = vmalloc_to_page(si->swap_map + offset);
3354 if (page_private(head) != SWP_CONTINUED) {
3355 BUG_ON(count & COUNT_CONTINUED);
3356 return false; /* need to add count continuation */
3359 offset &= ~PAGE_MASK;
3360 page = list_entry(head->lru.next, struct page, lru);
3361 map = kmap_atomic(page) + offset;
3363 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3364 goto init_map; /* jump over SWAP_CONT_MAX checks */
3366 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3368 * Think of how you add 1 to 999
3370 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3372 page = list_entry(page->lru.next, struct page, lru);
3373 BUG_ON(page == head);
3374 map = kmap_atomic(page) + offset;
3376 if (*map == SWAP_CONT_MAX) {
3378 page = list_entry(page->lru.next, struct page, lru);
3380 return false; /* add count continuation */
3381 map = kmap_atomic(page) + offset;
3382 init_map: *map = 0; /* we didn't zero the page */
3386 page = list_entry(page->lru.prev, struct page, lru);
3387 while (page != head) {
3388 map = kmap_atomic(page) + offset;
3389 *map = COUNT_CONTINUED;
3391 page = list_entry(page->lru.prev, struct page, lru);
3393 return true; /* incremented */
3395 } else { /* decrementing */
3397 * Think of how you subtract 1 from 1000
3399 BUG_ON(count != COUNT_CONTINUED);
3400 while (*map == COUNT_CONTINUED) {
3402 page = list_entry(page->lru.next, struct page, lru);
3403 BUG_ON(page == head);
3404 map = kmap_atomic(page) + offset;
3411 page = list_entry(page->lru.prev, struct page, lru);
3412 while (page != head) {
3413 map = kmap_atomic(page) + offset;
3414 *map = SWAP_CONT_MAX | count;
3415 count = COUNT_CONTINUED;
3417 page = list_entry(page->lru.prev, struct page, lru);
3419 return count == COUNT_CONTINUED;
3424 * free_swap_count_continuations - swapoff free all the continuation pages
3425 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3427 static void free_swap_count_continuations(struct swap_info_struct *si)
3431 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3433 head = vmalloc_to_page(si->swap_map + offset);
3434 if (page_private(head)) {
3435 struct page *page, *next;
3437 list_for_each_entry_safe(page, next, &head->lru, lru) {
3438 list_del(&page->lru);