4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
9 #include <linux/sched/mm.h>
10 #include <linux/sched/task.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mman.h>
13 #include <linux/slab.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/swap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/pagemap.h>
18 #include <linux/namei.h>
19 #include <linux/shmem_fs.h>
20 #include <linux/blkdev.h>
21 #include <linux/random.h>
22 #include <linux/writeback.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/init.h>
26 #include <linux/ksm.h>
27 #include <linux/rmap.h>
28 #include <linux/security.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mutex.h>
31 #include <linux/capability.h>
32 #include <linux/syscalls.h>
33 #include <linux/memcontrol.h>
34 #include <linux/poll.h>
35 #include <linux/oom.h>
36 #include <linux/frontswap.h>
37 #include <linux/swapfile.h>
38 #include <linux/export.h>
39 #include <linux/swap_slots.h>
40 #include <linux/sort.h>
42 #include <asm/pgtable.h>
43 #include <asm/tlbflush.h>
44 #include <linux/swapops.h>
45 #include <linux/swap_cgroup.h>
47 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
49 static void free_swap_count_continuations(struct swap_info_struct *);
50 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
52 DEFINE_SPINLOCK(swap_lock);
53 static unsigned int nr_swapfiles;
54 atomic_long_t nr_swap_pages;
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
60 EXPORT_SYMBOL_GPL(nr_swap_pages);
61 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62 long total_swap_pages;
63 static int least_priority = -1;
65 static const char Bad_file[] = "Bad swap file entry ";
66 static const char Unused_file[] = "Unused swap file entry ";
67 static const char Bad_offset[] = "Bad swap offset entry ";
68 static const char Unused_offset[] = "Unused swap offset entry ";
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
74 PLIST_HEAD(swap_active_head);
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
88 static struct plist_head *swap_avail_heads;
89 static DEFINE_SPINLOCK(swap_avail_lock);
91 struct swap_info_struct *swap_info[MAX_SWAPFILES];
93 static DEFINE_MUTEX(swapon_mutex);
95 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96 /* Activity counter to indicate that a swapon or swapoff has occurred */
97 static atomic_t proc_poll_event = ATOMIC_INIT(0);
99 atomic_t nr_rotate_swap = ATOMIC_INIT(0);
101 static inline unsigned char swap_count(unsigned char ent)
103 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
106 /* returns 1 if swap entry is freed */
108 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
110 swp_entry_t entry = swp_entry(si->type, offset);
114 page = find_get_page(swap_address_space(entry), swp_offset(entry));
118 * This function is called from scan_swap_map() and it's called
119 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
120 * We have to use trylock for avoiding deadlock. This is a special
121 * case and you should use try_to_free_swap() with explicit lock_page()
122 * in usual operations.
124 if (trylock_page(page)) {
125 ret = try_to_free_swap(page);
133 * swapon tell device that all the old swap contents can be discarded,
134 * to allow the swap device to optimize its wear-levelling.
136 static int discard_swap(struct swap_info_struct *si)
138 struct swap_extent *se;
139 sector_t start_block;
143 /* Do not discard the swap header page! */
144 se = &si->first_swap_extent;
145 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
146 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
148 err = blkdev_issue_discard(si->bdev, start_block,
149 nr_blocks, GFP_KERNEL, 0);
155 list_for_each_entry(se, &si->first_swap_extent.list, list) {
156 start_block = se->start_block << (PAGE_SHIFT - 9);
157 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
159 err = blkdev_issue_discard(si->bdev, start_block,
160 nr_blocks, GFP_KERNEL, 0);
166 return err; /* That will often be -EOPNOTSUPP */
170 * swap allocation tell device that a cluster of swap can now be discarded,
171 * to allow the swap device to optimize its wear-levelling.
173 static void discard_swap_cluster(struct swap_info_struct *si,
174 pgoff_t start_page, pgoff_t nr_pages)
176 struct swap_extent *se = si->curr_swap_extent;
177 int found_extent = 0;
180 if (se->start_page <= start_page &&
181 start_page < se->start_page + se->nr_pages) {
182 pgoff_t offset = start_page - se->start_page;
183 sector_t start_block = se->start_block + offset;
184 sector_t nr_blocks = se->nr_pages - offset;
186 if (nr_blocks > nr_pages)
187 nr_blocks = nr_pages;
188 start_page += nr_blocks;
189 nr_pages -= nr_blocks;
192 si->curr_swap_extent = se;
194 start_block <<= PAGE_SHIFT - 9;
195 nr_blocks <<= PAGE_SHIFT - 9;
196 if (blkdev_issue_discard(si->bdev, start_block,
197 nr_blocks, GFP_NOIO, 0))
201 se = list_next_entry(se, list);
205 #ifdef CONFIG_THP_SWAP
206 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
208 #define SWAPFILE_CLUSTER 256
210 #define LATENCY_LIMIT 256
212 static inline void cluster_set_flag(struct swap_cluster_info *info,
218 static inline unsigned int cluster_count(struct swap_cluster_info *info)
223 static inline void cluster_set_count(struct swap_cluster_info *info,
229 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
230 unsigned int c, unsigned int f)
236 static inline unsigned int cluster_next(struct swap_cluster_info *info)
241 static inline void cluster_set_next(struct swap_cluster_info *info,
247 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
248 unsigned int n, unsigned int f)
254 static inline bool cluster_is_free(struct swap_cluster_info *info)
256 return info->flags & CLUSTER_FLAG_FREE;
259 static inline bool cluster_is_null(struct swap_cluster_info *info)
261 return info->flags & CLUSTER_FLAG_NEXT_NULL;
264 static inline void cluster_set_null(struct swap_cluster_info *info)
266 info->flags = CLUSTER_FLAG_NEXT_NULL;
270 static inline bool cluster_is_huge(struct swap_cluster_info *info)
272 return info->flags & CLUSTER_FLAG_HUGE;
275 static inline void cluster_clear_huge(struct swap_cluster_info *info)
277 info->flags &= ~CLUSTER_FLAG_HUGE;
280 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
281 unsigned long offset)
283 struct swap_cluster_info *ci;
285 ci = si->cluster_info;
287 ci += offset / SWAPFILE_CLUSTER;
288 spin_lock(&ci->lock);
293 static inline void unlock_cluster(struct swap_cluster_info *ci)
296 spin_unlock(&ci->lock);
300 * Determine the locking method in use for this device. Return
301 * swap_cluster_info if SSD-style cluster-based locking is in place.
303 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
304 struct swap_info_struct *si, unsigned long offset)
306 struct swap_cluster_info *ci;
308 /* Try to use fine-grained SSD-style locking if available: */
309 ci = lock_cluster(si, offset);
310 /* Otherwise, fall back to traditional, coarse locking: */
312 spin_lock(&si->lock);
317 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
318 struct swap_cluster_info *ci)
323 spin_unlock(&si->lock);
326 static inline bool cluster_list_empty(struct swap_cluster_list *list)
328 return cluster_is_null(&list->head);
331 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
333 return cluster_next(&list->head);
336 static void cluster_list_init(struct swap_cluster_list *list)
338 cluster_set_null(&list->head);
339 cluster_set_null(&list->tail);
342 static void cluster_list_add_tail(struct swap_cluster_list *list,
343 struct swap_cluster_info *ci,
346 if (cluster_list_empty(list)) {
347 cluster_set_next_flag(&list->head, idx, 0);
348 cluster_set_next_flag(&list->tail, idx, 0);
350 struct swap_cluster_info *ci_tail;
351 unsigned int tail = cluster_next(&list->tail);
354 * Nested cluster lock, but both cluster locks are
355 * only acquired when we held swap_info_struct->lock
358 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
359 cluster_set_next(ci_tail, idx);
360 spin_unlock(&ci_tail->lock);
361 cluster_set_next_flag(&list->tail, idx, 0);
365 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
366 struct swap_cluster_info *ci)
370 idx = cluster_next(&list->head);
371 if (cluster_next(&list->tail) == idx) {
372 cluster_set_null(&list->head);
373 cluster_set_null(&list->tail);
375 cluster_set_next_flag(&list->head,
376 cluster_next(&ci[idx]), 0);
381 /* Add a cluster to discard list and schedule it to do discard */
382 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
386 * If scan_swap_map() can't find a free cluster, it will check
387 * si->swap_map directly. To make sure the discarding cluster isn't
388 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
389 * will be cleared after discard
391 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
392 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
394 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
396 schedule_work(&si->discard_work);
399 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
401 struct swap_cluster_info *ci = si->cluster_info;
403 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
404 cluster_list_add_tail(&si->free_clusters, ci, idx);
408 * Doing discard actually. After a cluster discard is finished, the cluster
409 * will be added to free cluster list. caller should hold si->lock.
411 static void swap_do_scheduled_discard(struct swap_info_struct *si)
413 struct swap_cluster_info *info, *ci;
416 info = si->cluster_info;
418 while (!cluster_list_empty(&si->discard_clusters)) {
419 idx = cluster_list_del_first(&si->discard_clusters, info);
420 spin_unlock(&si->lock);
422 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
425 spin_lock(&si->lock);
426 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
427 __free_cluster(si, idx);
428 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
429 0, SWAPFILE_CLUSTER);
434 static void swap_discard_work(struct work_struct *work)
436 struct swap_info_struct *si;
438 si = container_of(work, struct swap_info_struct, discard_work);
440 spin_lock(&si->lock);
441 swap_do_scheduled_discard(si);
442 spin_unlock(&si->lock);
445 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
447 struct swap_cluster_info *ci = si->cluster_info;
449 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
450 cluster_list_del_first(&si->free_clusters, ci);
451 cluster_set_count_flag(ci + idx, 0, 0);
454 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
456 struct swap_cluster_info *ci = si->cluster_info + idx;
458 VM_BUG_ON(cluster_count(ci) != 0);
460 * If the swap is discardable, prepare discard the cluster
461 * instead of free it immediately. The cluster will be freed
464 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
465 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
466 swap_cluster_schedule_discard(si, idx);
470 __free_cluster(si, idx);
474 * The cluster corresponding to page_nr will be used. The cluster will be
475 * removed from free cluster list and its usage counter will be increased.
477 static void inc_cluster_info_page(struct swap_info_struct *p,
478 struct swap_cluster_info *cluster_info, unsigned long page_nr)
480 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
484 if (cluster_is_free(&cluster_info[idx]))
485 alloc_cluster(p, idx);
487 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
488 cluster_set_count(&cluster_info[idx],
489 cluster_count(&cluster_info[idx]) + 1);
493 * The cluster corresponding to page_nr decreases one usage. If the usage
494 * counter becomes 0, which means no page in the cluster is in using, we can
495 * optionally discard the cluster and add it to free cluster list.
497 static void dec_cluster_info_page(struct swap_info_struct *p,
498 struct swap_cluster_info *cluster_info, unsigned long page_nr)
500 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
505 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
506 cluster_set_count(&cluster_info[idx],
507 cluster_count(&cluster_info[idx]) - 1);
509 if (cluster_count(&cluster_info[idx]) == 0)
510 free_cluster(p, idx);
514 * It's possible scan_swap_map() uses a free cluster in the middle of free
515 * cluster list. Avoiding such abuse to avoid list corruption.
518 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
519 unsigned long offset)
521 struct percpu_cluster *percpu_cluster;
524 offset /= SWAPFILE_CLUSTER;
525 conflict = !cluster_list_empty(&si->free_clusters) &&
526 offset != cluster_list_first(&si->free_clusters) &&
527 cluster_is_free(&si->cluster_info[offset]);
532 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
533 cluster_set_null(&percpu_cluster->index);
538 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
539 * might involve allocating a new cluster for current CPU too.
541 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
542 unsigned long *offset, unsigned long *scan_base)
544 struct percpu_cluster *cluster;
545 struct swap_cluster_info *ci;
547 unsigned long tmp, max;
550 cluster = this_cpu_ptr(si->percpu_cluster);
551 if (cluster_is_null(&cluster->index)) {
552 if (!cluster_list_empty(&si->free_clusters)) {
553 cluster->index = si->free_clusters.head;
554 cluster->next = cluster_next(&cluster->index) *
556 } else if (!cluster_list_empty(&si->discard_clusters)) {
558 * we don't have free cluster but have some clusters in
559 * discarding, do discard now and reclaim them
561 swap_do_scheduled_discard(si);
562 *scan_base = *offset = si->cluster_next;
571 * Other CPUs can use our cluster if they can't find a free cluster,
572 * check if there is still free entry in the cluster
575 max = min_t(unsigned long, si->max,
576 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
578 cluster_set_null(&cluster->index);
581 ci = lock_cluster(si, tmp);
583 if (!si->swap_map[tmp]) {
591 cluster_set_null(&cluster->index);
594 cluster->next = tmp + 1;
600 static void __del_from_avail_list(struct swap_info_struct *p)
605 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
608 static void del_from_avail_list(struct swap_info_struct *p)
610 spin_lock(&swap_avail_lock);
611 __del_from_avail_list(p);
612 spin_unlock(&swap_avail_lock);
615 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
616 unsigned int nr_entries)
618 unsigned int end = offset + nr_entries - 1;
620 if (offset == si->lowest_bit)
621 si->lowest_bit += nr_entries;
622 if (end == si->highest_bit)
623 si->highest_bit -= nr_entries;
624 si->inuse_pages += nr_entries;
625 if (si->inuse_pages == si->pages) {
626 si->lowest_bit = si->max;
628 del_from_avail_list(si);
632 static void add_to_avail_list(struct swap_info_struct *p)
636 spin_lock(&swap_avail_lock);
638 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
639 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
641 spin_unlock(&swap_avail_lock);
644 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
645 unsigned int nr_entries)
647 unsigned long end = offset + nr_entries - 1;
648 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
650 if (offset < si->lowest_bit)
651 si->lowest_bit = offset;
652 if (end > si->highest_bit) {
653 bool was_full = !si->highest_bit;
655 si->highest_bit = end;
656 if (was_full && (si->flags & SWP_WRITEOK))
657 add_to_avail_list(si);
659 atomic_long_add(nr_entries, &nr_swap_pages);
660 si->inuse_pages -= nr_entries;
661 if (si->flags & SWP_BLKDEV)
662 swap_slot_free_notify =
663 si->bdev->bd_disk->fops->swap_slot_free_notify;
665 swap_slot_free_notify = NULL;
666 while (offset <= end) {
667 frontswap_invalidate_page(si->type, offset);
668 if (swap_slot_free_notify)
669 swap_slot_free_notify(si->bdev, offset);
674 static int scan_swap_map_slots(struct swap_info_struct *si,
675 unsigned char usage, int nr,
678 struct swap_cluster_info *ci;
679 unsigned long offset;
680 unsigned long scan_base;
681 unsigned long last_in_cluster = 0;
682 int latency_ration = LATENCY_LIMIT;
689 * We try to cluster swap pages by allocating them sequentially
690 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
691 * way, however, we resort to first-free allocation, starting
692 * a new cluster. This prevents us from scattering swap pages
693 * all over the entire swap partition, so that we reduce
694 * overall disk seek times between swap pages. -- sct
695 * But we do now try to find an empty cluster. -Andrea
696 * And we let swap pages go all over an SSD partition. Hugh
699 si->flags += SWP_SCANNING;
700 scan_base = offset = si->cluster_next;
703 if (si->cluster_info) {
704 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
710 if (unlikely(!si->cluster_nr--)) {
711 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
712 si->cluster_nr = SWAPFILE_CLUSTER - 1;
716 spin_unlock(&si->lock);
719 * If seek is expensive, start searching for new cluster from
720 * start of partition, to minimize the span of allocated swap.
721 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
722 * case, just handled by scan_swap_map_try_ssd_cluster() above.
724 scan_base = offset = si->lowest_bit;
725 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
727 /* Locate the first empty (unaligned) cluster */
728 for (; last_in_cluster <= si->highest_bit; offset++) {
729 if (si->swap_map[offset])
730 last_in_cluster = offset + SWAPFILE_CLUSTER;
731 else if (offset == last_in_cluster) {
732 spin_lock(&si->lock);
733 offset -= SWAPFILE_CLUSTER - 1;
734 si->cluster_next = offset;
735 si->cluster_nr = SWAPFILE_CLUSTER - 1;
738 if (unlikely(--latency_ration < 0)) {
740 latency_ration = LATENCY_LIMIT;
745 spin_lock(&si->lock);
746 si->cluster_nr = SWAPFILE_CLUSTER - 1;
750 if (si->cluster_info) {
751 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
752 /* take a break if we already got some slots */
755 if (!scan_swap_map_try_ssd_cluster(si, &offset,
760 if (!(si->flags & SWP_WRITEOK))
762 if (!si->highest_bit)
764 if (offset > si->highest_bit)
765 scan_base = offset = si->lowest_bit;
767 ci = lock_cluster(si, offset);
768 /* reuse swap entry of cache-only swap if not busy. */
769 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
772 spin_unlock(&si->lock);
773 swap_was_freed = __try_to_reclaim_swap(si, offset);
774 spin_lock(&si->lock);
775 /* entry was freed successfully, try to use this again */
778 goto scan; /* check next one */
781 if (si->swap_map[offset]) {
788 si->swap_map[offset] = usage;
789 inc_cluster_info_page(si, si->cluster_info, offset);
792 swap_range_alloc(si, offset, 1);
793 si->cluster_next = offset + 1;
794 slots[n_ret++] = swp_entry(si->type, offset);
796 /* got enough slots or reach max slots? */
797 if ((n_ret == nr) || (offset >= si->highest_bit))
800 /* search for next available slot */
802 /* time to take a break? */
803 if (unlikely(--latency_ration < 0)) {
806 spin_unlock(&si->lock);
808 spin_lock(&si->lock);
809 latency_ration = LATENCY_LIMIT;
812 /* try to get more slots in cluster */
813 if (si->cluster_info) {
814 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
822 /* non-ssd case, still more slots in cluster? */
823 if (si->cluster_nr && !si->swap_map[offset]) {
829 si->flags -= SWP_SCANNING;
833 spin_unlock(&si->lock);
834 while (++offset <= si->highest_bit) {
835 if (!si->swap_map[offset]) {
836 spin_lock(&si->lock);
839 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
840 spin_lock(&si->lock);
843 if (unlikely(--latency_ration < 0)) {
845 latency_ration = LATENCY_LIMIT;
848 offset = si->lowest_bit;
849 while (offset < scan_base) {
850 if (!si->swap_map[offset]) {
851 spin_lock(&si->lock);
854 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
855 spin_lock(&si->lock);
858 if (unlikely(--latency_ration < 0)) {
860 latency_ration = LATENCY_LIMIT;
864 spin_lock(&si->lock);
867 si->flags -= SWP_SCANNING;
871 #ifdef CONFIG_THP_SWAP
872 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
875 struct swap_cluster_info *ci;
876 unsigned long offset, i;
879 if (cluster_list_empty(&si->free_clusters))
882 idx = cluster_list_first(&si->free_clusters);
883 offset = idx * SWAPFILE_CLUSTER;
884 ci = lock_cluster(si, offset);
885 alloc_cluster(si, idx);
886 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
888 map = si->swap_map + offset;
889 for (i = 0; i < SWAPFILE_CLUSTER; i++)
890 map[i] = SWAP_HAS_CACHE;
892 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
893 *slot = swp_entry(si->type, offset);
898 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
900 unsigned long offset = idx * SWAPFILE_CLUSTER;
901 struct swap_cluster_info *ci;
903 ci = lock_cluster(si, offset);
904 cluster_set_count_flag(ci, 0, 0);
905 free_cluster(si, idx);
907 swap_range_free(si, offset, SWAPFILE_CLUSTER);
910 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
915 #endif /* CONFIG_THP_SWAP */
917 static unsigned long scan_swap_map(struct swap_info_struct *si,
923 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
926 return swp_offset(entry);
932 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
934 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
935 struct swap_info_struct *si, *next;
940 /* Only single cluster request supported */
941 WARN_ON_ONCE(n_goal > 1 && cluster);
943 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
947 if (n_goal > SWAP_BATCH)
950 if (n_goal > avail_pgs)
953 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
955 spin_lock(&swap_avail_lock);
958 node = numa_node_id();
959 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
960 /* requeue si to after same-priority siblings */
961 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
962 spin_unlock(&swap_avail_lock);
963 spin_lock(&si->lock);
964 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
965 spin_lock(&swap_avail_lock);
966 if (plist_node_empty(&si->avail_lists[node])) {
967 spin_unlock(&si->lock);
970 WARN(!si->highest_bit,
971 "swap_info %d in list but !highest_bit\n",
973 WARN(!(si->flags & SWP_WRITEOK),
974 "swap_info %d in list but !SWP_WRITEOK\n",
976 __del_from_avail_list(si);
977 spin_unlock(&si->lock);
981 if (!(si->flags & SWP_FILE))
982 n_ret = swap_alloc_cluster(si, swp_entries);
984 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
985 n_goal, swp_entries);
986 spin_unlock(&si->lock);
987 if (n_ret || cluster)
989 pr_debug("scan_swap_map of si %d failed to find offset\n",
992 spin_lock(&swap_avail_lock);
995 * if we got here, it's likely that si was almost full before,
996 * and since scan_swap_map() can drop the si->lock, multiple
997 * callers probably all tried to get a page from the same si
998 * and it filled up before we could get one; or, the si filled
999 * up between us dropping swap_avail_lock and taking si->lock.
1000 * Since we dropped the swap_avail_lock, the swap_avail_head
1001 * list may have been modified; so if next is still in the
1002 * swap_avail_head list then try it, otherwise start over
1003 * if we have not gotten any slots.
1005 if (plist_node_empty(&next->avail_lists[node]))
1009 spin_unlock(&swap_avail_lock);
1013 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1019 /* The only caller of this function is now suspend routine */
1020 swp_entry_t get_swap_page_of_type(int type)
1022 struct swap_info_struct *si;
1025 si = swap_info[type];
1026 spin_lock(&si->lock);
1027 if (si && (si->flags & SWP_WRITEOK)) {
1028 atomic_long_dec(&nr_swap_pages);
1029 /* This is called for allocating swap entry, not cache */
1030 offset = scan_swap_map(si, 1);
1032 spin_unlock(&si->lock);
1033 return swp_entry(type, offset);
1035 atomic_long_inc(&nr_swap_pages);
1037 spin_unlock(&si->lock);
1038 return (swp_entry_t) {0};
1041 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1043 struct swap_info_struct *p;
1044 unsigned long offset, type;
1048 type = swp_type(entry);
1049 if (type >= nr_swapfiles)
1051 p = swap_info[type];
1052 if (!(p->flags & SWP_USED))
1054 offset = swp_offset(entry);
1055 if (offset >= p->max)
1060 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1063 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1066 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1071 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1073 struct swap_info_struct *p;
1075 p = __swap_info_get(entry);
1078 if (!p->swap_map[swp_offset(entry)])
1083 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1089 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1091 struct swap_info_struct *p;
1093 p = _swap_info_get(entry);
1095 spin_lock(&p->lock);
1099 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1100 struct swap_info_struct *q)
1102 struct swap_info_struct *p;
1104 p = _swap_info_get(entry);
1108 spin_unlock(&q->lock);
1110 spin_lock(&p->lock);
1115 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1116 swp_entry_t entry, unsigned char usage)
1118 struct swap_cluster_info *ci;
1119 unsigned long offset = swp_offset(entry);
1120 unsigned char count;
1121 unsigned char has_cache;
1123 ci = lock_cluster_or_swap_info(p, offset);
1125 count = p->swap_map[offset];
1127 has_cache = count & SWAP_HAS_CACHE;
1128 count &= ~SWAP_HAS_CACHE;
1130 if (usage == SWAP_HAS_CACHE) {
1131 VM_BUG_ON(!has_cache);
1133 } else if (count == SWAP_MAP_SHMEM) {
1135 * Or we could insist on shmem.c using a special
1136 * swap_shmem_free() and free_shmem_swap_and_cache()...
1139 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1140 if (count == COUNT_CONTINUED) {
1141 if (swap_count_continued(p, offset, count))
1142 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1144 count = SWAP_MAP_MAX;
1149 usage = count | has_cache;
1150 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1152 unlock_cluster_or_swap_info(p, ci);
1157 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1159 struct swap_cluster_info *ci;
1160 unsigned long offset = swp_offset(entry);
1161 unsigned char count;
1163 ci = lock_cluster(p, offset);
1164 count = p->swap_map[offset];
1165 VM_BUG_ON(count != SWAP_HAS_CACHE);
1166 p->swap_map[offset] = 0;
1167 dec_cluster_info_page(p, p->cluster_info, offset);
1170 mem_cgroup_uncharge_swap(entry, 1);
1171 swap_range_free(p, offset, 1);
1175 * Caller has made sure that the swap device corresponding to entry
1176 * is still around or has not been recycled.
1178 void swap_free(swp_entry_t entry)
1180 struct swap_info_struct *p;
1182 p = _swap_info_get(entry);
1184 if (!__swap_entry_free(p, entry, 1))
1185 free_swap_slot(entry);
1190 * Called after dropping swapcache to decrease refcnt to swap entries.
1192 static void swapcache_free(swp_entry_t entry)
1194 struct swap_info_struct *p;
1196 p = _swap_info_get(entry);
1198 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1199 free_swap_slot(entry);
1203 #ifdef CONFIG_THP_SWAP
1204 static void swapcache_free_cluster(swp_entry_t entry)
1206 unsigned long offset = swp_offset(entry);
1207 unsigned long idx = offset / SWAPFILE_CLUSTER;
1208 struct swap_cluster_info *ci;
1209 struct swap_info_struct *si;
1211 unsigned int i, free_entries = 0;
1214 si = _swap_info_get(entry);
1218 ci = lock_cluster(si, offset);
1219 VM_BUG_ON(!cluster_is_huge(ci));
1220 map = si->swap_map + offset;
1221 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1223 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1224 if (val == SWAP_HAS_CACHE)
1227 if (!free_entries) {
1228 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1229 map[i] &= ~SWAP_HAS_CACHE;
1231 cluster_clear_huge(ci);
1233 if (free_entries == SWAPFILE_CLUSTER) {
1234 spin_lock(&si->lock);
1235 ci = lock_cluster(si, offset);
1236 memset(map, 0, SWAPFILE_CLUSTER);
1238 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1239 swap_free_cluster(si, idx);
1240 spin_unlock(&si->lock);
1241 } else if (free_entries) {
1242 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1243 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1244 free_swap_slot(entry);
1249 int split_swap_cluster(swp_entry_t entry)
1251 struct swap_info_struct *si;
1252 struct swap_cluster_info *ci;
1253 unsigned long offset = swp_offset(entry);
1255 si = _swap_info_get(entry);
1258 ci = lock_cluster(si, offset);
1259 cluster_clear_huge(ci);
1264 static inline void swapcache_free_cluster(swp_entry_t entry)
1267 #endif /* CONFIG_THP_SWAP */
1269 void put_swap_page(struct page *page, swp_entry_t entry)
1271 if (!PageTransHuge(page))
1272 swapcache_free(entry);
1274 swapcache_free_cluster(entry);
1277 static int swp_entry_cmp(const void *ent1, const void *ent2)
1279 const swp_entry_t *e1 = ent1, *e2 = ent2;
1281 return (int)swp_type(*e1) - (int)swp_type(*e2);
1284 void swapcache_free_entries(swp_entry_t *entries, int n)
1286 struct swap_info_struct *p, *prev;
1296 * Sort swap entries by swap device, so each lock is only taken once.
1297 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1298 * so low that it isn't necessary to optimize further.
1300 if (nr_swapfiles > 1)
1301 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1302 for (i = 0; i < n; ++i) {
1303 p = swap_info_get_cont(entries[i], prev);
1305 swap_entry_free(p, entries[i]);
1309 spin_unlock(&p->lock);
1313 * How many references to page are currently swapped out?
1314 * This does not give an exact answer when swap count is continued,
1315 * but does include the high COUNT_CONTINUED flag to allow for that.
1317 int page_swapcount(struct page *page)
1320 struct swap_info_struct *p;
1321 struct swap_cluster_info *ci;
1323 unsigned long offset;
1325 entry.val = page_private(page);
1326 p = _swap_info_get(entry);
1328 offset = swp_offset(entry);
1329 ci = lock_cluster_or_swap_info(p, offset);
1330 count = swap_count(p->swap_map[offset]);
1331 unlock_cluster_or_swap_info(p, ci);
1336 int __swap_count(struct swap_info_struct *si, swp_entry_t entry)
1338 pgoff_t offset = swp_offset(entry);
1340 return swap_count(si->swap_map[offset]);
1343 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1346 pgoff_t offset = swp_offset(entry);
1347 struct swap_cluster_info *ci;
1349 ci = lock_cluster_or_swap_info(si, offset);
1350 count = swap_count(si->swap_map[offset]);
1351 unlock_cluster_or_swap_info(si, ci);
1356 * How many references to @entry are currently swapped out?
1357 * This does not give an exact answer when swap count is continued,
1358 * but does include the high COUNT_CONTINUED flag to allow for that.
1360 int __swp_swapcount(swp_entry_t entry)
1363 struct swap_info_struct *si;
1365 si = __swap_info_get(entry);
1367 count = swap_swapcount(si, entry);
1372 * How many references to @entry are currently swapped out?
1373 * This considers COUNT_CONTINUED so it returns exact answer.
1375 int swp_swapcount(swp_entry_t entry)
1377 int count, tmp_count, n;
1378 struct swap_info_struct *p;
1379 struct swap_cluster_info *ci;
1384 p = _swap_info_get(entry);
1388 offset = swp_offset(entry);
1390 ci = lock_cluster_or_swap_info(p, offset);
1392 count = swap_count(p->swap_map[offset]);
1393 if (!(count & COUNT_CONTINUED))
1396 count &= ~COUNT_CONTINUED;
1397 n = SWAP_MAP_MAX + 1;
1399 page = vmalloc_to_page(p->swap_map + offset);
1400 offset &= ~PAGE_MASK;
1401 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1404 page = list_next_entry(page, lru);
1405 map = kmap_atomic(page);
1406 tmp_count = map[offset];
1409 count += (tmp_count & ~COUNT_CONTINUED) * n;
1410 n *= (SWAP_CONT_MAX + 1);
1411 } while (tmp_count & COUNT_CONTINUED);
1413 unlock_cluster_or_swap_info(p, ci);
1417 #ifdef CONFIG_THP_SWAP
1418 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1421 struct swap_cluster_info *ci;
1422 unsigned char *map = si->swap_map;
1423 unsigned long roffset = swp_offset(entry);
1424 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1428 ci = lock_cluster_or_swap_info(si, offset);
1429 if (!ci || !cluster_is_huge(ci)) {
1430 if (map[roffset] != SWAP_HAS_CACHE)
1434 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1435 if (map[offset + i] != SWAP_HAS_CACHE) {
1441 unlock_cluster_or_swap_info(si, ci);
1445 static bool page_swapped(struct page *page)
1448 struct swap_info_struct *si;
1450 if (likely(!PageTransCompound(page)))
1451 return page_swapcount(page) != 0;
1453 page = compound_head(page);
1454 entry.val = page_private(page);
1455 si = _swap_info_get(entry);
1457 return swap_page_trans_huge_swapped(si, entry);
1461 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1462 int *total_swapcount)
1464 int i, map_swapcount, _total_mapcount, _total_swapcount;
1465 unsigned long offset = 0;
1466 struct swap_info_struct *si;
1467 struct swap_cluster_info *ci = NULL;
1468 unsigned char *map = NULL;
1469 int mapcount, swapcount = 0;
1471 /* hugetlbfs shouldn't call it */
1472 VM_BUG_ON_PAGE(PageHuge(page), page);
1474 if (likely(!PageTransCompound(page))) {
1475 mapcount = atomic_read(&page->_mapcount) + 1;
1477 *total_mapcount = mapcount;
1478 if (PageSwapCache(page))
1479 swapcount = page_swapcount(page);
1480 if (total_swapcount)
1481 *total_swapcount = swapcount;
1482 return mapcount + swapcount;
1485 page = compound_head(page);
1487 _total_mapcount = _total_swapcount = map_swapcount = 0;
1488 if (PageSwapCache(page)) {
1491 entry.val = page_private(page);
1492 si = _swap_info_get(entry);
1495 offset = swp_offset(entry);
1499 ci = lock_cluster(si, offset);
1500 for (i = 0; i < HPAGE_PMD_NR; i++) {
1501 mapcount = atomic_read(&page[i]._mapcount) + 1;
1502 _total_mapcount += mapcount;
1504 swapcount = swap_count(map[offset + i]);
1505 _total_swapcount += swapcount;
1507 map_swapcount = max(map_swapcount, mapcount + swapcount);
1510 if (PageDoubleMap(page)) {
1512 _total_mapcount -= HPAGE_PMD_NR;
1514 mapcount = compound_mapcount(page);
1515 map_swapcount += mapcount;
1516 _total_mapcount += mapcount;
1518 *total_mapcount = _total_mapcount;
1519 if (total_swapcount)
1520 *total_swapcount = _total_swapcount;
1522 return map_swapcount;
1525 #define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1526 #define page_swapped(page) (page_swapcount(page) != 0)
1528 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1529 int *total_swapcount)
1531 int mapcount, swapcount = 0;
1533 /* hugetlbfs shouldn't call it */
1534 VM_BUG_ON_PAGE(PageHuge(page), page);
1536 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1537 if (PageSwapCache(page))
1538 swapcount = page_swapcount(page);
1539 if (total_swapcount)
1540 *total_swapcount = swapcount;
1541 return mapcount + swapcount;
1546 * We can write to an anon page without COW if there are no other references
1547 * to it. And as a side-effect, free up its swap: because the old content
1548 * on disk will never be read, and seeking back there to write new content
1549 * later would only waste time away from clustering.
1551 * NOTE: total_map_swapcount should not be relied upon by the caller if
1552 * reuse_swap_page() returns false, but it may be always overwritten
1553 * (see the other implementation for CONFIG_SWAP=n).
1555 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1557 int count, total_mapcount, total_swapcount;
1559 VM_BUG_ON_PAGE(!PageLocked(page), page);
1560 if (unlikely(PageKsm(page)))
1562 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1564 if (total_map_swapcount)
1565 *total_map_swapcount = total_mapcount + total_swapcount;
1566 if (count == 1 && PageSwapCache(page) &&
1567 (likely(!PageTransCompound(page)) ||
1568 /* The remaining swap count will be freed soon */
1569 total_swapcount == page_swapcount(page))) {
1570 if (!PageWriteback(page)) {
1571 page = compound_head(page);
1572 delete_from_swap_cache(page);
1576 struct swap_info_struct *p;
1578 entry.val = page_private(page);
1579 p = swap_info_get(entry);
1580 if (p->flags & SWP_STABLE_WRITES) {
1581 spin_unlock(&p->lock);
1584 spin_unlock(&p->lock);
1592 * If swap is getting full, or if there are no more mappings of this page,
1593 * then try_to_free_swap is called to free its swap space.
1595 int try_to_free_swap(struct page *page)
1597 VM_BUG_ON_PAGE(!PageLocked(page), page);
1599 if (!PageSwapCache(page))
1601 if (PageWriteback(page))
1603 if (page_swapped(page))
1607 * Once hibernation has begun to create its image of memory,
1608 * there's a danger that one of the calls to try_to_free_swap()
1609 * - most probably a call from __try_to_reclaim_swap() while
1610 * hibernation is allocating its own swap pages for the image,
1611 * but conceivably even a call from memory reclaim - will free
1612 * the swap from a page which has already been recorded in the
1613 * image as a clean swapcache page, and then reuse its swap for
1614 * another page of the image. On waking from hibernation, the
1615 * original page might be freed under memory pressure, then
1616 * later read back in from swap, now with the wrong data.
1618 * Hibernation suspends storage while it is writing the image
1619 * to disk so check that here.
1621 if (pm_suspended_storage())
1624 page = compound_head(page);
1625 delete_from_swap_cache(page);
1631 * Free the swap entry like above, but also try to
1632 * free the page cache entry if it is the last user.
1634 int free_swap_and_cache(swp_entry_t entry)
1636 struct swap_info_struct *p;
1637 struct page *page = NULL;
1638 unsigned char count;
1640 if (non_swap_entry(entry))
1643 p = _swap_info_get(entry);
1645 count = __swap_entry_free(p, entry, 1);
1646 if (count == SWAP_HAS_CACHE &&
1647 !swap_page_trans_huge_swapped(p, entry)) {
1648 page = find_get_page(swap_address_space(entry),
1650 if (page && !trylock_page(page)) {
1655 free_swap_slot(entry);
1659 * Not mapped elsewhere, or swap space full? Free it!
1660 * Also recheck PageSwapCache now page is locked (above).
1662 if (PageSwapCache(page) && !PageWriteback(page) &&
1663 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1664 !swap_page_trans_huge_swapped(p, entry)) {
1665 page = compound_head(page);
1666 delete_from_swap_cache(page);
1675 #ifdef CONFIG_HIBERNATION
1677 * Find the swap type that corresponds to given device (if any).
1679 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1680 * from 0, in which the swap header is expected to be located.
1682 * This is needed for the suspend to disk (aka swsusp).
1684 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1686 struct block_device *bdev = NULL;
1690 bdev = bdget(device);
1692 spin_lock(&swap_lock);
1693 for (type = 0; type < nr_swapfiles; type++) {
1694 struct swap_info_struct *sis = swap_info[type];
1696 if (!(sis->flags & SWP_WRITEOK))
1701 *bdev_p = bdgrab(sis->bdev);
1703 spin_unlock(&swap_lock);
1706 if (bdev == sis->bdev) {
1707 struct swap_extent *se = &sis->first_swap_extent;
1709 if (se->start_block == offset) {
1711 *bdev_p = bdgrab(sis->bdev);
1713 spin_unlock(&swap_lock);
1719 spin_unlock(&swap_lock);
1727 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1728 * corresponding to given index in swap_info (swap type).
1730 sector_t swapdev_block(int type, pgoff_t offset)
1732 struct block_device *bdev;
1734 if ((unsigned int)type >= nr_swapfiles)
1736 if (!(swap_info[type]->flags & SWP_WRITEOK))
1738 return map_swap_entry(swp_entry(type, offset), &bdev);
1742 * Return either the total number of swap pages of given type, or the number
1743 * of free pages of that type (depending on @free)
1745 * This is needed for software suspend
1747 unsigned int count_swap_pages(int type, int free)
1751 spin_lock(&swap_lock);
1752 if ((unsigned int)type < nr_swapfiles) {
1753 struct swap_info_struct *sis = swap_info[type];
1755 spin_lock(&sis->lock);
1756 if (sis->flags & SWP_WRITEOK) {
1759 n -= sis->inuse_pages;
1761 spin_unlock(&sis->lock);
1763 spin_unlock(&swap_lock);
1766 #endif /* CONFIG_HIBERNATION */
1768 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1770 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1774 * No need to decide whether this PTE shares the swap entry with others,
1775 * just let do_wp_page work it out if a write is requested later - to
1776 * force COW, vm_page_prot omits write permission from any private vma.
1778 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1779 unsigned long addr, swp_entry_t entry, struct page *page)
1781 struct page *swapcache;
1782 struct mem_cgroup *memcg;
1788 page = ksm_might_need_to_copy(page, vma, addr);
1789 if (unlikely(!page))
1792 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1798 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1799 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1800 mem_cgroup_cancel_charge(page, memcg, false);
1805 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1806 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1808 set_pte_at(vma->vm_mm, addr, pte,
1809 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1810 if (page == swapcache) {
1811 page_add_anon_rmap(page, vma, addr, false);
1812 mem_cgroup_commit_charge(page, memcg, true, false);
1813 } else { /* ksm created a completely new copy */
1814 page_add_new_anon_rmap(page, vma, addr, false);
1815 mem_cgroup_commit_charge(page, memcg, false, false);
1816 lru_cache_add_active_or_unevictable(page, vma);
1820 * Move the page to the active list so it is not
1821 * immediately swapped out again after swapon.
1823 activate_page(page);
1825 pte_unmap_unlock(pte, ptl);
1827 if (page != swapcache) {
1834 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1835 unsigned long addr, unsigned long end,
1836 swp_entry_t entry, struct page *page)
1838 pte_t swp_pte = swp_entry_to_pte(entry);
1843 * We don't actually need pte lock while scanning for swp_pte: since
1844 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1845 * page table while we're scanning; though it could get zapped, and on
1846 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1847 * of unmatched parts which look like swp_pte, so unuse_pte must
1848 * recheck under pte lock. Scanning without pte lock lets it be
1849 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1851 pte = pte_offset_map(pmd, addr);
1854 * swapoff spends a _lot_ of time in this loop!
1855 * Test inline before going to call unuse_pte.
1857 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1859 ret = unuse_pte(vma, pmd, addr, entry, page);
1862 pte = pte_offset_map(pmd, addr);
1864 } while (pte++, addr += PAGE_SIZE, addr != end);
1870 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1871 unsigned long addr, unsigned long end,
1872 swp_entry_t entry, struct page *page)
1878 pmd = pmd_offset(pud, addr);
1881 next = pmd_addr_end(addr, end);
1882 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1884 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1887 } while (pmd++, addr = next, addr != end);
1891 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1892 unsigned long addr, unsigned long end,
1893 swp_entry_t entry, struct page *page)
1899 pud = pud_offset(p4d, addr);
1901 next = pud_addr_end(addr, end);
1902 if (pud_none_or_clear_bad(pud))
1904 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1907 } while (pud++, addr = next, addr != end);
1911 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1912 unsigned long addr, unsigned long end,
1913 swp_entry_t entry, struct page *page)
1919 p4d = p4d_offset(pgd, addr);
1921 next = p4d_addr_end(addr, end);
1922 if (p4d_none_or_clear_bad(p4d))
1924 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1927 } while (p4d++, addr = next, addr != end);
1931 static int unuse_vma(struct vm_area_struct *vma,
1932 swp_entry_t entry, struct page *page)
1935 unsigned long addr, end, next;
1938 if (page_anon_vma(page)) {
1939 addr = page_address_in_vma(page, vma);
1940 if (addr == -EFAULT)
1943 end = addr + PAGE_SIZE;
1945 addr = vma->vm_start;
1949 pgd = pgd_offset(vma->vm_mm, addr);
1951 next = pgd_addr_end(addr, end);
1952 if (pgd_none_or_clear_bad(pgd))
1954 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1957 } while (pgd++, addr = next, addr != end);
1961 static int unuse_mm(struct mm_struct *mm,
1962 swp_entry_t entry, struct page *page)
1964 struct vm_area_struct *vma;
1967 if (!down_read_trylock(&mm->mmap_sem)) {
1969 * Activate page so shrink_inactive_list is unlikely to unmap
1970 * its ptes while lock is dropped, so swapoff can make progress.
1972 activate_page(page);
1974 down_read(&mm->mmap_sem);
1977 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1978 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1982 up_read(&mm->mmap_sem);
1983 return (ret < 0)? ret: 0;
1987 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1988 * from current position to next entry still in use.
1989 * Recycle to start on reaching the end, returning 0 when empty.
1991 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1992 unsigned int prev, bool frontswap)
1994 unsigned int max = si->max;
1995 unsigned int i = prev;
1996 unsigned char count;
1999 * No need for swap_lock here: we're just looking
2000 * for whether an entry is in use, not modifying it; false
2001 * hits are okay, and sys_swapoff() has already prevented new
2002 * allocations from this area (while holding swap_lock).
2011 * No entries in use at top of swap_map,
2012 * loop back to start and recheck there.
2018 count = READ_ONCE(si->swap_map[i]);
2019 if (count && swap_count(count) != SWAP_MAP_BAD)
2020 if (!frontswap || frontswap_test(si, i))
2022 if ((i % LATENCY_LIMIT) == 0)
2029 * We completely avoid races by reading each swap page in advance,
2030 * and then search for the process using it. All the necessary
2031 * page table adjustments can then be made atomically.
2033 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2034 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2036 int try_to_unuse(unsigned int type, bool frontswap,
2037 unsigned long pages_to_unuse)
2039 struct swap_info_struct *si = swap_info[type];
2040 struct mm_struct *start_mm;
2041 volatile unsigned char *swap_map; /* swap_map is accessed without
2042 * locking. Mark it as volatile
2043 * to prevent compiler doing
2046 unsigned char swcount;
2053 * When searching mms for an entry, a good strategy is to
2054 * start at the first mm we freed the previous entry from
2055 * (though actually we don't notice whether we or coincidence
2056 * freed the entry). Initialize this start_mm with a hold.
2058 * A simpler strategy would be to start at the last mm we
2059 * freed the previous entry from; but that would take less
2060 * advantage of mmlist ordering, which clusters forked mms
2061 * together, child after parent. If we race with dup_mmap(), we
2062 * prefer to resolve parent before child, lest we miss entries
2063 * duplicated after we scanned child: using last mm would invert
2066 start_mm = &init_mm;
2070 * Keep on scanning until all entries have gone. Usually,
2071 * one pass through swap_map is enough, but not necessarily:
2072 * there are races when an instance of an entry might be missed.
2074 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2075 if (signal_pending(current)) {
2081 * Get a page for the entry, using the existing swap
2082 * cache page if there is one. Otherwise, get a clean
2083 * page and read the swap into it.
2085 swap_map = &si->swap_map[i];
2086 entry = swp_entry(type, i);
2087 page = read_swap_cache_async(entry,
2088 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2091 * Either swap_duplicate() failed because entry
2092 * has been freed independently, and will not be
2093 * reused since sys_swapoff() already disabled
2094 * allocation from here, or alloc_page() failed.
2096 swcount = *swap_map;
2098 * We don't hold lock here, so the swap entry could be
2099 * SWAP_MAP_BAD (when the cluster is discarding).
2100 * Instead of fail out, We can just skip the swap
2101 * entry because swapoff will wait for discarding
2104 if (!swcount || swcount == SWAP_MAP_BAD)
2111 * Don't hold on to start_mm if it looks like exiting.
2113 if (atomic_read(&start_mm->mm_users) == 1) {
2115 start_mm = &init_mm;
2120 * Wait for and lock page. When do_swap_page races with
2121 * try_to_unuse, do_swap_page can handle the fault much
2122 * faster than try_to_unuse can locate the entry. This
2123 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2124 * defer to do_swap_page in such a case - in some tests,
2125 * do_swap_page and try_to_unuse repeatedly compete.
2127 wait_on_page_locked(page);
2128 wait_on_page_writeback(page);
2130 wait_on_page_writeback(page);
2133 * Remove all references to entry.
2135 swcount = *swap_map;
2136 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2137 retval = shmem_unuse(entry, page);
2138 /* page has already been unlocked and released */
2143 if (swap_count(swcount) && start_mm != &init_mm)
2144 retval = unuse_mm(start_mm, entry, page);
2146 if (swap_count(*swap_map)) {
2147 int set_start_mm = (*swap_map >= swcount);
2148 struct list_head *p = &start_mm->mmlist;
2149 struct mm_struct *new_start_mm = start_mm;
2150 struct mm_struct *prev_mm = start_mm;
2151 struct mm_struct *mm;
2153 mmget(new_start_mm);
2155 spin_lock(&mmlist_lock);
2156 while (swap_count(*swap_map) && !retval &&
2157 (p = p->next) != &start_mm->mmlist) {
2158 mm = list_entry(p, struct mm_struct, mmlist);
2159 if (!mmget_not_zero(mm))
2161 spin_unlock(&mmlist_lock);
2167 swcount = *swap_map;
2168 if (!swap_count(swcount)) /* any usage ? */
2170 else if (mm == &init_mm)
2173 retval = unuse_mm(mm, entry, page);
2175 if (set_start_mm && *swap_map < swcount) {
2176 mmput(new_start_mm);
2181 spin_lock(&mmlist_lock);
2183 spin_unlock(&mmlist_lock);
2186 start_mm = new_start_mm;
2195 * If a reference remains (rare), we would like to leave
2196 * the page in the swap cache; but try_to_unmap could
2197 * then re-duplicate the entry once we drop page lock,
2198 * so we might loop indefinitely; also, that page could
2199 * not be swapped out to other storage meanwhile. So:
2200 * delete from cache even if there's another reference,
2201 * after ensuring that the data has been saved to disk -
2202 * since if the reference remains (rarer), it will be
2203 * read from disk into another page. Splitting into two
2204 * pages would be incorrect if swap supported "shared
2205 * private" pages, but they are handled by tmpfs files.
2207 * Given how unuse_vma() targets one particular offset
2208 * in an anon_vma, once the anon_vma has been determined,
2209 * this splitting happens to be just what is needed to
2210 * handle where KSM pages have been swapped out: re-reading
2211 * is unnecessarily slow, but we can fix that later on.
2213 if (swap_count(*swap_map) &&
2214 PageDirty(page) && PageSwapCache(page)) {
2215 struct writeback_control wbc = {
2216 .sync_mode = WB_SYNC_NONE,
2219 swap_writepage(compound_head(page), &wbc);
2221 wait_on_page_writeback(page);
2225 * It is conceivable that a racing task removed this page from
2226 * swap cache just before we acquired the page lock at the top,
2227 * or while we dropped it in unuse_mm(). The page might even
2228 * be back in swap cache on another swap area: that we must not
2229 * delete, since it may not have been written out to swap yet.
2231 if (PageSwapCache(page) &&
2232 likely(page_private(page) == entry.val) &&
2233 !page_swapped(page))
2234 delete_from_swap_cache(compound_head(page));
2237 * So we could skip searching mms once swap count went
2238 * to 1, we did not mark any present ptes as dirty: must
2239 * mark page dirty so shrink_page_list will preserve it.
2246 * Make sure that we aren't completely killing
2247 * interactive performance.
2250 if (frontswap && pages_to_unuse > 0) {
2251 if (!--pages_to_unuse)
2261 * After a successful try_to_unuse, if no swap is now in use, we know
2262 * we can empty the mmlist. swap_lock must be held on entry and exit.
2263 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2264 * added to the mmlist just after page_duplicate - before would be racy.
2266 static void drain_mmlist(void)
2268 struct list_head *p, *next;
2271 for (type = 0; type < nr_swapfiles; type++)
2272 if (swap_info[type]->inuse_pages)
2274 spin_lock(&mmlist_lock);
2275 list_for_each_safe(p, next, &init_mm.mmlist)
2277 spin_unlock(&mmlist_lock);
2281 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2282 * corresponds to page offset for the specified swap entry.
2283 * Note that the type of this function is sector_t, but it returns page offset
2284 * into the bdev, not sector offset.
2286 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2288 struct swap_info_struct *sis;
2289 struct swap_extent *start_se;
2290 struct swap_extent *se;
2293 sis = swap_info[swp_type(entry)];
2296 offset = swp_offset(entry);
2297 start_se = sis->curr_swap_extent;
2301 if (se->start_page <= offset &&
2302 offset < (se->start_page + se->nr_pages)) {
2303 return se->start_block + (offset - se->start_page);
2305 se = list_next_entry(se, list);
2306 sis->curr_swap_extent = se;
2307 BUG_ON(se == start_se); /* It *must* be present */
2312 * Returns the page offset into bdev for the specified page's swap entry.
2314 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2317 entry.val = page_private(page);
2318 return map_swap_entry(entry, bdev);
2322 * Free all of a swapdev's extent information
2324 static void destroy_swap_extents(struct swap_info_struct *sis)
2326 while (!list_empty(&sis->first_swap_extent.list)) {
2327 struct swap_extent *se;
2329 se = list_first_entry(&sis->first_swap_extent.list,
2330 struct swap_extent, list);
2331 list_del(&se->list);
2335 if (sis->flags & SWP_FILE) {
2336 struct file *swap_file = sis->swap_file;
2337 struct address_space *mapping = swap_file->f_mapping;
2339 sis->flags &= ~SWP_FILE;
2340 mapping->a_ops->swap_deactivate(swap_file);
2345 * Add a block range (and the corresponding page range) into this swapdev's
2346 * extent list. The extent list is kept sorted in page order.
2348 * This function rather assumes that it is called in ascending page order.
2351 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2352 unsigned long nr_pages, sector_t start_block)
2354 struct swap_extent *se;
2355 struct swap_extent *new_se;
2356 struct list_head *lh;
2358 if (start_page == 0) {
2359 se = &sis->first_swap_extent;
2360 sis->curr_swap_extent = se;
2362 se->nr_pages = nr_pages;
2363 se->start_block = start_block;
2366 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2367 se = list_entry(lh, struct swap_extent, list);
2368 BUG_ON(se->start_page + se->nr_pages != start_page);
2369 if (se->start_block + se->nr_pages == start_block) {
2371 se->nr_pages += nr_pages;
2377 * No merge. Insert a new extent, preserving ordering.
2379 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2382 new_se->start_page = start_page;
2383 new_se->nr_pages = nr_pages;
2384 new_se->start_block = start_block;
2386 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2391 * A `swap extent' is a simple thing which maps a contiguous range of pages
2392 * onto a contiguous range of disk blocks. An ordered list of swap extents
2393 * is built at swapon time and is then used at swap_writepage/swap_readpage
2394 * time for locating where on disk a page belongs.
2396 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2397 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2398 * swap files identically.
2400 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2401 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2402 * swapfiles are handled *identically* after swapon time.
2404 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2405 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2406 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2407 * requirements, they are simply tossed out - we will never use those blocks
2410 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2411 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2412 * which will scribble on the fs.
2414 * The amount of disk space which a single swap extent represents varies.
2415 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2416 * extents in the list. To avoid much list walking, we cache the previous
2417 * search location in `curr_swap_extent', and start new searches from there.
2418 * This is extremely effective. The average number of iterations in
2419 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2421 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2423 struct file *swap_file = sis->swap_file;
2424 struct address_space *mapping = swap_file->f_mapping;
2425 struct inode *inode = mapping->host;
2428 if (S_ISBLK(inode->i_mode)) {
2429 ret = add_swap_extent(sis, 0, sis->max, 0);
2434 if (mapping->a_ops->swap_activate) {
2435 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2437 sis->flags |= SWP_FILE;
2438 ret = add_swap_extent(sis, 0, sis->max, 0);
2444 return generic_swapfile_activate(sis, swap_file, span);
2447 static int swap_node(struct swap_info_struct *p)
2449 struct block_device *bdev;
2454 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2456 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2459 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2460 unsigned char *swap_map,
2461 struct swap_cluster_info *cluster_info)
2468 p->prio = --least_priority;
2470 * the plist prio is negated because plist ordering is
2471 * low-to-high, while swap ordering is high-to-low
2473 p->list.prio = -p->prio;
2476 p->avail_lists[i].prio = -p->prio;
2478 if (swap_node(p) == i)
2479 p->avail_lists[i].prio = 1;
2481 p->avail_lists[i].prio = -p->prio;
2484 p->swap_map = swap_map;
2485 p->cluster_info = cluster_info;
2486 p->flags |= SWP_WRITEOK;
2487 atomic_long_add(p->pages, &nr_swap_pages);
2488 total_swap_pages += p->pages;
2490 assert_spin_locked(&swap_lock);
2492 * both lists are plists, and thus priority ordered.
2493 * swap_active_head needs to be priority ordered for swapoff(),
2494 * which on removal of any swap_info_struct with an auto-assigned
2495 * (i.e. negative) priority increments the auto-assigned priority
2496 * of any lower-priority swap_info_structs.
2497 * swap_avail_head needs to be priority ordered for get_swap_page(),
2498 * which allocates swap pages from the highest available priority
2501 plist_add(&p->list, &swap_active_head);
2502 add_to_avail_list(p);
2505 static void enable_swap_info(struct swap_info_struct *p, int prio,
2506 unsigned char *swap_map,
2507 struct swap_cluster_info *cluster_info,
2508 unsigned long *frontswap_map)
2510 frontswap_init(p->type, frontswap_map);
2511 spin_lock(&swap_lock);
2512 spin_lock(&p->lock);
2513 _enable_swap_info(p, prio, swap_map, cluster_info);
2514 spin_unlock(&p->lock);
2515 spin_unlock(&swap_lock);
2518 static void reinsert_swap_info(struct swap_info_struct *p)
2520 spin_lock(&swap_lock);
2521 spin_lock(&p->lock);
2522 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2523 spin_unlock(&p->lock);
2524 spin_unlock(&swap_lock);
2527 bool has_usable_swap(void)
2531 spin_lock(&swap_lock);
2532 if (plist_head_empty(&swap_active_head))
2534 spin_unlock(&swap_lock);
2538 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2540 struct swap_info_struct *p = NULL;
2541 unsigned char *swap_map;
2542 struct swap_cluster_info *cluster_info;
2543 unsigned long *frontswap_map;
2544 struct file *swap_file, *victim;
2545 struct address_space *mapping;
2546 struct inode *inode;
2547 struct filename *pathname;
2549 unsigned int old_block_size;
2551 if (!capable(CAP_SYS_ADMIN))
2554 BUG_ON(!current->mm);
2556 pathname = getname(specialfile);
2557 if (IS_ERR(pathname))
2558 return PTR_ERR(pathname);
2560 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2561 err = PTR_ERR(victim);
2565 mapping = victim->f_mapping;
2566 spin_lock(&swap_lock);
2567 plist_for_each_entry(p, &swap_active_head, list) {
2568 if (p->flags & SWP_WRITEOK) {
2569 if (p->swap_file->f_mapping == mapping) {
2577 spin_unlock(&swap_lock);
2580 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2581 vm_unacct_memory(p->pages);
2584 spin_unlock(&swap_lock);
2587 del_from_avail_list(p);
2588 spin_lock(&p->lock);
2590 struct swap_info_struct *si = p;
2593 plist_for_each_entry_continue(si, &swap_active_head, list) {
2596 for_each_node(nid) {
2597 if (si->avail_lists[nid].prio != 1)
2598 si->avail_lists[nid].prio--;
2603 plist_del(&p->list, &swap_active_head);
2604 atomic_long_sub(p->pages, &nr_swap_pages);
2605 total_swap_pages -= p->pages;
2606 p->flags &= ~SWP_WRITEOK;
2607 spin_unlock(&p->lock);
2608 spin_unlock(&swap_lock);
2610 disable_swap_slots_cache_lock();
2612 set_current_oom_origin();
2613 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2614 clear_current_oom_origin();
2617 /* re-insert swap space back into swap_list */
2618 reinsert_swap_info(p);
2619 reenable_swap_slots_cache_unlock();
2623 reenable_swap_slots_cache_unlock();
2625 flush_work(&p->discard_work);
2627 destroy_swap_extents(p);
2628 if (p->flags & SWP_CONTINUED)
2629 free_swap_count_continuations(p);
2631 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2632 atomic_dec(&nr_rotate_swap);
2634 mutex_lock(&swapon_mutex);
2635 spin_lock(&swap_lock);
2636 spin_lock(&p->lock);
2639 /* wait for anyone still in scan_swap_map */
2640 p->highest_bit = 0; /* cuts scans short */
2641 while (p->flags >= SWP_SCANNING) {
2642 spin_unlock(&p->lock);
2643 spin_unlock(&swap_lock);
2644 schedule_timeout_uninterruptible(1);
2645 spin_lock(&swap_lock);
2646 spin_lock(&p->lock);
2649 swap_file = p->swap_file;
2650 old_block_size = p->old_block_size;
2651 p->swap_file = NULL;
2653 swap_map = p->swap_map;
2655 cluster_info = p->cluster_info;
2656 p->cluster_info = NULL;
2657 frontswap_map = frontswap_map_get(p);
2658 spin_unlock(&p->lock);
2659 spin_unlock(&swap_lock);
2660 frontswap_invalidate_area(p->type);
2661 frontswap_map_set(p, NULL);
2662 mutex_unlock(&swapon_mutex);
2663 free_percpu(p->percpu_cluster);
2664 p->percpu_cluster = NULL;
2666 kvfree(cluster_info);
2667 kvfree(frontswap_map);
2668 /* Destroy swap account information */
2669 swap_cgroup_swapoff(p->type);
2670 exit_swap_address_space(p->type);
2672 inode = mapping->host;
2673 if (S_ISBLK(inode->i_mode)) {
2674 struct block_device *bdev = I_BDEV(inode);
2675 set_blocksize(bdev, old_block_size);
2676 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2679 inode->i_flags &= ~S_SWAPFILE;
2680 inode_unlock(inode);
2682 filp_close(swap_file, NULL);
2685 * Clear the SWP_USED flag after all resources are freed so that swapon
2686 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2687 * not hold p->lock after we cleared its SWP_WRITEOK.
2689 spin_lock(&swap_lock);
2691 spin_unlock(&swap_lock);
2694 atomic_inc(&proc_poll_event);
2695 wake_up_interruptible(&proc_poll_wait);
2698 filp_close(victim, NULL);
2704 #ifdef CONFIG_PROC_FS
2705 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2707 struct seq_file *seq = file->private_data;
2709 poll_wait(file, &proc_poll_wait, wait);
2711 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2712 seq->poll_event = atomic_read(&proc_poll_event);
2713 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2716 return EPOLLIN | EPOLLRDNORM;
2720 static void *swap_start(struct seq_file *swap, loff_t *pos)
2722 struct swap_info_struct *si;
2726 mutex_lock(&swapon_mutex);
2729 return SEQ_START_TOKEN;
2731 for (type = 0; type < nr_swapfiles; type++) {
2732 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2733 si = swap_info[type];
2734 if (!(si->flags & SWP_USED) || !si->swap_map)
2743 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2745 struct swap_info_struct *si = v;
2748 if (v == SEQ_START_TOKEN)
2751 type = si->type + 1;
2753 for (; type < nr_swapfiles; type++) {
2754 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2755 si = swap_info[type];
2756 if (!(si->flags & SWP_USED) || !si->swap_map)
2765 static void swap_stop(struct seq_file *swap, void *v)
2767 mutex_unlock(&swapon_mutex);
2770 static int swap_show(struct seq_file *swap, void *v)
2772 struct swap_info_struct *si = v;
2776 if (si == SEQ_START_TOKEN) {
2777 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2781 file = si->swap_file;
2782 len = seq_file_path(swap, file, " \t\n\\");
2783 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2784 len < 40 ? 40 - len : 1, " ",
2785 S_ISBLK(file_inode(file)->i_mode) ?
2786 "partition" : "file\t",
2787 si->pages << (PAGE_SHIFT - 10),
2788 si->inuse_pages << (PAGE_SHIFT - 10),
2793 static const struct seq_operations swaps_op = {
2794 .start = swap_start,
2800 static int swaps_open(struct inode *inode, struct file *file)
2802 struct seq_file *seq;
2805 ret = seq_open(file, &swaps_op);
2809 seq = file->private_data;
2810 seq->poll_event = atomic_read(&proc_poll_event);
2814 static const struct file_operations proc_swaps_operations = {
2817 .llseek = seq_lseek,
2818 .release = seq_release,
2822 static int __init procswaps_init(void)
2824 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2827 __initcall(procswaps_init);
2828 #endif /* CONFIG_PROC_FS */
2830 #ifdef MAX_SWAPFILES_CHECK
2831 static int __init max_swapfiles_check(void)
2833 MAX_SWAPFILES_CHECK();
2836 late_initcall(max_swapfiles_check);
2839 static struct swap_info_struct *alloc_swap_info(void)
2841 struct swap_info_struct *p;
2845 p = kzalloc(sizeof(*p), GFP_KERNEL);
2847 return ERR_PTR(-ENOMEM);
2849 spin_lock(&swap_lock);
2850 for (type = 0; type < nr_swapfiles; type++) {
2851 if (!(swap_info[type]->flags & SWP_USED))
2854 if (type >= MAX_SWAPFILES) {
2855 spin_unlock(&swap_lock);
2857 return ERR_PTR(-EPERM);
2859 if (type >= nr_swapfiles) {
2861 swap_info[type] = p;
2863 * Write swap_info[type] before nr_swapfiles, in case a
2864 * racing procfs swap_start() or swap_next() is reading them.
2865 * (We never shrink nr_swapfiles, we never free this entry.)
2871 p = swap_info[type];
2873 * Do not memset this entry: a racing procfs swap_next()
2874 * would be relying on p->type to remain valid.
2877 INIT_LIST_HEAD(&p->first_swap_extent.list);
2878 plist_node_init(&p->list, 0);
2880 plist_node_init(&p->avail_lists[i], 0);
2881 p->flags = SWP_USED;
2882 spin_unlock(&swap_lock);
2883 spin_lock_init(&p->lock);
2884 spin_lock_init(&p->cont_lock);
2889 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2893 if (S_ISBLK(inode->i_mode)) {
2894 p->bdev = bdgrab(I_BDEV(inode));
2895 error = blkdev_get(p->bdev,
2896 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2901 p->old_block_size = block_size(p->bdev);
2902 error = set_blocksize(p->bdev, PAGE_SIZE);
2905 p->flags |= SWP_BLKDEV;
2906 } else if (S_ISREG(inode->i_mode)) {
2907 p->bdev = inode->i_sb->s_bdev;
2909 if (IS_SWAPFILE(inode))
2919 * Find out how many pages are allowed for a single swap device. There
2920 * are two limiting factors:
2921 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2922 * 2) the number of bits in the swap pte, as defined by the different
2925 * In order to find the largest possible bit mask, a swap entry with
2926 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2927 * decoded to a swp_entry_t again, and finally the swap offset is
2930 * This will mask all the bits from the initial ~0UL mask that can't
2931 * be encoded in either the swp_entry_t or the architecture definition
2934 unsigned long generic_max_swapfile_size(void)
2936 return swp_offset(pte_to_swp_entry(
2937 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2940 /* Can be overridden by an architecture for additional checks. */
2941 __weak unsigned long max_swapfile_size(void)
2943 return generic_max_swapfile_size();
2946 static unsigned long read_swap_header(struct swap_info_struct *p,
2947 union swap_header *swap_header,
2948 struct inode *inode)
2951 unsigned long maxpages;
2952 unsigned long swapfilepages;
2953 unsigned long last_page;
2955 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2956 pr_err("Unable to find swap-space signature\n");
2960 /* swap partition endianess hack... */
2961 if (swab32(swap_header->info.version) == 1) {
2962 swab32s(&swap_header->info.version);
2963 swab32s(&swap_header->info.last_page);
2964 swab32s(&swap_header->info.nr_badpages);
2965 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2967 for (i = 0; i < swap_header->info.nr_badpages; i++)
2968 swab32s(&swap_header->info.badpages[i]);
2970 /* Check the swap header's sub-version */
2971 if (swap_header->info.version != 1) {
2972 pr_warn("Unable to handle swap header version %d\n",
2973 swap_header->info.version);
2978 p->cluster_next = 1;
2981 maxpages = max_swapfile_size();
2982 last_page = swap_header->info.last_page;
2984 pr_warn("Empty swap-file\n");
2987 if (last_page > maxpages) {
2988 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2989 maxpages << (PAGE_SHIFT - 10),
2990 last_page << (PAGE_SHIFT - 10));
2992 if (maxpages > last_page) {
2993 maxpages = last_page + 1;
2994 /* p->max is an unsigned int: don't overflow it */
2995 if ((unsigned int)maxpages == 0)
2996 maxpages = UINT_MAX;
2998 p->highest_bit = maxpages - 1;
3002 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3003 if (swapfilepages && maxpages > swapfilepages) {
3004 pr_warn("Swap area shorter than signature indicates\n");
3007 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3009 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3015 #define SWAP_CLUSTER_INFO_COLS \
3016 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3017 #define SWAP_CLUSTER_SPACE_COLS \
3018 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3019 #define SWAP_CLUSTER_COLS \
3020 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3022 static int setup_swap_map_and_extents(struct swap_info_struct *p,
3023 union swap_header *swap_header,
3024 unsigned char *swap_map,
3025 struct swap_cluster_info *cluster_info,
3026 unsigned long maxpages,
3030 unsigned int nr_good_pages;
3032 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3033 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3034 unsigned long i, idx;
3036 nr_good_pages = maxpages - 1; /* omit header page */
3038 cluster_list_init(&p->free_clusters);
3039 cluster_list_init(&p->discard_clusters);
3041 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3042 unsigned int page_nr = swap_header->info.badpages[i];
3043 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3045 if (page_nr < maxpages) {
3046 swap_map[page_nr] = SWAP_MAP_BAD;
3049 * Haven't marked the cluster free yet, no list
3050 * operation involved
3052 inc_cluster_info_page(p, cluster_info, page_nr);
3056 /* Haven't marked the cluster free yet, no list operation involved */
3057 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3058 inc_cluster_info_page(p, cluster_info, i);
3060 if (nr_good_pages) {
3061 swap_map[0] = SWAP_MAP_BAD;
3063 * Not mark the cluster free yet, no list
3064 * operation involved
3066 inc_cluster_info_page(p, cluster_info, 0);
3068 p->pages = nr_good_pages;
3069 nr_extents = setup_swap_extents(p, span);
3072 nr_good_pages = p->pages;
3074 if (!nr_good_pages) {
3075 pr_warn("Empty swap-file\n");
3084 * Reduce false cache line sharing between cluster_info and
3085 * sharing same address space.
3087 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3088 j = (k + col) % SWAP_CLUSTER_COLS;
3089 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3090 idx = i * SWAP_CLUSTER_COLS + j;
3091 if (idx >= nr_clusters)
3093 if (cluster_count(&cluster_info[idx]))
3095 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3096 cluster_list_add_tail(&p->free_clusters, cluster_info,
3104 * Helper to sys_swapon determining if a given swap
3105 * backing device queue supports DISCARD operations.
3107 static bool swap_discardable(struct swap_info_struct *si)
3109 struct request_queue *q = bdev_get_queue(si->bdev);
3111 if (!q || !blk_queue_discard(q))
3117 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3119 struct swap_info_struct *p;
3120 struct filename *name;
3121 struct file *swap_file = NULL;
3122 struct address_space *mapping;
3125 union swap_header *swap_header;
3128 unsigned long maxpages;
3129 unsigned char *swap_map = NULL;
3130 struct swap_cluster_info *cluster_info = NULL;
3131 unsigned long *frontswap_map = NULL;
3132 struct page *page = NULL;
3133 struct inode *inode = NULL;
3134 bool inced_nr_rotate_swap = false;
3136 if (swap_flags & ~SWAP_FLAGS_VALID)
3139 if (!capable(CAP_SYS_ADMIN))
3142 if (!swap_avail_heads)
3145 p = alloc_swap_info();
3149 INIT_WORK(&p->discard_work, swap_discard_work);
3151 name = getname(specialfile);
3153 error = PTR_ERR(name);
3157 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3158 if (IS_ERR(swap_file)) {
3159 error = PTR_ERR(swap_file);
3164 p->swap_file = swap_file;
3165 mapping = swap_file->f_mapping;
3166 inode = mapping->host;
3168 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3169 error = claim_swapfile(p, inode);
3170 if (unlikely(error))
3174 * Read the swap header.
3176 if (!mapping->a_ops->readpage) {
3180 page = read_mapping_page(mapping, 0, swap_file);
3182 error = PTR_ERR(page);
3185 swap_header = kmap(page);
3187 maxpages = read_swap_header(p, swap_header, inode);
3188 if (unlikely(!maxpages)) {
3193 /* OK, set up the swap map and apply the bad block list */
3194 swap_map = vzalloc(maxpages);
3200 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3201 p->flags |= SWP_STABLE_WRITES;
3203 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3204 p->flags |= SWP_SYNCHRONOUS_IO;
3206 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3208 unsigned long ci, nr_cluster;
3210 p->flags |= SWP_SOLIDSTATE;
3212 * select a random position to start with to help wear leveling
3215 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3216 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3218 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3220 if (!cluster_info) {
3225 for (ci = 0; ci < nr_cluster; ci++)
3226 spin_lock_init(&((cluster_info + ci)->lock));
3228 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3229 if (!p->percpu_cluster) {
3233 for_each_possible_cpu(cpu) {
3234 struct percpu_cluster *cluster;
3235 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3236 cluster_set_null(&cluster->index);
3239 atomic_inc(&nr_rotate_swap);
3240 inced_nr_rotate_swap = true;
3243 error = swap_cgroup_swapon(p->type, maxpages);
3247 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3248 cluster_info, maxpages, &span);
3249 if (unlikely(nr_extents < 0)) {
3253 /* frontswap enabled? set up bit-per-page map for frontswap */
3254 if (IS_ENABLED(CONFIG_FRONTSWAP))
3255 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3259 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3261 * When discard is enabled for swap with no particular
3262 * policy flagged, we set all swap discard flags here in
3263 * order to sustain backward compatibility with older
3264 * swapon(8) releases.
3266 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3270 * By flagging sys_swapon, a sysadmin can tell us to
3271 * either do single-time area discards only, or to just
3272 * perform discards for released swap page-clusters.
3273 * Now it's time to adjust the p->flags accordingly.
3275 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3276 p->flags &= ~SWP_PAGE_DISCARD;
3277 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3278 p->flags &= ~SWP_AREA_DISCARD;
3280 /* issue a swapon-time discard if it's still required */
3281 if (p->flags & SWP_AREA_DISCARD) {
3282 int err = discard_swap(p);
3284 pr_err("swapon: discard_swap(%p): %d\n",
3289 error = init_swap_address_space(p->type, maxpages);
3293 mutex_lock(&swapon_mutex);
3295 if (swap_flags & SWAP_FLAG_PREFER)
3297 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3298 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3300 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3301 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3302 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3303 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3304 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3305 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3306 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3307 (frontswap_map) ? "FS" : "");
3309 mutex_unlock(&swapon_mutex);
3310 atomic_inc(&proc_poll_event);
3311 wake_up_interruptible(&proc_poll_wait);
3313 if (S_ISREG(inode->i_mode))
3314 inode->i_flags |= S_SWAPFILE;
3318 free_percpu(p->percpu_cluster);
3319 p->percpu_cluster = NULL;
3320 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3321 set_blocksize(p->bdev, p->old_block_size);
3322 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3324 destroy_swap_extents(p);
3325 swap_cgroup_swapoff(p->type);
3326 spin_lock(&swap_lock);
3327 p->swap_file = NULL;
3329 spin_unlock(&swap_lock);
3331 kvfree(cluster_info);
3332 kvfree(frontswap_map);
3333 if (inced_nr_rotate_swap)
3334 atomic_dec(&nr_rotate_swap);
3336 if (inode && S_ISREG(inode->i_mode)) {
3337 inode_unlock(inode);
3340 filp_close(swap_file, NULL);
3343 if (page && !IS_ERR(page)) {
3349 if (inode && S_ISREG(inode->i_mode))
3350 inode_unlock(inode);
3352 enable_swap_slots_cache();
3356 void si_swapinfo(struct sysinfo *val)
3359 unsigned long nr_to_be_unused = 0;
3361 spin_lock(&swap_lock);
3362 for (type = 0; type < nr_swapfiles; type++) {
3363 struct swap_info_struct *si = swap_info[type];
3365 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3366 nr_to_be_unused += si->inuse_pages;
3368 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3369 val->totalswap = total_swap_pages + nr_to_be_unused;
3370 spin_unlock(&swap_lock);
3374 * Verify that a swap entry is valid and increment its swap map count.
3376 * Returns error code in following case.
3378 * - swp_entry is invalid -> EINVAL
3379 * - swp_entry is migration entry -> EINVAL
3380 * - swap-cache reference is requested but there is already one. -> EEXIST
3381 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3382 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3384 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3386 struct swap_info_struct *p;
3387 struct swap_cluster_info *ci;
3388 unsigned long offset, type;
3389 unsigned char count;
3390 unsigned char has_cache;
3393 if (non_swap_entry(entry))
3396 type = swp_type(entry);
3397 if (type >= nr_swapfiles)
3399 p = swap_info[type];
3400 offset = swp_offset(entry);
3401 if (unlikely(offset >= p->max))
3404 ci = lock_cluster_or_swap_info(p, offset);
3406 count = p->swap_map[offset];
3409 * swapin_readahead() doesn't check if a swap entry is valid, so the
3410 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3412 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3417 has_cache = count & SWAP_HAS_CACHE;
3418 count &= ~SWAP_HAS_CACHE;
3421 if (usage == SWAP_HAS_CACHE) {
3423 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3424 if (!has_cache && count)
3425 has_cache = SWAP_HAS_CACHE;
3426 else if (has_cache) /* someone else added cache */
3428 else /* no users remaining */
3431 } else if (count || has_cache) {
3433 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3435 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3437 else if (swap_count_continued(p, offset, count))
3438 count = COUNT_CONTINUED;
3442 err = -ENOENT; /* unused swap entry */
3444 p->swap_map[offset] = count | has_cache;
3447 unlock_cluster_or_swap_info(p, ci);
3452 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3457 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3458 * (in which case its reference count is never incremented).
3460 void swap_shmem_alloc(swp_entry_t entry)
3462 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3466 * Increase reference count of swap entry by 1.
3467 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3468 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3469 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3470 * might occur if a page table entry has got corrupted.
3472 int swap_duplicate(swp_entry_t entry)
3476 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3477 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3482 * @entry: swap entry for which we allocate swap cache.
3484 * Called when allocating swap cache for existing swap entry,
3485 * This can return error codes. Returns 0 at success.
3486 * -EBUSY means there is a swap cache.
3487 * Note: return code is different from swap_duplicate().
3489 int swapcache_prepare(swp_entry_t entry)
3491 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3494 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3496 return swap_info[swp_type(entry)];
3499 struct swap_info_struct *page_swap_info(struct page *page)
3501 swp_entry_t entry = { .val = page_private(page) };
3502 return swp_swap_info(entry);
3506 * out-of-line __page_file_ methods to avoid include hell.
3508 struct address_space *__page_file_mapping(struct page *page)
3510 return page_swap_info(page)->swap_file->f_mapping;
3512 EXPORT_SYMBOL_GPL(__page_file_mapping);
3514 pgoff_t __page_file_index(struct page *page)
3516 swp_entry_t swap = { .val = page_private(page) };
3517 return swp_offset(swap);
3519 EXPORT_SYMBOL_GPL(__page_file_index);
3522 * add_swap_count_continuation - called when a swap count is duplicated
3523 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3524 * page of the original vmalloc'ed swap_map, to hold the continuation count
3525 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3526 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3528 * These continuation pages are seldom referenced: the common paths all work
3529 * on the original swap_map, only referring to a continuation page when the
3530 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3532 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3533 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3534 * can be called after dropping locks.
3536 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3538 struct swap_info_struct *si;
3539 struct swap_cluster_info *ci;
3542 struct page *list_page;
3544 unsigned char count;
3547 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3548 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3550 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3552 si = swap_info_get(entry);
3555 * An acceptable race has occurred since the failing
3556 * __swap_duplicate(): the swap entry has been freed,
3557 * perhaps even the whole swap_map cleared for swapoff.
3562 offset = swp_offset(entry);
3564 ci = lock_cluster(si, offset);
3566 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3568 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3570 * The higher the swap count, the more likely it is that tasks
3571 * will race to add swap count continuation: we need to avoid
3572 * over-provisioning.
3579 spin_unlock(&si->lock);
3584 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3585 * no architecture is using highmem pages for kernel page tables: so it
3586 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3588 head = vmalloc_to_page(si->swap_map + offset);
3589 offset &= ~PAGE_MASK;
3591 spin_lock(&si->cont_lock);
3593 * Page allocation does not initialize the page's lru field,
3594 * but it does always reset its private field.
3596 if (!page_private(head)) {
3597 BUG_ON(count & COUNT_CONTINUED);
3598 INIT_LIST_HEAD(&head->lru);
3599 set_page_private(head, SWP_CONTINUED);
3600 si->flags |= SWP_CONTINUED;
3603 list_for_each_entry(list_page, &head->lru, lru) {
3607 * If the previous map said no continuation, but we've found
3608 * a continuation page, free our allocation and use this one.
3610 if (!(count & COUNT_CONTINUED))
3611 goto out_unlock_cont;
3613 map = kmap_atomic(list_page) + offset;
3618 * If this continuation count now has some space in it,
3619 * free our allocation and use this one.
3621 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3622 goto out_unlock_cont;
3625 list_add_tail(&page->lru, &head->lru);
3626 page = NULL; /* now it's attached, don't free it */
3628 spin_unlock(&si->cont_lock);
3631 spin_unlock(&si->lock);
3639 * swap_count_continued - when the original swap_map count is incremented
3640 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3641 * into, carry if so, or else fail until a new continuation page is allocated;
3642 * when the original swap_map count is decremented from 0 with continuation,
3643 * borrow from the continuation and report whether it still holds more.
3644 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3647 static bool swap_count_continued(struct swap_info_struct *si,
3648 pgoff_t offset, unsigned char count)
3655 head = vmalloc_to_page(si->swap_map + offset);
3656 if (page_private(head) != SWP_CONTINUED) {
3657 BUG_ON(count & COUNT_CONTINUED);
3658 return false; /* need to add count continuation */
3661 spin_lock(&si->cont_lock);
3662 offset &= ~PAGE_MASK;
3663 page = list_entry(head->lru.next, struct page, lru);
3664 map = kmap_atomic(page) + offset;
3666 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3667 goto init_map; /* jump over SWAP_CONT_MAX checks */
3669 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3671 * Think of how you add 1 to 999
3673 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3675 page = list_entry(page->lru.next, struct page, lru);
3676 BUG_ON(page == head);
3677 map = kmap_atomic(page) + offset;
3679 if (*map == SWAP_CONT_MAX) {
3681 page = list_entry(page->lru.next, struct page, lru);
3683 ret = false; /* add count continuation */
3686 map = kmap_atomic(page) + offset;
3687 init_map: *map = 0; /* we didn't zero the page */
3691 page = list_entry(page->lru.prev, struct page, lru);
3692 while (page != head) {
3693 map = kmap_atomic(page) + offset;
3694 *map = COUNT_CONTINUED;
3696 page = list_entry(page->lru.prev, struct page, lru);
3698 ret = true; /* incremented */
3700 } else { /* decrementing */
3702 * Think of how you subtract 1 from 1000
3704 BUG_ON(count != COUNT_CONTINUED);
3705 while (*map == COUNT_CONTINUED) {
3707 page = list_entry(page->lru.next, struct page, lru);
3708 BUG_ON(page == head);
3709 map = kmap_atomic(page) + offset;
3716 page = list_entry(page->lru.prev, struct page, lru);
3717 while (page != head) {
3718 map = kmap_atomic(page) + offset;
3719 *map = SWAP_CONT_MAX | count;
3720 count = COUNT_CONTINUED;
3722 page = list_entry(page->lru.prev, struct page, lru);
3724 ret = count == COUNT_CONTINUED;
3727 spin_unlock(&si->cont_lock);
3732 * free_swap_count_continuations - swapoff free all the continuation pages
3733 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3735 static void free_swap_count_continuations(struct swap_info_struct *si)
3739 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3741 head = vmalloc_to_page(si->swap_map + offset);
3742 if (page_private(head)) {
3743 struct page *page, *next;
3745 list_for_each_entry_safe(page, next, &head->lru, lru) {
3746 list_del(&page->lru);
3753 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3754 void mem_cgroup_throttle_swaprate(struct mem_cgroup *memcg, int node,
3757 struct swap_info_struct *si, *next;
3758 if (!(gfp_mask & __GFP_IO) || !memcg)
3761 if (!blk_cgroup_congested())
3765 * We've already scheduled a throttle, avoid taking the global swap
3768 if (current->throttle_queue)
3771 spin_lock(&swap_avail_lock);
3772 plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
3773 avail_lists[node]) {
3775 blkcg_schedule_throttle(bdev_get_queue(si->bdev),
3780 spin_unlock(&swap_avail_lock);
3784 static int __init swapfile_init(void)
3788 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3790 if (!swap_avail_heads) {
3791 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3796 plist_head_init(&swap_avail_heads[nid]);
3800 subsys_initcall(swapfile_init);