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;
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 PLIST_HEAD(swap_avail_head);
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 static inline unsigned char swap_count(unsigned char ent)
101 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
104 /* returns 1 if swap entry is freed */
106 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
108 swp_entry_t entry = swp_entry(si->type, offset);
112 page = find_get_page(swap_address_space(entry), swp_offset(entry));
116 * This function is called from scan_swap_map() and it's called
117 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
118 * We have to use trylock for avoiding deadlock. This is a special
119 * case and you should use try_to_free_swap() with explicit lock_page()
120 * in usual operations.
122 if (trylock_page(page)) {
123 ret = try_to_free_swap(page);
131 * swapon tell device that all the old swap contents can be discarded,
132 * to allow the swap device to optimize its wear-levelling.
134 static int discard_swap(struct swap_info_struct *si)
136 struct swap_extent *se;
137 sector_t start_block;
141 /* Do not discard the swap header page! */
142 se = &si->first_swap_extent;
143 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
144 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
146 err = blkdev_issue_discard(si->bdev, start_block,
147 nr_blocks, GFP_KERNEL, 0);
153 list_for_each_entry(se, &si->first_swap_extent.list, list) {
154 start_block = se->start_block << (PAGE_SHIFT - 9);
155 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
157 err = blkdev_issue_discard(si->bdev, start_block,
158 nr_blocks, GFP_KERNEL, 0);
164 return err; /* That will often be -EOPNOTSUPP */
168 * swap allocation tell device that a cluster of swap can now be discarded,
169 * to allow the swap device to optimize its wear-levelling.
171 static void discard_swap_cluster(struct swap_info_struct *si,
172 pgoff_t start_page, pgoff_t nr_pages)
174 struct swap_extent *se = si->curr_swap_extent;
175 int found_extent = 0;
178 if (se->start_page <= start_page &&
179 start_page < se->start_page + se->nr_pages) {
180 pgoff_t offset = start_page - se->start_page;
181 sector_t start_block = se->start_block + offset;
182 sector_t nr_blocks = se->nr_pages - offset;
184 if (nr_blocks > nr_pages)
185 nr_blocks = nr_pages;
186 start_page += nr_blocks;
187 nr_pages -= nr_blocks;
190 si->curr_swap_extent = se;
192 start_block <<= PAGE_SHIFT - 9;
193 nr_blocks <<= PAGE_SHIFT - 9;
194 if (blkdev_issue_discard(si->bdev, start_block,
195 nr_blocks, GFP_NOIO, 0))
199 se = list_next_entry(se, list);
203 #ifdef CONFIG_THP_SWAP
204 #define SWAPFILE_CLUSTER HPAGE_PMD_NR
206 #define SWAPFILE_CLUSTER 256
208 #define LATENCY_LIMIT 256
210 static inline void cluster_set_flag(struct swap_cluster_info *info,
216 static inline unsigned int cluster_count(struct swap_cluster_info *info)
221 static inline void cluster_set_count(struct swap_cluster_info *info,
227 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
228 unsigned int c, unsigned int f)
234 static inline unsigned int cluster_next(struct swap_cluster_info *info)
239 static inline void cluster_set_next(struct swap_cluster_info *info,
245 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
246 unsigned int n, unsigned int f)
252 static inline bool cluster_is_free(struct swap_cluster_info *info)
254 return info->flags & CLUSTER_FLAG_FREE;
257 static inline bool cluster_is_null(struct swap_cluster_info *info)
259 return info->flags & CLUSTER_FLAG_NEXT_NULL;
262 static inline void cluster_set_null(struct swap_cluster_info *info)
264 info->flags = CLUSTER_FLAG_NEXT_NULL;
268 static inline bool cluster_is_huge(struct swap_cluster_info *info)
270 return info->flags & CLUSTER_FLAG_HUGE;
273 static inline void cluster_clear_huge(struct swap_cluster_info *info)
275 info->flags &= ~CLUSTER_FLAG_HUGE;
278 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
279 unsigned long offset)
281 struct swap_cluster_info *ci;
283 ci = si->cluster_info;
285 ci += offset / SWAPFILE_CLUSTER;
286 spin_lock(&ci->lock);
291 static inline void unlock_cluster(struct swap_cluster_info *ci)
294 spin_unlock(&ci->lock);
297 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
298 struct swap_info_struct *si,
299 unsigned long offset)
301 struct swap_cluster_info *ci;
303 ci = lock_cluster(si, offset);
305 spin_lock(&si->lock);
310 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
311 struct swap_cluster_info *ci)
316 spin_unlock(&si->lock);
319 static inline bool cluster_list_empty(struct swap_cluster_list *list)
321 return cluster_is_null(&list->head);
324 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
326 return cluster_next(&list->head);
329 static void cluster_list_init(struct swap_cluster_list *list)
331 cluster_set_null(&list->head);
332 cluster_set_null(&list->tail);
335 static void cluster_list_add_tail(struct swap_cluster_list *list,
336 struct swap_cluster_info *ci,
339 if (cluster_list_empty(list)) {
340 cluster_set_next_flag(&list->head, idx, 0);
341 cluster_set_next_flag(&list->tail, idx, 0);
343 struct swap_cluster_info *ci_tail;
344 unsigned int tail = cluster_next(&list->tail);
347 * Nested cluster lock, but both cluster locks are
348 * only acquired when we held swap_info_struct->lock
351 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
352 cluster_set_next(ci_tail, idx);
353 spin_unlock(&ci_tail->lock);
354 cluster_set_next_flag(&list->tail, idx, 0);
358 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
359 struct swap_cluster_info *ci)
363 idx = cluster_next(&list->head);
364 if (cluster_next(&list->tail) == idx) {
365 cluster_set_null(&list->head);
366 cluster_set_null(&list->tail);
368 cluster_set_next_flag(&list->head,
369 cluster_next(&ci[idx]), 0);
374 /* Add a cluster to discard list and schedule it to do discard */
375 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
379 * If scan_swap_map() can't find a free cluster, it will check
380 * si->swap_map directly. To make sure the discarding cluster isn't
381 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
382 * will be cleared after discard
384 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
385 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
387 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
389 schedule_work(&si->discard_work);
392 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
394 struct swap_cluster_info *ci = si->cluster_info;
396 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
397 cluster_list_add_tail(&si->free_clusters, ci, idx);
401 * Doing discard actually. After a cluster discard is finished, the cluster
402 * will be added to free cluster list. caller should hold si->lock.
404 static void swap_do_scheduled_discard(struct swap_info_struct *si)
406 struct swap_cluster_info *info, *ci;
409 info = si->cluster_info;
411 while (!cluster_list_empty(&si->discard_clusters)) {
412 idx = cluster_list_del_first(&si->discard_clusters, info);
413 spin_unlock(&si->lock);
415 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
418 spin_lock(&si->lock);
419 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
420 __free_cluster(si, idx);
421 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
422 0, SWAPFILE_CLUSTER);
427 static void swap_discard_work(struct work_struct *work)
429 struct swap_info_struct *si;
431 si = container_of(work, struct swap_info_struct, discard_work);
433 spin_lock(&si->lock);
434 swap_do_scheduled_discard(si);
435 spin_unlock(&si->lock);
438 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
440 struct swap_cluster_info *ci = si->cluster_info;
442 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
443 cluster_list_del_first(&si->free_clusters, ci);
444 cluster_set_count_flag(ci + idx, 0, 0);
447 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
449 struct swap_cluster_info *ci = si->cluster_info + idx;
451 VM_BUG_ON(cluster_count(ci) != 0);
453 * If the swap is discardable, prepare discard the cluster
454 * instead of free it immediately. The cluster will be freed
457 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
458 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
459 swap_cluster_schedule_discard(si, idx);
463 __free_cluster(si, idx);
467 * The cluster corresponding to page_nr will be used. The cluster will be
468 * removed from free cluster list and its usage counter will be increased.
470 static void inc_cluster_info_page(struct swap_info_struct *p,
471 struct swap_cluster_info *cluster_info, unsigned long page_nr)
473 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
477 if (cluster_is_free(&cluster_info[idx]))
478 alloc_cluster(p, idx);
480 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
481 cluster_set_count(&cluster_info[idx],
482 cluster_count(&cluster_info[idx]) + 1);
486 * The cluster corresponding to page_nr decreases one usage. If the usage
487 * counter becomes 0, which means no page in the cluster is in using, we can
488 * optionally discard the cluster and add it to free cluster list.
490 static void dec_cluster_info_page(struct swap_info_struct *p,
491 struct swap_cluster_info *cluster_info, unsigned long page_nr)
493 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
498 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
499 cluster_set_count(&cluster_info[idx],
500 cluster_count(&cluster_info[idx]) - 1);
502 if (cluster_count(&cluster_info[idx]) == 0)
503 free_cluster(p, idx);
507 * It's possible scan_swap_map() uses a free cluster in the middle of free
508 * cluster list. Avoiding such abuse to avoid list corruption.
511 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
512 unsigned long offset)
514 struct percpu_cluster *percpu_cluster;
517 offset /= SWAPFILE_CLUSTER;
518 conflict = !cluster_list_empty(&si->free_clusters) &&
519 offset != cluster_list_first(&si->free_clusters) &&
520 cluster_is_free(&si->cluster_info[offset]);
525 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
526 cluster_set_null(&percpu_cluster->index);
531 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
532 * might involve allocating a new cluster for current CPU too.
534 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
535 unsigned long *offset, unsigned long *scan_base)
537 struct percpu_cluster *cluster;
538 struct swap_cluster_info *ci;
540 unsigned long tmp, max;
543 cluster = this_cpu_ptr(si->percpu_cluster);
544 if (cluster_is_null(&cluster->index)) {
545 if (!cluster_list_empty(&si->free_clusters)) {
546 cluster->index = si->free_clusters.head;
547 cluster->next = cluster_next(&cluster->index) *
549 } else if (!cluster_list_empty(&si->discard_clusters)) {
551 * we don't have free cluster but have some clusters in
552 * discarding, do discard now and reclaim them
554 swap_do_scheduled_discard(si);
555 *scan_base = *offset = si->cluster_next;
564 * Other CPUs can use our cluster if they can't find a free cluster,
565 * check if there is still free entry in the cluster
568 max = min_t(unsigned long, si->max,
569 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
571 cluster_set_null(&cluster->index);
574 ci = lock_cluster(si, tmp);
576 if (!si->swap_map[tmp]) {
584 cluster_set_null(&cluster->index);
587 cluster->next = tmp + 1;
593 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
594 unsigned int nr_entries)
596 unsigned int end = offset + nr_entries - 1;
598 if (offset == si->lowest_bit)
599 si->lowest_bit += nr_entries;
600 if (end == si->highest_bit)
601 si->highest_bit -= nr_entries;
602 si->inuse_pages += nr_entries;
603 if (si->inuse_pages == si->pages) {
604 si->lowest_bit = si->max;
606 spin_lock(&swap_avail_lock);
607 plist_del(&si->avail_list, &swap_avail_head);
608 spin_unlock(&swap_avail_lock);
612 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
613 unsigned int nr_entries)
615 unsigned long end = offset + nr_entries - 1;
616 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
618 if (offset < si->lowest_bit)
619 si->lowest_bit = offset;
620 if (end > si->highest_bit) {
621 bool was_full = !si->highest_bit;
623 si->highest_bit = end;
624 if (was_full && (si->flags & SWP_WRITEOK)) {
625 spin_lock(&swap_avail_lock);
626 WARN_ON(!plist_node_empty(&si->avail_list));
627 if (plist_node_empty(&si->avail_list))
628 plist_add(&si->avail_list, &swap_avail_head);
629 spin_unlock(&swap_avail_lock);
632 atomic_long_add(nr_entries, &nr_swap_pages);
633 si->inuse_pages -= nr_entries;
634 if (si->flags & SWP_BLKDEV)
635 swap_slot_free_notify =
636 si->bdev->bd_disk->fops->swap_slot_free_notify;
638 swap_slot_free_notify = NULL;
639 while (offset <= end) {
640 frontswap_invalidate_page(si->type, offset);
641 if (swap_slot_free_notify)
642 swap_slot_free_notify(si->bdev, offset);
647 static int scan_swap_map_slots(struct swap_info_struct *si,
648 unsigned char usage, int nr,
651 struct swap_cluster_info *ci;
652 unsigned long offset;
653 unsigned long scan_base;
654 unsigned long last_in_cluster = 0;
655 int latency_ration = LATENCY_LIMIT;
662 * We try to cluster swap pages by allocating them sequentially
663 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
664 * way, however, we resort to first-free allocation, starting
665 * a new cluster. This prevents us from scattering swap pages
666 * all over the entire swap partition, so that we reduce
667 * overall disk seek times between swap pages. -- sct
668 * But we do now try to find an empty cluster. -Andrea
669 * And we let swap pages go all over an SSD partition. Hugh
672 si->flags += SWP_SCANNING;
673 scan_base = offset = si->cluster_next;
676 if (si->cluster_info) {
677 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
683 if (unlikely(!si->cluster_nr--)) {
684 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
685 si->cluster_nr = SWAPFILE_CLUSTER - 1;
689 spin_unlock(&si->lock);
692 * If seek is expensive, start searching for new cluster from
693 * start of partition, to minimize the span of allocated swap.
694 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
695 * case, just handled by scan_swap_map_try_ssd_cluster() above.
697 scan_base = offset = si->lowest_bit;
698 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
700 /* Locate the first empty (unaligned) cluster */
701 for (; last_in_cluster <= si->highest_bit; offset++) {
702 if (si->swap_map[offset])
703 last_in_cluster = offset + SWAPFILE_CLUSTER;
704 else if (offset == last_in_cluster) {
705 spin_lock(&si->lock);
706 offset -= SWAPFILE_CLUSTER - 1;
707 si->cluster_next = offset;
708 si->cluster_nr = SWAPFILE_CLUSTER - 1;
711 if (unlikely(--latency_ration < 0)) {
713 latency_ration = LATENCY_LIMIT;
718 spin_lock(&si->lock);
719 si->cluster_nr = SWAPFILE_CLUSTER - 1;
723 if (si->cluster_info) {
724 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
725 /* take a break if we already got some slots */
728 if (!scan_swap_map_try_ssd_cluster(si, &offset,
733 if (!(si->flags & SWP_WRITEOK))
735 if (!si->highest_bit)
737 if (offset > si->highest_bit)
738 scan_base = offset = si->lowest_bit;
740 ci = lock_cluster(si, offset);
741 /* reuse swap entry of cache-only swap if not busy. */
742 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
745 spin_unlock(&si->lock);
746 swap_was_freed = __try_to_reclaim_swap(si, offset);
747 spin_lock(&si->lock);
748 /* entry was freed successfully, try to use this again */
751 goto scan; /* check next one */
754 if (si->swap_map[offset]) {
761 si->swap_map[offset] = usage;
762 inc_cluster_info_page(si, si->cluster_info, offset);
765 swap_range_alloc(si, offset, 1);
766 si->cluster_next = offset + 1;
767 slots[n_ret++] = swp_entry(si->type, offset);
769 /* got enough slots or reach max slots? */
770 if ((n_ret == nr) || (offset >= si->highest_bit))
773 /* search for next available slot */
775 /* time to take a break? */
776 if (unlikely(--latency_ration < 0)) {
779 spin_unlock(&si->lock);
781 spin_lock(&si->lock);
782 latency_ration = LATENCY_LIMIT;
785 /* try to get more slots in cluster */
786 if (si->cluster_info) {
787 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
795 /* non-ssd case, still more slots in cluster? */
796 if (si->cluster_nr && !si->swap_map[offset]) {
802 si->flags -= SWP_SCANNING;
806 spin_unlock(&si->lock);
807 while (++offset <= si->highest_bit) {
808 if (!si->swap_map[offset]) {
809 spin_lock(&si->lock);
812 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
813 spin_lock(&si->lock);
816 if (unlikely(--latency_ration < 0)) {
818 latency_ration = LATENCY_LIMIT;
821 offset = si->lowest_bit;
822 while (offset < scan_base) {
823 if (!si->swap_map[offset]) {
824 spin_lock(&si->lock);
827 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
828 spin_lock(&si->lock);
831 if (unlikely(--latency_ration < 0)) {
833 latency_ration = LATENCY_LIMIT;
837 spin_lock(&si->lock);
840 si->flags -= SWP_SCANNING;
844 #ifdef CONFIG_THP_SWAP
845 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
848 struct swap_cluster_info *ci;
849 unsigned long offset, i;
852 if (cluster_list_empty(&si->free_clusters))
855 idx = cluster_list_first(&si->free_clusters);
856 offset = idx * SWAPFILE_CLUSTER;
857 ci = lock_cluster(si, offset);
858 alloc_cluster(si, idx);
859 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
861 map = si->swap_map + offset;
862 for (i = 0; i < SWAPFILE_CLUSTER; i++)
863 map[i] = SWAP_HAS_CACHE;
865 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
866 *slot = swp_entry(si->type, offset);
871 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
873 unsigned long offset = idx * SWAPFILE_CLUSTER;
874 struct swap_cluster_info *ci;
876 ci = lock_cluster(si, offset);
877 cluster_set_count_flag(ci, 0, 0);
878 free_cluster(si, idx);
880 swap_range_free(si, offset, SWAPFILE_CLUSTER);
883 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
888 #endif /* CONFIG_THP_SWAP */
890 static unsigned long scan_swap_map(struct swap_info_struct *si,
896 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
899 return swp_offset(entry);
905 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
907 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
908 struct swap_info_struct *si, *next;
912 /* Only single cluster request supported */
913 WARN_ON_ONCE(n_goal > 1 && cluster);
915 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
919 if (n_goal > SWAP_BATCH)
922 if (n_goal > avail_pgs)
925 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
927 spin_lock(&swap_avail_lock);
930 plist_for_each_entry_safe(si, next, &swap_avail_head, avail_list) {
931 /* requeue si to after same-priority siblings */
932 plist_requeue(&si->avail_list, &swap_avail_head);
933 spin_unlock(&swap_avail_lock);
934 spin_lock(&si->lock);
935 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
936 spin_lock(&swap_avail_lock);
937 if (plist_node_empty(&si->avail_list)) {
938 spin_unlock(&si->lock);
941 WARN(!si->highest_bit,
942 "swap_info %d in list but !highest_bit\n",
944 WARN(!(si->flags & SWP_WRITEOK),
945 "swap_info %d in list but !SWP_WRITEOK\n",
947 plist_del(&si->avail_list, &swap_avail_head);
948 spin_unlock(&si->lock);
952 if (!(si->flags & SWP_FILE))
953 n_ret = swap_alloc_cluster(si, swp_entries);
955 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
956 n_goal, swp_entries);
957 spin_unlock(&si->lock);
958 if (n_ret || cluster)
960 pr_debug("scan_swap_map of si %d failed to find offset\n",
963 spin_lock(&swap_avail_lock);
966 * if we got here, it's likely that si was almost full before,
967 * and since scan_swap_map() can drop the si->lock, multiple
968 * callers probably all tried to get a page from the same si
969 * and it filled up before we could get one; or, the si filled
970 * up between us dropping swap_avail_lock and taking si->lock.
971 * Since we dropped the swap_avail_lock, the swap_avail_head
972 * list may have been modified; so if next is still in the
973 * swap_avail_head list then try it, otherwise start over
974 * if we have not gotten any slots.
976 if (plist_node_empty(&next->avail_list))
980 spin_unlock(&swap_avail_lock);
984 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
990 /* The only caller of this function is now suspend routine */
991 swp_entry_t get_swap_page_of_type(int type)
993 struct swap_info_struct *si;
996 si = swap_info[type];
997 spin_lock(&si->lock);
998 if (si && (si->flags & SWP_WRITEOK)) {
999 atomic_long_dec(&nr_swap_pages);
1000 /* This is called for allocating swap entry, not cache */
1001 offset = scan_swap_map(si, 1);
1003 spin_unlock(&si->lock);
1004 return swp_entry(type, offset);
1006 atomic_long_inc(&nr_swap_pages);
1008 spin_unlock(&si->lock);
1009 return (swp_entry_t) {0};
1012 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1014 struct swap_info_struct *p;
1015 unsigned long offset, type;
1019 type = swp_type(entry);
1020 if (type >= nr_swapfiles)
1022 p = swap_info[type];
1023 if (!(p->flags & SWP_USED))
1025 offset = swp_offset(entry);
1026 if (offset >= p->max)
1031 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1034 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1037 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1042 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1044 struct swap_info_struct *p;
1046 p = __swap_info_get(entry);
1049 if (!p->swap_map[swp_offset(entry)])
1054 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1060 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1062 struct swap_info_struct *p;
1064 p = _swap_info_get(entry);
1066 spin_lock(&p->lock);
1070 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1071 struct swap_info_struct *q)
1073 struct swap_info_struct *p;
1075 p = _swap_info_get(entry);
1079 spin_unlock(&q->lock);
1081 spin_lock(&p->lock);
1086 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1087 swp_entry_t entry, unsigned char usage)
1089 struct swap_cluster_info *ci;
1090 unsigned long offset = swp_offset(entry);
1091 unsigned char count;
1092 unsigned char has_cache;
1094 ci = lock_cluster_or_swap_info(p, offset);
1096 count = p->swap_map[offset];
1098 has_cache = count & SWAP_HAS_CACHE;
1099 count &= ~SWAP_HAS_CACHE;
1101 if (usage == SWAP_HAS_CACHE) {
1102 VM_BUG_ON(!has_cache);
1104 } else if (count == SWAP_MAP_SHMEM) {
1106 * Or we could insist on shmem.c using a special
1107 * swap_shmem_free() and free_shmem_swap_and_cache()...
1110 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1111 if (count == COUNT_CONTINUED) {
1112 if (swap_count_continued(p, offset, count))
1113 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1115 count = SWAP_MAP_MAX;
1120 usage = count | has_cache;
1121 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1123 unlock_cluster_or_swap_info(p, ci);
1128 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1130 struct swap_cluster_info *ci;
1131 unsigned long offset = swp_offset(entry);
1132 unsigned char count;
1134 ci = lock_cluster(p, offset);
1135 count = p->swap_map[offset];
1136 VM_BUG_ON(count != SWAP_HAS_CACHE);
1137 p->swap_map[offset] = 0;
1138 dec_cluster_info_page(p, p->cluster_info, offset);
1141 mem_cgroup_uncharge_swap(entry, 1);
1142 swap_range_free(p, offset, 1);
1146 * Caller has made sure that the swap device corresponding to entry
1147 * is still around or has not been recycled.
1149 void swap_free(swp_entry_t entry)
1151 struct swap_info_struct *p;
1153 p = _swap_info_get(entry);
1155 if (!__swap_entry_free(p, entry, 1))
1156 free_swap_slot(entry);
1161 * Called after dropping swapcache to decrease refcnt to swap entries.
1163 static void swapcache_free(swp_entry_t entry)
1165 struct swap_info_struct *p;
1167 p = _swap_info_get(entry);
1169 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1170 free_swap_slot(entry);
1174 #ifdef CONFIG_THP_SWAP
1175 static void swapcache_free_cluster(swp_entry_t entry)
1177 unsigned long offset = swp_offset(entry);
1178 unsigned long idx = offset / SWAPFILE_CLUSTER;
1179 struct swap_cluster_info *ci;
1180 struct swap_info_struct *si;
1182 unsigned int i, free_entries = 0;
1185 si = _swap_info_get(entry);
1189 ci = lock_cluster(si, offset);
1190 VM_BUG_ON(!cluster_is_huge(ci));
1191 map = si->swap_map + offset;
1192 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1194 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1195 if (val == SWAP_HAS_CACHE)
1198 if (!free_entries) {
1199 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1200 map[i] &= ~SWAP_HAS_CACHE;
1202 cluster_clear_huge(ci);
1204 if (free_entries == SWAPFILE_CLUSTER) {
1205 spin_lock(&si->lock);
1206 ci = lock_cluster(si, offset);
1207 memset(map, 0, SWAPFILE_CLUSTER);
1209 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1210 swap_free_cluster(si, idx);
1211 spin_unlock(&si->lock);
1212 } else if (free_entries) {
1213 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1214 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1215 free_swap_slot(entry);
1220 static inline void swapcache_free_cluster(swp_entry_t entry)
1223 #endif /* CONFIG_THP_SWAP */
1225 void put_swap_page(struct page *page, swp_entry_t entry)
1227 if (!PageTransHuge(page))
1228 swapcache_free(entry);
1230 swapcache_free_cluster(entry);
1233 static int swp_entry_cmp(const void *ent1, const void *ent2)
1235 const swp_entry_t *e1 = ent1, *e2 = ent2;
1237 return (int)swp_type(*e1) - (int)swp_type(*e2);
1240 void swapcache_free_entries(swp_entry_t *entries, int n)
1242 struct swap_info_struct *p, *prev;
1252 * Sort swap entries by swap device, so each lock is only taken once.
1253 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1254 * so low that it isn't necessary to optimize further.
1256 if (nr_swapfiles > 1)
1257 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1258 for (i = 0; i < n; ++i) {
1259 p = swap_info_get_cont(entries[i], prev);
1261 swap_entry_free(p, entries[i]);
1265 spin_unlock(&p->lock);
1269 * How many references to page are currently swapped out?
1270 * This does not give an exact answer when swap count is continued,
1271 * but does include the high COUNT_CONTINUED flag to allow for that.
1273 int page_swapcount(struct page *page)
1276 struct swap_info_struct *p;
1277 struct swap_cluster_info *ci;
1279 unsigned long offset;
1281 entry.val = page_private(page);
1282 p = _swap_info_get(entry);
1284 offset = swp_offset(entry);
1285 ci = lock_cluster_or_swap_info(p, offset);
1286 count = swap_count(p->swap_map[offset]);
1287 unlock_cluster_or_swap_info(p, ci);
1292 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1295 pgoff_t offset = swp_offset(entry);
1296 struct swap_cluster_info *ci;
1298 ci = lock_cluster_or_swap_info(si, offset);
1299 count = swap_count(si->swap_map[offset]);
1300 unlock_cluster_or_swap_info(si, ci);
1305 * How many references to @entry are currently swapped out?
1306 * This does not give an exact answer when swap count is continued,
1307 * but does include the high COUNT_CONTINUED flag to allow for that.
1309 int __swp_swapcount(swp_entry_t entry)
1312 struct swap_info_struct *si;
1314 si = __swap_info_get(entry);
1316 count = swap_swapcount(si, entry);
1321 * How many references to @entry are currently swapped out?
1322 * This considers COUNT_CONTINUED so it returns exact answer.
1324 int swp_swapcount(swp_entry_t entry)
1326 int count, tmp_count, n;
1327 struct swap_info_struct *p;
1328 struct swap_cluster_info *ci;
1333 p = _swap_info_get(entry);
1337 offset = swp_offset(entry);
1339 ci = lock_cluster_or_swap_info(p, offset);
1341 count = swap_count(p->swap_map[offset]);
1342 if (!(count & COUNT_CONTINUED))
1345 count &= ~COUNT_CONTINUED;
1346 n = SWAP_MAP_MAX + 1;
1348 page = vmalloc_to_page(p->swap_map + offset);
1349 offset &= ~PAGE_MASK;
1350 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1353 page = list_next_entry(page, lru);
1354 map = kmap_atomic(page);
1355 tmp_count = map[offset];
1358 count += (tmp_count & ~COUNT_CONTINUED) * n;
1359 n *= (SWAP_CONT_MAX + 1);
1360 } while (tmp_count & COUNT_CONTINUED);
1362 unlock_cluster_or_swap_info(p, ci);
1366 #ifdef CONFIG_THP_SWAP
1367 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1370 struct swap_cluster_info *ci;
1371 unsigned char *map = si->swap_map;
1372 unsigned long roffset = swp_offset(entry);
1373 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1377 ci = lock_cluster_or_swap_info(si, offset);
1378 if (!ci || !cluster_is_huge(ci)) {
1379 if (map[roffset] != SWAP_HAS_CACHE)
1383 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1384 if (map[offset + i] != SWAP_HAS_CACHE) {
1390 unlock_cluster_or_swap_info(si, ci);
1394 static bool page_swapped(struct page *page)
1397 struct swap_info_struct *si;
1399 if (likely(!PageTransCompound(page)))
1400 return page_swapcount(page) != 0;
1402 page = compound_head(page);
1403 entry.val = page_private(page);
1404 si = _swap_info_get(entry);
1406 return swap_page_trans_huge_swapped(si, entry);
1410 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1411 int *total_swapcount)
1413 int i, map_swapcount, _total_mapcount, _total_swapcount;
1414 unsigned long offset = 0;
1415 struct swap_info_struct *si;
1416 struct swap_cluster_info *ci = NULL;
1417 unsigned char *map = NULL;
1418 int mapcount, swapcount = 0;
1420 /* hugetlbfs shouldn't call it */
1421 VM_BUG_ON_PAGE(PageHuge(page), page);
1423 if (likely(!PageTransCompound(page))) {
1424 mapcount = atomic_read(&page->_mapcount) + 1;
1426 *total_mapcount = mapcount;
1427 if (PageSwapCache(page))
1428 swapcount = page_swapcount(page);
1429 if (total_swapcount)
1430 *total_swapcount = swapcount;
1431 return mapcount + swapcount;
1434 page = compound_head(page);
1436 _total_mapcount = _total_swapcount = map_swapcount = 0;
1437 if (PageSwapCache(page)) {
1440 entry.val = page_private(page);
1441 si = _swap_info_get(entry);
1444 offset = swp_offset(entry);
1448 ci = lock_cluster(si, offset);
1449 for (i = 0; i < HPAGE_PMD_NR; i++) {
1450 mapcount = atomic_read(&page[i]._mapcount) + 1;
1451 _total_mapcount += mapcount;
1453 swapcount = swap_count(map[offset + i]);
1454 _total_swapcount += swapcount;
1456 map_swapcount = max(map_swapcount, mapcount + swapcount);
1459 if (PageDoubleMap(page)) {
1461 _total_mapcount -= HPAGE_PMD_NR;
1463 mapcount = compound_mapcount(page);
1464 map_swapcount += mapcount;
1465 _total_mapcount += mapcount;
1467 *total_mapcount = _total_mapcount;
1468 if (total_swapcount)
1469 *total_swapcount = _total_swapcount;
1471 return map_swapcount;
1474 #define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1475 #define page_swapped(page) (page_swapcount(page) != 0)
1477 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1478 int *total_swapcount)
1480 int mapcount, swapcount = 0;
1482 /* hugetlbfs shouldn't call it */
1483 VM_BUG_ON_PAGE(PageHuge(page), page);
1485 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1486 if (PageSwapCache(page))
1487 swapcount = page_swapcount(page);
1488 if (total_swapcount)
1489 *total_swapcount = swapcount;
1490 return mapcount + swapcount;
1495 * We can write to an anon page without COW if there are no other references
1496 * to it. And as a side-effect, free up its swap: because the old content
1497 * on disk will never be read, and seeking back there to write new content
1498 * later would only waste time away from clustering.
1500 * NOTE: total_map_swapcount should not be relied upon by the caller if
1501 * reuse_swap_page() returns false, but it may be always overwritten
1502 * (see the other implementation for CONFIG_SWAP=n).
1504 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1506 int count, total_mapcount, total_swapcount;
1508 VM_BUG_ON_PAGE(!PageLocked(page), page);
1509 if (unlikely(PageKsm(page)))
1511 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1513 if (total_map_swapcount)
1514 *total_map_swapcount = total_mapcount + total_swapcount;
1515 if (count == 1 && PageSwapCache(page) &&
1516 (likely(!PageTransCompound(page)) ||
1517 /* The remaining swap count will be freed soon */
1518 total_swapcount == page_swapcount(page))) {
1519 if (!PageWriteback(page)) {
1520 page = compound_head(page);
1521 delete_from_swap_cache(page);
1525 struct swap_info_struct *p;
1527 entry.val = page_private(page);
1528 p = swap_info_get(entry);
1529 if (p->flags & SWP_STABLE_WRITES) {
1530 spin_unlock(&p->lock);
1533 spin_unlock(&p->lock);
1541 * If swap is getting full, or if there are no more mappings of this page,
1542 * then try_to_free_swap is called to free its swap space.
1544 int try_to_free_swap(struct page *page)
1546 VM_BUG_ON_PAGE(!PageLocked(page), page);
1548 if (!PageSwapCache(page))
1550 if (PageWriteback(page))
1552 if (page_swapped(page))
1556 * Once hibernation has begun to create its image of memory,
1557 * there's a danger that one of the calls to try_to_free_swap()
1558 * - most probably a call from __try_to_reclaim_swap() while
1559 * hibernation is allocating its own swap pages for the image,
1560 * but conceivably even a call from memory reclaim - will free
1561 * the swap from a page which has already been recorded in the
1562 * image as a clean swapcache page, and then reuse its swap for
1563 * another page of the image. On waking from hibernation, the
1564 * original page might be freed under memory pressure, then
1565 * later read back in from swap, now with the wrong data.
1567 * Hibernation suspends storage while it is writing the image
1568 * to disk so check that here.
1570 if (pm_suspended_storage())
1573 page = compound_head(page);
1574 delete_from_swap_cache(page);
1580 * Free the swap entry like above, but also try to
1581 * free the page cache entry if it is the last user.
1583 int free_swap_and_cache(swp_entry_t entry)
1585 struct swap_info_struct *p;
1586 struct page *page = NULL;
1587 unsigned char count;
1589 if (non_swap_entry(entry))
1592 p = _swap_info_get(entry);
1594 count = __swap_entry_free(p, entry, 1);
1595 if (count == SWAP_HAS_CACHE &&
1596 !swap_page_trans_huge_swapped(p, entry)) {
1597 page = find_get_page(swap_address_space(entry),
1599 if (page && !trylock_page(page)) {
1604 free_swap_slot(entry);
1608 * Not mapped elsewhere, or swap space full? Free it!
1609 * Also recheck PageSwapCache now page is locked (above).
1611 if (PageSwapCache(page) && !PageWriteback(page) &&
1612 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1613 !swap_page_trans_huge_swapped(p, entry)) {
1614 page = compound_head(page);
1615 delete_from_swap_cache(page);
1624 #ifdef CONFIG_HIBERNATION
1626 * Find the swap type that corresponds to given device (if any).
1628 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1629 * from 0, in which the swap header is expected to be located.
1631 * This is needed for the suspend to disk (aka swsusp).
1633 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1635 struct block_device *bdev = NULL;
1639 bdev = bdget(device);
1641 spin_lock(&swap_lock);
1642 for (type = 0; type < nr_swapfiles; type++) {
1643 struct swap_info_struct *sis = swap_info[type];
1645 if (!(sis->flags & SWP_WRITEOK))
1650 *bdev_p = bdgrab(sis->bdev);
1652 spin_unlock(&swap_lock);
1655 if (bdev == sis->bdev) {
1656 struct swap_extent *se = &sis->first_swap_extent;
1658 if (se->start_block == offset) {
1660 *bdev_p = bdgrab(sis->bdev);
1662 spin_unlock(&swap_lock);
1668 spin_unlock(&swap_lock);
1676 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1677 * corresponding to given index in swap_info (swap type).
1679 sector_t swapdev_block(int type, pgoff_t offset)
1681 struct block_device *bdev;
1683 if ((unsigned int)type >= nr_swapfiles)
1685 if (!(swap_info[type]->flags & SWP_WRITEOK))
1687 return map_swap_entry(swp_entry(type, offset), &bdev);
1691 * Return either the total number of swap pages of given type, or the number
1692 * of free pages of that type (depending on @free)
1694 * This is needed for software suspend
1696 unsigned int count_swap_pages(int type, int free)
1700 spin_lock(&swap_lock);
1701 if ((unsigned int)type < nr_swapfiles) {
1702 struct swap_info_struct *sis = swap_info[type];
1704 spin_lock(&sis->lock);
1705 if (sis->flags & SWP_WRITEOK) {
1708 n -= sis->inuse_pages;
1710 spin_unlock(&sis->lock);
1712 spin_unlock(&swap_lock);
1715 #endif /* CONFIG_HIBERNATION */
1717 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1719 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1723 * No need to decide whether this PTE shares the swap entry with others,
1724 * just let do_wp_page work it out if a write is requested later - to
1725 * force COW, vm_page_prot omits write permission from any private vma.
1727 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1728 unsigned long addr, swp_entry_t entry, struct page *page)
1730 struct page *swapcache;
1731 struct mem_cgroup *memcg;
1737 page = ksm_might_need_to_copy(page, vma, addr);
1738 if (unlikely(!page))
1741 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1747 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1748 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1749 mem_cgroup_cancel_charge(page, memcg, false);
1754 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1755 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1757 set_pte_at(vma->vm_mm, addr, pte,
1758 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1759 if (page == swapcache) {
1760 page_add_anon_rmap(page, vma, addr, false);
1761 mem_cgroup_commit_charge(page, memcg, true, false);
1762 } else { /* ksm created a completely new copy */
1763 page_add_new_anon_rmap(page, vma, addr, false);
1764 mem_cgroup_commit_charge(page, memcg, false, false);
1765 lru_cache_add_active_or_unevictable(page, vma);
1769 * Move the page to the active list so it is not
1770 * immediately swapped out again after swapon.
1772 activate_page(page);
1774 pte_unmap_unlock(pte, ptl);
1776 if (page != swapcache) {
1783 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1784 unsigned long addr, unsigned long end,
1785 swp_entry_t entry, struct page *page)
1787 pte_t swp_pte = swp_entry_to_pte(entry);
1792 * We don't actually need pte lock while scanning for swp_pte: since
1793 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1794 * page table while we're scanning; though it could get zapped, and on
1795 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1796 * of unmatched parts which look like swp_pte, so unuse_pte must
1797 * recheck under pte lock. Scanning without pte lock lets it be
1798 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1800 pte = pte_offset_map(pmd, addr);
1803 * swapoff spends a _lot_ of time in this loop!
1804 * Test inline before going to call unuse_pte.
1806 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1808 ret = unuse_pte(vma, pmd, addr, entry, page);
1811 pte = pte_offset_map(pmd, addr);
1813 } while (pte++, addr += PAGE_SIZE, addr != end);
1819 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1820 unsigned long addr, unsigned long end,
1821 swp_entry_t entry, struct page *page)
1827 pmd = pmd_offset(pud, addr);
1830 next = pmd_addr_end(addr, end);
1831 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1833 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1836 } while (pmd++, addr = next, addr != end);
1840 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1841 unsigned long addr, unsigned long end,
1842 swp_entry_t entry, struct page *page)
1848 pud = pud_offset(p4d, addr);
1850 next = pud_addr_end(addr, end);
1851 if (pud_none_or_clear_bad(pud))
1853 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1856 } while (pud++, addr = next, addr != end);
1860 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1861 unsigned long addr, unsigned long end,
1862 swp_entry_t entry, struct page *page)
1868 p4d = p4d_offset(pgd, addr);
1870 next = p4d_addr_end(addr, end);
1871 if (p4d_none_or_clear_bad(p4d))
1873 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1876 } while (p4d++, addr = next, addr != end);
1880 static int unuse_vma(struct vm_area_struct *vma,
1881 swp_entry_t entry, struct page *page)
1884 unsigned long addr, end, next;
1887 if (page_anon_vma(page)) {
1888 addr = page_address_in_vma(page, vma);
1889 if (addr == -EFAULT)
1892 end = addr + PAGE_SIZE;
1894 addr = vma->vm_start;
1898 pgd = pgd_offset(vma->vm_mm, addr);
1900 next = pgd_addr_end(addr, end);
1901 if (pgd_none_or_clear_bad(pgd))
1903 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1906 } while (pgd++, addr = next, addr != end);
1910 static int unuse_mm(struct mm_struct *mm,
1911 swp_entry_t entry, struct page *page)
1913 struct vm_area_struct *vma;
1916 if (!down_read_trylock(&mm->mmap_sem)) {
1918 * Activate page so shrink_inactive_list is unlikely to unmap
1919 * its ptes while lock is dropped, so swapoff can make progress.
1921 activate_page(page);
1923 down_read(&mm->mmap_sem);
1926 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1927 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1931 up_read(&mm->mmap_sem);
1932 return (ret < 0)? ret: 0;
1936 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1937 * from current position to next entry still in use.
1938 * Recycle to start on reaching the end, returning 0 when empty.
1940 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1941 unsigned int prev, bool frontswap)
1943 unsigned int max = si->max;
1944 unsigned int i = prev;
1945 unsigned char count;
1948 * No need for swap_lock here: we're just looking
1949 * for whether an entry is in use, not modifying it; false
1950 * hits are okay, and sys_swapoff() has already prevented new
1951 * allocations from this area (while holding swap_lock).
1960 * No entries in use at top of swap_map,
1961 * loop back to start and recheck there.
1967 count = READ_ONCE(si->swap_map[i]);
1968 if (count && swap_count(count) != SWAP_MAP_BAD)
1969 if (!frontswap || frontswap_test(si, i))
1971 if ((i % LATENCY_LIMIT) == 0)
1978 * We completely avoid races by reading each swap page in advance,
1979 * and then search for the process using it. All the necessary
1980 * page table adjustments can then be made atomically.
1982 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1983 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1985 int try_to_unuse(unsigned int type, bool frontswap,
1986 unsigned long pages_to_unuse)
1988 struct swap_info_struct *si = swap_info[type];
1989 struct mm_struct *start_mm;
1990 volatile unsigned char *swap_map; /* swap_map is accessed without
1991 * locking. Mark it as volatile
1992 * to prevent compiler doing
1995 unsigned char swcount;
2002 * When searching mms for an entry, a good strategy is to
2003 * start at the first mm we freed the previous entry from
2004 * (though actually we don't notice whether we or coincidence
2005 * freed the entry). Initialize this start_mm with a hold.
2007 * A simpler strategy would be to start at the last mm we
2008 * freed the previous entry from; but that would take less
2009 * advantage of mmlist ordering, which clusters forked mms
2010 * together, child after parent. If we race with dup_mmap(), we
2011 * prefer to resolve parent before child, lest we miss entries
2012 * duplicated after we scanned child: using last mm would invert
2015 start_mm = &init_mm;
2019 * Keep on scanning until all entries have gone. Usually,
2020 * one pass through swap_map is enough, but not necessarily:
2021 * there are races when an instance of an entry might be missed.
2023 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2024 if (signal_pending(current)) {
2030 * Get a page for the entry, using the existing swap
2031 * cache page if there is one. Otherwise, get a clean
2032 * page and read the swap into it.
2034 swap_map = &si->swap_map[i];
2035 entry = swp_entry(type, i);
2036 page = read_swap_cache_async(entry,
2037 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2040 * Either swap_duplicate() failed because entry
2041 * has been freed independently, and will not be
2042 * reused since sys_swapoff() already disabled
2043 * allocation from here, or alloc_page() failed.
2045 swcount = *swap_map;
2047 * We don't hold lock here, so the swap entry could be
2048 * SWAP_MAP_BAD (when the cluster is discarding).
2049 * Instead of fail out, We can just skip the swap
2050 * entry because swapoff will wait for discarding
2053 if (!swcount || swcount == SWAP_MAP_BAD)
2060 * Don't hold on to start_mm if it looks like exiting.
2062 if (atomic_read(&start_mm->mm_users) == 1) {
2064 start_mm = &init_mm;
2069 * Wait for and lock page. When do_swap_page races with
2070 * try_to_unuse, do_swap_page can handle the fault much
2071 * faster than try_to_unuse can locate the entry. This
2072 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2073 * defer to do_swap_page in such a case - in some tests,
2074 * do_swap_page and try_to_unuse repeatedly compete.
2076 wait_on_page_locked(page);
2077 wait_on_page_writeback(page);
2079 wait_on_page_writeback(page);
2082 * Remove all references to entry.
2084 swcount = *swap_map;
2085 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2086 retval = shmem_unuse(entry, page);
2087 /* page has already been unlocked and released */
2092 if (swap_count(swcount) && start_mm != &init_mm)
2093 retval = unuse_mm(start_mm, entry, page);
2095 if (swap_count(*swap_map)) {
2096 int set_start_mm = (*swap_map >= swcount);
2097 struct list_head *p = &start_mm->mmlist;
2098 struct mm_struct *new_start_mm = start_mm;
2099 struct mm_struct *prev_mm = start_mm;
2100 struct mm_struct *mm;
2102 mmget(new_start_mm);
2104 spin_lock(&mmlist_lock);
2105 while (swap_count(*swap_map) && !retval &&
2106 (p = p->next) != &start_mm->mmlist) {
2107 mm = list_entry(p, struct mm_struct, mmlist);
2108 if (!mmget_not_zero(mm))
2110 spin_unlock(&mmlist_lock);
2116 swcount = *swap_map;
2117 if (!swap_count(swcount)) /* any usage ? */
2119 else if (mm == &init_mm)
2122 retval = unuse_mm(mm, entry, page);
2124 if (set_start_mm && *swap_map < swcount) {
2125 mmput(new_start_mm);
2130 spin_lock(&mmlist_lock);
2132 spin_unlock(&mmlist_lock);
2135 start_mm = new_start_mm;
2144 * If a reference remains (rare), we would like to leave
2145 * the page in the swap cache; but try_to_unmap could
2146 * then re-duplicate the entry once we drop page lock,
2147 * so we might loop indefinitely; also, that page could
2148 * not be swapped out to other storage meanwhile. So:
2149 * delete from cache even if there's another reference,
2150 * after ensuring that the data has been saved to disk -
2151 * since if the reference remains (rarer), it will be
2152 * read from disk into another page. Splitting into two
2153 * pages would be incorrect if swap supported "shared
2154 * private" pages, but they are handled by tmpfs files.
2156 * Given how unuse_vma() targets one particular offset
2157 * in an anon_vma, once the anon_vma has been determined,
2158 * this splitting happens to be just what is needed to
2159 * handle where KSM pages have been swapped out: re-reading
2160 * is unnecessarily slow, but we can fix that later on.
2162 if (swap_count(*swap_map) &&
2163 PageDirty(page) && PageSwapCache(page)) {
2164 struct writeback_control wbc = {
2165 .sync_mode = WB_SYNC_NONE,
2168 swap_writepage(compound_head(page), &wbc);
2170 wait_on_page_writeback(page);
2174 * It is conceivable that a racing task removed this page from
2175 * swap cache just before we acquired the page lock at the top,
2176 * or while we dropped it in unuse_mm(). The page might even
2177 * be back in swap cache on another swap area: that we must not
2178 * delete, since it may not have been written out to swap yet.
2180 if (PageSwapCache(page) &&
2181 likely(page_private(page) == entry.val) &&
2182 !page_swapped(page))
2183 delete_from_swap_cache(compound_head(page));
2186 * So we could skip searching mms once swap count went
2187 * to 1, we did not mark any present ptes as dirty: must
2188 * mark page dirty so shrink_page_list will preserve it.
2195 * Make sure that we aren't completely killing
2196 * interactive performance.
2199 if (frontswap && pages_to_unuse > 0) {
2200 if (!--pages_to_unuse)
2210 * After a successful try_to_unuse, if no swap is now in use, we know
2211 * we can empty the mmlist. swap_lock must be held on entry and exit.
2212 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2213 * added to the mmlist just after page_duplicate - before would be racy.
2215 static void drain_mmlist(void)
2217 struct list_head *p, *next;
2220 for (type = 0; type < nr_swapfiles; type++)
2221 if (swap_info[type]->inuse_pages)
2223 spin_lock(&mmlist_lock);
2224 list_for_each_safe(p, next, &init_mm.mmlist)
2226 spin_unlock(&mmlist_lock);
2230 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2231 * corresponds to page offset for the specified swap entry.
2232 * Note that the type of this function is sector_t, but it returns page offset
2233 * into the bdev, not sector offset.
2235 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2237 struct swap_info_struct *sis;
2238 struct swap_extent *start_se;
2239 struct swap_extent *se;
2242 sis = swap_info[swp_type(entry)];
2245 offset = swp_offset(entry);
2246 start_se = sis->curr_swap_extent;
2250 if (se->start_page <= offset &&
2251 offset < (se->start_page + se->nr_pages)) {
2252 return se->start_block + (offset - se->start_page);
2254 se = list_next_entry(se, list);
2255 sis->curr_swap_extent = se;
2256 BUG_ON(se == start_se); /* It *must* be present */
2261 * Returns the page offset into bdev for the specified page's swap entry.
2263 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2266 entry.val = page_private(page);
2267 return map_swap_entry(entry, bdev);
2271 * Free all of a swapdev's extent information
2273 static void destroy_swap_extents(struct swap_info_struct *sis)
2275 while (!list_empty(&sis->first_swap_extent.list)) {
2276 struct swap_extent *se;
2278 se = list_first_entry(&sis->first_swap_extent.list,
2279 struct swap_extent, list);
2280 list_del(&se->list);
2284 if (sis->flags & SWP_FILE) {
2285 struct file *swap_file = sis->swap_file;
2286 struct address_space *mapping = swap_file->f_mapping;
2288 sis->flags &= ~SWP_FILE;
2289 mapping->a_ops->swap_deactivate(swap_file);
2294 * Add a block range (and the corresponding page range) into this swapdev's
2295 * extent list. The extent list is kept sorted in page order.
2297 * This function rather assumes that it is called in ascending page order.
2300 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2301 unsigned long nr_pages, sector_t start_block)
2303 struct swap_extent *se;
2304 struct swap_extent *new_se;
2305 struct list_head *lh;
2307 if (start_page == 0) {
2308 se = &sis->first_swap_extent;
2309 sis->curr_swap_extent = se;
2311 se->nr_pages = nr_pages;
2312 se->start_block = start_block;
2315 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2316 se = list_entry(lh, struct swap_extent, list);
2317 BUG_ON(se->start_page + se->nr_pages != start_page);
2318 if (se->start_block + se->nr_pages == start_block) {
2320 se->nr_pages += nr_pages;
2326 * No merge. Insert a new extent, preserving ordering.
2328 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2331 new_se->start_page = start_page;
2332 new_se->nr_pages = nr_pages;
2333 new_se->start_block = start_block;
2335 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2340 * A `swap extent' is a simple thing which maps a contiguous range of pages
2341 * onto a contiguous range of disk blocks. An ordered list of swap extents
2342 * is built at swapon time and is then used at swap_writepage/swap_readpage
2343 * time for locating where on disk a page belongs.
2345 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2346 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2347 * swap files identically.
2349 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2350 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2351 * swapfiles are handled *identically* after swapon time.
2353 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2354 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2355 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2356 * requirements, they are simply tossed out - we will never use those blocks
2359 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2360 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2361 * which will scribble on the fs.
2363 * The amount of disk space which a single swap extent represents varies.
2364 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2365 * extents in the list. To avoid much list walking, we cache the previous
2366 * search location in `curr_swap_extent', and start new searches from there.
2367 * This is extremely effective. The average number of iterations in
2368 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2370 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2372 struct file *swap_file = sis->swap_file;
2373 struct address_space *mapping = swap_file->f_mapping;
2374 struct inode *inode = mapping->host;
2377 if (S_ISBLK(inode->i_mode)) {
2378 ret = add_swap_extent(sis, 0, sis->max, 0);
2383 if (mapping->a_ops->swap_activate) {
2384 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2386 sis->flags |= SWP_FILE;
2387 ret = add_swap_extent(sis, 0, sis->max, 0);
2393 return generic_swapfile_activate(sis, swap_file, span);
2396 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2397 unsigned char *swap_map,
2398 struct swap_cluster_info *cluster_info)
2403 p->prio = --least_priority;
2405 * the plist prio is negated because plist ordering is
2406 * low-to-high, while swap ordering is high-to-low
2408 p->list.prio = -p->prio;
2409 p->avail_list.prio = -p->prio;
2410 p->swap_map = swap_map;
2411 p->cluster_info = cluster_info;
2412 p->flags |= SWP_WRITEOK;
2413 atomic_long_add(p->pages, &nr_swap_pages);
2414 total_swap_pages += p->pages;
2416 assert_spin_locked(&swap_lock);
2418 * both lists are plists, and thus priority ordered.
2419 * swap_active_head needs to be priority ordered for swapoff(),
2420 * which on removal of any swap_info_struct with an auto-assigned
2421 * (i.e. negative) priority increments the auto-assigned priority
2422 * of any lower-priority swap_info_structs.
2423 * swap_avail_head needs to be priority ordered for get_swap_page(),
2424 * which allocates swap pages from the highest available priority
2427 plist_add(&p->list, &swap_active_head);
2428 spin_lock(&swap_avail_lock);
2429 plist_add(&p->avail_list, &swap_avail_head);
2430 spin_unlock(&swap_avail_lock);
2433 static void enable_swap_info(struct swap_info_struct *p, int prio,
2434 unsigned char *swap_map,
2435 struct swap_cluster_info *cluster_info,
2436 unsigned long *frontswap_map)
2438 frontswap_init(p->type, frontswap_map);
2439 spin_lock(&swap_lock);
2440 spin_lock(&p->lock);
2441 _enable_swap_info(p, prio, swap_map, cluster_info);
2442 spin_unlock(&p->lock);
2443 spin_unlock(&swap_lock);
2446 static void reinsert_swap_info(struct swap_info_struct *p)
2448 spin_lock(&swap_lock);
2449 spin_lock(&p->lock);
2450 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2451 spin_unlock(&p->lock);
2452 spin_unlock(&swap_lock);
2455 bool has_usable_swap(void)
2459 spin_lock(&swap_lock);
2460 if (plist_head_empty(&swap_active_head))
2462 spin_unlock(&swap_lock);
2466 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2468 struct swap_info_struct *p = NULL;
2469 unsigned char *swap_map;
2470 struct swap_cluster_info *cluster_info;
2471 unsigned long *frontswap_map;
2472 struct file *swap_file, *victim;
2473 struct address_space *mapping;
2474 struct inode *inode;
2475 struct filename *pathname;
2477 unsigned int old_block_size;
2479 if (!capable(CAP_SYS_ADMIN))
2482 BUG_ON(!current->mm);
2484 pathname = getname(specialfile);
2485 if (IS_ERR(pathname))
2486 return PTR_ERR(pathname);
2488 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2489 err = PTR_ERR(victim);
2493 mapping = victim->f_mapping;
2494 spin_lock(&swap_lock);
2495 plist_for_each_entry(p, &swap_active_head, list) {
2496 if (p->flags & SWP_WRITEOK) {
2497 if (p->swap_file->f_mapping == mapping) {
2505 spin_unlock(&swap_lock);
2508 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2509 vm_unacct_memory(p->pages);
2512 spin_unlock(&swap_lock);
2515 spin_lock(&swap_avail_lock);
2516 plist_del(&p->avail_list, &swap_avail_head);
2517 spin_unlock(&swap_avail_lock);
2518 spin_lock(&p->lock);
2520 struct swap_info_struct *si = p;
2522 plist_for_each_entry_continue(si, &swap_active_head, list) {
2525 si->avail_list.prio--;
2529 plist_del(&p->list, &swap_active_head);
2530 atomic_long_sub(p->pages, &nr_swap_pages);
2531 total_swap_pages -= p->pages;
2532 p->flags &= ~SWP_WRITEOK;
2533 spin_unlock(&p->lock);
2534 spin_unlock(&swap_lock);
2536 disable_swap_slots_cache_lock();
2538 set_current_oom_origin();
2539 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2540 clear_current_oom_origin();
2543 /* re-insert swap space back into swap_list */
2544 reinsert_swap_info(p);
2545 reenable_swap_slots_cache_unlock();
2549 reenable_swap_slots_cache_unlock();
2551 flush_work(&p->discard_work);
2553 destroy_swap_extents(p);
2554 if (p->flags & SWP_CONTINUED)
2555 free_swap_count_continuations(p);
2557 mutex_lock(&swapon_mutex);
2558 spin_lock(&swap_lock);
2559 spin_lock(&p->lock);
2562 /* wait for anyone still in scan_swap_map */
2563 p->highest_bit = 0; /* cuts scans short */
2564 while (p->flags >= SWP_SCANNING) {
2565 spin_unlock(&p->lock);
2566 spin_unlock(&swap_lock);
2567 schedule_timeout_uninterruptible(1);
2568 spin_lock(&swap_lock);
2569 spin_lock(&p->lock);
2572 swap_file = p->swap_file;
2573 old_block_size = p->old_block_size;
2574 p->swap_file = NULL;
2576 swap_map = p->swap_map;
2578 cluster_info = p->cluster_info;
2579 p->cluster_info = NULL;
2580 frontswap_map = frontswap_map_get(p);
2581 spin_unlock(&p->lock);
2582 spin_unlock(&swap_lock);
2583 frontswap_invalidate_area(p->type);
2584 frontswap_map_set(p, NULL);
2585 mutex_unlock(&swapon_mutex);
2586 free_percpu(p->percpu_cluster);
2587 p->percpu_cluster = NULL;
2589 kvfree(cluster_info);
2590 kvfree(frontswap_map);
2591 /* Destroy swap account information */
2592 swap_cgroup_swapoff(p->type);
2593 exit_swap_address_space(p->type);
2595 inode = mapping->host;
2596 if (S_ISBLK(inode->i_mode)) {
2597 struct block_device *bdev = I_BDEV(inode);
2598 set_blocksize(bdev, old_block_size);
2599 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2602 inode->i_flags &= ~S_SWAPFILE;
2603 inode_unlock(inode);
2605 filp_close(swap_file, NULL);
2608 * Clear the SWP_USED flag after all resources are freed so that swapon
2609 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2610 * not hold p->lock after we cleared its SWP_WRITEOK.
2612 spin_lock(&swap_lock);
2614 spin_unlock(&swap_lock);
2617 atomic_inc(&proc_poll_event);
2618 wake_up_interruptible(&proc_poll_wait);
2621 filp_close(victim, NULL);
2627 #ifdef CONFIG_PROC_FS
2628 static unsigned swaps_poll(struct file *file, poll_table *wait)
2630 struct seq_file *seq = file->private_data;
2632 poll_wait(file, &proc_poll_wait, wait);
2634 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2635 seq->poll_event = atomic_read(&proc_poll_event);
2636 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2639 return POLLIN | POLLRDNORM;
2643 static void *swap_start(struct seq_file *swap, loff_t *pos)
2645 struct swap_info_struct *si;
2649 mutex_lock(&swapon_mutex);
2652 return SEQ_START_TOKEN;
2654 for (type = 0; type < nr_swapfiles; type++) {
2655 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2656 si = swap_info[type];
2657 if (!(si->flags & SWP_USED) || !si->swap_map)
2666 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2668 struct swap_info_struct *si = v;
2671 if (v == SEQ_START_TOKEN)
2674 type = si->type + 1;
2676 for (; type < nr_swapfiles; type++) {
2677 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2678 si = swap_info[type];
2679 if (!(si->flags & SWP_USED) || !si->swap_map)
2688 static void swap_stop(struct seq_file *swap, void *v)
2690 mutex_unlock(&swapon_mutex);
2693 static int swap_show(struct seq_file *swap, void *v)
2695 struct swap_info_struct *si = v;
2699 if (si == SEQ_START_TOKEN) {
2700 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2704 file = si->swap_file;
2705 len = seq_file_path(swap, file, " \t\n\\");
2706 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2707 len < 40 ? 40 - len : 1, " ",
2708 S_ISBLK(file_inode(file)->i_mode) ?
2709 "partition" : "file\t",
2710 si->pages << (PAGE_SHIFT - 10),
2711 si->inuse_pages << (PAGE_SHIFT - 10),
2716 static const struct seq_operations swaps_op = {
2717 .start = swap_start,
2723 static int swaps_open(struct inode *inode, struct file *file)
2725 struct seq_file *seq;
2728 ret = seq_open(file, &swaps_op);
2732 seq = file->private_data;
2733 seq->poll_event = atomic_read(&proc_poll_event);
2737 static const struct file_operations proc_swaps_operations = {
2740 .llseek = seq_lseek,
2741 .release = seq_release,
2745 static int __init procswaps_init(void)
2747 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2750 __initcall(procswaps_init);
2751 #endif /* CONFIG_PROC_FS */
2753 #ifdef MAX_SWAPFILES_CHECK
2754 static int __init max_swapfiles_check(void)
2756 MAX_SWAPFILES_CHECK();
2759 late_initcall(max_swapfiles_check);
2762 static struct swap_info_struct *alloc_swap_info(void)
2764 struct swap_info_struct *p;
2767 p = kzalloc(sizeof(*p), GFP_KERNEL);
2769 return ERR_PTR(-ENOMEM);
2771 spin_lock(&swap_lock);
2772 for (type = 0; type < nr_swapfiles; type++) {
2773 if (!(swap_info[type]->flags & SWP_USED))
2776 if (type >= MAX_SWAPFILES) {
2777 spin_unlock(&swap_lock);
2779 return ERR_PTR(-EPERM);
2781 if (type >= nr_swapfiles) {
2783 swap_info[type] = p;
2785 * Write swap_info[type] before nr_swapfiles, in case a
2786 * racing procfs swap_start() or swap_next() is reading them.
2787 * (We never shrink nr_swapfiles, we never free this entry.)
2793 p = swap_info[type];
2795 * Do not memset this entry: a racing procfs swap_next()
2796 * would be relying on p->type to remain valid.
2799 INIT_LIST_HEAD(&p->first_swap_extent.list);
2800 plist_node_init(&p->list, 0);
2801 plist_node_init(&p->avail_list, 0);
2802 p->flags = SWP_USED;
2803 spin_unlock(&swap_lock);
2804 spin_lock_init(&p->lock);
2809 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2813 if (S_ISBLK(inode->i_mode)) {
2814 p->bdev = bdgrab(I_BDEV(inode));
2815 error = blkdev_get(p->bdev,
2816 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2821 p->old_block_size = block_size(p->bdev);
2822 error = set_blocksize(p->bdev, PAGE_SIZE);
2825 p->flags |= SWP_BLKDEV;
2826 } else if (S_ISREG(inode->i_mode)) {
2827 p->bdev = inode->i_sb->s_bdev;
2829 if (IS_SWAPFILE(inode))
2837 static unsigned long read_swap_header(struct swap_info_struct *p,
2838 union swap_header *swap_header,
2839 struct inode *inode)
2842 unsigned long maxpages;
2843 unsigned long swapfilepages;
2844 unsigned long last_page;
2846 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2847 pr_err("Unable to find swap-space signature\n");
2851 /* swap partition endianess hack... */
2852 if (swab32(swap_header->info.version) == 1) {
2853 swab32s(&swap_header->info.version);
2854 swab32s(&swap_header->info.last_page);
2855 swab32s(&swap_header->info.nr_badpages);
2856 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2858 for (i = 0; i < swap_header->info.nr_badpages; i++)
2859 swab32s(&swap_header->info.badpages[i]);
2861 /* Check the swap header's sub-version */
2862 if (swap_header->info.version != 1) {
2863 pr_warn("Unable to handle swap header version %d\n",
2864 swap_header->info.version);
2869 p->cluster_next = 1;
2873 * Find out how many pages are allowed for a single swap
2874 * device. There are two limiting factors: 1) the number
2875 * of bits for the swap offset in the swp_entry_t type, and
2876 * 2) the number of bits in the swap pte as defined by the
2877 * different architectures. In order to find the
2878 * largest possible bit mask, a swap entry with swap type 0
2879 * and swap offset ~0UL is created, encoded to a swap pte,
2880 * decoded to a swp_entry_t again, and finally the swap
2881 * offset is extracted. This will mask all the bits from
2882 * the initial ~0UL mask that can't be encoded in either
2883 * the swp_entry_t or the architecture definition of a
2886 maxpages = swp_offset(pte_to_swp_entry(
2887 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2888 last_page = swap_header->info.last_page;
2889 if (last_page > maxpages) {
2890 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2891 maxpages << (PAGE_SHIFT - 10),
2892 last_page << (PAGE_SHIFT - 10));
2894 if (maxpages > last_page) {
2895 maxpages = last_page + 1;
2896 /* p->max is an unsigned int: don't overflow it */
2897 if ((unsigned int)maxpages == 0)
2898 maxpages = UINT_MAX;
2900 p->highest_bit = maxpages - 1;
2904 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2905 if (swapfilepages && maxpages > swapfilepages) {
2906 pr_warn("Swap area shorter than signature indicates\n");
2909 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2911 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2917 #define SWAP_CLUSTER_INFO_COLS \
2918 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2919 #define SWAP_CLUSTER_SPACE_COLS \
2920 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2921 #define SWAP_CLUSTER_COLS \
2922 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2924 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2925 union swap_header *swap_header,
2926 unsigned char *swap_map,
2927 struct swap_cluster_info *cluster_info,
2928 unsigned long maxpages,
2932 unsigned int nr_good_pages;
2934 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2935 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2936 unsigned long i, idx;
2938 nr_good_pages = maxpages - 1; /* omit header page */
2940 cluster_list_init(&p->free_clusters);
2941 cluster_list_init(&p->discard_clusters);
2943 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2944 unsigned int page_nr = swap_header->info.badpages[i];
2945 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2947 if (page_nr < maxpages) {
2948 swap_map[page_nr] = SWAP_MAP_BAD;
2951 * Haven't marked the cluster free yet, no list
2952 * operation involved
2954 inc_cluster_info_page(p, cluster_info, page_nr);
2958 /* Haven't marked the cluster free yet, no list operation involved */
2959 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2960 inc_cluster_info_page(p, cluster_info, i);
2962 if (nr_good_pages) {
2963 swap_map[0] = SWAP_MAP_BAD;
2965 * Not mark the cluster free yet, no list
2966 * operation involved
2968 inc_cluster_info_page(p, cluster_info, 0);
2970 p->pages = nr_good_pages;
2971 nr_extents = setup_swap_extents(p, span);
2974 nr_good_pages = p->pages;
2976 if (!nr_good_pages) {
2977 pr_warn("Empty swap-file\n");
2986 * Reduce false cache line sharing between cluster_info and
2987 * sharing same address space.
2989 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2990 j = (k + col) % SWAP_CLUSTER_COLS;
2991 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2992 idx = i * SWAP_CLUSTER_COLS + j;
2993 if (idx >= nr_clusters)
2995 if (cluster_count(&cluster_info[idx]))
2997 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2998 cluster_list_add_tail(&p->free_clusters, cluster_info,
3006 * Helper to sys_swapon determining if a given swap
3007 * backing device queue supports DISCARD operations.
3009 static bool swap_discardable(struct swap_info_struct *si)
3011 struct request_queue *q = bdev_get_queue(si->bdev);
3013 if (!q || !blk_queue_discard(q))
3019 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3021 struct swap_info_struct *p;
3022 struct filename *name;
3023 struct file *swap_file = NULL;
3024 struct address_space *mapping;
3027 union swap_header *swap_header;
3030 unsigned long maxpages;
3031 unsigned char *swap_map = NULL;
3032 struct swap_cluster_info *cluster_info = NULL;
3033 unsigned long *frontswap_map = NULL;
3034 struct page *page = NULL;
3035 struct inode *inode = NULL;
3037 if (swap_flags & ~SWAP_FLAGS_VALID)
3040 if (!capable(CAP_SYS_ADMIN))
3043 p = alloc_swap_info();
3047 INIT_WORK(&p->discard_work, swap_discard_work);
3049 name = getname(specialfile);
3051 error = PTR_ERR(name);
3055 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3056 if (IS_ERR(swap_file)) {
3057 error = PTR_ERR(swap_file);
3062 p->swap_file = swap_file;
3063 mapping = swap_file->f_mapping;
3064 inode = mapping->host;
3066 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3067 error = claim_swapfile(p, inode);
3068 if (unlikely(error))
3072 * Read the swap header.
3074 if (!mapping->a_ops->readpage) {
3078 page = read_mapping_page(mapping, 0, swap_file);
3080 error = PTR_ERR(page);
3083 swap_header = kmap(page);
3085 maxpages = read_swap_header(p, swap_header, inode);
3086 if (unlikely(!maxpages)) {
3091 /* OK, set up the swap map and apply the bad block list */
3092 swap_map = vzalloc(maxpages);
3098 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3099 p->flags |= SWP_STABLE_WRITES;
3101 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3103 unsigned long ci, nr_cluster;
3105 p->flags |= SWP_SOLIDSTATE;
3107 * select a random position to start with to help wear leveling
3110 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3111 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3113 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3115 if (!cluster_info) {
3120 for (ci = 0; ci < nr_cluster; ci++)
3121 spin_lock_init(&((cluster_info + ci)->lock));
3123 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3124 if (!p->percpu_cluster) {
3128 for_each_possible_cpu(cpu) {
3129 struct percpu_cluster *cluster;
3130 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3131 cluster_set_null(&cluster->index);
3135 error = swap_cgroup_swapon(p->type, maxpages);
3139 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3140 cluster_info, maxpages, &span);
3141 if (unlikely(nr_extents < 0)) {
3145 /* frontswap enabled? set up bit-per-page map for frontswap */
3146 if (IS_ENABLED(CONFIG_FRONTSWAP))
3147 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3150 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3152 * When discard is enabled for swap with no particular
3153 * policy flagged, we set all swap discard flags here in
3154 * order to sustain backward compatibility with older
3155 * swapon(8) releases.
3157 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3161 * By flagging sys_swapon, a sysadmin can tell us to
3162 * either do single-time area discards only, or to just
3163 * perform discards for released swap page-clusters.
3164 * Now it's time to adjust the p->flags accordingly.
3166 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3167 p->flags &= ~SWP_PAGE_DISCARD;
3168 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3169 p->flags &= ~SWP_AREA_DISCARD;
3171 /* issue a swapon-time discard if it's still required */
3172 if (p->flags & SWP_AREA_DISCARD) {
3173 int err = discard_swap(p);
3175 pr_err("swapon: discard_swap(%p): %d\n",
3180 error = init_swap_address_space(p->type, maxpages);
3184 mutex_lock(&swapon_mutex);
3186 if (swap_flags & SWAP_FLAG_PREFER)
3188 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3189 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3191 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3192 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3193 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3194 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3195 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3196 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3197 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3198 (frontswap_map) ? "FS" : "");
3200 mutex_unlock(&swapon_mutex);
3201 atomic_inc(&proc_poll_event);
3202 wake_up_interruptible(&proc_poll_wait);
3204 if (S_ISREG(inode->i_mode))
3205 inode->i_flags |= S_SWAPFILE;
3209 free_percpu(p->percpu_cluster);
3210 p->percpu_cluster = NULL;
3211 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3212 set_blocksize(p->bdev, p->old_block_size);
3213 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3215 destroy_swap_extents(p);
3216 swap_cgroup_swapoff(p->type);
3217 spin_lock(&swap_lock);
3218 p->swap_file = NULL;
3220 spin_unlock(&swap_lock);
3222 vfree(cluster_info);
3224 if (inode && S_ISREG(inode->i_mode)) {
3225 inode_unlock(inode);
3228 filp_close(swap_file, NULL);
3231 if (page && !IS_ERR(page)) {
3237 if (inode && S_ISREG(inode->i_mode))
3238 inode_unlock(inode);
3240 enable_swap_slots_cache();
3244 void si_swapinfo(struct sysinfo *val)
3247 unsigned long nr_to_be_unused = 0;
3249 spin_lock(&swap_lock);
3250 for (type = 0; type < nr_swapfiles; type++) {
3251 struct swap_info_struct *si = swap_info[type];
3253 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3254 nr_to_be_unused += si->inuse_pages;
3256 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3257 val->totalswap = total_swap_pages + nr_to_be_unused;
3258 spin_unlock(&swap_lock);
3262 * Verify that a swap entry is valid and increment its swap map count.
3264 * Returns error code in following case.
3266 * - swp_entry is invalid -> EINVAL
3267 * - swp_entry is migration entry -> EINVAL
3268 * - swap-cache reference is requested but there is already one. -> EEXIST
3269 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3270 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3272 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3274 struct swap_info_struct *p;
3275 struct swap_cluster_info *ci;
3276 unsigned long offset, type;
3277 unsigned char count;
3278 unsigned char has_cache;
3281 if (non_swap_entry(entry))
3284 type = swp_type(entry);
3285 if (type >= nr_swapfiles)
3287 p = swap_info[type];
3288 offset = swp_offset(entry);
3289 if (unlikely(offset >= p->max))
3292 ci = lock_cluster_or_swap_info(p, offset);
3294 count = p->swap_map[offset];
3297 * swapin_readahead() doesn't check if a swap entry is valid, so the
3298 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3300 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3305 has_cache = count & SWAP_HAS_CACHE;
3306 count &= ~SWAP_HAS_CACHE;
3309 if (usage == SWAP_HAS_CACHE) {
3311 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3312 if (!has_cache && count)
3313 has_cache = SWAP_HAS_CACHE;
3314 else if (has_cache) /* someone else added cache */
3316 else /* no users remaining */
3319 } else if (count || has_cache) {
3321 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3323 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3325 else if (swap_count_continued(p, offset, count))
3326 count = COUNT_CONTINUED;
3330 err = -ENOENT; /* unused swap entry */
3332 p->swap_map[offset] = count | has_cache;
3335 unlock_cluster_or_swap_info(p, ci);
3340 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3345 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3346 * (in which case its reference count is never incremented).
3348 void swap_shmem_alloc(swp_entry_t entry)
3350 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3354 * Increase reference count of swap entry by 1.
3355 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3356 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3357 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3358 * might occur if a page table entry has got corrupted.
3360 int swap_duplicate(swp_entry_t entry)
3364 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3365 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3370 * @entry: swap entry for which we allocate swap cache.
3372 * Called when allocating swap cache for existing swap entry,
3373 * This can return error codes. Returns 0 at success.
3374 * -EBUSY means there is a swap cache.
3375 * Note: return code is different from swap_duplicate().
3377 int swapcache_prepare(swp_entry_t entry)
3379 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3382 struct swap_info_struct *page_swap_info(struct page *page)
3384 swp_entry_t swap = { .val = page_private(page) };
3385 return swap_info[swp_type(swap)];
3389 * out-of-line __page_file_ methods to avoid include hell.
3391 struct address_space *__page_file_mapping(struct page *page)
3393 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3394 return page_swap_info(page)->swap_file->f_mapping;
3396 EXPORT_SYMBOL_GPL(__page_file_mapping);
3398 pgoff_t __page_file_index(struct page *page)
3400 swp_entry_t swap = { .val = page_private(page) };
3401 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3402 return swp_offset(swap);
3404 EXPORT_SYMBOL_GPL(__page_file_index);
3407 * add_swap_count_continuation - called when a swap count is duplicated
3408 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3409 * page of the original vmalloc'ed swap_map, to hold the continuation count
3410 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3411 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3413 * These continuation pages are seldom referenced: the common paths all work
3414 * on the original swap_map, only referring to a continuation page when the
3415 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3417 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3418 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3419 * can be called after dropping locks.
3421 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3423 struct swap_info_struct *si;
3424 struct swap_cluster_info *ci;
3427 struct page *list_page;
3429 unsigned char count;
3432 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3433 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3435 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3437 si = swap_info_get(entry);
3440 * An acceptable race has occurred since the failing
3441 * __swap_duplicate(): the swap entry has been freed,
3442 * perhaps even the whole swap_map cleared for swapoff.
3447 offset = swp_offset(entry);
3449 ci = lock_cluster(si, offset);
3451 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3453 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3455 * The higher the swap count, the more likely it is that tasks
3456 * will race to add swap count continuation: we need to avoid
3457 * over-provisioning.
3464 spin_unlock(&si->lock);
3469 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3470 * no architecture is using highmem pages for kernel page tables: so it
3471 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3473 head = vmalloc_to_page(si->swap_map + offset);
3474 offset &= ~PAGE_MASK;
3477 * Page allocation does not initialize the page's lru field,
3478 * but it does always reset its private field.
3480 if (!page_private(head)) {
3481 BUG_ON(count & COUNT_CONTINUED);
3482 INIT_LIST_HEAD(&head->lru);
3483 set_page_private(head, SWP_CONTINUED);
3484 si->flags |= SWP_CONTINUED;
3487 list_for_each_entry(list_page, &head->lru, lru) {
3491 * If the previous map said no continuation, but we've found
3492 * a continuation page, free our allocation and use this one.
3494 if (!(count & COUNT_CONTINUED))
3497 map = kmap_atomic(list_page) + offset;
3502 * If this continuation count now has some space in it,
3503 * free our allocation and use this one.
3505 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3509 list_add_tail(&page->lru, &head->lru);
3510 page = NULL; /* now it's attached, don't free it */
3513 spin_unlock(&si->lock);
3521 * swap_count_continued - when the original swap_map count is incremented
3522 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3523 * into, carry if so, or else fail until a new continuation page is allocated;
3524 * when the original swap_map count is decremented from 0 with continuation,
3525 * borrow from the continuation and report whether it still holds more.
3526 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3529 static bool swap_count_continued(struct swap_info_struct *si,
3530 pgoff_t offset, unsigned char count)
3536 head = vmalloc_to_page(si->swap_map + offset);
3537 if (page_private(head) != SWP_CONTINUED) {
3538 BUG_ON(count & COUNT_CONTINUED);
3539 return false; /* need to add count continuation */
3542 offset &= ~PAGE_MASK;
3543 page = list_entry(head->lru.next, struct page, lru);
3544 map = kmap_atomic(page) + offset;
3546 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3547 goto init_map; /* jump over SWAP_CONT_MAX checks */
3549 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3551 * Think of how you add 1 to 999
3553 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3555 page = list_entry(page->lru.next, struct page, lru);
3556 BUG_ON(page == head);
3557 map = kmap_atomic(page) + offset;
3559 if (*map == SWAP_CONT_MAX) {
3561 page = list_entry(page->lru.next, struct page, lru);
3563 return false; /* add count continuation */
3564 map = kmap_atomic(page) + offset;
3565 init_map: *map = 0; /* we didn't zero the page */
3569 page = list_entry(page->lru.prev, struct page, lru);
3570 while (page != head) {
3571 map = kmap_atomic(page) + offset;
3572 *map = COUNT_CONTINUED;
3574 page = list_entry(page->lru.prev, struct page, lru);
3576 return true; /* incremented */
3578 } else { /* decrementing */
3580 * Think of how you subtract 1 from 1000
3582 BUG_ON(count != COUNT_CONTINUED);
3583 while (*map == COUNT_CONTINUED) {
3585 page = list_entry(page->lru.next, struct page, lru);
3586 BUG_ON(page == head);
3587 map = kmap_atomic(page) + offset;
3594 page = list_entry(page->lru.prev, struct page, lru);
3595 while (page != head) {
3596 map = kmap_atomic(page) + offset;
3597 *map = SWAP_CONT_MAX | count;
3598 count = COUNT_CONTINUED;
3600 page = list_entry(page->lru.prev, struct page, lru);
3602 return count == COUNT_CONTINUED;
3607 * free_swap_count_continuations - swapoff free all the continuation pages
3608 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3610 static void free_swap_count_continuations(struct swap_info_struct *si)
3614 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3616 head = vmalloc_to_page(si->swap_map + offset);
3617 if (page_private(head)) {
3618 struct page *page, *next;
3620 list_for_each_entry_safe(page, next, &head->lru, lru) {
3621 list_del(&page->lru);