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 n_ret = swap_alloc_cluster(si, swp_entries);
954 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
955 n_goal, swp_entries);
956 spin_unlock(&si->lock);
957 if (n_ret || cluster)
959 pr_debug("scan_swap_map of si %d failed to find offset\n",
962 spin_lock(&swap_avail_lock);
965 * if we got here, it's likely that si was almost full before,
966 * and since scan_swap_map() can drop the si->lock, multiple
967 * callers probably all tried to get a page from the same si
968 * and it filled up before we could get one; or, the si filled
969 * up between us dropping swap_avail_lock and taking si->lock.
970 * Since we dropped the swap_avail_lock, the swap_avail_head
971 * list may have been modified; so if next is still in the
972 * swap_avail_head list then try it, otherwise start over
973 * if we have not gotten any slots.
975 if (plist_node_empty(&next->avail_list))
979 spin_unlock(&swap_avail_lock);
983 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
989 /* The only caller of this function is now suspend routine */
990 swp_entry_t get_swap_page_of_type(int type)
992 struct swap_info_struct *si;
995 si = swap_info[type];
996 spin_lock(&si->lock);
997 if (si && (si->flags & SWP_WRITEOK)) {
998 atomic_long_dec(&nr_swap_pages);
999 /* This is called for allocating swap entry, not cache */
1000 offset = scan_swap_map(si, 1);
1002 spin_unlock(&si->lock);
1003 return swp_entry(type, offset);
1005 atomic_long_inc(&nr_swap_pages);
1007 spin_unlock(&si->lock);
1008 return (swp_entry_t) {0};
1011 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1013 struct swap_info_struct *p;
1014 unsigned long offset, type;
1018 type = swp_type(entry);
1019 if (type >= nr_swapfiles)
1021 p = swap_info[type];
1022 if (!(p->flags & SWP_USED))
1024 offset = swp_offset(entry);
1025 if (offset >= p->max)
1030 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1033 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1036 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1041 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1043 struct swap_info_struct *p;
1045 p = __swap_info_get(entry);
1048 if (!p->swap_map[swp_offset(entry)])
1053 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1059 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1061 struct swap_info_struct *p;
1063 p = _swap_info_get(entry);
1065 spin_lock(&p->lock);
1069 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1070 struct swap_info_struct *q)
1072 struct swap_info_struct *p;
1074 p = _swap_info_get(entry);
1078 spin_unlock(&q->lock);
1080 spin_lock(&p->lock);
1085 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1086 swp_entry_t entry, unsigned char usage)
1088 struct swap_cluster_info *ci;
1089 unsigned long offset = swp_offset(entry);
1090 unsigned char count;
1091 unsigned char has_cache;
1093 ci = lock_cluster_or_swap_info(p, offset);
1095 count = p->swap_map[offset];
1097 has_cache = count & SWAP_HAS_CACHE;
1098 count &= ~SWAP_HAS_CACHE;
1100 if (usage == SWAP_HAS_CACHE) {
1101 VM_BUG_ON(!has_cache);
1103 } else if (count == SWAP_MAP_SHMEM) {
1105 * Or we could insist on shmem.c using a special
1106 * swap_shmem_free() and free_shmem_swap_and_cache()...
1109 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1110 if (count == COUNT_CONTINUED) {
1111 if (swap_count_continued(p, offset, count))
1112 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1114 count = SWAP_MAP_MAX;
1119 usage = count | has_cache;
1120 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1122 unlock_cluster_or_swap_info(p, ci);
1127 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1129 struct swap_cluster_info *ci;
1130 unsigned long offset = swp_offset(entry);
1131 unsigned char count;
1133 ci = lock_cluster(p, offset);
1134 count = p->swap_map[offset];
1135 VM_BUG_ON(count != SWAP_HAS_CACHE);
1136 p->swap_map[offset] = 0;
1137 dec_cluster_info_page(p, p->cluster_info, offset);
1140 mem_cgroup_uncharge_swap(entry, 1);
1141 swap_range_free(p, offset, 1);
1145 * Caller has made sure that the swap device corresponding to entry
1146 * is still around or has not been recycled.
1148 void swap_free(swp_entry_t entry)
1150 struct swap_info_struct *p;
1152 p = _swap_info_get(entry);
1154 if (!__swap_entry_free(p, entry, 1))
1155 free_swap_slot(entry);
1160 * Called after dropping swapcache to decrease refcnt to swap entries.
1162 static void swapcache_free(swp_entry_t entry)
1164 struct swap_info_struct *p;
1166 p = _swap_info_get(entry);
1168 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1169 free_swap_slot(entry);
1173 #ifdef CONFIG_THP_SWAP
1174 static void swapcache_free_cluster(swp_entry_t entry)
1176 unsigned long offset = swp_offset(entry);
1177 unsigned long idx = offset / SWAPFILE_CLUSTER;
1178 struct swap_cluster_info *ci;
1179 struct swap_info_struct *si;
1181 unsigned int i, free_entries = 0;
1184 si = _swap_info_get(entry);
1188 ci = lock_cluster(si, offset);
1189 VM_BUG_ON(!cluster_is_huge(ci));
1190 map = si->swap_map + offset;
1191 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1193 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1194 if (val == SWAP_HAS_CACHE)
1197 if (!free_entries) {
1198 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1199 map[i] &= ~SWAP_HAS_CACHE;
1201 cluster_clear_huge(ci);
1203 if (free_entries == SWAPFILE_CLUSTER) {
1204 spin_lock(&si->lock);
1205 ci = lock_cluster(si, offset);
1206 memset(map, 0, SWAPFILE_CLUSTER);
1208 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1209 swap_free_cluster(si, idx);
1210 spin_unlock(&si->lock);
1211 } else if (free_entries) {
1212 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1213 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1214 free_swap_slot(entry);
1219 static inline void swapcache_free_cluster(swp_entry_t entry)
1222 #endif /* CONFIG_THP_SWAP */
1224 void put_swap_page(struct page *page, swp_entry_t entry)
1226 if (!PageTransHuge(page))
1227 swapcache_free(entry);
1229 swapcache_free_cluster(entry);
1232 static int swp_entry_cmp(const void *ent1, const void *ent2)
1234 const swp_entry_t *e1 = ent1, *e2 = ent2;
1236 return (int)swp_type(*e1) - (int)swp_type(*e2);
1239 void swapcache_free_entries(swp_entry_t *entries, int n)
1241 struct swap_info_struct *p, *prev;
1251 * Sort swap entries by swap device, so each lock is only taken once.
1252 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1253 * so low that it isn't necessary to optimize further.
1255 if (nr_swapfiles > 1)
1256 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1257 for (i = 0; i < n; ++i) {
1258 p = swap_info_get_cont(entries[i], prev);
1260 swap_entry_free(p, entries[i]);
1264 spin_unlock(&p->lock);
1268 * How many references to page are currently swapped out?
1269 * This does not give an exact answer when swap count is continued,
1270 * but does include the high COUNT_CONTINUED flag to allow for that.
1272 int page_swapcount(struct page *page)
1275 struct swap_info_struct *p;
1276 struct swap_cluster_info *ci;
1278 unsigned long offset;
1280 entry.val = page_private(page);
1281 p = _swap_info_get(entry);
1283 offset = swp_offset(entry);
1284 ci = lock_cluster_or_swap_info(p, offset);
1285 count = swap_count(p->swap_map[offset]);
1286 unlock_cluster_or_swap_info(p, ci);
1291 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1294 pgoff_t offset = swp_offset(entry);
1295 struct swap_cluster_info *ci;
1297 ci = lock_cluster_or_swap_info(si, offset);
1298 count = swap_count(si->swap_map[offset]);
1299 unlock_cluster_or_swap_info(si, ci);
1304 * How many references to @entry are currently swapped out?
1305 * This does not give an exact answer when swap count is continued,
1306 * but does include the high COUNT_CONTINUED flag to allow for that.
1308 int __swp_swapcount(swp_entry_t entry)
1311 struct swap_info_struct *si;
1313 si = __swap_info_get(entry);
1315 count = swap_swapcount(si, entry);
1320 * How many references to @entry are currently swapped out?
1321 * This considers COUNT_CONTINUED so it returns exact answer.
1323 int swp_swapcount(swp_entry_t entry)
1325 int count, tmp_count, n;
1326 struct swap_info_struct *p;
1327 struct swap_cluster_info *ci;
1332 p = _swap_info_get(entry);
1336 offset = swp_offset(entry);
1338 ci = lock_cluster_or_swap_info(p, offset);
1340 count = swap_count(p->swap_map[offset]);
1341 if (!(count & COUNT_CONTINUED))
1344 count &= ~COUNT_CONTINUED;
1345 n = SWAP_MAP_MAX + 1;
1347 page = vmalloc_to_page(p->swap_map + offset);
1348 offset &= ~PAGE_MASK;
1349 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1352 page = list_next_entry(page, lru);
1353 map = kmap_atomic(page);
1354 tmp_count = map[offset];
1357 count += (tmp_count & ~COUNT_CONTINUED) * n;
1358 n *= (SWAP_CONT_MAX + 1);
1359 } while (tmp_count & COUNT_CONTINUED);
1361 unlock_cluster_or_swap_info(p, ci);
1365 #ifdef CONFIG_THP_SWAP
1366 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1369 struct swap_cluster_info *ci;
1370 unsigned char *map = si->swap_map;
1371 unsigned long roffset = swp_offset(entry);
1372 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1376 ci = lock_cluster_or_swap_info(si, offset);
1377 if (!ci || !cluster_is_huge(ci)) {
1378 if (map[roffset] != SWAP_HAS_CACHE)
1382 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1383 if (map[offset + i] != SWAP_HAS_CACHE) {
1389 unlock_cluster_or_swap_info(si, ci);
1393 static bool page_swapped(struct page *page)
1396 struct swap_info_struct *si;
1398 if (likely(!PageTransCompound(page)))
1399 return page_swapcount(page) != 0;
1401 page = compound_head(page);
1402 entry.val = page_private(page);
1403 si = _swap_info_get(entry);
1405 return swap_page_trans_huge_swapped(si, entry);
1409 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1410 int *total_swapcount)
1412 int i, map_swapcount, _total_mapcount, _total_swapcount;
1413 unsigned long offset = 0;
1414 struct swap_info_struct *si;
1415 struct swap_cluster_info *ci = NULL;
1416 unsigned char *map = NULL;
1417 int mapcount, swapcount = 0;
1419 /* hugetlbfs shouldn't call it */
1420 VM_BUG_ON_PAGE(PageHuge(page), page);
1422 if (likely(!PageTransCompound(page))) {
1423 mapcount = atomic_read(&page->_mapcount) + 1;
1425 *total_mapcount = mapcount;
1426 if (PageSwapCache(page))
1427 swapcount = page_swapcount(page);
1428 if (total_swapcount)
1429 *total_swapcount = swapcount;
1430 return mapcount + swapcount;
1433 page = compound_head(page);
1435 _total_mapcount = _total_swapcount = map_swapcount = 0;
1436 if (PageSwapCache(page)) {
1439 entry.val = page_private(page);
1440 si = _swap_info_get(entry);
1443 offset = swp_offset(entry);
1447 ci = lock_cluster(si, offset);
1448 for (i = 0; i < HPAGE_PMD_NR; i++) {
1449 mapcount = atomic_read(&page[i]._mapcount) + 1;
1450 _total_mapcount += mapcount;
1452 swapcount = swap_count(map[offset + i]);
1453 _total_swapcount += swapcount;
1455 map_swapcount = max(map_swapcount, mapcount + swapcount);
1458 if (PageDoubleMap(page)) {
1460 _total_mapcount -= HPAGE_PMD_NR;
1462 mapcount = compound_mapcount(page);
1463 map_swapcount += mapcount;
1464 _total_mapcount += mapcount;
1466 *total_mapcount = _total_mapcount;
1467 if (total_swapcount)
1468 *total_swapcount = _total_swapcount;
1470 return map_swapcount;
1473 #define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1474 #define page_swapped(page) (page_swapcount(page) != 0)
1476 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1477 int *total_swapcount)
1479 int mapcount, swapcount = 0;
1481 /* hugetlbfs shouldn't call it */
1482 VM_BUG_ON_PAGE(PageHuge(page), page);
1484 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1485 if (PageSwapCache(page))
1486 swapcount = page_swapcount(page);
1487 if (total_swapcount)
1488 *total_swapcount = swapcount;
1489 return mapcount + swapcount;
1494 * We can write to an anon page without COW if there are no other references
1495 * to it. And as a side-effect, free up its swap: because the old content
1496 * on disk will never be read, and seeking back there to write new content
1497 * later would only waste time away from clustering.
1499 * NOTE: total_map_swapcount should not be relied upon by the caller if
1500 * reuse_swap_page() returns false, but it may be always overwritten
1501 * (see the other implementation for CONFIG_SWAP=n).
1503 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1505 int count, total_mapcount, total_swapcount;
1507 VM_BUG_ON_PAGE(!PageLocked(page), page);
1508 if (unlikely(PageKsm(page)))
1510 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1512 if (total_map_swapcount)
1513 *total_map_swapcount = total_mapcount + total_swapcount;
1514 if (count == 1 && PageSwapCache(page) &&
1515 (likely(!PageTransCompound(page)) ||
1516 /* The remaining swap count will be freed soon */
1517 total_swapcount == page_swapcount(page))) {
1518 if (!PageWriteback(page)) {
1519 page = compound_head(page);
1520 delete_from_swap_cache(page);
1524 struct swap_info_struct *p;
1526 entry.val = page_private(page);
1527 p = swap_info_get(entry);
1528 if (p->flags & SWP_STABLE_WRITES) {
1529 spin_unlock(&p->lock);
1532 spin_unlock(&p->lock);
1540 * If swap is getting full, or if there are no more mappings of this page,
1541 * then try_to_free_swap is called to free its swap space.
1543 int try_to_free_swap(struct page *page)
1545 VM_BUG_ON_PAGE(!PageLocked(page), page);
1547 if (!PageSwapCache(page))
1549 if (PageWriteback(page))
1551 if (page_swapped(page))
1555 * Once hibernation has begun to create its image of memory,
1556 * there's a danger that one of the calls to try_to_free_swap()
1557 * - most probably a call from __try_to_reclaim_swap() while
1558 * hibernation is allocating its own swap pages for the image,
1559 * but conceivably even a call from memory reclaim - will free
1560 * the swap from a page which has already been recorded in the
1561 * image as a clean swapcache page, and then reuse its swap for
1562 * another page of the image. On waking from hibernation, the
1563 * original page might be freed under memory pressure, then
1564 * later read back in from swap, now with the wrong data.
1566 * Hibernation suspends storage while it is writing the image
1567 * to disk so check that here.
1569 if (pm_suspended_storage())
1572 page = compound_head(page);
1573 delete_from_swap_cache(page);
1579 * Free the swap entry like above, but also try to
1580 * free the page cache entry if it is the last user.
1582 int free_swap_and_cache(swp_entry_t entry)
1584 struct swap_info_struct *p;
1585 struct page *page = NULL;
1586 unsigned char count;
1588 if (non_swap_entry(entry))
1591 p = _swap_info_get(entry);
1593 count = __swap_entry_free(p, entry, 1);
1594 if (count == SWAP_HAS_CACHE &&
1595 !swap_page_trans_huge_swapped(p, entry)) {
1596 page = find_get_page(swap_address_space(entry),
1598 if (page && !trylock_page(page)) {
1603 free_swap_slot(entry);
1607 * Not mapped elsewhere, or swap space full? Free it!
1608 * Also recheck PageSwapCache now page is locked (above).
1610 if (PageSwapCache(page) && !PageWriteback(page) &&
1611 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1612 !swap_page_trans_huge_swapped(p, entry)) {
1613 page = compound_head(page);
1614 delete_from_swap_cache(page);
1623 #ifdef CONFIG_HIBERNATION
1625 * Find the swap type that corresponds to given device (if any).
1627 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1628 * from 0, in which the swap header is expected to be located.
1630 * This is needed for the suspend to disk (aka swsusp).
1632 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1634 struct block_device *bdev = NULL;
1638 bdev = bdget(device);
1640 spin_lock(&swap_lock);
1641 for (type = 0; type < nr_swapfiles; type++) {
1642 struct swap_info_struct *sis = swap_info[type];
1644 if (!(sis->flags & SWP_WRITEOK))
1649 *bdev_p = bdgrab(sis->bdev);
1651 spin_unlock(&swap_lock);
1654 if (bdev == sis->bdev) {
1655 struct swap_extent *se = &sis->first_swap_extent;
1657 if (se->start_block == offset) {
1659 *bdev_p = bdgrab(sis->bdev);
1661 spin_unlock(&swap_lock);
1667 spin_unlock(&swap_lock);
1675 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1676 * corresponding to given index in swap_info (swap type).
1678 sector_t swapdev_block(int type, pgoff_t offset)
1680 struct block_device *bdev;
1682 if ((unsigned int)type >= nr_swapfiles)
1684 if (!(swap_info[type]->flags & SWP_WRITEOK))
1686 return map_swap_entry(swp_entry(type, offset), &bdev);
1690 * Return either the total number of swap pages of given type, or the number
1691 * of free pages of that type (depending on @free)
1693 * This is needed for software suspend
1695 unsigned int count_swap_pages(int type, int free)
1699 spin_lock(&swap_lock);
1700 if ((unsigned int)type < nr_swapfiles) {
1701 struct swap_info_struct *sis = swap_info[type];
1703 spin_lock(&sis->lock);
1704 if (sis->flags & SWP_WRITEOK) {
1707 n -= sis->inuse_pages;
1709 spin_unlock(&sis->lock);
1711 spin_unlock(&swap_lock);
1714 #endif /* CONFIG_HIBERNATION */
1716 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1718 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1722 * No need to decide whether this PTE shares the swap entry with others,
1723 * just let do_wp_page work it out if a write is requested later - to
1724 * force COW, vm_page_prot omits write permission from any private vma.
1726 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1727 unsigned long addr, swp_entry_t entry, struct page *page)
1729 struct page *swapcache;
1730 struct mem_cgroup *memcg;
1736 page = ksm_might_need_to_copy(page, vma, addr);
1737 if (unlikely(!page))
1740 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1746 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1747 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1748 mem_cgroup_cancel_charge(page, memcg, false);
1753 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1754 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1756 set_pte_at(vma->vm_mm, addr, pte,
1757 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1758 if (page == swapcache) {
1759 page_add_anon_rmap(page, vma, addr, false);
1760 mem_cgroup_commit_charge(page, memcg, true, false);
1761 } else { /* ksm created a completely new copy */
1762 page_add_new_anon_rmap(page, vma, addr, false);
1763 mem_cgroup_commit_charge(page, memcg, false, false);
1764 lru_cache_add_active_or_unevictable(page, vma);
1768 * Move the page to the active list so it is not
1769 * immediately swapped out again after swapon.
1771 activate_page(page);
1773 pte_unmap_unlock(pte, ptl);
1775 if (page != swapcache) {
1782 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1783 unsigned long addr, unsigned long end,
1784 swp_entry_t entry, struct page *page)
1786 pte_t swp_pte = swp_entry_to_pte(entry);
1791 * We don't actually need pte lock while scanning for swp_pte: since
1792 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1793 * page table while we're scanning; though it could get zapped, and on
1794 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1795 * of unmatched parts which look like swp_pte, so unuse_pte must
1796 * recheck under pte lock. Scanning without pte lock lets it be
1797 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1799 pte = pte_offset_map(pmd, addr);
1802 * swapoff spends a _lot_ of time in this loop!
1803 * Test inline before going to call unuse_pte.
1805 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1807 ret = unuse_pte(vma, pmd, addr, entry, page);
1810 pte = pte_offset_map(pmd, addr);
1812 } while (pte++, addr += PAGE_SIZE, addr != end);
1818 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1819 unsigned long addr, unsigned long end,
1820 swp_entry_t entry, struct page *page)
1826 pmd = pmd_offset(pud, addr);
1829 next = pmd_addr_end(addr, end);
1830 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1832 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1835 } while (pmd++, addr = next, addr != end);
1839 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1840 unsigned long addr, unsigned long end,
1841 swp_entry_t entry, struct page *page)
1847 pud = pud_offset(p4d, addr);
1849 next = pud_addr_end(addr, end);
1850 if (pud_none_or_clear_bad(pud))
1852 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1855 } while (pud++, addr = next, addr != end);
1859 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1860 unsigned long addr, unsigned long end,
1861 swp_entry_t entry, struct page *page)
1867 p4d = p4d_offset(pgd, addr);
1869 next = p4d_addr_end(addr, end);
1870 if (p4d_none_or_clear_bad(p4d))
1872 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1875 } while (p4d++, addr = next, addr != end);
1879 static int unuse_vma(struct vm_area_struct *vma,
1880 swp_entry_t entry, struct page *page)
1883 unsigned long addr, end, next;
1886 if (page_anon_vma(page)) {
1887 addr = page_address_in_vma(page, vma);
1888 if (addr == -EFAULT)
1891 end = addr + PAGE_SIZE;
1893 addr = vma->vm_start;
1897 pgd = pgd_offset(vma->vm_mm, addr);
1899 next = pgd_addr_end(addr, end);
1900 if (pgd_none_or_clear_bad(pgd))
1902 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1905 } while (pgd++, addr = next, addr != end);
1909 static int unuse_mm(struct mm_struct *mm,
1910 swp_entry_t entry, struct page *page)
1912 struct vm_area_struct *vma;
1915 if (!down_read_trylock(&mm->mmap_sem)) {
1917 * Activate page so shrink_inactive_list is unlikely to unmap
1918 * its ptes while lock is dropped, so swapoff can make progress.
1920 activate_page(page);
1922 down_read(&mm->mmap_sem);
1925 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1926 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1930 up_read(&mm->mmap_sem);
1931 return (ret < 0)? ret: 0;
1935 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1936 * from current position to next entry still in use.
1937 * Recycle to start on reaching the end, returning 0 when empty.
1939 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1940 unsigned int prev, bool frontswap)
1942 unsigned int max = si->max;
1943 unsigned int i = prev;
1944 unsigned char count;
1947 * No need for swap_lock here: we're just looking
1948 * for whether an entry is in use, not modifying it; false
1949 * hits are okay, and sys_swapoff() has already prevented new
1950 * allocations from this area (while holding swap_lock).
1959 * No entries in use at top of swap_map,
1960 * loop back to start and recheck there.
1966 count = READ_ONCE(si->swap_map[i]);
1967 if (count && swap_count(count) != SWAP_MAP_BAD)
1968 if (!frontswap || frontswap_test(si, i))
1970 if ((i % LATENCY_LIMIT) == 0)
1977 * We completely avoid races by reading each swap page in advance,
1978 * and then search for the process using it. All the necessary
1979 * page table adjustments can then be made atomically.
1981 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1982 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1984 int try_to_unuse(unsigned int type, bool frontswap,
1985 unsigned long pages_to_unuse)
1987 struct swap_info_struct *si = swap_info[type];
1988 struct mm_struct *start_mm;
1989 volatile unsigned char *swap_map; /* swap_map is accessed without
1990 * locking. Mark it as volatile
1991 * to prevent compiler doing
1994 unsigned char swcount;
2001 * When searching mms for an entry, a good strategy is to
2002 * start at the first mm we freed the previous entry from
2003 * (though actually we don't notice whether we or coincidence
2004 * freed the entry). Initialize this start_mm with a hold.
2006 * A simpler strategy would be to start at the last mm we
2007 * freed the previous entry from; but that would take less
2008 * advantage of mmlist ordering, which clusters forked mms
2009 * together, child after parent. If we race with dup_mmap(), we
2010 * prefer to resolve parent before child, lest we miss entries
2011 * duplicated after we scanned child: using last mm would invert
2014 start_mm = &init_mm;
2018 * Keep on scanning until all entries have gone. Usually,
2019 * one pass through swap_map is enough, but not necessarily:
2020 * there are races when an instance of an entry might be missed.
2022 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2023 if (signal_pending(current)) {
2029 * Get a page for the entry, using the existing swap
2030 * cache page if there is one. Otherwise, get a clean
2031 * page and read the swap into it.
2033 swap_map = &si->swap_map[i];
2034 entry = swp_entry(type, i);
2035 page = read_swap_cache_async(entry,
2036 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2039 * Either swap_duplicate() failed because entry
2040 * has been freed independently, and will not be
2041 * reused since sys_swapoff() already disabled
2042 * allocation from here, or alloc_page() failed.
2044 swcount = *swap_map;
2046 * We don't hold lock here, so the swap entry could be
2047 * SWAP_MAP_BAD (when the cluster is discarding).
2048 * Instead of fail out, We can just skip the swap
2049 * entry because swapoff will wait for discarding
2052 if (!swcount || swcount == SWAP_MAP_BAD)
2059 * Don't hold on to start_mm if it looks like exiting.
2061 if (atomic_read(&start_mm->mm_users) == 1) {
2063 start_mm = &init_mm;
2068 * Wait for and lock page. When do_swap_page races with
2069 * try_to_unuse, do_swap_page can handle the fault much
2070 * faster than try_to_unuse can locate the entry. This
2071 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2072 * defer to do_swap_page in such a case - in some tests,
2073 * do_swap_page and try_to_unuse repeatedly compete.
2075 wait_on_page_locked(page);
2076 wait_on_page_writeback(page);
2078 wait_on_page_writeback(page);
2081 * Remove all references to entry.
2083 swcount = *swap_map;
2084 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2085 retval = shmem_unuse(entry, page);
2086 /* page has already been unlocked and released */
2091 if (swap_count(swcount) && start_mm != &init_mm)
2092 retval = unuse_mm(start_mm, entry, page);
2094 if (swap_count(*swap_map)) {
2095 int set_start_mm = (*swap_map >= swcount);
2096 struct list_head *p = &start_mm->mmlist;
2097 struct mm_struct *new_start_mm = start_mm;
2098 struct mm_struct *prev_mm = start_mm;
2099 struct mm_struct *mm;
2101 mmget(new_start_mm);
2103 spin_lock(&mmlist_lock);
2104 while (swap_count(*swap_map) && !retval &&
2105 (p = p->next) != &start_mm->mmlist) {
2106 mm = list_entry(p, struct mm_struct, mmlist);
2107 if (!mmget_not_zero(mm))
2109 spin_unlock(&mmlist_lock);
2115 swcount = *swap_map;
2116 if (!swap_count(swcount)) /* any usage ? */
2118 else if (mm == &init_mm)
2121 retval = unuse_mm(mm, entry, page);
2123 if (set_start_mm && *swap_map < swcount) {
2124 mmput(new_start_mm);
2129 spin_lock(&mmlist_lock);
2131 spin_unlock(&mmlist_lock);
2134 start_mm = new_start_mm;
2143 * If a reference remains (rare), we would like to leave
2144 * the page in the swap cache; but try_to_unmap could
2145 * then re-duplicate the entry once we drop page lock,
2146 * so we might loop indefinitely; also, that page could
2147 * not be swapped out to other storage meanwhile. So:
2148 * delete from cache even if there's another reference,
2149 * after ensuring that the data has been saved to disk -
2150 * since if the reference remains (rarer), it will be
2151 * read from disk into another page. Splitting into two
2152 * pages would be incorrect if swap supported "shared
2153 * private" pages, but they are handled by tmpfs files.
2155 * Given how unuse_vma() targets one particular offset
2156 * in an anon_vma, once the anon_vma has been determined,
2157 * this splitting happens to be just what is needed to
2158 * handle where KSM pages have been swapped out: re-reading
2159 * is unnecessarily slow, but we can fix that later on.
2161 if (swap_count(*swap_map) &&
2162 PageDirty(page) && PageSwapCache(page)) {
2163 struct writeback_control wbc = {
2164 .sync_mode = WB_SYNC_NONE,
2167 swap_writepage(compound_head(page), &wbc);
2169 wait_on_page_writeback(page);
2173 * It is conceivable that a racing task removed this page from
2174 * swap cache just before we acquired the page lock at the top,
2175 * or while we dropped it in unuse_mm(). The page might even
2176 * be back in swap cache on another swap area: that we must not
2177 * delete, since it may not have been written out to swap yet.
2179 if (PageSwapCache(page) &&
2180 likely(page_private(page) == entry.val) &&
2181 !page_swapped(page))
2182 delete_from_swap_cache(compound_head(page));
2185 * So we could skip searching mms once swap count went
2186 * to 1, we did not mark any present ptes as dirty: must
2187 * mark page dirty so shrink_page_list will preserve it.
2194 * Make sure that we aren't completely killing
2195 * interactive performance.
2198 if (frontswap && pages_to_unuse > 0) {
2199 if (!--pages_to_unuse)
2209 * After a successful try_to_unuse, if no swap is now in use, we know
2210 * we can empty the mmlist. swap_lock must be held on entry and exit.
2211 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2212 * added to the mmlist just after page_duplicate - before would be racy.
2214 static void drain_mmlist(void)
2216 struct list_head *p, *next;
2219 for (type = 0; type < nr_swapfiles; type++)
2220 if (swap_info[type]->inuse_pages)
2222 spin_lock(&mmlist_lock);
2223 list_for_each_safe(p, next, &init_mm.mmlist)
2225 spin_unlock(&mmlist_lock);
2229 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2230 * corresponds to page offset for the specified swap entry.
2231 * Note that the type of this function is sector_t, but it returns page offset
2232 * into the bdev, not sector offset.
2234 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2236 struct swap_info_struct *sis;
2237 struct swap_extent *start_se;
2238 struct swap_extent *se;
2241 sis = swap_info[swp_type(entry)];
2244 offset = swp_offset(entry);
2245 start_se = sis->curr_swap_extent;
2249 if (se->start_page <= offset &&
2250 offset < (se->start_page + se->nr_pages)) {
2251 return se->start_block + (offset - se->start_page);
2253 se = list_next_entry(se, list);
2254 sis->curr_swap_extent = se;
2255 BUG_ON(se == start_se); /* It *must* be present */
2260 * Returns the page offset into bdev for the specified page's swap entry.
2262 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2265 entry.val = page_private(page);
2266 return map_swap_entry(entry, bdev);
2270 * Free all of a swapdev's extent information
2272 static void destroy_swap_extents(struct swap_info_struct *sis)
2274 while (!list_empty(&sis->first_swap_extent.list)) {
2275 struct swap_extent *se;
2277 se = list_first_entry(&sis->first_swap_extent.list,
2278 struct swap_extent, list);
2279 list_del(&se->list);
2283 if (sis->flags & SWP_FILE) {
2284 struct file *swap_file = sis->swap_file;
2285 struct address_space *mapping = swap_file->f_mapping;
2287 sis->flags &= ~SWP_FILE;
2288 mapping->a_ops->swap_deactivate(swap_file);
2293 * Add a block range (and the corresponding page range) into this swapdev's
2294 * extent list. The extent list is kept sorted in page order.
2296 * This function rather assumes that it is called in ascending page order.
2299 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2300 unsigned long nr_pages, sector_t start_block)
2302 struct swap_extent *se;
2303 struct swap_extent *new_se;
2304 struct list_head *lh;
2306 if (start_page == 0) {
2307 se = &sis->first_swap_extent;
2308 sis->curr_swap_extent = se;
2310 se->nr_pages = nr_pages;
2311 se->start_block = start_block;
2314 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2315 se = list_entry(lh, struct swap_extent, list);
2316 BUG_ON(se->start_page + se->nr_pages != start_page);
2317 if (se->start_block + se->nr_pages == start_block) {
2319 se->nr_pages += nr_pages;
2325 * No merge. Insert a new extent, preserving ordering.
2327 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2330 new_se->start_page = start_page;
2331 new_se->nr_pages = nr_pages;
2332 new_se->start_block = start_block;
2334 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2339 * A `swap extent' is a simple thing which maps a contiguous range of pages
2340 * onto a contiguous range of disk blocks. An ordered list of swap extents
2341 * is built at swapon time and is then used at swap_writepage/swap_readpage
2342 * time for locating where on disk a page belongs.
2344 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2345 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2346 * swap files identically.
2348 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2349 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2350 * swapfiles are handled *identically* after swapon time.
2352 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2353 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2354 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2355 * requirements, they are simply tossed out - we will never use those blocks
2358 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2359 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2360 * which will scribble on the fs.
2362 * The amount of disk space which a single swap extent represents varies.
2363 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2364 * extents in the list. To avoid much list walking, we cache the previous
2365 * search location in `curr_swap_extent', and start new searches from there.
2366 * This is extremely effective. The average number of iterations in
2367 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2369 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2371 struct file *swap_file = sis->swap_file;
2372 struct address_space *mapping = swap_file->f_mapping;
2373 struct inode *inode = mapping->host;
2376 if (S_ISBLK(inode->i_mode)) {
2377 ret = add_swap_extent(sis, 0, sis->max, 0);
2382 if (mapping->a_ops->swap_activate) {
2383 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2385 sis->flags |= SWP_FILE;
2386 ret = add_swap_extent(sis, 0, sis->max, 0);
2392 return generic_swapfile_activate(sis, swap_file, span);
2395 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2396 unsigned char *swap_map,
2397 struct swap_cluster_info *cluster_info)
2402 p->prio = --least_priority;
2404 * the plist prio is negated because plist ordering is
2405 * low-to-high, while swap ordering is high-to-low
2407 p->list.prio = -p->prio;
2408 p->avail_list.prio = -p->prio;
2409 p->swap_map = swap_map;
2410 p->cluster_info = cluster_info;
2411 p->flags |= SWP_WRITEOK;
2412 atomic_long_add(p->pages, &nr_swap_pages);
2413 total_swap_pages += p->pages;
2415 assert_spin_locked(&swap_lock);
2417 * both lists are plists, and thus priority ordered.
2418 * swap_active_head needs to be priority ordered for swapoff(),
2419 * which on removal of any swap_info_struct with an auto-assigned
2420 * (i.e. negative) priority increments the auto-assigned priority
2421 * of any lower-priority swap_info_structs.
2422 * swap_avail_head needs to be priority ordered for get_swap_page(),
2423 * which allocates swap pages from the highest available priority
2426 plist_add(&p->list, &swap_active_head);
2427 spin_lock(&swap_avail_lock);
2428 plist_add(&p->avail_list, &swap_avail_head);
2429 spin_unlock(&swap_avail_lock);
2432 static void enable_swap_info(struct swap_info_struct *p, int prio,
2433 unsigned char *swap_map,
2434 struct swap_cluster_info *cluster_info,
2435 unsigned long *frontswap_map)
2437 frontswap_init(p->type, frontswap_map);
2438 spin_lock(&swap_lock);
2439 spin_lock(&p->lock);
2440 _enable_swap_info(p, prio, swap_map, cluster_info);
2441 spin_unlock(&p->lock);
2442 spin_unlock(&swap_lock);
2445 static void reinsert_swap_info(struct swap_info_struct *p)
2447 spin_lock(&swap_lock);
2448 spin_lock(&p->lock);
2449 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2450 spin_unlock(&p->lock);
2451 spin_unlock(&swap_lock);
2454 bool has_usable_swap(void)
2458 spin_lock(&swap_lock);
2459 if (plist_head_empty(&swap_active_head))
2461 spin_unlock(&swap_lock);
2465 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2467 struct swap_info_struct *p = NULL;
2468 unsigned char *swap_map;
2469 struct swap_cluster_info *cluster_info;
2470 unsigned long *frontswap_map;
2471 struct file *swap_file, *victim;
2472 struct address_space *mapping;
2473 struct inode *inode;
2474 struct filename *pathname;
2476 unsigned int old_block_size;
2478 if (!capable(CAP_SYS_ADMIN))
2481 BUG_ON(!current->mm);
2483 pathname = getname(specialfile);
2484 if (IS_ERR(pathname))
2485 return PTR_ERR(pathname);
2487 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2488 err = PTR_ERR(victim);
2492 mapping = victim->f_mapping;
2493 spin_lock(&swap_lock);
2494 plist_for_each_entry(p, &swap_active_head, list) {
2495 if (p->flags & SWP_WRITEOK) {
2496 if (p->swap_file->f_mapping == mapping) {
2504 spin_unlock(&swap_lock);
2507 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2508 vm_unacct_memory(p->pages);
2511 spin_unlock(&swap_lock);
2514 spin_lock(&swap_avail_lock);
2515 plist_del(&p->avail_list, &swap_avail_head);
2516 spin_unlock(&swap_avail_lock);
2517 spin_lock(&p->lock);
2519 struct swap_info_struct *si = p;
2521 plist_for_each_entry_continue(si, &swap_active_head, list) {
2524 si->avail_list.prio--;
2528 plist_del(&p->list, &swap_active_head);
2529 atomic_long_sub(p->pages, &nr_swap_pages);
2530 total_swap_pages -= p->pages;
2531 p->flags &= ~SWP_WRITEOK;
2532 spin_unlock(&p->lock);
2533 spin_unlock(&swap_lock);
2535 disable_swap_slots_cache_lock();
2537 set_current_oom_origin();
2538 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2539 clear_current_oom_origin();
2542 /* re-insert swap space back into swap_list */
2543 reinsert_swap_info(p);
2544 reenable_swap_slots_cache_unlock();
2548 reenable_swap_slots_cache_unlock();
2550 flush_work(&p->discard_work);
2552 destroy_swap_extents(p);
2553 if (p->flags & SWP_CONTINUED)
2554 free_swap_count_continuations(p);
2556 mutex_lock(&swapon_mutex);
2557 spin_lock(&swap_lock);
2558 spin_lock(&p->lock);
2561 /* wait for anyone still in scan_swap_map */
2562 p->highest_bit = 0; /* cuts scans short */
2563 while (p->flags >= SWP_SCANNING) {
2564 spin_unlock(&p->lock);
2565 spin_unlock(&swap_lock);
2566 schedule_timeout_uninterruptible(1);
2567 spin_lock(&swap_lock);
2568 spin_lock(&p->lock);
2571 swap_file = p->swap_file;
2572 old_block_size = p->old_block_size;
2573 p->swap_file = NULL;
2575 swap_map = p->swap_map;
2577 cluster_info = p->cluster_info;
2578 p->cluster_info = NULL;
2579 frontswap_map = frontswap_map_get(p);
2580 spin_unlock(&p->lock);
2581 spin_unlock(&swap_lock);
2582 frontswap_invalidate_area(p->type);
2583 frontswap_map_set(p, NULL);
2584 mutex_unlock(&swapon_mutex);
2585 free_percpu(p->percpu_cluster);
2586 p->percpu_cluster = NULL;
2588 kvfree(cluster_info);
2589 kvfree(frontswap_map);
2590 /* Destroy swap account information */
2591 swap_cgroup_swapoff(p->type);
2592 exit_swap_address_space(p->type);
2594 inode = mapping->host;
2595 if (S_ISBLK(inode->i_mode)) {
2596 struct block_device *bdev = I_BDEV(inode);
2597 set_blocksize(bdev, old_block_size);
2598 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2601 inode->i_flags &= ~S_SWAPFILE;
2602 inode_unlock(inode);
2604 filp_close(swap_file, NULL);
2607 * Clear the SWP_USED flag after all resources are freed so that swapon
2608 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2609 * not hold p->lock after we cleared its SWP_WRITEOK.
2611 spin_lock(&swap_lock);
2613 spin_unlock(&swap_lock);
2616 atomic_inc(&proc_poll_event);
2617 wake_up_interruptible(&proc_poll_wait);
2620 filp_close(victim, NULL);
2626 #ifdef CONFIG_PROC_FS
2627 static unsigned swaps_poll(struct file *file, poll_table *wait)
2629 struct seq_file *seq = file->private_data;
2631 poll_wait(file, &proc_poll_wait, wait);
2633 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2634 seq->poll_event = atomic_read(&proc_poll_event);
2635 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2638 return POLLIN | POLLRDNORM;
2642 static void *swap_start(struct seq_file *swap, loff_t *pos)
2644 struct swap_info_struct *si;
2648 mutex_lock(&swapon_mutex);
2651 return SEQ_START_TOKEN;
2653 for (type = 0; type < nr_swapfiles; type++) {
2654 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2655 si = swap_info[type];
2656 if (!(si->flags & SWP_USED) || !si->swap_map)
2665 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2667 struct swap_info_struct *si = v;
2670 if (v == SEQ_START_TOKEN)
2673 type = si->type + 1;
2675 for (; type < nr_swapfiles; type++) {
2676 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2677 si = swap_info[type];
2678 if (!(si->flags & SWP_USED) || !si->swap_map)
2687 static void swap_stop(struct seq_file *swap, void *v)
2689 mutex_unlock(&swapon_mutex);
2692 static int swap_show(struct seq_file *swap, void *v)
2694 struct swap_info_struct *si = v;
2698 if (si == SEQ_START_TOKEN) {
2699 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2703 file = si->swap_file;
2704 len = seq_file_path(swap, file, " \t\n\\");
2705 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2706 len < 40 ? 40 - len : 1, " ",
2707 S_ISBLK(file_inode(file)->i_mode) ?
2708 "partition" : "file\t",
2709 si->pages << (PAGE_SHIFT - 10),
2710 si->inuse_pages << (PAGE_SHIFT - 10),
2715 static const struct seq_operations swaps_op = {
2716 .start = swap_start,
2722 static int swaps_open(struct inode *inode, struct file *file)
2724 struct seq_file *seq;
2727 ret = seq_open(file, &swaps_op);
2731 seq = file->private_data;
2732 seq->poll_event = atomic_read(&proc_poll_event);
2736 static const struct file_operations proc_swaps_operations = {
2739 .llseek = seq_lseek,
2740 .release = seq_release,
2744 static int __init procswaps_init(void)
2746 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2749 __initcall(procswaps_init);
2750 #endif /* CONFIG_PROC_FS */
2752 #ifdef MAX_SWAPFILES_CHECK
2753 static int __init max_swapfiles_check(void)
2755 MAX_SWAPFILES_CHECK();
2758 late_initcall(max_swapfiles_check);
2761 static struct swap_info_struct *alloc_swap_info(void)
2763 struct swap_info_struct *p;
2766 p = kzalloc(sizeof(*p), GFP_KERNEL);
2768 return ERR_PTR(-ENOMEM);
2770 spin_lock(&swap_lock);
2771 for (type = 0; type < nr_swapfiles; type++) {
2772 if (!(swap_info[type]->flags & SWP_USED))
2775 if (type >= MAX_SWAPFILES) {
2776 spin_unlock(&swap_lock);
2778 return ERR_PTR(-EPERM);
2780 if (type >= nr_swapfiles) {
2782 swap_info[type] = p;
2784 * Write swap_info[type] before nr_swapfiles, in case a
2785 * racing procfs swap_start() or swap_next() is reading them.
2786 * (We never shrink nr_swapfiles, we never free this entry.)
2792 p = swap_info[type];
2794 * Do not memset this entry: a racing procfs swap_next()
2795 * would be relying on p->type to remain valid.
2798 INIT_LIST_HEAD(&p->first_swap_extent.list);
2799 plist_node_init(&p->list, 0);
2800 plist_node_init(&p->avail_list, 0);
2801 p->flags = SWP_USED;
2802 spin_unlock(&swap_lock);
2803 spin_lock_init(&p->lock);
2808 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2812 if (S_ISBLK(inode->i_mode)) {
2813 p->bdev = bdgrab(I_BDEV(inode));
2814 error = blkdev_get(p->bdev,
2815 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2820 p->old_block_size = block_size(p->bdev);
2821 error = set_blocksize(p->bdev, PAGE_SIZE);
2824 p->flags |= SWP_BLKDEV;
2825 } else if (S_ISREG(inode->i_mode)) {
2826 p->bdev = inode->i_sb->s_bdev;
2828 if (IS_SWAPFILE(inode))
2836 static unsigned long read_swap_header(struct swap_info_struct *p,
2837 union swap_header *swap_header,
2838 struct inode *inode)
2841 unsigned long maxpages;
2842 unsigned long swapfilepages;
2843 unsigned long last_page;
2845 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2846 pr_err("Unable to find swap-space signature\n");
2850 /* swap partition endianess hack... */
2851 if (swab32(swap_header->info.version) == 1) {
2852 swab32s(&swap_header->info.version);
2853 swab32s(&swap_header->info.last_page);
2854 swab32s(&swap_header->info.nr_badpages);
2855 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2857 for (i = 0; i < swap_header->info.nr_badpages; i++)
2858 swab32s(&swap_header->info.badpages[i]);
2860 /* Check the swap header's sub-version */
2861 if (swap_header->info.version != 1) {
2862 pr_warn("Unable to handle swap header version %d\n",
2863 swap_header->info.version);
2868 p->cluster_next = 1;
2872 * Find out how many pages are allowed for a single swap
2873 * device. There are two limiting factors: 1) the number
2874 * of bits for the swap offset in the swp_entry_t type, and
2875 * 2) the number of bits in the swap pte as defined by the
2876 * different architectures. In order to find the
2877 * largest possible bit mask, a swap entry with swap type 0
2878 * and swap offset ~0UL is created, encoded to a swap pte,
2879 * decoded to a swp_entry_t again, and finally the swap
2880 * offset is extracted. This will mask all the bits from
2881 * the initial ~0UL mask that can't be encoded in either
2882 * the swp_entry_t or the architecture definition of a
2885 maxpages = swp_offset(pte_to_swp_entry(
2886 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2887 last_page = swap_header->info.last_page;
2888 if (last_page > maxpages) {
2889 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2890 maxpages << (PAGE_SHIFT - 10),
2891 last_page << (PAGE_SHIFT - 10));
2893 if (maxpages > last_page) {
2894 maxpages = last_page + 1;
2895 /* p->max is an unsigned int: don't overflow it */
2896 if ((unsigned int)maxpages == 0)
2897 maxpages = UINT_MAX;
2899 p->highest_bit = maxpages - 1;
2903 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2904 if (swapfilepages && maxpages > swapfilepages) {
2905 pr_warn("Swap area shorter than signature indicates\n");
2908 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2910 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2916 #define SWAP_CLUSTER_INFO_COLS \
2917 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2918 #define SWAP_CLUSTER_SPACE_COLS \
2919 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2920 #define SWAP_CLUSTER_COLS \
2921 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2923 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2924 union swap_header *swap_header,
2925 unsigned char *swap_map,
2926 struct swap_cluster_info *cluster_info,
2927 unsigned long maxpages,
2931 unsigned int nr_good_pages;
2933 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2934 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2935 unsigned long i, idx;
2937 nr_good_pages = maxpages - 1; /* omit header page */
2939 cluster_list_init(&p->free_clusters);
2940 cluster_list_init(&p->discard_clusters);
2942 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2943 unsigned int page_nr = swap_header->info.badpages[i];
2944 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2946 if (page_nr < maxpages) {
2947 swap_map[page_nr] = SWAP_MAP_BAD;
2950 * Haven't marked the cluster free yet, no list
2951 * operation involved
2953 inc_cluster_info_page(p, cluster_info, page_nr);
2957 /* Haven't marked the cluster free yet, no list operation involved */
2958 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2959 inc_cluster_info_page(p, cluster_info, i);
2961 if (nr_good_pages) {
2962 swap_map[0] = SWAP_MAP_BAD;
2964 * Not mark the cluster free yet, no list
2965 * operation involved
2967 inc_cluster_info_page(p, cluster_info, 0);
2969 p->pages = nr_good_pages;
2970 nr_extents = setup_swap_extents(p, span);
2973 nr_good_pages = p->pages;
2975 if (!nr_good_pages) {
2976 pr_warn("Empty swap-file\n");
2985 * Reduce false cache line sharing between cluster_info and
2986 * sharing same address space.
2988 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2989 j = (k + col) % SWAP_CLUSTER_COLS;
2990 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2991 idx = i * SWAP_CLUSTER_COLS + j;
2992 if (idx >= nr_clusters)
2994 if (cluster_count(&cluster_info[idx]))
2996 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2997 cluster_list_add_tail(&p->free_clusters, cluster_info,
3005 * Helper to sys_swapon determining if a given swap
3006 * backing device queue supports DISCARD operations.
3008 static bool swap_discardable(struct swap_info_struct *si)
3010 struct request_queue *q = bdev_get_queue(si->bdev);
3012 if (!q || !blk_queue_discard(q))
3018 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3020 struct swap_info_struct *p;
3021 struct filename *name;
3022 struct file *swap_file = NULL;
3023 struct address_space *mapping;
3026 union swap_header *swap_header;
3029 unsigned long maxpages;
3030 unsigned char *swap_map = NULL;
3031 struct swap_cluster_info *cluster_info = NULL;
3032 unsigned long *frontswap_map = NULL;
3033 struct page *page = NULL;
3034 struct inode *inode = NULL;
3036 if (swap_flags & ~SWAP_FLAGS_VALID)
3039 if (!capable(CAP_SYS_ADMIN))
3042 p = alloc_swap_info();
3046 INIT_WORK(&p->discard_work, swap_discard_work);
3048 name = getname(specialfile);
3050 error = PTR_ERR(name);
3054 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3055 if (IS_ERR(swap_file)) {
3056 error = PTR_ERR(swap_file);
3061 p->swap_file = swap_file;
3062 mapping = swap_file->f_mapping;
3063 inode = mapping->host;
3065 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3066 error = claim_swapfile(p, inode);
3067 if (unlikely(error))
3071 * Read the swap header.
3073 if (!mapping->a_ops->readpage) {
3077 page = read_mapping_page(mapping, 0, swap_file);
3079 error = PTR_ERR(page);
3082 swap_header = kmap(page);
3084 maxpages = read_swap_header(p, swap_header, inode);
3085 if (unlikely(!maxpages)) {
3090 /* OK, set up the swap map and apply the bad block list */
3091 swap_map = vzalloc(maxpages);
3097 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3098 p->flags |= SWP_STABLE_WRITES;
3100 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3102 unsigned long ci, nr_cluster;
3104 p->flags |= SWP_SOLIDSTATE;
3106 * select a random position to start with to help wear leveling
3109 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3110 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3112 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3114 if (!cluster_info) {
3119 for (ci = 0; ci < nr_cluster; ci++)
3120 spin_lock_init(&((cluster_info + ci)->lock));
3122 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3123 if (!p->percpu_cluster) {
3127 for_each_possible_cpu(cpu) {
3128 struct percpu_cluster *cluster;
3129 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3130 cluster_set_null(&cluster->index);
3134 error = swap_cgroup_swapon(p->type, maxpages);
3138 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3139 cluster_info, maxpages, &span);
3140 if (unlikely(nr_extents < 0)) {
3144 /* frontswap enabled? set up bit-per-page map for frontswap */
3145 if (IS_ENABLED(CONFIG_FRONTSWAP))
3146 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3149 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3151 * When discard is enabled for swap with no particular
3152 * policy flagged, we set all swap discard flags here in
3153 * order to sustain backward compatibility with older
3154 * swapon(8) releases.
3156 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3160 * By flagging sys_swapon, a sysadmin can tell us to
3161 * either do single-time area discards only, or to just
3162 * perform discards for released swap page-clusters.
3163 * Now it's time to adjust the p->flags accordingly.
3165 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3166 p->flags &= ~SWP_PAGE_DISCARD;
3167 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3168 p->flags &= ~SWP_AREA_DISCARD;
3170 /* issue a swapon-time discard if it's still required */
3171 if (p->flags & SWP_AREA_DISCARD) {
3172 int err = discard_swap(p);
3174 pr_err("swapon: discard_swap(%p): %d\n",
3179 error = init_swap_address_space(p->type, maxpages);
3183 mutex_lock(&swapon_mutex);
3185 if (swap_flags & SWAP_FLAG_PREFER)
3187 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3188 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3190 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3191 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3192 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3193 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3194 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3195 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3196 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3197 (frontswap_map) ? "FS" : "");
3199 mutex_unlock(&swapon_mutex);
3200 atomic_inc(&proc_poll_event);
3201 wake_up_interruptible(&proc_poll_wait);
3203 if (S_ISREG(inode->i_mode))
3204 inode->i_flags |= S_SWAPFILE;
3208 free_percpu(p->percpu_cluster);
3209 p->percpu_cluster = NULL;
3210 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3211 set_blocksize(p->bdev, p->old_block_size);
3212 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3214 destroy_swap_extents(p);
3215 swap_cgroup_swapoff(p->type);
3216 spin_lock(&swap_lock);
3217 p->swap_file = NULL;
3219 spin_unlock(&swap_lock);
3221 vfree(cluster_info);
3223 if (inode && S_ISREG(inode->i_mode)) {
3224 inode_unlock(inode);
3227 filp_close(swap_file, NULL);
3230 if (page && !IS_ERR(page)) {
3236 if (inode && S_ISREG(inode->i_mode))
3237 inode_unlock(inode);
3239 enable_swap_slots_cache();
3243 void si_swapinfo(struct sysinfo *val)
3246 unsigned long nr_to_be_unused = 0;
3248 spin_lock(&swap_lock);
3249 for (type = 0; type < nr_swapfiles; type++) {
3250 struct swap_info_struct *si = swap_info[type];
3252 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3253 nr_to_be_unused += si->inuse_pages;
3255 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3256 val->totalswap = total_swap_pages + nr_to_be_unused;
3257 spin_unlock(&swap_lock);
3261 * Verify that a swap entry is valid and increment its swap map count.
3263 * Returns error code in following case.
3265 * - swp_entry is invalid -> EINVAL
3266 * - swp_entry is migration entry -> EINVAL
3267 * - swap-cache reference is requested but there is already one. -> EEXIST
3268 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3269 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3271 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3273 struct swap_info_struct *p;
3274 struct swap_cluster_info *ci;
3275 unsigned long offset, type;
3276 unsigned char count;
3277 unsigned char has_cache;
3280 if (non_swap_entry(entry))
3283 type = swp_type(entry);
3284 if (type >= nr_swapfiles)
3286 p = swap_info[type];
3287 offset = swp_offset(entry);
3288 if (unlikely(offset >= p->max))
3291 ci = lock_cluster_or_swap_info(p, offset);
3293 count = p->swap_map[offset];
3296 * swapin_readahead() doesn't check if a swap entry is valid, so the
3297 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3299 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3304 has_cache = count & SWAP_HAS_CACHE;
3305 count &= ~SWAP_HAS_CACHE;
3308 if (usage == SWAP_HAS_CACHE) {
3310 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3311 if (!has_cache && count)
3312 has_cache = SWAP_HAS_CACHE;
3313 else if (has_cache) /* someone else added cache */
3315 else /* no users remaining */
3318 } else if (count || has_cache) {
3320 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3322 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3324 else if (swap_count_continued(p, offset, count))
3325 count = COUNT_CONTINUED;
3329 err = -ENOENT; /* unused swap entry */
3331 p->swap_map[offset] = count | has_cache;
3334 unlock_cluster_or_swap_info(p, ci);
3339 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3344 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3345 * (in which case its reference count is never incremented).
3347 void swap_shmem_alloc(swp_entry_t entry)
3349 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3353 * Increase reference count of swap entry by 1.
3354 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3355 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3356 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3357 * might occur if a page table entry has got corrupted.
3359 int swap_duplicate(swp_entry_t entry)
3363 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3364 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3369 * @entry: swap entry for which we allocate swap cache.
3371 * Called when allocating swap cache for existing swap entry,
3372 * This can return error codes. Returns 0 at success.
3373 * -EBUSY means there is a swap cache.
3374 * Note: return code is different from swap_duplicate().
3376 int swapcache_prepare(swp_entry_t entry)
3378 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3381 struct swap_info_struct *page_swap_info(struct page *page)
3383 swp_entry_t swap = { .val = page_private(page) };
3384 return swap_info[swp_type(swap)];
3388 * out-of-line __page_file_ methods to avoid include hell.
3390 struct address_space *__page_file_mapping(struct page *page)
3392 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3393 return page_swap_info(page)->swap_file->f_mapping;
3395 EXPORT_SYMBOL_GPL(__page_file_mapping);
3397 pgoff_t __page_file_index(struct page *page)
3399 swp_entry_t swap = { .val = page_private(page) };
3400 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3401 return swp_offset(swap);
3403 EXPORT_SYMBOL_GPL(__page_file_index);
3406 * add_swap_count_continuation - called when a swap count is duplicated
3407 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3408 * page of the original vmalloc'ed swap_map, to hold the continuation count
3409 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3410 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3412 * These continuation pages are seldom referenced: the common paths all work
3413 * on the original swap_map, only referring to a continuation page when the
3414 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3416 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3417 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3418 * can be called after dropping locks.
3420 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3422 struct swap_info_struct *si;
3423 struct swap_cluster_info *ci;
3426 struct page *list_page;
3428 unsigned char count;
3431 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3432 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3434 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3436 si = swap_info_get(entry);
3439 * An acceptable race has occurred since the failing
3440 * __swap_duplicate(): the swap entry has been freed,
3441 * perhaps even the whole swap_map cleared for swapoff.
3446 offset = swp_offset(entry);
3448 ci = lock_cluster(si, offset);
3450 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3452 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3454 * The higher the swap count, the more likely it is that tasks
3455 * will race to add swap count continuation: we need to avoid
3456 * over-provisioning.
3463 spin_unlock(&si->lock);
3468 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3469 * no architecture is using highmem pages for kernel page tables: so it
3470 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3472 head = vmalloc_to_page(si->swap_map + offset);
3473 offset &= ~PAGE_MASK;
3476 * Page allocation does not initialize the page's lru field,
3477 * but it does always reset its private field.
3479 if (!page_private(head)) {
3480 BUG_ON(count & COUNT_CONTINUED);
3481 INIT_LIST_HEAD(&head->lru);
3482 set_page_private(head, SWP_CONTINUED);
3483 si->flags |= SWP_CONTINUED;
3486 list_for_each_entry(list_page, &head->lru, lru) {
3490 * If the previous map said no continuation, but we've found
3491 * a continuation page, free our allocation and use this one.
3493 if (!(count & COUNT_CONTINUED))
3496 map = kmap_atomic(list_page) + offset;
3501 * If this continuation count now has some space in it,
3502 * free our allocation and use this one.
3504 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3508 list_add_tail(&page->lru, &head->lru);
3509 page = NULL; /* now it's attached, don't free it */
3512 spin_unlock(&si->lock);
3520 * swap_count_continued - when the original swap_map count is incremented
3521 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3522 * into, carry if so, or else fail until a new continuation page is allocated;
3523 * when the original swap_map count is decremented from 0 with continuation,
3524 * borrow from the continuation and report whether it still holds more.
3525 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3528 static bool swap_count_continued(struct swap_info_struct *si,
3529 pgoff_t offset, unsigned char count)
3535 head = vmalloc_to_page(si->swap_map + offset);
3536 if (page_private(head) != SWP_CONTINUED) {
3537 BUG_ON(count & COUNT_CONTINUED);
3538 return false; /* need to add count continuation */
3541 offset &= ~PAGE_MASK;
3542 page = list_entry(head->lru.next, struct page, lru);
3543 map = kmap_atomic(page) + offset;
3545 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3546 goto init_map; /* jump over SWAP_CONT_MAX checks */
3548 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3550 * Think of how you add 1 to 999
3552 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3554 page = list_entry(page->lru.next, struct page, lru);
3555 BUG_ON(page == head);
3556 map = kmap_atomic(page) + offset;
3558 if (*map == SWAP_CONT_MAX) {
3560 page = list_entry(page->lru.next, struct page, lru);
3562 return false; /* add count continuation */
3563 map = kmap_atomic(page) + offset;
3564 init_map: *map = 0; /* we didn't zero the page */
3568 page = list_entry(page->lru.prev, struct page, lru);
3569 while (page != head) {
3570 map = kmap_atomic(page) + offset;
3571 *map = COUNT_CONTINUED;
3573 page = list_entry(page->lru.prev, struct page, lru);
3575 return true; /* incremented */
3577 } else { /* decrementing */
3579 * Think of how you subtract 1 from 1000
3581 BUG_ON(count != COUNT_CONTINUED);
3582 while (*map == COUNT_CONTINUED) {
3584 page = list_entry(page->lru.next, struct page, lru);
3585 BUG_ON(page == head);
3586 map = kmap_atomic(page) + offset;
3593 page = list_entry(page->lru.prev, struct page, lru);
3594 while (page != head) {
3595 map = kmap_atomic(page) + offset;
3596 *map = SWAP_CONT_MAX | count;
3597 count = COUNT_CONTINUED;
3599 page = list_entry(page->lru.prev, struct page, lru);
3601 return count == COUNT_CONTINUED;
3606 * free_swap_count_continuations - swapoff free all the continuation pages
3607 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3609 static void free_swap_count_continuations(struct swap_info_struct *si)
3613 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3615 head = vmalloc_to_page(si->swap_map + offset);
3616 if (page_private(head)) {
3617 struct page *page, *next;
3619 list_for_each_entry_safe(page, next, &head->lru, lru) {
3620 list_del(&page->lru);