4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shmem_fs.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32 #include <linux/poll.h>
33 #include <linux/oom.h>
34 #include <linux/frontswap.h>
35 #include <linux/swapfile.h>
36 #include <linux/export.h>
38 #include <asm/pgtable.h>
39 #include <asm/tlbflush.h>
40 #include <linux/swapops.h>
41 #include <linux/page_cgroup.h>
43 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
45 static void free_swap_count_continuations(struct swap_info_struct *);
46 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
48 DEFINE_SPINLOCK(swap_lock);
49 static unsigned int nr_swapfiles;
50 atomic_long_t nr_swap_pages;
51 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
52 long total_swap_pages;
53 static int least_priority;
54 static atomic_t highest_priority_index = ATOMIC_INIT(-1);
56 static const char Bad_file[] = "Bad swap file entry ";
57 static const char Unused_file[] = "Unused swap file entry ";
58 static const char Bad_offset[] = "Bad swap offset entry ";
59 static const char Unused_offset[] = "Unused swap offset entry ";
61 struct swap_list_t swap_list = {-1, -1};
63 struct swap_info_struct *swap_info[MAX_SWAPFILES];
65 static DEFINE_MUTEX(swapon_mutex);
67 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
68 /* Activity counter to indicate that a swapon or swapoff has occurred */
69 static atomic_t proc_poll_event = ATOMIC_INIT(0);
71 static inline unsigned char swap_count(unsigned char ent)
73 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
76 /* returns 1 if swap entry is freed */
78 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
80 swp_entry_t entry = swp_entry(si->type, offset);
84 page = find_get_page(swap_address_space(entry), entry.val);
88 * This function is called from scan_swap_map() and it's called
89 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
90 * We have to use trylock for avoiding deadlock. This is a special
91 * case and you should use try_to_free_swap() with explicit lock_page()
92 * in usual operations.
94 if (trylock_page(page)) {
95 ret = try_to_free_swap(page);
98 page_cache_release(page);
103 * swapon tell device that all the old swap contents can be discarded,
104 * to allow the swap device to optimize its wear-levelling.
106 static int discard_swap(struct swap_info_struct *si)
108 struct swap_extent *se;
109 sector_t start_block;
113 /* Do not discard the swap header page! */
114 se = &si->first_swap_extent;
115 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
116 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
118 err = blkdev_issue_discard(si->bdev, start_block,
119 nr_blocks, GFP_KERNEL, 0);
125 list_for_each_entry(se, &si->first_swap_extent.list, list) {
126 start_block = se->start_block << (PAGE_SHIFT - 9);
127 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
129 err = blkdev_issue_discard(si->bdev, start_block,
130 nr_blocks, GFP_KERNEL, 0);
136 return err; /* That will often be -EOPNOTSUPP */
140 * swap allocation tell device that a cluster of swap can now be discarded,
141 * to allow the swap device to optimize its wear-levelling.
143 static void discard_swap_cluster(struct swap_info_struct *si,
144 pgoff_t start_page, pgoff_t nr_pages)
146 struct swap_extent *se = si->curr_swap_extent;
147 int found_extent = 0;
150 struct list_head *lh;
152 if (se->start_page <= start_page &&
153 start_page < se->start_page + se->nr_pages) {
154 pgoff_t offset = start_page - se->start_page;
155 sector_t start_block = se->start_block + offset;
156 sector_t nr_blocks = se->nr_pages - offset;
158 if (nr_blocks > nr_pages)
159 nr_blocks = nr_pages;
160 start_page += nr_blocks;
161 nr_pages -= nr_blocks;
164 si->curr_swap_extent = se;
166 start_block <<= PAGE_SHIFT - 9;
167 nr_blocks <<= PAGE_SHIFT - 9;
168 if (blkdev_issue_discard(si->bdev, start_block,
169 nr_blocks, GFP_NOIO, 0))
174 se = list_entry(lh, struct swap_extent, list);
178 static int wait_for_discard(void *word)
184 #define SWAPFILE_CLUSTER 256
185 #define LATENCY_LIMIT 256
187 static inline void cluster_set_flag(struct swap_cluster_info *info,
193 static inline unsigned int cluster_count(struct swap_cluster_info *info)
198 static inline void cluster_set_count(struct swap_cluster_info *info,
204 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
205 unsigned int c, unsigned int f)
211 static inline unsigned int cluster_next(struct swap_cluster_info *info)
216 static inline void cluster_set_next(struct swap_cluster_info *info,
222 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
223 unsigned int n, unsigned int f)
229 static inline bool cluster_is_free(struct swap_cluster_info *info)
231 return info->flags & CLUSTER_FLAG_FREE;
234 static inline bool cluster_is_null(struct swap_cluster_info *info)
236 return info->flags & CLUSTER_FLAG_NEXT_NULL;
239 static inline void cluster_set_null(struct swap_cluster_info *info)
241 info->flags = CLUSTER_FLAG_NEXT_NULL;
246 * The cluster corresponding to page_nr will be used. The cluster will be
247 * removed from free cluster list and its usage counter will be increased.
249 static void inc_cluster_info_page(struct swap_info_struct *p,
250 struct swap_cluster_info *cluster_info, unsigned long page_nr)
252 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
256 if (cluster_is_free(&cluster_info[idx])) {
257 VM_BUG_ON(cluster_next(&p->free_cluster_head) != idx);
258 cluster_set_next_flag(&p->free_cluster_head,
259 cluster_next(&cluster_info[idx]), 0);
260 if (cluster_next(&p->free_cluster_tail) == idx) {
261 cluster_set_null(&p->free_cluster_tail);
262 cluster_set_null(&p->free_cluster_head);
264 cluster_set_count_flag(&cluster_info[idx], 0, 0);
267 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
268 cluster_set_count(&cluster_info[idx],
269 cluster_count(&cluster_info[idx]) + 1);
273 * The cluster corresponding to page_nr decreases one usage. If the usage
274 * counter becomes 0, which means no page in the cluster is in using, we can
275 * optionally discard the cluster and add it to free cluster list.
277 static void dec_cluster_info_page(struct swap_info_struct *p,
278 struct swap_cluster_info *cluster_info, unsigned long page_nr)
280 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
285 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
286 cluster_set_count(&cluster_info[idx],
287 cluster_count(&cluster_info[idx]) - 1);
289 if (cluster_count(&cluster_info[idx]) == 0) {
290 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
291 if (cluster_is_null(&p->free_cluster_head)) {
292 cluster_set_next_flag(&p->free_cluster_head, idx, 0);
293 cluster_set_next_flag(&p->free_cluster_tail, idx, 0);
295 unsigned int tail = cluster_next(&p->free_cluster_tail);
296 cluster_set_next(&cluster_info[tail], idx);
297 cluster_set_next_flag(&p->free_cluster_tail, idx, 0);
303 * It's possible scan_swap_map() uses a free cluster in the middle of free
304 * cluster list. Avoiding such abuse to avoid list corruption.
306 static inline bool scan_swap_map_recheck_cluster(struct swap_info_struct *si,
307 unsigned long offset)
309 offset /= SWAPFILE_CLUSTER;
310 return !cluster_is_null(&si->free_cluster_head) &&
311 offset != cluster_next(&si->free_cluster_head) &&
312 cluster_is_free(&si->cluster_info[offset]);
315 static unsigned long scan_swap_map(struct swap_info_struct *si,
318 unsigned long offset;
319 unsigned long scan_base;
320 unsigned long last_in_cluster = 0;
321 int latency_ration = LATENCY_LIMIT;
322 int found_free_cluster = 0;
325 * We try to cluster swap pages by allocating them sequentially
326 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
327 * way, however, we resort to first-free allocation, starting
328 * a new cluster. This prevents us from scattering swap pages
329 * all over the entire swap partition, so that we reduce
330 * overall disk seek times between swap pages. -- sct
331 * But we do now try to find an empty cluster. -Andrea
332 * And we let swap pages go all over an SSD partition. Hugh
335 si->flags += SWP_SCANNING;
336 scan_base = offset = si->cluster_next;
338 if (unlikely(!si->cluster_nr--)) {
339 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
340 si->cluster_nr = SWAPFILE_CLUSTER - 1;
343 if (si->flags & SWP_PAGE_DISCARD) {
345 * Start range check on racing allocations, in case
346 * they overlap the cluster we eventually decide on
347 * (we scan without swap_lock to allow preemption).
348 * It's hardly conceivable that cluster_nr could be
349 * wrapped during our scan, but don't depend on it.
351 if (si->lowest_alloc)
353 si->lowest_alloc = si->max;
354 si->highest_alloc = 0;
357 if (!cluster_is_null(&si->free_cluster_head)) {
358 offset = cluster_next(&si->free_cluster_head) *
360 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
361 si->cluster_next = offset;
362 si->cluster_nr = SWAPFILE_CLUSTER - 1;
363 found_free_cluster = 1;
365 } else if (si->cluster_info) {
367 * Checking free cluster is fast enough, we can do the
371 si->lowest_alloc = 0;
375 spin_unlock(&si->lock);
378 * If seek is expensive, start searching for new cluster from
379 * start of partition, to minimize the span of allocated swap.
380 * But if seek is cheap, search from our current position, so
381 * that swap is allocated from all over the partition: if the
382 * Flash Translation Layer only remaps within limited zones,
383 * we don't want to wear out the first zone too quickly.
385 if (!(si->flags & SWP_SOLIDSTATE))
386 scan_base = offset = si->lowest_bit;
387 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
389 /* Locate the first empty (unaligned) cluster */
390 for (; last_in_cluster <= si->highest_bit; offset++) {
391 if (si->swap_map[offset])
392 last_in_cluster = offset + SWAPFILE_CLUSTER;
393 else if (offset == last_in_cluster) {
394 spin_lock(&si->lock);
395 offset -= SWAPFILE_CLUSTER - 1;
396 si->cluster_next = offset;
397 si->cluster_nr = SWAPFILE_CLUSTER - 1;
398 found_free_cluster = 1;
401 if (unlikely(--latency_ration < 0)) {
403 latency_ration = LATENCY_LIMIT;
407 offset = si->lowest_bit;
408 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
410 /* Locate the first empty (unaligned) cluster */
411 for (; last_in_cluster < scan_base; offset++) {
412 if (si->swap_map[offset])
413 last_in_cluster = offset + SWAPFILE_CLUSTER;
414 else if (offset == last_in_cluster) {
415 spin_lock(&si->lock);
416 offset -= SWAPFILE_CLUSTER - 1;
417 si->cluster_next = offset;
418 si->cluster_nr = SWAPFILE_CLUSTER - 1;
419 found_free_cluster = 1;
422 if (unlikely(--latency_ration < 0)) {
424 latency_ration = LATENCY_LIMIT;
429 spin_lock(&si->lock);
430 si->cluster_nr = SWAPFILE_CLUSTER - 1;
431 si->lowest_alloc = 0;
435 if (scan_swap_map_recheck_cluster(si, offset))
437 if (!(si->flags & SWP_WRITEOK))
439 if (!si->highest_bit)
441 if (offset > si->highest_bit)
442 scan_base = offset = si->lowest_bit;
444 /* reuse swap entry of cache-only swap if not busy. */
445 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
447 spin_unlock(&si->lock);
448 swap_was_freed = __try_to_reclaim_swap(si, offset);
449 spin_lock(&si->lock);
450 /* entry was freed successfully, try to use this again */
453 goto scan; /* check next one */
456 if (si->swap_map[offset])
459 if (offset == si->lowest_bit)
461 if (offset == si->highest_bit)
464 if (si->inuse_pages == si->pages) {
465 si->lowest_bit = si->max;
468 si->swap_map[offset] = usage;
469 inc_cluster_info_page(si, si->cluster_info, offset);
470 si->cluster_next = offset + 1;
471 si->flags -= SWP_SCANNING;
473 if (si->lowest_alloc) {
475 * Only set when SWP_PAGE_DISCARD, and there's a scan
476 * for a free cluster in progress or just completed.
478 if (found_free_cluster) {
480 * To optimize wear-levelling, discard the
481 * old data of the cluster, taking care not to
482 * discard any of its pages that have already
483 * been allocated by racing tasks (offset has
484 * already stepped over any at the beginning).
486 if (offset < si->highest_alloc &&
487 si->lowest_alloc <= last_in_cluster)
488 last_in_cluster = si->lowest_alloc - 1;
489 si->flags |= SWP_DISCARDING;
490 spin_unlock(&si->lock);
492 if (offset < last_in_cluster)
493 discard_swap_cluster(si, offset,
494 last_in_cluster - offset + 1);
496 spin_lock(&si->lock);
497 si->lowest_alloc = 0;
498 si->flags &= ~SWP_DISCARDING;
500 smp_mb(); /* wake_up_bit advises this */
501 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
503 } else if (si->flags & SWP_DISCARDING) {
505 * Delay using pages allocated by racing tasks
506 * until the whole discard has been issued. We
507 * could defer that delay until swap_writepage,
508 * but it's easier to keep this self-contained.
510 spin_unlock(&si->lock);
511 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
512 wait_for_discard, TASK_UNINTERRUPTIBLE);
513 spin_lock(&si->lock);
516 * Note pages allocated by racing tasks while
517 * scan for a free cluster is in progress, so
518 * that its final discard can exclude them.
520 if (offset < si->lowest_alloc)
521 si->lowest_alloc = offset;
522 if (offset > si->highest_alloc)
523 si->highest_alloc = offset;
529 spin_unlock(&si->lock);
530 while (++offset <= si->highest_bit) {
531 if (!si->swap_map[offset]) {
532 spin_lock(&si->lock);
535 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
536 spin_lock(&si->lock);
539 if (unlikely(--latency_ration < 0)) {
541 latency_ration = LATENCY_LIMIT;
544 offset = si->lowest_bit;
545 while (++offset < scan_base) {
546 if (!si->swap_map[offset]) {
547 spin_lock(&si->lock);
550 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
551 spin_lock(&si->lock);
554 if (unlikely(--latency_ration < 0)) {
556 latency_ration = LATENCY_LIMIT;
559 spin_lock(&si->lock);
562 si->flags -= SWP_SCANNING;
566 swp_entry_t get_swap_page(void)
568 struct swap_info_struct *si;
574 spin_lock(&swap_lock);
575 if (atomic_long_read(&nr_swap_pages) <= 0)
577 atomic_long_dec(&nr_swap_pages);
579 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
580 hp_index = atomic_xchg(&highest_priority_index, -1);
582 * highest_priority_index records current highest priority swap
583 * type which just frees swap entries. If its priority is
584 * higher than that of swap_list.next swap type, we use it. It
585 * isn't protected by swap_lock, so it can be an invalid value
586 * if the corresponding swap type is swapoff. We double check
587 * the flags here. It's even possible the swap type is swapoff
588 * and swapon again and its priority is changed. In such rare
589 * case, low prority swap type might be used, but eventually
590 * high priority swap will be used after several rounds of
593 if (hp_index != -1 && hp_index != type &&
594 swap_info[type]->prio < swap_info[hp_index]->prio &&
595 (swap_info[hp_index]->flags & SWP_WRITEOK)) {
597 swap_list.next = type;
600 si = swap_info[type];
603 (!wrapped && si->prio != swap_info[next]->prio)) {
604 next = swap_list.head;
608 spin_lock(&si->lock);
609 if (!si->highest_bit) {
610 spin_unlock(&si->lock);
613 if (!(si->flags & SWP_WRITEOK)) {
614 spin_unlock(&si->lock);
618 swap_list.next = next;
620 spin_unlock(&swap_lock);
621 /* This is called for allocating swap entry for cache */
622 offset = scan_swap_map(si, SWAP_HAS_CACHE);
623 spin_unlock(&si->lock);
625 return swp_entry(type, offset);
626 spin_lock(&swap_lock);
627 next = swap_list.next;
630 atomic_long_inc(&nr_swap_pages);
632 spin_unlock(&swap_lock);
633 return (swp_entry_t) {0};
636 /* The only caller of this function is now susupend routine */
637 swp_entry_t get_swap_page_of_type(int type)
639 struct swap_info_struct *si;
642 si = swap_info[type];
643 spin_lock(&si->lock);
644 if (si && (si->flags & SWP_WRITEOK)) {
645 atomic_long_dec(&nr_swap_pages);
646 /* This is called for allocating swap entry, not cache */
647 offset = scan_swap_map(si, 1);
649 spin_unlock(&si->lock);
650 return swp_entry(type, offset);
652 atomic_long_inc(&nr_swap_pages);
654 spin_unlock(&si->lock);
655 return (swp_entry_t) {0};
658 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
660 struct swap_info_struct *p;
661 unsigned long offset, type;
665 type = swp_type(entry);
666 if (type >= nr_swapfiles)
669 if (!(p->flags & SWP_USED))
671 offset = swp_offset(entry);
672 if (offset >= p->max)
674 if (!p->swap_map[offset])
680 pr_err("swap_free: %s%08lx\n", Unused_offset, entry.val);
683 pr_err("swap_free: %s%08lx\n", Bad_offset, entry.val);
686 pr_err("swap_free: %s%08lx\n", Unused_file, entry.val);
689 pr_err("swap_free: %s%08lx\n", Bad_file, entry.val);
695 * This swap type frees swap entry, check if it is the highest priority swap
696 * type which just frees swap entry. get_swap_page() uses
697 * highest_priority_index to search highest priority swap type. The
698 * swap_info_struct.lock can't protect us if there are multiple swap types
699 * active, so we use atomic_cmpxchg.
701 static void set_highest_priority_index(int type)
703 int old_hp_index, new_hp_index;
706 old_hp_index = atomic_read(&highest_priority_index);
707 if (old_hp_index != -1 &&
708 swap_info[old_hp_index]->prio >= swap_info[type]->prio)
711 } while (atomic_cmpxchg(&highest_priority_index,
712 old_hp_index, new_hp_index) != old_hp_index);
715 static unsigned char swap_entry_free(struct swap_info_struct *p,
716 swp_entry_t entry, unsigned char usage)
718 unsigned long offset = swp_offset(entry);
720 unsigned char has_cache;
722 count = p->swap_map[offset];
723 has_cache = count & SWAP_HAS_CACHE;
724 count &= ~SWAP_HAS_CACHE;
726 if (usage == SWAP_HAS_CACHE) {
727 VM_BUG_ON(!has_cache);
729 } else if (count == SWAP_MAP_SHMEM) {
731 * Or we could insist on shmem.c using a special
732 * swap_shmem_free() and free_shmem_swap_and_cache()...
735 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
736 if (count == COUNT_CONTINUED) {
737 if (swap_count_continued(p, offset, count))
738 count = SWAP_MAP_MAX | COUNT_CONTINUED;
740 count = SWAP_MAP_MAX;
746 mem_cgroup_uncharge_swap(entry);
748 usage = count | has_cache;
749 p->swap_map[offset] = usage;
751 /* free if no reference */
753 dec_cluster_info_page(p, p->cluster_info, offset);
754 if (offset < p->lowest_bit)
755 p->lowest_bit = offset;
756 if (offset > p->highest_bit)
757 p->highest_bit = offset;
758 set_highest_priority_index(p->type);
759 atomic_long_inc(&nr_swap_pages);
761 frontswap_invalidate_page(p->type, offset);
762 if (p->flags & SWP_BLKDEV) {
763 struct gendisk *disk = p->bdev->bd_disk;
764 if (disk->fops->swap_slot_free_notify)
765 disk->fops->swap_slot_free_notify(p->bdev,
774 * Caller has made sure that the swapdevice corresponding to entry
775 * is still around or has not been recycled.
777 void swap_free(swp_entry_t entry)
779 struct swap_info_struct *p;
781 p = swap_info_get(entry);
783 swap_entry_free(p, entry, 1);
784 spin_unlock(&p->lock);
789 * Called after dropping swapcache to decrease refcnt to swap entries.
791 void swapcache_free(swp_entry_t entry, struct page *page)
793 struct swap_info_struct *p;
796 p = swap_info_get(entry);
798 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
800 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
801 spin_unlock(&p->lock);
806 * How many references to page are currently swapped out?
807 * This does not give an exact answer when swap count is continued,
808 * but does include the high COUNT_CONTINUED flag to allow for that.
810 int page_swapcount(struct page *page)
813 struct swap_info_struct *p;
816 entry.val = page_private(page);
817 p = swap_info_get(entry);
819 count = swap_count(p->swap_map[swp_offset(entry)]);
820 spin_unlock(&p->lock);
826 * We can write to an anon page without COW if there are no other references
827 * to it. And as a side-effect, free up its swap: because the old content
828 * on disk will never be read, and seeking back there to write new content
829 * later would only waste time away from clustering.
831 int reuse_swap_page(struct page *page)
835 VM_BUG_ON(!PageLocked(page));
836 if (unlikely(PageKsm(page)))
838 count = page_mapcount(page);
839 if (count <= 1 && PageSwapCache(page)) {
840 count += page_swapcount(page);
841 if (count == 1 && !PageWriteback(page)) {
842 delete_from_swap_cache(page);
850 * If swap is getting full, or if there are no more mappings of this page,
851 * then try_to_free_swap is called to free its swap space.
853 int try_to_free_swap(struct page *page)
855 VM_BUG_ON(!PageLocked(page));
857 if (!PageSwapCache(page))
859 if (PageWriteback(page))
861 if (page_swapcount(page))
865 * Once hibernation has begun to create its image of memory,
866 * there's a danger that one of the calls to try_to_free_swap()
867 * - most probably a call from __try_to_reclaim_swap() while
868 * hibernation is allocating its own swap pages for the image,
869 * but conceivably even a call from memory reclaim - will free
870 * the swap from a page which has already been recorded in the
871 * image as a clean swapcache page, and then reuse its swap for
872 * another page of the image. On waking from hibernation, the
873 * original page might be freed under memory pressure, then
874 * later read back in from swap, now with the wrong data.
876 * Hibration suspends storage while it is writing the image
877 * to disk so check that here.
879 if (pm_suspended_storage())
882 delete_from_swap_cache(page);
888 * Free the swap entry like above, but also try to
889 * free the page cache entry if it is the last user.
891 int free_swap_and_cache(swp_entry_t entry)
893 struct swap_info_struct *p;
894 struct page *page = NULL;
896 if (non_swap_entry(entry))
899 p = swap_info_get(entry);
901 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
902 page = find_get_page(swap_address_space(entry),
904 if (page && !trylock_page(page)) {
905 page_cache_release(page);
909 spin_unlock(&p->lock);
913 * Not mapped elsewhere, or swap space full? Free it!
914 * Also recheck PageSwapCache now page is locked (above).
916 if (PageSwapCache(page) && !PageWriteback(page) &&
917 (!page_mapped(page) || vm_swap_full())) {
918 delete_from_swap_cache(page);
922 page_cache_release(page);
927 #ifdef CONFIG_HIBERNATION
929 * Find the swap type that corresponds to given device (if any).
931 * @offset - number of the PAGE_SIZE-sized block of the device, starting
932 * from 0, in which the swap header is expected to be located.
934 * This is needed for the suspend to disk (aka swsusp).
936 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
938 struct block_device *bdev = NULL;
942 bdev = bdget(device);
944 spin_lock(&swap_lock);
945 for (type = 0; type < nr_swapfiles; type++) {
946 struct swap_info_struct *sis = swap_info[type];
948 if (!(sis->flags & SWP_WRITEOK))
953 *bdev_p = bdgrab(sis->bdev);
955 spin_unlock(&swap_lock);
958 if (bdev == sis->bdev) {
959 struct swap_extent *se = &sis->first_swap_extent;
961 if (se->start_block == offset) {
963 *bdev_p = bdgrab(sis->bdev);
965 spin_unlock(&swap_lock);
971 spin_unlock(&swap_lock);
979 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
980 * corresponding to given index in swap_info (swap type).
982 sector_t swapdev_block(int type, pgoff_t offset)
984 struct block_device *bdev;
986 if ((unsigned int)type >= nr_swapfiles)
988 if (!(swap_info[type]->flags & SWP_WRITEOK))
990 return map_swap_entry(swp_entry(type, offset), &bdev);
994 * Return either the total number of swap pages of given type, or the number
995 * of free pages of that type (depending on @free)
997 * This is needed for software suspend
999 unsigned int count_swap_pages(int type, int free)
1003 spin_lock(&swap_lock);
1004 if ((unsigned int)type < nr_swapfiles) {
1005 struct swap_info_struct *sis = swap_info[type];
1007 spin_lock(&sis->lock);
1008 if (sis->flags & SWP_WRITEOK) {
1011 n -= sis->inuse_pages;
1013 spin_unlock(&sis->lock);
1015 spin_unlock(&swap_lock);
1018 #endif /* CONFIG_HIBERNATION */
1020 static inline int maybe_same_pte(pte_t pte, pte_t swp_pte)
1022 #ifdef CONFIG_MEM_SOFT_DIRTY
1024 * When pte keeps soft dirty bit the pte generated
1025 * from swap entry does not has it, still it's same
1026 * pte from logical point of view.
1028 pte_t swp_pte_dirty = pte_swp_mksoft_dirty(swp_pte);
1029 return pte_same(pte, swp_pte) || pte_same(pte, swp_pte_dirty);
1031 return pte_same(pte, swp_pte);
1036 * No need to decide whether this PTE shares the swap entry with others,
1037 * just let do_wp_page work it out if a write is requested later - to
1038 * force COW, vm_page_prot omits write permission from any private vma.
1040 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1041 unsigned long addr, swp_entry_t entry, struct page *page)
1043 struct page *swapcache;
1044 struct mem_cgroup *memcg;
1050 page = ksm_might_need_to_copy(page, vma, addr);
1051 if (unlikely(!page))
1054 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
1055 GFP_KERNEL, &memcg)) {
1060 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1061 if (unlikely(!maybe_same_pte(*pte, swp_entry_to_pte(entry)))) {
1062 mem_cgroup_cancel_charge_swapin(memcg);
1067 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1068 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1070 set_pte_at(vma->vm_mm, addr, pte,
1071 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1072 if (page == swapcache)
1073 page_add_anon_rmap(page, vma, addr);
1074 else /* ksm created a completely new copy */
1075 page_add_new_anon_rmap(page, vma, addr);
1076 mem_cgroup_commit_charge_swapin(page, memcg);
1079 * Move the page to the active list so it is not
1080 * immediately swapped out again after swapon.
1082 activate_page(page);
1084 pte_unmap_unlock(pte, ptl);
1086 if (page != swapcache) {
1093 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1094 unsigned long addr, unsigned long end,
1095 swp_entry_t entry, struct page *page)
1097 pte_t swp_pte = swp_entry_to_pte(entry);
1102 * We don't actually need pte lock while scanning for swp_pte: since
1103 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1104 * page table while we're scanning; though it could get zapped, and on
1105 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1106 * of unmatched parts which look like swp_pte, so unuse_pte must
1107 * recheck under pte lock. Scanning without pte lock lets it be
1108 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1110 pte = pte_offset_map(pmd, addr);
1113 * swapoff spends a _lot_ of time in this loop!
1114 * Test inline before going to call unuse_pte.
1116 if (unlikely(maybe_same_pte(*pte, swp_pte))) {
1118 ret = unuse_pte(vma, pmd, addr, entry, page);
1121 pte = pte_offset_map(pmd, addr);
1123 } while (pte++, addr += PAGE_SIZE, addr != end);
1129 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1130 unsigned long addr, unsigned long end,
1131 swp_entry_t entry, struct page *page)
1137 pmd = pmd_offset(pud, addr);
1139 next = pmd_addr_end(addr, end);
1140 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1142 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1145 } while (pmd++, addr = next, addr != end);
1149 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
1150 unsigned long addr, unsigned long end,
1151 swp_entry_t entry, struct page *page)
1157 pud = pud_offset(pgd, addr);
1159 next = pud_addr_end(addr, end);
1160 if (pud_none_or_clear_bad(pud))
1162 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1165 } while (pud++, addr = next, addr != end);
1169 static int unuse_vma(struct vm_area_struct *vma,
1170 swp_entry_t entry, struct page *page)
1173 unsigned long addr, end, next;
1176 if (page_anon_vma(page)) {
1177 addr = page_address_in_vma(page, vma);
1178 if (addr == -EFAULT)
1181 end = addr + PAGE_SIZE;
1183 addr = vma->vm_start;
1187 pgd = pgd_offset(vma->vm_mm, addr);
1189 next = pgd_addr_end(addr, end);
1190 if (pgd_none_or_clear_bad(pgd))
1192 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1195 } while (pgd++, addr = next, addr != end);
1199 static int unuse_mm(struct mm_struct *mm,
1200 swp_entry_t entry, struct page *page)
1202 struct vm_area_struct *vma;
1205 if (!down_read_trylock(&mm->mmap_sem)) {
1207 * Activate page so shrink_inactive_list is unlikely to unmap
1208 * its ptes while lock is dropped, so swapoff can make progress.
1210 activate_page(page);
1212 down_read(&mm->mmap_sem);
1215 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1216 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1219 up_read(&mm->mmap_sem);
1220 return (ret < 0)? ret: 0;
1224 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1225 * from current position to next entry still in use.
1226 * Recycle to start on reaching the end, returning 0 when empty.
1228 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1229 unsigned int prev, bool frontswap)
1231 unsigned int max = si->max;
1232 unsigned int i = prev;
1233 unsigned char count;
1236 * No need for swap_lock here: we're just looking
1237 * for whether an entry is in use, not modifying it; false
1238 * hits are okay, and sys_swapoff() has already prevented new
1239 * allocations from this area (while holding swap_lock).
1248 * No entries in use at top of swap_map,
1249 * loop back to start and recheck there.
1256 if (frontswap_test(si, i))
1261 count = si->swap_map[i];
1262 if (count && swap_count(count) != SWAP_MAP_BAD)
1269 * We completely avoid races by reading each swap page in advance,
1270 * and then search for the process using it. All the necessary
1271 * page table adjustments can then be made atomically.
1273 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1274 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1276 int try_to_unuse(unsigned int type, bool frontswap,
1277 unsigned long pages_to_unuse)
1279 struct swap_info_struct *si = swap_info[type];
1280 struct mm_struct *start_mm;
1281 unsigned char *swap_map;
1282 unsigned char swcount;
1289 * When searching mms for an entry, a good strategy is to
1290 * start at the first mm we freed the previous entry from
1291 * (though actually we don't notice whether we or coincidence
1292 * freed the entry). Initialize this start_mm with a hold.
1294 * A simpler strategy would be to start at the last mm we
1295 * freed the previous entry from; but that would take less
1296 * advantage of mmlist ordering, which clusters forked mms
1297 * together, child after parent. If we race with dup_mmap(), we
1298 * prefer to resolve parent before child, lest we miss entries
1299 * duplicated after we scanned child: using last mm would invert
1302 start_mm = &init_mm;
1303 atomic_inc(&init_mm.mm_users);
1306 * Keep on scanning until all entries have gone. Usually,
1307 * one pass through swap_map is enough, but not necessarily:
1308 * there are races when an instance of an entry might be missed.
1310 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1311 if (signal_pending(current)) {
1317 * Get a page for the entry, using the existing swap
1318 * cache page if there is one. Otherwise, get a clean
1319 * page and read the swap into it.
1321 swap_map = &si->swap_map[i];
1322 entry = swp_entry(type, i);
1323 page = read_swap_cache_async(entry,
1324 GFP_HIGHUSER_MOVABLE, NULL, 0);
1327 * Either swap_duplicate() failed because entry
1328 * has been freed independently, and will not be
1329 * reused since sys_swapoff() already disabled
1330 * allocation from here, or alloc_page() failed.
1339 * Don't hold on to start_mm if it looks like exiting.
1341 if (atomic_read(&start_mm->mm_users) == 1) {
1343 start_mm = &init_mm;
1344 atomic_inc(&init_mm.mm_users);
1348 * Wait for and lock page. When do_swap_page races with
1349 * try_to_unuse, do_swap_page can handle the fault much
1350 * faster than try_to_unuse can locate the entry. This
1351 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1352 * defer to do_swap_page in such a case - in some tests,
1353 * do_swap_page and try_to_unuse repeatedly compete.
1355 wait_on_page_locked(page);
1356 wait_on_page_writeback(page);
1358 wait_on_page_writeback(page);
1361 * Remove all references to entry.
1363 swcount = *swap_map;
1364 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1365 retval = shmem_unuse(entry, page);
1366 /* page has already been unlocked and released */
1371 if (swap_count(swcount) && start_mm != &init_mm)
1372 retval = unuse_mm(start_mm, entry, page);
1374 if (swap_count(*swap_map)) {
1375 int set_start_mm = (*swap_map >= swcount);
1376 struct list_head *p = &start_mm->mmlist;
1377 struct mm_struct *new_start_mm = start_mm;
1378 struct mm_struct *prev_mm = start_mm;
1379 struct mm_struct *mm;
1381 atomic_inc(&new_start_mm->mm_users);
1382 atomic_inc(&prev_mm->mm_users);
1383 spin_lock(&mmlist_lock);
1384 while (swap_count(*swap_map) && !retval &&
1385 (p = p->next) != &start_mm->mmlist) {
1386 mm = list_entry(p, struct mm_struct, mmlist);
1387 if (!atomic_inc_not_zero(&mm->mm_users))
1389 spin_unlock(&mmlist_lock);
1395 swcount = *swap_map;
1396 if (!swap_count(swcount)) /* any usage ? */
1398 else if (mm == &init_mm)
1401 retval = unuse_mm(mm, entry, page);
1403 if (set_start_mm && *swap_map < swcount) {
1404 mmput(new_start_mm);
1405 atomic_inc(&mm->mm_users);
1409 spin_lock(&mmlist_lock);
1411 spin_unlock(&mmlist_lock);
1414 start_mm = new_start_mm;
1418 page_cache_release(page);
1423 * If a reference remains (rare), we would like to leave
1424 * the page in the swap cache; but try_to_unmap could
1425 * then re-duplicate the entry once we drop page lock,
1426 * so we might loop indefinitely; also, that page could
1427 * not be swapped out to other storage meanwhile. So:
1428 * delete from cache even if there's another reference,
1429 * after ensuring that the data has been saved to disk -
1430 * since if the reference remains (rarer), it will be
1431 * read from disk into another page. Splitting into two
1432 * pages would be incorrect if swap supported "shared
1433 * private" pages, but they are handled by tmpfs files.
1435 * Given how unuse_vma() targets one particular offset
1436 * in an anon_vma, once the anon_vma has been determined,
1437 * this splitting happens to be just what is needed to
1438 * handle where KSM pages have been swapped out: re-reading
1439 * is unnecessarily slow, but we can fix that later on.
1441 if (swap_count(*swap_map) &&
1442 PageDirty(page) && PageSwapCache(page)) {
1443 struct writeback_control wbc = {
1444 .sync_mode = WB_SYNC_NONE,
1447 swap_writepage(page, &wbc);
1449 wait_on_page_writeback(page);
1453 * It is conceivable that a racing task removed this page from
1454 * swap cache just before we acquired the page lock at the top,
1455 * or while we dropped it in unuse_mm(). The page might even
1456 * be back in swap cache on another swap area: that we must not
1457 * delete, since it may not have been written out to swap yet.
1459 if (PageSwapCache(page) &&
1460 likely(page_private(page) == entry.val))
1461 delete_from_swap_cache(page);
1464 * So we could skip searching mms once swap count went
1465 * to 1, we did not mark any present ptes as dirty: must
1466 * mark page dirty so shrink_page_list will preserve it.
1470 page_cache_release(page);
1473 * Make sure that we aren't completely killing
1474 * interactive performance.
1477 if (frontswap && pages_to_unuse > 0) {
1478 if (!--pages_to_unuse)
1488 * After a successful try_to_unuse, if no swap is now in use, we know
1489 * we can empty the mmlist. swap_lock must be held on entry and exit.
1490 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1491 * added to the mmlist just after page_duplicate - before would be racy.
1493 static void drain_mmlist(void)
1495 struct list_head *p, *next;
1498 for (type = 0; type < nr_swapfiles; type++)
1499 if (swap_info[type]->inuse_pages)
1501 spin_lock(&mmlist_lock);
1502 list_for_each_safe(p, next, &init_mm.mmlist)
1504 spin_unlock(&mmlist_lock);
1508 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1509 * corresponds to page offset for the specified swap entry.
1510 * Note that the type of this function is sector_t, but it returns page offset
1511 * into the bdev, not sector offset.
1513 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1515 struct swap_info_struct *sis;
1516 struct swap_extent *start_se;
1517 struct swap_extent *se;
1520 sis = swap_info[swp_type(entry)];
1523 offset = swp_offset(entry);
1524 start_se = sis->curr_swap_extent;
1528 struct list_head *lh;
1530 if (se->start_page <= offset &&
1531 offset < (se->start_page + se->nr_pages)) {
1532 return se->start_block + (offset - se->start_page);
1535 se = list_entry(lh, struct swap_extent, list);
1536 sis->curr_swap_extent = se;
1537 BUG_ON(se == start_se); /* It *must* be present */
1542 * Returns the page offset into bdev for the specified page's swap entry.
1544 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1547 entry.val = page_private(page);
1548 return map_swap_entry(entry, bdev);
1552 * Free all of a swapdev's extent information
1554 static void destroy_swap_extents(struct swap_info_struct *sis)
1556 while (!list_empty(&sis->first_swap_extent.list)) {
1557 struct swap_extent *se;
1559 se = list_entry(sis->first_swap_extent.list.next,
1560 struct swap_extent, list);
1561 list_del(&se->list);
1565 if (sis->flags & SWP_FILE) {
1566 struct file *swap_file = sis->swap_file;
1567 struct address_space *mapping = swap_file->f_mapping;
1569 sis->flags &= ~SWP_FILE;
1570 mapping->a_ops->swap_deactivate(swap_file);
1575 * Add a block range (and the corresponding page range) into this swapdev's
1576 * extent list. The extent list is kept sorted in page order.
1578 * This function rather assumes that it is called in ascending page order.
1581 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1582 unsigned long nr_pages, sector_t start_block)
1584 struct swap_extent *se;
1585 struct swap_extent *new_se;
1586 struct list_head *lh;
1588 if (start_page == 0) {
1589 se = &sis->first_swap_extent;
1590 sis->curr_swap_extent = se;
1592 se->nr_pages = nr_pages;
1593 se->start_block = start_block;
1596 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1597 se = list_entry(lh, struct swap_extent, list);
1598 BUG_ON(se->start_page + se->nr_pages != start_page);
1599 if (se->start_block + se->nr_pages == start_block) {
1601 se->nr_pages += nr_pages;
1607 * No merge. Insert a new extent, preserving ordering.
1609 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1612 new_se->start_page = start_page;
1613 new_se->nr_pages = nr_pages;
1614 new_se->start_block = start_block;
1616 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1621 * A `swap extent' is a simple thing which maps a contiguous range of pages
1622 * onto a contiguous range of disk blocks. An ordered list of swap extents
1623 * is built at swapon time and is then used at swap_writepage/swap_readpage
1624 * time for locating where on disk a page belongs.
1626 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1627 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1628 * swap files identically.
1630 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1631 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1632 * swapfiles are handled *identically* after swapon time.
1634 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1635 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1636 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1637 * requirements, they are simply tossed out - we will never use those blocks
1640 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1641 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1642 * which will scribble on the fs.
1644 * The amount of disk space which a single swap extent represents varies.
1645 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1646 * extents in the list. To avoid much list walking, we cache the previous
1647 * search location in `curr_swap_extent', and start new searches from there.
1648 * This is extremely effective. The average number of iterations in
1649 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1651 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1653 struct file *swap_file = sis->swap_file;
1654 struct address_space *mapping = swap_file->f_mapping;
1655 struct inode *inode = mapping->host;
1658 if (S_ISBLK(inode->i_mode)) {
1659 ret = add_swap_extent(sis, 0, sis->max, 0);
1664 if (mapping->a_ops->swap_activate) {
1665 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1667 sis->flags |= SWP_FILE;
1668 ret = add_swap_extent(sis, 0, sis->max, 0);
1674 return generic_swapfile_activate(sis, swap_file, span);
1677 static void _enable_swap_info(struct swap_info_struct *p, int prio,
1678 unsigned char *swap_map,
1679 struct swap_cluster_info *cluster_info)
1686 p->prio = --least_priority;
1687 p->swap_map = swap_map;
1688 p->cluster_info = cluster_info;
1689 p->flags |= SWP_WRITEOK;
1690 atomic_long_add(p->pages, &nr_swap_pages);
1691 total_swap_pages += p->pages;
1693 /* insert swap space into swap_list: */
1695 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1696 if (p->prio >= swap_info[i]->prio)
1702 swap_list.head = swap_list.next = p->type;
1704 swap_info[prev]->next = p->type;
1707 static void enable_swap_info(struct swap_info_struct *p, int prio,
1708 unsigned char *swap_map,
1709 struct swap_cluster_info *cluster_info,
1710 unsigned long *frontswap_map)
1712 frontswap_init(p->type, frontswap_map);
1713 spin_lock(&swap_lock);
1714 spin_lock(&p->lock);
1715 _enable_swap_info(p, prio, swap_map, cluster_info);
1716 spin_unlock(&p->lock);
1717 spin_unlock(&swap_lock);
1720 static void reinsert_swap_info(struct swap_info_struct *p)
1722 spin_lock(&swap_lock);
1723 spin_lock(&p->lock);
1724 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
1725 spin_unlock(&p->lock);
1726 spin_unlock(&swap_lock);
1729 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1731 struct swap_info_struct *p = NULL;
1732 unsigned char *swap_map;
1733 struct swap_cluster_info *cluster_info;
1734 unsigned long *frontswap_map;
1735 struct file *swap_file, *victim;
1736 struct address_space *mapping;
1737 struct inode *inode;
1738 struct filename *pathname;
1742 if (!capable(CAP_SYS_ADMIN))
1745 BUG_ON(!current->mm);
1747 pathname = getname(specialfile);
1748 if (IS_ERR(pathname))
1749 return PTR_ERR(pathname);
1751 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
1752 err = PTR_ERR(victim);
1756 mapping = victim->f_mapping;
1758 spin_lock(&swap_lock);
1759 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1760 p = swap_info[type];
1761 if (p->flags & SWP_WRITEOK) {
1762 if (p->swap_file->f_mapping == mapping)
1769 spin_unlock(&swap_lock);
1772 if (!security_vm_enough_memory_mm(current->mm, p->pages))
1773 vm_unacct_memory(p->pages);
1776 spin_unlock(&swap_lock);
1780 swap_list.head = p->next;
1782 swap_info[prev]->next = p->next;
1783 if (type == swap_list.next) {
1784 /* just pick something that's safe... */
1785 swap_list.next = swap_list.head;
1787 spin_lock(&p->lock);
1789 for (i = p->next; i >= 0; i = swap_info[i]->next)
1790 swap_info[i]->prio = p->prio--;
1793 atomic_long_sub(p->pages, &nr_swap_pages);
1794 total_swap_pages -= p->pages;
1795 p->flags &= ~SWP_WRITEOK;
1796 spin_unlock(&p->lock);
1797 spin_unlock(&swap_lock);
1799 set_current_oom_origin();
1800 err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1801 clear_current_oom_origin();
1804 /* re-insert swap space back into swap_list */
1805 reinsert_swap_info(p);
1809 destroy_swap_extents(p);
1810 if (p->flags & SWP_CONTINUED)
1811 free_swap_count_continuations(p);
1813 mutex_lock(&swapon_mutex);
1814 spin_lock(&swap_lock);
1815 spin_lock(&p->lock);
1818 /* wait for anyone still in scan_swap_map */
1819 p->highest_bit = 0; /* cuts scans short */
1820 while (p->flags >= SWP_SCANNING) {
1821 spin_unlock(&p->lock);
1822 spin_unlock(&swap_lock);
1823 schedule_timeout_uninterruptible(1);
1824 spin_lock(&swap_lock);
1825 spin_lock(&p->lock);
1828 swap_file = p->swap_file;
1829 p->swap_file = NULL;
1831 swap_map = p->swap_map;
1833 cluster_info = p->cluster_info;
1834 p->cluster_info = NULL;
1836 frontswap_map = frontswap_map_get(p);
1837 frontswap_map_set(p, NULL);
1838 spin_unlock(&p->lock);
1839 spin_unlock(&swap_lock);
1840 frontswap_invalidate_area(type);
1841 mutex_unlock(&swapon_mutex);
1843 vfree(cluster_info);
1844 vfree(frontswap_map);
1845 /* Destroy swap account informatin */
1846 swap_cgroup_swapoff(type);
1848 inode = mapping->host;
1849 if (S_ISBLK(inode->i_mode)) {
1850 struct block_device *bdev = I_BDEV(inode);
1851 set_blocksize(bdev, p->old_block_size);
1852 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1854 mutex_lock(&inode->i_mutex);
1855 inode->i_flags &= ~S_SWAPFILE;
1856 mutex_unlock(&inode->i_mutex);
1858 filp_close(swap_file, NULL);
1860 atomic_inc(&proc_poll_event);
1861 wake_up_interruptible(&proc_poll_wait);
1864 filp_close(victim, NULL);
1870 #ifdef CONFIG_PROC_FS
1871 static unsigned swaps_poll(struct file *file, poll_table *wait)
1873 struct seq_file *seq = file->private_data;
1875 poll_wait(file, &proc_poll_wait, wait);
1877 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1878 seq->poll_event = atomic_read(&proc_poll_event);
1879 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1882 return POLLIN | POLLRDNORM;
1886 static void *swap_start(struct seq_file *swap, loff_t *pos)
1888 struct swap_info_struct *si;
1892 mutex_lock(&swapon_mutex);
1895 return SEQ_START_TOKEN;
1897 for (type = 0; type < nr_swapfiles; type++) {
1898 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1899 si = swap_info[type];
1900 if (!(si->flags & SWP_USED) || !si->swap_map)
1909 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1911 struct swap_info_struct *si = v;
1914 if (v == SEQ_START_TOKEN)
1917 type = si->type + 1;
1919 for (; type < nr_swapfiles; type++) {
1920 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1921 si = swap_info[type];
1922 if (!(si->flags & SWP_USED) || !si->swap_map)
1931 static void swap_stop(struct seq_file *swap, void *v)
1933 mutex_unlock(&swapon_mutex);
1936 static int swap_show(struct seq_file *swap, void *v)
1938 struct swap_info_struct *si = v;
1942 if (si == SEQ_START_TOKEN) {
1943 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1947 file = si->swap_file;
1948 len = seq_path(swap, &file->f_path, " \t\n\\");
1949 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1950 len < 40 ? 40 - len : 1, " ",
1951 S_ISBLK(file_inode(file)->i_mode) ?
1952 "partition" : "file\t",
1953 si->pages << (PAGE_SHIFT - 10),
1954 si->inuse_pages << (PAGE_SHIFT - 10),
1959 static const struct seq_operations swaps_op = {
1960 .start = swap_start,
1966 static int swaps_open(struct inode *inode, struct file *file)
1968 struct seq_file *seq;
1971 ret = seq_open(file, &swaps_op);
1975 seq = file->private_data;
1976 seq->poll_event = atomic_read(&proc_poll_event);
1980 static const struct file_operations proc_swaps_operations = {
1983 .llseek = seq_lseek,
1984 .release = seq_release,
1988 static int __init procswaps_init(void)
1990 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1993 __initcall(procswaps_init);
1994 #endif /* CONFIG_PROC_FS */
1996 #ifdef MAX_SWAPFILES_CHECK
1997 static int __init max_swapfiles_check(void)
1999 MAX_SWAPFILES_CHECK();
2002 late_initcall(max_swapfiles_check);
2005 static struct swap_info_struct *alloc_swap_info(void)
2007 struct swap_info_struct *p;
2010 p = kzalloc(sizeof(*p), GFP_KERNEL);
2012 return ERR_PTR(-ENOMEM);
2014 spin_lock(&swap_lock);
2015 for (type = 0; type < nr_swapfiles; type++) {
2016 if (!(swap_info[type]->flags & SWP_USED))
2019 if (type >= MAX_SWAPFILES) {
2020 spin_unlock(&swap_lock);
2022 return ERR_PTR(-EPERM);
2024 if (type >= nr_swapfiles) {
2026 swap_info[type] = p;
2028 * Write swap_info[type] before nr_swapfiles, in case a
2029 * racing procfs swap_start() or swap_next() is reading them.
2030 * (We never shrink nr_swapfiles, we never free this entry.)
2036 p = swap_info[type];
2038 * Do not memset this entry: a racing procfs swap_next()
2039 * would be relying on p->type to remain valid.
2042 INIT_LIST_HEAD(&p->first_swap_extent.list);
2043 p->flags = SWP_USED;
2045 spin_unlock(&swap_lock);
2046 spin_lock_init(&p->lock);
2051 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2055 if (S_ISBLK(inode->i_mode)) {
2056 p->bdev = bdgrab(I_BDEV(inode));
2057 error = blkdev_get(p->bdev,
2058 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
2064 p->old_block_size = block_size(p->bdev);
2065 error = set_blocksize(p->bdev, PAGE_SIZE);
2068 p->flags |= SWP_BLKDEV;
2069 } else if (S_ISREG(inode->i_mode)) {
2070 p->bdev = inode->i_sb->s_bdev;
2071 mutex_lock(&inode->i_mutex);
2072 if (IS_SWAPFILE(inode))
2080 static unsigned long read_swap_header(struct swap_info_struct *p,
2081 union swap_header *swap_header,
2082 struct inode *inode)
2085 unsigned long maxpages;
2086 unsigned long swapfilepages;
2087 unsigned long last_page;
2089 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2090 pr_err("Unable to find swap-space signature\n");
2094 /* swap partition endianess hack... */
2095 if (swab32(swap_header->info.version) == 1) {
2096 swab32s(&swap_header->info.version);
2097 swab32s(&swap_header->info.last_page);
2098 swab32s(&swap_header->info.nr_badpages);
2099 for (i = 0; i < swap_header->info.nr_badpages; i++)
2100 swab32s(&swap_header->info.badpages[i]);
2102 /* Check the swap header's sub-version */
2103 if (swap_header->info.version != 1) {
2104 pr_warn("Unable to handle swap header version %d\n",
2105 swap_header->info.version);
2110 p->cluster_next = 1;
2114 * Find out how many pages are allowed for a single swap
2115 * device. There are two limiting factors: 1) the number
2116 * of bits for the swap offset in the swp_entry_t type, and
2117 * 2) the number of bits in the swap pte as defined by the
2118 * different architectures. In order to find the
2119 * largest possible bit mask, a swap entry with swap type 0
2120 * and swap offset ~0UL is created, encoded to a swap pte,
2121 * decoded to a swp_entry_t again, and finally the swap
2122 * offset is extracted. This will mask all the bits from
2123 * the initial ~0UL mask that can't be encoded in either
2124 * the swp_entry_t or the architecture definition of a
2127 maxpages = swp_offset(pte_to_swp_entry(
2128 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2129 last_page = swap_header->info.last_page;
2130 if (last_page > maxpages) {
2131 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2132 maxpages << (PAGE_SHIFT - 10),
2133 last_page << (PAGE_SHIFT - 10));
2135 if (maxpages > last_page) {
2136 maxpages = last_page + 1;
2137 /* p->max is an unsigned int: don't overflow it */
2138 if ((unsigned int)maxpages == 0)
2139 maxpages = UINT_MAX;
2141 p->highest_bit = maxpages - 1;
2145 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2146 if (swapfilepages && maxpages > swapfilepages) {
2147 pr_warn("Swap area shorter than signature indicates\n");
2150 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2152 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2158 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2159 union swap_header *swap_header,
2160 unsigned char *swap_map,
2161 struct swap_cluster_info *cluster_info,
2162 unsigned long maxpages,
2166 unsigned int nr_good_pages;
2168 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2169 unsigned long idx = p->cluster_next / SWAPFILE_CLUSTER;
2171 nr_good_pages = maxpages - 1; /* omit header page */
2173 cluster_set_null(&p->free_cluster_head);
2174 cluster_set_null(&p->free_cluster_tail);
2176 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2177 unsigned int page_nr = swap_header->info.badpages[i];
2178 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2180 if (page_nr < maxpages) {
2181 swap_map[page_nr] = SWAP_MAP_BAD;
2184 * Haven't marked the cluster free yet, no list
2185 * operation involved
2187 inc_cluster_info_page(p, cluster_info, page_nr);
2191 /* Haven't marked the cluster free yet, no list operation involved */
2192 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2193 inc_cluster_info_page(p, cluster_info, i);
2195 if (nr_good_pages) {
2196 swap_map[0] = SWAP_MAP_BAD;
2198 * Not mark the cluster free yet, no list
2199 * operation involved
2201 inc_cluster_info_page(p, cluster_info, 0);
2203 p->pages = nr_good_pages;
2204 nr_extents = setup_swap_extents(p, span);
2207 nr_good_pages = p->pages;
2209 if (!nr_good_pages) {
2210 pr_warn("Empty swap-file\n");
2217 for (i = 0; i < nr_clusters; i++) {
2218 if (!cluster_count(&cluster_info[idx])) {
2219 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2220 if (cluster_is_null(&p->free_cluster_head)) {
2221 cluster_set_next_flag(&p->free_cluster_head,
2223 cluster_set_next_flag(&p->free_cluster_tail,
2228 tail = cluster_next(&p->free_cluster_tail);
2229 cluster_set_next(&cluster_info[tail], idx);
2230 cluster_set_next_flag(&p->free_cluster_tail,
2235 if (idx == nr_clusters)
2242 * Helper to sys_swapon determining if a given swap
2243 * backing device queue supports DISCARD operations.
2245 static bool swap_discardable(struct swap_info_struct *si)
2247 struct request_queue *q = bdev_get_queue(si->bdev);
2249 if (!q || !blk_queue_discard(q))
2255 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2257 struct swap_info_struct *p;
2258 struct filename *name;
2259 struct file *swap_file = NULL;
2260 struct address_space *mapping;
2264 union swap_header *swap_header;
2267 unsigned long maxpages;
2268 unsigned char *swap_map = NULL;
2269 struct swap_cluster_info *cluster_info = NULL;
2270 unsigned long *frontswap_map = NULL;
2271 struct page *page = NULL;
2272 struct inode *inode = NULL;
2274 if (swap_flags & ~SWAP_FLAGS_VALID)
2277 if (!capable(CAP_SYS_ADMIN))
2280 p = alloc_swap_info();
2284 name = getname(specialfile);
2286 error = PTR_ERR(name);
2290 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2291 if (IS_ERR(swap_file)) {
2292 error = PTR_ERR(swap_file);
2297 p->swap_file = swap_file;
2298 mapping = swap_file->f_mapping;
2300 for (i = 0; i < nr_swapfiles; i++) {
2301 struct swap_info_struct *q = swap_info[i];
2303 if (q == p || !q->swap_file)
2305 if (mapping == q->swap_file->f_mapping) {
2311 inode = mapping->host;
2312 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2313 error = claim_swapfile(p, inode);
2314 if (unlikely(error))
2318 * Read the swap header.
2320 if (!mapping->a_ops->readpage) {
2324 page = read_mapping_page(mapping, 0, swap_file);
2326 error = PTR_ERR(page);
2329 swap_header = kmap(page);
2331 maxpages = read_swap_header(p, swap_header, inode);
2332 if (unlikely(!maxpages)) {
2337 /* OK, set up the swap map and apply the bad block list */
2338 swap_map = vzalloc(maxpages);
2343 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2344 p->flags |= SWP_SOLIDSTATE;
2346 * select a random position to start with to help wear leveling
2349 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2351 cluster_info = vzalloc(DIV_ROUND_UP(maxpages,
2352 SWAPFILE_CLUSTER) * sizeof(*cluster_info));
2353 if (!cluster_info) {
2359 error = swap_cgroup_swapon(p->type, maxpages);
2363 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2364 cluster_info, maxpages, &span);
2365 if (unlikely(nr_extents < 0)) {
2369 /* frontswap enabled? set up bit-per-page map for frontswap */
2370 if (frontswap_enabled)
2371 frontswap_map = vzalloc(BITS_TO_LONGS(maxpages) * sizeof(long));
2373 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2375 * When discard is enabled for swap with no particular
2376 * policy flagged, we set all swap discard flags here in
2377 * order to sustain backward compatibility with older
2378 * swapon(8) releases.
2380 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2384 * By flagging sys_swapon, a sysadmin can tell us to
2385 * either do single-time area discards only, or to just
2386 * perform discards for released swap page-clusters.
2387 * Now it's time to adjust the p->flags accordingly.
2389 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2390 p->flags &= ~SWP_PAGE_DISCARD;
2391 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2392 p->flags &= ~SWP_AREA_DISCARD;
2394 /* issue a swapon-time discard if it's still required */
2395 if (p->flags & SWP_AREA_DISCARD) {
2396 int err = discard_swap(p);
2398 pr_err("swapon: discard_swap(%p): %d\n",
2403 mutex_lock(&swapon_mutex);
2405 if (swap_flags & SWAP_FLAG_PREFER)
2407 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2408 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
2410 pr_info("Adding %uk swap on %s. "
2411 "Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
2412 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2413 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2414 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2415 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2416 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
2417 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
2418 (frontswap_map) ? "FS" : "");
2420 mutex_unlock(&swapon_mutex);
2421 atomic_inc(&proc_poll_event);
2422 wake_up_interruptible(&proc_poll_wait);
2424 if (S_ISREG(inode->i_mode))
2425 inode->i_flags |= S_SWAPFILE;
2429 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2430 set_blocksize(p->bdev, p->old_block_size);
2431 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2433 destroy_swap_extents(p);
2434 swap_cgroup_swapoff(p->type);
2435 spin_lock(&swap_lock);
2436 p->swap_file = NULL;
2438 spin_unlock(&swap_lock);
2440 vfree(cluster_info);
2442 if (inode && S_ISREG(inode->i_mode)) {
2443 mutex_unlock(&inode->i_mutex);
2446 filp_close(swap_file, NULL);
2449 if (page && !IS_ERR(page)) {
2451 page_cache_release(page);
2455 if (inode && S_ISREG(inode->i_mode))
2456 mutex_unlock(&inode->i_mutex);
2460 void si_swapinfo(struct sysinfo *val)
2463 unsigned long nr_to_be_unused = 0;
2465 spin_lock(&swap_lock);
2466 for (type = 0; type < nr_swapfiles; type++) {
2467 struct swap_info_struct *si = swap_info[type];
2469 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2470 nr_to_be_unused += si->inuse_pages;
2472 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
2473 val->totalswap = total_swap_pages + nr_to_be_unused;
2474 spin_unlock(&swap_lock);
2478 * Verify that a swap entry is valid and increment its swap map count.
2480 * Returns error code in following case.
2482 * - swp_entry is invalid -> EINVAL
2483 * - swp_entry is migration entry -> EINVAL
2484 * - swap-cache reference is requested but there is already one. -> EEXIST
2485 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2486 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2488 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2490 struct swap_info_struct *p;
2491 unsigned long offset, type;
2492 unsigned char count;
2493 unsigned char has_cache;
2496 if (non_swap_entry(entry))
2499 type = swp_type(entry);
2500 if (type >= nr_swapfiles)
2502 p = swap_info[type];
2503 offset = swp_offset(entry);
2505 spin_lock(&p->lock);
2506 if (unlikely(offset >= p->max))
2509 count = p->swap_map[offset];
2510 has_cache = count & SWAP_HAS_CACHE;
2511 count &= ~SWAP_HAS_CACHE;
2514 if (usage == SWAP_HAS_CACHE) {
2516 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2517 if (!has_cache && count)
2518 has_cache = SWAP_HAS_CACHE;
2519 else if (has_cache) /* someone else added cache */
2521 else /* no users remaining */
2524 } else if (count || has_cache) {
2526 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2528 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2530 else if (swap_count_continued(p, offset, count))
2531 count = COUNT_CONTINUED;
2535 err = -ENOENT; /* unused swap entry */
2537 p->swap_map[offset] = count | has_cache;
2540 spin_unlock(&p->lock);
2545 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
2550 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2551 * (in which case its reference count is never incremented).
2553 void swap_shmem_alloc(swp_entry_t entry)
2555 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2559 * Increase reference count of swap entry by 1.
2560 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2561 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2562 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2563 * might occur if a page table entry has got corrupted.
2565 int swap_duplicate(swp_entry_t entry)
2569 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2570 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2575 * @entry: swap entry for which we allocate swap cache.
2577 * Called when allocating swap cache for existing swap entry,
2578 * This can return error codes. Returns 0 at success.
2579 * -EBUSY means there is a swap cache.
2580 * Note: return code is different from swap_duplicate().
2582 int swapcache_prepare(swp_entry_t entry)
2584 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2587 struct swap_info_struct *page_swap_info(struct page *page)
2589 swp_entry_t swap = { .val = page_private(page) };
2590 BUG_ON(!PageSwapCache(page));
2591 return swap_info[swp_type(swap)];
2595 * out-of-line __page_file_ methods to avoid include hell.
2597 struct address_space *__page_file_mapping(struct page *page)
2599 VM_BUG_ON(!PageSwapCache(page));
2600 return page_swap_info(page)->swap_file->f_mapping;
2602 EXPORT_SYMBOL_GPL(__page_file_mapping);
2604 pgoff_t __page_file_index(struct page *page)
2606 swp_entry_t swap = { .val = page_private(page) };
2607 VM_BUG_ON(!PageSwapCache(page));
2608 return swp_offset(swap);
2610 EXPORT_SYMBOL_GPL(__page_file_index);
2613 * add_swap_count_continuation - called when a swap count is duplicated
2614 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2615 * page of the original vmalloc'ed swap_map, to hold the continuation count
2616 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2617 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2619 * These continuation pages are seldom referenced: the common paths all work
2620 * on the original swap_map, only referring to a continuation page when the
2621 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2623 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2624 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2625 * can be called after dropping locks.
2627 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2629 struct swap_info_struct *si;
2632 struct page *list_page;
2634 unsigned char count;
2637 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2638 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2640 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2642 si = swap_info_get(entry);
2645 * An acceptable race has occurred since the failing
2646 * __swap_duplicate(): the swap entry has been freed,
2647 * perhaps even the whole swap_map cleared for swapoff.
2652 offset = swp_offset(entry);
2653 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2655 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2657 * The higher the swap count, the more likely it is that tasks
2658 * will race to add swap count continuation: we need to avoid
2659 * over-provisioning.
2665 spin_unlock(&si->lock);
2670 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2671 * no architecture is using highmem pages for kernel pagetables: so it
2672 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2674 head = vmalloc_to_page(si->swap_map + offset);
2675 offset &= ~PAGE_MASK;
2678 * Page allocation does not initialize the page's lru field,
2679 * but it does always reset its private field.
2681 if (!page_private(head)) {
2682 BUG_ON(count & COUNT_CONTINUED);
2683 INIT_LIST_HEAD(&head->lru);
2684 set_page_private(head, SWP_CONTINUED);
2685 si->flags |= SWP_CONTINUED;
2688 list_for_each_entry(list_page, &head->lru, lru) {
2692 * If the previous map said no continuation, but we've found
2693 * a continuation page, free our allocation and use this one.
2695 if (!(count & COUNT_CONTINUED))
2698 map = kmap_atomic(list_page) + offset;
2703 * If this continuation count now has some space in it,
2704 * free our allocation and use this one.
2706 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2710 list_add_tail(&page->lru, &head->lru);
2711 page = NULL; /* now it's attached, don't free it */
2713 spin_unlock(&si->lock);
2721 * swap_count_continued - when the original swap_map count is incremented
2722 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2723 * into, carry if so, or else fail until a new continuation page is allocated;
2724 * when the original swap_map count is decremented from 0 with continuation,
2725 * borrow from the continuation and report whether it still holds more.
2726 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2728 static bool swap_count_continued(struct swap_info_struct *si,
2729 pgoff_t offset, unsigned char count)
2735 head = vmalloc_to_page(si->swap_map + offset);
2736 if (page_private(head) != SWP_CONTINUED) {
2737 BUG_ON(count & COUNT_CONTINUED);
2738 return false; /* need to add count continuation */
2741 offset &= ~PAGE_MASK;
2742 page = list_entry(head->lru.next, struct page, lru);
2743 map = kmap_atomic(page) + offset;
2745 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2746 goto init_map; /* jump over SWAP_CONT_MAX checks */
2748 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2750 * Think of how you add 1 to 999
2752 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2754 page = list_entry(page->lru.next, struct page, lru);
2755 BUG_ON(page == head);
2756 map = kmap_atomic(page) + offset;
2758 if (*map == SWAP_CONT_MAX) {
2760 page = list_entry(page->lru.next, struct page, lru);
2762 return false; /* add count continuation */
2763 map = kmap_atomic(page) + offset;
2764 init_map: *map = 0; /* we didn't zero the page */
2768 page = list_entry(page->lru.prev, struct page, lru);
2769 while (page != head) {
2770 map = kmap_atomic(page) + offset;
2771 *map = COUNT_CONTINUED;
2773 page = list_entry(page->lru.prev, struct page, lru);
2775 return true; /* incremented */
2777 } else { /* decrementing */
2779 * Think of how you subtract 1 from 1000
2781 BUG_ON(count != COUNT_CONTINUED);
2782 while (*map == COUNT_CONTINUED) {
2784 page = list_entry(page->lru.next, struct page, lru);
2785 BUG_ON(page == head);
2786 map = kmap_atomic(page) + offset;
2793 page = list_entry(page->lru.prev, struct page, lru);
2794 while (page != head) {
2795 map = kmap_atomic(page) + offset;
2796 *map = SWAP_CONT_MAX | count;
2797 count = COUNT_CONTINUED;
2799 page = list_entry(page->lru.prev, struct page, lru);
2801 return count == COUNT_CONTINUED;
2806 * free_swap_count_continuations - swapoff free all the continuation pages
2807 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2809 static void free_swap_count_continuations(struct swap_info_struct *si)
2813 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2815 head = vmalloc_to_page(si->swap_map + offset);
2816 if (page_private(head)) {
2817 struct list_head *this, *next;
2818 list_for_each_safe(this, next, &head->lru) {
2820 page = list_entry(this, struct page, lru);