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/shm.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/module.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/security.h>
28 #include <linux/backing-dev.h>
29 #include <linux/mutex.h>
30 #include <linux/capability.h>
31 #include <linux/syscalls.h>
32 #include <linux/memcontrol.h>
33 #include <linux/poll.h>
35 #include <asm/pgtable.h>
36 #include <asm/tlbflush.h>
37 #include <linux/swapops.h>
38 #include <linux/page_cgroup.h>
40 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
42 static void free_swap_count_continuations(struct swap_info_struct *);
43 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
45 static DEFINE_SPINLOCK(swap_lock);
46 static unsigned int nr_swapfiles;
48 long total_swap_pages;
49 static int least_priority;
51 static const char Bad_file[] = "Bad swap file entry ";
52 static const char Unused_file[] = "Unused swap file entry ";
53 static const char Bad_offset[] = "Bad swap offset entry ";
54 static const char Unused_offset[] = "Unused swap offset entry ";
56 static struct swap_list_t swap_list = {-1, -1};
58 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
60 static DEFINE_MUTEX(swapon_mutex);
62 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
63 /* Activity counter to indicate that a swapon or swapoff has occurred */
64 static atomic_t proc_poll_event = ATOMIC_INIT(0);
66 static inline unsigned char swap_count(unsigned char ent)
68 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
71 /* returns 1 if swap entry is freed */
73 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
75 swp_entry_t entry = swp_entry(si->type, offset);
79 page = find_get_page(&swapper_space, entry.val);
83 * This function is called from scan_swap_map() and it's called
84 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
85 * We have to use trylock for avoiding deadlock. This is a special
86 * case and you should use try_to_free_swap() with explicit lock_page()
87 * in usual operations.
89 if (trylock_page(page)) {
90 ret = try_to_free_swap(page);
93 page_cache_release(page);
98 * We need this because the bdev->unplug_fn can sleep and we cannot
99 * hold swap_lock while calling the unplug_fn. And swap_lock
100 * cannot be turned into a mutex.
102 static DECLARE_RWSEM(swap_unplug_sem);
104 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
108 down_read(&swap_unplug_sem);
109 entry.val = page_private(page);
110 if (PageSwapCache(page)) {
111 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
112 struct backing_dev_info *bdi;
115 * If the page is removed from swapcache from under us (with a
116 * racy try_to_unuse/swapoff) we need an additional reference
117 * count to avoid reading garbage from page_private(page) above.
118 * If the WARN_ON triggers during a swapoff it maybe the race
119 * condition and it's harmless. However if it triggers without
120 * swapoff it signals a problem.
122 WARN_ON(page_count(page) <= 1);
124 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
125 blk_run_backing_dev(bdi, page);
127 up_read(&swap_unplug_sem);
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 struct list_head *lh;
180 if (se->start_page <= start_page &&
181 start_page < se->start_page + se->nr_pages) {
182 pgoff_t offset = start_page - se->start_page;
183 sector_t start_block = se->start_block + offset;
184 sector_t nr_blocks = se->nr_pages - offset;
186 if (nr_blocks > nr_pages)
187 nr_blocks = nr_pages;
188 start_page += nr_blocks;
189 nr_pages -= nr_blocks;
192 si->curr_swap_extent = se;
194 start_block <<= PAGE_SHIFT - 9;
195 nr_blocks <<= PAGE_SHIFT - 9;
196 if (blkdev_issue_discard(si->bdev, start_block,
197 nr_blocks, GFP_NOIO, 0))
202 se = list_entry(lh, struct swap_extent, list);
206 static int wait_for_discard(void *word)
212 #define SWAPFILE_CLUSTER 256
213 #define LATENCY_LIMIT 256
215 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
218 unsigned long offset;
219 unsigned long scan_base;
220 unsigned long last_in_cluster = 0;
221 int latency_ration = LATENCY_LIMIT;
222 int found_free_cluster = 0;
225 * We try to cluster swap pages by allocating them sequentially
226 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
227 * way, however, we resort to first-free allocation, starting
228 * a new cluster. This prevents us from scattering swap pages
229 * all over the entire swap partition, so that we reduce
230 * overall disk seek times between swap pages. -- sct
231 * But we do now try to find an empty cluster. -Andrea
232 * And we let swap pages go all over an SSD partition. Hugh
235 si->flags += SWP_SCANNING;
236 scan_base = offset = si->cluster_next;
238 if (unlikely(!si->cluster_nr--)) {
239 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
240 si->cluster_nr = SWAPFILE_CLUSTER - 1;
243 if (si->flags & SWP_DISCARDABLE) {
245 * Start range check on racing allocations, in case
246 * they overlap the cluster we eventually decide on
247 * (we scan without swap_lock to allow preemption).
248 * It's hardly conceivable that cluster_nr could be
249 * wrapped during our scan, but don't depend on it.
251 if (si->lowest_alloc)
253 si->lowest_alloc = si->max;
254 si->highest_alloc = 0;
256 spin_unlock(&swap_lock);
259 * If seek is expensive, start searching for new cluster from
260 * start of partition, to minimize the span of allocated swap.
261 * But if seek is cheap, search from our current position, so
262 * that swap is allocated from all over the partition: if the
263 * Flash Translation Layer only remaps within limited zones,
264 * we don't want to wear out the first zone too quickly.
266 if (!(si->flags & SWP_SOLIDSTATE))
267 scan_base = offset = si->lowest_bit;
268 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
270 /* Locate the first empty (unaligned) cluster */
271 for (; last_in_cluster <= si->highest_bit; offset++) {
272 if (si->swap_map[offset])
273 last_in_cluster = offset + SWAPFILE_CLUSTER;
274 else if (offset == last_in_cluster) {
275 spin_lock(&swap_lock);
276 offset -= SWAPFILE_CLUSTER - 1;
277 si->cluster_next = offset;
278 si->cluster_nr = SWAPFILE_CLUSTER - 1;
279 found_free_cluster = 1;
282 if (unlikely(--latency_ration < 0)) {
284 latency_ration = LATENCY_LIMIT;
288 offset = si->lowest_bit;
289 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
291 /* Locate the first empty (unaligned) cluster */
292 for (; last_in_cluster < scan_base; offset++) {
293 if (si->swap_map[offset])
294 last_in_cluster = offset + SWAPFILE_CLUSTER;
295 else if (offset == last_in_cluster) {
296 spin_lock(&swap_lock);
297 offset -= SWAPFILE_CLUSTER - 1;
298 si->cluster_next = offset;
299 si->cluster_nr = SWAPFILE_CLUSTER - 1;
300 found_free_cluster = 1;
303 if (unlikely(--latency_ration < 0)) {
305 latency_ration = LATENCY_LIMIT;
310 spin_lock(&swap_lock);
311 si->cluster_nr = SWAPFILE_CLUSTER - 1;
312 si->lowest_alloc = 0;
316 if (!(si->flags & SWP_WRITEOK))
318 if (!si->highest_bit)
320 if (offset > si->highest_bit)
321 scan_base = offset = si->lowest_bit;
323 /* reuse swap entry of cache-only swap if not busy. */
324 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
326 spin_unlock(&swap_lock);
327 swap_was_freed = __try_to_reclaim_swap(si, offset);
328 spin_lock(&swap_lock);
329 /* entry was freed successfully, try to use this again */
332 goto scan; /* check next one */
335 if (si->swap_map[offset])
338 if (offset == si->lowest_bit)
340 if (offset == si->highest_bit)
343 if (si->inuse_pages == si->pages) {
344 si->lowest_bit = si->max;
347 si->swap_map[offset] = usage;
348 si->cluster_next = offset + 1;
349 si->flags -= SWP_SCANNING;
351 if (si->lowest_alloc) {
353 * Only set when SWP_DISCARDABLE, and there's a scan
354 * for a free cluster in progress or just completed.
356 if (found_free_cluster) {
358 * To optimize wear-levelling, discard the
359 * old data of the cluster, taking care not to
360 * discard any of its pages that have already
361 * been allocated by racing tasks (offset has
362 * already stepped over any at the beginning).
364 if (offset < si->highest_alloc &&
365 si->lowest_alloc <= last_in_cluster)
366 last_in_cluster = si->lowest_alloc - 1;
367 si->flags |= SWP_DISCARDING;
368 spin_unlock(&swap_lock);
370 if (offset < last_in_cluster)
371 discard_swap_cluster(si, offset,
372 last_in_cluster - offset + 1);
374 spin_lock(&swap_lock);
375 si->lowest_alloc = 0;
376 si->flags &= ~SWP_DISCARDING;
378 smp_mb(); /* wake_up_bit advises this */
379 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
381 } else if (si->flags & SWP_DISCARDING) {
383 * Delay using pages allocated by racing tasks
384 * until the whole discard has been issued. We
385 * could defer that delay until swap_writepage,
386 * but it's easier to keep this self-contained.
388 spin_unlock(&swap_lock);
389 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
390 wait_for_discard, TASK_UNINTERRUPTIBLE);
391 spin_lock(&swap_lock);
394 * Note pages allocated by racing tasks while
395 * scan for a free cluster is in progress, so
396 * that its final discard can exclude them.
398 if (offset < si->lowest_alloc)
399 si->lowest_alloc = offset;
400 if (offset > si->highest_alloc)
401 si->highest_alloc = offset;
407 spin_unlock(&swap_lock);
408 while (++offset <= si->highest_bit) {
409 if (!si->swap_map[offset]) {
410 spin_lock(&swap_lock);
413 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
414 spin_lock(&swap_lock);
417 if (unlikely(--latency_ration < 0)) {
419 latency_ration = LATENCY_LIMIT;
422 offset = si->lowest_bit;
423 while (++offset < scan_base) {
424 if (!si->swap_map[offset]) {
425 spin_lock(&swap_lock);
428 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
429 spin_lock(&swap_lock);
432 if (unlikely(--latency_ration < 0)) {
434 latency_ration = LATENCY_LIMIT;
437 spin_lock(&swap_lock);
440 si->flags -= SWP_SCANNING;
444 swp_entry_t get_swap_page(void)
446 struct swap_info_struct *si;
451 spin_lock(&swap_lock);
452 if (nr_swap_pages <= 0)
456 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
457 si = swap_info[type];
460 (!wrapped && si->prio != swap_info[next]->prio)) {
461 next = swap_list.head;
465 if (!si->highest_bit)
467 if (!(si->flags & SWP_WRITEOK))
470 swap_list.next = next;
471 /* This is called for allocating swap entry for cache */
472 offset = scan_swap_map(si, SWAP_HAS_CACHE);
474 spin_unlock(&swap_lock);
475 return swp_entry(type, offset);
477 next = swap_list.next;
482 spin_unlock(&swap_lock);
483 return (swp_entry_t) {0};
486 /* The only caller of this function is now susupend routine */
487 swp_entry_t get_swap_page_of_type(int type)
489 struct swap_info_struct *si;
492 spin_lock(&swap_lock);
493 si = swap_info[type];
494 if (si && (si->flags & SWP_WRITEOK)) {
496 /* This is called for allocating swap entry, not cache */
497 offset = scan_swap_map(si, 1);
499 spin_unlock(&swap_lock);
500 return swp_entry(type, offset);
504 spin_unlock(&swap_lock);
505 return (swp_entry_t) {0};
508 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
510 struct swap_info_struct *p;
511 unsigned long offset, type;
515 type = swp_type(entry);
516 if (type >= nr_swapfiles)
519 if (!(p->flags & SWP_USED))
521 offset = swp_offset(entry);
522 if (offset >= p->max)
524 if (!p->swap_map[offset])
526 spin_lock(&swap_lock);
530 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
533 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
536 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
539 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
544 static unsigned char swap_entry_free(struct swap_info_struct *p,
545 swp_entry_t entry, unsigned char usage)
547 unsigned long offset = swp_offset(entry);
549 unsigned char has_cache;
551 count = p->swap_map[offset];
552 has_cache = count & SWAP_HAS_CACHE;
553 count &= ~SWAP_HAS_CACHE;
555 if (usage == SWAP_HAS_CACHE) {
556 VM_BUG_ON(!has_cache);
558 } else if (count == SWAP_MAP_SHMEM) {
560 * Or we could insist on shmem.c using a special
561 * swap_shmem_free() and free_shmem_swap_and_cache()...
564 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
565 if (count == COUNT_CONTINUED) {
566 if (swap_count_continued(p, offset, count))
567 count = SWAP_MAP_MAX | COUNT_CONTINUED;
569 count = SWAP_MAP_MAX;
575 mem_cgroup_uncharge_swap(entry);
577 usage = count | has_cache;
578 p->swap_map[offset] = usage;
580 /* free if no reference */
582 struct gendisk *disk = p->bdev->bd_disk;
583 if (offset < p->lowest_bit)
584 p->lowest_bit = offset;
585 if (offset > p->highest_bit)
586 p->highest_bit = offset;
587 if (swap_list.next >= 0 &&
588 p->prio > swap_info[swap_list.next]->prio)
589 swap_list.next = p->type;
592 if ((p->flags & SWP_BLKDEV) &&
593 disk->fops->swap_slot_free_notify)
594 disk->fops->swap_slot_free_notify(p->bdev, offset);
597 if (!swap_count(count)) {
598 mem_cgroup_uncharge_swap(entry);
599 if (p->notify_swap_entry_free_fn)
600 p->notify_swap_entry_free_fn(offset);
607 * Caller has made sure that the swapdevice corresponding to entry
608 * is still around or has not been recycled.
610 void swap_free(swp_entry_t entry)
612 struct swap_info_struct *p;
614 p = swap_info_get(entry);
616 swap_entry_free(p, entry, 1);
617 spin_unlock(&swap_lock);
622 * Called after dropping swapcache to decrease refcnt to swap entries.
624 void swapcache_free(swp_entry_t entry, struct page *page)
626 struct swap_info_struct *p;
629 p = swap_info_get(entry);
631 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
633 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
634 spin_unlock(&swap_lock);
639 * How many references to page are currently swapped out?
640 * This does not give an exact answer when swap count is continued,
641 * but does include the high COUNT_CONTINUED flag to allow for that.
643 static inline int page_swapcount(struct page *page)
646 struct swap_info_struct *p;
649 entry.val = page_private(page);
650 p = swap_info_get(entry);
652 count = swap_count(p->swap_map[swp_offset(entry)]);
653 spin_unlock(&swap_lock);
659 * We can write to an anon page without COW if there are no other references
660 * to it. And as a side-effect, free up its swap: because the old content
661 * on disk will never be read, and seeking back there to write new content
662 * later would only waste time away from clustering.
664 int reuse_swap_page(struct page *page)
668 VM_BUG_ON(!PageLocked(page));
669 if (unlikely(PageKsm(page)))
671 count = page_mapcount(page);
672 if (count <= 1 && PageSwapCache(page)) {
673 count += page_swapcount(page);
674 if (count == 1 && !PageWriteback(page)) {
675 delete_from_swap_cache(page);
683 * If swap is getting full, or if there are no more mappings of this page,
684 * then try_to_free_swap is called to free its swap space.
686 int try_to_free_swap(struct page *page)
688 VM_BUG_ON(!PageLocked(page));
690 if (!PageSwapCache(page))
692 if (PageWriteback(page))
694 if (page_swapcount(page))
698 * Once hibernation has begun to create its image of memory,
699 * there's a danger that one of the calls to try_to_free_swap()
700 * - most probably a call from __try_to_reclaim_swap() while
701 * hibernation is allocating its own swap pages for the image,
702 * but conceivably even a call from memory reclaim - will free
703 * the swap from a page which has already been recorded in the
704 * image as a clean swapcache page, and then reuse its swap for
705 * another page of the image. On waking from hibernation, the
706 * original page might be freed under memory pressure, then
707 * later read back in from swap, now with the wrong data.
709 * Hibernation clears bits from gfp_allowed_mask to prevent
710 * memory reclaim from writing to disk, so check that here.
712 if (!(gfp_allowed_mask & __GFP_IO))
715 delete_from_swap_cache(page);
721 * Free the swap entry like above, but also try to
722 * free the page cache entry if it is the last user.
724 int free_swap_and_cache(swp_entry_t entry)
726 struct swap_info_struct *p;
727 struct page *page = NULL;
729 if (non_swap_entry(entry))
732 p = swap_info_get(entry);
734 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
735 page = find_get_page(&swapper_space, entry.val);
736 if (page && !trylock_page(page)) {
737 page_cache_release(page);
741 spin_unlock(&swap_lock);
745 * Not mapped elsewhere, or swap space full? Free it!
746 * Also recheck PageSwapCache now page is locked (above).
748 if (PageSwapCache(page) && !PageWriteback(page) &&
749 (!page_mapped(page) || vm_swap_full())) {
750 delete_from_swap_cache(page);
754 page_cache_release(page);
759 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
761 * mem_cgroup_count_swap_user - count the user of a swap entry
762 * @ent: the swap entry to be checked
763 * @pagep: the pointer for the swap cache page of the entry to be stored
765 * Returns the number of the user of the swap entry. The number is valid only
766 * for swaps of anonymous pages.
767 * If the entry is found on swap cache, the page is stored to pagep with
768 * refcount of it being incremented.
770 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
773 struct swap_info_struct *p;
776 page = find_get_page(&swapper_space, ent.val);
778 count += page_mapcount(page);
779 p = swap_info_get(ent);
781 count += swap_count(p->swap_map[swp_offset(ent)]);
782 spin_unlock(&swap_lock);
790 #ifdef CONFIG_HIBERNATION
792 * Find the swap type that corresponds to given device (if any).
794 * @offset - number of the PAGE_SIZE-sized block of the device, starting
795 * from 0, in which the swap header is expected to be located.
797 * This is needed for the suspend to disk (aka swsusp).
799 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
801 struct block_device *bdev = NULL;
805 bdev = bdget(device);
807 spin_lock(&swap_lock);
808 for (type = 0; type < nr_swapfiles; type++) {
809 struct swap_info_struct *sis = swap_info[type];
811 if (!(sis->flags & SWP_WRITEOK))
816 *bdev_p = bdgrab(sis->bdev);
818 spin_unlock(&swap_lock);
821 if (bdev == sis->bdev) {
822 struct swap_extent *se = &sis->first_swap_extent;
824 if (se->start_block == offset) {
826 *bdev_p = bdgrab(sis->bdev);
828 spin_unlock(&swap_lock);
834 spin_unlock(&swap_lock);
842 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
843 * corresponding to given index in swap_info (swap type).
845 sector_t swapdev_block(int type, pgoff_t offset)
847 struct block_device *bdev;
849 if ((unsigned int)type >= nr_swapfiles)
851 if (!(swap_info[type]->flags & SWP_WRITEOK))
853 return map_swap_entry(swp_entry(type, offset), &bdev);
857 * Return either the total number of swap pages of given type, or the number
858 * of free pages of that type (depending on @free)
860 * This is needed for software suspend
862 unsigned int count_swap_pages(int type, int free)
866 spin_lock(&swap_lock);
867 if ((unsigned int)type < nr_swapfiles) {
868 struct swap_info_struct *sis = swap_info[type];
870 if (sis->flags & SWP_WRITEOK) {
873 n -= sis->inuse_pages;
876 spin_unlock(&swap_lock);
879 #endif /* CONFIG_HIBERNATION */
882 * No need to decide whether this PTE shares the swap entry with others,
883 * just let do_wp_page work it out if a write is requested later - to
884 * force COW, vm_page_prot omits write permission from any private vma.
886 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
887 unsigned long addr, swp_entry_t entry, struct page *page)
889 struct mem_cgroup *ptr = NULL;
894 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
899 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
900 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
902 mem_cgroup_cancel_charge_swapin(ptr);
907 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
908 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
910 set_pte_at(vma->vm_mm, addr, pte,
911 pte_mkold(mk_pte(page, vma->vm_page_prot)));
912 page_add_anon_rmap(page, vma, addr);
913 mem_cgroup_commit_charge_swapin(page, ptr);
916 * Move the page to the active list so it is not
917 * immediately swapped out again after swapon.
921 pte_unmap_unlock(pte, ptl);
926 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
927 unsigned long addr, unsigned long end,
928 swp_entry_t entry, struct page *page)
930 pte_t swp_pte = swp_entry_to_pte(entry);
935 * We don't actually need pte lock while scanning for swp_pte: since
936 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
937 * page table while we're scanning; though it could get zapped, and on
938 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
939 * of unmatched parts which look like swp_pte, so unuse_pte must
940 * recheck under pte lock. Scanning without pte lock lets it be
941 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
943 pte = pte_offset_map(pmd, addr);
946 * swapoff spends a _lot_ of time in this loop!
947 * Test inline before going to call unuse_pte.
949 if (unlikely(pte_same(*pte, swp_pte))) {
951 ret = unuse_pte(vma, pmd, addr, entry, page);
954 pte = pte_offset_map(pmd, addr);
956 } while (pte++, addr += PAGE_SIZE, addr != end);
962 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
963 unsigned long addr, unsigned long end,
964 swp_entry_t entry, struct page *page)
970 pmd = pmd_offset(pud, addr);
972 next = pmd_addr_end(addr, end);
973 if (unlikely(pmd_trans_huge(*pmd)))
975 if (pmd_none_or_clear_bad(pmd))
977 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
980 } while (pmd++, addr = next, addr != end);
984 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
985 unsigned long addr, unsigned long end,
986 swp_entry_t entry, struct page *page)
992 pud = pud_offset(pgd, addr);
994 next = pud_addr_end(addr, end);
995 if (pud_none_or_clear_bad(pud))
997 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1000 } while (pud++, addr = next, addr != end);
1004 static int unuse_vma(struct vm_area_struct *vma,
1005 swp_entry_t entry, struct page *page)
1008 unsigned long addr, end, next;
1011 if (page_anon_vma(page)) {
1012 addr = page_address_in_vma(page, vma);
1013 if (addr == -EFAULT)
1016 end = addr + PAGE_SIZE;
1018 addr = vma->vm_start;
1022 pgd = pgd_offset(vma->vm_mm, addr);
1024 next = pgd_addr_end(addr, end);
1025 if (pgd_none_or_clear_bad(pgd))
1027 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1030 } while (pgd++, addr = next, addr != end);
1034 static int unuse_mm(struct mm_struct *mm,
1035 swp_entry_t entry, struct page *page)
1037 struct vm_area_struct *vma;
1040 if (!down_read_trylock(&mm->mmap_sem)) {
1042 * Activate page so shrink_inactive_list is unlikely to unmap
1043 * its ptes while lock is dropped, so swapoff can make progress.
1045 activate_page(page);
1047 down_read(&mm->mmap_sem);
1050 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1051 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1054 up_read(&mm->mmap_sem);
1055 return (ret < 0)? ret: 0;
1059 * Scan swap_map from current position to next entry still in use.
1060 * Recycle to start on reaching the end, returning 0 when empty.
1062 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1065 unsigned int max = si->max;
1066 unsigned int i = prev;
1067 unsigned char count;
1070 * No need for swap_lock here: we're just looking
1071 * for whether an entry is in use, not modifying it; false
1072 * hits are okay, and sys_swapoff() has already prevented new
1073 * allocations from this area (while holding swap_lock).
1082 * No entries in use at top of swap_map,
1083 * loop back to start and recheck there.
1089 count = si->swap_map[i];
1090 if (count && swap_count(count) != SWAP_MAP_BAD)
1097 * We completely avoid races by reading each swap page in advance,
1098 * and then search for the process using it. All the necessary
1099 * page table adjustments can then be made atomically.
1101 static int try_to_unuse(unsigned int type)
1103 struct swap_info_struct *si = swap_info[type];
1104 struct mm_struct *start_mm;
1105 unsigned char *swap_map;
1106 unsigned char swcount;
1113 * When searching mms for an entry, a good strategy is to
1114 * start at the first mm we freed the previous entry from
1115 * (though actually we don't notice whether we or coincidence
1116 * freed the entry). Initialize this start_mm with a hold.
1118 * A simpler strategy would be to start at the last mm we
1119 * freed the previous entry from; but that would take less
1120 * advantage of mmlist ordering, which clusters forked mms
1121 * together, child after parent. If we race with dup_mmap(), we
1122 * prefer to resolve parent before child, lest we miss entries
1123 * duplicated after we scanned child: using last mm would invert
1126 start_mm = &init_mm;
1127 atomic_inc(&init_mm.mm_users);
1130 * Keep on scanning until all entries have gone. Usually,
1131 * one pass through swap_map is enough, but not necessarily:
1132 * there are races when an instance of an entry might be missed.
1134 while ((i = find_next_to_unuse(si, i)) != 0) {
1135 if (signal_pending(current)) {
1141 * Get a page for the entry, using the existing swap
1142 * cache page if there is one. Otherwise, get a clean
1143 * page and read the swap into it.
1145 swap_map = &si->swap_map[i];
1146 entry = swp_entry(type, i);
1147 page = read_swap_cache_async(entry,
1148 GFP_HIGHUSER_MOVABLE, NULL, 0);
1151 * Either swap_duplicate() failed because entry
1152 * has been freed independently, and will not be
1153 * reused since sys_swapoff() already disabled
1154 * allocation from here, or alloc_page() failed.
1163 * Don't hold on to start_mm if it looks like exiting.
1165 if (atomic_read(&start_mm->mm_users) == 1) {
1167 start_mm = &init_mm;
1168 atomic_inc(&init_mm.mm_users);
1172 * Wait for and lock page. When do_swap_page races with
1173 * try_to_unuse, do_swap_page can handle the fault much
1174 * faster than try_to_unuse can locate the entry. This
1175 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1176 * defer to do_swap_page in such a case - in some tests,
1177 * do_swap_page and try_to_unuse repeatedly compete.
1179 wait_on_page_locked(page);
1180 wait_on_page_writeback(page);
1182 wait_on_page_writeback(page);
1185 * Remove all references to entry.
1187 swcount = *swap_map;
1188 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1189 retval = shmem_unuse(entry, page);
1190 /* page has already been unlocked and released */
1195 if (swap_count(swcount) && start_mm != &init_mm)
1196 retval = unuse_mm(start_mm, entry, page);
1198 if (swap_count(*swap_map)) {
1199 int set_start_mm = (*swap_map >= swcount);
1200 struct list_head *p = &start_mm->mmlist;
1201 struct mm_struct *new_start_mm = start_mm;
1202 struct mm_struct *prev_mm = start_mm;
1203 struct mm_struct *mm;
1205 atomic_inc(&new_start_mm->mm_users);
1206 atomic_inc(&prev_mm->mm_users);
1207 spin_lock(&mmlist_lock);
1208 while (swap_count(*swap_map) && !retval &&
1209 (p = p->next) != &start_mm->mmlist) {
1210 mm = list_entry(p, struct mm_struct, mmlist);
1211 if (!atomic_inc_not_zero(&mm->mm_users))
1213 spin_unlock(&mmlist_lock);
1219 swcount = *swap_map;
1220 if (!swap_count(swcount)) /* any usage ? */
1222 else if (mm == &init_mm)
1225 retval = unuse_mm(mm, entry, page);
1227 if (set_start_mm && *swap_map < swcount) {
1228 mmput(new_start_mm);
1229 atomic_inc(&mm->mm_users);
1233 spin_lock(&mmlist_lock);
1235 spin_unlock(&mmlist_lock);
1238 start_mm = new_start_mm;
1242 page_cache_release(page);
1247 * If a reference remains (rare), we would like to leave
1248 * the page in the swap cache; but try_to_unmap could
1249 * then re-duplicate the entry once we drop page lock,
1250 * so we might loop indefinitely; also, that page could
1251 * not be swapped out to other storage meanwhile. So:
1252 * delete from cache even if there's another reference,
1253 * after ensuring that the data has been saved to disk -
1254 * since if the reference remains (rarer), it will be
1255 * read from disk into another page. Splitting into two
1256 * pages would be incorrect if swap supported "shared
1257 * private" pages, but they are handled by tmpfs files.
1259 * Given how unuse_vma() targets one particular offset
1260 * in an anon_vma, once the anon_vma has been determined,
1261 * this splitting happens to be just what is needed to
1262 * handle where KSM pages have been swapped out: re-reading
1263 * is unnecessarily slow, but we can fix that later on.
1265 if (swap_count(*swap_map) &&
1266 PageDirty(page) && PageSwapCache(page)) {
1267 struct writeback_control wbc = {
1268 .sync_mode = WB_SYNC_NONE,
1271 swap_writepage(page, &wbc);
1273 wait_on_page_writeback(page);
1277 * It is conceivable that a racing task removed this page from
1278 * swap cache just before we acquired the page lock at the top,
1279 * or while we dropped it in unuse_mm(). The page might even
1280 * be back in swap cache on another swap area: that we must not
1281 * delete, since it may not have been written out to swap yet.
1283 if (PageSwapCache(page) &&
1284 likely(page_private(page) == entry.val))
1285 delete_from_swap_cache(page);
1288 * So we could skip searching mms once swap count went
1289 * to 1, we did not mark any present ptes as dirty: must
1290 * mark page dirty so shrink_page_list will preserve it.
1294 page_cache_release(page);
1297 * Make sure that we aren't completely killing
1298 * interactive performance.
1308 * After a successful try_to_unuse, if no swap is now in use, we know
1309 * we can empty the mmlist. swap_lock must be held on entry and exit.
1310 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1311 * added to the mmlist just after page_duplicate - before would be racy.
1313 static void drain_mmlist(void)
1315 struct list_head *p, *next;
1318 for (type = 0; type < nr_swapfiles; type++)
1319 if (swap_info[type]->inuse_pages)
1321 spin_lock(&mmlist_lock);
1322 list_for_each_safe(p, next, &init_mm.mmlist)
1324 spin_unlock(&mmlist_lock);
1328 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1329 * corresponds to page offset for the specified swap entry.
1330 * Note that the type of this function is sector_t, but it returns page offset
1331 * into the bdev, not sector offset.
1333 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1335 struct swap_info_struct *sis;
1336 struct swap_extent *start_se;
1337 struct swap_extent *se;
1340 sis = swap_info[swp_type(entry)];
1343 offset = swp_offset(entry);
1344 start_se = sis->curr_swap_extent;
1348 struct list_head *lh;
1350 if (se->start_page <= offset &&
1351 offset < (se->start_page + se->nr_pages)) {
1352 return se->start_block + (offset - se->start_page);
1355 se = list_entry(lh, struct swap_extent, list);
1356 sis->curr_swap_extent = se;
1357 BUG_ON(se == start_se); /* It *must* be present */
1362 * Returns the page offset into bdev for the specified page's swap entry.
1364 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1367 entry.val = page_private(page);
1368 return map_swap_entry(entry, bdev);
1372 * Free all of a swapdev's extent information
1374 static void destroy_swap_extents(struct swap_info_struct *sis)
1376 while (!list_empty(&sis->first_swap_extent.list)) {
1377 struct swap_extent *se;
1379 se = list_entry(sis->first_swap_extent.list.next,
1380 struct swap_extent, list);
1381 list_del(&se->list);
1387 * Add a block range (and the corresponding page range) into this swapdev's
1388 * extent list. The extent list is kept sorted in page order.
1390 * This function rather assumes that it is called in ascending page order.
1393 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1394 unsigned long nr_pages, sector_t start_block)
1396 struct swap_extent *se;
1397 struct swap_extent *new_se;
1398 struct list_head *lh;
1400 if (start_page == 0) {
1401 se = &sis->first_swap_extent;
1402 sis->curr_swap_extent = se;
1404 se->nr_pages = nr_pages;
1405 se->start_block = start_block;
1408 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1409 se = list_entry(lh, struct swap_extent, list);
1410 BUG_ON(se->start_page + se->nr_pages != start_page);
1411 if (se->start_block + se->nr_pages == start_block) {
1413 se->nr_pages += nr_pages;
1419 * No merge. Insert a new extent, preserving ordering.
1421 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1424 new_se->start_page = start_page;
1425 new_se->nr_pages = nr_pages;
1426 new_se->start_block = start_block;
1428 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1433 * A `swap extent' is a simple thing which maps a contiguous range of pages
1434 * onto a contiguous range of disk blocks. An ordered list of swap extents
1435 * is built at swapon time and is then used at swap_writepage/swap_readpage
1436 * time for locating where on disk a page belongs.
1438 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1439 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1440 * swap files identically.
1442 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1443 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1444 * swapfiles are handled *identically* after swapon time.
1446 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1447 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1448 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1449 * requirements, they are simply tossed out - we will never use those blocks
1452 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1453 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1454 * which will scribble on the fs.
1456 * The amount of disk space which a single swap extent represents varies.
1457 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1458 * extents in the list. To avoid much list walking, we cache the previous
1459 * search location in `curr_swap_extent', and start new searches from there.
1460 * This is extremely effective. The average number of iterations in
1461 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1463 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1465 struct inode *inode;
1466 unsigned blocks_per_page;
1467 unsigned long page_no;
1469 sector_t probe_block;
1470 sector_t last_block;
1471 sector_t lowest_block = -1;
1472 sector_t highest_block = 0;
1476 inode = sis->swap_file->f_mapping->host;
1477 if (S_ISBLK(inode->i_mode)) {
1478 ret = add_swap_extent(sis, 0, sis->max, 0);
1483 blkbits = inode->i_blkbits;
1484 blocks_per_page = PAGE_SIZE >> blkbits;
1487 * Map all the blocks into the extent list. This code doesn't try
1492 last_block = i_size_read(inode) >> blkbits;
1493 while ((probe_block + blocks_per_page) <= last_block &&
1494 page_no < sis->max) {
1495 unsigned block_in_page;
1496 sector_t first_block;
1498 first_block = bmap(inode, probe_block);
1499 if (first_block == 0)
1503 * It must be PAGE_SIZE aligned on-disk
1505 if (first_block & (blocks_per_page - 1)) {
1510 for (block_in_page = 1; block_in_page < blocks_per_page;
1514 block = bmap(inode, probe_block + block_in_page);
1517 if (block != first_block + block_in_page) {
1524 first_block >>= (PAGE_SHIFT - blkbits);
1525 if (page_no) { /* exclude the header page */
1526 if (first_block < lowest_block)
1527 lowest_block = first_block;
1528 if (first_block > highest_block)
1529 highest_block = first_block;
1533 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1535 ret = add_swap_extent(sis, page_no, 1, first_block);
1540 probe_block += blocks_per_page;
1545 *span = 1 + highest_block - lowest_block;
1547 page_no = 1; /* force Empty message */
1549 sis->pages = page_no - 1;
1550 sis->highest_bit = page_no - 1;
1554 printk(KERN_ERR "swapon: swapfile has holes\n");
1559 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1561 struct swap_info_struct *p = NULL;
1562 unsigned char *swap_map;
1563 struct file *swap_file, *victim;
1564 struct address_space *mapping;
1565 struct inode *inode;
1570 if (!capable(CAP_SYS_ADMIN))
1573 pathname = getname(specialfile);
1574 err = PTR_ERR(pathname);
1575 if (IS_ERR(pathname))
1578 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1580 err = PTR_ERR(victim);
1584 mapping = victim->f_mapping;
1586 spin_lock(&swap_lock);
1587 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1588 p = swap_info[type];
1589 if (p->flags & SWP_WRITEOK) {
1590 if (p->swap_file->f_mapping == mapping)
1597 spin_unlock(&swap_lock);
1600 if (!security_vm_enough_memory(p->pages))
1601 vm_unacct_memory(p->pages);
1604 spin_unlock(&swap_lock);
1608 swap_list.head = p->next;
1610 swap_info[prev]->next = p->next;
1611 if (type == swap_list.next) {
1612 /* just pick something that's safe... */
1613 swap_list.next = swap_list.head;
1616 for (i = p->next; i >= 0; i = swap_info[i]->next)
1617 swap_info[i]->prio = p->prio--;
1620 nr_swap_pages -= p->pages;
1621 total_swap_pages -= p->pages;
1622 p->flags &= ~SWP_WRITEOK;
1623 spin_unlock(&swap_lock);
1625 current->flags |= PF_OOM_ORIGIN;
1626 err = try_to_unuse(type);
1627 current->flags &= ~PF_OOM_ORIGIN;
1630 /* re-insert swap space back into swap_list */
1631 spin_lock(&swap_lock);
1633 p->prio = --least_priority;
1635 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1636 if (p->prio >= swap_info[i]->prio)
1642 swap_list.head = swap_list.next = type;
1644 swap_info[prev]->next = type;
1645 nr_swap_pages += p->pages;
1646 total_swap_pages += p->pages;
1647 p->flags |= SWP_WRITEOK;
1648 spin_unlock(&swap_lock);
1652 /* wait for any unplug function to finish */
1653 down_write(&swap_unplug_sem);
1654 up_write(&swap_unplug_sem);
1656 destroy_swap_extents(p);
1657 if (p->flags & SWP_CONTINUED)
1658 free_swap_count_continuations(p);
1660 mutex_lock(&swapon_mutex);
1661 spin_lock(&swap_lock);
1664 /* wait for anyone still in scan_swap_map */
1665 p->highest_bit = 0; /* cuts scans short */
1666 while (p->flags >= SWP_SCANNING) {
1667 spin_unlock(&swap_lock);
1668 schedule_timeout_uninterruptible(1);
1669 spin_lock(&swap_lock);
1672 swap_file = p->swap_file;
1673 p->swap_file = NULL;
1675 swap_map = p->swap_map;
1678 spin_unlock(&swap_lock);
1679 mutex_unlock(&swapon_mutex);
1681 /* Destroy swap account informatin */
1682 swap_cgroup_swapoff(type);
1684 inode = mapping->host;
1685 if (S_ISBLK(inode->i_mode)) {
1686 struct block_device *bdev = I_BDEV(inode);
1687 set_blocksize(bdev, p->old_block_size);
1688 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1690 mutex_lock(&inode->i_mutex);
1691 inode->i_flags &= ~S_SWAPFILE;
1692 mutex_unlock(&inode->i_mutex);
1694 filp_close(swap_file, NULL);
1696 atomic_inc(&proc_poll_event);
1697 wake_up_interruptible(&proc_poll_wait);
1700 filp_close(victim, NULL);
1705 #ifdef CONFIG_PROC_FS
1707 struct seq_file seq;
1711 static unsigned swaps_poll(struct file *file, poll_table *wait)
1713 struct proc_swaps *s = file->private_data;
1715 poll_wait(file, &proc_poll_wait, wait);
1717 if (s->event != atomic_read(&proc_poll_event)) {
1718 s->event = atomic_read(&proc_poll_event);
1719 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1722 return POLLIN | POLLRDNORM;
1726 static void *swap_start(struct seq_file *swap, loff_t *pos)
1728 struct swap_info_struct *si;
1732 mutex_lock(&swapon_mutex);
1735 return SEQ_START_TOKEN;
1737 for (type = 0; type < nr_swapfiles; type++) {
1738 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1739 si = swap_info[type];
1740 if (!(si->flags & SWP_USED) || !si->swap_map)
1749 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1751 struct swap_info_struct *si = v;
1754 if (v == SEQ_START_TOKEN)
1757 type = si->type + 1;
1759 for (; type < nr_swapfiles; type++) {
1760 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1761 si = swap_info[type];
1762 if (!(si->flags & SWP_USED) || !si->swap_map)
1771 static void swap_stop(struct seq_file *swap, void *v)
1773 mutex_unlock(&swapon_mutex);
1776 static int swap_show(struct seq_file *swap, void *v)
1778 struct swap_info_struct *si = v;
1782 if (si == SEQ_START_TOKEN) {
1783 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1787 file = si->swap_file;
1788 len = seq_path(swap, &file->f_path, " \t\n\\");
1789 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1790 len < 40 ? 40 - len : 1, " ",
1791 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1792 "partition" : "file\t",
1793 si->pages << (PAGE_SHIFT - 10),
1794 si->inuse_pages << (PAGE_SHIFT - 10),
1799 static const struct seq_operations swaps_op = {
1800 .start = swap_start,
1806 static int swaps_open(struct inode *inode, struct file *file)
1808 struct proc_swaps *s;
1811 s = kmalloc(sizeof(struct proc_swaps), GFP_KERNEL);
1815 file->private_data = s;
1817 ret = seq_open(file, &swaps_op);
1824 s->event = atomic_read(&proc_poll_event);
1828 static const struct file_operations proc_swaps_operations = {
1831 .llseek = seq_lseek,
1832 .release = seq_release,
1836 static int __init procswaps_init(void)
1838 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1841 __initcall(procswaps_init);
1842 #endif /* CONFIG_PROC_FS */
1844 #ifdef MAX_SWAPFILES_CHECK
1845 static int __init max_swapfiles_check(void)
1847 MAX_SWAPFILES_CHECK();
1850 late_initcall(max_swapfiles_check);
1854 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1856 * The swapon system call
1858 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1860 struct swap_info_struct *p;
1862 struct block_device *bdev = NULL;
1863 struct file *swap_file = NULL;
1864 struct address_space *mapping;
1868 union swap_header *swap_header;
1869 unsigned int nr_good_pages;
1872 unsigned long maxpages;
1873 unsigned long swapfilepages;
1874 unsigned char *swap_map = NULL;
1875 struct page *page = NULL;
1876 struct inode *inode = NULL;
1879 if (!capable(CAP_SYS_ADMIN))
1882 p = kzalloc(sizeof(*p), GFP_KERNEL);
1886 spin_lock(&swap_lock);
1887 for (type = 0; type < nr_swapfiles; type++) {
1888 if (!(swap_info[type]->flags & SWP_USED))
1892 if (type >= MAX_SWAPFILES) {
1893 spin_unlock(&swap_lock);
1897 if (type >= nr_swapfiles) {
1899 swap_info[type] = p;
1901 * Write swap_info[type] before nr_swapfiles, in case a
1902 * racing procfs swap_start() or swap_next() is reading them.
1903 * (We never shrink nr_swapfiles, we never free this entry.)
1909 p = swap_info[type];
1911 * Do not memset this entry: a racing procfs swap_next()
1912 * would be relying on p->type to remain valid.
1915 INIT_LIST_HEAD(&p->first_swap_extent.list);
1916 p->flags = SWP_USED;
1918 spin_unlock(&swap_lock);
1920 name = getname(specialfile);
1921 error = PTR_ERR(name);
1926 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1927 error = PTR_ERR(swap_file);
1928 if (IS_ERR(swap_file)) {
1933 p->swap_file = swap_file;
1934 mapping = swap_file->f_mapping;
1935 inode = mapping->host;
1938 for (i = 0; i < nr_swapfiles; i++) {
1939 struct swap_info_struct *q = swap_info[i];
1941 if (i == type || !q->swap_file)
1943 if (mapping == q->swap_file->f_mapping)
1948 if (S_ISBLK(inode->i_mode)) {
1949 bdev = bdgrab(I_BDEV(inode));
1950 error = blkdev_get(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1957 p->old_block_size = block_size(bdev);
1958 error = set_blocksize(bdev, PAGE_SIZE);
1962 p->flags |= SWP_BLKDEV;
1963 } else if (S_ISREG(inode->i_mode)) {
1964 p->bdev = inode->i_sb->s_bdev;
1965 mutex_lock(&inode->i_mutex);
1967 if (IS_SWAPFILE(inode)) {
1975 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1978 * Read the swap header.
1980 if (!mapping->a_ops->readpage) {
1984 page = read_mapping_page(mapping, 0, swap_file);
1986 error = PTR_ERR(page);
1989 swap_header = kmap(page);
1991 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1992 printk(KERN_ERR "Unable to find swap-space signature\n");
1997 /* swap partition endianess hack... */
1998 if (swab32(swap_header->info.version) == 1) {
1999 swab32s(&swap_header->info.version);
2000 swab32s(&swap_header->info.last_page);
2001 swab32s(&swap_header->info.nr_badpages);
2002 for (i = 0; i < swap_header->info.nr_badpages; i++)
2003 swab32s(&swap_header->info.badpages[i]);
2005 /* Check the swap header's sub-version */
2006 if (swap_header->info.version != 1) {
2008 "Unable to handle swap header version %d\n",
2009 swap_header->info.version);
2015 p->cluster_next = 1;
2019 * Find out how many pages are allowed for a single swap
2020 * device. There are two limiting factors: 1) the number of
2021 * bits for the swap offset in the swp_entry_t type and
2022 * 2) the number of bits in the a swap pte as defined by
2023 * the different architectures. In order to find the
2024 * largest possible bit mask a swap entry with swap type 0
2025 * and swap offset ~0UL is created, encoded to a swap pte,
2026 * decoded to a swp_entry_t again and finally the swap
2027 * offset is extracted. This will mask all the bits from
2028 * the initial ~0UL mask that can't be encoded in either
2029 * the swp_entry_t or the architecture definition of a
2032 maxpages = swp_offset(pte_to_swp_entry(
2033 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2034 if (maxpages > swap_header->info.last_page) {
2035 maxpages = swap_header->info.last_page + 1;
2036 /* p->max is an unsigned int: don't overflow it */
2037 if ((unsigned int)maxpages == 0)
2038 maxpages = UINT_MAX;
2040 p->highest_bit = maxpages - 1;
2045 if (swapfilepages && maxpages > swapfilepages) {
2047 "Swap area shorter than signature indicates\n");
2050 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2052 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2055 /* OK, set up the swap map and apply the bad block list */
2056 swap_map = vmalloc(maxpages);
2062 memset(swap_map, 0, maxpages);
2063 nr_good_pages = maxpages - 1; /* omit header page */
2065 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2066 unsigned int page_nr = swap_header->info.badpages[i];
2067 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2071 if (page_nr < maxpages) {
2072 swap_map[page_nr] = SWAP_MAP_BAD;
2077 error = swap_cgroup_swapon(type, maxpages);
2081 if (nr_good_pages) {
2082 swap_map[0] = SWAP_MAP_BAD;
2084 p->pages = nr_good_pages;
2085 nr_extents = setup_swap_extents(p, &span);
2086 if (nr_extents < 0) {
2090 nr_good_pages = p->pages;
2092 if (!nr_good_pages) {
2093 printk(KERN_WARNING "Empty swap-file\n");
2099 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2100 p->flags |= SWP_SOLIDSTATE;
2101 p->cluster_next = 1 + (random32() % p->highest_bit);
2103 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2104 p->flags |= SWP_DISCARDABLE;
2107 mutex_lock(&swapon_mutex);
2108 spin_lock(&swap_lock);
2109 if (swap_flags & SWAP_FLAG_PREFER)
2111 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2113 p->prio = --least_priority;
2114 p->swap_map = swap_map;
2115 p->flags |= SWP_WRITEOK;
2116 nr_swap_pages += nr_good_pages;
2117 total_swap_pages += nr_good_pages;
2119 printk(KERN_INFO "Adding %uk swap on %s. "
2120 "Priority:%d extents:%d across:%lluk %s%s\n",
2121 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2122 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2123 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2124 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2126 /* insert swap space into swap_list: */
2128 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2129 if (p->prio >= swap_info[i]->prio)
2135 swap_list.head = swap_list.next = type;
2137 swap_info[prev]->next = type;
2138 spin_unlock(&swap_lock);
2139 mutex_unlock(&swapon_mutex);
2140 atomic_inc(&proc_poll_event);
2141 wake_up_interruptible(&proc_poll_wait);
2147 set_blocksize(bdev, p->old_block_size);
2148 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2150 destroy_swap_extents(p);
2151 swap_cgroup_swapoff(type);
2153 spin_lock(&swap_lock);
2154 p->swap_file = NULL;
2156 spin_unlock(&swap_lock);
2160 mutex_unlock(&inode->i_mutex);
2163 filp_close(swap_file, NULL);
2166 if (page && !IS_ERR(page)) {
2168 page_cache_release(page);
2174 inode->i_flags |= S_SWAPFILE;
2175 mutex_unlock(&inode->i_mutex);
2180 void si_swapinfo(struct sysinfo *val)
2183 unsigned long nr_to_be_unused = 0;
2185 spin_lock(&swap_lock);
2186 for (type = 0; type < nr_swapfiles; type++) {
2187 struct swap_info_struct *si = swap_info[type];
2189 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2190 nr_to_be_unused += si->inuse_pages;
2192 val->freeswap = nr_swap_pages + nr_to_be_unused;
2193 val->totalswap = total_swap_pages + nr_to_be_unused;
2194 spin_unlock(&swap_lock);
2198 * Verify that a swap entry is valid and increment its swap map count.
2200 * Returns error code in following case.
2202 * - swp_entry is invalid -> EINVAL
2203 * - swp_entry is migration entry -> EINVAL
2204 * - swap-cache reference is requested but there is already one. -> EEXIST
2205 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2206 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2208 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2210 struct swap_info_struct *p;
2211 unsigned long offset, type;
2212 unsigned char count;
2213 unsigned char has_cache;
2216 if (non_swap_entry(entry))
2219 type = swp_type(entry);
2220 if (type >= nr_swapfiles)
2222 p = swap_info[type];
2223 offset = swp_offset(entry);
2225 spin_lock(&swap_lock);
2226 if (unlikely(offset >= p->max))
2229 count = p->swap_map[offset];
2230 has_cache = count & SWAP_HAS_CACHE;
2231 count &= ~SWAP_HAS_CACHE;
2234 if (usage == SWAP_HAS_CACHE) {
2236 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2237 if (!has_cache && count)
2238 has_cache = SWAP_HAS_CACHE;
2239 else if (has_cache) /* someone else added cache */
2241 else /* no users remaining */
2244 } else if (count || has_cache) {
2246 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2248 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2250 else if (swap_count_continued(p, offset, count))
2251 count = COUNT_CONTINUED;
2255 err = -ENOENT; /* unused swap entry */
2257 p->swap_map[offset] = count | has_cache;
2260 spin_unlock(&swap_lock);
2265 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2270 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2271 * (in which case its reference count is never incremented).
2273 void swap_shmem_alloc(swp_entry_t entry)
2275 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2279 * Increase reference count of swap entry by 1.
2280 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2281 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2282 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2283 * might occur if a page table entry has got corrupted.
2285 int swap_duplicate(swp_entry_t entry)
2289 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2290 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2295 * @entry: swap entry for which we allocate swap cache.
2297 * Called when allocating swap cache for existing swap entry,
2298 * This can return error codes. Returns 0 at success.
2299 * -EBUSY means there is a swap cache.
2300 * Note: return code is different from swap_duplicate().
2302 int swapcache_prepare(swp_entry_t entry)
2304 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2308 * Sets callback for event when swap_map[offset] == 0
2309 * i.e. page at this swap offset is not longer used.
2311 * type: identifies swap file
2312 * fn: callback function
2314 void set_notify_swap_entry_free(unsigned type, void (*fn) (unsigned long))
2316 struct swap_info_struct *sis;
2317 sis = swap_info[type];
2319 sis->notify_swap_entry_free_fn = fn;
2322 EXPORT_SYMBOL(set_notify_swap_entry_free);
2325 * swap_lock prevents swap_map being freed. Don't grab an extra
2326 * reference on the swaphandle, it doesn't matter if it becomes unused.
2328 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2330 struct swap_info_struct *si;
2331 int our_page_cluster = page_cluster;
2332 pgoff_t target, toff;
2336 if (!our_page_cluster) /* no readahead */
2339 si = swap_info[swp_type(entry)];
2340 target = swp_offset(entry);
2341 base = (target >> our_page_cluster) << our_page_cluster;
2342 end = base + (1 << our_page_cluster);
2343 if (!base) /* first page is swap header */
2346 spin_lock(&swap_lock);
2347 if (end > si->max) /* don't go beyond end of map */
2350 /* Count contiguous allocated slots above our target */
2351 for (toff = target; ++toff < end; nr_pages++) {
2352 /* Don't read in free or bad pages */
2353 if (!si->swap_map[toff])
2355 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2358 /* Count contiguous allocated slots below our target */
2359 for (toff = target; --toff >= base; nr_pages++) {
2360 /* Don't read in free or bad pages */
2361 if (!si->swap_map[toff])
2363 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2366 spin_unlock(&swap_lock);
2369 * Indicate starting offset, and return number of pages to get:
2370 * if only 1, say 0, since there's then no readahead to be done.
2373 return nr_pages? ++nr_pages: 0;
2377 * add_swap_count_continuation - called when a swap count is duplicated
2378 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2379 * page of the original vmalloc'ed swap_map, to hold the continuation count
2380 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2381 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2383 * These continuation pages are seldom referenced: the common paths all work
2384 * on the original swap_map, only referring to a continuation page when the
2385 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2387 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2388 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2389 * can be called after dropping locks.
2391 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2393 struct swap_info_struct *si;
2396 struct page *list_page;
2398 unsigned char count;
2401 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2402 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2404 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2406 si = swap_info_get(entry);
2409 * An acceptable race has occurred since the failing
2410 * __swap_duplicate(): the swap entry has been freed,
2411 * perhaps even the whole swap_map cleared for swapoff.
2416 offset = swp_offset(entry);
2417 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2419 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2421 * The higher the swap count, the more likely it is that tasks
2422 * will race to add swap count continuation: we need to avoid
2423 * over-provisioning.
2429 spin_unlock(&swap_lock);
2434 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2435 * no architecture is using highmem pages for kernel pagetables: so it
2436 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2438 head = vmalloc_to_page(si->swap_map + offset);
2439 offset &= ~PAGE_MASK;
2442 * Page allocation does not initialize the page's lru field,
2443 * but it does always reset its private field.
2445 if (!page_private(head)) {
2446 BUG_ON(count & COUNT_CONTINUED);
2447 INIT_LIST_HEAD(&head->lru);
2448 set_page_private(head, SWP_CONTINUED);
2449 si->flags |= SWP_CONTINUED;
2452 list_for_each_entry(list_page, &head->lru, lru) {
2456 * If the previous map said no continuation, but we've found
2457 * a continuation page, free our allocation and use this one.
2459 if (!(count & COUNT_CONTINUED))
2462 map = kmap_atomic(list_page, KM_USER0) + offset;
2464 kunmap_atomic(map, KM_USER0);
2467 * If this continuation count now has some space in it,
2468 * free our allocation and use this one.
2470 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2474 list_add_tail(&page->lru, &head->lru);
2475 page = NULL; /* now it's attached, don't free it */
2477 spin_unlock(&swap_lock);
2485 * swap_count_continued - when the original swap_map count is incremented
2486 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2487 * into, carry if so, or else fail until a new continuation page is allocated;
2488 * when the original swap_map count is decremented from 0 with continuation,
2489 * borrow from the continuation and report whether it still holds more.
2490 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2492 static bool swap_count_continued(struct swap_info_struct *si,
2493 pgoff_t offset, unsigned char count)
2499 head = vmalloc_to_page(si->swap_map + offset);
2500 if (page_private(head) != SWP_CONTINUED) {
2501 BUG_ON(count & COUNT_CONTINUED);
2502 return false; /* need to add count continuation */
2505 offset &= ~PAGE_MASK;
2506 page = list_entry(head->lru.next, struct page, lru);
2507 map = kmap_atomic(page, KM_USER0) + offset;
2509 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2510 goto init_map; /* jump over SWAP_CONT_MAX checks */
2512 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2514 * Think of how you add 1 to 999
2516 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2517 kunmap_atomic(map, KM_USER0);
2518 page = list_entry(page->lru.next, struct page, lru);
2519 BUG_ON(page == head);
2520 map = kmap_atomic(page, KM_USER0) + offset;
2522 if (*map == SWAP_CONT_MAX) {
2523 kunmap_atomic(map, KM_USER0);
2524 page = list_entry(page->lru.next, struct page, lru);
2526 return false; /* add count continuation */
2527 map = kmap_atomic(page, KM_USER0) + offset;
2528 init_map: *map = 0; /* we didn't zero the page */
2531 kunmap_atomic(map, KM_USER0);
2532 page = list_entry(page->lru.prev, struct page, lru);
2533 while (page != head) {
2534 map = kmap_atomic(page, KM_USER0) + offset;
2535 *map = COUNT_CONTINUED;
2536 kunmap_atomic(map, KM_USER0);
2537 page = list_entry(page->lru.prev, struct page, lru);
2539 return true; /* incremented */
2541 } else { /* decrementing */
2543 * Think of how you subtract 1 from 1000
2545 BUG_ON(count != COUNT_CONTINUED);
2546 while (*map == COUNT_CONTINUED) {
2547 kunmap_atomic(map, KM_USER0);
2548 page = list_entry(page->lru.next, struct page, lru);
2549 BUG_ON(page == head);
2550 map = kmap_atomic(page, KM_USER0) + offset;
2556 kunmap_atomic(map, KM_USER0);
2557 page = list_entry(page->lru.prev, struct page, lru);
2558 while (page != head) {
2559 map = kmap_atomic(page, KM_USER0) + offset;
2560 *map = SWAP_CONT_MAX | count;
2561 count = COUNT_CONTINUED;
2562 kunmap_atomic(map, KM_USER0);
2563 page = list_entry(page->lru.prev, struct page, lru);
2565 return count == COUNT_CONTINUED;
2570 * free_swap_count_continuations - swapoff free all the continuation pages
2571 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2573 static void free_swap_count_continuations(struct swap_info_struct *si)
2577 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2579 head = vmalloc_to_page(si->swap_map + offset);
2580 if (page_private(head)) {
2581 struct list_head *this, *next;
2582 list_for_each_safe(this, next, &head->lru) {
2584 page = list_entry(this, struct page, lru);