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>
34 #include <asm/pgtable.h>
35 #include <asm/tlbflush.h>
36 #include <linux/swapops.h>
37 #include <linux/page_cgroup.h>
39 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
41 static void free_swap_count_continuations(struct swap_info_struct *);
42 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
44 static DEFINE_SPINLOCK(swap_lock);
45 static unsigned int nr_swapfiles;
47 long total_swap_pages;
48 static int least_priority;
50 static const char Bad_file[] = "Bad swap file entry ";
51 static const char Unused_file[] = "Unused swap file entry ";
52 static const char Bad_offset[] = "Bad swap offset entry ";
53 static const char Unused_offset[] = "Unused swap offset entry ";
55 static struct swap_list_t swap_list = {-1, -1};
57 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
59 static DEFINE_MUTEX(swapon_mutex);
61 static inline unsigned char swap_count(unsigned char ent)
63 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
66 /* returns 1 if swap entry is freed */
68 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
70 swp_entry_t entry = swp_entry(si->type, offset);
74 page = find_get_page(&swapper_space, entry.val);
78 * This function is called from scan_swap_map() and it's called
79 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
80 * We have to use trylock for avoiding deadlock. This is a special
81 * case and you should use try_to_free_swap() with explicit lock_page()
82 * in usual operations.
84 if (trylock_page(page)) {
85 ret = try_to_free_swap(page);
88 page_cache_release(page);
93 * We need this because the bdev->unplug_fn can sleep and we cannot
94 * hold swap_lock while calling the unplug_fn. And swap_lock
95 * cannot be turned into a mutex.
97 static DECLARE_RWSEM(swap_unplug_sem);
99 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
103 down_read(&swap_unplug_sem);
104 entry.val = page_private(page);
105 if (PageSwapCache(page)) {
106 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
107 struct backing_dev_info *bdi;
110 * If the page is removed from swapcache from under us (with a
111 * racy try_to_unuse/swapoff) we need an additional reference
112 * count to avoid reading garbage from page_private(page) above.
113 * If the WARN_ON triggers during a swapoff it maybe the race
114 * condition and it's harmless. However if it triggers without
115 * swapoff it signals a problem.
117 WARN_ON(page_count(page) <= 1);
119 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
120 blk_run_backing_dev(bdi, page);
122 up_read(&swap_unplug_sem);
126 * swapon tell device that all the old swap contents can be discarded,
127 * to allow the swap device to optimize its wear-levelling.
129 static int discard_swap(struct swap_info_struct *si)
131 struct swap_extent *se;
132 sector_t start_block;
136 /* Do not discard the swap header page! */
137 se = &si->first_swap_extent;
138 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
139 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
141 err = blkdev_issue_discard(si->bdev, start_block,
142 nr_blocks, GFP_KERNEL,
143 BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
149 list_for_each_entry(se, &si->first_swap_extent.list, list) {
150 start_block = se->start_block << (PAGE_SHIFT - 9);
151 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
153 err = blkdev_issue_discard(si->bdev, start_block,
154 nr_blocks, GFP_KERNEL,
155 BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
161 return err; /* That will often be -EOPNOTSUPP */
165 * swap allocation tell device that a cluster of swap can now be discarded,
166 * to allow the swap device to optimize its wear-levelling.
168 static void discard_swap_cluster(struct swap_info_struct *si,
169 pgoff_t start_page, pgoff_t nr_pages)
171 struct swap_extent *se = si->curr_swap_extent;
172 int found_extent = 0;
175 struct list_head *lh;
177 if (se->start_page <= start_page &&
178 start_page < se->start_page + se->nr_pages) {
179 pgoff_t offset = start_page - se->start_page;
180 sector_t start_block = se->start_block + offset;
181 sector_t nr_blocks = se->nr_pages - offset;
183 if (nr_blocks > nr_pages)
184 nr_blocks = nr_pages;
185 start_page += nr_blocks;
186 nr_pages -= nr_blocks;
189 si->curr_swap_extent = se;
191 start_block <<= PAGE_SHIFT - 9;
192 nr_blocks <<= PAGE_SHIFT - 9;
193 if (blkdev_issue_discard(si->bdev, start_block,
194 nr_blocks, GFP_NOIO, BLKDEV_IFL_WAIT |
200 se = list_entry(lh, struct swap_extent, list);
204 static int wait_for_discard(void *word)
210 #define SWAPFILE_CLUSTER 256
211 #define LATENCY_LIMIT 256
213 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
216 unsigned long offset;
217 unsigned long scan_base;
218 unsigned long last_in_cluster = 0;
219 int latency_ration = LATENCY_LIMIT;
220 int found_free_cluster = 0;
223 * We try to cluster swap pages by allocating them sequentially
224 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
225 * way, however, we resort to first-free allocation, starting
226 * a new cluster. This prevents us from scattering swap pages
227 * all over the entire swap partition, so that we reduce
228 * overall disk seek times between swap pages. -- sct
229 * But we do now try to find an empty cluster. -Andrea
230 * And we let swap pages go all over an SSD partition. Hugh
233 si->flags += SWP_SCANNING;
234 scan_base = offset = si->cluster_next;
236 if (unlikely(!si->cluster_nr--)) {
237 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
238 si->cluster_nr = SWAPFILE_CLUSTER - 1;
241 if (si->flags & SWP_DISCARDABLE) {
243 * Start range check on racing allocations, in case
244 * they overlap the cluster we eventually decide on
245 * (we scan without swap_lock to allow preemption).
246 * It's hardly conceivable that cluster_nr could be
247 * wrapped during our scan, but don't depend on it.
249 if (si->lowest_alloc)
251 si->lowest_alloc = si->max;
252 si->highest_alloc = 0;
254 spin_unlock(&swap_lock);
257 * If seek is expensive, start searching for new cluster from
258 * start of partition, to minimize the span of allocated swap.
259 * But if seek is cheap, search from our current position, so
260 * that swap is allocated from all over the partition: if the
261 * Flash Translation Layer only remaps within limited zones,
262 * we don't want to wear out the first zone too quickly.
264 if (!(si->flags & SWP_SOLIDSTATE))
265 scan_base = offset = si->lowest_bit;
266 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
268 /* Locate the first empty (unaligned) cluster */
269 for (; last_in_cluster <= si->highest_bit; offset++) {
270 if (si->swap_map[offset])
271 last_in_cluster = offset + SWAPFILE_CLUSTER;
272 else if (offset == last_in_cluster) {
273 spin_lock(&swap_lock);
274 offset -= SWAPFILE_CLUSTER - 1;
275 si->cluster_next = offset;
276 si->cluster_nr = SWAPFILE_CLUSTER - 1;
277 found_free_cluster = 1;
280 if (unlikely(--latency_ration < 0)) {
282 latency_ration = LATENCY_LIMIT;
286 offset = si->lowest_bit;
287 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
289 /* Locate the first empty (unaligned) cluster */
290 for (; last_in_cluster < scan_base; offset++) {
291 if (si->swap_map[offset])
292 last_in_cluster = offset + SWAPFILE_CLUSTER;
293 else if (offset == last_in_cluster) {
294 spin_lock(&swap_lock);
295 offset -= SWAPFILE_CLUSTER - 1;
296 si->cluster_next = offset;
297 si->cluster_nr = SWAPFILE_CLUSTER - 1;
298 found_free_cluster = 1;
301 if (unlikely(--latency_ration < 0)) {
303 latency_ration = LATENCY_LIMIT;
308 spin_lock(&swap_lock);
309 si->cluster_nr = SWAPFILE_CLUSTER - 1;
310 si->lowest_alloc = 0;
314 if (!(si->flags & SWP_WRITEOK))
316 if (!si->highest_bit)
318 if (offset > si->highest_bit)
319 scan_base = offset = si->lowest_bit;
321 /* reuse swap entry of cache-only swap if not busy. */
322 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
324 spin_unlock(&swap_lock);
325 swap_was_freed = __try_to_reclaim_swap(si, offset);
326 spin_lock(&swap_lock);
327 /* entry was freed successfully, try to use this again */
330 goto scan; /* check next one */
333 if (si->swap_map[offset])
336 if (offset == si->lowest_bit)
338 if (offset == si->highest_bit)
341 if (si->inuse_pages == si->pages) {
342 si->lowest_bit = si->max;
345 si->swap_map[offset] = usage;
346 si->cluster_next = offset + 1;
347 si->flags -= SWP_SCANNING;
349 if (si->lowest_alloc) {
351 * Only set when SWP_DISCARDABLE, and there's a scan
352 * for a free cluster in progress or just completed.
354 if (found_free_cluster) {
356 * To optimize wear-levelling, discard the
357 * old data of the cluster, taking care not to
358 * discard any of its pages that have already
359 * been allocated by racing tasks (offset has
360 * already stepped over any at the beginning).
362 if (offset < si->highest_alloc &&
363 si->lowest_alloc <= last_in_cluster)
364 last_in_cluster = si->lowest_alloc - 1;
365 si->flags |= SWP_DISCARDING;
366 spin_unlock(&swap_lock);
368 if (offset < last_in_cluster)
369 discard_swap_cluster(si, offset,
370 last_in_cluster - offset + 1);
372 spin_lock(&swap_lock);
373 si->lowest_alloc = 0;
374 si->flags &= ~SWP_DISCARDING;
376 smp_mb(); /* wake_up_bit advises this */
377 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
379 } else if (si->flags & SWP_DISCARDING) {
381 * Delay using pages allocated by racing tasks
382 * until the whole discard has been issued. We
383 * could defer that delay until swap_writepage,
384 * but it's easier to keep this self-contained.
386 spin_unlock(&swap_lock);
387 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
388 wait_for_discard, TASK_UNINTERRUPTIBLE);
389 spin_lock(&swap_lock);
392 * Note pages allocated by racing tasks while
393 * scan for a free cluster is in progress, so
394 * that its final discard can exclude them.
396 if (offset < si->lowest_alloc)
397 si->lowest_alloc = offset;
398 if (offset > si->highest_alloc)
399 si->highest_alloc = offset;
405 spin_unlock(&swap_lock);
406 while (++offset <= si->highest_bit) {
407 if (!si->swap_map[offset]) {
408 spin_lock(&swap_lock);
411 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
412 spin_lock(&swap_lock);
415 if (unlikely(--latency_ration < 0)) {
417 latency_ration = LATENCY_LIMIT;
420 offset = si->lowest_bit;
421 while (++offset < scan_base) {
422 if (!si->swap_map[offset]) {
423 spin_lock(&swap_lock);
426 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
427 spin_lock(&swap_lock);
430 if (unlikely(--latency_ration < 0)) {
432 latency_ration = LATENCY_LIMIT;
435 spin_lock(&swap_lock);
438 si->flags -= SWP_SCANNING;
442 swp_entry_t get_swap_page(void)
444 struct swap_info_struct *si;
449 spin_lock(&swap_lock);
450 if (nr_swap_pages <= 0)
454 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
455 si = swap_info[type];
458 (!wrapped && si->prio != swap_info[next]->prio)) {
459 next = swap_list.head;
463 if (!si->highest_bit)
465 if (!(si->flags & SWP_WRITEOK))
468 swap_list.next = next;
469 /* This is called for allocating swap entry for cache */
470 offset = scan_swap_map(si, SWAP_HAS_CACHE);
472 spin_unlock(&swap_lock);
473 return swp_entry(type, offset);
475 next = swap_list.next;
480 spin_unlock(&swap_lock);
481 return (swp_entry_t) {0};
484 /* The only caller of this function is now susupend routine */
485 swp_entry_t get_swap_page_of_type(int type)
487 struct swap_info_struct *si;
490 spin_lock(&swap_lock);
491 si = swap_info[type];
492 if (si && (si->flags & SWP_WRITEOK)) {
494 /* This is called for allocating swap entry, not cache */
495 offset = scan_swap_map(si, 1);
497 spin_unlock(&swap_lock);
498 return swp_entry(type, offset);
502 spin_unlock(&swap_lock);
503 return (swp_entry_t) {0};
506 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
508 struct swap_info_struct *p;
509 unsigned long offset, type;
513 type = swp_type(entry);
514 if (type >= nr_swapfiles)
517 if (!(p->flags & SWP_USED))
519 offset = swp_offset(entry);
520 if (offset >= p->max)
522 if (!p->swap_map[offset])
524 spin_lock(&swap_lock);
528 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
531 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
534 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
537 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
542 static unsigned char swap_entry_free(struct swap_info_struct *p,
543 swp_entry_t entry, unsigned char usage)
545 unsigned long offset = swp_offset(entry);
547 unsigned char has_cache;
549 count = p->swap_map[offset];
550 has_cache = count & SWAP_HAS_CACHE;
551 count &= ~SWAP_HAS_CACHE;
553 if (usage == SWAP_HAS_CACHE) {
554 VM_BUG_ON(!has_cache);
556 } else if (count == SWAP_MAP_SHMEM) {
558 * Or we could insist on shmem.c using a special
559 * swap_shmem_free() and free_shmem_swap_and_cache()...
562 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
563 if (count == COUNT_CONTINUED) {
564 if (swap_count_continued(p, offset, count))
565 count = SWAP_MAP_MAX | COUNT_CONTINUED;
567 count = SWAP_MAP_MAX;
573 mem_cgroup_uncharge_swap(entry);
575 usage = count | has_cache;
576 p->swap_map[offset] = usage;
578 /* free if no reference */
580 if (offset < p->lowest_bit)
581 p->lowest_bit = offset;
582 if (offset > p->highest_bit)
583 p->highest_bit = offset;
584 if (swap_list.next >= 0 &&
585 p->prio > swap_info[swap_list.next]->prio)
586 swap_list.next = p->type;
595 * Caller has made sure that the swapdevice corresponding to entry
596 * is still around or has not been recycled.
598 void swap_free(swp_entry_t entry)
600 struct swap_info_struct *p;
602 p = swap_info_get(entry);
604 swap_entry_free(p, entry, 1);
605 spin_unlock(&swap_lock);
610 * Called after dropping swapcache to decrease refcnt to swap entries.
612 void swapcache_free(swp_entry_t entry, struct page *page)
614 struct swap_info_struct *p;
617 p = swap_info_get(entry);
619 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
621 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
622 spin_unlock(&swap_lock);
627 * How many references to page are currently swapped out?
628 * This does not give an exact answer when swap count is continued,
629 * but does include the high COUNT_CONTINUED flag to allow for that.
631 static inline int page_swapcount(struct page *page)
634 struct swap_info_struct *p;
637 entry.val = page_private(page);
638 p = swap_info_get(entry);
640 count = swap_count(p->swap_map[swp_offset(entry)]);
641 spin_unlock(&swap_lock);
647 * We can write to an anon page without COW if there are no other references
648 * to it. And as a side-effect, free up its swap: because the old content
649 * on disk will never be read, and seeking back there to write new content
650 * later would only waste time away from clustering.
652 int reuse_swap_page(struct page *page)
656 VM_BUG_ON(!PageLocked(page));
657 if (unlikely(PageKsm(page)))
659 count = page_mapcount(page);
660 if (count <= 1 && PageSwapCache(page)) {
661 count += page_swapcount(page);
662 if (count == 1 && !PageWriteback(page)) {
663 delete_from_swap_cache(page);
671 * If swap is getting full, or if there are no more mappings of this page,
672 * then try_to_free_swap is called to free its swap space.
674 int try_to_free_swap(struct page *page)
676 VM_BUG_ON(!PageLocked(page));
678 if (!PageSwapCache(page))
680 if (PageWriteback(page))
682 if (page_swapcount(page))
685 delete_from_swap_cache(page);
691 * Free the swap entry like above, but also try to
692 * free the page cache entry if it is the last user.
694 int free_swap_and_cache(swp_entry_t entry)
696 struct swap_info_struct *p;
697 struct page *page = NULL;
699 if (non_swap_entry(entry))
702 p = swap_info_get(entry);
704 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
705 page = find_get_page(&swapper_space, entry.val);
706 if (page && !trylock_page(page)) {
707 page_cache_release(page);
711 spin_unlock(&swap_lock);
715 * Not mapped elsewhere, or swap space full? Free it!
716 * Also recheck PageSwapCache now page is locked (above).
718 if (PageSwapCache(page) && !PageWriteback(page) &&
719 (!page_mapped(page) || vm_swap_full())) {
720 delete_from_swap_cache(page);
724 page_cache_release(page);
729 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
731 * mem_cgroup_count_swap_user - count the user of a swap entry
732 * @ent: the swap entry to be checked
733 * @pagep: the pointer for the swap cache page of the entry to be stored
735 * Returns the number of the user of the swap entry. The number is valid only
736 * for swaps of anonymous pages.
737 * If the entry is found on swap cache, the page is stored to pagep with
738 * refcount of it being incremented.
740 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
743 struct swap_info_struct *p;
746 page = find_get_page(&swapper_space, ent.val);
748 count += page_mapcount(page);
749 p = swap_info_get(ent);
751 count += swap_count(p->swap_map[swp_offset(ent)]);
752 spin_unlock(&swap_lock);
760 #ifdef CONFIG_HIBERNATION
762 * Find the swap type that corresponds to given device (if any).
764 * @offset - number of the PAGE_SIZE-sized block of the device, starting
765 * from 0, in which the swap header is expected to be located.
767 * This is needed for the suspend to disk (aka swsusp).
769 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
771 struct block_device *bdev = NULL;
775 bdev = bdget(device);
777 spin_lock(&swap_lock);
778 for (type = 0; type < nr_swapfiles; type++) {
779 struct swap_info_struct *sis = swap_info[type];
781 if (!(sis->flags & SWP_WRITEOK))
786 *bdev_p = bdgrab(sis->bdev);
788 spin_unlock(&swap_lock);
791 if (bdev == sis->bdev) {
792 struct swap_extent *se = &sis->first_swap_extent;
794 if (se->start_block == offset) {
796 *bdev_p = bdgrab(sis->bdev);
798 spin_unlock(&swap_lock);
804 spin_unlock(&swap_lock);
812 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
813 * corresponding to given index in swap_info (swap type).
815 sector_t swapdev_block(int type, pgoff_t offset)
817 struct block_device *bdev;
819 if ((unsigned int)type >= nr_swapfiles)
821 if (!(swap_info[type]->flags & SWP_WRITEOK))
823 return map_swap_entry(swp_entry(type, offset), &bdev);
827 * Return either the total number of swap pages of given type, or the number
828 * of free pages of that type (depending on @free)
830 * This is needed for software suspend
832 unsigned int count_swap_pages(int type, int free)
836 spin_lock(&swap_lock);
837 if ((unsigned int)type < nr_swapfiles) {
838 struct swap_info_struct *sis = swap_info[type];
840 if (sis->flags & SWP_WRITEOK) {
843 n -= sis->inuse_pages;
846 spin_unlock(&swap_lock);
849 #endif /* CONFIG_HIBERNATION */
852 * No need to decide whether this PTE shares the swap entry with others,
853 * just let do_wp_page work it out if a write is requested later - to
854 * force COW, vm_page_prot omits write permission from any private vma.
856 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
857 unsigned long addr, swp_entry_t entry, struct page *page)
859 struct mem_cgroup *ptr = NULL;
864 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
869 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
870 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
872 mem_cgroup_cancel_charge_swapin(ptr);
877 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
878 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
880 set_pte_at(vma->vm_mm, addr, pte,
881 pte_mkold(mk_pte(page, vma->vm_page_prot)));
882 page_add_anon_rmap(page, vma, addr);
883 mem_cgroup_commit_charge_swapin(page, ptr);
886 * Move the page to the active list so it is not
887 * immediately swapped out again after swapon.
891 pte_unmap_unlock(pte, ptl);
896 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
897 unsigned long addr, unsigned long end,
898 swp_entry_t entry, struct page *page)
900 pte_t swp_pte = swp_entry_to_pte(entry);
905 * We don't actually need pte lock while scanning for swp_pte: since
906 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
907 * page table while we're scanning; though it could get zapped, and on
908 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
909 * of unmatched parts which look like swp_pte, so unuse_pte must
910 * recheck under pte lock. Scanning without pte lock lets it be
911 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
913 pte = pte_offset_map(pmd, addr);
916 * swapoff spends a _lot_ of time in this loop!
917 * Test inline before going to call unuse_pte.
919 if (unlikely(pte_same(*pte, swp_pte))) {
921 ret = unuse_pte(vma, pmd, addr, entry, page);
924 pte = pte_offset_map(pmd, addr);
926 } while (pte++, addr += PAGE_SIZE, addr != end);
932 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
933 unsigned long addr, unsigned long end,
934 swp_entry_t entry, struct page *page)
940 pmd = pmd_offset(pud, addr);
942 next = pmd_addr_end(addr, end);
943 if (pmd_none_or_clear_bad(pmd))
945 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
948 } while (pmd++, addr = next, addr != end);
952 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
953 unsigned long addr, unsigned long end,
954 swp_entry_t entry, struct page *page)
960 pud = pud_offset(pgd, addr);
962 next = pud_addr_end(addr, end);
963 if (pud_none_or_clear_bad(pud))
965 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
968 } while (pud++, addr = next, addr != end);
972 static int unuse_vma(struct vm_area_struct *vma,
973 swp_entry_t entry, struct page *page)
976 unsigned long addr, end, next;
979 if (page_anon_vma(page)) {
980 addr = page_address_in_vma(page, vma);
984 end = addr + PAGE_SIZE;
986 addr = vma->vm_start;
990 pgd = pgd_offset(vma->vm_mm, addr);
992 next = pgd_addr_end(addr, end);
993 if (pgd_none_or_clear_bad(pgd))
995 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
998 } while (pgd++, addr = next, addr != end);
1002 static int unuse_mm(struct mm_struct *mm,
1003 swp_entry_t entry, struct page *page)
1005 struct vm_area_struct *vma;
1008 if (!down_read_trylock(&mm->mmap_sem)) {
1010 * Activate page so shrink_inactive_list is unlikely to unmap
1011 * its ptes while lock is dropped, so swapoff can make progress.
1013 activate_page(page);
1015 down_read(&mm->mmap_sem);
1018 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1019 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1022 up_read(&mm->mmap_sem);
1023 return (ret < 0)? ret: 0;
1027 * Scan swap_map from current position to next entry still in use.
1028 * Recycle to start on reaching the end, returning 0 when empty.
1030 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1033 unsigned int max = si->max;
1034 unsigned int i = prev;
1035 unsigned char count;
1038 * No need for swap_lock here: we're just looking
1039 * for whether an entry is in use, not modifying it; false
1040 * hits are okay, and sys_swapoff() has already prevented new
1041 * allocations from this area (while holding swap_lock).
1050 * No entries in use at top of swap_map,
1051 * loop back to start and recheck there.
1057 count = si->swap_map[i];
1058 if (count && swap_count(count) != SWAP_MAP_BAD)
1065 * We completely avoid races by reading each swap page in advance,
1066 * and then search for the process using it. All the necessary
1067 * page table adjustments can then be made atomically.
1069 static int try_to_unuse(unsigned int type)
1071 struct swap_info_struct *si = swap_info[type];
1072 struct mm_struct *start_mm;
1073 unsigned char *swap_map;
1074 unsigned char swcount;
1081 * When searching mms for an entry, a good strategy is to
1082 * start at the first mm we freed the previous entry from
1083 * (though actually we don't notice whether we or coincidence
1084 * freed the entry). Initialize this start_mm with a hold.
1086 * A simpler strategy would be to start at the last mm we
1087 * freed the previous entry from; but that would take less
1088 * advantage of mmlist ordering, which clusters forked mms
1089 * together, child after parent. If we race with dup_mmap(), we
1090 * prefer to resolve parent before child, lest we miss entries
1091 * duplicated after we scanned child: using last mm would invert
1094 start_mm = &init_mm;
1095 atomic_inc(&init_mm.mm_users);
1098 * Keep on scanning until all entries have gone. Usually,
1099 * one pass through swap_map is enough, but not necessarily:
1100 * there are races when an instance of an entry might be missed.
1102 while ((i = find_next_to_unuse(si, i)) != 0) {
1103 if (signal_pending(current)) {
1109 * Get a page for the entry, using the existing swap
1110 * cache page if there is one. Otherwise, get a clean
1111 * page and read the swap into it.
1113 swap_map = &si->swap_map[i];
1114 entry = swp_entry(type, i);
1115 page = read_swap_cache_async(entry,
1116 GFP_HIGHUSER_MOVABLE, NULL, 0);
1119 * Either swap_duplicate() failed because entry
1120 * has been freed independently, and will not be
1121 * reused since sys_swapoff() already disabled
1122 * allocation from here, or alloc_page() failed.
1131 * Don't hold on to start_mm if it looks like exiting.
1133 if (atomic_read(&start_mm->mm_users) == 1) {
1135 start_mm = &init_mm;
1136 atomic_inc(&init_mm.mm_users);
1140 * Wait for and lock page. When do_swap_page races with
1141 * try_to_unuse, do_swap_page can handle the fault much
1142 * faster than try_to_unuse can locate the entry. This
1143 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1144 * defer to do_swap_page in such a case - in some tests,
1145 * do_swap_page and try_to_unuse repeatedly compete.
1147 wait_on_page_locked(page);
1148 wait_on_page_writeback(page);
1150 wait_on_page_writeback(page);
1153 * Remove all references to entry.
1155 swcount = *swap_map;
1156 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1157 retval = shmem_unuse(entry, page);
1158 /* page has already been unlocked and released */
1163 if (swap_count(swcount) && start_mm != &init_mm)
1164 retval = unuse_mm(start_mm, entry, page);
1166 if (swap_count(*swap_map)) {
1167 int set_start_mm = (*swap_map >= swcount);
1168 struct list_head *p = &start_mm->mmlist;
1169 struct mm_struct *new_start_mm = start_mm;
1170 struct mm_struct *prev_mm = start_mm;
1171 struct mm_struct *mm;
1173 atomic_inc(&new_start_mm->mm_users);
1174 atomic_inc(&prev_mm->mm_users);
1175 spin_lock(&mmlist_lock);
1176 while (swap_count(*swap_map) && !retval &&
1177 (p = p->next) != &start_mm->mmlist) {
1178 mm = list_entry(p, struct mm_struct, mmlist);
1179 if (!atomic_inc_not_zero(&mm->mm_users))
1181 spin_unlock(&mmlist_lock);
1187 swcount = *swap_map;
1188 if (!swap_count(swcount)) /* any usage ? */
1190 else if (mm == &init_mm)
1193 retval = unuse_mm(mm, entry, page);
1195 if (set_start_mm && *swap_map < swcount) {
1196 mmput(new_start_mm);
1197 atomic_inc(&mm->mm_users);
1201 spin_lock(&mmlist_lock);
1203 spin_unlock(&mmlist_lock);
1206 start_mm = new_start_mm;
1210 page_cache_release(page);
1215 * If a reference remains (rare), we would like to leave
1216 * the page in the swap cache; but try_to_unmap could
1217 * then re-duplicate the entry once we drop page lock,
1218 * so we might loop indefinitely; also, that page could
1219 * not be swapped out to other storage meanwhile. So:
1220 * delete from cache even if there's another reference,
1221 * after ensuring that the data has been saved to disk -
1222 * since if the reference remains (rarer), it will be
1223 * read from disk into another page. Splitting into two
1224 * pages would be incorrect if swap supported "shared
1225 * private" pages, but they are handled by tmpfs files.
1227 * Given how unuse_vma() targets one particular offset
1228 * in an anon_vma, once the anon_vma has been determined,
1229 * this splitting happens to be just what is needed to
1230 * handle where KSM pages have been swapped out: re-reading
1231 * is unnecessarily slow, but we can fix that later on.
1233 if (swap_count(*swap_map) &&
1234 PageDirty(page) && PageSwapCache(page)) {
1235 struct writeback_control wbc = {
1236 .sync_mode = WB_SYNC_NONE,
1239 swap_writepage(page, &wbc);
1241 wait_on_page_writeback(page);
1245 * It is conceivable that a racing task removed this page from
1246 * swap cache just before we acquired the page lock at the top,
1247 * or while we dropped it in unuse_mm(). The page might even
1248 * be back in swap cache on another swap area: that we must not
1249 * delete, since it may not have been written out to swap yet.
1251 if (PageSwapCache(page) &&
1252 likely(page_private(page) == entry.val))
1253 delete_from_swap_cache(page);
1256 * So we could skip searching mms once swap count went
1257 * to 1, we did not mark any present ptes as dirty: must
1258 * mark page dirty so shrink_page_list will preserve it.
1262 page_cache_release(page);
1265 * Make sure that we aren't completely killing
1266 * interactive performance.
1276 * After a successful try_to_unuse, if no swap is now in use, we know
1277 * we can empty the mmlist. swap_lock must be held on entry and exit.
1278 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1279 * added to the mmlist just after page_duplicate - before would be racy.
1281 static void drain_mmlist(void)
1283 struct list_head *p, *next;
1286 for (type = 0; type < nr_swapfiles; type++)
1287 if (swap_info[type]->inuse_pages)
1289 spin_lock(&mmlist_lock);
1290 list_for_each_safe(p, next, &init_mm.mmlist)
1292 spin_unlock(&mmlist_lock);
1296 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1297 * corresponds to page offset for the specified swap entry.
1298 * Note that the type of this function is sector_t, but it returns page offset
1299 * into the bdev, not sector offset.
1301 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1303 struct swap_info_struct *sis;
1304 struct swap_extent *start_se;
1305 struct swap_extent *se;
1308 sis = swap_info[swp_type(entry)];
1311 offset = swp_offset(entry);
1312 start_se = sis->curr_swap_extent;
1316 struct list_head *lh;
1318 if (se->start_page <= offset &&
1319 offset < (se->start_page + se->nr_pages)) {
1320 return se->start_block + (offset - se->start_page);
1323 se = list_entry(lh, struct swap_extent, list);
1324 sis->curr_swap_extent = se;
1325 BUG_ON(se == start_se); /* It *must* be present */
1330 * Returns the page offset into bdev for the specified page's swap entry.
1332 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1335 entry.val = page_private(page);
1336 return map_swap_entry(entry, bdev);
1340 * Free all of a swapdev's extent information
1342 static void destroy_swap_extents(struct swap_info_struct *sis)
1344 while (!list_empty(&sis->first_swap_extent.list)) {
1345 struct swap_extent *se;
1347 se = list_entry(sis->first_swap_extent.list.next,
1348 struct swap_extent, list);
1349 list_del(&se->list);
1355 * Add a block range (and the corresponding page range) into this swapdev's
1356 * extent list. The extent list is kept sorted in page order.
1358 * This function rather assumes that it is called in ascending page order.
1361 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1362 unsigned long nr_pages, sector_t start_block)
1364 struct swap_extent *se;
1365 struct swap_extent *new_se;
1366 struct list_head *lh;
1368 if (start_page == 0) {
1369 se = &sis->first_swap_extent;
1370 sis->curr_swap_extent = se;
1372 se->nr_pages = nr_pages;
1373 se->start_block = start_block;
1376 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1377 se = list_entry(lh, struct swap_extent, list);
1378 BUG_ON(se->start_page + se->nr_pages != start_page);
1379 if (se->start_block + se->nr_pages == start_block) {
1381 se->nr_pages += nr_pages;
1387 * No merge. Insert a new extent, preserving ordering.
1389 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1392 new_se->start_page = start_page;
1393 new_se->nr_pages = nr_pages;
1394 new_se->start_block = start_block;
1396 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1401 * A `swap extent' is a simple thing which maps a contiguous range of pages
1402 * onto a contiguous range of disk blocks. An ordered list of swap extents
1403 * is built at swapon time and is then used at swap_writepage/swap_readpage
1404 * time for locating where on disk a page belongs.
1406 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1407 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1408 * swap files identically.
1410 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1411 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1412 * swapfiles are handled *identically* after swapon time.
1414 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1415 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1416 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1417 * requirements, they are simply tossed out - we will never use those blocks
1420 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1421 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1422 * which will scribble on the fs.
1424 * The amount of disk space which a single swap extent represents varies.
1425 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1426 * extents in the list. To avoid much list walking, we cache the previous
1427 * search location in `curr_swap_extent', and start new searches from there.
1428 * This is extremely effective. The average number of iterations in
1429 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1431 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1433 struct inode *inode;
1434 unsigned blocks_per_page;
1435 unsigned long page_no;
1437 sector_t probe_block;
1438 sector_t last_block;
1439 sector_t lowest_block = -1;
1440 sector_t highest_block = 0;
1444 inode = sis->swap_file->f_mapping->host;
1445 if (S_ISBLK(inode->i_mode)) {
1446 ret = add_swap_extent(sis, 0, sis->max, 0);
1451 blkbits = inode->i_blkbits;
1452 blocks_per_page = PAGE_SIZE >> blkbits;
1455 * Map all the blocks into the extent list. This code doesn't try
1460 last_block = i_size_read(inode) >> blkbits;
1461 while ((probe_block + blocks_per_page) <= last_block &&
1462 page_no < sis->max) {
1463 unsigned block_in_page;
1464 sector_t first_block;
1466 first_block = bmap(inode, probe_block);
1467 if (first_block == 0)
1471 * It must be PAGE_SIZE aligned on-disk
1473 if (first_block & (blocks_per_page - 1)) {
1478 for (block_in_page = 1; block_in_page < blocks_per_page;
1482 block = bmap(inode, probe_block + block_in_page);
1485 if (block != first_block + block_in_page) {
1492 first_block >>= (PAGE_SHIFT - blkbits);
1493 if (page_no) { /* exclude the header page */
1494 if (first_block < lowest_block)
1495 lowest_block = first_block;
1496 if (first_block > highest_block)
1497 highest_block = first_block;
1501 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1503 ret = add_swap_extent(sis, page_no, 1, first_block);
1508 probe_block += blocks_per_page;
1513 *span = 1 + highest_block - lowest_block;
1515 page_no = 1; /* force Empty message */
1517 sis->pages = page_no - 1;
1518 sis->highest_bit = page_no - 1;
1522 printk(KERN_ERR "swapon: swapfile has holes\n");
1527 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1529 struct swap_info_struct *p = NULL;
1530 unsigned char *swap_map;
1531 struct file *swap_file, *victim;
1532 struct address_space *mapping;
1533 struct inode *inode;
1538 if (!capable(CAP_SYS_ADMIN))
1541 pathname = getname(specialfile);
1542 err = PTR_ERR(pathname);
1543 if (IS_ERR(pathname))
1546 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1548 err = PTR_ERR(victim);
1552 mapping = victim->f_mapping;
1554 spin_lock(&swap_lock);
1555 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1556 p = swap_info[type];
1557 if (p->flags & SWP_WRITEOK) {
1558 if (p->swap_file->f_mapping == mapping)
1565 spin_unlock(&swap_lock);
1568 if (!security_vm_enough_memory(p->pages))
1569 vm_unacct_memory(p->pages);
1572 spin_unlock(&swap_lock);
1576 swap_list.head = p->next;
1578 swap_info[prev]->next = p->next;
1579 if (type == swap_list.next) {
1580 /* just pick something that's safe... */
1581 swap_list.next = swap_list.head;
1584 for (i = p->next; i >= 0; i = swap_info[i]->next)
1585 swap_info[i]->prio = p->prio--;
1588 nr_swap_pages -= p->pages;
1589 total_swap_pages -= p->pages;
1590 p->flags &= ~SWP_WRITEOK;
1591 spin_unlock(&swap_lock);
1593 current->flags |= PF_OOM_ORIGIN;
1594 err = try_to_unuse(type);
1595 current->flags &= ~PF_OOM_ORIGIN;
1598 /* re-insert swap space back into swap_list */
1599 spin_lock(&swap_lock);
1601 p->prio = --least_priority;
1603 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1604 if (p->prio >= swap_info[i]->prio)
1610 swap_list.head = swap_list.next = type;
1612 swap_info[prev]->next = type;
1613 nr_swap_pages += p->pages;
1614 total_swap_pages += p->pages;
1615 p->flags |= SWP_WRITEOK;
1616 spin_unlock(&swap_lock);
1620 /* wait for any unplug function to finish */
1621 down_write(&swap_unplug_sem);
1622 up_write(&swap_unplug_sem);
1624 destroy_swap_extents(p);
1625 if (p->flags & SWP_CONTINUED)
1626 free_swap_count_continuations(p);
1628 mutex_lock(&swapon_mutex);
1629 spin_lock(&swap_lock);
1632 /* wait for anyone still in scan_swap_map */
1633 p->highest_bit = 0; /* cuts scans short */
1634 while (p->flags >= SWP_SCANNING) {
1635 spin_unlock(&swap_lock);
1636 schedule_timeout_uninterruptible(1);
1637 spin_lock(&swap_lock);
1640 swap_file = p->swap_file;
1641 p->swap_file = NULL;
1643 swap_map = p->swap_map;
1646 spin_unlock(&swap_lock);
1647 mutex_unlock(&swapon_mutex);
1649 /* Destroy swap account informatin */
1650 swap_cgroup_swapoff(type);
1652 inode = mapping->host;
1653 if (S_ISBLK(inode->i_mode)) {
1654 struct block_device *bdev = I_BDEV(inode);
1655 set_blocksize(bdev, p->old_block_size);
1658 mutex_lock(&inode->i_mutex);
1659 inode->i_flags &= ~S_SWAPFILE;
1660 mutex_unlock(&inode->i_mutex);
1662 filp_close(swap_file, NULL);
1666 filp_close(victim, NULL);
1671 #ifdef CONFIG_PROC_FS
1673 static void *swap_start(struct seq_file *swap, loff_t *pos)
1675 struct swap_info_struct *si;
1679 mutex_lock(&swapon_mutex);
1682 return SEQ_START_TOKEN;
1684 for (type = 0; type < nr_swapfiles; type++) {
1685 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1686 si = swap_info[type];
1687 if (!(si->flags & SWP_USED) || !si->swap_map)
1696 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1698 struct swap_info_struct *si = v;
1701 if (v == SEQ_START_TOKEN)
1704 type = si->type + 1;
1706 for (; type < nr_swapfiles; type++) {
1707 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1708 si = swap_info[type];
1709 if (!(si->flags & SWP_USED) || !si->swap_map)
1718 static void swap_stop(struct seq_file *swap, void *v)
1720 mutex_unlock(&swapon_mutex);
1723 static int swap_show(struct seq_file *swap, void *v)
1725 struct swap_info_struct *si = v;
1729 if (si == SEQ_START_TOKEN) {
1730 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1734 file = si->swap_file;
1735 len = seq_path(swap, &file->f_path, " \t\n\\");
1736 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1737 len < 40 ? 40 - len : 1, " ",
1738 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1739 "partition" : "file\t",
1740 si->pages << (PAGE_SHIFT - 10),
1741 si->inuse_pages << (PAGE_SHIFT - 10),
1746 static const struct seq_operations swaps_op = {
1747 .start = swap_start,
1753 static int swaps_open(struct inode *inode, struct file *file)
1755 return seq_open(file, &swaps_op);
1758 static const struct file_operations proc_swaps_operations = {
1761 .llseek = seq_lseek,
1762 .release = seq_release,
1765 static int __init procswaps_init(void)
1767 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1770 __initcall(procswaps_init);
1771 #endif /* CONFIG_PROC_FS */
1773 #ifdef MAX_SWAPFILES_CHECK
1774 static int __init max_swapfiles_check(void)
1776 MAX_SWAPFILES_CHECK();
1779 late_initcall(max_swapfiles_check);
1783 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1785 * The swapon system call
1787 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1789 struct swap_info_struct *p;
1791 struct block_device *bdev = NULL;
1792 struct file *swap_file = NULL;
1793 struct address_space *mapping;
1797 union swap_header *swap_header;
1798 unsigned int nr_good_pages;
1801 unsigned long maxpages;
1802 unsigned long swapfilepages;
1803 unsigned char *swap_map = NULL;
1804 struct page *page = NULL;
1805 struct inode *inode = NULL;
1808 if (!capable(CAP_SYS_ADMIN))
1811 p = kzalloc(sizeof(*p), GFP_KERNEL);
1815 spin_lock(&swap_lock);
1816 for (type = 0; type < nr_swapfiles; type++) {
1817 if (!(swap_info[type]->flags & SWP_USED))
1821 if (type >= MAX_SWAPFILES) {
1822 spin_unlock(&swap_lock);
1826 if (type >= nr_swapfiles) {
1828 swap_info[type] = p;
1830 * Write swap_info[type] before nr_swapfiles, in case a
1831 * racing procfs swap_start() or swap_next() is reading them.
1832 * (We never shrink nr_swapfiles, we never free this entry.)
1838 p = swap_info[type];
1840 * Do not memset this entry: a racing procfs swap_next()
1841 * would be relying on p->type to remain valid.
1844 INIT_LIST_HEAD(&p->first_swap_extent.list);
1845 p->flags = SWP_USED;
1847 spin_unlock(&swap_lock);
1849 name = getname(specialfile);
1850 error = PTR_ERR(name);
1855 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1856 error = PTR_ERR(swap_file);
1857 if (IS_ERR(swap_file)) {
1862 p->swap_file = swap_file;
1863 mapping = swap_file->f_mapping;
1864 inode = mapping->host;
1867 for (i = 0; i < nr_swapfiles; i++) {
1868 struct swap_info_struct *q = swap_info[i];
1870 if (i == type || !q->swap_file)
1872 if (mapping == q->swap_file->f_mapping)
1877 if (S_ISBLK(inode->i_mode)) {
1878 bdev = I_BDEV(inode);
1879 error = bd_claim(bdev, sys_swapon);
1885 p->old_block_size = block_size(bdev);
1886 error = set_blocksize(bdev, PAGE_SIZE);
1890 } else if (S_ISREG(inode->i_mode)) {
1891 p->bdev = inode->i_sb->s_bdev;
1892 mutex_lock(&inode->i_mutex);
1894 if (IS_SWAPFILE(inode)) {
1902 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1905 * Read the swap header.
1907 if (!mapping->a_ops->readpage) {
1911 page = read_mapping_page(mapping, 0, swap_file);
1913 error = PTR_ERR(page);
1916 swap_header = kmap(page);
1918 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1919 printk(KERN_ERR "Unable to find swap-space signature\n");
1924 /* swap partition endianess hack... */
1925 if (swab32(swap_header->info.version) == 1) {
1926 swab32s(&swap_header->info.version);
1927 swab32s(&swap_header->info.last_page);
1928 swab32s(&swap_header->info.nr_badpages);
1929 for (i = 0; i < swap_header->info.nr_badpages; i++)
1930 swab32s(&swap_header->info.badpages[i]);
1932 /* Check the swap header's sub-version */
1933 if (swap_header->info.version != 1) {
1935 "Unable to handle swap header version %d\n",
1936 swap_header->info.version);
1942 p->cluster_next = 1;
1946 * Find out how many pages are allowed for a single swap
1947 * device. There are two limiting factors: 1) the number of
1948 * bits for the swap offset in the swp_entry_t type and
1949 * 2) the number of bits in the a swap pte as defined by
1950 * the different architectures. In order to find the
1951 * largest possible bit mask a swap entry with swap type 0
1952 * and swap offset ~0UL is created, encoded to a swap pte,
1953 * decoded to a swp_entry_t again and finally the swap
1954 * offset is extracted. This will mask all the bits from
1955 * the initial ~0UL mask that can't be encoded in either
1956 * the swp_entry_t or the architecture definition of a
1959 maxpages = swp_offset(pte_to_swp_entry(
1960 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1961 if (maxpages > swap_header->info.last_page) {
1962 maxpages = swap_header->info.last_page + 1;
1963 /* p->max is an unsigned int: don't overflow it */
1964 if ((unsigned int)maxpages == 0)
1965 maxpages = UINT_MAX;
1967 p->highest_bit = maxpages - 1;
1972 if (swapfilepages && maxpages > swapfilepages) {
1974 "Swap area shorter than signature indicates\n");
1977 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1979 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1982 /* OK, set up the swap map and apply the bad block list */
1983 swap_map = vmalloc(maxpages);
1989 memset(swap_map, 0, maxpages);
1990 nr_good_pages = maxpages - 1; /* omit header page */
1992 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1993 unsigned int page_nr = swap_header->info.badpages[i];
1994 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
1998 if (page_nr < maxpages) {
1999 swap_map[page_nr] = SWAP_MAP_BAD;
2004 error = swap_cgroup_swapon(type, maxpages);
2008 if (nr_good_pages) {
2009 swap_map[0] = SWAP_MAP_BAD;
2011 p->pages = nr_good_pages;
2012 nr_extents = setup_swap_extents(p, &span);
2013 if (nr_extents < 0) {
2017 nr_good_pages = p->pages;
2019 if (!nr_good_pages) {
2020 printk(KERN_WARNING "Empty swap-file\n");
2026 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2027 p->flags |= SWP_SOLIDSTATE;
2028 p->cluster_next = 1 + (random32() % p->highest_bit);
2030 if (discard_swap(p) == 0)
2031 p->flags |= SWP_DISCARDABLE;
2034 mutex_lock(&swapon_mutex);
2035 spin_lock(&swap_lock);
2036 if (swap_flags & SWAP_FLAG_PREFER)
2038 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2040 p->prio = --least_priority;
2041 p->swap_map = swap_map;
2042 p->flags |= SWP_WRITEOK;
2043 nr_swap_pages += nr_good_pages;
2044 total_swap_pages += nr_good_pages;
2046 printk(KERN_INFO "Adding %uk swap on %s. "
2047 "Priority:%d extents:%d across:%lluk %s%s\n",
2048 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2049 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2050 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2051 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2053 /* insert swap space into swap_list: */
2055 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2056 if (p->prio >= swap_info[i]->prio)
2062 swap_list.head = swap_list.next = type;
2064 swap_info[prev]->next = type;
2065 spin_unlock(&swap_lock);
2066 mutex_unlock(&swapon_mutex);
2071 set_blocksize(bdev, p->old_block_size);
2074 destroy_swap_extents(p);
2075 swap_cgroup_swapoff(type);
2077 spin_lock(&swap_lock);
2078 p->swap_file = NULL;
2080 spin_unlock(&swap_lock);
2083 filp_close(swap_file, NULL);
2085 if (page && !IS_ERR(page)) {
2087 page_cache_release(page);
2093 inode->i_flags |= S_SWAPFILE;
2094 mutex_unlock(&inode->i_mutex);
2099 void si_swapinfo(struct sysinfo *val)
2102 unsigned long nr_to_be_unused = 0;
2104 spin_lock(&swap_lock);
2105 for (type = 0; type < nr_swapfiles; type++) {
2106 struct swap_info_struct *si = swap_info[type];
2108 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2109 nr_to_be_unused += si->inuse_pages;
2111 val->freeswap = nr_swap_pages + nr_to_be_unused;
2112 val->totalswap = total_swap_pages + nr_to_be_unused;
2113 spin_unlock(&swap_lock);
2117 * Verify that a swap entry is valid and increment its swap map count.
2119 * Returns error code in following case.
2121 * - swp_entry is invalid -> EINVAL
2122 * - swp_entry is migration entry -> EINVAL
2123 * - swap-cache reference is requested but there is already one. -> EEXIST
2124 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2125 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2127 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2129 struct swap_info_struct *p;
2130 unsigned long offset, type;
2131 unsigned char count;
2132 unsigned char has_cache;
2135 if (non_swap_entry(entry))
2138 type = swp_type(entry);
2139 if (type >= nr_swapfiles)
2141 p = swap_info[type];
2142 offset = swp_offset(entry);
2144 spin_lock(&swap_lock);
2145 if (unlikely(offset >= p->max))
2148 count = p->swap_map[offset];
2149 has_cache = count & SWAP_HAS_CACHE;
2150 count &= ~SWAP_HAS_CACHE;
2153 if (usage == SWAP_HAS_CACHE) {
2155 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2156 if (!has_cache && count)
2157 has_cache = SWAP_HAS_CACHE;
2158 else if (has_cache) /* someone else added cache */
2160 else /* no users remaining */
2163 } else if (count || has_cache) {
2165 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2167 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2169 else if (swap_count_continued(p, offset, count))
2170 count = COUNT_CONTINUED;
2174 err = -ENOENT; /* unused swap entry */
2176 p->swap_map[offset] = count | has_cache;
2179 spin_unlock(&swap_lock);
2184 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2189 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2190 * (in which case its reference count is never incremented).
2192 void swap_shmem_alloc(swp_entry_t entry)
2194 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2198 * Increase reference count of swap entry by 1.
2199 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2200 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2201 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2202 * might occur if a page table entry has got corrupted.
2204 int swap_duplicate(swp_entry_t entry)
2208 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2209 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2214 * @entry: swap entry for which we allocate swap cache.
2216 * Called when allocating swap cache for existing swap entry,
2217 * This can return error codes. Returns 0 at success.
2218 * -EBUSY means there is a swap cache.
2219 * Note: return code is different from swap_duplicate().
2221 int swapcache_prepare(swp_entry_t entry)
2223 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2227 * swap_lock prevents swap_map being freed. Don't grab an extra
2228 * reference on the swaphandle, it doesn't matter if it becomes unused.
2230 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2232 struct swap_info_struct *si;
2233 int our_page_cluster = page_cluster;
2234 pgoff_t target, toff;
2238 if (!our_page_cluster) /* no readahead */
2241 si = swap_info[swp_type(entry)];
2242 target = swp_offset(entry);
2243 base = (target >> our_page_cluster) << our_page_cluster;
2244 end = base + (1 << our_page_cluster);
2245 if (!base) /* first page is swap header */
2248 spin_lock(&swap_lock);
2249 if (end > si->max) /* don't go beyond end of map */
2252 /* Count contiguous allocated slots above our target */
2253 for (toff = target; ++toff < end; nr_pages++) {
2254 /* Don't read in free or bad pages */
2255 if (!si->swap_map[toff])
2257 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2260 /* Count contiguous allocated slots below our target */
2261 for (toff = target; --toff >= base; nr_pages++) {
2262 /* Don't read in free or bad pages */
2263 if (!si->swap_map[toff])
2265 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2268 spin_unlock(&swap_lock);
2271 * Indicate starting offset, and return number of pages to get:
2272 * if only 1, say 0, since there's then no readahead to be done.
2275 return nr_pages? ++nr_pages: 0;
2279 * add_swap_count_continuation - called when a swap count is duplicated
2280 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2281 * page of the original vmalloc'ed swap_map, to hold the continuation count
2282 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2283 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2285 * These continuation pages are seldom referenced: the common paths all work
2286 * on the original swap_map, only referring to a continuation page when the
2287 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2289 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2290 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2291 * can be called after dropping locks.
2293 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2295 struct swap_info_struct *si;
2298 struct page *list_page;
2300 unsigned char count;
2303 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2304 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2306 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2308 si = swap_info_get(entry);
2311 * An acceptable race has occurred since the failing
2312 * __swap_duplicate(): the swap entry has been freed,
2313 * perhaps even the whole swap_map cleared for swapoff.
2318 offset = swp_offset(entry);
2319 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2321 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2323 * The higher the swap count, the more likely it is that tasks
2324 * will race to add swap count continuation: we need to avoid
2325 * over-provisioning.
2331 spin_unlock(&swap_lock);
2336 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2337 * no architecture is using highmem pages for kernel pagetables: so it
2338 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2340 head = vmalloc_to_page(si->swap_map + offset);
2341 offset &= ~PAGE_MASK;
2344 * Page allocation does not initialize the page's lru field,
2345 * but it does always reset its private field.
2347 if (!page_private(head)) {
2348 BUG_ON(count & COUNT_CONTINUED);
2349 INIT_LIST_HEAD(&head->lru);
2350 set_page_private(head, SWP_CONTINUED);
2351 si->flags |= SWP_CONTINUED;
2354 list_for_each_entry(list_page, &head->lru, lru) {
2358 * If the previous map said no continuation, but we've found
2359 * a continuation page, free our allocation and use this one.
2361 if (!(count & COUNT_CONTINUED))
2364 map = kmap_atomic(list_page, KM_USER0) + offset;
2366 kunmap_atomic(map, KM_USER0);
2369 * If this continuation count now has some space in it,
2370 * free our allocation and use this one.
2372 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2376 list_add_tail(&page->lru, &head->lru);
2377 page = NULL; /* now it's attached, don't free it */
2379 spin_unlock(&swap_lock);
2387 * swap_count_continued - when the original swap_map count is incremented
2388 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2389 * into, carry if so, or else fail until a new continuation page is allocated;
2390 * when the original swap_map count is decremented from 0 with continuation,
2391 * borrow from the continuation and report whether it still holds more.
2392 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2394 static bool swap_count_continued(struct swap_info_struct *si,
2395 pgoff_t offset, unsigned char count)
2401 head = vmalloc_to_page(si->swap_map + offset);
2402 if (page_private(head) != SWP_CONTINUED) {
2403 BUG_ON(count & COUNT_CONTINUED);
2404 return false; /* need to add count continuation */
2407 offset &= ~PAGE_MASK;
2408 page = list_entry(head->lru.next, struct page, lru);
2409 map = kmap_atomic(page, KM_USER0) + offset;
2411 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2412 goto init_map; /* jump over SWAP_CONT_MAX checks */
2414 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2416 * Think of how you add 1 to 999
2418 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2419 kunmap_atomic(map, KM_USER0);
2420 page = list_entry(page->lru.next, struct page, lru);
2421 BUG_ON(page == head);
2422 map = kmap_atomic(page, KM_USER0) + offset;
2424 if (*map == SWAP_CONT_MAX) {
2425 kunmap_atomic(map, KM_USER0);
2426 page = list_entry(page->lru.next, struct page, lru);
2428 return false; /* add count continuation */
2429 map = kmap_atomic(page, KM_USER0) + offset;
2430 init_map: *map = 0; /* we didn't zero the page */
2433 kunmap_atomic(map, KM_USER0);
2434 page = list_entry(page->lru.prev, struct page, lru);
2435 while (page != head) {
2436 map = kmap_atomic(page, KM_USER0) + offset;
2437 *map = COUNT_CONTINUED;
2438 kunmap_atomic(map, KM_USER0);
2439 page = list_entry(page->lru.prev, struct page, lru);
2441 return true; /* incremented */
2443 } else { /* decrementing */
2445 * Think of how you subtract 1 from 1000
2447 BUG_ON(count != COUNT_CONTINUED);
2448 while (*map == COUNT_CONTINUED) {
2449 kunmap_atomic(map, KM_USER0);
2450 page = list_entry(page->lru.next, struct page, lru);
2451 BUG_ON(page == head);
2452 map = kmap_atomic(page, KM_USER0) + offset;
2458 kunmap_atomic(map, KM_USER0);
2459 page = list_entry(page->lru.prev, struct page, lru);
2460 while (page != head) {
2461 map = kmap_atomic(page, KM_USER0) + offset;
2462 *map = SWAP_CONT_MAX | count;
2463 count = COUNT_CONTINUED;
2464 kunmap_atomic(map, KM_USER0);
2465 page = list_entry(page->lru.prev, struct page, lru);
2467 return count == COUNT_CONTINUED;
2472 * free_swap_count_continuations - swapoff free all the continuation pages
2473 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2475 static void free_swap_count_continuations(struct swap_info_struct *si)
2479 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2481 head = vmalloc_to_page(si->swap_map + offset);
2482 if (page_private(head)) {
2483 struct list_head *this, *next;
2484 list_for_each_safe(this, next, &head->lru) {
2486 page = list_entry(this, struct page, lru);