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, DISCARD_FL_BARRIER);
148 list_for_each_entry(se, &si->first_swap_extent.list, list) {
149 start_block = se->start_block << (PAGE_SHIFT - 9);
150 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
152 err = blkdev_issue_discard(si->bdev, start_block,
153 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
159 return err; /* That will often be -EOPNOTSUPP */
163 * swap allocation tell device that a cluster of swap can now be discarded,
164 * to allow the swap device to optimize its wear-levelling.
166 static void discard_swap_cluster(struct swap_info_struct *si,
167 pgoff_t start_page, pgoff_t nr_pages)
169 struct swap_extent *se = si->curr_swap_extent;
170 int found_extent = 0;
173 struct list_head *lh;
175 if (se->start_page <= start_page &&
176 start_page < se->start_page + se->nr_pages) {
177 pgoff_t offset = start_page - se->start_page;
178 sector_t start_block = se->start_block + offset;
179 sector_t nr_blocks = se->nr_pages - offset;
181 if (nr_blocks > nr_pages)
182 nr_blocks = nr_pages;
183 start_page += nr_blocks;
184 nr_pages -= nr_blocks;
187 si->curr_swap_extent = se;
189 start_block <<= PAGE_SHIFT - 9;
190 nr_blocks <<= PAGE_SHIFT - 9;
191 if (blkdev_issue_discard(si->bdev, start_block,
192 nr_blocks, GFP_NOIO, DISCARD_FL_BARRIER))
197 se = list_entry(lh, struct swap_extent, list);
201 static int wait_for_discard(void *word)
207 #define SWAPFILE_CLUSTER 256
208 #define LATENCY_LIMIT 256
210 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
213 unsigned long offset;
214 unsigned long scan_base;
215 unsigned long last_in_cluster = 0;
216 int latency_ration = LATENCY_LIMIT;
217 int found_free_cluster = 0;
220 * We try to cluster swap pages by allocating them sequentially
221 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
222 * way, however, we resort to first-free allocation, starting
223 * a new cluster. This prevents us from scattering swap pages
224 * all over the entire swap partition, so that we reduce
225 * overall disk seek times between swap pages. -- sct
226 * But we do now try to find an empty cluster. -Andrea
227 * And we let swap pages go all over an SSD partition. Hugh
230 si->flags += SWP_SCANNING;
231 scan_base = offset = si->cluster_next;
233 if (unlikely(!si->cluster_nr--)) {
234 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
235 si->cluster_nr = SWAPFILE_CLUSTER - 1;
238 if (si->flags & SWP_DISCARDABLE) {
240 * Start range check on racing allocations, in case
241 * they overlap the cluster we eventually decide on
242 * (we scan without swap_lock to allow preemption).
243 * It's hardly conceivable that cluster_nr could be
244 * wrapped during our scan, but don't depend on it.
246 if (si->lowest_alloc)
248 si->lowest_alloc = si->max;
249 si->highest_alloc = 0;
251 spin_unlock(&swap_lock);
254 * If seek is expensive, start searching for new cluster from
255 * start of partition, to minimize the span of allocated swap.
256 * But if seek is cheap, search from our current position, so
257 * that swap is allocated from all over the partition: if the
258 * Flash Translation Layer only remaps within limited zones,
259 * we don't want to wear out the first zone too quickly.
261 if (!(si->flags & SWP_SOLIDSTATE))
262 scan_base = offset = si->lowest_bit;
263 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
265 /* Locate the first empty (unaligned) cluster */
266 for (; last_in_cluster <= si->highest_bit; offset++) {
267 if (si->swap_map[offset])
268 last_in_cluster = offset + SWAPFILE_CLUSTER;
269 else if (offset == last_in_cluster) {
270 spin_lock(&swap_lock);
271 offset -= SWAPFILE_CLUSTER - 1;
272 si->cluster_next = offset;
273 si->cluster_nr = SWAPFILE_CLUSTER - 1;
274 found_free_cluster = 1;
277 if (unlikely(--latency_ration < 0)) {
279 latency_ration = LATENCY_LIMIT;
283 offset = si->lowest_bit;
284 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
286 /* Locate the first empty (unaligned) cluster */
287 for (; last_in_cluster < scan_base; offset++) {
288 if (si->swap_map[offset])
289 last_in_cluster = offset + SWAPFILE_CLUSTER;
290 else if (offset == last_in_cluster) {
291 spin_lock(&swap_lock);
292 offset -= SWAPFILE_CLUSTER - 1;
293 si->cluster_next = offset;
294 si->cluster_nr = SWAPFILE_CLUSTER - 1;
295 found_free_cluster = 1;
298 if (unlikely(--latency_ration < 0)) {
300 latency_ration = LATENCY_LIMIT;
305 spin_lock(&swap_lock);
306 si->cluster_nr = SWAPFILE_CLUSTER - 1;
307 si->lowest_alloc = 0;
311 if (!(si->flags & SWP_WRITEOK))
313 if (!si->highest_bit)
315 if (offset > si->highest_bit)
316 scan_base = offset = si->lowest_bit;
318 /* reuse swap entry of cache-only swap if not busy. */
319 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
321 spin_unlock(&swap_lock);
322 swap_was_freed = __try_to_reclaim_swap(si, offset);
323 spin_lock(&swap_lock);
324 /* entry was freed successfully, try to use this again */
327 goto scan; /* check next one */
330 if (si->swap_map[offset])
333 if (offset == si->lowest_bit)
335 if (offset == si->highest_bit)
338 if (si->inuse_pages == si->pages) {
339 si->lowest_bit = si->max;
342 si->swap_map[offset] = usage;
343 si->cluster_next = offset + 1;
344 si->flags -= SWP_SCANNING;
346 if (si->lowest_alloc) {
348 * Only set when SWP_DISCARDABLE, and there's a scan
349 * for a free cluster in progress or just completed.
351 if (found_free_cluster) {
353 * To optimize wear-levelling, discard the
354 * old data of the cluster, taking care not to
355 * discard any of its pages that have already
356 * been allocated by racing tasks (offset has
357 * already stepped over any at the beginning).
359 if (offset < si->highest_alloc &&
360 si->lowest_alloc <= last_in_cluster)
361 last_in_cluster = si->lowest_alloc - 1;
362 si->flags |= SWP_DISCARDING;
363 spin_unlock(&swap_lock);
365 if (offset < last_in_cluster)
366 discard_swap_cluster(si, offset,
367 last_in_cluster - offset + 1);
369 spin_lock(&swap_lock);
370 si->lowest_alloc = 0;
371 si->flags &= ~SWP_DISCARDING;
373 smp_mb(); /* wake_up_bit advises this */
374 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
376 } else if (si->flags & SWP_DISCARDING) {
378 * Delay using pages allocated by racing tasks
379 * until the whole discard has been issued. We
380 * could defer that delay until swap_writepage,
381 * but it's easier to keep this self-contained.
383 spin_unlock(&swap_lock);
384 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
385 wait_for_discard, TASK_UNINTERRUPTIBLE);
386 spin_lock(&swap_lock);
389 * Note pages allocated by racing tasks while
390 * scan for a free cluster is in progress, so
391 * that its final discard can exclude them.
393 if (offset < si->lowest_alloc)
394 si->lowest_alloc = offset;
395 if (offset > si->highest_alloc)
396 si->highest_alloc = offset;
402 spin_unlock(&swap_lock);
403 while (++offset <= si->highest_bit) {
404 if (!si->swap_map[offset]) {
405 spin_lock(&swap_lock);
408 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
409 spin_lock(&swap_lock);
412 if (unlikely(--latency_ration < 0)) {
414 latency_ration = LATENCY_LIMIT;
417 offset = si->lowest_bit;
418 while (++offset < scan_base) {
419 if (!si->swap_map[offset]) {
420 spin_lock(&swap_lock);
423 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
424 spin_lock(&swap_lock);
427 if (unlikely(--latency_ration < 0)) {
429 latency_ration = LATENCY_LIMIT;
432 spin_lock(&swap_lock);
435 si->flags -= SWP_SCANNING;
439 swp_entry_t get_swap_page(void)
441 struct swap_info_struct *si;
446 spin_lock(&swap_lock);
447 if (nr_swap_pages <= 0)
451 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
452 si = swap_info[type];
455 (!wrapped && si->prio != swap_info[next]->prio)) {
456 next = swap_list.head;
460 if (!si->highest_bit)
462 if (!(si->flags & SWP_WRITEOK))
465 swap_list.next = next;
466 /* This is called for allocating swap entry for cache */
467 offset = scan_swap_map(si, SWAP_HAS_CACHE);
469 spin_unlock(&swap_lock);
470 return swp_entry(type, offset);
472 next = swap_list.next;
477 spin_unlock(&swap_lock);
478 return (swp_entry_t) {0};
481 /* The only caller of this function is now susupend routine */
482 swp_entry_t get_swap_page_of_type(int type)
484 struct swap_info_struct *si;
487 spin_lock(&swap_lock);
488 si = swap_info[type];
489 if (si && (si->flags & SWP_WRITEOK)) {
491 /* This is called for allocating swap entry, not cache */
492 offset = scan_swap_map(si, 1);
494 spin_unlock(&swap_lock);
495 return swp_entry(type, offset);
499 spin_unlock(&swap_lock);
500 return (swp_entry_t) {0};
503 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
505 struct swap_info_struct *p;
506 unsigned long offset, type;
510 type = swp_type(entry);
511 if (type >= nr_swapfiles)
514 if (!(p->flags & SWP_USED))
516 offset = swp_offset(entry);
517 if (offset >= p->max)
519 if (!p->swap_map[offset])
521 spin_lock(&swap_lock);
525 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
528 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
531 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
534 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
539 static unsigned char swap_entry_free(struct swap_info_struct *p,
540 swp_entry_t entry, unsigned char usage)
542 unsigned long offset = swp_offset(entry);
544 unsigned char has_cache;
546 count = p->swap_map[offset];
547 has_cache = count & SWAP_HAS_CACHE;
548 count &= ~SWAP_HAS_CACHE;
550 if (usage == SWAP_HAS_CACHE) {
551 VM_BUG_ON(!has_cache);
553 } else if (count == SWAP_MAP_SHMEM) {
555 * Or we could insist on shmem.c using a special
556 * swap_shmem_free() and free_shmem_swap_and_cache()...
559 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
560 if (count == COUNT_CONTINUED) {
561 if (swap_count_continued(p, offset, count))
562 count = SWAP_MAP_MAX | COUNT_CONTINUED;
564 count = SWAP_MAP_MAX;
570 mem_cgroup_uncharge_swap(entry);
572 usage = count | has_cache;
573 p->swap_map[offset] = usage;
575 /* free if no reference */
577 struct gendisk *disk = p->bdev->bd_disk;
578 if (offset < p->lowest_bit)
579 p->lowest_bit = offset;
580 if (offset > p->highest_bit)
581 p->highest_bit = offset;
582 if (swap_list.next >= 0 &&
583 p->prio > swap_info[swap_list.next]->prio)
584 swap_list.next = p->type;
587 if ((p->flags & SWP_BLKDEV) &&
588 disk->fops->swap_slot_free_notify)
589 disk->fops->swap_slot_free_notify(p->bdev, offset);
596 * Caller has made sure that the swapdevice corresponding to entry
597 * is still around or has not been recycled.
599 void swap_free(swp_entry_t entry)
601 struct swap_info_struct *p;
603 p = swap_info_get(entry);
605 swap_entry_free(p, entry, 1);
606 spin_unlock(&swap_lock);
611 * Called after dropping swapcache to decrease refcnt to swap entries.
613 void swapcache_free(swp_entry_t entry, struct page *page)
615 struct swap_info_struct *p;
618 p = swap_info_get(entry);
620 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
622 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
623 spin_unlock(&swap_lock);
628 * How many references to page are currently swapped out?
629 * This does not give an exact answer when swap count is continued,
630 * but does include the high COUNT_CONTINUED flag to allow for that.
632 static inline int page_swapcount(struct page *page)
635 struct swap_info_struct *p;
638 entry.val = page_private(page);
639 p = swap_info_get(entry);
641 count = swap_count(p->swap_map[swp_offset(entry)]);
642 spin_unlock(&swap_lock);
648 * We can write to an anon page without COW if there are no other references
649 * to it. And as a side-effect, free up its swap: because the old content
650 * on disk will never be read, and seeking back there to write new content
651 * later would only waste time away from clustering.
653 int reuse_swap_page(struct page *page)
657 VM_BUG_ON(!PageLocked(page));
658 if (unlikely(PageKsm(page)))
660 count = page_mapcount(page);
661 if (count <= 1 && PageSwapCache(page)) {
662 count += page_swapcount(page);
663 if (count == 1 && !PageWriteback(page)) {
664 delete_from_swap_cache(page);
672 * If swap is getting full, or if there are no more mappings of this page,
673 * then try_to_free_swap is called to free its swap space.
675 int try_to_free_swap(struct page *page)
677 VM_BUG_ON(!PageLocked(page));
679 if (!PageSwapCache(page))
681 if (PageWriteback(page))
683 if (page_swapcount(page))
686 delete_from_swap_cache(page);
692 * Free the swap entry like above, but also try to
693 * free the page cache entry if it is the last user.
695 int free_swap_and_cache(swp_entry_t entry)
697 struct swap_info_struct *p;
698 struct page *page = NULL;
700 if (non_swap_entry(entry))
703 p = swap_info_get(entry);
705 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
706 page = find_get_page(&swapper_space, entry.val);
707 if (page && !trylock_page(page)) {
708 page_cache_release(page);
712 spin_unlock(&swap_lock);
716 * Not mapped elsewhere, or swap space full? Free it!
717 * Also recheck PageSwapCache now page is locked (above).
719 if (PageSwapCache(page) && !PageWriteback(page) &&
720 (!page_mapped(page) || vm_swap_full())) {
721 delete_from_swap_cache(page);
725 page_cache_release(page);
730 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
732 * mem_cgroup_count_swap_user - count the user of a swap entry
733 * @ent: the swap entry to be checked
734 * @pagep: the pointer for the swap cache page of the entry to be stored
736 * Returns the number of the user of the swap entry. The number is valid only
737 * for swaps of anonymous pages.
738 * If the entry is found on swap cache, the page is stored to pagep with
739 * refcount of it being incremented.
741 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
744 struct swap_info_struct *p;
747 page = find_get_page(&swapper_space, ent.val);
749 count += page_mapcount(page);
750 p = swap_info_get(ent);
752 count += swap_count(p->swap_map[swp_offset(ent)]);
753 spin_unlock(&swap_lock);
761 #ifdef CONFIG_HIBERNATION
763 * Find the swap type that corresponds to given device (if any).
765 * @offset - number of the PAGE_SIZE-sized block of the device, starting
766 * from 0, in which the swap header is expected to be located.
768 * This is needed for the suspend to disk (aka swsusp).
770 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
772 struct block_device *bdev = NULL;
776 bdev = bdget(device);
778 spin_lock(&swap_lock);
779 for (type = 0; type < nr_swapfiles; type++) {
780 struct swap_info_struct *sis = swap_info[type];
782 if (!(sis->flags & SWP_WRITEOK))
787 *bdev_p = bdgrab(sis->bdev);
789 spin_unlock(&swap_lock);
792 if (bdev == sis->bdev) {
793 struct swap_extent *se = &sis->first_swap_extent;
795 if (se->start_block == offset) {
797 *bdev_p = bdgrab(sis->bdev);
799 spin_unlock(&swap_lock);
805 spin_unlock(&swap_lock);
813 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
814 * corresponding to given index in swap_info (swap type).
816 sector_t swapdev_block(int type, pgoff_t offset)
818 struct block_device *bdev;
820 if ((unsigned int)type >= nr_swapfiles)
822 if (!(swap_info[type]->flags & SWP_WRITEOK))
824 return map_swap_entry(swp_entry(type, offset), &bdev);
828 * Return either the total number of swap pages of given type, or the number
829 * of free pages of that type (depending on @free)
831 * This is needed for software suspend
833 unsigned int count_swap_pages(int type, int free)
837 spin_lock(&swap_lock);
838 if ((unsigned int)type < nr_swapfiles) {
839 struct swap_info_struct *sis = swap_info[type];
841 if (sis->flags & SWP_WRITEOK) {
844 n -= sis->inuse_pages;
847 spin_unlock(&swap_lock);
850 #endif /* CONFIG_HIBERNATION */
853 * No need to decide whether this PTE shares the swap entry with others,
854 * just let do_wp_page work it out if a write is requested later - to
855 * force COW, vm_page_prot omits write permission from any private vma.
857 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
858 unsigned long addr, swp_entry_t entry, struct page *page)
860 struct mem_cgroup *ptr = NULL;
865 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
870 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
871 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
873 mem_cgroup_cancel_charge_swapin(ptr);
878 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
879 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
881 set_pte_at(vma->vm_mm, addr, pte,
882 pte_mkold(mk_pte(page, vma->vm_page_prot)));
883 page_add_anon_rmap(page, vma, addr);
884 mem_cgroup_commit_charge_swapin(page, ptr);
887 * Move the page to the active list so it is not
888 * immediately swapped out again after swapon.
892 pte_unmap_unlock(pte, ptl);
897 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
898 unsigned long addr, unsigned long end,
899 swp_entry_t entry, struct page *page)
901 pte_t swp_pte = swp_entry_to_pte(entry);
906 * We don't actually need pte lock while scanning for swp_pte: since
907 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
908 * page table while we're scanning; though it could get zapped, and on
909 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
910 * of unmatched parts which look like swp_pte, so unuse_pte must
911 * recheck under pte lock. Scanning without pte lock lets it be
912 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
914 pte = pte_offset_map(pmd, addr);
917 * swapoff spends a _lot_ of time in this loop!
918 * Test inline before going to call unuse_pte.
920 if (unlikely(pte_same(*pte, swp_pte))) {
922 ret = unuse_pte(vma, pmd, addr, entry, page);
925 pte = pte_offset_map(pmd, addr);
927 } while (pte++, addr += PAGE_SIZE, addr != end);
933 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
934 unsigned long addr, unsigned long end,
935 swp_entry_t entry, struct page *page)
941 pmd = pmd_offset(pud, addr);
943 next = pmd_addr_end(addr, end);
944 if (pmd_none_or_clear_bad(pmd))
946 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
949 } while (pmd++, addr = next, addr != end);
953 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
954 unsigned long addr, unsigned long end,
955 swp_entry_t entry, struct page *page)
961 pud = pud_offset(pgd, addr);
963 next = pud_addr_end(addr, end);
964 if (pud_none_or_clear_bad(pud))
966 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
969 } while (pud++, addr = next, addr != end);
973 static int unuse_vma(struct vm_area_struct *vma,
974 swp_entry_t entry, struct page *page)
977 unsigned long addr, end, next;
980 if (page_anon_vma(page)) {
981 addr = page_address_in_vma(page, vma);
985 end = addr + PAGE_SIZE;
987 addr = vma->vm_start;
991 pgd = pgd_offset(vma->vm_mm, addr);
993 next = pgd_addr_end(addr, end);
994 if (pgd_none_or_clear_bad(pgd))
996 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
999 } while (pgd++, addr = next, addr != end);
1003 static int unuse_mm(struct mm_struct *mm,
1004 swp_entry_t entry, struct page *page)
1006 struct vm_area_struct *vma;
1009 if (!down_read_trylock(&mm->mmap_sem)) {
1011 * Activate page so shrink_inactive_list is unlikely to unmap
1012 * its ptes while lock is dropped, so swapoff can make progress.
1014 activate_page(page);
1016 down_read(&mm->mmap_sem);
1019 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1020 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1023 up_read(&mm->mmap_sem);
1024 return (ret < 0)? ret: 0;
1028 * Scan swap_map from current position to next entry still in use.
1029 * Recycle to start on reaching the end, returning 0 when empty.
1031 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1034 unsigned int max = si->max;
1035 unsigned int i = prev;
1036 unsigned char count;
1039 * No need for swap_lock here: we're just looking
1040 * for whether an entry is in use, not modifying it; false
1041 * hits are okay, and sys_swapoff() has already prevented new
1042 * allocations from this area (while holding swap_lock).
1051 * No entries in use at top of swap_map,
1052 * loop back to start and recheck there.
1058 count = si->swap_map[i];
1059 if (count && swap_count(count) != SWAP_MAP_BAD)
1066 * We completely avoid races by reading each swap page in advance,
1067 * and then search for the process using it. All the necessary
1068 * page table adjustments can then be made atomically.
1070 static int try_to_unuse(unsigned int type)
1072 struct swap_info_struct *si = swap_info[type];
1073 struct mm_struct *start_mm;
1074 unsigned char *swap_map;
1075 unsigned char swcount;
1082 * When searching mms for an entry, a good strategy is to
1083 * start at the first mm we freed the previous entry from
1084 * (though actually we don't notice whether we or coincidence
1085 * freed the entry). Initialize this start_mm with a hold.
1087 * A simpler strategy would be to start at the last mm we
1088 * freed the previous entry from; but that would take less
1089 * advantage of mmlist ordering, which clusters forked mms
1090 * together, child after parent. If we race with dup_mmap(), we
1091 * prefer to resolve parent before child, lest we miss entries
1092 * duplicated after we scanned child: using last mm would invert
1095 start_mm = &init_mm;
1096 atomic_inc(&init_mm.mm_users);
1099 * Keep on scanning until all entries have gone. Usually,
1100 * one pass through swap_map is enough, but not necessarily:
1101 * there are races when an instance of an entry might be missed.
1103 while ((i = find_next_to_unuse(si, i)) != 0) {
1104 if (signal_pending(current)) {
1110 * Get a page for the entry, using the existing swap
1111 * cache page if there is one. Otherwise, get a clean
1112 * page and read the swap into it.
1114 swap_map = &si->swap_map[i];
1115 entry = swp_entry(type, i);
1116 page = read_swap_cache_async(entry,
1117 GFP_HIGHUSER_MOVABLE, NULL, 0);
1120 * Either swap_duplicate() failed because entry
1121 * has been freed independently, and will not be
1122 * reused since sys_swapoff() already disabled
1123 * allocation from here, or alloc_page() failed.
1132 * Don't hold on to start_mm if it looks like exiting.
1134 if (atomic_read(&start_mm->mm_users) == 1) {
1136 start_mm = &init_mm;
1137 atomic_inc(&init_mm.mm_users);
1141 * Wait for and lock page. When do_swap_page races with
1142 * try_to_unuse, do_swap_page can handle the fault much
1143 * faster than try_to_unuse can locate the entry. This
1144 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1145 * defer to do_swap_page in such a case - in some tests,
1146 * do_swap_page and try_to_unuse repeatedly compete.
1148 wait_on_page_locked(page);
1149 wait_on_page_writeback(page);
1151 wait_on_page_writeback(page);
1154 * Remove all references to entry.
1156 swcount = *swap_map;
1157 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1158 retval = shmem_unuse(entry, page);
1159 /* page has already been unlocked and released */
1164 if (swap_count(swcount) && start_mm != &init_mm)
1165 retval = unuse_mm(start_mm, entry, page);
1167 if (swap_count(*swap_map)) {
1168 int set_start_mm = (*swap_map >= swcount);
1169 struct list_head *p = &start_mm->mmlist;
1170 struct mm_struct *new_start_mm = start_mm;
1171 struct mm_struct *prev_mm = start_mm;
1172 struct mm_struct *mm;
1174 atomic_inc(&new_start_mm->mm_users);
1175 atomic_inc(&prev_mm->mm_users);
1176 spin_lock(&mmlist_lock);
1177 while (swap_count(*swap_map) && !retval &&
1178 (p = p->next) != &start_mm->mmlist) {
1179 mm = list_entry(p, struct mm_struct, mmlist);
1180 if (!atomic_inc_not_zero(&mm->mm_users))
1182 spin_unlock(&mmlist_lock);
1188 swcount = *swap_map;
1189 if (!swap_count(swcount)) /* any usage ? */
1191 else if (mm == &init_mm)
1194 retval = unuse_mm(mm, entry, page);
1196 if (set_start_mm && *swap_map < swcount) {
1197 mmput(new_start_mm);
1198 atomic_inc(&mm->mm_users);
1202 spin_lock(&mmlist_lock);
1204 spin_unlock(&mmlist_lock);
1207 start_mm = new_start_mm;
1211 page_cache_release(page);
1216 * If a reference remains (rare), we would like to leave
1217 * the page in the swap cache; but try_to_unmap could
1218 * then re-duplicate the entry once we drop page lock,
1219 * so we might loop indefinitely; also, that page could
1220 * not be swapped out to other storage meanwhile. So:
1221 * delete from cache even if there's another reference,
1222 * after ensuring that the data has been saved to disk -
1223 * since if the reference remains (rarer), it will be
1224 * read from disk into another page. Splitting into two
1225 * pages would be incorrect if swap supported "shared
1226 * private" pages, but they are handled by tmpfs files.
1228 * Given how unuse_vma() targets one particular offset
1229 * in an anon_vma, once the anon_vma has been determined,
1230 * this splitting happens to be just what is needed to
1231 * handle where KSM pages have been swapped out: re-reading
1232 * is unnecessarily slow, but we can fix that later on.
1234 if (swap_count(*swap_map) &&
1235 PageDirty(page) && PageSwapCache(page)) {
1236 struct writeback_control wbc = {
1237 .sync_mode = WB_SYNC_NONE,
1240 swap_writepage(page, &wbc);
1242 wait_on_page_writeback(page);
1246 * It is conceivable that a racing task removed this page from
1247 * swap cache just before we acquired the page lock at the top,
1248 * or while we dropped it in unuse_mm(). The page might even
1249 * be back in swap cache on another swap area: that we must not
1250 * delete, since it may not have been written out to swap yet.
1252 if (PageSwapCache(page) &&
1253 likely(page_private(page) == entry.val))
1254 delete_from_swap_cache(page);
1257 * So we could skip searching mms once swap count went
1258 * to 1, we did not mark any present ptes as dirty: must
1259 * mark page dirty so shrink_page_list will preserve it.
1263 page_cache_release(page);
1266 * Make sure that we aren't completely killing
1267 * interactive performance.
1277 * After a successful try_to_unuse, if no swap is now in use, we know
1278 * we can empty the mmlist. swap_lock must be held on entry and exit.
1279 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1280 * added to the mmlist just after page_duplicate - before would be racy.
1282 static void drain_mmlist(void)
1284 struct list_head *p, *next;
1287 for (type = 0; type < nr_swapfiles; type++)
1288 if (swap_info[type]->inuse_pages)
1290 spin_lock(&mmlist_lock);
1291 list_for_each_safe(p, next, &init_mm.mmlist)
1293 spin_unlock(&mmlist_lock);
1297 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1298 * corresponds to page offset for the specified swap entry.
1299 * Note that the type of this function is sector_t, but it returns page offset
1300 * into the bdev, not sector offset.
1302 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1304 struct swap_info_struct *sis;
1305 struct swap_extent *start_se;
1306 struct swap_extent *se;
1309 sis = swap_info[swp_type(entry)];
1312 offset = swp_offset(entry);
1313 start_se = sis->curr_swap_extent;
1317 struct list_head *lh;
1319 if (se->start_page <= offset &&
1320 offset < (se->start_page + se->nr_pages)) {
1321 return se->start_block + (offset - se->start_page);
1324 se = list_entry(lh, struct swap_extent, list);
1325 sis->curr_swap_extent = se;
1326 BUG_ON(se == start_se); /* It *must* be present */
1331 * Returns the page offset into bdev for the specified page's swap entry.
1333 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1336 entry.val = page_private(page);
1337 return map_swap_entry(entry, bdev);
1341 * Free all of a swapdev's extent information
1343 static void destroy_swap_extents(struct swap_info_struct *sis)
1345 while (!list_empty(&sis->first_swap_extent.list)) {
1346 struct swap_extent *se;
1348 se = list_entry(sis->first_swap_extent.list.next,
1349 struct swap_extent, list);
1350 list_del(&se->list);
1356 * Add a block range (and the corresponding page range) into this swapdev's
1357 * extent list. The extent list is kept sorted in page order.
1359 * This function rather assumes that it is called in ascending page order.
1362 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1363 unsigned long nr_pages, sector_t start_block)
1365 struct swap_extent *se;
1366 struct swap_extent *new_se;
1367 struct list_head *lh;
1369 if (start_page == 0) {
1370 se = &sis->first_swap_extent;
1371 sis->curr_swap_extent = se;
1373 se->nr_pages = nr_pages;
1374 se->start_block = start_block;
1377 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1378 se = list_entry(lh, struct swap_extent, list);
1379 BUG_ON(se->start_page + se->nr_pages != start_page);
1380 if (se->start_block + se->nr_pages == start_block) {
1382 se->nr_pages += nr_pages;
1388 * No merge. Insert a new extent, preserving ordering.
1390 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1393 new_se->start_page = start_page;
1394 new_se->nr_pages = nr_pages;
1395 new_se->start_block = start_block;
1397 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1402 * A `swap extent' is a simple thing which maps a contiguous range of pages
1403 * onto a contiguous range of disk blocks. An ordered list of swap extents
1404 * is built at swapon time and is then used at swap_writepage/swap_readpage
1405 * time for locating where on disk a page belongs.
1407 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1408 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1409 * swap files identically.
1411 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1412 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1413 * swapfiles are handled *identically* after swapon time.
1415 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1416 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1417 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1418 * requirements, they are simply tossed out - we will never use those blocks
1421 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1422 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1423 * which will scribble on the fs.
1425 * The amount of disk space which a single swap extent represents varies.
1426 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1427 * extents in the list. To avoid much list walking, we cache the previous
1428 * search location in `curr_swap_extent', and start new searches from there.
1429 * This is extremely effective. The average number of iterations in
1430 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1432 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1434 struct inode *inode;
1435 unsigned blocks_per_page;
1436 unsigned long page_no;
1438 sector_t probe_block;
1439 sector_t last_block;
1440 sector_t lowest_block = -1;
1441 sector_t highest_block = 0;
1445 inode = sis->swap_file->f_mapping->host;
1446 if (S_ISBLK(inode->i_mode)) {
1447 ret = add_swap_extent(sis, 0, sis->max, 0);
1452 blkbits = inode->i_blkbits;
1453 blocks_per_page = PAGE_SIZE >> blkbits;
1456 * Map all the blocks into the extent list. This code doesn't try
1461 last_block = i_size_read(inode) >> blkbits;
1462 while ((probe_block + blocks_per_page) <= last_block &&
1463 page_no < sis->max) {
1464 unsigned block_in_page;
1465 sector_t first_block;
1467 first_block = bmap(inode, probe_block);
1468 if (first_block == 0)
1472 * It must be PAGE_SIZE aligned on-disk
1474 if (first_block & (blocks_per_page - 1)) {
1479 for (block_in_page = 1; block_in_page < blocks_per_page;
1483 block = bmap(inode, probe_block + block_in_page);
1486 if (block != first_block + block_in_page) {
1493 first_block >>= (PAGE_SHIFT - blkbits);
1494 if (page_no) { /* exclude the header page */
1495 if (first_block < lowest_block)
1496 lowest_block = first_block;
1497 if (first_block > highest_block)
1498 highest_block = first_block;
1502 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1504 ret = add_swap_extent(sis, page_no, 1, first_block);
1509 probe_block += blocks_per_page;
1514 *span = 1 + highest_block - lowest_block;
1516 page_no = 1; /* force Empty message */
1518 sis->pages = page_no - 1;
1519 sis->highest_bit = page_no - 1;
1523 printk(KERN_ERR "swapon: swapfile has holes\n");
1528 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1530 struct swap_info_struct *p = NULL;
1531 unsigned char *swap_map;
1532 struct file *swap_file, *victim;
1533 struct address_space *mapping;
1534 struct inode *inode;
1539 if (!capable(CAP_SYS_ADMIN))
1542 pathname = getname(specialfile);
1543 err = PTR_ERR(pathname);
1544 if (IS_ERR(pathname))
1547 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1549 err = PTR_ERR(victim);
1553 mapping = victim->f_mapping;
1555 spin_lock(&swap_lock);
1556 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1557 p = swap_info[type];
1558 if (p->flags & SWP_WRITEOK) {
1559 if (p->swap_file->f_mapping == mapping)
1566 spin_unlock(&swap_lock);
1569 if (!security_vm_enough_memory(p->pages))
1570 vm_unacct_memory(p->pages);
1573 spin_unlock(&swap_lock);
1577 swap_list.head = p->next;
1579 swap_info[prev]->next = p->next;
1580 if (type == swap_list.next) {
1581 /* just pick something that's safe... */
1582 swap_list.next = swap_list.head;
1585 for (i = p->next; i >= 0; i = swap_info[i]->next)
1586 swap_info[i]->prio = p->prio--;
1589 nr_swap_pages -= p->pages;
1590 total_swap_pages -= p->pages;
1591 p->flags &= ~SWP_WRITEOK;
1592 spin_unlock(&swap_lock);
1594 current->flags |= PF_OOM_ORIGIN;
1595 err = try_to_unuse(type);
1596 current->flags &= ~PF_OOM_ORIGIN;
1599 /* re-insert swap space back into swap_list */
1600 spin_lock(&swap_lock);
1602 p->prio = --least_priority;
1604 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1605 if (p->prio >= swap_info[i]->prio)
1611 swap_list.head = swap_list.next = type;
1613 swap_info[prev]->next = type;
1614 nr_swap_pages += p->pages;
1615 total_swap_pages += p->pages;
1616 p->flags |= SWP_WRITEOK;
1617 spin_unlock(&swap_lock);
1621 /* wait for any unplug function to finish */
1622 down_write(&swap_unplug_sem);
1623 up_write(&swap_unplug_sem);
1625 destroy_swap_extents(p);
1626 if (p->flags & SWP_CONTINUED)
1627 free_swap_count_continuations(p);
1629 mutex_lock(&swapon_mutex);
1630 spin_lock(&swap_lock);
1633 /* wait for anyone still in scan_swap_map */
1634 p->highest_bit = 0; /* cuts scans short */
1635 while (p->flags >= SWP_SCANNING) {
1636 spin_unlock(&swap_lock);
1637 schedule_timeout_uninterruptible(1);
1638 spin_lock(&swap_lock);
1641 swap_file = p->swap_file;
1642 p->swap_file = NULL;
1644 swap_map = p->swap_map;
1647 spin_unlock(&swap_lock);
1648 mutex_unlock(&swapon_mutex);
1650 /* Destroy swap account informatin */
1651 swap_cgroup_swapoff(type);
1653 inode = mapping->host;
1654 if (S_ISBLK(inode->i_mode)) {
1655 struct block_device *bdev = I_BDEV(inode);
1656 set_blocksize(bdev, p->old_block_size);
1659 mutex_lock(&inode->i_mutex);
1660 inode->i_flags &= ~S_SWAPFILE;
1661 mutex_unlock(&inode->i_mutex);
1663 filp_close(swap_file, NULL);
1667 filp_close(victim, NULL);
1672 #ifdef CONFIG_PROC_FS
1674 static void *swap_start(struct seq_file *swap, loff_t *pos)
1676 struct swap_info_struct *si;
1680 mutex_lock(&swapon_mutex);
1683 return SEQ_START_TOKEN;
1685 for (type = 0; type < nr_swapfiles; type++) {
1686 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1687 si = swap_info[type];
1688 if (!(si->flags & SWP_USED) || !si->swap_map)
1697 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1699 struct swap_info_struct *si = v;
1702 if (v == SEQ_START_TOKEN)
1705 type = si->type + 1;
1707 for (; type < nr_swapfiles; type++) {
1708 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1709 si = swap_info[type];
1710 if (!(si->flags & SWP_USED) || !si->swap_map)
1719 static void swap_stop(struct seq_file *swap, void *v)
1721 mutex_unlock(&swapon_mutex);
1724 static int swap_show(struct seq_file *swap, void *v)
1726 struct swap_info_struct *si = v;
1730 if (si == SEQ_START_TOKEN) {
1731 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1735 file = si->swap_file;
1736 len = seq_path(swap, &file->f_path, " \t\n\\");
1737 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1738 len < 40 ? 40 - len : 1, " ",
1739 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1740 "partition" : "file\t",
1741 si->pages << (PAGE_SHIFT - 10),
1742 si->inuse_pages << (PAGE_SHIFT - 10),
1747 static const struct seq_operations swaps_op = {
1748 .start = swap_start,
1754 static int swaps_open(struct inode *inode, struct file *file)
1756 return seq_open(file, &swaps_op);
1759 static const struct file_operations proc_swaps_operations = {
1762 .llseek = seq_lseek,
1763 .release = seq_release,
1766 static int __init procswaps_init(void)
1768 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1771 __initcall(procswaps_init);
1772 #endif /* CONFIG_PROC_FS */
1774 #ifdef MAX_SWAPFILES_CHECK
1775 static int __init max_swapfiles_check(void)
1777 MAX_SWAPFILES_CHECK();
1780 late_initcall(max_swapfiles_check);
1784 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1786 * The swapon system call
1788 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1790 struct swap_info_struct *p;
1792 struct block_device *bdev = NULL;
1793 struct file *swap_file = NULL;
1794 struct address_space *mapping;
1798 union swap_header *swap_header;
1799 unsigned int nr_good_pages;
1802 unsigned long maxpages;
1803 unsigned long swapfilepages;
1804 unsigned char *swap_map = NULL;
1805 struct page *page = NULL;
1806 struct inode *inode = NULL;
1809 if (!capable(CAP_SYS_ADMIN))
1812 p = kzalloc(sizeof(*p), GFP_KERNEL);
1816 spin_lock(&swap_lock);
1817 for (type = 0; type < nr_swapfiles; type++) {
1818 if (!(swap_info[type]->flags & SWP_USED))
1822 if (type >= MAX_SWAPFILES) {
1823 spin_unlock(&swap_lock);
1827 if (type >= nr_swapfiles) {
1829 swap_info[type] = p;
1831 * Write swap_info[type] before nr_swapfiles, in case a
1832 * racing procfs swap_start() or swap_next() is reading them.
1833 * (We never shrink nr_swapfiles, we never free this entry.)
1839 p = swap_info[type];
1841 * Do not memset this entry: a racing procfs swap_next()
1842 * would be relying on p->type to remain valid.
1845 INIT_LIST_HEAD(&p->first_swap_extent.list);
1846 p->flags = SWP_USED;
1848 spin_unlock(&swap_lock);
1850 name = getname(specialfile);
1851 error = PTR_ERR(name);
1856 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1857 error = PTR_ERR(swap_file);
1858 if (IS_ERR(swap_file)) {
1863 p->swap_file = swap_file;
1864 mapping = swap_file->f_mapping;
1865 inode = mapping->host;
1868 for (i = 0; i < nr_swapfiles; i++) {
1869 struct swap_info_struct *q = swap_info[i];
1871 if (i == type || !q->swap_file)
1873 if (mapping == q->swap_file->f_mapping)
1878 if (S_ISBLK(inode->i_mode)) {
1879 bdev = I_BDEV(inode);
1880 error = bd_claim(bdev, sys_swapon);
1886 p->old_block_size = block_size(bdev);
1887 error = set_blocksize(bdev, PAGE_SIZE);
1891 p->flags |= SWP_BLKDEV;
1892 } else if (S_ISREG(inode->i_mode)) {
1893 p->bdev = inode->i_sb->s_bdev;
1894 mutex_lock(&inode->i_mutex);
1896 if (IS_SWAPFILE(inode)) {
1904 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1907 * Read the swap header.
1909 if (!mapping->a_ops->readpage) {
1913 page = read_mapping_page(mapping, 0, swap_file);
1915 error = PTR_ERR(page);
1918 swap_header = kmap(page);
1920 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1921 printk(KERN_ERR "Unable to find swap-space signature\n");
1926 /* swap partition endianess hack... */
1927 if (swab32(swap_header->info.version) == 1) {
1928 swab32s(&swap_header->info.version);
1929 swab32s(&swap_header->info.last_page);
1930 swab32s(&swap_header->info.nr_badpages);
1931 for (i = 0; i < swap_header->info.nr_badpages; i++)
1932 swab32s(&swap_header->info.badpages[i]);
1934 /* Check the swap header's sub-version */
1935 if (swap_header->info.version != 1) {
1937 "Unable to handle swap header version %d\n",
1938 swap_header->info.version);
1944 p->cluster_next = 1;
1948 * Find out how many pages are allowed for a single swap
1949 * device. There are two limiting factors: 1) the number of
1950 * bits for the swap offset in the swp_entry_t type and
1951 * 2) the number of bits in the a swap pte as defined by
1952 * the different architectures. In order to find the
1953 * largest possible bit mask a swap entry with swap type 0
1954 * and swap offset ~0UL is created, encoded to a swap pte,
1955 * decoded to a swp_entry_t again and finally the swap
1956 * offset is extracted. This will mask all the bits from
1957 * the initial ~0UL mask that can't be encoded in either
1958 * the swp_entry_t or the architecture definition of a
1961 maxpages = swp_offset(pte_to_swp_entry(
1962 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1963 if (maxpages > swap_header->info.last_page) {
1964 maxpages = swap_header->info.last_page + 1;
1965 /* p->max is an unsigned int: don't overflow it */
1966 if ((unsigned int)maxpages == 0)
1967 maxpages = UINT_MAX;
1969 p->highest_bit = maxpages - 1;
1974 if (swapfilepages && maxpages > swapfilepages) {
1976 "Swap area shorter than signature indicates\n");
1979 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1981 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1984 /* OK, set up the swap map and apply the bad block list */
1985 swap_map = vmalloc(maxpages);
1991 memset(swap_map, 0, maxpages);
1992 nr_good_pages = maxpages - 1; /* omit header page */
1994 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1995 unsigned int page_nr = swap_header->info.badpages[i];
1996 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2000 if (page_nr < maxpages) {
2001 swap_map[page_nr] = SWAP_MAP_BAD;
2006 error = swap_cgroup_swapon(type, maxpages);
2010 if (nr_good_pages) {
2011 swap_map[0] = SWAP_MAP_BAD;
2013 p->pages = nr_good_pages;
2014 nr_extents = setup_swap_extents(p, &span);
2015 if (nr_extents < 0) {
2019 nr_good_pages = p->pages;
2021 if (!nr_good_pages) {
2022 printk(KERN_WARNING "Empty swap-file\n");
2028 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2029 p->flags |= SWP_SOLIDSTATE;
2030 p->cluster_next = 1 + (random32() % p->highest_bit);
2032 if (discard_swap(p) == 0)
2033 p->flags |= SWP_DISCARDABLE;
2036 mutex_lock(&swapon_mutex);
2037 spin_lock(&swap_lock);
2038 if (swap_flags & SWAP_FLAG_PREFER)
2040 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2042 p->prio = --least_priority;
2043 p->swap_map = swap_map;
2044 p->flags |= SWP_WRITEOK;
2045 nr_swap_pages += nr_good_pages;
2046 total_swap_pages += nr_good_pages;
2048 printk(KERN_INFO "Adding %uk swap on %s. "
2049 "Priority:%d extents:%d across:%lluk %s%s\n",
2050 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2051 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2052 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2053 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2055 /* insert swap space into swap_list: */
2057 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2058 if (p->prio >= swap_info[i]->prio)
2064 swap_list.head = swap_list.next = type;
2066 swap_info[prev]->next = type;
2067 spin_unlock(&swap_lock);
2068 mutex_unlock(&swapon_mutex);
2073 set_blocksize(bdev, p->old_block_size);
2076 destroy_swap_extents(p);
2077 swap_cgroup_swapoff(type);
2079 spin_lock(&swap_lock);
2080 p->swap_file = NULL;
2082 spin_unlock(&swap_lock);
2085 filp_close(swap_file, NULL);
2087 if (page && !IS_ERR(page)) {
2089 page_cache_release(page);
2095 inode->i_flags |= S_SWAPFILE;
2096 mutex_unlock(&inode->i_mutex);
2101 void si_swapinfo(struct sysinfo *val)
2104 unsigned long nr_to_be_unused = 0;
2106 spin_lock(&swap_lock);
2107 for (type = 0; type < nr_swapfiles; type++) {
2108 struct swap_info_struct *si = swap_info[type];
2110 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2111 nr_to_be_unused += si->inuse_pages;
2113 val->freeswap = nr_swap_pages + nr_to_be_unused;
2114 val->totalswap = total_swap_pages + nr_to_be_unused;
2115 spin_unlock(&swap_lock);
2119 * Verify that a swap entry is valid and increment its swap map count.
2121 * Returns error code in following case.
2123 * - swp_entry is invalid -> EINVAL
2124 * - swp_entry is migration entry -> EINVAL
2125 * - swap-cache reference is requested but there is already one. -> EEXIST
2126 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2127 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2129 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2131 struct swap_info_struct *p;
2132 unsigned long offset, type;
2133 unsigned char count;
2134 unsigned char has_cache;
2137 if (non_swap_entry(entry))
2140 type = swp_type(entry);
2141 if (type >= nr_swapfiles)
2143 p = swap_info[type];
2144 offset = swp_offset(entry);
2146 spin_lock(&swap_lock);
2147 if (unlikely(offset >= p->max))
2150 count = p->swap_map[offset];
2151 has_cache = count & SWAP_HAS_CACHE;
2152 count &= ~SWAP_HAS_CACHE;
2155 if (usage == SWAP_HAS_CACHE) {
2157 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2158 if (!has_cache && count)
2159 has_cache = SWAP_HAS_CACHE;
2160 else if (has_cache) /* someone else added cache */
2162 else /* no users remaining */
2165 } else if (count || has_cache) {
2167 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2169 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2171 else if (swap_count_continued(p, offset, count))
2172 count = COUNT_CONTINUED;
2176 err = -ENOENT; /* unused swap entry */
2178 p->swap_map[offset] = count | has_cache;
2181 spin_unlock(&swap_lock);
2186 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2191 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2192 * (in which case its reference count is never incremented).
2194 void swap_shmem_alloc(swp_entry_t entry)
2196 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2200 * Increase reference count of swap entry by 1.
2201 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2202 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2203 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2204 * might occur if a page table entry has got corrupted.
2206 int swap_duplicate(swp_entry_t entry)
2210 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2211 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2216 * @entry: swap entry for which we allocate swap cache.
2218 * Called when allocating swap cache for existing swap entry,
2219 * This can return error codes. Returns 0 at success.
2220 * -EBUSY means there is a swap cache.
2221 * Note: return code is different from swap_duplicate().
2223 int swapcache_prepare(swp_entry_t entry)
2225 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2229 * swap_lock prevents swap_map being freed. Don't grab an extra
2230 * reference on the swaphandle, it doesn't matter if it becomes unused.
2232 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2234 struct swap_info_struct *si;
2235 int our_page_cluster = page_cluster;
2236 pgoff_t target, toff;
2240 if (!our_page_cluster) /* no readahead */
2243 si = swap_info[swp_type(entry)];
2244 target = swp_offset(entry);
2245 base = (target >> our_page_cluster) << our_page_cluster;
2246 end = base + (1 << our_page_cluster);
2247 if (!base) /* first page is swap header */
2250 spin_lock(&swap_lock);
2251 if (end > si->max) /* don't go beyond end of map */
2254 /* Count contiguous allocated slots above our target */
2255 for (toff = target; ++toff < end; nr_pages++) {
2256 /* Don't read in free or bad pages */
2257 if (!si->swap_map[toff])
2259 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2262 /* Count contiguous allocated slots below our target */
2263 for (toff = target; --toff >= base; nr_pages++) {
2264 /* Don't read in free or bad pages */
2265 if (!si->swap_map[toff])
2267 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2270 spin_unlock(&swap_lock);
2273 * Indicate starting offset, and return number of pages to get:
2274 * if only 1, say 0, since there's then no readahead to be done.
2277 return nr_pages? ++nr_pages: 0;
2281 * add_swap_count_continuation - called when a swap count is duplicated
2282 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2283 * page of the original vmalloc'ed swap_map, to hold the continuation count
2284 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2285 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2287 * These continuation pages are seldom referenced: the common paths all work
2288 * on the original swap_map, only referring to a continuation page when the
2289 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2291 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2292 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2293 * can be called after dropping locks.
2295 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2297 struct swap_info_struct *si;
2300 struct page *list_page;
2302 unsigned char count;
2305 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2306 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2308 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2310 si = swap_info_get(entry);
2313 * An acceptable race has occurred since the failing
2314 * __swap_duplicate(): the swap entry has been freed,
2315 * perhaps even the whole swap_map cleared for swapoff.
2320 offset = swp_offset(entry);
2321 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2323 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2325 * The higher the swap count, the more likely it is that tasks
2326 * will race to add swap count continuation: we need to avoid
2327 * over-provisioning.
2333 spin_unlock(&swap_lock);
2338 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2339 * no architecture is using highmem pages for kernel pagetables: so it
2340 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2342 head = vmalloc_to_page(si->swap_map + offset);
2343 offset &= ~PAGE_MASK;
2346 * Page allocation does not initialize the page's lru field,
2347 * but it does always reset its private field.
2349 if (!page_private(head)) {
2350 BUG_ON(count & COUNT_CONTINUED);
2351 INIT_LIST_HEAD(&head->lru);
2352 set_page_private(head, SWP_CONTINUED);
2353 si->flags |= SWP_CONTINUED;
2356 list_for_each_entry(list_page, &head->lru, lru) {
2360 * If the previous map said no continuation, but we've found
2361 * a continuation page, free our allocation and use this one.
2363 if (!(count & COUNT_CONTINUED))
2366 map = kmap_atomic(list_page, KM_USER0) + offset;
2368 kunmap_atomic(map, KM_USER0);
2371 * If this continuation count now has some space in it,
2372 * free our allocation and use this one.
2374 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2378 list_add_tail(&page->lru, &head->lru);
2379 page = NULL; /* now it's attached, don't free it */
2381 spin_unlock(&swap_lock);
2389 * swap_count_continued - when the original swap_map count is incremented
2390 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2391 * into, carry if so, or else fail until a new continuation page is allocated;
2392 * when the original swap_map count is decremented from 0 with continuation,
2393 * borrow from the continuation and report whether it still holds more.
2394 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2396 static bool swap_count_continued(struct swap_info_struct *si,
2397 pgoff_t offset, unsigned char count)
2403 head = vmalloc_to_page(si->swap_map + offset);
2404 if (page_private(head) != SWP_CONTINUED) {
2405 BUG_ON(count & COUNT_CONTINUED);
2406 return false; /* need to add count continuation */
2409 offset &= ~PAGE_MASK;
2410 page = list_entry(head->lru.next, struct page, lru);
2411 map = kmap_atomic(page, KM_USER0) + offset;
2413 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2414 goto init_map; /* jump over SWAP_CONT_MAX checks */
2416 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2418 * Think of how you add 1 to 999
2420 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2421 kunmap_atomic(map, KM_USER0);
2422 page = list_entry(page->lru.next, struct page, lru);
2423 BUG_ON(page == head);
2424 map = kmap_atomic(page, KM_USER0) + offset;
2426 if (*map == SWAP_CONT_MAX) {
2427 kunmap_atomic(map, KM_USER0);
2428 page = list_entry(page->lru.next, struct page, lru);
2430 return false; /* add count continuation */
2431 map = kmap_atomic(page, KM_USER0) + offset;
2432 init_map: *map = 0; /* we didn't zero the page */
2435 kunmap_atomic(map, KM_USER0);
2436 page = list_entry(page->lru.prev, struct page, lru);
2437 while (page != head) {
2438 map = kmap_atomic(page, KM_USER0) + offset;
2439 *map = COUNT_CONTINUED;
2440 kunmap_atomic(map, KM_USER0);
2441 page = list_entry(page->lru.prev, struct page, lru);
2443 return true; /* incremented */
2445 } else { /* decrementing */
2447 * Think of how you subtract 1 from 1000
2449 BUG_ON(count != COUNT_CONTINUED);
2450 while (*map == COUNT_CONTINUED) {
2451 kunmap_atomic(map, KM_USER0);
2452 page = list_entry(page->lru.next, struct page, lru);
2453 BUG_ON(page == head);
2454 map = kmap_atomic(page, KM_USER0) + offset;
2460 kunmap_atomic(map, KM_USER0);
2461 page = list_entry(page->lru.prev, struct page, lru);
2462 while (page != head) {
2463 map = kmap_atomic(page, KM_USER0) + offset;
2464 *map = SWAP_CONT_MAX | count;
2465 count = COUNT_CONTINUED;
2466 kunmap_atomic(map, KM_USER0);
2467 page = list_entry(page->lru.prev, struct page, lru);
2469 return count == COUNT_CONTINUED;
2474 * free_swap_count_continuations - swapoff free all the continuation pages
2475 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2477 static void free_swap_count_continuations(struct swap_info_struct *si)
2481 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2483 head = vmalloc_to_page(si->swap_map + offset);
2484 if (page_private(head)) {
2485 struct list_head *this, *next;
2486 list_for_each_safe(this, next, &head->lru) {
2488 page = list_entry(this, struct page, lru);