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/writeback.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/init.h>
23 #include <linux/module.h>
24 #include <linux/rmap.h>
25 #include <linux/security.h>
26 #include <linux/backing-dev.h>
27 #include <linux/mutex.h>
28 #include <linux/capability.h>
29 #include <linux/syscalls.h>
30 #include <linux/memcontrol.h>
32 #include <asm/pgtable.h>
33 #include <asm/tlbflush.h>
34 #include <linux/swapops.h>
36 static DEFINE_SPINLOCK(swap_lock);
37 static unsigned int nr_swapfiles;
38 long total_swap_pages;
39 static int swap_overflow;
40 static int least_priority;
42 static const char Bad_file[] = "Bad swap file entry ";
43 static const char Unused_file[] = "Unused swap file entry ";
44 static const char Bad_offset[] = "Bad swap offset entry ";
45 static const char Unused_offset[] = "Unused swap offset entry ";
47 static struct swap_list_t swap_list = {-1, -1};
49 static struct swap_info_struct swap_info[MAX_SWAPFILES];
51 static DEFINE_MUTEX(swapon_mutex);
54 * We need this because the bdev->unplug_fn can sleep and we cannot
55 * hold swap_lock while calling the unplug_fn. And swap_lock
56 * cannot be turned into a mutex.
58 static DECLARE_RWSEM(swap_unplug_sem);
60 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
64 down_read(&swap_unplug_sem);
65 entry.val = page_private(page);
66 if (PageSwapCache(page)) {
67 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
68 struct backing_dev_info *bdi;
71 * If the page is removed from swapcache from under us (with a
72 * racy try_to_unuse/swapoff) we need an additional reference
73 * count to avoid reading garbage from page_private(page) above.
74 * If the WARN_ON triggers during a swapoff it maybe the race
75 * condition and it's harmless. However if it triggers without
76 * swapoff it signals a problem.
78 WARN_ON(page_count(page) <= 1);
80 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
81 blk_run_backing_dev(bdi, page);
83 up_read(&swap_unplug_sem);
86 #define SWAPFILE_CLUSTER 256
87 #define LATENCY_LIMIT 256
89 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
91 unsigned long offset, last_in_cluster;
92 int latency_ration = LATENCY_LIMIT;
95 * We try to cluster swap pages by allocating them sequentially
96 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
97 * way, however, we resort to first-free allocation, starting
98 * a new cluster. This prevents us from scattering swap pages
99 * all over the entire swap partition, so that we reduce
100 * overall disk seek times between swap pages. -- sct
101 * But we do now try to find an empty cluster. -Andrea
104 si->flags += SWP_SCANNING;
105 if (unlikely(!si->cluster_nr)) {
106 si->cluster_nr = SWAPFILE_CLUSTER - 1;
107 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
109 spin_unlock(&swap_lock);
111 offset = si->lowest_bit;
112 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
114 /* Locate the first empty (unaligned) cluster */
115 for (; last_in_cluster <= si->highest_bit; offset++) {
116 if (si->swap_map[offset])
117 last_in_cluster = offset + SWAPFILE_CLUSTER;
118 else if (offset == last_in_cluster) {
119 spin_lock(&swap_lock);
120 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
123 if (unlikely(--latency_ration < 0)) {
125 latency_ration = LATENCY_LIMIT;
128 spin_lock(&swap_lock);
134 offset = si->cluster_next;
135 if (offset > si->highest_bit)
136 lowest: offset = si->lowest_bit;
137 checks: if (!(si->flags & SWP_WRITEOK))
139 if (!si->highest_bit)
141 if (!si->swap_map[offset]) {
142 if (offset == si->lowest_bit)
144 if (offset == si->highest_bit)
147 if (si->inuse_pages == si->pages) {
148 si->lowest_bit = si->max;
151 si->swap_map[offset] = 1;
152 si->cluster_next = offset + 1;
153 si->flags -= SWP_SCANNING;
157 spin_unlock(&swap_lock);
158 while (++offset <= si->highest_bit) {
159 if (!si->swap_map[offset]) {
160 spin_lock(&swap_lock);
163 if (unlikely(--latency_ration < 0)) {
165 latency_ration = LATENCY_LIMIT;
168 spin_lock(&swap_lock);
172 si->flags -= SWP_SCANNING;
176 swp_entry_t get_swap_page(void)
178 struct swap_info_struct *si;
183 spin_lock(&swap_lock);
184 if (nr_swap_pages <= 0)
188 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
189 si = swap_info + type;
192 (!wrapped && si->prio != swap_info[next].prio)) {
193 next = swap_list.head;
197 if (!si->highest_bit)
199 if (!(si->flags & SWP_WRITEOK))
202 swap_list.next = next;
203 offset = scan_swap_map(si);
205 spin_unlock(&swap_lock);
206 return swp_entry(type, offset);
208 next = swap_list.next;
213 spin_unlock(&swap_lock);
214 return (swp_entry_t) {0};
217 swp_entry_t get_swap_page_of_type(int type)
219 struct swap_info_struct *si;
222 spin_lock(&swap_lock);
223 si = swap_info + type;
224 if (si->flags & SWP_WRITEOK) {
226 offset = scan_swap_map(si);
228 spin_unlock(&swap_lock);
229 return swp_entry(type, offset);
233 spin_unlock(&swap_lock);
234 return (swp_entry_t) {0};
237 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
239 struct swap_info_struct * p;
240 unsigned long offset, type;
244 type = swp_type(entry);
245 if (type >= nr_swapfiles)
247 p = & swap_info[type];
248 if (!(p->flags & SWP_USED))
250 offset = swp_offset(entry);
251 if (offset >= p->max)
253 if (!p->swap_map[offset])
255 spin_lock(&swap_lock);
259 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
262 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
265 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
268 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
273 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
275 int count = p->swap_map[offset];
277 if (count < SWAP_MAP_MAX) {
279 p->swap_map[offset] = count;
281 if (offset < p->lowest_bit)
282 p->lowest_bit = offset;
283 if (offset > p->highest_bit)
284 p->highest_bit = offset;
285 if (p->prio > swap_info[swap_list.next].prio)
286 swap_list.next = p - swap_info;
295 * Caller has made sure that the swapdevice corresponding to entry
296 * is still around or has not been recycled.
298 void swap_free(swp_entry_t entry)
300 struct swap_info_struct * p;
302 p = swap_info_get(entry);
304 swap_entry_free(p, swp_offset(entry));
305 spin_unlock(&swap_lock);
310 * How many references to page are currently swapped out?
312 static inline int page_swapcount(struct page *page)
315 struct swap_info_struct *p;
318 entry.val = page_private(page);
319 p = swap_info_get(entry);
321 /* Subtract the 1 for the swap cache itself */
322 count = p->swap_map[swp_offset(entry)] - 1;
323 spin_unlock(&swap_lock);
329 * We can write to an anon page without COW if there are no other references
330 * to it. And as a side-effect, free up its swap: because the old content
331 * on disk will never be read, and seeking back there to write new content
332 * later would only waste time away from clustering.
334 int reuse_swap_page(struct page *page)
338 VM_BUG_ON(!PageLocked(page));
339 count = page_mapcount(page);
340 if (count <= 1 && PageSwapCache(page)) {
341 count += page_swapcount(page);
342 if (count == 1 && !PageWriteback(page)) {
343 delete_from_swap_cache(page);
351 * Work out if there are any other processes sharing this
352 * swap cache page. Free it if you can. Return success.
354 static int remove_exclusive_swap_page_count(struct page *page, int count)
357 struct swap_info_struct * p;
360 VM_BUG_ON(!PageLocked(page));
362 if (!PageSwapCache(page))
364 if (PageWriteback(page))
366 if (page_count(page) != count) /* us + cache + ptes */
369 entry.val = page_private(page);
370 p = swap_info_get(entry);
374 /* Is the only swap cache user the cache itself? */
376 if (p->swap_map[swp_offset(entry)] == 1) {
377 /* Recheck the page count with the swapcache lock held.. */
378 spin_lock_irq(&swapper_space.tree_lock);
379 if ((page_count(page) == count) && !PageWriteback(page)) {
380 __delete_from_swap_cache(page);
384 spin_unlock_irq(&swapper_space.tree_lock);
386 spin_unlock(&swap_lock);
390 page_cache_release(page);
397 * Most of the time the page should have two references: one for the
398 * process and one for the swap cache.
400 int remove_exclusive_swap_page(struct page *page)
402 return remove_exclusive_swap_page_count(page, 2);
406 * The pageout code holds an extra reference to the page. That raises
407 * the reference count to test for to 2 for a page that is only in the
408 * swap cache plus 1 for each process that maps the page.
410 int remove_exclusive_swap_page_ref(struct page *page)
412 return remove_exclusive_swap_page_count(page, 2 + page_mapcount(page));
416 * Free the swap entry like above, but also try to
417 * free the page cache entry if it is the last user.
419 void free_swap_and_cache(swp_entry_t entry)
421 struct swap_info_struct * p;
422 struct page *page = NULL;
424 if (is_migration_entry(entry))
427 p = swap_info_get(entry);
429 if (swap_entry_free(p, swp_offset(entry)) == 1) {
430 page = find_get_page(&swapper_space, entry.val);
431 if (page && !trylock_page(page)) {
432 page_cache_release(page);
436 spin_unlock(&swap_lock);
441 one_user = (page_count(page) == 2);
442 /* Only cache user (+us), or swap space full? Free it! */
443 /* Also recheck PageSwapCache after page is locked (above) */
444 if (PageSwapCache(page) && !PageWriteback(page) &&
445 (one_user || vm_swap_full())) {
446 delete_from_swap_cache(page);
450 page_cache_release(page);
454 #ifdef CONFIG_HIBERNATION
456 * Find the swap type that corresponds to given device (if any).
458 * @offset - number of the PAGE_SIZE-sized block of the device, starting
459 * from 0, in which the swap header is expected to be located.
461 * This is needed for the suspend to disk (aka swsusp).
463 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
465 struct block_device *bdev = NULL;
469 bdev = bdget(device);
471 spin_lock(&swap_lock);
472 for (i = 0; i < nr_swapfiles; i++) {
473 struct swap_info_struct *sis = swap_info + i;
475 if (!(sis->flags & SWP_WRITEOK))
482 spin_unlock(&swap_lock);
485 if (bdev == sis->bdev) {
486 struct swap_extent *se;
488 se = list_entry(sis->extent_list.next,
489 struct swap_extent, list);
490 if (se->start_block == offset) {
494 spin_unlock(&swap_lock);
500 spin_unlock(&swap_lock);
508 * Return either the total number of swap pages of given type, or the number
509 * of free pages of that type (depending on @free)
511 * This is needed for software suspend
513 unsigned int count_swap_pages(int type, int free)
517 if (type < nr_swapfiles) {
518 spin_lock(&swap_lock);
519 if (swap_info[type].flags & SWP_WRITEOK) {
520 n = swap_info[type].pages;
522 n -= swap_info[type].inuse_pages;
524 spin_unlock(&swap_lock);
531 * No need to decide whether this PTE shares the swap entry with others,
532 * just let do_wp_page work it out if a write is requested later - to
533 * force COW, vm_page_prot omits write permission from any private vma.
535 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
536 unsigned long addr, swp_entry_t entry, struct page *page)
542 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
545 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
546 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
548 mem_cgroup_uncharge_page(page);
553 inc_mm_counter(vma->vm_mm, anon_rss);
555 set_pte_at(vma->vm_mm, addr, pte,
556 pte_mkold(mk_pte(page, vma->vm_page_prot)));
557 page_add_anon_rmap(page, vma, addr);
560 * Move the page to the active list so it is not
561 * immediately swapped out again after swapon.
565 pte_unmap_unlock(pte, ptl);
569 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
570 unsigned long addr, unsigned long end,
571 swp_entry_t entry, struct page *page)
573 pte_t swp_pte = swp_entry_to_pte(entry);
578 * We don't actually need pte lock while scanning for swp_pte: since
579 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
580 * page table while we're scanning; though it could get zapped, and on
581 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
582 * of unmatched parts which look like swp_pte, so unuse_pte must
583 * recheck under pte lock. Scanning without pte lock lets it be
584 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
586 pte = pte_offset_map(pmd, addr);
589 * swapoff spends a _lot_ of time in this loop!
590 * Test inline before going to call unuse_pte.
592 if (unlikely(pte_same(*pte, swp_pte))) {
594 ret = unuse_pte(vma, pmd, addr, entry, page);
597 pte = pte_offset_map(pmd, addr);
599 } while (pte++, addr += PAGE_SIZE, addr != end);
605 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
606 unsigned long addr, unsigned long end,
607 swp_entry_t entry, struct page *page)
613 pmd = pmd_offset(pud, addr);
615 next = pmd_addr_end(addr, end);
616 if (pmd_none_or_clear_bad(pmd))
618 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
621 } while (pmd++, addr = next, addr != end);
625 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
626 unsigned long addr, unsigned long end,
627 swp_entry_t entry, struct page *page)
633 pud = pud_offset(pgd, addr);
635 next = pud_addr_end(addr, end);
636 if (pud_none_or_clear_bad(pud))
638 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
641 } while (pud++, addr = next, addr != end);
645 static int unuse_vma(struct vm_area_struct *vma,
646 swp_entry_t entry, struct page *page)
649 unsigned long addr, end, next;
653 addr = page_address_in_vma(page, vma);
657 end = addr + PAGE_SIZE;
659 addr = vma->vm_start;
663 pgd = pgd_offset(vma->vm_mm, addr);
665 next = pgd_addr_end(addr, end);
666 if (pgd_none_or_clear_bad(pgd))
668 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
671 } while (pgd++, addr = next, addr != end);
675 static int unuse_mm(struct mm_struct *mm,
676 swp_entry_t entry, struct page *page)
678 struct vm_area_struct *vma;
681 if (!down_read_trylock(&mm->mmap_sem)) {
683 * Activate page so shrink_inactive_list is unlikely to unmap
684 * its ptes while lock is dropped, so swapoff can make progress.
688 down_read(&mm->mmap_sem);
691 for (vma = mm->mmap; vma; vma = vma->vm_next) {
692 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
695 up_read(&mm->mmap_sem);
696 return (ret < 0)? ret: 0;
700 * Scan swap_map from current position to next entry still in use.
701 * Recycle to start on reaching the end, returning 0 when empty.
703 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
706 unsigned int max = si->max;
707 unsigned int i = prev;
711 * No need for swap_lock here: we're just looking
712 * for whether an entry is in use, not modifying it; false
713 * hits are okay, and sys_swapoff() has already prevented new
714 * allocations from this area (while holding swap_lock).
723 * No entries in use at top of swap_map,
724 * loop back to start and recheck there.
730 count = si->swap_map[i];
731 if (count && count != SWAP_MAP_BAD)
738 * We completely avoid races by reading each swap page in advance,
739 * and then search for the process using it. All the necessary
740 * page table adjustments can then be made atomically.
742 static int try_to_unuse(unsigned int type)
744 struct swap_info_struct * si = &swap_info[type];
745 struct mm_struct *start_mm;
746 unsigned short *swap_map;
747 unsigned short swcount;
752 int reset_overflow = 0;
756 * When searching mms for an entry, a good strategy is to
757 * start at the first mm we freed the previous entry from
758 * (though actually we don't notice whether we or coincidence
759 * freed the entry). Initialize this start_mm with a hold.
761 * A simpler strategy would be to start at the last mm we
762 * freed the previous entry from; but that would take less
763 * advantage of mmlist ordering, which clusters forked mms
764 * together, child after parent. If we race with dup_mmap(), we
765 * prefer to resolve parent before child, lest we miss entries
766 * duplicated after we scanned child: using last mm would invert
767 * that. Though it's only a serious concern when an overflowed
768 * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
771 atomic_inc(&init_mm.mm_users);
774 * Keep on scanning until all entries have gone. Usually,
775 * one pass through swap_map is enough, but not necessarily:
776 * there are races when an instance of an entry might be missed.
778 while ((i = find_next_to_unuse(si, i)) != 0) {
779 if (signal_pending(current)) {
785 * Get a page for the entry, using the existing swap
786 * cache page if there is one. Otherwise, get a clean
787 * page and read the swap into it.
789 swap_map = &si->swap_map[i];
790 entry = swp_entry(type, i);
791 page = read_swap_cache_async(entry,
792 GFP_HIGHUSER_MOVABLE, NULL, 0);
795 * Either swap_duplicate() failed because entry
796 * has been freed independently, and will not be
797 * reused since sys_swapoff() already disabled
798 * allocation from here, or alloc_page() failed.
807 * Don't hold on to start_mm if it looks like exiting.
809 if (atomic_read(&start_mm->mm_users) == 1) {
812 atomic_inc(&init_mm.mm_users);
816 * Wait for and lock page. When do_swap_page races with
817 * try_to_unuse, do_swap_page can handle the fault much
818 * faster than try_to_unuse can locate the entry. This
819 * apparently redundant "wait_on_page_locked" lets try_to_unuse
820 * defer to do_swap_page in such a case - in some tests,
821 * do_swap_page and try_to_unuse repeatedly compete.
823 wait_on_page_locked(page);
824 wait_on_page_writeback(page);
826 wait_on_page_writeback(page);
829 * Remove all references to entry.
830 * Whenever we reach init_mm, there's no address space
831 * to search, but use it as a reminder to search shmem.
836 if (start_mm == &init_mm)
837 shmem = shmem_unuse(entry, page);
839 retval = unuse_mm(start_mm, entry, page);
842 int set_start_mm = (*swap_map >= swcount);
843 struct list_head *p = &start_mm->mmlist;
844 struct mm_struct *new_start_mm = start_mm;
845 struct mm_struct *prev_mm = start_mm;
846 struct mm_struct *mm;
848 atomic_inc(&new_start_mm->mm_users);
849 atomic_inc(&prev_mm->mm_users);
850 spin_lock(&mmlist_lock);
851 while (*swap_map > 1 && !retval && !shmem &&
852 (p = p->next) != &start_mm->mmlist) {
853 mm = list_entry(p, struct mm_struct, mmlist);
854 if (!atomic_inc_not_zero(&mm->mm_users))
856 spin_unlock(&mmlist_lock);
865 else if (mm == &init_mm) {
867 shmem = shmem_unuse(entry, page);
869 retval = unuse_mm(mm, entry, page);
870 if (set_start_mm && *swap_map < swcount) {
872 atomic_inc(&mm->mm_users);
876 spin_lock(&mmlist_lock);
878 spin_unlock(&mmlist_lock);
881 start_mm = new_start_mm;
884 /* page has already been unlocked and released */
892 page_cache_release(page);
897 * How could swap count reach 0x7fff when the maximum
898 * pid is 0x7fff, and there's no way to repeat a swap
899 * page within an mm (except in shmem, where it's the
900 * shared object which takes the reference count)?
901 * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
903 * If that's wrong, then we should worry more about
904 * exit_mmap() and do_munmap() cases described above:
905 * we might be resetting SWAP_MAP_MAX too early here.
906 * We know "Undead"s can happen, they're okay, so don't
907 * report them; but do report if we reset SWAP_MAP_MAX.
909 if (*swap_map == SWAP_MAP_MAX) {
910 spin_lock(&swap_lock);
912 spin_unlock(&swap_lock);
917 * If a reference remains (rare), we would like to leave
918 * the page in the swap cache; but try_to_unmap could
919 * then re-duplicate the entry once we drop page lock,
920 * so we might loop indefinitely; also, that page could
921 * not be swapped out to other storage meanwhile. So:
922 * delete from cache even if there's another reference,
923 * after ensuring that the data has been saved to disk -
924 * since if the reference remains (rarer), it will be
925 * read from disk into another page. Splitting into two
926 * pages would be incorrect if swap supported "shared
927 * private" pages, but they are handled by tmpfs files.
929 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
930 struct writeback_control wbc = {
931 .sync_mode = WB_SYNC_NONE,
934 swap_writepage(page, &wbc);
936 wait_on_page_writeback(page);
938 if (PageSwapCache(page))
939 delete_from_swap_cache(page);
942 * So we could skip searching mms once swap count went
943 * to 1, we did not mark any present ptes as dirty: must
944 * mark page dirty so shrink_page_list will preserve it.
948 page_cache_release(page);
951 * Make sure that we aren't completely killing
952 * interactive performance.
958 if (reset_overflow) {
959 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
966 * After a successful try_to_unuse, if no swap is now in use, we know
967 * we can empty the mmlist. swap_lock must be held on entry and exit.
968 * Note that mmlist_lock nests inside swap_lock, and an mm must be
969 * added to the mmlist just after page_duplicate - before would be racy.
971 static void drain_mmlist(void)
973 struct list_head *p, *next;
976 for (i = 0; i < nr_swapfiles; i++)
977 if (swap_info[i].inuse_pages)
979 spin_lock(&mmlist_lock);
980 list_for_each_safe(p, next, &init_mm.mmlist)
982 spin_unlock(&mmlist_lock);
986 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
987 * corresponds to page offset `offset'.
989 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
991 struct swap_extent *se = sis->curr_swap_extent;
992 struct swap_extent *start_se = se;
995 struct list_head *lh;
997 if (se->start_page <= offset &&
998 offset < (se->start_page + se->nr_pages)) {
999 return se->start_block + (offset - se->start_page);
1002 if (lh == &sis->extent_list)
1004 se = list_entry(lh, struct swap_extent, list);
1005 sis->curr_swap_extent = se;
1006 BUG_ON(se == start_se); /* It *must* be present */
1010 #ifdef CONFIG_HIBERNATION
1012 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1013 * corresponding to given index in swap_info (swap type).
1015 sector_t swapdev_block(int swap_type, pgoff_t offset)
1017 struct swap_info_struct *sis;
1019 if (swap_type >= nr_swapfiles)
1022 sis = swap_info + swap_type;
1023 return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1025 #endif /* CONFIG_HIBERNATION */
1028 * Free all of a swapdev's extent information
1030 static void destroy_swap_extents(struct swap_info_struct *sis)
1032 while (!list_empty(&sis->extent_list)) {
1033 struct swap_extent *se;
1035 se = list_entry(sis->extent_list.next,
1036 struct swap_extent, list);
1037 list_del(&se->list);
1043 * Add a block range (and the corresponding page range) into this swapdev's
1044 * extent list. The extent list is kept sorted in page order.
1046 * This function rather assumes that it is called in ascending page order.
1049 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1050 unsigned long nr_pages, sector_t start_block)
1052 struct swap_extent *se;
1053 struct swap_extent *new_se;
1054 struct list_head *lh;
1056 lh = sis->extent_list.prev; /* The highest page extent */
1057 if (lh != &sis->extent_list) {
1058 se = list_entry(lh, struct swap_extent, list);
1059 BUG_ON(se->start_page + se->nr_pages != start_page);
1060 if (se->start_block + se->nr_pages == start_block) {
1062 se->nr_pages += nr_pages;
1068 * No merge. Insert a new extent, preserving ordering.
1070 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1073 new_se->start_page = start_page;
1074 new_se->nr_pages = nr_pages;
1075 new_se->start_block = start_block;
1077 list_add_tail(&new_se->list, &sis->extent_list);
1082 * A `swap extent' is a simple thing which maps a contiguous range of pages
1083 * onto a contiguous range of disk blocks. An ordered list of swap extents
1084 * is built at swapon time and is then used at swap_writepage/swap_readpage
1085 * time for locating where on disk a page belongs.
1087 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1088 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1089 * swap files identically.
1091 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1092 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1093 * swapfiles are handled *identically* after swapon time.
1095 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1096 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1097 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1098 * requirements, they are simply tossed out - we will never use those blocks
1101 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1102 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1103 * which will scribble on the fs.
1105 * The amount of disk space which a single swap extent represents varies.
1106 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1107 * extents in the list. To avoid much list walking, we cache the previous
1108 * search location in `curr_swap_extent', and start new searches from there.
1109 * This is extremely effective. The average number of iterations in
1110 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1112 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1114 struct inode *inode;
1115 unsigned blocks_per_page;
1116 unsigned long page_no;
1118 sector_t probe_block;
1119 sector_t last_block;
1120 sector_t lowest_block = -1;
1121 sector_t highest_block = 0;
1125 inode = sis->swap_file->f_mapping->host;
1126 if (S_ISBLK(inode->i_mode)) {
1127 ret = add_swap_extent(sis, 0, sis->max, 0);
1132 blkbits = inode->i_blkbits;
1133 blocks_per_page = PAGE_SIZE >> blkbits;
1136 * Map all the blocks into the extent list. This code doesn't try
1141 last_block = i_size_read(inode) >> blkbits;
1142 while ((probe_block + blocks_per_page) <= last_block &&
1143 page_no < sis->max) {
1144 unsigned block_in_page;
1145 sector_t first_block;
1147 first_block = bmap(inode, probe_block);
1148 if (first_block == 0)
1152 * It must be PAGE_SIZE aligned on-disk
1154 if (first_block & (blocks_per_page - 1)) {
1159 for (block_in_page = 1; block_in_page < blocks_per_page;
1163 block = bmap(inode, probe_block + block_in_page);
1166 if (block != first_block + block_in_page) {
1173 first_block >>= (PAGE_SHIFT - blkbits);
1174 if (page_no) { /* exclude the header page */
1175 if (first_block < lowest_block)
1176 lowest_block = first_block;
1177 if (first_block > highest_block)
1178 highest_block = first_block;
1182 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1184 ret = add_swap_extent(sis, page_no, 1, first_block);
1189 probe_block += blocks_per_page;
1194 *span = 1 + highest_block - lowest_block;
1196 page_no = 1; /* force Empty message */
1198 sis->pages = page_no - 1;
1199 sis->highest_bit = page_no - 1;
1201 sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1202 struct swap_extent, list);
1205 printk(KERN_ERR "swapon: swapfile has holes\n");
1211 #if 0 /* We don't need this yet */
1212 #include <linux/backing-dev.h>
1213 int page_queue_congested(struct page *page)
1215 struct backing_dev_info *bdi;
1217 VM_BUG_ON(!PageLocked(page)); /* It pins the swap_info_struct */
1219 if (PageSwapCache(page)) {
1220 swp_entry_t entry = { .val = page_private(page) };
1221 struct swap_info_struct *sis;
1223 sis = get_swap_info_struct(swp_type(entry));
1224 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1226 bdi = page->mapping->backing_dev_info;
1227 return bdi_write_congested(bdi);
1231 asmlinkage long sys_swapoff(const char __user * specialfile)
1233 struct swap_info_struct * p = NULL;
1234 unsigned short *swap_map;
1235 struct file *swap_file, *victim;
1236 struct address_space *mapping;
1237 struct inode *inode;
1242 if (!capable(CAP_SYS_ADMIN))
1245 pathname = getname(specialfile);
1246 err = PTR_ERR(pathname);
1247 if (IS_ERR(pathname))
1250 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1252 err = PTR_ERR(victim);
1256 mapping = victim->f_mapping;
1258 spin_lock(&swap_lock);
1259 for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1260 p = swap_info + type;
1261 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1262 if (p->swap_file->f_mapping == mapping)
1269 spin_unlock(&swap_lock);
1272 if (!security_vm_enough_memory(p->pages))
1273 vm_unacct_memory(p->pages);
1276 spin_unlock(&swap_lock);
1280 swap_list.head = p->next;
1282 swap_info[prev].next = p->next;
1284 if (type == swap_list.next) {
1285 /* just pick something that's safe... */
1286 swap_list.next = swap_list.head;
1289 for (i = p->next; i >= 0; i = swap_info[i].next)
1290 swap_info[i].prio = p->prio--;
1293 nr_swap_pages -= p->pages;
1294 total_swap_pages -= p->pages;
1295 p->flags &= ~SWP_WRITEOK;
1296 spin_unlock(&swap_lock);
1298 current->flags |= PF_SWAPOFF;
1299 err = try_to_unuse(type);
1300 current->flags &= ~PF_SWAPOFF;
1303 /* re-insert swap space back into swap_list */
1304 spin_lock(&swap_lock);
1306 p->prio = --least_priority;
1308 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1309 if (p->prio >= swap_info[i].prio)
1315 swap_list.head = swap_list.next = p - swap_info;
1317 swap_info[prev].next = p - swap_info;
1318 nr_swap_pages += p->pages;
1319 total_swap_pages += p->pages;
1320 p->flags |= SWP_WRITEOK;
1321 spin_unlock(&swap_lock);
1325 /* wait for any unplug function to finish */
1326 down_write(&swap_unplug_sem);
1327 up_write(&swap_unplug_sem);
1329 destroy_swap_extents(p);
1330 mutex_lock(&swapon_mutex);
1331 spin_lock(&swap_lock);
1334 /* wait for anyone still in scan_swap_map */
1335 p->highest_bit = 0; /* cuts scans short */
1336 while (p->flags >= SWP_SCANNING) {
1337 spin_unlock(&swap_lock);
1338 schedule_timeout_uninterruptible(1);
1339 spin_lock(&swap_lock);
1342 swap_file = p->swap_file;
1343 p->swap_file = NULL;
1345 swap_map = p->swap_map;
1348 spin_unlock(&swap_lock);
1349 mutex_unlock(&swapon_mutex);
1351 inode = mapping->host;
1352 if (S_ISBLK(inode->i_mode)) {
1353 struct block_device *bdev = I_BDEV(inode);
1354 set_blocksize(bdev, p->old_block_size);
1357 mutex_lock(&inode->i_mutex);
1358 inode->i_flags &= ~S_SWAPFILE;
1359 mutex_unlock(&inode->i_mutex);
1361 filp_close(swap_file, NULL);
1365 filp_close(victim, NULL);
1370 #ifdef CONFIG_PROC_FS
1372 static void *swap_start(struct seq_file *swap, loff_t *pos)
1374 struct swap_info_struct *ptr = swap_info;
1378 mutex_lock(&swapon_mutex);
1381 return SEQ_START_TOKEN;
1383 for (i = 0; i < nr_swapfiles; i++, ptr++) {
1384 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1393 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1395 struct swap_info_struct *ptr;
1396 struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1398 if (v == SEQ_START_TOKEN)
1405 for (; ptr < endptr; ptr++) {
1406 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1415 static void swap_stop(struct seq_file *swap, void *v)
1417 mutex_unlock(&swapon_mutex);
1420 static int swap_show(struct seq_file *swap, void *v)
1422 struct swap_info_struct *ptr = v;
1426 if (ptr == SEQ_START_TOKEN) {
1427 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1431 file = ptr->swap_file;
1432 len = seq_path(swap, &file->f_path, " \t\n\\");
1433 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1434 len < 40 ? 40 - len : 1, " ",
1435 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1436 "partition" : "file\t",
1437 ptr->pages << (PAGE_SHIFT - 10),
1438 ptr->inuse_pages << (PAGE_SHIFT - 10),
1443 static const struct seq_operations swaps_op = {
1444 .start = swap_start,
1450 static int swaps_open(struct inode *inode, struct file *file)
1452 return seq_open(file, &swaps_op);
1455 static const struct file_operations proc_swaps_operations = {
1458 .llseek = seq_lseek,
1459 .release = seq_release,
1462 static int __init procswaps_init(void)
1464 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1467 __initcall(procswaps_init);
1468 #endif /* CONFIG_PROC_FS */
1470 #ifdef MAX_SWAPFILES_CHECK
1471 static int __init max_swapfiles_check(void)
1473 MAX_SWAPFILES_CHECK();
1476 late_initcall(max_swapfiles_check);
1480 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1482 * The swapon system call
1484 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1486 struct swap_info_struct * p;
1488 struct block_device *bdev = NULL;
1489 struct file *swap_file = NULL;
1490 struct address_space *mapping;
1494 union swap_header *swap_header = NULL;
1495 int swap_header_version;
1496 unsigned int nr_good_pages = 0;
1499 unsigned long maxpages = 1;
1501 unsigned short *swap_map = NULL;
1502 struct page *page = NULL;
1503 struct inode *inode = NULL;
1506 if (!capable(CAP_SYS_ADMIN))
1508 spin_lock(&swap_lock);
1510 for (type = 0 ; type < nr_swapfiles ; type++,p++)
1511 if (!(p->flags & SWP_USED))
1514 if (type >= MAX_SWAPFILES) {
1515 spin_unlock(&swap_lock);
1518 if (type >= nr_swapfiles)
1519 nr_swapfiles = type+1;
1520 memset(p, 0, sizeof(*p));
1521 INIT_LIST_HEAD(&p->extent_list);
1522 p->flags = SWP_USED;
1524 spin_unlock(&swap_lock);
1525 name = getname(specialfile);
1526 error = PTR_ERR(name);
1531 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1532 error = PTR_ERR(swap_file);
1533 if (IS_ERR(swap_file)) {
1538 p->swap_file = swap_file;
1539 mapping = swap_file->f_mapping;
1540 inode = mapping->host;
1543 for (i = 0; i < nr_swapfiles; i++) {
1544 struct swap_info_struct *q = &swap_info[i];
1546 if (i == type || !q->swap_file)
1548 if (mapping == q->swap_file->f_mapping)
1553 if (S_ISBLK(inode->i_mode)) {
1554 bdev = I_BDEV(inode);
1555 error = bd_claim(bdev, sys_swapon);
1561 p->old_block_size = block_size(bdev);
1562 error = set_blocksize(bdev, PAGE_SIZE);
1566 } else if (S_ISREG(inode->i_mode)) {
1567 p->bdev = inode->i_sb->s_bdev;
1568 mutex_lock(&inode->i_mutex);
1570 if (IS_SWAPFILE(inode)) {
1578 swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1581 * Read the swap header.
1583 if (!mapping->a_ops->readpage) {
1587 page = read_mapping_page(mapping, 0, swap_file);
1589 error = PTR_ERR(page);
1593 swap_header = page_address(page);
1595 if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1596 swap_header_version = 1;
1597 else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1598 swap_header_version = 2;
1600 printk(KERN_ERR "Unable to find swap-space signature\n");
1605 switch (swap_header_version) {
1607 printk(KERN_ERR "version 0 swap is no longer supported. "
1608 "Use mkswap -v1 %s\n", name);
1612 /* swap partition endianess hack... */
1613 if (swab32(swap_header->info.version) == 1) {
1614 swab32s(&swap_header->info.version);
1615 swab32s(&swap_header->info.last_page);
1616 swab32s(&swap_header->info.nr_badpages);
1617 for (i = 0; i < swap_header->info.nr_badpages; i++)
1618 swab32s(&swap_header->info.badpages[i]);
1620 /* Check the swap header's sub-version and the size of
1621 the swap file and bad block lists */
1622 if (swap_header->info.version != 1) {
1624 "Unable to handle swap header version %d\n",
1625 swap_header->info.version);
1631 p->cluster_next = 1;
1634 * Find out how many pages are allowed for a single swap
1635 * device. There are two limiting factors: 1) the number of
1636 * bits for the swap offset in the swp_entry_t type and
1637 * 2) the number of bits in the a swap pte as defined by
1638 * the different architectures. In order to find the
1639 * largest possible bit mask a swap entry with swap type 0
1640 * and swap offset ~0UL is created, encoded to a swap pte,
1641 * decoded to a swp_entry_t again and finally the swap
1642 * offset is extracted. This will mask all the bits from
1643 * the initial ~0UL mask that can't be encoded in either
1644 * the swp_entry_t or the architecture definition of a
1647 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1648 if (maxpages > swap_header->info.last_page)
1649 maxpages = swap_header->info.last_page;
1650 p->highest_bit = maxpages - 1;
1655 if (swapfilesize && maxpages > swapfilesize) {
1657 "Swap area shorter than signature indicates\n");
1660 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1662 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1665 /* OK, set up the swap map and apply the bad block list */
1666 swap_map = vmalloc(maxpages * sizeof(short));
1673 memset(swap_map, 0, maxpages * sizeof(short));
1674 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1675 int page_nr = swap_header->info.badpages[i];
1676 if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1679 swap_map[page_nr] = SWAP_MAP_BAD;
1681 nr_good_pages = swap_header->info.last_page -
1682 swap_header->info.nr_badpages -
1683 1 /* header page */;
1688 if (nr_good_pages) {
1689 swap_map[0] = SWAP_MAP_BAD;
1691 p->pages = nr_good_pages;
1692 nr_extents = setup_swap_extents(p, &span);
1693 if (nr_extents < 0) {
1697 nr_good_pages = p->pages;
1699 if (!nr_good_pages) {
1700 printk(KERN_WARNING "Empty swap-file\n");
1705 mutex_lock(&swapon_mutex);
1706 spin_lock(&swap_lock);
1707 if (swap_flags & SWAP_FLAG_PREFER)
1709 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1711 p->prio = --least_priority;
1712 p->swap_map = swap_map;
1713 p->flags = SWP_ACTIVE;
1714 nr_swap_pages += nr_good_pages;
1715 total_swap_pages += nr_good_pages;
1717 printk(KERN_INFO "Adding %uk swap on %s. "
1718 "Priority:%d extents:%d across:%lluk\n",
1719 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1720 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1722 /* insert swap space into swap_list: */
1724 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1725 if (p->prio >= swap_info[i].prio) {
1732 swap_list.head = swap_list.next = p - swap_info;
1734 swap_info[prev].next = p - swap_info;
1736 spin_unlock(&swap_lock);
1737 mutex_unlock(&swapon_mutex);
1742 set_blocksize(bdev, p->old_block_size);
1745 destroy_swap_extents(p);
1747 spin_lock(&swap_lock);
1748 p->swap_file = NULL;
1750 spin_unlock(&swap_lock);
1753 filp_close(swap_file, NULL);
1755 if (page && !IS_ERR(page)) {
1757 page_cache_release(page);
1763 inode->i_flags |= S_SWAPFILE;
1764 mutex_unlock(&inode->i_mutex);
1769 void si_swapinfo(struct sysinfo *val)
1772 unsigned long nr_to_be_unused = 0;
1774 spin_lock(&swap_lock);
1775 for (i = 0; i < nr_swapfiles; i++) {
1776 if (!(swap_info[i].flags & SWP_USED) ||
1777 (swap_info[i].flags & SWP_WRITEOK))
1779 nr_to_be_unused += swap_info[i].inuse_pages;
1781 val->freeswap = nr_swap_pages + nr_to_be_unused;
1782 val->totalswap = total_swap_pages + nr_to_be_unused;
1783 spin_unlock(&swap_lock);
1787 * Verify that a swap entry is valid and increment its swap map count.
1789 * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1790 * "permanent", but will be reclaimed by the next swapoff.
1792 int swap_duplicate(swp_entry_t entry)
1794 struct swap_info_struct * p;
1795 unsigned long offset, type;
1798 if (is_migration_entry(entry))
1801 type = swp_type(entry);
1802 if (type >= nr_swapfiles)
1804 p = type + swap_info;
1805 offset = swp_offset(entry);
1807 spin_lock(&swap_lock);
1808 if (offset < p->max && p->swap_map[offset]) {
1809 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1810 p->swap_map[offset]++;
1812 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1813 if (swap_overflow++ < 5)
1814 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1815 p->swap_map[offset] = SWAP_MAP_MAX;
1819 spin_unlock(&swap_lock);
1824 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1828 struct swap_info_struct *
1829 get_swap_info_struct(unsigned type)
1831 return &swap_info[type];
1835 * swap_lock prevents swap_map being freed. Don't grab an extra
1836 * reference on the swaphandle, it doesn't matter if it becomes unused.
1838 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1840 struct swap_info_struct *si;
1841 int our_page_cluster = page_cluster;
1842 pgoff_t target, toff;
1846 if (!our_page_cluster) /* no readahead */
1849 si = &swap_info[swp_type(entry)];
1850 target = swp_offset(entry);
1851 base = (target >> our_page_cluster) << our_page_cluster;
1852 end = base + (1 << our_page_cluster);
1853 if (!base) /* first page is swap header */
1856 spin_lock(&swap_lock);
1857 if (end > si->max) /* don't go beyond end of map */
1860 /* Count contiguous allocated slots above our target */
1861 for (toff = target; ++toff < end; nr_pages++) {
1862 /* Don't read in free or bad pages */
1863 if (!si->swap_map[toff])
1865 if (si->swap_map[toff] == SWAP_MAP_BAD)
1868 /* Count contiguous allocated slots below our target */
1869 for (toff = target; --toff >= base; nr_pages++) {
1870 /* Don't read in free or bad pages */
1871 if (!si->swap_map[toff])
1873 if (si->swap_map[toff] == SWAP_MAP_BAD)
1876 spin_unlock(&swap_lock);
1879 * Indicate starting offset, and return number of pages to get:
1880 * if only 1, say 0, since there's then no readahead to be done.
1883 return nr_pages? ++nr_pages: 0;