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;
39 long total_swap_pages;
40 static int swap_overflow;
41 static int least_priority;
43 static const char Bad_file[] = "Bad swap file entry ";
44 static const char Unused_file[] = "Unused swap file entry ";
45 static const char Bad_offset[] = "Bad swap offset entry ";
46 static const char Unused_offset[] = "Unused swap offset entry ";
48 static struct swap_list_t swap_list = {-1, -1};
50 static struct swap_info_struct swap_info[MAX_SWAPFILES];
52 static DEFINE_MUTEX(swapon_mutex);
55 * We need this because the bdev->unplug_fn can sleep and we cannot
56 * hold swap_lock while calling the unplug_fn. And swap_lock
57 * cannot be turned into a mutex.
59 static DECLARE_RWSEM(swap_unplug_sem);
61 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
65 down_read(&swap_unplug_sem);
66 entry.val = page_private(page);
67 if (PageSwapCache(page)) {
68 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
69 struct backing_dev_info *bdi;
72 * If the page is removed from swapcache from under us (with a
73 * racy try_to_unuse/swapoff) we need an additional reference
74 * count to avoid reading garbage from page_private(page) above.
75 * If the WARN_ON triggers during a swapoff it maybe the race
76 * condition and it's harmless. However if it triggers without
77 * swapoff it signals a problem.
79 WARN_ON(page_count(page) <= 1);
81 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
82 blk_run_backing_dev(bdi, page);
84 up_read(&swap_unplug_sem);
87 #define SWAPFILE_CLUSTER 256
88 #define LATENCY_LIMIT 256
90 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
92 unsigned long offset, last_in_cluster;
93 int latency_ration = LATENCY_LIMIT;
96 * We try to cluster swap pages by allocating them sequentially
97 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
98 * way, however, we resort to first-free allocation, starting
99 * a new cluster. This prevents us from scattering swap pages
100 * all over the entire swap partition, so that we reduce
101 * overall disk seek times between swap pages. -- sct
102 * But we do now try to find an empty cluster. -Andrea
105 si->flags += SWP_SCANNING;
106 if (unlikely(!si->cluster_nr)) {
107 si->cluster_nr = SWAPFILE_CLUSTER - 1;
108 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
110 spin_unlock(&swap_lock);
112 offset = si->lowest_bit;
113 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
115 /* Locate the first empty (unaligned) cluster */
116 for (; last_in_cluster <= si->highest_bit; offset++) {
117 if (si->swap_map[offset])
118 last_in_cluster = offset + SWAPFILE_CLUSTER;
119 else if (offset == last_in_cluster) {
120 spin_lock(&swap_lock);
121 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
124 if (unlikely(--latency_ration < 0)) {
126 latency_ration = LATENCY_LIMIT;
129 spin_lock(&swap_lock);
135 offset = si->cluster_next;
136 if (offset > si->highest_bit)
137 lowest: offset = si->lowest_bit;
138 checks: if (!(si->flags & SWP_WRITEOK))
140 if (!si->highest_bit)
142 if (!si->swap_map[offset]) {
143 if (offset == si->lowest_bit)
145 if (offset == si->highest_bit)
148 if (si->inuse_pages == si->pages) {
149 si->lowest_bit = si->max;
152 si->swap_map[offset] = 1;
153 si->cluster_next = offset + 1;
154 si->flags -= SWP_SCANNING;
158 spin_unlock(&swap_lock);
159 while (++offset <= si->highest_bit) {
160 if (!si->swap_map[offset]) {
161 spin_lock(&swap_lock);
164 if (unlikely(--latency_ration < 0)) {
166 latency_ration = LATENCY_LIMIT;
169 spin_lock(&swap_lock);
173 si->flags -= SWP_SCANNING;
177 swp_entry_t get_swap_page(void)
179 struct swap_info_struct *si;
184 spin_lock(&swap_lock);
185 if (nr_swap_pages <= 0)
189 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
190 si = swap_info + type;
193 (!wrapped && si->prio != swap_info[next].prio)) {
194 next = swap_list.head;
198 if (!si->highest_bit)
200 if (!(si->flags & SWP_WRITEOK))
203 swap_list.next = next;
204 offset = scan_swap_map(si);
206 spin_unlock(&swap_lock);
207 return swp_entry(type, offset);
209 next = swap_list.next;
214 spin_unlock(&swap_lock);
215 return (swp_entry_t) {0};
218 swp_entry_t get_swap_page_of_type(int type)
220 struct swap_info_struct *si;
223 spin_lock(&swap_lock);
224 si = swap_info + type;
225 if (si->flags & SWP_WRITEOK) {
227 offset = scan_swap_map(si);
229 spin_unlock(&swap_lock);
230 return swp_entry(type, offset);
234 spin_unlock(&swap_lock);
235 return (swp_entry_t) {0};
238 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
240 struct swap_info_struct * p;
241 unsigned long offset, type;
245 type = swp_type(entry);
246 if (type >= nr_swapfiles)
248 p = & swap_info[type];
249 if (!(p->flags & SWP_USED))
251 offset = swp_offset(entry);
252 if (offset >= p->max)
254 if (!p->swap_map[offset])
256 spin_lock(&swap_lock);
260 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
263 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
266 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
269 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
274 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
276 int count = p->swap_map[offset];
278 if (count < SWAP_MAP_MAX) {
280 p->swap_map[offset] = count;
282 if (offset < p->lowest_bit)
283 p->lowest_bit = offset;
284 if (offset > p->highest_bit)
285 p->highest_bit = offset;
286 if (p->prio > swap_info[swap_list.next].prio)
287 swap_list.next = p - swap_info;
296 * Caller has made sure that the swapdevice corresponding to entry
297 * is still around or has not been recycled.
299 void swap_free(swp_entry_t entry)
301 struct swap_info_struct * p;
303 p = swap_info_get(entry);
305 swap_entry_free(p, swp_offset(entry));
306 spin_unlock(&swap_lock);
311 * How many references to page are currently swapped out?
313 static inline int page_swapcount(struct page *page)
316 struct swap_info_struct *p;
319 entry.val = page_private(page);
320 p = swap_info_get(entry);
322 /* Subtract the 1 for the swap cache itself */
323 count = p->swap_map[swp_offset(entry)] - 1;
324 spin_unlock(&swap_lock);
330 * We can write to an anon page without COW if there are no other references
331 * to it. And as a side-effect, free up its swap: because the old content
332 * on disk will never be read, and seeking back there to write new content
333 * later would only waste time away from clustering.
335 int reuse_swap_page(struct page *page)
339 VM_BUG_ON(!PageLocked(page));
340 count = page_mapcount(page);
341 if (count <= 1 && PageSwapCache(page)) {
342 count += page_swapcount(page);
343 if (count == 1 && !PageWriteback(page)) {
344 delete_from_swap_cache(page);
352 * If swap is getting full, or if there are no more mappings of this page,
353 * then try_to_free_swap is called to free its swap space.
355 int try_to_free_swap(struct page *page)
357 VM_BUG_ON(!PageLocked(page));
359 if (!PageSwapCache(page))
361 if (PageWriteback(page))
363 if (page_swapcount(page))
366 delete_from_swap_cache(page);
372 * Free the swap entry like above, but also try to
373 * free the page cache entry if it is the last user.
375 void free_swap_and_cache(swp_entry_t entry)
377 struct swap_info_struct * p;
378 struct page *page = NULL;
380 if (is_migration_entry(entry))
383 p = swap_info_get(entry);
385 if (swap_entry_free(p, swp_offset(entry)) == 1) {
386 page = find_get_page(&swapper_space, entry.val);
387 if (page && !trylock_page(page)) {
388 page_cache_release(page);
392 spin_unlock(&swap_lock);
396 * Not mapped elsewhere, or swap space full? Free it!
397 * Also recheck PageSwapCache now page is locked (above).
399 if (PageSwapCache(page) && !PageWriteback(page) &&
400 (!page_mapped(page) || vm_swap_full())) {
401 delete_from_swap_cache(page);
405 page_cache_release(page);
409 #ifdef CONFIG_HIBERNATION
411 * Find the swap type that corresponds to given device (if any).
413 * @offset - number of the PAGE_SIZE-sized block of the device, starting
414 * from 0, in which the swap header is expected to be located.
416 * This is needed for the suspend to disk (aka swsusp).
418 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
420 struct block_device *bdev = NULL;
424 bdev = bdget(device);
426 spin_lock(&swap_lock);
427 for (i = 0; i < nr_swapfiles; i++) {
428 struct swap_info_struct *sis = swap_info + i;
430 if (!(sis->flags & SWP_WRITEOK))
437 spin_unlock(&swap_lock);
440 if (bdev == sis->bdev) {
441 struct swap_extent *se;
443 se = list_entry(sis->extent_list.next,
444 struct swap_extent, list);
445 if (se->start_block == offset) {
449 spin_unlock(&swap_lock);
455 spin_unlock(&swap_lock);
463 * Return either the total number of swap pages of given type, or the number
464 * of free pages of that type (depending on @free)
466 * This is needed for software suspend
468 unsigned int count_swap_pages(int type, int free)
472 if (type < nr_swapfiles) {
473 spin_lock(&swap_lock);
474 if (swap_info[type].flags & SWP_WRITEOK) {
475 n = swap_info[type].pages;
477 n -= swap_info[type].inuse_pages;
479 spin_unlock(&swap_lock);
486 * No need to decide whether this PTE shares the swap entry with others,
487 * just let do_wp_page work it out if a write is requested later - to
488 * force COW, vm_page_prot omits write permission from any private vma.
490 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
491 unsigned long addr, swp_entry_t entry, struct page *page)
497 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
500 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
501 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
503 mem_cgroup_uncharge_page(page);
508 inc_mm_counter(vma->vm_mm, anon_rss);
510 set_pte_at(vma->vm_mm, addr, pte,
511 pte_mkold(mk_pte(page, vma->vm_page_prot)));
512 page_add_anon_rmap(page, vma, addr);
515 * Move the page to the active list so it is not
516 * immediately swapped out again after swapon.
520 pte_unmap_unlock(pte, ptl);
524 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
525 unsigned long addr, unsigned long end,
526 swp_entry_t entry, struct page *page)
528 pte_t swp_pte = swp_entry_to_pte(entry);
533 * We don't actually need pte lock while scanning for swp_pte: since
534 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
535 * page table while we're scanning; though it could get zapped, and on
536 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
537 * of unmatched parts which look like swp_pte, so unuse_pte must
538 * recheck under pte lock. Scanning without pte lock lets it be
539 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
541 pte = pte_offset_map(pmd, addr);
544 * swapoff spends a _lot_ of time in this loop!
545 * Test inline before going to call unuse_pte.
547 if (unlikely(pte_same(*pte, swp_pte))) {
549 ret = unuse_pte(vma, pmd, addr, entry, page);
552 pte = pte_offset_map(pmd, addr);
554 } while (pte++, addr += PAGE_SIZE, addr != end);
560 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
561 unsigned long addr, unsigned long end,
562 swp_entry_t entry, struct page *page)
568 pmd = pmd_offset(pud, addr);
570 next = pmd_addr_end(addr, end);
571 if (pmd_none_or_clear_bad(pmd))
573 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
576 } while (pmd++, addr = next, addr != end);
580 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
581 unsigned long addr, unsigned long end,
582 swp_entry_t entry, struct page *page)
588 pud = pud_offset(pgd, addr);
590 next = pud_addr_end(addr, end);
591 if (pud_none_or_clear_bad(pud))
593 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
596 } while (pud++, addr = next, addr != end);
600 static int unuse_vma(struct vm_area_struct *vma,
601 swp_entry_t entry, struct page *page)
604 unsigned long addr, end, next;
608 addr = page_address_in_vma(page, vma);
612 end = addr + PAGE_SIZE;
614 addr = vma->vm_start;
618 pgd = pgd_offset(vma->vm_mm, addr);
620 next = pgd_addr_end(addr, end);
621 if (pgd_none_or_clear_bad(pgd))
623 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
626 } while (pgd++, addr = next, addr != end);
630 static int unuse_mm(struct mm_struct *mm,
631 swp_entry_t entry, struct page *page)
633 struct vm_area_struct *vma;
636 if (!down_read_trylock(&mm->mmap_sem)) {
638 * Activate page so shrink_inactive_list is unlikely to unmap
639 * its ptes while lock is dropped, so swapoff can make progress.
643 down_read(&mm->mmap_sem);
646 for (vma = mm->mmap; vma; vma = vma->vm_next) {
647 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
650 up_read(&mm->mmap_sem);
651 return (ret < 0)? ret: 0;
655 * Scan swap_map from current position to next entry still in use.
656 * Recycle to start on reaching the end, returning 0 when empty.
658 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
661 unsigned int max = si->max;
662 unsigned int i = prev;
666 * No need for swap_lock here: we're just looking
667 * for whether an entry is in use, not modifying it; false
668 * hits are okay, and sys_swapoff() has already prevented new
669 * allocations from this area (while holding swap_lock).
678 * No entries in use at top of swap_map,
679 * loop back to start and recheck there.
685 count = si->swap_map[i];
686 if (count && count != SWAP_MAP_BAD)
693 * We completely avoid races by reading each swap page in advance,
694 * and then search for the process using it. All the necessary
695 * page table adjustments can then be made atomically.
697 static int try_to_unuse(unsigned int type)
699 struct swap_info_struct * si = &swap_info[type];
700 struct mm_struct *start_mm;
701 unsigned short *swap_map;
702 unsigned short swcount;
707 int reset_overflow = 0;
711 * When searching mms for an entry, a good strategy is to
712 * start at the first mm we freed the previous entry from
713 * (though actually we don't notice whether we or coincidence
714 * freed the entry). Initialize this start_mm with a hold.
716 * A simpler strategy would be to start at the last mm we
717 * freed the previous entry from; but that would take less
718 * advantage of mmlist ordering, which clusters forked mms
719 * together, child after parent. If we race with dup_mmap(), we
720 * prefer to resolve parent before child, lest we miss entries
721 * duplicated after we scanned child: using last mm would invert
722 * that. Though it's only a serious concern when an overflowed
723 * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
726 atomic_inc(&init_mm.mm_users);
729 * Keep on scanning until all entries have gone. Usually,
730 * one pass through swap_map is enough, but not necessarily:
731 * there are races when an instance of an entry might be missed.
733 while ((i = find_next_to_unuse(si, i)) != 0) {
734 if (signal_pending(current)) {
740 * Get a page for the entry, using the existing swap
741 * cache page if there is one. Otherwise, get a clean
742 * page and read the swap into it.
744 swap_map = &si->swap_map[i];
745 entry = swp_entry(type, i);
746 page = read_swap_cache_async(entry,
747 GFP_HIGHUSER_MOVABLE, NULL, 0);
750 * Either swap_duplicate() failed because entry
751 * has been freed independently, and will not be
752 * reused since sys_swapoff() already disabled
753 * allocation from here, or alloc_page() failed.
762 * Don't hold on to start_mm if it looks like exiting.
764 if (atomic_read(&start_mm->mm_users) == 1) {
767 atomic_inc(&init_mm.mm_users);
771 * Wait for and lock page. When do_swap_page races with
772 * try_to_unuse, do_swap_page can handle the fault much
773 * faster than try_to_unuse can locate the entry. This
774 * apparently redundant "wait_on_page_locked" lets try_to_unuse
775 * defer to do_swap_page in such a case - in some tests,
776 * do_swap_page and try_to_unuse repeatedly compete.
778 wait_on_page_locked(page);
779 wait_on_page_writeback(page);
781 wait_on_page_writeback(page);
784 * Remove all references to entry.
785 * Whenever we reach init_mm, there's no address space
786 * to search, but use it as a reminder to search shmem.
791 if (start_mm == &init_mm)
792 shmem = shmem_unuse(entry, page);
794 retval = unuse_mm(start_mm, entry, page);
797 int set_start_mm = (*swap_map >= swcount);
798 struct list_head *p = &start_mm->mmlist;
799 struct mm_struct *new_start_mm = start_mm;
800 struct mm_struct *prev_mm = start_mm;
801 struct mm_struct *mm;
803 atomic_inc(&new_start_mm->mm_users);
804 atomic_inc(&prev_mm->mm_users);
805 spin_lock(&mmlist_lock);
806 while (*swap_map > 1 && !retval && !shmem &&
807 (p = p->next) != &start_mm->mmlist) {
808 mm = list_entry(p, struct mm_struct, mmlist);
809 if (!atomic_inc_not_zero(&mm->mm_users))
811 spin_unlock(&mmlist_lock);
820 else if (mm == &init_mm) {
822 shmem = shmem_unuse(entry, page);
824 retval = unuse_mm(mm, entry, page);
825 if (set_start_mm && *swap_map < swcount) {
827 atomic_inc(&mm->mm_users);
831 spin_lock(&mmlist_lock);
833 spin_unlock(&mmlist_lock);
836 start_mm = new_start_mm;
839 /* page has already been unlocked and released */
847 page_cache_release(page);
852 * How could swap count reach 0x7fff when the maximum
853 * pid is 0x7fff, and there's no way to repeat a swap
854 * page within an mm (except in shmem, where it's the
855 * shared object which takes the reference count)?
856 * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
858 * If that's wrong, then we should worry more about
859 * exit_mmap() and do_munmap() cases described above:
860 * we might be resetting SWAP_MAP_MAX too early here.
861 * We know "Undead"s can happen, they're okay, so don't
862 * report them; but do report if we reset SWAP_MAP_MAX.
864 if (*swap_map == SWAP_MAP_MAX) {
865 spin_lock(&swap_lock);
867 spin_unlock(&swap_lock);
872 * If a reference remains (rare), we would like to leave
873 * the page in the swap cache; but try_to_unmap could
874 * then re-duplicate the entry once we drop page lock,
875 * so we might loop indefinitely; also, that page could
876 * not be swapped out to other storage meanwhile. So:
877 * delete from cache even if there's another reference,
878 * after ensuring that the data has been saved to disk -
879 * since if the reference remains (rarer), it will be
880 * read from disk into another page. Splitting into two
881 * pages would be incorrect if swap supported "shared
882 * private" pages, but they are handled by tmpfs files.
884 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
885 struct writeback_control wbc = {
886 .sync_mode = WB_SYNC_NONE,
889 swap_writepage(page, &wbc);
891 wait_on_page_writeback(page);
895 * It is conceivable that a racing task removed this page from
896 * swap cache just before we acquired the page lock at the top,
897 * or while we dropped it in unuse_mm(). The page might even
898 * be back in swap cache on another swap area: that we must not
899 * delete, since it may not have been written out to swap yet.
901 if (PageSwapCache(page) &&
902 likely(page_private(page) == entry.val))
903 delete_from_swap_cache(page);
906 * So we could skip searching mms once swap count went
907 * to 1, we did not mark any present ptes as dirty: must
908 * mark page dirty so shrink_page_list will preserve it.
912 page_cache_release(page);
915 * Make sure that we aren't completely killing
916 * interactive performance.
922 if (reset_overflow) {
923 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
930 * After a successful try_to_unuse, if no swap is now in use, we know
931 * we can empty the mmlist. swap_lock must be held on entry and exit.
932 * Note that mmlist_lock nests inside swap_lock, and an mm must be
933 * added to the mmlist just after page_duplicate - before would be racy.
935 static void drain_mmlist(void)
937 struct list_head *p, *next;
940 for (i = 0; i < nr_swapfiles; i++)
941 if (swap_info[i].inuse_pages)
943 spin_lock(&mmlist_lock);
944 list_for_each_safe(p, next, &init_mm.mmlist)
946 spin_unlock(&mmlist_lock);
950 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
951 * corresponds to page offset `offset'.
953 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
955 struct swap_extent *se = sis->curr_swap_extent;
956 struct swap_extent *start_se = se;
959 struct list_head *lh;
961 if (se->start_page <= offset &&
962 offset < (se->start_page + se->nr_pages)) {
963 return se->start_block + (offset - se->start_page);
966 if (lh == &sis->extent_list)
968 se = list_entry(lh, struct swap_extent, list);
969 sis->curr_swap_extent = se;
970 BUG_ON(se == start_se); /* It *must* be present */
974 #ifdef CONFIG_HIBERNATION
976 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
977 * corresponding to given index in swap_info (swap type).
979 sector_t swapdev_block(int swap_type, pgoff_t offset)
981 struct swap_info_struct *sis;
983 if (swap_type >= nr_swapfiles)
986 sis = swap_info + swap_type;
987 return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
989 #endif /* CONFIG_HIBERNATION */
992 * Free all of a swapdev's extent information
994 static void destroy_swap_extents(struct swap_info_struct *sis)
996 while (!list_empty(&sis->extent_list)) {
997 struct swap_extent *se;
999 se = list_entry(sis->extent_list.next,
1000 struct swap_extent, list);
1001 list_del(&se->list);
1007 * Add a block range (and the corresponding page range) into this swapdev's
1008 * extent list. The extent list is kept sorted in page order.
1010 * This function rather assumes that it is called in ascending page order.
1013 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1014 unsigned long nr_pages, sector_t start_block)
1016 struct swap_extent *se;
1017 struct swap_extent *new_se;
1018 struct list_head *lh;
1020 lh = sis->extent_list.prev; /* The highest page extent */
1021 if (lh != &sis->extent_list) {
1022 se = list_entry(lh, struct swap_extent, list);
1023 BUG_ON(se->start_page + se->nr_pages != start_page);
1024 if (se->start_block + se->nr_pages == start_block) {
1026 se->nr_pages += nr_pages;
1032 * No merge. Insert a new extent, preserving ordering.
1034 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1037 new_se->start_page = start_page;
1038 new_se->nr_pages = nr_pages;
1039 new_se->start_block = start_block;
1041 list_add_tail(&new_se->list, &sis->extent_list);
1046 * A `swap extent' is a simple thing which maps a contiguous range of pages
1047 * onto a contiguous range of disk blocks. An ordered list of swap extents
1048 * is built at swapon time and is then used at swap_writepage/swap_readpage
1049 * time for locating where on disk a page belongs.
1051 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1052 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1053 * swap files identically.
1055 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1056 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1057 * swapfiles are handled *identically* after swapon time.
1059 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1060 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1061 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1062 * requirements, they are simply tossed out - we will never use those blocks
1065 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1066 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1067 * which will scribble on the fs.
1069 * The amount of disk space which a single swap extent represents varies.
1070 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1071 * extents in the list. To avoid much list walking, we cache the previous
1072 * search location in `curr_swap_extent', and start new searches from there.
1073 * This is extremely effective. The average number of iterations in
1074 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1076 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1078 struct inode *inode;
1079 unsigned blocks_per_page;
1080 unsigned long page_no;
1082 sector_t probe_block;
1083 sector_t last_block;
1084 sector_t lowest_block = -1;
1085 sector_t highest_block = 0;
1089 inode = sis->swap_file->f_mapping->host;
1090 if (S_ISBLK(inode->i_mode)) {
1091 ret = add_swap_extent(sis, 0, sis->max, 0);
1096 blkbits = inode->i_blkbits;
1097 blocks_per_page = PAGE_SIZE >> blkbits;
1100 * Map all the blocks into the extent list. This code doesn't try
1105 last_block = i_size_read(inode) >> blkbits;
1106 while ((probe_block + blocks_per_page) <= last_block &&
1107 page_no < sis->max) {
1108 unsigned block_in_page;
1109 sector_t first_block;
1111 first_block = bmap(inode, probe_block);
1112 if (first_block == 0)
1116 * It must be PAGE_SIZE aligned on-disk
1118 if (first_block & (blocks_per_page - 1)) {
1123 for (block_in_page = 1; block_in_page < blocks_per_page;
1127 block = bmap(inode, probe_block + block_in_page);
1130 if (block != first_block + block_in_page) {
1137 first_block >>= (PAGE_SHIFT - blkbits);
1138 if (page_no) { /* exclude the header page */
1139 if (first_block < lowest_block)
1140 lowest_block = first_block;
1141 if (first_block > highest_block)
1142 highest_block = first_block;
1146 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1148 ret = add_swap_extent(sis, page_no, 1, first_block);
1153 probe_block += blocks_per_page;
1158 *span = 1 + highest_block - lowest_block;
1160 page_no = 1; /* force Empty message */
1162 sis->pages = page_no - 1;
1163 sis->highest_bit = page_no - 1;
1165 sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1166 struct swap_extent, list);
1169 printk(KERN_ERR "swapon: swapfile has holes\n");
1175 #if 0 /* We don't need this yet */
1176 #include <linux/backing-dev.h>
1177 int page_queue_congested(struct page *page)
1179 struct backing_dev_info *bdi;
1181 VM_BUG_ON(!PageLocked(page)); /* It pins the swap_info_struct */
1183 if (PageSwapCache(page)) {
1184 swp_entry_t entry = { .val = page_private(page) };
1185 struct swap_info_struct *sis;
1187 sis = get_swap_info_struct(swp_type(entry));
1188 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1190 bdi = page->mapping->backing_dev_info;
1191 return bdi_write_congested(bdi);
1195 asmlinkage long sys_swapoff(const char __user * specialfile)
1197 struct swap_info_struct * p = NULL;
1198 unsigned short *swap_map;
1199 struct file *swap_file, *victim;
1200 struct address_space *mapping;
1201 struct inode *inode;
1206 if (!capable(CAP_SYS_ADMIN))
1209 pathname = getname(specialfile);
1210 err = PTR_ERR(pathname);
1211 if (IS_ERR(pathname))
1214 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1216 err = PTR_ERR(victim);
1220 mapping = victim->f_mapping;
1222 spin_lock(&swap_lock);
1223 for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1224 p = swap_info + type;
1225 if (p->flags & SWP_WRITEOK) {
1226 if (p->swap_file->f_mapping == mapping)
1233 spin_unlock(&swap_lock);
1236 if (!security_vm_enough_memory(p->pages))
1237 vm_unacct_memory(p->pages);
1240 spin_unlock(&swap_lock);
1244 swap_list.head = p->next;
1246 swap_info[prev].next = p->next;
1248 if (type == swap_list.next) {
1249 /* just pick something that's safe... */
1250 swap_list.next = swap_list.head;
1253 for (i = p->next; i >= 0; i = swap_info[i].next)
1254 swap_info[i].prio = p->prio--;
1257 nr_swap_pages -= p->pages;
1258 total_swap_pages -= p->pages;
1259 p->flags &= ~SWP_WRITEOK;
1260 spin_unlock(&swap_lock);
1262 current->flags |= PF_SWAPOFF;
1263 err = try_to_unuse(type);
1264 current->flags &= ~PF_SWAPOFF;
1267 /* re-insert swap space back into swap_list */
1268 spin_lock(&swap_lock);
1270 p->prio = --least_priority;
1272 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1273 if (p->prio >= swap_info[i].prio)
1279 swap_list.head = swap_list.next = p - swap_info;
1281 swap_info[prev].next = p - swap_info;
1282 nr_swap_pages += p->pages;
1283 total_swap_pages += p->pages;
1284 p->flags |= SWP_WRITEOK;
1285 spin_unlock(&swap_lock);
1289 /* wait for any unplug function to finish */
1290 down_write(&swap_unplug_sem);
1291 up_write(&swap_unplug_sem);
1293 destroy_swap_extents(p);
1294 mutex_lock(&swapon_mutex);
1295 spin_lock(&swap_lock);
1298 /* wait for anyone still in scan_swap_map */
1299 p->highest_bit = 0; /* cuts scans short */
1300 while (p->flags >= SWP_SCANNING) {
1301 spin_unlock(&swap_lock);
1302 schedule_timeout_uninterruptible(1);
1303 spin_lock(&swap_lock);
1306 swap_file = p->swap_file;
1307 p->swap_file = NULL;
1309 swap_map = p->swap_map;
1312 spin_unlock(&swap_lock);
1313 mutex_unlock(&swapon_mutex);
1315 inode = mapping->host;
1316 if (S_ISBLK(inode->i_mode)) {
1317 struct block_device *bdev = I_BDEV(inode);
1318 set_blocksize(bdev, p->old_block_size);
1321 mutex_lock(&inode->i_mutex);
1322 inode->i_flags &= ~S_SWAPFILE;
1323 mutex_unlock(&inode->i_mutex);
1325 filp_close(swap_file, NULL);
1329 filp_close(victim, NULL);
1334 #ifdef CONFIG_PROC_FS
1336 static void *swap_start(struct seq_file *swap, loff_t *pos)
1338 struct swap_info_struct *ptr = swap_info;
1342 mutex_lock(&swapon_mutex);
1345 return SEQ_START_TOKEN;
1347 for (i = 0; i < nr_swapfiles; i++, ptr++) {
1348 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1357 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1359 struct swap_info_struct *ptr;
1360 struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1362 if (v == SEQ_START_TOKEN)
1369 for (; ptr < endptr; ptr++) {
1370 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1379 static void swap_stop(struct seq_file *swap, void *v)
1381 mutex_unlock(&swapon_mutex);
1384 static int swap_show(struct seq_file *swap, void *v)
1386 struct swap_info_struct *ptr = v;
1390 if (ptr == SEQ_START_TOKEN) {
1391 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1395 file = ptr->swap_file;
1396 len = seq_path(swap, &file->f_path, " \t\n\\");
1397 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1398 len < 40 ? 40 - len : 1, " ",
1399 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1400 "partition" : "file\t",
1401 ptr->pages << (PAGE_SHIFT - 10),
1402 ptr->inuse_pages << (PAGE_SHIFT - 10),
1407 static const struct seq_operations swaps_op = {
1408 .start = swap_start,
1414 static int swaps_open(struct inode *inode, struct file *file)
1416 return seq_open(file, &swaps_op);
1419 static const struct file_operations proc_swaps_operations = {
1422 .llseek = seq_lseek,
1423 .release = seq_release,
1426 static int __init procswaps_init(void)
1428 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1431 __initcall(procswaps_init);
1432 #endif /* CONFIG_PROC_FS */
1434 #ifdef MAX_SWAPFILES_CHECK
1435 static int __init max_swapfiles_check(void)
1437 MAX_SWAPFILES_CHECK();
1440 late_initcall(max_swapfiles_check);
1444 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1446 * The swapon system call
1448 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1450 struct swap_info_struct * p;
1452 struct block_device *bdev = NULL;
1453 struct file *swap_file = NULL;
1454 struct address_space *mapping;
1458 union swap_header *swap_header = NULL;
1459 unsigned int nr_good_pages = 0;
1462 unsigned long maxpages = 1;
1463 unsigned long swapfilepages;
1464 unsigned short *swap_map = NULL;
1465 struct page *page = NULL;
1466 struct inode *inode = NULL;
1469 if (!capable(CAP_SYS_ADMIN))
1471 spin_lock(&swap_lock);
1473 for (type = 0 ; type < nr_swapfiles ; type++,p++)
1474 if (!(p->flags & SWP_USED))
1477 if (type >= MAX_SWAPFILES) {
1478 spin_unlock(&swap_lock);
1481 if (type >= nr_swapfiles)
1482 nr_swapfiles = type+1;
1483 memset(p, 0, sizeof(*p));
1484 INIT_LIST_HEAD(&p->extent_list);
1485 p->flags = SWP_USED;
1487 spin_unlock(&swap_lock);
1488 name = getname(specialfile);
1489 error = PTR_ERR(name);
1494 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1495 error = PTR_ERR(swap_file);
1496 if (IS_ERR(swap_file)) {
1501 p->swap_file = swap_file;
1502 mapping = swap_file->f_mapping;
1503 inode = mapping->host;
1506 for (i = 0; i < nr_swapfiles; i++) {
1507 struct swap_info_struct *q = &swap_info[i];
1509 if (i == type || !q->swap_file)
1511 if (mapping == q->swap_file->f_mapping)
1516 if (S_ISBLK(inode->i_mode)) {
1517 bdev = I_BDEV(inode);
1518 error = bd_claim(bdev, sys_swapon);
1524 p->old_block_size = block_size(bdev);
1525 error = set_blocksize(bdev, PAGE_SIZE);
1529 } else if (S_ISREG(inode->i_mode)) {
1530 p->bdev = inode->i_sb->s_bdev;
1531 mutex_lock(&inode->i_mutex);
1533 if (IS_SWAPFILE(inode)) {
1541 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1544 * Read the swap header.
1546 if (!mapping->a_ops->readpage) {
1550 page = read_mapping_page(mapping, 0, swap_file);
1552 error = PTR_ERR(page);
1555 swap_header = kmap(page);
1557 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1558 printk(KERN_ERR "Unable to find swap-space signature\n");
1563 /* swap partition endianess hack... */
1564 if (swab32(swap_header->info.version) == 1) {
1565 swab32s(&swap_header->info.version);
1566 swab32s(&swap_header->info.last_page);
1567 swab32s(&swap_header->info.nr_badpages);
1568 for (i = 0; i < swap_header->info.nr_badpages; i++)
1569 swab32s(&swap_header->info.badpages[i]);
1571 /* Check the swap header's sub-version */
1572 if (swap_header->info.version != 1) {
1574 "Unable to handle swap header version %d\n",
1575 swap_header->info.version);
1581 p->cluster_next = 1;
1584 * Find out how many pages are allowed for a single swap
1585 * device. There are two limiting factors: 1) the number of
1586 * bits for the swap offset in the swp_entry_t type and
1587 * 2) the number of bits in the a swap pte as defined by
1588 * the different architectures. In order to find the
1589 * largest possible bit mask a swap entry with swap type 0
1590 * and swap offset ~0UL is created, encoded to a swap pte,
1591 * decoded to a swp_entry_t again and finally the swap
1592 * offset is extracted. This will mask all the bits from
1593 * the initial ~0UL mask that can't be encoded in either
1594 * the swp_entry_t or the architecture definition of a
1597 maxpages = swp_offset(pte_to_swp_entry(
1598 swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1599 if (maxpages > swap_header->info.last_page)
1600 maxpages = swap_header->info.last_page;
1601 p->highest_bit = maxpages - 1;
1606 if (swapfilepages && maxpages > swapfilepages) {
1608 "Swap area shorter than signature indicates\n");
1611 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1613 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1616 /* OK, set up the swap map and apply the bad block list */
1617 swap_map = vmalloc(maxpages * sizeof(short));
1623 memset(swap_map, 0, maxpages * sizeof(short));
1624 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1625 int page_nr = swap_header->info.badpages[i];
1626 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1630 swap_map[page_nr] = SWAP_MAP_BAD;
1632 nr_good_pages = swap_header->info.last_page -
1633 swap_header->info.nr_badpages -
1634 1 /* header page */;
1636 if (nr_good_pages) {
1637 swap_map[0] = SWAP_MAP_BAD;
1639 p->pages = nr_good_pages;
1640 nr_extents = setup_swap_extents(p, &span);
1641 if (nr_extents < 0) {
1645 nr_good_pages = p->pages;
1647 if (!nr_good_pages) {
1648 printk(KERN_WARNING "Empty swap-file\n");
1653 mutex_lock(&swapon_mutex);
1654 spin_lock(&swap_lock);
1655 if (swap_flags & SWAP_FLAG_PREFER)
1657 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1659 p->prio = --least_priority;
1660 p->swap_map = swap_map;
1661 p->flags |= SWP_WRITEOK;
1662 nr_swap_pages += nr_good_pages;
1663 total_swap_pages += nr_good_pages;
1665 printk(KERN_INFO "Adding %uk swap on %s. "
1666 "Priority:%d extents:%d across:%lluk\n",
1667 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1668 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1670 /* insert swap space into swap_list: */
1672 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1673 if (p->prio >= swap_info[i].prio) {
1680 swap_list.head = swap_list.next = p - swap_info;
1682 swap_info[prev].next = p - swap_info;
1684 spin_unlock(&swap_lock);
1685 mutex_unlock(&swapon_mutex);
1690 set_blocksize(bdev, p->old_block_size);
1693 destroy_swap_extents(p);
1695 spin_lock(&swap_lock);
1696 p->swap_file = NULL;
1698 spin_unlock(&swap_lock);
1701 filp_close(swap_file, NULL);
1703 if (page && !IS_ERR(page)) {
1705 page_cache_release(page);
1711 inode->i_flags |= S_SWAPFILE;
1712 mutex_unlock(&inode->i_mutex);
1717 void si_swapinfo(struct sysinfo *val)
1720 unsigned long nr_to_be_unused = 0;
1722 spin_lock(&swap_lock);
1723 for (i = 0; i < nr_swapfiles; i++) {
1724 if (!(swap_info[i].flags & SWP_USED) ||
1725 (swap_info[i].flags & SWP_WRITEOK))
1727 nr_to_be_unused += swap_info[i].inuse_pages;
1729 val->freeswap = nr_swap_pages + nr_to_be_unused;
1730 val->totalswap = total_swap_pages + nr_to_be_unused;
1731 spin_unlock(&swap_lock);
1735 * Verify that a swap entry is valid and increment its swap map count.
1737 * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1738 * "permanent", but will be reclaimed by the next swapoff.
1740 int swap_duplicate(swp_entry_t entry)
1742 struct swap_info_struct * p;
1743 unsigned long offset, type;
1746 if (is_migration_entry(entry))
1749 type = swp_type(entry);
1750 if (type >= nr_swapfiles)
1752 p = type + swap_info;
1753 offset = swp_offset(entry);
1755 spin_lock(&swap_lock);
1756 if (offset < p->max && p->swap_map[offset]) {
1757 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1758 p->swap_map[offset]++;
1760 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1761 if (swap_overflow++ < 5)
1762 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1763 p->swap_map[offset] = SWAP_MAP_MAX;
1767 spin_unlock(&swap_lock);
1772 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1776 struct swap_info_struct *
1777 get_swap_info_struct(unsigned type)
1779 return &swap_info[type];
1783 * swap_lock prevents swap_map being freed. Don't grab an extra
1784 * reference on the swaphandle, it doesn't matter if it becomes unused.
1786 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1788 struct swap_info_struct *si;
1789 int our_page_cluster = page_cluster;
1790 pgoff_t target, toff;
1794 if (!our_page_cluster) /* no readahead */
1797 si = &swap_info[swp_type(entry)];
1798 target = swp_offset(entry);
1799 base = (target >> our_page_cluster) << our_page_cluster;
1800 end = base + (1 << our_page_cluster);
1801 if (!base) /* first page is swap header */
1804 spin_lock(&swap_lock);
1805 if (end > si->max) /* don't go beyond end of map */
1808 /* Count contiguous allocated slots above our target */
1809 for (toff = target; ++toff < end; nr_pages++) {
1810 /* Don't read in free or bad pages */
1811 if (!si->swap_map[toff])
1813 if (si->swap_map[toff] == SWAP_MAP_BAD)
1816 /* Count contiguous allocated slots below our target */
1817 for (toff = target; --toff >= base; nr_pages++) {
1818 /* Don't read in free or bad pages */
1819 if (!si->swap_map[toff])
1821 if (si->swap_map[toff] == SWAP_MAP_BAD)
1824 spin_unlock(&swap_lock);
1827 * Indicate starting offset, and return number of pages to get:
1828 * if only 1, say 0, since there's then no readahead to be done.
1831 return nr_pages? ++nr_pages: 0;