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/shmem_fs.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/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32 #include <linux/poll.h>
33 #include <linux/oom.h>
34 #include <linux/frontswap.h>
35 #include <linux/swapfile.h>
36 #include <linux/export.h>
38 #include <asm/pgtable.h>
39 #include <asm/tlbflush.h>
40 #include <linux/swapops.h>
41 #include <linux/page_cgroup.h>
43 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
45 static void free_swap_count_continuations(struct swap_info_struct *);
46 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
48 DEFINE_SPINLOCK(swap_lock);
49 static unsigned int nr_swapfiles;
51 long total_swap_pages;
52 static int least_priority;
54 static const char Bad_file[] = "Bad swap file entry ";
55 static const char Unused_file[] = "Unused swap file entry ";
56 static const char Bad_offset[] = "Bad swap offset entry ";
57 static const char Unused_offset[] = "Unused swap offset entry ";
59 struct swap_list_t swap_list = {-1, -1};
61 struct swap_info_struct *swap_info[MAX_SWAPFILES];
63 static DEFINE_MUTEX(swapon_mutex);
65 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
66 /* Activity counter to indicate that a swapon or swapoff has occurred */
67 static atomic_t proc_poll_event = ATOMIC_INIT(0);
69 static inline unsigned char swap_count(unsigned char ent)
71 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
74 /* returns 1 if swap entry is freed */
76 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
78 swp_entry_t entry = swp_entry(si->type, offset);
82 page = find_get_page(&swapper_space, entry.val);
86 * This function is called from scan_swap_map() and it's called
87 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
88 * We have to use trylock for avoiding deadlock. This is a special
89 * case and you should use try_to_free_swap() with explicit lock_page()
90 * in usual operations.
92 if (trylock_page(page)) {
93 ret = try_to_free_swap(page);
96 page_cache_release(page);
101 * swapon tell device that all the old swap contents can be discarded,
102 * to allow the swap device to optimize its wear-levelling.
104 static int discard_swap(struct swap_info_struct *si)
106 struct swap_extent *se;
107 sector_t start_block;
111 /* Do not discard the swap header page! */
112 se = &si->first_swap_extent;
113 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
114 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
116 err = blkdev_issue_discard(si->bdev, start_block,
117 nr_blocks, GFP_KERNEL, 0);
123 list_for_each_entry(se, &si->first_swap_extent.list, list) {
124 start_block = se->start_block << (PAGE_SHIFT - 9);
125 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
127 err = blkdev_issue_discard(si->bdev, start_block,
128 nr_blocks, GFP_KERNEL, 0);
134 return err; /* That will often be -EOPNOTSUPP */
138 * swap allocation tell device that a cluster of swap can now be discarded,
139 * to allow the swap device to optimize its wear-levelling.
141 static void discard_swap_cluster(struct swap_info_struct *si,
142 pgoff_t start_page, pgoff_t nr_pages)
144 struct swap_extent *se = si->curr_swap_extent;
145 int found_extent = 0;
148 struct list_head *lh;
150 if (se->start_page <= start_page &&
151 start_page < se->start_page + se->nr_pages) {
152 pgoff_t offset = start_page - se->start_page;
153 sector_t start_block = se->start_block + offset;
154 sector_t nr_blocks = se->nr_pages - offset;
156 if (nr_blocks > nr_pages)
157 nr_blocks = nr_pages;
158 start_page += nr_blocks;
159 nr_pages -= nr_blocks;
162 si->curr_swap_extent = se;
164 start_block <<= PAGE_SHIFT - 9;
165 nr_blocks <<= PAGE_SHIFT - 9;
166 if (blkdev_issue_discard(si->bdev, start_block,
167 nr_blocks, GFP_NOIO, 0))
172 se = list_entry(lh, struct swap_extent, list);
176 static int wait_for_discard(void *word)
182 #define SWAPFILE_CLUSTER 256
183 #define LATENCY_LIMIT 256
185 static unsigned long scan_swap_map(struct swap_info_struct *si,
188 unsigned long offset;
189 unsigned long scan_base;
190 unsigned long last_in_cluster = 0;
191 int latency_ration = LATENCY_LIMIT;
192 int found_free_cluster = 0;
195 * We try to cluster swap pages by allocating them sequentially
196 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
197 * way, however, we resort to first-free allocation, starting
198 * a new cluster. This prevents us from scattering swap pages
199 * all over the entire swap partition, so that we reduce
200 * overall disk seek times between swap pages. -- sct
201 * But we do now try to find an empty cluster. -Andrea
202 * And we let swap pages go all over an SSD partition. Hugh
205 si->flags += SWP_SCANNING;
206 scan_base = offset = si->cluster_next;
208 if (unlikely(!si->cluster_nr--)) {
209 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
210 si->cluster_nr = SWAPFILE_CLUSTER - 1;
213 if (si->flags & SWP_DISCARDABLE) {
215 * Start range check on racing allocations, in case
216 * they overlap the cluster we eventually decide on
217 * (we scan without swap_lock to allow preemption).
218 * It's hardly conceivable that cluster_nr could be
219 * wrapped during our scan, but don't depend on it.
221 if (si->lowest_alloc)
223 si->lowest_alloc = si->max;
224 si->highest_alloc = 0;
226 spin_unlock(&swap_lock);
229 * If seek is expensive, start searching for new cluster from
230 * start of partition, to minimize the span of allocated swap.
231 * But if seek is cheap, search from our current position, so
232 * that swap is allocated from all over the partition: if the
233 * Flash Translation Layer only remaps within limited zones,
234 * we don't want to wear out the first zone too quickly.
236 if (!(si->flags & SWP_SOLIDSTATE))
237 scan_base = offset = si->lowest_bit;
238 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
240 /* Locate the first empty (unaligned) cluster */
241 for (; last_in_cluster <= si->highest_bit; offset++) {
242 if (si->swap_map[offset])
243 last_in_cluster = offset + SWAPFILE_CLUSTER;
244 else if (offset == last_in_cluster) {
245 spin_lock(&swap_lock);
246 offset -= SWAPFILE_CLUSTER - 1;
247 si->cluster_next = offset;
248 si->cluster_nr = SWAPFILE_CLUSTER - 1;
249 found_free_cluster = 1;
252 if (unlikely(--latency_ration < 0)) {
254 latency_ration = LATENCY_LIMIT;
258 offset = si->lowest_bit;
259 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
261 /* Locate the first empty (unaligned) cluster */
262 for (; last_in_cluster < scan_base; offset++) {
263 if (si->swap_map[offset])
264 last_in_cluster = offset + SWAPFILE_CLUSTER;
265 else if (offset == last_in_cluster) {
266 spin_lock(&swap_lock);
267 offset -= SWAPFILE_CLUSTER - 1;
268 si->cluster_next = offset;
269 si->cluster_nr = SWAPFILE_CLUSTER - 1;
270 found_free_cluster = 1;
273 if (unlikely(--latency_ration < 0)) {
275 latency_ration = LATENCY_LIMIT;
280 spin_lock(&swap_lock);
281 si->cluster_nr = SWAPFILE_CLUSTER - 1;
282 si->lowest_alloc = 0;
286 if (!(si->flags & SWP_WRITEOK))
288 if (!si->highest_bit)
290 if (offset > si->highest_bit)
291 scan_base = offset = si->lowest_bit;
293 /* reuse swap entry of cache-only swap if not busy. */
294 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
296 spin_unlock(&swap_lock);
297 swap_was_freed = __try_to_reclaim_swap(si, offset);
298 spin_lock(&swap_lock);
299 /* entry was freed successfully, try to use this again */
302 goto scan; /* check next one */
305 if (si->swap_map[offset])
308 if (offset == si->lowest_bit)
310 if (offset == si->highest_bit)
313 if (si->inuse_pages == si->pages) {
314 si->lowest_bit = si->max;
317 si->swap_map[offset] = usage;
318 si->cluster_next = offset + 1;
319 si->flags -= SWP_SCANNING;
321 if (si->lowest_alloc) {
323 * Only set when SWP_DISCARDABLE, and there's a scan
324 * for a free cluster in progress or just completed.
326 if (found_free_cluster) {
328 * To optimize wear-levelling, discard the
329 * old data of the cluster, taking care not to
330 * discard any of its pages that have already
331 * been allocated by racing tasks (offset has
332 * already stepped over any at the beginning).
334 if (offset < si->highest_alloc &&
335 si->lowest_alloc <= last_in_cluster)
336 last_in_cluster = si->lowest_alloc - 1;
337 si->flags |= SWP_DISCARDING;
338 spin_unlock(&swap_lock);
340 if (offset < last_in_cluster)
341 discard_swap_cluster(si, offset,
342 last_in_cluster - offset + 1);
344 spin_lock(&swap_lock);
345 si->lowest_alloc = 0;
346 si->flags &= ~SWP_DISCARDING;
348 smp_mb(); /* wake_up_bit advises this */
349 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
351 } else if (si->flags & SWP_DISCARDING) {
353 * Delay using pages allocated by racing tasks
354 * until the whole discard has been issued. We
355 * could defer that delay until swap_writepage,
356 * but it's easier to keep this self-contained.
358 spin_unlock(&swap_lock);
359 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
360 wait_for_discard, TASK_UNINTERRUPTIBLE);
361 spin_lock(&swap_lock);
364 * Note pages allocated by racing tasks while
365 * scan for a free cluster is in progress, so
366 * that its final discard can exclude them.
368 if (offset < si->lowest_alloc)
369 si->lowest_alloc = offset;
370 if (offset > si->highest_alloc)
371 si->highest_alloc = offset;
377 spin_unlock(&swap_lock);
378 while (++offset <= si->highest_bit) {
379 if (!si->swap_map[offset]) {
380 spin_lock(&swap_lock);
383 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
384 spin_lock(&swap_lock);
387 if (unlikely(--latency_ration < 0)) {
389 latency_ration = LATENCY_LIMIT;
392 offset = si->lowest_bit;
393 while (++offset < scan_base) {
394 if (!si->swap_map[offset]) {
395 spin_lock(&swap_lock);
398 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
399 spin_lock(&swap_lock);
402 if (unlikely(--latency_ration < 0)) {
404 latency_ration = LATENCY_LIMIT;
407 spin_lock(&swap_lock);
410 si->flags -= SWP_SCANNING;
414 swp_entry_t get_swap_page(void)
416 struct swap_info_struct *si;
421 spin_lock(&swap_lock);
422 if (nr_swap_pages <= 0)
426 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
427 si = swap_info[type];
430 (!wrapped && si->prio != swap_info[next]->prio)) {
431 next = swap_list.head;
435 if (!si->highest_bit)
437 if (!(si->flags & SWP_WRITEOK))
440 swap_list.next = next;
441 /* This is called for allocating swap entry for cache */
442 offset = scan_swap_map(si, SWAP_HAS_CACHE);
444 spin_unlock(&swap_lock);
445 return swp_entry(type, offset);
447 next = swap_list.next;
452 spin_unlock(&swap_lock);
453 return (swp_entry_t) {0};
456 /* The only caller of this function is now susupend routine */
457 swp_entry_t get_swap_page_of_type(int type)
459 struct swap_info_struct *si;
462 spin_lock(&swap_lock);
463 si = swap_info[type];
464 if (si && (si->flags & SWP_WRITEOK)) {
466 /* This is called for allocating swap entry, not cache */
467 offset = scan_swap_map(si, 1);
469 spin_unlock(&swap_lock);
470 return swp_entry(type, offset);
474 spin_unlock(&swap_lock);
475 return (swp_entry_t) {0};
478 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
480 struct swap_info_struct *p;
481 unsigned long offset, type;
485 type = swp_type(entry);
486 if (type >= nr_swapfiles)
489 if (!(p->flags & SWP_USED))
491 offset = swp_offset(entry);
492 if (offset >= p->max)
494 if (!p->swap_map[offset])
496 spin_lock(&swap_lock);
500 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
503 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
506 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
509 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
514 static unsigned char swap_entry_free(struct swap_info_struct *p,
515 swp_entry_t entry, unsigned char usage)
517 unsigned long offset = swp_offset(entry);
519 unsigned char has_cache;
521 count = p->swap_map[offset];
522 has_cache = count & SWAP_HAS_CACHE;
523 count &= ~SWAP_HAS_CACHE;
525 if (usage == SWAP_HAS_CACHE) {
526 VM_BUG_ON(!has_cache);
528 } else if (count == SWAP_MAP_SHMEM) {
530 * Or we could insist on shmem.c using a special
531 * swap_shmem_free() and free_shmem_swap_and_cache()...
534 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
535 if (count == COUNT_CONTINUED) {
536 if (swap_count_continued(p, offset, count))
537 count = SWAP_MAP_MAX | COUNT_CONTINUED;
539 count = SWAP_MAP_MAX;
545 mem_cgroup_uncharge_swap(entry);
547 usage = count | has_cache;
548 p->swap_map[offset] = usage;
550 /* free if no reference */
552 struct gendisk *disk = p->bdev->bd_disk;
553 if (offset < p->lowest_bit)
554 p->lowest_bit = offset;
555 if (offset > p->highest_bit)
556 p->highest_bit = offset;
557 if (swap_list.next >= 0 &&
558 p->prio > swap_info[swap_list.next]->prio)
559 swap_list.next = p->type;
562 frontswap_invalidate_page(p->type, offset);
563 if ((p->flags & SWP_BLKDEV) &&
564 disk->fops->swap_slot_free_notify)
565 disk->fops->swap_slot_free_notify(p->bdev, offset);
572 * Caller has made sure that the swapdevice corresponding to entry
573 * is still around or has not been recycled.
575 void swap_free(swp_entry_t entry)
577 struct swap_info_struct *p;
579 p = swap_info_get(entry);
581 swap_entry_free(p, entry, 1);
582 spin_unlock(&swap_lock);
587 * Called after dropping swapcache to decrease refcnt to swap entries.
589 void swapcache_free(swp_entry_t entry, struct page *page)
591 struct swap_info_struct *p;
594 p = swap_info_get(entry);
596 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
598 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
599 spin_unlock(&swap_lock);
604 * How many references to page are currently swapped out?
605 * This does not give an exact answer when swap count is continued,
606 * but does include the high COUNT_CONTINUED flag to allow for that.
608 int page_swapcount(struct page *page)
611 struct swap_info_struct *p;
614 entry.val = page_private(page);
615 p = swap_info_get(entry);
617 count = swap_count(p->swap_map[swp_offset(entry)]);
618 spin_unlock(&swap_lock);
624 * We can write to an anon page without COW if there are no other references
625 * to it. And as a side-effect, free up its swap: because the old content
626 * on disk will never be read, and seeking back there to write new content
627 * later would only waste time away from clustering.
629 int reuse_swap_page(struct page *page)
633 VM_BUG_ON(!PageLocked(page));
634 if (unlikely(PageKsm(page)))
636 count = page_mapcount(page);
637 if (count <= 1 && PageSwapCache(page)) {
638 count += page_swapcount(page);
639 if (count == 1 && !PageWriteback(page)) {
640 delete_from_swap_cache(page);
648 * If swap is getting full, or if there are no more mappings of this page,
649 * then try_to_free_swap is called to free its swap space.
651 int try_to_free_swap(struct page *page)
653 VM_BUG_ON(!PageLocked(page));
655 if (!PageSwapCache(page))
657 if (PageWriteback(page))
659 if (page_swapcount(page))
663 * Once hibernation has begun to create its image of memory,
664 * there's a danger that one of the calls to try_to_free_swap()
665 * - most probably a call from __try_to_reclaim_swap() while
666 * hibernation is allocating its own swap pages for the image,
667 * but conceivably even a call from memory reclaim - will free
668 * the swap from a page which has already been recorded in the
669 * image as a clean swapcache page, and then reuse its swap for
670 * another page of the image. On waking from hibernation, the
671 * original page might be freed under memory pressure, then
672 * later read back in from swap, now with the wrong data.
674 * Hibration suspends storage while it is writing the image
675 * to disk so check that here.
677 if (pm_suspended_storage())
680 delete_from_swap_cache(page);
686 * Free the swap entry like above, but also try to
687 * free the page cache entry if it is the last user.
689 int free_swap_and_cache(swp_entry_t entry)
691 struct swap_info_struct *p;
692 struct page *page = NULL;
694 if (non_swap_entry(entry))
697 p = swap_info_get(entry);
699 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
700 page = find_get_page(&swapper_space, entry.val);
701 if (page && !trylock_page(page)) {
702 page_cache_release(page);
706 spin_unlock(&swap_lock);
710 * Not mapped elsewhere, or swap space full? Free it!
711 * Also recheck PageSwapCache now page is locked (above).
713 if (PageSwapCache(page) && !PageWriteback(page) &&
714 (!page_mapped(page) || vm_swap_full())) {
715 delete_from_swap_cache(page);
719 page_cache_release(page);
724 #ifdef CONFIG_HIBERNATION
726 * Find the swap type that corresponds to given device (if any).
728 * @offset - number of the PAGE_SIZE-sized block of the device, starting
729 * from 0, in which the swap header is expected to be located.
731 * This is needed for the suspend to disk (aka swsusp).
733 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
735 struct block_device *bdev = NULL;
739 bdev = bdget(device);
741 spin_lock(&swap_lock);
742 for (type = 0; type < nr_swapfiles; type++) {
743 struct swap_info_struct *sis = swap_info[type];
745 if (!(sis->flags & SWP_WRITEOK))
750 *bdev_p = bdgrab(sis->bdev);
752 spin_unlock(&swap_lock);
755 if (bdev == sis->bdev) {
756 struct swap_extent *se = &sis->first_swap_extent;
758 if (se->start_block == offset) {
760 *bdev_p = bdgrab(sis->bdev);
762 spin_unlock(&swap_lock);
768 spin_unlock(&swap_lock);
776 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
777 * corresponding to given index in swap_info (swap type).
779 sector_t swapdev_block(int type, pgoff_t offset)
781 struct block_device *bdev;
783 if ((unsigned int)type >= nr_swapfiles)
785 if (!(swap_info[type]->flags & SWP_WRITEOK))
787 return map_swap_entry(swp_entry(type, offset), &bdev);
791 * Return either the total number of swap pages of given type, or the number
792 * of free pages of that type (depending on @free)
794 * This is needed for software suspend
796 unsigned int count_swap_pages(int type, int free)
800 spin_lock(&swap_lock);
801 if ((unsigned int)type < nr_swapfiles) {
802 struct swap_info_struct *sis = swap_info[type];
804 if (sis->flags & SWP_WRITEOK) {
807 n -= sis->inuse_pages;
810 spin_unlock(&swap_lock);
813 #endif /* CONFIG_HIBERNATION */
816 * No need to decide whether this PTE shares the swap entry with others,
817 * just let do_wp_page work it out if a write is requested later - to
818 * force COW, vm_page_prot omits write permission from any private vma.
820 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
821 unsigned long addr, swp_entry_t entry, struct page *page)
823 struct mem_cgroup *memcg;
828 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
829 GFP_KERNEL, &memcg)) {
834 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
835 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
837 mem_cgroup_cancel_charge_swapin(memcg);
842 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
843 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
845 set_pte_at(vma->vm_mm, addr, pte,
846 pte_mkold(mk_pte(page, vma->vm_page_prot)));
847 page_add_anon_rmap(page, vma, addr);
848 mem_cgroup_commit_charge_swapin(page, memcg);
851 * Move the page to the active list so it is not
852 * immediately swapped out again after swapon.
856 pte_unmap_unlock(pte, ptl);
861 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
862 unsigned long addr, unsigned long end,
863 swp_entry_t entry, struct page *page)
865 pte_t swp_pte = swp_entry_to_pte(entry);
870 * We don't actually need pte lock while scanning for swp_pte: since
871 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
872 * page table while we're scanning; though it could get zapped, and on
873 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
874 * of unmatched parts which look like swp_pte, so unuse_pte must
875 * recheck under pte lock. Scanning without pte lock lets it be
876 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
878 pte = pte_offset_map(pmd, addr);
881 * swapoff spends a _lot_ of time in this loop!
882 * Test inline before going to call unuse_pte.
884 if (unlikely(pte_same(*pte, swp_pte))) {
886 ret = unuse_pte(vma, pmd, addr, entry, page);
889 pte = pte_offset_map(pmd, addr);
891 } while (pte++, addr += PAGE_SIZE, addr != end);
897 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
898 unsigned long addr, unsigned long end,
899 swp_entry_t entry, struct page *page)
905 pmd = pmd_offset(pud, addr);
907 next = pmd_addr_end(addr, end);
908 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
910 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
913 } while (pmd++, addr = next, addr != end);
917 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
918 unsigned long addr, unsigned long end,
919 swp_entry_t entry, struct page *page)
925 pud = pud_offset(pgd, addr);
927 next = pud_addr_end(addr, end);
928 if (pud_none_or_clear_bad(pud))
930 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
933 } while (pud++, addr = next, addr != end);
937 static int unuse_vma(struct vm_area_struct *vma,
938 swp_entry_t entry, struct page *page)
941 unsigned long addr, end, next;
944 if (page_anon_vma(page)) {
945 addr = page_address_in_vma(page, vma);
949 end = addr + PAGE_SIZE;
951 addr = vma->vm_start;
955 pgd = pgd_offset(vma->vm_mm, addr);
957 next = pgd_addr_end(addr, end);
958 if (pgd_none_or_clear_bad(pgd))
960 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
963 } while (pgd++, addr = next, addr != end);
967 static int unuse_mm(struct mm_struct *mm,
968 swp_entry_t entry, struct page *page)
970 struct vm_area_struct *vma;
973 if (!down_read_trylock(&mm->mmap_sem)) {
975 * Activate page so shrink_inactive_list is unlikely to unmap
976 * its ptes while lock is dropped, so swapoff can make progress.
980 down_read(&mm->mmap_sem);
983 for (vma = mm->mmap; vma; vma = vma->vm_next) {
984 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
987 up_read(&mm->mmap_sem);
988 return (ret < 0)? ret: 0;
992 * Scan swap_map (or frontswap_map if frontswap parameter is true)
993 * from current position to next entry still in use.
994 * Recycle to start on reaching the end, returning 0 when empty.
996 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
997 unsigned int prev, bool frontswap)
999 unsigned int max = si->max;
1000 unsigned int i = prev;
1001 unsigned char count;
1004 * No need for swap_lock here: we're just looking
1005 * for whether an entry is in use, not modifying it; false
1006 * hits are okay, and sys_swapoff() has already prevented new
1007 * allocations from this area (while holding swap_lock).
1016 * No entries in use at top of swap_map,
1017 * loop back to start and recheck there.
1024 if (frontswap_test(si, i))
1029 count = si->swap_map[i];
1030 if (count && swap_count(count) != SWAP_MAP_BAD)
1037 * We completely avoid races by reading each swap page in advance,
1038 * and then search for the process using it. All the necessary
1039 * page table adjustments can then be made atomically.
1041 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1042 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1044 int try_to_unuse(unsigned int type, bool frontswap,
1045 unsigned long pages_to_unuse)
1047 struct swap_info_struct *si = swap_info[type];
1048 struct mm_struct *start_mm;
1049 unsigned char *swap_map;
1050 unsigned char swcount;
1057 * When searching mms for an entry, a good strategy is to
1058 * start at the first mm we freed the previous entry from
1059 * (though actually we don't notice whether we or coincidence
1060 * freed the entry). Initialize this start_mm with a hold.
1062 * A simpler strategy would be to start at the last mm we
1063 * freed the previous entry from; but that would take less
1064 * advantage of mmlist ordering, which clusters forked mms
1065 * together, child after parent. If we race with dup_mmap(), we
1066 * prefer to resolve parent before child, lest we miss entries
1067 * duplicated after we scanned child: using last mm would invert
1070 start_mm = &init_mm;
1071 atomic_inc(&init_mm.mm_users);
1074 * Keep on scanning until all entries have gone. Usually,
1075 * one pass through swap_map is enough, but not necessarily:
1076 * there are races when an instance of an entry might be missed.
1078 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1079 if (signal_pending(current)) {
1085 * Get a page for the entry, using the existing swap
1086 * cache page if there is one. Otherwise, get a clean
1087 * page and read the swap into it.
1089 swap_map = &si->swap_map[i];
1090 entry = swp_entry(type, i);
1091 page = read_swap_cache_async(entry,
1092 GFP_HIGHUSER_MOVABLE, NULL, 0);
1095 * Either swap_duplicate() failed because entry
1096 * has been freed independently, and will not be
1097 * reused since sys_swapoff() already disabled
1098 * allocation from here, or alloc_page() failed.
1107 * Don't hold on to start_mm if it looks like exiting.
1109 if (atomic_read(&start_mm->mm_users) == 1) {
1111 start_mm = &init_mm;
1112 atomic_inc(&init_mm.mm_users);
1116 * Wait for and lock page. When do_swap_page races with
1117 * try_to_unuse, do_swap_page can handle the fault much
1118 * faster than try_to_unuse can locate the entry. This
1119 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1120 * defer to do_swap_page in such a case - in some tests,
1121 * do_swap_page and try_to_unuse repeatedly compete.
1123 wait_on_page_locked(page);
1124 wait_on_page_writeback(page);
1126 wait_on_page_writeback(page);
1129 * Remove all references to entry.
1131 swcount = *swap_map;
1132 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1133 retval = shmem_unuse(entry, page);
1134 /* page has already been unlocked and released */
1139 if (swap_count(swcount) && start_mm != &init_mm)
1140 retval = unuse_mm(start_mm, entry, page);
1142 if (swap_count(*swap_map)) {
1143 int set_start_mm = (*swap_map >= swcount);
1144 struct list_head *p = &start_mm->mmlist;
1145 struct mm_struct *new_start_mm = start_mm;
1146 struct mm_struct *prev_mm = start_mm;
1147 struct mm_struct *mm;
1149 atomic_inc(&new_start_mm->mm_users);
1150 atomic_inc(&prev_mm->mm_users);
1151 spin_lock(&mmlist_lock);
1152 while (swap_count(*swap_map) && !retval &&
1153 (p = p->next) != &start_mm->mmlist) {
1154 mm = list_entry(p, struct mm_struct, mmlist);
1155 if (!atomic_inc_not_zero(&mm->mm_users))
1157 spin_unlock(&mmlist_lock);
1163 swcount = *swap_map;
1164 if (!swap_count(swcount)) /* any usage ? */
1166 else if (mm == &init_mm)
1169 retval = unuse_mm(mm, entry, page);
1171 if (set_start_mm && *swap_map < swcount) {
1172 mmput(new_start_mm);
1173 atomic_inc(&mm->mm_users);
1177 spin_lock(&mmlist_lock);
1179 spin_unlock(&mmlist_lock);
1182 start_mm = new_start_mm;
1186 page_cache_release(page);
1191 * If a reference remains (rare), we would like to leave
1192 * the page in the swap cache; but try_to_unmap could
1193 * then re-duplicate the entry once we drop page lock,
1194 * so we might loop indefinitely; also, that page could
1195 * not be swapped out to other storage meanwhile. So:
1196 * delete from cache even if there's another reference,
1197 * after ensuring that the data has been saved to disk -
1198 * since if the reference remains (rarer), it will be
1199 * read from disk into another page. Splitting into two
1200 * pages would be incorrect if swap supported "shared
1201 * private" pages, but they are handled by tmpfs files.
1203 * Given how unuse_vma() targets one particular offset
1204 * in an anon_vma, once the anon_vma has been determined,
1205 * this splitting happens to be just what is needed to
1206 * handle where KSM pages have been swapped out: re-reading
1207 * is unnecessarily slow, but we can fix that later on.
1209 if (swap_count(*swap_map) &&
1210 PageDirty(page) && PageSwapCache(page)) {
1211 struct writeback_control wbc = {
1212 .sync_mode = WB_SYNC_NONE,
1215 swap_writepage(page, &wbc);
1217 wait_on_page_writeback(page);
1221 * It is conceivable that a racing task removed this page from
1222 * swap cache just before we acquired the page lock at the top,
1223 * or while we dropped it in unuse_mm(). The page might even
1224 * be back in swap cache on another swap area: that we must not
1225 * delete, since it may not have been written out to swap yet.
1227 if (PageSwapCache(page) &&
1228 likely(page_private(page) == entry.val))
1229 delete_from_swap_cache(page);
1232 * So we could skip searching mms once swap count went
1233 * to 1, we did not mark any present ptes as dirty: must
1234 * mark page dirty so shrink_page_list will preserve it.
1238 page_cache_release(page);
1241 * Make sure that we aren't completely killing
1242 * interactive performance.
1245 if (frontswap && pages_to_unuse > 0) {
1246 if (!--pages_to_unuse)
1256 * After a successful try_to_unuse, if no swap is now in use, we know
1257 * we can empty the mmlist. swap_lock must be held on entry and exit.
1258 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1259 * added to the mmlist just after page_duplicate - before would be racy.
1261 static void drain_mmlist(void)
1263 struct list_head *p, *next;
1266 for (type = 0; type < nr_swapfiles; type++)
1267 if (swap_info[type]->inuse_pages)
1269 spin_lock(&mmlist_lock);
1270 list_for_each_safe(p, next, &init_mm.mmlist)
1272 spin_unlock(&mmlist_lock);
1276 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1277 * corresponds to page offset for the specified swap entry.
1278 * Note that the type of this function is sector_t, but it returns page offset
1279 * into the bdev, not sector offset.
1281 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1283 struct swap_info_struct *sis;
1284 struct swap_extent *start_se;
1285 struct swap_extent *se;
1288 sis = swap_info[swp_type(entry)];
1291 offset = swp_offset(entry);
1292 start_se = sis->curr_swap_extent;
1296 struct list_head *lh;
1298 if (se->start_page <= offset &&
1299 offset < (se->start_page + se->nr_pages)) {
1300 return se->start_block + (offset - se->start_page);
1303 se = list_entry(lh, struct swap_extent, list);
1304 sis->curr_swap_extent = se;
1305 BUG_ON(se == start_se); /* It *must* be present */
1310 * Returns the page offset into bdev for the specified page's swap entry.
1312 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1315 entry.val = page_private(page);
1316 return map_swap_entry(entry, bdev);
1320 * Free all of a swapdev's extent information
1322 static void destroy_swap_extents(struct swap_info_struct *sis)
1324 while (!list_empty(&sis->first_swap_extent.list)) {
1325 struct swap_extent *se;
1327 se = list_entry(sis->first_swap_extent.list.next,
1328 struct swap_extent, list);
1329 list_del(&se->list);
1333 if (sis->flags & SWP_FILE) {
1334 struct file *swap_file = sis->swap_file;
1335 struct address_space *mapping = swap_file->f_mapping;
1337 sis->flags &= ~SWP_FILE;
1338 mapping->a_ops->swap_deactivate(swap_file);
1343 * Add a block range (and the corresponding page range) into this swapdev's
1344 * extent list. The extent list is kept sorted in page order.
1346 * This function rather assumes that it is called in ascending page order.
1349 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1350 unsigned long nr_pages, sector_t start_block)
1352 struct swap_extent *se;
1353 struct swap_extent *new_se;
1354 struct list_head *lh;
1356 if (start_page == 0) {
1357 se = &sis->first_swap_extent;
1358 sis->curr_swap_extent = se;
1360 se->nr_pages = nr_pages;
1361 se->start_block = start_block;
1364 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1365 se = list_entry(lh, struct swap_extent, list);
1366 BUG_ON(se->start_page + se->nr_pages != start_page);
1367 if (se->start_block + se->nr_pages == start_block) {
1369 se->nr_pages += nr_pages;
1375 * No merge. Insert a new extent, preserving ordering.
1377 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1380 new_se->start_page = start_page;
1381 new_se->nr_pages = nr_pages;
1382 new_se->start_block = start_block;
1384 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1389 * A `swap extent' is a simple thing which maps a contiguous range of pages
1390 * onto a contiguous range of disk blocks. An ordered list of swap extents
1391 * is built at swapon time and is then used at swap_writepage/swap_readpage
1392 * time for locating where on disk a page belongs.
1394 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1395 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1396 * swap files identically.
1398 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1399 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1400 * swapfiles are handled *identically* after swapon time.
1402 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1403 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1404 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1405 * requirements, they are simply tossed out - we will never use those blocks
1408 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1409 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1410 * which will scribble on the fs.
1412 * The amount of disk space which a single swap extent represents varies.
1413 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1414 * extents in the list. To avoid much list walking, we cache the previous
1415 * search location in `curr_swap_extent', and start new searches from there.
1416 * This is extremely effective. The average number of iterations in
1417 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1419 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1421 struct file *swap_file = sis->swap_file;
1422 struct address_space *mapping = swap_file->f_mapping;
1423 struct inode *inode = mapping->host;
1426 if (S_ISBLK(inode->i_mode)) {
1427 ret = add_swap_extent(sis, 0, sis->max, 0);
1432 if (mapping->a_ops->swap_activate) {
1433 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1435 sis->flags |= SWP_FILE;
1436 ret = add_swap_extent(sis, 0, sis->max, 0);
1442 return generic_swapfile_activate(sis, swap_file, span);
1445 static void enable_swap_info(struct swap_info_struct *p, int prio,
1446 unsigned char *swap_map,
1447 unsigned long *frontswap_map)
1451 spin_lock(&swap_lock);
1455 p->prio = --least_priority;
1456 p->swap_map = swap_map;
1457 frontswap_map_set(p, frontswap_map);
1458 p->flags |= SWP_WRITEOK;
1459 nr_swap_pages += p->pages;
1460 total_swap_pages += p->pages;
1462 /* insert swap space into swap_list: */
1464 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1465 if (p->prio >= swap_info[i]->prio)
1471 swap_list.head = swap_list.next = p->type;
1473 swap_info[prev]->next = p->type;
1474 frontswap_init(p->type);
1475 spin_unlock(&swap_lock);
1478 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1480 struct swap_info_struct *p = NULL;
1481 unsigned char *swap_map;
1482 struct file *swap_file, *victim;
1483 struct address_space *mapping;
1484 struct inode *inode;
1490 if (!capable(CAP_SYS_ADMIN))
1493 BUG_ON(!current->mm);
1495 pathname = getname(specialfile);
1496 err = PTR_ERR(pathname);
1497 if (IS_ERR(pathname))
1500 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1502 err = PTR_ERR(victim);
1506 mapping = victim->f_mapping;
1508 spin_lock(&swap_lock);
1509 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1510 p = swap_info[type];
1511 if (p->flags & SWP_WRITEOK) {
1512 if (p->swap_file->f_mapping == mapping)
1519 spin_unlock(&swap_lock);
1522 if (!security_vm_enough_memory_mm(current->mm, p->pages))
1523 vm_unacct_memory(p->pages);
1526 spin_unlock(&swap_lock);
1530 swap_list.head = p->next;
1532 swap_info[prev]->next = p->next;
1533 if (type == swap_list.next) {
1534 /* just pick something that's safe... */
1535 swap_list.next = swap_list.head;
1538 for (i = p->next; i >= 0; i = swap_info[i]->next)
1539 swap_info[i]->prio = p->prio--;
1542 nr_swap_pages -= p->pages;
1543 total_swap_pages -= p->pages;
1544 p->flags &= ~SWP_WRITEOK;
1545 spin_unlock(&swap_lock);
1547 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1548 err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1549 compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj);
1553 * reading p->prio and p->swap_map outside the lock is
1554 * safe here because only sys_swapon and sys_swapoff
1555 * change them, and there can be no other sys_swapon or
1556 * sys_swapoff for this swap_info_struct at this point.
1558 /* re-insert swap space back into swap_list */
1559 enable_swap_info(p, p->prio, p->swap_map, frontswap_map_get(p));
1563 destroy_swap_extents(p);
1564 if (p->flags & SWP_CONTINUED)
1565 free_swap_count_continuations(p);
1567 mutex_lock(&swapon_mutex);
1568 spin_lock(&swap_lock);
1571 /* wait for anyone still in scan_swap_map */
1572 p->highest_bit = 0; /* cuts scans short */
1573 while (p->flags >= SWP_SCANNING) {
1574 spin_unlock(&swap_lock);
1575 schedule_timeout_uninterruptible(1);
1576 spin_lock(&swap_lock);
1579 swap_file = p->swap_file;
1580 p->swap_file = NULL;
1582 swap_map = p->swap_map;
1585 frontswap_invalidate_area(type);
1586 spin_unlock(&swap_lock);
1587 mutex_unlock(&swapon_mutex);
1589 vfree(frontswap_map_get(p));
1590 /* Destroy swap account informatin */
1591 swap_cgroup_swapoff(type);
1593 inode = mapping->host;
1594 if (S_ISBLK(inode->i_mode)) {
1595 struct block_device *bdev = I_BDEV(inode);
1596 set_blocksize(bdev, p->old_block_size);
1597 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1599 mutex_lock(&inode->i_mutex);
1600 inode->i_flags &= ~S_SWAPFILE;
1601 mutex_unlock(&inode->i_mutex);
1603 filp_close(swap_file, NULL);
1605 atomic_inc(&proc_poll_event);
1606 wake_up_interruptible(&proc_poll_wait);
1609 filp_close(victim, NULL);
1614 #ifdef CONFIG_PROC_FS
1615 static unsigned swaps_poll(struct file *file, poll_table *wait)
1617 struct seq_file *seq = file->private_data;
1619 poll_wait(file, &proc_poll_wait, wait);
1621 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1622 seq->poll_event = atomic_read(&proc_poll_event);
1623 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1626 return POLLIN | POLLRDNORM;
1630 static void *swap_start(struct seq_file *swap, loff_t *pos)
1632 struct swap_info_struct *si;
1636 mutex_lock(&swapon_mutex);
1639 return SEQ_START_TOKEN;
1641 for (type = 0; type < nr_swapfiles; type++) {
1642 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1643 si = swap_info[type];
1644 if (!(si->flags & SWP_USED) || !si->swap_map)
1653 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1655 struct swap_info_struct *si = v;
1658 if (v == SEQ_START_TOKEN)
1661 type = si->type + 1;
1663 for (; type < nr_swapfiles; type++) {
1664 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1665 si = swap_info[type];
1666 if (!(si->flags & SWP_USED) || !si->swap_map)
1675 static void swap_stop(struct seq_file *swap, void *v)
1677 mutex_unlock(&swapon_mutex);
1680 static int swap_show(struct seq_file *swap, void *v)
1682 struct swap_info_struct *si = v;
1686 if (si == SEQ_START_TOKEN) {
1687 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1691 file = si->swap_file;
1692 len = seq_path(swap, &file->f_path, " \t\n\\");
1693 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1694 len < 40 ? 40 - len : 1, " ",
1695 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1696 "partition" : "file\t",
1697 si->pages << (PAGE_SHIFT - 10),
1698 si->inuse_pages << (PAGE_SHIFT - 10),
1703 static const struct seq_operations swaps_op = {
1704 .start = swap_start,
1710 static int swaps_open(struct inode *inode, struct file *file)
1712 struct seq_file *seq;
1715 ret = seq_open(file, &swaps_op);
1719 seq = file->private_data;
1720 seq->poll_event = atomic_read(&proc_poll_event);
1724 static const struct file_operations proc_swaps_operations = {
1727 .llseek = seq_lseek,
1728 .release = seq_release,
1732 static int __init procswaps_init(void)
1734 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1737 __initcall(procswaps_init);
1738 #endif /* CONFIG_PROC_FS */
1740 #ifdef MAX_SWAPFILES_CHECK
1741 static int __init max_swapfiles_check(void)
1743 MAX_SWAPFILES_CHECK();
1746 late_initcall(max_swapfiles_check);
1749 static struct swap_info_struct *alloc_swap_info(void)
1751 struct swap_info_struct *p;
1754 p = kzalloc(sizeof(*p), GFP_KERNEL);
1756 return ERR_PTR(-ENOMEM);
1758 spin_lock(&swap_lock);
1759 for (type = 0; type < nr_swapfiles; type++) {
1760 if (!(swap_info[type]->flags & SWP_USED))
1763 if (type >= MAX_SWAPFILES) {
1764 spin_unlock(&swap_lock);
1766 return ERR_PTR(-EPERM);
1768 if (type >= nr_swapfiles) {
1770 swap_info[type] = p;
1772 * Write swap_info[type] before nr_swapfiles, in case a
1773 * racing procfs swap_start() or swap_next() is reading them.
1774 * (We never shrink nr_swapfiles, we never free this entry.)
1780 p = swap_info[type];
1782 * Do not memset this entry: a racing procfs swap_next()
1783 * would be relying on p->type to remain valid.
1786 INIT_LIST_HEAD(&p->first_swap_extent.list);
1787 p->flags = SWP_USED;
1789 spin_unlock(&swap_lock);
1794 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1798 if (S_ISBLK(inode->i_mode)) {
1799 p->bdev = bdgrab(I_BDEV(inode));
1800 error = blkdev_get(p->bdev,
1801 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1807 p->old_block_size = block_size(p->bdev);
1808 error = set_blocksize(p->bdev, PAGE_SIZE);
1811 p->flags |= SWP_BLKDEV;
1812 } else if (S_ISREG(inode->i_mode)) {
1813 p->bdev = inode->i_sb->s_bdev;
1814 mutex_lock(&inode->i_mutex);
1815 if (IS_SWAPFILE(inode))
1823 static unsigned long read_swap_header(struct swap_info_struct *p,
1824 union swap_header *swap_header,
1825 struct inode *inode)
1828 unsigned long maxpages;
1829 unsigned long swapfilepages;
1831 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1832 printk(KERN_ERR "Unable to find swap-space signature\n");
1836 /* swap partition endianess hack... */
1837 if (swab32(swap_header->info.version) == 1) {
1838 swab32s(&swap_header->info.version);
1839 swab32s(&swap_header->info.last_page);
1840 swab32s(&swap_header->info.nr_badpages);
1841 for (i = 0; i < swap_header->info.nr_badpages; i++)
1842 swab32s(&swap_header->info.badpages[i]);
1844 /* Check the swap header's sub-version */
1845 if (swap_header->info.version != 1) {
1847 "Unable to handle swap header version %d\n",
1848 swap_header->info.version);
1853 p->cluster_next = 1;
1857 * Find out how many pages are allowed for a single swap
1858 * device. There are two limiting factors: 1) the number
1859 * of bits for the swap offset in the swp_entry_t type, and
1860 * 2) the number of bits in the swap pte as defined by the
1861 * different architectures. In order to find the
1862 * largest possible bit mask, a swap entry with swap type 0
1863 * and swap offset ~0UL is created, encoded to a swap pte,
1864 * decoded to a swp_entry_t again, and finally the swap
1865 * offset is extracted. This will mask all the bits from
1866 * the initial ~0UL mask that can't be encoded in either
1867 * the swp_entry_t or the architecture definition of a
1870 maxpages = swp_offset(pte_to_swp_entry(
1871 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1872 if (maxpages > swap_header->info.last_page) {
1873 maxpages = swap_header->info.last_page + 1;
1874 /* p->max is an unsigned int: don't overflow it */
1875 if ((unsigned int)maxpages == 0)
1876 maxpages = UINT_MAX;
1878 p->highest_bit = maxpages - 1;
1882 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1883 if (swapfilepages && maxpages > swapfilepages) {
1885 "Swap area shorter than signature indicates\n");
1888 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1890 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1896 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1897 union swap_header *swap_header,
1898 unsigned char *swap_map,
1899 unsigned long maxpages,
1903 unsigned int nr_good_pages;
1906 nr_good_pages = maxpages - 1; /* omit header page */
1908 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1909 unsigned int page_nr = swap_header->info.badpages[i];
1910 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1912 if (page_nr < maxpages) {
1913 swap_map[page_nr] = SWAP_MAP_BAD;
1918 if (nr_good_pages) {
1919 swap_map[0] = SWAP_MAP_BAD;
1921 p->pages = nr_good_pages;
1922 nr_extents = setup_swap_extents(p, span);
1925 nr_good_pages = p->pages;
1927 if (!nr_good_pages) {
1928 printk(KERN_WARNING "Empty swap-file\n");
1935 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1937 struct swap_info_struct *p;
1939 struct file *swap_file = NULL;
1940 struct address_space *mapping;
1944 union swap_header *swap_header;
1947 unsigned long maxpages;
1948 unsigned char *swap_map = NULL;
1949 unsigned long *frontswap_map = NULL;
1950 struct page *page = NULL;
1951 struct inode *inode = NULL;
1953 if (swap_flags & ~SWAP_FLAGS_VALID)
1956 if (!capable(CAP_SYS_ADMIN))
1959 p = alloc_swap_info();
1963 name = getname(specialfile);
1965 error = PTR_ERR(name);
1969 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1970 if (IS_ERR(swap_file)) {
1971 error = PTR_ERR(swap_file);
1976 p->swap_file = swap_file;
1977 mapping = swap_file->f_mapping;
1979 for (i = 0; i < nr_swapfiles; i++) {
1980 struct swap_info_struct *q = swap_info[i];
1982 if (q == p || !q->swap_file)
1984 if (mapping == q->swap_file->f_mapping) {
1990 inode = mapping->host;
1991 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
1992 error = claim_swapfile(p, inode);
1993 if (unlikely(error))
1997 * Read the swap header.
1999 if (!mapping->a_ops->readpage) {
2003 page = read_mapping_page(mapping, 0, swap_file);
2005 error = PTR_ERR(page);
2008 swap_header = kmap(page);
2010 maxpages = read_swap_header(p, swap_header, inode);
2011 if (unlikely(!maxpages)) {
2016 /* OK, set up the swap map and apply the bad block list */
2017 swap_map = vzalloc(maxpages);
2023 error = swap_cgroup_swapon(p->type, maxpages);
2027 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2029 if (unlikely(nr_extents < 0)) {
2033 /* frontswap enabled? set up bit-per-page map for frontswap */
2034 if (frontswap_enabled)
2035 frontswap_map = vzalloc(maxpages / sizeof(long));
2038 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2039 p->flags |= SWP_SOLIDSTATE;
2040 p->cluster_next = 1 + (random32() % p->highest_bit);
2042 if ((swap_flags & SWAP_FLAG_DISCARD) && discard_swap(p) == 0)
2043 p->flags |= SWP_DISCARDABLE;
2046 mutex_lock(&swapon_mutex);
2048 if (swap_flags & SWAP_FLAG_PREFER)
2050 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2051 enable_swap_info(p, prio, swap_map, frontswap_map);
2053 printk(KERN_INFO "Adding %uk swap on %s. "
2054 "Priority:%d extents:%d across:%lluk %s%s%s\n",
2055 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2056 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2057 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2058 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2059 (frontswap_map) ? "FS" : "");
2061 mutex_unlock(&swapon_mutex);
2062 atomic_inc(&proc_poll_event);
2063 wake_up_interruptible(&proc_poll_wait);
2065 if (S_ISREG(inode->i_mode))
2066 inode->i_flags |= S_SWAPFILE;
2070 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2071 set_blocksize(p->bdev, p->old_block_size);
2072 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2074 destroy_swap_extents(p);
2075 swap_cgroup_swapoff(p->type);
2076 spin_lock(&swap_lock);
2077 p->swap_file = NULL;
2079 spin_unlock(&swap_lock);
2082 if (inode && S_ISREG(inode->i_mode)) {
2083 mutex_unlock(&inode->i_mutex);
2086 filp_close(swap_file, NULL);
2089 if (page && !IS_ERR(page)) {
2091 page_cache_release(page);
2095 if (inode && S_ISREG(inode->i_mode))
2096 mutex_unlock(&inode->i_mutex);
2100 void si_swapinfo(struct sysinfo *val)
2103 unsigned long nr_to_be_unused = 0;
2105 spin_lock(&swap_lock);
2106 for (type = 0; type < nr_swapfiles; type++) {
2107 struct swap_info_struct *si = swap_info[type];
2109 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2110 nr_to_be_unused += si->inuse_pages;
2112 val->freeswap = nr_swap_pages + nr_to_be_unused;
2113 val->totalswap = total_swap_pages + nr_to_be_unused;
2114 spin_unlock(&swap_lock);
2118 * Verify that a swap entry is valid and increment its swap map count.
2120 * Returns error code in following case.
2122 * - swp_entry is invalid -> EINVAL
2123 * - swp_entry is migration entry -> EINVAL
2124 * - swap-cache reference is requested but there is already one. -> EEXIST
2125 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2126 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2128 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2130 struct swap_info_struct *p;
2131 unsigned long offset, type;
2132 unsigned char count;
2133 unsigned char has_cache;
2136 if (non_swap_entry(entry))
2139 type = swp_type(entry);
2140 if (type >= nr_swapfiles)
2142 p = swap_info[type];
2143 offset = swp_offset(entry);
2145 spin_lock(&swap_lock);
2146 if (unlikely(offset >= p->max))
2149 count = p->swap_map[offset];
2150 has_cache = count & SWAP_HAS_CACHE;
2151 count &= ~SWAP_HAS_CACHE;
2154 if (usage == SWAP_HAS_CACHE) {
2156 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2157 if (!has_cache && count)
2158 has_cache = SWAP_HAS_CACHE;
2159 else if (has_cache) /* someone else added cache */
2161 else /* no users remaining */
2164 } else if (count || has_cache) {
2166 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2168 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2170 else if (swap_count_continued(p, offset, count))
2171 count = COUNT_CONTINUED;
2175 err = -ENOENT; /* unused swap entry */
2177 p->swap_map[offset] = count | has_cache;
2180 spin_unlock(&swap_lock);
2185 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2190 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2191 * (in which case its reference count is never incremented).
2193 void swap_shmem_alloc(swp_entry_t entry)
2195 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2199 * Increase reference count of swap entry by 1.
2200 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2201 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2202 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2203 * might occur if a page table entry has got corrupted.
2205 int swap_duplicate(swp_entry_t entry)
2209 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2210 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2215 * @entry: swap entry for which we allocate swap cache.
2217 * Called when allocating swap cache for existing swap entry,
2218 * This can return error codes. Returns 0 at success.
2219 * -EBUSY means there is a swap cache.
2220 * Note: return code is different from swap_duplicate().
2222 int swapcache_prepare(swp_entry_t entry)
2224 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2227 struct swap_info_struct *page_swap_info(struct page *page)
2229 swp_entry_t swap = { .val = page_private(page) };
2230 BUG_ON(!PageSwapCache(page));
2231 return swap_info[swp_type(swap)];
2235 * out-of-line __page_file_ methods to avoid include hell.
2237 struct address_space *__page_file_mapping(struct page *page)
2239 VM_BUG_ON(!PageSwapCache(page));
2240 return page_swap_info(page)->swap_file->f_mapping;
2242 EXPORT_SYMBOL_GPL(__page_file_mapping);
2244 pgoff_t __page_file_index(struct page *page)
2246 swp_entry_t swap = { .val = page_private(page) };
2247 VM_BUG_ON(!PageSwapCache(page));
2248 return swp_offset(swap);
2250 EXPORT_SYMBOL_GPL(__page_file_index);
2253 * add_swap_count_continuation - called when a swap count is duplicated
2254 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2255 * page of the original vmalloc'ed swap_map, to hold the continuation count
2256 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2257 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2259 * These continuation pages are seldom referenced: the common paths all work
2260 * on the original swap_map, only referring to a continuation page when the
2261 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2263 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2264 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2265 * can be called after dropping locks.
2267 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2269 struct swap_info_struct *si;
2272 struct page *list_page;
2274 unsigned char count;
2277 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2278 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2280 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2282 si = swap_info_get(entry);
2285 * An acceptable race has occurred since the failing
2286 * __swap_duplicate(): the swap entry has been freed,
2287 * perhaps even the whole swap_map cleared for swapoff.
2292 offset = swp_offset(entry);
2293 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2295 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2297 * The higher the swap count, the more likely it is that tasks
2298 * will race to add swap count continuation: we need to avoid
2299 * over-provisioning.
2305 spin_unlock(&swap_lock);
2310 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2311 * no architecture is using highmem pages for kernel pagetables: so it
2312 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2314 head = vmalloc_to_page(si->swap_map + offset);
2315 offset &= ~PAGE_MASK;
2318 * Page allocation does not initialize the page's lru field,
2319 * but it does always reset its private field.
2321 if (!page_private(head)) {
2322 BUG_ON(count & COUNT_CONTINUED);
2323 INIT_LIST_HEAD(&head->lru);
2324 set_page_private(head, SWP_CONTINUED);
2325 si->flags |= SWP_CONTINUED;
2328 list_for_each_entry(list_page, &head->lru, lru) {
2332 * If the previous map said no continuation, but we've found
2333 * a continuation page, free our allocation and use this one.
2335 if (!(count & COUNT_CONTINUED))
2338 map = kmap_atomic(list_page) + offset;
2343 * If this continuation count now has some space in it,
2344 * free our allocation and use this one.
2346 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2350 list_add_tail(&page->lru, &head->lru);
2351 page = NULL; /* now it's attached, don't free it */
2353 spin_unlock(&swap_lock);
2361 * swap_count_continued - when the original swap_map count is incremented
2362 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2363 * into, carry if so, or else fail until a new continuation page is allocated;
2364 * when the original swap_map count is decremented from 0 with continuation,
2365 * borrow from the continuation and report whether it still holds more.
2366 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2368 static bool swap_count_continued(struct swap_info_struct *si,
2369 pgoff_t offset, unsigned char count)
2375 head = vmalloc_to_page(si->swap_map + offset);
2376 if (page_private(head) != SWP_CONTINUED) {
2377 BUG_ON(count & COUNT_CONTINUED);
2378 return false; /* need to add count continuation */
2381 offset &= ~PAGE_MASK;
2382 page = list_entry(head->lru.next, struct page, lru);
2383 map = kmap_atomic(page) + offset;
2385 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2386 goto init_map; /* jump over SWAP_CONT_MAX checks */
2388 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2390 * Think of how you add 1 to 999
2392 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2394 page = list_entry(page->lru.next, struct page, lru);
2395 BUG_ON(page == head);
2396 map = kmap_atomic(page) + offset;
2398 if (*map == SWAP_CONT_MAX) {
2400 page = list_entry(page->lru.next, struct page, lru);
2402 return false; /* add count continuation */
2403 map = kmap_atomic(page) + offset;
2404 init_map: *map = 0; /* we didn't zero the page */
2408 page = list_entry(page->lru.prev, struct page, lru);
2409 while (page != head) {
2410 map = kmap_atomic(page) + offset;
2411 *map = COUNT_CONTINUED;
2413 page = list_entry(page->lru.prev, struct page, lru);
2415 return true; /* incremented */
2417 } else { /* decrementing */
2419 * Think of how you subtract 1 from 1000
2421 BUG_ON(count != COUNT_CONTINUED);
2422 while (*map == COUNT_CONTINUED) {
2424 page = list_entry(page->lru.next, struct page, lru);
2425 BUG_ON(page == head);
2426 map = kmap_atomic(page) + offset;
2433 page = list_entry(page->lru.prev, struct page, lru);
2434 while (page != head) {
2435 map = kmap_atomic(page) + offset;
2436 *map = SWAP_CONT_MAX | count;
2437 count = COUNT_CONTINUED;
2439 page = list_entry(page->lru.prev, struct page, lru);
2441 return count == COUNT_CONTINUED;
2446 * free_swap_count_continuations - swapoff free all the continuation pages
2447 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2449 static void free_swap_count_continuations(struct swap_info_struct *si)
2453 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2455 head = vmalloc_to_page(si->swap_map + offset);
2456 if (page_private(head)) {
2457 struct list_head *this, *next;
2458 list_for_each_safe(this, next, &head->lru) {
2460 page = list_entry(this, struct page, lru);