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/module.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/security.h>
28 #include <linux/backing-dev.h>
29 #include <linux/mutex.h>
30 #include <linux/capability.h>
31 #include <linux/syscalls.h>
32 #include <linux/memcontrol.h>
33 #include <linux/poll.h>
34 #include <linux/oom.h>
36 #include <asm/pgtable.h>
37 #include <asm/tlbflush.h>
38 #include <linux/swapops.h>
39 #include <linux/page_cgroup.h>
41 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
43 static void free_swap_count_continuations(struct swap_info_struct *);
44 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
46 static DEFINE_SPINLOCK(swap_lock);
47 static unsigned int nr_swapfiles;
49 long total_swap_pages;
50 static int least_priority;
52 static const char Bad_file[] = "Bad swap file entry ";
53 static const char Unused_file[] = "Unused swap file entry ";
54 static const char Bad_offset[] = "Bad swap offset entry ";
55 static const char Unused_offset[] = "Unused swap offset entry ";
57 static struct swap_list_t swap_list = {-1, -1};
59 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
61 static DEFINE_MUTEX(swapon_mutex);
63 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
64 /* Activity counter to indicate that a swapon or swapoff has occurred */
65 static atomic_t proc_poll_event = ATOMIC_INIT(0);
67 static inline unsigned char swap_count(unsigned char ent)
69 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
72 /* returns 1 if swap entry is freed */
74 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
76 swp_entry_t entry = swp_entry(si->type, offset);
80 page = find_get_page(&swapper_space, entry.val);
84 * This function is called from scan_swap_map() and it's called
85 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
86 * We have to use trylock for avoiding deadlock. This is a special
87 * case and you should use try_to_free_swap() with explicit lock_page()
88 * in usual operations.
90 if (trylock_page(page)) {
91 ret = try_to_free_swap(page);
94 page_cache_release(page);
99 * swapon tell device that all the old swap contents can be discarded,
100 * to allow the swap device to optimize its wear-levelling.
102 static int discard_swap(struct swap_info_struct *si)
104 struct swap_extent *se;
105 sector_t start_block;
109 /* Do not discard the swap header page! */
110 se = &si->first_swap_extent;
111 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
112 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
114 err = blkdev_issue_discard(si->bdev, start_block,
115 nr_blocks, GFP_KERNEL, 0);
121 list_for_each_entry(se, &si->first_swap_extent.list, list) {
122 start_block = se->start_block << (PAGE_SHIFT - 9);
123 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
125 err = blkdev_issue_discard(si->bdev, start_block,
126 nr_blocks, GFP_KERNEL, 0);
132 return err; /* That will often be -EOPNOTSUPP */
136 * swap allocation tell device that a cluster of swap can now be discarded,
137 * to allow the swap device to optimize its wear-levelling.
139 static void discard_swap_cluster(struct swap_info_struct *si,
140 pgoff_t start_page, pgoff_t nr_pages)
142 struct swap_extent *se = si->curr_swap_extent;
143 int found_extent = 0;
146 struct list_head *lh;
148 if (se->start_page <= start_page &&
149 start_page < se->start_page + se->nr_pages) {
150 pgoff_t offset = start_page - se->start_page;
151 sector_t start_block = se->start_block + offset;
152 sector_t nr_blocks = se->nr_pages - offset;
154 if (nr_blocks > nr_pages)
155 nr_blocks = nr_pages;
156 start_page += nr_blocks;
157 nr_pages -= nr_blocks;
160 si->curr_swap_extent = se;
162 start_block <<= PAGE_SHIFT - 9;
163 nr_blocks <<= PAGE_SHIFT - 9;
164 if (blkdev_issue_discard(si->bdev, start_block,
165 nr_blocks, GFP_NOIO, 0))
170 se = list_entry(lh, struct swap_extent, list);
174 static int wait_for_discard(void *word)
180 #define SWAPFILE_CLUSTER 256
181 #define LATENCY_LIMIT 256
183 static unsigned long scan_swap_map(struct swap_info_struct *si,
186 unsigned long offset;
187 unsigned long scan_base;
188 unsigned long last_in_cluster = 0;
189 int latency_ration = LATENCY_LIMIT;
190 int found_free_cluster = 0;
193 * We try to cluster swap pages by allocating them sequentially
194 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
195 * way, however, we resort to first-free allocation, starting
196 * a new cluster. This prevents us from scattering swap pages
197 * all over the entire swap partition, so that we reduce
198 * overall disk seek times between swap pages. -- sct
199 * But we do now try to find an empty cluster. -Andrea
200 * And we let swap pages go all over an SSD partition. Hugh
203 si->flags += SWP_SCANNING;
204 scan_base = offset = si->cluster_next;
206 if (unlikely(!si->cluster_nr--)) {
207 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
208 si->cluster_nr = SWAPFILE_CLUSTER - 1;
211 if (si->flags & SWP_DISCARDABLE) {
213 * Start range check on racing allocations, in case
214 * they overlap the cluster we eventually decide on
215 * (we scan without swap_lock to allow preemption).
216 * It's hardly conceivable that cluster_nr could be
217 * wrapped during our scan, but don't depend on it.
219 if (si->lowest_alloc)
221 si->lowest_alloc = si->max;
222 si->highest_alloc = 0;
224 spin_unlock(&swap_lock);
227 * If seek is expensive, start searching for new cluster from
228 * start of partition, to minimize the span of allocated swap.
229 * But if seek is cheap, search from our current position, so
230 * that swap is allocated from all over the partition: if the
231 * Flash Translation Layer only remaps within limited zones,
232 * we don't want to wear out the first zone too quickly.
234 if (!(si->flags & SWP_SOLIDSTATE))
235 scan_base = offset = si->lowest_bit;
236 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
238 /* Locate the first empty (unaligned) cluster */
239 for (; last_in_cluster <= si->highest_bit; offset++) {
240 if (si->swap_map[offset])
241 last_in_cluster = offset + SWAPFILE_CLUSTER;
242 else if (offset == last_in_cluster) {
243 spin_lock(&swap_lock);
244 offset -= SWAPFILE_CLUSTER - 1;
245 si->cluster_next = offset;
246 si->cluster_nr = SWAPFILE_CLUSTER - 1;
247 found_free_cluster = 1;
250 if (unlikely(--latency_ration < 0)) {
252 latency_ration = LATENCY_LIMIT;
256 offset = si->lowest_bit;
257 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
259 /* Locate the first empty (unaligned) cluster */
260 for (; last_in_cluster < scan_base; offset++) {
261 if (si->swap_map[offset])
262 last_in_cluster = offset + SWAPFILE_CLUSTER;
263 else if (offset == last_in_cluster) {
264 spin_lock(&swap_lock);
265 offset -= SWAPFILE_CLUSTER - 1;
266 si->cluster_next = offset;
267 si->cluster_nr = SWAPFILE_CLUSTER - 1;
268 found_free_cluster = 1;
271 if (unlikely(--latency_ration < 0)) {
273 latency_ration = LATENCY_LIMIT;
278 spin_lock(&swap_lock);
279 si->cluster_nr = SWAPFILE_CLUSTER - 1;
280 si->lowest_alloc = 0;
284 if (!(si->flags & SWP_WRITEOK))
286 if (!si->highest_bit)
288 if (offset > si->highest_bit)
289 scan_base = offset = si->lowest_bit;
291 /* reuse swap entry of cache-only swap if not busy. */
292 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
294 spin_unlock(&swap_lock);
295 swap_was_freed = __try_to_reclaim_swap(si, offset);
296 spin_lock(&swap_lock);
297 /* entry was freed successfully, try to use this again */
300 goto scan; /* check next one */
303 if (si->swap_map[offset])
306 if (offset == si->lowest_bit)
308 if (offset == si->highest_bit)
311 if (si->inuse_pages == si->pages) {
312 si->lowest_bit = si->max;
315 si->swap_map[offset] = usage;
316 si->cluster_next = offset + 1;
317 si->flags -= SWP_SCANNING;
319 if (si->lowest_alloc) {
321 * Only set when SWP_DISCARDABLE, and there's a scan
322 * for a free cluster in progress or just completed.
324 if (found_free_cluster) {
326 * To optimize wear-levelling, discard the
327 * old data of the cluster, taking care not to
328 * discard any of its pages that have already
329 * been allocated by racing tasks (offset has
330 * already stepped over any at the beginning).
332 if (offset < si->highest_alloc &&
333 si->lowest_alloc <= last_in_cluster)
334 last_in_cluster = si->lowest_alloc - 1;
335 si->flags |= SWP_DISCARDING;
336 spin_unlock(&swap_lock);
338 if (offset < last_in_cluster)
339 discard_swap_cluster(si, offset,
340 last_in_cluster - offset + 1);
342 spin_lock(&swap_lock);
343 si->lowest_alloc = 0;
344 si->flags &= ~SWP_DISCARDING;
346 smp_mb(); /* wake_up_bit advises this */
347 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
349 } else if (si->flags & SWP_DISCARDING) {
351 * Delay using pages allocated by racing tasks
352 * until the whole discard has been issued. We
353 * could defer that delay until swap_writepage,
354 * but it's easier to keep this self-contained.
356 spin_unlock(&swap_lock);
357 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
358 wait_for_discard, TASK_UNINTERRUPTIBLE);
359 spin_lock(&swap_lock);
362 * Note pages allocated by racing tasks while
363 * scan for a free cluster is in progress, so
364 * that its final discard can exclude them.
366 if (offset < si->lowest_alloc)
367 si->lowest_alloc = offset;
368 if (offset > si->highest_alloc)
369 si->highest_alloc = offset;
375 spin_unlock(&swap_lock);
376 while (++offset <= si->highest_bit) {
377 if (!si->swap_map[offset]) {
378 spin_lock(&swap_lock);
381 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
382 spin_lock(&swap_lock);
385 if (unlikely(--latency_ration < 0)) {
387 latency_ration = LATENCY_LIMIT;
390 offset = si->lowest_bit;
391 while (++offset < scan_base) {
392 if (!si->swap_map[offset]) {
393 spin_lock(&swap_lock);
396 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
397 spin_lock(&swap_lock);
400 if (unlikely(--latency_ration < 0)) {
402 latency_ration = LATENCY_LIMIT;
405 spin_lock(&swap_lock);
408 si->flags -= SWP_SCANNING;
412 swp_entry_t get_swap_page(void)
414 struct swap_info_struct *si;
419 spin_lock(&swap_lock);
420 if (nr_swap_pages <= 0)
424 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
425 si = swap_info[type];
428 (!wrapped && si->prio != swap_info[next]->prio)) {
429 next = swap_list.head;
433 if (!si->highest_bit)
435 if (!(si->flags & SWP_WRITEOK))
438 swap_list.next = next;
439 /* This is called for allocating swap entry for cache */
440 offset = scan_swap_map(si, SWAP_HAS_CACHE);
442 spin_unlock(&swap_lock);
443 return swp_entry(type, offset);
445 next = swap_list.next;
450 spin_unlock(&swap_lock);
451 return (swp_entry_t) {0};
454 /* The only caller of this function is now susupend routine */
455 swp_entry_t get_swap_page_of_type(int type)
457 struct swap_info_struct *si;
460 spin_lock(&swap_lock);
461 si = swap_info[type];
462 if (si && (si->flags & SWP_WRITEOK)) {
464 /* This is called for allocating swap entry, not cache */
465 offset = scan_swap_map(si, 1);
467 spin_unlock(&swap_lock);
468 return swp_entry(type, offset);
472 spin_unlock(&swap_lock);
473 return (swp_entry_t) {0};
476 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
478 struct swap_info_struct *p;
479 unsigned long offset, type;
483 type = swp_type(entry);
484 if (type >= nr_swapfiles)
487 if (!(p->flags & SWP_USED))
489 offset = swp_offset(entry);
490 if (offset >= p->max)
492 if (!p->swap_map[offset])
494 spin_lock(&swap_lock);
498 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
501 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
504 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
507 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
512 static unsigned char swap_entry_free(struct swap_info_struct *p,
513 swp_entry_t entry, unsigned char usage)
515 unsigned long offset = swp_offset(entry);
517 unsigned char has_cache;
519 count = p->swap_map[offset];
520 has_cache = count & SWAP_HAS_CACHE;
521 count &= ~SWAP_HAS_CACHE;
523 if (usage == SWAP_HAS_CACHE) {
524 VM_BUG_ON(!has_cache);
526 } else if (count == SWAP_MAP_SHMEM) {
528 * Or we could insist on shmem.c using a special
529 * swap_shmem_free() and free_shmem_swap_and_cache()...
532 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
533 if (count == COUNT_CONTINUED) {
534 if (swap_count_continued(p, offset, count))
535 count = SWAP_MAP_MAX | COUNT_CONTINUED;
537 count = SWAP_MAP_MAX;
543 mem_cgroup_uncharge_swap(entry);
545 usage = count | has_cache;
546 p->swap_map[offset] = usage;
548 /* free if no reference */
550 struct gendisk *disk = p->bdev->bd_disk;
551 if (offset < p->lowest_bit)
552 p->lowest_bit = offset;
553 if (offset > p->highest_bit)
554 p->highest_bit = offset;
555 if (swap_list.next >= 0 &&
556 p->prio > swap_info[swap_list.next]->prio)
557 swap_list.next = p->type;
560 if ((p->flags & SWP_BLKDEV) &&
561 disk->fops->swap_slot_free_notify)
562 disk->fops->swap_slot_free_notify(p->bdev, offset);
569 * Caller has made sure that the swapdevice corresponding to entry
570 * is still around or has not been recycled.
572 void swap_free(swp_entry_t entry)
574 struct swap_info_struct *p;
576 p = swap_info_get(entry);
578 swap_entry_free(p, entry, 1);
579 spin_unlock(&swap_lock);
584 * Called after dropping swapcache to decrease refcnt to swap entries.
586 void swapcache_free(swp_entry_t entry, struct page *page)
588 struct swap_info_struct *p;
591 p = swap_info_get(entry);
593 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
595 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
596 spin_unlock(&swap_lock);
601 * How many references to page are currently swapped out?
602 * This does not give an exact answer when swap count is continued,
603 * but does include the high COUNT_CONTINUED flag to allow for that.
605 static inline int page_swapcount(struct page *page)
608 struct swap_info_struct *p;
611 entry.val = page_private(page);
612 p = swap_info_get(entry);
614 count = swap_count(p->swap_map[swp_offset(entry)]);
615 spin_unlock(&swap_lock);
621 * We can write to an anon page without COW if there are no other references
622 * to it. And as a side-effect, free up its swap: because the old content
623 * on disk will never be read, and seeking back there to write new content
624 * later would only waste time away from clustering.
626 int reuse_swap_page(struct page *page)
630 VM_BUG_ON(!PageLocked(page));
631 if (unlikely(PageKsm(page)))
633 count = page_mapcount(page);
634 if (count <= 1 && PageSwapCache(page)) {
635 count += page_swapcount(page);
636 if (count == 1 && !PageWriteback(page)) {
637 delete_from_swap_cache(page);
645 * If swap is getting full, or if there are no more mappings of this page,
646 * then try_to_free_swap is called to free its swap space.
648 int try_to_free_swap(struct page *page)
650 VM_BUG_ON(!PageLocked(page));
652 if (!PageSwapCache(page))
654 if (PageWriteback(page))
656 if (page_swapcount(page))
660 * Once hibernation has begun to create its image of memory,
661 * there's a danger that one of the calls to try_to_free_swap()
662 * - most probably a call from __try_to_reclaim_swap() while
663 * hibernation is allocating its own swap pages for the image,
664 * but conceivably even a call from memory reclaim - will free
665 * the swap from a page which has already been recorded in the
666 * image as a clean swapcache page, and then reuse its swap for
667 * another page of the image. On waking from hibernation, the
668 * original page might be freed under memory pressure, then
669 * later read back in from swap, now with the wrong data.
671 * Hibernation clears bits from gfp_allowed_mask to prevent
672 * memory reclaim from writing to disk, so check that here.
674 if (!(gfp_allowed_mask & __GFP_IO))
677 delete_from_swap_cache(page);
683 * Free the swap entry like above, but also try to
684 * free the page cache entry if it is the last user.
686 int free_swap_and_cache(swp_entry_t entry)
688 struct swap_info_struct *p;
689 struct page *page = NULL;
691 if (non_swap_entry(entry))
694 p = swap_info_get(entry);
696 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
697 page = find_get_page(&swapper_space, entry.val);
698 if (page && !trylock_page(page)) {
699 page_cache_release(page);
703 spin_unlock(&swap_lock);
707 * Not mapped elsewhere, or swap space full? Free it!
708 * Also recheck PageSwapCache now page is locked (above).
710 if (PageSwapCache(page) && !PageWriteback(page) &&
711 (!page_mapped(page) || vm_swap_full())) {
712 delete_from_swap_cache(page);
716 page_cache_release(page);
721 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
723 * mem_cgroup_count_swap_user - count the user of a swap entry
724 * @ent: the swap entry to be checked
725 * @pagep: the pointer for the swap cache page of the entry to be stored
727 * Returns the number of the user of the swap entry. The number is valid only
728 * for swaps of anonymous pages.
729 * If the entry is found on swap cache, the page is stored to pagep with
730 * refcount of it being incremented.
732 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
735 struct swap_info_struct *p;
738 page = find_get_page(&swapper_space, ent.val);
740 count += page_mapcount(page);
741 p = swap_info_get(ent);
743 count += swap_count(p->swap_map[swp_offset(ent)]);
744 spin_unlock(&swap_lock);
752 #ifdef CONFIG_HIBERNATION
754 * Find the swap type that corresponds to given device (if any).
756 * @offset - number of the PAGE_SIZE-sized block of the device, starting
757 * from 0, in which the swap header is expected to be located.
759 * This is needed for the suspend to disk (aka swsusp).
761 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
763 struct block_device *bdev = NULL;
767 bdev = bdget(device);
769 spin_lock(&swap_lock);
770 for (type = 0; type < nr_swapfiles; type++) {
771 struct swap_info_struct *sis = swap_info[type];
773 if (!(sis->flags & SWP_WRITEOK))
778 *bdev_p = bdgrab(sis->bdev);
780 spin_unlock(&swap_lock);
783 if (bdev == sis->bdev) {
784 struct swap_extent *se = &sis->first_swap_extent;
786 if (se->start_block == offset) {
788 *bdev_p = bdgrab(sis->bdev);
790 spin_unlock(&swap_lock);
796 spin_unlock(&swap_lock);
804 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
805 * corresponding to given index in swap_info (swap type).
807 sector_t swapdev_block(int type, pgoff_t offset)
809 struct block_device *bdev;
811 if ((unsigned int)type >= nr_swapfiles)
813 if (!(swap_info[type]->flags & SWP_WRITEOK))
815 return map_swap_entry(swp_entry(type, offset), &bdev);
819 * Return either the total number of swap pages of given type, or the number
820 * of free pages of that type (depending on @free)
822 * This is needed for software suspend
824 unsigned int count_swap_pages(int type, int free)
828 spin_lock(&swap_lock);
829 if ((unsigned int)type < nr_swapfiles) {
830 struct swap_info_struct *sis = swap_info[type];
832 if (sis->flags & SWP_WRITEOK) {
835 n -= sis->inuse_pages;
838 spin_unlock(&swap_lock);
841 #endif /* CONFIG_HIBERNATION */
844 * No need to decide whether this PTE shares the swap entry with others,
845 * just let do_wp_page work it out if a write is requested later - to
846 * force COW, vm_page_prot omits write permission from any private vma.
848 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
849 unsigned long addr, swp_entry_t entry, struct page *page)
851 struct mem_cgroup *ptr;
856 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
861 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
862 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
864 mem_cgroup_cancel_charge_swapin(ptr);
869 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
870 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
872 set_pte_at(vma->vm_mm, addr, pte,
873 pte_mkold(mk_pte(page, vma->vm_page_prot)));
874 page_add_anon_rmap(page, vma, addr);
875 mem_cgroup_commit_charge_swapin(page, ptr);
878 * Move the page to the active list so it is not
879 * immediately swapped out again after swapon.
883 pte_unmap_unlock(pte, ptl);
888 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
889 unsigned long addr, unsigned long end,
890 swp_entry_t entry, struct page *page)
892 pte_t swp_pte = swp_entry_to_pte(entry);
897 * We don't actually need pte lock while scanning for swp_pte: since
898 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
899 * page table while we're scanning; though it could get zapped, and on
900 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
901 * of unmatched parts which look like swp_pte, so unuse_pte must
902 * recheck under pte lock. Scanning without pte lock lets it be
903 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
905 pte = pte_offset_map(pmd, addr);
908 * swapoff spends a _lot_ of time in this loop!
909 * Test inline before going to call unuse_pte.
911 if (unlikely(pte_same(*pte, swp_pte))) {
913 ret = unuse_pte(vma, pmd, addr, entry, page);
916 pte = pte_offset_map(pmd, addr);
918 } while (pte++, addr += PAGE_SIZE, addr != end);
924 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
925 unsigned long addr, unsigned long end,
926 swp_entry_t entry, struct page *page)
932 pmd = pmd_offset(pud, addr);
934 next = pmd_addr_end(addr, end);
935 if (unlikely(pmd_trans_huge(*pmd)))
937 if (pmd_none_or_clear_bad(pmd))
939 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
942 } while (pmd++, addr = next, addr != end);
946 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
947 unsigned long addr, unsigned long end,
948 swp_entry_t entry, struct page *page)
954 pud = pud_offset(pgd, addr);
956 next = pud_addr_end(addr, end);
957 if (pud_none_or_clear_bad(pud))
959 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
962 } while (pud++, addr = next, addr != end);
966 static int unuse_vma(struct vm_area_struct *vma,
967 swp_entry_t entry, struct page *page)
970 unsigned long addr, end, next;
973 if (page_anon_vma(page)) {
974 addr = page_address_in_vma(page, vma);
978 end = addr + PAGE_SIZE;
980 addr = vma->vm_start;
984 pgd = pgd_offset(vma->vm_mm, addr);
986 next = pgd_addr_end(addr, end);
987 if (pgd_none_or_clear_bad(pgd))
989 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
992 } while (pgd++, addr = next, addr != end);
996 static int unuse_mm(struct mm_struct *mm,
997 swp_entry_t entry, struct page *page)
999 struct vm_area_struct *vma;
1002 if (!down_read_trylock(&mm->mmap_sem)) {
1004 * Activate page so shrink_inactive_list is unlikely to unmap
1005 * its ptes while lock is dropped, so swapoff can make progress.
1007 activate_page(page);
1009 down_read(&mm->mmap_sem);
1012 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1013 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1016 up_read(&mm->mmap_sem);
1017 return (ret < 0)? ret: 0;
1021 * Scan swap_map from current position to next entry still in use.
1022 * Recycle to start on reaching the end, returning 0 when empty.
1024 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1027 unsigned int max = si->max;
1028 unsigned int i = prev;
1029 unsigned char count;
1032 * No need for swap_lock here: we're just looking
1033 * for whether an entry is in use, not modifying it; false
1034 * hits are okay, and sys_swapoff() has already prevented new
1035 * allocations from this area (while holding swap_lock).
1044 * No entries in use at top of swap_map,
1045 * loop back to start and recheck there.
1051 count = si->swap_map[i];
1052 if (count && swap_count(count) != SWAP_MAP_BAD)
1059 * We completely avoid races by reading each swap page in advance,
1060 * and then search for the process using it. All the necessary
1061 * page table adjustments can then be made atomically.
1063 static int try_to_unuse(unsigned int type)
1065 struct swap_info_struct *si = swap_info[type];
1066 struct mm_struct *start_mm;
1067 unsigned char *swap_map;
1068 unsigned char swcount;
1075 * When searching mms for an entry, a good strategy is to
1076 * start at the first mm we freed the previous entry from
1077 * (though actually we don't notice whether we or coincidence
1078 * freed the entry). Initialize this start_mm with a hold.
1080 * A simpler strategy would be to start at the last mm we
1081 * freed the previous entry from; but that would take less
1082 * advantage of mmlist ordering, which clusters forked mms
1083 * together, child after parent. If we race with dup_mmap(), we
1084 * prefer to resolve parent before child, lest we miss entries
1085 * duplicated after we scanned child: using last mm would invert
1088 start_mm = &init_mm;
1089 atomic_inc(&init_mm.mm_users);
1092 * Keep on scanning until all entries have gone. Usually,
1093 * one pass through swap_map is enough, but not necessarily:
1094 * there are races when an instance of an entry might be missed.
1096 while ((i = find_next_to_unuse(si, i)) != 0) {
1097 if (signal_pending(current)) {
1103 * Get a page for the entry, using the existing swap
1104 * cache page if there is one. Otherwise, get a clean
1105 * page and read the swap into it.
1107 swap_map = &si->swap_map[i];
1108 entry = swp_entry(type, i);
1109 page = read_swap_cache_async(entry,
1110 GFP_HIGHUSER_MOVABLE, NULL, 0);
1113 * Either swap_duplicate() failed because entry
1114 * has been freed independently, and will not be
1115 * reused since sys_swapoff() already disabled
1116 * allocation from here, or alloc_page() failed.
1125 * Don't hold on to start_mm if it looks like exiting.
1127 if (atomic_read(&start_mm->mm_users) == 1) {
1129 start_mm = &init_mm;
1130 atomic_inc(&init_mm.mm_users);
1134 * Wait for and lock page. When do_swap_page races with
1135 * try_to_unuse, do_swap_page can handle the fault much
1136 * faster than try_to_unuse can locate the entry. This
1137 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1138 * defer to do_swap_page in such a case - in some tests,
1139 * do_swap_page and try_to_unuse repeatedly compete.
1141 wait_on_page_locked(page);
1142 wait_on_page_writeback(page);
1144 wait_on_page_writeback(page);
1147 * Remove all references to entry.
1149 swcount = *swap_map;
1150 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1151 retval = shmem_unuse(entry, page);
1152 /* page has already been unlocked and released */
1157 if (swap_count(swcount) && start_mm != &init_mm)
1158 retval = unuse_mm(start_mm, entry, page);
1160 if (swap_count(*swap_map)) {
1161 int set_start_mm = (*swap_map >= swcount);
1162 struct list_head *p = &start_mm->mmlist;
1163 struct mm_struct *new_start_mm = start_mm;
1164 struct mm_struct *prev_mm = start_mm;
1165 struct mm_struct *mm;
1167 atomic_inc(&new_start_mm->mm_users);
1168 atomic_inc(&prev_mm->mm_users);
1169 spin_lock(&mmlist_lock);
1170 while (swap_count(*swap_map) && !retval &&
1171 (p = p->next) != &start_mm->mmlist) {
1172 mm = list_entry(p, struct mm_struct, mmlist);
1173 if (!atomic_inc_not_zero(&mm->mm_users))
1175 spin_unlock(&mmlist_lock);
1181 swcount = *swap_map;
1182 if (!swap_count(swcount)) /* any usage ? */
1184 else if (mm == &init_mm)
1187 retval = unuse_mm(mm, entry, page);
1189 if (set_start_mm && *swap_map < swcount) {
1190 mmput(new_start_mm);
1191 atomic_inc(&mm->mm_users);
1195 spin_lock(&mmlist_lock);
1197 spin_unlock(&mmlist_lock);
1200 start_mm = new_start_mm;
1204 page_cache_release(page);
1209 * If a reference remains (rare), we would like to leave
1210 * the page in the swap cache; but try_to_unmap could
1211 * then re-duplicate the entry once we drop page lock,
1212 * so we might loop indefinitely; also, that page could
1213 * not be swapped out to other storage meanwhile. So:
1214 * delete from cache even if there's another reference,
1215 * after ensuring that the data has been saved to disk -
1216 * since if the reference remains (rarer), it will be
1217 * read from disk into another page. Splitting into two
1218 * pages would be incorrect if swap supported "shared
1219 * private" pages, but they are handled by tmpfs files.
1221 * Given how unuse_vma() targets one particular offset
1222 * in an anon_vma, once the anon_vma has been determined,
1223 * this splitting happens to be just what is needed to
1224 * handle where KSM pages have been swapped out: re-reading
1225 * is unnecessarily slow, but we can fix that later on.
1227 if (swap_count(*swap_map) &&
1228 PageDirty(page) && PageSwapCache(page)) {
1229 struct writeback_control wbc = {
1230 .sync_mode = WB_SYNC_NONE,
1233 swap_writepage(page, &wbc);
1235 wait_on_page_writeback(page);
1239 * It is conceivable that a racing task removed this page from
1240 * swap cache just before we acquired the page lock at the top,
1241 * or while we dropped it in unuse_mm(). The page might even
1242 * be back in swap cache on another swap area: that we must not
1243 * delete, since it may not have been written out to swap yet.
1245 if (PageSwapCache(page) &&
1246 likely(page_private(page) == entry.val))
1247 delete_from_swap_cache(page);
1250 * So we could skip searching mms once swap count went
1251 * to 1, we did not mark any present ptes as dirty: must
1252 * mark page dirty so shrink_page_list will preserve it.
1256 page_cache_release(page);
1259 * Make sure that we aren't completely killing
1260 * interactive performance.
1270 * After a successful try_to_unuse, if no swap is now in use, we know
1271 * we can empty the mmlist. swap_lock must be held on entry and exit.
1272 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1273 * added to the mmlist just after page_duplicate - before would be racy.
1275 static void drain_mmlist(void)
1277 struct list_head *p, *next;
1280 for (type = 0; type < nr_swapfiles; type++)
1281 if (swap_info[type]->inuse_pages)
1283 spin_lock(&mmlist_lock);
1284 list_for_each_safe(p, next, &init_mm.mmlist)
1286 spin_unlock(&mmlist_lock);
1290 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1291 * corresponds to page offset for the specified swap entry.
1292 * Note that the type of this function is sector_t, but it returns page offset
1293 * into the bdev, not sector offset.
1295 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1297 struct swap_info_struct *sis;
1298 struct swap_extent *start_se;
1299 struct swap_extent *se;
1302 sis = swap_info[swp_type(entry)];
1305 offset = swp_offset(entry);
1306 start_se = sis->curr_swap_extent;
1310 struct list_head *lh;
1312 if (se->start_page <= offset &&
1313 offset < (se->start_page + se->nr_pages)) {
1314 return se->start_block + (offset - se->start_page);
1317 se = list_entry(lh, struct swap_extent, list);
1318 sis->curr_swap_extent = se;
1319 BUG_ON(se == start_se); /* It *must* be present */
1324 * Returns the page offset into bdev for the specified page's swap entry.
1326 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1329 entry.val = page_private(page);
1330 return map_swap_entry(entry, bdev);
1334 * Free all of a swapdev's extent information
1336 static void destroy_swap_extents(struct swap_info_struct *sis)
1338 while (!list_empty(&sis->first_swap_extent.list)) {
1339 struct swap_extent *se;
1341 se = list_entry(sis->first_swap_extent.list.next,
1342 struct swap_extent, list);
1343 list_del(&se->list);
1349 * Add a block range (and the corresponding page range) into this swapdev's
1350 * extent list. The extent list is kept sorted in page order.
1352 * This function rather assumes that it is called in ascending page order.
1355 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1356 unsigned long nr_pages, sector_t start_block)
1358 struct swap_extent *se;
1359 struct swap_extent *new_se;
1360 struct list_head *lh;
1362 if (start_page == 0) {
1363 se = &sis->first_swap_extent;
1364 sis->curr_swap_extent = se;
1366 se->nr_pages = nr_pages;
1367 se->start_block = start_block;
1370 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1371 se = list_entry(lh, struct swap_extent, list);
1372 BUG_ON(se->start_page + se->nr_pages != start_page);
1373 if (se->start_block + se->nr_pages == start_block) {
1375 se->nr_pages += nr_pages;
1381 * No merge. Insert a new extent, preserving ordering.
1383 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1386 new_se->start_page = start_page;
1387 new_se->nr_pages = nr_pages;
1388 new_se->start_block = start_block;
1390 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1395 * A `swap extent' is a simple thing which maps a contiguous range of pages
1396 * onto a contiguous range of disk blocks. An ordered list of swap extents
1397 * is built at swapon time and is then used at swap_writepage/swap_readpage
1398 * time for locating where on disk a page belongs.
1400 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1401 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1402 * swap files identically.
1404 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1405 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1406 * swapfiles are handled *identically* after swapon time.
1408 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1409 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1410 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1411 * requirements, they are simply tossed out - we will never use those blocks
1414 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1415 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1416 * which will scribble on the fs.
1418 * The amount of disk space which a single swap extent represents varies.
1419 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1420 * extents in the list. To avoid much list walking, we cache the previous
1421 * search location in `curr_swap_extent', and start new searches from there.
1422 * This is extremely effective. The average number of iterations in
1423 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1425 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1427 struct inode *inode;
1428 unsigned blocks_per_page;
1429 unsigned long page_no;
1431 sector_t probe_block;
1432 sector_t last_block;
1433 sector_t lowest_block = -1;
1434 sector_t highest_block = 0;
1438 inode = sis->swap_file->f_mapping->host;
1439 if (S_ISBLK(inode->i_mode)) {
1440 ret = add_swap_extent(sis, 0, sis->max, 0);
1445 blkbits = inode->i_blkbits;
1446 blocks_per_page = PAGE_SIZE >> blkbits;
1449 * Map all the blocks into the extent list. This code doesn't try
1454 last_block = i_size_read(inode) >> blkbits;
1455 while ((probe_block + blocks_per_page) <= last_block &&
1456 page_no < sis->max) {
1457 unsigned block_in_page;
1458 sector_t first_block;
1460 first_block = bmap(inode, probe_block);
1461 if (first_block == 0)
1465 * It must be PAGE_SIZE aligned on-disk
1467 if (first_block & (blocks_per_page - 1)) {
1472 for (block_in_page = 1; block_in_page < blocks_per_page;
1476 block = bmap(inode, probe_block + block_in_page);
1479 if (block != first_block + block_in_page) {
1486 first_block >>= (PAGE_SHIFT - blkbits);
1487 if (page_no) { /* exclude the header page */
1488 if (first_block < lowest_block)
1489 lowest_block = first_block;
1490 if (first_block > highest_block)
1491 highest_block = first_block;
1495 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1497 ret = add_swap_extent(sis, page_no, 1, first_block);
1502 probe_block += blocks_per_page;
1507 *span = 1 + highest_block - lowest_block;
1509 page_no = 1; /* force Empty message */
1511 sis->pages = page_no - 1;
1512 sis->highest_bit = page_no - 1;
1516 printk(KERN_ERR "swapon: swapfile has holes\n");
1521 static void enable_swap_info(struct swap_info_struct *p, int prio,
1522 unsigned char *swap_map)
1526 spin_lock(&swap_lock);
1530 p->prio = --least_priority;
1531 p->swap_map = swap_map;
1532 p->flags |= SWP_WRITEOK;
1533 nr_swap_pages += p->pages;
1534 total_swap_pages += p->pages;
1536 /* insert swap space into swap_list: */
1538 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1539 if (p->prio >= swap_info[i]->prio)
1545 swap_list.head = swap_list.next = p->type;
1547 swap_info[prev]->next = p->type;
1548 spin_unlock(&swap_lock);
1551 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1553 struct swap_info_struct *p = NULL;
1554 unsigned char *swap_map;
1555 struct file *swap_file, *victim;
1556 struct address_space *mapping;
1557 struct inode *inode;
1563 if (!capable(CAP_SYS_ADMIN))
1566 pathname = getname(specialfile);
1567 err = PTR_ERR(pathname);
1568 if (IS_ERR(pathname))
1571 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1573 err = PTR_ERR(victim);
1577 mapping = victim->f_mapping;
1579 spin_lock(&swap_lock);
1580 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1581 p = swap_info[type];
1582 if (p->flags & SWP_WRITEOK) {
1583 if (p->swap_file->f_mapping == mapping)
1590 spin_unlock(&swap_lock);
1593 if (!security_vm_enough_memory(p->pages))
1594 vm_unacct_memory(p->pages);
1597 spin_unlock(&swap_lock);
1601 swap_list.head = p->next;
1603 swap_info[prev]->next = p->next;
1604 if (type == swap_list.next) {
1605 /* just pick something that's safe... */
1606 swap_list.next = swap_list.head;
1609 for (i = p->next; i >= 0; i = swap_info[i]->next)
1610 swap_info[i]->prio = p->prio--;
1613 nr_swap_pages -= p->pages;
1614 total_swap_pages -= p->pages;
1615 p->flags &= ~SWP_WRITEOK;
1616 spin_unlock(&swap_lock);
1618 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1619 err = try_to_unuse(type);
1620 test_set_oom_score_adj(oom_score_adj);
1624 * reading p->prio and p->swap_map outside the lock is
1625 * safe here because only sys_swapon and sys_swapoff
1626 * change them, and there can be no other sys_swapon or
1627 * sys_swapoff for this swap_info_struct at this point.
1629 /* re-insert swap space back into swap_list */
1630 enable_swap_info(p, p->prio, p->swap_map);
1634 destroy_swap_extents(p);
1635 if (p->flags & SWP_CONTINUED)
1636 free_swap_count_continuations(p);
1638 mutex_lock(&swapon_mutex);
1639 spin_lock(&swap_lock);
1642 /* wait for anyone still in scan_swap_map */
1643 p->highest_bit = 0; /* cuts scans short */
1644 while (p->flags >= SWP_SCANNING) {
1645 spin_unlock(&swap_lock);
1646 schedule_timeout_uninterruptible(1);
1647 spin_lock(&swap_lock);
1650 swap_file = p->swap_file;
1651 p->swap_file = NULL;
1653 swap_map = p->swap_map;
1656 spin_unlock(&swap_lock);
1657 mutex_unlock(&swapon_mutex);
1659 /* Destroy swap account informatin */
1660 swap_cgroup_swapoff(type);
1662 inode = mapping->host;
1663 if (S_ISBLK(inode->i_mode)) {
1664 struct block_device *bdev = I_BDEV(inode);
1665 set_blocksize(bdev, p->old_block_size);
1666 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1668 mutex_lock(&inode->i_mutex);
1669 inode->i_flags &= ~S_SWAPFILE;
1670 mutex_unlock(&inode->i_mutex);
1672 filp_close(swap_file, NULL);
1674 atomic_inc(&proc_poll_event);
1675 wake_up_interruptible(&proc_poll_wait);
1678 filp_close(victim, NULL);
1683 #ifdef CONFIG_PROC_FS
1684 static unsigned swaps_poll(struct file *file, poll_table *wait)
1686 struct seq_file *seq = file->private_data;
1688 poll_wait(file, &proc_poll_wait, wait);
1690 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1691 seq->poll_event = atomic_read(&proc_poll_event);
1692 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1695 return POLLIN | POLLRDNORM;
1699 static void *swap_start(struct seq_file *swap, loff_t *pos)
1701 struct swap_info_struct *si;
1705 mutex_lock(&swapon_mutex);
1708 return SEQ_START_TOKEN;
1710 for (type = 0; type < nr_swapfiles; type++) {
1711 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1712 si = swap_info[type];
1713 if (!(si->flags & SWP_USED) || !si->swap_map)
1722 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1724 struct swap_info_struct *si = v;
1727 if (v == SEQ_START_TOKEN)
1730 type = si->type + 1;
1732 for (; type < nr_swapfiles; type++) {
1733 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1734 si = swap_info[type];
1735 if (!(si->flags & SWP_USED) || !si->swap_map)
1744 static void swap_stop(struct seq_file *swap, void *v)
1746 mutex_unlock(&swapon_mutex);
1749 static int swap_show(struct seq_file *swap, void *v)
1751 struct swap_info_struct *si = v;
1755 if (si == SEQ_START_TOKEN) {
1756 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1760 file = si->swap_file;
1761 len = seq_path(swap, &file->f_path, " \t\n\\");
1762 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1763 len < 40 ? 40 - len : 1, " ",
1764 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1765 "partition" : "file\t",
1766 si->pages << (PAGE_SHIFT - 10),
1767 si->inuse_pages << (PAGE_SHIFT - 10),
1772 static const struct seq_operations swaps_op = {
1773 .start = swap_start,
1779 static int swaps_open(struct inode *inode, struct file *file)
1781 struct seq_file *seq;
1784 ret = seq_open(file, &swaps_op);
1788 seq = file->private_data;
1789 seq->poll_event = atomic_read(&proc_poll_event);
1793 static const struct file_operations proc_swaps_operations = {
1796 .llseek = seq_lseek,
1797 .release = seq_release,
1801 static int __init procswaps_init(void)
1803 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1806 __initcall(procswaps_init);
1807 #endif /* CONFIG_PROC_FS */
1809 #ifdef MAX_SWAPFILES_CHECK
1810 static int __init max_swapfiles_check(void)
1812 MAX_SWAPFILES_CHECK();
1815 late_initcall(max_swapfiles_check);
1818 static struct swap_info_struct *alloc_swap_info(void)
1820 struct swap_info_struct *p;
1823 p = kzalloc(sizeof(*p), GFP_KERNEL);
1825 return ERR_PTR(-ENOMEM);
1827 spin_lock(&swap_lock);
1828 for (type = 0; type < nr_swapfiles; type++) {
1829 if (!(swap_info[type]->flags & SWP_USED))
1832 if (type >= MAX_SWAPFILES) {
1833 spin_unlock(&swap_lock);
1835 return ERR_PTR(-EPERM);
1837 if (type >= nr_swapfiles) {
1839 swap_info[type] = p;
1841 * Write swap_info[type] before nr_swapfiles, in case a
1842 * racing procfs swap_start() or swap_next() is reading them.
1843 * (We never shrink nr_swapfiles, we never free this entry.)
1849 p = swap_info[type];
1851 * Do not memset this entry: a racing procfs swap_next()
1852 * would be relying on p->type to remain valid.
1855 INIT_LIST_HEAD(&p->first_swap_extent.list);
1856 p->flags = SWP_USED;
1858 spin_unlock(&swap_lock);
1863 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1867 if (S_ISBLK(inode->i_mode)) {
1868 p->bdev = bdgrab(I_BDEV(inode));
1869 error = blkdev_get(p->bdev,
1870 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1876 p->old_block_size = block_size(p->bdev);
1877 error = set_blocksize(p->bdev, PAGE_SIZE);
1880 p->flags |= SWP_BLKDEV;
1881 } else if (S_ISREG(inode->i_mode)) {
1882 p->bdev = inode->i_sb->s_bdev;
1883 mutex_lock(&inode->i_mutex);
1884 if (IS_SWAPFILE(inode))
1892 static unsigned long read_swap_header(struct swap_info_struct *p,
1893 union swap_header *swap_header,
1894 struct inode *inode)
1897 unsigned long maxpages;
1898 unsigned long swapfilepages;
1900 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1901 printk(KERN_ERR "Unable to find swap-space signature\n");
1905 /* swap partition endianess hack... */
1906 if (swab32(swap_header->info.version) == 1) {
1907 swab32s(&swap_header->info.version);
1908 swab32s(&swap_header->info.last_page);
1909 swab32s(&swap_header->info.nr_badpages);
1910 for (i = 0; i < swap_header->info.nr_badpages; i++)
1911 swab32s(&swap_header->info.badpages[i]);
1913 /* Check the swap header's sub-version */
1914 if (swap_header->info.version != 1) {
1916 "Unable to handle swap header version %d\n",
1917 swap_header->info.version);
1922 p->cluster_next = 1;
1926 * Find out how many pages are allowed for a single swap
1927 * device. There are three limiting factors: 1) the number
1928 * of bits for the swap offset in the swp_entry_t type, and
1929 * 2) the number of bits in the swap pte as defined by the
1930 * the different architectures, and 3) the number of free bits
1931 * in an exceptional radix_tree entry. In order to find the
1932 * largest possible bit mask, a swap entry with swap type 0
1933 * and swap offset ~0UL is created, encoded to a swap pte,
1934 * decoded to a swp_entry_t again, and finally the swap
1935 * offset is extracted. This will mask all the bits from
1936 * the initial ~0UL mask that can't be encoded in either
1937 * the swp_entry_t or the architecture definition of a
1938 * swap pte. Then the same is done for a radix_tree entry.
1940 maxpages = swp_offset(pte_to_swp_entry(
1941 swp_entry_to_pte(swp_entry(0, ~0UL))));
1942 maxpages = swp_offset(radix_to_swp_entry(
1943 swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1945 if (maxpages > swap_header->info.last_page) {
1946 maxpages = swap_header->info.last_page + 1;
1947 /* p->max is an unsigned int: don't overflow it */
1948 if ((unsigned int)maxpages == 0)
1949 maxpages = UINT_MAX;
1951 p->highest_bit = maxpages - 1;
1955 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1956 if (swapfilepages && maxpages > swapfilepages) {
1958 "Swap area shorter than signature indicates\n");
1961 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1963 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1969 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1970 union swap_header *swap_header,
1971 unsigned char *swap_map,
1972 unsigned long maxpages,
1976 unsigned int nr_good_pages;
1979 nr_good_pages = maxpages - 1; /* omit header page */
1981 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1982 unsigned int page_nr = swap_header->info.badpages[i];
1983 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1985 if (page_nr < maxpages) {
1986 swap_map[page_nr] = SWAP_MAP_BAD;
1991 if (nr_good_pages) {
1992 swap_map[0] = SWAP_MAP_BAD;
1994 p->pages = nr_good_pages;
1995 nr_extents = setup_swap_extents(p, span);
1998 nr_good_pages = p->pages;
2000 if (!nr_good_pages) {
2001 printk(KERN_WARNING "Empty swap-file\n");
2008 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2010 struct swap_info_struct *p;
2012 struct file *swap_file = NULL;
2013 struct address_space *mapping;
2017 union swap_header *swap_header;
2020 unsigned long maxpages;
2021 unsigned char *swap_map = NULL;
2022 struct page *page = NULL;
2023 struct inode *inode = NULL;
2025 if (!capable(CAP_SYS_ADMIN))
2028 p = alloc_swap_info();
2032 name = getname(specialfile);
2034 error = PTR_ERR(name);
2038 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2039 if (IS_ERR(swap_file)) {
2040 error = PTR_ERR(swap_file);
2045 p->swap_file = swap_file;
2046 mapping = swap_file->f_mapping;
2048 for (i = 0; i < nr_swapfiles; i++) {
2049 struct swap_info_struct *q = swap_info[i];
2051 if (q == p || !q->swap_file)
2053 if (mapping == q->swap_file->f_mapping) {
2059 inode = mapping->host;
2060 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2061 error = claim_swapfile(p, inode);
2062 if (unlikely(error))
2066 * Read the swap header.
2068 if (!mapping->a_ops->readpage) {
2072 page = read_mapping_page(mapping, 0, swap_file);
2074 error = PTR_ERR(page);
2077 swap_header = kmap(page);
2079 maxpages = read_swap_header(p, swap_header, inode);
2080 if (unlikely(!maxpages)) {
2085 /* OK, set up the swap map and apply the bad block list */
2086 swap_map = vzalloc(maxpages);
2092 error = swap_cgroup_swapon(p->type, maxpages);
2096 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2098 if (unlikely(nr_extents < 0)) {
2104 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2105 p->flags |= SWP_SOLIDSTATE;
2106 p->cluster_next = 1 + (random32() % p->highest_bit);
2108 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2109 p->flags |= SWP_DISCARDABLE;
2112 mutex_lock(&swapon_mutex);
2114 if (swap_flags & SWAP_FLAG_PREFER)
2116 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2117 enable_swap_info(p, prio, swap_map);
2119 printk(KERN_INFO "Adding %uk swap on %s. "
2120 "Priority:%d extents:%d across:%lluk %s%s\n",
2121 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2122 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2123 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2124 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2126 mutex_unlock(&swapon_mutex);
2127 atomic_inc(&proc_poll_event);
2128 wake_up_interruptible(&proc_poll_wait);
2130 if (S_ISREG(inode->i_mode))
2131 inode->i_flags |= S_SWAPFILE;
2135 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2136 set_blocksize(p->bdev, p->old_block_size);
2137 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2139 destroy_swap_extents(p);
2140 swap_cgroup_swapoff(p->type);
2141 spin_lock(&swap_lock);
2142 p->swap_file = NULL;
2144 spin_unlock(&swap_lock);
2147 if (inode && S_ISREG(inode->i_mode)) {
2148 mutex_unlock(&inode->i_mutex);
2151 filp_close(swap_file, NULL);
2154 if (page && !IS_ERR(page)) {
2156 page_cache_release(page);
2160 if (inode && S_ISREG(inode->i_mode))
2161 mutex_unlock(&inode->i_mutex);
2165 void si_swapinfo(struct sysinfo *val)
2168 unsigned long nr_to_be_unused = 0;
2170 spin_lock(&swap_lock);
2171 for (type = 0; type < nr_swapfiles; type++) {
2172 struct swap_info_struct *si = swap_info[type];
2174 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2175 nr_to_be_unused += si->inuse_pages;
2177 val->freeswap = nr_swap_pages + nr_to_be_unused;
2178 val->totalswap = total_swap_pages + nr_to_be_unused;
2179 spin_unlock(&swap_lock);
2183 * Verify that a swap entry is valid and increment its swap map count.
2185 * Returns error code in following case.
2187 * - swp_entry is invalid -> EINVAL
2188 * - swp_entry is migration entry -> EINVAL
2189 * - swap-cache reference is requested but there is already one. -> EEXIST
2190 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2191 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2193 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2195 struct swap_info_struct *p;
2196 unsigned long offset, type;
2197 unsigned char count;
2198 unsigned char has_cache;
2201 if (non_swap_entry(entry))
2204 type = swp_type(entry);
2205 if (type >= nr_swapfiles)
2207 p = swap_info[type];
2208 offset = swp_offset(entry);
2210 spin_lock(&swap_lock);
2211 if (unlikely(offset >= p->max))
2214 count = p->swap_map[offset];
2215 has_cache = count & SWAP_HAS_CACHE;
2216 count &= ~SWAP_HAS_CACHE;
2219 if (usage == SWAP_HAS_CACHE) {
2221 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2222 if (!has_cache && count)
2223 has_cache = SWAP_HAS_CACHE;
2224 else if (has_cache) /* someone else added cache */
2226 else /* no users remaining */
2229 } else if (count || has_cache) {
2231 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2233 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2235 else if (swap_count_continued(p, offset, count))
2236 count = COUNT_CONTINUED;
2240 err = -ENOENT; /* unused swap entry */
2242 p->swap_map[offset] = count | has_cache;
2245 spin_unlock(&swap_lock);
2250 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2255 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2256 * (in which case its reference count is never incremented).
2258 void swap_shmem_alloc(swp_entry_t entry)
2260 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2264 * Increase reference count of swap entry by 1.
2265 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2266 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2267 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2268 * might occur if a page table entry has got corrupted.
2270 int swap_duplicate(swp_entry_t entry)
2274 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2275 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2280 * @entry: swap entry for which we allocate swap cache.
2282 * Called when allocating swap cache for existing swap entry,
2283 * This can return error codes. Returns 0 at success.
2284 * -EBUSY means there is a swap cache.
2285 * Note: return code is different from swap_duplicate().
2287 int swapcache_prepare(swp_entry_t entry)
2289 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2293 * swap_lock prevents swap_map being freed. Don't grab an extra
2294 * reference on the swaphandle, it doesn't matter if it becomes unused.
2296 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2298 struct swap_info_struct *si;
2299 int our_page_cluster = page_cluster;
2300 pgoff_t target, toff;
2304 if (!our_page_cluster) /* no readahead */
2307 si = swap_info[swp_type(entry)];
2308 target = swp_offset(entry);
2309 base = (target >> our_page_cluster) << our_page_cluster;
2310 end = base + (1 << our_page_cluster);
2311 if (!base) /* first page is swap header */
2314 spin_lock(&swap_lock);
2315 if (end > si->max) /* don't go beyond end of map */
2318 /* Count contiguous allocated slots above our target */
2319 for (toff = target; ++toff < end; nr_pages++) {
2320 /* Don't read in free or bad pages */
2321 if (!si->swap_map[toff])
2323 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2326 /* Count contiguous allocated slots below our target */
2327 for (toff = target; --toff >= base; nr_pages++) {
2328 /* Don't read in free or bad pages */
2329 if (!si->swap_map[toff])
2331 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2334 spin_unlock(&swap_lock);
2337 * Indicate starting offset, and return number of pages to get:
2338 * if only 1, say 0, since there's then no readahead to be done.
2341 return nr_pages? ++nr_pages: 0;
2345 * add_swap_count_continuation - called when a swap count is duplicated
2346 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2347 * page of the original vmalloc'ed swap_map, to hold the continuation count
2348 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2349 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2351 * These continuation pages are seldom referenced: the common paths all work
2352 * on the original swap_map, only referring to a continuation page when the
2353 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2355 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2356 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2357 * can be called after dropping locks.
2359 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2361 struct swap_info_struct *si;
2364 struct page *list_page;
2366 unsigned char count;
2369 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2370 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2372 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2374 si = swap_info_get(entry);
2377 * An acceptable race has occurred since the failing
2378 * __swap_duplicate(): the swap entry has been freed,
2379 * perhaps even the whole swap_map cleared for swapoff.
2384 offset = swp_offset(entry);
2385 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2387 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2389 * The higher the swap count, the more likely it is that tasks
2390 * will race to add swap count continuation: we need to avoid
2391 * over-provisioning.
2397 spin_unlock(&swap_lock);
2402 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2403 * no architecture is using highmem pages for kernel pagetables: so it
2404 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2406 head = vmalloc_to_page(si->swap_map + offset);
2407 offset &= ~PAGE_MASK;
2410 * Page allocation does not initialize the page's lru field,
2411 * but it does always reset its private field.
2413 if (!page_private(head)) {
2414 BUG_ON(count & COUNT_CONTINUED);
2415 INIT_LIST_HEAD(&head->lru);
2416 set_page_private(head, SWP_CONTINUED);
2417 si->flags |= SWP_CONTINUED;
2420 list_for_each_entry(list_page, &head->lru, lru) {
2424 * If the previous map said no continuation, but we've found
2425 * a continuation page, free our allocation and use this one.
2427 if (!(count & COUNT_CONTINUED))
2430 map = kmap_atomic(list_page, KM_USER0) + offset;
2432 kunmap_atomic(map, KM_USER0);
2435 * If this continuation count now has some space in it,
2436 * free our allocation and use this one.
2438 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2442 list_add_tail(&page->lru, &head->lru);
2443 page = NULL; /* now it's attached, don't free it */
2445 spin_unlock(&swap_lock);
2453 * swap_count_continued - when the original swap_map count is incremented
2454 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2455 * into, carry if so, or else fail until a new continuation page is allocated;
2456 * when the original swap_map count is decremented from 0 with continuation,
2457 * borrow from the continuation and report whether it still holds more.
2458 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2460 static bool swap_count_continued(struct swap_info_struct *si,
2461 pgoff_t offset, unsigned char count)
2467 head = vmalloc_to_page(si->swap_map + offset);
2468 if (page_private(head) != SWP_CONTINUED) {
2469 BUG_ON(count & COUNT_CONTINUED);
2470 return false; /* need to add count continuation */
2473 offset &= ~PAGE_MASK;
2474 page = list_entry(head->lru.next, struct page, lru);
2475 map = kmap_atomic(page, KM_USER0) + offset;
2477 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2478 goto init_map; /* jump over SWAP_CONT_MAX checks */
2480 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2482 * Think of how you add 1 to 999
2484 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2485 kunmap_atomic(map, KM_USER0);
2486 page = list_entry(page->lru.next, struct page, lru);
2487 BUG_ON(page == head);
2488 map = kmap_atomic(page, KM_USER0) + offset;
2490 if (*map == SWAP_CONT_MAX) {
2491 kunmap_atomic(map, KM_USER0);
2492 page = list_entry(page->lru.next, struct page, lru);
2494 return false; /* add count continuation */
2495 map = kmap_atomic(page, KM_USER0) + offset;
2496 init_map: *map = 0; /* we didn't zero the page */
2499 kunmap_atomic(map, KM_USER0);
2500 page = list_entry(page->lru.prev, struct page, lru);
2501 while (page != head) {
2502 map = kmap_atomic(page, KM_USER0) + offset;
2503 *map = COUNT_CONTINUED;
2504 kunmap_atomic(map, KM_USER0);
2505 page = list_entry(page->lru.prev, struct page, lru);
2507 return true; /* incremented */
2509 } else { /* decrementing */
2511 * Think of how you subtract 1 from 1000
2513 BUG_ON(count != COUNT_CONTINUED);
2514 while (*map == COUNT_CONTINUED) {
2515 kunmap_atomic(map, KM_USER0);
2516 page = list_entry(page->lru.next, struct page, lru);
2517 BUG_ON(page == head);
2518 map = kmap_atomic(page, KM_USER0) + offset;
2524 kunmap_atomic(map, KM_USER0);
2525 page = list_entry(page->lru.prev, struct page, lru);
2526 while (page != head) {
2527 map = kmap_atomic(page, KM_USER0) + offset;
2528 *map = SWAP_CONT_MAX | count;
2529 count = COUNT_CONTINUED;
2530 kunmap_atomic(map, KM_USER0);
2531 page = list_entry(page->lru.prev, struct page, lru);
2533 return count == COUNT_CONTINUED;
2538 * free_swap_count_continuations - swapoff free all the continuation pages
2539 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2541 static void free_swap_count_continuations(struct swap_info_struct *si)
2545 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2547 head = vmalloc_to_page(si->swap_map + offset);
2548 if (page_private(head)) {
2549 struct list_head *this, *next;
2550 list_for_each_safe(this, next, &head->lru) {
2552 page = list_entry(this, struct page, lru);