4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/security.h>
28 #include <linux/backing-dev.h>
29 #include <linux/mutex.h>
30 #include <linux/capability.h>
31 #include <linux/syscalls.h>
32 #include <linux/memcontrol.h>
34 #include <asm/pgtable.h>
35 #include <asm/tlbflush.h>
36 #include <linux/swapops.h>
37 #include <linux/page_cgroup.h>
39 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
41 static void free_swap_count_continuations(struct swap_info_struct *);
42 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
44 static DEFINE_SPINLOCK(swap_lock);
45 static unsigned int nr_swapfiles;
47 long total_swap_pages;
48 static int least_priority;
50 static bool swap_for_hibernation;
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 inline unsigned char swap_count(unsigned char ent)
65 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
68 /* returns 1 if swap entry is freed */
70 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
72 swp_entry_t entry = swp_entry(si->type, offset);
76 page = find_get_page(&swapper_space, entry.val);
80 * This function is called from scan_swap_map() and it's called
81 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
82 * We have to use trylock for avoiding deadlock. This is a special
83 * case and you should use try_to_free_swap() with explicit lock_page()
84 * in usual operations.
86 if (trylock_page(page)) {
87 ret = try_to_free_swap(page);
90 page_cache_release(page);
95 * We need this because the bdev->unplug_fn can sleep and we cannot
96 * hold swap_lock while calling the unplug_fn. And swap_lock
97 * cannot be turned into a mutex.
99 static DECLARE_RWSEM(swap_unplug_sem);
101 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
105 down_read(&swap_unplug_sem);
106 entry.val = page_private(page);
107 if (PageSwapCache(page)) {
108 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
109 struct backing_dev_info *bdi;
112 * If the page is removed from swapcache from under us (with a
113 * racy try_to_unuse/swapoff) we need an additional reference
114 * count to avoid reading garbage from page_private(page) above.
115 * If the WARN_ON triggers during a swapoff it maybe the race
116 * condition and it's harmless. However if it triggers without
117 * swapoff it signals a problem.
119 WARN_ON(page_count(page) <= 1);
121 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
122 blk_run_backing_dev(bdi, page);
124 up_read(&swap_unplug_sem);
128 * swapon tell device that all the old swap contents can be discarded,
129 * to allow the swap device to optimize its wear-levelling.
131 static int discard_swap(struct swap_info_struct *si)
133 struct swap_extent *se;
134 sector_t start_block;
138 /* Do not discard the swap header page! */
139 se = &si->first_swap_extent;
140 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
141 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
143 err = blkdev_issue_discard(si->bdev, start_block,
144 nr_blocks, GFP_KERNEL, 0);
150 list_for_each_entry(se, &si->first_swap_extent.list, list) {
151 start_block = se->start_block << (PAGE_SHIFT - 9);
152 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
154 err = blkdev_issue_discard(si->bdev, start_block,
155 nr_blocks, GFP_KERNEL, 0);
161 return err; /* That will often be -EOPNOTSUPP */
165 * swap allocation tell device that a cluster of swap can now be discarded,
166 * to allow the swap device to optimize its wear-levelling.
168 static void discard_swap_cluster(struct swap_info_struct *si,
169 pgoff_t start_page, pgoff_t nr_pages)
171 struct swap_extent *se = si->curr_swap_extent;
172 int found_extent = 0;
175 struct list_head *lh;
177 if (se->start_page <= start_page &&
178 start_page < se->start_page + se->nr_pages) {
179 pgoff_t offset = start_page - se->start_page;
180 sector_t start_block = se->start_block + offset;
181 sector_t nr_blocks = se->nr_pages - offset;
183 if (nr_blocks > nr_pages)
184 nr_blocks = nr_pages;
185 start_page += nr_blocks;
186 nr_pages -= nr_blocks;
189 si->curr_swap_extent = se;
191 start_block <<= PAGE_SHIFT - 9;
192 nr_blocks <<= PAGE_SHIFT - 9;
193 if (blkdev_issue_discard(si->bdev, start_block,
194 nr_blocks, GFP_NOIO, 0))
199 se = list_entry(lh, struct swap_extent, list);
203 static int wait_for_discard(void *word)
209 #define SWAPFILE_CLUSTER 256
210 #define LATENCY_LIMIT 256
212 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
215 unsigned long offset;
216 unsigned long scan_base;
217 unsigned long last_in_cluster = 0;
218 int latency_ration = LATENCY_LIMIT;
219 int found_free_cluster = 0;
222 * We try to cluster swap pages by allocating them sequentially
223 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
224 * way, however, we resort to first-free allocation, starting
225 * a new cluster. This prevents us from scattering swap pages
226 * all over the entire swap partition, so that we reduce
227 * overall disk seek times between swap pages. -- sct
228 * But we do now try to find an empty cluster. -Andrea
229 * And we let swap pages go all over an SSD partition. Hugh
232 si->flags += SWP_SCANNING;
233 scan_base = offset = si->cluster_next;
235 if (unlikely(!si->cluster_nr--)) {
236 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
237 si->cluster_nr = SWAPFILE_CLUSTER - 1;
240 if (si->flags & SWP_DISCARDABLE) {
242 * Start range check on racing allocations, in case
243 * they overlap the cluster we eventually decide on
244 * (we scan without swap_lock to allow preemption).
245 * It's hardly conceivable that cluster_nr could be
246 * wrapped during our scan, but don't depend on it.
248 if (si->lowest_alloc)
250 si->lowest_alloc = si->max;
251 si->highest_alloc = 0;
253 spin_unlock(&swap_lock);
256 * If seek is expensive, start searching for new cluster from
257 * start of partition, to minimize the span of allocated swap.
258 * But if seek is cheap, search from our current position, so
259 * that swap is allocated from all over the partition: if the
260 * Flash Translation Layer only remaps within limited zones,
261 * we don't want to wear out the first zone too quickly.
263 if (!(si->flags & SWP_SOLIDSTATE))
264 scan_base = offset = si->lowest_bit;
265 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
267 /* Locate the first empty (unaligned) cluster */
268 for (; last_in_cluster <= si->highest_bit; offset++) {
269 if (si->swap_map[offset])
270 last_in_cluster = offset + SWAPFILE_CLUSTER;
271 else if (offset == last_in_cluster) {
272 spin_lock(&swap_lock);
273 offset -= SWAPFILE_CLUSTER - 1;
274 si->cluster_next = offset;
275 si->cluster_nr = SWAPFILE_CLUSTER - 1;
276 found_free_cluster = 1;
279 if (unlikely(--latency_ration < 0)) {
281 latency_ration = LATENCY_LIMIT;
285 offset = si->lowest_bit;
286 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
288 /* Locate the first empty (unaligned) cluster */
289 for (; last_in_cluster < scan_base; offset++) {
290 if (si->swap_map[offset])
291 last_in_cluster = offset + SWAPFILE_CLUSTER;
292 else if (offset == last_in_cluster) {
293 spin_lock(&swap_lock);
294 offset -= SWAPFILE_CLUSTER - 1;
295 si->cluster_next = offset;
296 si->cluster_nr = SWAPFILE_CLUSTER - 1;
297 found_free_cluster = 1;
300 if (unlikely(--latency_ration < 0)) {
302 latency_ration = LATENCY_LIMIT;
307 spin_lock(&swap_lock);
308 si->cluster_nr = SWAPFILE_CLUSTER - 1;
309 si->lowest_alloc = 0;
313 if (!(si->flags & SWP_WRITEOK))
315 if (!si->highest_bit)
317 if (offset > si->highest_bit)
318 scan_base = offset = si->lowest_bit;
320 /* reuse swap entry of cache-only swap if not hibernation. */
322 && usage == SWAP_HAS_CACHE
323 && si->swap_map[offset] == SWAP_HAS_CACHE) {
325 spin_unlock(&swap_lock);
326 swap_was_freed = __try_to_reclaim_swap(si, offset);
327 spin_lock(&swap_lock);
328 /* entry was freed successfully, try to use this again */
331 goto scan; /* check next one */
334 if (si->swap_map[offset])
337 if (offset == si->lowest_bit)
339 if (offset == si->highest_bit)
342 if (si->inuse_pages == si->pages) {
343 si->lowest_bit = si->max;
346 si->swap_map[offset] = usage;
347 si->cluster_next = offset + 1;
348 si->flags -= SWP_SCANNING;
350 if (si->lowest_alloc) {
352 * Only set when SWP_DISCARDABLE, and there's a scan
353 * for a free cluster in progress or just completed.
355 if (found_free_cluster) {
357 * To optimize wear-levelling, discard the
358 * old data of the cluster, taking care not to
359 * discard any of its pages that have already
360 * been allocated by racing tasks (offset has
361 * already stepped over any at the beginning).
363 if (offset < si->highest_alloc &&
364 si->lowest_alloc <= last_in_cluster)
365 last_in_cluster = si->lowest_alloc - 1;
366 si->flags |= SWP_DISCARDING;
367 spin_unlock(&swap_lock);
369 if (offset < last_in_cluster)
370 discard_swap_cluster(si, offset,
371 last_in_cluster - offset + 1);
373 spin_lock(&swap_lock);
374 si->lowest_alloc = 0;
375 si->flags &= ~SWP_DISCARDING;
377 smp_mb(); /* wake_up_bit advises this */
378 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
380 } else if (si->flags & SWP_DISCARDING) {
382 * Delay using pages allocated by racing tasks
383 * until the whole discard has been issued. We
384 * could defer that delay until swap_writepage,
385 * but it's easier to keep this self-contained.
387 spin_unlock(&swap_lock);
388 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
389 wait_for_discard, TASK_UNINTERRUPTIBLE);
390 spin_lock(&swap_lock);
393 * Note pages allocated by racing tasks while
394 * scan for a free cluster is in progress, so
395 * that its final discard can exclude them.
397 if (offset < si->lowest_alloc)
398 si->lowest_alloc = offset;
399 if (offset > si->highest_alloc)
400 si->highest_alloc = offset;
406 spin_unlock(&swap_lock);
407 while (++offset <= si->highest_bit) {
408 if (!si->swap_map[offset]) {
409 spin_lock(&swap_lock);
412 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
413 spin_lock(&swap_lock);
416 if (unlikely(--latency_ration < 0)) {
418 latency_ration = LATENCY_LIMIT;
421 offset = si->lowest_bit;
422 while (++offset < scan_base) {
423 if (!si->swap_map[offset]) {
424 spin_lock(&swap_lock);
427 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
428 spin_lock(&swap_lock);
431 if (unlikely(--latency_ration < 0)) {
433 latency_ration = LATENCY_LIMIT;
436 spin_lock(&swap_lock);
439 si->flags -= SWP_SCANNING;
443 swp_entry_t get_swap_page(void)
445 struct swap_info_struct *si;
450 spin_lock(&swap_lock);
451 if (nr_swap_pages <= 0)
453 if (swap_for_hibernation)
457 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
458 si = swap_info[type];
461 (!wrapped && si->prio != swap_info[next]->prio)) {
462 next = swap_list.head;
466 if (!si->highest_bit)
468 if (!(si->flags & SWP_WRITEOK))
471 swap_list.next = next;
472 /* This is called for allocating swap entry for cache */
473 offset = scan_swap_map(si, SWAP_HAS_CACHE);
475 spin_unlock(&swap_lock);
476 return swp_entry(type, offset);
478 next = swap_list.next;
483 spin_unlock(&swap_lock);
484 return (swp_entry_t) {0};
487 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
489 struct swap_info_struct *p;
490 unsigned long offset, type;
494 type = swp_type(entry);
495 if (type >= nr_swapfiles)
498 if (!(p->flags & SWP_USED))
500 offset = swp_offset(entry);
501 if (offset >= p->max)
503 if (!p->swap_map[offset])
505 spin_lock(&swap_lock);
509 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
512 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
515 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
518 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
523 static unsigned char swap_entry_free(struct swap_info_struct *p,
524 swp_entry_t entry, unsigned char usage)
526 unsigned long offset = swp_offset(entry);
528 unsigned char has_cache;
530 count = p->swap_map[offset];
531 has_cache = count & SWAP_HAS_CACHE;
532 count &= ~SWAP_HAS_CACHE;
534 if (usage == SWAP_HAS_CACHE) {
535 VM_BUG_ON(!has_cache);
537 } else if (count == SWAP_MAP_SHMEM) {
539 * Or we could insist on shmem.c using a special
540 * swap_shmem_free() and free_shmem_swap_and_cache()...
543 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
544 if (count == COUNT_CONTINUED) {
545 if (swap_count_continued(p, offset, count))
546 count = SWAP_MAP_MAX | COUNT_CONTINUED;
548 count = SWAP_MAP_MAX;
554 mem_cgroup_uncharge_swap(entry);
556 usage = count | has_cache;
557 p->swap_map[offset] = usage;
559 /* free if no reference */
561 struct gendisk *disk = p->bdev->bd_disk;
562 if (offset < p->lowest_bit)
563 p->lowest_bit = offset;
564 if (offset > p->highest_bit)
565 p->highest_bit = offset;
566 if (swap_list.next >= 0 &&
567 p->prio > swap_info[swap_list.next]->prio)
568 swap_list.next = p->type;
571 if ((p->flags & SWP_BLKDEV) &&
572 disk->fops->swap_slot_free_notify)
573 disk->fops->swap_slot_free_notify(p->bdev, offset);
580 * Caller has made sure that the swapdevice corresponding to entry
581 * is still around or has not been recycled.
583 void swap_free(swp_entry_t entry)
585 struct swap_info_struct *p;
587 p = swap_info_get(entry);
589 swap_entry_free(p, entry, 1);
590 spin_unlock(&swap_lock);
595 * Called after dropping swapcache to decrease refcnt to swap entries.
597 void swapcache_free(swp_entry_t entry, struct page *page)
599 struct swap_info_struct *p;
602 p = swap_info_get(entry);
604 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
606 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
607 spin_unlock(&swap_lock);
612 * How many references to page are currently swapped out?
613 * This does not give an exact answer when swap count is continued,
614 * but does include the high COUNT_CONTINUED flag to allow for that.
616 static inline int page_swapcount(struct page *page)
619 struct swap_info_struct *p;
622 entry.val = page_private(page);
623 p = swap_info_get(entry);
625 count = swap_count(p->swap_map[swp_offset(entry)]);
626 spin_unlock(&swap_lock);
632 * We can write to an anon page without COW if there are no other references
633 * to it. And as a side-effect, free up its swap: because the old content
634 * on disk will never be read, and seeking back there to write new content
635 * later would only waste time away from clustering.
637 int reuse_swap_page(struct page *page)
641 VM_BUG_ON(!PageLocked(page));
642 if (unlikely(PageKsm(page)))
644 count = page_mapcount(page);
645 if (count <= 1 && PageSwapCache(page)) {
646 count += page_swapcount(page);
647 if (count == 1 && !PageWriteback(page)) {
648 delete_from_swap_cache(page);
656 * If swap is getting full, or if there are no more mappings of this page,
657 * then try_to_free_swap is called to free its swap space.
659 int try_to_free_swap(struct page *page)
661 VM_BUG_ON(!PageLocked(page));
663 if (!PageSwapCache(page))
665 if (PageWriteback(page))
667 if (page_swapcount(page))
670 delete_from_swap_cache(page);
676 * Free the swap entry like above, but also try to
677 * free the page cache entry if it is the last user.
679 int free_swap_and_cache(swp_entry_t entry)
681 struct swap_info_struct *p;
682 struct page *page = NULL;
684 if (non_swap_entry(entry))
687 p = swap_info_get(entry);
689 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
690 page = find_get_page(&swapper_space, entry.val);
691 if (page && !trylock_page(page)) {
692 page_cache_release(page);
696 spin_unlock(&swap_lock);
700 * Not mapped elsewhere, or swap space full? Free it!
701 * Also recheck PageSwapCache now page is locked (above).
703 if (PageSwapCache(page) && !PageWriteback(page) &&
704 (!page_mapped(page) || vm_swap_full())) {
705 delete_from_swap_cache(page);
709 page_cache_release(page);
714 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
716 * mem_cgroup_count_swap_user - count the user of a swap entry
717 * @ent: the swap entry to be checked
718 * @pagep: the pointer for the swap cache page of the entry to be stored
720 * Returns the number of the user of the swap entry. The number is valid only
721 * for swaps of anonymous pages.
722 * If the entry is found on swap cache, the page is stored to pagep with
723 * refcount of it being incremented.
725 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
728 struct swap_info_struct *p;
731 page = find_get_page(&swapper_space, ent.val);
733 count += page_mapcount(page);
734 p = swap_info_get(ent);
736 count += swap_count(p->swap_map[swp_offset(ent)]);
737 spin_unlock(&swap_lock);
745 #ifdef CONFIG_HIBERNATION
747 static pgoff_t hibernation_offset[MAX_SWAPFILES];
749 * Once hibernation starts to use swap, we freeze swap_map[]. Otherwise,
750 * saved swap_map[] image to the disk will be an incomplete because it's
751 * changing without synchronization with hibernation snap shot.
752 * At resume, we just make swap_for_hibernation=false. We can forget
755 void hibernation_freeze_swap(void)
759 spin_lock(&swap_lock);
761 printk(KERN_INFO "PM: Freeze Swap\n");
762 swap_for_hibernation = true;
763 for (i = 0; i < MAX_SWAPFILES; i++)
764 hibernation_offset[i] = 1;
765 spin_unlock(&swap_lock);
768 void hibernation_thaw_swap(void)
770 spin_lock(&swap_lock);
771 if (swap_for_hibernation) {
772 printk(KERN_INFO "PM: Thaw Swap\n");
773 swap_for_hibernation = false;
775 spin_unlock(&swap_lock);
779 * Because updateing swap_map[] can make not-saved-status-change,
780 * we use our own easy allocator.
781 * Please see kernel/power/swap.c, Used swaps are recorded into
784 swp_entry_t get_swap_for_hibernation(int type)
787 swp_entry_t val = {0};
788 struct swap_info_struct *si;
790 spin_lock(&swap_lock);
792 si = swap_info[type];
793 if (!si || !(si->flags & SWP_WRITEOK))
796 for (off = hibernation_offset[type]; off < si->max; ++off) {
797 if (!si->swap_map[off])
801 val = swp_entry(type, off);
802 hibernation_offset[type] = off + 1;
805 spin_unlock(&swap_lock);
809 void swap_free_for_hibernation(swp_entry_t ent)
815 * Find the swap type that corresponds to given device (if any).
817 * @offset - number of the PAGE_SIZE-sized block of the device, starting
818 * from 0, in which the swap header is expected to be located.
820 * This is needed for the suspend to disk (aka swsusp).
822 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
824 struct block_device *bdev = NULL;
828 bdev = bdget(device);
830 spin_lock(&swap_lock);
831 for (type = 0; type < nr_swapfiles; type++) {
832 struct swap_info_struct *sis = swap_info[type];
834 if (!(sis->flags & SWP_WRITEOK))
839 *bdev_p = bdgrab(sis->bdev);
841 spin_unlock(&swap_lock);
844 if (bdev == sis->bdev) {
845 struct swap_extent *se = &sis->first_swap_extent;
847 if (se->start_block == offset) {
849 *bdev_p = bdgrab(sis->bdev);
851 spin_unlock(&swap_lock);
857 spin_unlock(&swap_lock);
865 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
866 * corresponding to given index in swap_info (swap type).
868 sector_t swapdev_block(int type, pgoff_t offset)
870 struct block_device *bdev;
872 if ((unsigned int)type >= nr_swapfiles)
874 if (!(swap_info[type]->flags & SWP_WRITEOK))
876 return map_swap_entry(swp_entry(type, offset), &bdev);
880 * Return either the total number of swap pages of given type, or the number
881 * of free pages of that type (depending on @free)
883 * This is needed for software suspend
885 unsigned int count_swap_pages(int type, int free)
889 spin_lock(&swap_lock);
890 if ((unsigned int)type < nr_swapfiles) {
891 struct swap_info_struct *sis = swap_info[type];
893 if (sis->flags & SWP_WRITEOK) {
896 n -= sis->inuse_pages;
899 spin_unlock(&swap_lock);
902 #endif /* CONFIG_HIBERNATION */
905 * No need to decide whether this PTE shares the swap entry with others,
906 * just let do_wp_page work it out if a write is requested later - to
907 * force COW, vm_page_prot omits write permission from any private vma.
909 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
910 unsigned long addr, swp_entry_t entry, struct page *page)
912 struct mem_cgroup *ptr = NULL;
917 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
922 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
923 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
925 mem_cgroup_cancel_charge_swapin(ptr);
930 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
931 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
933 set_pte_at(vma->vm_mm, addr, pte,
934 pte_mkold(mk_pte(page, vma->vm_page_prot)));
935 page_add_anon_rmap(page, vma, addr);
936 mem_cgroup_commit_charge_swapin(page, ptr);
939 * Move the page to the active list so it is not
940 * immediately swapped out again after swapon.
944 pte_unmap_unlock(pte, ptl);
949 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
950 unsigned long addr, unsigned long end,
951 swp_entry_t entry, struct page *page)
953 pte_t swp_pte = swp_entry_to_pte(entry);
958 * We don't actually need pte lock while scanning for swp_pte: since
959 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
960 * page table while we're scanning; though it could get zapped, and on
961 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
962 * of unmatched parts which look like swp_pte, so unuse_pte must
963 * recheck under pte lock. Scanning without pte lock lets it be
964 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
966 pte = pte_offset_map(pmd, addr);
969 * swapoff spends a _lot_ of time in this loop!
970 * Test inline before going to call unuse_pte.
972 if (unlikely(pte_same(*pte, swp_pte))) {
974 ret = unuse_pte(vma, pmd, addr, entry, page);
977 pte = pte_offset_map(pmd, addr);
979 } while (pte++, addr += PAGE_SIZE, addr != end);
985 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
986 unsigned long addr, unsigned long end,
987 swp_entry_t entry, struct page *page)
993 pmd = pmd_offset(pud, addr);
995 next = pmd_addr_end(addr, end);
996 if (pmd_none_or_clear_bad(pmd))
998 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1001 } while (pmd++, addr = next, addr != end);
1005 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
1006 unsigned long addr, unsigned long end,
1007 swp_entry_t entry, struct page *page)
1013 pud = pud_offset(pgd, addr);
1015 next = pud_addr_end(addr, end);
1016 if (pud_none_or_clear_bad(pud))
1018 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1021 } while (pud++, addr = next, addr != end);
1025 static int unuse_vma(struct vm_area_struct *vma,
1026 swp_entry_t entry, struct page *page)
1029 unsigned long addr, end, next;
1032 if (page_anon_vma(page)) {
1033 addr = page_address_in_vma(page, vma);
1034 if (addr == -EFAULT)
1037 end = addr + PAGE_SIZE;
1039 addr = vma->vm_start;
1043 pgd = pgd_offset(vma->vm_mm, addr);
1045 next = pgd_addr_end(addr, end);
1046 if (pgd_none_or_clear_bad(pgd))
1048 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1051 } while (pgd++, addr = next, addr != end);
1055 static int unuse_mm(struct mm_struct *mm,
1056 swp_entry_t entry, struct page *page)
1058 struct vm_area_struct *vma;
1061 if (!down_read_trylock(&mm->mmap_sem)) {
1063 * Activate page so shrink_inactive_list is unlikely to unmap
1064 * its ptes while lock is dropped, so swapoff can make progress.
1066 activate_page(page);
1068 down_read(&mm->mmap_sem);
1071 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1072 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1075 up_read(&mm->mmap_sem);
1076 return (ret < 0)? ret: 0;
1080 * Scan swap_map from current position to next entry still in use.
1081 * Recycle to start on reaching the end, returning 0 when empty.
1083 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1086 unsigned int max = si->max;
1087 unsigned int i = prev;
1088 unsigned char count;
1091 * No need for swap_lock here: we're just looking
1092 * for whether an entry is in use, not modifying it; false
1093 * hits are okay, and sys_swapoff() has already prevented new
1094 * allocations from this area (while holding swap_lock).
1103 * No entries in use at top of swap_map,
1104 * loop back to start and recheck there.
1110 count = si->swap_map[i];
1111 if (count && swap_count(count) != SWAP_MAP_BAD)
1118 * We completely avoid races by reading each swap page in advance,
1119 * and then search for the process using it. All the necessary
1120 * page table adjustments can then be made atomically.
1122 static int try_to_unuse(unsigned int type)
1124 struct swap_info_struct *si = swap_info[type];
1125 struct mm_struct *start_mm;
1126 unsigned char *swap_map;
1127 unsigned char swcount;
1134 * When searching mms for an entry, a good strategy is to
1135 * start at the first mm we freed the previous entry from
1136 * (though actually we don't notice whether we or coincidence
1137 * freed the entry). Initialize this start_mm with a hold.
1139 * A simpler strategy would be to start at the last mm we
1140 * freed the previous entry from; but that would take less
1141 * advantage of mmlist ordering, which clusters forked mms
1142 * together, child after parent. If we race with dup_mmap(), we
1143 * prefer to resolve parent before child, lest we miss entries
1144 * duplicated after we scanned child: using last mm would invert
1147 start_mm = &init_mm;
1148 atomic_inc(&init_mm.mm_users);
1151 * Keep on scanning until all entries have gone. Usually,
1152 * one pass through swap_map is enough, but not necessarily:
1153 * there are races when an instance of an entry might be missed.
1155 while ((i = find_next_to_unuse(si, i)) != 0) {
1156 if (signal_pending(current)) {
1162 * Get a page for the entry, using the existing swap
1163 * cache page if there is one. Otherwise, get a clean
1164 * page and read the swap into it.
1166 swap_map = &si->swap_map[i];
1167 entry = swp_entry(type, i);
1168 page = read_swap_cache_async(entry,
1169 GFP_HIGHUSER_MOVABLE, NULL, 0);
1172 * Either swap_duplicate() failed because entry
1173 * has been freed independently, and will not be
1174 * reused since sys_swapoff() already disabled
1175 * allocation from here, or alloc_page() failed.
1184 * Don't hold on to start_mm if it looks like exiting.
1186 if (atomic_read(&start_mm->mm_users) == 1) {
1188 start_mm = &init_mm;
1189 atomic_inc(&init_mm.mm_users);
1193 * Wait for and lock page. When do_swap_page races with
1194 * try_to_unuse, do_swap_page can handle the fault much
1195 * faster than try_to_unuse can locate the entry. This
1196 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1197 * defer to do_swap_page in such a case - in some tests,
1198 * do_swap_page and try_to_unuse repeatedly compete.
1200 wait_on_page_locked(page);
1201 wait_on_page_writeback(page);
1203 wait_on_page_writeback(page);
1206 * Remove all references to entry.
1208 swcount = *swap_map;
1209 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1210 retval = shmem_unuse(entry, page);
1211 /* page has already been unlocked and released */
1216 if (swap_count(swcount) && start_mm != &init_mm)
1217 retval = unuse_mm(start_mm, entry, page);
1219 if (swap_count(*swap_map)) {
1220 int set_start_mm = (*swap_map >= swcount);
1221 struct list_head *p = &start_mm->mmlist;
1222 struct mm_struct *new_start_mm = start_mm;
1223 struct mm_struct *prev_mm = start_mm;
1224 struct mm_struct *mm;
1226 atomic_inc(&new_start_mm->mm_users);
1227 atomic_inc(&prev_mm->mm_users);
1228 spin_lock(&mmlist_lock);
1229 while (swap_count(*swap_map) && !retval &&
1230 (p = p->next) != &start_mm->mmlist) {
1231 mm = list_entry(p, struct mm_struct, mmlist);
1232 if (!atomic_inc_not_zero(&mm->mm_users))
1234 spin_unlock(&mmlist_lock);
1240 swcount = *swap_map;
1241 if (!swap_count(swcount)) /* any usage ? */
1243 else if (mm == &init_mm)
1246 retval = unuse_mm(mm, entry, page);
1248 if (set_start_mm && *swap_map < swcount) {
1249 mmput(new_start_mm);
1250 atomic_inc(&mm->mm_users);
1254 spin_lock(&mmlist_lock);
1256 spin_unlock(&mmlist_lock);
1259 start_mm = new_start_mm;
1263 page_cache_release(page);
1268 * If a reference remains (rare), we would like to leave
1269 * the page in the swap cache; but try_to_unmap could
1270 * then re-duplicate the entry once we drop page lock,
1271 * so we might loop indefinitely; also, that page could
1272 * not be swapped out to other storage meanwhile. So:
1273 * delete from cache even if there's another reference,
1274 * after ensuring that the data has been saved to disk -
1275 * since if the reference remains (rarer), it will be
1276 * read from disk into another page. Splitting into two
1277 * pages would be incorrect if swap supported "shared
1278 * private" pages, but they are handled by tmpfs files.
1280 * Given how unuse_vma() targets one particular offset
1281 * in an anon_vma, once the anon_vma has been determined,
1282 * this splitting happens to be just what is needed to
1283 * handle where KSM pages have been swapped out: re-reading
1284 * is unnecessarily slow, but we can fix that later on.
1286 if (swap_count(*swap_map) &&
1287 PageDirty(page) && PageSwapCache(page)) {
1288 struct writeback_control wbc = {
1289 .sync_mode = WB_SYNC_NONE,
1292 swap_writepage(page, &wbc);
1294 wait_on_page_writeback(page);
1298 * It is conceivable that a racing task removed this page from
1299 * swap cache just before we acquired the page lock at the top,
1300 * or while we dropped it in unuse_mm(). The page might even
1301 * be back in swap cache on another swap area: that we must not
1302 * delete, since it may not have been written out to swap yet.
1304 if (PageSwapCache(page) &&
1305 likely(page_private(page) == entry.val))
1306 delete_from_swap_cache(page);
1309 * So we could skip searching mms once swap count went
1310 * to 1, we did not mark any present ptes as dirty: must
1311 * mark page dirty so shrink_page_list will preserve it.
1315 page_cache_release(page);
1318 * Make sure that we aren't completely killing
1319 * interactive performance.
1329 * After a successful try_to_unuse, if no swap is now in use, we know
1330 * we can empty the mmlist. swap_lock must be held on entry and exit.
1331 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1332 * added to the mmlist just after page_duplicate - before would be racy.
1334 static void drain_mmlist(void)
1336 struct list_head *p, *next;
1339 for (type = 0; type < nr_swapfiles; type++)
1340 if (swap_info[type]->inuse_pages)
1342 spin_lock(&mmlist_lock);
1343 list_for_each_safe(p, next, &init_mm.mmlist)
1345 spin_unlock(&mmlist_lock);
1349 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1350 * corresponds to page offset for the specified swap entry.
1351 * Note that the type of this function is sector_t, but it returns page offset
1352 * into the bdev, not sector offset.
1354 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1356 struct swap_info_struct *sis;
1357 struct swap_extent *start_se;
1358 struct swap_extent *se;
1361 sis = swap_info[swp_type(entry)];
1364 offset = swp_offset(entry);
1365 start_se = sis->curr_swap_extent;
1369 struct list_head *lh;
1371 if (se->start_page <= offset &&
1372 offset < (se->start_page + se->nr_pages)) {
1373 return se->start_block + (offset - se->start_page);
1376 se = list_entry(lh, struct swap_extent, list);
1377 sis->curr_swap_extent = se;
1378 BUG_ON(se == start_se); /* It *must* be present */
1383 * Returns the page offset into bdev for the specified page's swap entry.
1385 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1388 entry.val = page_private(page);
1389 return map_swap_entry(entry, bdev);
1393 * Free all of a swapdev's extent information
1395 static void destroy_swap_extents(struct swap_info_struct *sis)
1397 while (!list_empty(&sis->first_swap_extent.list)) {
1398 struct swap_extent *se;
1400 se = list_entry(sis->first_swap_extent.list.next,
1401 struct swap_extent, list);
1402 list_del(&se->list);
1408 * Add a block range (and the corresponding page range) into this swapdev's
1409 * extent list. The extent list is kept sorted in page order.
1411 * This function rather assumes that it is called in ascending page order.
1414 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1415 unsigned long nr_pages, sector_t start_block)
1417 struct swap_extent *se;
1418 struct swap_extent *new_se;
1419 struct list_head *lh;
1421 if (start_page == 0) {
1422 se = &sis->first_swap_extent;
1423 sis->curr_swap_extent = se;
1425 se->nr_pages = nr_pages;
1426 se->start_block = start_block;
1429 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1430 se = list_entry(lh, struct swap_extent, list);
1431 BUG_ON(se->start_page + se->nr_pages != start_page);
1432 if (se->start_block + se->nr_pages == start_block) {
1434 se->nr_pages += nr_pages;
1440 * No merge. Insert a new extent, preserving ordering.
1442 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1445 new_se->start_page = start_page;
1446 new_se->nr_pages = nr_pages;
1447 new_se->start_block = start_block;
1449 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1454 * A `swap extent' is a simple thing which maps a contiguous range of pages
1455 * onto a contiguous range of disk blocks. An ordered list of swap extents
1456 * is built at swapon time and is then used at swap_writepage/swap_readpage
1457 * time for locating where on disk a page belongs.
1459 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1460 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1461 * swap files identically.
1463 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1464 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1465 * swapfiles are handled *identically* after swapon time.
1467 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1468 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1469 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1470 * requirements, they are simply tossed out - we will never use those blocks
1473 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1474 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1475 * which will scribble on the fs.
1477 * The amount of disk space which a single swap extent represents varies.
1478 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1479 * extents in the list. To avoid much list walking, we cache the previous
1480 * search location in `curr_swap_extent', and start new searches from there.
1481 * This is extremely effective. The average number of iterations in
1482 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1484 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1486 struct inode *inode;
1487 unsigned blocks_per_page;
1488 unsigned long page_no;
1490 sector_t probe_block;
1491 sector_t last_block;
1492 sector_t lowest_block = -1;
1493 sector_t highest_block = 0;
1497 inode = sis->swap_file->f_mapping->host;
1498 if (S_ISBLK(inode->i_mode)) {
1499 ret = add_swap_extent(sis, 0, sis->max, 0);
1504 blkbits = inode->i_blkbits;
1505 blocks_per_page = PAGE_SIZE >> blkbits;
1508 * Map all the blocks into the extent list. This code doesn't try
1513 last_block = i_size_read(inode) >> blkbits;
1514 while ((probe_block + blocks_per_page) <= last_block &&
1515 page_no < sis->max) {
1516 unsigned block_in_page;
1517 sector_t first_block;
1519 first_block = bmap(inode, probe_block);
1520 if (first_block == 0)
1524 * It must be PAGE_SIZE aligned on-disk
1526 if (first_block & (blocks_per_page - 1)) {
1531 for (block_in_page = 1; block_in_page < blocks_per_page;
1535 block = bmap(inode, probe_block + block_in_page);
1538 if (block != first_block + block_in_page) {
1545 first_block >>= (PAGE_SHIFT - blkbits);
1546 if (page_no) { /* exclude the header page */
1547 if (first_block < lowest_block)
1548 lowest_block = first_block;
1549 if (first_block > highest_block)
1550 highest_block = first_block;
1554 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1556 ret = add_swap_extent(sis, page_no, 1, first_block);
1561 probe_block += blocks_per_page;
1566 *span = 1 + highest_block - lowest_block;
1568 page_no = 1; /* force Empty message */
1570 sis->pages = page_no - 1;
1571 sis->highest_bit = page_no - 1;
1575 printk(KERN_ERR "swapon: swapfile has holes\n");
1580 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1582 struct swap_info_struct *p = NULL;
1583 unsigned char *swap_map;
1584 struct file *swap_file, *victim;
1585 struct address_space *mapping;
1586 struct inode *inode;
1591 if (!capable(CAP_SYS_ADMIN))
1594 pathname = getname(specialfile);
1595 err = PTR_ERR(pathname);
1596 if (IS_ERR(pathname))
1599 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1601 err = PTR_ERR(victim);
1605 mapping = victim->f_mapping;
1607 spin_lock(&swap_lock);
1608 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1609 p = swap_info[type];
1610 if (p->flags & SWP_WRITEOK) {
1611 if (p->swap_file->f_mapping == mapping)
1618 spin_unlock(&swap_lock);
1621 if (!security_vm_enough_memory(p->pages))
1622 vm_unacct_memory(p->pages);
1625 spin_unlock(&swap_lock);
1629 swap_list.head = p->next;
1631 swap_info[prev]->next = p->next;
1632 if (type == swap_list.next) {
1633 /* just pick something that's safe... */
1634 swap_list.next = swap_list.head;
1637 for (i = p->next; i >= 0; i = swap_info[i]->next)
1638 swap_info[i]->prio = p->prio--;
1641 nr_swap_pages -= p->pages;
1642 total_swap_pages -= p->pages;
1643 p->flags &= ~SWP_WRITEOK;
1644 spin_unlock(&swap_lock);
1646 current->flags |= PF_OOM_ORIGIN;
1647 err = try_to_unuse(type);
1648 current->flags &= ~PF_OOM_ORIGIN;
1651 /* re-insert swap space back into swap_list */
1652 spin_lock(&swap_lock);
1654 p->prio = --least_priority;
1656 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1657 if (p->prio >= swap_info[i]->prio)
1663 swap_list.head = swap_list.next = type;
1665 swap_info[prev]->next = type;
1666 nr_swap_pages += p->pages;
1667 total_swap_pages += p->pages;
1668 p->flags |= SWP_WRITEOK;
1669 spin_unlock(&swap_lock);
1673 /* wait for any unplug function to finish */
1674 down_write(&swap_unplug_sem);
1675 up_write(&swap_unplug_sem);
1677 destroy_swap_extents(p);
1678 if (p->flags & SWP_CONTINUED)
1679 free_swap_count_continuations(p);
1681 mutex_lock(&swapon_mutex);
1682 spin_lock(&swap_lock);
1685 /* wait for anyone still in scan_swap_map */
1686 p->highest_bit = 0; /* cuts scans short */
1687 while (p->flags >= SWP_SCANNING) {
1688 spin_unlock(&swap_lock);
1689 schedule_timeout_uninterruptible(1);
1690 spin_lock(&swap_lock);
1693 swap_file = p->swap_file;
1694 p->swap_file = NULL;
1696 swap_map = p->swap_map;
1699 spin_unlock(&swap_lock);
1700 mutex_unlock(&swapon_mutex);
1702 /* Destroy swap account informatin */
1703 swap_cgroup_swapoff(type);
1705 inode = mapping->host;
1706 if (S_ISBLK(inode->i_mode)) {
1707 struct block_device *bdev = I_BDEV(inode);
1708 set_blocksize(bdev, p->old_block_size);
1711 mutex_lock(&inode->i_mutex);
1712 inode->i_flags &= ~S_SWAPFILE;
1713 mutex_unlock(&inode->i_mutex);
1715 filp_close(swap_file, NULL);
1719 filp_close(victim, NULL);
1724 #ifdef CONFIG_PROC_FS
1726 static void *swap_start(struct seq_file *swap, loff_t *pos)
1728 struct swap_info_struct *si;
1732 mutex_lock(&swapon_mutex);
1735 return SEQ_START_TOKEN;
1737 for (type = 0; type < nr_swapfiles; type++) {
1738 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1739 si = swap_info[type];
1740 if (!(si->flags & SWP_USED) || !si->swap_map)
1749 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1751 struct swap_info_struct *si = v;
1754 if (v == SEQ_START_TOKEN)
1757 type = si->type + 1;
1759 for (; type < nr_swapfiles; type++) {
1760 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1761 si = swap_info[type];
1762 if (!(si->flags & SWP_USED) || !si->swap_map)
1771 static void swap_stop(struct seq_file *swap, void *v)
1773 mutex_unlock(&swapon_mutex);
1776 static int swap_show(struct seq_file *swap, void *v)
1778 struct swap_info_struct *si = v;
1782 if (si == SEQ_START_TOKEN) {
1783 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1787 file = si->swap_file;
1788 len = seq_path(swap, &file->f_path, " \t\n\\");
1789 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1790 len < 40 ? 40 - len : 1, " ",
1791 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1792 "partition" : "file\t",
1793 si->pages << (PAGE_SHIFT - 10),
1794 si->inuse_pages << (PAGE_SHIFT - 10),
1799 static const struct seq_operations swaps_op = {
1800 .start = swap_start,
1806 static int swaps_open(struct inode *inode, struct file *file)
1808 return seq_open(file, &swaps_op);
1811 static const struct file_operations proc_swaps_operations = {
1814 .llseek = seq_lseek,
1815 .release = seq_release,
1818 static int __init procswaps_init(void)
1820 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1823 __initcall(procswaps_init);
1824 #endif /* CONFIG_PROC_FS */
1826 #ifdef MAX_SWAPFILES_CHECK
1827 static int __init max_swapfiles_check(void)
1829 MAX_SWAPFILES_CHECK();
1832 late_initcall(max_swapfiles_check);
1836 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1838 * The swapon system call
1840 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1842 struct swap_info_struct *p;
1844 struct block_device *bdev = NULL;
1845 struct file *swap_file = NULL;
1846 struct address_space *mapping;
1850 union swap_header *swap_header;
1851 unsigned int nr_good_pages;
1854 unsigned long maxpages;
1855 unsigned long swapfilepages;
1856 unsigned char *swap_map = NULL;
1857 struct page *page = NULL;
1858 struct inode *inode = NULL;
1861 if (!capable(CAP_SYS_ADMIN))
1864 p = kzalloc(sizeof(*p), GFP_KERNEL);
1868 spin_lock(&swap_lock);
1869 for (type = 0; type < nr_swapfiles; type++) {
1870 if (!(swap_info[type]->flags & SWP_USED))
1874 if (type >= MAX_SWAPFILES) {
1875 spin_unlock(&swap_lock);
1879 if (type >= nr_swapfiles) {
1881 swap_info[type] = p;
1883 * Write swap_info[type] before nr_swapfiles, in case a
1884 * racing procfs swap_start() or swap_next() is reading them.
1885 * (We never shrink nr_swapfiles, we never free this entry.)
1891 p = swap_info[type];
1893 * Do not memset this entry: a racing procfs swap_next()
1894 * would be relying on p->type to remain valid.
1897 INIT_LIST_HEAD(&p->first_swap_extent.list);
1898 p->flags = SWP_USED;
1900 spin_unlock(&swap_lock);
1902 name = getname(specialfile);
1903 error = PTR_ERR(name);
1908 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1909 error = PTR_ERR(swap_file);
1910 if (IS_ERR(swap_file)) {
1915 p->swap_file = swap_file;
1916 mapping = swap_file->f_mapping;
1917 inode = mapping->host;
1920 for (i = 0; i < nr_swapfiles; i++) {
1921 struct swap_info_struct *q = swap_info[i];
1923 if (i == type || !q->swap_file)
1925 if (mapping == q->swap_file->f_mapping)
1930 if (S_ISBLK(inode->i_mode)) {
1931 bdev = I_BDEV(inode);
1932 error = bd_claim(bdev, sys_swapon);
1938 p->old_block_size = block_size(bdev);
1939 error = set_blocksize(bdev, PAGE_SIZE);
1943 p->flags |= SWP_BLKDEV;
1944 } else if (S_ISREG(inode->i_mode)) {
1945 p->bdev = inode->i_sb->s_bdev;
1946 mutex_lock(&inode->i_mutex);
1948 if (IS_SWAPFILE(inode)) {
1956 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1959 * Read the swap header.
1961 if (!mapping->a_ops->readpage) {
1965 page = read_mapping_page(mapping, 0, swap_file);
1967 error = PTR_ERR(page);
1970 swap_header = kmap(page);
1972 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1973 printk(KERN_ERR "Unable to find swap-space signature\n");
1978 /* swap partition endianess hack... */
1979 if (swab32(swap_header->info.version) == 1) {
1980 swab32s(&swap_header->info.version);
1981 swab32s(&swap_header->info.last_page);
1982 swab32s(&swap_header->info.nr_badpages);
1983 for (i = 0; i < swap_header->info.nr_badpages; i++)
1984 swab32s(&swap_header->info.badpages[i]);
1986 /* Check the swap header's sub-version */
1987 if (swap_header->info.version != 1) {
1989 "Unable to handle swap header version %d\n",
1990 swap_header->info.version);
1996 p->cluster_next = 1;
2000 * Find out how many pages are allowed for a single swap
2001 * device. There are two limiting factors: 1) the number of
2002 * bits for the swap offset in the swp_entry_t type and
2003 * 2) the number of bits in the a swap pte as defined by
2004 * the different architectures. In order to find the
2005 * largest possible bit mask a swap entry with swap type 0
2006 * and swap offset ~0UL is created, encoded to a swap pte,
2007 * decoded to a swp_entry_t again and finally the swap
2008 * offset is extracted. This will mask all the bits from
2009 * the initial ~0UL mask that can't be encoded in either
2010 * the swp_entry_t or the architecture definition of a
2013 maxpages = swp_offset(pte_to_swp_entry(
2014 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2015 if (maxpages > swap_header->info.last_page) {
2016 maxpages = swap_header->info.last_page + 1;
2017 /* p->max is an unsigned int: don't overflow it */
2018 if ((unsigned int)maxpages == 0)
2019 maxpages = UINT_MAX;
2021 p->highest_bit = maxpages - 1;
2026 if (swapfilepages && maxpages > swapfilepages) {
2028 "Swap area shorter than signature indicates\n");
2031 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2033 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2036 /* OK, set up the swap map and apply the bad block list */
2037 swap_map = vmalloc(maxpages);
2043 memset(swap_map, 0, maxpages);
2044 nr_good_pages = maxpages - 1; /* omit header page */
2046 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2047 unsigned int page_nr = swap_header->info.badpages[i];
2048 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2052 if (page_nr < maxpages) {
2053 swap_map[page_nr] = SWAP_MAP_BAD;
2058 error = swap_cgroup_swapon(type, maxpages);
2062 if (nr_good_pages) {
2063 swap_map[0] = SWAP_MAP_BAD;
2065 p->pages = nr_good_pages;
2066 nr_extents = setup_swap_extents(p, &span);
2067 if (nr_extents < 0) {
2071 nr_good_pages = p->pages;
2073 if (!nr_good_pages) {
2074 printk(KERN_WARNING "Empty swap-file\n");
2080 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2081 p->flags |= SWP_SOLIDSTATE;
2082 p->cluster_next = 1 + (random32() % p->highest_bit);
2084 if (discard_swap(p) == 0)
2085 p->flags |= SWP_DISCARDABLE;
2088 mutex_lock(&swapon_mutex);
2089 spin_lock(&swap_lock);
2090 if (swap_flags & SWAP_FLAG_PREFER)
2092 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2094 p->prio = --least_priority;
2095 p->swap_map = swap_map;
2096 p->flags |= SWP_WRITEOK;
2097 nr_swap_pages += nr_good_pages;
2098 total_swap_pages += nr_good_pages;
2100 printk(KERN_INFO "Adding %uk swap on %s. "
2101 "Priority:%d extents:%d across:%lluk %s%s\n",
2102 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2103 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2104 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2105 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2107 /* insert swap space into swap_list: */
2109 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2110 if (p->prio >= swap_info[i]->prio)
2116 swap_list.head = swap_list.next = type;
2118 swap_info[prev]->next = type;
2119 spin_unlock(&swap_lock);
2120 mutex_unlock(&swapon_mutex);
2125 set_blocksize(bdev, p->old_block_size);
2128 destroy_swap_extents(p);
2129 swap_cgroup_swapoff(type);
2131 spin_lock(&swap_lock);
2132 p->swap_file = NULL;
2134 spin_unlock(&swap_lock);
2137 filp_close(swap_file, NULL);
2139 if (page && !IS_ERR(page)) {
2141 page_cache_release(page);
2147 inode->i_flags |= S_SWAPFILE;
2148 mutex_unlock(&inode->i_mutex);
2153 void si_swapinfo(struct sysinfo *val)
2156 unsigned long nr_to_be_unused = 0;
2158 spin_lock(&swap_lock);
2159 for (type = 0; type < nr_swapfiles; type++) {
2160 struct swap_info_struct *si = swap_info[type];
2162 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2163 nr_to_be_unused += si->inuse_pages;
2165 val->freeswap = nr_swap_pages + nr_to_be_unused;
2166 val->totalswap = total_swap_pages + nr_to_be_unused;
2167 spin_unlock(&swap_lock);
2171 * Verify that a swap entry is valid and increment its swap map count.
2173 * Returns error code in following case.
2175 * - swp_entry is invalid -> EINVAL
2176 * - swp_entry is migration entry -> EINVAL
2177 * - swap-cache reference is requested but there is already one. -> EEXIST
2178 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2179 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2181 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2183 struct swap_info_struct *p;
2184 unsigned long offset, type;
2185 unsigned char count;
2186 unsigned char has_cache;
2189 if (non_swap_entry(entry))
2192 type = swp_type(entry);
2193 if (type >= nr_swapfiles)
2195 p = swap_info[type];
2196 offset = swp_offset(entry);
2198 spin_lock(&swap_lock);
2199 if (unlikely(offset >= p->max))
2202 count = p->swap_map[offset];
2203 has_cache = count & SWAP_HAS_CACHE;
2204 count &= ~SWAP_HAS_CACHE;
2207 if (usage == SWAP_HAS_CACHE) {
2209 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2210 if (!has_cache && count)
2211 has_cache = SWAP_HAS_CACHE;
2212 else if (has_cache) /* someone else added cache */
2214 else /* no users remaining */
2217 } else if (count || has_cache) {
2219 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2221 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2223 else if (swap_count_continued(p, offset, count))
2224 count = COUNT_CONTINUED;
2228 err = -ENOENT; /* unused swap entry */
2230 p->swap_map[offset] = count | has_cache;
2233 spin_unlock(&swap_lock);
2238 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2243 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2244 * (in which case its reference count is never incremented).
2246 void swap_shmem_alloc(swp_entry_t entry)
2248 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2252 * Increase reference count of swap entry by 1.
2253 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2254 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2255 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2256 * might occur if a page table entry has got corrupted.
2258 int swap_duplicate(swp_entry_t entry)
2262 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2263 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2268 * @entry: swap entry for which we allocate swap cache.
2270 * Called when allocating swap cache for existing swap entry,
2271 * This can return error codes. Returns 0 at success.
2272 * -EBUSY means there is a swap cache.
2273 * Note: return code is different from swap_duplicate().
2275 int swapcache_prepare(swp_entry_t entry)
2277 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2281 * swap_lock prevents swap_map being freed. Don't grab an extra
2282 * reference on the swaphandle, it doesn't matter if it becomes unused.
2284 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2286 struct swap_info_struct *si;
2287 int our_page_cluster = page_cluster;
2288 pgoff_t target, toff;
2292 if (!our_page_cluster) /* no readahead */
2295 si = swap_info[swp_type(entry)];
2296 target = swp_offset(entry);
2297 base = (target >> our_page_cluster) << our_page_cluster;
2298 end = base + (1 << our_page_cluster);
2299 if (!base) /* first page is swap header */
2302 spin_lock(&swap_lock);
2303 if (end > si->max) /* don't go beyond end of map */
2306 /* Count contiguous allocated slots above our target */
2307 for (toff = target; ++toff < end; nr_pages++) {
2308 /* Don't read in free or bad pages */
2309 if (!si->swap_map[toff])
2311 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2314 /* Count contiguous allocated slots below our target */
2315 for (toff = target; --toff >= base; nr_pages++) {
2316 /* Don't read in free or bad pages */
2317 if (!si->swap_map[toff])
2319 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2322 spin_unlock(&swap_lock);
2325 * Indicate starting offset, and return number of pages to get:
2326 * if only 1, say 0, since there's then no readahead to be done.
2329 return nr_pages? ++nr_pages: 0;
2333 * add_swap_count_continuation - called when a swap count is duplicated
2334 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2335 * page of the original vmalloc'ed swap_map, to hold the continuation count
2336 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2337 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2339 * These continuation pages are seldom referenced: the common paths all work
2340 * on the original swap_map, only referring to a continuation page when the
2341 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2343 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2344 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2345 * can be called after dropping locks.
2347 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2349 struct swap_info_struct *si;
2352 struct page *list_page;
2354 unsigned char count;
2357 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2358 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2360 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2362 si = swap_info_get(entry);
2365 * An acceptable race has occurred since the failing
2366 * __swap_duplicate(): the swap entry has been freed,
2367 * perhaps even the whole swap_map cleared for swapoff.
2372 offset = swp_offset(entry);
2373 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2375 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2377 * The higher the swap count, the more likely it is that tasks
2378 * will race to add swap count continuation: we need to avoid
2379 * over-provisioning.
2385 spin_unlock(&swap_lock);
2390 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2391 * no architecture is using highmem pages for kernel pagetables: so it
2392 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2394 head = vmalloc_to_page(si->swap_map + offset);
2395 offset &= ~PAGE_MASK;
2398 * Page allocation does not initialize the page's lru field,
2399 * but it does always reset its private field.
2401 if (!page_private(head)) {
2402 BUG_ON(count & COUNT_CONTINUED);
2403 INIT_LIST_HEAD(&head->lru);
2404 set_page_private(head, SWP_CONTINUED);
2405 si->flags |= SWP_CONTINUED;
2408 list_for_each_entry(list_page, &head->lru, lru) {
2412 * If the previous map said no continuation, but we've found
2413 * a continuation page, free our allocation and use this one.
2415 if (!(count & COUNT_CONTINUED))
2418 map = kmap_atomic(list_page, KM_USER0) + offset;
2420 kunmap_atomic(map, KM_USER0);
2423 * If this continuation count now has some space in it,
2424 * free our allocation and use this one.
2426 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2430 list_add_tail(&page->lru, &head->lru);
2431 page = NULL; /* now it's attached, don't free it */
2433 spin_unlock(&swap_lock);
2441 * swap_count_continued - when the original swap_map count is incremented
2442 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2443 * into, carry if so, or else fail until a new continuation page is allocated;
2444 * when the original swap_map count is decremented from 0 with continuation,
2445 * borrow from the continuation and report whether it still holds more.
2446 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2448 static bool swap_count_continued(struct swap_info_struct *si,
2449 pgoff_t offset, unsigned char count)
2455 head = vmalloc_to_page(si->swap_map + offset);
2456 if (page_private(head) != SWP_CONTINUED) {
2457 BUG_ON(count & COUNT_CONTINUED);
2458 return false; /* need to add count continuation */
2461 offset &= ~PAGE_MASK;
2462 page = list_entry(head->lru.next, struct page, lru);
2463 map = kmap_atomic(page, KM_USER0) + offset;
2465 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2466 goto init_map; /* jump over SWAP_CONT_MAX checks */
2468 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2470 * Think of how you add 1 to 999
2472 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2473 kunmap_atomic(map, KM_USER0);
2474 page = list_entry(page->lru.next, struct page, lru);
2475 BUG_ON(page == head);
2476 map = kmap_atomic(page, KM_USER0) + offset;
2478 if (*map == SWAP_CONT_MAX) {
2479 kunmap_atomic(map, KM_USER0);
2480 page = list_entry(page->lru.next, struct page, lru);
2482 return false; /* add count continuation */
2483 map = kmap_atomic(page, KM_USER0) + offset;
2484 init_map: *map = 0; /* we didn't zero the page */
2487 kunmap_atomic(map, KM_USER0);
2488 page = list_entry(page->lru.prev, struct page, lru);
2489 while (page != head) {
2490 map = kmap_atomic(page, KM_USER0) + offset;
2491 *map = COUNT_CONTINUED;
2492 kunmap_atomic(map, KM_USER0);
2493 page = list_entry(page->lru.prev, struct page, lru);
2495 return true; /* incremented */
2497 } else { /* decrementing */
2499 * Think of how you subtract 1 from 1000
2501 BUG_ON(count != COUNT_CONTINUED);
2502 while (*map == COUNT_CONTINUED) {
2503 kunmap_atomic(map, KM_USER0);
2504 page = list_entry(page->lru.next, struct page, lru);
2505 BUG_ON(page == head);
2506 map = kmap_atomic(page, KM_USER0) + offset;
2512 kunmap_atomic(map, KM_USER0);
2513 page = list_entry(page->lru.prev, struct page, lru);
2514 while (page != head) {
2515 map = kmap_atomic(page, KM_USER0) + offset;
2516 *map = SWAP_CONT_MAX | count;
2517 count = COUNT_CONTINUED;
2518 kunmap_atomic(map, KM_USER0);
2519 page = list_entry(page->lru.prev, struct page, lru);
2521 return count == COUNT_CONTINUED;
2526 * free_swap_count_continuations - swapoff free all the continuation pages
2527 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2529 static void free_swap_count_continuations(struct swap_info_struct *si)
2533 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2535 head = vmalloc_to_page(si->swap_map + offset);
2536 if (page_private(head)) {
2537 struct list_head *this, *next;
2538 list_for_each_safe(this, next, &head->lru) {
2540 page = list_entry(this, struct page, lru);