The current implementation of the /dev/hpet driver couples opening the
[linux-2.6/next.git] / mm / swapfile.c
blob630c33db4bff560770a12fbe9734086e6de73e81
1 /*
2 * linux/mm/swapfile.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
6 */
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shmem_fs.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32 #include <linux/poll.h>
33 #include <linux/oom.h>
34 #include <linux/frontswap.h>
35 #include <linux/swapfile.h>
37 #include <asm/pgtable.h>
38 #include <asm/tlbflush.h>
39 #include <linux/swapops.h>
40 #include <linux/page_cgroup.h>
42 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
43 unsigned char);
44 static void free_swap_count_continuations(struct swap_info_struct *);
45 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
47 DEFINE_SPINLOCK(swap_lock);
48 static unsigned int nr_swapfiles;
49 long nr_swap_pages;
50 long total_swap_pages;
51 static int least_priority;
53 static const char Bad_file[] = "Bad swap file entry ";
54 static const char Unused_file[] = "Unused swap file entry ";
55 static const char Bad_offset[] = "Bad swap offset entry ";
56 static const char Unused_offset[] = "Unused swap offset entry ";
58 struct swap_list_t swap_list = {-1, -1};
60 struct swap_info_struct *swap_info[MAX_SWAPFILES];
62 static DEFINE_MUTEX(swapon_mutex);
64 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
65 /* Activity counter to indicate that a swapon or swapoff has occurred */
66 static atomic_t proc_poll_event = ATOMIC_INIT(0);
68 static inline unsigned char swap_count(unsigned char ent)
70 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
73 /* returns 1 if swap entry is freed */
74 static int
75 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
77 swp_entry_t entry = swp_entry(si->type, offset);
78 struct page *page;
79 int ret = 0;
81 page = find_get_page(&swapper_space, entry.val);
82 if (!page)
83 return 0;
85 * This function is called from scan_swap_map() and it's called
86 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
87 * We have to use trylock for avoiding deadlock. This is a special
88 * case and you should use try_to_free_swap() with explicit lock_page()
89 * in usual operations.
91 if (trylock_page(page)) {
92 ret = try_to_free_swap(page);
93 unlock_page(page);
95 page_cache_release(page);
96 return ret;
100 * swapon tell device that all the old swap contents can be discarded,
101 * to allow the swap device to optimize its wear-levelling.
103 static int discard_swap(struct swap_info_struct *si)
105 struct swap_extent *se;
106 sector_t start_block;
107 sector_t nr_blocks;
108 int err = 0;
110 /* Do not discard the swap header page! */
111 se = &si->first_swap_extent;
112 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
113 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
114 if (nr_blocks) {
115 err = blkdev_issue_discard(si->bdev, start_block,
116 nr_blocks, GFP_KERNEL, 0);
117 if (err)
118 return err;
119 cond_resched();
122 list_for_each_entry(se, &si->first_swap_extent.list, list) {
123 start_block = se->start_block << (PAGE_SHIFT - 9);
124 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
126 err = blkdev_issue_discard(si->bdev, start_block,
127 nr_blocks, GFP_KERNEL, 0);
128 if (err)
129 break;
131 cond_resched();
133 return err; /* That will often be -EOPNOTSUPP */
137 * swap allocation tell device that a cluster of swap can now be discarded,
138 * to allow the swap device to optimize its wear-levelling.
140 static void discard_swap_cluster(struct swap_info_struct *si,
141 pgoff_t start_page, pgoff_t nr_pages)
143 struct swap_extent *se = si->curr_swap_extent;
144 int found_extent = 0;
146 while (nr_pages) {
147 struct list_head *lh;
149 if (se->start_page <= start_page &&
150 start_page < se->start_page + se->nr_pages) {
151 pgoff_t offset = start_page - se->start_page;
152 sector_t start_block = se->start_block + offset;
153 sector_t nr_blocks = se->nr_pages - offset;
155 if (nr_blocks > nr_pages)
156 nr_blocks = nr_pages;
157 start_page += nr_blocks;
158 nr_pages -= nr_blocks;
160 if (!found_extent++)
161 si->curr_swap_extent = se;
163 start_block <<= PAGE_SHIFT - 9;
164 nr_blocks <<= PAGE_SHIFT - 9;
165 if (blkdev_issue_discard(si->bdev, start_block,
166 nr_blocks, GFP_NOIO, 0))
167 break;
170 lh = se->list.next;
171 se = list_entry(lh, struct swap_extent, list);
175 static int wait_for_discard(void *word)
177 schedule();
178 return 0;
181 #define SWAPFILE_CLUSTER 256
182 #define LATENCY_LIMIT 256
184 static unsigned long scan_swap_map(struct swap_info_struct *si,
185 unsigned char usage)
187 unsigned long offset;
188 unsigned long scan_base;
189 unsigned long last_in_cluster = 0;
190 int latency_ration = LATENCY_LIMIT;
191 int found_free_cluster = 0;
194 * We try to cluster swap pages by allocating them sequentially
195 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
196 * way, however, we resort to first-free allocation, starting
197 * a new cluster. This prevents us from scattering swap pages
198 * all over the entire swap partition, so that we reduce
199 * overall disk seek times between swap pages. -- sct
200 * But we do now try to find an empty cluster. -Andrea
201 * And we let swap pages go all over an SSD partition. Hugh
204 si->flags += SWP_SCANNING;
205 scan_base = offset = si->cluster_next;
207 if (unlikely(!si->cluster_nr--)) {
208 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
209 si->cluster_nr = SWAPFILE_CLUSTER - 1;
210 goto checks;
212 if (si->flags & SWP_DISCARDABLE) {
214 * Start range check on racing allocations, in case
215 * they overlap the cluster we eventually decide on
216 * (we scan without swap_lock to allow preemption).
217 * It's hardly conceivable that cluster_nr could be
218 * wrapped during our scan, but don't depend on it.
220 if (si->lowest_alloc)
221 goto checks;
222 si->lowest_alloc = si->max;
223 si->highest_alloc = 0;
225 spin_unlock(&swap_lock);
228 * If seek is expensive, start searching for new cluster from
229 * start of partition, to minimize the span of allocated swap.
230 * But if seek is cheap, search from our current position, so
231 * that swap is allocated from all over the partition: if the
232 * Flash Translation Layer only remaps within limited zones,
233 * we don't want to wear out the first zone too quickly.
235 if (!(si->flags & SWP_SOLIDSTATE))
236 scan_base = offset = si->lowest_bit;
237 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
239 /* Locate the first empty (unaligned) cluster */
240 for (; last_in_cluster <= si->highest_bit; offset++) {
241 if (si->swap_map[offset])
242 last_in_cluster = offset + SWAPFILE_CLUSTER;
243 else if (offset == last_in_cluster) {
244 spin_lock(&swap_lock);
245 offset -= SWAPFILE_CLUSTER - 1;
246 si->cluster_next = offset;
247 si->cluster_nr = SWAPFILE_CLUSTER - 1;
248 found_free_cluster = 1;
249 goto checks;
251 if (unlikely(--latency_ration < 0)) {
252 cond_resched();
253 latency_ration = LATENCY_LIMIT;
257 offset = si->lowest_bit;
258 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
260 /* Locate the first empty (unaligned) cluster */
261 for (; last_in_cluster < scan_base; offset++) {
262 if (si->swap_map[offset])
263 last_in_cluster = offset + SWAPFILE_CLUSTER;
264 else if (offset == last_in_cluster) {
265 spin_lock(&swap_lock);
266 offset -= SWAPFILE_CLUSTER - 1;
267 si->cluster_next = offset;
268 si->cluster_nr = SWAPFILE_CLUSTER - 1;
269 found_free_cluster = 1;
270 goto checks;
272 if (unlikely(--latency_ration < 0)) {
273 cond_resched();
274 latency_ration = LATENCY_LIMIT;
278 offset = scan_base;
279 spin_lock(&swap_lock);
280 si->cluster_nr = SWAPFILE_CLUSTER - 1;
281 si->lowest_alloc = 0;
284 checks:
285 if (!(si->flags & SWP_WRITEOK))
286 goto no_page;
287 if (!si->highest_bit)
288 goto no_page;
289 if (offset > si->highest_bit)
290 scan_base = offset = si->lowest_bit;
292 /* reuse swap entry of cache-only swap if not busy. */
293 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
294 int swap_was_freed;
295 spin_unlock(&swap_lock);
296 swap_was_freed = __try_to_reclaim_swap(si, offset);
297 spin_lock(&swap_lock);
298 /* entry was freed successfully, try to use this again */
299 if (swap_was_freed)
300 goto checks;
301 goto scan; /* check next one */
304 if (si->swap_map[offset])
305 goto scan;
307 if (offset == si->lowest_bit)
308 si->lowest_bit++;
309 if (offset == si->highest_bit)
310 si->highest_bit--;
311 si->inuse_pages++;
312 if (si->inuse_pages == si->pages) {
313 si->lowest_bit = si->max;
314 si->highest_bit = 0;
316 si->swap_map[offset] = usage;
317 si->cluster_next = offset + 1;
318 si->flags -= SWP_SCANNING;
320 if (si->lowest_alloc) {
322 * Only set when SWP_DISCARDABLE, and there's a scan
323 * for a free cluster in progress or just completed.
325 if (found_free_cluster) {
327 * To optimize wear-levelling, discard the
328 * old data of the cluster, taking care not to
329 * discard any of its pages that have already
330 * been allocated by racing tasks (offset has
331 * already stepped over any at the beginning).
333 if (offset < si->highest_alloc &&
334 si->lowest_alloc <= last_in_cluster)
335 last_in_cluster = si->lowest_alloc - 1;
336 si->flags |= SWP_DISCARDING;
337 spin_unlock(&swap_lock);
339 if (offset < last_in_cluster)
340 discard_swap_cluster(si, offset,
341 last_in_cluster - offset + 1);
343 spin_lock(&swap_lock);
344 si->lowest_alloc = 0;
345 si->flags &= ~SWP_DISCARDING;
347 smp_mb(); /* wake_up_bit advises this */
348 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
350 } else if (si->flags & SWP_DISCARDING) {
352 * Delay using pages allocated by racing tasks
353 * until the whole discard has been issued. We
354 * could defer that delay until swap_writepage,
355 * but it's easier to keep this self-contained.
357 spin_unlock(&swap_lock);
358 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
359 wait_for_discard, TASK_UNINTERRUPTIBLE);
360 spin_lock(&swap_lock);
361 } else {
363 * Note pages allocated by racing tasks while
364 * scan for a free cluster is in progress, so
365 * that its final discard can exclude them.
367 if (offset < si->lowest_alloc)
368 si->lowest_alloc = offset;
369 if (offset > si->highest_alloc)
370 si->highest_alloc = offset;
373 return offset;
375 scan:
376 spin_unlock(&swap_lock);
377 while (++offset <= si->highest_bit) {
378 if (!si->swap_map[offset]) {
379 spin_lock(&swap_lock);
380 goto checks;
382 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
383 spin_lock(&swap_lock);
384 goto checks;
386 if (unlikely(--latency_ration < 0)) {
387 cond_resched();
388 latency_ration = LATENCY_LIMIT;
391 offset = si->lowest_bit;
392 while (++offset < scan_base) {
393 if (!si->swap_map[offset]) {
394 spin_lock(&swap_lock);
395 goto checks;
397 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
398 spin_lock(&swap_lock);
399 goto checks;
401 if (unlikely(--latency_ration < 0)) {
402 cond_resched();
403 latency_ration = LATENCY_LIMIT;
406 spin_lock(&swap_lock);
408 no_page:
409 si->flags -= SWP_SCANNING;
410 return 0;
413 swp_entry_t get_swap_page(void)
415 struct swap_info_struct *si;
416 pgoff_t offset;
417 int type, next;
418 int wrapped = 0;
420 spin_lock(&swap_lock);
421 if (nr_swap_pages <= 0)
422 goto noswap;
423 nr_swap_pages--;
425 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
426 si = swap_info[type];
427 next = si->next;
428 if (next < 0 ||
429 (!wrapped && si->prio != swap_info[next]->prio)) {
430 next = swap_list.head;
431 wrapped++;
434 if (!si->highest_bit)
435 continue;
436 if (!(si->flags & SWP_WRITEOK))
437 continue;
439 swap_list.next = next;
440 /* This is called for allocating swap entry for cache */
441 offset = scan_swap_map(si, SWAP_HAS_CACHE);
442 if (offset) {
443 spin_unlock(&swap_lock);
444 return swp_entry(type, offset);
446 next = swap_list.next;
449 nr_swap_pages++;
450 noswap:
451 spin_unlock(&swap_lock);
452 return (swp_entry_t) {0};
455 /* The only caller of this function is now susupend routine */
456 swp_entry_t get_swap_page_of_type(int type)
458 struct swap_info_struct *si;
459 pgoff_t offset;
461 spin_lock(&swap_lock);
462 si = swap_info[type];
463 if (si && (si->flags & SWP_WRITEOK)) {
464 nr_swap_pages--;
465 /* This is called for allocating swap entry, not cache */
466 offset = scan_swap_map(si, 1);
467 if (offset) {
468 spin_unlock(&swap_lock);
469 return swp_entry(type, offset);
471 nr_swap_pages++;
473 spin_unlock(&swap_lock);
474 return (swp_entry_t) {0};
477 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
479 struct swap_info_struct *p;
480 unsigned long offset, type;
482 if (!entry.val)
483 goto out;
484 type = swp_type(entry);
485 if (type >= nr_swapfiles)
486 goto bad_nofile;
487 p = swap_info[type];
488 if (!(p->flags & SWP_USED))
489 goto bad_device;
490 offset = swp_offset(entry);
491 if (offset >= p->max)
492 goto bad_offset;
493 if (!p->swap_map[offset])
494 goto bad_free;
495 spin_lock(&swap_lock);
496 return p;
498 bad_free:
499 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
500 goto out;
501 bad_offset:
502 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
503 goto out;
504 bad_device:
505 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
506 goto out;
507 bad_nofile:
508 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
509 out:
510 return NULL;
513 static unsigned char swap_entry_free(struct swap_info_struct *p,
514 swp_entry_t entry, unsigned char usage)
516 unsigned long offset = swp_offset(entry);
517 unsigned char count;
518 unsigned char has_cache;
520 count = p->swap_map[offset];
521 has_cache = count & SWAP_HAS_CACHE;
522 count &= ~SWAP_HAS_CACHE;
524 if (usage == SWAP_HAS_CACHE) {
525 VM_BUG_ON(!has_cache);
526 has_cache = 0;
527 } else if (count == SWAP_MAP_SHMEM) {
529 * Or we could insist on shmem.c using a special
530 * swap_shmem_free() and free_shmem_swap_and_cache()...
532 count = 0;
533 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
534 if (count == COUNT_CONTINUED) {
535 if (swap_count_continued(p, offset, count))
536 count = SWAP_MAP_MAX | COUNT_CONTINUED;
537 else
538 count = SWAP_MAP_MAX;
539 } else
540 count--;
543 if (!count)
544 mem_cgroup_uncharge_swap(entry);
546 usage = count | has_cache;
547 p->swap_map[offset] = usage;
549 /* free if no reference */
550 if (!usage) {
551 struct gendisk *disk = p->bdev->bd_disk;
552 if (offset < p->lowest_bit)
553 p->lowest_bit = offset;
554 if (offset > p->highest_bit)
555 p->highest_bit = offset;
556 if (swap_list.next >= 0 &&
557 p->prio > swap_info[swap_list.next]->prio)
558 swap_list.next = p->type;
559 nr_swap_pages++;
560 p->inuse_pages--;
561 frontswap_flush_page(p->type, offset);
562 if ((p->flags & SWP_BLKDEV) &&
563 disk->fops->swap_slot_free_notify)
564 disk->fops->swap_slot_free_notify(p->bdev, offset);
567 return usage;
571 * Caller has made sure that the swapdevice corresponding to entry
572 * is still around or has not been recycled.
574 void swap_free(swp_entry_t entry)
576 struct swap_info_struct *p;
578 p = swap_info_get(entry);
579 if (p) {
580 swap_entry_free(p, entry, 1);
581 spin_unlock(&swap_lock);
586 * Called after dropping swapcache to decrease refcnt to swap entries.
588 void swapcache_free(swp_entry_t entry, struct page *page)
590 struct swap_info_struct *p;
591 unsigned char count;
593 p = swap_info_get(entry);
594 if (p) {
595 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
596 if (page)
597 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
598 spin_unlock(&swap_lock);
603 * How many references to page are currently swapped out?
604 * This does not give an exact answer when swap count is continued,
605 * but does include the high COUNT_CONTINUED flag to allow for that.
607 static inline int page_swapcount(struct page *page)
609 int count = 0;
610 struct swap_info_struct *p;
611 swp_entry_t entry;
613 entry.val = page_private(page);
614 p = swap_info_get(entry);
615 if (p) {
616 count = swap_count(p->swap_map[swp_offset(entry)]);
617 spin_unlock(&swap_lock);
619 return count;
623 * We can write to an anon page without COW if there are no other references
624 * to it. And as a side-effect, free up its swap: because the old content
625 * on disk will never be read, and seeking back there to write new content
626 * later would only waste time away from clustering.
628 int reuse_swap_page(struct page *page)
630 int count;
632 VM_BUG_ON(!PageLocked(page));
633 if (unlikely(PageKsm(page)))
634 return 0;
635 count = page_mapcount(page);
636 if (count <= 1 && PageSwapCache(page)) {
637 count += page_swapcount(page);
638 if (count == 1 && !PageWriteback(page)) {
639 delete_from_swap_cache(page);
640 SetPageDirty(page);
643 return count <= 1;
647 * If swap is getting full, or if there are no more mappings of this page,
648 * then try_to_free_swap is called to free its swap space.
650 int try_to_free_swap(struct page *page)
652 VM_BUG_ON(!PageLocked(page));
654 if (!PageSwapCache(page))
655 return 0;
656 if (PageWriteback(page))
657 return 0;
658 if (page_swapcount(page))
659 return 0;
662 * Once hibernation has begun to create its image of memory,
663 * there's a danger that one of the calls to try_to_free_swap()
664 * - most probably a call from __try_to_reclaim_swap() while
665 * hibernation is allocating its own swap pages for the image,
666 * but conceivably even a call from memory reclaim - will free
667 * the swap from a page which has already been recorded in the
668 * image as a clean swapcache page, and then reuse its swap for
669 * another page of the image. On waking from hibernation, the
670 * original page might be freed under memory pressure, then
671 * later read back in from swap, now with the wrong data.
673 * Hibernation clears bits from gfp_allowed_mask to prevent
674 * memory reclaim from writing to disk, so check that here.
676 if (!(gfp_allowed_mask & __GFP_IO))
677 return 0;
679 delete_from_swap_cache(page);
680 SetPageDirty(page);
681 return 1;
685 * Free the swap entry like above, but also try to
686 * free the page cache entry if it is the last user.
688 int free_swap_and_cache(swp_entry_t entry)
690 struct swap_info_struct *p;
691 struct page *page = NULL;
693 if (non_swap_entry(entry))
694 return 1;
696 p = swap_info_get(entry);
697 if (p) {
698 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
699 page = find_get_page(&swapper_space, entry.val);
700 if (page && !trylock_page(page)) {
701 page_cache_release(page);
702 page = NULL;
705 spin_unlock(&swap_lock);
707 if (page) {
709 * Not mapped elsewhere, or swap space full? Free it!
710 * Also recheck PageSwapCache now page is locked (above).
712 if (PageSwapCache(page) && !PageWriteback(page) &&
713 (!page_mapped(page) || vm_swap_full())) {
714 delete_from_swap_cache(page);
715 SetPageDirty(page);
717 unlock_page(page);
718 page_cache_release(page);
720 return p != NULL;
723 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
725 * mem_cgroup_count_swap_user - count the user of a swap entry
726 * @ent: the swap entry to be checked
727 * @pagep: the pointer for the swap cache page of the entry to be stored
729 * Returns the number of the user of the swap entry. The number is valid only
730 * for swaps of anonymous pages.
731 * If the entry is found on swap cache, the page is stored to pagep with
732 * refcount of it being incremented.
734 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
736 struct page *page;
737 struct swap_info_struct *p;
738 int count = 0;
740 page = find_get_page(&swapper_space, ent.val);
741 if (page)
742 count += page_mapcount(page);
743 p = swap_info_get(ent);
744 if (p) {
745 count += swap_count(p->swap_map[swp_offset(ent)]);
746 spin_unlock(&swap_lock);
749 *pagep = page;
750 return count;
752 #endif
754 #ifdef CONFIG_HIBERNATION
756 * Find the swap type that corresponds to given device (if any).
758 * @offset - number of the PAGE_SIZE-sized block of the device, starting
759 * from 0, in which the swap header is expected to be located.
761 * This is needed for the suspend to disk (aka swsusp).
763 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
765 struct block_device *bdev = NULL;
766 int type;
768 if (device)
769 bdev = bdget(device);
771 spin_lock(&swap_lock);
772 for (type = 0; type < nr_swapfiles; type++) {
773 struct swap_info_struct *sis = swap_info[type];
775 if (!(sis->flags & SWP_WRITEOK))
776 continue;
778 if (!bdev) {
779 if (bdev_p)
780 *bdev_p = bdgrab(sis->bdev);
782 spin_unlock(&swap_lock);
783 return type;
785 if (bdev == sis->bdev) {
786 struct swap_extent *se = &sis->first_swap_extent;
788 if (se->start_block == offset) {
789 if (bdev_p)
790 *bdev_p = bdgrab(sis->bdev);
792 spin_unlock(&swap_lock);
793 bdput(bdev);
794 return type;
798 spin_unlock(&swap_lock);
799 if (bdev)
800 bdput(bdev);
802 return -ENODEV;
806 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
807 * corresponding to given index in swap_info (swap type).
809 sector_t swapdev_block(int type, pgoff_t offset)
811 struct block_device *bdev;
813 if ((unsigned int)type >= nr_swapfiles)
814 return 0;
815 if (!(swap_info[type]->flags & SWP_WRITEOK))
816 return 0;
817 return map_swap_entry(swp_entry(type, offset), &bdev);
821 * Return either the total number of swap pages of given type, or the number
822 * of free pages of that type (depending on @free)
824 * This is needed for software suspend
826 unsigned int count_swap_pages(int type, int free)
828 unsigned int n = 0;
830 spin_lock(&swap_lock);
831 if ((unsigned int)type < nr_swapfiles) {
832 struct swap_info_struct *sis = swap_info[type];
834 if (sis->flags & SWP_WRITEOK) {
835 n = sis->pages;
836 if (free)
837 n -= sis->inuse_pages;
840 spin_unlock(&swap_lock);
841 return n;
843 #endif /* CONFIG_HIBERNATION */
846 * No need to decide whether this PTE shares the swap entry with others,
847 * just let do_wp_page work it out if a write is requested later - to
848 * force COW, vm_page_prot omits write permission from any private vma.
850 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
851 unsigned long addr, swp_entry_t entry, struct page *page)
853 struct mem_cgroup *ptr;
854 spinlock_t *ptl;
855 pte_t *pte;
856 int ret = 1;
858 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
859 ret = -ENOMEM;
860 goto out_nolock;
863 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
864 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
865 if (ret > 0)
866 mem_cgroup_cancel_charge_swapin(ptr);
867 ret = 0;
868 goto out;
871 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
872 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
873 get_page(page);
874 set_pte_at(vma->vm_mm, addr, pte,
875 pte_mkold(mk_pte(page, vma->vm_page_prot)));
876 page_add_anon_rmap(page, vma, addr);
877 mem_cgroup_commit_charge_swapin(page, ptr);
878 swap_free(entry);
880 * Move the page to the active list so it is not
881 * immediately swapped out again after swapon.
883 activate_page(page);
884 out:
885 pte_unmap_unlock(pte, ptl);
886 out_nolock:
887 return ret;
890 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
891 unsigned long addr, unsigned long end,
892 swp_entry_t entry, struct page *page)
894 pte_t swp_pte = swp_entry_to_pte(entry);
895 pte_t *pte;
896 int ret = 0;
899 * We don't actually need pte lock while scanning for swp_pte: since
900 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
901 * page table while we're scanning; though it could get zapped, and on
902 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
903 * of unmatched parts which look like swp_pte, so unuse_pte must
904 * recheck under pte lock. Scanning without pte lock lets it be
905 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
907 pte = pte_offset_map(pmd, addr);
908 do {
910 * swapoff spends a _lot_ of time in this loop!
911 * Test inline before going to call unuse_pte.
913 if (unlikely(pte_same(*pte, swp_pte))) {
914 pte_unmap(pte);
915 ret = unuse_pte(vma, pmd, addr, entry, page);
916 if (ret)
917 goto out;
918 pte = pte_offset_map(pmd, addr);
920 } while (pte++, addr += PAGE_SIZE, addr != end);
921 pte_unmap(pte - 1);
922 out:
923 return ret;
926 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
927 unsigned long addr, unsigned long end,
928 swp_entry_t entry, struct page *page)
930 pmd_t *pmd;
931 unsigned long next;
932 int ret;
934 pmd = pmd_offset(pud, addr);
935 do {
936 next = pmd_addr_end(addr, end);
937 if (unlikely(pmd_trans_huge(*pmd)))
938 continue;
939 if (pmd_none_or_clear_bad(pmd))
940 continue;
941 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
942 if (ret)
943 return ret;
944 } while (pmd++, addr = next, addr != end);
945 return 0;
948 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
949 unsigned long addr, unsigned long end,
950 swp_entry_t entry, struct page *page)
952 pud_t *pud;
953 unsigned long next;
954 int ret;
956 pud = pud_offset(pgd, addr);
957 do {
958 next = pud_addr_end(addr, end);
959 if (pud_none_or_clear_bad(pud))
960 continue;
961 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
962 if (ret)
963 return ret;
964 } while (pud++, addr = next, addr != end);
965 return 0;
968 static int unuse_vma(struct vm_area_struct *vma,
969 swp_entry_t entry, struct page *page)
971 pgd_t *pgd;
972 unsigned long addr, end, next;
973 int ret;
975 if (page_anon_vma(page)) {
976 addr = page_address_in_vma(page, vma);
977 if (addr == -EFAULT)
978 return 0;
979 else
980 end = addr + PAGE_SIZE;
981 } else {
982 addr = vma->vm_start;
983 end = vma->vm_end;
986 pgd = pgd_offset(vma->vm_mm, addr);
987 do {
988 next = pgd_addr_end(addr, end);
989 if (pgd_none_or_clear_bad(pgd))
990 continue;
991 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
992 if (ret)
993 return ret;
994 } while (pgd++, addr = next, addr != end);
995 return 0;
998 static int unuse_mm(struct mm_struct *mm,
999 swp_entry_t entry, struct page *page)
1001 struct vm_area_struct *vma;
1002 int ret = 0;
1004 if (!down_read_trylock(&mm->mmap_sem)) {
1006 * Activate page so shrink_inactive_list is unlikely to unmap
1007 * its ptes while lock is dropped, so swapoff can make progress.
1009 activate_page(page);
1010 unlock_page(page);
1011 down_read(&mm->mmap_sem);
1012 lock_page(page);
1014 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1015 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1016 break;
1018 up_read(&mm->mmap_sem);
1019 return (ret < 0)? ret: 0;
1023 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1024 * from current position to next entry still in use.
1025 * Recycle to start on reaching the end, returning 0 when empty.
1027 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1028 unsigned int prev, bool frontswap)
1030 unsigned int max = si->max;
1031 unsigned int i = prev;
1032 unsigned char count;
1035 * No need for swap_lock here: we're just looking
1036 * for whether an entry is in use, not modifying it; false
1037 * hits are okay, and sys_swapoff() has already prevented new
1038 * allocations from this area (while holding swap_lock).
1040 for (;;) {
1041 if (++i >= max) {
1042 if (!prev) {
1043 i = 0;
1044 break;
1047 * No entries in use at top of swap_map,
1048 * loop back to start and recheck there.
1050 max = prev + 1;
1051 prev = 0;
1052 i = 1;
1054 if (frontswap) {
1055 if (frontswap_test(si, i))
1056 break;
1057 else
1058 continue;
1060 count = si->swap_map[i];
1061 if (count && swap_count(count) != SWAP_MAP_BAD)
1062 break;
1064 return i;
1068 * We completely avoid races by reading each swap page in advance,
1069 * and then search for the process using it. All the necessary
1070 * page table adjustments can then be made atomically.
1072 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1073 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1075 int try_to_unuse(unsigned int type, bool frontswap,
1076 unsigned long pages_to_unuse)
1078 struct swap_info_struct *si = swap_info[type];
1079 struct mm_struct *start_mm;
1080 unsigned char *swap_map;
1081 unsigned char swcount;
1082 struct page *page;
1083 swp_entry_t entry;
1084 unsigned int i = 0;
1085 int retval = 0;
1088 * When searching mms for an entry, a good strategy is to
1089 * start at the first mm we freed the previous entry from
1090 * (though actually we don't notice whether we or coincidence
1091 * freed the entry). Initialize this start_mm with a hold.
1093 * A simpler strategy would be to start at the last mm we
1094 * freed the previous entry from; but that would take less
1095 * advantage of mmlist ordering, which clusters forked mms
1096 * together, child after parent. If we race with dup_mmap(), we
1097 * prefer to resolve parent before child, lest we miss entries
1098 * duplicated after we scanned child: using last mm would invert
1099 * that.
1101 start_mm = &init_mm;
1102 atomic_inc(&init_mm.mm_users);
1105 * Keep on scanning until all entries have gone. Usually,
1106 * one pass through swap_map is enough, but not necessarily:
1107 * there are races when an instance of an entry might be missed.
1109 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1110 if (signal_pending(current)) {
1111 retval = -EINTR;
1112 break;
1116 * Get a page for the entry, using the existing swap
1117 * cache page if there is one. Otherwise, get a clean
1118 * page and read the swap into it.
1120 swap_map = &si->swap_map[i];
1121 entry = swp_entry(type, i);
1122 page = read_swap_cache_async(entry,
1123 GFP_HIGHUSER_MOVABLE, NULL, 0);
1124 if (!page) {
1126 * Either swap_duplicate() failed because entry
1127 * has been freed independently, and will not be
1128 * reused since sys_swapoff() already disabled
1129 * allocation from here, or alloc_page() failed.
1131 if (!*swap_map)
1132 continue;
1133 retval = -ENOMEM;
1134 break;
1138 * Don't hold on to start_mm if it looks like exiting.
1140 if (atomic_read(&start_mm->mm_users) == 1) {
1141 mmput(start_mm);
1142 start_mm = &init_mm;
1143 atomic_inc(&init_mm.mm_users);
1147 * Wait for and lock page. When do_swap_page races with
1148 * try_to_unuse, do_swap_page can handle the fault much
1149 * faster than try_to_unuse can locate the entry. This
1150 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1151 * defer to do_swap_page in such a case - in some tests,
1152 * do_swap_page and try_to_unuse repeatedly compete.
1154 wait_on_page_locked(page);
1155 wait_on_page_writeback(page);
1156 lock_page(page);
1157 wait_on_page_writeback(page);
1160 * Remove all references to entry.
1162 swcount = *swap_map;
1163 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1164 retval = shmem_unuse(entry, page);
1165 /* page has already been unlocked and released */
1166 if (retval < 0)
1167 break;
1168 continue;
1170 if (swap_count(swcount) && start_mm != &init_mm)
1171 retval = unuse_mm(start_mm, entry, page);
1173 if (swap_count(*swap_map)) {
1174 int set_start_mm = (*swap_map >= swcount);
1175 struct list_head *p = &start_mm->mmlist;
1176 struct mm_struct *new_start_mm = start_mm;
1177 struct mm_struct *prev_mm = start_mm;
1178 struct mm_struct *mm;
1180 atomic_inc(&new_start_mm->mm_users);
1181 atomic_inc(&prev_mm->mm_users);
1182 spin_lock(&mmlist_lock);
1183 while (swap_count(*swap_map) && !retval &&
1184 (p = p->next) != &start_mm->mmlist) {
1185 mm = list_entry(p, struct mm_struct, mmlist);
1186 if (!atomic_inc_not_zero(&mm->mm_users))
1187 continue;
1188 spin_unlock(&mmlist_lock);
1189 mmput(prev_mm);
1190 prev_mm = mm;
1192 cond_resched();
1194 swcount = *swap_map;
1195 if (!swap_count(swcount)) /* any usage ? */
1197 else if (mm == &init_mm)
1198 set_start_mm = 1;
1199 else
1200 retval = unuse_mm(mm, entry, page);
1202 if (set_start_mm && *swap_map < swcount) {
1203 mmput(new_start_mm);
1204 atomic_inc(&mm->mm_users);
1205 new_start_mm = mm;
1206 set_start_mm = 0;
1208 spin_lock(&mmlist_lock);
1210 spin_unlock(&mmlist_lock);
1211 mmput(prev_mm);
1212 mmput(start_mm);
1213 start_mm = new_start_mm;
1215 if (retval) {
1216 unlock_page(page);
1217 page_cache_release(page);
1218 break;
1222 * If a reference remains (rare), we would like to leave
1223 * the page in the swap cache; but try_to_unmap could
1224 * then re-duplicate the entry once we drop page lock,
1225 * so we might loop indefinitely; also, that page could
1226 * not be swapped out to other storage meanwhile. So:
1227 * delete from cache even if there's another reference,
1228 * after ensuring that the data has been saved to disk -
1229 * since if the reference remains (rarer), it will be
1230 * read from disk into another page. Splitting into two
1231 * pages would be incorrect if swap supported "shared
1232 * private" pages, but they are handled by tmpfs files.
1234 * Given how unuse_vma() targets one particular offset
1235 * in an anon_vma, once the anon_vma has been determined,
1236 * this splitting happens to be just what is needed to
1237 * handle where KSM pages have been swapped out: re-reading
1238 * is unnecessarily slow, but we can fix that later on.
1240 if (swap_count(*swap_map) &&
1241 PageDirty(page) && PageSwapCache(page)) {
1242 struct writeback_control wbc = {
1243 .sync_mode = WB_SYNC_NONE,
1246 swap_writepage(page, &wbc);
1247 lock_page(page);
1248 wait_on_page_writeback(page);
1252 * It is conceivable that a racing task removed this page from
1253 * swap cache just before we acquired the page lock at the top,
1254 * or while we dropped it in unuse_mm(). The page might even
1255 * be back in swap cache on another swap area: that we must not
1256 * delete, since it may not have been written out to swap yet.
1258 if (PageSwapCache(page) &&
1259 likely(page_private(page) == entry.val))
1260 delete_from_swap_cache(page);
1263 * So we could skip searching mms once swap count went
1264 * to 1, we did not mark any present ptes as dirty: must
1265 * mark page dirty so shrink_page_list will preserve it.
1267 SetPageDirty(page);
1268 unlock_page(page);
1269 page_cache_release(page);
1272 * Make sure that we aren't completely killing
1273 * interactive performance.
1275 cond_resched();
1276 if (frontswap && pages_to_unuse > 0) {
1277 if (!--pages_to_unuse)
1278 break;
1282 mmput(start_mm);
1283 return retval;
1287 * After a successful try_to_unuse, if no swap is now in use, we know
1288 * we can empty the mmlist. swap_lock must be held on entry and exit.
1289 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1290 * added to the mmlist just after page_duplicate - before would be racy.
1292 static void drain_mmlist(void)
1294 struct list_head *p, *next;
1295 unsigned int type;
1297 for (type = 0; type < nr_swapfiles; type++)
1298 if (swap_info[type]->inuse_pages)
1299 return;
1300 spin_lock(&mmlist_lock);
1301 list_for_each_safe(p, next, &init_mm.mmlist)
1302 list_del_init(p);
1303 spin_unlock(&mmlist_lock);
1307 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1308 * corresponds to page offset for the specified swap entry.
1309 * Note that the type of this function is sector_t, but it returns page offset
1310 * into the bdev, not sector offset.
1312 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1314 struct swap_info_struct *sis;
1315 struct swap_extent *start_se;
1316 struct swap_extent *se;
1317 pgoff_t offset;
1319 sis = swap_info[swp_type(entry)];
1320 *bdev = sis->bdev;
1322 offset = swp_offset(entry);
1323 start_se = sis->curr_swap_extent;
1324 se = start_se;
1326 for ( ; ; ) {
1327 struct list_head *lh;
1329 if (se->start_page <= offset &&
1330 offset < (se->start_page + se->nr_pages)) {
1331 return se->start_block + (offset - se->start_page);
1333 lh = se->list.next;
1334 se = list_entry(lh, struct swap_extent, list);
1335 sis->curr_swap_extent = se;
1336 BUG_ON(se == start_se); /* It *must* be present */
1341 * Returns the page offset into bdev for the specified page's swap entry.
1343 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1345 swp_entry_t entry;
1346 entry.val = page_private(page);
1347 return map_swap_entry(entry, bdev);
1351 * Free all of a swapdev's extent information
1353 static void destroy_swap_extents(struct swap_info_struct *sis)
1355 while (!list_empty(&sis->first_swap_extent.list)) {
1356 struct swap_extent *se;
1358 se = list_entry(sis->first_swap_extent.list.next,
1359 struct swap_extent, list);
1360 list_del(&se->list);
1361 kfree(se);
1366 * Add a block range (and the corresponding page range) into this swapdev's
1367 * extent list. The extent list is kept sorted in page order.
1369 * This function rather assumes that it is called in ascending page order.
1371 static int
1372 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1373 unsigned long nr_pages, sector_t start_block)
1375 struct swap_extent *se;
1376 struct swap_extent *new_se;
1377 struct list_head *lh;
1379 if (start_page == 0) {
1380 se = &sis->first_swap_extent;
1381 sis->curr_swap_extent = se;
1382 se->start_page = 0;
1383 se->nr_pages = nr_pages;
1384 se->start_block = start_block;
1385 return 1;
1386 } else {
1387 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1388 se = list_entry(lh, struct swap_extent, list);
1389 BUG_ON(se->start_page + se->nr_pages != start_page);
1390 if (se->start_block + se->nr_pages == start_block) {
1391 /* Merge it */
1392 se->nr_pages += nr_pages;
1393 return 0;
1398 * No merge. Insert a new extent, preserving ordering.
1400 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1401 if (new_se == NULL)
1402 return -ENOMEM;
1403 new_se->start_page = start_page;
1404 new_se->nr_pages = nr_pages;
1405 new_se->start_block = start_block;
1407 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1408 return 1;
1412 * A `swap extent' is a simple thing which maps a contiguous range of pages
1413 * onto a contiguous range of disk blocks. An ordered list of swap extents
1414 * is built at swapon time and is then used at swap_writepage/swap_readpage
1415 * time for locating where on disk a page belongs.
1417 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1418 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1419 * swap files identically.
1421 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1422 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1423 * swapfiles are handled *identically* after swapon time.
1425 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1426 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1427 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1428 * requirements, they are simply tossed out - we will never use those blocks
1429 * for swapping.
1431 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1432 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1433 * which will scribble on the fs.
1435 * The amount of disk space which a single swap extent represents varies.
1436 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1437 * extents in the list. To avoid much list walking, we cache the previous
1438 * search location in `curr_swap_extent', and start new searches from there.
1439 * This is extremely effective. The average number of iterations in
1440 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1442 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1444 struct inode *inode;
1445 unsigned blocks_per_page;
1446 unsigned long page_no;
1447 unsigned blkbits;
1448 sector_t probe_block;
1449 sector_t last_block;
1450 sector_t lowest_block = -1;
1451 sector_t highest_block = 0;
1452 int nr_extents = 0;
1453 int ret;
1455 inode = sis->swap_file->f_mapping->host;
1456 if (S_ISBLK(inode->i_mode)) {
1457 ret = add_swap_extent(sis, 0, sis->max, 0);
1458 *span = sis->pages;
1459 goto out;
1462 blkbits = inode->i_blkbits;
1463 blocks_per_page = PAGE_SIZE >> blkbits;
1466 * Map all the blocks into the extent list. This code doesn't try
1467 * to be very smart.
1469 probe_block = 0;
1470 page_no = 0;
1471 last_block = i_size_read(inode) >> blkbits;
1472 while ((probe_block + blocks_per_page) <= last_block &&
1473 page_no < sis->max) {
1474 unsigned block_in_page;
1475 sector_t first_block;
1477 first_block = bmap(inode, probe_block);
1478 if (first_block == 0)
1479 goto bad_bmap;
1482 * It must be PAGE_SIZE aligned on-disk
1484 if (first_block & (blocks_per_page - 1)) {
1485 probe_block++;
1486 goto reprobe;
1489 for (block_in_page = 1; block_in_page < blocks_per_page;
1490 block_in_page++) {
1491 sector_t block;
1493 block = bmap(inode, probe_block + block_in_page);
1494 if (block == 0)
1495 goto bad_bmap;
1496 if (block != first_block + block_in_page) {
1497 /* Discontiguity */
1498 probe_block++;
1499 goto reprobe;
1503 first_block >>= (PAGE_SHIFT - blkbits);
1504 if (page_no) { /* exclude the header page */
1505 if (first_block < lowest_block)
1506 lowest_block = first_block;
1507 if (first_block > highest_block)
1508 highest_block = first_block;
1512 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1514 ret = add_swap_extent(sis, page_no, 1, first_block);
1515 if (ret < 0)
1516 goto out;
1517 nr_extents += ret;
1518 page_no++;
1519 probe_block += blocks_per_page;
1520 reprobe:
1521 continue;
1523 ret = nr_extents;
1524 *span = 1 + highest_block - lowest_block;
1525 if (page_no == 0)
1526 page_no = 1; /* force Empty message */
1527 sis->max = page_no;
1528 sis->pages = page_no - 1;
1529 sis->highest_bit = page_no - 1;
1530 out:
1531 return ret;
1532 bad_bmap:
1533 printk(KERN_ERR "swapon: swapfile has holes\n");
1534 ret = -EINVAL;
1535 goto out;
1538 static void enable_swap_info(struct swap_info_struct *p, int prio,
1539 unsigned char *swap_map,
1540 unsigned long *frontswap_map)
1542 int i, prev;
1544 spin_lock(&swap_lock);
1545 if (prio >= 0)
1546 p->prio = prio;
1547 else
1548 p->prio = --least_priority;
1549 p->swap_map = swap_map;
1550 frontswap_map_set(p, frontswap_map);
1551 p->flags |= SWP_WRITEOK;
1552 nr_swap_pages += p->pages;
1553 total_swap_pages += p->pages;
1555 /* insert swap space into swap_list: */
1556 prev = -1;
1557 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1558 if (p->prio >= swap_info[i]->prio)
1559 break;
1560 prev = i;
1562 p->next = i;
1563 if (prev < 0)
1564 swap_list.head = swap_list.next = p->type;
1565 else
1566 swap_info[prev]->next = p->type;
1567 frontswap_init(p->type);
1568 spin_unlock(&swap_lock);
1571 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1573 struct swap_info_struct *p = NULL;
1574 unsigned char *swap_map;
1575 struct file *swap_file, *victim;
1576 struct address_space *mapping;
1577 struct inode *inode;
1578 char *pathname;
1579 int oom_score_adj;
1580 int i, type, prev;
1581 int err;
1583 if (!capable(CAP_SYS_ADMIN))
1584 return -EPERM;
1586 pathname = getname(specialfile);
1587 err = PTR_ERR(pathname);
1588 if (IS_ERR(pathname))
1589 goto out;
1591 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1592 putname(pathname);
1593 err = PTR_ERR(victim);
1594 if (IS_ERR(victim))
1595 goto out;
1597 mapping = victim->f_mapping;
1598 prev = -1;
1599 spin_lock(&swap_lock);
1600 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1601 p = swap_info[type];
1602 if (p->flags & SWP_WRITEOK) {
1603 if (p->swap_file->f_mapping == mapping)
1604 break;
1606 prev = type;
1608 if (type < 0) {
1609 err = -EINVAL;
1610 spin_unlock(&swap_lock);
1611 goto out_dput;
1613 if (!security_vm_enough_memory(p->pages))
1614 vm_unacct_memory(p->pages);
1615 else {
1616 err = -ENOMEM;
1617 spin_unlock(&swap_lock);
1618 goto out_dput;
1620 if (prev < 0)
1621 swap_list.head = p->next;
1622 else
1623 swap_info[prev]->next = p->next;
1624 if (type == swap_list.next) {
1625 /* just pick something that's safe... */
1626 swap_list.next = swap_list.head;
1628 if (p->prio < 0) {
1629 for (i = p->next; i >= 0; i = swap_info[i]->next)
1630 swap_info[i]->prio = p->prio--;
1631 least_priority++;
1633 nr_swap_pages -= p->pages;
1634 total_swap_pages -= p->pages;
1635 p->flags &= ~SWP_WRITEOK;
1636 spin_unlock(&swap_lock);
1638 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1639 err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1640 test_set_oom_score_adj(oom_score_adj);
1642 if (err) {
1644 * reading p->prio and p->swap_map outside the lock is
1645 * safe here because only sys_swapon and sys_swapoff
1646 * change them, and there can be no other sys_swapon or
1647 * sys_swapoff for this swap_info_struct at this point.
1649 /* re-insert swap space back into swap_list */
1650 enable_swap_info(p, p->prio, p->swap_map, frontswap_map_get(p));
1651 goto out_dput;
1654 destroy_swap_extents(p);
1655 if (p->flags & SWP_CONTINUED)
1656 free_swap_count_continuations(p);
1658 mutex_lock(&swapon_mutex);
1659 spin_lock(&swap_lock);
1660 drain_mmlist();
1662 /* wait for anyone still in scan_swap_map */
1663 p->highest_bit = 0; /* cuts scans short */
1664 while (p->flags >= SWP_SCANNING) {
1665 spin_unlock(&swap_lock);
1666 schedule_timeout_uninterruptible(1);
1667 spin_lock(&swap_lock);
1670 swap_file = p->swap_file;
1671 p->swap_file = NULL;
1672 p->max = 0;
1673 swap_map = p->swap_map;
1674 p->swap_map = NULL;
1675 p->flags = 0;
1676 frontswap_flush_area(type);
1677 spin_unlock(&swap_lock);
1678 mutex_unlock(&swapon_mutex);
1679 vfree(swap_map);
1680 vfree(frontswap_map_get(p));
1681 /* Destroy swap account informatin */
1682 swap_cgroup_swapoff(type);
1684 inode = mapping->host;
1685 if (S_ISBLK(inode->i_mode)) {
1686 struct block_device *bdev = I_BDEV(inode);
1687 set_blocksize(bdev, p->old_block_size);
1688 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1689 } else {
1690 mutex_lock(&inode->i_mutex);
1691 inode->i_flags &= ~S_SWAPFILE;
1692 mutex_unlock(&inode->i_mutex);
1694 filp_close(swap_file, NULL);
1695 err = 0;
1696 atomic_inc(&proc_poll_event);
1697 wake_up_interruptible(&proc_poll_wait);
1699 out_dput:
1700 filp_close(victim, NULL);
1701 out:
1702 return err;
1705 #ifdef CONFIG_PROC_FS
1706 static unsigned swaps_poll(struct file *file, poll_table *wait)
1708 struct seq_file *seq = file->private_data;
1710 poll_wait(file, &proc_poll_wait, wait);
1712 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1713 seq->poll_event = atomic_read(&proc_poll_event);
1714 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1717 return POLLIN | POLLRDNORM;
1720 /* iterator */
1721 static void *swap_start(struct seq_file *swap, loff_t *pos)
1723 struct swap_info_struct *si;
1724 int type;
1725 loff_t l = *pos;
1727 mutex_lock(&swapon_mutex);
1729 if (!l)
1730 return SEQ_START_TOKEN;
1732 for (type = 0; 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)
1736 continue;
1737 if (!--l)
1738 return si;
1741 return NULL;
1744 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1746 struct swap_info_struct *si = v;
1747 int type;
1749 if (v == SEQ_START_TOKEN)
1750 type = 0;
1751 else
1752 type = si->type + 1;
1754 for (; type < nr_swapfiles; type++) {
1755 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1756 si = swap_info[type];
1757 if (!(si->flags & SWP_USED) || !si->swap_map)
1758 continue;
1759 ++*pos;
1760 return si;
1763 return NULL;
1766 static void swap_stop(struct seq_file *swap, void *v)
1768 mutex_unlock(&swapon_mutex);
1771 static int swap_show(struct seq_file *swap, void *v)
1773 struct swap_info_struct *si = v;
1774 struct file *file;
1775 int len;
1777 if (si == SEQ_START_TOKEN) {
1778 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1779 return 0;
1782 file = si->swap_file;
1783 len = seq_path(swap, &file->f_path, " \t\n\\");
1784 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1785 len < 40 ? 40 - len : 1, " ",
1786 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1787 "partition" : "file\t",
1788 si->pages << (PAGE_SHIFT - 10),
1789 si->inuse_pages << (PAGE_SHIFT - 10),
1790 si->prio);
1791 return 0;
1794 static const struct seq_operations swaps_op = {
1795 .start = swap_start,
1796 .next = swap_next,
1797 .stop = swap_stop,
1798 .show = swap_show
1801 static int swaps_open(struct inode *inode, struct file *file)
1803 struct seq_file *seq;
1804 int ret;
1806 ret = seq_open(file, &swaps_op);
1807 if (ret)
1808 return ret;
1810 seq = file->private_data;
1811 seq->poll_event = atomic_read(&proc_poll_event);
1812 return 0;
1815 static const struct file_operations proc_swaps_operations = {
1816 .open = swaps_open,
1817 .read = seq_read,
1818 .llseek = seq_lseek,
1819 .release = seq_release,
1820 .poll = swaps_poll,
1823 static int __init procswaps_init(void)
1825 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1826 return 0;
1828 __initcall(procswaps_init);
1829 #endif /* CONFIG_PROC_FS */
1831 #ifdef MAX_SWAPFILES_CHECK
1832 static int __init max_swapfiles_check(void)
1834 MAX_SWAPFILES_CHECK();
1835 return 0;
1837 late_initcall(max_swapfiles_check);
1838 #endif
1840 static struct swap_info_struct *alloc_swap_info(void)
1842 struct swap_info_struct *p;
1843 unsigned int type;
1845 p = kzalloc(sizeof(*p), GFP_KERNEL);
1846 if (!p)
1847 return ERR_PTR(-ENOMEM);
1849 spin_lock(&swap_lock);
1850 for (type = 0; type < nr_swapfiles; type++) {
1851 if (!(swap_info[type]->flags & SWP_USED))
1852 break;
1854 if (type >= MAX_SWAPFILES) {
1855 spin_unlock(&swap_lock);
1856 kfree(p);
1857 return ERR_PTR(-EPERM);
1859 if (type >= nr_swapfiles) {
1860 p->type = type;
1861 swap_info[type] = p;
1863 * Write swap_info[type] before nr_swapfiles, in case a
1864 * racing procfs swap_start() or swap_next() is reading them.
1865 * (We never shrink nr_swapfiles, we never free this entry.)
1867 smp_wmb();
1868 nr_swapfiles++;
1869 } else {
1870 kfree(p);
1871 p = swap_info[type];
1873 * Do not memset this entry: a racing procfs swap_next()
1874 * would be relying on p->type to remain valid.
1877 INIT_LIST_HEAD(&p->first_swap_extent.list);
1878 p->flags = SWP_USED;
1879 p->next = -1;
1880 spin_unlock(&swap_lock);
1882 return p;
1885 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1887 int error;
1889 if (S_ISBLK(inode->i_mode)) {
1890 p->bdev = bdgrab(I_BDEV(inode));
1891 error = blkdev_get(p->bdev,
1892 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1893 sys_swapon);
1894 if (error < 0) {
1895 p->bdev = NULL;
1896 return -EINVAL;
1898 p->old_block_size = block_size(p->bdev);
1899 error = set_blocksize(p->bdev, PAGE_SIZE);
1900 if (error < 0)
1901 return error;
1902 p->flags |= SWP_BLKDEV;
1903 } else if (S_ISREG(inode->i_mode)) {
1904 p->bdev = inode->i_sb->s_bdev;
1905 mutex_lock(&inode->i_mutex);
1906 if (IS_SWAPFILE(inode))
1907 return -EBUSY;
1908 } else
1909 return -EINVAL;
1911 return 0;
1914 static unsigned long read_swap_header(struct swap_info_struct *p,
1915 union swap_header *swap_header,
1916 struct inode *inode)
1918 int i;
1919 unsigned long maxpages;
1920 unsigned long swapfilepages;
1922 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1923 printk(KERN_ERR "Unable to find swap-space signature\n");
1924 return 0;
1927 /* swap partition endianess hack... */
1928 if (swab32(swap_header->info.version) == 1) {
1929 swab32s(&swap_header->info.version);
1930 swab32s(&swap_header->info.last_page);
1931 swab32s(&swap_header->info.nr_badpages);
1932 for (i = 0; i < swap_header->info.nr_badpages; i++)
1933 swab32s(&swap_header->info.badpages[i]);
1935 /* Check the swap header's sub-version */
1936 if (swap_header->info.version != 1) {
1937 printk(KERN_WARNING
1938 "Unable to handle swap header version %d\n",
1939 swap_header->info.version);
1940 return 0;
1943 p->lowest_bit = 1;
1944 p->cluster_next = 1;
1945 p->cluster_nr = 0;
1948 * Find out how many pages are allowed for a single swap
1949 * device. There are three limiting factors: 1) the number
1950 * of bits for the swap offset in the swp_entry_t type, and
1951 * 2) the number of bits in the swap pte as defined by the
1952 * the different architectures, and 3) the number of free bits
1953 * in an exceptional radix_tree entry. In order to find the
1954 * largest possible bit mask, a swap entry with swap type 0
1955 * and swap offset ~0UL is created, encoded to a swap pte,
1956 * decoded to a swp_entry_t again, and finally the swap
1957 * offset is extracted. This will mask all the bits from
1958 * the initial ~0UL mask that can't be encoded in either
1959 * the swp_entry_t or the architecture definition of a
1960 * swap pte. Then the same is done for a radix_tree entry.
1962 maxpages = swp_offset(pte_to_swp_entry(
1963 swp_entry_to_pte(swp_entry(0, ~0UL))));
1964 maxpages = swp_offset(radix_to_swp_entry(
1965 swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1967 if (maxpages > swap_header->info.last_page) {
1968 maxpages = swap_header->info.last_page + 1;
1969 /* p->max is an unsigned int: don't overflow it */
1970 if ((unsigned int)maxpages == 0)
1971 maxpages = UINT_MAX;
1973 p->highest_bit = maxpages - 1;
1975 if (!maxpages)
1976 return 0;
1977 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1978 if (swapfilepages && maxpages > swapfilepages) {
1979 printk(KERN_WARNING
1980 "Swap area shorter than signature indicates\n");
1981 return 0;
1983 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1984 return 0;
1985 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1986 return 0;
1988 return maxpages;
1991 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1992 union swap_header *swap_header,
1993 unsigned char *swap_map,
1994 unsigned long maxpages,
1995 sector_t *span)
1997 int i;
1998 unsigned int nr_good_pages;
1999 int nr_extents;
2001 nr_good_pages = maxpages - 1; /* omit header page */
2003 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2004 unsigned int page_nr = swap_header->info.badpages[i];
2005 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2006 return -EINVAL;
2007 if (page_nr < maxpages) {
2008 swap_map[page_nr] = SWAP_MAP_BAD;
2009 nr_good_pages--;
2013 if (nr_good_pages) {
2014 swap_map[0] = SWAP_MAP_BAD;
2015 p->max = maxpages;
2016 p->pages = nr_good_pages;
2017 nr_extents = setup_swap_extents(p, span);
2018 if (nr_extents < 0)
2019 return nr_extents;
2020 nr_good_pages = p->pages;
2022 if (!nr_good_pages) {
2023 printk(KERN_WARNING "Empty swap-file\n");
2024 return -EINVAL;
2027 return nr_extents;
2030 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2032 struct swap_info_struct *p;
2033 char *name;
2034 struct file *swap_file = NULL;
2035 struct address_space *mapping;
2036 int i;
2037 int prio;
2038 int error;
2039 union swap_header *swap_header;
2040 int nr_extents;
2041 sector_t span;
2042 unsigned long maxpages;
2043 unsigned char *swap_map = NULL;
2044 unsigned long *frontswap_map = NULL;
2045 struct page *page = NULL;
2046 struct inode *inode = NULL;
2048 if (!capable(CAP_SYS_ADMIN))
2049 return -EPERM;
2051 p = alloc_swap_info();
2052 if (IS_ERR(p))
2053 return PTR_ERR(p);
2055 name = getname(specialfile);
2056 if (IS_ERR(name)) {
2057 error = PTR_ERR(name);
2058 name = NULL;
2059 goto bad_swap;
2061 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2062 if (IS_ERR(swap_file)) {
2063 error = PTR_ERR(swap_file);
2064 swap_file = NULL;
2065 goto bad_swap;
2068 p->swap_file = swap_file;
2069 mapping = swap_file->f_mapping;
2071 for (i = 0; i < nr_swapfiles; i++) {
2072 struct swap_info_struct *q = swap_info[i];
2074 if (q == p || !q->swap_file)
2075 continue;
2076 if (mapping == q->swap_file->f_mapping) {
2077 error = -EBUSY;
2078 goto bad_swap;
2082 inode = mapping->host;
2083 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2084 error = claim_swapfile(p, inode);
2085 if (unlikely(error))
2086 goto bad_swap;
2089 * Read the swap header.
2091 if (!mapping->a_ops->readpage) {
2092 error = -EINVAL;
2093 goto bad_swap;
2095 page = read_mapping_page(mapping, 0, swap_file);
2096 if (IS_ERR(page)) {
2097 error = PTR_ERR(page);
2098 goto bad_swap;
2100 swap_header = kmap(page);
2102 maxpages = read_swap_header(p, swap_header, inode);
2103 if (unlikely(!maxpages)) {
2104 error = -EINVAL;
2105 goto bad_swap;
2108 /* OK, set up the swap map and apply the bad block list */
2109 swap_map = vzalloc(maxpages);
2110 if (!swap_map) {
2111 error = -ENOMEM;
2112 goto bad_swap;
2115 error = swap_cgroup_swapon(p->type, maxpages);
2116 if (error)
2117 goto bad_swap;
2119 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2120 maxpages, &span);
2121 if (unlikely(nr_extents < 0)) {
2122 error = nr_extents;
2123 goto bad_swap;
2125 /* frontswap enabled? set up bit-per-page map for frontswap */
2126 if (frontswap_enabled)
2127 frontswap_map = vzalloc(maxpages / sizeof(long));
2129 if (p->bdev) {
2130 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2131 p->flags |= SWP_SOLIDSTATE;
2132 p->cluster_next = 1 + (random32() % p->highest_bit);
2134 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2135 p->flags |= SWP_DISCARDABLE;
2138 mutex_lock(&swapon_mutex);
2139 prio = -1;
2140 if (swap_flags & SWAP_FLAG_PREFER)
2141 prio =
2142 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2143 enable_swap_info(p, prio, swap_map, frontswap_map);
2145 printk(KERN_INFO "Adding %uk swap on %s. "
2146 "Priority:%d extents:%d across:%lluk %s%s%s\n",
2147 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2148 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2149 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2150 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2151 (frontswap_map) ? "FS" : "");
2153 mutex_unlock(&swapon_mutex);
2154 atomic_inc(&proc_poll_event);
2155 wake_up_interruptible(&proc_poll_wait);
2157 if (S_ISREG(inode->i_mode))
2158 inode->i_flags |= S_SWAPFILE;
2159 error = 0;
2160 goto out;
2161 bad_swap:
2162 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2163 set_blocksize(p->bdev, p->old_block_size);
2164 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2166 destroy_swap_extents(p);
2167 swap_cgroup_swapoff(p->type);
2168 spin_lock(&swap_lock);
2169 p->swap_file = NULL;
2170 p->flags = 0;
2171 spin_unlock(&swap_lock);
2172 vfree(swap_map);
2173 if (swap_file) {
2174 if (inode && S_ISREG(inode->i_mode)) {
2175 mutex_unlock(&inode->i_mutex);
2176 inode = NULL;
2178 filp_close(swap_file, NULL);
2180 out:
2181 if (page && !IS_ERR(page)) {
2182 kunmap(page);
2183 page_cache_release(page);
2185 if (name)
2186 putname(name);
2187 if (inode && S_ISREG(inode->i_mode))
2188 mutex_unlock(&inode->i_mutex);
2189 return error;
2192 void si_swapinfo(struct sysinfo *val)
2194 unsigned int type;
2195 unsigned long nr_to_be_unused = 0;
2197 spin_lock(&swap_lock);
2198 for (type = 0; type < nr_swapfiles; type++) {
2199 struct swap_info_struct *si = swap_info[type];
2201 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2202 nr_to_be_unused += si->inuse_pages;
2204 val->freeswap = nr_swap_pages + nr_to_be_unused;
2205 val->totalswap = total_swap_pages + nr_to_be_unused;
2206 spin_unlock(&swap_lock);
2210 * Verify that a swap entry is valid and increment its swap map count.
2212 * Returns error code in following case.
2213 * - success -> 0
2214 * - swp_entry is invalid -> EINVAL
2215 * - swp_entry is migration entry -> EINVAL
2216 * - swap-cache reference is requested but there is already one. -> EEXIST
2217 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2218 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2220 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2222 struct swap_info_struct *p;
2223 unsigned long offset, type;
2224 unsigned char count;
2225 unsigned char has_cache;
2226 int err = -EINVAL;
2228 if (non_swap_entry(entry))
2229 goto out;
2231 type = swp_type(entry);
2232 if (type >= nr_swapfiles)
2233 goto bad_file;
2234 p = swap_info[type];
2235 offset = swp_offset(entry);
2237 spin_lock(&swap_lock);
2238 if (unlikely(offset >= p->max))
2239 goto unlock_out;
2241 count = p->swap_map[offset];
2242 has_cache = count & SWAP_HAS_CACHE;
2243 count &= ~SWAP_HAS_CACHE;
2244 err = 0;
2246 if (usage == SWAP_HAS_CACHE) {
2248 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2249 if (!has_cache && count)
2250 has_cache = SWAP_HAS_CACHE;
2251 else if (has_cache) /* someone else added cache */
2252 err = -EEXIST;
2253 else /* no users remaining */
2254 err = -ENOENT;
2256 } else if (count || has_cache) {
2258 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2259 count += usage;
2260 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2261 err = -EINVAL;
2262 else if (swap_count_continued(p, offset, count))
2263 count = COUNT_CONTINUED;
2264 else
2265 err = -ENOMEM;
2266 } else
2267 err = -ENOENT; /* unused swap entry */
2269 p->swap_map[offset] = count | has_cache;
2271 unlock_out:
2272 spin_unlock(&swap_lock);
2273 out:
2274 return err;
2276 bad_file:
2277 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2278 goto out;
2282 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2283 * (in which case its reference count is never incremented).
2285 void swap_shmem_alloc(swp_entry_t entry)
2287 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2291 * Increase reference count of swap entry by 1.
2292 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2293 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2294 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2295 * might occur if a page table entry has got corrupted.
2297 int swap_duplicate(swp_entry_t entry)
2299 int err = 0;
2301 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2302 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2303 return err;
2307 * @entry: swap entry for which we allocate swap cache.
2309 * Called when allocating swap cache for existing swap entry,
2310 * This can return error codes. Returns 0 at success.
2311 * -EBUSY means there is a swap cache.
2312 * Note: return code is different from swap_duplicate().
2314 int swapcache_prepare(swp_entry_t entry)
2316 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2320 * swap_lock prevents swap_map being freed. Don't grab an extra
2321 * reference on the swaphandle, it doesn't matter if it becomes unused.
2323 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2325 struct swap_info_struct *si;
2326 int our_page_cluster = page_cluster;
2327 pgoff_t target, toff;
2328 pgoff_t base, end;
2329 int nr_pages = 0;
2331 if (!our_page_cluster) /* no readahead */
2332 return 0;
2334 si = swap_info[swp_type(entry)];
2335 target = swp_offset(entry);
2336 base = (target >> our_page_cluster) << our_page_cluster;
2337 end = base + (1 << our_page_cluster);
2338 if (!base) /* first page is swap header */
2339 base++;
2341 spin_lock(&swap_lock);
2342 if (frontswap_test(si, target)) {
2343 spin_unlock(&swap_lock);
2344 return 0;
2346 if (end > si->max) /* don't go beyond end of map */
2347 end = si->max;
2349 /* Count contiguous allocated slots above our target */
2350 for (toff = target; ++toff < end; nr_pages++) {
2351 /* Don't read in free or bad pages */
2352 if (!si->swap_map[toff])
2353 break;
2354 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2355 break;
2356 /* Don't read in frontswap pages */
2357 if (frontswap_test(si, toff))
2358 break;
2360 /* Count contiguous allocated slots below our target */
2361 for (toff = target; --toff >= base; nr_pages++) {
2362 /* Don't read in free or bad pages */
2363 if (!si->swap_map[toff])
2364 break;
2365 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2366 break;
2367 /* Don't read in frontswap pages */
2368 if (frontswap_test(si, toff))
2369 break;
2371 spin_unlock(&swap_lock);
2374 * Indicate starting offset, and return number of pages to get:
2375 * if only 1, say 0, since there's then no readahead to be done.
2377 *offset = ++toff;
2378 return nr_pages? ++nr_pages: 0;
2382 * add_swap_count_continuation - called when a swap count is duplicated
2383 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2384 * page of the original vmalloc'ed swap_map, to hold the continuation count
2385 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2386 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2388 * These continuation pages are seldom referenced: the common paths all work
2389 * on the original swap_map, only referring to a continuation page when the
2390 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2392 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2393 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2394 * can be called after dropping locks.
2396 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2398 struct swap_info_struct *si;
2399 struct page *head;
2400 struct page *page;
2401 struct page *list_page;
2402 pgoff_t offset;
2403 unsigned char count;
2406 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2407 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2409 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2411 si = swap_info_get(entry);
2412 if (!si) {
2414 * An acceptable race has occurred since the failing
2415 * __swap_duplicate(): the swap entry has been freed,
2416 * perhaps even the whole swap_map cleared for swapoff.
2418 goto outer;
2421 offset = swp_offset(entry);
2422 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2424 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2426 * The higher the swap count, the more likely it is that tasks
2427 * will race to add swap count continuation: we need to avoid
2428 * over-provisioning.
2430 goto out;
2433 if (!page) {
2434 spin_unlock(&swap_lock);
2435 return -ENOMEM;
2439 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2440 * no architecture is using highmem pages for kernel pagetables: so it
2441 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2443 head = vmalloc_to_page(si->swap_map + offset);
2444 offset &= ~PAGE_MASK;
2447 * Page allocation does not initialize the page's lru field,
2448 * but it does always reset its private field.
2450 if (!page_private(head)) {
2451 BUG_ON(count & COUNT_CONTINUED);
2452 INIT_LIST_HEAD(&head->lru);
2453 set_page_private(head, SWP_CONTINUED);
2454 si->flags |= SWP_CONTINUED;
2457 list_for_each_entry(list_page, &head->lru, lru) {
2458 unsigned char *map;
2461 * If the previous map said no continuation, but we've found
2462 * a continuation page, free our allocation and use this one.
2464 if (!(count & COUNT_CONTINUED))
2465 goto out;
2467 map = kmap_atomic(list_page, KM_USER0) + offset;
2468 count = *map;
2469 kunmap_atomic(map, KM_USER0);
2472 * If this continuation count now has some space in it,
2473 * free our allocation and use this one.
2475 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2476 goto out;
2479 list_add_tail(&page->lru, &head->lru);
2480 page = NULL; /* now it's attached, don't free it */
2481 out:
2482 spin_unlock(&swap_lock);
2483 outer:
2484 if (page)
2485 __free_page(page);
2486 return 0;
2490 * swap_count_continued - when the original swap_map count is incremented
2491 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2492 * into, carry if so, or else fail until a new continuation page is allocated;
2493 * when the original swap_map count is decremented from 0 with continuation,
2494 * borrow from the continuation and report whether it still holds more.
2495 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2497 static bool swap_count_continued(struct swap_info_struct *si,
2498 pgoff_t offset, unsigned char count)
2500 struct page *head;
2501 struct page *page;
2502 unsigned char *map;
2504 head = vmalloc_to_page(si->swap_map + offset);
2505 if (page_private(head) != SWP_CONTINUED) {
2506 BUG_ON(count & COUNT_CONTINUED);
2507 return false; /* need to add count continuation */
2510 offset &= ~PAGE_MASK;
2511 page = list_entry(head->lru.next, struct page, lru);
2512 map = kmap_atomic(page, KM_USER0) + offset;
2514 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2515 goto init_map; /* jump over SWAP_CONT_MAX checks */
2517 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2519 * Think of how you add 1 to 999
2521 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2522 kunmap_atomic(map, KM_USER0);
2523 page = list_entry(page->lru.next, struct page, lru);
2524 BUG_ON(page == head);
2525 map = kmap_atomic(page, KM_USER0) + offset;
2527 if (*map == SWAP_CONT_MAX) {
2528 kunmap_atomic(map, KM_USER0);
2529 page = list_entry(page->lru.next, struct page, lru);
2530 if (page == head)
2531 return false; /* add count continuation */
2532 map = kmap_atomic(page, KM_USER0) + offset;
2533 init_map: *map = 0; /* we didn't zero the page */
2535 *map += 1;
2536 kunmap_atomic(map, KM_USER0);
2537 page = list_entry(page->lru.prev, struct page, lru);
2538 while (page != head) {
2539 map = kmap_atomic(page, KM_USER0) + offset;
2540 *map = COUNT_CONTINUED;
2541 kunmap_atomic(map, KM_USER0);
2542 page = list_entry(page->lru.prev, struct page, lru);
2544 return true; /* incremented */
2546 } else { /* decrementing */
2548 * Think of how you subtract 1 from 1000
2550 BUG_ON(count != COUNT_CONTINUED);
2551 while (*map == COUNT_CONTINUED) {
2552 kunmap_atomic(map, KM_USER0);
2553 page = list_entry(page->lru.next, struct page, lru);
2554 BUG_ON(page == head);
2555 map = kmap_atomic(page, KM_USER0) + offset;
2557 BUG_ON(*map == 0);
2558 *map -= 1;
2559 if (*map == 0)
2560 count = 0;
2561 kunmap_atomic(map, KM_USER0);
2562 page = list_entry(page->lru.prev, struct page, lru);
2563 while (page != head) {
2564 map = kmap_atomic(page, KM_USER0) + offset;
2565 *map = SWAP_CONT_MAX | count;
2566 count = COUNT_CONTINUED;
2567 kunmap_atomic(map, KM_USER0);
2568 page = list_entry(page->lru.prev, struct page, lru);
2570 return count == COUNT_CONTINUED;
2575 * free_swap_count_continuations - swapoff free all the continuation pages
2576 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2578 static void free_swap_count_continuations(struct swap_info_struct *si)
2580 pgoff_t offset;
2582 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2583 struct page *head;
2584 head = vmalloc_to_page(si->swap_map + offset);
2585 if (page_private(head)) {
2586 struct list_head *this, *next;
2587 list_for_each_safe(this, next, &head->lru) {
2588 struct page *page;
2589 page = list_entry(this, struct page, lru);
2590 list_del(this);
2591 __free_page(page);