powerpc/spufs: Use llseek in all file operations
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / swapfile.c
blob7c703ff2f36f0b760b79eb36149084f07621a0a1
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/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,
40 unsigned char);
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;
46 long nr_swap_pages;
47 long total_swap_pages;
48 static int least_priority;
50 static const char Bad_file[] = "Bad swap file entry ";
51 static const char Unused_file[] = "Unused swap file entry ";
52 static const char Bad_offset[] = "Bad swap offset entry ";
53 static const char Unused_offset[] = "Unused swap offset entry ";
55 static struct swap_list_t swap_list = {-1, -1};
57 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
59 static DEFINE_MUTEX(swapon_mutex);
61 static inline unsigned char swap_count(unsigned char ent)
63 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
66 /* returns 1 if swap entry is freed */
67 static int
68 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
70 swp_entry_t entry = swp_entry(si->type, offset);
71 struct page *page;
72 int ret = 0;
74 page = find_get_page(&swapper_space, entry.val);
75 if (!page)
76 return 0;
78 * This function is called from scan_swap_map() and it's called
79 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
80 * We have to use trylock for avoiding deadlock. This is a special
81 * case and you should use try_to_free_swap() with explicit lock_page()
82 * in usual operations.
84 if (trylock_page(page)) {
85 ret = try_to_free_swap(page);
86 unlock_page(page);
88 page_cache_release(page);
89 return ret;
93 * We need this because the bdev->unplug_fn can sleep and we cannot
94 * hold swap_lock while calling the unplug_fn. And swap_lock
95 * cannot be turned into a mutex.
97 static DECLARE_RWSEM(swap_unplug_sem);
99 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
101 swp_entry_t entry;
103 down_read(&swap_unplug_sem);
104 entry.val = page_private(page);
105 if (PageSwapCache(page)) {
106 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
107 struct backing_dev_info *bdi;
110 * If the page is removed from swapcache from under us (with a
111 * racy try_to_unuse/swapoff) we need an additional reference
112 * count to avoid reading garbage from page_private(page) above.
113 * If the WARN_ON triggers during a swapoff it maybe the race
114 * condition and it's harmless. However if it triggers without
115 * swapoff it signals a problem.
117 WARN_ON(page_count(page) <= 1);
119 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
120 blk_run_backing_dev(bdi, page);
122 up_read(&swap_unplug_sem);
126 * swapon tell device that all the old swap contents can be discarded,
127 * to allow the swap device to optimize its wear-levelling.
129 static int discard_swap(struct swap_info_struct *si)
131 struct swap_extent *se;
132 sector_t start_block;
133 sector_t nr_blocks;
134 int err = 0;
136 /* Do not discard the swap header page! */
137 se = &si->first_swap_extent;
138 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
139 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
140 if (nr_blocks) {
141 err = blkdev_issue_discard(si->bdev, start_block,
142 nr_blocks, GFP_KERNEL, BLKDEV_IFL_WAIT);
143 if (err)
144 return err;
145 cond_resched();
148 list_for_each_entry(se, &si->first_swap_extent.list, list) {
149 start_block = se->start_block << (PAGE_SHIFT - 9);
150 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
152 err = blkdev_issue_discard(si->bdev, start_block,
153 nr_blocks, GFP_KERNEL, BLKDEV_IFL_WAIT);
154 if (err)
155 break;
157 cond_resched();
159 return err; /* That will often be -EOPNOTSUPP */
163 * swap allocation tell device that a cluster of swap can now be discarded,
164 * to allow the swap device to optimize its wear-levelling.
166 static void discard_swap_cluster(struct swap_info_struct *si,
167 pgoff_t start_page, pgoff_t nr_pages)
169 struct swap_extent *se = si->curr_swap_extent;
170 int found_extent = 0;
172 while (nr_pages) {
173 struct list_head *lh;
175 if (se->start_page <= start_page &&
176 start_page < se->start_page + se->nr_pages) {
177 pgoff_t offset = start_page - se->start_page;
178 sector_t start_block = se->start_block + offset;
179 sector_t nr_blocks = se->nr_pages - offset;
181 if (nr_blocks > nr_pages)
182 nr_blocks = nr_pages;
183 start_page += nr_blocks;
184 nr_pages -= nr_blocks;
186 if (!found_extent++)
187 si->curr_swap_extent = se;
189 start_block <<= PAGE_SHIFT - 9;
190 nr_blocks <<= PAGE_SHIFT - 9;
191 if (blkdev_issue_discard(si->bdev, start_block,
192 nr_blocks, GFP_NOIO, BLKDEV_IFL_WAIT))
193 break;
196 lh = se->list.next;
197 se = list_entry(lh, struct swap_extent, list);
201 static int wait_for_discard(void *word)
203 schedule();
204 return 0;
207 #define SWAPFILE_CLUSTER 256
208 #define LATENCY_LIMIT 256
210 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
211 unsigned char usage)
213 unsigned long offset;
214 unsigned long scan_base;
215 unsigned long last_in_cluster = 0;
216 int latency_ration = LATENCY_LIMIT;
217 int found_free_cluster = 0;
220 * We try to cluster swap pages by allocating them sequentially
221 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
222 * way, however, we resort to first-free allocation, starting
223 * a new cluster. This prevents us from scattering swap pages
224 * all over the entire swap partition, so that we reduce
225 * overall disk seek times between swap pages. -- sct
226 * But we do now try to find an empty cluster. -Andrea
227 * And we let swap pages go all over an SSD partition. Hugh
230 si->flags += SWP_SCANNING;
231 scan_base = offset = si->cluster_next;
233 if (unlikely(!si->cluster_nr--)) {
234 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
235 si->cluster_nr = SWAPFILE_CLUSTER - 1;
236 goto checks;
238 if (si->flags & SWP_DISCARDABLE) {
240 * Start range check on racing allocations, in case
241 * they overlap the cluster we eventually decide on
242 * (we scan without swap_lock to allow preemption).
243 * It's hardly conceivable that cluster_nr could be
244 * wrapped during our scan, but don't depend on it.
246 if (si->lowest_alloc)
247 goto checks;
248 si->lowest_alloc = si->max;
249 si->highest_alloc = 0;
251 spin_unlock(&swap_lock);
254 * If seek is expensive, start searching for new cluster from
255 * start of partition, to minimize the span of allocated swap.
256 * But if seek is cheap, search from our current position, so
257 * that swap is allocated from all over the partition: if the
258 * Flash Translation Layer only remaps within limited zones,
259 * we don't want to wear out the first zone too quickly.
261 if (!(si->flags & SWP_SOLIDSTATE))
262 scan_base = offset = si->lowest_bit;
263 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
265 /* Locate the first empty (unaligned) cluster */
266 for (; last_in_cluster <= si->highest_bit; offset++) {
267 if (si->swap_map[offset])
268 last_in_cluster = offset + SWAPFILE_CLUSTER;
269 else if (offset == last_in_cluster) {
270 spin_lock(&swap_lock);
271 offset -= SWAPFILE_CLUSTER - 1;
272 si->cluster_next = offset;
273 si->cluster_nr = SWAPFILE_CLUSTER - 1;
274 found_free_cluster = 1;
275 goto checks;
277 if (unlikely(--latency_ration < 0)) {
278 cond_resched();
279 latency_ration = LATENCY_LIMIT;
283 offset = si->lowest_bit;
284 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
286 /* Locate the first empty (unaligned) cluster */
287 for (; last_in_cluster < scan_base; offset++) {
288 if (si->swap_map[offset])
289 last_in_cluster = offset + SWAPFILE_CLUSTER;
290 else if (offset == last_in_cluster) {
291 spin_lock(&swap_lock);
292 offset -= SWAPFILE_CLUSTER - 1;
293 si->cluster_next = offset;
294 si->cluster_nr = SWAPFILE_CLUSTER - 1;
295 found_free_cluster = 1;
296 goto checks;
298 if (unlikely(--latency_ration < 0)) {
299 cond_resched();
300 latency_ration = LATENCY_LIMIT;
304 offset = scan_base;
305 spin_lock(&swap_lock);
306 si->cluster_nr = SWAPFILE_CLUSTER - 1;
307 si->lowest_alloc = 0;
310 checks:
311 if (!(si->flags & SWP_WRITEOK))
312 goto no_page;
313 if (!si->highest_bit)
314 goto no_page;
315 if (offset > si->highest_bit)
316 scan_base = offset = si->lowest_bit;
318 /* reuse swap entry of cache-only swap if not busy. */
319 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
320 int swap_was_freed;
321 spin_unlock(&swap_lock);
322 swap_was_freed = __try_to_reclaim_swap(si, offset);
323 spin_lock(&swap_lock);
324 /* entry was freed successfully, try to use this again */
325 if (swap_was_freed)
326 goto checks;
327 goto scan; /* check next one */
330 if (si->swap_map[offset])
331 goto scan;
333 if (offset == si->lowest_bit)
334 si->lowest_bit++;
335 if (offset == si->highest_bit)
336 si->highest_bit--;
337 si->inuse_pages++;
338 if (si->inuse_pages == si->pages) {
339 si->lowest_bit = si->max;
340 si->highest_bit = 0;
342 si->swap_map[offset] = usage;
343 si->cluster_next = offset + 1;
344 si->flags -= SWP_SCANNING;
346 if (si->lowest_alloc) {
348 * Only set when SWP_DISCARDABLE, and there's a scan
349 * for a free cluster in progress or just completed.
351 if (found_free_cluster) {
353 * To optimize wear-levelling, discard the
354 * old data of the cluster, taking care not to
355 * discard any of its pages that have already
356 * been allocated by racing tasks (offset has
357 * already stepped over any at the beginning).
359 if (offset < si->highest_alloc &&
360 si->lowest_alloc <= last_in_cluster)
361 last_in_cluster = si->lowest_alloc - 1;
362 si->flags |= SWP_DISCARDING;
363 spin_unlock(&swap_lock);
365 if (offset < last_in_cluster)
366 discard_swap_cluster(si, offset,
367 last_in_cluster - offset + 1);
369 spin_lock(&swap_lock);
370 si->lowest_alloc = 0;
371 si->flags &= ~SWP_DISCARDING;
373 smp_mb(); /* wake_up_bit advises this */
374 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
376 } else if (si->flags & SWP_DISCARDING) {
378 * Delay using pages allocated by racing tasks
379 * until the whole discard has been issued. We
380 * could defer that delay until swap_writepage,
381 * but it's easier to keep this self-contained.
383 spin_unlock(&swap_lock);
384 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
385 wait_for_discard, TASK_UNINTERRUPTIBLE);
386 spin_lock(&swap_lock);
387 } else {
389 * Note pages allocated by racing tasks while
390 * scan for a free cluster is in progress, so
391 * that its final discard can exclude them.
393 if (offset < si->lowest_alloc)
394 si->lowest_alloc = offset;
395 if (offset > si->highest_alloc)
396 si->highest_alloc = offset;
399 return offset;
401 scan:
402 spin_unlock(&swap_lock);
403 while (++offset <= si->highest_bit) {
404 if (!si->swap_map[offset]) {
405 spin_lock(&swap_lock);
406 goto checks;
408 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
409 spin_lock(&swap_lock);
410 goto checks;
412 if (unlikely(--latency_ration < 0)) {
413 cond_resched();
414 latency_ration = LATENCY_LIMIT;
417 offset = si->lowest_bit;
418 while (++offset < scan_base) {
419 if (!si->swap_map[offset]) {
420 spin_lock(&swap_lock);
421 goto checks;
423 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
424 spin_lock(&swap_lock);
425 goto checks;
427 if (unlikely(--latency_ration < 0)) {
428 cond_resched();
429 latency_ration = LATENCY_LIMIT;
432 spin_lock(&swap_lock);
434 no_page:
435 si->flags -= SWP_SCANNING;
436 return 0;
439 swp_entry_t get_swap_page(void)
441 struct swap_info_struct *si;
442 pgoff_t offset;
443 int type, next;
444 int wrapped = 0;
446 spin_lock(&swap_lock);
447 if (nr_swap_pages <= 0)
448 goto noswap;
449 nr_swap_pages--;
451 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
452 si = swap_info[type];
453 next = si->next;
454 if (next < 0 ||
455 (!wrapped && si->prio != swap_info[next]->prio)) {
456 next = swap_list.head;
457 wrapped++;
460 if (!si->highest_bit)
461 continue;
462 if (!(si->flags & SWP_WRITEOK))
463 continue;
465 swap_list.next = next;
466 /* This is called for allocating swap entry for cache */
467 offset = scan_swap_map(si, SWAP_HAS_CACHE);
468 if (offset) {
469 spin_unlock(&swap_lock);
470 return swp_entry(type, offset);
472 next = swap_list.next;
475 nr_swap_pages++;
476 noswap:
477 spin_unlock(&swap_lock);
478 return (swp_entry_t) {0};
481 /* The only caller of this function is now susupend routine */
482 swp_entry_t get_swap_page_of_type(int type)
484 struct swap_info_struct *si;
485 pgoff_t offset;
487 spin_lock(&swap_lock);
488 si = swap_info[type];
489 if (si && (si->flags & SWP_WRITEOK)) {
490 nr_swap_pages--;
491 /* This is called for allocating swap entry, not cache */
492 offset = scan_swap_map(si, 1);
493 if (offset) {
494 spin_unlock(&swap_lock);
495 return swp_entry(type, offset);
497 nr_swap_pages++;
499 spin_unlock(&swap_lock);
500 return (swp_entry_t) {0};
503 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
505 struct swap_info_struct *p;
506 unsigned long offset, type;
508 if (!entry.val)
509 goto out;
510 type = swp_type(entry);
511 if (type >= nr_swapfiles)
512 goto bad_nofile;
513 p = swap_info[type];
514 if (!(p->flags & SWP_USED))
515 goto bad_device;
516 offset = swp_offset(entry);
517 if (offset >= p->max)
518 goto bad_offset;
519 if (!p->swap_map[offset])
520 goto bad_free;
521 spin_lock(&swap_lock);
522 return p;
524 bad_free:
525 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
526 goto out;
527 bad_offset:
528 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
529 goto out;
530 bad_device:
531 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
532 goto out;
533 bad_nofile:
534 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
535 out:
536 return NULL;
539 static unsigned char swap_entry_free(struct swap_info_struct *p,
540 swp_entry_t entry, unsigned char usage)
542 unsigned long offset = swp_offset(entry);
543 unsigned char count;
544 unsigned char has_cache;
546 count = p->swap_map[offset];
547 has_cache = count & SWAP_HAS_CACHE;
548 count &= ~SWAP_HAS_CACHE;
550 if (usage == SWAP_HAS_CACHE) {
551 VM_BUG_ON(!has_cache);
552 has_cache = 0;
553 } else if (count == SWAP_MAP_SHMEM) {
555 * Or we could insist on shmem.c using a special
556 * swap_shmem_free() and free_shmem_swap_and_cache()...
558 count = 0;
559 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
560 if (count == COUNT_CONTINUED) {
561 if (swap_count_continued(p, offset, count))
562 count = SWAP_MAP_MAX | COUNT_CONTINUED;
563 else
564 count = SWAP_MAP_MAX;
565 } else
566 count--;
569 if (!count)
570 mem_cgroup_uncharge_swap(entry);
572 usage = count | has_cache;
573 p->swap_map[offset] = usage;
575 /* free if no reference */
576 if (!usage) {
577 struct gendisk *disk = p->bdev->bd_disk;
578 if (offset < p->lowest_bit)
579 p->lowest_bit = offset;
580 if (offset > p->highest_bit)
581 p->highest_bit = offset;
582 if (swap_list.next >= 0 &&
583 p->prio > swap_info[swap_list.next]->prio)
584 swap_list.next = p->type;
585 nr_swap_pages++;
586 p->inuse_pages--;
587 if ((p->flags & SWP_BLKDEV) &&
588 disk->fops->swap_slot_free_notify)
589 disk->fops->swap_slot_free_notify(p->bdev, offset);
592 return usage;
596 * Caller has made sure that the swapdevice corresponding to entry
597 * is still around or has not been recycled.
599 void swap_free(swp_entry_t entry)
601 struct swap_info_struct *p;
603 p = swap_info_get(entry);
604 if (p) {
605 swap_entry_free(p, entry, 1);
606 spin_unlock(&swap_lock);
611 * Called after dropping swapcache to decrease refcnt to swap entries.
613 void swapcache_free(swp_entry_t entry, struct page *page)
615 struct swap_info_struct *p;
616 unsigned char count;
618 p = swap_info_get(entry);
619 if (p) {
620 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
621 if (page)
622 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
623 spin_unlock(&swap_lock);
628 * How many references to page are currently swapped out?
629 * This does not give an exact answer when swap count is continued,
630 * but does include the high COUNT_CONTINUED flag to allow for that.
632 static inline int page_swapcount(struct page *page)
634 int count = 0;
635 struct swap_info_struct *p;
636 swp_entry_t entry;
638 entry.val = page_private(page);
639 p = swap_info_get(entry);
640 if (p) {
641 count = swap_count(p->swap_map[swp_offset(entry)]);
642 spin_unlock(&swap_lock);
644 return count;
648 * We can write to an anon page without COW if there are no other references
649 * to it. And as a side-effect, free up its swap: because the old content
650 * on disk will never be read, and seeking back there to write new content
651 * later would only waste time away from clustering.
653 int reuse_swap_page(struct page *page)
655 int count;
657 VM_BUG_ON(!PageLocked(page));
658 if (unlikely(PageKsm(page)))
659 return 0;
660 count = page_mapcount(page);
661 if (count <= 1 && PageSwapCache(page)) {
662 count += page_swapcount(page);
663 if (count == 1 && !PageWriteback(page)) {
664 delete_from_swap_cache(page);
665 SetPageDirty(page);
668 return count <= 1;
672 * If swap is getting full, or if there are no more mappings of this page,
673 * then try_to_free_swap is called to free its swap space.
675 int try_to_free_swap(struct page *page)
677 VM_BUG_ON(!PageLocked(page));
679 if (!PageSwapCache(page))
680 return 0;
681 if (PageWriteback(page))
682 return 0;
683 if (page_swapcount(page))
684 return 0;
687 * Once hibernation has begun to create its image of memory,
688 * there's a danger that one of the calls to try_to_free_swap()
689 * - most probably a call from __try_to_reclaim_swap() while
690 * hibernation is allocating its own swap pages for the image,
691 * but conceivably even a call from memory reclaim - will free
692 * the swap from a page which has already been recorded in the
693 * image as a clean swapcache page, and then reuse its swap for
694 * another page of the image. On waking from hibernation, the
695 * original page might be freed under memory pressure, then
696 * later read back in from swap, now with the wrong data.
698 * Hibernation clears bits from gfp_allowed_mask to prevent
699 * memory reclaim from writing to disk, so check that here.
701 if (!(gfp_allowed_mask & __GFP_IO))
702 return 0;
704 delete_from_swap_cache(page);
705 SetPageDirty(page);
706 return 1;
710 * Free the swap entry like above, but also try to
711 * free the page cache entry if it is the last user.
713 int free_swap_and_cache(swp_entry_t entry)
715 struct swap_info_struct *p;
716 struct page *page = NULL;
718 if (non_swap_entry(entry))
719 return 1;
721 p = swap_info_get(entry);
722 if (p) {
723 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
724 page = find_get_page(&swapper_space, entry.val);
725 if (page && !trylock_page(page)) {
726 page_cache_release(page);
727 page = NULL;
730 spin_unlock(&swap_lock);
732 if (page) {
734 * Not mapped elsewhere, or swap space full? Free it!
735 * Also recheck PageSwapCache now page is locked (above).
737 if (PageSwapCache(page) && !PageWriteback(page) &&
738 (!page_mapped(page) || vm_swap_full())) {
739 delete_from_swap_cache(page);
740 SetPageDirty(page);
742 unlock_page(page);
743 page_cache_release(page);
745 return p != NULL;
748 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
750 * mem_cgroup_count_swap_user - count the user of a swap entry
751 * @ent: the swap entry to be checked
752 * @pagep: the pointer for the swap cache page of the entry to be stored
754 * Returns the number of the user of the swap entry. The number is valid only
755 * for swaps of anonymous pages.
756 * If the entry is found on swap cache, the page is stored to pagep with
757 * refcount of it being incremented.
759 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
761 struct page *page;
762 struct swap_info_struct *p;
763 int count = 0;
765 page = find_get_page(&swapper_space, ent.val);
766 if (page)
767 count += page_mapcount(page);
768 p = swap_info_get(ent);
769 if (p) {
770 count += swap_count(p->swap_map[swp_offset(ent)]);
771 spin_unlock(&swap_lock);
774 *pagep = page;
775 return count;
777 #endif
779 #ifdef CONFIG_HIBERNATION
781 * Find the swap type that corresponds to given device (if any).
783 * @offset - number of the PAGE_SIZE-sized block of the device, starting
784 * from 0, in which the swap header is expected to be located.
786 * This is needed for the suspend to disk (aka swsusp).
788 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
790 struct block_device *bdev = NULL;
791 int type;
793 if (device)
794 bdev = bdget(device);
796 spin_lock(&swap_lock);
797 for (type = 0; type < nr_swapfiles; type++) {
798 struct swap_info_struct *sis = swap_info[type];
800 if (!(sis->flags & SWP_WRITEOK))
801 continue;
803 if (!bdev) {
804 if (bdev_p)
805 *bdev_p = bdgrab(sis->bdev);
807 spin_unlock(&swap_lock);
808 return type;
810 if (bdev == sis->bdev) {
811 struct swap_extent *se = &sis->first_swap_extent;
813 if (se->start_block == offset) {
814 if (bdev_p)
815 *bdev_p = bdgrab(sis->bdev);
817 spin_unlock(&swap_lock);
818 bdput(bdev);
819 return type;
823 spin_unlock(&swap_lock);
824 if (bdev)
825 bdput(bdev);
827 return -ENODEV;
831 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
832 * corresponding to given index in swap_info (swap type).
834 sector_t swapdev_block(int type, pgoff_t offset)
836 struct block_device *bdev;
838 if ((unsigned int)type >= nr_swapfiles)
839 return 0;
840 if (!(swap_info[type]->flags & SWP_WRITEOK))
841 return 0;
842 return map_swap_entry(swp_entry(type, offset), &bdev);
846 * Return either the total number of swap pages of given type, or the number
847 * of free pages of that type (depending on @free)
849 * This is needed for software suspend
851 unsigned int count_swap_pages(int type, int free)
853 unsigned int n = 0;
855 spin_lock(&swap_lock);
856 if ((unsigned int)type < nr_swapfiles) {
857 struct swap_info_struct *sis = swap_info[type];
859 if (sis->flags & SWP_WRITEOK) {
860 n = sis->pages;
861 if (free)
862 n -= sis->inuse_pages;
865 spin_unlock(&swap_lock);
866 return n;
868 #endif /* CONFIG_HIBERNATION */
871 * No need to decide whether this PTE shares the swap entry with others,
872 * just let do_wp_page work it out if a write is requested later - to
873 * force COW, vm_page_prot omits write permission from any private vma.
875 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
876 unsigned long addr, swp_entry_t entry, struct page *page)
878 struct mem_cgroup *ptr = NULL;
879 spinlock_t *ptl;
880 pte_t *pte;
881 int ret = 1;
883 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
884 ret = -ENOMEM;
885 goto out_nolock;
888 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
889 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
890 if (ret > 0)
891 mem_cgroup_cancel_charge_swapin(ptr);
892 ret = 0;
893 goto out;
896 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
897 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
898 get_page(page);
899 set_pte_at(vma->vm_mm, addr, pte,
900 pte_mkold(mk_pte(page, vma->vm_page_prot)));
901 page_add_anon_rmap(page, vma, addr);
902 mem_cgroup_commit_charge_swapin(page, ptr);
903 swap_free(entry);
905 * Move the page to the active list so it is not
906 * immediately swapped out again after swapon.
908 activate_page(page);
909 out:
910 pte_unmap_unlock(pte, ptl);
911 out_nolock:
912 return ret;
915 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
916 unsigned long addr, unsigned long end,
917 swp_entry_t entry, struct page *page)
919 pte_t swp_pte = swp_entry_to_pte(entry);
920 pte_t *pte;
921 int ret = 0;
924 * We don't actually need pte lock while scanning for swp_pte: since
925 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
926 * page table while we're scanning; though it could get zapped, and on
927 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
928 * of unmatched parts which look like swp_pte, so unuse_pte must
929 * recheck under pte lock. Scanning without pte lock lets it be
930 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
932 pte = pte_offset_map(pmd, addr);
933 do {
935 * swapoff spends a _lot_ of time in this loop!
936 * Test inline before going to call unuse_pte.
938 if (unlikely(pte_same(*pte, swp_pte))) {
939 pte_unmap(pte);
940 ret = unuse_pte(vma, pmd, addr, entry, page);
941 if (ret)
942 goto out;
943 pte = pte_offset_map(pmd, addr);
945 } while (pte++, addr += PAGE_SIZE, addr != end);
946 pte_unmap(pte - 1);
947 out:
948 return ret;
951 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
952 unsigned long addr, unsigned long end,
953 swp_entry_t entry, struct page *page)
955 pmd_t *pmd;
956 unsigned long next;
957 int ret;
959 pmd = pmd_offset(pud, addr);
960 do {
961 next = pmd_addr_end(addr, end);
962 if (pmd_none_or_clear_bad(pmd))
963 continue;
964 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
965 if (ret)
966 return ret;
967 } while (pmd++, addr = next, addr != end);
968 return 0;
971 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
972 unsigned long addr, unsigned long end,
973 swp_entry_t entry, struct page *page)
975 pud_t *pud;
976 unsigned long next;
977 int ret;
979 pud = pud_offset(pgd, addr);
980 do {
981 next = pud_addr_end(addr, end);
982 if (pud_none_or_clear_bad(pud))
983 continue;
984 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
985 if (ret)
986 return ret;
987 } while (pud++, addr = next, addr != end);
988 return 0;
991 static int unuse_vma(struct vm_area_struct *vma,
992 swp_entry_t entry, struct page *page)
994 pgd_t *pgd;
995 unsigned long addr, end, next;
996 int ret;
998 if (page_anon_vma(page)) {
999 addr = page_address_in_vma(page, vma);
1000 if (addr == -EFAULT)
1001 return 0;
1002 else
1003 end = addr + PAGE_SIZE;
1004 } else {
1005 addr = vma->vm_start;
1006 end = vma->vm_end;
1009 pgd = pgd_offset(vma->vm_mm, addr);
1010 do {
1011 next = pgd_addr_end(addr, end);
1012 if (pgd_none_or_clear_bad(pgd))
1013 continue;
1014 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1015 if (ret)
1016 return ret;
1017 } while (pgd++, addr = next, addr != end);
1018 return 0;
1021 static int unuse_mm(struct mm_struct *mm,
1022 swp_entry_t entry, struct page *page)
1024 struct vm_area_struct *vma;
1025 int ret = 0;
1027 if (!down_read_trylock(&mm->mmap_sem)) {
1029 * Activate page so shrink_inactive_list is unlikely to unmap
1030 * its ptes while lock is dropped, so swapoff can make progress.
1032 activate_page(page);
1033 unlock_page(page);
1034 down_read(&mm->mmap_sem);
1035 lock_page(page);
1037 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1038 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1039 break;
1041 up_read(&mm->mmap_sem);
1042 return (ret < 0)? ret: 0;
1046 * Scan swap_map from current position to next entry still in use.
1047 * Recycle to start on reaching the end, returning 0 when empty.
1049 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1050 unsigned int prev)
1052 unsigned int max = si->max;
1053 unsigned int i = prev;
1054 unsigned char count;
1057 * No need for swap_lock here: we're just looking
1058 * for whether an entry is in use, not modifying it; false
1059 * hits are okay, and sys_swapoff() has already prevented new
1060 * allocations from this area (while holding swap_lock).
1062 for (;;) {
1063 if (++i >= max) {
1064 if (!prev) {
1065 i = 0;
1066 break;
1069 * No entries in use at top of swap_map,
1070 * loop back to start and recheck there.
1072 max = prev + 1;
1073 prev = 0;
1074 i = 1;
1076 count = si->swap_map[i];
1077 if (count && swap_count(count) != SWAP_MAP_BAD)
1078 break;
1080 return i;
1084 * We completely avoid races by reading each swap page in advance,
1085 * and then search for the process using it. All the necessary
1086 * page table adjustments can then be made atomically.
1088 static int try_to_unuse(unsigned int type)
1090 struct swap_info_struct *si = swap_info[type];
1091 struct mm_struct *start_mm;
1092 unsigned char *swap_map;
1093 unsigned char swcount;
1094 struct page *page;
1095 swp_entry_t entry;
1096 unsigned int i = 0;
1097 int retval = 0;
1100 * When searching mms for an entry, a good strategy is to
1101 * start at the first mm we freed the previous entry from
1102 * (though actually we don't notice whether we or coincidence
1103 * freed the entry). Initialize this start_mm with a hold.
1105 * A simpler strategy would be to start at the last mm we
1106 * freed the previous entry from; but that would take less
1107 * advantage of mmlist ordering, which clusters forked mms
1108 * together, child after parent. If we race with dup_mmap(), we
1109 * prefer to resolve parent before child, lest we miss entries
1110 * duplicated after we scanned child: using last mm would invert
1111 * that.
1113 start_mm = &init_mm;
1114 atomic_inc(&init_mm.mm_users);
1117 * Keep on scanning until all entries have gone. Usually,
1118 * one pass through swap_map is enough, but not necessarily:
1119 * there are races when an instance of an entry might be missed.
1121 while ((i = find_next_to_unuse(si, i)) != 0) {
1122 if (signal_pending(current)) {
1123 retval = -EINTR;
1124 break;
1128 * Get a page for the entry, using the existing swap
1129 * cache page if there is one. Otherwise, get a clean
1130 * page and read the swap into it.
1132 swap_map = &si->swap_map[i];
1133 entry = swp_entry(type, i);
1134 page = read_swap_cache_async(entry,
1135 GFP_HIGHUSER_MOVABLE, NULL, 0);
1136 if (!page) {
1138 * Either swap_duplicate() failed because entry
1139 * has been freed independently, and will not be
1140 * reused since sys_swapoff() already disabled
1141 * allocation from here, or alloc_page() failed.
1143 if (!*swap_map)
1144 continue;
1145 retval = -ENOMEM;
1146 break;
1150 * Don't hold on to start_mm if it looks like exiting.
1152 if (atomic_read(&start_mm->mm_users) == 1) {
1153 mmput(start_mm);
1154 start_mm = &init_mm;
1155 atomic_inc(&init_mm.mm_users);
1159 * Wait for and lock page. When do_swap_page races with
1160 * try_to_unuse, do_swap_page can handle the fault much
1161 * faster than try_to_unuse can locate the entry. This
1162 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1163 * defer to do_swap_page in such a case - in some tests,
1164 * do_swap_page and try_to_unuse repeatedly compete.
1166 wait_on_page_locked(page);
1167 wait_on_page_writeback(page);
1168 lock_page(page);
1169 wait_on_page_writeback(page);
1172 * Remove all references to entry.
1174 swcount = *swap_map;
1175 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1176 retval = shmem_unuse(entry, page);
1177 /* page has already been unlocked and released */
1178 if (retval < 0)
1179 break;
1180 continue;
1182 if (swap_count(swcount) && start_mm != &init_mm)
1183 retval = unuse_mm(start_mm, entry, page);
1185 if (swap_count(*swap_map)) {
1186 int set_start_mm = (*swap_map >= swcount);
1187 struct list_head *p = &start_mm->mmlist;
1188 struct mm_struct *new_start_mm = start_mm;
1189 struct mm_struct *prev_mm = start_mm;
1190 struct mm_struct *mm;
1192 atomic_inc(&new_start_mm->mm_users);
1193 atomic_inc(&prev_mm->mm_users);
1194 spin_lock(&mmlist_lock);
1195 while (swap_count(*swap_map) && !retval &&
1196 (p = p->next) != &start_mm->mmlist) {
1197 mm = list_entry(p, struct mm_struct, mmlist);
1198 if (!atomic_inc_not_zero(&mm->mm_users))
1199 continue;
1200 spin_unlock(&mmlist_lock);
1201 mmput(prev_mm);
1202 prev_mm = mm;
1204 cond_resched();
1206 swcount = *swap_map;
1207 if (!swap_count(swcount)) /* any usage ? */
1209 else if (mm == &init_mm)
1210 set_start_mm = 1;
1211 else
1212 retval = unuse_mm(mm, entry, page);
1214 if (set_start_mm && *swap_map < swcount) {
1215 mmput(new_start_mm);
1216 atomic_inc(&mm->mm_users);
1217 new_start_mm = mm;
1218 set_start_mm = 0;
1220 spin_lock(&mmlist_lock);
1222 spin_unlock(&mmlist_lock);
1223 mmput(prev_mm);
1224 mmput(start_mm);
1225 start_mm = new_start_mm;
1227 if (retval) {
1228 unlock_page(page);
1229 page_cache_release(page);
1230 break;
1234 * If a reference remains (rare), we would like to leave
1235 * the page in the swap cache; but try_to_unmap could
1236 * then re-duplicate the entry once we drop page lock,
1237 * so we might loop indefinitely; also, that page could
1238 * not be swapped out to other storage meanwhile. So:
1239 * delete from cache even if there's another reference,
1240 * after ensuring that the data has been saved to disk -
1241 * since if the reference remains (rarer), it will be
1242 * read from disk into another page. Splitting into two
1243 * pages would be incorrect if swap supported "shared
1244 * private" pages, but they are handled by tmpfs files.
1246 * Given how unuse_vma() targets one particular offset
1247 * in an anon_vma, once the anon_vma has been determined,
1248 * this splitting happens to be just what is needed to
1249 * handle where KSM pages have been swapped out: re-reading
1250 * is unnecessarily slow, but we can fix that later on.
1252 if (swap_count(*swap_map) &&
1253 PageDirty(page) && PageSwapCache(page)) {
1254 struct writeback_control wbc = {
1255 .sync_mode = WB_SYNC_NONE,
1258 swap_writepage(page, &wbc);
1259 lock_page(page);
1260 wait_on_page_writeback(page);
1264 * It is conceivable that a racing task removed this page from
1265 * swap cache just before we acquired the page lock at the top,
1266 * or while we dropped it in unuse_mm(). The page might even
1267 * be back in swap cache on another swap area: that we must not
1268 * delete, since it may not have been written out to swap yet.
1270 if (PageSwapCache(page) &&
1271 likely(page_private(page) == entry.val))
1272 delete_from_swap_cache(page);
1275 * So we could skip searching mms once swap count went
1276 * to 1, we did not mark any present ptes as dirty: must
1277 * mark page dirty so shrink_page_list will preserve it.
1279 SetPageDirty(page);
1280 unlock_page(page);
1281 page_cache_release(page);
1284 * Make sure that we aren't completely killing
1285 * interactive performance.
1287 cond_resched();
1290 mmput(start_mm);
1291 return retval;
1295 * After a successful try_to_unuse, if no swap is now in use, we know
1296 * we can empty the mmlist. swap_lock must be held on entry and exit.
1297 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1298 * added to the mmlist just after page_duplicate - before would be racy.
1300 static void drain_mmlist(void)
1302 struct list_head *p, *next;
1303 unsigned int type;
1305 for (type = 0; type < nr_swapfiles; type++)
1306 if (swap_info[type]->inuse_pages)
1307 return;
1308 spin_lock(&mmlist_lock);
1309 list_for_each_safe(p, next, &init_mm.mmlist)
1310 list_del_init(p);
1311 spin_unlock(&mmlist_lock);
1315 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1316 * corresponds to page offset for the specified swap entry.
1317 * Note that the type of this function is sector_t, but it returns page offset
1318 * into the bdev, not sector offset.
1320 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1322 struct swap_info_struct *sis;
1323 struct swap_extent *start_se;
1324 struct swap_extent *se;
1325 pgoff_t offset;
1327 sis = swap_info[swp_type(entry)];
1328 *bdev = sis->bdev;
1330 offset = swp_offset(entry);
1331 start_se = sis->curr_swap_extent;
1332 se = start_se;
1334 for ( ; ; ) {
1335 struct list_head *lh;
1337 if (se->start_page <= offset &&
1338 offset < (se->start_page + se->nr_pages)) {
1339 return se->start_block + (offset - se->start_page);
1341 lh = se->list.next;
1342 se = list_entry(lh, struct swap_extent, list);
1343 sis->curr_swap_extent = se;
1344 BUG_ON(se == start_se); /* It *must* be present */
1349 * Returns the page offset into bdev for the specified page's swap entry.
1351 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1353 swp_entry_t entry;
1354 entry.val = page_private(page);
1355 return map_swap_entry(entry, bdev);
1359 * Free all of a swapdev's extent information
1361 static void destroy_swap_extents(struct swap_info_struct *sis)
1363 while (!list_empty(&sis->first_swap_extent.list)) {
1364 struct swap_extent *se;
1366 se = list_entry(sis->first_swap_extent.list.next,
1367 struct swap_extent, list);
1368 list_del(&se->list);
1369 kfree(se);
1374 * Add a block range (and the corresponding page range) into this swapdev's
1375 * extent list. The extent list is kept sorted in page order.
1377 * This function rather assumes that it is called in ascending page order.
1379 static int
1380 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1381 unsigned long nr_pages, sector_t start_block)
1383 struct swap_extent *se;
1384 struct swap_extent *new_se;
1385 struct list_head *lh;
1387 if (start_page == 0) {
1388 se = &sis->first_swap_extent;
1389 sis->curr_swap_extent = se;
1390 se->start_page = 0;
1391 se->nr_pages = nr_pages;
1392 se->start_block = start_block;
1393 return 1;
1394 } else {
1395 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1396 se = list_entry(lh, struct swap_extent, list);
1397 BUG_ON(se->start_page + se->nr_pages != start_page);
1398 if (se->start_block + se->nr_pages == start_block) {
1399 /* Merge it */
1400 se->nr_pages += nr_pages;
1401 return 0;
1406 * No merge. Insert a new extent, preserving ordering.
1408 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1409 if (new_se == NULL)
1410 return -ENOMEM;
1411 new_se->start_page = start_page;
1412 new_se->nr_pages = nr_pages;
1413 new_se->start_block = start_block;
1415 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1416 return 1;
1420 * A `swap extent' is a simple thing which maps a contiguous range of pages
1421 * onto a contiguous range of disk blocks. An ordered list of swap extents
1422 * is built at swapon time and is then used at swap_writepage/swap_readpage
1423 * time for locating where on disk a page belongs.
1425 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1426 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1427 * swap files identically.
1429 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1430 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1431 * swapfiles are handled *identically* after swapon time.
1433 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1434 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1435 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1436 * requirements, they are simply tossed out - we will never use those blocks
1437 * for swapping.
1439 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1440 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1441 * which will scribble on the fs.
1443 * The amount of disk space which a single swap extent represents varies.
1444 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1445 * extents in the list. To avoid much list walking, we cache the previous
1446 * search location in `curr_swap_extent', and start new searches from there.
1447 * This is extremely effective. The average number of iterations in
1448 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1450 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1452 struct inode *inode;
1453 unsigned blocks_per_page;
1454 unsigned long page_no;
1455 unsigned blkbits;
1456 sector_t probe_block;
1457 sector_t last_block;
1458 sector_t lowest_block = -1;
1459 sector_t highest_block = 0;
1460 int nr_extents = 0;
1461 int ret;
1463 inode = sis->swap_file->f_mapping->host;
1464 if (S_ISBLK(inode->i_mode)) {
1465 ret = add_swap_extent(sis, 0, sis->max, 0);
1466 *span = sis->pages;
1467 goto out;
1470 blkbits = inode->i_blkbits;
1471 blocks_per_page = PAGE_SIZE >> blkbits;
1474 * Map all the blocks into the extent list. This code doesn't try
1475 * to be very smart.
1477 probe_block = 0;
1478 page_no = 0;
1479 last_block = i_size_read(inode) >> blkbits;
1480 while ((probe_block + blocks_per_page) <= last_block &&
1481 page_no < sis->max) {
1482 unsigned block_in_page;
1483 sector_t first_block;
1485 first_block = bmap(inode, probe_block);
1486 if (first_block == 0)
1487 goto bad_bmap;
1490 * It must be PAGE_SIZE aligned on-disk
1492 if (first_block & (blocks_per_page - 1)) {
1493 probe_block++;
1494 goto reprobe;
1497 for (block_in_page = 1; block_in_page < blocks_per_page;
1498 block_in_page++) {
1499 sector_t block;
1501 block = bmap(inode, probe_block + block_in_page);
1502 if (block == 0)
1503 goto bad_bmap;
1504 if (block != first_block + block_in_page) {
1505 /* Discontiguity */
1506 probe_block++;
1507 goto reprobe;
1511 first_block >>= (PAGE_SHIFT - blkbits);
1512 if (page_no) { /* exclude the header page */
1513 if (first_block < lowest_block)
1514 lowest_block = first_block;
1515 if (first_block > highest_block)
1516 highest_block = first_block;
1520 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1522 ret = add_swap_extent(sis, page_no, 1, first_block);
1523 if (ret < 0)
1524 goto out;
1525 nr_extents += ret;
1526 page_no++;
1527 probe_block += blocks_per_page;
1528 reprobe:
1529 continue;
1531 ret = nr_extents;
1532 *span = 1 + highest_block - lowest_block;
1533 if (page_no == 0)
1534 page_no = 1; /* force Empty message */
1535 sis->max = page_no;
1536 sis->pages = page_no - 1;
1537 sis->highest_bit = page_no - 1;
1538 out:
1539 return ret;
1540 bad_bmap:
1541 printk(KERN_ERR "swapon: swapfile has holes\n");
1542 ret = -EINVAL;
1543 goto out;
1546 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1548 struct swap_info_struct *p = NULL;
1549 unsigned char *swap_map;
1550 struct file *swap_file, *victim;
1551 struct address_space *mapping;
1552 struct inode *inode;
1553 char *pathname;
1554 int i, type, prev;
1555 int err;
1557 if (!capable(CAP_SYS_ADMIN))
1558 return -EPERM;
1560 pathname = getname(specialfile);
1561 err = PTR_ERR(pathname);
1562 if (IS_ERR(pathname))
1563 goto out;
1565 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1566 putname(pathname);
1567 err = PTR_ERR(victim);
1568 if (IS_ERR(victim))
1569 goto out;
1571 mapping = victim->f_mapping;
1572 prev = -1;
1573 spin_lock(&swap_lock);
1574 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1575 p = swap_info[type];
1576 if (p->flags & SWP_WRITEOK) {
1577 if (p->swap_file->f_mapping == mapping)
1578 break;
1580 prev = type;
1582 if (type < 0) {
1583 err = -EINVAL;
1584 spin_unlock(&swap_lock);
1585 goto out_dput;
1587 if (!security_vm_enough_memory(p->pages))
1588 vm_unacct_memory(p->pages);
1589 else {
1590 err = -ENOMEM;
1591 spin_unlock(&swap_lock);
1592 goto out_dput;
1594 if (prev < 0)
1595 swap_list.head = p->next;
1596 else
1597 swap_info[prev]->next = p->next;
1598 if (type == swap_list.next) {
1599 /* just pick something that's safe... */
1600 swap_list.next = swap_list.head;
1602 if (p->prio < 0) {
1603 for (i = p->next; i >= 0; i = swap_info[i]->next)
1604 swap_info[i]->prio = p->prio--;
1605 least_priority++;
1607 nr_swap_pages -= p->pages;
1608 total_swap_pages -= p->pages;
1609 p->flags &= ~SWP_WRITEOK;
1610 spin_unlock(&swap_lock);
1612 current->flags |= PF_OOM_ORIGIN;
1613 err = try_to_unuse(type);
1614 current->flags &= ~PF_OOM_ORIGIN;
1616 if (err) {
1617 /* re-insert swap space back into swap_list */
1618 spin_lock(&swap_lock);
1619 if (p->prio < 0)
1620 p->prio = --least_priority;
1621 prev = -1;
1622 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1623 if (p->prio >= swap_info[i]->prio)
1624 break;
1625 prev = i;
1627 p->next = i;
1628 if (prev < 0)
1629 swap_list.head = swap_list.next = type;
1630 else
1631 swap_info[prev]->next = type;
1632 nr_swap_pages += p->pages;
1633 total_swap_pages += p->pages;
1634 p->flags |= SWP_WRITEOK;
1635 spin_unlock(&swap_lock);
1636 goto out_dput;
1639 /* wait for any unplug function to finish */
1640 down_write(&swap_unplug_sem);
1641 up_write(&swap_unplug_sem);
1643 destroy_swap_extents(p);
1644 if (p->flags & SWP_CONTINUED)
1645 free_swap_count_continuations(p);
1647 mutex_lock(&swapon_mutex);
1648 spin_lock(&swap_lock);
1649 drain_mmlist();
1651 /* wait for anyone still in scan_swap_map */
1652 p->highest_bit = 0; /* cuts scans short */
1653 while (p->flags >= SWP_SCANNING) {
1654 spin_unlock(&swap_lock);
1655 schedule_timeout_uninterruptible(1);
1656 spin_lock(&swap_lock);
1659 swap_file = p->swap_file;
1660 p->swap_file = NULL;
1661 p->max = 0;
1662 swap_map = p->swap_map;
1663 p->swap_map = NULL;
1664 p->flags = 0;
1665 spin_unlock(&swap_lock);
1666 mutex_unlock(&swapon_mutex);
1667 vfree(swap_map);
1668 /* Destroy swap account informatin */
1669 swap_cgroup_swapoff(type);
1671 inode = mapping->host;
1672 if (S_ISBLK(inode->i_mode)) {
1673 struct block_device *bdev = I_BDEV(inode);
1674 set_blocksize(bdev, p->old_block_size);
1675 bd_release(bdev);
1676 } else {
1677 mutex_lock(&inode->i_mutex);
1678 inode->i_flags &= ~S_SWAPFILE;
1679 mutex_unlock(&inode->i_mutex);
1681 filp_close(swap_file, NULL);
1682 err = 0;
1684 out_dput:
1685 filp_close(victim, NULL);
1686 out:
1687 return err;
1690 #ifdef CONFIG_PROC_FS
1691 /* iterator */
1692 static void *swap_start(struct seq_file *swap, loff_t *pos)
1694 struct swap_info_struct *si;
1695 int type;
1696 loff_t l = *pos;
1698 mutex_lock(&swapon_mutex);
1700 if (!l)
1701 return SEQ_START_TOKEN;
1703 for (type = 0; type < nr_swapfiles; type++) {
1704 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1705 si = swap_info[type];
1706 if (!(si->flags & SWP_USED) || !si->swap_map)
1707 continue;
1708 if (!--l)
1709 return si;
1712 return NULL;
1715 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1717 struct swap_info_struct *si = v;
1718 int type;
1720 if (v == SEQ_START_TOKEN)
1721 type = 0;
1722 else
1723 type = si->type + 1;
1725 for (; type < nr_swapfiles; type++) {
1726 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1727 si = swap_info[type];
1728 if (!(si->flags & SWP_USED) || !si->swap_map)
1729 continue;
1730 ++*pos;
1731 return si;
1734 return NULL;
1737 static void swap_stop(struct seq_file *swap, void *v)
1739 mutex_unlock(&swapon_mutex);
1742 static int swap_show(struct seq_file *swap, void *v)
1744 struct swap_info_struct *si = v;
1745 struct file *file;
1746 int len;
1748 if (si == SEQ_START_TOKEN) {
1749 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1750 return 0;
1753 file = si->swap_file;
1754 len = seq_path(swap, &file->f_path, " \t\n\\");
1755 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1756 len < 40 ? 40 - len : 1, " ",
1757 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1758 "partition" : "file\t",
1759 si->pages << (PAGE_SHIFT - 10),
1760 si->inuse_pages << (PAGE_SHIFT - 10),
1761 si->prio);
1762 return 0;
1765 static const struct seq_operations swaps_op = {
1766 .start = swap_start,
1767 .next = swap_next,
1768 .stop = swap_stop,
1769 .show = swap_show
1772 static int swaps_open(struct inode *inode, struct file *file)
1774 return seq_open(file, &swaps_op);
1777 static const struct file_operations proc_swaps_operations = {
1778 .open = swaps_open,
1779 .read = seq_read,
1780 .llseek = seq_lseek,
1781 .release = seq_release,
1784 static int __init procswaps_init(void)
1786 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1787 return 0;
1789 __initcall(procswaps_init);
1790 #endif /* CONFIG_PROC_FS */
1792 #ifdef MAX_SWAPFILES_CHECK
1793 static int __init max_swapfiles_check(void)
1795 MAX_SWAPFILES_CHECK();
1796 return 0;
1798 late_initcall(max_swapfiles_check);
1799 #endif
1802 * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1804 * The swapon system call
1806 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1808 struct swap_info_struct *p;
1809 char *name = NULL;
1810 struct block_device *bdev = NULL;
1811 struct file *swap_file = NULL;
1812 struct address_space *mapping;
1813 unsigned int type;
1814 int i, prev;
1815 int error;
1816 union swap_header *swap_header;
1817 unsigned int nr_good_pages;
1818 int nr_extents = 0;
1819 sector_t span;
1820 unsigned long maxpages;
1821 unsigned long swapfilepages;
1822 unsigned char *swap_map = NULL;
1823 struct page *page = NULL;
1824 struct inode *inode = NULL;
1825 int did_down = 0;
1827 if (!capable(CAP_SYS_ADMIN))
1828 return -EPERM;
1830 p = kzalloc(sizeof(*p), GFP_KERNEL);
1831 if (!p)
1832 return -ENOMEM;
1834 spin_lock(&swap_lock);
1835 for (type = 0; type < nr_swapfiles; type++) {
1836 if (!(swap_info[type]->flags & SWP_USED))
1837 break;
1839 error = -EPERM;
1840 if (type >= MAX_SWAPFILES) {
1841 spin_unlock(&swap_lock);
1842 kfree(p);
1843 goto out;
1845 if (type >= nr_swapfiles) {
1846 p->type = type;
1847 swap_info[type] = p;
1849 * Write swap_info[type] before nr_swapfiles, in case a
1850 * racing procfs swap_start() or swap_next() is reading them.
1851 * (We never shrink nr_swapfiles, we never free this entry.)
1853 smp_wmb();
1854 nr_swapfiles++;
1855 } else {
1856 kfree(p);
1857 p = swap_info[type];
1859 * Do not memset this entry: a racing procfs swap_next()
1860 * would be relying on p->type to remain valid.
1863 INIT_LIST_HEAD(&p->first_swap_extent.list);
1864 p->flags = SWP_USED;
1865 p->next = -1;
1866 spin_unlock(&swap_lock);
1868 name = getname(specialfile);
1869 error = PTR_ERR(name);
1870 if (IS_ERR(name)) {
1871 name = NULL;
1872 goto bad_swap_2;
1874 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1875 error = PTR_ERR(swap_file);
1876 if (IS_ERR(swap_file)) {
1877 swap_file = NULL;
1878 goto bad_swap_2;
1881 p->swap_file = swap_file;
1882 mapping = swap_file->f_mapping;
1883 inode = mapping->host;
1885 error = -EBUSY;
1886 for (i = 0; i < nr_swapfiles; i++) {
1887 struct swap_info_struct *q = swap_info[i];
1889 if (i == type || !q->swap_file)
1890 continue;
1891 if (mapping == q->swap_file->f_mapping)
1892 goto bad_swap;
1895 error = -EINVAL;
1896 if (S_ISBLK(inode->i_mode)) {
1897 bdev = I_BDEV(inode);
1898 error = bd_claim(bdev, sys_swapon);
1899 if (error < 0) {
1900 bdev = NULL;
1901 error = -EINVAL;
1902 goto bad_swap;
1904 p->old_block_size = block_size(bdev);
1905 error = set_blocksize(bdev, PAGE_SIZE);
1906 if (error < 0)
1907 goto bad_swap;
1908 p->bdev = bdev;
1909 p->flags |= SWP_BLKDEV;
1910 } else if (S_ISREG(inode->i_mode)) {
1911 p->bdev = inode->i_sb->s_bdev;
1912 mutex_lock(&inode->i_mutex);
1913 did_down = 1;
1914 if (IS_SWAPFILE(inode)) {
1915 error = -EBUSY;
1916 goto bad_swap;
1918 } else {
1919 goto bad_swap;
1922 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1925 * Read the swap header.
1927 if (!mapping->a_ops->readpage) {
1928 error = -EINVAL;
1929 goto bad_swap;
1931 page = read_mapping_page(mapping, 0, swap_file);
1932 if (IS_ERR(page)) {
1933 error = PTR_ERR(page);
1934 goto bad_swap;
1936 swap_header = kmap(page);
1938 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1939 printk(KERN_ERR "Unable to find swap-space signature\n");
1940 error = -EINVAL;
1941 goto bad_swap;
1944 /* swap partition endianess hack... */
1945 if (swab32(swap_header->info.version) == 1) {
1946 swab32s(&swap_header->info.version);
1947 swab32s(&swap_header->info.last_page);
1948 swab32s(&swap_header->info.nr_badpages);
1949 for (i = 0; i < swap_header->info.nr_badpages; i++)
1950 swab32s(&swap_header->info.badpages[i]);
1952 /* Check the swap header's sub-version */
1953 if (swap_header->info.version != 1) {
1954 printk(KERN_WARNING
1955 "Unable to handle swap header version %d\n",
1956 swap_header->info.version);
1957 error = -EINVAL;
1958 goto bad_swap;
1961 p->lowest_bit = 1;
1962 p->cluster_next = 1;
1963 p->cluster_nr = 0;
1966 * Find out how many pages are allowed for a single swap
1967 * device. There are two limiting factors: 1) the number of
1968 * bits for the swap offset in the swp_entry_t type and
1969 * 2) the number of bits in the a swap pte as defined by
1970 * the different architectures. In order to find the
1971 * largest possible bit mask a swap entry with swap type 0
1972 * and swap offset ~0UL is created, encoded to a swap pte,
1973 * decoded to a swp_entry_t again and finally the swap
1974 * offset is extracted. This will mask all the bits from
1975 * the initial ~0UL mask that can't be encoded in either
1976 * the swp_entry_t or the architecture definition of a
1977 * swap pte.
1979 maxpages = swp_offset(pte_to_swp_entry(
1980 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1981 if (maxpages > swap_header->info.last_page) {
1982 maxpages = swap_header->info.last_page + 1;
1983 /* p->max is an unsigned int: don't overflow it */
1984 if ((unsigned int)maxpages == 0)
1985 maxpages = UINT_MAX;
1987 p->highest_bit = maxpages - 1;
1989 error = -EINVAL;
1990 if (!maxpages)
1991 goto bad_swap;
1992 if (swapfilepages && maxpages > swapfilepages) {
1993 printk(KERN_WARNING
1994 "Swap area shorter than signature indicates\n");
1995 goto bad_swap;
1997 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1998 goto bad_swap;
1999 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2000 goto bad_swap;
2002 /* OK, set up the swap map and apply the bad block list */
2003 swap_map = vmalloc(maxpages);
2004 if (!swap_map) {
2005 error = -ENOMEM;
2006 goto bad_swap;
2009 memset(swap_map, 0, maxpages);
2010 nr_good_pages = maxpages - 1; /* omit header page */
2012 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2013 unsigned int page_nr = swap_header->info.badpages[i];
2014 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2015 error = -EINVAL;
2016 goto bad_swap;
2018 if (page_nr < maxpages) {
2019 swap_map[page_nr] = SWAP_MAP_BAD;
2020 nr_good_pages--;
2024 error = swap_cgroup_swapon(type, maxpages);
2025 if (error)
2026 goto bad_swap;
2028 if (nr_good_pages) {
2029 swap_map[0] = SWAP_MAP_BAD;
2030 p->max = maxpages;
2031 p->pages = nr_good_pages;
2032 nr_extents = setup_swap_extents(p, &span);
2033 if (nr_extents < 0) {
2034 error = nr_extents;
2035 goto bad_swap;
2037 nr_good_pages = p->pages;
2039 if (!nr_good_pages) {
2040 printk(KERN_WARNING "Empty swap-file\n");
2041 error = -EINVAL;
2042 goto bad_swap;
2045 if (p->bdev) {
2046 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2047 p->flags |= SWP_SOLIDSTATE;
2048 p->cluster_next = 1 + (random32() % p->highest_bit);
2050 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2051 p->flags |= SWP_DISCARDABLE;
2054 mutex_lock(&swapon_mutex);
2055 spin_lock(&swap_lock);
2056 if (swap_flags & SWAP_FLAG_PREFER)
2057 p->prio =
2058 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2059 else
2060 p->prio = --least_priority;
2061 p->swap_map = swap_map;
2062 p->flags |= SWP_WRITEOK;
2063 nr_swap_pages += nr_good_pages;
2064 total_swap_pages += nr_good_pages;
2066 printk(KERN_INFO "Adding %uk swap on %s. "
2067 "Priority:%d extents:%d across:%lluk %s%s\n",
2068 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2069 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2070 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2071 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2073 /* insert swap space into swap_list: */
2074 prev = -1;
2075 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2076 if (p->prio >= swap_info[i]->prio)
2077 break;
2078 prev = i;
2080 p->next = i;
2081 if (prev < 0)
2082 swap_list.head = swap_list.next = type;
2083 else
2084 swap_info[prev]->next = type;
2085 spin_unlock(&swap_lock);
2086 mutex_unlock(&swapon_mutex);
2087 error = 0;
2088 goto out;
2089 bad_swap:
2090 if (bdev) {
2091 set_blocksize(bdev, p->old_block_size);
2092 bd_release(bdev);
2094 destroy_swap_extents(p);
2095 swap_cgroup_swapoff(type);
2096 bad_swap_2:
2097 spin_lock(&swap_lock);
2098 p->swap_file = NULL;
2099 p->flags = 0;
2100 spin_unlock(&swap_lock);
2101 vfree(swap_map);
2102 if (swap_file)
2103 filp_close(swap_file, NULL);
2104 out:
2105 if (page && !IS_ERR(page)) {
2106 kunmap(page);
2107 page_cache_release(page);
2109 if (name)
2110 putname(name);
2111 if (did_down) {
2112 if (!error)
2113 inode->i_flags |= S_SWAPFILE;
2114 mutex_unlock(&inode->i_mutex);
2116 return error;
2119 void si_swapinfo(struct sysinfo *val)
2121 unsigned int type;
2122 unsigned long nr_to_be_unused = 0;
2124 spin_lock(&swap_lock);
2125 for (type = 0; type < nr_swapfiles; type++) {
2126 struct swap_info_struct *si = swap_info[type];
2128 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2129 nr_to_be_unused += si->inuse_pages;
2131 val->freeswap = nr_swap_pages + nr_to_be_unused;
2132 val->totalswap = total_swap_pages + nr_to_be_unused;
2133 spin_unlock(&swap_lock);
2137 * Verify that a swap entry is valid and increment its swap map count.
2139 * Returns error code in following case.
2140 * - success -> 0
2141 * - swp_entry is invalid -> EINVAL
2142 * - swp_entry is migration entry -> EINVAL
2143 * - swap-cache reference is requested but there is already one. -> EEXIST
2144 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2145 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2147 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2149 struct swap_info_struct *p;
2150 unsigned long offset, type;
2151 unsigned char count;
2152 unsigned char has_cache;
2153 int err = -EINVAL;
2155 if (non_swap_entry(entry))
2156 goto out;
2158 type = swp_type(entry);
2159 if (type >= nr_swapfiles)
2160 goto bad_file;
2161 p = swap_info[type];
2162 offset = swp_offset(entry);
2164 spin_lock(&swap_lock);
2165 if (unlikely(offset >= p->max))
2166 goto unlock_out;
2168 count = p->swap_map[offset];
2169 has_cache = count & SWAP_HAS_CACHE;
2170 count &= ~SWAP_HAS_CACHE;
2171 err = 0;
2173 if (usage == SWAP_HAS_CACHE) {
2175 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2176 if (!has_cache && count)
2177 has_cache = SWAP_HAS_CACHE;
2178 else if (has_cache) /* someone else added cache */
2179 err = -EEXIST;
2180 else /* no users remaining */
2181 err = -ENOENT;
2183 } else if (count || has_cache) {
2185 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2186 count += usage;
2187 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2188 err = -EINVAL;
2189 else if (swap_count_continued(p, offset, count))
2190 count = COUNT_CONTINUED;
2191 else
2192 err = -ENOMEM;
2193 } else
2194 err = -ENOENT; /* unused swap entry */
2196 p->swap_map[offset] = count | has_cache;
2198 unlock_out:
2199 spin_unlock(&swap_lock);
2200 out:
2201 return err;
2203 bad_file:
2204 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2205 goto out;
2209 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2210 * (in which case its reference count is never incremented).
2212 void swap_shmem_alloc(swp_entry_t entry)
2214 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2218 * Increase reference count of swap entry by 1.
2219 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2220 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2221 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2222 * might occur if a page table entry has got corrupted.
2224 int swap_duplicate(swp_entry_t entry)
2226 int err = 0;
2228 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2229 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2230 return err;
2234 * @entry: swap entry for which we allocate swap cache.
2236 * Called when allocating swap cache for existing swap entry,
2237 * This can return error codes. Returns 0 at success.
2238 * -EBUSY means there is a swap cache.
2239 * Note: return code is different from swap_duplicate().
2241 int swapcache_prepare(swp_entry_t entry)
2243 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2247 * swap_lock prevents swap_map being freed. Don't grab an extra
2248 * reference on the swaphandle, it doesn't matter if it becomes unused.
2250 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2252 struct swap_info_struct *si;
2253 int our_page_cluster = page_cluster;
2254 pgoff_t target, toff;
2255 pgoff_t base, end;
2256 int nr_pages = 0;
2258 if (!our_page_cluster) /* no readahead */
2259 return 0;
2261 si = swap_info[swp_type(entry)];
2262 target = swp_offset(entry);
2263 base = (target >> our_page_cluster) << our_page_cluster;
2264 end = base + (1 << our_page_cluster);
2265 if (!base) /* first page is swap header */
2266 base++;
2268 spin_lock(&swap_lock);
2269 if (end > si->max) /* don't go beyond end of map */
2270 end = si->max;
2272 /* Count contiguous allocated slots above our target */
2273 for (toff = target; ++toff < end; nr_pages++) {
2274 /* Don't read in free or bad pages */
2275 if (!si->swap_map[toff])
2276 break;
2277 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2278 break;
2280 /* Count contiguous allocated slots below our target */
2281 for (toff = target; --toff >= base; nr_pages++) {
2282 /* Don't read in free or bad pages */
2283 if (!si->swap_map[toff])
2284 break;
2285 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2286 break;
2288 spin_unlock(&swap_lock);
2291 * Indicate starting offset, and return number of pages to get:
2292 * if only 1, say 0, since there's then no readahead to be done.
2294 *offset = ++toff;
2295 return nr_pages? ++nr_pages: 0;
2299 * add_swap_count_continuation - called when a swap count is duplicated
2300 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2301 * page of the original vmalloc'ed swap_map, to hold the continuation count
2302 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2303 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2305 * These continuation pages are seldom referenced: the common paths all work
2306 * on the original swap_map, only referring to a continuation page when the
2307 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2309 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2310 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2311 * can be called after dropping locks.
2313 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2315 struct swap_info_struct *si;
2316 struct page *head;
2317 struct page *page;
2318 struct page *list_page;
2319 pgoff_t offset;
2320 unsigned char count;
2323 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2324 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2326 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2328 si = swap_info_get(entry);
2329 if (!si) {
2331 * An acceptable race has occurred since the failing
2332 * __swap_duplicate(): the swap entry has been freed,
2333 * perhaps even the whole swap_map cleared for swapoff.
2335 goto outer;
2338 offset = swp_offset(entry);
2339 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2341 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2343 * The higher the swap count, the more likely it is that tasks
2344 * will race to add swap count continuation: we need to avoid
2345 * over-provisioning.
2347 goto out;
2350 if (!page) {
2351 spin_unlock(&swap_lock);
2352 return -ENOMEM;
2356 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2357 * no architecture is using highmem pages for kernel pagetables: so it
2358 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2360 head = vmalloc_to_page(si->swap_map + offset);
2361 offset &= ~PAGE_MASK;
2364 * Page allocation does not initialize the page's lru field,
2365 * but it does always reset its private field.
2367 if (!page_private(head)) {
2368 BUG_ON(count & COUNT_CONTINUED);
2369 INIT_LIST_HEAD(&head->lru);
2370 set_page_private(head, SWP_CONTINUED);
2371 si->flags |= SWP_CONTINUED;
2374 list_for_each_entry(list_page, &head->lru, lru) {
2375 unsigned char *map;
2378 * If the previous map said no continuation, but we've found
2379 * a continuation page, free our allocation and use this one.
2381 if (!(count & COUNT_CONTINUED))
2382 goto out;
2384 map = kmap_atomic(list_page, KM_USER0) + offset;
2385 count = *map;
2386 kunmap_atomic(map, KM_USER0);
2389 * If this continuation count now has some space in it,
2390 * free our allocation and use this one.
2392 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2393 goto out;
2396 list_add_tail(&page->lru, &head->lru);
2397 page = NULL; /* now it's attached, don't free it */
2398 out:
2399 spin_unlock(&swap_lock);
2400 outer:
2401 if (page)
2402 __free_page(page);
2403 return 0;
2407 * swap_count_continued - when the original swap_map count is incremented
2408 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2409 * into, carry if so, or else fail until a new continuation page is allocated;
2410 * when the original swap_map count is decremented from 0 with continuation,
2411 * borrow from the continuation and report whether it still holds more.
2412 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2414 static bool swap_count_continued(struct swap_info_struct *si,
2415 pgoff_t offset, unsigned char count)
2417 struct page *head;
2418 struct page *page;
2419 unsigned char *map;
2421 head = vmalloc_to_page(si->swap_map + offset);
2422 if (page_private(head) != SWP_CONTINUED) {
2423 BUG_ON(count & COUNT_CONTINUED);
2424 return false; /* need to add count continuation */
2427 offset &= ~PAGE_MASK;
2428 page = list_entry(head->lru.next, struct page, lru);
2429 map = kmap_atomic(page, KM_USER0) + offset;
2431 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2432 goto init_map; /* jump over SWAP_CONT_MAX checks */
2434 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2436 * Think of how you add 1 to 999
2438 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2439 kunmap_atomic(map, KM_USER0);
2440 page = list_entry(page->lru.next, struct page, lru);
2441 BUG_ON(page == head);
2442 map = kmap_atomic(page, KM_USER0) + offset;
2444 if (*map == SWAP_CONT_MAX) {
2445 kunmap_atomic(map, KM_USER0);
2446 page = list_entry(page->lru.next, struct page, lru);
2447 if (page == head)
2448 return false; /* add count continuation */
2449 map = kmap_atomic(page, KM_USER0) + offset;
2450 init_map: *map = 0; /* we didn't zero the page */
2452 *map += 1;
2453 kunmap_atomic(map, KM_USER0);
2454 page = list_entry(page->lru.prev, struct page, lru);
2455 while (page != head) {
2456 map = kmap_atomic(page, KM_USER0) + offset;
2457 *map = COUNT_CONTINUED;
2458 kunmap_atomic(map, KM_USER0);
2459 page = list_entry(page->lru.prev, struct page, lru);
2461 return true; /* incremented */
2463 } else { /* decrementing */
2465 * Think of how you subtract 1 from 1000
2467 BUG_ON(count != COUNT_CONTINUED);
2468 while (*map == COUNT_CONTINUED) {
2469 kunmap_atomic(map, KM_USER0);
2470 page = list_entry(page->lru.next, struct page, lru);
2471 BUG_ON(page == head);
2472 map = kmap_atomic(page, KM_USER0) + offset;
2474 BUG_ON(*map == 0);
2475 *map -= 1;
2476 if (*map == 0)
2477 count = 0;
2478 kunmap_atomic(map, KM_USER0);
2479 page = list_entry(page->lru.prev, struct page, lru);
2480 while (page != head) {
2481 map = kmap_atomic(page, KM_USER0) + offset;
2482 *map = SWAP_CONT_MAX | count;
2483 count = COUNT_CONTINUED;
2484 kunmap_atomic(map, KM_USER0);
2485 page = list_entry(page->lru.prev, struct page, lru);
2487 return count == COUNT_CONTINUED;
2492 * free_swap_count_continuations - swapoff free all the continuation pages
2493 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2495 static void free_swap_count_continuations(struct swap_info_struct *si)
2497 pgoff_t offset;
2499 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2500 struct page *head;
2501 head = vmalloc_to_page(si->swap_map + offset);
2502 if (page_private(head)) {
2503 struct list_head *this, *next;
2504 list_for_each_safe(this, next, &head->lru) {
2505 struct page *page;
2506 page = list_entry(this, struct page, lru);
2507 list_del(this);
2508 __free_page(page);