Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
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>
33
34 #include <asm/pgtable.h>
35 #include <asm/tlbflush.h>
36 #include <linux/swapops.h>
37 #include <linux/page_cgroup.h>
38
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**);
43
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;
49
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 ";
54
55 static struct swap_list_t swap_list = {-1, -1};
56
57 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
58
59 static DEFINE_MUTEX(swapon_mutex);
60
61 static inline unsigned char swap_count(unsigned char ent)
62 {
63         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
64 }
65
66 /* returns 1 if swap entry is freed */
67 static int
68 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
69 {
70         swp_entry_t entry = swp_entry(si->type, offset);
71         struct page *page;
72         int ret = 0;
73
74         page = find_get_page(&swapper_space, entry.val);
75         if (!page)
76                 return 0;
77         /*
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.
83          */
84         if (trylock_page(page)) {
85                 ret = try_to_free_swap(page);
86                 unlock_page(page);
87         }
88         page_cache_release(page);
89         return ret;
90 }
91
92 /*
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.
96  */
97 static DECLARE_RWSEM(swap_unplug_sem);
98
99 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
100 {
101         swp_entry_t entry;
102
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;
108
109                 /*
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.
116                  */
117                 WARN_ON(page_count(page) <= 1);
118
119                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
120                 blk_run_backing_dev(bdi, page);
121         }
122         up_read(&swap_unplug_sem);
123 }
124
125 /*
126  * swapon tell device that all the old swap contents can be discarded,
127  * to allow the swap device to optimize its wear-levelling.
128  */
129 static int discard_swap(struct swap_info_struct *si)
130 {
131         struct swap_extent *se;
132         sector_t start_block;
133         sector_t nr_blocks;
134         int err = 0;
135
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, DISCARD_FL_BARRIER);
143                 if (err)
144                         return err;
145                 cond_resched();
146         }
147
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);
151
152                 err = blkdev_issue_discard(si->bdev, start_block,
153                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
154                 if (err)
155                         break;
156
157                 cond_resched();
158         }
159         return err;             /* That will often be -EOPNOTSUPP */
160 }
161
162 /*
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.
165  */
166 static void discard_swap_cluster(struct swap_info_struct *si,
167                                  pgoff_t start_page, pgoff_t nr_pages)
168 {
169         struct swap_extent *se = si->curr_swap_extent;
170         int found_extent = 0;
171
172         while (nr_pages) {
173                 struct list_head *lh;
174
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;
180
181                         if (nr_blocks > nr_pages)
182                                 nr_blocks = nr_pages;
183                         start_page += nr_blocks;
184                         nr_pages -= nr_blocks;
185
186                         if (!found_extent++)
187                                 si->curr_swap_extent = se;
188
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, DISCARD_FL_BARRIER))
193                                 break;
194                 }
195
196                 lh = se->list.next;
197                 se = list_entry(lh, struct swap_extent, list);
198         }
199 }
200
201 static int wait_for_discard(void *word)
202 {
203         schedule();
204         return 0;
205 }
206
207 #define SWAPFILE_CLUSTER        256
208 #define LATENCY_LIMIT           256
209
210 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
211                                           unsigned char usage)
212 {
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;
218
219         /*
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
228          */
229
230         si->flags += SWP_SCANNING;
231         scan_base = offset = si->cluster_next;
232
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;
237                 }
238                 if (si->flags & SWP_DISCARDABLE) {
239                         /*
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.
245                          */
246                         if (si->lowest_alloc)
247                                 goto checks;
248                         si->lowest_alloc = si->max;
249                         si->highest_alloc = 0;
250                 }
251                 spin_unlock(&swap_lock);
252
253                 /*
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.
260                  */
261                 if (!(si->flags & SWP_SOLIDSTATE))
262                         scan_base = offset = si->lowest_bit;
263                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
264
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;
276                         }
277                         if (unlikely(--latency_ration < 0)) {
278                                 cond_resched();
279                                 latency_ration = LATENCY_LIMIT;
280                         }
281                 }
282
283                 offset = si->lowest_bit;
284                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
285
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;
297                         }
298                         if (unlikely(--latency_ration < 0)) {
299                                 cond_resched();
300                                 latency_ration = LATENCY_LIMIT;
301                         }
302                 }
303
304                 offset = scan_base;
305                 spin_lock(&swap_lock);
306                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
307                 si->lowest_alloc = 0;
308         }
309
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;
317
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 */
328         }
329
330         if (si->swap_map[offset])
331                 goto scan;
332
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;
341         }
342         si->swap_map[offset] = usage;
343         si->cluster_next = offset + 1;
344         si->flags -= SWP_SCANNING;
345
346         if (si->lowest_alloc) {
347                 /*
348                  * Only set when SWP_DISCARDABLE, and there's a scan
349                  * for a free cluster in progress or just completed.
350                  */
351                 if (found_free_cluster) {
352                         /*
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).
358                          */
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);
364
365                         if (offset < last_in_cluster)
366                                 discard_swap_cluster(si, offset,
367                                         last_in_cluster - offset + 1);
368
369                         spin_lock(&swap_lock);
370                         si->lowest_alloc = 0;
371                         si->flags &= ~SWP_DISCARDING;
372
373                         smp_mb();       /* wake_up_bit advises this */
374                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
375
376                 } else if (si->flags & SWP_DISCARDING) {
377                         /*
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.
382                          */
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 {
388                         /*
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.
392                          */
393                         if (offset < si->lowest_alloc)
394                                 si->lowest_alloc = offset;
395                         if (offset > si->highest_alloc)
396                                 si->highest_alloc = offset;
397                 }
398         }
399         return offset;
400
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;
407                 }
408                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
409                         spin_lock(&swap_lock);
410                         goto checks;
411                 }
412                 if (unlikely(--latency_ration < 0)) {
413                         cond_resched();
414                         latency_ration = LATENCY_LIMIT;
415                 }
416         }
417         offset = si->lowest_bit;
418         while (++offset < scan_base) {
419                 if (!si->swap_map[offset]) {
420                         spin_lock(&swap_lock);
421                         goto checks;
422                 }
423                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
424                         spin_lock(&swap_lock);
425                         goto checks;
426                 }
427                 if (unlikely(--latency_ration < 0)) {
428                         cond_resched();
429                         latency_ration = LATENCY_LIMIT;
430                 }
431         }
432         spin_lock(&swap_lock);
433
434 no_page:
435         si->flags -= SWP_SCANNING;
436         return 0;
437 }
438
439 swp_entry_t get_swap_page(void)
440 {
441         struct swap_info_struct *si;
442         pgoff_t offset;
443         int type, next;
444         int wrapped = 0;
445
446         spin_lock(&swap_lock);
447         if (nr_swap_pages <= 0)
448                 goto noswap;
449         nr_swap_pages--;
450
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++;
458                 }
459
460                 if (!si->highest_bit)
461                         continue;
462                 if (!(si->flags & SWP_WRITEOK))
463                         continue;
464
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);
471                 }
472                 next = swap_list.next;
473         }
474
475         nr_swap_pages++;
476 noswap:
477         spin_unlock(&swap_lock);
478         return (swp_entry_t) {0};
479 }
480
481 /* The only caller of this function is now susupend routine */
482 swp_entry_t get_swap_page_of_type(int type)
483 {
484         struct swap_info_struct *si;
485         pgoff_t offset;
486
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);
496                 }
497                 nr_swap_pages++;
498         }
499         spin_unlock(&swap_lock);
500         return (swp_entry_t) {0};
501 }
502
503 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
504 {
505         struct swap_info_struct *p;
506         unsigned long offset, type;
507
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;
523
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;
537 }
538
539 static unsigned char swap_entry_free(struct swap_info_struct *p,
540                                      swp_entry_t entry, unsigned char usage)
541 {
542         unsigned long offset = swp_offset(entry);
543         unsigned char count;
544         unsigned char has_cache;
545
546         count = p->swap_map[offset];
547         has_cache = count & SWAP_HAS_CACHE;
548         count &= ~SWAP_HAS_CACHE;
549
550         if (usage == SWAP_HAS_CACHE) {
551                 VM_BUG_ON(!has_cache);
552                 has_cache = 0;
553         } else if (count == SWAP_MAP_SHMEM) {
554                 /*
555                  * Or we could insist on shmem.c using a special
556                  * swap_shmem_free() and free_shmem_swap_and_cache()...
557                  */
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--;
567         }
568
569         if (!count)
570                 mem_cgroup_uncharge_swap(entry);
571
572         usage = count | has_cache;
573         p->swap_map[offset] = usage;
574
575         /* free if no reference */
576         if (!usage) {
577                 if (offset < p->lowest_bit)
578                         p->lowest_bit = offset;
579                 if (offset > p->highest_bit)
580                         p->highest_bit = offset;
581                 if (swap_list.next >= 0 &&
582                     p->prio > swap_info[swap_list.next]->prio)
583                         swap_list.next = p->type;
584                 nr_swap_pages++;
585                 p->inuse_pages--;
586         }
587
588         return usage;
589 }
590
591 /*
592  * Caller has made sure that the swapdevice corresponding to entry
593  * is still around or has not been recycled.
594  */
595 void swap_free(swp_entry_t entry)
596 {
597         struct swap_info_struct *p;
598
599         p = swap_info_get(entry);
600         if (p) {
601                 swap_entry_free(p, entry, 1);
602                 spin_unlock(&swap_lock);
603         }
604 }
605
606 /*
607  * Called after dropping swapcache to decrease refcnt to swap entries.
608  */
609 void swapcache_free(swp_entry_t entry, struct page *page)
610 {
611         struct swap_info_struct *p;
612         unsigned char count;
613
614         p = swap_info_get(entry);
615         if (p) {
616                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
617                 if (page)
618                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
619                 spin_unlock(&swap_lock);
620         }
621 }
622
623 /*
624  * How many references to page are currently swapped out?
625  * This does not give an exact answer when swap count is continued,
626  * but does include the high COUNT_CONTINUED flag to allow for that.
627  */
628 static inline int page_swapcount(struct page *page)
629 {
630         int count = 0;
631         struct swap_info_struct *p;
632         swp_entry_t entry;
633
634         entry.val = page_private(page);
635         p = swap_info_get(entry);
636         if (p) {
637                 count = swap_count(p->swap_map[swp_offset(entry)]);
638                 spin_unlock(&swap_lock);
639         }
640         return count;
641 }
642
643 /*
644  * We can write to an anon page without COW if there are no other references
645  * to it.  And as a side-effect, free up its swap: because the old content
646  * on disk will never be read, and seeking back there to write new content
647  * later would only waste time away from clustering.
648  */
649 int reuse_swap_page(struct page *page)
650 {
651         int count;
652
653         VM_BUG_ON(!PageLocked(page));
654         if (unlikely(PageKsm(page)))
655                 return 0;
656         count = page_mapcount(page);
657         if (count <= 1 && PageSwapCache(page)) {
658                 count += page_swapcount(page);
659                 if (count == 1 && !PageWriteback(page)) {
660                         delete_from_swap_cache(page);
661                         SetPageDirty(page);
662                 }
663         }
664         return count <= 1;
665 }
666
667 /*
668  * If swap is getting full, or if there are no more mappings of this page,
669  * then try_to_free_swap is called to free its swap space.
670  */
671 int try_to_free_swap(struct page *page)
672 {
673         VM_BUG_ON(!PageLocked(page));
674
675         if (!PageSwapCache(page))
676                 return 0;
677         if (PageWriteback(page))
678                 return 0;
679         if (page_swapcount(page))
680                 return 0;
681
682         delete_from_swap_cache(page);
683         SetPageDirty(page);
684         return 1;
685 }
686
687 /*
688  * Free the swap entry like above, but also try to
689  * free the page cache entry if it is the last user.
690  */
691 int free_swap_and_cache(swp_entry_t entry)
692 {
693         struct swap_info_struct *p;
694         struct page *page = NULL;
695
696         if (non_swap_entry(entry))
697                 return 1;
698
699         p = swap_info_get(entry);
700         if (p) {
701                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
702                         page = find_get_page(&swapper_space, entry.val);
703                         if (page && !trylock_page(page)) {
704                                 page_cache_release(page);
705                                 page = NULL;
706                         }
707                 }
708                 spin_unlock(&swap_lock);
709         }
710         if (page) {
711                 /*
712                  * Not mapped elsewhere, or swap space full? Free it!
713                  * Also recheck PageSwapCache now page is locked (above).
714                  */
715                 if (PageSwapCache(page) && !PageWriteback(page) &&
716                                 (!page_mapped(page) || vm_swap_full())) {
717                         delete_from_swap_cache(page);
718                         SetPageDirty(page);
719                 }
720                 unlock_page(page);
721                 page_cache_release(page);
722         }
723         return p != NULL;
724 }
725
726 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
727 /**
728  * mem_cgroup_count_swap_user - count the user of a swap entry
729  * @ent: the swap entry to be checked
730  * @pagep: the pointer for the swap cache page of the entry to be stored
731  *
732  * Returns the number of the user of the swap entry. The number is valid only
733  * for swaps of anonymous pages.
734  * If the entry is found on swap cache, the page is stored to pagep with
735  * refcount of it being incremented.
736  */
737 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
738 {
739         struct page *page;
740         struct swap_info_struct *p;
741         int count = 0;
742
743         page = find_get_page(&swapper_space, ent.val);
744         if (page)
745                 count += page_mapcount(page);
746         p = swap_info_get(ent);
747         if (p) {
748                 count += swap_count(p->swap_map[swp_offset(ent)]);
749                 spin_unlock(&swap_lock);
750         }
751
752         *pagep = page;
753         return count;
754 }
755 #endif
756
757 #ifdef CONFIG_HIBERNATION
758 /*
759  * Find the swap type that corresponds to given device (if any).
760  *
761  * @offset - number of the PAGE_SIZE-sized block of the device, starting
762  * from 0, in which the swap header is expected to be located.
763  *
764  * This is needed for the suspend to disk (aka swsusp).
765  */
766 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
767 {
768         struct block_device *bdev = NULL;
769         int type;
770
771         if (device)
772                 bdev = bdget(device);
773
774         spin_lock(&swap_lock);
775         for (type = 0; type < nr_swapfiles; type++) {
776                 struct swap_info_struct *sis = swap_info[type];
777
778                 if (!(sis->flags & SWP_WRITEOK))
779                         continue;
780
781                 if (!bdev) {
782                         if (bdev_p)
783                                 *bdev_p = bdgrab(sis->bdev);
784
785                         spin_unlock(&swap_lock);
786                         return type;
787                 }
788                 if (bdev == sis->bdev) {
789                         struct swap_extent *se = &sis->first_swap_extent;
790
791                         if (se->start_block == offset) {
792                                 if (bdev_p)
793                                         *bdev_p = bdgrab(sis->bdev);
794
795                                 spin_unlock(&swap_lock);
796                                 bdput(bdev);
797                                 return type;
798                         }
799                 }
800         }
801         spin_unlock(&swap_lock);
802         if (bdev)
803                 bdput(bdev);
804
805         return -ENODEV;
806 }
807
808 /*
809  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
810  * corresponding to given index in swap_info (swap type).
811  */
812 sector_t swapdev_block(int type, pgoff_t offset)
813 {
814         struct block_device *bdev;
815
816         if ((unsigned int)type >= nr_swapfiles)
817                 return 0;
818         if (!(swap_info[type]->flags & SWP_WRITEOK))
819                 return 0;
820         return map_swap_entry(swp_entry(type, offset), &bdev);
821 }
822
823 /*
824  * Return either the total number of swap pages of given type, or the number
825  * of free pages of that type (depending on @free)
826  *
827  * This is needed for software suspend
828  */
829 unsigned int count_swap_pages(int type, int free)
830 {
831         unsigned int n = 0;
832
833         spin_lock(&swap_lock);
834         if ((unsigned int)type < nr_swapfiles) {
835                 struct swap_info_struct *sis = swap_info[type];
836
837                 if (sis->flags & SWP_WRITEOK) {
838                         n = sis->pages;
839                         if (free)
840                                 n -= sis->inuse_pages;
841                 }
842         }
843         spin_unlock(&swap_lock);
844         return n;
845 }
846 #endif /* CONFIG_HIBERNATION */
847
848 /*
849  * No need to decide whether this PTE shares the swap entry with others,
850  * just let do_wp_page work it out if a write is requested later - to
851  * force COW, vm_page_prot omits write permission from any private vma.
852  */
853 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
854                 unsigned long addr, swp_entry_t entry, struct page *page)
855 {
856         struct mem_cgroup *ptr = NULL;
857         spinlock_t *ptl;
858         pte_t *pte;
859         int ret = 1;
860
861         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
862                 ret = -ENOMEM;
863                 goto out_nolock;
864         }
865
866         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
867         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
868                 if (ret > 0)
869                         mem_cgroup_cancel_charge_swapin(ptr);
870                 ret = 0;
871                 goto out;
872         }
873
874         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
875         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
876         get_page(page);
877         set_pte_at(vma->vm_mm, addr, pte,
878                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
879         page_add_anon_rmap(page, vma, addr);
880         mem_cgroup_commit_charge_swapin(page, ptr);
881         swap_free(entry);
882         /*
883          * Move the page to the active list so it is not
884          * immediately swapped out again after swapon.
885          */
886         activate_page(page);
887 out:
888         pte_unmap_unlock(pte, ptl);
889 out_nolock:
890         return ret;
891 }
892
893 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
894                                 unsigned long addr, unsigned long end,
895                                 swp_entry_t entry, struct page *page)
896 {
897         pte_t swp_pte = swp_entry_to_pte(entry);
898         pte_t *pte;
899         int ret = 0;
900
901         /*
902          * We don't actually need pte lock while scanning for swp_pte: since
903          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
904          * page table while we're scanning; though it could get zapped, and on
905          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
906          * of unmatched parts which look like swp_pte, so unuse_pte must
907          * recheck under pte lock.  Scanning without pte lock lets it be
908          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
909          */
910         pte = pte_offset_map(pmd, addr);
911         do {
912                 /*
913                  * swapoff spends a _lot_ of time in this loop!
914                  * Test inline before going to call unuse_pte.
915                  */
916                 if (unlikely(pte_same(*pte, swp_pte))) {
917                         pte_unmap(pte);
918                         ret = unuse_pte(vma, pmd, addr, entry, page);
919                         if (ret)
920                                 goto out;
921                         pte = pte_offset_map(pmd, addr);
922                 }
923         } while (pte++, addr += PAGE_SIZE, addr != end);
924         pte_unmap(pte - 1);
925 out:
926         return ret;
927 }
928
929 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
930                                 unsigned long addr, unsigned long end,
931                                 swp_entry_t entry, struct page *page)
932 {
933         pmd_t *pmd;
934         unsigned long next;
935         int ret;
936
937         pmd = pmd_offset(pud, addr);
938         do {
939                 next = pmd_addr_end(addr, end);
940                 if (pmd_none_or_clear_bad(pmd))
941                         continue;
942                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
943                 if (ret)
944                         return ret;
945         } while (pmd++, addr = next, addr != end);
946         return 0;
947 }
948
949 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
950                                 unsigned long addr, unsigned long end,
951                                 swp_entry_t entry, struct page *page)
952 {
953         pud_t *pud;
954         unsigned long next;
955         int ret;
956
957         pud = pud_offset(pgd, addr);
958         do {
959                 next = pud_addr_end(addr, end);
960                 if (pud_none_or_clear_bad(pud))
961                         continue;
962                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
963                 if (ret)
964                         return ret;
965         } while (pud++, addr = next, addr != end);
966         return 0;
967 }
968
969 static int unuse_vma(struct vm_area_struct *vma,
970                                 swp_entry_t entry, struct page *page)
971 {
972         pgd_t *pgd;
973         unsigned long addr, end, next;
974         int ret;
975
976         if (page_anon_vma(page)) {
977                 addr = page_address_in_vma(page, vma);
978                 if (addr == -EFAULT)
979                         return 0;
980                 else
981                         end = addr + PAGE_SIZE;
982         } else {
983                 addr = vma->vm_start;
984                 end = vma->vm_end;
985         }
986
987         pgd = pgd_offset(vma->vm_mm, addr);
988         do {
989                 next = pgd_addr_end(addr, end);
990                 if (pgd_none_or_clear_bad(pgd))
991                         continue;
992                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
993                 if (ret)
994                         return ret;
995         } while (pgd++, addr = next, addr != end);
996         return 0;
997 }
998
999 static int unuse_mm(struct mm_struct *mm,
1000                                 swp_entry_t entry, struct page *page)
1001 {
1002         struct vm_area_struct *vma;
1003         int ret = 0;
1004
1005         if (!down_read_trylock(&mm->mmap_sem)) {
1006                 /*
1007                  * Activate page so shrink_inactive_list is unlikely to unmap
1008                  * its ptes while lock is dropped, so swapoff can make progress.
1009                  */
1010                 activate_page(page);
1011                 unlock_page(page);
1012                 down_read(&mm->mmap_sem);
1013                 lock_page(page);
1014         }
1015         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1016                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1017                         break;
1018         }
1019         up_read(&mm->mmap_sem);
1020         return (ret < 0)? ret: 0;
1021 }
1022
1023 /*
1024  * Scan swap_map from current position to next entry still in use.
1025  * Recycle to start on reaching the end, returning 0 when empty.
1026  */
1027 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1028                                         unsigned int prev)
1029 {
1030         unsigned int max = si->max;
1031         unsigned int i = prev;
1032         unsigned char count;
1033
1034         /*
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).
1039          */
1040         for (;;) {
1041                 if (++i >= max) {
1042                         if (!prev) {
1043                                 i = 0;
1044                                 break;
1045                         }
1046                         /*
1047                          * No entries in use at top of swap_map,
1048                          * loop back to start and recheck there.
1049                          */
1050                         max = prev + 1;
1051                         prev = 0;
1052                         i = 1;
1053                 }
1054                 count = si->swap_map[i];
1055                 if (count && swap_count(count) != SWAP_MAP_BAD)
1056                         break;
1057         }
1058         return i;
1059 }
1060
1061 /*
1062  * We completely avoid races by reading each swap page in advance,
1063  * and then search for the process using it.  All the necessary
1064  * page table adjustments can then be made atomically.
1065  */
1066 static int try_to_unuse(unsigned int type)
1067 {
1068         struct swap_info_struct *si = swap_info[type];
1069         struct mm_struct *start_mm;
1070         unsigned char *swap_map;
1071         unsigned char swcount;
1072         struct page *page;
1073         swp_entry_t entry;
1074         unsigned int i = 0;
1075         int retval = 0;
1076
1077         /*
1078          * When searching mms for an entry, a good strategy is to
1079          * start at the first mm we freed the previous entry from
1080          * (though actually we don't notice whether we or coincidence
1081          * freed the entry).  Initialize this start_mm with a hold.
1082          *
1083          * A simpler strategy would be to start at the last mm we
1084          * freed the previous entry from; but that would take less
1085          * advantage of mmlist ordering, which clusters forked mms
1086          * together, child after parent.  If we race with dup_mmap(), we
1087          * prefer to resolve parent before child, lest we miss entries
1088          * duplicated after we scanned child: using last mm would invert
1089          * that.
1090          */
1091         start_mm = &init_mm;
1092         atomic_inc(&init_mm.mm_users);
1093
1094         /*
1095          * Keep on scanning until all entries have gone.  Usually,
1096          * one pass through swap_map is enough, but not necessarily:
1097          * there are races when an instance of an entry might be missed.
1098          */
1099         while ((i = find_next_to_unuse(si, i)) != 0) {
1100                 if (signal_pending(current)) {
1101                         retval = -EINTR;
1102                         break;
1103                 }
1104
1105                 /*
1106                  * Get a page for the entry, using the existing swap
1107                  * cache page if there is one.  Otherwise, get a clean
1108                  * page and read the swap into it.
1109                  */
1110                 swap_map = &si->swap_map[i];
1111                 entry = swp_entry(type, i);
1112                 page = read_swap_cache_async(entry,
1113                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1114                 if (!page) {
1115                         /*
1116                          * Either swap_duplicate() failed because entry
1117                          * has been freed independently, and will not be
1118                          * reused since sys_swapoff() already disabled
1119                          * allocation from here, or alloc_page() failed.
1120                          */
1121                         if (!*swap_map)
1122                                 continue;
1123                         retval = -ENOMEM;
1124                         break;
1125                 }
1126
1127                 /*
1128                  * Don't hold on to start_mm if it looks like exiting.
1129                  */
1130                 if (atomic_read(&start_mm->mm_users) == 1) {
1131                         mmput(start_mm);
1132                         start_mm = &init_mm;
1133                         atomic_inc(&init_mm.mm_users);
1134                 }
1135
1136                 /*
1137                  * Wait for and lock page.  When do_swap_page races with
1138                  * try_to_unuse, do_swap_page can handle the fault much
1139                  * faster than try_to_unuse can locate the entry.  This
1140                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1141                  * defer to do_swap_page in such a case - in some tests,
1142                  * do_swap_page and try_to_unuse repeatedly compete.
1143                  */
1144                 wait_on_page_locked(page);
1145                 wait_on_page_writeback(page);
1146                 lock_page(page);
1147                 wait_on_page_writeback(page);
1148
1149                 /*
1150                  * Remove all references to entry.
1151                  */
1152                 swcount = *swap_map;
1153                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1154                         retval = shmem_unuse(entry, page);
1155                         /* page has already been unlocked and released */
1156                         if (retval < 0)
1157                                 break;
1158                         continue;
1159                 }
1160                 if (swap_count(swcount) && start_mm != &init_mm)
1161                         retval = unuse_mm(start_mm, entry, page);
1162
1163                 if (swap_count(*swap_map)) {
1164                         int set_start_mm = (*swap_map >= swcount);
1165                         struct list_head *p = &start_mm->mmlist;
1166                         struct mm_struct *new_start_mm = start_mm;
1167                         struct mm_struct *prev_mm = start_mm;
1168                         struct mm_struct *mm;
1169
1170                         atomic_inc(&new_start_mm->mm_users);
1171                         atomic_inc(&prev_mm->mm_users);
1172                         spin_lock(&mmlist_lock);
1173                         while (swap_count(*swap_map) && !retval &&
1174                                         (p = p->next) != &start_mm->mmlist) {
1175                                 mm = list_entry(p, struct mm_struct, mmlist);
1176                                 if (!atomic_inc_not_zero(&mm->mm_users))
1177                                         continue;
1178                                 spin_unlock(&mmlist_lock);
1179                                 mmput(prev_mm);
1180                                 prev_mm = mm;
1181
1182                                 cond_resched();
1183
1184                                 swcount = *swap_map;
1185                                 if (!swap_count(swcount)) /* any usage ? */
1186                                         ;
1187                                 else if (mm == &init_mm)
1188                                         set_start_mm = 1;
1189                                 else
1190                                         retval = unuse_mm(mm, entry, page);
1191
1192                                 if (set_start_mm && *swap_map < swcount) {
1193                                         mmput(new_start_mm);
1194                                         atomic_inc(&mm->mm_users);
1195                                         new_start_mm = mm;
1196                                         set_start_mm = 0;
1197                                 }
1198                                 spin_lock(&mmlist_lock);
1199                         }
1200                         spin_unlock(&mmlist_lock);
1201                         mmput(prev_mm);
1202                         mmput(start_mm);
1203                         start_mm = new_start_mm;
1204                 }
1205                 if (retval) {
1206                         unlock_page(page);
1207                         page_cache_release(page);
1208                         break;
1209                 }
1210
1211                 /*
1212                  * If a reference remains (rare), we would like to leave
1213                  * the page in the swap cache; but try_to_unmap could
1214                  * then re-duplicate the entry once we drop page lock,
1215                  * so we might loop indefinitely; also, that page could
1216                  * not be swapped out to other storage meanwhile.  So:
1217                  * delete from cache even if there's another reference,
1218                  * after ensuring that the data has been saved to disk -
1219                  * since if the reference remains (rarer), it will be
1220                  * read from disk into another page.  Splitting into two
1221                  * pages would be incorrect if swap supported "shared
1222                  * private" pages, but they are handled by tmpfs files.
1223                  *
1224                  * Given how unuse_vma() targets one particular offset
1225                  * in an anon_vma, once the anon_vma has been determined,
1226                  * this splitting happens to be just what is needed to
1227                  * handle where KSM pages have been swapped out: re-reading
1228                  * is unnecessarily slow, but we can fix that later on.
1229                  */
1230                 if (swap_count(*swap_map) &&
1231                      PageDirty(page) && PageSwapCache(page)) {
1232                         struct writeback_control wbc = {
1233                                 .sync_mode = WB_SYNC_NONE,
1234                         };
1235
1236                         swap_writepage(page, &wbc);
1237                         lock_page(page);
1238                         wait_on_page_writeback(page);
1239                 }
1240
1241                 /*
1242                  * It is conceivable that a racing task removed this page from
1243                  * swap cache just before we acquired the page lock at the top,
1244                  * or while we dropped it in unuse_mm().  The page might even
1245                  * be back in swap cache on another swap area: that we must not
1246                  * delete, since it may not have been written out to swap yet.
1247                  */
1248                 if (PageSwapCache(page) &&
1249                     likely(page_private(page) == entry.val))
1250                         delete_from_swap_cache(page);
1251
1252                 /*
1253                  * So we could skip searching mms once swap count went
1254                  * to 1, we did not mark any present ptes as dirty: must
1255                  * mark page dirty so shrink_page_list will preserve it.
1256                  */
1257                 SetPageDirty(page);
1258                 unlock_page(page);
1259                 page_cache_release(page);
1260
1261                 /*
1262                  * Make sure that we aren't completely killing
1263                  * interactive performance.
1264                  */
1265                 cond_resched();
1266         }
1267
1268         mmput(start_mm);
1269         return retval;
1270 }
1271
1272 /*
1273  * After a successful try_to_unuse, if no swap is now in use, we know
1274  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1275  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1276  * added to the mmlist just after page_duplicate - before would be racy.
1277  */
1278 static void drain_mmlist(void)
1279 {
1280         struct list_head *p, *next;
1281         unsigned int type;
1282
1283         for (type = 0; type < nr_swapfiles; type++)
1284                 if (swap_info[type]->inuse_pages)
1285                         return;
1286         spin_lock(&mmlist_lock);
1287         list_for_each_safe(p, next, &init_mm.mmlist)
1288                 list_del_init(p);
1289         spin_unlock(&mmlist_lock);
1290 }
1291
1292 /*
1293  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1294  * corresponds to page offset for the specified swap entry.
1295  * Note that the type of this function is sector_t, but it returns page offset
1296  * into the bdev, not sector offset.
1297  */
1298 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1299 {
1300         struct swap_info_struct *sis;
1301         struct swap_extent *start_se;
1302         struct swap_extent *se;
1303         pgoff_t offset;
1304
1305         sis = swap_info[swp_type(entry)];
1306         *bdev = sis->bdev;
1307
1308         offset = swp_offset(entry);
1309         start_se = sis->curr_swap_extent;
1310         se = start_se;
1311
1312         for ( ; ; ) {
1313                 struct list_head *lh;
1314
1315                 if (se->start_page <= offset &&
1316                                 offset < (se->start_page + se->nr_pages)) {
1317                         return se->start_block + (offset - se->start_page);
1318                 }
1319                 lh = se->list.next;
1320                 se = list_entry(lh, struct swap_extent, list);
1321                 sis->curr_swap_extent = se;
1322                 BUG_ON(se == start_se);         /* It *must* be present */
1323         }
1324 }
1325
1326 /*
1327  * Returns the page offset into bdev for the specified page's swap entry.
1328  */
1329 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1330 {
1331         swp_entry_t entry;
1332         entry.val = page_private(page);
1333         return map_swap_entry(entry, bdev);
1334 }
1335
1336 /*
1337  * Free all of a swapdev's extent information
1338  */
1339 static void destroy_swap_extents(struct swap_info_struct *sis)
1340 {
1341         while (!list_empty(&sis->first_swap_extent.list)) {
1342                 struct swap_extent *se;
1343
1344                 se = list_entry(sis->first_swap_extent.list.next,
1345                                 struct swap_extent, list);
1346                 list_del(&se->list);
1347                 kfree(se);
1348         }
1349 }
1350
1351 /*
1352  * Add a block range (and the corresponding page range) into this swapdev's
1353  * extent list.  The extent list is kept sorted in page order.
1354  *
1355  * This function rather assumes that it is called in ascending page order.
1356  */
1357 static int
1358 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1359                 unsigned long nr_pages, sector_t start_block)
1360 {
1361         struct swap_extent *se;
1362         struct swap_extent *new_se;
1363         struct list_head *lh;
1364
1365         if (start_page == 0) {
1366                 se = &sis->first_swap_extent;
1367                 sis->curr_swap_extent = se;
1368                 se->start_page = 0;
1369                 se->nr_pages = nr_pages;
1370                 se->start_block = start_block;
1371                 return 1;
1372         } else {
1373                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1374                 se = list_entry(lh, struct swap_extent, list);
1375                 BUG_ON(se->start_page + se->nr_pages != start_page);
1376                 if (se->start_block + se->nr_pages == start_block) {
1377                         /* Merge it */
1378                         se->nr_pages += nr_pages;
1379                         return 0;
1380                 }
1381         }
1382
1383         /*
1384          * No merge.  Insert a new extent, preserving ordering.
1385          */
1386         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1387         if (new_se == NULL)
1388                 return -ENOMEM;
1389         new_se->start_page = start_page;
1390         new_se->nr_pages = nr_pages;
1391         new_se->start_block = start_block;
1392
1393         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1394         return 1;
1395 }
1396
1397 /*
1398  * A `swap extent' is a simple thing which maps a contiguous range of pages
1399  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1400  * is built at swapon time and is then used at swap_writepage/swap_readpage
1401  * time for locating where on disk a page belongs.
1402  *
1403  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1404  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1405  * swap files identically.
1406  *
1407  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1408  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1409  * swapfiles are handled *identically* after swapon time.
1410  *
1411  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1412  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1413  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1414  * requirements, they are simply tossed out - we will never use those blocks
1415  * for swapping.
1416  *
1417  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1418  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1419  * which will scribble on the fs.
1420  *
1421  * The amount of disk space which a single swap extent represents varies.
1422  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1423  * extents in the list.  To avoid much list walking, we cache the previous
1424  * search location in `curr_swap_extent', and start new searches from there.
1425  * This is extremely effective.  The average number of iterations in
1426  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1427  */
1428 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1429 {
1430         struct inode *inode;
1431         unsigned blocks_per_page;
1432         unsigned long page_no;
1433         unsigned blkbits;
1434         sector_t probe_block;
1435         sector_t last_block;
1436         sector_t lowest_block = -1;
1437         sector_t highest_block = 0;
1438         int nr_extents = 0;
1439         int ret;
1440
1441         inode = sis->swap_file->f_mapping->host;
1442         if (S_ISBLK(inode->i_mode)) {
1443                 ret = add_swap_extent(sis, 0, sis->max, 0);
1444                 *span = sis->pages;
1445                 goto out;
1446         }
1447
1448         blkbits = inode->i_blkbits;
1449         blocks_per_page = PAGE_SIZE >> blkbits;
1450
1451         /*
1452          * Map all the blocks into the extent list.  This code doesn't try
1453          * to be very smart.
1454          */
1455         probe_block = 0;
1456         page_no = 0;
1457         last_block = i_size_read(inode) >> blkbits;
1458         while ((probe_block + blocks_per_page) <= last_block &&
1459                         page_no < sis->max) {
1460                 unsigned block_in_page;
1461                 sector_t first_block;
1462
1463                 first_block = bmap(inode, probe_block);
1464                 if (first_block == 0)
1465                         goto bad_bmap;
1466
1467                 /*
1468                  * It must be PAGE_SIZE aligned on-disk
1469                  */
1470                 if (first_block & (blocks_per_page - 1)) {
1471                         probe_block++;
1472                         goto reprobe;
1473                 }
1474
1475                 for (block_in_page = 1; block_in_page < blocks_per_page;
1476                                         block_in_page++) {
1477                         sector_t block;
1478
1479                         block = bmap(inode, probe_block + block_in_page);
1480                         if (block == 0)
1481                                 goto bad_bmap;
1482                         if (block != first_block + block_in_page) {
1483                                 /* Discontiguity */
1484                                 probe_block++;
1485                                 goto reprobe;
1486                         }
1487                 }
1488
1489                 first_block >>= (PAGE_SHIFT - blkbits);
1490                 if (page_no) {  /* exclude the header page */
1491                         if (first_block < lowest_block)
1492                                 lowest_block = first_block;
1493                         if (first_block > highest_block)
1494                                 highest_block = first_block;
1495                 }
1496
1497                 /*
1498                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1499                  */
1500                 ret = add_swap_extent(sis, page_no, 1, first_block);
1501                 if (ret < 0)
1502                         goto out;
1503                 nr_extents += ret;
1504                 page_no++;
1505                 probe_block += blocks_per_page;
1506 reprobe:
1507                 continue;
1508         }
1509         ret = nr_extents;
1510         *span = 1 + highest_block - lowest_block;
1511         if (page_no == 0)
1512                 page_no = 1;    /* force Empty message */
1513         sis->max = page_no;
1514         sis->pages = page_no - 1;
1515         sis->highest_bit = page_no - 1;
1516 out:
1517         return ret;
1518 bad_bmap:
1519         printk(KERN_ERR "swapon: swapfile has holes\n");
1520         ret = -EINVAL;
1521         goto out;
1522 }
1523
1524 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1525 {
1526         struct swap_info_struct *p = NULL;
1527         unsigned char *swap_map;
1528         struct file *swap_file, *victim;
1529         struct address_space *mapping;
1530         struct inode *inode;
1531         char *pathname;
1532         int i, type, prev;
1533         int err;
1534
1535         if (!capable(CAP_SYS_ADMIN))
1536                 return -EPERM;
1537
1538         pathname = getname(specialfile);
1539         err = PTR_ERR(pathname);
1540         if (IS_ERR(pathname))
1541                 goto out;
1542
1543         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1544         putname(pathname);
1545         err = PTR_ERR(victim);
1546         if (IS_ERR(victim))
1547                 goto out;
1548
1549         mapping = victim->f_mapping;
1550         prev = -1;
1551         spin_lock(&swap_lock);
1552         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1553                 p = swap_info[type];
1554                 if (p->flags & SWP_WRITEOK) {
1555                         if (p->swap_file->f_mapping == mapping)
1556                                 break;
1557                 }
1558                 prev = type;
1559         }
1560         if (type < 0) {
1561                 err = -EINVAL;
1562                 spin_unlock(&swap_lock);
1563                 goto out_dput;
1564         }
1565         if (!security_vm_enough_memory(p->pages))
1566                 vm_unacct_memory(p->pages);
1567         else {
1568                 err = -ENOMEM;
1569                 spin_unlock(&swap_lock);
1570                 goto out_dput;
1571         }
1572         if (prev < 0)
1573                 swap_list.head = p->next;
1574         else
1575                 swap_info[prev]->next = p->next;
1576         if (type == swap_list.next) {
1577                 /* just pick something that's safe... */
1578                 swap_list.next = swap_list.head;
1579         }
1580         if (p->prio < 0) {
1581                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1582                         swap_info[i]->prio = p->prio--;
1583                 least_priority++;
1584         }
1585         nr_swap_pages -= p->pages;
1586         total_swap_pages -= p->pages;
1587         p->flags &= ~SWP_WRITEOK;
1588         spin_unlock(&swap_lock);
1589
1590         current->flags |= PF_OOM_ORIGIN;
1591         err = try_to_unuse(type);
1592         current->flags &= ~PF_OOM_ORIGIN;
1593
1594         if (err) {
1595                 /* re-insert swap space back into swap_list */
1596                 spin_lock(&swap_lock);
1597                 if (p->prio < 0)
1598                         p->prio = --least_priority;
1599                 prev = -1;
1600                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1601                         if (p->prio >= swap_info[i]->prio)
1602                                 break;
1603                         prev = i;
1604                 }
1605                 p->next = i;
1606                 if (prev < 0)
1607                         swap_list.head = swap_list.next = type;
1608                 else
1609                         swap_info[prev]->next = type;
1610                 nr_swap_pages += p->pages;
1611                 total_swap_pages += p->pages;
1612                 p->flags |= SWP_WRITEOK;
1613                 spin_unlock(&swap_lock);
1614                 goto out_dput;
1615         }
1616
1617         /* wait for any unplug function to finish */
1618         down_write(&swap_unplug_sem);
1619         up_write(&swap_unplug_sem);
1620
1621         destroy_swap_extents(p);
1622         if (p->flags & SWP_CONTINUED)
1623                 free_swap_count_continuations(p);
1624
1625         mutex_lock(&swapon_mutex);
1626         spin_lock(&swap_lock);
1627         drain_mmlist();
1628
1629         /* wait for anyone still in scan_swap_map */
1630         p->highest_bit = 0;             /* cuts scans short */
1631         while (p->flags >= SWP_SCANNING) {
1632                 spin_unlock(&swap_lock);
1633                 schedule_timeout_uninterruptible(1);
1634                 spin_lock(&swap_lock);
1635         }
1636
1637         swap_file = p->swap_file;
1638         p->swap_file = NULL;
1639         p->max = 0;
1640         swap_map = p->swap_map;
1641         p->swap_map = NULL;
1642         p->flags = 0;
1643         spin_unlock(&swap_lock);
1644         mutex_unlock(&swapon_mutex);
1645         vfree(swap_map);
1646         /* Destroy swap account informatin */
1647         swap_cgroup_swapoff(type);
1648
1649         inode = mapping->host;
1650         if (S_ISBLK(inode->i_mode)) {
1651                 struct block_device *bdev = I_BDEV(inode);
1652                 set_blocksize(bdev, p->old_block_size);
1653                 bd_release(bdev);
1654         } else {
1655                 mutex_lock(&inode->i_mutex);
1656                 inode->i_flags &= ~S_SWAPFILE;
1657                 mutex_unlock(&inode->i_mutex);
1658         }
1659         filp_close(swap_file, NULL);
1660         err = 0;
1661
1662 out_dput:
1663         filp_close(victim, NULL);
1664 out:
1665         return err;
1666 }
1667
1668 #ifdef CONFIG_PROC_FS
1669 /* iterator */
1670 static void *swap_start(struct seq_file *swap, loff_t *pos)
1671 {
1672         struct swap_info_struct *si;
1673         int type;
1674         loff_t l = *pos;
1675
1676         mutex_lock(&swapon_mutex);
1677
1678         if (!l)
1679                 return SEQ_START_TOKEN;
1680
1681         for (type = 0; type < nr_swapfiles; type++) {
1682                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1683                 si = swap_info[type];
1684                 if (!(si->flags & SWP_USED) || !si->swap_map)
1685                         continue;
1686                 if (!--l)
1687                         return si;
1688         }
1689
1690         return NULL;
1691 }
1692
1693 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1694 {
1695         struct swap_info_struct *si = v;
1696         int type;
1697
1698         if (v == SEQ_START_TOKEN)
1699                 type = 0;
1700         else
1701                 type = si->type + 1;
1702
1703         for (; 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                 ++*pos;
1709                 return si;
1710         }
1711
1712         return NULL;
1713 }
1714
1715 static void swap_stop(struct seq_file *swap, void *v)
1716 {
1717         mutex_unlock(&swapon_mutex);
1718 }
1719
1720 static int swap_show(struct seq_file *swap, void *v)
1721 {
1722         struct swap_info_struct *si = v;
1723         struct file *file;
1724         int len;
1725
1726         if (si == SEQ_START_TOKEN) {
1727                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1728                 return 0;
1729         }
1730
1731         file = si->swap_file;
1732         len = seq_path(swap, &file->f_path, " \t\n\\");
1733         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1734                         len < 40 ? 40 - len : 1, " ",
1735                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1736                                 "partition" : "file\t",
1737                         si->pages << (PAGE_SHIFT - 10),
1738                         si->inuse_pages << (PAGE_SHIFT - 10),
1739                         si->prio);
1740         return 0;
1741 }
1742
1743 static const struct seq_operations swaps_op = {
1744         .start =        swap_start,
1745         .next =         swap_next,
1746         .stop =         swap_stop,
1747         .show =         swap_show
1748 };
1749
1750 static int swaps_open(struct inode *inode, struct file *file)
1751 {
1752         return seq_open(file, &swaps_op);
1753 }
1754
1755 static const struct file_operations proc_swaps_operations = {
1756         .open           = swaps_open,
1757         .read           = seq_read,
1758         .llseek         = seq_lseek,
1759         .release        = seq_release,
1760 };
1761
1762 static int __init procswaps_init(void)
1763 {
1764         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1765         return 0;
1766 }
1767 __initcall(procswaps_init);
1768 #endif /* CONFIG_PROC_FS */
1769
1770 #ifdef MAX_SWAPFILES_CHECK
1771 static int __init max_swapfiles_check(void)
1772 {
1773         MAX_SWAPFILES_CHECK();
1774         return 0;
1775 }
1776 late_initcall(max_swapfiles_check);
1777 #endif
1778
1779 /*
1780  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1781  *
1782  * The swapon system call
1783  */
1784 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1785 {
1786         struct swap_info_struct *p;
1787         char *name = NULL;
1788         struct block_device *bdev = NULL;
1789         struct file *swap_file = NULL;
1790         struct address_space *mapping;
1791         unsigned int type;
1792         int i, prev;
1793         int error;
1794         union swap_header *swap_header;
1795         unsigned int nr_good_pages;
1796         int nr_extents = 0;
1797         sector_t span;
1798         unsigned long maxpages;
1799         unsigned long swapfilepages;
1800         unsigned char *swap_map = NULL;
1801         struct page *page = NULL;
1802         struct inode *inode = NULL;
1803         int did_down = 0;
1804
1805         if (!capable(CAP_SYS_ADMIN))
1806                 return -EPERM;
1807
1808         p = kzalloc(sizeof(*p), GFP_KERNEL);
1809         if (!p)
1810                 return -ENOMEM;
1811
1812         spin_lock(&swap_lock);
1813         for (type = 0; type < nr_swapfiles; type++) {
1814                 if (!(swap_info[type]->flags & SWP_USED))
1815                         break;
1816         }
1817         error = -EPERM;
1818         if (type >= MAX_SWAPFILES) {
1819                 spin_unlock(&swap_lock);
1820                 kfree(p);
1821                 goto out;
1822         }
1823         if (type >= nr_swapfiles) {
1824                 p->type = type;
1825                 swap_info[type] = p;
1826                 /*
1827                  * Write swap_info[type] before nr_swapfiles, in case a
1828                  * racing procfs swap_start() or swap_next() is reading them.
1829                  * (We never shrink nr_swapfiles, we never free this entry.)
1830                  */
1831                 smp_wmb();
1832                 nr_swapfiles++;
1833         } else {
1834                 kfree(p);
1835                 p = swap_info[type];
1836                 /*
1837                  * Do not memset this entry: a racing procfs swap_next()
1838                  * would be relying on p->type to remain valid.
1839                  */
1840         }
1841         INIT_LIST_HEAD(&p->first_swap_extent.list);
1842         p->flags = SWP_USED;
1843         p->next = -1;
1844         spin_unlock(&swap_lock);
1845
1846         name = getname(specialfile);
1847         error = PTR_ERR(name);
1848         if (IS_ERR(name)) {
1849                 name = NULL;
1850                 goto bad_swap_2;
1851         }
1852         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1853         error = PTR_ERR(swap_file);
1854         if (IS_ERR(swap_file)) {
1855                 swap_file = NULL;
1856                 goto bad_swap_2;
1857         }
1858
1859         p->swap_file = swap_file;
1860         mapping = swap_file->f_mapping;
1861         inode = mapping->host;
1862
1863         error = -EBUSY;
1864         for (i = 0; i < nr_swapfiles; i++) {
1865                 struct swap_info_struct *q = swap_info[i];
1866
1867                 if (i == type || !q->swap_file)
1868                         continue;
1869                 if (mapping == q->swap_file->f_mapping)
1870                         goto bad_swap;
1871         }
1872
1873         error = -EINVAL;
1874         if (S_ISBLK(inode->i_mode)) {
1875                 bdev = I_BDEV(inode);
1876                 error = bd_claim(bdev, sys_swapon);
1877                 if (error < 0) {
1878                         bdev = NULL;
1879                         error = -EINVAL;
1880                         goto bad_swap;
1881                 }
1882                 p->old_block_size = block_size(bdev);
1883                 error = set_blocksize(bdev, PAGE_SIZE);
1884                 if (error < 0)
1885                         goto bad_swap;
1886                 p->bdev = bdev;
1887         } else if (S_ISREG(inode->i_mode)) {
1888                 p->bdev = inode->i_sb->s_bdev;
1889                 mutex_lock(&inode->i_mutex);
1890                 did_down = 1;
1891                 if (IS_SWAPFILE(inode)) {
1892                         error = -EBUSY;
1893                         goto bad_swap;
1894                 }
1895         } else {
1896                 goto bad_swap;
1897         }
1898
1899         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1900
1901         /*
1902          * Read the swap header.
1903          */
1904         if (!mapping->a_ops->readpage) {
1905                 error = -EINVAL;
1906                 goto bad_swap;
1907         }
1908         page = read_mapping_page(mapping, 0, swap_file);
1909         if (IS_ERR(page)) {
1910                 error = PTR_ERR(page);
1911                 goto bad_swap;
1912         }
1913         swap_header = kmap(page);
1914
1915         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1916                 printk(KERN_ERR "Unable to find swap-space signature\n");
1917                 error = -EINVAL;
1918                 goto bad_swap;
1919         }
1920
1921         /* swap partition endianess hack... */
1922         if (swab32(swap_header->info.version) == 1) {
1923                 swab32s(&swap_header->info.version);
1924                 swab32s(&swap_header->info.last_page);
1925                 swab32s(&swap_header->info.nr_badpages);
1926                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1927                         swab32s(&swap_header->info.badpages[i]);
1928         }
1929         /* Check the swap header's sub-version */
1930         if (swap_header->info.version != 1) {
1931                 printk(KERN_WARNING
1932                        "Unable to handle swap header version %d\n",
1933                        swap_header->info.version);
1934                 error = -EINVAL;
1935                 goto bad_swap;
1936         }
1937
1938         p->lowest_bit  = 1;
1939         p->cluster_next = 1;
1940         p->cluster_nr = 0;
1941
1942         /*
1943          * Find out how many pages are allowed for a single swap
1944          * device. There are two limiting factors: 1) the number of
1945          * bits for the swap offset in the swp_entry_t type and
1946          * 2) the number of bits in the a swap pte as defined by
1947          * the different architectures. In order to find the
1948          * largest possible bit mask a swap entry with swap type 0
1949          * and swap offset ~0UL is created, encoded to a swap pte,
1950          * decoded to a swp_entry_t again and finally the swap
1951          * offset is extracted. This will mask all the bits from
1952          * the initial ~0UL mask that can't be encoded in either
1953          * the swp_entry_t or the architecture definition of a
1954          * swap pte.
1955          */
1956         maxpages = swp_offset(pte_to_swp_entry(
1957                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1958         if (maxpages > swap_header->info.last_page) {
1959                 maxpages = swap_header->info.last_page + 1;
1960                 /* p->max is an unsigned int: don't overflow it */
1961                 if ((unsigned int)maxpages == 0)
1962                         maxpages = UINT_MAX;
1963         }
1964         p->highest_bit = maxpages - 1;
1965
1966         error = -EINVAL;
1967         if (!maxpages)
1968                 goto bad_swap;
1969         if (swapfilepages && maxpages > swapfilepages) {
1970                 printk(KERN_WARNING
1971                        "Swap area shorter than signature indicates\n");
1972                 goto bad_swap;
1973         }
1974         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1975                 goto bad_swap;
1976         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1977                 goto bad_swap;
1978
1979         /* OK, set up the swap map and apply the bad block list */
1980         swap_map = vmalloc(maxpages);
1981         if (!swap_map) {
1982                 error = -ENOMEM;
1983                 goto bad_swap;
1984         }
1985
1986         memset(swap_map, 0, maxpages);
1987         nr_good_pages = maxpages - 1;   /* omit header page */
1988
1989         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1990                 unsigned int page_nr = swap_header->info.badpages[i];
1991                 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
1992                         error = -EINVAL;
1993                         goto bad_swap;
1994                 }
1995                 if (page_nr < maxpages) {
1996                         swap_map[page_nr] = SWAP_MAP_BAD;
1997                         nr_good_pages--;
1998                 }
1999         }
2000
2001         error = swap_cgroup_swapon(type, maxpages);
2002         if (error)
2003                 goto bad_swap;
2004
2005         if (nr_good_pages) {
2006                 swap_map[0] = SWAP_MAP_BAD;
2007                 p->max = maxpages;
2008                 p->pages = nr_good_pages;
2009                 nr_extents = setup_swap_extents(p, &span);
2010                 if (nr_extents < 0) {
2011                         error = nr_extents;
2012                         goto bad_swap;
2013                 }
2014                 nr_good_pages = p->pages;
2015         }
2016         if (!nr_good_pages) {
2017                 printk(KERN_WARNING "Empty swap-file\n");
2018                 error = -EINVAL;
2019                 goto bad_swap;
2020         }
2021
2022         if (p->bdev) {
2023                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2024                         p->flags |= SWP_SOLIDSTATE;
2025                         p->cluster_next = 1 + (random32() % p->highest_bit);
2026                 }
2027                 if (discard_swap(p) == 0)
2028                         p->flags |= SWP_DISCARDABLE;
2029         }
2030
2031         mutex_lock(&swapon_mutex);
2032         spin_lock(&swap_lock);
2033         if (swap_flags & SWAP_FLAG_PREFER)
2034                 p->prio =
2035                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2036         else
2037                 p->prio = --least_priority;
2038         p->swap_map = swap_map;
2039         p->flags |= SWP_WRITEOK;
2040         nr_swap_pages += nr_good_pages;
2041         total_swap_pages += nr_good_pages;
2042
2043         printk(KERN_INFO "Adding %uk swap on %s.  "
2044                         "Priority:%d extents:%d across:%lluk %s%s\n",
2045                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2046                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2047                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2048                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2049
2050         /* insert swap space into swap_list: */
2051         prev = -1;
2052         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2053                 if (p->prio >= swap_info[i]->prio)
2054                         break;
2055                 prev = i;
2056         }
2057         p->next = i;
2058         if (prev < 0)
2059                 swap_list.head = swap_list.next = type;
2060         else
2061                 swap_info[prev]->next = type;
2062         spin_unlock(&swap_lock);
2063         mutex_unlock(&swapon_mutex);
2064         error = 0;
2065         goto out;
2066 bad_swap:
2067         if (bdev) {
2068                 set_blocksize(bdev, p->old_block_size);
2069                 bd_release(bdev);
2070         }
2071         destroy_swap_extents(p);
2072         swap_cgroup_swapoff(type);
2073 bad_swap_2:
2074         spin_lock(&swap_lock);
2075         p->swap_file = NULL;
2076         p->flags = 0;
2077         spin_unlock(&swap_lock);
2078         vfree(swap_map);
2079         if (swap_file)
2080                 filp_close(swap_file, NULL);
2081 out:
2082         if (page && !IS_ERR(page)) {
2083                 kunmap(page);
2084                 page_cache_release(page);
2085         }
2086         if (name)
2087                 putname(name);
2088         if (did_down) {
2089                 if (!error)
2090                         inode->i_flags |= S_SWAPFILE;
2091                 mutex_unlock(&inode->i_mutex);
2092         }
2093         return error;
2094 }
2095
2096 void si_swapinfo(struct sysinfo *val)
2097 {
2098         unsigned int type;
2099         unsigned long nr_to_be_unused = 0;
2100
2101         spin_lock(&swap_lock);
2102         for (type = 0; type < nr_swapfiles; type++) {
2103                 struct swap_info_struct *si = swap_info[type];
2104
2105                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2106                         nr_to_be_unused += si->inuse_pages;
2107         }
2108         val->freeswap = nr_swap_pages + nr_to_be_unused;
2109         val->totalswap = total_swap_pages + nr_to_be_unused;
2110         spin_unlock(&swap_lock);
2111 }
2112
2113 /*
2114  * Verify that a swap entry is valid and increment its swap map count.
2115  *
2116  * Returns error code in following case.
2117  * - success -> 0
2118  * - swp_entry is invalid -> EINVAL
2119  * - swp_entry is migration entry -> EINVAL
2120  * - swap-cache reference is requested but there is already one. -> EEXIST
2121  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2122  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2123  */
2124 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2125 {
2126         struct swap_info_struct *p;
2127         unsigned long offset, type;
2128         unsigned char count;
2129         unsigned char has_cache;
2130         int err = -EINVAL;
2131
2132         if (non_swap_entry(entry))
2133                 goto out;
2134
2135         type = swp_type(entry);
2136         if (type >= nr_swapfiles)
2137                 goto bad_file;
2138         p = swap_info[type];
2139         offset = swp_offset(entry);
2140
2141         spin_lock(&swap_lock);
2142         if (unlikely(offset >= p->max))
2143                 goto unlock_out;
2144
2145         count = p->swap_map[offset];
2146         has_cache = count & SWAP_HAS_CACHE;
2147         count &= ~SWAP_HAS_CACHE;
2148         err = 0;
2149
2150         if (usage == SWAP_HAS_CACHE) {
2151
2152                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2153                 if (!has_cache && count)
2154                         has_cache = SWAP_HAS_CACHE;
2155                 else if (has_cache)             /* someone else added cache */
2156                         err = -EEXIST;
2157                 else                            /* no users remaining */
2158                         err = -ENOENT;
2159
2160         } else if (count || has_cache) {
2161
2162                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2163                         count += usage;
2164                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2165                         err = -EINVAL;
2166                 else if (swap_count_continued(p, offset, count))
2167                         count = COUNT_CONTINUED;
2168                 else
2169                         err = -ENOMEM;
2170         } else
2171                 err = -ENOENT;                  /* unused swap entry */
2172
2173         p->swap_map[offset] = count | has_cache;
2174
2175 unlock_out:
2176         spin_unlock(&swap_lock);
2177 out:
2178         return err;
2179
2180 bad_file:
2181         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2182         goto out;
2183 }
2184
2185 /*
2186  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2187  * (in which case its reference count is never incremented).
2188  */
2189 void swap_shmem_alloc(swp_entry_t entry)
2190 {
2191         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2192 }
2193
2194 /*
2195  * Increase reference count of swap entry by 1.
2196  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2197  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2198  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2199  * might occur if a page table entry has got corrupted.
2200  */
2201 int swap_duplicate(swp_entry_t entry)
2202 {
2203         int err = 0;
2204
2205         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2206                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2207         return err;
2208 }
2209
2210 /*
2211  * @entry: swap entry for which we allocate swap cache.
2212  *
2213  * Called when allocating swap cache for existing swap entry,
2214  * This can return error codes. Returns 0 at success.
2215  * -EBUSY means there is a swap cache.
2216  * Note: return code is different from swap_duplicate().
2217  */
2218 int swapcache_prepare(swp_entry_t entry)
2219 {
2220         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2221 }
2222
2223 /*
2224  * swap_lock prevents swap_map being freed. Don't grab an extra
2225  * reference on the swaphandle, it doesn't matter if it becomes unused.
2226  */
2227 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2228 {
2229         struct swap_info_struct *si;
2230         int our_page_cluster = page_cluster;
2231         pgoff_t target, toff;
2232         pgoff_t base, end;
2233         int nr_pages = 0;
2234
2235         if (!our_page_cluster)  /* no readahead */
2236                 return 0;
2237
2238         si = swap_info[swp_type(entry)];
2239         target = swp_offset(entry);
2240         base = (target >> our_page_cluster) << our_page_cluster;
2241         end = base + (1 << our_page_cluster);
2242         if (!base)              /* first page is swap header */
2243                 base++;
2244
2245         spin_lock(&swap_lock);
2246         if (end > si->max)      /* don't go beyond end of map */
2247                 end = si->max;
2248
2249         /* Count contiguous allocated slots above our target */
2250         for (toff = target; ++toff < end; nr_pages++) {
2251                 /* Don't read in free or bad pages */
2252                 if (!si->swap_map[toff])
2253                         break;
2254                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2255                         break;
2256         }
2257         /* Count contiguous allocated slots below our target */
2258         for (toff = target; --toff >= base; nr_pages++) {
2259                 /* Don't read in free or bad pages */
2260                 if (!si->swap_map[toff])
2261                         break;
2262                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2263                         break;
2264         }
2265         spin_unlock(&swap_lock);
2266
2267         /*
2268          * Indicate starting offset, and return number of pages to get:
2269          * if only 1, say 0, since there's then no readahead to be done.
2270          */
2271         *offset = ++toff;
2272         return nr_pages? ++nr_pages: 0;
2273 }
2274
2275 /*
2276  * add_swap_count_continuation - called when a swap count is duplicated
2277  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2278  * page of the original vmalloc'ed swap_map, to hold the continuation count
2279  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2280  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2281  *
2282  * These continuation pages are seldom referenced: the common paths all work
2283  * on the original swap_map, only referring to a continuation page when the
2284  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2285  *
2286  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2287  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2288  * can be called after dropping locks.
2289  */
2290 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2291 {
2292         struct swap_info_struct *si;
2293         struct page *head;
2294         struct page *page;
2295         struct page *list_page;
2296         pgoff_t offset;
2297         unsigned char count;
2298
2299         /*
2300          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2301          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2302          */
2303         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2304
2305         si = swap_info_get(entry);
2306         if (!si) {
2307                 /*
2308                  * An acceptable race has occurred since the failing
2309                  * __swap_duplicate(): the swap entry has been freed,
2310                  * perhaps even the whole swap_map cleared for swapoff.
2311                  */
2312                 goto outer;
2313         }
2314
2315         offset = swp_offset(entry);
2316         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2317
2318         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2319                 /*
2320                  * The higher the swap count, the more likely it is that tasks
2321                  * will race to add swap count continuation: we need to avoid
2322                  * over-provisioning.
2323                  */
2324                 goto out;
2325         }
2326
2327         if (!page) {
2328                 spin_unlock(&swap_lock);
2329                 return -ENOMEM;
2330         }
2331
2332         /*
2333          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2334          * no architecture is using highmem pages for kernel pagetables: so it
2335          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2336          */
2337         head = vmalloc_to_page(si->swap_map + offset);
2338         offset &= ~PAGE_MASK;
2339
2340         /*
2341          * Page allocation does not initialize the page's lru field,
2342          * but it does always reset its private field.
2343          */
2344         if (!page_private(head)) {
2345                 BUG_ON(count & COUNT_CONTINUED);
2346                 INIT_LIST_HEAD(&head->lru);
2347                 set_page_private(head, SWP_CONTINUED);
2348                 si->flags |= SWP_CONTINUED;
2349         }
2350
2351         list_for_each_entry(list_page, &head->lru, lru) {
2352                 unsigned char *map;
2353
2354                 /*
2355                  * If the previous map said no continuation, but we've found
2356                  * a continuation page, free our allocation and use this one.
2357                  */
2358                 if (!(count & COUNT_CONTINUED))
2359                         goto out;
2360
2361                 map = kmap_atomic(list_page, KM_USER0) + offset;
2362                 count = *map;
2363                 kunmap_atomic(map, KM_USER0);
2364
2365                 /*
2366                  * If this continuation count now has some space in it,
2367                  * free our allocation and use this one.
2368                  */
2369                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2370                         goto out;
2371         }
2372
2373         list_add_tail(&page->lru, &head->lru);
2374         page = NULL;                    /* now it's attached, don't free it */
2375 out:
2376         spin_unlock(&swap_lock);
2377 outer:
2378         if (page)
2379                 __free_page(page);
2380         return 0;
2381 }
2382
2383 /*
2384  * swap_count_continued - when the original swap_map count is incremented
2385  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2386  * into, carry if so, or else fail until a new continuation page is allocated;
2387  * when the original swap_map count is decremented from 0 with continuation,
2388  * borrow from the continuation and report whether it still holds more.
2389  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2390  */
2391 static bool swap_count_continued(struct swap_info_struct *si,
2392                                  pgoff_t offset, unsigned char count)
2393 {
2394         struct page *head;
2395         struct page *page;
2396         unsigned char *map;
2397
2398         head = vmalloc_to_page(si->swap_map + offset);
2399         if (page_private(head) != SWP_CONTINUED) {
2400                 BUG_ON(count & COUNT_CONTINUED);
2401                 return false;           /* need to add count continuation */
2402         }
2403
2404         offset &= ~PAGE_MASK;
2405         page = list_entry(head->lru.next, struct page, lru);
2406         map = kmap_atomic(page, KM_USER0) + offset;
2407
2408         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2409                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2410
2411         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2412                 /*
2413                  * Think of how you add 1 to 999
2414                  */
2415                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2416                         kunmap_atomic(map, KM_USER0);
2417                         page = list_entry(page->lru.next, struct page, lru);
2418                         BUG_ON(page == head);
2419                         map = kmap_atomic(page, KM_USER0) + offset;
2420                 }
2421                 if (*map == SWAP_CONT_MAX) {
2422                         kunmap_atomic(map, KM_USER0);
2423                         page = list_entry(page->lru.next, struct page, lru);
2424                         if (page == head)
2425                                 return false;   /* add count continuation */
2426                         map = kmap_atomic(page, KM_USER0) + offset;
2427 init_map:               *map = 0;               /* we didn't zero the page */
2428                 }
2429                 *map += 1;
2430                 kunmap_atomic(map, KM_USER0);
2431                 page = list_entry(page->lru.prev, struct page, lru);
2432                 while (page != head) {
2433                         map = kmap_atomic(page, KM_USER0) + offset;
2434                         *map = COUNT_CONTINUED;
2435                         kunmap_atomic(map, KM_USER0);
2436                         page = list_entry(page->lru.prev, struct page, lru);
2437                 }
2438                 return true;                    /* incremented */
2439
2440         } else {                                /* decrementing */
2441                 /*
2442                  * Think of how you subtract 1 from 1000
2443                  */
2444                 BUG_ON(count != COUNT_CONTINUED);
2445                 while (*map == COUNT_CONTINUED) {
2446                         kunmap_atomic(map, KM_USER0);
2447                         page = list_entry(page->lru.next, struct page, lru);
2448                         BUG_ON(page == head);
2449                         map = kmap_atomic(page, KM_USER0) + offset;
2450                 }
2451                 BUG_ON(*map == 0);
2452                 *map -= 1;
2453                 if (*map == 0)
2454                         count = 0;
2455                 kunmap_atomic(map, KM_USER0);
2456                 page = list_entry(page->lru.prev, struct page, lru);
2457                 while (page != head) {
2458                         map = kmap_atomic(page, KM_USER0) + offset;
2459                         *map = SWAP_CONT_MAX | count;
2460                         count = COUNT_CONTINUED;
2461                         kunmap_atomic(map, KM_USER0);
2462                         page = list_entry(page->lru.prev, struct page, lru);
2463                 }
2464                 return count == COUNT_CONTINUED;
2465         }
2466 }
2467
2468 /*
2469  * free_swap_count_continuations - swapoff free all the continuation pages
2470  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2471  */
2472 static void free_swap_count_continuations(struct swap_info_struct *si)
2473 {
2474         pgoff_t offset;
2475
2476         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2477                 struct page *head;
2478                 head = vmalloc_to_page(si->swap_map + offset);
2479                 if (page_private(head)) {
2480                         struct list_head *this, *next;
2481                         list_for_each_safe(this, next, &head->lru) {
2482                                 struct page *page;
2483                                 page = list_entry(this, struct page, lru);
2484                                 list_del(this);
2485                                 __free_page(page);
2486                         }
2487                 }
2488         }
2489 }