swap_info: miscellaneous minor cleanups
[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/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32
33 #include <asm/pgtable.h>
34 #include <asm/tlbflush.h>
35 #include <linux/swapops.h>
36 #include <linux/page_cgroup.h>
37
38 static DEFINE_SPINLOCK(swap_lock);
39 static unsigned int nr_swapfiles;
40 long nr_swap_pages;
41 long total_swap_pages;
42 static int swap_overflow;
43 static int least_priority;
44
45 static const char Bad_file[] = "Bad swap file entry ";
46 static const char Unused_file[] = "Unused swap file entry ";
47 static const char Bad_offset[] = "Bad swap offset entry ";
48 static const char Unused_offset[] = "Unused swap offset entry ";
49
50 static struct swap_list_t swap_list = {-1, -1};
51
52 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
53
54 static DEFINE_MUTEX(swapon_mutex);
55
56 /* For reference count accounting in swap_map */
57 /* enum for swap_map[] handling. internal use only */
58 enum {
59         SWAP_MAP = 0,   /* ops for reference from swap users */
60         SWAP_CACHE,     /* ops for reference from swap cache */
61 };
62
63 static inline int swap_count(unsigned short ent)
64 {
65         return ent & SWAP_COUNT_MASK;
66 }
67
68 static inline bool swap_has_cache(unsigned short ent)
69 {
70         return !!(ent & SWAP_HAS_CACHE);
71 }
72
73 static inline unsigned short encode_swapmap(int count, bool has_cache)
74 {
75         unsigned short ret = count;
76
77         if (has_cache)
78                 return SWAP_HAS_CACHE | ret;
79         return ret;
80 }
81
82 /* returns 1 if swap entry is freed */
83 static int
84 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
85 {
86         swp_entry_t entry = swp_entry(si->type, offset);
87         struct page *page;
88         int ret = 0;
89
90         page = find_get_page(&swapper_space, entry.val);
91         if (!page)
92                 return 0;
93         /*
94          * This function is called from scan_swap_map() and it's called
95          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
96          * We have to use trylock for avoiding deadlock. This is a special
97          * case and you should use try_to_free_swap() with explicit lock_page()
98          * in usual operations.
99          */
100         if (trylock_page(page)) {
101                 ret = try_to_free_swap(page);
102                 unlock_page(page);
103         }
104         page_cache_release(page);
105         return ret;
106 }
107
108 /*
109  * We need this because the bdev->unplug_fn can sleep and we cannot
110  * hold swap_lock while calling the unplug_fn. And swap_lock
111  * cannot be turned into a mutex.
112  */
113 static DECLARE_RWSEM(swap_unplug_sem);
114
115 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
116 {
117         swp_entry_t entry;
118
119         down_read(&swap_unplug_sem);
120         entry.val = page_private(page);
121         if (PageSwapCache(page)) {
122                 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
123                 struct backing_dev_info *bdi;
124
125                 /*
126                  * If the page is removed from swapcache from under us (with a
127                  * racy try_to_unuse/swapoff) we need an additional reference
128                  * count to avoid reading garbage from page_private(page) above.
129                  * If the WARN_ON triggers during a swapoff it maybe the race
130                  * condition and it's harmless. However if it triggers without
131                  * swapoff it signals a problem.
132                  */
133                 WARN_ON(page_count(page) <= 1);
134
135                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
136                 blk_run_backing_dev(bdi, page);
137         }
138         up_read(&swap_unplug_sem);
139 }
140
141 /*
142  * swapon tell device that all the old swap contents can be discarded,
143  * to allow the swap device to optimize its wear-levelling.
144  */
145 static int discard_swap(struct swap_info_struct *si)
146 {
147         struct swap_extent *se;
148         sector_t start_block;
149         sector_t nr_blocks;
150         int err = 0;
151
152         /* Do not discard the swap header page! */
153         se = &si->first_swap_extent;
154         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
155         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
156         if (nr_blocks) {
157                 err = blkdev_issue_discard(si->bdev, start_block,
158                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
159                 if (err)
160                         return err;
161                 cond_resched();
162         }
163
164         list_for_each_entry(se, &si->first_swap_extent.list, list) {
165                 start_block = se->start_block << (PAGE_SHIFT - 9);
166                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
167
168                 err = blkdev_issue_discard(si->bdev, start_block,
169                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
170                 if (err)
171                         break;
172
173                 cond_resched();
174         }
175         return err;             /* That will often be -EOPNOTSUPP */
176 }
177
178 /*
179  * swap allocation tell device that a cluster of swap can now be discarded,
180  * to allow the swap device to optimize its wear-levelling.
181  */
182 static void discard_swap_cluster(struct swap_info_struct *si,
183                                  pgoff_t start_page, pgoff_t nr_pages)
184 {
185         struct swap_extent *se = si->curr_swap_extent;
186         int found_extent = 0;
187
188         while (nr_pages) {
189                 struct list_head *lh;
190
191                 if (se->start_page <= start_page &&
192                     start_page < se->start_page + se->nr_pages) {
193                         pgoff_t offset = start_page - se->start_page;
194                         sector_t start_block = se->start_block + offset;
195                         sector_t nr_blocks = se->nr_pages - offset;
196
197                         if (nr_blocks > nr_pages)
198                                 nr_blocks = nr_pages;
199                         start_page += nr_blocks;
200                         nr_pages -= nr_blocks;
201
202                         if (!found_extent++)
203                                 si->curr_swap_extent = se;
204
205                         start_block <<= PAGE_SHIFT - 9;
206                         nr_blocks <<= PAGE_SHIFT - 9;
207                         if (blkdev_issue_discard(si->bdev, start_block,
208                                     nr_blocks, GFP_NOIO, DISCARD_FL_BARRIER))
209                                 break;
210                 }
211
212                 lh = se->list.next;
213                 se = list_entry(lh, struct swap_extent, list);
214         }
215 }
216
217 static int wait_for_discard(void *word)
218 {
219         schedule();
220         return 0;
221 }
222
223 #define SWAPFILE_CLUSTER        256
224 #define LATENCY_LIMIT           256
225
226 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
227                                           int cache)
228 {
229         unsigned long offset;
230         unsigned long scan_base;
231         unsigned long last_in_cluster = 0;
232         int latency_ration = LATENCY_LIMIT;
233         int found_free_cluster = 0;
234
235         /*
236          * We try to cluster swap pages by allocating them sequentially
237          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
238          * way, however, we resort to first-free allocation, starting
239          * a new cluster.  This prevents us from scattering swap pages
240          * all over the entire swap partition, so that we reduce
241          * overall disk seek times between swap pages.  -- sct
242          * But we do now try to find an empty cluster.  -Andrea
243          * And we let swap pages go all over an SSD partition.  Hugh
244          */
245
246         si->flags += SWP_SCANNING;
247         scan_base = offset = si->cluster_next;
248
249         if (unlikely(!si->cluster_nr--)) {
250                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
251                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
252                         goto checks;
253                 }
254                 if (si->flags & SWP_DISCARDABLE) {
255                         /*
256                          * Start range check on racing allocations, in case
257                          * they overlap the cluster we eventually decide on
258                          * (we scan without swap_lock to allow preemption).
259                          * It's hardly conceivable that cluster_nr could be
260                          * wrapped during our scan, but don't depend on it.
261                          */
262                         if (si->lowest_alloc)
263                                 goto checks;
264                         si->lowest_alloc = si->max;
265                         si->highest_alloc = 0;
266                 }
267                 spin_unlock(&swap_lock);
268
269                 /*
270                  * If seek is expensive, start searching for new cluster from
271                  * start of partition, to minimize the span of allocated swap.
272                  * But if seek is cheap, search from our current position, so
273                  * that swap is allocated from all over the partition: if the
274                  * Flash Translation Layer only remaps within limited zones,
275                  * we don't want to wear out the first zone too quickly.
276                  */
277                 if (!(si->flags & SWP_SOLIDSTATE))
278                         scan_base = offset = si->lowest_bit;
279                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
280
281                 /* Locate the first empty (unaligned) cluster */
282                 for (; last_in_cluster <= si->highest_bit; offset++) {
283                         if (si->swap_map[offset])
284                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
285                         else if (offset == last_in_cluster) {
286                                 spin_lock(&swap_lock);
287                                 offset -= SWAPFILE_CLUSTER - 1;
288                                 si->cluster_next = offset;
289                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
290                                 found_free_cluster = 1;
291                                 goto checks;
292                         }
293                         if (unlikely(--latency_ration < 0)) {
294                                 cond_resched();
295                                 latency_ration = LATENCY_LIMIT;
296                         }
297                 }
298
299                 offset = si->lowest_bit;
300                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
301
302                 /* Locate the first empty (unaligned) cluster */
303                 for (; last_in_cluster < scan_base; offset++) {
304                         if (si->swap_map[offset])
305                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
306                         else if (offset == last_in_cluster) {
307                                 spin_lock(&swap_lock);
308                                 offset -= SWAPFILE_CLUSTER - 1;
309                                 si->cluster_next = offset;
310                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
311                                 found_free_cluster = 1;
312                                 goto checks;
313                         }
314                         if (unlikely(--latency_ration < 0)) {
315                                 cond_resched();
316                                 latency_ration = LATENCY_LIMIT;
317                         }
318                 }
319
320                 offset = scan_base;
321                 spin_lock(&swap_lock);
322                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
323                 si->lowest_alloc = 0;
324         }
325
326 checks:
327         if (!(si->flags & SWP_WRITEOK))
328                 goto no_page;
329         if (!si->highest_bit)
330                 goto no_page;
331         if (offset > si->highest_bit)
332                 scan_base = offset = si->lowest_bit;
333
334         /* reuse swap entry of cache-only swap if not busy. */
335         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
336                 int swap_was_freed;
337                 spin_unlock(&swap_lock);
338                 swap_was_freed = __try_to_reclaim_swap(si, offset);
339                 spin_lock(&swap_lock);
340                 /* entry was freed successfully, try to use this again */
341                 if (swap_was_freed)
342                         goto checks;
343                 goto scan; /* check next one */
344         }
345
346         if (si->swap_map[offset])
347                 goto scan;
348
349         if (offset == si->lowest_bit)
350                 si->lowest_bit++;
351         if (offset == si->highest_bit)
352                 si->highest_bit--;
353         si->inuse_pages++;
354         if (si->inuse_pages == si->pages) {
355                 si->lowest_bit = si->max;
356                 si->highest_bit = 0;
357         }
358         if (cache == SWAP_CACHE) /* at usual swap-out via vmscan.c */
359                 si->swap_map[offset] = encode_swapmap(0, true);
360         else /* at suspend */
361                 si->swap_map[offset] = encode_swapmap(1, false);
362         si->cluster_next = offset + 1;
363         si->flags -= SWP_SCANNING;
364
365         if (si->lowest_alloc) {
366                 /*
367                  * Only set when SWP_DISCARDABLE, and there's a scan
368                  * for a free cluster in progress or just completed.
369                  */
370                 if (found_free_cluster) {
371                         /*
372                          * To optimize wear-levelling, discard the
373                          * old data of the cluster, taking care not to
374                          * discard any of its pages that have already
375                          * been allocated by racing tasks (offset has
376                          * already stepped over any at the beginning).
377                          */
378                         if (offset < si->highest_alloc &&
379                             si->lowest_alloc <= last_in_cluster)
380                                 last_in_cluster = si->lowest_alloc - 1;
381                         si->flags |= SWP_DISCARDING;
382                         spin_unlock(&swap_lock);
383
384                         if (offset < last_in_cluster)
385                                 discard_swap_cluster(si, offset,
386                                         last_in_cluster - offset + 1);
387
388                         spin_lock(&swap_lock);
389                         si->lowest_alloc = 0;
390                         si->flags &= ~SWP_DISCARDING;
391
392                         smp_mb();       /* wake_up_bit advises this */
393                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
394
395                 } else if (si->flags & SWP_DISCARDING) {
396                         /*
397                          * Delay using pages allocated by racing tasks
398                          * until the whole discard has been issued. We
399                          * could defer that delay until swap_writepage,
400                          * but it's easier to keep this self-contained.
401                          */
402                         spin_unlock(&swap_lock);
403                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
404                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
405                         spin_lock(&swap_lock);
406                 } else {
407                         /*
408                          * Note pages allocated by racing tasks while
409                          * scan for a free cluster is in progress, so
410                          * that its final discard can exclude them.
411                          */
412                         if (offset < si->lowest_alloc)
413                                 si->lowest_alloc = offset;
414                         if (offset > si->highest_alloc)
415                                 si->highest_alloc = offset;
416                 }
417         }
418         return offset;
419
420 scan:
421         spin_unlock(&swap_lock);
422         while (++offset <= si->highest_bit) {
423                 if (!si->swap_map[offset]) {
424                         spin_lock(&swap_lock);
425                         goto checks;
426                 }
427                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
428                         spin_lock(&swap_lock);
429                         goto checks;
430                 }
431                 if (unlikely(--latency_ration < 0)) {
432                         cond_resched();
433                         latency_ration = LATENCY_LIMIT;
434                 }
435         }
436         offset = si->lowest_bit;
437         while (++offset < scan_base) {
438                 if (!si->swap_map[offset]) {
439                         spin_lock(&swap_lock);
440                         goto checks;
441                 }
442                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
443                         spin_lock(&swap_lock);
444                         goto checks;
445                 }
446                 if (unlikely(--latency_ration < 0)) {
447                         cond_resched();
448                         latency_ration = LATENCY_LIMIT;
449                 }
450         }
451         spin_lock(&swap_lock);
452
453 no_page:
454         si->flags -= SWP_SCANNING;
455         return 0;
456 }
457
458 swp_entry_t get_swap_page(void)
459 {
460         struct swap_info_struct *si;
461         pgoff_t offset;
462         int type, next;
463         int wrapped = 0;
464
465         spin_lock(&swap_lock);
466         if (nr_swap_pages <= 0)
467                 goto noswap;
468         nr_swap_pages--;
469
470         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
471                 si = swap_info[type];
472                 next = si->next;
473                 if (next < 0 ||
474                     (!wrapped && si->prio != swap_info[next]->prio)) {
475                         next = swap_list.head;
476                         wrapped++;
477                 }
478
479                 if (!si->highest_bit)
480                         continue;
481                 if (!(si->flags & SWP_WRITEOK))
482                         continue;
483
484                 swap_list.next = next;
485                 /* This is called for allocating swap entry for cache */
486                 offset = scan_swap_map(si, SWAP_CACHE);
487                 if (offset) {
488                         spin_unlock(&swap_lock);
489                         return swp_entry(type, offset);
490                 }
491                 next = swap_list.next;
492         }
493
494         nr_swap_pages++;
495 noswap:
496         spin_unlock(&swap_lock);
497         return (swp_entry_t) {0};
498 }
499
500 /* The only caller of this function is now susupend routine */
501 swp_entry_t get_swap_page_of_type(int type)
502 {
503         struct swap_info_struct *si;
504         pgoff_t offset;
505
506         spin_lock(&swap_lock);
507         si = swap_info[type];
508         if (si && (si->flags & SWP_WRITEOK)) {
509                 nr_swap_pages--;
510                 /* This is called for allocating swap entry, not cache */
511                 offset = scan_swap_map(si, SWAP_MAP);
512                 if (offset) {
513                         spin_unlock(&swap_lock);
514                         return swp_entry(type, offset);
515                 }
516                 nr_swap_pages++;
517         }
518         spin_unlock(&swap_lock);
519         return (swp_entry_t) {0};
520 }
521
522 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
523 {
524         struct swap_info_struct *p;
525         unsigned long offset, type;
526
527         if (!entry.val)
528                 goto out;
529         type = swp_type(entry);
530         if (type >= nr_swapfiles)
531                 goto bad_nofile;
532         p = swap_info[type];
533         if (!(p->flags & SWP_USED))
534                 goto bad_device;
535         offset = swp_offset(entry);
536         if (offset >= p->max)
537                 goto bad_offset;
538         if (!p->swap_map[offset])
539                 goto bad_free;
540         spin_lock(&swap_lock);
541         return p;
542
543 bad_free:
544         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
545         goto out;
546 bad_offset:
547         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
548         goto out;
549 bad_device:
550         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
551         goto out;
552 bad_nofile:
553         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
554 out:
555         return NULL;
556 }
557
558 static int swap_entry_free(struct swap_info_struct *p,
559                            swp_entry_t ent, int cache)
560 {
561         unsigned long offset = swp_offset(ent);
562         int count = swap_count(p->swap_map[offset]);
563         bool has_cache;
564
565         has_cache = swap_has_cache(p->swap_map[offset]);
566
567         if (cache == SWAP_MAP) { /* dropping usage count of swap */
568                 if (count < SWAP_MAP_MAX) {
569                         count--;
570                         p->swap_map[offset] = encode_swapmap(count, has_cache);
571                 }
572         } else { /* dropping swap cache flag */
573                 VM_BUG_ON(!has_cache);
574                 p->swap_map[offset] = encode_swapmap(count, false);
575
576         }
577         /* return code. */
578         count = p->swap_map[offset];
579         /* free if no reference */
580         if (!count) {
581                 if (offset < p->lowest_bit)
582                         p->lowest_bit = offset;
583                 if (offset > p->highest_bit)
584                         p->highest_bit = offset;
585                 if (swap_list.next >= 0 &&
586                     p->prio > swap_info[swap_list.next]->prio)
587                         swap_list.next = p->type;
588                 nr_swap_pages++;
589                 p->inuse_pages--;
590         }
591         if (!swap_count(count))
592                 mem_cgroup_uncharge_swap(ent);
593         return count;
594 }
595
596 /*
597  * Caller has made sure that the swapdevice corresponding to entry
598  * is still around or has not been recycled.
599  */
600 void swap_free(swp_entry_t entry)
601 {
602         struct swap_info_struct *p;
603
604         p = swap_info_get(entry);
605         if (p) {
606                 swap_entry_free(p, entry, SWAP_MAP);
607                 spin_unlock(&swap_lock);
608         }
609 }
610
611 /*
612  * Called after dropping swapcache to decrease refcnt to swap entries.
613  */
614 void swapcache_free(swp_entry_t entry, struct page *page)
615 {
616         struct swap_info_struct *p;
617         int ret;
618
619         p = swap_info_get(entry);
620         if (p) {
621                 ret = swap_entry_free(p, entry, SWAP_CACHE);
622                 if (page) {
623                         bool swapout;
624                         if (ret)
625                                 swapout = true; /* the end of swap out */
626                         else
627                                 swapout = false; /* no more swap users! */
628                         mem_cgroup_uncharge_swapcache(page, entry, swapout);
629                 }
630                 spin_unlock(&swap_lock);
631         }
632 }
633
634 /*
635  * How many references to page are currently swapped out?
636  */
637 static inline int page_swapcount(struct page *page)
638 {
639         int count = 0;
640         struct swap_info_struct *p;
641         swp_entry_t entry;
642
643         entry.val = page_private(page);
644         p = swap_info_get(entry);
645         if (p) {
646                 count = swap_count(p->swap_map[swp_offset(entry)]);
647                 spin_unlock(&swap_lock);
648         }
649         return count;
650 }
651
652 /*
653  * We can write to an anon page without COW if there are no other references
654  * to it.  And as a side-effect, free up its swap: because the old content
655  * on disk will never be read, and seeking back there to write new content
656  * later would only waste time away from clustering.
657  */
658 int reuse_swap_page(struct page *page)
659 {
660         int count;
661
662         VM_BUG_ON(!PageLocked(page));
663         count = page_mapcount(page);
664         if (count <= 1 && PageSwapCache(page)) {
665                 count += page_swapcount(page);
666                 if (count == 1 && !PageWriteback(page)) {
667                         delete_from_swap_cache(page);
668                         SetPageDirty(page);
669                 }
670         }
671         return count == 1;
672 }
673
674 /*
675  * If swap is getting full, or if there are no more mappings of this page,
676  * then try_to_free_swap is called to free its swap space.
677  */
678 int try_to_free_swap(struct page *page)
679 {
680         VM_BUG_ON(!PageLocked(page));
681
682         if (!PageSwapCache(page))
683                 return 0;
684         if (PageWriteback(page))
685                 return 0;
686         if (page_swapcount(page))
687                 return 0;
688
689         delete_from_swap_cache(page);
690         SetPageDirty(page);
691         return 1;
692 }
693
694 /*
695  * Free the swap entry like above, but also try to
696  * free the page cache entry if it is the last user.
697  */
698 int free_swap_and_cache(swp_entry_t entry)
699 {
700         struct swap_info_struct *p;
701         struct page *page = NULL;
702
703         if (non_swap_entry(entry))
704                 return 1;
705
706         p = swap_info_get(entry);
707         if (p) {
708                 if (swap_entry_free(p, entry, SWAP_MAP) == SWAP_HAS_CACHE) {
709                         page = find_get_page(&swapper_space, entry.val);
710                         if (page && !trylock_page(page)) {
711                                 page_cache_release(page);
712                                 page = NULL;
713                         }
714                 }
715                 spin_unlock(&swap_lock);
716         }
717         if (page) {
718                 /*
719                  * Not mapped elsewhere, or swap space full? Free it!
720                  * Also recheck PageSwapCache now page is locked (above).
721                  */
722                 if (PageSwapCache(page) && !PageWriteback(page) &&
723                                 (!page_mapped(page) || vm_swap_full())) {
724                         delete_from_swap_cache(page);
725                         SetPageDirty(page);
726                 }
727                 unlock_page(page);
728                 page_cache_release(page);
729         }
730         return p != NULL;
731 }
732
733 #ifdef CONFIG_HIBERNATION
734 /*
735  * Find the swap type that corresponds to given device (if any).
736  *
737  * @offset - number of the PAGE_SIZE-sized block of the device, starting
738  * from 0, in which the swap header is expected to be located.
739  *
740  * This is needed for the suspend to disk (aka swsusp).
741  */
742 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
743 {
744         struct block_device *bdev = NULL;
745         int type;
746
747         if (device)
748                 bdev = bdget(device);
749
750         spin_lock(&swap_lock);
751         for (type = 0; type < nr_swapfiles; type++) {
752                 struct swap_info_struct *sis = swap_info[type];
753
754                 if (!(sis->flags & SWP_WRITEOK))
755                         continue;
756
757                 if (!bdev) {
758                         if (bdev_p)
759                                 *bdev_p = bdgrab(sis->bdev);
760
761                         spin_unlock(&swap_lock);
762                         return type;
763                 }
764                 if (bdev == sis->bdev) {
765                         struct swap_extent *se = &sis->first_swap_extent;
766
767                         if (se->start_block == offset) {
768                                 if (bdev_p)
769                                         *bdev_p = bdgrab(sis->bdev);
770
771                                 spin_unlock(&swap_lock);
772                                 bdput(bdev);
773                                 return type;
774                         }
775                 }
776         }
777         spin_unlock(&swap_lock);
778         if (bdev)
779                 bdput(bdev);
780
781         return -ENODEV;
782 }
783
784 /*
785  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
786  * corresponding to given index in swap_info (swap type).
787  */
788 sector_t swapdev_block(int type, pgoff_t offset)
789 {
790         struct block_device *bdev;
791
792         if ((unsigned int)type >= nr_swapfiles)
793                 return 0;
794         if (!(swap_info[type]->flags & SWP_WRITEOK))
795                 return 0;
796         return map_swap_page(swp_entry(type, offset), &bdev);
797 }
798
799 /*
800  * Return either the total number of swap pages of given type, or the number
801  * of free pages of that type (depending on @free)
802  *
803  * This is needed for software suspend
804  */
805 unsigned int count_swap_pages(int type, int free)
806 {
807         unsigned int n = 0;
808
809         spin_lock(&swap_lock);
810         if ((unsigned int)type < nr_swapfiles) {
811                 struct swap_info_struct *sis = swap_info[type];
812
813                 if (sis->flags & SWP_WRITEOK) {
814                         n = sis->pages;
815                         if (free)
816                                 n -= sis->inuse_pages;
817                 }
818         }
819         spin_unlock(&swap_lock);
820         return n;
821 }
822 #endif /* CONFIG_HIBERNATION */
823
824 /*
825  * No need to decide whether this PTE shares the swap entry with others,
826  * just let do_wp_page work it out if a write is requested later - to
827  * force COW, vm_page_prot omits write permission from any private vma.
828  */
829 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
830                 unsigned long addr, swp_entry_t entry, struct page *page)
831 {
832         struct mem_cgroup *ptr = NULL;
833         spinlock_t *ptl;
834         pte_t *pte;
835         int ret = 1;
836
837         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
838                 ret = -ENOMEM;
839                 goto out_nolock;
840         }
841
842         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
843         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
844                 if (ret > 0)
845                         mem_cgroup_cancel_charge_swapin(ptr);
846                 ret = 0;
847                 goto out;
848         }
849
850         inc_mm_counter(vma->vm_mm, anon_rss);
851         get_page(page);
852         set_pte_at(vma->vm_mm, addr, pte,
853                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
854         page_add_anon_rmap(page, vma, addr);
855         mem_cgroup_commit_charge_swapin(page, ptr);
856         swap_free(entry);
857         /*
858          * Move the page to the active list so it is not
859          * immediately swapped out again after swapon.
860          */
861         activate_page(page);
862 out:
863         pte_unmap_unlock(pte, ptl);
864 out_nolock:
865         return ret;
866 }
867
868 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
869                                 unsigned long addr, unsigned long end,
870                                 swp_entry_t entry, struct page *page)
871 {
872         pte_t swp_pte = swp_entry_to_pte(entry);
873         pte_t *pte;
874         int ret = 0;
875
876         /*
877          * We don't actually need pte lock while scanning for swp_pte: since
878          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
879          * page table while we're scanning; though it could get zapped, and on
880          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
881          * of unmatched parts which look like swp_pte, so unuse_pte must
882          * recheck under pte lock.  Scanning without pte lock lets it be
883          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
884          */
885         pte = pte_offset_map(pmd, addr);
886         do {
887                 /*
888                  * swapoff spends a _lot_ of time in this loop!
889                  * Test inline before going to call unuse_pte.
890                  */
891                 if (unlikely(pte_same(*pte, swp_pte))) {
892                         pte_unmap(pte);
893                         ret = unuse_pte(vma, pmd, addr, entry, page);
894                         if (ret)
895                                 goto out;
896                         pte = pte_offset_map(pmd, addr);
897                 }
898         } while (pte++, addr += PAGE_SIZE, addr != end);
899         pte_unmap(pte - 1);
900 out:
901         return ret;
902 }
903
904 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
905                                 unsigned long addr, unsigned long end,
906                                 swp_entry_t entry, struct page *page)
907 {
908         pmd_t *pmd;
909         unsigned long next;
910         int ret;
911
912         pmd = pmd_offset(pud, addr);
913         do {
914                 next = pmd_addr_end(addr, end);
915                 if (pmd_none_or_clear_bad(pmd))
916                         continue;
917                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
918                 if (ret)
919                         return ret;
920         } while (pmd++, addr = next, addr != end);
921         return 0;
922 }
923
924 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
925                                 unsigned long addr, unsigned long end,
926                                 swp_entry_t entry, struct page *page)
927 {
928         pud_t *pud;
929         unsigned long next;
930         int ret;
931
932         pud = pud_offset(pgd, addr);
933         do {
934                 next = pud_addr_end(addr, end);
935                 if (pud_none_or_clear_bad(pud))
936                         continue;
937                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
938                 if (ret)
939                         return ret;
940         } while (pud++, addr = next, addr != end);
941         return 0;
942 }
943
944 static int unuse_vma(struct vm_area_struct *vma,
945                                 swp_entry_t entry, struct page *page)
946 {
947         pgd_t *pgd;
948         unsigned long addr, end, next;
949         int ret;
950
951         if (page->mapping) {
952                 addr = page_address_in_vma(page, vma);
953                 if (addr == -EFAULT)
954                         return 0;
955                 else
956                         end = addr + PAGE_SIZE;
957         } else {
958                 addr = vma->vm_start;
959                 end = vma->vm_end;
960         }
961
962         pgd = pgd_offset(vma->vm_mm, addr);
963         do {
964                 next = pgd_addr_end(addr, end);
965                 if (pgd_none_or_clear_bad(pgd))
966                         continue;
967                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
968                 if (ret)
969                         return ret;
970         } while (pgd++, addr = next, addr != end);
971         return 0;
972 }
973
974 static int unuse_mm(struct mm_struct *mm,
975                                 swp_entry_t entry, struct page *page)
976 {
977         struct vm_area_struct *vma;
978         int ret = 0;
979
980         if (!down_read_trylock(&mm->mmap_sem)) {
981                 /*
982                  * Activate page so shrink_inactive_list is unlikely to unmap
983                  * its ptes while lock is dropped, so swapoff can make progress.
984                  */
985                 activate_page(page);
986                 unlock_page(page);
987                 down_read(&mm->mmap_sem);
988                 lock_page(page);
989         }
990         for (vma = mm->mmap; vma; vma = vma->vm_next) {
991                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
992                         break;
993         }
994         up_read(&mm->mmap_sem);
995         return (ret < 0)? ret: 0;
996 }
997
998 /*
999  * Scan swap_map from current position to next entry still in use.
1000  * Recycle to start on reaching the end, returning 0 when empty.
1001  */
1002 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1003                                         unsigned int prev)
1004 {
1005         unsigned int max = si->max;
1006         unsigned int i = prev;
1007         int count;
1008
1009         /*
1010          * No need for swap_lock here: we're just looking
1011          * for whether an entry is in use, not modifying it; false
1012          * hits are okay, and sys_swapoff() has already prevented new
1013          * allocations from this area (while holding swap_lock).
1014          */
1015         for (;;) {
1016                 if (++i >= max) {
1017                         if (!prev) {
1018                                 i = 0;
1019                                 break;
1020                         }
1021                         /*
1022                          * No entries in use at top of swap_map,
1023                          * loop back to start and recheck there.
1024                          */
1025                         max = prev + 1;
1026                         prev = 0;
1027                         i = 1;
1028                 }
1029                 count = si->swap_map[i];
1030                 if (count && swap_count(count) != SWAP_MAP_BAD)
1031                         break;
1032         }
1033         return i;
1034 }
1035
1036 /*
1037  * We completely avoid races by reading each swap page in advance,
1038  * and then search for the process using it.  All the necessary
1039  * page table adjustments can then be made atomically.
1040  */
1041 static int try_to_unuse(unsigned int type)
1042 {
1043         struct swap_info_struct *si = swap_info[type];
1044         struct mm_struct *start_mm;
1045         unsigned short *swap_map;
1046         unsigned short swcount;
1047         struct page *page;
1048         swp_entry_t entry;
1049         unsigned int i = 0;
1050         int retval = 0;
1051         int reset_overflow = 0;
1052         int shmem;
1053
1054         /*
1055          * When searching mms for an entry, a good strategy is to
1056          * start at the first mm we freed the previous entry from
1057          * (though actually we don't notice whether we or coincidence
1058          * freed the entry).  Initialize this start_mm with a hold.
1059          *
1060          * A simpler strategy would be to start at the last mm we
1061          * freed the previous entry from; but that would take less
1062          * advantage of mmlist ordering, which clusters forked mms
1063          * together, child after parent.  If we race with dup_mmap(), we
1064          * prefer to resolve parent before child, lest we miss entries
1065          * duplicated after we scanned child: using last mm would invert
1066          * that.  Though it's only a serious concern when an overflowed
1067          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
1068          */
1069         start_mm = &init_mm;
1070         atomic_inc(&init_mm.mm_users);
1071
1072         /*
1073          * Keep on scanning until all entries have gone.  Usually,
1074          * one pass through swap_map is enough, but not necessarily:
1075          * there are races when an instance of an entry might be missed.
1076          */
1077         while ((i = find_next_to_unuse(si, i)) != 0) {
1078                 if (signal_pending(current)) {
1079                         retval = -EINTR;
1080                         break;
1081                 }
1082
1083                 /*
1084                  * Get a page for the entry, using the existing swap
1085                  * cache page if there is one.  Otherwise, get a clean
1086                  * page and read the swap into it.
1087                  */
1088                 swap_map = &si->swap_map[i];
1089                 entry = swp_entry(type, i);
1090                 page = read_swap_cache_async(entry,
1091                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1092                 if (!page) {
1093                         /*
1094                          * Either swap_duplicate() failed because entry
1095                          * has been freed independently, and will not be
1096                          * reused since sys_swapoff() already disabled
1097                          * allocation from here, or alloc_page() failed.
1098                          */
1099                         if (!*swap_map)
1100                                 continue;
1101                         retval = -ENOMEM;
1102                         break;
1103                 }
1104
1105                 /*
1106                  * Don't hold on to start_mm if it looks like exiting.
1107                  */
1108                 if (atomic_read(&start_mm->mm_users) == 1) {
1109                         mmput(start_mm);
1110                         start_mm = &init_mm;
1111                         atomic_inc(&init_mm.mm_users);
1112                 }
1113
1114                 /*
1115                  * Wait for and lock page.  When do_swap_page races with
1116                  * try_to_unuse, do_swap_page can handle the fault much
1117                  * faster than try_to_unuse can locate the entry.  This
1118                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1119                  * defer to do_swap_page in such a case - in some tests,
1120                  * do_swap_page and try_to_unuse repeatedly compete.
1121                  */
1122                 wait_on_page_locked(page);
1123                 wait_on_page_writeback(page);
1124                 lock_page(page);
1125                 wait_on_page_writeback(page);
1126
1127                 /*
1128                  * Remove all references to entry.
1129                  * Whenever we reach init_mm, there's no address space
1130                  * to search, but use it as a reminder to search shmem.
1131                  */
1132                 shmem = 0;
1133                 swcount = *swap_map;
1134                 if (swap_count(swcount)) {
1135                         if (start_mm == &init_mm)
1136                                 shmem = shmem_unuse(entry, page);
1137                         else
1138                                 retval = unuse_mm(start_mm, entry, page);
1139                 }
1140                 if (swap_count(*swap_map)) {
1141                         int set_start_mm = (*swap_map >= swcount);
1142                         struct list_head *p = &start_mm->mmlist;
1143                         struct mm_struct *new_start_mm = start_mm;
1144                         struct mm_struct *prev_mm = start_mm;
1145                         struct mm_struct *mm;
1146
1147                         atomic_inc(&new_start_mm->mm_users);
1148                         atomic_inc(&prev_mm->mm_users);
1149                         spin_lock(&mmlist_lock);
1150                         while (swap_count(*swap_map) && !retval && !shmem &&
1151                                         (p = p->next) != &start_mm->mmlist) {
1152                                 mm = list_entry(p, struct mm_struct, mmlist);
1153                                 if (!atomic_inc_not_zero(&mm->mm_users))
1154                                         continue;
1155                                 spin_unlock(&mmlist_lock);
1156                                 mmput(prev_mm);
1157                                 prev_mm = mm;
1158
1159                                 cond_resched();
1160
1161                                 swcount = *swap_map;
1162                                 if (!swap_count(swcount)) /* any usage ? */
1163                                         ;
1164                                 else if (mm == &init_mm) {
1165                                         set_start_mm = 1;
1166                                         shmem = shmem_unuse(entry, page);
1167                                 } else
1168                                         retval = unuse_mm(mm, entry, page);
1169
1170                                 if (set_start_mm && *swap_map < swcount) {
1171                                         mmput(new_start_mm);
1172                                         atomic_inc(&mm->mm_users);
1173                                         new_start_mm = mm;
1174                                         set_start_mm = 0;
1175                                 }
1176                                 spin_lock(&mmlist_lock);
1177                         }
1178                         spin_unlock(&mmlist_lock);
1179                         mmput(prev_mm);
1180                         mmput(start_mm);
1181                         start_mm = new_start_mm;
1182                 }
1183                 if (shmem) {
1184                         /* page has already been unlocked and released */
1185                         if (shmem > 0)
1186                                 continue;
1187                         retval = shmem;
1188                         break;
1189                 }
1190                 if (retval) {
1191                         unlock_page(page);
1192                         page_cache_release(page);
1193                         break;
1194                 }
1195
1196                 /*
1197                  * How could swap count reach 0x7ffe ?
1198                  * There's no way to repeat a swap page within an mm
1199                  * (except in shmem, where it's the shared object which takes
1200                  * the reference count)?
1201                  * We believe SWAP_MAP_MAX cannot occur.(if occur, unsigned
1202                  * short is too small....)
1203                  * If that's wrong, then we should worry more about
1204                  * exit_mmap() and do_munmap() cases described above:
1205                  * we might be resetting SWAP_MAP_MAX too early here.
1206                  * We know "Undead"s can happen, they're okay, so don't
1207                  * report them; but do report if we reset SWAP_MAP_MAX.
1208                  */
1209                 /* We might release the lock_page() in unuse_mm(). */
1210                 if (!PageSwapCache(page) || page_private(page) != entry.val)
1211                         goto retry;
1212
1213                 if (swap_count(*swap_map) == SWAP_MAP_MAX) {
1214                         spin_lock(&swap_lock);
1215                         *swap_map = encode_swapmap(0, true);
1216                         spin_unlock(&swap_lock);
1217                         reset_overflow = 1;
1218                 }
1219
1220                 /*
1221                  * If a reference remains (rare), we would like to leave
1222                  * the page in the swap cache; but try_to_unmap could
1223                  * then re-duplicate the entry once we drop page lock,
1224                  * so we might loop indefinitely; also, that page could
1225                  * not be swapped out to other storage meanwhile.  So:
1226                  * delete from cache even if there's another reference,
1227                  * after ensuring that the data has been saved to disk -
1228                  * since if the reference remains (rarer), it will be
1229                  * read from disk into another page.  Splitting into two
1230                  * pages would be incorrect if swap supported "shared
1231                  * private" pages, but they are handled by tmpfs files.
1232                  */
1233                 if (swap_count(*swap_map) &&
1234                      PageDirty(page) && PageSwapCache(page)) {
1235                         struct writeback_control wbc = {
1236                                 .sync_mode = WB_SYNC_NONE,
1237                         };
1238
1239                         swap_writepage(page, &wbc);
1240                         lock_page(page);
1241                         wait_on_page_writeback(page);
1242                 }
1243
1244                 /*
1245                  * It is conceivable that a racing task removed this page from
1246                  * swap cache just before we acquired the page lock at the top,
1247                  * or while we dropped it in unuse_mm().  The page might even
1248                  * be back in swap cache on another swap area: that we must not
1249                  * delete, since it may not have been written out to swap yet.
1250                  */
1251                 if (PageSwapCache(page) &&
1252                     likely(page_private(page) == entry.val))
1253                         delete_from_swap_cache(page);
1254
1255                 /*
1256                  * So we could skip searching mms once swap count went
1257                  * to 1, we did not mark any present ptes as dirty: must
1258                  * mark page dirty so shrink_page_list will preserve it.
1259                  */
1260                 SetPageDirty(page);
1261 retry:
1262                 unlock_page(page);
1263                 page_cache_release(page);
1264
1265                 /*
1266                  * Make sure that we aren't completely killing
1267                  * interactive performance.
1268                  */
1269                 cond_resched();
1270         }
1271
1272         mmput(start_mm);
1273         if (reset_overflow) {
1274                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1275                 swap_overflow = 0;
1276         }
1277         return retval;
1278 }
1279
1280 /*
1281  * After a successful try_to_unuse, if no swap is now in use, we know
1282  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1283  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1284  * added to the mmlist just after page_duplicate - before would be racy.
1285  */
1286 static void drain_mmlist(void)
1287 {
1288         struct list_head *p, *next;
1289         unsigned int type;
1290
1291         for (type = 0; type < nr_swapfiles; type++)
1292                 if (swap_info[type]->inuse_pages)
1293                         return;
1294         spin_lock(&mmlist_lock);
1295         list_for_each_safe(p, next, &init_mm.mmlist)
1296                 list_del_init(p);
1297         spin_unlock(&mmlist_lock);
1298 }
1299
1300 /*
1301  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1302  * corresponds to page offset `offset'.  Note that the type of this function
1303  * is sector_t, but it returns page offset into the bdev, not sector offset.
1304  */
1305 sector_t map_swap_page(swp_entry_t entry, struct block_device **bdev)
1306 {
1307         struct swap_info_struct *sis;
1308         struct swap_extent *start_se;
1309         struct swap_extent *se;
1310         pgoff_t offset;
1311
1312         sis = swap_info[swp_type(entry)];
1313         *bdev = sis->bdev;
1314
1315         offset = swp_offset(entry);
1316         start_se = sis->curr_swap_extent;
1317         se = start_se;
1318
1319         for ( ; ; ) {
1320                 struct list_head *lh;
1321
1322                 if (se->start_page <= offset &&
1323                                 offset < (se->start_page + se->nr_pages)) {
1324                         return se->start_block + (offset - se->start_page);
1325                 }
1326                 lh = se->list.next;
1327                 se = list_entry(lh, struct swap_extent, list);
1328                 sis->curr_swap_extent = se;
1329                 BUG_ON(se == start_se);         /* It *must* be present */
1330         }
1331 }
1332
1333 /*
1334  * Free all of a swapdev's extent information
1335  */
1336 static void destroy_swap_extents(struct swap_info_struct *sis)
1337 {
1338         while (!list_empty(&sis->first_swap_extent.list)) {
1339                 struct swap_extent *se;
1340
1341                 se = list_entry(sis->first_swap_extent.list.next,
1342                                 struct swap_extent, list);
1343                 list_del(&se->list);
1344                 kfree(se);
1345         }
1346 }
1347
1348 /*
1349  * Add a block range (and the corresponding page range) into this swapdev's
1350  * extent list.  The extent list is kept sorted in page order.
1351  *
1352  * This function rather assumes that it is called in ascending page order.
1353  */
1354 static int
1355 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1356                 unsigned long nr_pages, sector_t start_block)
1357 {
1358         struct swap_extent *se;
1359         struct swap_extent *new_se;
1360         struct list_head *lh;
1361
1362         if (start_page == 0) {
1363                 se = &sis->first_swap_extent;
1364                 sis->curr_swap_extent = se;
1365                 se->start_page = 0;
1366                 se->nr_pages = nr_pages;
1367                 se->start_block = start_block;
1368                 return 1;
1369         } else {
1370                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1371                 se = list_entry(lh, struct swap_extent, list);
1372                 BUG_ON(se->start_page + se->nr_pages != start_page);
1373                 if (se->start_block + se->nr_pages == start_block) {
1374                         /* Merge it */
1375                         se->nr_pages += nr_pages;
1376                         return 0;
1377                 }
1378         }
1379
1380         /*
1381          * No merge.  Insert a new extent, preserving ordering.
1382          */
1383         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1384         if (new_se == NULL)
1385                 return -ENOMEM;
1386         new_se->start_page = start_page;
1387         new_se->nr_pages = nr_pages;
1388         new_se->start_block = start_block;
1389
1390         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1391         return 1;
1392 }
1393
1394 /*
1395  * A `swap extent' is a simple thing which maps a contiguous range of pages
1396  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1397  * is built at swapon time and is then used at swap_writepage/swap_readpage
1398  * time for locating where on disk a page belongs.
1399  *
1400  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1401  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1402  * swap files identically.
1403  *
1404  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1405  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1406  * swapfiles are handled *identically* after swapon time.
1407  *
1408  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1409  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1410  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1411  * requirements, they are simply tossed out - we will never use those blocks
1412  * for swapping.
1413  *
1414  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1415  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1416  * which will scribble on the fs.
1417  *
1418  * The amount of disk space which a single swap extent represents varies.
1419  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1420  * extents in the list.  To avoid much list walking, we cache the previous
1421  * search location in `curr_swap_extent', and start new searches from there.
1422  * This is extremely effective.  The average number of iterations in
1423  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1424  */
1425 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1426 {
1427         struct inode *inode;
1428         unsigned blocks_per_page;
1429         unsigned long page_no;
1430         unsigned blkbits;
1431         sector_t probe_block;
1432         sector_t last_block;
1433         sector_t lowest_block = -1;
1434         sector_t highest_block = 0;
1435         int nr_extents = 0;
1436         int ret;
1437
1438         inode = sis->swap_file->f_mapping->host;
1439         if (S_ISBLK(inode->i_mode)) {
1440                 ret = add_swap_extent(sis, 0, sis->max, 0);
1441                 *span = sis->pages;
1442                 goto out;
1443         }
1444
1445         blkbits = inode->i_blkbits;
1446         blocks_per_page = PAGE_SIZE >> blkbits;
1447
1448         /*
1449          * Map all the blocks into the extent list.  This code doesn't try
1450          * to be very smart.
1451          */
1452         probe_block = 0;
1453         page_no = 0;
1454         last_block = i_size_read(inode) >> blkbits;
1455         while ((probe_block + blocks_per_page) <= last_block &&
1456                         page_no < sis->max) {
1457                 unsigned block_in_page;
1458                 sector_t first_block;
1459
1460                 first_block = bmap(inode, probe_block);
1461                 if (first_block == 0)
1462                         goto bad_bmap;
1463
1464                 /*
1465                  * It must be PAGE_SIZE aligned on-disk
1466                  */
1467                 if (first_block & (blocks_per_page - 1)) {
1468                         probe_block++;
1469                         goto reprobe;
1470                 }
1471
1472                 for (block_in_page = 1; block_in_page < blocks_per_page;
1473                                         block_in_page++) {
1474                         sector_t block;
1475
1476                         block = bmap(inode, probe_block + block_in_page);
1477                         if (block == 0)
1478                                 goto bad_bmap;
1479                         if (block != first_block + block_in_page) {
1480                                 /* Discontiguity */
1481                                 probe_block++;
1482                                 goto reprobe;
1483                         }
1484                 }
1485
1486                 first_block >>= (PAGE_SHIFT - blkbits);
1487                 if (page_no) {  /* exclude the header page */
1488                         if (first_block < lowest_block)
1489                                 lowest_block = first_block;
1490                         if (first_block > highest_block)
1491                                 highest_block = first_block;
1492                 }
1493
1494                 /*
1495                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1496                  */
1497                 ret = add_swap_extent(sis, page_no, 1, first_block);
1498                 if (ret < 0)
1499                         goto out;
1500                 nr_extents += ret;
1501                 page_no++;
1502                 probe_block += blocks_per_page;
1503 reprobe:
1504                 continue;
1505         }
1506         ret = nr_extents;
1507         *span = 1 + highest_block - lowest_block;
1508         if (page_no == 0)
1509                 page_no = 1;    /* force Empty message */
1510         sis->max = page_no;
1511         sis->pages = page_no - 1;
1512         sis->highest_bit = page_no - 1;
1513 out:
1514         return ret;
1515 bad_bmap:
1516         printk(KERN_ERR "swapon: swapfile has holes\n");
1517         ret = -EINVAL;
1518         goto out;
1519 }
1520
1521 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1522 {
1523         struct swap_info_struct *p = NULL;
1524         unsigned short *swap_map;
1525         struct file *swap_file, *victim;
1526         struct address_space *mapping;
1527         struct inode *inode;
1528         char *pathname;
1529         int i, type, prev;
1530         int err;
1531
1532         if (!capable(CAP_SYS_ADMIN))
1533                 return -EPERM;
1534
1535         pathname = getname(specialfile);
1536         err = PTR_ERR(pathname);
1537         if (IS_ERR(pathname))
1538                 goto out;
1539
1540         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1541         putname(pathname);
1542         err = PTR_ERR(victim);
1543         if (IS_ERR(victim))
1544                 goto out;
1545
1546         mapping = victim->f_mapping;
1547         prev = -1;
1548         spin_lock(&swap_lock);
1549         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1550                 p = swap_info[type];
1551                 if (p->flags & SWP_WRITEOK) {
1552                         if (p->swap_file->f_mapping == mapping)
1553                                 break;
1554                 }
1555                 prev = type;
1556         }
1557         if (type < 0) {
1558                 err = -EINVAL;
1559                 spin_unlock(&swap_lock);
1560                 goto out_dput;
1561         }
1562         if (!security_vm_enough_memory(p->pages))
1563                 vm_unacct_memory(p->pages);
1564         else {
1565                 err = -ENOMEM;
1566                 spin_unlock(&swap_lock);
1567                 goto out_dput;
1568         }
1569         if (prev < 0)
1570                 swap_list.head = p->next;
1571         else
1572                 swap_info[prev]->next = p->next;
1573         if (type == swap_list.next) {
1574                 /* just pick something that's safe... */
1575                 swap_list.next = swap_list.head;
1576         }
1577         if (p->prio < 0) {
1578                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1579                         swap_info[i]->prio = p->prio--;
1580                 least_priority++;
1581         }
1582         nr_swap_pages -= p->pages;
1583         total_swap_pages -= p->pages;
1584         p->flags &= ~SWP_WRITEOK;
1585         spin_unlock(&swap_lock);
1586
1587         current->flags |= PF_OOM_ORIGIN;
1588         err = try_to_unuse(type);
1589         current->flags &= ~PF_OOM_ORIGIN;
1590
1591         if (err) {
1592                 /* re-insert swap space back into swap_list */
1593                 spin_lock(&swap_lock);
1594                 if (p->prio < 0)
1595                         p->prio = --least_priority;
1596                 prev = -1;
1597                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1598                         if (p->prio >= swap_info[i]->prio)
1599                                 break;
1600                         prev = i;
1601                 }
1602                 p->next = i;
1603                 if (prev < 0)
1604                         swap_list.head = swap_list.next = type;
1605                 else
1606                         swap_info[prev]->next = type;
1607                 nr_swap_pages += p->pages;
1608                 total_swap_pages += p->pages;
1609                 p->flags |= SWP_WRITEOK;
1610                 spin_unlock(&swap_lock);
1611                 goto out_dput;
1612         }
1613
1614         /* wait for any unplug function to finish */
1615         down_write(&swap_unplug_sem);
1616         up_write(&swap_unplug_sem);
1617
1618         destroy_swap_extents(p);
1619         mutex_lock(&swapon_mutex);
1620         spin_lock(&swap_lock);
1621         drain_mmlist();
1622
1623         /* wait for anyone still in scan_swap_map */
1624         p->highest_bit = 0;             /* cuts scans short */
1625         while (p->flags >= SWP_SCANNING) {
1626                 spin_unlock(&swap_lock);
1627                 schedule_timeout_uninterruptible(1);
1628                 spin_lock(&swap_lock);
1629         }
1630
1631         swap_file = p->swap_file;
1632         p->swap_file = NULL;
1633         p->max = 0;
1634         swap_map = p->swap_map;
1635         p->swap_map = NULL;
1636         p->flags = 0;
1637         spin_unlock(&swap_lock);
1638         mutex_unlock(&swapon_mutex);
1639         vfree(swap_map);
1640         /* Destroy swap account informatin */
1641         swap_cgroup_swapoff(type);
1642
1643         inode = mapping->host;
1644         if (S_ISBLK(inode->i_mode)) {
1645                 struct block_device *bdev = I_BDEV(inode);
1646                 set_blocksize(bdev, p->old_block_size);
1647                 bd_release(bdev);
1648         } else {
1649                 mutex_lock(&inode->i_mutex);
1650                 inode->i_flags &= ~S_SWAPFILE;
1651                 mutex_unlock(&inode->i_mutex);
1652         }
1653         filp_close(swap_file, NULL);
1654         err = 0;
1655
1656 out_dput:
1657         filp_close(victim, NULL);
1658 out:
1659         return err;
1660 }
1661
1662 #ifdef CONFIG_PROC_FS
1663 /* iterator */
1664 static void *swap_start(struct seq_file *swap, loff_t *pos)
1665 {
1666         struct swap_info_struct *si;
1667         int type;
1668         loff_t l = *pos;
1669
1670         mutex_lock(&swapon_mutex);
1671
1672         if (!l)
1673                 return SEQ_START_TOKEN;
1674
1675         for (type = 0; type < nr_swapfiles; type++) {
1676                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1677                 si = swap_info[type];
1678                 if (!(si->flags & SWP_USED) || !si->swap_map)
1679                         continue;
1680                 if (!--l)
1681                         return si;
1682         }
1683
1684         return NULL;
1685 }
1686
1687 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1688 {
1689         struct swap_info_struct *si = v;
1690         int type;
1691
1692         if (v == SEQ_START_TOKEN)
1693                 type = 0;
1694         else
1695                 type = si->type + 1;
1696
1697         for (; type < nr_swapfiles; type++) {
1698                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1699                 si = swap_info[type];
1700                 if (!(si->flags & SWP_USED) || !si->swap_map)
1701                         continue;
1702                 ++*pos;
1703                 return si;
1704         }
1705
1706         return NULL;
1707 }
1708
1709 static void swap_stop(struct seq_file *swap, void *v)
1710 {
1711         mutex_unlock(&swapon_mutex);
1712 }
1713
1714 static int swap_show(struct seq_file *swap, void *v)
1715 {
1716         struct swap_info_struct *si = v;
1717         struct file *file;
1718         int len;
1719
1720         if (si == SEQ_START_TOKEN) {
1721                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1722                 return 0;
1723         }
1724
1725         file = si->swap_file;
1726         len = seq_path(swap, &file->f_path, " \t\n\\");
1727         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1728                         len < 40 ? 40 - len : 1, " ",
1729                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1730                                 "partition" : "file\t",
1731                         si->pages << (PAGE_SHIFT - 10),
1732                         si->inuse_pages << (PAGE_SHIFT - 10),
1733                         si->prio);
1734         return 0;
1735 }
1736
1737 static const struct seq_operations swaps_op = {
1738         .start =        swap_start,
1739         .next =         swap_next,
1740         .stop =         swap_stop,
1741         .show =         swap_show
1742 };
1743
1744 static int swaps_open(struct inode *inode, struct file *file)
1745 {
1746         return seq_open(file, &swaps_op);
1747 }
1748
1749 static const struct file_operations proc_swaps_operations = {
1750         .open           = swaps_open,
1751         .read           = seq_read,
1752         .llseek         = seq_lseek,
1753         .release        = seq_release,
1754 };
1755
1756 static int __init procswaps_init(void)
1757 {
1758         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1759         return 0;
1760 }
1761 __initcall(procswaps_init);
1762 #endif /* CONFIG_PROC_FS */
1763
1764 #ifdef MAX_SWAPFILES_CHECK
1765 static int __init max_swapfiles_check(void)
1766 {
1767         MAX_SWAPFILES_CHECK();
1768         return 0;
1769 }
1770 late_initcall(max_swapfiles_check);
1771 #endif
1772
1773 /*
1774  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1775  *
1776  * The swapon system call
1777  */
1778 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1779 {
1780         struct swap_info_struct *p;
1781         char *name = NULL;
1782         struct block_device *bdev = NULL;
1783         struct file *swap_file = NULL;
1784         struct address_space *mapping;
1785         unsigned int type;
1786         int i, prev;
1787         int error;
1788         union swap_header *swap_header = NULL;
1789         unsigned int nr_good_pages = 0;
1790         int nr_extents = 0;
1791         sector_t span;
1792         unsigned long maxpages = 1;
1793         unsigned long swapfilepages;
1794         unsigned short *swap_map = NULL;
1795         struct page *page = NULL;
1796         struct inode *inode = NULL;
1797         int did_down = 0;
1798
1799         if (!capable(CAP_SYS_ADMIN))
1800                 return -EPERM;
1801
1802         p = kzalloc(sizeof(*p), GFP_KERNEL);
1803         if (!p)
1804                 return -ENOMEM;
1805
1806         spin_lock(&swap_lock);
1807         for (type = 0; type < nr_swapfiles; type++) {
1808                 if (!(swap_info[type]->flags & SWP_USED))
1809                         break;
1810         }
1811         error = -EPERM;
1812         if (type >= MAX_SWAPFILES) {
1813                 spin_unlock(&swap_lock);
1814                 kfree(p);
1815                 goto out;
1816         }
1817         if (type >= nr_swapfiles) {
1818                 p->type = type;
1819                 swap_info[type] = p;
1820                 /*
1821                  * Write swap_info[type] before nr_swapfiles, in case a
1822                  * racing procfs swap_start() or swap_next() is reading them.
1823                  * (We never shrink nr_swapfiles, we never free this entry.)
1824                  */
1825                 smp_wmb();
1826                 nr_swapfiles++;
1827         } else {
1828                 kfree(p);
1829                 p = swap_info[type];
1830                 /*
1831                  * Do not memset this entry: a racing procfs swap_next()
1832                  * would be relying on p->type to remain valid.
1833                  */
1834         }
1835         INIT_LIST_HEAD(&p->first_swap_extent.list);
1836         p->flags = SWP_USED;
1837         p->next = -1;
1838         spin_unlock(&swap_lock);
1839
1840         name = getname(specialfile);
1841         error = PTR_ERR(name);
1842         if (IS_ERR(name)) {
1843                 name = NULL;
1844                 goto bad_swap_2;
1845         }
1846         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1847         error = PTR_ERR(swap_file);
1848         if (IS_ERR(swap_file)) {
1849                 swap_file = NULL;
1850                 goto bad_swap_2;
1851         }
1852
1853         p->swap_file = swap_file;
1854         mapping = swap_file->f_mapping;
1855         inode = mapping->host;
1856
1857         error = -EBUSY;
1858         for (i = 0; i < nr_swapfiles; i++) {
1859                 struct swap_info_struct *q = swap_info[i];
1860
1861                 if (i == type || !q->swap_file)
1862                         continue;
1863                 if (mapping == q->swap_file->f_mapping)
1864                         goto bad_swap;
1865         }
1866
1867         error = -EINVAL;
1868         if (S_ISBLK(inode->i_mode)) {
1869                 bdev = I_BDEV(inode);
1870                 error = bd_claim(bdev, sys_swapon);
1871                 if (error < 0) {
1872                         bdev = NULL;
1873                         error = -EINVAL;
1874                         goto bad_swap;
1875                 }
1876                 p->old_block_size = block_size(bdev);
1877                 error = set_blocksize(bdev, PAGE_SIZE);
1878                 if (error < 0)
1879                         goto bad_swap;
1880                 p->bdev = bdev;
1881         } else if (S_ISREG(inode->i_mode)) {
1882                 p->bdev = inode->i_sb->s_bdev;
1883                 mutex_lock(&inode->i_mutex);
1884                 did_down = 1;
1885                 if (IS_SWAPFILE(inode)) {
1886                         error = -EBUSY;
1887                         goto bad_swap;
1888                 }
1889         } else {
1890                 goto bad_swap;
1891         }
1892
1893         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1894
1895         /*
1896          * Read the swap header.
1897          */
1898         if (!mapping->a_ops->readpage) {
1899                 error = -EINVAL;
1900                 goto bad_swap;
1901         }
1902         page = read_mapping_page(mapping, 0, swap_file);
1903         if (IS_ERR(page)) {
1904                 error = PTR_ERR(page);
1905                 goto bad_swap;
1906         }
1907         swap_header = kmap(page);
1908
1909         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1910                 printk(KERN_ERR "Unable to find swap-space signature\n");
1911                 error = -EINVAL;
1912                 goto bad_swap;
1913         }
1914
1915         /* swap partition endianess hack... */
1916         if (swab32(swap_header->info.version) == 1) {
1917                 swab32s(&swap_header->info.version);
1918                 swab32s(&swap_header->info.last_page);
1919                 swab32s(&swap_header->info.nr_badpages);
1920                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1921                         swab32s(&swap_header->info.badpages[i]);
1922         }
1923         /* Check the swap header's sub-version */
1924         if (swap_header->info.version != 1) {
1925                 printk(KERN_WARNING
1926                        "Unable to handle swap header version %d\n",
1927                        swap_header->info.version);
1928                 error = -EINVAL;
1929                 goto bad_swap;
1930         }
1931
1932         p->lowest_bit  = 1;
1933         p->cluster_next = 1;
1934         p->cluster_nr = 0;
1935
1936         /*
1937          * Find out how many pages are allowed for a single swap
1938          * device. There are two limiting factors: 1) the number of
1939          * bits for the swap offset in the swp_entry_t type and
1940          * 2) the number of bits in the a swap pte as defined by
1941          * the different architectures. In order to find the
1942          * largest possible bit mask a swap entry with swap type 0
1943          * and swap offset ~0UL is created, encoded to a swap pte,
1944          * decoded to a swp_entry_t again and finally the swap
1945          * offset is extracted. This will mask all the bits from
1946          * the initial ~0UL mask that can't be encoded in either
1947          * the swp_entry_t or the architecture definition of a
1948          * swap pte.
1949          */
1950         maxpages = swp_offset(pte_to_swp_entry(
1951                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1952         if (maxpages > swap_header->info.last_page)
1953                 maxpages = swap_header->info.last_page;
1954         p->highest_bit = maxpages - 1;
1955
1956         error = -EINVAL;
1957         if (!maxpages)
1958                 goto bad_swap;
1959         if (swapfilepages && maxpages > swapfilepages) {
1960                 printk(KERN_WARNING
1961                        "Swap area shorter than signature indicates\n");
1962                 goto bad_swap;
1963         }
1964         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1965                 goto bad_swap;
1966         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1967                 goto bad_swap;
1968
1969         /* OK, set up the swap map and apply the bad block list */
1970         swap_map = vmalloc(maxpages * sizeof(short));
1971         if (!swap_map) {
1972                 error = -ENOMEM;
1973                 goto bad_swap;
1974         }
1975
1976         memset(swap_map, 0, maxpages * sizeof(short));
1977         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1978                 int page_nr = swap_header->info.badpages[i];
1979                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1980                         error = -EINVAL;
1981                         goto bad_swap;
1982                 }
1983                 swap_map[page_nr] = SWAP_MAP_BAD;
1984         }
1985
1986         error = swap_cgroup_swapon(type, maxpages);
1987         if (error)
1988                 goto bad_swap;
1989
1990         nr_good_pages = swap_header->info.last_page -
1991                         swap_header->info.nr_badpages -
1992                         1 /* header page */;
1993
1994         if (nr_good_pages) {
1995                 swap_map[0] = SWAP_MAP_BAD;
1996                 p->max = maxpages;
1997                 p->pages = nr_good_pages;
1998                 nr_extents = setup_swap_extents(p, &span);
1999                 if (nr_extents < 0) {
2000                         error = nr_extents;
2001                         goto bad_swap;
2002                 }
2003                 nr_good_pages = p->pages;
2004         }
2005         if (!nr_good_pages) {
2006                 printk(KERN_WARNING "Empty swap-file\n");
2007                 error = -EINVAL;
2008                 goto bad_swap;
2009         }
2010
2011         if (p->bdev) {
2012                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2013                         p->flags |= SWP_SOLIDSTATE;
2014                         p->cluster_next = 1 + (random32() % p->highest_bit);
2015                 }
2016                 if (discard_swap(p) == 0)
2017                         p->flags |= SWP_DISCARDABLE;
2018         }
2019
2020         mutex_lock(&swapon_mutex);
2021         spin_lock(&swap_lock);
2022         if (swap_flags & SWAP_FLAG_PREFER)
2023                 p->prio =
2024                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2025         else
2026                 p->prio = --least_priority;
2027         p->swap_map = swap_map;
2028         p->flags |= SWP_WRITEOK;
2029         nr_swap_pages += nr_good_pages;
2030         total_swap_pages += nr_good_pages;
2031
2032         printk(KERN_INFO "Adding %uk swap on %s.  "
2033                         "Priority:%d extents:%d across:%lluk %s%s\n",
2034                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2035                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2036                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2037                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2038
2039         /* insert swap space into swap_list: */
2040         prev = -1;
2041         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2042                 if (p->prio >= swap_info[i]->prio)
2043                         break;
2044                 prev = i;
2045         }
2046         p->next = i;
2047         if (prev < 0)
2048                 swap_list.head = swap_list.next = type;
2049         else
2050                 swap_info[prev]->next = type;
2051         spin_unlock(&swap_lock);
2052         mutex_unlock(&swapon_mutex);
2053         error = 0;
2054         goto out;
2055 bad_swap:
2056         if (bdev) {
2057                 set_blocksize(bdev, p->old_block_size);
2058                 bd_release(bdev);
2059         }
2060         destroy_swap_extents(p);
2061         swap_cgroup_swapoff(type);
2062 bad_swap_2:
2063         spin_lock(&swap_lock);
2064         p->swap_file = NULL;
2065         p->flags = 0;
2066         spin_unlock(&swap_lock);
2067         vfree(swap_map);
2068         if (swap_file)
2069                 filp_close(swap_file, NULL);
2070 out:
2071         if (page && !IS_ERR(page)) {
2072                 kunmap(page);
2073                 page_cache_release(page);
2074         }
2075         if (name)
2076                 putname(name);
2077         if (did_down) {
2078                 if (!error)
2079                         inode->i_flags |= S_SWAPFILE;
2080                 mutex_unlock(&inode->i_mutex);
2081         }
2082         return error;
2083 }
2084
2085 void si_swapinfo(struct sysinfo *val)
2086 {
2087         unsigned int type;
2088         unsigned long nr_to_be_unused = 0;
2089
2090         spin_lock(&swap_lock);
2091         for (type = 0; type < nr_swapfiles; type++) {
2092                 struct swap_info_struct *si = swap_info[type];
2093
2094                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2095                         nr_to_be_unused += si->inuse_pages;
2096         }
2097         val->freeswap = nr_swap_pages + nr_to_be_unused;
2098         val->totalswap = total_swap_pages + nr_to_be_unused;
2099         spin_unlock(&swap_lock);
2100 }
2101
2102 /*
2103  * Verify that a swap entry is valid and increment its swap map count.
2104  *
2105  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
2106  * "permanent", but will be reclaimed by the next swapoff.
2107  * Returns error code in following case.
2108  * - success -> 0
2109  * - swp_entry is invalid -> EINVAL
2110  * - swp_entry is migration entry -> EINVAL
2111  * - swap-cache reference is requested but there is already one. -> EEXIST
2112  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2113  */
2114 static int __swap_duplicate(swp_entry_t entry, bool cache)
2115 {
2116         struct swap_info_struct *p;
2117         unsigned long offset, type;
2118         int result = -EINVAL;
2119         int count;
2120         bool has_cache;
2121
2122         if (non_swap_entry(entry))
2123                 return -EINVAL;
2124
2125         type = swp_type(entry);
2126         if (type >= nr_swapfiles)
2127                 goto bad_file;
2128         p = swap_info[type];
2129         offset = swp_offset(entry);
2130
2131         spin_lock(&swap_lock);
2132
2133         if (unlikely(offset >= p->max))
2134                 goto unlock_out;
2135
2136         count = swap_count(p->swap_map[offset]);
2137         has_cache = swap_has_cache(p->swap_map[offset]);
2138
2139         if (cache == SWAP_CACHE) { /* called for swapcache/swapin-readahead */
2140
2141                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2142                 if (!has_cache && count) {
2143                         p->swap_map[offset] = encode_swapmap(count, true);
2144                         result = 0;
2145                 } else if (has_cache) /* someone added cache */
2146                         result = -EEXIST;
2147                 else if (!count) /* no users */
2148                         result = -ENOENT;
2149
2150         } else if (count || has_cache) {
2151                 if (count < SWAP_MAP_MAX - 1) {
2152                         p->swap_map[offset] = encode_swapmap(count + 1,
2153                                                              has_cache);
2154                         result = 0;
2155                 } else if (count <= SWAP_MAP_MAX) {
2156                         if (swap_overflow++ < 5)
2157                                 printk(KERN_WARNING
2158                                        "swap_dup: swap entry overflow\n");
2159                         p->swap_map[offset] = encode_swapmap(SWAP_MAP_MAX,
2160                                                               has_cache);
2161                         result = 0;
2162                 }
2163         } else
2164                 result = -ENOENT; /* unused swap entry */
2165 unlock_out:
2166         spin_unlock(&swap_lock);
2167 out:
2168         return result;
2169
2170 bad_file:
2171         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2172         goto out;
2173 }
2174 /*
2175  * increase reference count of swap entry by 1.
2176  */
2177 void swap_duplicate(swp_entry_t entry)
2178 {
2179         __swap_duplicate(entry, SWAP_MAP);
2180 }
2181
2182 /*
2183  * @entry: swap entry for which we allocate swap cache.
2184  *
2185  * Called when allocating swap cache for existing swap entry,
2186  * This can return error codes. Returns 0 at success.
2187  * -EBUSY means there is a swap cache.
2188  * Note: return code is different from swap_duplicate().
2189  */
2190 int swapcache_prepare(swp_entry_t entry)
2191 {
2192         return __swap_duplicate(entry, SWAP_CACHE);
2193 }
2194
2195 /*
2196  * swap_lock prevents swap_map being freed. Don't grab an extra
2197  * reference on the swaphandle, it doesn't matter if it becomes unused.
2198  */
2199 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2200 {
2201         struct swap_info_struct *si;
2202         int our_page_cluster = page_cluster;
2203         pgoff_t target, toff;
2204         pgoff_t base, end;
2205         int nr_pages = 0;
2206
2207         if (!our_page_cluster)  /* no readahead */
2208                 return 0;
2209
2210         si = swap_info[swp_type(entry)];
2211         target = swp_offset(entry);
2212         base = (target >> our_page_cluster) << our_page_cluster;
2213         end = base + (1 << our_page_cluster);
2214         if (!base)              /* first page is swap header */
2215                 base++;
2216
2217         spin_lock(&swap_lock);
2218         if (end > si->max)      /* don't go beyond end of map */
2219                 end = si->max;
2220
2221         /* Count contiguous allocated slots above our target */
2222         for (toff = target; ++toff < end; nr_pages++) {
2223                 /* Don't read in free or bad pages */
2224                 if (!si->swap_map[toff])
2225                         break;
2226                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2227                         break;
2228         }
2229         /* Count contiguous allocated slots below our target */
2230         for (toff = target; --toff >= base; nr_pages++) {
2231                 /* Don't read in free or bad pages */
2232                 if (!si->swap_map[toff])
2233                         break;
2234                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2235                         break;
2236         }
2237         spin_unlock(&swap_lock);
2238
2239         /*
2240          * Indicate starting offset, and return number of pages to get:
2241          * if only 1, say 0, since there's then no readahead to be done.
2242          */
2243         *offset = ++toff;
2244         return nr_pages? ++nr_pages: 0;
2245 }