swap_info: change to array of pointers
[kernel.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         int err = 0;
149
150         list_for_each_entry(se, &si->extent_list, list) {
151                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
152                 sector_t nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
153
154                 if (se->start_page == 0) {
155                         /* Do not discard the swap header page! */
156                         start_block += 1 << (PAGE_SHIFT - 9);
157                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
158                         if (!nr_blocks)
159                                 continue;
160                 }
161
162                 err = blkdev_issue_discard(si->bdev, start_block,
163                                                 nr_blocks, GFP_KERNEL,
164                                                 DISCARD_FL_BARRIER);
165                 if (err)
166                         break;
167
168                 cond_resched();
169         }
170         return err;             /* That will often be -EOPNOTSUPP */
171 }
172
173 /*
174  * swap allocation tell device that a cluster of swap can now be discarded,
175  * to allow the swap device to optimize its wear-levelling.
176  */
177 static void discard_swap_cluster(struct swap_info_struct *si,
178                                  pgoff_t start_page, pgoff_t nr_pages)
179 {
180         struct swap_extent *se = si->curr_swap_extent;
181         int found_extent = 0;
182
183         while (nr_pages) {
184                 struct list_head *lh;
185
186                 if (se->start_page <= start_page &&
187                     start_page < se->start_page + se->nr_pages) {
188                         pgoff_t offset = start_page - se->start_page;
189                         sector_t start_block = se->start_block + offset;
190                         sector_t nr_blocks = se->nr_pages - offset;
191
192                         if (nr_blocks > nr_pages)
193                                 nr_blocks = nr_pages;
194                         start_page += nr_blocks;
195                         nr_pages -= nr_blocks;
196
197                         if (!found_extent++)
198                                 si->curr_swap_extent = se;
199
200                         start_block <<= PAGE_SHIFT - 9;
201                         nr_blocks <<= PAGE_SHIFT - 9;
202                         if (blkdev_issue_discard(si->bdev, start_block,
203                                                         nr_blocks, GFP_NOIO,
204                                                         DISCARD_FL_BARRIER))
205                                 break;
206                 }
207
208                 lh = se->list.next;
209                 if (lh == &si->extent_list)
210                         lh = lh->next;
211                 se = list_entry(lh, struct swap_extent, list);
212         }
213 }
214
215 static int wait_for_discard(void *word)
216 {
217         schedule();
218         return 0;
219 }
220
221 #define SWAPFILE_CLUSTER        256
222 #define LATENCY_LIMIT           256
223
224 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
225                                           int cache)
226 {
227         unsigned long offset;
228         unsigned long scan_base;
229         unsigned long last_in_cluster = 0;
230         int latency_ration = LATENCY_LIMIT;
231         int found_free_cluster = 0;
232
233         /*
234          * We try to cluster swap pages by allocating them sequentially
235          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
236          * way, however, we resort to first-free allocation, starting
237          * a new cluster.  This prevents us from scattering swap pages
238          * all over the entire swap partition, so that we reduce
239          * overall disk seek times between swap pages.  -- sct
240          * But we do now try to find an empty cluster.  -Andrea
241          * And we let swap pages go all over an SSD partition.  Hugh
242          */
243
244         si->flags += SWP_SCANNING;
245         scan_base = offset = si->cluster_next;
246
247         if (unlikely(!si->cluster_nr--)) {
248                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
249                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
250                         goto checks;
251                 }
252                 if (si->flags & SWP_DISCARDABLE) {
253                         /*
254                          * Start range check on racing allocations, in case
255                          * they overlap the cluster we eventually decide on
256                          * (we scan without swap_lock to allow preemption).
257                          * It's hardly conceivable that cluster_nr could be
258                          * wrapped during our scan, but don't depend on it.
259                          */
260                         if (si->lowest_alloc)
261                                 goto checks;
262                         si->lowest_alloc = si->max;
263                         si->highest_alloc = 0;
264                 }
265                 spin_unlock(&swap_lock);
266
267                 /*
268                  * If seek is expensive, start searching for new cluster from
269                  * start of partition, to minimize the span of allocated swap.
270                  * But if seek is cheap, search from our current position, so
271                  * that swap is allocated from all over the partition: if the
272                  * Flash Translation Layer only remaps within limited zones,
273                  * we don't want to wear out the first zone too quickly.
274                  */
275                 if (!(si->flags & SWP_SOLIDSTATE))
276                         scan_base = offset = si->lowest_bit;
277                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
278
279                 /* Locate the first empty (unaligned) cluster */
280                 for (; last_in_cluster <= si->highest_bit; offset++) {
281                         if (si->swap_map[offset])
282                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
283                         else if (offset == last_in_cluster) {
284                                 spin_lock(&swap_lock);
285                                 offset -= SWAPFILE_CLUSTER - 1;
286                                 si->cluster_next = offset;
287                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
288                                 found_free_cluster = 1;
289                                 goto checks;
290                         }
291                         if (unlikely(--latency_ration < 0)) {
292                                 cond_resched();
293                                 latency_ration = LATENCY_LIMIT;
294                         }
295                 }
296
297                 offset = si->lowest_bit;
298                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
299
300                 /* Locate the first empty (unaligned) cluster */
301                 for (; last_in_cluster < scan_base; offset++) {
302                         if (si->swap_map[offset])
303                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
304                         else if (offset == last_in_cluster) {
305                                 spin_lock(&swap_lock);
306                                 offset -= SWAPFILE_CLUSTER - 1;
307                                 si->cluster_next = offset;
308                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
309                                 found_free_cluster = 1;
310                                 goto checks;
311                         }
312                         if (unlikely(--latency_ration < 0)) {
313                                 cond_resched();
314                                 latency_ration = LATENCY_LIMIT;
315                         }
316                 }
317
318                 offset = scan_base;
319                 spin_lock(&swap_lock);
320                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
321                 si->lowest_alloc = 0;
322         }
323
324 checks:
325         if (!(si->flags & SWP_WRITEOK))
326                 goto no_page;
327         if (!si->highest_bit)
328                 goto no_page;
329         if (offset > si->highest_bit)
330                 scan_base = offset = si->lowest_bit;
331
332         /* reuse swap entry of cache-only swap if not busy. */
333         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
334                 int swap_was_freed;
335                 spin_unlock(&swap_lock);
336                 swap_was_freed = __try_to_reclaim_swap(si, offset);
337                 spin_lock(&swap_lock);
338                 /* entry was freed successfully, try to use this again */
339                 if (swap_was_freed)
340                         goto checks;
341                 goto scan; /* check next one */
342         }
343
344         if (si->swap_map[offset])
345                 goto scan;
346
347         if (offset == si->lowest_bit)
348                 si->lowest_bit++;
349         if (offset == si->highest_bit)
350                 si->highest_bit--;
351         si->inuse_pages++;
352         if (si->inuse_pages == si->pages) {
353                 si->lowest_bit = si->max;
354                 si->highest_bit = 0;
355         }
356         if (cache == SWAP_CACHE) /* at usual swap-out via vmscan.c */
357                 si->swap_map[offset] = encode_swapmap(0, true);
358         else /* at suspend */
359                 si->swap_map[offset] = encode_swapmap(1, false);
360         si->cluster_next = offset + 1;
361         si->flags -= SWP_SCANNING;
362
363         if (si->lowest_alloc) {
364                 /*
365                  * Only set when SWP_DISCARDABLE, and there's a scan
366                  * for a free cluster in progress or just completed.
367                  */
368                 if (found_free_cluster) {
369                         /*
370                          * To optimize wear-levelling, discard the
371                          * old data of the cluster, taking care not to
372                          * discard any of its pages that have already
373                          * been allocated by racing tasks (offset has
374                          * already stepped over any at the beginning).
375                          */
376                         if (offset < si->highest_alloc &&
377                             si->lowest_alloc <= last_in_cluster)
378                                 last_in_cluster = si->lowest_alloc - 1;
379                         si->flags |= SWP_DISCARDING;
380                         spin_unlock(&swap_lock);
381
382                         if (offset < last_in_cluster)
383                                 discard_swap_cluster(si, offset,
384                                         last_in_cluster - offset + 1);
385
386                         spin_lock(&swap_lock);
387                         si->lowest_alloc = 0;
388                         si->flags &= ~SWP_DISCARDING;
389
390                         smp_mb();       /* wake_up_bit advises this */
391                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
392
393                 } else if (si->flags & SWP_DISCARDING) {
394                         /*
395                          * Delay using pages allocated by racing tasks
396                          * until the whole discard has been issued. We
397                          * could defer that delay until swap_writepage,
398                          * but it's easier to keep this self-contained.
399                          */
400                         spin_unlock(&swap_lock);
401                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
402                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
403                         spin_lock(&swap_lock);
404                 } else {
405                         /*
406                          * Note pages allocated by racing tasks while
407                          * scan for a free cluster is in progress, so
408                          * that its final discard can exclude them.
409                          */
410                         if (offset < si->lowest_alloc)
411                                 si->lowest_alloc = offset;
412                         if (offset > si->highest_alloc)
413                                 si->highest_alloc = offset;
414                 }
415         }
416         return offset;
417
418 scan:
419         spin_unlock(&swap_lock);
420         while (++offset <= si->highest_bit) {
421                 if (!si->swap_map[offset]) {
422                         spin_lock(&swap_lock);
423                         goto checks;
424                 }
425                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
426                         spin_lock(&swap_lock);
427                         goto checks;
428                 }
429                 if (unlikely(--latency_ration < 0)) {
430                         cond_resched();
431                         latency_ration = LATENCY_LIMIT;
432                 }
433         }
434         offset = si->lowest_bit;
435         while (++offset < scan_base) {
436                 if (!si->swap_map[offset]) {
437                         spin_lock(&swap_lock);
438                         goto checks;
439                 }
440                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
441                         spin_lock(&swap_lock);
442                         goto checks;
443                 }
444                 if (unlikely(--latency_ration < 0)) {
445                         cond_resched();
446                         latency_ration = LATENCY_LIMIT;
447                 }
448         }
449         spin_lock(&swap_lock);
450
451 no_page:
452         si->flags -= SWP_SCANNING;
453         return 0;
454 }
455
456 swp_entry_t get_swap_page(void)
457 {
458         struct swap_info_struct *si;
459         pgoff_t offset;
460         int type, next;
461         int wrapped = 0;
462
463         spin_lock(&swap_lock);
464         if (nr_swap_pages <= 0)
465                 goto noswap;
466         nr_swap_pages--;
467
468         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
469                 si = swap_info[type];
470                 next = si->next;
471                 if (next < 0 ||
472                     (!wrapped && si->prio != swap_info[next]->prio)) {
473                         next = swap_list.head;
474                         wrapped++;
475                 }
476
477                 if (!si->highest_bit)
478                         continue;
479                 if (!(si->flags & SWP_WRITEOK))
480                         continue;
481
482                 swap_list.next = next;
483                 /* This is called for allocating swap entry for cache */
484                 offset = scan_swap_map(si, SWAP_CACHE);
485                 if (offset) {
486                         spin_unlock(&swap_lock);
487                         return swp_entry(type, offset);
488                 }
489                 next = swap_list.next;
490         }
491
492         nr_swap_pages++;
493 noswap:
494         spin_unlock(&swap_lock);
495         return (swp_entry_t) {0};
496 }
497
498 /* The only caller of this function is now susupend routine */
499 swp_entry_t get_swap_page_of_type(int type)
500 {
501         struct swap_info_struct *si;
502         pgoff_t offset;
503
504         spin_lock(&swap_lock);
505         si = swap_info[type];
506         if (si && (si->flags & SWP_WRITEOK)) {
507                 nr_swap_pages--;
508                 /* This is called for allocating swap entry, not cache */
509                 offset = scan_swap_map(si, SWAP_MAP);
510                 if (offset) {
511                         spin_unlock(&swap_lock);
512                         return swp_entry(type, offset);
513                 }
514                 nr_swap_pages++;
515         }
516         spin_unlock(&swap_lock);
517         return (swp_entry_t) {0};
518 }
519
520 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
521 {
522         struct swap_info_struct * p;
523         unsigned long offset, type;
524
525         if (!entry.val)
526                 goto out;
527         type = swp_type(entry);
528         if (type >= nr_swapfiles)
529                 goto bad_nofile;
530         p = swap_info[type];
531         if (!(p->flags & SWP_USED))
532                 goto bad_device;
533         offset = swp_offset(entry);
534         if (offset >= p->max)
535                 goto bad_offset;
536         if (!p->swap_map[offset])
537                 goto bad_free;
538         spin_lock(&swap_lock);
539         return p;
540
541 bad_free:
542         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
543         goto out;
544 bad_offset:
545         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
546         goto out;
547 bad_device:
548         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
549         goto out;
550 bad_nofile:
551         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
552 out:
553         return NULL;
554 }
555
556 static int swap_entry_free(struct swap_info_struct *p,
557                            swp_entry_t ent, int cache)
558 {
559         unsigned long offset = swp_offset(ent);
560         int count = swap_count(p->swap_map[offset]);
561         bool has_cache;
562
563         has_cache = swap_has_cache(p->swap_map[offset]);
564
565         if (cache == SWAP_MAP) { /* dropping usage count of swap */
566                 if (count < SWAP_MAP_MAX) {
567                         count--;
568                         p->swap_map[offset] = encode_swapmap(count, has_cache);
569                 }
570         } else { /* dropping swap cache flag */
571                 VM_BUG_ON(!has_cache);
572                 p->swap_map[offset] = encode_swapmap(count, false);
573
574         }
575         /* return code. */
576         count = p->swap_map[offset];
577         /* free if no reference */
578         if (!count) {
579                 if (offset < p->lowest_bit)
580                         p->lowest_bit = offset;
581                 if (offset > p->highest_bit)
582                         p->highest_bit = offset;
583                 if (swap_list.next >= 0 &&
584                     p->prio > swap_info[swap_list.next]->prio)
585                         swap_list.next = p->type;
586                 nr_swap_pages++;
587                 p->inuse_pages--;
588         }
589         if (!swap_count(count))
590                 mem_cgroup_uncharge_swap(ent);
591         return count;
592 }
593
594 /*
595  * Caller has made sure that the swapdevice corresponding to entry
596  * is still around or has not been recycled.
597  */
598 void swap_free(swp_entry_t entry)
599 {
600         struct swap_info_struct * p;
601
602         p = swap_info_get(entry);
603         if (p) {
604                 swap_entry_free(p, entry, SWAP_MAP);
605                 spin_unlock(&swap_lock);
606         }
607 }
608
609 /*
610  * Called after dropping swapcache to decrease refcnt to swap entries.
611  */
612 void swapcache_free(swp_entry_t entry, struct page *page)
613 {
614         struct swap_info_struct *p;
615         int ret;
616
617         p = swap_info_get(entry);
618         if (p) {
619                 ret = swap_entry_free(p, entry, SWAP_CACHE);
620                 if (page) {
621                         bool swapout;
622                         if (ret)
623                                 swapout = true; /* the end of swap out */
624                         else
625                                 swapout = false; /* no more swap users! */
626                         mem_cgroup_uncharge_swapcache(page, entry, swapout);
627                 }
628                 spin_unlock(&swap_lock);
629         }
630         return;
631 }
632
633 /*
634  * How many references to page are currently swapped out?
635  */
636 static inline int page_swapcount(struct page *page)
637 {
638         int count = 0;
639         struct swap_info_struct *p;
640         swp_entry_t entry;
641
642         entry.val = page_private(page);
643         p = swap_info_get(entry);
644         if (p) {
645                 count = swap_count(p->swap_map[swp_offset(entry)]);
646                 spin_unlock(&swap_lock);
647         }
648         return count;
649 }
650
651 /*
652  * We can write to an anon page without COW if there are no other references
653  * to it.  And as a side-effect, free up its swap: because the old content
654  * on disk will never be read, and seeking back there to write new content
655  * later would only waste time away from clustering.
656  */
657 int reuse_swap_page(struct page *page)
658 {
659         int count;
660
661         VM_BUG_ON(!PageLocked(page));
662         count = page_mapcount(page);
663         if (count <= 1 && PageSwapCache(page)) {
664                 count += page_swapcount(page);
665                 if (count == 1 && !PageWriteback(page)) {
666                         delete_from_swap_cache(page);
667                         SetPageDirty(page);
668                 }
669         }
670         return count == 1;
671 }
672
673 /*
674  * If swap is getting full, or if there are no more mappings of this page,
675  * then try_to_free_swap is called to free its swap space.
676  */
677 int try_to_free_swap(struct page *page)
678 {
679         VM_BUG_ON(!PageLocked(page));
680
681         if (!PageSwapCache(page))
682                 return 0;
683         if (PageWriteback(page))
684                 return 0;
685         if (page_swapcount(page))
686                 return 0;
687
688         delete_from_swap_cache(page);
689         SetPageDirty(page);
690         return 1;
691 }
692
693 /*
694  * Free the swap entry like above, but also try to
695  * free the page cache entry if it is the last user.
696  */
697 int free_swap_and_cache(swp_entry_t entry)
698 {
699         struct swap_info_struct *p;
700         struct page *page = NULL;
701
702         if (non_swap_entry(entry))
703                 return 1;
704
705         p = swap_info_get(entry);
706         if (p) {
707                 if (swap_entry_free(p, entry, SWAP_MAP) == SWAP_HAS_CACHE) {
708                         page = find_get_page(&swapper_space, entry.val);
709                         if (page && !trylock_page(page)) {
710                                 page_cache_release(page);
711                                 page = NULL;
712                         }
713                 }
714                 spin_unlock(&swap_lock);
715         }
716         if (page) {
717                 /*
718                  * Not mapped elsewhere, or swap space full? Free it!
719                  * Also recheck PageSwapCache now page is locked (above).
720                  */
721                 if (PageSwapCache(page) && !PageWriteback(page) &&
722                                 (!page_mapped(page) || vm_swap_full())) {
723                         delete_from_swap_cache(page);
724                         SetPageDirty(page);
725                 }
726                 unlock_page(page);
727                 page_cache_release(page);
728         }
729         return p != NULL;
730 }
731
732 #ifdef CONFIG_HIBERNATION
733 /*
734  * Find the swap type that corresponds to given device (if any).
735  *
736  * @offset - number of the PAGE_SIZE-sized block of the device, starting
737  * from 0, in which the swap header is expected to be located.
738  *
739  * This is needed for the suspend to disk (aka swsusp).
740  */
741 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
742 {
743         struct block_device *bdev = NULL;
744         int type;
745
746         if (device)
747                 bdev = bdget(device);
748
749         spin_lock(&swap_lock);
750         for (type = 0; type < nr_swapfiles; type++) {
751                 struct swap_info_struct *sis = swap_info[type];
752
753                 if (!(sis->flags & SWP_WRITEOK))
754                         continue;
755
756                 if (!bdev) {
757                         if (bdev_p)
758                                 *bdev_p = bdgrab(sis->bdev);
759
760                         spin_unlock(&swap_lock);
761                         return type;
762                 }
763                 if (bdev == sis->bdev) {
764                         struct swap_extent *se;
765
766                         se = list_entry(sis->extent_list.next,
767                                         struct swap_extent, list);
768                         if (se->start_block == offset) {
769                                 if (bdev_p)
770                                         *bdev_p = bdgrab(sis->bdev);
771
772                                 spin_unlock(&swap_lock);
773                                 bdput(bdev);
774                                 return type;
775                         }
776                 }
777         }
778         spin_unlock(&swap_lock);
779         if (bdev)
780                 bdput(bdev);
781
782         return -ENODEV;
783 }
784
785 /*
786  * Return either the total number of swap pages of given type, or the number
787  * of free pages of that type (depending on @free)
788  *
789  * This is needed for software suspend
790  */
791 unsigned int count_swap_pages(int type, int free)
792 {
793         unsigned int n = 0;
794
795         spin_lock(&swap_lock);
796         if ((unsigned int)type < nr_swapfiles) {
797                 struct swap_info_struct *sis = swap_info[type];
798
799                 if (sis->flags & SWP_WRITEOK) {
800                         n = sis->pages;
801                         if (free)
802                                 n -= sis->inuse_pages;
803                 }
804         }
805         spin_unlock(&swap_lock);
806         return n;
807 }
808 #endif
809
810 /*
811  * No need to decide whether this PTE shares the swap entry with others,
812  * just let do_wp_page work it out if a write is requested later - to
813  * force COW, vm_page_prot omits write permission from any private vma.
814  */
815 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
816                 unsigned long addr, swp_entry_t entry, struct page *page)
817 {
818         struct mem_cgroup *ptr = NULL;
819         spinlock_t *ptl;
820         pte_t *pte;
821         int ret = 1;
822
823         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
824                 ret = -ENOMEM;
825                 goto out_nolock;
826         }
827
828         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
829         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
830                 if (ret > 0)
831                         mem_cgroup_cancel_charge_swapin(ptr);
832                 ret = 0;
833                 goto out;
834         }
835
836         inc_mm_counter(vma->vm_mm, anon_rss);
837         get_page(page);
838         set_pte_at(vma->vm_mm, addr, pte,
839                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
840         page_add_anon_rmap(page, vma, addr);
841         mem_cgroup_commit_charge_swapin(page, ptr);
842         swap_free(entry);
843         /*
844          * Move the page to the active list so it is not
845          * immediately swapped out again after swapon.
846          */
847         activate_page(page);
848 out:
849         pte_unmap_unlock(pte, ptl);
850 out_nolock:
851         return ret;
852 }
853
854 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
855                                 unsigned long addr, unsigned long end,
856                                 swp_entry_t entry, struct page *page)
857 {
858         pte_t swp_pte = swp_entry_to_pte(entry);
859         pte_t *pte;
860         int ret = 0;
861
862         /*
863          * We don't actually need pte lock while scanning for swp_pte: since
864          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
865          * page table while we're scanning; though it could get zapped, and on
866          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
867          * of unmatched parts which look like swp_pte, so unuse_pte must
868          * recheck under pte lock.  Scanning without pte lock lets it be
869          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
870          */
871         pte = pte_offset_map(pmd, addr);
872         do {
873                 /*
874                  * swapoff spends a _lot_ of time in this loop!
875                  * Test inline before going to call unuse_pte.
876                  */
877                 if (unlikely(pte_same(*pte, swp_pte))) {
878                         pte_unmap(pte);
879                         ret = unuse_pte(vma, pmd, addr, entry, page);
880                         if (ret)
881                                 goto out;
882                         pte = pte_offset_map(pmd, addr);
883                 }
884         } while (pte++, addr += PAGE_SIZE, addr != end);
885         pte_unmap(pte - 1);
886 out:
887         return ret;
888 }
889
890 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
891                                 unsigned long addr, unsigned long end,
892                                 swp_entry_t entry, struct page *page)
893 {
894         pmd_t *pmd;
895         unsigned long next;
896         int ret;
897
898         pmd = pmd_offset(pud, addr);
899         do {
900                 next = pmd_addr_end(addr, end);
901                 if (pmd_none_or_clear_bad(pmd))
902                         continue;
903                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
904                 if (ret)
905                         return ret;
906         } while (pmd++, addr = next, addr != end);
907         return 0;
908 }
909
910 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
911                                 unsigned long addr, unsigned long end,
912                                 swp_entry_t entry, struct page *page)
913 {
914         pud_t *pud;
915         unsigned long next;
916         int ret;
917
918         pud = pud_offset(pgd, addr);
919         do {
920                 next = pud_addr_end(addr, end);
921                 if (pud_none_or_clear_bad(pud))
922                         continue;
923                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
924                 if (ret)
925                         return ret;
926         } while (pud++, addr = next, addr != end);
927         return 0;
928 }
929
930 static int unuse_vma(struct vm_area_struct *vma,
931                                 swp_entry_t entry, struct page *page)
932 {
933         pgd_t *pgd;
934         unsigned long addr, end, next;
935         int ret;
936
937         if (page->mapping) {
938                 addr = page_address_in_vma(page, vma);
939                 if (addr == -EFAULT)
940                         return 0;
941                 else
942                         end = addr + PAGE_SIZE;
943         } else {
944                 addr = vma->vm_start;
945                 end = vma->vm_end;
946         }
947
948         pgd = pgd_offset(vma->vm_mm, addr);
949         do {
950                 next = pgd_addr_end(addr, end);
951                 if (pgd_none_or_clear_bad(pgd))
952                         continue;
953                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
954                 if (ret)
955                         return ret;
956         } while (pgd++, addr = next, addr != end);
957         return 0;
958 }
959
960 static int unuse_mm(struct mm_struct *mm,
961                                 swp_entry_t entry, struct page *page)
962 {
963         struct vm_area_struct *vma;
964         int ret = 0;
965
966         if (!down_read_trylock(&mm->mmap_sem)) {
967                 /*
968                  * Activate page so shrink_inactive_list is unlikely to unmap
969                  * its ptes while lock is dropped, so swapoff can make progress.
970                  */
971                 activate_page(page);
972                 unlock_page(page);
973                 down_read(&mm->mmap_sem);
974                 lock_page(page);
975         }
976         for (vma = mm->mmap; vma; vma = vma->vm_next) {
977                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
978                         break;
979         }
980         up_read(&mm->mmap_sem);
981         return (ret < 0)? ret: 0;
982 }
983
984 /*
985  * Scan swap_map from current position to next entry still in use.
986  * Recycle to start on reaching the end, returning 0 when empty.
987  */
988 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
989                                         unsigned int prev)
990 {
991         unsigned int max = si->max;
992         unsigned int i = prev;
993         int count;
994
995         /*
996          * No need for swap_lock here: we're just looking
997          * for whether an entry is in use, not modifying it; false
998          * hits are okay, and sys_swapoff() has already prevented new
999          * allocations from this area (while holding swap_lock).
1000          */
1001         for (;;) {
1002                 if (++i >= max) {
1003                         if (!prev) {
1004                                 i = 0;
1005                                 break;
1006                         }
1007                         /*
1008                          * No entries in use at top of swap_map,
1009                          * loop back to start and recheck there.
1010                          */
1011                         max = prev + 1;
1012                         prev = 0;
1013                         i = 1;
1014                 }
1015                 count = si->swap_map[i];
1016                 if (count && swap_count(count) != SWAP_MAP_BAD)
1017                         break;
1018         }
1019         return i;
1020 }
1021
1022 /*
1023  * We completely avoid races by reading each swap page in advance,
1024  * and then search for the process using it.  All the necessary
1025  * page table adjustments can then be made atomically.
1026  */
1027 static int try_to_unuse(unsigned int type)
1028 {
1029         struct swap_info_struct *si = swap_info[type];
1030         struct mm_struct *start_mm;
1031         unsigned short *swap_map;
1032         unsigned short swcount;
1033         struct page *page;
1034         swp_entry_t entry;
1035         unsigned int i = 0;
1036         int retval = 0;
1037         int reset_overflow = 0;
1038         int shmem;
1039
1040         /*
1041          * When searching mms for an entry, a good strategy is to
1042          * start at the first mm we freed the previous entry from
1043          * (though actually we don't notice whether we or coincidence
1044          * freed the entry).  Initialize this start_mm with a hold.
1045          *
1046          * A simpler strategy would be to start at the last mm we
1047          * freed the previous entry from; but that would take less
1048          * advantage of mmlist ordering, which clusters forked mms
1049          * together, child after parent.  If we race with dup_mmap(), we
1050          * prefer to resolve parent before child, lest we miss entries
1051          * duplicated after we scanned child: using last mm would invert
1052          * that.  Though it's only a serious concern when an overflowed
1053          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
1054          */
1055         start_mm = &init_mm;
1056         atomic_inc(&init_mm.mm_users);
1057
1058         /*
1059          * Keep on scanning until all entries have gone.  Usually,
1060          * one pass through swap_map is enough, but not necessarily:
1061          * there are races when an instance of an entry might be missed.
1062          */
1063         while ((i = find_next_to_unuse(si, i)) != 0) {
1064                 if (signal_pending(current)) {
1065                         retval = -EINTR;
1066                         break;
1067                 }
1068
1069                 /*
1070                  * Get a page for the entry, using the existing swap
1071                  * cache page if there is one.  Otherwise, get a clean
1072                  * page and read the swap into it.
1073                  */
1074                 swap_map = &si->swap_map[i];
1075                 entry = swp_entry(type, i);
1076                 page = read_swap_cache_async(entry,
1077                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1078                 if (!page) {
1079                         /*
1080                          * Either swap_duplicate() failed because entry
1081                          * has been freed independently, and will not be
1082                          * reused since sys_swapoff() already disabled
1083                          * allocation from here, or alloc_page() failed.
1084                          */
1085                         if (!*swap_map)
1086                                 continue;
1087                         retval = -ENOMEM;
1088                         break;
1089                 }
1090
1091                 /*
1092                  * Don't hold on to start_mm if it looks like exiting.
1093                  */
1094                 if (atomic_read(&start_mm->mm_users) == 1) {
1095                         mmput(start_mm);
1096                         start_mm = &init_mm;
1097                         atomic_inc(&init_mm.mm_users);
1098                 }
1099
1100                 /*
1101                  * Wait for and lock page.  When do_swap_page races with
1102                  * try_to_unuse, do_swap_page can handle the fault much
1103                  * faster than try_to_unuse can locate the entry.  This
1104                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1105                  * defer to do_swap_page in such a case - in some tests,
1106                  * do_swap_page and try_to_unuse repeatedly compete.
1107                  */
1108                 wait_on_page_locked(page);
1109                 wait_on_page_writeback(page);
1110                 lock_page(page);
1111                 wait_on_page_writeback(page);
1112
1113                 /*
1114                  * Remove all references to entry.
1115                  * Whenever we reach init_mm, there's no address space
1116                  * to search, but use it as a reminder to search shmem.
1117                  */
1118                 shmem = 0;
1119                 swcount = *swap_map;
1120                 if (swap_count(swcount)) {
1121                         if (start_mm == &init_mm)
1122                                 shmem = shmem_unuse(entry, page);
1123                         else
1124                                 retval = unuse_mm(start_mm, entry, page);
1125                 }
1126                 if (swap_count(*swap_map)) {
1127                         int set_start_mm = (*swap_map >= swcount);
1128                         struct list_head *p = &start_mm->mmlist;
1129                         struct mm_struct *new_start_mm = start_mm;
1130                         struct mm_struct *prev_mm = start_mm;
1131                         struct mm_struct *mm;
1132
1133                         atomic_inc(&new_start_mm->mm_users);
1134                         atomic_inc(&prev_mm->mm_users);
1135                         spin_lock(&mmlist_lock);
1136                         while (swap_count(*swap_map) && !retval && !shmem &&
1137                                         (p = p->next) != &start_mm->mmlist) {
1138                                 mm = list_entry(p, struct mm_struct, mmlist);
1139                                 if (!atomic_inc_not_zero(&mm->mm_users))
1140                                         continue;
1141                                 spin_unlock(&mmlist_lock);
1142                                 mmput(prev_mm);
1143                                 prev_mm = mm;
1144
1145                                 cond_resched();
1146
1147                                 swcount = *swap_map;
1148                                 if (!swap_count(swcount)) /* any usage ? */
1149                                         ;
1150                                 else if (mm == &init_mm) {
1151                                         set_start_mm = 1;
1152                                         shmem = shmem_unuse(entry, page);
1153                                 } else
1154                                         retval = unuse_mm(mm, entry, page);
1155
1156                                 if (set_start_mm && *swap_map < swcount) {
1157                                         mmput(new_start_mm);
1158                                         atomic_inc(&mm->mm_users);
1159                                         new_start_mm = mm;
1160                                         set_start_mm = 0;
1161                                 }
1162                                 spin_lock(&mmlist_lock);
1163                         }
1164                         spin_unlock(&mmlist_lock);
1165                         mmput(prev_mm);
1166                         mmput(start_mm);
1167                         start_mm = new_start_mm;
1168                 }
1169                 if (shmem) {
1170                         /* page has already been unlocked and released */
1171                         if (shmem > 0)
1172                                 continue;
1173                         retval = shmem;
1174                         break;
1175                 }
1176                 if (retval) {
1177                         unlock_page(page);
1178                         page_cache_release(page);
1179                         break;
1180                 }
1181
1182                 /*
1183                  * How could swap count reach 0x7ffe ?
1184                  * There's no way to repeat a swap page within an mm
1185                  * (except in shmem, where it's the shared object which takes
1186                  * the reference count)?
1187                  * We believe SWAP_MAP_MAX cannot occur.(if occur, unsigned
1188                  * short is too small....)
1189                  * If that's wrong, then we should worry more about
1190                  * exit_mmap() and do_munmap() cases described above:
1191                  * we might be resetting SWAP_MAP_MAX too early here.
1192                  * We know "Undead"s can happen, they're okay, so don't
1193                  * report them; but do report if we reset SWAP_MAP_MAX.
1194                  */
1195                 /* We might release the lock_page() in unuse_mm(). */
1196                 if (!PageSwapCache(page) || page_private(page) != entry.val)
1197                         goto retry;
1198
1199                 if (swap_count(*swap_map) == SWAP_MAP_MAX) {
1200                         spin_lock(&swap_lock);
1201                         *swap_map = encode_swapmap(0, true);
1202                         spin_unlock(&swap_lock);
1203                         reset_overflow = 1;
1204                 }
1205
1206                 /*
1207                  * If a reference remains (rare), we would like to leave
1208                  * the page in the swap cache; but try_to_unmap could
1209                  * then re-duplicate the entry once we drop page lock,
1210                  * so we might loop indefinitely; also, that page could
1211                  * not be swapped out to other storage meanwhile.  So:
1212                  * delete from cache even if there's another reference,
1213                  * after ensuring that the data has been saved to disk -
1214                  * since if the reference remains (rarer), it will be
1215                  * read from disk into another page.  Splitting into two
1216                  * pages would be incorrect if swap supported "shared
1217                  * private" pages, but they are handled by tmpfs files.
1218                  */
1219                 if (swap_count(*swap_map) &&
1220                      PageDirty(page) && PageSwapCache(page)) {
1221                         struct writeback_control wbc = {
1222                                 .sync_mode = WB_SYNC_NONE,
1223                         };
1224
1225                         swap_writepage(page, &wbc);
1226                         lock_page(page);
1227                         wait_on_page_writeback(page);
1228                 }
1229
1230                 /*
1231                  * It is conceivable that a racing task removed this page from
1232                  * swap cache just before we acquired the page lock at the top,
1233                  * or while we dropped it in unuse_mm().  The page might even
1234                  * be back in swap cache on another swap area: that we must not
1235                  * delete, since it may not have been written out to swap yet.
1236                  */
1237                 if (PageSwapCache(page) &&
1238                     likely(page_private(page) == entry.val))
1239                         delete_from_swap_cache(page);
1240
1241                 /*
1242                  * So we could skip searching mms once swap count went
1243                  * to 1, we did not mark any present ptes as dirty: must
1244                  * mark page dirty so shrink_page_list will preserve it.
1245                  */
1246                 SetPageDirty(page);
1247 retry:
1248                 unlock_page(page);
1249                 page_cache_release(page);
1250
1251                 /*
1252                  * Make sure that we aren't completely killing
1253                  * interactive performance.
1254                  */
1255                 cond_resched();
1256         }
1257
1258         mmput(start_mm);
1259         if (reset_overflow) {
1260                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1261                 swap_overflow = 0;
1262         }
1263         return retval;
1264 }
1265
1266 /*
1267  * After a successful try_to_unuse, if no swap is now in use, we know
1268  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1269  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1270  * added to the mmlist just after page_duplicate - before would be racy.
1271  */
1272 static void drain_mmlist(void)
1273 {
1274         struct list_head *p, *next;
1275         unsigned int type;
1276
1277         for (type = 0; type < nr_swapfiles; type++)
1278                 if (swap_info[type]->inuse_pages)
1279                         return;
1280         spin_lock(&mmlist_lock);
1281         list_for_each_safe(p, next, &init_mm.mmlist)
1282                 list_del_init(p);
1283         spin_unlock(&mmlist_lock);
1284 }
1285
1286 /*
1287  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1288  * corresponds to page offset `offset'.  Note that the type of this function
1289  * is sector_t, but it returns page offset into the bdev, not sector offset.
1290  */
1291 sector_t map_swap_page(swp_entry_t entry, struct block_device **bdev)
1292 {
1293         struct swap_info_struct *sis;
1294         struct swap_extent *start_se;
1295         struct swap_extent *se;
1296         pgoff_t offset;
1297
1298         sis = swap_info[swp_type(entry)];
1299         *bdev = sis->bdev;
1300
1301         offset = swp_offset(entry);
1302         start_se = sis->curr_swap_extent;
1303         se = start_se;
1304
1305         for ( ; ; ) {
1306                 struct list_head *lh;
1307
1308                 if (se->start_page <= offset &&
1309                                 offset < (se->start_page + se->nr_pages)) {
1310                         return se->start_block + (offset - se->start_page);
1311                 }
1312                 lh = se->list.next;
1313                 if (lh == &sis->extent_list)
1314                         lh = lh->next;
1315                 se = list_entry(lh, struct swap_extent, list);
1316                 sis->curr_swap_extent = se;
1317                 BUG_ON(se == start_se);         /* It *must* be present */
1318         }
1319 }
1320
1321 #ifdef CONFIG_HIBERNATION
1322 /*
1323  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1324  * corresponding to given index in swap_info (swap type).
1325  */
1326 sector_t swapdev_block(int type, pgoff_t offset)
1327 {
1328         struct block_device *bdev;
1329
1330         if ((unsigned int)type >= nr_swapfiles)
1331                 return 0;
1332         if (!(swap_info[type]->flags & SWP_WRITEOK))
1333                 return 0;
1334         return map_swap_page(swp_entry(type, offset), &bdev);
1335 }
1336 #endif /* CONFIG_HIBERNATION */
1337
1338 /*
1339  * Free all of a swapdev's extent information
1340  */
1341 static void destroy_swap_extents(struct swap_info_struct *sis)
1342 {
1343         while (!list_empty(&sis->extent_list)) {
1344                 struct swap_extent *se;
1345
1346                 se = list_entry(sis->extent_list.next,
1347                                 struct swap_extent, list);
1348                 list_del(&se->list);
1349                 kfree(se);
1350         }
1351 }
1352
1353 /*
1354  * Add a block range (and the corresponding page range) into this swapdev's
1355  * extent list.  The extent list is kept sorted in page order.
1356  *
1357  * This function rather assumes that it is called in ascending page order.
1358  */
1359 static int
1360 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1361                 unsigned long nr_pages, sector_t start_block)
1362 {
1363         struct swap_extent *se;
1364         struct swap_extent *new_se;
1365         struct list_head *lh;
1366
1367         lh = sis->extent_list.prev;     /* The highest page extent */
1368         if (lh != &sis->extent_list) {
1369                 se = list_entry(lh, struct swap_extent, list);
1370                 BUG_ON(se->start_page + se->nr_pages != start_page);
1371                 if (se->start_block + se->nr_pages == start_block) {
1372                         /* Merge it */
1373                         se->nr_pages += nr_pages;
1374                         return 0;
1375                 }
1376         }
1377
1378         /*
1379          * No merge.  Insert a new extent, preserving ordering.
1380          */
1381         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1382         if (new_se == NULL)
1383                 return -ENOMEM;
1384         new_se->start_page = start_page;
1385         new_se->nr_pages = nr_pages;
1386         new_se->start_block = start_block;
1387
1388         list_add_tail(&new_se->list, &sis->extent_list);
1389         return 1;
1390 }
1391
1392 /*
1393  * A `swap extent' is a simple thing which maps a contiguous range of pages
1394  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1395  * is built at swapon time and is then used at swap_writepage/swap_readpage
1396  * time for locating where on disk a page belongs.
1397  *
1398  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1399  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1400  * swap files identically.
1401  *
1402  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1403  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1404  * swapfiles are handled *identically* after swapon time.
1405  *
1406  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1407  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1408  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1409  * requirements, they are simply tossed out - we will never use those blocks
1410  * for swapping.
1411  *
1412  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1413  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1414  * which will scribble on the fs.
1415  *
1416  * The amount of disk space which a single swap extent represents varies.
1417  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1418  * extents in the list.  To avoid much list walking, we cache the previous
1419  * search location in `curr_swap_extent', and start new searches from there.
1420  * This is extremely effective.  The average number of iterations in
1421  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1422  */
1423 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1424 {
1425         struct inode *inode;
1426         unsigned blocks_per_page;
1427         unsigned long page_no;
1428         unsigned blkbits;
1429         sector_t probe_block;
1430         sector_t last_block;
1431         sector_t lowest_block = -1;
1432         sector_t highest_block = 0;
1433         int nr_extents = 0;
1434         int ret;
1435
1436         inode = sis->swap_file->f_mapping->host;
1437         if (S_ISBLK(inode->i_mode)) {
1438                 ret = add_swap_extent(sis, 0, sis->max, 0);
1439                 *span = sis->pages;
1440                 goto done;
1441         }
1442
1443         blkbits = inode->i_blkbits;
1444         blocks_per_page = PAGE_SIZE >> blkbits;
1445
1446         /*
1447          * Map all the blocks into the extent list.  This code doesn't try
1448          * to be very smart.
1449          */
1450         probe_block = 0;
1451         page_no = 0;
1452         last_block = i_size_read(inode) >> blkbits;
1453         while ((probe_block + blocks_per_page) <= last_block &&
1454                         page_no < sis->max) {
1455                 unsigned block_in_page;
1456                 sector_t first_block;
1457
1458                 first_block = bmap(inode, probe_block);
1459                 if (first_block == 0)
1460                         goto bad_bmap;
1461
1462                 /*
1463                  * It must be PAGE_SIZE aligned on-disk
1464                  */
1465                 if (first_block & (blocks_per_page - 1)) {
1466                         probe_block++;
1467                         goto reprobe;
1468                 }
1469
1470                 for (block_in_page = 1; block_in_page < blocks_per_page;
1471                                         block_in_page++) {
1472                         sector_t block;
1473
1474                         block = bmap(inode, probe_block + block_in_page);
1475                         if (block == 0)
1476                                 goto bad_bmap;
1477                         if (block != first_block + block_in_page) {
1478                                 /* Discontiguity */
1479                                 probe_block++;
1480                                 goto reprobe;
1481                         }
1482                 }
1483
1484                 first_block >>= (PAGE_SHIFT - blkbits);
1485                 if (page_no) {  /* exclude the header page */
1486                         if (first_block < lowest_block)
1487                                 lowest_block = first_block;
1488                         if (first_block > highest_block)
1489                                 highest_block = first_block;
1490                 }
1491
1492                 /*
1493                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1494                  */
1495                 ret = add_swap_extent(sis, page_no, 1, first_block);
1496                 if (ret < 0)
1497                         goto out;
1498                 nr_extents += ret;
1499                 page_no++;
1500                 probe_block += blocks_per_page;
1501 reprobe:
1502                 continue;
1503         }
1504         ret = nr_extents;
1505         *span = 1 + highest_block - lowest_block;
1506         if (page_no == 0)
1507                 page_no = 1;    /* force Empty message */
1508         sis->max = page_no;
1509         sis->pages = page_no - 1;
1510         sis->highest_bit = page_no - 1;
1511 done:
1512         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1513                                         struct swap_extent, list);
1514         goto out;
1515 bad_bmap:
1516         printk(KERN_ERR "swapon: swapfile has holes\n");
1517         ret = -EINVAL;
1518 out:
1519         return ret;
1520 }
1521
1522 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1523 {
1524         struct swap_info_struct * p = NULL;
1525         unsigned short *swap_map;
1526         struct file *swap_file, *victim;
1527         struct address_space *mapping;
1528         struct inode *inode;
1529         char * pathname;
1530         int i, type, prev;
1531         int err;
1532
1533         if (!capable(CAP_SYS_ADMIN))
1534                 return -EPERM;
1535
1536         pathname = getname(specialfile);
1537         err = PTR_ERR(pathname);
1538         if (IS_ERR(pathname))
1539                 goto out;
1540
1541         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1542         putname(pathname);
1543         err = PTR_ERR(victim);
1544         if (IS_ERR(victim))
1545                 goto out;
1546
1547         mapping = victim->f_mapping;
1548         prev = -1;
1549         spin_lock(&swap_lock);
1550         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1551                 p = swap_info[type];
1552                 if (p->flags & SWP_WRITEOK) {
1553                         if (p->swap_file->f_mapping == mapping)
1554                                 break;
1555                 }
1556                 prev = type;
1557         }
1558         if (type < 0) {
1559                 err = -EINVAL;
1560                 spin_unlock(&swap_lock);
1561                 goto out_dput;
1562         }
1563         if (!security_vm_enough_memory(p->pages))
1564                 vm_unacct_memory(p->pages);
1565         else {
1566                 err = -ENOMEM;
1567                 spin_unlock(&swap_lock);
1568                 goto out_dput;
1569         }
1570         if (prev < 0)
1571                 swap_list.head = p->next;
1572         else
1573                 swap_info[prev]->next = p->next;
1574         if (type == swap_list.next) {
1575                 /* just pick something that's safe... */
1576                 swap_list.next = swap_list.head;
1577         }
1578         if (p->prio < 0) {
1579                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1580                         swap_info[i]->prio = p->prio--;
1581                 least_priority++;
1582         }
1583         nr_swap_pages -= p->pages;
1584         total_swap_pages -= p->pages;
1585         p->flags &= ~SWP_WRITEOK;
1586         spin_unlock(&swap_lock);
1587
1588         current->flags |= PF_OOM_ORIGIN;
1589         err = try_to_unuse(type);
1590         current->flags &= ~PF_OOM_ORIGIN;
1591
1592         if (err) {
1593                 /* re-insert swap space back into swap_list */
1594                 spin_lock(&swap_lock);
1595                 if (p->prio < 0)
1596                         p->prio = --least_priority;
1597                 prev = -1;
1598                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1599                         if (p->prio >= swap_info[i]->prio)
1600                                 break;
1601                         prev = i;
1602                 }
1603                 p->next = i;
1604                 if (prev < 0)
1605                         swap_list.head = swap_list.next = type;
1606                 else
1607                         swap_info[prev]->next = type;
1608                 nr_swap_pages += p->pages;
1609                 total_swap_pages += p->pages;
1610                 p->flags |= SWP_WRITEOK;
1611                 spin_unlock(&swap_lock);
1612                 goto out_dput;
1613         }
1614
1615         /* wait for any unplug function to finish */
1616         down_write(&swap_unplug_sem);
1617         up_write(&swap_unplug_sem);
1618
1619         destroy_swap_extents(p);
1620         mutex_lock(&swapon_mutex);
1621         spin_lock(&swap_lock);
1622         drain_mmlist();
1623
1624         /* wait for anyone still in scan_swap_map */
1625         p->highest_bit = 0;             /* cuts scans short */
1626         while (p->flags >= SWP_SCANNING) {
1627                 spin_unlock(&swap_lock);
1628                 schedule_timeout_uninterruptible(1);
1629                 spin_lock(&swap_lock);
1630         }
1631
1632         swap_file = p->swap_file;
1633         p->swap_file = NULL;
1634         p->max = 0;
1635         swap_map = p->swap_map;
1636         p->swap_map = NULL;
1637         p->flags = 0;
1638         spin_unlock(&swap_lock);
1639         mutex_unlock(&swapon_mutex);
1640         vfree(swap_map);
1641         /* Destroy swap account informatin */
1642         swap_cgroup_swapoff(type);
1643
1644         inode = mapping->host;
1645         if (S_ISBLK(inode->i_mode)) {
1646                 struct block_device *bdev = I_BDEV(inode);
1647                 set_blocksize(bdev, p->old_block_size);
1648                 bd_release(bdev);
1649         } else {
1650                 mutex_lock(&inode->i_mutex);
1651                 inode->i_flags &= ~S_SWAPFILE;
1652                 mutex_unlock(&inode->i_mutex);
1653         }
1654         filp_close(swap_file, NULL);
1655         err = 0;
1656
1657 out_dput:
1658         filp_close(victim, NULL);
1659 out:
1660         return err;
1661 }
1662
1663 #ifdef CONFIG_PROC_FS
1664 /* iterator */
1665 static void *swap_start(struct seq_file *swap, loff_t *pos)
1666 {
1667         struct swap_info_struct *si;
1668         int type;
1669         loff_t l = *pos;
1670
1671         mutex_lock(&swapon_mutex);
1672
1673         if (!l)
1674                 return SEQ_START_TOKEN;
1675
1676         for (type = 0; type < nr_swapfiles; type++) {
1677                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1678                 si = swap_info[type];
1679                 if (!(si->flags & SWP_USED) || !si->swap_map)
1680                         continue;
1681                 if (!--l)
1682                         return si;
1683         }
1684
1685         return NULL;
1686 }
1687
1688 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1689 {
1690         struct swap_info_struct *si = v;
1691         int type;
1692
1693         if (v == SEQ_START_TOKEN)
1694                 type = 0;
1695         else
1696                 type = si->type + 1;
1697
1698         for (; type < nr_swapfiles; type++) {
1699                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1700                 si = swap_info[type];
1701                 if (!(si->flags & SWP_USED) || !si->swap_map)
1702                         continue;
1703                 ++*pos;
1704                 return si;
1705         }
1706
1707         return NULL;
1708 }
1709
1710 static void swap_stop(struct seq_file *swap, void *v)
1711 {
1712         mutex_unlock(&swapon_mutex);
1713 }
1714
1715 static int swap_show(struct seq_file *swap, void *v)
1716 {
1717         struct swap_info_struct *si = v;
1718         struct file *file;
1719         int len;
1720
1721         if (si == SEQ_START_TOKEN) {
1722                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1723                 return 0;
1724         }
1725
1726         file = si->swap_file;
1727         len = seq_path(swap, &file->f_path, " \t\n\\");
1728         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1729                         len < 40 ? 40 - len : 1, " ",
1730                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1731                                 "partition" : "file\t",
1732                         si->pages << (PAGE_SHIFT - 10),
1733                         si->inuse_pages << (PAGE_SHIFT - 10),
1734                         si->prio);
1735         return 0;
1736 }
1737
1738 static const struct seq_operations swaps_op = {
1739         .start =        swap_start,
1740         .next =         swap_next,
1741         .stop =         swap_stop,
1742         .show =         swap_show
1743 };
1744
1745 static int swaps_open(struct inode *inode, struct file *file)
1746 {
1747         return seq_open(file, &swaps_op);
1748 }
1749
1750 static const struct file_operations proc_swaps_operations = {
1751         .open           = swaps_open,
1752         .read           = seq_read,
1753         .llseek         = seq_lseek,
1754         .release        = seq_release,
1755 };
1756
1757 static int __init procswaps_init(void)
1758 {
1759         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1760         return 0;
1761 }
1762 __initcall(procswaps_init);
1763 #endif /* CONFIG_PROC_FS */
1764
1765 #ifdef MAX_SWAPFILES_CHECK
1766 static int __init max_swapfiles_check(void)
1767 {
1768         MAX_SWAPFILES_CHECK();
1769         return 0;
1770 }
1771 late_initcall(max_swapfiles_check);
1772 #endif
1773
1774 /*
1775  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1776  *
1777  * The swapon system call
1778  */
1779 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1780 {
1781         struct swap_info_struct * p;
1782         char *name = NULL;
1783         struct block_device *bdev = NULL;
1784         struct file *swap_file = NULL;
1785         struct address_space *mapping;
1786         unsigned int type;
1787         int i, prev;
1788         int error;
1789         union swap_header *swap_header = NULL;
1790         unsigned int nr_good_pages = 0;
1791         int nr_extents = 0;
1792         sector_t span;
1793         unsigned long maxpages = 1;
1794         unsigned long swapfilepages;
1795         unsigned short *swap_map = NULL;
1796         struct page *page = NULL;
1797         struct inode *inode = NULL;
1798         int did_down = 0;
1799
1800         if (!capable(CAP_SYS_ADMIN))
1801                 return -EPERM;
1802
1803         p = kzalloc(sizeof(*p), GFP_KERNEL);
1804         if (!p)
1805                 return -ENOMEM;
1806
1807         spin_lock(&swap_lock);
1808         for (type = 0; type < nr_swapfiles; type++) {
1809                 if (!(swap_info[type]->flags & SWP_USED))
1810                         break;
1811         }
1812         error = -EPERM;
1813         if (type >= MAX_SWAPFILES) {
1814                 spin_unlock(&swap_lock);
1815                 kfree(p);
1816                 goto out;
1817         }
1818         INIT_LIST_HEAD(&p->extent_list);
1819         if (type >= nr_swapfiles) {
1820                 p->type = type;
1821                 swap_info[type] = p;
1822                 /*
1823                  * Write swap_info[type] before nr_swapfiles, in case a
1824                  * racing procfs swap_start() or swap_next() is reading them.
1825                  * (We never shrink nr_swapfiles, we never free this entry.)
1826                  */
1827                 smp_wmb();
1828                 nr_swapfiles++;
1829         } else {
1830                 kfree(p);
1831                 p = swap_info[type];
1832                 /*
1833                  * Do not memset this entry: a racing procfs swap_next()
1834                  * would be relying on p->type to remain valid.
1835                  */
1836         }
1837         p->flags = SWP_USED;
1838         p->next = -1;
1839         spin_unlock(&swap_lock);
1840
1841         name = getname(specialfile);
1842         error = PTR_ERR(name);
1843         if (IS_ERR(name)) {
1844                 name = NULL;
1845                 goto bad_swap_2;
1846         }
1847         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1848         error = PTR_ERR(swap_file);
1849         if (IS_ERR(swap_file)) {
1850                 swap_file = NULL;
1851                 goto bad_swap_2;
1852         }
1853
1854         p->swap_file = swap_file;
1855         mapping = swap_file->f_mapping;
1856         inode = mapping->host;
1857
1858         error = -EBUSY;
1859         for (i = 0; i < nr_swapfiles; i++) {
1860                 struct swap_info_struct *q = swap_info[i];
1861
1862                 if (i == type || !q->swap_file)
1863                         continue;
1864                 if (mapping == q->swap_file->f_mapping)
1865                         goto bad_swap;
1866         }
1867
1868         error = -EINVAL;
1869         if (S_ISBLK(inode->i_mode)) {
1870                 bdev = I_BDEV(inode);
1871                 error = bd_claim(bdev, sys_swapon);
1872                 if (error < 0) {
1873                         bdev = NULL;
1874                         error = -EINVAL;
1875                         goto bad_swap;
1876                 }
1877                 p->old_block_size = block_size(bdev);
1878                 error = set_blocksize(bdev, PAGE_SIZE);
1879                 if (error < 0)
1880                         goto bad_swap;
1881                 p->bdev = bdev;
1882         } else if (S_ISREG(inode->i_mode)) {
1883                 p->bdev = inode->i_sb->s_bdev;
1884                 mutex_lock(&inode->i_mutex);
1885                 did_down = 1;
1886                 if (IS_SWAPFILE(inode)) {
1887                         error = -EBUSY;
1888                         goto bad_swap;
1889                 }
1890         } else {
1891                 goto bad_swap;
1892         }
1893
1894         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1895
1896         /*
1897          * Read the swap header.
1898          */
1899         if (!mapping->a_ops->readpage) {
1900                 error = -EINVAL;
1901                 goto bad_swap;
1902         }
1903         page = read_mapping_page(mapping, 0, swap_file);
1904         if (IS_ERR(page)) {
1905                 error = PTR_ERR(page);
1906                 goto bad_swap;
1907         }
1908         swap_header = kmap(page);
1909
1910         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1911                 printk(KERN_ERR "Unable to find swap-space signature\n");
1912                 error = -EINVAL;
1913                 goto bad_swap;
1914         }
1915
1916         /* swap partition endianess hack... */
1917         if (swab32(swap_header->info.version) == 1) {
1918                 swab32s(&swap_header->info.version);
1919                 swab32s(&swap_header->info.last_page);
1920                 swab32s(&swap_header->info.nr_badpages);
1921                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1922                         swab32s(&swap_header->info.badpages[i]);
1923         }
1924         /* Check the swap header's sub-version */
1925         if (swap_header->info.version != 1) {
1926                 printk(KERN_WARNING
1927                        "Unable to handle swap header version %d\n",
1928                        swap_header->info.version);
1929                 error = -EINVAL;
1930                 goto bad_swap;
1931         }
1932
1933         p->lowest_bit  = 1;
1934         p->cluster_next = 1;
1935         p->cluster_nr = 0;
1936
1937         /*
1938          * Find out how many pages are allowed for a single swap
1939          * device. There are two limiting factors: 1) the number of
1940          * bits for the swap offset in the swp_entry_t type and
1941          * 2) the number of bits in the a swap pte as defined by
1942          * the different architectures. In order to find the
1943          * largest possible bit mask a swap entry with swap type 0
1944          * and swap offset ~0UL is created, encoded to a swap pte,
1945          * decoded to a swp_entry_t again and finally the swap
1946          * offset is extracted. This will mask all the bits from
1947          * the initial ~0UL mask that can't be encoded in either
1948          * the swp_entry_t or the architecture definition of a
1949          * swap pte.
1950          */
1951         maxpages = swp_offset(pte_to_swp_entry(
1952                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1953         if (maxpages > swap_header->info.last_page)
1954                 maxpages = swap_header->info.last_page;
1955         p->highest_bit = maxpages - 1;
1956
1957         error = -EINVAL;
1958         if (!maxpages)
1959                 goto bad_swap;
1960         if (swapfilepages && maxpages > swapfilepages) {
1961                 printk(KERN_WARNING
1962                        "Swap area shorter than signature indicates\n");
1963                 goto bad_swap;
1964         }
1965         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1966                 goto bad_swap;
1967         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1968                 goto bad_swap;
1969
1970         /* OK, set up the swap map and apply the bad block list */
1971         swap_map = vmalloc(maxpages * sizeof(short));
1972         if (!swap_map) {
1973                 error = -ENOMEM;
1974                 goto bad_swap;
1975         }
1976
1977         memset(swap_map, 0, maxpages * sizeof(short));
1978         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1979                 int page_nr = swap_header->info.badpages[i];
1980                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1981                         error = -EINVAL;
1982                         goto bad_swap;
1983                 }
1984                 swap_map[page_nr] = SWAP_MAP_BAD;
1985         }
1986
1987         error = swap_cgroup_swapon(type, maxpages);
1988         if (error)
1989                 goto bad_swap;
1990
1991         nr_good_pages = swap_header->info.last_page -
1992                         swap_header->info.nr_badpages -
1993                         1 /* header page */;
1994
1995         if (nr_good_pages) {
1996                 swap_map[0] = SWAP_MAP_BAD;
1997                 p->max = maxpages;
1998                 p->pages = nr_good_pages;
1999                 nr_extents = setup_swap_extents(p, &span);
2000                 if (nr_extents < 0) {
2001                         error = nr_extents;
2002                         goto bad_swap;
2003                 }
2004                 nr_good_pages = p->pages;
2005         }
2006         if (!nr_good_pages) {
2007                 printk(KERN_WARNING "Empty swap-file\n");
2008                 error = -EINVAL;
2009                 goto bad_swap;
2010         }
2011
2012         if (p->bdev) {
2013                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2014                         p->flags |= SWP_SOLIDSTATE;
2015                         p->cluster_next = 1 + (random32() % p->highest_bit);
2016                 }
2017                 if (discard_swap(p) == 0)
2018                         p->flags |= SWP_DISCARDABLE;
2019         }
2020
2021         mutex_lock(&swapon_mutex);
2022         spin_lock(&swap_lock);
2023         if (swap_flags & SWAP_FLAG_PREFER)
2024                 p->prio =
2025                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2026         else
2027                 p->prio = --least_priority;
2028         p->swap_map = swap_map;
2029         p->flags |= SWP_WRITEOK;
2030         nr_swap_pages += nr_good_pages;
2031         total_swap_pages += nr_good_pages;
2032
2033         printk(KERN_INFO "Adding %uk swap on %s.  "
2034                         "Priority:%d extents:%d across:%lluk %s%s\n",
2035                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2036                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2037                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2038                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2039
2040         /* insert swap space into swap_list: */
2041         prev = -1;
2042         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2043                 if (p->prio >= swap_info[i]->prio)
2044                         break;
2045                 prev = i;
2046         }
2047         p->next = i;
2048         if (prev < 0)
2049                 swap_list.head = swap_list.next = type;
2050         else
2051                 swap_info[prev]->next = type;
2052         spin_unlock(&swap_lock);
2053         mutex_unlock(&swapon_mutex);
2054         error = 0;
2055         goto out;
2056 bad_swap:
2057         if (bdev) {
2058                 set_blocksize(bdev, p->old_block_size);
2059                 bd_release(bdev);
2060         }
2061         destroy_swap_extents(p);
2062         swap_cgroup_swapoff(type);
2063 bad_swap_2:
2064         spin_lock(&swap_lock);
2065         p->swap_file = NULL;
2066         p->flags = 0;
2067         spin_unlock(&swap_lock);
2068         vfree(swap_map);
2069         if (swap_file)
2070                 filp_close(swap_file, NULL);
2071 out:
2072         if (page && !IS_ERR(page)) {
2073                 kunmap(page);
2074                 page_cache_release(page);
2075         }
2076         if (name)
2077                 putname(name);
2078         if (did_down) {
2079                 if (!error)
2080                         inode->i_flags |= S_SWAPFILE;
2081                 mutex_unlock(&inode->i_mutex);
2082         }
2083         return error;
2084 }
2085
2086 void si_swapinfo(struct sysinfo *val)
2087 {
2088         unsigned int type;
2089         unsigned long nr_to_be_unused = 0;
2090
2091         spin_lock(&swap_lock);
2092         for (type = 0; type < nr_swapfiles; type++) {
2093                 struct swap_info_struct *si = swap_info[type];
2094
2095                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2096                         nr_to_be_unused += si->inuse_pages;
2097         }
2098         val->freeswap = nr_swap_pages + nr_to_be_unused;
2099         val->totalswap = total_swap_pages + nr_to_be_unused;
2100         spin_unlock(&swap_lock);
2101 }
2102
2103 /*
2104  * Verify that a swap entry is valid and increment its swap map count.
2105  *
2106  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
2107  * "permanent", but will be reclaimed by the next swapoff.
2108  * Returns error code in following case.
2109  * - success -> 0
2110  * - swp_entry is invalid -> EINVAL
2111  * - swp_entry is migration entry -> EINVAL
2112  * - swap-cache reference is requested but there is already one. -> EEXIST
2113  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2114  */
2115 static int __swap_duplicate(swp_entry_t entry, bool cache)
2116 {
2117         struct swap_info_struct * p;
2118         unsigned long offset, type;
2119         int result = -EINVAL;
2120         int count;
2121         bool has_cache;
2122
2123         if (non_swap_entry(entry))
2124                 return -EINVAL;
2125
2126         type = swp_type(entry);
2127         if (type >= nr_swapfiles)
2128                 goto bad_file;
2129         p = swap_info[type];
2130         offset = swp_offset(entry);
2131
2132         spin_lock(&swap_lock);
2133
2134         if (unlikely(offset >= p->max))
2135                 goto unlock_out;
2136
2137         count = swap_count(p->swap_map[offset]);
2138         has_cache = swap_has_cache(p->swap_map[offset]);
2139
2140         if (cache == SWAP_CACHE) { /* called for swapcache/swapin-readahead */
2141
2142                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2143                 if (!has_cache && count) {
2144                         p->swap_map[offset] = encode_swapmap(count, true);
2145                         result = 0;
2146                 } else if (has_cache) /* someone added cache */
2147                         result = -EEXIST;
2148                 else if (!count) /* no users */
2149                         result = -ENOENT;
2150
2151         } else if (count || has_cache) {
2152                 if (count < SWAP_MAP_MAX - 1) {
2153                         p->swap_map[offset] = encode_swapmap(count + 1,
2154                                                              has_cache);
2155                         result = 0;
2156                 } else if (count <= SWAP_MAP_MAX) {
2157                         if (swap_overflow++ < 5)
2158                                 printk(KERN_WARNING
2159                                        "swap_dup: swap entry overflow\n");
2160                         p->swap_map[offset] = encode_swapmap(SWAP_MAP_MAX,
2161                                                               has_cache);
2162                         result = 0;
2163                 }
2164         } else
2165                 result = -ENOENT; /* unused swap entry */
2166 unlock_out:
2167         spin_unlock(&swap_lock);
2168 out:
2169         return result;
2170
2171 bad_file:
2172         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2173         goto out;
2174 }
2175 /*
2176  * increase reference count of swap entry by 1.
2177  */
2178 void swap_duplicate(swp_entry_t entry)
2179 {
2180         __swap_duplicate(entry, SWAP_MAP);
2181 }
2182
2183 /*
2184  * @entry: swap entry for which we allocate swap cache.
2185  *
2186  * Called when allocating swap cache for exising swap entry,
2187  * This can return error codes. Returns 0 at success.
2188  * -EBUSY means there is a swap cache.
2189  * Note: return code is different from swap_duplicate().
2190  */
2191 int swapcache_prepare(swp_entry_t entry)
2192 {
2193         return __swap_duplicate(entry, SWAP_CACHE);
2194 }
2195
2196 /*
2197  * swap_lock prevents swap_map being freed. Don't grab an extra
2198  * reference on the swaphandle, it doesn't matter if it becomes unused.
2199  */
2200 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2201 {
2202         struct swap_info_struct *si;
2203         int our_page_cluster = page_cluster;
2204         pgoff_t target, toff;
2205         pgoff_t base, end;
2206         int nr_pages = 0;
2207
2208         if (!our_page_cluster)  /* no readahead */
2209                 return 0;
2210
2211         si = swap_info[swp_type(entry)];
2212         target = swp_offset(entry);
2213         base = (target >> our_page_cluster) << our_page_cluster;
2214         end = base + (1 << our_page_cluster);
2215         if (!base)              /* first page is swap header */
2216                 base++;
2217
2218         spin_lock(&swap_lock);
2219         if (end > si->max)      /* don't go beyond end of map */
2220                 end = si->max;
2221
2222         /* Count contiguous allocated slots above our target */
2223         for (toff = target; ++toff < end; nr_pages++) {
2224                 /* Don't read in free or bad pages */
2225                 if (!si->swap_map[toff])
2226                         break;
2227                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2228                         break;
2229         }
2230         /* Count contiguous allocated slots below our target */
2231         for (toff = target; --toff >= base; nr_pages++) {
2232                 /* Don't read in free or bad pages */
2233                 if (!si->swap_map[toff])
2234                         break;
2235                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2236                         break;
2237         }
2238         spin_unlock(&swap_lock);
2239
2240         /*
2241          * Indicate starting offset, and return number of pages to get:
2242          * if only 1, say 0, since there's then no readahead to be done.
2243          */
2244         *offset = ++toff;
2245         return nr_pages? ++nr_pages: 0;
2246 }