f83f1c6f6196490052426c1e46e3e2ce506565dc
[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 /* returnes 1 if swap entry is freed */
83 static int
84 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
85 {
86         int type = si - swap_info;
87         swp_entry_t entry = swp_entry(type, offset);
88         struct page *page;
89         int ret = 0;
90
91         page = find_get_page(&swapper_space, entry.val);
92         if (!page)
93                 return 0;
94         /*
95          * This function is called from scan_swap_map() and it's called
96          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
97          * We have to use trylock for avoiding deadlock. This is a special
98          * case and you should use try_to_free_swap() with explicit lock_page()
99          * in usual operations.
100          */
101         if (trylock_page(page)) {
102                 ret = try_to_free_swap(page);
103                 unlock_page(page);
104         }
105         page_cache_release(page);
106         return ret;
107 }
108
109 /*
110  * We need this because the bdev->unplug_fn can sleep and we cannot
111  * hold swap_lock while calling the unplug_fn. And swap_lock
112  * cannot be turned into a mutex.
113  */
114 static DECLARE_RWSEM(swap_unplug_sem);
115
116 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
117 {
118         swp_entry_t entry;
119
120         down_read(&swap_unplug_sem);
121         entry.val = page_private(page);
122         if (PageSwapCache(page)) {
123                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
124                 struct backing_dev_info *bdi;
125
126                 /*
127                  * If the page is removed from swapcache from under us (with a
128                  * racy try_to_unuse/swapoff) we need an additional reference
129                  * count to avoid reading garbage from page_private(page) above.
130                  * If the WARN_ON triggers during a swapoff it maybe the race
131                  * condition and it's harmless. However if it triggers without
132                  * swapoff it signals a problem.
133                  */
134                 WARN_ON(page_count(page) <= 1);
135
136                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
137                 blk_run_backing_dev(bdi, page);
138         }
139         up_read(&swap_unplug_sem);
140 }
141
142 /*
143  * swapon tell device that all the old swap contents can be discarded,
144  * to allow the swap device to optimize its wear-levelling.
145  */
146 static int discard_swap(struct swap_info_struct *si)
147 {
148         struct swap_extent *se;
149         int err = 0;
150
151         list_for_each_entry(se, &si->extent_list, list) {
152                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
153                 sector_t nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
154
155                 if (se->start_page == 0) {
156                         /* Do not discard the swap header page! */
157                         start_block += 1 << (PAGE_SHIFT - 9);
158                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
159                         if (!nr_blocks)
160                                 continue;
161                 }
162
163                 err = blkdev_issue_discard(si->bdev, start_block,
164                                                 nr_blocks, GFP_KERNEL,
165                                                 DISCARD_FL_BARRIER);
166                 if (err)
167                         break;
168
169                 cond_resched();
170         }
171         return err;             /* That will often be -EOPNOTSUPP */
172 }
173
174 /*
175  * swap allocation tell device that a cluster of swap can now be discarded,
176  * to allow the swap device to optimize its wear-levelling.
177  */
178 static void discard_swap_cluster(struct swap_info_struct *si,
179                                  pgoff_t start_page, pgoff_t nr_pages)
180 {
181         struct swap_extent *se = si->curr_swap_extent;
182         int found_extent = 0;
183
184         while (nr_pages) {
185                 struct list_head *lh;
186
187                 if (se->start_page <= start_page &&
188                     start_page < se->start_page + se->nr_pages) {
189                         pgoff_t offset = start_page - se->start_page;
190                         sector_t start_block = se->start_block + offset;
191                         sector_t nr_blocks = se->nr_pages - offset;
192
193                         if (nr_blocks > nr_pages)
194                                 nr_blocks = nr_pages;
195                         start_page += nr_blocks;
196                         nr_pages -= nr_blocks;
197
198                         if (!found_extent++)
199                                 si->curr_swap_extent = se;
200
201                         start_block <<= PAGE_SHIFT - 9;
202                         nr_blocks <<= PAGE_SHIFT - 9;
203                         if (blkdev_issue_discard(si->bdev, start_block,
204                                                         nr_blocks, GFP_NOIO,
205                                                         DISCARD_FL_BARRIER))
206                                 break;
207                 }
208
209                 lh = se->list.next;
210                 if (lh == &si->extent_list)
211                         lh = lh->next;
212                 se = list_entry(lh, struct swap_extent, list);
213         }
214 }
215
216 static int wait_for_discard(void *word)
217 {
218         schedule();
219         return 0;
220 }
221
222 #define SWAPFILE_CLUSTER        256
223 #define LATENCY_LIMIT           256
224
225 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
226                                           int cache)
227 {
228         unsigned long offset;
229         unsigned long scan_base;
230         unsigned long last_in_cluster = 0;
231         int latency_ration = LATENCY_LIMIT;
232         int found_free_cluster = 0;
233
234         /*
235          * We try to cluster swap pages by allocating them sequentially
236          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
237          * way, however, we resort to first-free allocation, starting
238          * a new cluster.  This prevents us from scattering swap pages
239          * all over the entire swap partition, so that we reduce
240          * overall disk seek times between swap pages.  -- sct
241          * But we do now try to find an empty cluster.  -Andrea
242          * And we let swap pages go all over an SSD partition.  Hugh
243          */
244
245         si->flags += SWP_SCANNING;
246         scan_base = offset = si->cluster_next;
247
248         if (unlikely(!si->cluster_nr--)) {
249                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
250                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
251                         goto checks;
252                 }
253                 if (si->flags & SWP_DISCARDABLE) {
254                         /*
255                          * Start range check on racing allocations, in case
256                          * they overlap the cluster we eventually decide on
257                          * (we scan without swap_lock to allow preemption).
258                          * It's hardly conceivable that cluster_nr could be
259                          * wrapped during our scan, but don't depend on it.
260                          */
261                         if (si->lowest_alloc)
262                                 goto checks;
263                         si->lowest_alloc = si->max;
264                         si->highest_alloc = 0;
265                 }
266                 spin_unlock(&swap_lock);
267
268                 /*
269                  * If seek is expensive, start searching for new cluster from
270                  * start of partition, to minimize the span of allocated swap.
271                  * But if seek is cheap, search from our current position, so
272                  * that swap is allocated from all over the partition: if the
273                  * Flash Translation Layer only remaps within limited zones,
274                  * we don't want to wear out the first zone too quickly.
275                  */
276                 if (!(si->flags & SWP_SOLIDSTATE))
277                         scan_base = offset = si->lowest_bit;
278                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
279
280                 /* Locate the first empty (unaligned) cluster */
281                 for (; last_in_cluster <= si->highest_bit; offset++) {
282                         if (si->swap_map[offset])
283                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
284                         else if (offset == last_in_cluster) {
285                                 spin_lock(&swap_lock);
286                                 offset -= SWAPFILE_CLUSTER - 1;
287                                 si->cluster_next = offset;
288                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
289                                 found_free_cluster = 1;
290                                 goto checks;
291                         }
292                         if (unlikely(--latency_ration < 0)) {
293                                 cond_resched();
294                                 latency_ration = LATENCY_LIMIT;
295                         }
296                 }
297
298                 offset = si->lowest_bit;
299                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
300
301                 /* Locate the first empty (unaligned) cluster */
302                 for (; last_in_cluster < scan_base; offset++) {
303                         if (si->swap_map[offset])
304                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
305                         else if (offset == last_in_cluster) {
306                                 spin_lock(&swap_lock);
307                                 offset -= SWAPFILE_CLUSTER - 1;
308                                 si->cluster_next = offset;
309                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
310                                 found_free_cluster = 1;
311                                 goto checks;
312                         }
313                         if (unlikely(--latency_ration < 0)) {
314                                 cond_resched();
315                                 latency_ration = LATENCY_LIMIT;
316                         }
317                 }
318
319                 offset = scan_base;
320                 spin_lock(&swap_lock);
321                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
322                 si->lowest_alloc = 0;
323         }
324
325 checks:
326         if (!(si->flags & SWP_WRITEOK))
327                 goto no_page;
328         if (!si->highest_bit)
329                 goto no_page;
330         if (offset > si->highest_bit)
331                 scan_base = offset = si->lowest_bit;
332
333         /* reuse swap entry of cache-only swap if not busy. */
334         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
335                 int swap_was_freed;
336                 spin_unlock(&swap_lock);
337                 swap_was_freed = __try_to_reclaim_swap(si, offset);
338                 spin_lock(&swap_lock);
339                 /* entry was freed successfully, try to use this again */
340                 if (swap_was_freed)
341                         goto checks;
342                 goto scan; /* check next one */
343         }
344
345         if (si->swap_map[offset])
346                 goto scan;
347
348         if (offset == si->lowest_bit)
349                 si->lowest_bit++;
350         if (offset == si->highest_bit)
351                 si->highest_bit--;
352         si->inuse_pages++;
353         if (si->inuse_pages == si->pages) {
354                 si->lowest_bit = si->max;
355                 si->highest_bit = 0;
356         }
357         if (cache == SWAP_CACHE) /* at usual swap-out via vmscan.c */
358                 si->swap_map[offset] = encode_swapmap(0, true);
359         else /* at suspend */
360                 si->swap_map[offset] = encode_swapmap(1, false);
361         si->cluster_next = offset + 1;
362         si->flags -= SWP_SCANNING;
363
364         if (si->lowest_alloc) {
365                 /*
366                  * Only set when SWP_DISCARDABLE, and there's a scan
367                  * for a free cluster in progress or just completed.
368                  */
369                 if (found_free_cluster) {
370                         /*
371                          * To optimize wear-levelling, discard the
372                          * old data of the cluster, taking care not to
373                          * discard any of its pages that have already
374                          * been allocated by racing tasks (offset has
375                          * already stepped over any at the beginning).
376                          */
377                         if (offset < si->highest_alloc &&
378                             si->lowest_alloc <= last_in_cluster)
379                                 last_in_cluster = si->lowest_alloc - 1;
380                         si->flags |= SWP_DISCARDING;
381                         spin_unlock(&swap_lock);
382
383                         if (offset < last_in_cluster)
384                                 discard_swap_cluster(si, offset,
385                                         last_in_cluster - offset + 1);
386
387                         spin_lock(&swap_lock);
388                         si->lowest_alloc = 0;
389                         si->flags &= ~SWP_DISCARDING;
390
391                         smp_mb();       /* wake_up_bit advises this */
392                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
393
394                 } else if (si->flags & SWP_DISCARDING) {
395                         /*
396                          * Delay using pages allocated by racing tasks
397                          * until the whole discard has been issued. We
398                          * could defer that delay until swap_writepage,
399                          * but it's easier to keep this self-contained.
400                          */
401                         spin_unlock(&swap_lock);
402                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
403                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
404                         spin_lock(&swap_lock);
405                 } else {
406                         /*
407                          * Note pages allocated by racing tasks while
408                          * scan for a free cluster is in progress, so
409                          * that its final discard can exclude them.
410                          */
411                         if (offset < si->lowest_alloc)
412                                 si->lowest_alloc = offset;
413                         if (offset > si->highest_alloc)
414                                 si->highest_alloc = offset;
415                 }
416         }
417         return offset;
418
419 scan:
420         spin_unlock(&swap_lock);
421         while (++offset <= si->highest_bit) {
422                 if (!si->swap_map[offset]) {
423                         spin_lock(&swap_lock);
424                         goto checks;
425                 }
426                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
427                         spin_lock(&swap_lock);
428                         goto checks;
429                 }
430                 if (unlikely(--latency_ration < 0)) {
431                         cond_resched();
432                         latency_ration = LATENCY_LIMIT;
433                 }
434         }
435         offset = si->lowest_bit;
436         while (++offset < scan_base) {
437                 if (!si->swap_map[offset]) {
438                         spin_lock(&swap_lock);
439                         goto checks;
440                 }
441                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
442                         spin_lock(&swap_lock);
443                         goto checks;
444                 }
445                 if (unlikely(--latency_ration < 0)) {
446                         cond_resched();
447                         latency_ration = LATENCY_LIMIT;
448                 }
449         }
450         spin_lock(&swap_lock);
451
452 no_page:
453         si->flags -= SWP_SCANNING;
454         return 0;
455 }
456
457 swp_entry_t get_swap_page(void)
458 {
459         struct swap_info_struct *si;
460         pgoff_t offset;
461         int type, next;
462         int wrapped = 0;
463
464         spin_lock(&swap_lock);
465         if (nr_swap_pages <= 0)
466                 goto noswap;
467         nr_swap_pages--;
468
469         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
470                 si = swap_info + type;
471                 next = si->next;
472                 if (next < 0 ||
473                     (!wrapped && si->prio != swap_info[next].prio)) {
474                         next = swap_list.head;
475                         wrapped++;
476                 }
477
478                 if (!si->highest_bit)
479                         continue;
480                 if (!(si->flags & SWP_WRITEOK))
481                         continue;
482
483                 swap_list.next = next;
484                 /* This is called for allocating swap entry for cache */
485                 offset = scan_swap_map(si, SWAP_CACHE);
486                 if (offset) {
487                         spin_unlock(&swap_lock);
488                         return swp_entry(type, offset);
489                 }
490                 next = swap_list.next;
491         }
492
493         nr_swap_pages++;
494 noswap:
495         spin_unlock(&swap_lock);
496         return (swp_entry_t) {0};
497 }
498
499 /* The only caller of this function is now susupend routine */
500 swp_entry_t get_swap_page_of_type(int type)
501 {
502         struct swap_info_struct *si;
503         pgoff_t offset;
504
505         spin_lock(&swap_lock);
506         si = swap_info + type;
507         if (si->flags & SWP_WRITEOK) {
508                 nr_swap_pages--;
509                 /* This is called for allocating swap entry, not cache */
510                 offset = scan_swap_map(si, SWAP_MAP);
511                 if (offset) {
512                         spin_unlock(&swap_lock);
513                         return swp_entry(type, offset);
514                 }
515                 nr_swap_pages++;
516         }
517         spin_unlock(&swap_lock);
518         return (swp_entry_t) {0};
519 }
520
521 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
522 {
523         struct swap_info_struct * p;
524         unsigned long offset, type;
525
526         if (!entry.val)
527                 goto out;
528         type = swp_type(entry);
529         if (type >= nr_swapfiles)
530                 goto bad_nofile;
531         p = & swap_info[type];
532         if (!(p->flags & SWP_USED))
533                 goto bad_device;
534         offset = swp_offset(entry);
535         if (offset >= p->max)
536                 goto bad_offset;
537         if (!p->swap_map[offset])
538                 goto bad_free;
539         spin_lock(&swap_lock);
540         return p;
541
542 bad_free:
543         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
544         goto out;
545 bad_offset:
546         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
547         goto out;
548 bad_device:
549         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
550         goto out;
551 bad_nofile:
552         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
553 out:
554         return NULL;
555 }
556
557 static int swap_entry_free(struct swap_info_struct *p,
558                            swp_entry_t ent, int cache)
559 {
560         unsigned long offset = swp_offset(ent);
561         int count = swap_count(p->swap_map[offset]);
562         bool has_cache;
563
564         has_cache = swap_has_cache(p->swap_map[offset]);
565
566         if (cache == SWAP_MAP) { /* dropping usage count of swap */
567                 if (count < SWAP_MAP_MAX) {
568                         count--;
569                         p->swap_map[offset] = encode_swapmap(count, has_cache);
570                 }
571         } else { /* dropping swap cache flag */
572                 VM_BUG_ON(!has_cache);
573                 p->swap_map[offset] = encode_swapmap(count, false);
574
575         }
576         /* return code. */
577         count = p->swap_map[offset];
578         /* free if no reference */
579         if (!count) {
580                 if (offset < p->lowest_bit)
581                         p->lowest_bit = offset;
582                 if (offset > p->highest_bit)
583                         p->highest_bit = offset;
584                 if (p->prio > swap_info[swap_list.next].prio)
585                         swap_list.next = p - swap_info;
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 i;
745
746         if (device)
747                 bdev = bdget(device);
748
749         spin_lock(&swap_lock);
750         for (i = 0; i < nr_swapfiles; i++) {
751                 struct swap_info_struct *sis = swap_info + i;
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 i;
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 i;
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         if (type < nr_swapfiles) {
796                 spin_lock(&swap_lock);
797                 if (swap_info[type].flags & SWP_WRITEOK) {
798                         n = swap_info[type].pages;
799                         if (free)
800                                 n -= swap_info[type].inuse_pages;
801                 }
802                 spin_unlock(&swap_lock);
803         }
804         return n;
805 }
806 #endif
807
808 /*
809  * No need to decide whether this PTE shares the swap entry with others,
810  * just let do_wp_page work it out if a write is requested later - to
811  * force COW, vm_page_prot omits write permission from any private vma.
812  */
813 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
814                 unsigned long addr, swp_entry_t entry, struct page *page)
815 {
816         struct mem_cgroup *ptr = NULL;
817         spinlock_t *ptl;
818         pte_t *pte;
819         int ret = 1;
820
821         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
822                 ret = -ENOMEM;
823                 goto out_nolock;
824         }
825
826         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
827         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
828                 if (ret > 0)
829                         mem_cgroup_cancel_charge_swapin(ptr);
830                 ret = 0;
831                 goto out;
832         }
833
834         inc_mm_counter(vma->vm_mm, anon_rss);
835         get_page(page);
836         set_pte_at(vma->vm_mm, addr, pte,
837                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
838         page_add_anon_rmap(page, vma, addr);
839         mem_cgroup_commit_charge_swapin(page, ptr);
840         swap_free(entry);
841         /*
842          * Move the page to the active list so it is not
843          * immediately swapped out again after swapon.
844          */
845         activate_page(page);
846 out:
847         pte_unmap_unlock(pte, ptl);
848 out_nolock:
849         return ret;
850 }
851
852 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
853                                 unsigned long addr, unsigned long end,
854                                 swp_entry_t entry, struct page *page)
855 {
856         pte_t swp_pte = swp_entry_to_pte(entry);
857         pte_t *pte;
858         int ret = 0;
859
860         /*
861          * We don't actually need pte lock while scanning for swp_pte: since
862          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
863          * page table while we're scanning; though it could get zapped, and on
864          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
865          * of unmatched parts which look like swp_pte, so unuse_pte must
866          * recheck under pte lock.  Scanning without pte lock lets it be
867          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
868          */
869         pte = pte_offset_map(pmd, addr);
870         do {
871                 /*
872                  * swapoff spends a _lot_ of time in this loop!
873                  * Test inline before going to call unuse_pte.
874                  */
875                 if (unlikely(pte_same(*pte, swp_pte))) {
876                         pte_unmap(pte);
877                         ret = unuse_pte(vma, pmd, addr, entry, page);
878                         if (ret)
879                                 goto out;
880                         pte = pte_offset_map(pmd, addr);
881                 }
882         } while (pte++, addr += PAGE_SIZE, addr != end);
883         pte_unmap(pte - 1);
884 out:
885         return ret;
886 }
887
888 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
889                                 unsigned long addr, unsigned long end,
890                                 swp_entry_t entry, struct page *page)
891 {
892         pmd_t *pmd;
893         unsigned long next;
894         int ret;
895
896         pmd = pmd_offset(pud, addr);
897         do {
898                 next = pmd_addr_end(addr, end);
899                 if (pmd_none_or_clear_bad(pmd))
900                         continue;
901                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
902                 if (ret)
903                         return ret;
904         } while (pmd++, addr = next, addr != end);
905         return 0;
906 }
907
908 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
909                                 unsigned long addr, unsigned long end,
910                                 swp_entry_t entry, struct page *page)
911 {
912         pud_t *pud;
913         unsigned long next;
914         int ret;
915
916         pud = pud_offset(pgd, addr);
917         do {
918                 next = pud_addr_end(addr, end);
919                 if (pud_none_or_clear_bad(pud))
920                         continue;
921                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
922                 if (ret)
923                         return ret;
924         } while (pud++, addr = next, addr != end);
925         return 0;
926 }
927
928 static int unuse_vma(struct vm_area_struct *vma,
929                                 swp_entry_t entry, struct page *page)
930 {
931         pgd_t *pgd;
932         unsigned long addr, end, next;
933         int ret;
934
935         if (page->mapping) {
936                 addr = page_address_in_vma(page, vma);
937                 if (addr == -EFAULT)
938                         return 0;
939                 else
940                         end = addr + PAGE_SIZE;
941         } else {
942                 addr = vma->vm_start;
943                 end = vma->vm_end;
944         }
945
946         pgd = pgd_offset(vma->vm_mm, addr);
947         do {
948                 next = pgd_addr_end(addr, end);
949                 if (pgd_none_or_clear_bad(pgd))
950                         continue;
951                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
952                 if (ret)
953                         return ret;
954         } while (pgd++, addr = next, addr != end);
955         return 0;
956 }
957
958 static int unuse_mm(struct mm_struct *mm,
959                                 swp_entry_t entry, struct page *page)
960 {
961         struct vm_area_struct *vma;
962         int ret = 0;
963
964         if (!down_read_trylock(&mm->mmap_sem)) {
965                 /*
966                  * Activate page so shrink_inactive_list is unlikely to unmap
967                  * its ptes while lock is dropped, so swapoff can make progress.
968                  */
969                 activate_page(page);
970                 unlock_page(page);
971                 down_read(&mm->mmap_sem);
972                 lock_page(page);
973         }
974         for (vma = mm->mmap; vma; vma = vma->vm_next) {
975                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
976                         break;
977         }
978         up_read(&mm->mmap_sem);
979         return (ret < 0)? ret: 0;
980 }
981
982 /*
983  * Scan swap_map from current position to next entry still in use.
984  * Recycle to start on reaching the end, returning 0 when empty.
985  */
986 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
987                                         unsigned int prev)
988 {
989         unsigned int max = si->max;
990         unsigned int i = prev;
991         int count;
992
993         /*
994          * No need for swap_lock here: we're just looking
995          * for whether an entry is in use, not modifying it; false
996          * hits are okay, and sys_swapoff() has already prevented new
997          * allocations from this area (while holding swap_lock).
998          */
999         for (;;) {
1000                 if (++i >= max) {
1001                         if (!prev) {
1002                                 i = 0;
1003                                 break;
1004                         }
1005                         /*
1006                          * No entries in use at top of swap_map,
1007                          * loop back to start and recheck there.
1008                          */
1009                         max = prev + 1;
1010                         prev = 0;
1011                         i = 1;
1012                 }
1013                 count = si->swap_map[i];
1014                 if (count && swap_count(count) != SWAP_MAP_BAD)
1015                         break;
1016         }
1017         return i;
1018 }
1019
1020 /*
1021  * We completely avoid races by reading each swap page in advance,
1022  * and then search for the process using it.  All the necessary
1023  * page table adjustments can then be made atomically.
1024  */
1025 static int try_to_unuse(unsigned int type)
1026 {
1027         struct swap_info_struct * si = &swap_info[type];
1028         struct mm_struct *start_mm;
1029         unsigned short *swap_map;
1030         unsigned short swcount;
1031         struct page *page;
1032         swp_entry_t entry;
1033         unsigned int i = 0;
1034         int retval = 0;
1035         int reset_overflow = 0;
1036         int shmem;
1037
1038         /*
1039          * When searching mms for an entry, a good strategy is to
1040          * start at the first mm we freed the previous entry from
1041          * (though actually we don't notice whether we or coincidence
1042          * freed the entry).  Initialize this start_mm with a hold.
1043          *
1044          * A simpler strategy would be to start at the last mm we
1045          * freed the previous entry from; but that would take less
1046          * advantage of mmlist ordering, which clusters forked mms
1047          * together, child after parent.  If we race with dup_mmap(), we
1048          * prefer to resolve parent before child, lest we miss entries
1049          * duplicated after we scanned child: using last mm would invert
1050          * that.  Though it's only a serious concern when an overflowed
1051          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
1052          */
1053         start_mm = &init_mm;
1054         atomic_inc(&init_mm.mm_users);
1055
1056         /*
1057          * Keep on scanning until all entries have gone.  Usually,
1058          * one pass through swap_map is enough, but not necessarily:
1059          * there are races when an instance of an entry might be missed.
1060          */
1061         while ((i = find_next_to_unuse(si, i)) != 0) {
1062                 if (signal_pending(current)) {
1063                         retval = -EINTR;
1064                         break;
1065                 }
1066
1067                 /*
1068                  * Get a page for the entry, using the existing swap
1069                  * cache page if there is one.  Otherwise, get a clean
1070                  * page and read the swap into it.
1071                  */
1072                 swap_map = &si->swap_map[i];
1073                 entry = swp_entry(type, i);
1074                 page = read_swap_cache_async(entry,
1075                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1076                 if (!page) {
1077                         /*
1078                          * Either swap_duplicate() failed because entry
1079                          * has been freed independently, and will not be
1080                          * reused since sys_swapoff() already disabled
1081                          * allocation from here, or alloc_page() failed.
1082                          */
1083                         if (!*swap_map)
1084                                 continue;
1085                         retval = -ENOMEM;
1086                         break;
1087                 }
1088
1089                 /*
1090                  * Don't hold on to start_mm if it looks like exiting.
1091                  */
1092                 if (atomic_read(&start_mm->mm_users) == 1) {
1093                         mmput(start_mm);
1094                         start_mm = &init_mm;
1095                         atomic_inc(&init_mm.mm_users);
1096                 }
1097
1098                 /*
1099                  * Wait for and lock page.  When do_swap_page races with
1100                  * try_to_unuse, do_swap_page can handle the fault much
1101                  * faster than try_to_unuse can locate the entry.  This
1102                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1103                  * defer to do_swap_page in such a case - in some tests,
1104                  * do_swap_page and try_to_unuse repeatedly compete.
1105                  */
1106                 wait_on_page_locked(page);
1107                 wait_on_page_writeback(page);
1108                 lock_page(page);
1109                 wait_on_page_writeback(page);
1110
1111                 /*
1112                  * Remove all references to entry.
1113                  * Whenever we reach init_mm, there's no address space
1114                  * to search, but use it as a reminder to search shmem.
1115                  */
1116                 shmem = 0;
1117                 swcount = *swap_map;
1118                 if (swap_count(swcount)) {
1119                         if (start_mm == &init_mm)
1120                                 shmem = shmem_unuse(entry, page);
1121                         else
1122                                 retval = unuse_mm(start_mm, entry, page);
1123                 }
1124                 if (swap_count(*swap_map)) {
1125                         int set_start_mm = (*swap_map >= swcount);
1126                         struct list_head *p = &start_mm->mmlist;
1127                         struct mm_struct *new_start_mm = start_mm;
1128                         struct mm_struct *prev_mm = start_mm;
1129                         struct mm_struct *mm;
1130
1131                         atomic_inc(&new_start_mm->mm_users);
1132                         atomic_inc(&prev_mm->mm_users);
1133                         spin_lock(&mmlist_lock);
1134                         while (swap_count(*swap_map) && !retval && !shmem &&
1135                                         (p = p->next) != &start_mm->mmlist) {
1136                                 mm = list_entry(p, struct mm_struct, mmlist);
1137                                 if (!atomic_inc_not_zero(&mm->mm_users))
1138                                         continue;
1139                                 spin_unlock(&mmlist_lock);
1140                                 mmput(prev_mm);
1141                                 prev_mm = mm;
1142
1143                                 cond_resched();
1144
1145                                 swcount = *swap_map;
1146                                 if (!swap_count(swcount)) /* any usage ? */
1147                                         ;
1148                                 else if (mm == &init_mm) {
1149                                         set_start_mm = 1;
1150                                         shmem = shmem_unuse(entry, page);
1151                                 } else
1152                                         retval = unuse_mm(mm, entry, page);
1153
1154                                 if (set_start_mm && *swap_map < swcount) {
1155                                         mmput(new_start_mm);
1156                                         atomic_inc(&mm->mm_users);
1157                                         new_start_mm = mm;
1158                                         set_start_mm = 0;
1159                                 }
1160                                 spin_lock(&mmlist_lock);
1161                         }
1162                         spin_unlock(&mmlist_lock);
1163                         mmput(prev_mm);
1164                         mmput(start_mm);
1165                         start_mm = new_start_mm;
1166                 }
1167                 if (shmem) {
1168                         /* page has already been unlocked and released */
1169                         if (shmem > 0)
1170                                 continue;
1171                         retval = shmem;
1172                         break;
1173                 }
1174                 if (retval) {
1175                         unlock_page(page);
1176                         page_cache_release(page);
1177                         break;
1178                 }
1179
1180                 /*
1181                  * How could swap count reach 0x7ffe ?
1182                  * There's no way to repeat a swap page within an mm
1183                  * (except in shmem, where it's the shared object which takes
1184                  * the reference count)?
1185                  * We believe SWAP_MAP_MAX cannot occur.(if occur, unsigned
1186                  * short is too small....)
1187                  * If that's wrong, then we should worry more about
1188                  * exit_mmap() and do_munmap() cases described above:
1189                  * we might be resetting SWAP_MAP_MAX too early here.
1190                  * We know "Undead"s can happen, they're okay, so don't
1191                  * report them; but do report if we reset SWAP_MAP_MAX.
1192                  */
1193                 /* We might release the lock_page() in unuse_mm(). */
1194                 if (!PageSwapCache(page) || page_private(page) != entry.val)
1195                         goto retry;
1196
1197                 if (swap_count(*swap_map) == SWAP_MAP_MAX) {
1198                         spin_lock(&swap_lock);
1199                         *swap_map = encode_swapmap(0, true);
1200                         spin_unlock(&swap_lock);
1201                         reset_overflow = 1;
1202                 }
1203
1204                 /*
1205                  * If a reference remains (rare), we would like to leave
1206                  * the page in the swap cache; but try_to_unmap could
1207                  * then re-duplicate the entry once we drop page lock,
1208                  * so we might loop indefinitely; also, that page could
1209                  * not be swapped out to other storage meanwhile.  So:
1210                  * delete from cache even if there's another reference,
1211                  * after ensuring that the data has been saved to disk -
1212                  * since if the reference remains (rarer), it will be
1213                  * read from disk into another page.  Splitting into two
1214                  * pages would be incorrect if swap supported "shared
1215                  * private" pages, but they are handled by tmpfs files.
1216                  */
1217                 if (swap_count(*swap_map) &&
1218                      PageDirty(page) && PageSwapCache(page)) {
1219                         struct writeback_control wbc = {
1220                                 .sync_mode = WB_SYNC_NONE,
1221                         };
1222
1223                         swap_writepage(page, &wbc);
1224                         lock_page(page);
1225                         wait_on_page_writeback(page);
1226                 }
1227
1228                 /*
1229                  * It is conceivable that a racing task removed this page from
1230                  * swap cache just before we acquired the page lock at the top,
1231                  * or while we dropped it in unuse_mm().  The page might even
1232                  * be back in swap cache on another swap area: that we must not
1233                  * delete, since it may not have been written out to swap yet.
1234                  */
1235                 if (PageSwapCache(page) &&
1236                     likely(page_private(page) == entry.val))
1237                         delete_from_swap_cache(page);
1238
1239                 /*
1240                  * So we could skip searching mms once swap count went
1241                  * to 1, we did not mark any present ptes as dirty: must
1242                  * mark page dirty so shrink_page_list will preserve it.
1243                  */
1244                 SetPageDirty(page);
1245 retry:
1246                 unlock_page(page);
1247                 page_cache_release(page);
1248
1249                 /*
1250                  * Make sure that we aren't completely killing
1251                  * interactive performance.
1252                  */
1253                 cond_resched();
1254         }
1255
1256         mmput(start_mm);
1257         if (reset_overflow) {
1258                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1259                 swap_overflow = 0;
1260         }
1261         return retval;
1262 }
1263
1264 /*
1265  * After a successful try_to_unuse, if no swap is now in use, we know
1266  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1267  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1268  * added to the mmlist just after page_duplicate - before would be racy.
1269  */
1270 static void drain_mmlist(void)
1271 {
1272         struct list_head *p, *next;
1273         unsigned int i;
1274
1275         for (i = 0; i < nr_swapfiles; i++)
1276                 if (swap_info[i].inuse_pages)
1277                         return;
1278         spin_lock(&mmlist_lock);
1279         list_for_each_safe(p, next, &init_mm.mmlist)
1280                 list_del_init(p);
1281         spin_unlock(&mmlist_lock);
1282 }
1283
1284 /*
1285  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1286  * corresponds to page offset `offset'.  Note that the type of this function
1287  * is sector_t, but it returns page offset into the bdev, not sector offset.
1288  */
1289 sector_t map_swap_page(swp_entry_t entry, struct block_device **bdev)
1290 {
1291         struct swap_info_struct *sis;
1292         struct swap_extent *start_se;
1293         struct swap_extent *se;
1294         pgoff_t offset;
1295
1296         sis = swap_info + swp_type(entry);
1297         *bdev = sis->bdev;
1298
1299         offset = swp_offset(entry);
1300         start_se = sis->curr_swap_extent;
1301         se = start_se;
1302
1303         for ( ; ; ) {
1304                 struct list_head *lh;
1305
1306                 if (se->start_page <= offset &&
1307                                 offset < (se->start_page + se->nr_pages)) {
1308                         return se->start_block + (offset - se->start_page);
1309                 }
1310                 lh = se->list.next;
1311                 if (lh == &sis->extent_list)
1312                         lh = lh->next;
1313                 se = list_entry(lh, struct swap_extent, list);
1314                 sis->curr_swap_extent = se;
1315                 BUG_ON(se == start_se);         /* It *must* be present */
1316         }
1317 }
1318
1319 #ifdef CONFIG_HIBERNATION
1320 /*
1321  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1322  * corresponding to given index in swap_info (swap type).
1323  */
1324 sector_t swapdev_block(int swap_type, pgoff_t offset)
1325 {
1326         struct swap_info_struct *sis;
1327         struct block_device *bdev;
1328
1329         if (swap_type >= nr_swapfiles)
1330                 return 0;
1331
1332         sis = swap_info + swap_type;
1333         return (sis->flags & SWP_WRITEOK) ?
1334                 map_swap_page(swp_entry(swap_type, offset), &bdev) : 0;
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         }
1575         if (type == swap_list.next) {
1576                 /* just pick something that's safe... */
1577                 swap_list.next = swap_list.head;
1578         }
1579         if (p->prio < 0) {
1580                 for (i = p->next; i >= 0; i = swap_info[i].next)
1581                         swap_info[i].prio = p->prio--;
1582                 least_priority++;
1583         }
1584         nr_swap_pages -= p->pages;
1585         total_swap_pages -= p->pages;
1586         p->flags &= ~SWP_WRITEOK;
1587         spin_unlock(&swap_lock);
1588
1589         current->flags |= PF_OOM_ORIGIN;
1590         err = try_to_unuse(type);
1591         current->flags &= ~PF_OOM_ORIGIN;
1592
1593         if (err) {
1594                 /* re-insert swap space back into swap_list */
1595                 spin_lock(&swap_lock);
1596                 if (p->prio < 0)
1597                         p->prio = --least_priority;
1598                 prev = -1;
1599                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1600                         if (p->prio >= swap_info[i].prio)
1601                                 break;
1602                         prev = i;
1603                 }
1604                 p->next = i;
1605                 if (prev < 0)
1606                         swap_list.head = swap_list.next = p - swap_info;
1607                 else
1608                         swap_info[prev].next = p - swap_info;
1609                 nr_swap_pages += p->pages;
1610                 total_swap_pages += p->pages;
1611                 p->flags |= SWP_WRITEOK;
1612                 spin_unlock(&swap_lock);
1613                 goto out_dput;
1614         }
1615
1616         /* wait for any unplug function to finish */
1617         down_write(&swap_unplug_sem);
1618         up_write(&swap_unplug_sem);
1619
1620         destroy_swap_extents(p);
1621         mutex_lock(&swapon_mutex);
1622         spin_lock(&swap_lock);
1623         drain_mmlist();
1624
1625         /* wait for anyone still in scan_swap_map */
1626         p->highest_bit = 0;             /* cuts scans short */
1627         while (p->flags >= SWP_SCANNING) {
1628                 spin_unlock(&swap_lock);
1629                 schedule_timeout_uninterruptible(1);
1630                 spin_lock(&swap_lock);
1631         }
1632
1633         swap_file = p->swap_file;
1634         p->swap_file = NULL;
1635         p->max = 0;
1636         swap_map = p->swap_map;
1637         p->swap_map = NULL;
1638         p->flags = 0;
1639         spin_unlock(&swap_lock);
1640         mutex_unlock(&swapon_mutex);
1641         vfree(swap_map);
1642         /* Destroy swap account informatin */
1643         swap_cgroup_swapoff(type);
1644
1645         inode = mapping->host;
1646         if (S_ISBLK(inode->i_mode)) {
1647                 struct block_device *bdev = I_BDEV(inode);
1648                 set_blocksize(bdev, p->old_block_size);
1649                 bd_release(bdev);
1650         } else {
1651                 mutex_lock(&inode->i_mutex);
1652                 inode->i_flags &= ~S_SWAPFILE;
1653                 mutex_unlock(&inode->i_mutex);
1654         }
1655         filp_close(swap_file, NULL);
1656         err = 0;
1657
1658 out_dput:
1659         filp_close(victim, NULL);
1660 out:
1661         return err;
1662 }
1663
1664 #ifdef CONFIG_PROC_FS
1665 /* iterator */
1666 static void *swap_start(struct seq_file *swap, loff_t *pos)
1667 {
1668         struct swap_info_struct *ptr = swap_info;
1669         int i;
1670         loff_t l = *pos;
1671
1672         mutex_lock(&swapon_mutex);
1673
1674         if (!l)
1675                 return SEQ_START_TOKEN;
1676
1677         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1678                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1679                         continue;
1680                 if (!--l)
1681                         return ptr;
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 *ptr;
1690         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1691
1692         if (v == SEQ_START_TOKEN)
1693                 ptr = swap_info;
1694         else {
1695                 ptr = v;
1696                 ptr++;
1697         }
1698
1699         for (; ptr < endptr; ptr++) {
1700                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1701                         continue;
1702                 ++*pos;
1703                 return ptr;
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 *ptr = v;
1717         struct file *file;
1718         int len;
1719
1720         if (ptr == SEQ_START_TOKEN) {
1721                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1722                 return 0;
1723         }
1724
1725         file = ptr->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                         ptr->pages << (PAGE_SHIFT - 10),
1732                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1733                         ptr->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         spin_lock(&swap_lock);
1802         p = swap_info;
1803         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1804                 if (!(p->flags & SWP_USED))
1805                         break;
1806         error = -EPERM;
1807         if (type >= MAX_SWAPFILES) {
1808                 spin_unlock(&swap_lock);
1809                 goto out;
1810         }
1811         if (type >= nr_swapfiles)
1812                 nr_swapfiles = type+1;
1813         memset(p, 0, sizeof(*p));
1814         INIT_LIST_HEAD(&p->extent_list);
1815         p->flags = SWP_USED;
1816         p->next = -1;
1817         spin_unlock(&swap_lock);
1818         name = getname(specialfile);
1819         error = PTR_ERR(name);
1820         if (IS_ERR(name)) {
1821                 name = NULL;
1822                 goto bad_swap_2;
1823         }
1824         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1825         error = PTR_ERR(swap_file);
1826         if (IS_ERR(swap_file)) {
1827                 swap_file = NULL;
1828                 goto bad_swap_2;
1829         }
1830
1831         p->swap_file = swap_file;
1832         mapping = swap_file->f_mapping;
1833         inode = mapping->host;
1834
1835         error = -EBUSY;
1836         for (i = 0; i < nr_swapfiles; i++) {
1837                 struct swap_info_struct *q = &swap_info[i];
1838
1839                 if (i == type || !q->swap_file)
1840                         continue;
1841                 if (mapping == q->swap_file->f_mapping)
1842                         goto bad_swap;
1843         }
1844
1845         error = -EINVAL;
1846         if (S_ISBLK(inode->i_mode)) {
1847                 bdev = I_BDEV(inode);
1848                 error = bd_claim(bdev, sys_swapon);
1849                 if (error < 0) {
1850                         bdev = NULL;
1851                         error = -EINVAL;
1852                         goto bad_swap;
1853                 }
1854                 p->old_block_size = block_size(bdev);
1855                 error = set_blocksize(bdev, PAGE_SIZE);
1856                 if (error < 0)
1857                         goto bad_swap;
1858                 p->bdev = bdev;
1859         } else if (S_ISREG(inode->i_mode)) {
1860                 p->bdev = inode->i_sb->s_bdev;
1861                 mutex_lock(&inode->i_mutex);
1862                 did_down = 1;
1863                 if (IS_SWAPFILE(inode)) {
1864                         error = -EBUSY;
1865                         goto bad_swap;
1866                 }
1867         } else {
1868                 goto bad_swap;
1869         }
1870
1871         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1872
1873         /*
1874          * Read the swap header.
1875          */
1876         if (!mapping->a_ops->readpage) {
1877                 error = -EINVAL;
1878                 goto bad_swap;
1879         }
1880         page = read_mapping_page(mapping, 0, swap_file);
1881         if (IS_ERR(page)) {
1882                 error = PTR_ERR(page);
1883                 goto bad_swap;
1884         }
1885         swap_header = kmap(page);
1886
1887         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1888                 printk(KERN_ERR "Unable to find swap-space signature\n");
1889                 error = -EINVAL;
1890                 goto bad_swap;
1891         }
1892
1893         /* swap partition endianess hack... */
1894         if (swab32(swap_header->info.version) == 1) {
1895                 swab32s(&swap_header->info.version);
1896                 swab32s(&swap_header->info.last_page);
1897                 swab32s(&swap_header->info.nr_badpages);
1898                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1899                         swab32s(&swap_header->info.badpages[i]);
1900         }
1901         /* Check the swap header's sub-version */
1902         if (swap_header->info.version != 1) {
1903                 printk(KERN_WARNING
1904                        "Unable to handle swap header version %d\n",
1905                        swap_header->info.version);
1906                 error = -EINVAL;
1907                 goto bad_swap;
1908         }
1909
1910         p->lowest_bit  = 1;
1911         p->cluster_next = 1;
1912
1913         /*
1914          * Find out how many pages are allowed for a single swap
1915          * device. There are two limiting factors: 1) the number of
1916          * bits for the swap offset in the swp_entry_t type and
1917          * 2) the number of bits in the a swap pte as defined by
1918          * the different architectures. In order to find the
1919          * largest possible bit mask a swap entry with swap type 0
1920          * and swap offset ~0UL is created, encoded to a swap pte,
1921          * decoded to a swp_entry_t again and finally the swap
1922          * offset is extracted. This will mask all the bits from
1923          * the initial ~0UL mask that can't be encoded in either
1924          * the swp_entry_t or the architecture definition of a
1925          * swap pte.
1926          */
1927         maxpages = swp_offset(pte_to_swp_entry(
1928                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1929         if (maxpages > swap_header->info.last_page)
1930                 maxpages = swap_header->info.last_page;
1931         p->highest_bit = maxpages - 1;
1932
1933         error = -EINVAL;
1934         if (!maxpages)
1935                 goto bad_swap;
1936         if (swapfilepages && maxpages > swapfilepages) {
1937                 printk(KERN_WARNING
1938                        "Swap area shorter than signature indicates\n");
1939                 goto bad_swap;
1940         }
1941         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1942                 goto bad_swap;
1943         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1944                 goto bad_swap;
1945
1946         /* OK, set up the swap map and apply the bad block list */
1947         swap_map = vmalloc(maxpages * sizeof(short));
1948         if (!swap_map) {
1949                 error = -ENOMEM;
1950                 goto bad_swap;
1951         }
1952
1953         memset(swap_map, 0, maxpages * sizeof(short));
1954         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1955                 int page_nr = swap_header->info.badpages[i];
1956                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1957                         error = -EINVAL;
1958                         goto bad_swap;
1959                 }
1960                 swap_map[page_nr] = SWAP_MAP_BAD;
1961         }
1962
1963         error = swap_cgroup_swapon(type, maxpages);
1964         if (error)
1965                 goto bad_swap;
1966
1967         nr_good_pages = swap_header->info.last_page -
1968                         swap_header->info.nr_badpages -
1969                         1 /* header page */;
1970
1971         if (nr_good_pages) {
1972                 swap_map[0] = SWAP_MAP_BAD;
1973                 p->max = maxpages;
1974                 p->pages = nr_good_pages;
1975                 nr_extents = setup_swap_extents(p, &span);
1976                 if (nr_extents < 0) {
1977                         error = nr_extents;
1978                         goto bad_swap;
1979                 }
1980                 nr_good_pages = p->pages;
1981         }
1982         if (!nr_good_pages) {
1983                 printk(KERN_WARNING "Empty swap-file\n");
1984                 error = -EINVAL;
1985                 goto bad_swap;
1986         }
1987
1988         if (p->bdev) {
1989                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1990                         p->flags |= SWP_SOLIDSTATE;
1991                         p->cluster_next = 1 + (random32() % p->highest_bit);
1992                 }
1993                 if (discard_swap(p) == 0)
1994                         p->flags |= SWP_DISCARDABLE;
1995         }
1996
1997         mutex_lock(&swapon_mutex);
1998         spin_lock(&swap_lock);
1999         if (swap_flags & SWAP_FLAG_PREFER)
2000                 p->prio =
2001                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2002         else
2003                 p->prio = --least_priority;
2004         p->swap_map = swap_map;
2005         p->flags |= SWP_WRITEOK;
2006         nr_swap_pages += nr_good_pages;
2007         total_swap_pages += nr_good_pages;
2008
2009         printk(KERN_INFO "Adding %uk swap on %s.  "
2010                         "Priority:%d extents:%d across:%lluk %s%s\n",
2011                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2012                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2013                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2014                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2015
2016         /* insert swap space into swap_list: */
2017         prev = -1;
2018         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
2019                 if (p->prio >= swap_info[i].prio) {
2020                         break;
2021                 }
2022                 prev = i;
2023         }
2024         p->next = i;
2025         if (prev < 0) {
2026                 swap_list.head = swap_list.next = p - swap_info;
2027         } else {
2028                 swap_info[prev].next = p - swap_info;
2029         }
2030         spin_unlock(&swap_lock);
2031         mutex_unlock(&swapon_mutex);
2032         error = 0;
2033         goto out;
2034 bad_swap:
2035         if (bdev) {
2036                 set_blocksize(bdev, p->old_block_size);
2037                 bd_release(bdev);
2038         }
2039         destroy_swap_extents(p);
2040         swap_cgroup_swapoff(type);
2041 bad_swap_2:
2042         spin_lock(&swap_lock);
2043         p->swap_file = NULL;
2044         p->flags = 0;
2045         spin_unlock(&swap_lock);
2046         vfree(swap_map);
2047         if (swap_file)
2048                 filp_close(swap_file, NULL);
2049 out:
2050         if (page && !IS_ERR(page)) {
2051                 kunmap(page);
2052                 page_cache_release(page);
2053         }
2054         if (name)
2055                 putname(name);
2056         if (did_down) {
2057                 if (!error)
2058                         inode->i_flags |= S_SWAPFILE;
2059                 mutex_unlock(&inode->i_mutex);
2060         }
2061         return error;
2062 }
2063
2064 void si_swapinfo(struct sysinfo *val)
2065 {
2066         unsigned int i;
2067         unsigned long nr_to_be_unused = 0;
2068
2069         spin_lock(&swap_lock);
2070         for (i = 0; i < nr_swapfiles; i++) {
2071                 if (!(swap_info[i].flags & SWP_USED) ||
2072                      (swap_info[i].flags & SWP_WRITEOK))
2073                         continue;
2074                 nr_to_be_unused += swap_info[i].inuse_pages;
2075         }
2076         val->freeswap = nr_swap_pages + nr_to_be_unused;
2077         val->totalswap = total_swap_pages + nr_to_be_unused;
2078         spin_unlock(&swap_lock);
2079 }
2080
2081 /*
2082  * Verify that a swap entry is valid and increment its swap map count.
2083  *
2084  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
2085  * "permanent", but will be reclaimed by the next swapoff.
2086  * Returns error code in following case.
2087  * - success -> 0
2088  * - swp_entry is invalid -> EINVAL
2089  * - swp_entry is migration entry -> EINVAL
2090  * - swap-cache reference is requested but there is already one. -> EEXIST
2091  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2092  */
2093 static int __swap_duplicate(swp_entry_t entry, bool cache)
2094 {
2095         struct swap_info_struct * p;
2096         unsigned long offset, type;
2097         int result = -EINVAL;
2098         int count;
2099         bool has_cache;
2100
2101         if (non_swap_entry(entry))
2102                 return -EINVAL;
2103
2104         type = swp_type(entry);
2105         if (type >= nr_swapfiles)
2106                 goto bad_file;
2107         p = type + swap_info;
2108         offset = swp_offset(entry);
2109
2110         spin_lock(&swap_lock);
2111
2112         if (unlikely(offset >= p->max))
2113                 goto unlock_out;
2114
2115         count = swap_count(p->swap_map[offset]);
2116         has_cache = swap_has_cache(p->swap_map[offset]);
2117
2118         if (cache == SWAP_CACHE) { /* called for swapcache/swapin-readahead */
2119
2120                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2121                 if (!has_cache && count) {
2122                         p->swap_map[offset] = encode_swapmap(count, true);
2123                         result = 0;
2124                 } else if (has_cache) /* someone added cache */
2125                         result = -EEXIST;
2126                 else if (!count) /* no users */
2127                         result = -ENOENT;
2128
2129         } else if (count || has_cache) {
2130                 if (count < SWAP_MAP_MAX - 1) {
2131                         p->swap_map[offset] = encode_swapmap(count + 1,
2132                                                              has_cache);
2133                         result = 0;
2134                 } else if (count <= SWAP_MAP_MAX) {
2135                         if (swap_overflow++ < 5)
2136                                 printk(KERN_WARNING
2137                                        "swap_dup: swap entry overflow\n");
2138                         p->swap_map[offset] = encode_swapmap(SWAP_MAP_MAX,
2139                                                               has_cache);
2140                         result = 0;
2141                 }
2142         } else
2143                 result = -ENOENT; /* unused swap entry */
2144 unlock_out:
2145         spin_unlock(&swap_lock);
2146 out:
2147         return result;
2148
2149 bad_file:
2150         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2151         goto out;
2152 }
2153 /*
2154  * increase reference count of swap entry by 1.
2155  */
2156 void swap_duplicate(swp_entry_t entry)
2157 {
2158         __swap_duplicate(entry, SWAP_MAP);
2159 }
2160
2161 /*
2162  * @entry: swap entry for which we allocate swap cache.
2163  *
2164  * Called when allocating swap cache for exising swap entry,
2165  * This can return error codes. Returns 0 at success.
2166  * -EBUSY means there is a swap cache.
2167  * Note: return code is different from swap_duplicate().
2168  */
2169 int swapcache_prepare(swp_entry_t entry)
2170 {
2171         return __swap_duplicate(entry, SWAP_CACHE);
2172 }
2173
2174 /*
2175  * swap_lock prevents swap_map being freed. Don't grab an extra
2176  * reference on the swaphandle, it doesn't matter if it becomes unused.
2177  */
2178 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2179 {
2180         struct swap_info_struct *si;
2181         int our_page_cluster = page_cluster;
2182         pgoff_t target, toff;
2183         pgoff_t base, end;
2184         int nr_pages = 0;
2185
2186         if (!our_page_cluster)  /* no readahead */
2187                 return 0;
2188
2189         si = &swap_info[swp_type(entry)];
2190         target = swp_offset(entry);
2191         base = (target >> our_page_cluster) << our_page_cluster;
2192         end = base + (1 << our_page_cluster);
2193         if (!base)              /* first page is swap header */
2194                 base++;
2195
2196         spin_lock(&swap_lock);
2197         if (end > si->max)      /* don't go beyond end of map */
2198                 end = si->max;
2199
2200         /* Count contiguous allocated slots above our target */
2201         for (toff = target; ++toff < end; nr_pages++) {
2202                 /* Don't read in free or bad pages */
2203                 if (!si->swap_map[toff])
2204                         break;
2205                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2206                         break;
2207         }
2208         /* Count contiguous allocated slots below our target */
2209         for (toff = target; --toff >= base; nr_pages++) {
2210                 /* Don't read in free or bad pages */
2211                 if (!si->swap_map[toff])
2212                         break;
2213                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2214                         break;
2215         }
2216         spin_unlock(&swap_lock);
2217
2218         /*
2219          * Indicate starting offset, and return number of pages to get:
2220          * if only 1, say 0, since there's then no readahead to be done.
2221          */
2222         *offset = ++toff;
2223         return nr_pages? ++nr_pages: 0;
2224 }