Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph...
[platform/adaptation/renesas_rcar/renesas_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/shmem_fs.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/ksm.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 #include <linux/poll.h>
33 #include <linux/oom.h>
34
35 #include <asm/pgtable.h>
36 #include <asm/tlbflush.h>
37 #include <linux/swapops.h>
38 #include <linux/page_cgroup.h>
39
40 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
41                                  unsigned char);
42 static void free_swap_count_continuations(struct swap_info_struct *);
43 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
44
45 static DEFINE_SPINLOCK(swap_lock);
46 static unsigned int nr_swapfiles;
47 long nr_swap_pages;
48 long total_swap_pages;
49 static int least_priority;
50
51 static const char Bad_file[] = "Bad swap file entry ";
52 static const char Unused_file[] = "Unused swap file entry ";
53 static const char Bad_offset[] = "Bad swap offset entry ";
54 static const char Unused_offset[] = "Unused swap offset entry ";
55
56 static struct swap_list_t swap_list = {-1, -1};
57
58 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
59
60 static DEFINE_MUTEX(swapon_mutex);
61
62 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
63 /* Activity counter to indicate that a swapon or swapoff has occurred */
64 static atomic_t proc_poll_event = ATOMIC_INIT(0);
65
66 static inline unsigned char swap_count(unsigned char ent)
67 {
68         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
69 }
70
71 /* returns 1 if swap entry is freed */
72 static int
73 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
74 {
75         swp_entry_t entry = swp_entry(si->type, offset);
76         struct page *page;
77         int ret = 0;
78
79         page = find_get_page(&swapper_space, entry.val);
80         if (!page)
81                 return 0;
82         /*
83          * This function is called from scan_swap_map() and it's called
84          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
85          * We have to use trylock for avoiding deadlock. This is a special
86          * case and you should use try_to_free_swap() with explicit lock_page()
87          * in usual operations.
88          */
89         if (trylock_page(page)) {
90                 ret = try_to_free_swap(page);
91                 unlock_page(page);
92         }
93         page_cache_release(page);
94         return ret;
95 }
96
97 /*
98  * swapon tell device that all the old swap contents can be discarded,
99  * to allow the swap device to optimize its wear-levelling.
100  */
101 static int discard_swap(struct swap_info_struct *si)
102 {
103         struct swap_extent *se;
104         sector_t start_block;
105         sector_t nr_blocks;
106         int err = 0;
107
108         /* Do not discard the swap header page! */
109         se = &si->first_swap_extent;
110         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
111         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
112         if (nr_blocks) {
113                 err = blkdev_issue_discard(si->bdev, start_block,
114                                 nr_blocks, GFP_KERNEL, 0);
115                 if (err)
116                         return err;
117                 cond_resched();
118         }
119
120         list_for_each_entry(se, &si->first_swap_extent.list, list) {
121                 start_block = se->start_block << (PAGE_SHIFT - 9);
122                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
123
124                 err = blkdev_issue_discard(si->bdev, start_block,
125                                 nr_blocks, GFP_KERNEL, 0);
126                 if (err)
127                         break;
128
129                 cond_resched();
130         }
131         return err;             /* That will often be -EOPNOTSUPP */
132 }
133
134 /*
135  * swap allocation tell device that a cluster of swap can now be discarded,
136  * to allow the swap device to optimize its wear-levelling.
137  */
138 static void discard_swap_cluster(struct swap_info_struct *si,
139                                  pgoff_t start_page, pgoff_t nr_pages)
140 {
141         struct swap_extent *se = si->curr_swap_extent;
142         int found_extent = 0;
143
144         while (nr_pages) {
145                 struct list_head *lh;
146
147                 if (se->start_page <= start_page &&
148                     start_page < se->start_page + se->nr_pages) {
149                         pgoff_t offset = start_page - se->start_page;
150                         sector_t start_block = se->start_block + offset;
151                         sector_t nr_blocks = se->nr_pages - offset;
152
153                         if (nr_blocks > nr_pages)
154                                 nr_blocks = nr_pages;
155                         start_page += nr_blocks;
156                         nr_pages -= nr_blocks;
157
158                         if (!found_extent++)
159                                 si->curr_swap_extent = se;
160
161                         start_block <<= PAGE_SHIFT - 9;
162                         nr_blocks <<= PAGE_SHIFT - 9;
163                         if (blkdev_issue_discard(si->bdev, start_block,
164                                     nr_blocks, GFP_NOIO, 0))
165                                 break;
166                 }
167
168                 lh = se->list.next;
169                 se = list_entry(lh, struct swap_extent, list);
170         }
171 }
172
173 static int wait_for_discard(void *word)
174 {
175         schedule();
176         return 0;
177 }
178
179 #define SWAPFILE_CLUSTER        256
180 #define LATENCY_LIMIT           256
181
182 static unsigned long scan_swap_map(struct swap_info_struct *si,
183                                    unsigned char usage)
184 {
185         unsigned long offset;
186         unsigned long scan_base;
187         unsigned long last_in_cluster = 0;
188         int latency_ration = LATENCY_LIMIT;
189         int found_free_cluster = 0;
190
191         /*
192          * We try to cluster swap pages by allocating them sequentially
193          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
194          * way, however, we resort to first-free allocation, starting
195          * a new cluster.  This prevents us from scattering swap pages
196          * all over the entire swap partition, so that we reduce
197          * overall disk seek times between swap pages.  -- sct
198          * But we do now try to find an empty cluster.  -Andrea
199          * And we let swap pages go all over an SSD partition.  Hugh
200          */
201
202         si->flags += SWP_SCANNING;
203         scan_base = offset = si->cluster_next;
204
205         if (unlikely(!si->cluster_nr--)) {
206                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
207                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
208                         goto checks;
209                 }
210                 if (si->flags & SWP_DISCARDABLE) {
211                         /*
212                          * Start range check on racing allocations, in case
213                          * they overlap the cluster we eventually decide on
214                          * (we scan without swap_lock to allow preemption).
215                          * It's hardly conceivable that cluster_nr could be
216                          * wrapped during our scan, but don't depend on it.
217                          */
218                         if (si->lowest_alloc)
219                                 goto checks;
220                         si->lowest_alloc = si->max;
221                         si->highest_alloc = 0;
222                 }
223                 spin_unlock(&swap_lock);
224
225                 /*
226                  * If seek is expensive, start searching for new cluster from
227                  * start of partition, to minimize the span of allocated swap.
228                  * But if seek is cheap, search from our current position, so
229                  * that swap is allocated from all over the partition: if the
230                  * Flash Translation Layer only remaps within limited zones,
231                  * we don't want to wear out the first zone too quickly.
232                  */
233                 if (!(si->flags & SWP_SOLIDSTATE))
234                         scan_base = offset = si->lowest_bit;
235                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
236
237                 /* Locate the first empty (unaligned) cluster */
238                 for (; last_in_cluster <= si->highest_bit; offset++) {
239                         if (si->swap_map[offset])
240                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
241                         else if (offset == last_in_cluster) {
242                                 spin_lock(&swap_lock);
243                                 offset -= SWAPFILE_CLUSTER - 1;
244                                 si->cluster_next = offset;
245                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
246                                 found_free_cluster = 1;
247                                 goto checks;
248                         }
249                         if (unlikely(--latency_ration < 0)) {
250                                 cond_resched();
251                                 latency_ration = LATENCY_LIMIT;
252                         }
253                 }
254
255                 offset = si->lowest_bit;
256                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
257
258                 /* Locate the first empty (unaligned) cluster */
259                 for (; last_in_cluster < scan_base; offset++) {
260                         if (si->swap_map[offset])
261                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
262                         else if (offset == last_in_cluster) {
263                                 spin_lock(&swap_lock);
264                                 offset -= SWAPFILE_CLUSTER - 1;
265                                 si->cluster_next = offset;
266                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
267                                 found_free_cluster = 1;
268                                 goto checks;
269                         }
270                         if (unlikely(--latency_ration < 0)) {
271                                 cond_resched();
272                                 latency_ration = LATENCY_LIMIT;
273                         }
274                 }
275
276                 offset = scan_base;
277                 spin_lock(&swap_lock);
278                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
279                 si->lowest_alloc = 0;
280         }
281
282 checks:
283         if (!(si->flags & SWP_WRITEOK))
284                 goto no_page;
285         if (!si->highest_bit)
286                 goto no_page;
287         if (offset > si->highest_bit)
288                 scan_base = offset = si->lowest_bit;
289
290         /* reuse swap entry of cache-only swap if not busy. */
291         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
292                 int swap_was_freed;
293                 spin_unlock(&swap_lock);
294                 swap_was_freed = __try_to_reclaim_swap(si, offset);
295                 spin_lock(&swap_lock);
296                 /* entry was freed successfully, try to use this again */
297                 if (swap_was_freed)
298                         goto checks;
299                 goto scan; /* check next one */
300         }
301
302         if (si->swap_map[offset])
303                 goto scan;
304
305         if (offset == si->lowest_bit)
306                 si->lowest_bit++;
307         if (offset == si->highest_bit)
308                 si->highest_bit--;
309         si->inuse_pages++;
310         if (si->inuse_pages == si->pages) {
311                 si->lowest_bit = si->max;
312                 si->highest_bit = 0;
313         }
314         si->swap_map[offset] = usage;
315         si->cluster_next = offset + 1;
316         si->flags -= SWP_SCANNING;
317
318         if (si->lowest_alloc) {
319                 /*
320                  * Only set when SWP_DISCARDABLE, and there's a scan
321                  * for a free cluster in progress or just completed.
322                  */
323                 if (found_free_cluster) {
324                         /*
325                          * To optimize wear-levelling, discard the
326                          * old data of the cluster, taking care not to
327                          * discard any of its pages that have already
328                          * been allocated by racing tasks (offset has
329                          * already stepped over any at the beginning).
330                          */
331                         if (offset < si->highest_alloc &&
332                             si->lowest_alloc <= last_in_cluster)
333                                 last_in_cluster = si->lowest_alloc - 1;
334                         si->flags |= SWP_DISCARDING;
335                         spin_unlock(&swap_lock);
336
337                         if (offset < last_in_cluster)
338                                 discard_swap_cluster(si, offset,
339                                         last_in_cluster - offset + 1);
340
341                         spin_lock(&swap_lock);
342                         si->lowest_alloc = 0;
343                         si->flags &= ~SWP_DISCARDING;
344
345                         smp_mb();       /* wake_up_bit advises this */
346                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
347
348                 } else if (si->flags & SWP_DISCARDING) {
349                         /*
350                          * Delay using pages allocated by racing tasks
351                          * until the whole discard has been issued. We
352                          * could defer that delay until swap_writepage,
353                          * but it's easier to keep this self-contained.
354                          */
355                         spin_unlock(&swap_lock);
356                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
357                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
358                         spin_lock(&swap_lock);
359                 } else {
360                         /*
361                          * Note pages allocated by racing tasks while
362                          * scan for a free cluster is in progress, so
363                          * that its final discard can exclude them.
364                          */
365                         if (offset < si->lowest_alloc)
366                                 si->lowest_alloc = offset;
367                         if (offset > si->highest_alloc)
368                                 si->highest_alloc = offset;
369                 }
370         }
371         return offset;
372
373 scan:
374         spin_unlock(&swap_lock);
375         while (++offset <= si->highest_bit) {
376                 if (!si->swap_map[offset]) {
377                         spin_lock(&swap_lock);
378                         goto checks;
379                 }
380                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
381                         spin_lock(&swap_lock);
382                         goto checks;
383                 }
384                 if (unlikely(--latency_ration < 0)) {
385                         cond_resched();
386                         latency_ration = LATENCY_LIMIT;
387                 }
388         }
389         offset = si->lowest_bit;
390         while (++offset < scan_base) {
391                 if (!si->swap_map[offset]) {
392                         spin_lock(&swap_lock);
393                         goto checks;
394                 }
395                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
396                         spin_lock(&swap_lock);
397                         goto checks;
398                 }
399                 if (unlikely(--latency_ration < 0)) {
400                         cond_resched();
401                         latency_ration = LATENCY_LIMIT;
402                 }
403         }
404         spin_lock(&swap_lock);
405
406 no_page:
407         si->flags -= SWP_SCANNING;
408         return 0;
409 }
410
411 swp_entry_t get_swap_page(void)
412 {
413         struct swap_info_struct *si;
414         pgoff_t offset;
415         int type, next;
416         int wrapped = 0;
417
418         spin_lock(&swap_lock);
419         if (nr_swap_pages <= 0)
420                 goto noswap;
421         nr_swap_pages--;
422
423         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
424                 si = swap_info[type];
425                 next = si->next;
426                 if (next < 0 ||
427                     (!wrapped && si->prio != swap_info[next]->prio)) {
428                         next = swap_list.head;
429                         wrapped++;
430                 }
431
432                 if (!si->highest_bit)
433                         continue;
434                 if (!(si->flags & SWP_WRITEOK))
435                         continue;
436
437                 swap_list.next = next;
438                 /* This is called for allocating swap entry for cache */
439                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
440                 if (offset) {
441                         spin_unlock(&swap_lock);
442                         return swp_entry(type, offset);
443                 }
444                 next = swap_list.next;
445         }
446
447         nr_swap_pages++;
448 noswap:
449         spin_unlock(&swap_lock);
450         return (swp_entry_t) {0};
451 }
452
453 /* The only caller of this function is now susupend routine */
454 swp_entry_t get_swap_page_of_type(int type)
455 {
456         struct swap_info_struct *si;
457         pgoff_t offset;
458
459         spin_lock(&swap_lock);
460         si = swap_info[type];
461         if (si && (si->flags & SWP_WRITEOK)) {
462                 nr_swap_pages--;
463                 /* This is called for allocating swap entry, not cache */
464                 offset = scan_swap_map(si, 1);
465                 if (offset) {
466                         spin_unlock(&swap_lock);
467                         return swp_entry(type, offset);
468                 }
469                 nr_swap_pages++;
470         }
471         spin_unlock(&swap_lock);
472         return (swp_entry_t) {0};
473 }
474
475 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
476 {
477         struct swap_info_struct *p;
478         unsigned long offset, type;
479
480         if (!entry.val)
481                 goto out;
482         type = swp_type(entry);
483         if (type >= nr_swapfiles)
484                 goto bad_nofile;
485         p = swap_info[type];
486         if (!(p->flags & SWP_USED))
487                 goto bad_device;
488         offset = swp_offset(entry);
489         if (offset >= p->max)
490                 goto bad_offset;
491         if (!p->swap_map[offset])
492                 goto bad_free;
493         spin_lock(&swap_lock);
494         return p;
495
496 bad_free:
497         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
498         goto out;
499 bad_offset:
500         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
501         goto out;
502 bad_device:
503         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
504         goto out;
505 bad_nofile:
506         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
507 out:
508         return NULL;
509 }
510
511 static unsigned char swap_entry_free(struct swap_info_struct *p,
512                                      swp_entry_t entry, unsigned char usage)
513 {
514         unsigned long offset = swp_offset(entry);
515         unsigned char count;
516         unsigned char has_cache;
517
518         count = p->swap_map[offset];
519         has_cache = count & SWAP_HAS_CACHE;
520         count &= ~SWAP_HAS_CACHE;
521
522         if (usage == SWAP_HAS_CACHE) {
523                 VM_BUG_ON(!has_cache);
524                 has_cache = 0;
525         } else if (count == SWAP_MAP_SHMEM) {
526                 /*
527                  * Or we could insist on shmem.c using a special
528                  * swap_shmem_free() and free_shmem_swap_and_cache()...
529                  */
530                 count = 0;
531         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
532                 if (count == COUNT_CONTINUED) {
533                         if (swap_count_continued(p, offset, count))
534                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
535                         else
536                                 count = SWAP_MAP_MAX;
537                 } else
538                         count--;
539         }
540
541         if (!count)
542                 mem_cgroup_uncharge_swap(entry);
543
544         usage = count | has_cache;
545         p->swap_map[offset] = usage;
546
547         /* free if no reference */
548         if (!usage) {
549                 struct gendisk *disk = p->bdev->bd_disk;
550                 if (offset < p->lowest_bit)
551                         p->lowest_bit = offset;
552                 if (offset > p->highest_bit)
553                         p->highest_bit = offset;
554                 if (swap_list.next >= 0 &&
555                     p->prio > swap_info[swap_list.next]->prio)
556                         swap_list.next = p->type;
557                 nr_swap_pages++;
558                 p->inuse_pages--;
559                 if ((p->flags & SWP_BLKDEV) &&
560                                 disk->fops->swap_slot_free_notify)
561                         disk->fops->swap_slot_free_notify(p->bdev, offset);
562         }
563
564         return usage;
565 }
566
567 /*
568  * Caller has made sure that the swapdevice corresponding to entry
569  * is still around or has not been recycled.
570  */
571 void swap_free(swp_entry_t entry)
572 {
573         struct swap_info_struct *p;
574
575         p = swap_info_get(entry);
576         if (p) {
577                 swap_entry_free(p, entry, 1);
578                 spin_unlock(&swap_lock);
579         }
580 }
581
582 /*
583  * Called after dropping swapcache to decrease refcnt to swap entries.
584  */
585 void swapcache_free(swp_entry_t entry, struct page *page)
586 {
587         struct swap_info_struct *p;
588         unsigned char count;
589
590         p = swap_info_get(entry);
591         if (p) {
592                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
593                 if (page)
594                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
595                 spin_unlock(&swap_lock);
596         }
597 }
598
599 /*
600  * How many references to page are currently swapped out?
601  * This does not give an exact answer when swap count is continued,
602  * but does include the high COUNT_CONTINUED flag to allow for that.
603  */
604 static inline int page_swapcount(struct page *page)
605 {
606         int count = 0;
607         struct swap_info_struct *p;
608         swp_entry_t entry;
609
610         entry.val = page_private(page);
611         p = swap_info_get(entry);
612         if (p) {
613                 count = swap_count(p->swap_map[swp_offset(entry)]);
614                 spin_unlock(&swap_lock);
615         }
616         return count;
617 }
618
619 /*
620  * We can write to an anon page without COW if there are no other references
621  * to it.  And as a side-effect, free up its swap: because the old content
622  * on disk will never be read, and seeking back there to write new content
623  * later would only waste time away from clustering.
624  */
625 int reuse_swap_page(struct page *page)
626 {
627         int count;
628
629         VM_BUG_ON(!PageLocked(page));
630         if (unlikely(PageKsm(page)))
631                 return 0;
632         count = page_mapcount(page);
633         if (count <= 1 && PageSwapCache(page)) {
634                 count += page_swapcount(page);
635                 if (count == 1 && !PageWriteback(page)) {
636                         delete_from_swap_cache(page);
637                         SetPageDirty(page);
638                 }
639         }
640         return count <= 1;
641 }
642
643 /*
644  * If swap is getting full, or if there are no more mappings of this page,
645  * then try_to_free_swap is called to free its swap space.
646  */
647 int try_to_free_swap(struct page *page)
648 {
649         VM_BUG_ON(!PageLocked(page));
650
651         if (!PageSwapCache(page))
652                 return 0;
653         if (PageWriteback(page))
654                 return 0;
655         if (page_swapcount(page))
656                 return 0;
657
658         /*
659          * Once hibernation has begun to create its image of memory,
660          * there's a danger that one of the calls to try_to_free_swap()
661          * - most probably a call from __try_to_reclaim_swap() while
662          * hibernation is allocating its own swap pages for the image,
663          * but conceivably even a call from memory reclaim - will free
664          * the swap from a page which has already been recorded in the
665          * image as a clean swapcache page, and then reuse its swap for
666          * another page of the image.  On waking from hibernation, the
667          * original page might be freed under memory pressure, then
668          * later read back in from swap, now with the wrong data.
669          *
670          * Hibernation clears bits from gfp_allowed_mask to prevent
671          * memory reclaim from writing to disk, so check that here.
672          */
673         if (!(gfp_allowed_mask & __GFP_IO))
674                 return 0;
675
676         delete_from_swap_cache(page);
677         SetPageDirty(page);
678         return 1;
679 }
680
681 /*
682  * Free the swap entry like above, but also try to
683  * free the page cache entry if it is the last user.
684  */
685 int free_swap_and_cache(swp_entry_t entry)
686 {
687         struct swap_info_struct *p;
688         struct page *page = NULL;
689
690         if (non_swap_entry(entry))
691                 return 1;
692
693         p = swap_info_get(entry);
694         if (p) {
695                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
696                         page = find_get_page(&swapper_space, entry.val);
697                         if (page && !trylock_page(page)) {
698                                 page_cache_release(page);
699                                 page = NULL;
700                         }
701                 }
702                 spin_unlock(&swap_lock);
703         }
704         if (page) {
705                 /*
706                  * Not mapped elsewhere, or swap space full? Free it!
707                  * Also recheck PageSwapCache now page is locked (above).
708                  */
709                 if (PageSwapCache(page) && !PageWriteback(page) &&
710                                 (!page_mapped(page) || vm_swap_full())) {
711                         delete_from_swap_cache(page);
712                         SetPageDirty(page);
713                 }
714                 unlock_page(page);
715                 page_cache_release(page);
716         }
717         return p != NULL;
718 }
719
720 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
721 /**
722  * mem_cgroup_count_swap_user - count the user of a swap entry
723  * @ent: the swap entry to be checked
724  * @pagep: the pointer for the swap cache page of the entry to be stored
725  *
726  * Returns the number of the user of the swap entry. The number is valid only
727  * for swaps of anonymous pages.
728  * If the entry is found on swap cache, the page is stored to pagep with
729  * refcount of it being incremented.
730  */
731 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
732 {
733         struct page *page;
734         struct swap_info_struct *p;
735         int count = 0;
736
737         page = find_get_page(&swapper_space, ent.val);
738         if (page)
739                 count += page_mapcount(page);
740         p = swap_info_get(ent);
741         if (p) {
742                 count += swap_count(p->swap_map[swp_offset(ent)]);
743                 spin_unlock(&swap_lock);
744         }
745
746         *pagep = page;
747         return count;
748 }
749 #endif
750
751 #ifdef CONFIG_HIBERNATION
752 /*
753  * Find the swap type that corresponds to given device (if any).
754  *
755  * @offset - number of the PAGE_SIZE-sized block of the device, starting
756  * from 0, in which the swap header is expected to be located.
757  *
758  * This is needed for the suspend to disk (aka swsusp).
759  */
760 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
761 {
762         struct block_device *bdev = NULL;
763         int type;
764
765         if (device)
766                 bdev = bdget(device);
767
768         spin_lock(&swap_lock);
769         for (type = 0; type < nr_swapfiles; type++) {
770                 struct swap_info_struct *sis = swap_info[type];
771
772                 if (!(sis->flags & SWP_WRITEOK))
773                         continue;
774
775                 if (!bdev) {
776                         if (bdev_p)
777                                 *bdev_p = bdgrab(sis->bdev);
778
779                         spin_unlock(&swap_lock);
780                         return type;
781                 }
782                 if (bdev == sis->bdev) {
783                         struct swap_extent *se = &sis->first_swap_extent;
784
785                         if (se->start_block == offset) {
786                                 if (bdev_p)
787                                         *bdev_p = bdgrab(sis->bdev);
788
789                                 spin_unlock(&swap_lock);
790                                 bdput(bdev);
791                                 return type;
792                         }
793                 }
794         }
795         spin_unlock(&swap_lock);
796         if (bdev)
797                 bdput(bdev);
798
799         return -ENODEV;
800 }
801
802 /*
803  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
804  * corresponding to given index in swap_info (swap type).
805  */
806 sector_t swapdev_block(int type, pgoff_t offset)
807 {
808         struct block_device *bdev;
809
810         if ((unsigned int)type >= nr_swapfiles)
811                 return 0;
812         if (!(swap_info[type]->flags & SWP_WRITEOK))
813                 return 0;
814         return map_swap_entry(swp_entry(type, offset), &bdev);
815 }
816
817 /*
818  * Return either the total number of swap pages of given type, or the number
819  * of free pages of that type (depending on @free)
820  *
821  * This is needed for software suspend
822  */
823 unsigned int count_swap_pages(int type, int free)
824 {
825         unsigned int n = 0;
826
827         spin_lock(&swap_lock);
828         if ((unsigned int)type < nr_swapfiles) {
829                 struct swap_info_struct *sis = swap_info[type];
830
831                 if (sis->flags & SWP_WRITEOK) {
832                         n = sis->pages;
833                         if (free)
834                                 n -= sis->inuse_pages;
835                 }
836         }
837         spin_unlock(&swap_lock);
838         return n;
839 }
840 #endif /* CONFIG_HIBERNATION */
841
842 /*
843  * No need to decide whether this PTE shares the swap entry with others,
844  * just let do_wp_page work it out if a write is requested later - to
845  * force COW, vm_page_prot omits write permission from any private vma.
846  */
847 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
848                 unsigned long addr, swp_entry_t entry, struct page *page)
849 {
850         struct mem_cgroup *ptr;
851         spinlock_t *ptl;
852         pte_t *pte;
853         int ret = 1;
854
855         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
856                 ret = -ENOMEM;
857                 goto out_nolock;
858         }
859
860         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
861         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
862                 if (ret > 0)
863                         mem_cgroup_cancel_charge_swapin(ptr);
864                 ret = 0;
865                 goto out;
866         }
867
868         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
869         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
870         get_page(page);
871         set_pte_at(vma->vm_mm, addr, pte,
872                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
873         page_add_anon_rmap(page, vma, addr);
874         mem_cgroup_commit_charge_swapin(page, ptr);
875         swap_free(entry);
876         /*
877          * Move the page to the active list so it is not
878          * immediately swapped out again after swapon.
879          */
880         activate_page(page);
881 out:
882         pte_unmap_unlock(pte, ptl);
883 out_nolock:
884         return ret;
885 }
886
887 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
888                                 unsigned long addr, unsigned long end,
889                                 swp_entry_t entry, struct page *page)
890 {
891         pte_t swp_pte = swp_entry_to_pte(entry);
892         pte_t *pte;
893         int ret = 0;
894
895         /*
896          * We don't actually need pte lock while scanning for swp_pte: since
897          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
898          * page table while we're scanning; though it could get zapped, and on
899          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
900          * of unmatched parts which look like swp_pte, so unuse_pte must
901          * recheck under pte lock.  Scanning without pte lock lets it be
902          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
903          */
904         pte = pte_offset_map(pmd, addr);
905         do {
906                 /*
907                  * swapoff spends a _lot_ of time in this loop!
908                  * Test inline before going to call unuse_pte.
909                  */
910                 if (unlikely(pte_same(*pte, swp_pte))) {
911                         pte_unmap(pte);
912                         ret = unuse_pte(vma, pmd, addr, entry, page);
913                         if (ret)
914                                 goto out;
915                         pte = pte_offset_map(pmd, addr);
916                 }
917         } while (pte++, addr += PAGE_SIZE, addr != end);
918         pte_unmap(pte - 1);
919 out:
920         return ret;
921 }
922
923 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
924                                 unsigned long addr, unsigned long end,
925                                 swp_entry_t entry, struct page *page)
926 {
927         pmd_t *pmd;
928         unsigned long next;
929         int ret;
930
931         pmd = pmd_offset(pud, addr);
932         do {
933                 next = pmd_addr_end(addr, end);
934                 if (unlikely(pmd_trans_huge(*pmd)))
935                         continue;
936                 if (pmd_none_or_clear_bad(pmd))
937                         continue;
938                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
939                 if (ret)
940                         return ret;
941         } while (pmd++, addr = next, addr != end);
942         return 0;
943 }
944
945 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
946                                 unsigned long addr, unsigned long end,
947                                 swp_entry_t entry, struct page *page)
948 {
949         pud_t *pud;
950         unsigned long next;
951         int ret;
952
953         pud = pud_offset(pgd, addr);
954         do {
955                 next = pud_addr_end(addr, end);
956                 if (pud_none_or_clear_bad(pud))
957                         continue;
958                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
959                 if (ret)
960                         return ret;
961         } while (pud++, addr = next, addr != end);
962         return 0;
963 }
964
965 static int unuse_vma(struct vm_area_struct *vma,
966                                 swp_entry_t entry, struct page *page)
967 {
968         pgd_t *pgd;
969         unsigned long addr, end, next;
970         int ret;
971
972         if (page_anon_vma(page)) {
973                 addr = page_address_in_vma(page, vma);
974                 if (addr == -EFAULT)
975                         return 0;
976                 else
977                         end = addr + PAGE_SIZE;
978         } else {
979                 addr = vma->vm_start;
980                 end = vma->vm_end;
981         }
982
983         pgd = pgd_offset(vma->vm_mm, addr);
984         do {
985                 next = pgd_addr_end(addr, end);
986                 if (pgd_none_or_clear_bad(pgd))
987                         continue;
988                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
989                 if (ret)
990                         return ret;
991         } while (pgd++, addr = next, addr != end);
992         return 0;
993 }
994
995 static int unuse_mm(struct mm_struct *mm,
996                                 swp_entry_t entry, struct page *page)
997 {
998         struct vm_area_struct *vma;
999         int ret = 0;
1000
1001         if (!down_read_trylock(&mm->mmap_sem)) {
1002                 /*
1003                  * Activate page so shrink_inactive_list is unlikely to unmap
1004                  * its ptes while lock is dropped, so swapoff can make progress.
1005                  */
1006                 activate_page(page);
1007                 unlock_page(page);
1008                 down_read(&mm->mmap_sem);
1009                 lock_page(page);
1010         }
1011         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1012                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1013                         break;
1014         }
1015         up_read(&mm->mmap_sem);
1016         return (ret < 0)? ret: 0;
1017 }
1018
1019 /*
1020  * Scan swap_map from current position to next entry still in use.
1021  * Recycle to start on reaching the end, returning 0 when empty.
1022  */
1023 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1024                                         unsigned int prev)
1025 {
1026         unsigned int max = si->max;
1027         unsigned int i = prev;
1028         unsigned char count;
1029
1030         /*
1031          * No need for swap_lock here: we're just looking
1032          * for whether an entry is in use, not modifying it; false
1033          * hits are okay, and sys_swapoff() has already prevented new
1034          * allocations from this area (while holding swap_lock).
1035          */
1036         for (;;) {
1037                 if (++i >= max) {
1038                         if (!prev) {
1039                                 i = 0;
1040                                 break;
1041                         }
1042                         /*
1043                          * No entries in use at top of swap_map,
1044                          * loop back to start and recheck there.
1045                          */
1046                         max = prev + 1;
1047                         prev = 0;
1048                         i = 1;
1049                 }
1050                 count = si->swap_map[i];
1051                 if (count && swap_count(count) != SWAP_MAP_BAD)
1052                         break;
1053         }
1054         return i;
1055 }
1056
1057 /*
1058  * We completely avoid races by reading each swap page in advance,
1059  * and then search for the process using it.  All the necessary
1060  * page table adjustments can then be made atomically.
1061  */
1062 static int try_to_unuse(unsigned int type)
1063 {
1064         struct swap_info_struct *si = swap_info[type];
1065         struct mm_struct *start_mm;
1066         unsigned char *swap_map;
1067         unsigned char swcount;
1068         struct page *page;
1069         swp_entry_t entry;
1070         unsigned int i = 0;
1071         int retval = 0;
1072
1073         /*
1074          * When searching mms for an entry, a good strategy is to
1075          * start at the first mm we freed the previous entry from
1076          * (though actually we don't notice whether we or coincidence
1077          * freed the entry).  Initialize this start_mm with a hold.
1078          *
1079          * A simpler strategy would be to start at the last mm we
1080          * freed the previous entry from; but that would take less
1081          * advantage of mmlist ordering, which clusters forked mms
1082          * together, child after parent.  If we race with dup_mmap(), we
1083          * prefer to resolve parent before child, lest we miss entries
1084          * duplicated after we scanned child: using last mm would invert
1085          * that.
1086          */
1087         start_mm = &init_mm;
1088         atomic_inc(&init_mm.mm_users);
1089
1090         /*
1091          * Keep on scanning until all entries have gone.  Usually,
1092          * one pass through swap_map is enough, but not necessarily:
1093          * there are races when an instance of an entry might be missed.
1094          */
1095         while ((i = find_next_to_unuse(si, i)) != 0) {
1096                 if (signal_pending(current)) {
1097                         retval = -EINTR;
1098                         break;
1099                 }
1100
1101                 /*
1102                  * Get a page for the entry, using the existing swap
1103                  * cache page if there is one.  Otherwise, get a clean
1104                  * page and read the swap into it.
1105                  */
1106                 swap_map = &si->swap_map[i];
1107                 entry = swp_entry(type, i);
1108                 page = read_swap_cache_async(entry,
1109                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1110                 if (!page) {
1111                         /*
1112                          * Either swap_duplicate() failed because entry
1113                          * has been freed independently, and will not be
1114                          * reused since sys_swapoff() already disabled
1115                          * allocation from here, or alloc_page() failed.
1116                          */
1117                         if (!*swap_map)
1118                                 continue;
1119                         retval = -ENOMEM;
1120                         break;
1121                 }
1122
1123                 /*
1124                  * Don't hold on to start_mm if it looks like exiting.
1125                  */
1126                 if (atomic_read(&start_mm->mm_users) == 1) {
1127                         mmput(start_mm);
1128                         start_mm = &init_mm;
1129                         atomic_inc(&init_mm.mm_users);
1130                 }
1131
1132                 /*
1133                  * Wait for and lock page.  When do_swap_page races with
1134                  * try_to_unuse, do_swap_page can handle the fault much
1135                  * faster than try_to_unuse can locate the entry.  This
1136                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1137                  * defer to do_swap_page in such a case - in some tests,
1138                  * do_swap_page and try_to_unuse repeatedly compete.
1139                  */
1140                 wait_on_page_locked(page);
1141                 wait_on_page_writeback(page);
1142                 lock_page(page);
1143                 wait_on_page_writeback(page);
1144
1145                 /*
1146                  * Remove all references to entry.
1147                  */
1148                 swcount = *swap_map;
1149                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1150                         retval = shmem_unuse(entry, page);
1151                         /* page has already been unlocked and released */
1152                         if (retval < 0)
1153                                 break;
1154                         continue;
1155                 }
1156                 if (swap_count(swcount) && start_mm != &init_mm)
1157                         retval = unuse_mm(start_mm, entry, page);
1158
1159                 if (swap_count(*swap_map)) {
1160                         int set_start_mm = (*swap_map >= swcount);
1161                         struct list_head *p = &start_mm->mmlist;
1162                         struct mm_struct *new_start_mm = start_mm;
1163                         struct mm_struct *prev_mm = start_mm;
1164                         struct mm_struct *mm;
1165
1166                         atomic_inc(&new_start_mm->mm_users);
1167                         atomic_inc(&prev_mm->mm_users);
1168                         spin_lock(&mmlist_lock);
1169                         while (swap_count(*swap_map) && !retval &&
1170                                         (p = p->next) != &start_mm->mmlist) {
1171                                 mm = list_entry(p, struct mm_struct, mmlist);
1172                                 if (!atomic_inc_not_zero(&mm->mm_users))
1173                                         continue;
1174                                 spin_unlock(&mmlist_lock);
1175                                 mmput(prev_mm);
1176                                 prev_mm = mm;
1177
1178                                 cond_resched();
1179
1180                                 swcount = *swap_map;
1181                                 if (!swap_count(swcount)) /* any usage ? */
1182                                         ;
1183                                 else if (mm == &init_mm)
1184                                         set_start_mm = 1;
1185                                 else
1186                                         retval = unuse_mm(mm, entry, page);
1187
1188                                 if (set_start_mm && *swap_map < swcount) {
1189                                         mmput(new_start_mm);
1190                                         atomic_inc(&mm->mm_users);
1191                                         new_start_mm = mm;
1192                                         set_start_mm = 0;
1193                                 }
1194                                 spin_lock(&mmlist_lock);
1195                         }
1196                         spin_unlock(&mmlist_lock);
1197                         mmput(prev_mm);
1198                         mmput(start_mm);
1199                         start_mm = new_start_mm;
1200                 }
1201                 if (retval) {
1202                         unlock_page(page);
1203                         page_cache_release(page);
1204                         break;
1205                 }
1206
1207                 /*
1208                  * If a reference remains (rare), we would like to leave
1209                  * the page in the swap cache; but try_to_unmap could
1210                  * then re-duplicate the entry once we drop page lock,
1211                  * so we might loop indefinitely; also, that page could
1212                  * not be swapped out to other storage meanwhile.  So:
1213                  * delete from cache even if there's another reference,
1214                  * after ensuring that the data has been saved to disk -
1215                  * since if the reference remains (rarer), it will be
1216                  * read from disk into another page.  Splitting into two
1217                  * pages would be incorrect if swap supported "shared
1218                  * private" pages, but they are handled by tmpfs files.
1219                  *
1220                  * Given how unuse_vma() targets one particular offset
1221                  * in an anon_vma, once the anon_vma has been determined,
1222                  * this splitting happens to be just what is needed to
1223                  * handle where KSM pages have been swapped out: re-reading
1224                  * is unnecessarily slow, but we can fix that later on.
1225                  */
1226                 if (swap_count(*swap_map) &&
1227                      PageDirty(page) && PageSwapCache(page)) {
1228                         struct writeback_control wbc = {
1229                                 .sync_mode = WB_SYNC_NONE,
1230                         };
1231
1232                         swap_writepage(page, &wbc);
1233                         lock_page(page);
1234                         wait_on_page_writeback(page);
1235                 }
1236
1237                 /*
1238                  * It is conceivable that a racing task removed this page from
1239                  * swap cache just before we acquired the page lock at the top,
1240                  * or while we dropped it in unuse_mm().  The page might even
1241                  * be back in swap cache on another swap area: that we must not
1242                  * delete, since it may not have been written out to swap yet.
1243                  */
1244                 if (PageSwapCache(page) &&
1245                     likely(page_private(page) == entry.val))
1246                         delete_from_swap_cache(page);
1247
1248                 /*
1249                  * So we could skip searching mms once swap count went
1250                  * to 1, we did not mark any present ptes as dirty: must
1251                  * mark page dirty so shrink_page_list will preserve it.
1252                  */
1253                 SetPageDirty(page);
1254                 unlock_page(page);
1255                 page_cache_release(page);
1256
1257                 /*
1258                  * Make sure that we aren't completely killing
1259                  * interactive performance.
1260                  */
1261                 cond_resched();
1262         }
1263
1264         mmput(start_mm);
1265         return retval;
1266 }
1267
1268 /*
1269  * After a successful try_to_unuse, if no swap is now in use, we know
1270  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1271  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1272  * added to the mmlist just after page_duplicate - before would be racy.
1273  */
1274 static void drain_mmlist(void)
1275 {
1276         struct list_head *p, *next;
1277         unsigned int type;
1278
1279         for (type = 0; type < nr_swapfiles; type++)
1280                 if (swap_info[type]->inuse_pages)
1281                         return;
1282         spin_lock(&mmlist_lock);
1283         list_for_each_safe(p, next, &init_mm.mmlist)
1284                 list_del_init(p);
1285         spin_unlock(&mmlist_lock);
1286 }
1287
1288 /*
1289  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1290  * corresponds to page offset for the specified swap entry.
1291  * Note that the type of this function is sector_t, but it returns page offset
1292  * into the bdev, not sector offset.
1293  */
1294 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1295 {
1296         struct swap_info_struct *sis;
1297         struct swap_extent *start_se;
1298         struct swap_extent *se;
1299         pgoff_t offset;
1300
1301         sis = swap_info[swp_type(entry)];
1302         *bdev = sis->bdev;
1303
1304         offset = swp_offset(entry);
1305         start_se = sis->curr_swap_extent;
1306         se = start_se;
1307
1308         for ( ; ; ) {
1309                 struct list_head *lh;
1310
1311                 if (se->start_page <= offset &&
1312                                 offset < (se->start_page + se->nr_pages)) {
1313                         return se->start_block + (offset - se->start_page);
1314                 }
1315                 lh = se->list.next;
1316                 se = list_entry(lh, struct swap_extent, list);
1317                 sis->curr_swap_extent = se;
1318                 BUG_ON(se == start_se);         /* It *must* be present */
1319         }
1320 }
1321
1322 /*
1323  * Returns the page offset into bdev for the specified page's swap entry.
1324  */
1325 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1326 {
1327         swp_entry_t entry;
1328         entry.val = page_private(page);
1329         return map_swap_entry(entry, bdev);
1330 }
1331
1332 /*
1333  * Free all of a swapdev's extent information
1334  */
1335 static void destroy_swap_extents(struct swap_info_struct *sis)
1336 {
1337         while (!list_empty(&sis->first_swap_extent.list)) {
1338                 struct swap_extent *se;
1339
1340                 se = list_entry(sis->first_swap_extent.list.next,
1341                                 struct swap_extent, list);
1342                 list_del(&se->list);
1343                 kfree(se);
1344         }
1345 }
1346
1347 /*
1348  * Add a block range (and the corresponding page range) into this swapdev's
1349  * extent list.  The extent list is kept sorted in page order.
1350  *
1351  * This function rather assumes that it is called in ascending page order.
1352  */
1353 static int
1354 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1355                 unsigned long nr_pages, sector_t start_block)
1356 {
1357         struct swap_extent *se;
1358         struct swap_extent *new_se;
1359         struct list_head *lh;
1360
1361         if (start_page == 0) {
1362                 se = &sis->first_swap_extent;
1363                 sis->curr_swap_extent = se;
1364                 se->start_page = 0;
1365                 se->nr_pages = nr_pages;
1366                 se->start_block = start_block;
1367                 return 1;
1368         } else {
1369                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1370                 se = list_entry(lh, struct swap_extent, list);
1371                 BUG_ON(se->start_page + se->nr_pages != start_page);
1372                 if (se->start_block + se->nr_pages == start_block) {
1373                         /* Merge it */
1374                         se->nr_pages += nr_pages;
1375                         return 0;
1376                 }
1377         }
1378
1379         /*
1380          * No merge.  Insert a new extent, preserving ordering.
1381          */
1382         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1383         if (new_se == NULL)
1384                 return -ENOMEM;
1385         new_se->start_page = start_page;
1386         new_se->nr_pages = nr_pages;
1387         new_se->start_block = start_block;
1388
1389         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1390         return 1;
1391 }
1392
1393 /*
1394  * A `swap extent' is a simple thing which maps a contiguous range of pages
1395  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1396  * is built at swapon time and is then used at swap_writepage/swap_readpage
1397  * time for locating where on disk a page belongs.
1398  *
1399  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1400  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1401  * swap files identically.
1402  *
1403  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1404  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1405  * swapfiles are handled *identically* after swapon time.
1406  *
1407  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1408  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1409  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1410  * requirements, they are simply tossed out - we will never use those blocks
1411  * for swapping.
1412  *
1413  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1414  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1415  * which will scribble on the fs.
1416  *
1417  * The amount of disk space which a single swap extent represents varies.
1418  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1419  * extents in the list.  To avoid much list walking, we cache the previous
1420  * search location in `curr_swap_extent', and start new searches from there.
1421  * This is extremely effective.  The average number of iterations in
1422  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1423  */
1424 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1425 {
1426         struct inode *inode;
1427         unsigned blocks_per_page;
1428         unsigned long page_no;
1429         unsigned blkbits;
1430         sector_t probe_block;
1431         sector_t last_block;
1432         sector_t lowest_block = -1;
1433         sector_t highest_block = 0;
1434         int nr_extents = 0;
1435         int ret;
1436
1437         inode = sis->swap_file->f_mapping->host;
1438         if (S_ISBLK(inode->i_mode)) {
1439                 ret = add_swap_extent(sis, 0, sis->max, 0);
1440                 *span = sis->pages;
1441                 goto out;
1442         }
1443
1444         blkbits = inode->i_blkbits;
1445         blocks_per_page = PAGE_SIZE >> blkbits;
1446
1447         /*
1448          * Map all the blocks into the extent list.  This code doesn't try
1449          * to be very smart.
1450          */
1451         probe_block = 0;
1452         page_no = 0;
1453         last_block = i_size_read(inode) >> blkbits;
1454         while ((probe_block + blocks_per_page) <= last_block &&
1455                         page_no < sis->max) {
1456                 unsigned block_in_page;
1457                 sector_t first_block;
1458
1459                 first_block = bmap(inode, probe_block);
1460                 if (first_block == 0)
1461                         goto bad_bmap;
1462
1463                 /*
1464                  * It must be PAGE_SIZE aligned on-disk
1465                  */
1466                 if (first_block & (blocks_per_page - 1)) {
1467                         probe_block++;
1468                         goto reprobe;
1469                 }
1470
1471                 for (block_in_page = 1; block_in_page < blocks_per_page;
1472                                         block_in_page++) {
1473                         sector_t block;
1474
1475                         block = bmap(inode, probe_block + block_in_page);
1476                         if (block == 0)
1477                                 goto bad_bmap;
1478                         if (block != first_block + block_in_page) {
1479                                 /* Discontiguity */
1480                                 probe_block++;
1481                                 goto reprobe;
1482                         }
1483                 }
1484
1485                 first_block >>= (PAGE_SHIFT - blkbits);
1486                 if (page_no) {  /* exclude the header page */
1487                         if (first_block < lowest_block)
1488                                 lowest_block = first_block;
1489                         if (first_block > highest_block)
1490                                 highest_block = first_block;
1491                 }
1492
1493                 /*
1494                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1495                  */
1496                 ret = add_swap_extent(sis, page_no, 1, first_block);
1497                 if (ret < 0)
1498                         goto out;
1499                 nr_extents += ret;
1500                 page_no++;
1501                 probe_block += blocks_per_page;
1502 reprobe:
1503                 continue;
1504         }
1505         ret = nr_extents;
1506         *span = 1 + highest_block - lowest_block;
1507         if (page_no == 0)
1508                 page_no = 1;    /* force Empty message */
1509         sis->max = page_no;
1510         sis->pages = page_no - 1;
1511         sis->highest_bit = page_no - 1;
1512 out:
1513         return ret;
1514 bad_bmap:
1515         printk(KERN_ERR "swapon: swapfile has holes\n");
1516         ret = -EINVAL;
1517         goto out;
1518 }
1519
1520 static void enable_swap_info(struct swap_info_struct *p, int prio,
1521                                 unsigned char *swap_map)
1522 {
1523         int i, prev;
1524
1525         spin_lock(&swap_lock);
1526         if (prio >= 0)
1527                 p->prio = prio;
1528         else
1529                 p->prio = --least_priority;
1530         p->swap_map = swap_map;
1531         p->flags |= SWP_WRITEOK;
1532         nr_swap_pages += p->pages;
1533         total_swap_pages += p->pages;
1534
1535         /* insert swap space into swap_list: */
1536         prev = -1;
1537         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1538                 if (p->prio >= swap_info[i]->prio)
1539                         break;
1540                 prev = i;
1541         }
1542         p->next = i;
1543         if (prev < 0)
1544                 swap_list.head = swap_list.next = p->type;
1545         else
1546                 swap_info[prev]->next = p->type;
1547         spin_unlock(&swap_lock);
1548 }
1549
1550 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1551 {
1552         struct swap_info_struct *p = NULL;
1553         unsigned char *swap_map;
1554         struct file *swap_file, *victim;
1555         struct address_space *mapping;
1556         struct inode *inode;
1557         char *pathname;
1558         int oom_score_adj;
1559         int i, type, prev;
1560         int err;
1561
1562         if (!capable(CAP_SYS_ADMIN))
1563                 return -EPERM;
1564
1565         pathname = getname(specialfile);
1566         err = PTR_ERR(pathname);
1567         if (IS_ERR(pathname))
1568                 goto out;
1569
1570         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1571         putname(pathname);
1572         err = PTR_ERR(victim);
1573         if (IS_ERR(victim))
1574                 goto out;
1575
1576         mapping = victim->f_mapping;
1577         prev = -1;
1578         spin_lock(&swap_lock);
1579         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1580                 p = swap_info[type];
1581                 if (p->flags & SWP_WRITEOK) {
1582                         if (p->swap_file->f_mapping == mapping)
1583                                 break;
1584                 }
1585                 prev = type;
1586         }
1587         if (type < 0) {
1588                 err = -EINVAL;
1589                 spin_unlock(&swap_lock);
1590                 goto out_dput;
1591         }
1592         if (!security_vm_enough_memory(p->pages))
1593                 vm_unacct_memory(p->pages);
1594         else {
1595                 err = -ENOMEM;
1596                 spin_unlock(&swap_lock);
1597                 goto out_dput;
1598         }
1599         if (prev < 0)
1600                 swap_list.head = p->next;
1601         else
1602                 swap_info[prev]->next = p->next;
1603         if (type == swap_list.next) {
1604                 /* just pick something that's safe... */
1605                 swap_list.next = swap_list.head;
1606         }
1607         if (p->prio < 0) {
1608                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1609                         swap_info[i]->prio = p->prio--;
1610                 least_priority++;
1611         }
1612         nr_swap_pages -= p->pages;
1613         total_swap_pages -= p->pages;
1614         p->flags &= ~SWP_WRITEOK;
1615         spin_unlock(&swap_lock);
1616
1617         oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1618         err = try_to_unuse(type);
1619         compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj);
1620
1621         if (err) {
1622                 /*
1623                  * reading p->prio and p->swap_map outside the lock is
1624                  * safe here because only sys_swapon and sys_swapoff
1625                  * change them, and there can be no other sys_swapon or
1626                  * sys_swapoff for this swap_info_struct at this point.
1627                  */
1628                 /* re-insert swap space back into swap_list */
1629                 enable_swap_info(p, p->prio, p->swap_map);
1630                 goto out_dput;
1631         }
1632
1633         destroy_swap_extents(p);
1634         if (p->flags & SWP_CONTINUED)
1635                 free_swap_count_continuations(p);
1636
1637         mutex_lock(&swapon_mutex);
1638         spin_lock(&swap_lock);
1639         drain_mmlist();
1640
1641         /* wait for anyone still in scan_swap_map */
1642         p->highest_bit = 0;             /* cuts scans short */
1643         while (p->flags >= SWP_SCANNING) {
1644                 spin_unlock(&swap_lock);
1645                 schedule_timeout_uninterruptible(1);
1646                 spin_lock(&swap_lock);
1647         }
1648
1649         swap_file = p->swap_file;
1650         p->swap_file = NULL;
1651         p->max = 0;
1652         swap_map = p->swap_map;
1653         p->swap_map = NULL;
1654         p->flags = 0;
1655         spin_unlock(&swap_lock);
1656         mutex_unlock(&swapon_mutex);
1657         vfree(swap_map);
1658         /* Destroy swap account informatin */
1659         swap_cgroup_swapoff(type);
1660
1661         inode = mapping->host;
1662         if (S_ISBLK(inode->i_mode)) {
1663                 struct block_device *bdev = I_BDEV(inode);
1664                 set_blocksize(bdev, p->old_block_size);
1665                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1666         } else {
1667                 mutex_lock(&inode->i_mutex);
1668                 inode->i_flags &= ~S_SWAPFILE;
1669                 mutex_unlock(&inode->i_mutex);
1670         }
1671         filp_close(swap_file, NULL);
1672         err = 0;
1673         atomic_inc(&proc_poll_event);
1674         wake_up_interruptible(&proc_poll_wait);
1675
1676 out_dput:
1677         filp_close(victim, NULL);
1678 out:
1679         return err;
1680 }
1681
1682 #ifdef CONFIG_PROC_FS
1683 static unsigned swaps_poll(struct file *file, poll_table *wait)
1684 {
1685         struct seq_file *seq = file->private_data;
1686
1687         poll_wait(file, &proc_poll_wait, wait);
1688
1689         if (seq->poll_event != atomic_read(&proc_poll_event)) {
1690                 seq->poll_event = atomic_read(&proc_poll_event);
1691                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1692         }
1693
1694         return POLLIN | POLLRDNORM;
1695 }
1696
1697 /* iterator */
1698 static void *swap_start(struct seq_file *swap, loff_t *pos)
1699 {
1700         struct swap_info_struct *si;
1701         int type;
1702         loff_t l = *pos;
1703
1704         mutex_lock(&swapon_mutex);
1705
1706         if (!l)
1707                 return SEQ_START_TOKEN;
1708
1709         for (type = 0; type < nr_swapfiles; type++) {
1710                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1711                 si = swap_info[type];
1712                 if (!(si->flags & SWP_USED) || !si->swap_map)
1713                         continue;
1714                 if (!--l)
1715                         return si;
1716         }
1717
1718         return NULL;
1719 }
1720
1721 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1722 {
1723         struct swap_info_struct *si = v;
1724         int type;
1725
1726         if (v == SEQ_START_TOKEN)
1727                 type = 0;
1728         else
1729                 type = si->type + 1;
1730
1731         for (; type < nr_swapfiles; type++) {
1732                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1733                 si = swap_info[type];
1734                 if (!(si->flags & SWP_USED) || !si->swap_map)
1735                         continue;
1736                 ++*pos;
1737                 return si;
1738         }
1739
1740         return NULL;
1741 }
1742
1743 static void swap_stop(struct seq_file *swap, void *v)
1744 {
1745         mutex_unlock(&swapon_mutex);
1746 }
1747
1748 static int swap_show(struct seq_file *swap, void *v)
1749 {
1750         struct swap_info_struct *si = v;
1751         struct file *file;
1752         int len;
1753
1754         if (si == SEQ_START_TOKEN) {
1755                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1756                 return 0;
1757         }
1758
1759         file = si->swap_file;
1760         len = seq_path(swap, &file->f_path, " \t\n\\");
1761         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1762                         len < 40 ? 40 - len : 1, " ",
1763                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1764                                 "partition" : "file\t",
1765                         si->pages << (PAGE_SHIFT - 10),
1766                         si->inuse_pages << (PAGE_SHIFT - 10),
1767                         si->prio);
1768         return 0;
1769 }
1770
1771 static const struct seq_operations swaps_op = {
1772         .start =        swap_start,
1773         .next =         swap_next,
1774         .stop =         swap_stop,
1775         .show =         swap_show
1776 };
1777
1778 static int swaps_open(struct inode *inode, struct file *file)
1779 {
1780         struct seq_file *seq;
1781         int ret;
1782
1783         ret = seq_open(file, &swaps_op);
1784         if (ret)
1785                 return ret;
1786
1787         seq = file->private_data;
1788         seq->poll_event = atomic_read(&proc_poll_event);
1789         return 0;
1790 }
1791
1792 static const struct file_operations proc_swaps_operations = {
1793         .open           = swaps_open,
1794         .read           = seq_read,
1795         .llseek         = seq_lseek,
1796         .release        = seq_release,
1797         .poll           = swaps_poll,
1798 };
1799
1800 static int __init procswaps_init(void)
1801 {
1802         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1803         return 0;
1804 }
1805 __initcall(procswaps_init);
1806 #endif /* CONFIG_PROC_FS */
1807
1808 #ifdef MAX_SWAPFILES_CHECK
1809 static int __init max_swapfiles_check(void)
1810 {
1811         MAX_SWAPFILES_CHECK();
1812         return 0;
1813 }
1814 late_initcall(max_swapfiles_check);
1815 #endif
1816
1817 static struct swap_info_struct *alloc_swap_info(void)
1818 {
1819         struct swap_info_struct *p;
1820         unsigned int type;
1821
1822         p = kzalloc(sizeof(*p), GFP_KERNEL);
1823         if (!p)
1824                 return ERR_PTR(-ENOMEM);
1825
1826         spin_lock(&swap_lock);
1827         for (type = 0; type < nr_swapfiles; type++) {
1828                 if (!(swap_info[type]->flags & SWP_USED))
1829                         break;
1830         }
1831         if (type >= MAX_SWAPFILES) {
1832                 spin_unlock(&swap_lock);
1833                 kfree(p);
1834                 return ERR_PTR(-EPERM);
1835         }
1836         if (type >= nr_swapfiles) {
1837                 p->type = type;
1838                 swap_info[type] = p;
1839                 /*
1840                  * Write swap_info[type] before nr_swapfiles, in case a
1841                  * racing procfs swap_start() or swap_next() is reading them.
1842                  * (We never shrink nr_swapfiles, we never free this entry.)
1843                  */
1844                 smp_wmb();
1845                 nr_swapfiles++;
1846         } else {
1847                 kfree(p);
1848                 p = swap_info[type];
1849                 /*
1850                  * Do not memset this entry: a racing procfs swap_next()
1851                  * would be relying on p->type to remain valid.
1852                  */
1853         }
1854         INIT_LIST_HEAD(&p->first_swap_extent.list);
1855         p->flags = SWP_USED;
1856         p->next = -1;
1857         spin_unlock(&swap_lock);
1858
1859         return p;
1860 }
1861
1862 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1863 {
1864         int error;
1865
1866         if (S_ISBLK(inode->i_mode)) {
1867                 p->bdev = bdgrab(I_BDEV(inode));
1868                 error = blkdev_get(p->bdev,
1869                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1870                                    sys_swapon);
1871                 if (error < 0) {
1872                         p->bdev = NULL;
1873                         return -EINVAL;
1874                 }
1875                 p->old_block_size = block_size(p->bdev);
1876                 error = set_blocksize(p->bdev, PAGE_SIZE);
1877                 if (error < 0)
1878                         return error;
1879                 p->flags |= SWP_BLKDEV;
1880         } else if (S_ISREG(inode->i_mode)) {
1881                 p->bdev = inode->i_sb->s_bdev;
1882                 mutex_lock(&inode->i_mutex);
1883                 if (IS_SWAPFILE(inode))
1884                         return -EBUSY;
1885         } else
1886                 return -EINVAL;
1887
1888         return 0;
1889 }
1890
1891 static unsigned long read_swap_header(struct swap_info_struct *p,
1892                                         union swap_header *swap_header,
1893                                         struct inode *inode)
1894 {
1895         int i;
1896         unsigned long maxpages;
1897         unsigned long swapfilepages;
1898
1899         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1900                 printk(KERN_ERR "Unable to find swap-space signature\n");
1901                 return 0;
1902         }
1903
1904         /* swap partition endianess hack... */
1905         if (swab32(swap_header->info.version) == 1) {
1906                 swab32s(&swap_header->info.version);
1907                 swab32s(&swap_header->info.last_page);
1908                 swab32s(&swap_header->info.nr_badpages);
1909                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1910                         swab32s(&swap_header->info.badpages[i]);
1911         }
1912         /* Check the swap header's sub-version */
1913         if (swap_header->info.version != 1) {
1914                 printk(KERN_WARNING
1915                        "Unable to handle swap header version %d\n",
1916                        swap_header->info.version);
1917                 return 0;
1918         }
1919
1920         p->lowest_bit  = 1;
1921         p->cluster_next = 1;
1922         p->cluster_nr = 0;
1923
1924         /*
1925          * Find out how many pages are allowed for a single swap
1926          * device. There are three limiting factors: 1) the number
1927          * of bits for the swap offset in the swp_entry_t type, and
1928          * 2) the number of bits in the swap pte as defined by the
1929          * the different architectures, and 3) the number of free bits
1930          * in an exceptional radix_tree entry. In order to find the
1931          * largest possible bit mask, a swap entry with swap type 0
1932          * and swap offset ~0UL is created, encoded to a swap pte,
1933          * decoded to a swp_entry_t again, and finally the swap
1934          * offset is extracted. This will mask all the bits from
1935          * the initial ~0UL mask that can't be encoded in either
1936          * the swp_entry_t or the architecture definition of a
1937          * swap pte.  Then the same is done for a radix_tree entry.
1938          */
1939         maxpages = swp_offset(pte_to_swp_entry(
1940                         swp_entry_to_pte(swp_entry(0, ~0UL))));
1941         maxpages = swp_offset(radix_to_swp_entry(
1942                         swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1943
1944         if (maxpages > swap_header->info.last_page) {
1945                 maxpages = swap_header->info.last_page + 1;
1946                 /* p->max is an unsigned int: don't overflow it */
1947                 if ((unsigned int)maxpages == 0)
1948                         maxpages = UINT_MAX;
1949         }
1950         p->highest_bit = maxpages - 1;
1951
1952         if (!maxpages)
1953                 return 0;
1954         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1955         if (swapfilepages && maxpages > swapfilepages) {
1956                 printk(KERN_WARNING
1957                        "Swap area shorter than signature indicates\n");
1958                 return 0;
1959         }
1960         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1961                 return 0;
1962         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1963                 return 0;
1964
1965         return maxpages;
1966 }
1967
1968 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1969                                         union swap_header *swap_header,
1970                                         unsigned char *swap_map,
1971                                         unsigned long maxpages,
1972                                         sector_t *span)
1973 {
1974         int i;
1975         unsigned int nr_good_pages;
1976         int nr_extents;
1977
1978         nr_good_pages = maxpages - 1;   /* omit header page */
1979
1980         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1981                 unsigned int page_nr = swap_header->info.badpages[i];
1982                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1983                         return -EINVAL;
1984                 if (page_nr < maxpages) {
1985                         swap_map[page_nr] = SWAP_MAP_BAD;
1986                         nr_good_pages--;
1987                 }
1988         }
1989
1990         if (nr_good_pages) {
1991                 swap_map[0] = SWAP_MAP_BAD;
1992                 p->max = maxpages;
1993                 p->pages = nr_good_pages;
1994                 nr_extents = setup_swap_extents(p, span);
1995                 if (nr_extents < 0)
1996                         return nr_extents;
1997                 nr_good_pages = p->pages;
1998         }
1999         if (!nr_good_pages) {
2000                 printk(KERN_WARNING "Empty swap-file\n");
2001                 return -EINVAL;
2002         }
2003
2004         return nr_extents;
2005 }
2006
2007 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2008 {
2009         struct swap_info_struct *p;
2010         char *name;
2011         struct file *swap_file = NULL;
2012         struct address_space *mapping;
2013         int i;
2014         int prio;
2015         int error;
2016         union swap_header *swap_header;
2017         int nr_extents;
2018         sector_t span;
2019         unsigned long maxpages;
2020         unsigned char *swap_map = NULL;
2021         struct page *page = NULL;
2022         struct inode *inode = NULL;
2023
2024         if (!capable(CAP_SYS_ADMIN))
2025                 return -EPERM;
2026
2027         p = alloc_swap_info();
2028         if (IS_ERR(p))
2029                 return PTR_ERR(p);
2030
2031         name = getname(specialfile);
2032         if (IS_ERR(name)) {
2033                 error = PTR_ERR(name);
2034                 name = NULL;
2035                 goto bad_swap;
2036         }
2037         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2038         if (IS_ERR(swap_file)) {
2039                 error = PTR_ERR(swap_file);
2040                 swap_file = NULL;
2041                 goto bad_swap;
2042         }
2043
2044         p->swap_file = swap_file;
2045         mapping = swap_file->f_mapping;
2046
2047         for (i = 0; i < nr_swapfiles; i++) {
2048                 struct swap_info_struct *q = swap_info[i];
2049
2050                 if (q == p || !q->swap_file)
2051                         continue;
2052                 if (mapping == q->swap_file->f_mapping) {
2053                         error = -EBUSY;
2054                         goto bad_swap;
2055                 }
2056         }
2057
2058         inode = mapping->host;
2059         /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2060         error = claim_swapfile(p, inode);
2061         if (unlikely(error))
2062                 goto bad_swap;
2063
2064         /*
2065          * Read the swap header.
2066          */
2067         if (!mapping->a_ops->readpage) {
2068                 error = -EINVAL;
2069                 goto bad_swap;
2070         }
2071         page = read_mapping_page(mapping, 0, swap_file);
2072         if (IS_ERR(page)) {
2073                 error = PTR_ERR(page);
2074                 goto bad_swap;
2075         }
2076         swap_header = kmap(page);
2077
2078         maxpages = read_swap_header(p, swap_header, inode);
2079         if (unlikely(!maxpages)) {
2080                 error = -EINVAL;
2081                 goto bad_swap;
2082         }
2083
2084         /* OK, set up the swap map and apply the bad block list */
2085         swap_map = vzalloc(maxpages);
2086         if (!swap_map) {
2087                 error = -ENOMEM;
2088                 goto bad_swap;
2089         }
2090
2091         error = swap_cgroup_swapon(p->type, maxpages);
2092         if (error)
2093                 goto bad_swap;
2094
2095         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2096                 maxpages, &span);
2097         if (unlikely(nr_extents < 0)) {
2098                 error = nr_extents;
2099                 goto bad_swap;
2100         }
2101
2102         if (p->bdev) {
2103                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2104                         p->flags |= SWP_SOLIDSTATE;
2105                         p->cluster_next = 1 + (random32() % p->highest_bit);
2106                 }
2107                 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2108                         p->flags |= SWP_DISCARDABLE;
2109         }
2110
2111         mutex_lock(&swapon_mutex);
2112         prio = -1;
2113         if (swap_flags & SWAP_FLAG_PREFER)
2114                 prio =
2115                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2116         enable_swap_info(p, prio, swap_map);
2117
2118         printk(KERN_INFO "Adding %uk swap on %s.  "
2119                         "Priority:%d extents:%d across:%lluk %s%s\n",
2120                 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2121                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2122                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2123                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2124
2125         mutex_unlock(&swapon_mutex);
2126         atomic_inc(&proc_poll_event);
2127         wake_up_interruptible(&proc_poll_wait);
2128
2129         if (S_ISREG(inode->i_mode))
2130                 inode->i_flags |= S_SWAPFILE;
2131         error = 0;
2132         goto out;
2133 bad_swap:
2134         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2135                 set_blocksize(p->bdev, p->old_block_size);
2136                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2137         }
2138         destroy_swap_extents(p);
2139         swap_cgroup_swapoff(p->type);
2140         spin_lock(&swap_lock);
2141         p->swap_file = NULL;
2142         p->flags = 0;
2143         spin_unlock(&swap_lock);
2144         vfree(swap_map);
2145         if (swap_file) {
2146                 if (inode && S_ISREG(inode->i_mode)) {
2147                         mutex_unlock(&inode->i_mutex);
2148                         inode = NULL;
2149                 }
2150                 filp_close(swap_file, NULL);
2151         }
2152 out:
2153         if (page && !IS_ERR(page)) {
2154                 kunmap(page);
2155                 page_cache_release(page);
2156         }
2157         if (name)
2158                 putname(name);
2159         if (inode && S_ISREG(inode->i_mode))
2160                 mutex_unlock(&inode->i_mutex);
2161         return error;
2162 }
2163
2164 void si_swapinfo(struct sysinfo *val)
2165 {
2166         unsigned int type;
2167         unsigned long nr_to_be_unused = 0;
2168
2169         spin_lock(&swap_lock);
2170         for (type = 0; type < nr_swapfiles; type++) {
2171                 struct swap_info_struct *si = swap_info[type];
2172
2173                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2174                         nr_to_be_unused += si->inuse_pages;
2175         }
2176         val->freeswap = nr_swap_pages + nr_to_be_unused;
2177         val->totalswap = total_swap_pages + nr_to_be_unused;
2178         spin_unlock(&swap_lock);
2179 }
2180
2181 /*
2182  * Verify that a swap entry is valid and increment its swap map count.
2183  *
2184  * Returns error code in following case.
2185  * - success -> 0
2186  * - swp_entry is invalid -> EINVAL
2187  * - swp_entry is migration entry -> EINVAL
2188  * - swap-cache reference is requested but there is already one. -> EEXIST
2189  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2190  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2191  */
2192 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2193 {
2194         struct swap_info_struct *p;
2195         unsigned long offset, type;
2196         unsigned char count;
2197         unsigned char has_cache;
2198         int err = -EINVAL;
2199
2200         if (non_swap_entry(entry))
2201                 goto out;
2202
2203         type = swp_type(entry);
2204         if (type >= nr_swapfiles)
2205                 goto bad_file;
2206         p = swap_info[type];
2207         offset = swp_offset(entry);
2208
2209         spin_lock(&swap_lock);
2210         if (unlikely(offset >= p->max))
2211                 goto unlock_out;
2212
2213         count = p->swap_map[offset];
2214         has_cache = count & SWAP_HAS_CACHE;
2215         count &= ~SWAP_HAS_CACHE;
2216         err = 0;
2217
2218         if (usage == SWAP_HAS_CACHE) {
2219
2220                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2221                 if (!has_cache && count)
2222                         has_cache = SWAP_HAS_CACHE;
2223                 else if (has_cache)             /* someone else added cache */
2224                         err = -EEXIST;
2225                 else                            /* no users remaining */
2226                         err = -ENOENT;
2227
2228         } else if (count || has_cache) {
2229
2230                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2231                         count += usage;
2232                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2233                         err = -EINVAL;
2234                 else if (swap_count_continued(p, offset, count))
2235                         count = COUNT_CONTINUED;
2236                 else
2237                         err = -ENOMEM;
2238         } else
2239                 err = -ENOENT;                  /* unused swap entry */
2240
2241         p->swap_map[offset] = count | has_cache;
2242
2243 unlock_out:
2244         spin_unlock(&swap_lock);
2245 out:
2246         return err;
2247
2248 bad_file:
2249         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2250         goto out;
2251 }
2252
2253 /*
2254  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2255  * (in which case its reference count is never incremented).
2256  */
2257 void swap_shmem_alloc(swp_entry_t entry)
2258 {
2259         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2260 }
2261
2262 /*
2263  * Increase reference count of swap entry by 1.
2264  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2265  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2266  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2267  * might occur if a page table entry has got corrupted.
2268  */
2269 int swap_duplicate(swp_entry_t entry)
2270 {
2271         int err = 0;
2272
2273         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2274                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2275         return err;
2276 }
2277
2278 /*
2279  * @entry: swap entry for which we allocate swap cache.
2280  *
2281  * Called when allocating swap cache for existing swap entry,
2282  * This can return error codes. Returns 0 at success.
2283  * -EBUSY means there is a swap cache.
2284  * Note: return code is different from swap_duplicate().
2285  */
2286 int swapcache_prepare(swp_entry_t entry)
2287 {
2288         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2289 }
2290
2291 /*
2292  * swap_lock prevents swap_map being freed. Don't grab an extra
2293  * reference on the swaphandle, it doesn't matter if it becomes unused.
2294  */
2295 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2296 {
2297         struct swap_info_struct *si;
2298         int our_page_cluster = page_cluster;
2299         pgoff_t target, toff;
2300         pgoff_t base, end;
2301         int nr_pages = 0;
2302
2303         if (!our_page_cluster)  /* no readahead */
2304                 return 0;
2305
2306         si = swap_info[swp_type(entry)];
2307         target = swp_offset(entry);
2308         base = (target >> our_page_cluster) << our_page_cluster;
2309         end = base + (1 << our_page_cluster);
2310         if (!base)              /* first page is swap header */
2311                 base++;
2312
2313         spin_lock(&swap_lock);
2314         if (end > si->max)      /* don't go beyond end of map */
2315                 end = si->max;
2316
2317         /* Count contiguous allocated slots above our target */
2318         for (toff = target; ++toff < end; nr_pages++) {
2319                 /* Don't read in free or bad pages */
2320                 if (!si->swap_map[toff])
2321                         break;
2322                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2323                         break;
2324         }
2325         /* Count contiguous allocated slots below our target */
2326         for (toff = target; --toff >= base; nr_pages++) {
2327                 /* Don't read in free or bad pages */
2328                 if (!si->swap_map[toff])
2329                         break;
2330                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2331                         break;
2332         }
2333         spin_unlock(&swap_lock);
2334
2335         /*
2336          * Indicate starting offset, and return number of pages to get:
2337          * if only 1, say 0, since there's then no readahead to be done.
2338          */
2339         *offset = ++toff;
2340         return nr_pages? ++nr_pages: 0;
2341 }
2342
2343 /*
2344  * add_swap_count_continuation - called when a swap count is duplicated
2345  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2346  * page of the original vmalloc'ed swap_map, to hold the continuation count
2347  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2348  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2349  *
2350  * These continuation pages are seldom referenced: the common paths all work
2351  * on the original swap_map, only referring to a continuation page when the
2352  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2353  *
2354  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2355  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2356  * can be called after dropping locks.
2357  */
2358 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2359 {
2360         struct swap_info_struct *si;
2361         struct page *head;
2362         struct page *page;
2363         struct page *list_page;
2364         pgoff_t offset;
2365         unsigned char count;
2366
2367         /*
2368          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2369          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2370          */
2371         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2372
2373         si = swap_info_get(entry);
2374         if (!si) {
2375                 /*
2376                  * An acceptable race has occurred since the failing
2377                  * __swap_duplicate(): the swap entry has been freed,
2378                  * perhaps even the whole swap_map cleared for swapoff.
2379                  */
2380                 goto outer;
2381         }
2382
2383         offset = swp_offset(entry);
2384         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2385
2386         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2387                 /*
2388                  * The higher the swap count, the more likely it is that tasks
2389                  * will race to add swap count continuation: we need to avoid
2390                  * over-provisioning.
2391                  */
2392                 goto out;
2393         }
2394
2395         if (!page) {
2396                 spin_unlock(&swap_lock);
2397                 return -ENOMEM;
2398         }
2399
2400         /*
2401          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2402          * no architecture is using highmem pages for kernel pagetables: so it
2403          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2404          */
2405         head = vmalloc_to_page(si->swap_map + offset);
2406         offset &= ~PAGE_MASK;
2407
2408         /*
2409          * Page allocation does not initialize the page's lru field,
2410          * but it does always reset its private field.
2411          */
2412         if (!page_private(head)) {
2413                 BUG_ON(count & COUNT_CONTINUED);
2414                 INIT_LIST_HEAD(&head->lru);
2415                 set_page_private(head, SWP_CONTINUED);
2416                 si->flags |= SWP_CONTINUED;
2417         }
2418
2419         list_for_each_entry(list_page, &head->lru, lru) {
2420                 unsigned char *map;
2421
2422                 /*
2423                  * If the previous map said no continuation, but we've found
2424                  * a continuation page, free our allocation and use this one.
2425                  */
2426                 if (!(count & COUNT_CONTINUED))
2427                         goto out;
2428
2429                 map = kmap_atomic(list_page, KM_USER0) + offset;
2430                 count = *map;
2431                 kunmap_atomic(map, KM_USER0);
2432
2433                 /*
2434                  * If this continuation count now has some space in it,
2435                  * free our allocation and use this one.
2436                  */
2437                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2438                         goto out;
2439         }
2440
2441         list_add_tail(&page->lru, &head->lru);
2442         page = NULL;                    /* now it's attached, don't free it */
2443 out:
2444         spin_unlock(&swap_lock);
2445 outer:
2446         if (page)
2447                 __free_page(page);
2448         return 0;
2449 }
2450
2451 /*
2452  * swap_count_continued - when the original swap_map count is incremented
2453  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2454  * into, carry if so, or else fail until a new continuation page is allocated;
2455  * when the original swap_map count is decremented from 0 with continuation,
2456  * borrow from the continuation and report whether it still holds more.
2457  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2458  */
2459 static bool swap_count_continued(struct swap_info_struct *si,
2460                                  pgoff_t offset, unsigned char count)
2461 {
2462         struct page *head;
2463         struct page *page;
2464         unsigned char *map;
2465
2466         head = vmalloc_to_page(si->swap_map + offset);
2467         if (page_private(head) != SWP_CONTINUED) {
2468                 BUG_ON(count & COUNT_CONTINUED);
2469                 return false;           /* need to add count continuation */
2470         }
2471
2472         offset &= ~PAGE_MASK;
2473         page = list_entry(head->lru.next, struct page, lru);
2474         map = kmap_atomic(page, KM_USER0) + offset;
2475
2476         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2477                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2478
2479         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2480                 /*
2481                  * Think of how you add 1 to 999
2482                  */
2483                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2484                         kunmap_atomic(map, KM_USER0);
2485                         page = list_entry(page->lru.next, struct page, lru);
2486                         BUG_ON(page == head);
2487                         map = kmap_atomic(page, KM_USER0) + offset;
2488                 }
2489                 if (*map == SWAP_CONT_MAX) {
2490                         kunmap_atomic(map, KM_USER0);
2491                         page = list_entry(page->lru.next, struct page, lru);
2492                         if (page == head)
2493                                 return false;   /* add count continuation */
2494                         map = kmap_atomic(page, KM_USER0) + offset;
2495 init_map:               *map = 0;               /* we didn't zero the page */
2496                 }
2497                 *map += 1;
2498                 kunmap_atomic(map, KM_USER0);
2499                 page = list_entry(page->lru.prev, struct page, lru);
2500                 while (page != head) {
2501                         map = kmap_atomic(page, KM_USER0) + offset;
2502                         *map = COUNT_CONTINUED;
2503                         kunmap_atomic(map, KM_USER0);
2504                         page = list_entry(page->lru.prev, struct page, lru);
2505                 }
2506                 return true;                    /* incremented */
2507
2508         } else {                                /* decrementing */
2509                 /*
2510                  * Think of how you subtract 1 from 1000
2511                  */
2512                 BUG_ON(count != COUNT_CONTINUED);
2513                 while (*map == COUNT_CONTINUED) {
2514                         kunmap_atomic(map, KM_USER0);
2515                         page = list_entry(page->lru.next, struct page, lru);
2516                         BUG_ON(page == head);
2517                         map = kmap_atomic(page, KM_USER0) + offset;
2518                 }
2519                 BUG_ON(*map == 0);
2520                 *map -= 1;
2521                 if (*map == 0)
2522                         count = 0;
2523                 kunmap_atomic(map, KM_USER0);
2524                 page = list_entry(page->lru.prev, struct page, lru);
2525                 while (page != head) {
2526                         map = kmap_atomic(page, KM_USER0) + offset;
2527                         *map = SWAP_CONT_MAX | count;
2528                         count = COUNT_CONTINUED;
2529                         kunmap_atomic(map, KM_USER0);
2530                         page = list_entry(page->lru.prev, struct page, lru);
2531                 }
2532                 return count == COUNT_CONTINUED;
2533         }
2534 }
2535
2536 /*
2537  * free_swap_count_continuations - swapoff free all the continuation pages
2538  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2539  */
2540 static void free_swap_count_continuations(struct swap_info_struct *si)
2541 {
2542         pgoff_t offset;
2543
2544         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2545                 struct page *head;
2546                 head = vmalloc_to_page(si->swap_map + offset);
2547                 if (page_private(head)) {
2548                         struct list_head *this, *next;
2549                         list_for_each_safe(this, next, &head->lru) {
2550                                 struct page *page;
2551                                 page = list_entry(this, struct page, lru);
2552                                 list_del(this);
2553                                 __free_page(page);
2554                         }
2555                 }
2556         }
2557 }