swap_info: note SWAP_MAP_SHMEM
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32
33 #include <asm/pgtable.h>
34 #include <asm/tlbflush.h>
35 #include <linux/swapops.h>
36 #include <linux/page_cgroup.h>
37
38 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
39                                  unsigned char);
40 static void free_swap_count_continuations(struct swap_info_struct *);
41
42 static DEFINE_SPINLOCK(swap_lock);
43 static unsigned int nr_swapfiles;
44 long nr_swap_pages;
45 long total_swap_pages;
46 static int least_priority;
47
48 static const char Bad_file[] = "Bad swap file entry ";
49 static const char Unused_file[] = "Unused swap file entry ";
50 static const char Bad_offset[] = "Bad swap offset entry ";
51 static const char Unused_offset[] = "Unused swap offset entry ";
52
53 static struct swap_list_t swap_list = {-1, -1};
54
55 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
56
57 static DEFINE_MUTEX(swapon_mutex);
58
59 static inline unsigned char swap_count(unsigned char ent)
60 {
61         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
62 }
63
64 /* returns 1 if swap entry is freed */
65 static int
66 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
67 {
68         swp_entry_t entry = swp_entry(si->type, offset);
69         struct page *page;
70         int ret = 0;
71
72         page = find_get_page(&swapper_space, entry.val);
73         if (!page)
74                 return 0;
75         /*
76          * This function is called from scan_swap_map() and it's called
77          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
78          * We have to use trylock for avoiding deadlock. This is a special
79          * case and you should use try_to_free_swap() with explicit lock_page()
80          * in usual operations.
81          */
82         if (trylock_page(page)) {
83                 ret = try_to_free_swap(page);
84                 unlock_page(page);
85         }
86         page_cache_release(page);
87         return ret;
88 }
89
90 /*
91  * We need this because the bdev->unplug_fn can sleep and we cannot
92  * hold swap_lock while calling the unplug_fn. And swap_lock
93  * cannot be turned into a mutex.
94  */
95 static DECLARE_RWSEM(swap_unplug_sem);
96
97 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
98 {
99         swp_entry_t entry;
100
101         down_read(&swap_unplug_sem);
102         entry.val = page_private(page);
103         if (PageSwapCache(page)) {
104                 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
105                 struct backing_dev_info *bdi;
106
107                 /*
108                  * If the page is removed from swapcache from under us (with a
109                  * racy try_to_unuse/swapoff) we need an additional reference
110                  * count to avoid reading garbage from page_private(page) above.
111                  * If the WARN_ON triggers during a swapoff it maybe the race
112                  * condition and it's harmless. However if it triggers without
113                  * swapoff it signals a problem.
114                  */
115                 WARN_ON(page_count(page) <= 1);
116
117                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
118                 blk_run_backing_dev(bdi, page);
119         }
120         up_read(&swap_unplug_sem);
121 }
122
123 /*
124  * swapon tell device that all the old swap contents can be discarded,
125  * to allow the swap device to optimize its wear-levelling.
126  */
127 static int discard_swap(struct swap_info_struct *si)
128 {
129         struct swap_extent *se;
130         sector_t start_block;
131         sector_t nr_blocks;
132         int err = 0;
133
134         /* Do not discard the swap header page! */
135         se = &si->first_swap_extent;
136         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
137         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
138         if (nr_blocks) {
139                 err = blkdev_issue_discard(si->bdev, start_block,
140                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
141                 if (err)
142                         return err;
143                 cond_resched();
144         }
145
146         list_for_each_entry(se, &si->first_swap_extent.list, list) {
147                 start_block = se->start_block << (PAGE_SHIFT - 9);
148                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
149
150                 err = blkdev_issue_discard(si->bdev, start_block,
151                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
152                 if (err)
153                         break;
154
155                 cond_resched();
156         }
157         return err;             /* That will often be -EOPNOTSUPP */
158 }
159
160 /*
161  * swap allocation tell device that a cluster of swap can now be discarded,
162  * to allow the swap device to optimize its wear-levelling.
163  */
164 static void discard_swap_cluster(struct swap_info_struct *si,
165                                  pgoff_t start_page, pgoff_t nr_pages)
166 {
167         struct swap_extent *se = si->curr_swap_extent;
168         int found_extent = 0;
169
170         while (nr_pages) {
171                 struct list_head *lh;
172
173                 if (se->start_page <= start_page &&
174                     start_page < se->start_page + se->nr_pages) {
175                         pgoff_t offset = start_page - se->start_page;
176                         sector_t start_block = se->start_block + offset;
177                         sector_t nr_blocks = se->nr_pages - offset;
178
179                         if (nr_blocks > nr_pages)
180                                 nr_blocks = nr_pages;
181                         start_page += nr_blocks;
182                         nr_pages -= nr_blocks;
183
184                         if (!found_extent++)
185                                 si->curr_swap_extent = se;
186
187                         start_block <<= PAGE_SHIFT - 9;
188                         nr_blocks <<= PAGE_SHIFT - 9;
189                         if (blkdev_issue_discard(si->bdev, start_block,
190                                     nr_blocks, GFP_NOIO, DISCARD_FL_BARRIER))
191                                 break;
192                 }
193
194                 lh = se->list.next;
195                 se = list_entry(lh, struct swap_extent, list);
196         }
197 }
198
199 static int wait_for_discard(void *word)
200 {
201         schedule();
202         return 0;
203 }
204
205 #define SWAPFILE_CLUSTER        256
206 #define LATENCY_LIMIT           256
207
208 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
209                                           unsigned char usage)
210 {
211         unsigned long offset;
212         unsigned long scan_base;
213         unsigned long last_in_cluster = 0;
214         int latency_ration = LATENCY_LIMIT;
215         int found_free_cluster = 0;
216
217         /*
218          * We try to cluster swap pages by allocating them sequentially
219          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
220          * way, however, we resort to first-free allocation, starting
221          * a new cluster.  This prevents us from scattering swap pages
222          * all over the entire swap partition, so that we reduce
223          * overall disk seek times between swap pages.  -- sct
224          * But we do now try to find an empty cluster.  -Andrea
225          * And we let swap pages go all over an SSD partition.  Hugh
226          */
227
228         si->flags += SWP_SCANNING;
229         scan_base = offset = si->cluster_next;
230
231         if (unlikely(!si->cluster_nr--)) {
232                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
233                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
234                         goto checks;
235                 }
236                 if (si->flags & SWP_DISCARDABLE) {
237                         /*
238                          * Start range check on racing allocations, in case
239                          * they overlap the cluster we eventually decide on
240                          * (we scan without swap_lock to allow preemption).
241                          * It's hardly conceivable that cluster_nr could be
242                          * wrapped during our scan, but don't depend on it.
243                          */
244                         if (si->lowest_alloc)
245                                 goto checks;
246                         si->lowest_alloc = si->max;
247                         si->highest_alloc = 0;
248                 }
249                 spin_unlock(&swap_lock);
250
251                 /*
252                  * If seek is expensive, start searching for new cluster from
253                  * start of partition, to minimize the span of allocated swap.
254                  * But if seek is cheap, search from our current position, so
255                  * that swap is allocated from all over the partition: if the
256                  * Flash Translation Layer only remaps within limited zones,
257                  * we don't want to wear out the first zone too quickly.
258                  */
259                 if (!(si->flags & SWP_SOLIDSTATE))
260                         scan_base = offset = si->lowest_bit;
261                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
262
263                 /* Locate the first empty (unaligned) cluster */
264                 for (; last_in_cluster <= si->highest_bit; offset++) {
265                         if (si->swap_map[offset])
266                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
267                         else if (offset == last_in_cluster) {
268                                 spin_lock(&swap_lock);
269                                 offset -= SWAPFILE_CLUSTER - 1;
270                                 si->cluster_next = offset;
271                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
272                                 found_free_cluster = 1;
273                                 goto checks;
274                         }
275                         if (unlikely(--latency_ration < 0)) {
276                                 cond_resched();
277                                 latency_ration = LATENCY_LIMIT;
278                         }
279                 }
280
281                 offset = si->lowest_bit;
282                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
283
284                 /* Locate the first empty (unaligned) cluster */
285                 for (; last_in_cluster < scan_base; offset++) {
286                         if (si->swap_map[offset])
287                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
288                         else if (offset == last_in_cluster) {
289                                 spin_lock(&swap_lock);
290                                 offset -= SWAPFILE_CLUSTER - 1;
291                                 si->cluster_next = offset;
292                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
293                                 found_free_cluster = 1;
294                                 goto checks;
295                         }
296                         if (unlikely(--latency_ration < 0)) {
297                                 cond_resched();
298                                 latency_ration = LATENCY_LIMIT;
299                         }
300                 }
301
302                 offset = scan_base;
303                 spin_lock(&swap_lock);
304                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
305                 si->lowest_alloc = 0;
306         }
307
308 checks:
309         if (!(si->flags & SWP_WRITEOK))
310                 goto no_page;
311         if (!si->highest_bit)
312                 goto no_page;
313         if (offset > si->highest_bit)
314                 scan_base = offset = si->lowest_bit;
315
316         /* reuse swap entry of cache-only swap if not busy. */
317         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
318                 int swap_was_freed;
319                 spin_unlock(&swap_lock);
320                 swap_was_freed = __try_to_reclaim_swap(si, offset);
321                 spin_lock(&swap_lock);
322                 /* entry was freed successfully, try to use this again */
323                 if (swap_was_freed)
324                         goto checks;
325                 goto scan; /* check next one */
326         }
327
328         if (si->swap_map[offset])
329                 goto scan;
330
331         if (offset == si->lowest_bit)
332                 si->lowest_bit++;
333         if (offset == si->highest_bit)
334                 si->highest_bit--;
335         si->inuse_pages++;
336         if (si->inuse_pages == si->pages) {
337                 si->lowest_bit = si->max;
338                 si->highest_bit = 0;
339         }
340         si->swap_map[offset] = usage;
341         si->cluster_next = offset + 1;
342         si->flags -= SWP_SCANNING;
343
344         if (si->lowest_alloc) {
345                 /*
346                  * Only set when SWP_DISCARDABLE, and there's a scan
347                  * for a free cluster in progress or just completed.
348                  */
349                 if (found_free_cluster) {
350                         /*
351                          * To optimize wear-levelling, discard the
352                          * old data of the cluster, taking care not to
353                          * discard any of its pages that have already
354                          * been allocated by racing tasks (offset has
355                          * already stepped over any at the beginning).
356                          */
357                         if (offset < si->highest_alloc &&
358                             si->lowest_alloc <= last_in_cluster)
359                                 last_in_cluster = si->lowest_alloc - 1;
360                         si->flags |= SWP_DISCARDING;
361                         spin_unlock(&swap_lock);
362
363                         if (offset < last_in_cluster)
364                                 discard_swap_cluster(si, offset,
365                                         last_in_cluster - offset + 1);
366
367                         spin_lock(&swap_lock);
368                         si->lowest_alloc = 0;
369                         si->flags &= ~SWP_DISCARDING;
370
371                         smp_mb();       /* wake_up_bit advises this */
372                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
373
374                 } else if (si->flags & SWP_DISCARDING) {
375                         /*
376                          * Delay using pages allocated by racing tasks
377                          * until the whole discard has been issued. We
378                          * could defer that delay until swap_writepage,
379                          * but it's easier to keep this self-contained.
380                          */
381                         spin_unlock(&swap_lock);
382                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
383                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
384                         spin_lock(&swap_lock);
385                 } else {
386                         /*
387                          * Note pages allocated by racing tasks while
388                          * scan for a free cluster is in progress, so
389                          * that its final discard can exclude them.
390                          */
391                         if (offset < si->lowest_alloc)
392                                 si->lowest_alloc = offset;
393                         if (offset > si->highest_alloc)
394                                 si->highest_alloc = offset;
395                 }
396         }
397         return offset;
398
399 scan:
400         spin_unlock(&swap_lock);
401         while (++offset <= si->highest_bit) {
402                 if (!si->swap_map[offset]) {
403                         spin_lock(&swap_lock);
404                         goto checks;
405                 }
406                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
407                         spin_lock(&swap_lock);
408                         goto checks;
409                 }
410                 if (unlikely(--latency_ration < 0)) {
411                         cond_resched();
412                         latency_ration = LATENCY_LIMIT;
413                 }
414         }
415         offset = si->lowest_bit;
416         while (++offset < scan_base) {
417                 if (!si->swap_map[offset]) {
418                         spin_lock(&swap_lock);
419                         goto checks;
420                 }
421                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
422                         spin_lock(&swap_lock);
423                         goto checks;
424                 }
425                 if (unlikely(--latency_ration < 0)) {
426                         cond_resched();
427                         latency_ration = LATENCY_LIMIT;
428                 }
429         }
430         spin_lock(&swap_lock);
431
432 no_page:
433         si->flags -= SWP_SCANNING;
434         return 0;
435 }
436
437 swp_entry_t get_swap_page(void)
438 {
439         struct swap_info_struct *si;
440         pgoff_t offset;
441         int type, next;
442         int wrapped = 0;
443
444         spin_lock(&swap_lock);
445         if (nr_swap_pages <= 0)
446                 goto noswap;
447         nr_swap_pages--;
448
449         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
450                 si = swap_info[type];
451                 next = si->next;
452                 if (next < 0 ||
453                     (!wrapped && si->prio != swap_info[next]->prio)) {
454                         next = swap_list.head;
455                         wrapped++;
456                 }
457
458                 if (!si->highest_bit)
459                         continue;
460                 if (!(si->flags & SWP_WRITEOK))
461                         continue;
462
463                 swap_list.next = next;
464                 /* This is called for allocating swap entry for cache */
465                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
466                 if (offset) {
467                         spin_unlock(&swap_lock);
468                         return swp_entry(type, offset);
469                 }
470                 next = swap_list.next;
471         }
472
473         nr_swap_pages++;
474 noswap:
475         spin_unlock(&swap_lock);
476         return (swp_entry_t) {0};
477 }
478
479 /* The only caller of this function is now susupend routine */
480 swp_entry_t get_swap_page_of_type(int type)
481 {
482         struct swap_info_struct *si;
483         pgoff_t offset;
484
485         spin_lock(&swap_lock);
486         si = swap_info[type];
487         if (si && (si->flags & SWP_WRITEOK)) {
488                 nr_swap_pages--;
489                 /* This is called for allocating swap entry, not cache */
490                 offset = scan_swap_map(si, 1);
491                 if (offset) {
492                         spin_unlock(&swap_lock);
493                         return swp_entry(type, offset);
494                 }
495                 nr_swap_pages++;
496         }
497         spin_unlock(&swap_lock);
498         return (swp_entry_t) {0};
499 }
500
501 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
502 {
503         struct swap_info_struct *p;
504         unsigned long offset, type;
505
506         if (!entry.val)
507                 goto out;
508         type = swp_type(entry);
509         if (type >= nr_swapfiles)
510                 goto bad_nofile;
511         p = swap_info[type];
512         if (!(p->flags & SWP_USED))
513                 goto bad_device;
514         offset = swp_offset(entry);
515         if (offset >= p->max)
516                 goto bad_offset;
517         if (!p->swap_map[offset])
518                 goto bad_free;
519         spin_lock(&swap_lock);
520         return p;
521
522 bad_free:
523         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
524         goto out;
525 bad_offset:
526         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
527         goto out;
528 bad_device:
529         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
530         goto out;
531 bad_nofile:
532         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
533 out:
534         return NULL;
535 }
536
537 static unsigned char swap_entry_free(struct swap_info_struct *p,
538                                      swp_entry_t entry, unsigned char usage)
539 {
540         unsigned long offset = swp_offset(entry);
541         unsigned char count;
542         unsigned char has_cache;
543
544         count = p->swap_map[offset];
545         has_cache = count & SWAP_HAS_CACHE;
546         count &= ~SWAP_HAS_CACHE;
547
548         if (usage == SWAP_HAS_CACHE) {
549                 VM_BUG_ON(!has_cache);
550                 has_cache = 0;
551         } else if (count == SWAP_MAP_SHMEM) {
552                 /*
553                  * Or we could insist on shmem.c using a special
554                  * swap_shmem_free() and free_shmem_swap_and_cache()...
555                  */
556                 count = 0;
557         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
558                 if (count == COUNT_CONTINUED) {
559                         if (swap_count_continued(p, offset, count))
560                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
561                         else
562                                 count = SWAP_MAP_MAX;
563                 } else
564                         count--;
565         }
566
567         if (!count)
568                 mem_cgroup_uncharge_swap(entry);
569
570         usage = count | has_cache;
571         p->swap_map[offset] = usage;
572
573         /* free if no reference */
574         if (!usage) {
575                 if (offset < p->lowest_bit)
576                         p->lowest_bit = offset;
577                 if (offset > p->highest_bit)
578                         p->highest_bit = offset;
579                 if (swap_list.next >= 0 &&
580                     p->prio > swap_info[swap_list.next]->prio)
581                         swap_list.next = p->type;
582                 nr_swap_pages++;
583                 p->inuse_pages--;
584         }
585
586         return usage;
587 }
588
589 /*
590  * Caller has made sure that the swapdevice corresponding to entry
591  * is still around or has not been recycled.
592  */
593 void swap_free(swp_entry_t entry)
594 {
595         struct swap_info_struct *p;
596
597         p = swap_info_get(entry);
598         if (p) {
599                 swap_entry_free(p, entry, 1);
600                 spin_unlock(&swap_lock);
601         }
602 }
603
604 /*
605  * Called after dropping swapcache to decrease refcnt to swap entries.
606  */
607 void swapcache_free(swp_entry_t entry, struct page *page)
608 {
609         struct swap_info_struct *p;
610         unsigned char count;
611
612         p = swap_info_get(entry);
613         if (p) {
614                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
615                 if (page)
616                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
617                 spin_unlock(&swap_lock);
618         }
619 }
620
621 /*
622  * How many references to page are currently swapped out?
623  * This does not give an exact answer when swap count is continued,
624  * but does include the high COUNT_CONTINUED flag to allow for that.
625  */
626 static inline int page_swapcount(struct page *page)
627 {
628         int count = 0;
629         struct swap_info_struct *p;
630         swp_entry_t entry;
631
632         entry.val = page_private(page);
633         p = swap_info_get(entry);
634         if (p) {
635                 count = swap_count(p->swap_map[swp_offset(entry)]);
636                 spin_unlock(&swap_lock);
637         }
638         return count;
639 }
640
641 /*
642  * We can write to an anon page without COW if there are no other references
643  * to it.  And as a side-effect, free up its swap: because the old content
644  * on disk will never be read, and seeking back there to write new content
645  * later would only waste time away from clustering.
646  */
647 int reuse_swap_page(struct page *page)
648 {
649         int count;
650
651         VM_BUG_ON(!PageLocked(page));
652         count = page_mapcount(page);
653         if (count <= 1 && PageSwapCache(page)) {
654                 count += page_swapcount(page);
655                 if (count == 1 && !PageWriteback(page)) {
656                         delete_from_swap_cache(page);
657                         SetPageDirty(page);
658                 }
659         }
660         return count == 1;
661 }
662
663 /*
664  * If swap is getting full, or if there are no more mappings of this page,
665  * then try_to_free_swap is called to free its swap space.
666  */
667 int try_to_free_swap(struct page *page)
668 {
669         VM_BUG_ON(!PageLocked(page));
670
671         if (!PageSwapCache(page))
672                 return 0;
673         if (PageWriteback(page))
674                 return 0;
675         if (page_swapcount(page))
676                 return 0;
677
678         delete_from_swap_cache(page);
679         SetPageDirty(page);
680         return 1;
681 }
682
683 /*
684  * Free the swap entry like above, but also try to
685  * free the page cache entry if it is the last user.
686  */
687 int free_swap_and_cache(swp_entry_t entry)
688 {
689         struct swap_info_struct *p;
690         struct page *page = NULL;
691
692         if (non_swap_entry(entry))
693                 return 1;
694
695         p = swap_info_get(entry);
696         if (p) {
697                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
698                         page = find_get_page(&swapper_space, entry.val);
699                         if (page && !trylock_page(page)) {
700                                 page_cache_release(page);
701                                 page = NULL;
702                         }
703                 }
704                 spin_unlock(&swap_lock);
705         }
706         if (page) {
707                 /*
708                  * Not mapped elsewhere, or swap space full? Free it!
709                  * Also recheck PageSwapCache now page is locked (above).
710                  */
711                 if (PageSwapCache(page) && !PageWriteback(page) &&
712                                 (!page_mapped(page) || vm_swap_full())) {
713                         delete_from_swap_cache(page);
714                         SetPageDirty(page);
715                 }
716                 unlock_page(page);
717                 page_cache_release(page);
718         }
719         return p != NULL;
720 }
721
722 #ifdef CONFIG_HIBERNATION
723 /*
724  * Find the swap type that corresponds to given device (if any).
725  *
726  * @offset - number of the PAGE_SIZE-sized block of the device, starting
727  * from 0, in which the swap header is expected to be located.
728  *
729  * This is needed for the suspend to disk (aka swsusp).
730  */
731 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
732 {
733         struct block_device *bdev = NULL;
734         int type;
735
736         if (device)
737                 bdev = bdget(device);
738
739         spin_lock(&swap_lock);
740         for (type = 0; type < nr_swapfiles; type++) {
741                 struct swap_info_struct *sis = swap_info[type];
742
743                 if (!(sis->flags & SWP_WRITEOK))
744                         continue;
745
746                 if (!bdev) {
747                         if (bdev_p)
748                                 *bdev_p = bdgrab(sis->bdev);
749
750                         spin_unlock(&swap_lock);
751                         return type;
752                 }
753                 if (bdev == sis->bdev) {
754                         struct swap_extent *se = &sis->first_swap_extent;
755
756                         if (se->start_block == offset) {
757                                 if (bdev_p)
758                                         *bdev_p = bdgrab(sis->bdev);
759
760                                 spin_unlock(&swap_lock);
761                                 bdput(bdev);
762                                 return type;
763                         }
764                 }
765         }
766         spin_unlock(&swap_lock);
767         if (bdev)
768                 bdput(bdev);
769
770         return -ENODEV;
771 }
772
773 /*
774  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
775  * corresponding to given index in swap_info (swap type).
776  */
777 sector_t swapdev_block(int type, pgoff_t offset)
778 {
779         struct block_device *bdev;
780
781         if ((unsigned int)type >= nr_swapfiles)
782                 return 0;
783         if (!(swap_info[type]->flags & SWP_WRITEOK))
784                 return 0;
785         return map_swap_page(swp_entry(type, offset), &bdev);
786 }
787
788 /*
789  * Return either the total number of swap pages of given type, or the number
790  * of free pages of that type (depending on @free)
791  *
792  * This is needed for software suspend
793  */
794 unsigned int count_swap_pages(int type, int free)
795 {
796         unsigned int n = 0;
797
798         spin_lock(&swap_lock);
799         if ((unsigned int)type < nr_swapfiles) {
800                 struct swap_info_struct *sis = swap_info[type];
801
802                 if (sis->flags & SWP_WRITEOK) {
803                         n = sis->pages;
804                         if (free)
805                                 n -= sis->inuse_pages;
806                 }
807         }
808         spin_unlock(&swap_lock);
809         return n;
810 }
811 #endif /* CONFIG_HIBERNATION */
812
813 /*
814  * No need to decide whether this PTE shares the swap entry with others,
815  * just let do_wp_page work it out if a write is requested later - to
816  * force COW, vm_page_prot omits write permission from any private vma.
817  */
818 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
819                 unsigned long addr, swp_entry_t entry, struct page *page)
820 {
821         struct mem_cgroup *ptr = NULL;
822         spinlock_t *ptl;
823         pte_t *pte;
824         int ret = 1;
825
826         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
827                 ret = -ENOMEM;
828                 goto out_nolock;
829         }
830
831         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
832         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
833                 if (ret > 0)
834                         mem_cgroup_cancel_charge_swapin(ptr);
835                 ret = 0;
836                 goto out;
837         }
838
839         inc_mm_counter(vma->vm_mm, anon_rss);
840         get_page(page);
841         set_pte_at(vma->vm_mm, addr, pte,
842                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
843         page_add_anon_rmap(page, vma, addr);
844         mem_cgroup_commit_charge_swapin(page, ptr);
845         swap_free(entry);
846         /*
847          * Move the page to the active list so it is not
848          * immediately swapped out again after swapon.
849          */
850         activate_page(page);
851 out:
852         pte_unmap_unlock(pte, ptl);
853 out_nolock:
854         return ret;
855 }
856
857 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
858                                 unsigned long addr, unsigned long end,
859                                 swp_entry_t entry, struct page *page)
860 {
861         pte_t swp_pte = swp_entry_to_pte(entry);
862         pte_t *pte;
863         int ret = 0;
864
865         /*
866          * We don't actually need pte lock while scanning for swp_pte: since
867          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
868          * page table while we're scanning; though it could get zapped, and on
869          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
870          * of unmatched parts which look like swp_pte, so unuse_pte must
871          * recheck under pte lock.  Scanning without pte lock lets it be
872          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
873          */
874         pte = pte_offset_map(pmd, addr);
875         do {
876                 /*
877                  * swapoff spends a _lot_ of time in this loop!
878                  * Test inline before going to call unuse_pte.
879                  */
880                 if (unlikely(pte_same(*pte, swp_pte))) {
881                         pte_unmap(pte);
882                         ret = unuse_pte(vma, pmd, addr, entry, page);
883                         if (ret)
884                                 goto out;
885                         pte = pte_offset_map(pmd, addr);
886                 }
887         } while (pte++, addr += PAGE_SIZE, addr != end);
888         pte_unmap(pte - 1);
889 out:
890         return ret;
891 }
892
893 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
894                                 unsigned long addr, unsigned long end,
895                                 swp_entry_t entry, struct page *page)
896 {
897         pmd_t *pmd;
898         unsigned long next;
899         int ret;
900
901         pmd = pmd_offset(pud, addr);
902         do {
903                 next = pmd_addr_end(addr, end);
904                 if (pmd_none_or_clear_bad(pmd))
905                         continue;
906                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
907                 if (ret)
908                         return ret;
909         } while (pmd++, addr = next, addr != end);
910         return 0;
911 }
912
913 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
914                                 unsigned long addr, unsigned long end,
915                                 swp_entry_t entry, struct page *page)
916 {
917         pud_t *pud;
918         unsigned long next;
919         int ret;
920
921         pud = pud_offset(pgd, addr);
922         do {
923                 next = pud_addr_end(addr, end);
924                 if (pud_none_or_clear_bad(pud))
925                         continue;
926                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
927                 if (ret)
928                         return ret;
929         } while (pud++, addr = next, addr != end);
930         return 0;
931 }
932
933 static int unuse_vma(struct vm_area_struct *vma,
934                                 swp_entry_t entry, struct page *page)
935 {
936         pgd_t *pgd;
937         unsigned long addr, end, next;
938         int ret;
939
940         if (page->mapping) {
941                 addr = page_address_in_vma(page, vma);
942                 if (addr == -EFAULT)
943                         return 0;
944                 else
945                         end = addr + PAGE_SIZE;
946         } else {
947                 addr = vma->vm_start;
948                 end = vma->vm_end;
949         }
950
951         pgd = pgd_offset(vma->vm_mm, addr);
952         do {
953                 next = pgd_addr_end(addr, end);
954                 if (pgd_none_or_clear_bad(pgd))
955                         continue;
956                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
957                 if (ret)
958                         return ret;
959         } while (pgd++, addr = next, addr != end);
960         return 0;
961 }
962
963 static int unuse_mm(struct mm_struct *mm,
964                                 swp_entry_t entry, struct page *page)
965 {
966         struct vm_area_struct *vma;
967         int ret = 0;
968
969         if (!down_read_trylock(&mm->mmap_sem)) {
970                 /*
971                  * Activate page so shrink_inactive_list is unlikely to unmap
972                  * its ptes while lock is dropped, so swapoff can make progress.
973                  */
974                 activate_page(page);
975                 unlock_page(page);
976                 down_read(&mm->mmap_sem);
977                 lock_page(page);
978         }
979         for (vma = mm->mmap; vma; vma = vma->vm_next) {
980                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
981                         break;
982         }
983         up_read(&mm->mmap_sem);
984         return (ret < 0)? ret: 0;
985 }
986
987 /*
988  * Scan swap_map from current position to next entry still in use.
989  * Recycle to start on reaching the end, returning 0 when empty.
990  */
991 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
992                                         unsigned int prev)
993 {
994         unsigned int max = si->max;
995         unsigned int i = prev;
996         unsigned char count;
997
998         /*
999          * No need for swap_lock here: we're just looking
1000          * for whether an entry is in use, not modifying it; false
1001          * hits are okay, and sys_swapoff() has already prevented new
1002          * allocations from this area (while holding swap_lock).
1003          */
1004         for (;;) {
1005                 if (++i >= max) {
1006                         if (!prev) {
1007                                 i = 0;
1008                                 break;
1009                         }
1010                         /*
1011                          * No entries in use at top of swap_map,
1012                          * loop back to start and recheck there.
1013                          */
1014                         max = prev + 1;
1015                         prev = 0;
1016                         i = 1;
1017                 }
1018                 count = si->swap_map[i];
1019                 if (count && swap_count(count) != SWAP_MAP_BAD)
1020                         break;
1021         }
1022         return i;
1023 }
1024
1025 /*
1026  * We completely avoid races by reading each swap page in advance,
1027  * and then search for the process using it.  All the necessary
1028  * page table adjustments can then be made atomically.
1029  */
1030 static int try_to_unuse(unsigned int type)
1031 {
1032         struct swap_info_struct *si = swap_info[type];
1033         struct mm_struct *start_mm;
1034         unsigned char *swap_map;
1035         unsigned char swcount;
1036         struct page *page;
1037         swp_entry_t entry;
1038         unsigned int i = 0;
1039         int retval = 0;
1040
1041         /*
1042          * When searching mms for an entry, a good strategy is to
1043          * start at the first mm we freed the previous entry from
1044          * (though actually we don't notice whether we or coincidence
1045          * freed the entry).  Initialize this start_mm with a hold.
1046          *
1047          * A simpler strategy would be to start at the last mm we
1048          * freed the previous entry from; but that would take less
1049          * advantage of mmlist ordering, which clusters forked mms
1050          * together, child after parent.  If we race with dup_mmap(), we
1051          * prefer to resolve parent before child, lest we miss entries
1052          * duplicated after we scanned child: using last mm would invert
1053          * that.
1054          */
1055         start_mm = &init_mm;
1056         atomic_inc(&init_mm.mm_users);
1057
1058         /*
1059          * Keep on scanning until all entries have gone.  Usually,
1060          * one pass through swap_map is enough, but not necessarily:
1061          * there are races when an instance of an entry might be missed.
1062          */
1063         while ((i = find_next_to_unuse(si, i)) != 0) {
1064                 if (signal_pending(current)) {
1065                         retval = -EINTR;
1066                         break;
1067                 }
1068
1069                 /*
1070                  * Get a page for the entry, using the existing swap
1071                  * cache page if there is one.  Otherwise, get a clean
1072                  * page and read the swap into it.
1073                  */
1074                 swap_map = &si->swap_map[i];
1075                 entry = swp_entry(type, i);
1076                 page = read_swap_cache_async(entry,
1077                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1078                 if (!page) {
1079                         /*
1080                          * Either swap_duplicate() failed because entry
1081                          * has been freed independently, and will not be
1082                          * reused since sys_swapoff() already disabled
1083                          * allocation from here, or alloc_page() failed.
1084                          */
1085                         if (!*swap_map)
1086                                 continue;
1087                         retval = -ENOMEM;
1088                         break;
1089                 }
1090
1091                 /*
1092                  * Don't hold on to start_mm if it looks like exiting.
1093                  */
1094                 if (atomic_read(&start_mm->mm_users) == 1) {
1095                         mmput(start_mm);
1096                         start_mm = &init_mm;
1097                         atomic_inc(&init_mm.mm_users);
1098                 }
1099
1100                 /*
1101                  * Wait for and lock page.  When do_swap_page races with
1102                  * try_to_unuse, do_swap_page can handle the fault much
1103                  * faster than try_to_unuse can locate the entry.  This
1104                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1105                  * defer to do_swap_page in such a case - in some tests,
1106                  * do_swap_page and try_to_unuse repeatedly compete.
1107                  */
1108                 wait_on_page_locked(page);
1109                 wait_on_page_writeback(page);
1110                 lock_page(page);
1111                 wait_on_page_writeback(page);
1112
1113                 /*
1114                  * Remove all references to entry.
1115                  */
1116                 swcount = *swap_map;
1117                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1118                         retval = shmem_unuse(entry, page);
1119                         /* page has already been unlocked and released */
1120                         if (retval < 0)
1121                                 break;
1122                         continue;
1123                 }
1124                 if (swap_count(swcount) && start_mm != &init_mm)
1125                         retval = unuse_mm(start_mm, entry, page);
1126
1127                 if (swap_count(*swap_map)) {
1128                         int set_start_mm = (*swap_map >= swcount);
1129                         struct list_head *p = &start_mm->mmlist;
1130                         struct mm_struct *new_start_mm = start_mm;
1131                         struct mm_struct *prev_mm = start_mm;
1132                         struct mm_struct *mm;
1133
1134                         atomic_inc(&new_start_mm->mm_users);
1135                         atomic_inc(&prev_mm->mm_users);
1136                         spin_lock(&mmlist_lock);
1137                         while (swap_count(*swap_map) && !retval &&
1138                                         (p = p->next) != &start_mm->mmlist) {
1139                                 mm = list_entry(p, struct mm_struct, mmlist);
1140                                 if (!atomic_inc_not_zero(&mm->mm_users))
1141                                         continue;
1142                                 spin_unlock(&mmlist_lock);
1143                                 mmput(prev_mm);
1144                                 prev_mm = mm;
1145
1146                                 cond_resched();
1147
1148                                 swcount = *swap_map;
1149                                 if (!swap_count(swcount)) /* any usage ? */
1150                                         ;
1151                                 else if (mm == &init_mm)
1152                                         set_start_mm = 1;
1153                                 else
1154                                         retval = unuse_mm(mm, entry, page);
1155
1156                                 if (set_start_mm && *swap_map < swcount) {
1157                                         mmput(new_start_mm);
1158                                         atomic_inc(&mm->mm_users);
1159                                         new_start_mm = mm;
1160                                         set_start_mm = 0;
1161                                 }
1162                                 spin_lock(&mmlist_lock);
1163                         }
1164                         spin_unlock(&mmlist_lock);
1165                         mmput(prev_mm);
1166                         mmput(start_mm);
1167                         start_mm = new_start_mm;
1168                 }
1169                 if (retval) {
1170                         unlock_page(page);
1171                         page_cache_release(page);
1172                         break;
1173                 }
1174
1175                 /*
1176                  * If a reference remains (rare), we would like to leave
1177                  * the page in the swap cache; but try_to_unmap could
1178                  * then re-duplicate the entry once we drop page lock,
1179                  * so we might loop indefinitely; also, that page could
1180                  * not be swapped out to other storage meanwhile.  So:
1181                  * delete from cache even if there's another reference,
1182                  * after ensuring that the data has been saved to disk -
1183                  * since if the reference remains (rarer), it will be
1184                  * read from disk into another page.  Splitting into two
1185                  * pages would be incorrect if swap supported "shared
1186                  * private" pages, but they are handled by tmpfs files.
1187                  */
1188                 if (swap_count(*swap_map) &&
1189                      PageDirty(page) && PageSwapCache(page)) {
1190                         struct writeback_control wbc = {
1191                                 .sync_mode = WB_SYNC_NONE,
1192                         };
1193
1194                         swap_writepage(page, &wbc);
1195                         lock_page(page);
1196                         wait_on_page_writeback(page);
1197                 }
1198
1199                 /*
1200                  * It is conceivable that a racing task removed this page from
1201                  * swap cache just before we acquired the page lock at the top,
1202                  * or while we dropped it in unuse_mm().  The page might even
1203                  * be back in swap cache on another swap area: that we must not
1204                  * delete, since it may not have been written out to swap yet.
1205                  */
1206                 if (PageSwapCache(page) &&
1207                     likely(page_private(page) == entry.val))
1208                         delete_from_swap_cache(page);
1209
1210                 /*
1211                  * So we could skip searching mms once swap count went
1212                  * to 1, we did not mark any present ptes as dirty: must
1213                  * mark page dirty so shrink_page_list will preserve it.
1214                  */
1215                 SetPageDirty(page);
1216                 unlock_page(page);
1217                 page_cache_release(page);
1218
1219                 /*
1220                  * Make sure that we aren't completely killing
1221                  * interactive performance.
1222                  */
1223                 cond_resched();
1224         }
1225
1226         mmput(start_mm);
1227         return retval;
1228 }
1229
1230 /*
1231  * After a successful try_to_unuse, if no swap is now in use, we know
1232  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1233  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1234  * added to the mmlist just after page_duplicate - before would be racy.
1235  */
1236 static void drain_mmlist(void)
1237 {
1238         struct list_head *p, *next;
1239         unsigned int type;
1240
1241         for (type = 0; type < nr_swapfiles; type++)
1242                 if (swap_info[type]->inuse_pages)
1243                         return;
1244         spin_lock(&mmlist_lock);
1245         list_for_each_safe(p, next, &init_mm.mmlist)
1246                 list_del_init(p);
1247         spin_unlock(&mmlist_lock);
1248 }
1249
1250 /*
1251  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1252  * corresponds to page offset `offset'.  Note that the type of this function
1253  * is sector_t, but it returns page offset into the bdev, not sector offset.
1254  */
1255 sector_t map_swap_page(swp_entry_t entry, struct block_device **bdev)
1256 {
1257         struct swap_info_struct *sis;
1258         struct swap_extent *start_se;
1259         struct swap_extent *se;
1260         pgoff_t offset;
1261
1262         sis = swap_info[swp_type(entry)];
1263         *bdev = sis->bdev;
1264
1265         offset = swp_offset(entry);
1266         start_se = sis->curr_swap_extent;
1267         se = start_se;
1268
1269         for ( ; ; ) {
1270                 struct list_head *lh;
1271
1272                 if (se->start_page <= offset &&
1273                                 offset < (se->start_page + se->nr_pages)) {
1274                         return se->start_block + (offset - se->start_page);
1275                 }
1276                 lh = se->list.next;
1277                 se = list_entry(lh, struct swap_extent, list);
1278                 sis->curr_swap_extent = se;
1279                 BUG_ON(se == start_se);         /* It *must* be present */
1280         }
1281 }
1282
1283 /*
1284  * Free all of a swapdev's extent information
1285  */
1286 static void destroy_swap_extents(struct swap_info_struct *sis)
1287 {
1288         while (!list_empty(&sis->first_swap_extent.list)) {
1289                 struct swap_extent *se;
1290
1291                 se = list_entry(sis->first_swap_extent.list.next,
1292                                 struct swap_extent, list);
1293                 list_del(&se->list);
1294                 kfree(se);
1295         }
1296 }
1297
1298 /*
1299  * Add a block range (and the corresponding page range) into this swapdev's
1300  * extent list.  The extent list is kept sorted in page order.
1301  *
1302  * This function rather assumes that it is called in ascending page order.
1303  */
1304 static int
1305 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1306                 unsigned long nr_pages, sector_t start_block)
1307 {
1308         struct swap_extent *se;
1309         struct swap_extent *new_se;
1310         struct list_head *lh;
1311
1312         if (start_page == 0) {
1313                 se = &sis->first_swap_extent;
1314                 sis->curr_swap_extent = se;
1315                 se->start_page = 0;
1316                 se->nr_pages = nr_pages;
1317                 se->start_block = start_block;
1318                 return 1;
1319         } else {
1320                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1321                 se = list_entry(lh, struct swap_extent, list);
1322                 BUG_ON(se->start_page + se->nr_pages != start_page);
1323                 if (se->start_block + se->nr_pages == start_block) {
1324                         /* Merge it */
1325                         se->nr_pages += nr_pages;
1326                         return 0;
1327                 }
1328         }
1329
1330         /*
1331          * No merge.  Insert a new extent, preserving ordering.
1332          */
1333         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1334         if (new_se == NULL)
1335                 return -ENOMEM;
1336         new_se->start_page = start_page;
1337         new_se->nr_pages = nr_pages;
1338         new_se->start_block = start_block;
1339
1340         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1341         return 1;
1342 }
1343
1344 /*
1345  * A `swap extent' is a simple thing which maps a contiguous range of pages
1346  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1347  * is built at swapon time and is then used at swap_writepage/swap_readpage
1348  * time for locating where on disk a page belongs.
1349  *
1350  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1351  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1352  * swap files identically.
1353  *
1354  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1355  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1356  * swapfiles are handled *identically* after swapon time.
1357  *
1358  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1359  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1360  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1361  * requirements, they are simply tossed out - we will never use those blocks
1362  * for swapping.
1363  *
1364  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1365  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1366  * which will scribble on the fs.
1367  *
1368  * The amount of disk space which a single swap extent represents varies.
1369  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1370  * extents in the list.  To avoid much list walking, we cache the previous
1371  * search location in `curr_swap_extent', and start new searches from there.
1372  * This is extremely effective.  The average number of iterations in
1373  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1374  */
1375 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1376 {
1377         struct inode *inode;
1378         unsigned blocks_per_page;
1379         unsigned long page_no;
1380         unsigned blkbits;
1381         sector_t probe_block;
1382         sector_t last_block;
1383         sector_t lowest_block = -1;
1384         sector_t highest_block = 0;
1385         int nr_extents = 0;
1386         int ret;
1387
1388         inode = sis->swap_file->f_mapping->host;
1389         if (S_ISBLK(inode->i_mode)) {
1390                 ret = add_swap_extent(sis, 0, sis->max, 0);
1391                 *span = sis->pages;
1392                 goto out;
1393         }
1394
1395         blkbits = inode->i_blkbits;
1396         blocks_per_page = PAGE_SIZE >> blkbits;
1397
1398         /*
1399          * Map all the blocks into the extent list.  This code doesn't try
1400          * to be very smart.
1401          */
1402         probe_block = 0;
1403         page_no = 0;
1404         last_block = i_size_read(inode) >> blkbits;
1405         while ((probe_block + blocks_per_page) <= last_block &&
1406                         page_no < sis->max) {
1407                 unsigned block_in_page;
1408                 sector_t first_block;
1409
1410                 first_block = bmap(inode, probe_block);
1411                 if (first_block == 0)
1412                         goto bad_bmap;
1413
1414                 /*
1415                  * It must be PAGE_SIZE aligned on-disk
1416                  */
1417                 if (first_block & (blocks_per_page - 1)) {
1418                         probe_block++;
1419                         goto reprobe;
1420                 }
1421
1422                 for (block_in_page = 1; block_in_page < blocks_per_page;
1423                                         block_in_page++) {
1424                         sector_t block;
1425
1426                         block = bmap(inode, probe_block + block_in_page);
1427                         if (block == 0)
1428                                 goto bad_bmap;
1429                         if (block != first_block + block_in_page) {
1430                                 /* Discontiguity */
1431                                 probe_block++;
1432                                 goto reprobe;
1433                         }
1434                 }
1435
1436                 first_block >>= (PAGE_SHIFT - blkbits);
1437                 if (page_no) {  /* exclude the header page */
1438                         if (first_block < lowest_block)
1439                                 lowest_block = first_block;
1440                         if (first_block > highest_block)
1441                                 highest_block = first_block;
1442                 }
1443
1444                 /*
1445                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1446                  */
1447                 ret = add_swap_extent(sis, page_no, 1, first_block);
1448                 if (ret < 0)
1449                         goto out;
1450                 nr_extents += ret;
1451                 page_no++;
1452                 probe_block += blocks_per_page;
1453 reprobe:
1454                 continue;
1455         }
1456         ret = nr_extents;
1457         *span = 1 + highest_block - lowest_block;
1458         if (page_no == 0)
1459                 page_no = 1;    /* force Empty message */
1460         sis->max = page_no;
1461         sis->pages = page_no - 1;
1462         sis->highest_bit = page_no - 1;
1463 out:
1464         return ret;
1465 bad_bmap:
1466         printk(KERN_ERR "swapon: swapfile has holes\n");
1467         ret = -EINVAL;
1468         goto out;
1469 }
1470
1471 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1472 {
1473         struct swap_info_struct *p = NULL;
1474         unsigned char *swap_map;
1475         struct file *swap_file, *victim;
1476         struct address_space *mapping;
1477         struct inode *inode;
1478         char *pathname;
1479         int i, type, prev;
1480         int err;
1481
1482         if (!capable(CAP_SYS_ADMIN))
1483                 return -EPERM;
1484
1485         pathname = getname(specialfile);
1486         err = PTR_ERR(pathname);
1487         if (IS_ERR(pathname))
1488                 goto out;
1489
1490         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1491         putname(pathname);
1492         err = PTR_ERR(victim);
1493         if (IS_ERR(victim))
1494                 goto out;
1495
1496         mapping = victim->f_mapping;
1497         prev = -1;
1498         spin_lock(&swap_lock);
1499         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1500                 p = swap_info[type];
1501                 if (p->flags & SWP_WRITEOK) {
1502                         if (p->swap_file->f_mapping == mapping)
1503                                 break;
1504                 }
1505                 prev = type;
1506         }
1507         if (type < 0) {
1508                 err = -EINVAL;
1509                 spin_unlock(&swap_lock);
1510                 goto out_dput;
1511         }
1512         if (!security_vm_enough_memory(p->pages))
1513                 vm_unacct_memory(p->pages);
1514         else {
1515                 err = -ENOMEM;
1516                 spin_unlock(&swap_lock);
1517                 goto out_dput;
1518         }
1519         if (prev < 0)
1520                 swap_list.head = p->next;
1521         else
1522                 swap_info[prev]->next = p->next;
1523         if (type == swap_list.next) {
1524                 /* just pick something that's safe... */
1525                 swap_list.next = swap_list.head;
1526         }
1527         if (p->prio < 0) {
1528                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1529                         swap_info[i]->prio = p->prio--;
1530                 least_priority++;
1531         }
1532         nr_swap_pages -= p->pages;
1533         total_swap_pages -= p->pages;
1534         p->flags &= ~SWP_WRITEOK;
1535         spin_unlock(&swap_lock);
1536
1537         current->flags |= PF_OOM_ORIGIN;
1538         err = try_to_unuse(type);
1539         current->flags &= ~PF_OOM_ORIGIN;
1540
1541         if (err) {
1542                 /* re-insert swap space back into swap_list */
1543                 spin_lock(&swap_lock);
1544                 if (p->prio < 0)
1545                         p->prio = --least_priority;
1546                 prev = -1;
1547                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1548                         if (p->prio >= swap_info[i]->prio)
1549                                 break;
1550                         prev = i;
1551                 }
1552                 p->next = i;
1553                 if (prev < 0)
1554                         swap_list.head = swap_list.next = type;
1555                 else
1556                         swap_info[prev]->next = type;
1557                 nr_swap_pages += p->pages;
1558                 total_swap_pages += p->pages;
1559                 p->flags |= SWP_WRITEOK;
1560                 spin_unlock(&swap_lock);
1561                 goto out_dput;
1562         }
1563
1564         /* wait for any unplug function to finish */
1565         down_write(&swap_unplug_sem);
1566         up_write(&swap_unplug_sem);
1567
1568         destroy_swap_extents(p);
1569         if (p->flags & SWP_CONTINUED)
1570                 free_swap_count_continuations(p);
1571
1572         mutex_lock(&swapon_mutex);
1573         spin_lock(&swap_lock);
1574         drain_mmlist();
1575
1576         /* wait for anyone still in scan_swap_map */
1577         p->highest_bit = 0;             /* cuts scans short */
1578         while (p->flags >= SWP_SCANNING) {
1579                 spin_unlock(&swap_lock);
1580                 schedule_timeout_uninterruptible(1);
1581                 spin_lock(&swap_lock);
1582         }
1583
1584         swap_file = p->swap_file;
1585         p->swap_file = NULL;
1586         p->max = 0;
1587         swap_map = p->swap_map;
1588         p->swap_map = NULL;
1589         p->flags = 0;
1590         spin_unlock(&swap_lock);
1591         mutex_unlock(&swapon_mutex);
1592         vfree(swap_map);
1593         /* Destroy swap account informatin */
1594         swap_cgroup_swapoff(type);
1595
1596         inode = mapping->host;
1597         if (S_ISBLK(inode->i_mode)) {
1598                 struct block_device *bdev = I_BDEV(inode);
1599                 set_blocksize(bdev, p->old_block_size);
1600                 bd_release(bdev);
1601         } else {
1602                 mutex_lock(&inode->i_mutex);
1603                 inode->i_flags &= ~S_SWAPFILE;
1604                 mutex_unlock(&inode->i_mutex);
1605         }
1606         filp_close(swap_file, NULL);
1607         err = 0;
1608
1609 out_dput:
1610         filp_close(victim, NULL);
1611 out:
1612         return err;
1613 }
1614
1615 #ifdef CONFIG_PROC_FS
1616 /* iterator */
1617 static void *swap_start(struct seq_file *swap, loff_t *pos)
1618 {
1619         struct swap_info_struct *si;
1620         int type;
1621         loff_t l = *pos;
1622
1623         mutex_lock(&swapon_mutex);
1624
1625         if (!l)
1626                 return SEQ_START_TOKEN;
1627
1628         for (type = 0; type < nr_swapfiles; type++) {
1629                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1630                 si = swap_info[type];
1631                 if (!(si->flags & SWP_USED) || !si->swap_map)
1632                         continue;
1633                 if (!--l)
1634                         return si;
1635         }
1636
1637         return NULL;
1638 }
1639
1640 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1641 {
1642         struct swap_info_struct *si = v;
1643         int type;
1644
1645         if (v == SEQ_START_TOKEN)
1646                 type = 0;
1647         else
1648                 type = si->type + 1;
1649
1650         for (; type < nr_swapfiles; type++) {
1651                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1652                 si = swap_info[type];
1653                 if (!(si->flags & SWP_USED) || !si->swap_map)
1654                         continue;
1655                 ++*pos;
1656                 return si;
1657         }
1658
1659         return NULL;
1660 }
1661
1662 static void swap_stop(struct seq_file *swap, void *v)
1663 {
1664         mutex_unlock(&swapon_mutex);
1665 }
1666
1667 static int swap_show(struct seq_file *swap, void *v)
1668 {
1669         struct swap_info_struct *si = v;
1670         struct file *file;
1671         int len;
1672
1673         if (si == SEQ_START_TOKEN) {
1674                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1675                 return 0;
1676         }
1677
1678         file = si->swap_file;
1679         len = seq_path(swap, &file->f_path, " \t\n\\");
1680         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1681                         len < 40 ? 40 - len : 1, " ",
1682                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1683                                 "partition" : "file\t",
1684                         si->pages << (PAGE_SHIFT - 10),
1685                         si->inuse_pages << (PAGE_SHIFT - 10),
1686                         si->prio);
1687         return 0;
1688 }
1689
1690 static const struct seq_operations swaps_op = {
1691         .start =        swap_start,
1692         .next =         swap_next,
1693         .stop =         swap_stop,
1694         .show =         swap_show
1695 };
1696
1697 static int swaps_open(struct inode *inode, struct file *file)
1698 {
1699         return seq_open(file, &swaps_op);
1700 }
1701
1702 static const struct file_operations proc_swaps_operations = {
1703         .open           = swaps_open,
1704         .read           = seq_read,
1705         .llseek         = seq_lseek,
1706         .release        = seq_release,
1707 };
1708
1709 static int __init procswaps_init(void)
1710 {
1711         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1712         return 0;
1713 }
1714 __initcall(procswaps_init);
1715 #endif /* CONFIG_PROC_FS */
1716
1717 #ifdef MAX_SWAPFILES_CHECK
1718 static int __init max_swapfiles_check(void)
1719 {
1720         MAX_SWAPFILES_CHECK();
1721         return 0;
1722 }
1723 late_initcall(max_swapfiles_check);
1724 #endif
1725
1726 /*
1727  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1728  *
1729  * The swapon system call
1730  */
1731 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1732 {
1733         struct swap_info_struct *p;
1734         char *name = NULL;
1735         struct block_device *bdev = NULL;
1736         struct file *swap_file = NULL;
1737         struct address_space *mapping;
1738         unsigned int type;
1739         int i, prev;
1740         int error;
1741         union swap_header *swap_header = NULL;
1742         unsigned int nr_good_pages = 0;
1743         int nr_extents = 0;
1744         sector_t span;
1745         unsigned long maxpages = 1;
1746         unsigned long swapfilepages;
1747         unsigned char *swap_map = NULL;
1748         struct page *page = NULL;
1749         struct inode *inode = NULL;
1750         int did_down = 0;
1751
1752         if (!capable(CAP_SYS_ADMIN))
1753                 return -EPERM;
1754
1755         p = kzalloc(sizeof(*p), GFP_KERNEL);
1756         if (!p)
1757                 return -ENOMEM;
1758
1759         spin_lock(&swap_lock);
1760         for (type = 0; type < nr_swapfiles; type++) {
1761                 if (!(swap_info[type]->flags & SWP_USED))
1762                         break;
1763         }
1764         error = -EPERM;
1765         if (type >= MAX_SWAPFILES) {
1766                 spin_unlock(&swap_lock);
1767                 kfree(p);
1768                 goto out;
1769         }
1770         if (type >= nr_swapfiles) {
1771                 p->type = type;
1772                 swap_info[type] = p;
1773                 /*
1774                  * Write swap_info[type] before nr_swapfiles, in case a
1775                  * racing procfs swap_start() or swap_next() is reading them.
1776                  * (We never shrink nr_swapfiles, we never free this entry.)
1777                  */
1778                 smp_wmb();
1779                 nr_swapfiles++;
1780         } else {
1781                 kfree(p);
1782                 p = swap_info[type];
1783                 /*
1784                  * Do not memset this entry: a racing procfs swap_next()
1785                  * would be relying on p->type to remain valid.
1786                  */
1787         }
1788         INIT_LIST_HEAD(&p->first_swap_extent.list);
1789         p->flags = SWP_USED;
1790         p->next = -1;
1791         spin_unlock(&swap_lock);
1792
1793         name = getname(specialfile);
1794         error = PTR_ERR(name);
1795         if (IS_ERR(name)) {
1796                 name = NULL;
1797                 goto bad_swap_2;
1798         }
1799         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1800         error = PTR_ERR(swap_file);
1801         if (IS_ERR(swap_file)) {
1802                 swap_file = NULL;
1803                 goto bad_swap_2;
1804         }
1805
1806         p->swap_file = swap_file;
1807         mapping = swap_file->f_mapping;
1808         inode = mapping->host;
1809
1810         error = -EBUSY;
1811         for (i = 0; i < nr_swapfiles; i++) {
1812                 struct swap_info_struct *q = swap_info[i];
1813
1814                 if (i == type || !q->swap_file)
1815                         continue;
1816                 if (mapping == q->swap_file->f_mapping)
1817                         goto bad_swap;
1818         }
1819
1820         error = -EINVAL;
1821         if (S_ISBLK(inode->i_mode)) {
1822                 bdev = I_BDEV(inode);
1823                 error = bd_claim(bdev, sys_swapon);
1824                 if (error < 0) {
1825                         bdev = NULL;
1826                         error = -EINVAL;
1827                         goto bad_swap;
1828                 }
1829                 p->old_block_size = block_size(bdev);
1830                 error = set_blocksize(bdev, PAGE_SIZE);
1831                 if (error < 0)
1832                         goto bad_swap;
1833                 p->bdev = bdev;
1834         } else if (S_ISREG(inode->i_mode)) {
1835                 p->bdev = inode->i_sb->s_bdev;
1836                 mutex_lock(&inode->i_mutex);
1837                 did_down = 1;
1838                 if (IS_SWAPFILE(inode)) {
1839                         error = -EBUSY;
1840                         goto bad_swap;
1841                 }
1842         } else {
1843                 goto bad_swap;
1844         }
1845
1846         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1847
1848         /*
1849          * Read the swap header.
1850          */
1851         if (!mapping->a_ops->readpage) {
1852                 error = -EINVAL;
1853                 goto bad_swap;
1854         }
1855         page = read_mapping_page(mapping, 0, swap_file);
1856         if (IS_ERR(page)) {
1857                 error = PTR_ERR(page);
1858                 goto bad_swap;
1859         }
1860         swap_header = kmap(page);
1861
1862         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1863                 printk(KERN_ERR "Unable to find swap-space signature\n");
1864                 error = -EINVAL;
1865                 goto bad_swap;
1866         }
1867
1868         /* swap partition endianess hack... */
1869         if (swab32(swap_header->info.version) == 1) {
1870                 swab32s(&swap_header->info.version);
1871                 swab32s(&swap_header->info.last_page);
1872                 swab32s(&swap_header->info.nr_badpages);
1873                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1874                         swab32s(&swap_header->info.badpages[i]);
1875         }
1876         /* Check the swap header's sub-version */
1877         if (swap_header->info.version != 1) {
1878                 printk(KERN_WARNING
1879                        "Unable to handle swap header version %d\n",
1880                        swap_header->info.version);
1881                 error = -EINVAL;
1882                 goto bad_swap;
1883         }
1884
1885         p->lowest_bit  = 1;
1886         p->cluster_next = 1;
1887         p->cluster_nr = 0;
1888
1889         /*
1890          * Find out how many pages are allowed for a single swap
1891          * device. There are two limiting factors: 1) the number of
1892          * bits for the swap offset in the swp_entry_t type and
1893          * 2) the number of bits in the a swap pte as defined by
1894          * the different architectures. In order to find the
1895          * largest possible bit mask a swap entry with swap type 0
1896          * and swap offset ~0UL is created, encoded to a swap pte,
1897          * decoded to a swp_entry_t again and finally the swap
1898          * offset is extracted. This will mask all the bits from
1899          * the initial ~0UL mask that can't be encoded in either
1900          * the swp_entry_t or the architecture definition of a
1901          * swap pte.
1902          */
1903         maxpages = swp_offset(pte_to_swp_entry(
1904                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1905         if (maxpages > swap_header->info.last_page)
1906                 maxpages = swap_header->info.last_page;
1907         p->highest_bit = maxpages - 1;
1908
1909         error = -EINVAL;
1910         if (!maxpages)
1911                 goto bad_swap;
1912         if (swapfilepages && maxpages > swapfilepages) {
1913                 printk(KERN_WARNING
1914                        "Swap area shorter than signature indicates\n");
1915                 goto bad_swap;
1916         }
1917         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1918                 goto bad_swap;
1919         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1920                 goto bad_swap;
1921
1922         /* OK, set up the swap map and apply the bad block list */
1923         swap_map = vmalloc(maxpages);
1924         if (!swap_map) {
1925                 error = -ENOMEM;
1926                 goto bad_swap;
1927         }
1928
1929         memset(swap_map, 0, maxpages);
1930         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1931                 int page_nr = swap_header->info.badpages[i];
1932                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1933                         error = -EINVAL;
1934                         goto bad_swap;
1935                 }
1936                 swap_map[page_nr] = SWAP_MAP_BAD;
1937         }
1938
1939         error = swap_cgroup_swapon(type, maxpages);
1940         if (error)
1941                 goto bad_swap;
1942
1943         nr_good_pages = swap_header->info.last_page -
1944                         swap_header->info.nr_badpages -
1945                         1 /* header page */;
1946
1947         if (nr_good_pages) {
1948                 swap_map[0] = SWAP_MAP_BAD;
1949                 p->max = maxpages;
1950                 p->pages = nr_good_pages;
1951                 nr_extents = setup_swap_extents(p, &span);
1952                 if (nr_extents < 0) {
1953                         error = nr_extents;
1954                         goto bad_swap;
1955                 }
1956                 nr_good_pages = p->pages;
1957         }
1958         if (!nr_good_pages) {
1959                 printk(KERN_WARNING "Empty swap-file\n");
1960                 error = -EINVAL;
1961                 goto bad_swap;
1962         }
1963
1964         if (p->bdev) {
1965                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1966                         p->flags |= SWP_SOLIDSTATE;
1967                         p->cluster_next = 1 + (random32() % p->highest_bit);
1968                 }
1969                 if (discard_swap(p) == 0)
1970                         p->flags |= SWP_DISCARDABLE;
1971         }
1972
1973         mutex_lock(&swapon_mutex);
1974         spin_lock(&swap_lock);
1975         if (swap_flags & SWAP_FLAG_PREFER)
1976                 p->prio =
1977                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1978         else
1979                 p->prio = --least_priority;
1980         p->swap_map = swap_map;
1981         p->flags |= SWP_WRITEOK;
1982         nr_swap_pages += nr_good_pages;
1983         total_swap_pages += nr_good_pages;
1984
1985         printk(KERN_INFO "Adding %uk swap on %s.  "
1986                         "Priority:%d extents:%d across:%lluk %s%s\n",
1987                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1988                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1989                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1990                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1991
1992         /* insert swap space into swap_list: */
1993         prev = -1;
1994         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1995                 if (p->prio >= swap_info[i]->prio)
1996                         break;
1997                 prev = i;
1998         }
1999         p->next = i;
2000         if (prev < 0)
2001                 swap_list.head = swap_list.next = type;
2002         else
2003                 swap_info[prev]->next = type;
2004         spin_unlock(&swap_lock);
2005         mutex_unlock(&swapon_mutex);
2006         error = 0;
2007         goto out;
2008 bad_swap:
2009         if (bdev) {
2010                 set_blocksize(bdev, p->old_block_size);
2011                 bd_release(bdev);
2012         }
2013         destroy_swap_extents(p);
2014         swap_cgroup_swapoff(type);
2015 bad_swap_2:
2016         spin_lock(&swap_lock);
2017         p->swap_file = NULL;
2018         p->flags = 0;
2019         spin_unlock(&swap_lock);
2020         vfree(swap_map);
2021         if (swap_file)
2022                 filp_close(swap_file, NULL);
2023 out:
2024         if (page && !IS_ERR(page)) {
2025                 kunmap(page);
2026                 page_cache_release(page);
2027         }
2028         if (name)
2029                 putname(name);
2030         if (did_down) {
2031                 if (!error)
2032                         inode->i_flags |= S_SWAPFILE;
2033                 mutex_unlock(&inode->i_mutex);
2034         }
2035         return error;
2036 }
2037
2038 void si_swapinfo(struct sysinfo *val)
2039 {
2040         unsigned int type;
2041         unsigned long nr_to_be_unused = 0;
2042
2043         spin_lock(&swap_lock);
2044         for (type = 0; type < nr_swapfiles; type++) {
2045                 struct swap_info_struct *si = swap_info[type];
2046
2047                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2048                         nr_to_be_unused += si->inuse_pages;
2049         }
2050         val->freeswap = nr_swap_pages + nr_to_be_unused;
2051         val->totalswap = total_swap_pages + nr_to_be_unused;
2052         spin_unlock(&swap_lock);
2053 }
2054
2055 /*
2056  * Verify that a swap entry is valid and increment its swap map count.
2057  *
2058  * Returns error code in following case.
2059  * - success -> 0
2060  * - swp_entry is invalid -> EINVAL
2061  * - swp_entry is migration entry -> EINVAL
2062  * - swap-cache reference is requested but there is already one. -> EEXIST
2063  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2064  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2065  */
2066 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2067 {
2068         struct swap_info_struct *p;
2069         unsigned long offset, type;
2070         unsigned char count;
2071         unsigned char has_cache;
2072         int err = -EINVAL;
2073
2074         if (non_swap_entry(entry))
2075                 goto out;
2076
2077         type = swp_type(entry);
2078         if (type >= nr_swapfiles)
2079                 goto bad_file;
2080         p = swap_info[type];
2081         offset = swp_offset(entry);
2082
2083         spin_lock(&swap_lock);
2084         if (unlikely(offset >= p->max))
2085                 goto unlock_out;
2086
2087         count = p->swap_map[offset];
2088         has_cache = count & SWAP_HAS_CACHE;
2089         count &= ~SWAP_HAS_CACHE;
2090         err = 0;
2091
2092         if (usage == SWAP_HAS_CACHE) {
2093
2094                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2095                 if (!has_cache && count)
2096                         has_cache = SWAP_HAS_CACHE;
2097                 else if (has_cache)             /* someone else added cache */
2098                         err = -EEXIST;
2099                 else                            /* no users remaining */
2100                         err = -ENOENT;
2101
2102         } else if (count || has_cache) {
2103
2104                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2105                         count += usage;
2106                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2107                         err = -EINVAL;
2108                 else if (swap_count_continued(p, offset, count))
2109                         count = COUNT_CONTINUED;
2110                 else
2111                         err = -ENOMEM;
2112         } else
2113                 err = -ENOENT;                  /* unused swap entry */
2114
2115         p->swap_map[offset] = count | has_cache;
2116
2117 unlock_out:
2118         spin_unlock(&swap_lock);
2119 out:
2120         return err;
2121
2122 bad_file:
2123         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2124         goto out;
2125 }
2126
2127 /*
2128  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2129  * (in which case its reference count is never incremented).
2130  */
2131 void swap_shmem_alloc(swp_entry_t entry)
2132 {
2133         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2134 }
2135
2136 /*
2137  * increase reference count of swap entry by 1.
2138  */
2139 int swap_duplicate(swp_entry_t entry)
2140 {
2141         int err = 0;
2142
2143         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2144                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2145         return err;
2146 }
2147
2148 /*
2149  * @entry: swap entry for which we allocate swap cache.
2150  *
2151  * Called when allocating swap cache for existing swap entry,
2152  * This can return error codes. Returns 0 at success.
2153  * -EBUSY means there is a swap cache.
2154  * Note: return code is different from swap_duplicate().
2155  */
2156 int swapcache_prepare(swp_entry_t entry)
2157 {
2158         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2159 }
2160
2161 /*
2162  * swap_lock prevents swap_map being freed. Don't grab an extra
2163  * reference on the swaphandle, it doesn't matter if it becomes unused.
2164  */
2165 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2166 {
2167         struct swap_info_struct *si;
2168         int our_page_cluster = page_cluster;
2169         pgoff_t target, toff;
2170         pgoff_t base, end;
2171         int nr_pages = 0;
2172
2173         if (!our_page_cluster)  /* no readahead */
2174                 return 0;
2175
2176         si = swap_info[swp_type(entry)];
2177         target = swp_offset(entry);
2178         base = (target >> our_page_cluster) << our_page_cluster;
2179         end = base + (1 << our_page_cluster);
2180         if (!base)              /* first page is swap header */
2181                 base++;
2182
2183         spin_lock(&swap_lock);
2184         if (end > si->max)      /* don't go beyond end of map */
2185                 end = si->max;
2186
2187         /* Count contiguous allocated slots above our target */
2188         for (toff = target; ++toff < end; nr_pages++) {
2189                 /* Don't read in free or bad pages */
2190                 if (!si->swap_map[toff])
2191                         break;
2192                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2193                         break;
2194         }
2195         /* Count contiguous allocated slots below our target */
2196         for (toff = target; --toff >= base; nr_pages++) {
2197                 /* Don't read in free or bad pages */
2198                 if (!si->swap_map[toff])
2199                         break;
2200                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2201                         break;
2202         }
2203         spin_unlock(&swap_lock);
2204
2205         /*
2206          * Indicate starting offset, and return number of pages to get:
2207          * if only 1, say 0, since there's then no readahead to be done.
2208          */
2209         *offset = ++toff;
2210         return nr_pages? ++nr_pages: 0;
2211 }
2212
2213 /*
2214  * add_swap_count_continuation - called when a swap count is duplicated
2215  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2216  * page of the original vmalloc'ed swap_map, to hold the continuation count
2217  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2218  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2219  *
2220  * These continuation pages are seldom referenced: the common paths all work
2221  * on the original swap_map, only referring to a continuation page when the
2222  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2223  *
2224  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2225  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2226  * can be called after dropping locks.
2227  */
2228 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2229 {
2230         struct swap_info_struct *si;
2231         struct page *head;
2232         struct page *page;
2233         struct page *list_page;
2234         pgoff_t offset;
2235         unsigned char count;
2236
2237         /*
2238          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2239          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2240          */
2241         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2242
2243         si = swap_info_get(entry);
2244         if (!si) {
2245                 /*
2246                  * An acceptable race has occurred since the failing
2247                  * __swap_duplicate(): the swap entry has been freed,
2248                  * perhaps even the whole swap_map cleared for swapoff.
2249                  */
2250                 goto outer;
2251         }
2252
2253         offset = swp_offset(entry);
2254         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2255
2256         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2257                 /*
2258                  * The higher the swap count, the more likely it is that tasks
2259                  * will race to add swap count continuation: we need to avoid
2260                  * over-provisioning.
2261                  */
2262                 goto out;
2263         }
2264
2265         if (!page) {
2266                 spin_unlock(&swap_lock);
2267                 return -ENOMEM;
2268         }
2269
2270         /*
2271          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2272          * no architecture is using highmem pages for kernel pagetables: so it
2273          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2274          */
2275         head = vmalloc_to_page(si->swap_map + offset);
2276         offset &= ~PAGE_MASK;
2277
2278         /*
2279          * Page allocation does not initialize the page's lru field,
2280          * but it does always reset its private field.
2281          */
2282         if (!page_private(head)) {
2283                 BUG_ON(count & COUNT_CONTINUED);
2284                 INIT_LIST_HEAD(&head->lru);
2285                 set_page_private(head, SWP_CONTINUED);
2286                 si->flags |= SWP_CONTINUED;
2287         }
2288
2289         list_for_each_entry(list_page, &head->lru, lru) {
2290                 unsigned char *map;
2291
2292                 /*
2293                  * If the previous map said no continuation, but we've found
2294                  * a continuation page, free our allocation and use this one.
2295                  */
2296                 if (!(count & COUNT_CONTINUED))
2297                         goto out;
2298
2299                 map = kmap_atomic(list_page, KM_USER0) + offset;
2300                 count = *map;
2301                 kunmap_atomic(map, KM_USER0);
2302
2303                 /*
2304                  * If this continuation count now has some space in it,
2305                  * free our allocation and use this one.
2306                  */
2307                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2308                         goto out;
2309         }
2310
2311         list_add_tail(&page->lru, &head->lru);
2312         page = NULL;                    /* now it's attached, don't free it */
2313 out:
2314         spin_unlock(&swap_lock);
2315 outer:
2316         if (page)
2317                 __free_page(page);
2318         return 0;
2319 }
2320
2321 /*
2322  * swap_count_continued - when the original swap_map count is incremented
2323  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2324  * into, carry if so, or else fail until a new continuation page is allocated;
2325  * when the original swap_map count is decremented from 0 with continuation,
2326  * borrow from the continuation and report whether it still holds more.
2327  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2328  */
2329 static bool swap_count_continued(struct swap_info_struct *si,
2330                                  pgoff_t offset, unsigned char count)
2331 {
2332         struct page *head;
2333         struct page *page;
2334         unsigned char *map;
2335
2336         head = vmalloc_to_page(si->swap_map + offset);
2337         if (page_private(head) != SWP_CONTINUED) {
2338                 BUG_ON(count & COUNT_CONTINUED);
2339                 return false;           /* need to add count continuation */
2340         }
2341
2342         offset &= ~PAGE_MASK;
2343         page = list_entry(head->lru.next, struct page, lru);
2344         map = kmap_atomic(page, KM_USER0) + offset;
2345
2346         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2347                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2348
2349         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2350                 /*
2351                  * Think of how you add 1 to 999
2352                  */
2353                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2354                         kunmap_atomic(map, KM_USER0);
2355                         page = list_entry(page->lru.next, struct page, lru);
2356                         BUG_ON(page == head);
2357                         map = kmap_atomic(page, KM_USER0) + offset;
2358                 }
2359                 if (*map == SWAP_CONT_MAX) {
2360                         kunmap_atomic(map, KM_USER0);
2361                         page = list_entry(page->lru.next, struct page, lru);
2362                         if (page == head)
2363                                 return false;   /* add count continuation */
2364                         map = kmap_atomic(page, KM_USER0) + offset;
2365 init_map:               *map = 0;               /* we didn't zero the page */
2366                 }
2367                 *map += 1;
2368                 kunmap_atomic(map, KM_USER0);
2369                 page = list_entry(page->lru.prev, struct page, lru);
2370                 while (page != head) {
2371                         map = kmap_atomic(page, KM_USER0) + offset;
2372                         *map = COUNT_CONTINUED;
2373                         kunmap_atomic(map, KM_USER0);
2374                         page = list_entry(page->lru.prev, struct page, lru);
2375                 }
2376                 return true;                    /* incremented */
2377
2378         } else {                                /* decrementing */
2379                 /*
2380                  * Think of how you subtract 1 from 1000
2381                  */
2382                 BUG_ON(count != COUNT_CONTINUED);
2383                 while (*map == COUNT_CONTINUED) {
2384                         kunmap_atomic(map, KM_USER0);
2385                         page = list_entry(page->lru.next, struct page, lru);
2386                         BUG_ON(page == head);
2387                         map = kmap_atomic(page, KM_USER0) + offset;
2388                 }
2389                 BUG_ON(*map == 0);
2390                 *map -= 1;
2391                 if (*map == 0)
2392                         count = 0;
2393                 kunmap_atomic(map, KM_USER0);
2394                 page = list_entry(page->lru.prev, struct page, lru);
2395                 while (page != head) {
2396                         map = kmap_atomic(page, KM_USER0) + offset;
2397                         *map = SWAP_CONT_MAX | count;
2398                         count = COUNT_CONTINUED;
2399                         kunmap_atomic(map, KM_USER0);
2400                         page = list_entry(page->lru.prev, struct page, lru);
2401                 }
2402                 return count == COUNT_CONTINUED;
2403         }
2404 }
2405
2406 /*
2407  * free_swap_count_continuations - swapoff free all the continuation pages
2408  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2409  */
2410 static void free_swap_count_continuations(struct swap_info_struct *si)
2411 {
2412         pgoff_t offset;
2413
2414         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2415                 struct page *head;
2416                 head = vmalloc_to_page(si->swap_map + offset);
2417                 if (page_private(head)) {
2418                         struct list_head *this, *next;
2419                         list_for_each_safe(this, next, &head->lru) {
2420                                 struct page *page;
2421                                 page = list_entry(this, struct page, lru);
2422                                 list_del(this);
2423                                 __free_page(page);
2424                         }
2425                 }
2426         }
2427 }