mm: fix corruption of hibernation caused by reusing swap during image saving
[sfrench/cifs-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/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/security.h>
28 #include <linux/backing-dev.h>
29 #include <linux/mutex.h>
30 #include <linux/capability.h>
31 #include <linux/syscalls.h>
32 #include <linux/memcontrol.h>
33
34 #include <asm/pgtable.h>
35 #include <asm/tlbflush.h>
36 #include <linux/swapops.h>
37 #include <linux/page_cgroup.h>
38
39 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
40                                  unsigned char);
41 static void free_swap_count_continuations(struct swap_info_struct *);
42 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
43
44 static DEFINE_SPINLOCK(swap_lock);
45 static unsigned int nr_swapfiles;
46 long nr_swap_pages;
47 long total_swap_pages;
48 static int least_priority;
49
50 static const char Bad_file[] = "Bad swap file entry ";
51 static const char Unused_file[] = "Unused swap file entry ";
52 static const char Bad_offset[] = "Bad swap offset entry ";
53 static const char Unused_offset[] = "Unused swap offset entry ";
54
55 static struct swap_list_t swap_list = {-1, -1};
56
57 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
58
59 static DEFINE_MUTEX(swapon_mutex);
60
61 static inline unsigned char swap_count(unsigned char ent)
62 {
63         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
64 }
65
66 /* returns 1 if swap entry is freed */
67 static int
68 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
69 {
70         swp_entry_t entry = swp_entry(si->type, offset);
71         struct page *page;
72         int ret = 0;
73
74         page = find_get_page(&swapper_space, entry.val);
75         if (!page)
76                 return 0;
77         /*
78          * This function is called from scan_swap_map() and it's called
79          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
80          * We have to use trylock for avoiding deadlock. This is a special
81          * case and you should use try_to_free_swap() with explicit lock_page()
82          * in usual operations.
83          */
84         if (trylock_page(page)) {
85                 ret = try_to_free_swap(page);
86                 unlock_page(page);
87         }
88         page_cache_release(page);
89         return ret;
90 }
91
92 /*
93  * We need this because the bdev->unplug_fn can sleep and we cannot
94  * hold swap_lock while calling the unplug_fn. And swap_lock
95  * cannot be turned into a mutex.
96  */
97 static DECLARE_RWSEM(swap_unplug_sem);
98
99 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
100 {
101         swp_entry_t entry;
102
103         down_read(&swap_unplug_sem);
104         entry.val = page_private(page);
105         if (PageSwapCache(page)) {
106                 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
107                 struct backing_dev_info *bdi;
108
109                 /*
110                  * If the page is removed from swapcache from under us (with a
111                  * racy try_to_unuse/swapoff) we need an additional reference
112                  * count to avoid reading garbage from page_private(page) above.
113                  * If the WARN_ON triggers during a swapoff it maybe the race
114                  * condition and it's harmless. However if it triggers without
115                  * swapoff it signals a problem.
116                  */
117                 WARN_ON(page_count(page) <= 1);
118
119                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
120                 blk_run_backing_dev(bdi, page);
121         }
122         up_read(&swap_unplug_sem);
123 }
124
125 /*
126  * swapon tell device that all the old swap contents can be discarded,
127  * to allow the swap device to optimize its wear-levelling.
128  */
129 static int discard_swap(struct swap_info_struct *si)
130 {
131         struct swap_extent *se;
132         sector_t start_block;
133         sector_t nr_blocks;
134         int err = 0;
135
136         /* Do not discard the swap header page! */
137         se = &si->first_swap_extent;
138         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
139         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
140         if (nr_blocks) {
141                 err = blkdev_issue_discard(si->bdev, start_block,
142                                 nr_blocks, GFP_KERNEL,
143                                 BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
144                 if (err)
145                         return err;
146                 cond_resched();
147         }
148
149         list_for_each_entry(se, &si->first_swap_extent.list, list) {
150                 start_block = se->start_block << (PAGE_SHIFT - 9);
151                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
152
153                 err = blkdev_issue_discard(si->bdev, start_block,
154                                 nr_blocks, GFP_KERNEL,
155                                 BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
156                 if (err)
157                         break;
158
159                 cond_resched();
160         }
161         return err;             /* That will often be -EOPNOTSUPP */
162 }
163
164 /*
165  * swap allocation tell device that a cluster of swap can now be discarded,
166  * to allow the swap device to optimize its wear-levelling.
167  */
168 static void discard_swap_cluster(struct swap_info_struct *si,
169                                  pgoff_t start_page, pgoff_t nr_pages)
170 {
171         struct swap_extent *se = si->curr_swap_extent;
172         int found_extent = 0;
173
174         while (nr_pages) {
175                 struct list_head *lh;
176
177                 if (se->start_page <= start_page &&
178                     start_page < se->start_page + se->nr_pages) {
179                         pgoff_t offset = start_page - se->start_page;
180                         sector_t start_block = se->start_block + offset;
181                         sector_t nr_blocks = se->nr_pages - offset;
182
183                         if (nr_blocks > nr_pages)
184                                 nr_blocks = nr_pages;
185                         start_page += nr_blocks;
186                         nr_pages -= nr_blocks;
187
188                         if (!found_extent++)
189                                 si->curr_swap_extent = se;
190
191                         start_block <<= PAGE_SHIFT - 9;
192                         nr_blocks <<= PAGE_SHIFT - 9;
193                         if (blkdev_issue_discard(si->bdev, start_block,
194                                     nr_blocks, GFP_NOIO, BLKDEV_IFL_WAIT |
195                                                         BLKDEV_IFL_BARRIER))
196                                 break;
197                 }
198
199                 lh = se->list.next;
200                 se = list_entry(lh, struct swap_extent, list);
201         }
202 }
203
204 static int wait_for_discard(void *word)
205 {
206         schedule();
207         return 0;
208 }
209
210 #define SWAPFILE_CLUSTER        256
211 #define LATENCY_LIMIT           256
212
213 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
214                                           unsigned char usage)
215 {
216         unsigned long offset;
217         unsigned long scan_base;
218         unsigned long last_in_cluster = 0;
219         int latency_ration = LATENCY_LIMIT;
220         int found_free_cluster = 0;
221
222         /*
223          * We try to cluster swap pages by allocating them sequentially
224          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
225          * way, however, we resort to first-free allocation, starting
226          * a new cluster.  This prevents us from scattering swap pages
227          * all over the entire swap partition, so that we reduce
228          * overall disk seek times between swap pages.  -- sct
229          * But we do now try to find an empty cluster.  -Andrea
230          * And we let swap pages go all over an SSD partition.  Hugh
231          */
232
233         si->flags += SWP_SCANNING;
234         scan_base = offset = si->cluster_next;
235
236         if (unlikely(!si->cluster_nr--)) {
237                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
238                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
239                         goto checks;
240                 }
241                 if (si->flags & SWP_DISCARDABLE) {
242                         /*
243                          * Start range check on racing allocations, in case
244                          * they overlap the cluster we eventually decide on
245                          * (we scan without swap_lock to allow preemption).
246                          * It's hardly conceivable that cluster_nr could be
247                          * wrapped during our scan, but don't depend on it.
248                          */
249                         if (si->lowest_alloc)
250                                 goto checks;
251                         si->lowest_alloc = si->max;
252                         si->highest_alloc = 0;
253                 }
254                 spin_unlock(&swap_lock);
255
256                 /*
257                  * If seek is expensive, start searching for new cluster from
258                  * start of partition, to minimize the span of allocated swap.
259                  * But if seek is cheap, search from our current position, so
260                  * that swap is allocated from all over the partition: if the
261                  * Flash Translation Layer only remaps within limited zones,
262                  * we don't want to wear out the first zone too quickly.
263                  */
264                 if (!(si->flags & SWP_SOLIDSTATE))
265                         scan_base = offset = si->lowest_bit;
266                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
267
268                 /* Locate the first empty (unaligned) cluster */
269                 for (; last_in_cluster <= si->highest_bit; offset++) {
270                         if (si->swap_map[offset])
271                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
272                         else if (offset == last_in_cluster) {
273                                 spin_lock(&swap_lock);
274                                 offset -= SWAPFILE_CLUSTER - 1;
275                                 si->cluster_next = offset;
276                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
277                                 found_free_cluster = 1;
278                                 goto checks;
279                         }
280                         if (unlikely(--latency_ration < 0)) {
281                                 cond_resched();
282                                 latency_ration = LATENCY_LIMIT;
283                         }
284                 }
285
286                 offset = si->lowest_bit;
287                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
288
289                 /* Locate the first empty (unaligned) cluster */
290                 for (; last_in_cluster < scan_base; offset++) {
291                         if (si->swap_map[offset])
292                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
293                         else if (offset == last_in_cluster) {
294                                 spin_lock(&swap_lock);
295                                 offset -= SWAPFILE_CLUSTER - 1;
296                                 si->cluster_next = offset;
297                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
298                                 found_free_cluster = 1;
299                                 goto checks;
300                         }
301                         if (unlikely(--latency_ration < 0)) {
302                                 cond_resched();
303                                 latency_ration = LATENCY_LIMIT;
304                         }
305                 }
306
307                 offset = scan_base;
308                 spin_lock(&swap_lock);
309                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
310                 si->lowest_alloc = 0;
311         }
312
313 checks:
314         if (!(si->flags & SWP_WRITEOK))
315                 goto no_page;
316         if (!si->highest_bit)
317                 goto no_page;
318         if (offset > si->highest_bit)
319                 scan_base = offset = si->lowest_bit;
320
321         /* reuse swap entry of cache-only swap if not hibernation. */
322         if (vm_swap_full()
323                 && usage == SWAP_HAS_CACHE
324                 && si->swap_map[offset] == SWAP_HAS_CACHE) {
325                 int swap_was_freed;
326                 spin_unlock(&swap_lock);
327                 swap_was_freed = __try_to_reclaim_swap(si, offset);
328                 spin_lock(&swap_lock);
329                 /* entry was freed successfully, try to use this again */
330                 if (swap_was_freed)
331                         goto checks;
332                 goto scan; /* check next one */
333         }
334
335         if (si->swap_map[offset])
336                 goto scan;
337
338         if (offset == si->lowest_bit)
339                 si->lowest_bit++;
340         if (offset == si->highest_bit)
341                 si->highest_bit--;
342         si->inuse_pages++;
343         if (si->inuse_pages == si->pages) {
344                 si->lowest_bit = si->max;
345                 si->highest_bit = 0;
346         }
347         si->swap_map[offset] = usage;
348         si->cluster_next = offset + 1;
349         si->flags -= SWP_SCANNING;
350
351         if (si->lowest_alloc) {
352                 /*
353                  * Only set when SWP_DISCARDABLE, and there's a scan
354                  * for a free cluster in progress or just completed.
355                  */
356                 if (found_free_cluster) {
357                         /*
358                          * To optimize wear-levelling, discard the
359                          * old data of the cluster, taking care not to
360                          * discard any of its pages that have already
361                          * been allocated by racing tasks (offset has
362                          * already stepped over any at the beginning).
363                          */
364                         if (offset < si->highest_alloc &&
365                             si->lowest_alloc <= last_in_cluster)
366                                 last_in_cluster = si->lowest_alloc - 1;
367                         si->flags |= SWP_DISCARDING;
368                         spin_unlock(&swap_lock);
369
370                         if (offset < last_in_cluster)
371                                 discard_swap_cluster(si, offset,
372                                         last_in_cluster - offset + 1);
373
374                         spin_lock(&swap_lock);
375                         si->lowest_alloc = 0;
376                         si->flags &= ~SWP_DISCARDING;
377
378                         smp_mb();       /* wake_up_bit advises this */
379                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
380
381                 } else if (si->flags & SWP_DISCARDING) {
382                         /*
383                          * Delay using pages allocated by racing tasks
384                          * until the whole discard has been issued. We
385                          * could defer that delay until swap_writepage,
386                          * but it's easier to keep this self-contained.
387                          */
388                         spin_unlock(&swap_lock);
389                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
390                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
391                         spin_lock(&swap_lock);
392                 } else {
393                         /*
394                          * Note pages allocated by racing tasks while
395                          * scan for a free cluster is in progress, so
396                          * that its final discard can exclude them.
397                          */
398                         if (offset < si->lowest_alloc)
399                                 si->lowest_alloc = offset;
400                         if (offset > si->highest_alloc)
401                                 si->highest_alloc = offset;
402                 }
403         }
404         return offset;
405
406 scan:
407         spin_unlock(&swap_lock);
408         while (++offset <= si->highest_bit) {
409                 if (!si->swap_map[offset]) {
410                         spin_lock(&swap_lock);
411                         goto checks;
412                 }
413                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
414                         spin_lock(&swap_lock);
415                         goto checks;
416                 }
417                 if (unlikely(--latency_ration < 0)) {
418                         cond_resched();
419                         latency_ration = LATENCY_LIMIT;
420                 }
421         }
422         offset = si->lowest_bit;
423         while (++offset < scan_base) {
424                 if (!si->swap_map[offset]) {
425                         spin_lock(&swap_lock);
426                         goto checks;
427                 }
428                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
429                         spin_lock(&swap_lock);
430                         goto checks;
431                 }
432                 if (unlikely(--latency_ration < 0)) {
433                         cond_resched();
434                         latency_ration = LATENCY_LIMIT;
435                 }
436         }
437         spin_lock(&swap_lock);
438
439 no_page:
440         si->flags -= SWP_SCANNING;
441         return 0;
442 }
443
444 swp_entry_t get_swap_page(void)
445 {
446         struct swap_info_struct *si;
447         pgoff_t offset;
448         int type, next;
449         int wrapped = 0;
450
451         spin_lock(&swap_lock);
452         if (nr_swap_pages <= 0)
453                 goto noswap;
454         nr_swap_pages--;
455
456         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
457                 si = swap_info[type];
458                 next = si->next;
459                 if (next < 0 ||
460                     (!wrapped && si->prio != swap_info[next]->prio)) {
461                         next = swap_list.head;
462                         wrapped++;
463                 }
464
465                 if (!si->highest_bit)
466                         continue;
467                 if (!(si->flags & SWP_WRITEOK))
468                         continue;
469
470                 swap_list.next = next;
471                 /* This is called for allocating swap entry for cache */
472                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
473                 if (offset) {
474                         spin_unlock(&swap_lock);
475                         return swp_entry(type, offset);
476                 }
477                 next = swap_list.next;
478         }
479
480         nr_swap_pages++;
481 noswap:
482         spin_unlock(&swap_lock);
483         return (swp_entry_t) {0};
484 }
485
486 /* The only caller of this function is now susupend routine */
487 swp_entry_t get_swap_page_of_type(int type)
488 {
489         struct swap_info_struct *si;
490         pgoff_t offset;
491
492         spin_lock(&swap_lock);
493         si = swap_info[type];
494         if (si && (si->flags & SWP_WRITEOK)) {
495                 nr_swap_pages--;
496                 /* This is called for allocating swap entry, not cache */
497                 offset = scan_swap_map(si, 1);
498                 if (offset) {
499                         spin_unlock(&swap_lock);
500                         return swp_entry(type, offset);
501                 }
502                 nr_swap_pages++;
503         }
504         spin_unlock(&swap_lock);
505         return (swp_entry_t) {0};
506 }
507
508 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
509 {
510         struct swap_info_struct *p;
511         unsigned long offset, type;
512
513         if (!entry.val)
514                 goto out;
515         type = swp_type(entry);
516         if (type >= nr_swapfiles)
517                 goto bad_nofile;
518         p = swap_info[type];
519         if (!(p->flags & SWP_USED))
520                 goto bad_device;
521         offset = swp_offset(entry);
522         if (offset >= p->max)
523                 goto bad_offset;
524         if (!p->swap_map[offset])
525                 goto bad_free;
526         spin_lock(&swap_lock);
527         return p;
528
529 bad_free:
530         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
531         goto out;
532 bad_offset:
533         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
534         goto out;
535 bad_device:
536         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
537         goto out;
538 bad_nofile:
539         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
540 out:
541         return NULL;
542 }
543
544 static unsigned char swap_entry_free(struct swap_info_struct *p,
545                                      swp_entry_t entry, unsigned char usage)
546 {
547         unsigned long offset = swp_offset(entry);
548         unsigned char count;
549         unsigned char has_cache;
550
551         count = p->swap_map[offset];
552         has_cache = count & SWAP_HAS_CACHE;
553         count &= ~SWAP_HAS_CACHE;
554
555         if (usage == SWAP_HAS_CACHE) {
556                 VM_BUG_ON(!has_cache);
557                 has_cache = 0;
558         } else if (count == SWAP_MAP_SHMEM) {
559                 /*
560                  * Or we could insist on shmem.c using a special
561                  * swap_shmem_free() and free_shmem_swap_and_cache()...
562                  */
563                 count = 0;
564         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
565                 if (count == COUNT_CONTINUED) {
566                         if (swap_count_continued(p, offset, count))
567                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
568                         else
569                                 count = SWAP_MAP_MAX;
570                 } else
571                         count--;
572         }
573
574         if (!count)
575                 mem_cgroup_uncharge_swap(entry);
576
577         usage = count | has_cache;
578         p->swap_map[offset] = usage;
579
580         /* free if no reference */
581         if (!usage) {
582                 struct gendisk *disk = p->bdev->bd_disk;
583                 if (offset < p->lowest_bit)
584                         p->lowest_bit = offset;
585                 if (offset > p->highest_bit)
586                         p->highest_bit = offset;
587                 if (swap_list.next >= 0 &&
588                     p->prio > swap_info[swap_list.next]->prio)
589                         swap_list.next = p->type;
590                 nr_swap_pages++;
591                 p->inuse_pages--;
592                 if ((p->flags & SWP_BLKDEV) &&
593                                 disk->fops->swap_slot_free_notify)
594                         disk->fops->swap_slot_free_notify(p->bdev, offset);
595         }
596
597         return usage;
598 }
599
600 /*
601  * Caller has made sure that the swapdevice corresponding to entry
602  * is still around or has not been recycled.
603  */
604 void swap_free(swp_entry_t entry)
605 {
606         struct swap_info_struct *p;
607
608         p = swap_info_get(entry);
609         if (p) {
610                 swap_entry_free(p, entry, 1);
611                 spin_unlock(&swap_lock);
612         }
613 }
614
615 /*
616  * Called after dropping swapcache to decrease refcnt to swap entries.
617  */
618 void swapcache_free(swp_entry_t entry, struct page *page)
619 {
620         struct swap_info_struct *p;
621         unsigned char count;
622
623         p = swap_info_get(entry);
624         if (p) {
625                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
626                 if (page)
627                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
628                 spin_unlock(&swap_lock);
629         }
630 }
631
632 /*
633  * How many references to page are currently swapped out?
634  * This does not give an exact answer when swap count is continued,
635  * but does include the high COUNT_CONTINUED flag to allow for that.
636  */
637 static inline int page_swapcount(struct page *page)
638 {
639         int count = 0;
640         struct swap_info_struct *p;
641         swp_entry_t entry;
642
643         entry.val = page_private(page);
644         p = swap_info_get(entry);
645         if (p) {
646                 count = swap_count(p->swap_map[swp_offset(entry)]);
647                 spin_unlock(&swap_lock);
648         }
649         return count;
650 }
651
652 /*
653  * We can write to an anon page without COW if there are no other references
654  * to it.  And as a side-effect, free up its swap: because the old content
655  * on disk will never be read, and seeking back there to write new content
656  * later would only waste time away from clustering.
657  */
658 int reuse_swap_page(struct page *page)
659 {
660         int count;
661
662         VM_BUG_ON(!PageLocked(page));
663         if (unlikely(PageKsm(page)))
664                 return 0;
665         count = page_mapcount(page);
666         if (count <= 1 && PageSwapCache(page)) {
667                 count += page_swapcount(page);
668                 if (count == 1 && !PageWriteback(page)) {
669                         delete_from_swap_cache(page);
670                         SetPageDirty(page);
671                 }
672         }
673         return count <= 1;
674 }
675
676 /*
677  * If swap is getting full, or if there are no more mappings of this page,
678  * then try_to_free_swap is called to free its swap space.
679  */
680 int try_to_free_swap(struct page *page)
681 {
682         VM_BUG_ON(!PageLocked(page));
683
684         if (!PageSwapCache(page))
685                 return 0;
686         if (PageWriteback(page))
687                 return 0;
688         if (page_swapcount(page))
689                 return 0;
690
691         delete_from_swap_cache(page);
692         SetPageDirty(page);
693         return 1;
694 }
695
696 /*
697  * Free the swap entry like above, but also try to
698  * free the page cache entry if it is the last user.
699  */
700 int free_swap_and_cache(swp_entry_t entry)
701 {
702         struct swap_info_struct *p;
703         struct page *page = NULL;
704
705         if (non_swap_entry(entry))
706                 return 1;
707
708         p = swap_info_get(entry);
709         if (p) {
710                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
711                         page = find_get_page(&swapper_space, entry.val);
712                         if (page && !trylock_page(page)) {
713                                 page_cache_release(page);
714                                 page = NULL;
715                         }
716                 }
717                 spin_unlock(&swap_lock);
718         }
719         if (page) {
720                 /*
721                  * Not mapped elsewhere, or swap space full? Free it!
722                  * Also recheck PageSwapCache now page is locked (above).
723                  */
724                 if (PageSwapCache(page) && !PageWriteback(page) &&
725                                 (!page_mapped(page) || vm_swap_full())) {
726                         delete_from_swap_cache(page);
727                         SetPageDirty(page);
728                 }
729                 unlock_page(page);
730                 page_cache_release(page);
731         }
732         return p != NULL;
733 }
734
735 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
736 /**
737  * mem_cgroup_count_swap_user - count the user of a swap entry
738  * @ent: the swap entry to be checked
739  * @pagep: the pointer for the swap cache page of the entry to be stored
740  *
741  * Returns the number of the user of the swap entry. The number is valid only
742  * for swaps of anonymous pages.
743  * If the entry is found on swap cache, the page is stored to pagep with
744  * refcount of it being incremented.
745  */
746 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
747 {
748         struct page *page;
749         struct swap_info_struct *p;
750         int count = 0;
751
752         page = find_get_page(&swapper_space, ent.val);
753         if (page)
754                 count += page_mapcount(page);
755         p = swap_info_get(ent);
756         if (p) {
757                 count += swap_count(p->swap_map[swp_offset(ent)]);
758                 spin_unlock(&swap_lock);
759         }
760
761         *pagep = page;
762         return count;
763 }
764 #endif
765
766 #ifdef CONFIG_HIBERNATION
767 /*
768  * Find the swap type that corresponds to given device (if any).
769  *
770  * @offset - number of the PAGE_SIZE-sized block of the device, starting
771  * from 0, in which the swap header is expected to be located.
772  *
773  * This is needed for the suspend to disk (aka swsusp).
774  */
775 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
776 {
777         struct block_device *bdev = NULL;
778         int type;
779
780         if (device)
781                 bdev = bdget(device);
782
783         spin_lock(&swap_lock);
784         for (type = 0; type < nr_swapfiles; type++) {
785                 struct swap_info_struct *sis = swap_info[type];
786
787                 if (!(sis->flags & SWP_WRITEOK))
788                         continue;
789
790                 if (!bdev) {
791                         if (bdev_p)
792                                 *bdev_p = bdgrab(sis->bdev);
793
794                         spin_unlock(&swap_lock);
795                         return type;
796                 }
797                 if (bdev == sis->bdev) {
798                         struct swap_extent *se = &sis->first_swap_extent;
799
800                         if (se->start_block == offset) {
801                                 if (bdev_p)
802                                         *bdev_p = bdgrab(sis->bdev);
803
804                                 spin_unlock(&swap_lock);
805                                 bdput(bdev);
806                                 return type;
807                         }
808                 }
809         }
810         spin_unlock(&swap_lock);
811         if (bdev)
812                 bdput(bdev);
813
814         return -ENODEV;
815 }
816
817 /*
818  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
819  * corresponding to given index in swap_info (swap type).
820  */
821 sector_t swapdev_block(int type, pgoff_t offset)
822 {
823         struct block_device *bdev;
824
825         if ((unsigned int)type >= nr_swapfiles)
826                 return 0;
827         if (!(swap_info[type]->flags & SWP_WRITEOK))
828                 return 0;
829         return map_swap_entry(swp_entry(type, offset), &bdev);
830 }
831
832 /*
833  * Return either the total number of swap pages of given type, or the number
834  * of free pages of that type (depending on @free)
835  *
836  * This is needed for software suspend
837  */
838 unsigned int count_swap_pages(int type, int free)
839 {
840         unsigned int n = 0;
841
842         spin_lock(&swap_lock);
843         if ((unsigned int)type < nr_swapfiles) {
844                 struct swap_info_struct *sis = swap_info[type];
845
846                 if (sis->flags & SWP_WRITEOK) {
847                         n = sis->pages;
848                         if (free)
849                                 n -= sis->inuse_pages;
850                 }
851         }
852         spin_unlock(&swap_lock);
853         return n;
854 }
855 #endif /* CONFIG_HIBERNATION */
856
857 /*
858  * No need to decide whether this PTE shares the swap entry with others,
859  * just let do_wp_page work it out if a write is requested later - to
860  * force COW, vm_page_prot omits write permission from any private vma.
861  */
862 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
863                 unsigned long addr, swp_entry_t entry, struct page *page)
864 {
865         struct mem_cgroup *ptr = NULL;
866         spinlock_t *ptl;
867         pte_t *pte;
868         int ret = 1;
869
870         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
871                 ret = -ENOMEM;
872                 goto out_nolock;
873         }
874
875         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
876         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
877                 if (ret > 0)
878                         mem_cgroup_cancel_charge_swapin(ptr);
879                 ret = 0;
880                 goto out;
881         }
882
883         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
884         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
885         get_page(page);
886         set_pte_at(vma->vm_mm, addr, pte,
887                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
888         page_add_anon_rmap(page, vma, addr);
889         mem_cgroup_commit_charge_swapin(page, ptr);
890         swap_free(entry);
891         /*
892          * Move the page to the active list so it is not
893          * immediately swapped out again after swapon.
894          */
895         activate_page(page);
896 out:
897         pte_unmap_unlock(pte, ptl);
898 out_nolock:
899         return ret;
900 }
901
902 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
903                                 unsigned long addr, unsigned long end,
904                                 swp_entry_t entry, struct page *page)
905 {
906         pte_t swp_pte = swp_entry_to_pte(entry);
907         pte_t *pte;
908         int ret = 0;
909
910         /*
911          * We don't actually need pte lock while scanning for swp_pte: since
912          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
913          * page table while we're scanning; though it could get zapped, and on
914          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
915          * of unmatched parts which look like swp_pte, so unuse_pte must
916          * recheck under pte lock.  Scanning without pte lock lets it be
917          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
918          */
919         pte = pte_offset_map(pmd, addr);
920         do {
921                 /*
922                  * swapoff spends a _lot_ of time in this loop!
923                  * Test inline before going to call unuse_pte.
924                  */
925                 if (unlikely(pte_same(*pte, swp_pte))) {
926                         pte_unmap(pte);
927                         ret = unuse_pte(vma, pmd, addr, entry, page);
928                         if (ret)
929                                 goto out;
930                         pte = pte_offset_map(pmd, addr);
931                 }
932         } while (pte++, addr += PAGE_SIZE, addr != end);
933         pte_unmap(pte - 1);
934 out:
935         return ret;
936 }
937
938 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
939                                 unsigned long addr, unsigned long end,
940                                 swp_entry_t entry, struct page *page)
941 {
942         pmd_t *pmd;
943         unsigned long next;
944         int ret;
945
946         pmd = pmd_offset(pud, addr);
947         do {
948                 next = pmd_addr_end(addr, end);
949                 if (pmd_none_or_clear_bad(pmd))
950                         continue;
951                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
952                 if (ret)
953                         return ret;
954         } while (pmd++, addr = next, addr != end);
955         return 0;
956 }
957
958 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
959                                 unsigned long addr, unsigned long end,
960                                 swp_entry_t entry, struct page *page)
961 {
962         pud_t *pud;
963         unsigned long next;
964         int ret;
965
966         pud = pud_offset(pgd, addr);
967         do {
968                 next = pud_addr_end(addr, end);
969                 if (pud_none_or_clear_bad(pud))
970                         continue;
971                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
972                 if (ret)
973                         return ret;
974         } while (pud++, addr = next, addr != end);
975         return 0;
976 }
977
978 static int unuse_vma(struct vm_area_struct *vma,
979                                 swp_entry_t entry, struct page *page)
980 {
981         pgd_t *pgd;
982         unsigned long addr, end, next;
983         int ret;
984
985         if (page_anon_vma(page)) {
986                 addr = page_address_in_vma(page, vma);
987                 if (addr == -EFAULT)
988                         return 0;
989                 else
990                         end = addr + PAGE_SIZE;
991         } else {
992                 addr = vma->vm_start;
993                 end = vma->vm_end;
994         }
995
996         pgd = pgd_offset(vma->vm_mm, addr);
997         do {
998                 next = pgd_addr_end(addr, end);
999                 if (pgd_none_or_clear_bad(pgd))
1000                         continue;
1001                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1002                 if (ret)
1003                         return ret;
1004         } while (pgd++, addr = next, addr != end);
1005         return 0;
1006 }
1007
1008 static int unuse_mm(struct mm_struct *mm,
1009                                 swp_entry_t entry, struct page *page)
1010 {
1011         struct vm_area_struct *vma;
1012         int ret = 0;
1013
1014         if (!down_read_trylock(&mm->mmap_sem)) {
1015                 /*
1016                  * Activate page so shrink_inactive_list is unlikely to unmap
1017                  * its ptes while lock is dropped, so swapoff can make progress.
1018                  */
1019                 activate_page(page);
1020                 unlock_page(page);
1021                 down_read(&mm->mmap_sem);
1022                 lock_page(page);
1023         }
1024         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1025                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1026                         break;
1027         }
1028         up_read(&mm->mmap_sem);
1029         return (ret < 0)? ret: 0;
1030 }
1031
1032 /*
1033  * Scan swap_map from current position to next entry still in use.
1034  * Recycle to start on reaching the end, returning 0 when empty.
1035  */
1036 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1037                                         unsigned int prev)
1038 {
1039         unsigned int max = si->max;
1040         unsigned int i = prev;
1041         unsigned char count;
1042
1043         /*
1044          * No need for swap_lock here: we're just looking
1045          * for whether an entry is in use, not modifying it; false
1046          * hits are okay, and sys_swapoff() has already prevented new
1047          * allocations from this area (while holding swap_lock).
1048          */
1049         for (;;) {
1050                 if (++i >= max) {
1051                         if (!prev) {
1052                                 i = 0;
1053                                 break;
1054                         }
1055                         /*
1056                          * No entries in use at top of swap_map,
1057                          * loop back to start and recheck there.
1058                          */
1059                         max = prev + 1;
1060                         prev = 0;
1061                         i = 1;
1062                 }
1063                 count = si->swap_map[i];
1064                 if (count && swap_count(count) != SWAP_MAP_BAD)
1065                         break;
1066         }
1067         return i;
1068 }
1069
1070 /*
1071  * We completely avoid races by reading each swap page in advance,
1072  * and then search for the process using it.  All the necessary
1073  * page table adjustments can then be made atomically.
1074  */
1075 static int try_to_unuse(unsigned int type)
1076 {
1077         struct swap_info_struct *si = swap_info[type];
1078         struct mm_struct *start_mm;
1079         unsigned char *swap_map;
1080         unsigned char swcount;
1081         struct page *page;
1082         swp_entry_t entry;
1083         unsigned int i = 0;
1084         int retval = 0;
1085
1086         /*
1087          * When searching mms for an entry, a good strategy is to
1088          * start at the first mm we freed the previous entry from
1089          * (though actually we don't notice whether we or coincidence
1090          * freed the entry).  Initialize this start_mm with a hold.
1091          *
1092          * A simpler strategy would be to start at the last mm we
1093          * freed the previous entry from; but that would take less
1094          * advantage of mmlist ordering, which clusters forked mms
1095          * together, child after parent.  If we race with dup_mmap(), we
1096          * prefer to resolve parent before child, lest we miss entries
1097          * duplicated after we scanned child: using last mm would invert
1098          * that.
1099          */
1100         start_mm = &init_mm;
1101         atomic_inc(&init_mm.mm_users);
1102
1103         /*
1104          * Keep on scanning until all entries have gone.  Usually,
1105          * one pass through swap_map is enough, but not necessarily:
1106          * there are races when an instance of an entry might be missed.
1107          */
1108         while ((i = find_next_to_unuse(si, i)) != 0) {
1109                 if (signal_pending(current)) {
1110                         retval = -EINTR;
1111                         break;
1112                 }
1113
1114                 /*
1115                  * Get a page for the entry, using the existing swap
1116                  * cache page if there is one.  Otherwise, get a clean
1117                  * page and read the swap into it.
1118                  */
1119                 swap_map = &si->swap_map[i];
1120                 entry = swp_entry(type, i);
1121                 page = read_swap_cache_async(entry,
1122                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1123                 if (!page) {
1124                         /*
1125                          * Either swap_duplicate() failed because entry
1126                          * has been freed independently, and will not be
1127                          * reused since sys_swapoff() already disabled
1128                          * allocation from here, or alloc_page() failed.
1129                          */
1130                         if (!*swap_map)
1131                                 continue;
1132                         retval = -ENOMEM;
1133                         break;
1134                 }
1135
1136                 /*
1137                  * Don't hold on to start_mm if it looks like exiting.
1138                  */
1139                 if (atomic_read(&start_mm->mm_users) == 1) {
1140                         mmput(start_mm);
1141                         start_mm = &init_mm;
1142                         atomic_inc(&init_mm.mm_users);
1143                 }
1144
1145                 /*
1146                  * Wait for and lock page.  When do_swap_page races with
1147                  * try_to_unuse, do_swap_page can handle the fault much
1148                  * faster than try_to_unuse can locate the entry.  This
1149                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1150                  * defer to do_swap_page in such a case - in some tests,
1151                  * do_swap_page and try_to_unuse repeatedly compete.
1152                  */
1153                 wait_on_page_locked(page);
1154                 wait_on_page_writeback(page);
1155                 lock_page(page);
1156                 wait_on_page_writeback(page);
1157
1158                 /*
1159                  * Remove all references to entry.
1160                  */
1161                 swcount = *swap_map;
1162                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1163                         retval = shmem_unuse(entry, page);
1164                         /* page has already been unlocked and released */
1165                         if (retval < 0)
1166                                 break;
1167                         continue;
1168                 }
1169                 if (swap_count(swcount) && start_mm != &init_mm)
1170                         retval = unuse_mm(start_mm, entry, page);
1171
1172                 if (swap_count(*swap_map)) {
1173                         int set_start_mm = (*swap_map >= swcount);
1174                         struct list_head *p = &start_mm->mmlist;
1175                         struct mm_struct *new_start_mm = start_mm;
1176                         struct mm_struct *prev_mm = start_mm;
1177                         struct mm_struct *mm;
1178
1179                         atomic_inc(&new_start_mm->mm_users);
1180                         atomic_inc(&prev_mm->mm_users);
1181                         spin_lock(&mmlist_lock);
1182                         while (swap_count(*swap_map) && !retval &&
1183                                         (p = p->next) != &start_mm->mmlist) {
1184                                 mm = list_entry(p, struct mm_struct, mmlist);
1185                                 if (!atomic_inc_not_zero(&mm->mm_users))
1186                                         continue;
1187                                 spin_unlock(&mmlist_lock);
1188                                 mmput(prev_mm);
1189                                 prev_mm = mm;
1190
1191                                 cond_resched();
1192
1193                                 swcount = *swap_map;
1194                                 if (!swap_count(swcount)) /* any usage ? */
1195                                         ;
1196                                 else if (mm == &init_mm)
1197                                         set_start_mm = 1;
1198                                 else
1199                                         retval = unuse_mm(mm, entry, page);
1200
1201                                 if (set_start_mm && *swap_map < swcount) {
1202                                         mmput(new_start_mm);
1203                                         atomic_inc(&mm->mm_users);
1204                                         new_start_mm = mm;
1205                                         set_start_mm = 0;
1206                                 }
1207                                 spin_lock(&mmlist_lock);
1208                         }
1209                         spin_unlock(&mmlist_lock);
1210                         mmput(prev_mm);
1211                         mmput(start_mm);
1212                         start_mm = new_start_mm;
1213                 }
1214                 if (retval) {
1215                         unlock_page(page);
1216                         page_cache_release(page);
1217                         break;
1218                 }
1219
1220                 /*
1221                  * If a reference remains (rare), we would like to leave
1222                  * the page in the swap cache; but try_to_unmap could
1223                  * then re-duplicate the entry once we drop page lock,
1224                  * so we might loop indefinitely; also, that page could
1225                  * not be swapped out to other storage meanwhile.  So:
1226                  * delete from cache even if there's another reference,
1227                  * after ensuring that the data has been saved to disk -
1228                  * since if the reference remains (rarer), it will be
1229                  * read from disk into another page.  Splitting into two
1230                  * pages would be incorrect if swap supported "shared
1231                  * private" pages, but they are handled by tmpfs files.
1232                  *
1233                  * Given how unuse_vma() targets one particular offset
1234                  * in an anon_vma, once the anon_vma has been determined,
1235                  * this splitting happens to be just what is needed to
1236                  * handle where KSM pages have been swapped out: re-reading
1237                  * is unnecessarily slow, but we can fix that later on.
1238                  */
1239                 if (swap_count(*swap_map) &&
1240                      PageDirty(page) && PageSwapCache(page)) {
1241                         struct writeback_control wbc = {
1242                                 .sync_mode = WB_SYNC_NONE,
1243                         };
1244
1245                         swap_writepage(page, &wbc);
1246                         lock_page(page);
1247                         wait_on_page_writeback(page);
1248                 }
1249
1250                 /*
1251                  * It is conceivable that a racing task removed this page from
1252                  * swap cache just before we acquired the page lock at the top,
1253                  * or while we dropped it in unuse_mm().  The page might even
1254                  * be back in swap cache on another swap area: that we must not
1255                  * delete, since it may not have been written out to swap yet.
1256                  */
1257                 if (PageSwapCache(page) &&
1258                     likely(page_private(page) == entry.val))
1259                         delete_from_swap_cache(page);
1260
1261                 /*
1262                  * So we could skip searching mms once swap count went
1263                  * to 1, we did not mark any present ptes as dirty: must
1264                  * mark page dirty so shrink_page_list will preserve it.
1265                  */
1266                 SetPageDirty(page);
1267                 unlock_page(page);
1268                 page_cache_release(page);
1269
1270                 /*
1271                  * Make sure that we aren't completely killing
1272                  * interactive performance.
1273                  */
1274                 cond_resched();
1275         }
1276
1277         mmput(start_mm);
1278         return retval;
1279 }
1280
1281 /*
1282  * After a successful try_to_unuse, if no swap is now in use, we know
1283  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1284  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1285  * added to the mmlist just after page_duplicate - before would be racy.
1286  */
1287 static void drain_mmlist(void)
1288 {
1289         struct list_head *p, *next;
1290         unsigned int type;
1291
1292         for (type = 0; type < nr_swapfiles; type++)
1293                 if (swap_info[type]->inuse_pages)
1294                         return;
1295         spin_lock(&mmlist_lock);
1296         list_for_each_safe(p, next, &init_mm.mmlist)
1297                 list_del_init(p);
1298         spin_unlock(&mmlist_lock);
1299 }
1300
1301 /*
1302  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1303  * corresponds to page offset for the specified swap entry.
1304  * Note that the type of this function is sector_t, but it returns page offset
1305  * into the bdev, not sector offset.
1306  */
1307 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1308 {
1309         struct swap_info_struct *sis;
1310         struct swap_extent *start_se;
1311         struct swap_extent *se;
1312         pgoff_t offset;
1313
1314         sis = swap_info[swp_type(entry)];
1315         *bdev = sis->bdev;
1316
1317         offset = swp_offset(entry);
1318         start_se = sis->curr_swap_extent;
1319         se = start_se;
1320
1321         for ( ; ; ) {
1322                 struct list_head *lh;
1323
1324                 if (se->start_page <= offset &&
1325                                 offset < (se->start_page + se->nr_pages)) {
1326                         return se->start_block + (offset - se->start_page);
1327                 }
1328                 lh = se->list.next;
1329                 se = list_entry(lh, struct swap_extent, list);
1330                 sis->curr_swap_extent = se;
1331                 BUG_ON(se == start_se);         /* It *must* be present */
1332         }
1333 }
1334
1335 /*
1336  * Returns the page offset into bdev for the specified page's swap entry.
1337  */
1338 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1339 {
1340         swp_entry_t entry;
1341         entry.val = page_private(page);
1342         return map_swap_entry(entry, bdev);
1343 }
1344
1345 /*
1346  * Free all of a swapdev's extent information
1347  */
1348 static void destroy_swap_extents(struct swap_info_struct *sis)
1349 {
1350         while (!list_empty(&sis->first_swap_extent.list)) {
1351                 struct swap_extent *se;
1352
1353                 se = list_entry(sis->first_swap_extent.list.next,
1354                                 struct swap_extent, list);
1355                 list_del(&se->list);
1356                 kfree(se);
1357         }
1358 }
1359
1360 /*
1361  * Add a block range (and the corresponding page range) into this swapdev's
1362  * extent list.  The extent list is kept sorted in page order.
1363  *
1364  * This function rather assumes that it is called in ascending page order.
1365  */
1366 static int
1367 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1368                 unsigned long nr_pages, sector_t start_block)
1369 {
1370         struct swap_extent *se;
1371         struct swap_extent *new_se;
1372         struct list_head *lh;
1373
1374         if (start_page == 0) {
1375                 se = &sis->first_swap_extent;
1376                 sis->curr_swap_extent = se;
1377                 se->start_page = 0;
1378                 se->nr_pages = nr_pages;
1379                 se->start_block = start_block;
1380                 return 1;
1381         } else {
1382                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1383                 se = list_entry(lh, struct swap_extent, list);
1384                 BUG_ON(se->start_page + se->nr_pages != start_page);
1385                 if (se->start_block + se->nr_pages == start_block) {
1386                         /* Merge it */
1387                         se->nr_pages += nr_pages;
1388                         return 0;
1389                 }
1390         }
1391
1392         /*
1393          * No merge.  Insert a new extent, preserving ordering.
1394          */
1395         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1396         if (new_se == NULL)
1397                 return -ENOMEM;
1398         new_se->start_page = start_page;
1399         new_se->nr_pages = nr_pages;
1400         new_se->start_block = start_block;
1401
1402         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1403         return 1;
1404 }
1405
1406 /*
1407  * A `swap extent' is a simple thing which maps a contiguous range of pages
1408  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1409  * is built at swapon time and is then used at swap_writepage/swap_readpage
1410  * time for locating where on disk a page belongs.
1411  *
1412  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1413  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1414  * swap files identically.
1415  *
1416  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1417  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1418  * swapfiles are handled *identically* after swapon time.
1419  *
1420  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1421  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1422  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1423  * requirements, they are simply tossed out - we will never use those blocks
1424  * for swapping.
1425  *
1426  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1427  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1428  * which will scribble on the fs.
1429  *
1430  * The amount of disk space which a single swap extent represents varies.
1431  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1432  * extents in the list.  To avoid much list walking, we cache the previous
1433  * search location in `curr_swap_extent', and start new searches from there.
1434  * This is extremely effective.  The average number of iterations in
1435  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1436  */
1437 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1438 {
1439         struct inode *inode;
1440         unsigned blocks_per_page;
1441         unsigned long page_no;
1442         unsigned blkbits;
1443         sector_t probe_block;
1444         sector_t last_block;
1445         sector_t lowest_block = -1;
1446         sector_t highest_block = 0;
1447         int nr_extents = 0;
1448         int ret;
1449
1450         inode = sis->swap_file->f_mapping->host;
1451         if (S_ISBLK(inode->i_mode)) {
1452                 ret = add_swap_extent(sis, 0, sis->max, 0);
1453                 *span = sis->pages;
1454                 goto out;
1455         }
1456
1457         blkbits = inode->i_blkbits;
1458         blocks_per_page = PAGE_SIZE >> blkbits;
1459
1460         /*
1461          * Map all the blocks into the extent list.  This code doesn't try
1462          * to be very smart.
1463          */
1464         probe_block = 0;
1465         page_no = 0;
1466         last_block = i_size_read(inode) >> blkbits;
1467         while ((probe_block + blocks_per_page) <= last_block &&
1468                         page_no < sis->max) {
1469                 unsigned block_in_page;
1470                 sector_t first_block;
1471
1472                 first_block = bmap(inode, probe_block);
1473                 if (first_block == 0)
1474                         goto bad_bmap;
1475
1476                 /*
1477                  * It must be PAGE_SIZE aligned on-disk
1478                  */
1479                 if (first_block & (blocks_per_page - 1)) {
1480                         probe_block++;
1481                         goto reprobe;
1482                 }
1483
1484                 for (block_in_page = 1; block_in_page < blocks_per_page;
1485                                         block_in_page++) {
1486                         sector_t block;
1487
1488                         block = bmap(inode, probe_block + block_in_page);
1489                         if (block == 0)
1490                                 goto bad_bmap;
1491                         if (block != first_block + block_in_page) {
1492                                 /* Discontiguity */
1493                                 probe_block++;
1494                                 goto reprobe;
1495                         }
1496                 }
1497
1498                 first_block >>= (PAGE_SHIFT - blkbits);
1499                 if (page_no) {  /* exclude the header page */
1500                         if (first_block < lowest_block)
1501                                 lowest_block = first_block;
1502                         if (first_block > highest_block)
1503                                 highest_block = first_block;
1504                 }
1505
1506                 /*
1507                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1508                  */
1509                 ret = add_swap_extent(sis, page_no, 1, first_block);
1510                 if (ret < 0)
1511                         goto out;
1512                 nr_extents += ret;
1513                 page_no++;
1514                 probe_block += blocks_per_page;
1515 reprobe:
1516                 continue;
1517         }
1518         ret = nr_extents;
1519         *span = 1 + highest_block - lowest_block;
1520         if (page_no == 0)
1521                 page_no = 1;    /* force Empty message */
1522         sis->max = page_no;
1523         sis->pages = page_no - 1;
1524         sis->highest_bit = page_no - 1;
1525 out:
1526         return ret;
1527 bad_bmap:
1528         printk(KERN_ERR "swapon: swapfile has holes\n");
1529         ret = -EINVAL;
1530         goto out;
1531 }
1532
1533 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1534 {
1535         struct swap_info_struct *p = NULL;
1536         unsigned char *swap_map;
1537         struct file *swap_file, *victim;
1538         struct address_space *mapping;
1539         struct inode *inode;
1540         char *pathname;
1541         int i, type, prev;
1542         int err;
1543
1544         if (!capable(CAP_SYS_ADMIN))
1545                 return -EPERM;
1546
1547         pathname = getname(specialfile);
1548         err = PTR_ERR(pathname);
1549         if (IS_ERR(pathname))
1550                 goto out;
1551
1552         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1553         putname(pathname);
1554         err = PTR_ERR(victim);
1555         if (IS_ERR(victim))
1556                 goto out;
1557
1558         mapping = victim->f_mapping;
1559         prev = -1;
1560         spin_lock(&swap_lock);
1561         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1562                 p = swap_info[type];
1563                 if (p->flags & SWP_WRITEOK) {
1564                         if (p->swap_file->f_mapping == mapping)
1565                                 break;
1566                 }
1567                 prev = type;
1568         }
1569         if (type < 0) {
1570                 err = -EINVAL;
1571                 spin_unlock(&swap_lock);
1572                 goto out_dput;
1573         }
1574         if (!security_vm_enough_memory(p->pages))
1575                 vm_unacct_memory(p->pages);
1576         else {
1577                 err = -ENOMEM;
1578                 spin_unlock(&swap_lock);
1579                 goto out_dput;
1580         }
1581         if (prev < 0)
1582                 swap_list.head = p->next;
1583         else
1584                 swap_info[prev]->next = p->next;
1585         if (type == swap_list.next) {
1586                 /* just pick something that's safe... */
1587                 swap_list.next = swap_list.head;
1588         }
1589         if (p->prio < 0) {
1590                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1591                         swap_info[i]->prio = p->prio--;
1592                 least_priority++;
1593         }
1594         nr_swap_pages -= p->pages;
1595         total_swap_pages -= p->pages;
1596         p->flags &= ~SWP_WRITEOK;
1597         spin_unlock(&swap_lock);
1598
1599         current->flags |= PF_OOM_ORIGIN;
1600         err = try_to_unuse(type);
1601         current->flags &= ~PF_OOM_ORIGIN;
1602
1603         if (err) {
1604                 /* re-insert swap space back into swap_list */
1605                 spin_lock(&swap_lock);
1606                 if (p->prio < 0)
1607                         p->prio = --least_priority;
1608                 prev = -1;
1609                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1610                         if (p->prio >= swap_info[i]->prio)
1611                                 break;
1612                         prev = i;
1613                 }
1614                 p->next = i;
1615                 if (prev < 0)
1616                         swap_list.head = swap_list.next = type;
1617                 else
1618                         swap_info[prev]->next = type;
1619                 nr_swap_pages += p->pages;
1620                 total_swap_pages += p->pages;
1621                 p->flags |= SWP_WRITEOK;
1622                 spin_unlock(&swap_lock);
1623                 goto out_dput;
1624         }
1625
1626         /* wait for any unplug function to finish */
1627         down_write(&swap_unplug_sem);
1628         up_write(&swap_unplug_sem);
1629
1630         destroy_swap_extents(p);
1631         if (p->flags & SWP_CONTINUED)
1632                 free_swap_count_continuations(p);
1633
1634         mutex_lock(&swapon_mutex);
1635         spin_lock(&swap_lock);
1636         drain_mmlist();
1637
1638         /* wait for anyone still in scan_swap_map */
1639         p->highest_bit = 0;             /* cuts scans short */
1640         while (p->flags >= SWP_SCANNING) {
1641                 spin_unlock(&swap_lock);
1642                 schedule_timeout_uninterruptible(1);
1643                 spin_lock(&swap_lock);
1644         }
1645
1646         swap_file = p->swap_file;
1647         p->swap_file = NULL;
1648         p->max = 0;
1649         swap_map = p->swap_map;
1650         p->swap_map = NULL;
1651         p->flags = 0;
1652         spin_unlock(&swap_lock);
1653         mutex_unlock(&swapon_mutex);
1654         vfree(swap_map);
1655         /* Destroy swap account informatin */
1656         swap_cgroup_swapoff(type);
1657
1658         inode = mapping->host;
1659         if (S_ISBLK(inode->i_mode)) {
1660                 struct block_device *bdev = I_BDEV(inode);
1661                 set_blocksize(bdev, p->old_block_size);
1662                 bd_release(bdev);
1663         } else {
1664                 mutex_lock(&inode->i_mutex);
1665                 inode->i_flags &= ~S_SWAPFILE;
1666                 mutex_unlock(&inode->i_mutex);
1667         }
1668         filp_close(swap_file, NULL);
1669         err = 0;
1670
1671 out_dput:
1672         filp_close(victim, NULL);
1673 out:
1674         return err;
1675 }
1676
1677 #ifdef CONFIG_PROC_FS
1678 /* iterator */
1679 static void *swap_start(struct seq_file *swap, loff_t *pos)
1680 {
1681         struct swap_info_struct *si;
1682         int type;
1683         loff_t l = *pos;
1684
1685         mutex_lock(&swapon_mutex);
1686
1687         if (!l)
1688                 return SEQ_START_TOKEN;
1689
1690         for (type = 0; type < nr_swapfiles; type++) {
1691                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1692                 si = swap_info[type];
1693                 if (!(si->flags & SWP_USED) || !si->swap_map)
1694                         continue;
1695                 if (!--l)
1696                         return si;
1697         }
1698
1699         return NULL;
1700 }
1701
1702 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1703 {
1704         struct swap_info_struct *si = v;
1705         int type;
1706
1707         if (v == SEQ_START_TOKEN)
1708                 type = 0;
1709         else
1710                 type = si->type + 1;
1711
1712         for (; type < nr_swapfiles; type++) {
1713                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1714                 si = swap_info[type];
1715                 if (!(si->flags & SWP_USED) || !si->swap_map)
1716                         continue;
1717                 ++*pos;
1718                 return si;
1719         }
1720
1721         return NULL;
1722 }
1723
1724 static void swap_stop(struct seq_file *swap, void *v)
1725 {
1726         mutex_unlock(&swapon_mutex);
1727 }
1728
1729 static int swap_show(struct seq_file *swap, void *v)
1730 {
1731         struct swap_info_struct *si = v;
1732         struct file *file;
1733         int len;
1734
1735         if (si == SEQ_START_TOKEN) {
1736                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1737                 return 0;
1738         }
1739
1740         file = si->swap_file;
1741         len = seq_path(swap, &file->f_path, " \t\n\\");
1742         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1743                         len < 40 ? 40 - len : 1, " ",
1744                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1745                                 "partition" : "file\t",
1746                         si->pages << (PAGE_SHIFT - 10),
1747                         si->inuse_pages << (PAGE_SHIFT - 10),
1748                         si->prio);
1749         return 0;
1750 }
1751
1752 static const struct seq_operations swaps_op = {
1753         .start =        swap_start,
1754         .next =         swap_next,
1755         .stop =         swap_stop,
1756         .show =         swap_show
1757 };
1758
1759 static int swaps_open(struct inode *inode, struct file *file)
1760 {
1761         return seq_open(file, &swaps_op);
1762 }
1763
1764 static const struct file_operations proc_swaps_operations = {
1765         .open           = swaps_open,
1766         .read           = seq_read,
1767         .llseek         = seq_lseek,
1768         .release        = seq_release,
1769 };
1770
1771 static int __init procswaps_init(void)
1772 {
1773         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1774         return 0;
1775 }
1776 __initcall(procswaps_init);
1777 #endif /* CONFIG_PROC_FS */
1778
1779 #ifdef MAX_SWAPFILES_CHECK
1780 static int __init max_swapfiles_check(void)
1781 {
1782         MAX_SWAPFILES_CHECK();
1783         return 0;
1784 }
1785 late_initcall(max_swapfiles_check);
1786 #endif
1787
1788 /*
1789  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1790  *
1791  * The swapon system call
1792  */
1793 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1794 {
1795         struct swap_info_struct *p;
1796         char *name = NULL;
1797         struct block_device *bdev = NULL;
1798         struct file *swap_file = NULL;
1799         struct address_space *mapping;
1800         unsigned int type;
1801         int i, prev;
1802         int error;
1803         union swap_header *swap_header;
1804         unsigned int nr_good_pages;
1805         int nr_extents = 0;
1806         sector_t span;
1807         unsigned long maxpages;
1808         unsigned long swapfilepages;
1809         unsigned char *swap_map = NULL;
1810         struct page *page = NULL;
1811         struct inode *inode = NULL;
1812         int did_down = 0;
1813
1814         if (!capable(CAP_SYS_ADMIN))
1815                 return -EPERM;
1816
1817         p = kzalloc(sizeof(*p), GFP_KERNEL);
1818         if (!p)
1819                 return -ENOMEM;
1820
1821         spin_lock(&swap_lock);
1822         for (type = 0; type < nr_swapfiles; type++) {
1823                 if (!(swap_info[type]->flags & SWP_USED))
1824                         break;
1825         }
1826         error = -EPERM;
1827         if (type >= MAX_SWAPFILES) {
1828                 spin_unlock(&swap_lock);
1829                 kfree(p);
1830                 goto out;
1831         }
1832         if (type >= nr_swapfiles) {
1833                 p->type = type;
1834                 swap_info[type] = p;
1835                 /*
1836                  * Write swap_info[type] before nr_swapfiles, in case a
1837                  * racing procfs swap_start() or swap_next() is reading them.
1838                  * (We never shrink nr_swapfiles, we never free this entry.)
1839                  */
1840                 smp_wmb();
1841                 nr_swapfiles++;
1842         } else {
1843                 kfree(p);
1844                 p = swap_info[type];
1845                 /*
1846                  * Do not memset this entry: a racing procfs swap_next()
1847                  * would be relying on p->type to remain valid.
1848                  */
1849         }
1850         INIT_LIST_HEAD(&p->first_swap_extent.list);
1851         p->flags = SWP_USED;
1852         p->next = -1;
1853         spin_unlock(&swap_lock);
1854
1855         name = getname(specialfile);
1856         error = PTR_ERR(name);
1857         if (IS_ERR(name)) {
1858                 name = NULL;
1859                 goto bad_swap_2;
1860         }
1861         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1862         error = PTR_ERR(swap_file);
1863         if (IS_ERR(swap_file)) {
1864                 swap_file = NULL;
1865                 goto bad_swap_2;
1866         }
1867
1868         p->swap_file = swap_file;
1869         mapping = swap_file->f_mapping;
1870         inode = mapping->host;
1871
1872         error = -EBUSY;
1873         for (i = 0; i < nr_swapfiles; i++) {
1874                 struct swap_info_struct *q = swap_info[i];
1875
1876                 if (i == type || !q->swap_file)
1877                         continue;
1878                 if (mapping == q->swap_file->f_mapping)
1879                         goto bad_swap;
1880         }
1881
1882         error = -EINVAL;
1883         if (S_ISBLK(inode->i_mode)) {
1884                 bdev = I_BDEV(inode);
1885                 error = bd_claim(bdev, sys_swapon);
1886                 if (error < 0) {
1887                         bdev = NULL;
1888                         error = -EINVAL;
1889                         goto bad_swap;
1890                 }
1891                 p->old_block_size = block_size(bdev);
1892                 error = set_blocksize(bdev, PAGE_SIZE);
1893                 if (error < 0)
1894                         goto bad_swap;
1895                 p->bdev = bdev;
1896                 p->flags |= SWP_BLKDEV;
1897         } else if (S_ISREG(inode->i_mode)) {
1898                 p->bdev = inode->i_sb->s_bdev;
1899                 mutex_lock(&inode->i_mutex);
1900                 did_down = 1;
1901                 if (IS_SWAPFILE(inode)) {
1902                         error = -EBUSY;
1903                         goto bad_swap;
1904                 }
1905         } else {
1906                 goto bad_swap;
1907         }
1908
1909         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1910
1911         /*
1912          * Read the swap header.
1913          */
1914         if (!mapping->a_ops->readpage) {
1915                 error = -EINVAL;
1916                 goto bad_swap;
1917         }
1918         page = read_mapping_page(mapping, 0, swap_file);
1919         if (IS_ERR(page)) {
1920                 error = PTR_ERR(page);
1921                 goto bad_swap;
1922         }
1923         swap_header = kmap(page);
1924
1925         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1926                 printk(KERN_ERR "Unable to find swap-space signature\n");
1927                 error = -EINVAL;
1928                 goto bad_swap;
1929         }
1930
1931         /* swap partition endianess hack... */
1932         if (swab32(swap_header->info.version) == 1) {
1933                 swab32s(&swap_header->info.version);
1934                 swab32s(&swap_header->info.last_page);
1935                 swab32s(&swap_header->info.nr_badpages);
1936                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1937                         swab32s(&swap_header->info.badpages[i]);
1938         }
1939         /* Check the swap header's sub-version */
1940         if (swap_header->info.version != 1) {
1941                 printk(KERN_WARNING
1942                        "Unable to handle swap header version %d\n",
1943                        swap_header->info.version);
1944                 error = -EINVAL;
1945                 goto bad_swap;
1946         }
1947
1948         p->lowest_bit  = 1;
1949         p->cluster_next = 1;
1950         p->cluster_nr = 0;
1951
1952         /*
1953          * Find out how many pages are allowed for a single swap
1954          * device. There are two limiting factors: 1) the number of
1955          * bits for the swap offset in the swp_entry_t type and
1956          * 2) the number of bits in the a swap pte as defined by
1957          * the different architectures. In order to find the
1958          * largest possible bit mask a swap entry with swap type 0
1959          * and swap offset ~0UL is created, encoded to a swap pte,
1960          * decoded to a swp_entry_t again and finally the swap
1961          * offset is extracted. This will mask all the bits from
1962          * the initial ~0UL mask that can't be encoded in either
1963          * the swp_entry_t or the architecture definition of a
1964          * swap pte.
1965          */
1966         maxpages = swp_offset(pte_to_swp_entry(
1967                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1968         if (maxpages > swap_header->info.last_page) {
1969                 maxpages = swap_header->info.last_page + 1;
1970                 /* p->max is an unsigned int: don't overflow it */
1971                 if ((unsigned int)maxpages == 0)
1972                         maxpages = UINT_MAX;
1973         }
1974         p->highest_bit = maxpages - 1;
1975
1976         error = -EINVAL;
1977         if (!maxpages)
1978                 goto bad_swap;
1979         if (swapfilepages && maxpages > swapfilepages) {
1980                 printk(KERN_WARNING
1981                        "Swap area shorter than signature indicates\n");
1982                 goto bad_swap;
1983         }
1984         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1985                 goto bad_swap;
1986         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1987                 goto bad_swap;
1988
1989         /* OK, set up the swap map and apply the bad block list */
1990         swap_map = vmalloc(maxpages);
1991         if (!swap_map) {
1992                 error = -ENOMEM;
1993                 goto bad_swap;
1994         }
1995
1996         memset(swap_map, 0, maxpages);
1997         nr_good_pages = maxpages - 1;   /* omit header page */
1998
1999         for (i = 0; i < swap_header->info.nr_badpages; i++) {
2000                 unsigned int page_nr = swap_header->info.badpages[i];
2001                 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2002                         error = -EINVAL;
2003                         goto bad_swap;
2004                 }
2005                 if (page_nr < maxpages) {
2006                         swap_map[page_nr] = SWAP_MAP_BAD;
2007                         nr_good_pages--;
2008                 }
2009         }
2010
2011         error = swap_cgroup_swapon(type, maxpages);
2012         if (error)
2013                 goto bad_swap;
2014
2015         if (nr_good_pages) {
2016                 swap_map[0] = SWAP_MAP_BAD;
2017                 p->max = maxpages;
2018                 p->pages = nr_good_pages;
2019                 nr_extents = setup_swap_extents(p, &span);
2020                 if (nr_extents < 0) {
2021                         error = nr_extents;
2022                         goto bad_swap;
2023                 }
2024                 nr_good_pages = p->pages;
2025         }
2026         if (!nr_good_pages) {
2027                 printk(KERN_WARNING "Empty swap-file\n");
2028                 error = -EINVAL;
2029                 goto bad_swap;
2030         }
2031
2032         if (p->bdev) {
2033                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2034                         p->flags |= SWP_SOLIDSTATE;
2035                         p->cluster_next = 1 + (random32() % p->highest_bit);
2036                 }
2037                 if (discard_swap(p) == 0)
2038                         p->flags |= SWP_DISCARDABLE;
2039         }
2040
2041         mutex_lock(&swapon_mutex);
2042         spin_lock(&swap_lock);
2043         if (swap_flags & SWAP_FLAG_PREFER)
2044                 p->prio =
2045                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2046         else
2047                 p->prio = --least_priority;
2048         p->swap_map = swap_map;
2049         p->flags |= SWP_WRITEOK;
2050         nr_swap_pages += nr_good_pages;
2051         total_swap_pages += nr_good_pages;
2052
2053         printk(KERN_INFO "Adding %uk swap on %s.  "
2054                         "Priority:%d extents:%d across:%lluk %s%s\n",
2055                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2056                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2057                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2058                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2059
2060         /* insert swap space into swap_list: */
2061         prev = -1;
2062         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2063                 if (p->prio >= swap_info[i]->prio)
2064                         break;
2065                 prev = i;
2066         }
2067         p->next = i;
2068         if (prev < 0)
2069                 swap_list.head = swap_list.next = type;
2070         else
2071                 swap_info[prev]->next = type;
2072         spin_unlock(&swap_lock);
2073         mutex_unlock(&swapon_mutex);
2074         error = 0;
2075         goto out;
2076 bad_swap:
2077         if (bdev) {
2078                 set_blocksize(bdev, p->old_block_size);
2079                 bd_release(bdev);
2080         }
2081         destroy_swap_extents(p);
2082         swap_cgroup_swapoff(type);
2083 bad_swap_2:
2084         spin_lock(&swap_lock);
2085         p->swap_file = NULL;
2086         p->flags = 0;
2087         spin_unlock(&swap_lock);
2088         vfree(swap_map);
2089         if (swap_file)
2090                 filp_close(swap_file, NULL);
2091 out:
2092         if (page && !IS_ERR(page)) {
2093                 kunmap(page);
2094                 page_cache_release(page);
2095         }
2096         if (name)
2097                 putname(name);
2098         if (did_down) {
2099                 if (!error)
2100                         inode->i_flags |= S_SWAPFILE;
2101                 mutex_unlock(&inode->i_mutex);
2102         }
2103         return error;
2104 }
2105
2106 void si_swapinfo(struct sysinfo *val)
2107 {
2108         unsigned int type;
2109         unsigned long nr_to_be_unused = 0;
2110
2111         spin_lock(&swap_lock);
2112         for (type = 0; type < nr_swapfiles; type++) {
2113                 struct swap_info_struct *si = swap_info[type];
2114
2115                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2116                         nr_to_be_unused += si->inuse_pages;
2117         }
2118         val->freeswap = nr_swap_pages + nr_to_be_unused;
2119         val->totalswap = total_swap_pages + nr_to_be_unused;
2120         spin_unlock(&swap_lock);
2121 }
2122
2123 /*
2124  * Verify that a swap entry is valid and increment its swap map count.
2125  *
2126  * Returns error code in following case.
2127  * - success -> 0
2128  * - swp_entry is invalid -> EINVAL
2129  * - swp_entry is migration entry -> EINVAL
2130  * - swap-cache reference is requested but there is already one. -> EEXIST
2131  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2132  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2133  */
2134 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2135 {
2136         struct swap_info_struct *p;
2137         unsigned long offset, type;
2138         unsigned char count;
2139         unsigned char has_cache;
2140         int err = -EINVAL;
2141
2142         if (non_swap_entry(entry))
2143                 goto out;
2144
2145         type = swp_type(entry);
2146         if (type >= nr_swapfiles)
2147                 goto bad_file;
2148         p = swap_info[type];
2149         offset = swp_offset(entry);
2150
2151         spin_lock(&swap_lock);
2152         if (unlikely(offset >= p->max))
2153                 goto unlock_out;
2154
2155         count = p->swap_map[offset];
2156         has_cache = count & SWAP_HAS_CACHE;
2157         count &= ~SWAP_HAS_CACHE;
2158         err = 0;
2159
2160         if (usage == SWAP_HAS_CACHE) {
2161
2162                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2163                 if (!has_cache && count)
2164                         has_cache = SWAP_HAS_CACHE;
2165                 else if (has_cache)             /* someone else added cache */
2166                         err = -EEXIST;
2167                 else                            /* no users remaining */
2168                         err = -ENOENT;
2169
2170         } else if (count || has_cache) {
2171
2172                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2173                         count += usage;
2174                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2175                         err = -EINVAL;
2176                 else if (swap_count_continued(p, offset, count))
2177                         count = COUNT_CONTINUED;
2178                 else
2179                         err = -ENOMEM;
2180         } else
2181                 err = -ENOENT;                  /* unused swap entry */
2182
2183         p->swap_map[offset] = count | has_cache;
2184
2185 unlock_out:
2186         spin_unlock(&swap_lock);
2187 out:
2188         return err;
2189
2190 bad_file:
2191         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2192         goto out;
2193 }
2194
2195 /*
2196  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2197  * (in which case its reference count is never incremented).
2198  */
2199 void swap_shmem_alloc(swp_entry_t entry)
2200 {
2201         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2202 }
2203
2204 /*
2205  * Increase reference count of swap entry by 1.
2206  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2207  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2208  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2209  * might occur if a page table entry has got corrupted.
2210  */
2211 int swap_duplicate(swp_entry_t entry)
2212 {
2213         int err = 0;
2214
2215         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2216                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2217         return err;
2218 }
2219
2220 /*
2221  * @entry: swap entry for which we allocate swap cache.
2222  *
2223  * Called when allocating swap cache for existing swap entry,
2224  * This can return error codes. Returns 0 at success.
2225  * -EBUSY means there is a swap cache.
2226  * Note: return code is different from swap_duplicate().
2227  */
2228 int swapcache_prepare(swp_entry_t entry)
2229 {
2230         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2231 }
2232
2233 /*
2234  * swap_lock prevents swap_map being freed. Don't grab an extra
2235  * reference on the swaphandle, it doesn't matter if it becomes unused.
2236  */
2237 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2238 {
2239         struct swap_info_struct *si;
2240         int our_page_cluster = page_cluster;
2241         pgoff_t target, toff;
2242         pgoff_t base, end;
2243         int nr_pages = 0;
2244
2245         if (!our_page_cluster)  /* no readahead */
2246                 return 0;
2247
2248         si = swap_info[swp_type(entry)];
2249         target = swp_offset(entry);
2250         base = (target >> our_page_cluster) << our_page_cluster;
2251         end = base + (1 << our_page_cluster);
2252         if (!base)              /* first page is swap header */
2253                 base++;
2254
2255         spin_lock(&swap_lock);
2256         if (end > si->max)      /* don't go beyond end of map */
2257                 end = si->max;
2258
2259         /* Count contiguous allocated slots above our target */
2260         for (toff = target; ++toff < end; nr_pages++) {
2261                 /* Don't read in free or bad pages */
2262                 if (!si->swap_map[toff])
2263                         break;
2264                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2265                         break;
2266         }
2267         /* Count contiguous allocated slots below our target */
2268         for (toff = target; --toff >= base; nr_pages++) {
2269                 /* Don't read in free or bad pages */
2270                 if (!si->swap_map[toff])
2271                         break;
2272                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2273                         break;
2274         }
2275         spin_unlock(&swap_lock);
2276
2277         /*
2278          * Indicate starting offset, and return number of pages to get:
2279          * if only 1, say 0, since there's then no readahead to be done.
2280          */
2281         *offset = ++toff;
2282         return nr_pages? ++nr_pages: 0;
2283 }
2284
2285 /*
2286  * add_swap_count_continuation - called when a swap count is duplicated
2287  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2288  * page of the original vmalloc'ed swap_map, to hold the continuation count
2289  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2290  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2291  *
2292  * These continuation pages are seldom referenced: the common paths all work
2293  * on the original swap_map, only referring to a continuation page when the
2294  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2295  *
2296  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2297  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2298  * can be called after dropping locks.
2299  */
2300 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2301 {
2302         struct swap_info_struct *si;
2303         struct page *head;
2304         struct page *page;
2305         struct page *list_page;
2306         pgoff_t offset;
2307         unsigned char count;
2308
2309         /*
2310          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2311          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2312          */
2313         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2314
2315         si = swap_info_get(entry);
2316         if (!si) {
2317                 /*
2318                  * An acceptable race has occurred since the failing
2319                  * __swap_duplicate(): the swap entry has been freed,
2320                  * perhaps even the whole swap_map cleared for swapoff.
2321                  */
2322                 goto outer;
2323         }
2324
2325         offset = swp_offset(entry);
2326         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2327
2328         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2329                 /*
2330                  * The higher the swap count, the more likely it is that tasks
2331                  * will race to add swap count continuation: we need to avoid
2332                  * over-provisioning.
2333                  */
2334                 goto out;
2335         }
2336
2337         if (!page) {
2338                 spin_unlock(&swap_lock);
2339                 return -ENOMEM;
2340         }
2341
2342         /*
2343          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2344          * no architecture is using highmem pages for kernel pagetables: so it
2345          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2346          */
2347         head = vmalloc_to_page(si->swap_map + offset);
2348         offset &= ~PAGE_MASK;
2349
2350         /*
2351          * Page allocation does not initialize the page's lru field,
2352          * but it does always reset its private field.
2353          */
2354         if (!page_private(head)) {
2355                 BUG_ON(count & COUNT_CONTINUED);
2356                 INIT_LIST_HEAD(&head->lru);
2357                 set_page_private(head, SWP_CONTINUED);
2358                 si->flags |= SWP_CONTINUED;
2359         }
2360
2361         list_for_each_entry(list_page, &head->lru, lru) {
2362                 unsigned char *map;
2363
2364                 /*
2365                  * If the previous map said no continuation, but we've found
2366                  * a continuation page, free our allocation and use this one.
2367                  */
2368                 if (!(count & COUNT_CONTINUED))
2369                         goto out;
2370
2371                 map = kmap_atomic(list_page, KM_USER0) + offset;
2372                 count = *map;
2373                 kunmap_atomic(map, KM_USER0);
2374
2375                 /*
2376                  * If this continuation count now has some space in it,
2377                  * free our allocation and use this one.
2378                  */
2379                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2380                         goto out;
2381         }
2382
2383         list_add_tail(&page->lru, &head->lru);
2384         page = NULL;                    /* now it's attached, don't free it */
2385 out:
2386         spin_unlock(&swap_lock);
2387 outer:
2388         if (page)
2389                 __free_page(page);
2390         return 0;
2391 }
2392
2393 /*
2394  * swap_count_continued - when the original swap_map count is incremented
2395  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2396  * into, carry if so, or else fail until a new continuation page is allocated;
2397  * when the original swap_map count is decremented from 0 with continuation,
2398  * borrow from the continuation and report whether it still holds more.
2399  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2400  */
2401 static bool swap_count_continued(struct swap_info_struct *si,
2402                                  pgoff_t offset, unsigned char count)
2403 {
2404         struct page *head;
2405         struct page *page;
2406         unsigned char *map;
2407
2408         head = vmalloc_to_page(si->swap_map + offset);
2409         if (page_private(head) != SWP_CONTINUED) {
2410                 BUG_ON(count & COUNT_CONTINUED);
2411                 return false;           /* need to add count continuation */
2412         }
2413
2414         offset &= ~PAGE_MASK;
2415         page = list_entry(head->lru.next, struct page, lru);
2416         map = kmap_atomic(page, KM_USER0) + offset;
2417
2418         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2419                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2420
2421         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2422                 /*
2423                  * Think of how you add 1 to 999
2424                  */
2425                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2426                         kunmap_atomic(map, KM_USER0);
2427                         page = list_entry(page->lru.next, struct page, lru);
2428                         BUG_ON(page == head);
2429                         map = kmap_atomic(page, KM_USER0) + offset;
2430                 }
2431                 if (*map == SWAP_CONT_MAX) {
2432                         kunmap_atomic(map, KM_USER0);
2433                         page = list_entry(page->lru.next, struct page, lru);
2434                         if (page == head)
2435                                 return false;   /* add count continuation */
2436                         map = kmap_atomic(page, KM_USER0) + offset;
2437 init_map:               *map = 0;               /* we didn't zero the page */
2438                 }
2439                 *map += 1;
2440                 kunmap_atomic(map, KM_USER0);
2441                 page = list_entry(page->lru.prev, struct page, lru);
2442                 while (page != head) {
2443                         map = kmap_atomic(page, KM_USER0) + offset;
2444                         *map = COUNT_CONTINUED;
2445                         kunmap_atomic(map, KM_USER0);
2446                         page = list_entry(page->lru.prev, struct page, lru);
2447                 }
2448                 return true;                    /* incremented */
2449
2450         } else {                                /* decrementing */
2451                 /*
2452                  * Think of how you subtract 1 from 1000
2453                  */
2454                 BUG_ON(count != COUNT_CONTINUED);
2455                 while (*map == COUNT_CONTINUED) {
2456                         kunmap_atomic(map, KM_USER0);
2457                         page = list_entry(page->lru.next, struct page, lru);
2458                         BUG_ON(page == head);
2459                         map = kmap_atomic(page, KM_USER0) + offset;
2460                 }
2461                 BUG_ON(*map == 0);
2462                 *map -= 1;
2463                 if (*map == 0)
2464                         count = 0;
2465                 kunmap_atomic(map, KM_USER0);
2466                 page = list_entry(page->lru.prev, struct page, lru);
2467                 while (page != head) {
2468                         map = kmap_atomic(page, KM_USER0) + offset;
2469                         *map = SWAP_CONT_MAX | count;
2470                         count = COUNT_CONTINUED;
2471                         kunmap_atomic(map, KM_USER0);
2472                         page = list_entry(page->lru.prev, struct page, lru);
2473                 }
2474                 return count == COUNT_CONTINUED;
2475         }
2476 }
2477
2478 /*
2479  * free_swap_count_continuations - swapoff free all the continuation pages
2480  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2481  */
2482 static void free_swap_count_continuations(struct swap_info_struct *si)
2483 {
2484         pgoff_t offset;
2485
2486         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2487                 struct page *head;
2488                 head = vmalloc_to_page(si->swap_map + offset);
2489                 if (page_private(head)) {
2490                         struct list_head *this, *next;
2491                         list_for_each_safe(this, next, &head->lru) {
2492                                 struct page *page;
2493                                 page = list_entry(this, struct page, lru);
2494                                 list_del(this);
2495                                 __free_page(page);
2496                         }
2497                 }
2498         }
2499 }