]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - mm/swapfile.c
mm: refactor reinsert of swap_info in sys_swapoff()
[~shefty/rdma-dev.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shmem_fs.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32 #include <linux/poll.h>
33 #include <linux/oom.h>
34 #include <linux/frontswap.h>
35 #include <linux/swapfile.h>
36 #include <linux/export.h>
37
38 #include <asm/pgtable.h>
39 #include <asm/tlbflush.h>
40 #include <linux/swapops.h>
41 #include <linux/page_cgroup.h>
42
43 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
44                                  unsigned char);
45 static void free_swap_count_continuations(struct swap_info_struct *);
46 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
47
48 DEFINE_SPINLOCK(swap_lock);
49 static unsigned int nr_swapfiles;
50 long nr_swap_pages;
51 long total_swap_pages;
52 static int least_priority;
53
54 static const char Bad_file[] = "Bad swap file entry ";
55 static const char Unused_file[] = "Unused swap file entry ";
56 static const char Bad_offset[] = "Bad swap offset entry ";
57 static const char Unused_offset[] = "Unused swap offset entry ";
58
59 struct swap_list_t swap_list = {-1, -1};
60
61 struct swap_info_struct *swap_info[MAX_SWAPFILES];
62
63 static DEFINE_MUTEX(swapon_mutex);
64
65 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
66 /* Activity counter to indicate that a swapon or swapoff has occurred */
67 static atomic_t proc_poll_event = ATOMIC_INIT(0);
68
69 static inline unsigned char swap_count(unsigned char ent)
70 {
71         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
72 }
73
74 /* returns 1 if swap entry is freed */
75 static int
76 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
77 {
78         swp_entry_t entry = swp_entry(si->type, offset);
79         struct page *page;
80         int ret = 0;
81
82         page = find_get_page(&swapper_space, entry.val);
83         if (!page)
84                 return 0;
85         /*
86          * This function is called from scan_swap_map() and it's called
87          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
88          * We have to use trylock for avoiding deadlock. This is a special
89          * case and you should use try_to_free_swap() with explicit lock_page()
90          * in usual operations.
91          */
92         if (trylock_page(page)) {
93                 ret = try_to_free_swap(page);
94                 unlock_page(page);
95         }
96         page_cache_release(page);
97         return ret;
98 }
99
100 /*
101  * swapon tell device that all the old swap contents can be discarded,
102  * to allow the swap device to optimize its wear-levelling.
103  */
104 static int discard_swap(struct swap_info_struct *si)
105 {
106         struct swap_extent *se;
107         sector_t start_block;
108         sector_t nr_blocks;
109         int err = 0;
110
111         /* Do not discard the swap header page! */
112         se = &si->first_swap_extent;
113         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
114         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
115         if (nr_blocks) {
116                 err = blkdev_issue_discard(si->bdev, start_block,
117                                 nr_blocks, GFP_KERNEL, 0);
118                 if (err)
119                         return err;
120                 cond_resched();
121         }
122
123         list_for_each_entry(se, &si->first_swap_extent.list, list) {
124                 start_block = se->start_block << (PAGE_SHIFT - 9);
125                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
126
127                 err = blkdev_issue_discard(si->bdev, start_block,
128                                 nr_blocks, GFP_KERNEL, 0);
129                 if (err)
130                         break;
131
132                 cond_resched();
133         }
134         return err;             /* That will often be -EOPNOTSUPP */
135 }
136
137 /*
138  * swap allocation tell device that a cluster of swap can now be discarded,
139  * to allow the swap device to optimize its wear-levelling.
140  */
141 static void discard_swap_cluster(struct swap_info_struct *si,
142                                  pgoff_t start_page, pgoff_t nr_pages)
143 {
144         struct swap_extent *se = si->curr_swap_extent;
145         int found_extent = 0;
146
147         while (nr_pages) {
148                 struct list_head *lh;
149
150                 if (se->start_page <= start_page &&
151                     start_page < se->start_page + se->nr_pages) {
152                         pgoff_t offset = start_page - se->start_page;
153                         sector_t start_block = se->start_block + offset;
154                         sector_t nr_blocks = se->nr_pages - offset;
155
156                         if (nr_blocks > nr_pages)
157                                 nr_blocks = nr_pages;
158                         start_page += nr_blocks;
159                         nr_pages -= nr_blocks;
160
161                         if (!found_extent++)
162                                 si->curr_swap_extent = se;
163
164                         start_block <<= PAGE_SHIFT - 9;
165                         nr_blocks <<= PAGE_SHIFT - 9;
166                         if (blkdev_issue_discard(si->bdev, start_block,
167                                     nr_blocks, GFP_NOIO, 0))
168                                 break;
169                 }
170
171                 lh = se->list.next;
172                 se = list_entry(lh, struct swap_extent, list);
173         }
174 }
175
176 static int wait_for_discard(void *word)
177 {
178         schedule();
179         return 0;
180 }
181
182 #define SWAPFILE_CLUSTER        256
183 #define LATENCY_LIMIT           256
184
185 static unsigned long scan_swap_map(struct swap_info_struct *si,
186                                    unsigned char usage)
187 {
188         unsigned long offset;
189         unsigned long scan_base;
190         unsigned long last_in_cluster = 0;
191         int latency_ration = LATENCY_LIMIT;
192         int found_free_cluster = 0;
193
194         /*
195          * We try to cluster swap pages by allocating them sequentially
196          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
197          * way, however, we resort to first-free allocation, starting
198          * a new cluster.  This prevents us from scattering swap pages
199          * all over the entire swap partition, so that we reduce
200          * overall disk seek times between swap pages.  -- sct
201          * But we do now try to find an empty cluster.  -Andrea
202          * And we let swap pages go all over an SSD partition.  Hugh
203          */
204
205         si->flags += SWP_SCANNING;
206         scan_base = offset = si->cluster_next;
207
208         if (unlikely(!si->cluster_nr--)) {
209                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
210                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
211                         goto checks;
212                 }
213                 if (si->flags & SWP_DISCARDABLE) {
214                         /*
215                          * Start range check on racing allocations, in case
216                          * they overlap the cluster we eventually decide on
217                          * (we scan without swap_lock to allow preemption).
218                          * It's hardly conceivable that cluster_nr could be
219                          * wrapped during our scan, but don't depend on it.
220                          */
221                         if (si->lowest_alloc)
222                                 goto checks;
223                         si->lowest_alloc = si->max;
224                         si->highest_alloc = 0;
225                 }
226                 spin_unlock(&swap_lock);
227
228                 /*
229                  * If seek is expensive, start searching for new cluster from
230                  * start of partition, to minimize the span of allocated swap.
231                  * But if seek is cheap, search from our current position, so
232                  * that swap is allocated from all over the partition: if the
233                  * Flash Translation Layer only remaps within limited zones,
234                  * we don't want to wear out the first zone too quickly.
235                  */
236                 if (!(si->flags & SWP_SOLIDSTATE))
237                         scan_base = offset = si->lowest_bit;
238                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
239
240                 /* Locate the first empty (unaligned) cluster */
241                 for (; last_in_cluster <= si->highest_bit; offset++) {
242                         if (si->swap_map[offset])
243                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
244                         else if (offset == last_in_cluster) {
245                                 spin_lock(&swap_lock);
246                                 offset -= SWAPFILE_CLUSTER - 1;
247                                 si->cluster_next = offset;
248                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
249                                 found_free_cluster = 1;
250                                 goto checks;
251                         }
252                         if (unlikely(--latency_ration < 0)) {
253                                 cond_resched();
254                                 latency_ration = LATENCY_LIMIT;
255                         }
256                 }
257
258                 offset = si->lowest_bit;
259                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
260
261                 /* Locate the first empty (unaligned) cluster */
262                 for (; last_in_cluster < scan_base; offset++) {
263                         if (si->swap_map[offset])
264                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
265                         else if (offset == last_in_cluster) {
266                                 spin_lock(&swap_lock);
267                                 offset -= SWAPFILE_CLUSTER - 1;
268                                 si->cluster_next = offset;
269                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
270                                 found_free_cluster = 1;
271                                 goto checks;
272                         }
273                         if (unlikely(--latency_ration < 0)) {
274                                 cond_resched();
275                                 latency_ration = LATENCY_LIMIT;
276                         }
277                 }
278
279                 offset = scan_base;
280                 spin_lock(&swap_lock);
281                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
282                 si->lowest_alloc = 0;
283         }
284
285 checks:
286         if (!(si->flags & SWP_WRITEOK))
287                 goto no_page;
288         if (!si->highest_bit)
289                 goto no_page;
290         if (offset > si->highest_bit)
291                 scan_base = offset = si->lowest_bit;
292
293         /* reuse swap entry of cache-only swap if not busy. */
294         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
295                 int swap_was_freed;
296                 spin_unlock(&swap_lock);
297                 swap_was_freed = __try_to_reclaim_swap(si, offset);
298                 spin_lock(&swap_lock);
299                 /* entry was freed successfully, try to use this again */
300                 if (swap_was_freed)
301                         goto checks;
302                 goto scan; /* check next one */
303         }
304
305         if (si->swap_map[offset])
306                 goto scan;
307
308         if (offset == si->lowest_bit)
309                 si->lowest_bit++;
310         if (offset == si->highest_bit)
311                 si->highest_bit--;
312         si->inuse_pages++;
313         if (si->inuse_pages == si->pages) {
314                 si->lowest_bit = si->max;
315                 si->highest_bit = 0;
316         }
317         si->swap_map[offset] = usage;
318         si->cluster_next = offset + 1;
319         si->flags -= SWP_SCANNING;
320
321         if (si->lowest_alloc) {
322                 /*
323                  * Only set when SWP_DISCARDABLE, and there's a scan
324                  * for a free cluster in progress or just completed.
325                  */
326                 if (found_free_cluster) {
327                         /*
328                          * To optimize wear-levelling, discard the
329                          * old data of the cluster, taking care not to
330                          * discard any of its pages that have already
331                          * been allocated by racing tasks (offset has
332                          * already stepped over any at the beginning).
333                          */
334                         if (offset < si->highest_alloc &&
335                             si->lowest_alloc <= last_in_cluster)
336                                 last_in_cluster = si->lowest_alloc - 1;
337                         si->flags |= SWP_DISCARDING;
338                         spin_unlock(&swap_lock);
339
340                         if (offset < last_in_cluster)
341                                 discard_swap_cluster(si, offset,
342                                         last_in_cluster - offset + 1);
343
344                         spin_lock(&swap_lock);
345                         si->lowest_alloc = 0;
346                         si->flags &= ~SWP_DISCARDING;
347
348                         smp_mb();       /* wake_up_bit advises this */
349                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
350
351                 } else if (si->flags & SWP_DISCARDING) {
352                         /*
353                          * Delay using pages allocated by racing tasks
354                          * until the whole discard has been issued. We
355                          * could defer that delay until swap_writepage,
356                          * but it's easier to keep this self-contained.
357                          */
358                         spin_unlock(&swap_lock);
359                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
360                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
361                         spin_lock(&swap_lock);
362                 } else {
363                         /*
364                          * Note pages allocated by racing tasks while
365                          * scan for a free cluster is in progress, so
366                          * that its final discard can exclude them.
367                          */
368                         if (offset < si->lowest_alloc)
369                                 si->lowest_alloc = offset;
370                         if (offset > si->highest_alloc)
371                                 si->highest_alloc = offset;
372                 }
373         }
374         return offset;
375
376 scan:
377         spin_unlock(&swap_lock);
378         while (++offset <= si->highest_bit) {
379                 if (!si->swap_map[offset]) {
380                         spin_lock(&swap_lock);
381                         goto checks;
382                 }
383                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
384                         spin_lock(&swap_lock);
385                         goto checks;
386                 }
387                 if (unlikely(--latency_ration < 0)) {
388                         cond_resched();
389                         latency_ration = LATENCY_LIMIT;
390                 }
391         }
392         offset = si->lowest_bit;
393         while (++offset < scan_base) {
394                 if (!si->swap_map[offset]) {
395                         spin_lock(&swap_lock);
396                         goto checks;
397                 }
398                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
399                         spin_lock(&swap_lock);
400                         goto checks;
401                 }
402                 if (unlikely(--latency_ration < 0)) {
403                         cond_resched();
404                         latency_ration = LATENCY_LIMIT;
405                 }
406         }
407         spin_lock(&swap_lock);
408
409 no_page:
410         si->flags -= SWP_SCANNING;
411         return 0;
412 }
413
414 swp_entry_t get_swap_page(void)
415 {
416         struct swap_info_struct *si;
417         pgoff_t offset;
418         int type, next;
419         int wrapped = 0;
420
421         spin_lock(&swap_lock);
422         if (nr_swap_pages <= 0)
423                 goto noswap;
424         nr_swap_pages--;
425
426         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
427                 si = swap_info[type];
428                 next = si->next;
429                 if (next < 0 ||
430                     (!wrapped && si->prio != swap_info[next]->prio)) {
431                         next = swap_list.head;
432                         wrapped++;
433                 }
434
435                 if (!si->highest_bit)
436                         continue;
437                 if (!(si->flags & SWP_WRITEOK))
438                         continue;
439
440                 swap_list.next = next;
441                 /* This is called for allocating swap entry for cache */
442                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
443                 if (offset) {
444                         spin_unlock(&swap_lock);
445                         return swp_entry(type, offset);
446                 }
447                 next = swap_list.next;
448         }
449
450         nr_swap_pages++;
451 noswap:
452         spin_unlock(&swap_lock);
453         return (swp_entry_t) {0};
454 }
455
456 /* The only caller of this function is now susupend routine */
457 swp_entry_t get_swap_page_of_type(int type)
458 {
459         struct swap_info_struct *si;
460         pgoff_t offset;
461
462         spin_lock(&swap_lock);
463         si = swap_info[type];
464         if (si && (si->flags & SWP_WRITEOK)) {
465                 nr_swap_pages--;
466                 /* This is called for allocating swap entry, not cache */
467                 offset = scan_swap_map(si, 1);
468                 if (offset) {
469                         spin_unlock(&swap_lock);
470                         return swp_entry(type, offset);
471                 }
472                 nr_swap_pages++;
473         }
474         spin_unlock(&swap_lock);
475         return (swp_entry_t) {0};
476 }
477
478 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
479 {
480         struct swap_info_struct *p;
481         unsigned long offset, type;
482
483         if (!entry.val)
484                 goto out;
485         type = swp_type(entry);
486         if (type >= nr_swapfiles)
487                 goto bad_nofile;
488         p = swap_info[type];
489         if (!(p->flags & SWP_USED))
490                 goto bad_device;
491         offset = swp_offset(entry);
492         if (offset >= p->max)
493                 goto bad_offset;
494         if (!p->swap_map[offset])
495                 goto bad_free;
496         spin_lock(&swap_lock);
497         return p;
498
499 bad_free:
500         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
501         goto out;
502 bad_offset:
503         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
504         goto out;
505 bad_device:
506         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
507         goto out;
508 bad_nofile:
509         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
510 out:
511         return NULL;
512 }
513
514 static unsigned char swap_entry_free(struct swap_info_struct *p,
515                                      swp_entry_t entry, unsigned char usage)
516 {
517         unsigned long offset = swp_offset(entry);
518         unsigned char count;
519         unsigned char has_cache;
520
521         count = p->swap_map[offset];
522         has_cache = count & SWAP_HAS_CACHE;
523         count &= ~SWAP_HAS_CACHE;
524
525         if (usage == SWAP_HAS_CACHE) {
526                 VM_BUG_ON(!has_cache);
527                 has_cache = 0;
528         } else if (count == SWAP_MAP_SHMEM) {
529                 /*
530                  * Or we could insist on shmem.c using a special
531                  * swap_shmem_free() and free_shmem_swap_and_cache()...
532                  */
533                 count = 0;
534         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
535                 if (count == COUNT_CONTINUED) {
536                         if (swap_count_continued(p, offset, count))
537                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
538                         else
539                                 count = SWAP_MAP_MAX;
540                 } else
541                         count--;
542         }
543
544         if (!count)
545                 mem_cgroup_uncharge_swap(entry);
546
547         usage = count | has_cache;
548         p->swap_map[offset] = usage;
549
550         /* free if no reference */
551         if (!usage) {
552                 if (offset < p->lowest_bit)
553                         p->lowest_bit = offset;
554                 if (offset > p->highest_bit)
555                         p->highest_bit = offset;
556                 if (swap_list.next >= 0 &&
557                     p->prio > swap_info[swap_list.next]->prio)
558                         swap_list.next = p->type;
559                 nr_swap_pages++;
560                 p->inuse_pages--;
561                 frontswap_invalidate_page(p->type, offset);
562                 if (p->flags & SWP_BLKDEV) {
563                         struct gendisk *disk = p->bdev->bd_disk;
564                         if (disk->fops->swap_slot_free_notify)
565                                 disk->fops->swap_slot_free_notify(p->bdev,
566                                                                   offset);
567                 }
568         }
569
570         return usage;
571 }
572
573 /*
574  * Caller has made sure that the swapdevice corresponding to entry
575  * is still around or has not been recycled.
576  */
577 void swap_free(swp_entry_t entry)
578 {
579         struct swap_info_struct *p;
580
581         p = swap_info_get(entry);
582         if (p) {
583                 swap_entry_free(p, entry, 1);
584                 spin_unlock(&swap_lock);
585         }
586 }
587
588 /*
589  * Called after dropping swapcache to decrease refcnt to swap entries.
590  */
591 void swapcache_free(swp_entry_t entry, struct page *page)
592 {
593         struct swap_info_struct *p;
594         unsigned char count;
595
596         p = swap_info_get(entry);
597         if (p) {
598                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
599                 if (page)
600                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
601                 spin_unlock(&swap_lock);
602         }
603 }
604
605 /*
606  * How many references to page are currently swapped out?
607  * This does not give an exact answer when swap count is continued,
608  * but does include the high COUNT_CONTINUED flag to allow for that.
609  */
610 int page_swapcount(struct page *page)
611 {
612         int count = 0;
613         struct swap_info_struct *p;
614         swp_entry_t entry;
615
616         entry.val = page_private(page);
617         p = swap_info_get(entry);
618         if (p) {
619                 count = swap_count(p->swap_map[swp_offset(entry)]);
620                 spin_unlock(&swap_lock);
621         }
622         return count;
623 }
624
625 /*
626  * We can write to an anon page without COW if there are no other references
627  * to it.  And as a side-effect, free up its swap: because the old content
628  * on disk will never be read, and seeking back there to write new content
629  * later would only waste time away from clustering.
630  */
631 int reuse_swap_page(struct page *page)
632 {
633         int count;
634
635         VM_BUG_ON(!PageLocked(page));
636         if (unlikely(PageKsm(page)))
637                 return 0;
638         count = page_mapcount(page);
639         if (count <= 1 && PageSwapCache(page)) {
640                 count += page_swapcount(page);
641                 if (count == 1 && !PageWriteback(page)) {
642                         delete_from_swap_cache(page);
643                         SetPageDirty(page);
644                 }
645         }
646         return count <= 1;
647 }
648
649 /*
650  * If swap is getting full, or if there are no more mappings of this page,
651  * then try_to_free_swap is called to free its swap space.
652  */
653 int try_to_free_swap(struct page *page)
654 {
655         VM_BUG_ON(!PageLocked(page));
656
657         if (!PageSwapCache(page))
658                 return 0;
659         if (PageWriteback(page))
660                 return 0;
661         if (page_swapcount(page))
662                 return 0;
663
664         /*
665          * Once hibernation has begun to create its image of memory,
666          * there's a danger that one of the calls to try_to_free_swap()
667          * - most probably a call from __try_to_reclaim_swap() while
668          * hibernation is allocating its own swap pages for the image,
669          * but conceivably even a call from memory reclaim - will free
670          * the swap from a page which has already been recorded in the
671          * image as a clean swapcache page, and then reuse its swap for
672          * another page of the image.  On waking from hibernation, the
673          * original page might be freed under memory pressure, then
674          * later read back in from swap, now with the wrong data.
675          *
676          * Hibration suspends storage while it is writing the image
677          * to disk so check that here.
678          */
679         if (pm_suspended_storage())
680                 return 0;
681
682         delete_from_swap_cache(page);
683         SetPageDirty(page);
684         return 1;
685 }
686
687 /*
688  * Free the swap entry like above, but also try to
689  * free the page cache entry if it is the last user.
690  */
691 int free_swap_and_cache(swp_entry_t entry)
692 {
693         struct swap_info_struct *p;
694         struct page *page = NULL;
695
696         if (non_swap_entry(entry))
697                 return 1;
698
699         p = swap_info_get(entry);
700         if (p) {
701                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
702                         page = find_get_page(&swapper_space, entry.val);
703                         if (page && !trylock_page(page)) {
704                                 page_cache_release(page);
705                                 page = NULL;
706                         }
707                 }
708                 spin_unlock(&swap_lock);
709         }
710         if (page) {
711                 /*
712                  * Not mapped elsewhere, or swap space full? Free it!
713                  * Also recheck PageSwapCache now page is locked (above).
714                  */
715                 if (PageSwapCache(page) && !PageWriteback(page) &&
716                                 (!page_mapped(page) || vm_swap_full())) {
717                         delete_from_swap_cache(page);
718                         SetPageDirty(page);
719                 }
720                 unlock_page(page);
721                 page_cache_release(page);
722         }
723         return p != NULL;
724 }
725
726 #ifdef CONFIG_HIBERNATION
727 /*
728  * Find the swap type that corresponds to given device (if any).
729  *
730  * @offset - number of the PAGE_SIZE-sized block of the device, starting
731  * from 0, in which the swap header is expected to be located.
732  *
733  * This is needed for the suspend to disk (aka swsusp).
734  */
735 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
736 {
737         struct block_device *bdev = NULL;
738         int type;
739
740         if (device)
741                 bdev = bdget(device);
742
743         spin_lock(&swap_lock);
744         for (type = 0; type < nr_swapfiles; type++) {
745                 struct swap_info_struct *sis = swap_info[type];
746
747                 if (!(sis->flags & SWP_WRITEOK))
748                         continue;
749
750                 if (!bdev) {
751                         if (bdev_p)
752                                 *bdev_p = bdgrab(sis->bdev);
753
754                         spin_unlock(&swap_lock);
755                         return type;
756                 }
757                 if (bdev == sis->bdev) {
758                         struct swap_extent *se = &sis->first_swap_extent;
759
760                         if (se->start_block == offset) {
761                                 if (bdev_p)
762                                         *bdev_p = bdgrab(sis->bdev);
763
764                                 spin_unlock(&swap_lock);
765                                 bdput(bdev);
766                                 return type;
767                         }
768                 }
769         }
770         spin_unlock(&swap_lock);
771         if (bdev)
772                 bdput(bdev);
773
774         return -ENODEV;
775 }
776
777 /*
778  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
779  * corresponding to given index in swap_info (swap type).
780  */
781 sector_t swapdev_block(int type, pgoff_t offset)
782 {
783         struct block_device *bdev;
784
785         if ((unsigned int)type >= nr_swapfiles)
786                 return 0;
787         if (!(swap_info[type]->flags & SWP_WRITEOK))
788                 return 0;
789         return map_swap_entry(swp_entry(type, offset), &bdev);
790 }
791
792 /*
793  * Return either the total number of swap pages of given type, or the number
794  * of free pages of that type (depending on @free)
795  *
796  * This is needed for software suspend
797  */
798 unsigned int count_swap_pages(int type, int free)
799 {
800         unsigned int n = 0;
801
802         spin_lock(&swap_lock);
803         if ((unsigned int)type < nr_swapfiles) {
804                 struct swap_info_struct *sis = swap_info[type];
805
806                 if (sis->flags & SWP_WRITEOK) {
807                         n = sis->pages;
808                         if (free)
809                                 n -= sis->inuse_pages;
810                 }
811         }
812         spin_unlock(&swap_lock);
813         return n;
814 }
815 #endif /* CONFIG_HIBERNATION */
816
817 /*
818  * No need to decide whether this PTE shares the swap entry with others,
819  * just let do_wp_page work it out if a write is requested later - to
820  * force COW, vm_page_prot omits write permission from any private vma.
821  */
822 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
823                 unsigned long addr, swp_entry_t entry, struct page *page)
824 {
825         struct mem_cgroup *memcg;
826         spinlock_t *ptl;
827         pte_t *pte;
828         int ret = 1;
829
830         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
831                                          GFP_KERNEL, &memcg)) {
832                 ret = -ENOMEM;
833                 goto out_nolock;
834         }
835
836         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
837         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
838                 mem_cgroup_cancel_charge_swapin(memcg);
839                 ret = 0;
840                 goto out;
841         }
842
843         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
844         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
845         get_page(page);
846         set_pte_at(vma->vm_mm, addr, pte,
847                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
848         page_add_anon_rmap(page, vma, addr);
849         mem_cgroup_commit_charge_swapin(page, memcg);
850         swap_free(entry);
851         /*
852          * Move the page to the active list so it is not
853          * immediately swapped out again after swapon.
854          */
855         activate_page(page);
856 out:
857         pte_unmap_unlock(pte, ptl);
858 out_nolock:
859         return ret;
860 }
861
862 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
863                                 unsigned long addr, unsigned long end,
864                                 swp_entry_t entry, struct page *page)
865 {
866         pte_t swp_pte = swp_entry_to_pte(entry);
867         pte_t *pte;
868         int ret = 0;
869
870         /*
871          * We don't actually need pte lock while scanning for swp_pte: since
872          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
873          * page table while we're scanning; though it could get zapped, and on
874          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
875          * of unmatched parts which look like swp_pte, so unuse_pte must
876          * recheck under pte lock.  Scanning without pte lock lets it be
877          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
878          */
879         pte = pte_offset_map(pmd, addr);
880         do {
881                 /*
882                  * swapoff spends a _lot_ of time in this loop!
883                  * Test inline before going to call unuse_pte.
884                  */
885                 if (unlikely(pte_same(*pte, swp_pte))) {
886                         pte_unmap(pte);
887                         ret = unuse_pte(vma, pmd, addr, entry, page);
888                         if (ret)
889                                 goto out;
890                         pte = pte_offset_map(pmd, addr);
891                 }
892         } while (pte++, addr += PAGE_SIZE, addr != end);
893         pte_unmap(pte - 1);
894 out:
895         return ret;
896 }
897
898 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
899                                 unsigned long addr, unsigned long end,
900                                 swp_entry_t entry, struct page *page)
901 {
902         pmd_t *pmd;
903         unsigned long next;
904         int ret;
905
906         pmd = pmd_offset(pud, addr);
907         do {
908                 next = pmd_addr_end(addr, end);
909                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
910                         continue;
911                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
912                 if (ret)
913                         return ret;
914         } while (pmd++, addr = next, addr != end);
915         return 0;
916 }
917
918 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
919                                 unsigned long addr, unsigned long end,
920                                 swp_entry_t entry, struct page *page)
921 {
922         pud_t *pud;
923         unsigned long next;
924         int ret;
925
926         pud = pud_offset(pgd, addr);
927         do {
928                 next = pud_addr_end(addr, end);
929                 if (pud_none_or_clear_bad(pud))
930                         continue;
931                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
932                 if (ret)
933                         return ret;
934         } while (pud++, addr = next, addr != end);
935         return 0;
936 }
937
938 static int unuse_vma(struct vm_area_struct *vma,
939                                 swp_entry_t entry, struct page *page)
940 {
941         pgd_t *pgd;
942         unsigned long addr, end, next;
943         int ret;
944
945         if (page_anon_vma(page)) {
946                 addr = page_address_in_vma(page, vma);
947                 if (addr == -EFAULT)
948                         return 0;
949                 else
950                         end = addr + PAGE_SIZE;
951         } else {
952                 addr = vma->vm_start;
953                 end = vma->vm_end;
954         }
955
956         pgd = pgd_offset(vma->vm_mm, addr);
957         do {
958                 next = pgd_addr_end(addr, end);
959                 if (pgd_none_or_clear_bad(pgd))
960                         continue;
961                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
962                 if (ret)
963                         return ret;
964         } while (pgd++, addr = next, addr != end);
965         return 0;
966 }
967
968 static int unuse_mm(struct mm_struct *mm,
969                                 swp_entry_t entry, struct page *page)
970 {
971         struct vm_area_struct *vma;
972         int ret = 0;
973
974         if (!down_read_trylock(&mm->mmap_sem)) {
975                 /*
976                  * Activate page so shrink_inactive_list is unlikely to unmap
977                  * its ptes while lock is dropped, so swapoff can make progress.
978                  */
979                 activate_page(page);
980                 unlock_page(page);
981                 down_read(&mm->mmap_sem);
982                 lock_page(page);
983         }
984         for (vma = mm->mmap; vma; vma = vma->vm_next) {
985                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
986                         break;
987         }
988         up_read(&mm->mmap_sem);
989         return (ret < 0)? ret: 0;
990 }
991
992 /*
993  * Scan swap_map (or frontswap_map if frontswap parameter is true)
994  * from current position to next entry still in use.
995  * Recycle to start on reaching the end, returning 0 when empty.
996  */
997 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
998                                         unsigned int prev, bool frontswap)
999 {
1000         unsigned int max = si->max;
1001         unsigned int i = prev;
1002         unsigned char count;
1003
1004         /*
1005          * No need for swap_lock here: we're just looking
1006          * for whether an entry is in use, not modifying it; false
1007          * hits are okay, and sys_swapoff() has already prevented new
1008          * allocations from this area (while holding swap_lock).
1009          */
1010         for (;;) {
1011                 if (++i >= max) {
1012                         if (!prev) {
1013                                 i = 0;
1014                                 break;
1015                         }
1016                         /*
1017                          * No entries in use at top of swap_map,
1018                          * loop back to start and recheck there.
1019                          */
1020                         max = prev + 1;
1021                         prev = 0;
1022                         i = 1;
1023                 }
1024                 if (frontswap) {
1025                         if (frontswap_test(si, i))
1026                                 break;
1027                         else
1028                                 continue;
1029                 }
1030                 count = si->swap_map[i];
1031                 if (count && swap_count(count) != SWAP_MAP_BAD)
1032                         break;
1033         }
1034         return i;
1035 }
1036
1037 /*
1038  * We completely avoid races by reading each swap page in advance,
1039  * and then search for the process using it.  All the necessary
1040  * page table adjustments can then be made atomically.
1041  *
1042  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1043  * pages_to_unuse==0 means all pages; ignored if frontswap is false
1044  */
1045 int try_to_unuse(unsigned int type, bool frontswap,
1046                  unsigned long pages_to_unuse)
1047 {
1048         struct swap_info_struct *si = swap_info[type];
1049         struct mm_struct *start_mm;
1050         unsigned char *swap_map;
1051         unsigned char swcount;
1052         struct page *page;
1053         swp_entry_t entry;
1054         unsigned int i = 0;
1055         int retval = 0;
1056
1057         /*
1058          * When searching mms for an entry, a good strategy is to
1059          * start at the first mm we freed the previous entry from
1060          * (though actually we don't notice whether we or coincidence
1061          * freed the entry).  Initialize this start_mm with a hold.
1062          *
1063          * A simpler strategy would be to start at the last mm we
1064          * freed the previous entry from; but that would take less
1065          * advantage of mmlist ordering, which clusters forked mms
1066          * together, child after parent.  If we race with dup_mmap(), we
1067          * prefer to resolve parent before child, lest we miss entries
1068          * duplicated after we scanned child: using last mm would invert
1069          * that.
1070          */
1071         start_mm = &init_mm;
1072         atomic_inc(&init_mm.mm_users);
1073
1074         /*
1075          * Keep on scanning until all entries have gone.  Usually,
1076          * one pass through swap_map is enough, but not necessarily:
1077          * there are races when an instance of an entry might be missed.
1078          */
1079         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1080                 if (signal_pending(current)) {
1081                         retval = -EINTR;
1082                         break;
1083                 }
1084
1085                 /*
1086                  * Get a page for the entry, using the existing swap
1087                  * cache page if there is one.  Otherwise, get a clean
1088                  * page and read the swap into it.
1089                  */
1090                 swap_map = &si->swap_map[i];
1091                 entry = swp_entry(type, i);
1092                 page = read_swap_cache_async(entry,
1093                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1094                 if (!page) {
1095                         /*
1096                          * Either swap_duplicate() failed because entry
1097                          * has been freed independently, and will not be
1098                          * reused since sys_swapoff() already disabled
1099                          * allocation from here, or alloc_page() failed.
1100                          */
1101                         if (!*swap_map)
1102                                 continue;
1103                         retval = -ENOMEM;
1104                         break;
1105                 }
1106
1107                 /*
1108                  * Don't hold on to start_mm if it looks like exiting.
1109                  */
1110                 if (atomic_read(&start_mm->mm_users) == 1) {
1111                         mmput(start_mm);
1112                         start_mm = &init_mm;
1113                         atomic_inc(&init_mm.mm_users);
1114                 }
1115
1116                 /*
1117                  * Wait for and lock page.  When do_swap_page races with
1118                  * try_to_unuse, do_swap_page can handle the fault much
1119                  * faster than try_to_unuse can locate the entry.  This
1120                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1121                  * defer to do_swap_page in such a case - in some tests,
1122                  * do_swap_page and try_to_unuse repeatedly compete.
1123                  */
1124                 wait_on_page_locked(page);
1125                 wait_on_page_writeback(page);
1126                 lock_page(page);
1127                 wait_on_page_writeback(page);
1128
1129                 /*
1130                  * Remove all references to entry.
1131                  */
1132                 swcount = *swap_map;
1133                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1134                         retval = shmem_unuse(entry, page);
1135                         /* page has already been unlocked and released */
1136                         if (retval < 0)
1137                                 break;
1138                         continue;
1139                 }
1140                 if (swap_count(swcount) && start_mm != &init_mm)
1141                         retval = unuse_mm(start_mm, entry, page);
1142
1143                 if (swap_count(*swap_map)) {
1144                         int set_start_mm = (*swap_map >= swcount);
1145                         struct list_head *p = &start_mm->mmlist;
1146                         struct mm_struct *new_start_mm = start_mm;
1147                         struct mm_struct *prev_mm = start_mm;
1148                         struct mm_struct *mm;
1149
1150                         atomic_inc(&new_start_mm->mm_users);
1151                         atomic_inc(&prev_mm->mm_users);
1152                         spin_lock(&mmlist_lock);
1153                         while (swap_count(*swap_map) && !retval &&
1154                                         (p = p->next) != &start_mm->mmlist) {
1155                                 mm = list_entry(p, struct mm_struct, mmlist);
1156                                 if (!atomic_inc_not_zero(&mm->mm_users))
1157                                         continue;
1158                                 spin_unlock(&mmlist_lock);
1159                                 mmput(prev_mm);
1160                                 prev_mm = mm;
1161
1162                                 cond_resched();
1163
1164                                 swcount = *swap_map;
1165                                 if (!swap_count(swcount)) /* any usage ? */
1166                                         ;
1167                                 else if (mm == &init_mm)
1168                                         set_start_mm = 1;
1169                                 else
1170                                         retval = unuse_mm(mm, entry, page);
1171
1172                                 if (set_start_mm && *swap_map < swcount) {
1173                                         mmput(new_start_mm);
1174                                         atomic_inc(&mm->mm_users);
1175                                         new_start_mm = mm;
1176                                         set_start_mm = 0;
1177                                 }
1178                                 spin_lock(&mmlist_lock);
1179                         }
1180                         spin_unlock(&mmlist_lock);
1181                         mmput(prev_mm);
1182                         mmput(start_mm);
1183                         start_mm = new_start_mm;
1184                 }
1185                 if (retval) {
1186                         unlock_page(page);
1187                         page_cache_release(page);
1188                         break;
1189                 }
1190
1191                 /*
1192                  * If a reference remains (rare), we would like to leave
1193                  * the page in the swap cache; but try_to_unmap could
1194                  * then re-duplicate the entry once we drop page lock,
1195                  * so we might loop indefinitely; also, that page could
1196                  * not be swapped out to other storage meanwhile.  So:
1197                  * delete from cache even if there's another reference,
1198                  * after ensuring that the data has been saved to disk -
1199                  * since if the reference remains (rarer), it will be
1200                  * read from disk into another page.  Splitting into two
1201                  * pages would be incorrect if swap supported "shared
1202                  * private" pages, but they are handled by tmpfs files.
1203                  *
1204                  * Given how unuse_vma() targets one particular offset
1205                  * in an anon_vma, once the anon_vma has been determined,
1206                  * this splitting happens to be just what is needed to
1207                  * handle where KSM pages have been swapped out: re-reading
1208                  * is unnecessarily slow, but we can fix that later on.
1209                  */
1210                 if (swap_count(*swap_map) &&
1211                      PageDirty(page) && PageSwapCache(page)) {
1212                         struct writeback_control wbc = {
1213                                 .sync_mode = WB_SYNC_NONE,
1214                         };
1215
1216                         swap_writepage(page, &wbc);
1217                         lock_page(page);
1218                         wait_on_page_writeback(page);
1219                 }
1220
1221                 /*
1222                  * It is conceivable that a racing task removed this page from
1223                  * swap cache just before we acquired the page lock at the top,
1224                  * or while we dropped it in unuse_mm().  The page might even
1225                  * be back in swap cache on another swap area: that we must not
1226                  * delete, since it may not have been written out to swap yet.
1227                  */
1228                 if (PageSwapCache(page) &&
1229                     likely(page_private(page) == entry.val))
1230                         delete_from_swap_cache(page);
1231
1232                 /*
1233                  * So we could skip searching mms once swap count went
1234                  * to 1, we did not mark any present ptes as dirty: must
1235                  * mark page dirty so shrink_page_list will preserve it.
1236                  */
1237                 SetPageDirty(page);
1238                 unlock_page(page);
1239                 page_cache_release(page);
1240
1241                 /*
1242                  * Make sure that we aren't completely killing
1243                  * interactive performance.
1244                  */
1245                 cond_resched();
1246                 if (frontswap && pages_to_unuse > 0) {
1247                         if (!--pages_to_unuse)
1248                                 break;
1249                 }
1250         }
1251
1252         mmput(start_mm);
1253         return retval;
1254 }
1255
1256 /*
1257  * After a successful try_to_unuse, if no swap is now in use, we know
1258  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1259  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1260  * added to the mmlist just after page_duplicate - before would be racy.
1261  */
1262 static void drain_mmlist(void)
1263 {
1264         struct list_head *p, *next;
1265         unsigned int type;
1266
1267         for (type = 0; type < nr_swapfiles; type++)
1268                 if (swap_info[type]->inuse_pages)
1269                         return;
1270         spin_lock(&mmlist_lock);
1271         list_for_each_safe(p, next, &init_mm.mmlist)
1272                 list_del_init(p);
1273         spin_unlock(&mmlist_lock);
1274 }
1275
1276 /*
1277  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1278  * corresponds to page offset for the specified swap entry.
1279  * Note that the type of this function is sector_t, but it returns page offset
1280  * into the bdev, not sector offset.
1281  */
1282 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1283 {
1284         struct swap_info_struct *sis;
1285         struct swap_extent *start_se;
1286         struct swap_extent *se;
1287         pgoff_t offset;
1288
1289         sis = swap_info[swp_type(entry)];
1290         *bdev = sis->bdev;
1291
1292         offset = swp_offset(entry);
1293         start_se = sis->curr_swap_extent;
1294         se = start_se;
1295
1296         for ( ; ; ) {
1297                 struct list_head *lh;
1298
1299                 if (se->start_page <= offset &&
1300                                 offset < (se->start_page + se->nr_pages)) {
1301                         return se->start_block + (offset - se->start_page);
1302                 }
1303                 lh = se->list.next;
1304                 se = list_entry(lh, struct swap_extent, list);
1305                 sis->curr_swap_extent = se;
1306                 BUG_ON(se == start_se);         /* It *must* be present */
1307         }
1308 }
1309
1310 /*
1311  * Returns the page offset into bdev for the specified page's swap entry.
1312  */
1313 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1314 {
1315         swp_entry_t entry;
1316         entry.val = page_private(page);
1317         return map_swap_entry(entry, bdev);
1318 }
1319
1320 /*
1321  * Free all of a swapdev's extent information
1322  */
1323 static void destroy_swap_extents(struct swap_info_struct *sis)
1324 {
1325         while (!list_empty(&sis->first_swap_extent.list)) {
1326                 struct swap_extent *se;
1327
1328                 se = list_entry(sis->first_swap_extent.list.next,
1329                                 struct swap_extent, list);
1330                 list_del(&se->list);
1331                 kfree(se);
1332         }
1333
1334         if (sis->flags & SWP_FILE) {
1335                 struct file *swap_file = sis->swap_file;
1336                 struct address_space *mapping = swap_file->f_mapping;
1337
1338                 sis->flags &= ~SWP_FILE;
1339                 mapping->a_ops->swap_deactivate(swap_file);
1340         }
1341 }
1342
1343 /*
1344  * Add a block range (and the corresponding page range) into this swapdev's
1345  * extent list.  The extent list is kept sorted in page order.
1346  *
1347  * This function rather assumes that it is called in ascending page order.
1348  */
1349 int
1350 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1351                 unsigned long nr_pages, sector_t start_block)
1352 {
1353         struct swap_extent *se;
1354         struct swap_extent *new_se;
1355         struct list_head *lh;
1356
1357         if (start_page == 0) {
1358                 se = &sis->first_swap_extent;
1359                 sis->curr_swap_extent = se;
1360                 se->start_page = 0;
1361                 se->nr_pages = nr_pages;
1362                 se->start_block = start_block;
1363                 return 1;
1364         } else {
1365                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1366                 se = list_entry(lh, struct swap_extent, list);
1367                 BUG_ON(se->start_page + se->nr_pages != start_page);
1368                 if (se->start_block + se->nr_pages == start_block) {
1369                         /* Merge it */
1370                         se->nr_pages += nr_pages;
1371                         return 0;
1372                 }
1373         }
1374
1375         /*
1376          * No merge.  Insert a new extent, preserving ordering.
1377          */
1378         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1379         if (new_se == NULL)
1380                 return -ENOMEM;
1381         new_se->start_page = start_page;
1382         new_se->nr_pages = nr_pages;
1383         new_se->start_block = start_block;
1384
1385         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1386         return 1;
1387 }
1388
1389 /*
1390  * A `swap extent' is a simple thing which maps a contiguous range of pages
1391  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1392  * is built at swapon time and is then used at swap_writepage/swap_readpage
1393  * time for locating where on disk a page belongs.
1394  *
1395  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1396  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1397  * swap files identically.
1398  *
1399  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1400  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1401  * swapfiles are handled *identically* after swapon time.
1402  *
1403  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1404  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1405  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1406  * requirements, they are simply tossed out - we will never use those blocks
1407  * for swapping.
1408  *
1409  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1410  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1411  * which will scribble on the fs.
1412  *
1413  * The amount of disk space which a single swap extent represents varies.
1414  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1415  * extents in the list.  To avoid much list walking, we cache the previous
1416  * search location in `curr_swap_extent', and start new searches from there.
1417  * This is extremely effective.  The average number of iterations in
1418  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1419  */
1420 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1421 {
1422         struct file *swap_file = sis->swap_file;
1423         struct address_space *mapping = swap_file->f_mapping;
1424         struct inode *inode = mapping->host;
1425         int ret;
1426
1427         if (S_ISBLK(inode->i_mode)) {
1428                 ret = add_swap_extent(sis, 0, sis->max, 0);
1429                 *span = sis->pages;
1430                 return ret;
1431         }
1432
1433         if (mapping->a_ops->swap_activate) {
1434                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1435                 if (!ret) {
1436                         sis->flags |= SWP_FILE;
1437                         ret = add_swap_extent(sis, 0, sis->max, 0);
1438                         *span = sis->pages;
1439                 }
1440                 return ret;
1441         }
1442
1443         return generic_swapfile_activate(sis, swap_file, span);
1444 }
1445
1446 static void _enable_swap_info(struct swap_info_struct *p, int prio,
1447                                 unsigned char *swap_map,
1448                                 unsigned long *frontswap_map)
1449 {
1450         int i, prev;
1451
1452         if (prio >= 0)
1453                 p->prio = prio;
1454         else
1455                 p->prio = --least_priority;
1456         p->swap_map = swap_map;
1457         frontswap_map_set(p, frontswap_map);
1458         p->flags |= SWP_WRITEOK;
1459         nr_swap_pages += p->pages;
1460         total_swap_pages += p->pages;
1461
1462         /* insert swap space into swap_list: */
1463         prev = -1;
1464         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1465                 if (p->prio >= swap_info[i]->prio)
1466                         break;
1467                 prev = i;
1468         }
1469         p->next = i;
1470         if (prev < 0)
1471                 swap_list.head = swap_list.next = p->type;
1472         else
1473                 swap_info[prev]->next = p->type;
1474         frontswap_init(p->type);
1475 }
1476
1477 static void enable_swap_info(struct swap_info_struct *p, int prio,
1478                                 unsigned char *swap_map,
1479                                 unsigned long *frontswap_map)
1480 {
1481         spin_lock(&swap_lock);
1482         _enable_swap_info(p, prio, swap_map, frontswap_map);
1483         spin_unlock(&swap_lock);
1484 }
1485
1486 static void reinsert_swap_info(struct swap_info_struct *p)
1487 {
1488         spin_lock(&swap_lock);
1489         _enable_swap_info(p, p->prio, p->swap_map, frontswap_map_get(p));
1490         spin_unlock(&swap_lock);
1491 }
1492
1493 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1494 {
1495         struct swap_info_struct *p = NULL;
1496         unsigned char *swap_map;
1497         struct file *swap_file, *victim;
1498         struct address_space *mapping;
1499         struct inode *inode;
1500         struct filename *pathname;
1501         int oom_score_adj;
1502         int i, type, prev;
1503         int err;
1504
1505         if (!capable(CAP_SYS_ADMIN))
1506                 return -EPERM;
1507
1508         BUG_ON(!current->mm);
1509
1510         pathname = getname(specialfile);
1511         if (IS_ERR(pathname))
1512                 return PTR_ERR(pathname);
1513
1514         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
1515         err = PTR_ERR(victim);
1516         if (IS_ERR(victim))
1517                 goto out;
1518
1519         mapping = victim->f_mapping;
1520         prev = -1;
1521         spin_lock(&swap_lock);
1522         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1523                 p = swap_info[type];
1524                 if (p->flags & SWP_WRITEOK) {
1525                         if (p->swap_file->f_mapping == mapping)
1526                                 break;
1527                 }
1528                 prev = type;
1529         }
1530         if (type < 0) {
1531                 err = -EINVAL;
1532                 spin_unlock(&swap_lock);
1533                 goto out_dput;
1534         }
1535         if (!security_vm_enough_memory_mm(current->mm, p->pages))
1536                 vm_unacct_memory(p->pages);
1537         else {
1538                 err = -ENOMEM;
1539                 spin_unlock(&swap_lock);
1540                 goto out_dput;
1541         }
1542         if (prev < 0)
1543                 swap_list.head = p->next;
1544         else
1545                 swap_info[prev]->next = p->next;
1546         if (type == swap_list.next) {
1547                 /* just pick something that's safe... */
1548                 swap_list.next = swap_list.head;
1549         }
1550         if (p->prio < 0) {
1551                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1552                         swap_info[i]->prio = p->prio--;
1553                 least_priority++;
1554         }
1555         nr_swap_pages -= p->pages;
1556         total_swap_pages -= p->pages;
1557         p->flags &= ~SWP_WRITEOK;
1558         spin_unlock(&swap_lock);
1559
1560         oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1561         err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1562         compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj);
1563
1564         if (err) {
1565                 /* re-insert swap space back into swap_list */
1566                 reinsert_swap_info(p);
1567                 goto out_dput;
1568         }
1569
1570         destroy_swap_extents(p);
1571         if (p->flags & SWP_CONTINUED)
1572                 free_swap_count_continuations(p);
1573
1574         mutex_lock(&swapon_mutex);
1575         spin_lock(&swap_lock);
1576         drain_mmlist();
1577
1578         /* wait for anyone still in scan_swap_map */
1579         p->highest_bit = 0;             /* cuts scans short */
1580         while (p->flags >= SWP_SCANNING) {
1581                 spin_unlock(&swap_lock);
1582                 schedule_timeout_uninterruptible(1);
1583                 spin_lock(&swap_lock);
1584         }
1585
1586         swap_file = p->swap_file;
1587         p->swap_file = NULL;
1588         p->max = 0;
1589         swap_map = p->swap_map;
1590         p->swap_map = NULL;
1591         p->flags = 0;
1592         frontswap_invalidate_area(type);
1593         spin_unlock(&swap_lock);
1594         mutex_unlock(&swapon_mutex);
1595         vfree(swap_map);
1596         vfree(frontswap_map_get(p));
1597         /* Destroy swap account informatin */
1598         swap_cgroup_swapoff(type);
1599
1600         inode = mapping->host;
1601         if (S_ISBLK(inode->i_mode)) {
1602                 struct block_device *bdev = I_BDEV(inode);
1603                 set_blocksize(bdev, p->old_block_size);
1604                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1605         } else {
1606                 mutex_lock(&inode->i_mutex);
1607                 inode->i_flags &= ~S_SWAPFILE;
1608                 mutex_unlock(&inode->i_mutex);
1609         }
1610         filp_close(swap_file, NULL);
1611         err = 0;
1612         atomic_inc(&proc_poll_event);
1613         wake_up_interruptible(&proc_poll_wait);
1614
1615 out_dput:
1616         filp_close(victim, NULL);
1617 out:
1618         putname(pathname);
1619         return err;
1620 }
1621
1622 #ifdef CONFIG_PROC_FS
1623 static unsigned swaps_poll(struct file *file, poll_table *wait)
1624 {
1625         struct seq_file *seq = file->private_data;
1626
1627         poll_wait(file, &proc_poll_wait, wait);
1628
1629         if (seq->poll_event != atomic_read(&proc_poll_event)) {
1630                 seq->poll_event = atomic_read(&proc_poll_event);
1631                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1632         }
1633
1634         return POLLIN | POLLRDNORM;
1635 }
1636
1637 /* iterator */
1638 static void *swap_start(struct seq_file *swap, loff_t *pos)
1639 {
1640         struct swap_info_struct *si;
1641         int type;
1642         loff_t l = *pos;
1643
1644         mutex_lock(&swapon_mutex);
1645
1646         if (!l)
1647                 return SEQ_START_TOKEN;
1648
1649         for (type = 0; type < nr_swapfiles; type++) {
1650                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1651                 si = swap_info[type];
1652                 if (!(si->flags & SWP_USED) || !si->swap_map)
1653                         continue;
1654                 if (!--l)
1655                         return si;
1656         }
1657
1658         return NULL;
1659 }
1660
1661 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1662 {
1663         struct swap_info_struct *si = v;
1664         int type;
1665
1666         if (v == SEQ_START_TOKEN)
1667                 type = 0;
1668         else
1669                 type = si->type + 1;
1670
1671         for (; type < nr_swapfiles; type++) {
1672                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1673                 si = swap_info[type];
1674                 if (!(si->flags & SWP_USED) || !si->swap_map)
1675                         continue;
1676                 ++*pos;
1677                 return si;
1678         }
1679
1680         return NULL;
1681 }
1682
1683 static void swap_stop(struct seq_file *swap, void *v)
1684 {
1685         mutex_unlock(&swapon_mutex);
1686 }
1687
1688 static int swap_show(struct seq_file *swap, void *v)
1689 {
1690         struct swap_info_struct *si = v;
1691         struct file *file;
1692         int len;
1693
1694         if (si == SEQ_START_TOKEN) {
1695                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1696                 return 0;
1697         }
1698
1699         file = si->swap_file;
1700         len = seq_path(swap, &file->f_path, " \t\n\\");
1701         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1702                         len < 40 ? 40 - len : 1, " ",
1703                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1704                                 "partition" : "file\t",
1705                         si->pages << (PAGE_SHIFT - 10),
1706                         si->inuse_pages << (PAGE_SHIFT - 10),
1707                         si->prio);
1708         return 0;
1709 }
1710
1711 static const struct seq_operations swaps_op = {
1712         .start =        swap_start,
1713         .next =         swap_next,
1714         .stop =         swap_stop,
1715         .show =         swap_show
1716 };
1717
1718 static int swaps_open(struct inode *inode, struct file *file)
1719 {
1720         struct seq_file *seq;
1721         int ret;
1722
1723         ret = seq_open(file, &swaps_op);
1724         if (ret)
1725                 return ret;
1726
1727         seq = file->private_data;
1728         seq->poll_event = atomic_read(&proc_poll_event);
1729         return 0;
1730 }
1731
1732 static const struct file_operations proc_swaps_operations = {
1733         .open           = swaps_open,
1734         .read           = seq_read,
1735         .llseek         = seq_lseek,
1736         .release        = seq_release,
1737         .poll           = swaps_poll,
1738 };
1739
1740 static int __init procswaps_init(void)
1741 {
1742         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1743         return 0;
1744 }
1745 __initcall(procswaps_init);
1746 #endif /* CONFIG_PROC_FS */
1747
1748 #ifdef MAX_SWAPFILES_CHECK
1749 static int __init max_swapfiles_check(void)
1750 {
1751         MAX_SWAPFILES_CHECK();
1752         return 0;
1753 }
1754 late_initcall(max_swapfiles_check);
1755 #endif
1756
1757 static struct swap_info_struct *alloc_swap_info(void)
1758 {
1759         struct swap_info_struct *p;
1760         unsigned int type;
1761
1762         p = kzalloc(sizeof(*p), GFP_KERNEL);
1763         if (!p)
1764                 return ERR_PTR(-ENOMEM);
1765
1766         spin_lock(&swap_lock);
1767         for (type = 0; type < nr_swapfiles; type++) {
1768                 if (!(swap_info[type]->flags & SWP_USED))
1769                         break;
1770         }
1771         if (type >= MAX_SWAPFILES) {
1772                 spin_unlock(&swap_lock);
1773                 kfree(p);
1774                 return ERR_PTR(-EPERM);
1775         }
1776         if (type >= nr_swapfiles) {
1777                 p->type = type;
1778                 swap_info[type] = p;
1779                 /*
1780                  * Write swap_info[type] before nr_swapfiles, in case a
1781                  * racing procfs swap_start() or swap_next() is reading them.
1782                  * (We never shrink nr_swapfiles, we never free this entry.)
1783                  */
1784                 smp_wmb();
1785                 nr_swapfiles++;
1786         } else {
1787                 kfree(p);
1788                 p = swap_info[type];
1789                 /*
1790                  * Do not memset this entry: a racing procfs swap_next()
1791                  * would be relying on p->type to remain valid.
1792                  */
1793         }
1794         INIT_LIST_HEAD(&p->first_swap_extent.list);
1795         p->flags = SWP_USED;
1796         p->next = -1;
1797         spin_unlock(&swap_lock);
1798
1799         return p;
1800 }
1801
1802 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1803 {
1804         int error;
1805
1806         if (S_ISBLK(inode->i_mode)) {
1807                 p->bdev = bdgrab(I_BDEV(inode));
1808                 error = blkdev_get(p->bdev,
1809                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1810                                    sys_swapon);
1811                 if (error < 0) {
1812                         p->bdev = NULL;
1813                         return -EINVAL;
1814                 }
1815                 p->old_block_size = block_size(p->bdev);
1816                 error = set_blocksize(p->bdev, PAGE_SIZE);
1817                 if (error < 0)
1818                         return error;
1819                 p->flags |= SWP_BLKDEV;
1820         } else if (S_ISREG(inode->i_mode)) {
1821                 p->bdev = inode->i_sb->s_bdev;
1822                 mutex_lock(&inode->i_mutex);
1823                 if (IS_SWAPFILE(inode))
1824                         return -EBUSY;
1825         } else
1826                 return -EINVAL;
1827
1828         return 0;
1829 }
1830
1831 static unsigned long read_swap_header(struct swap_info_struct *p,
1832                                         union swap_header *swap_header,
1833                                         struct inode *inode)
1834 {
1835         int i;
1836         unsigned long maxpages;
1837         unsigned long swapfilepages;
1838
1839         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1840                 printk(KERN_ERR "Unable to find swap-space signature\n");
1841                 return 0;
1842         }
1843
1844         /* swap partition endianess hack... */
1845         if (swab32(swap_header->info.version) == 1) {
1846                 swab32s(&swap_header->info.version);
1847                 swab32s(&swap_header->info.last_page);
1848                 swab32s(&swap_header->info.nr_badpages);
1849                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1850                         swab32s(&swap_header->info.badpages[i]);
1851         }
1852         /* Check the swap header's sub-version */
1853         if (swap_header->info.version != 1) {
1854                 printk(KERN_WARNING
1855                        "Unable to handle swap header version %d\n",
1856                        swap_header->info.version);
1857                 return 0;
1858         }
1859
1860         p->lowest_bit  = 1;
1861         p->cluster_next = 1;
1862         p->cluster_nr = 0;
1863
1864         /*
1865          * Find out how many pages are allowed for a single swap
1866          * device. There are two limiting factors: 1) the number
1867          * of bits for the swap offset in the swp_entry_t type, and
1868          * 2) the number of bits in the swap pte as defined by the
1869          * different architectures. In order to find the
1870          * largest possible bit mask, a swap entry with swap type 0
1871          * and swap offset ~0UL is created, encoded to a swap pte,
1872          * decoded to a swp_entry_t again, and finally the swap
1873          * offset is extracted. This will mask all the bits from
1874          * the initial ~0UL mask that can't be encoded in either
1875          * the swp_entry_t or the architecture definition of a
1876          * swap pte.
1877          */
1878         maxpages = swp_offset(pte_to_swp_entry(
1879                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1880         if (maxpages > swap_header->info.last_page) {
1881                 maxpages = swap_header->info.last_page + 1;
1882                 /* p->max is an unsigned int: don't overflow it */
1883                 if ((unsigned int)maxpages == 0)
1884                         maxpages = UINT_MAX;
1885         }
1886         p->highest_bit = maxpages - 1;
1887
1888         if (!maxpages)
1889                 return 0;
1890         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1891         if (swapfilepages && maxpages > swapfilepages) {
1892                 printk(KERN_WARNING
1893                        "Swap area shorter than signature indicates\n");
1894                 return 0;
1895         }
1896         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1897                 return 0;
1898         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1899                 return 0;
1900
1901         return maxpages;
1902 }
1903
1904 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1905                                         union swap_header *swap_header,
1906                                         unsigned char *swap_map,
1907                                         unsigned long maxpages,
1908                                         sector_t *span)
1909 {
1910         int i;
1911         unsigned int nr_good_pages;
1912         int nr_extents;
1913
1914         nr_good_pages = maxpages - 1;   /* omit header page */
1915
1916         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1917                 unsigned int page_nr = swap_header->info.badpages[i];
1918                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1919                         return -EINVAL;
1920                 if (page_nr < maxpages) {
1921                         swap_map[page_nr] = SWAP_MAP_BAD;
1922                         nr_good_pages--;
1923                 }
1924         }
1925
1926         if (nr_good_pages) {
1927                 swap_map[0] = SWAP_MAP_BAD;
1928                 p->max = maxpages;
1929                 p->pages = nr_good_pages;
1930                 nr_extents = setup_swap_extents(p, span);
1931                 if (nr_extents < 0)
1932                         return nr_extents;
1933                 nr_good_pages = p->pages;
1934         }
1935         if (!nr_good_pages) {
1936                 printk(KERN_WARNING "Empty swap-file\n");
1937                 return -EINVAL;
1938         }
1939
1940         return nr_extents;
1941 }
1942
1943 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1944 {
1945         struct swap_info_struct *p;
1946         struct filename *name;
1947         struct file *swap_file = NULL;
1948         struct address_space *mapping;
1949         int i;
1950         int prio;
1951         int error;
1952         union swap_header *swap_header;
1953         int nr_extents;
1954         sector_t span;
1955         unsigned long maxpages;
1956         unsigned char *swap_map = NULL;
1957         unsigned long *frontswap_map = NULL;
1958         struct page *page = NULL;
1959         struct inode *inode = NULL;
1960
1961         if (swap_flags & ~SWAP_FLAGS_VALID)
1962                 return -EINVAL;
1963
1964         if (!capable(CAP_SYS_ADMIN))
1965                 return -EPERM;
1966
1967         p = alloc_swap_info();
1968         if (IS_ERR(p))
1969                 return PTR_ERR(p);
1970
1971         name = getname(specialfile);
1972         if (IS_ERR(name)) {
1973                 error = PTR_ERR(name);
1974                 name = NULL;
1975                 goto bad_swap;
1976         }
1977         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
1978         if (IS_ERR(swap_file)) {
1979                 error = PTR_ERR(swap_file);
1980                 swap_file = NULL;
1981                 goto bad_swap;
1982         }
1983
1984         p->swap_file = swap_file;
1985         mapping = swap_file->f_mapping;
1986
1987         for (i = 0; i < nr_swapfiles; i++) {
1988                 struct swap_info_struct *q = swap_info[i];
1989
1990                 if (q == p || !q->swap_file)
1991                         continue;
1992                 if (mapping == q->swap_file->f_mapping) {
1993                         error = -EBUSY;
1994                         goto bad_swap;
1995                 }
1996         }
1997
1998         inode = mapping->host;
1999         /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2000         error = claim_swapfile(p, inode);
2001         if (unlikely(error))
2002                 goto bad_swap;
2003
2004         /*
2005          * Read the swap header.
2006          */
2007         if (!mapping->a_ops->readpage) {
2008                 error = -EINVAL;
2009                 goto bad_swap;
2010         }
2011         page = read_mapping_page(mapping, 0, swap_file);
2012         if (IS_ERR(page)) {
2013                 error = PTR_ERR(page);
2014                 goto bad_swap;
2015         }
2016         swap_header = kmap(page);
2017
2018         maxpages = read_swap_header(p, swap_header, inode);
2019         if (unlikely(!maxpages)) {
2020                 error = -EINVAL;
2021                 goto bad_swap;
2022         }
2023
2024         /* OK, set up the swap map and apply the bad block list */
2025         swap_map = vzalloc(maxpages);
2026         if (!swap_map) {
2027                 error = -ENOMEM;
2028                 goto bad_swap;
2029         }
2030
2031         error = swap_cgroup_swapon(p->type, maxpages);
2032         if (error)
2033                 goto bad_swap;
2034
2035         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2036                 maxpages, &span);
2037         if (unlikely(nr_extents < 0)) {
2038                 error = nr_extents;
2039                 goto bad_swap;
2040         }
2041         /* frontswap enabled? set up bit-per-page map for frontswap */
2042         if (frontswap_enabled)
2043                 frontswap_map = vzalloc(maxpages / sizeof(long));
2044
2045         if (p->bdev) {
2046                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2047                         p->flags |= SWP_SOLIDSTATE;
2048                         p->cluster_next = 1 + (random32() % p->highest_bit);
2049                 }
2050                 if ((swap_flags & SWAP_FLAG_DISCARD) && discard_swap(p) == 0)
2051                         p->flags |= SWP_DISCARDABLE;
2052         }
2053
2054         mutex_lock(&swapon_mutex);
2055         prio = -1;
2056         if (swap_flags & SWAP_FLAG_PREFER)
2057                 prio =
2058                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2059         enable_swap_info(p, prio, swap_map, frontswap_map);
2060
2061         printk(KERN_INFO "Adding %uk swap on %s.  "
2062                         "Priority:%d extents:%d across:%lluk %s%s%s\n",
2063                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2064                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2065                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2066                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2067                 (frontswap_map) ? "FS" : "");
2068
2069         mutex_unlock(&swapon_mutex);
2070         atomic_inc(&proc_poll_event);
2071         wake_up_interruptible(&proc_poll_wait);
2072
2073         if (S_ISREG(inode->i_mode))
2074                 inode->i_flags |= S_SWAPFILE;
2075         error = 0;
2076         goto out;
2077 bad_swap:
2078         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2079                 set_blocksize(p->bdev, p->old_block_size);
2080                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2081         }
2082         destroy_swap_extents(p);
2083         swap_cgroup_swapoff(p->type);
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                 if (inode && S_ISREG(inode->i_mode)) {
2091                         mutex_unlock(&inode->i_mutex);
2092                         inode = NULL;
2093                 }
2094                 filp_close(swap_file, NULL);
2095         }
2096 out:
2097         if (page && !IS_ERR(page)) {
2098                 kunmap(page);
2099                 page_cache_release(page);
2100         }
2101         if (name)
2102                 putname(name);
2103         if (inode && S_ISREG(inode->i_mode))
2104                 mutex_unlock(&inode->i_mutex);
2105         return error;
2106 }
2107
2108 void si_swapinfo(struct sysinfo *val)
2109 {
2110         unsigned int type;
2111         unsigned long nr_to_be_unused = 0;
2112
2113         spin_lock(&swap_lock);
2114         for (type = 0; type < nr_swapfiles; type++) {
2115                 struct swap_info_struct *si = swap_info[type];
2116
2117                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2118                         nr_to_be_unused += si->inuse_pages;
2119         }
2120         val->freeswap = nr_swap_pages + nr_to_be_unused;
2121         val->totalswap = total_swap_pages + nr_to_be_unused;
2122         spin_unlock(&swap_lock);
2123 }
2124
2125 /*
2126  * Verify that a swap entry is valid and increment its swap map count.
2127  *
2128  * Returns error code in following case.
2129  * - success -> 0
2130  * - swp_entry is invalid -> EINVAL
2131  * - swp_entry is migration entry -> EINVAL
2132  * - swap-cache reference is requested but there is already one. -> EEXIST
2133  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2134  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2135  */
2136 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2137 {
2138         struct swap_info_struct *p;
2139         unsigned long offset, type;
2140         unsigned char count;
2141         unsigned char has_cache;
2142         int err = -EINVAL;
2143
2144         if (non_swap_entry(entry))
2145                 goto out;
2146
2147         type = swp_type(entry);
2148         if (type >= nr_swapfiles)
2149                 goto bad_file;
2150         p = swap_info[type];
2151         offset = swp_offset(entry);
2152
2153         spin_lock(&swap_lock);
2154         if (unlikely(offset >= p->max))
2155                 goto unlock_out;
2156
2157         count = p->swap_map[offset];
2158         has_cache = count & SWAP_HAS_CACHE;
2159         count &= ~SWAP_HAS_CACHE;
2160         err = 0;
2161
2162         if (usage == SWAP_HAS_CACHE) {
2163
2164                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2165                 if (!has_cache && count)
2166                         has_cache = SWAP_HAS_CACHE;
2167                 else if (has_cache)             /* someone else added cache */
2168                         err = -EEXIST;
2169                 else                            /* no users remaining */
2170                         err = -ENOENT;
2171
2172         } else if (count || has_cache) {
2173
2174                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2175                         count += usage;
2176                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2177                         err = -EINVAL;
2178                 else if (swap_count_continued(p, offset, count))
2179                         count = COUNT_CONTINUED;
2180                 else
2181                         err = -ENOMEM;
2182         } else
2183                 err = -ENOENT;                  /* unused swap entry */
2184
2185         p->swap_map[offset] = count | has_cache;
2186
2187 unlock_out:
2188         spin_unlock(&swap_lock);
2189 out:
2190         return err;
2191
2192 bad_file:
2193         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2194         goto out;
2195 }
2196
2197 /*
2198  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2199  * (in which case its reference count is never incremented).
2200  */
2201 void swap_shmem_alloc(swp_entry_t entry)
2202 {
2203         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2204 }
2205
2206 /*
2207  * Increase reference count of swap entry by 1.
2208  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2209  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2210  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2211  * might occur if a page table entry has got corrupted.
2212  */
2213 int swap_duplicate(swp_entry_t entry)
2214 {
2215         int err = 0;
2216
2217         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2218                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2219         return err;
2220 }
2221
2222 /*
2223  * @entry: swap entry for which we allocate swap cache.
2224  *
2225  * Called when allocating swap cache for existing swap entry,
2226  * This can return error codes. Returns 0 at success.
2227  * -EBUSY means there is a swap cache.
2228  * Note: return code is different from swap_duplicate().
2229  */
2230 int swapcache_prepare(swp_entry_t entry)
2231 {
2232         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2233 }
2234
2235 struct swap_info_struct *page_swap_info(struct page *page)
2236 {
2237         swp_entry_t swap = { .val = page_private(page) };
2238         BUG_ON(!PageSwapCache(page));
2239         return swap_info[swp_type(swap)];
2240 }
2241
2242 /*
2243  * out-of-line __page_file_ methods to avoid include hell.
2244  */
2245 struct address_space *__page_file_mapping(struct page *page)
2246 {
2247         VM_BUG_ON(!PageSwapCache(page));
2248         return page_swap_info(page)->swap_file->f_mapping;
2249 }
2250 EXPORT_SYMBOL_GPL(__page_file_mapping);
2251
2252 pgoff_t __page_file_index(struct page *page)
2253 {
2254         swp_entry_t swap = { .val = page_private(page) };
2255         VM_BUG_ON(!PageSwapCache(page));
2256         return swp_offset(swap);
2257 }
2258 EXPORT_SYMBOL_GPL(__page_file_index);
2259
2260 /*
2261  * add_swap_count_continuation - called when a swap count is duplicated
2262  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2263  * page of the original vmalloc'ed swap_map, to hold the continuation count
2264  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2265  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2266  *
2267  * These continuation pages are seldom referenced: the common paths all work
2268  * on the original swap_map, only referring to a continuation page when the
2269  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2270  *
2271  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2272  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2273  * can be called after dropping locks.
2274  */
2275 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2276 {
2277         struct swap_info_struct *si;
2278         struct page *head;
2279         struct page *page;
2280         struct page *list_page;
2281         pgoff_t offset;
2282         unsigned char count;
2283
2284         /*
2285          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2286          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2287          */
2288         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2289
2290         si = swap_info_get(entry);
2291         if (!si) {
2292                 /*
2293                  * An acceptable race has occurred since the failing
2294                  * __swap_duplicate(): the swap entry has been freed,
2295                  * perhaps even the whole swap_map cleared for swapoff.
2296                  */
2297                 goto outer;
2298         }
2299
2300         offset = swp_offset(entry);
2301         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2302
2303         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2304                 /*
2305                  * The higher the swap count, the more likely it is that tasks
2306                  * will race to add swap count continuation: we need to avoid
2307                  * over-provisioning.
2308                  */
2309                 goto out;
2310         }
2311
2312         if (!page) {
2313                 spin_unlock(&swap_lock);
2314                 return -ENOMEM;
2315         }
2316
2317         /*
2318          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2319          * no architecture is using highmem pages for kernel pagetables: so it
2320          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2321          */
2322         head = vmalloc_to_page(si->swap_map + offset);
2323         offset &= ~PAGE_MASK;
2324
2325         /*
2326          * Page allocation does not initialize the page's lru field,
2327          * but it does always reset its private field.
2328          */
2329         if (!page_private(head)) {
2330                 BUG_ON(count & COUNT_CONTINUED);
2331                 INIT_LIST_HEAD(&head->lru);
2332                 set_page_private(head, SWP_CONTINUED);
2333                 si->flags |= SWP_CONTINUED;
2334         }
2335
2336         list_for_each_entry(list_page, &head->lru, lru) {
2337                 unsigned char *map;
2338
2339                 /*
2340                  * If the previous map said no continuation, but we've found
2341                  * a continuation page, free our allocation and use this one.
2342                  */
2343                 if (!(count & COUNT_CONTINUED))
2344                         goto out;
2345
2346                 map = kmap_atomic(list_page) + offset;
2347                 count = *map;
2348                 kunmap_atomic(map);
2349
2350                 /*
2351                  * If this continuation count now has some space in it,
2352                  * free our allocation and use this one.
2353                  */
2354                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2355                         goto out;
2356         }
2357
2358         list_add_tail(&page->lru, &head->lru);
2359         page = NULL;                    /* now it's attached, don't free it */
2360 out:
2361         spin_unlock(&swap_lock);
2362 outer:
2363         if (page)
2364                 __free_page(page);
2365         return 0;
2366 }
2367
2368 /*
2369  * swap_count_continued - when the original swap_map count is incremented
2370  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2371  * into, carry if so, or else fail until a new continuation page is allocated;
2372  * when the original swap_map count is decremented from 0 with continuation,
2373  * borrow from the continuation and report whether it still holds more.
2374  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2375  */
2376 static bool swap_count_continued(struct swap_info_struct *si,
2377                                  pgoff_t offset, unsigned char count)
2378 {
2379         struct page *head;
2380         struct page *page;
2381         unsigned char *map;
2382
2383         head = vmalloc_to_page(si->swap_map + offset);
2384         if (page_private(head) != SWP_CONTINUED) {
2385                 BUG_ON(count & COUNT_CONTINUED);
2386                 return false;           /* need to add count continuation */
2387         }
2388
2389         offset &= ~PAGE_MASK;
2390         page = list_entry(head->lru.next, struct page, lru);
2391         map = kmap_atomic(page) + offset;
2392
2393         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2394                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2395
2396         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2397                 /*
2398                  * Think of how you add 1 to 999
2399                  */
2400                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2401                         kunmap_atomic(map);
2402                         page = list_entry(page->lru.next, struct page, lru);
2403                         BUG_ON(page == head);
2404                         map = kmap_atomic(page) + offset;
2405                 }
2406                 if (*map == SWAP_CONT_MAX) {
2407                         kunmap_atomic(map);
2408                         page = list_entry(page->lru.next, struct page, lru);
2409                         if (page == head)
2410                                 return false;   /* add count continuation */
2411                         map = kmap_atomic(page) + offset;
2412 init_map:               *map = 0;               /* we didn't zero the page */
2413                 }
2414                 *map += 1;
2415                 kunmap_atomic(map);
2416                 page = list_entry(page->lru.prev, struct page, lru);
2417                 while (page != head) {
2418                         map = kmap_atomic(page) + offset;
2419                         *map = COUNT_CONTINUED;
2420                         kunmap_atomic(map);
2421                         page = list_entry(page->lru.prev, struct page, lru);
2422                 }
2423                 return true;                    /* incremented */
2424
2425         } else {                                /* decrementing */
2426                 /*
2427                  * Think of how you subtract 1 from 1000
2428                  */
2429                 BUG_ON(count != COUNT_CONTINUED);
2430                 while (*map == COUNT_CONTINUED) {
2431                         kunmap_atomic(map);
2432                         page = list_entry(page->lru.next, struct page, lru);
2433                         BUG_ON(page == head);
2434                         map = kmap_atomic(page) + offset;
2435                 }
2436                 BUG_ON(*map == 0);
2437                 *map -= 1;
2438                 if (*map == 0)
2439                         count = 0;
2440                 kunmap_atomic(map);
2441                 page = list_entry(page->lru.prev, struct page, lru);
2442                 while (page != head) {
2443                         map = kmap_atomic(page) + offset;
2444                         *map = SWAP_CONT_MAX | count;
2445                         count = COUNT_CONTINUED;
2446                         kunmap_atomic(map);
2447                         page = list_entry(page->lru.prev, struct page, lru);
2448                 }
2449                 return count == COUNT_CONTINUED;
2450         }
2451 }
2452
2453 /*
2454  * free_swap_count_continuations - swapoff free all the continuation pages
2455  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2456  */
2457 static void free_swap_count_continuations(struct swap_info_struct *si)
2458 {
2459         pgoff_t offset;
2460
2461         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2462                 struct page *head;
2463                 head = vmalloc_to_page(si->swap_map + offset);
2464                 if (page_private(head)) {
2465                         struct list_head *this, *next;
2466                         list_for_each_safe(this, next, &head->lru) {
2467                                 struct page *page;
2468                                 page = list_entry(this, struct page, lru);
2469                                 list_del(this);
2470                                 __free_page(page);
2471                         }
2472                 }
2473         }
2474 }