03aa2d55f1a2f54fc5eb17b099806edbc83a4049
[sfrench/cifs-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/security.h>
28 #include <linux/backing-dev.h>
29 #include <linux/mutex.h>
30 #include <linux/capability.h>
31 #include <linux/syscalls.h>
32 #include <linux/memcontrol.h>
33
34 #include <asm/pgtable.h>
35 #include <asm/tlbflush.h>
36 #include <linux/swapops.h>
37 #include <linux/page_cgroup.h>
38
39 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
40                                  unsigned char);
41 static void free_swap_count_continuations(struct swap_info_struct *);
42 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
43
44 static DEFINE_SPINLOCK(swap_lock);
45 static unsigned int nr_swapfiles;
46 long nr_swap_pages;
47 long total_swap_pages;
48 static int least_priority;
49
50 static const char Bad_file[] = "Bad swap file entry ";
51 static const char Unused_file[] = "Unused swap file entry ";
52 static const char Bad_offset[] = "Bad swap offset entry ";
53 static const char Unused_offset[] = "Unused swap offset entry ";
54
55 static struct swap_list_t swap_list = {-1, -1};
56
57 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
58
59 static DEFINE_MUTEX(swapon_mutex);
60
61 static inline unsigned char swap_count(unsigned char ent)
62 {
63         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
64 }
65
66 /* returns 1 if swap entry is freed */
67 static int
68 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
69 {
70         swp_entry_t entry = swp_entry(si->type, offset);
71         struct page *page;
72         int ret = 0;
73
74         page = find_get_page(&swapper_space, entry.val);
75         if (!page)
76                 return 0;
77         /*
78          * This function is called from scan_swap_map() and it's called
79          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
80          * We have to use trylock for avoiding deadlock. This is a special
81          * case and you should use try_to_free_swap() with explicit lock_page()
82          * in usual operations.
83          */
84         if (trylock_page(page)) {
85                 ret = try_to_free_swap(page);
86                 unlock_page(page);
87         }
88         page_cache_release(page);
89         return ret;
90 }
91
92 /*
93  * We need this because the bdev->unplug_fn can sleep and we cannot
94  * hold swap_lock while calling the unplug_fn. And swap_lock
95  * cannot be turned into a mutex.
96  */
97 static DECLARE_RWSEM(swap_unplug_sem);
98
99 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
100 {
101         swp_entry_t entry;
102
103         down_read(&swap_unplug_sem);
104         entry.val = page_private(page);
105         if (PageSwapCache(page)) {
106                 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
107                 struct backing_dev_info *bdi;
108
109                 /*
110                  * If the page is removed from swapcache from under us (with a
111                  * racy try_to_unuse/swapoff) we need an additional reference
112                  * count to avoid reading garbage from page_private(page) above.
113                  * If the WARN_ON triggers during a swapoff it maybe the race
114                  * condition and it's harmless. However if it triggers without
115                  * swapoff it signals a problem.
116                  */
117                 WARN_ON(page_count(page) <= 1);
118
119                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
120                 blk_run_backing_dev(bdi, page);
121         }
122         up_read(&swap_unplug_sem);
123 }
124
125 /*
126  * swapon tell device that all the old swap contents can be discarded,
127  * to allow the swap device to optimize its wear-levelling.
128  */
129 static int discard_swap(struct swap_info_struct *si)
130 {
131         struct swap_extent *se;
132         sector_t start_block;
133         sector_t nr_blocks;
134         int err = 0;
135
136         /* Do not discard the swap header page! */
137         se = &si->first_swap_extent;
138         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
139         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
140         if (nr_blocks) {
141                 err = blkdev_issue_discard(si->bdev, start_block,
142                                 nr_blocks, GFP_KERNEL,
143                                 BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
144                 if (err)
145                         return err;
146                 cond_resched();
147         }
148
149         list_for_each_entry(se, &si->first_swap_extent.list, list) {
150                 start_block = se->start_block << (PAGE_SHIFT - 9);
151                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
152
153                 err = blkdev_issue_discard(si->bdev, start_block,
154                                 nr_blocks, GFP_KERNEL,
155                                 BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
156                 if (err)
157                         break;
158
159                 cond_resched();
160         }
161         return err;             /* That will often be -EOPNOTSUPP */
162 }
163
164 /*
165  * swap allocation tell device that a cluster of swap can now be discarded,
166  * to allow the swap device to optimize its wear-levelling.
167  */
168 static void discard_swap_cluster(struct swap_info_struct *si,
169                                  pgoff_t start_page, pgoff_t nr_pages)
170 {
171         struct swap_extent *se = si->curr_swap_extent;
172         int found_extent = 0;
173
174         while (nr_pages) {
175                 struct list_head *lh;
176
177                 if (se->start_page <= start_page &&
178                     start_page < se->start_page + se->nr_pages) {
179                         pgoff_t offset = start_page - se->start_page;
180                         sector_t start_block = se->start_block + offset;
181                         sector_t nr_blocks = se->nr_pages - offset;
182
183                         if (nr_blocks > nr_pages)
184                                 nr_blocks = nr_pages;
185                         start_page += nr_blocks;
186                         nr_pages -= nr_blocks;
187
188                         if (!found_extent++)
189                                 si->curr_swap_extent = se;
190
191                         start_block <<= PAGE_SHIFT - 9;
192                         nr_blocks <<= PAGE_SHIFT - 9;
193                         if (blkdev_issue_discard(si->bdev, start_block,
194                                     nr_blocks, GFP_NOIO, BLKDEV_IFL_WAIT |
195                                                         BLKDEV_IFL_BARRIER))
196                                 break;
197                 }
198
199                 lh = se->list.next;
200                 se = list_entry(lh, struct swap_extent, list);
201         }
202 }
203
204 static int wait_for_discard(void *word)
205 {
206         schedule();
207         return 0;
208 }
209
210 #define SWAPFILE_CLUSTER        256
211 #define LATENCY_LIMIT           256
212
213 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
214                                           unsigned char usage)
215 {
216         unsigned long offset;
217         unsigned long scan_base;
218         unsigned long last_in_cluster = 0;
219         int latency_ration = LATENCY_LIMIT;
220         int found_free_cluster = 0;
221
222         /*
223          * We try to cluster swap pages by allocating them sequentially
224          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
225          * way, however, we resort to first-free allocation, starting
226          * a new cluster.  This prevents us from scattering swap pages
227          * all over the entire swap partition, so that we reduce
228          * overall disk seek times between swap pages.  -- sct
229          * But we do now try to find an empty cluster.  -Andrea
230          * And we let swap pages go all over an SSD partition.  Hugh
231          */
232
233         si->flags += SWP_SCANNING;
234         scan_base = offset = si->cluster_next;
235
236         if (unlikely(!si->cluster_nr--)) {
237                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
238                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
239                         goto checks;
240                 }
241                 if (si->flags & SWP_DISCARDABLE) {
242                         /*
243                          * Start range check on racing allocations, in case
244                          * they overlap the cluster we eventually decide on
245                          * (we scan without swap_lock to allow preemption).
246                          * It's hardly conceivable that cluster_nr could be
247                          * wrapped during our scan, but don't depend on it.
248                          */
249                         if (si->lowest_alloc)
250                                 goto checks;
251                         si->lowest_alloc = si->max;
252                         si->highest_alloc = 0;
253                 }
254                 spin_unlock(&swap_lock);
255
256                 /*
257                  * If seek is expensive, start searching for new cluster from
258                  * start of partition, to minimize the span of allocated swap.
259                  * But if seek is cheap, search from our current position, so
260                  * that swap is allocated from all over the partition: if the
261                  * Flash Translation Layer only remaps within limited zones,
262                  * we don't want to wear out the first zone too quickly.
263                  */
264                 if (!(si->flags & SWP_SOLIDSTATE))
265                         scan_base = offset = si->lowest_bit;
266                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
267
268                 /* Locate the first empty (unaligned) cluster */
269                 for (; last_in_cluster <= si->highest_bit; offset++) {
270                         if (si->swap_map[offset])
271                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
272                         else if (offset == last_in_cluster) {
273                                 spin_lock(&swap_lock);
274                                 offset -= SWAPFILE_CLUSTER - 1;
275                                 si->cluster_next = offset;
276                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
277                                 found_free_cluster = 1;
278                                 goto checks;
279                         }
280                         if (unlikely(--latency_ration < 0)) {
281                                 cond_resched();
282                                 latency_ration = LATENCY_LIMIT;
283                         }
284                 }
285
286                 offset = si->lowest_bit;
287                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
288
289                 /* Locate the first empty (unaligned) cluster */
290                 for (; last_in_cluster < scan_base; offset++) {
291                         if (si->swap_map[offset])
292                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
293                         else if (offset == last_in_cluster) {
294                                 spin_lock(&swap_lock);
295                                 offset -= SWAPFILE_CLUSTER - 1;
296                                 si->cluster_next = offset;
297                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
298                                 found_free_cluster = 1;
299                                 goto checks;
300                         }
301                         if (unlikely(--latency_ration < 0)) {
302                                 cond_resched();
303                                 latency_ration = LATENCY_LIMIT;
304                         }
305                 }
306
307                 offset = scan_base;
308                 spin_lock(&swap_lock);
309                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
310                 si->lowest_alloc = 0;
311         }
312
313 checks:
314         if (!(si->flags & SWP_WRITEOK))
315                 goto no_page;
316         if (!si->highest_bit)
317                 goto no_page;
318         if (offset > si->highest_bit)
319                 scan_base = offset = si->lowest_bit;
320
321         /* reuse swap entry of cache-only swap if not busy. */
322         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
323                 int swap_was_freed;
324                 spin_unlock(&swap_lock);
325                 swap_was_freed = __try_to_reclaim_swap(si, offset);
326                 spin_lock(&swap_lock);
327                 /* entry was freed successfully, try to use this again */
328                 if (swap_was_freed)
329                         goto checks;
330                 goto scan; /* check next one */
331         }
332
333         if (si->swap_map[offset])
334                 goto scan;
335
336         if (offset == si->lowest_bit)
337                 si->lowest_bit++;
338         if (offset == si->highest_bit)
339                 si->highest_bit--;
340         si->inuse_pages++;
341         if (si->inuse_pages == si->pages) {
342                 si->lowest_bit = si->max;
343                 si->highest_bit = 0;
344         }
345         si->swap_map[offset] = usage;
346         si->cluster_next = offset + 1;
347         si->flags -= SWP_SCANNING;
348
349         if (si->lowest_alloc) {
350                 /*
351                  * Only set when SWP_DISCARDABLE, and there's a scan
352                  * for a free cluster in progress or just completed.
353                  */
354                 if (found_free_cluster) {
355                         /*
356                          * To optimize wear-levelling, discard the
357                          * old data of the cluster, taking care not to
358                          * discard any of its pages that have already
359                          * been allocated by racing tasks (offset has
360                          * already stepped over any at the beginning).
361                          */
362                         if (offset < si->highest_alloc &&
363                             si->lowest_alloc <= last_in_cluster)
364                                 last_in_cluster = si->lowest_alloc - 1;
365                         si->flags |= SWP_DISCARDING;
366                         spin_unlock(&swap_lock);
367
368                         if (offset < last_in_cluster)
369                                 discard_swap_cluster(si, offset,
370                                         last_in_cluster - offset + 1);
371
372                         spin_lock(&swap_lock);
373                         si->lowest_alloc = 0;
374                         si->flags &= ~SWP_DISCARDING;
375
376                         smp_mb();       /* wake_up_bit advises this */
377                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
378
379                 } else if (si->flags & SWP_DISCARDING) {
380                         /*
381                          * Delay using pages allocated by racing tasks
382                          * until the whole discard has been issued. We
383                          * could defer that delay until swap_writepage,
384                          * but it's easier to keep this self-contained.
385                          */
386                         spin_unlock(&swap_lock);
387                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
388                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
389                         spin_lock(&swap_lock);
390                 } else {
391                         /*
392                          * Note pages allocated by racing tasks while
393                          * scan for a free cluster is in progress, so
394                          * that its final discard can exclude them.
395                          */
396                         if (offset < si->lowest_alloc)
397                                 si->lowest_alloc = offset;
398                         if (offset > si->highest_alloc)
399                                 si->highest_alloc = offset;
400                 }
401         }
402         return offset;
403
404 scan:
405         spin_unlock(&swap_lock);
406         while (++offset <= si->highest_bit) {
407                 if (!si->swap_map[offset]) {
408                         spin_lock(&swap_lock);
409                         goto checks;
410                 }
411                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
412                         spin_lock(&swap_lock);
413                         goto checks;
414                 }
415                 if (unlikely(--latency_ration < 0)) {
416                         cond_resched();
417                         latency_ration = LATENCY_LIMIT;
418                 }
419         }
420         offset = si->lowest_bit;
421         while (++offset < scan_base) {
422                 if (!si->swap_map[offset]) {
423                         spin_lock(&swap_lock);
424                         goto checks;
425                 }
426                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
427                         spin_lock(&swap_lock);
428                         goto checks;
429                 }
430                 if (unlikely(--latency_ration < 0)) {
431                         cond_resched();
432                         latency_ration = LATENCY_LIMIT;
433                 }
434         }
435         spin_lock(&swap_lock);
436
437 no_page:
438         si->flags -= SWP_SCANNING;
439         return 0;
440 }
441
442 swp_entry_t get_swap_page(void)
443 {
444         struct swap_info_struct *si;
445         pgoff_t offset;
446         int type, next;
447         int wrapped = 0;
448
449         spin_lock(&swap_lock);
450         if (nr_swap_pages <= 0)
451                 goto noswap;
452         nr_swap_pages--;
453
454         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
455                 si = swap_info[type];
456                 next = si->next;
457                 if (next < 0 ||
458                     (!wrapped && si->prio != swap_info[next]->prio)) {
459                         next = swap_list.head;
460                         wrapped++;
461                 }
462
463                 if (!si->highest_bit)
464                         continue;
465                 if (!(si->flags & SWP_WRITEOK))
466                         continue;
467
468                 swap_list.next = next;
469                 /* This is called for allocating swap entry for cache */
470                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
471                 if (offset) {
472                         spin_unlock(&swap_lock);
473                         return swp_entry(type, offset);
474                 }
475                 next = swap_list.next;
476         }
477
478         nr_swap_pages++;
479 noswap:
480         spin_unlock(&swap_lock);
481         return (swp_entry_t) {0};
482 }
483
484 /* The only caller of this function is now susupend routine */
485 swp_entry_t get_swap_page_of_type(int type)
486 {
487         struct swap_info_struct *si;
488         pgoff_t offset;
489
490         spin_lock(&swap_lock);
491         si = swap_info[type];
492         if (si && (si->flags & SWP_WRITEOK)) {
493                 nr_swap_pages--;
494                 /* This is called for allocating swap entry, not cache */
495                 offset = scan_swap_map(si, 1);
496                 if (offset) {
497                         spin_unlock(&swap_lock);
498                         return swp_entry(type, offset);
499                 }
500                 nr_swap_pages++;
501         }
502         spin_unlock(&swap_lock);
503         return (swp_entry_t) {0};
504 }
505
506 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
507 {
508         struct swap_info_struct *p;
509         unsigned long offset, type;
510
511         if (!entry.val)
512                 goto out;
513         type = swp_type(entry);
514         if (type >= nr_swapfiles)
515                 goto bad_nofile;
516         p = swap_info[type];
517         if (!(p->flags & SWP_USED))
518                 goto bad_device;
519         offset = swp_offset(entry);
520         if (offset >= p->max)
521                 goto bad_offset;
522         if (!p->swap_map[offset])
523                 goto bad_free;
524         spin_lock(&swap_lock);
525         return p;
526
527 bad_free:
528         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
529         goto out;
530 bad_offset:
531         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
532         goto out;
533 bad_device:
534         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
535         goto out;
536 bad_nofile:
537         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
538 out:
539         return NULL;
540 }
541
542 static unsigned char swap_entry_free(struct swap_info_struct *p,
543                                      swp_entry_t entry, unsigned char usage)
544 {
545         unsigned long offset = swp_offset(entry);
546         unsigned char count;
547         unsigned char has_cache;
548
549         count = p->swap_map[offset];
550         has_cache = count & SWAP_HAS_CACHE;
551         count &= ~SWAP_HAS_CACHE;
552
553         if (usage == SWAP_HAS_CACHE) {
554                 VM_BUG_ON(!has_cache);
555                 has_cache = 0;
556         } else if (count == SWAP_MAP_SHMEM) {
557                 /*
558                  * Or we could insist on shmem.c using a special
559                  * swap_shmem_free() and free_shmem_swap_and_cache()...
560                  */
561                 count = 0;
562         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
563                 if (count == COUNT_CONTINUED) {
564                         if (swap_count_continued(p, offset, count))
565                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
566                         else
567                                 count = SWAP_MAP_MAX;
568                 } else
569                         count--;
570         }
571
572         if (!count)
573                 mem_cgroup_uncharge_swap(entry);
574
575         usage = count | has_cache;
576         p->swap_map[offset] = usage;
577
578         /* free if no reference */
579         if (!usage) {
580                 struct gendisk *disk = p->bdev->bd_disk;
581                 if (offset < p->lowest_bit)
582                         p->lowest_bit = offset;
583                 if (offset > p->highest_bit)
584                         p->highest_bit = offset;
585                 if (swap_list.next >= 0 &&
586                     p->prio > swap_info[swap_list.next]->prio)
587                         swap_list.next = p->type;
588                 nr_swap_pages++;
589                 p->inuse_pages--;
590                 if ((p->flags & SWP_BLKDEV) &&
591                                 disk->fops->swap_slot_free_notify)
592                         disk->fops->swap_slot_free_notify(p->bdev, offset);
593         }
594
595         return usage;
596 }
597
598 /*
599  * Caller has made sure that the swapdevice corresponding to entry
600  * is still around or has not been recycled.
601  */
602 void swap_free(swp_entry_t entry)
603 {
604         struct swap_info_struct *p;
605
606         p = swap_info_get(entry);
607         if (p) {
608                 swap_entry_free(p, entry, 1);
609                 spin_unlock(&swap_lock);
610         }
611 }
612
613 /*
614  * Called after dropping swapcache to decrease refcnt to swap entries.
615  */
616 void swapcache_free(swp_entry_t entry, struct page *page)
617 {
618         struct swap_info_struct *p;
619         unsigned char count;
620
621         p = swap_info_get(entry);
622         if (p) {
623                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
624                 if (page)
625                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
626                 spin_unlock(&swap_lock);
627         }
628 }
629
630 /*
631  * How many references to page are currently swapped out?
632  * This does not give an exact answer when swap count is continued,
633  * but does include the high COUNT_CONTINUED flag to allow for that.
634  */
635 static inline int page_swapcount(struct page *page)
636 {
637         int count = 0;
638         struct swap_info_struct *p;
639         swp_entry_t entry;
640
641         entry.val = page_private(page);
642         p = swap_info_get(entry);
643         if (p) {
644                 count = swap_count(p->swap_map[swp_offset(entry)]);
645                 spin_unlock(&swap_lock);
646         }
647         return count;
648 }
649
650 /*
651  * We can write to an anon page without COW if there are no other references
652  * to it.  And as a side-effect, free up its swap: because the old content
653  * on disk will never be read, and seeking back there to write new content
654  * later would only waste time away from clustering.
655  */
656 int reuse_swap_page(struct page *page)
657 {
658         int count;
659
660         VM_BUG_ON(!PageLocked(page));
661         if (unlikely(PageKsm(page)))
662                 return 0;
663         count = page_mapcount(page);
664         if (count <= 1 && PageSwapCache(page)) {
665                 count += page_swapcount(page);
666                 if (count == 1 && !PageWriteback(page)) {
667                         delete_from_swap_cache(page);
668                         SetPageDirty(page);
669                 }
670         }
671         return count <= 1;
672 }
673
674 /*
675  * If swap is getting full, or if there are no more mappings of this page,
676  * then try_to_free_swap is called to free its swap space.
677  */
678 int try_to_free_swap(struct page *page)
679 {
680         VM_BUG_ON(!PageLocked(page));
681
682         if (!PageSwapCache(page))
683                 return 0;
684         if (PageWriteback(page))
685                 return 0;
686         if (page_swapcount(page))
687                 return 0;
688
689         delete_from_swap_cache(page);
690         SetPageDirty(page);
691         return 1;
692 }
693
694 /*
695  * Free the swap entry like above, but also try to
696  * free the page cache entry if it is the last user.
697  */
698 int free_swap_and_cache(swp_entry_t entry)
699 {
700         struct swap_info_struct *p;
701         struct page *page = NULL;
702
703         if (non_swap_entry(entry))
704                 return 1;
705
706         p = swap_info_get(entry);
707         if (p) {
708                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
709                         page = find_get_page(&swapper_space, entry.val);
710                         if (page && !trylock_page(page)) {
711                                 page_cache_release(page);
712                                 page = NULL;
713                         }
714                 }
715                 spin_unlock(&swap_lock);
716         }
717         if (page) {
718                 /*
719                  * Not mapped elsewhere, or swap space full? Free it!
720                  * Also recheck PageSwapCache now page is locked (above).
721                  */
722                 if (PageSwapCache(page) && !PageWriteback(page) &&
723                                 (!page_mapped(page) || vm_swap_full())) {
724                         delete_from_swap_cache(page);
725                         SetPageDirty(page);
726                 }
727                 unlock_page(page);
728                 page_cache_release(page);
729         }
730         return p != NULL;
731 }
732
733 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
734 /**
735  * mem_cgroup_count_swap_user - count the user of a swap entry
736  * @ent: the swap entry to be checked
737  * @pagep: the pointer for the swap cache page of the entry to be stored
738  *
739  * Returns the number of the user of the swap entry. The number is valid only
740  * for swaps of anonymous pages.
741  * If the entry is found on swap cache, the page is stored to pagep with
742  * refcount of it being incremented.
743  */
744 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
745 {
746         struct page *page;
747         struct swap_info_struct *p;
748         int count = 0;
749
750         page = find_get_page(&swapper_space, ent.val);
751         if (page)
752                 count += page_mapcount(page);
753         p = swap_info_get(ent);
754         if (p) {
755                 count += swap_count(p->swap_map[swp_offset(ent)]);
756                 spin_unlock(&swap_lock);
757         }
758
759         *pagep = page;
760         return count;
761 }
762 #endif
763
764 #ifdef CONFIG_HIBERNATION
765 /*
766  * Find the swap type that corresponds to given device (if any).
767  *
768  * @offset - number of the PAGE_SIZE-sized block of the device, starting
769  * from 0, in which the swap header is expected to be located.
770  *
771  * This is needed for the suspend to disk (aka swsusp).
772  */
773 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
774 {
775         struct block_device *bdev = NULL;
776         int type;
777
778         if (device)
779                 bdev = bdget(device);
780
781         spin_lock(&swap_lock);
782         for (type = 0; type < nr_swapfiles; type++) {
783                 struct swap_info_struct *sis = swap_info[type];
784
785                 if (!(sis->flags & SWP_WRITEOK))
786                         continue;
787
788                 if (!bdev) {
789                         if (bdev_p)
790                                 *bdev_p = bdgrab(sis->bdev);
791
792                         spin_unlock(&swap_lock);
793                         return type;
794                 }
795                 if (bdev == sis->bdev) {
796                         struct swap_extent *se = &sis->first_swap_extent;
797
798                         if (se->start_block == offset) {
799                                 if (bdev_p)
800                                         *bdev_p = bdgrab(sis->bdev);
801
802                                 spin_unlock(&swap_lock);
803                                 bdput(bdev);
804                                 return type;
805                         }
806                 }
807         }
808         spin_unlock(&swap_lock);
809         if (bdev)
810                 bdput(bdev);
811
812         return -ENODEV;
813 }
814
815 /*
816  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
817  * corresponding to given index in swap_info (swap type).
818  */
819 sector_t swapdev_block(int type, pgoff_t offset)
820 {
821         struct block_device *bdev;
822
823         if ((unsigned int)type >= nr_swapfiles)
824                 return 0;
825         if (!(swap_info[type]->flags & SWP_WRITEOK))
826                 return 0;
827         return map_swap_entry(swp_entry(type, offset), &bdev);
828 }
829
830 /*
831  * Return either the total number of swap pages of given type, or the number
832  * of free pages of that type (depending on @free)
833  *
834  * This is needed for software suspend
835  */
836 unsigned int count_swap_pages(int type, int free)
837 {
838         unsigned int n = 0;
839
840         spin_lock(&swap_lock);
841         if ((unsigned int)type < nr_swapfiles) {
842                 struct swap_info_struct *sis = swap_info[type];
843
844                 if (sis->flags & SWP_WRITEOK) {
845                         n = sis->pages;
846                         if (free)
847                                 n -= sis->inuse_pages;
848                 }
849         }
850         spin_unlock(&swap_lock);
851         return n;
852 }
853 #endif /* CONFIG_HIBERNATION */
854
855 /*
856  * No need to decide whether this PTE shares the swap entry with others,
857  * just let do_wp_page work it out if a write is requested later - to
858  * force COW, vm_page_prot omits write permission from any private vma.
859  */
860 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
861                 unsigned long addr, swp_entry_t entry, struct page *page)
862 {
863         struct mem_cgroup *ptr = NULL;
864         spinlock_t *ptl;
865         pte_t *pte;
866         int ret = 1;
867
868         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
869                 ret = -ENOMEM;
870                 goto out_nolock;
871         }
872
873         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
874         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
875                 if (ret > 0)
876                         mem_cgroup_cancel_charge_swapin(ptr);
877                 ret = 0;
878                 goto out;
879         }
880
881         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
882         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
883         get_page(page);
884         set_pte_at(vma->vm_mm, addr, pte,
885                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
886         page_add_anon_rmap(page, vma, addr);
887         mem_cgroup_commit_charge_swapin(page, ptr);
888         swap_free(entry);
889         /*
890          * Move the page to the active list so it is not
891          * immediately swapped out again after swapon.
892          */
893         activate_page(page);
894 out:
895         pte_unmap_unlock(pte, ptl);
896 out_nolock:
897         return ret;
898 }
899
900 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
901                                 unsigned long addr, unsigned long end,
902                                 swp_entry_t entry, struct page *page)
903 {
904         pte_t swp_pte = swp_entry_to_pte(entry);
905         pte_t *pte;
906         int ret = 0;
907
908         /*
909          * We don't actually need pte lock while scanning for swp_pte: since
910          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
911          * page table while we're scanning; though it could get zapped, and on
912          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
913          * of unmatched parts which look like swp_pte, so unuse_pte must
914          * recheck under pte lock.  Scanning without pte lock lets it be
915          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
916          */
917         pte = pte_offset_map(pmd, addr);
918         do {
919                 /*
920                  * swapoff spends a _lot_ of time in this loop!
921                  * Test inline before going to call unuse_pte.
922                  */
923                 if (unlikely(pte_same(*pte, swp_pte))) {
924                         pte_unmap(pte);
925                         ret = unuse_pte(vma, pmd, addr, entry, page);
926                         if (ret)
927                                 goto out;
928                         pte = pte_offset_map(pmd, addr);
929                 }
930         } while (pte++, addr += PAGE_SIZE, addr != end);
931         pte_unmap(pte - 1);
932 out:
933         return ret;
934 }
935
936 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
937                                 unsigned long addr, unsigned long end,
938                                 swp_entry_t entry, struct page *page)
939 {
940         pmd_t *pmd;
941         unsigned long next;
942         int ret;
943
944         pmd = pmd_offset(pud, addr);
945         do {
946                 next = pmd_addr_end(addr, end);
947                 if (pmd_none_or_clear_bad(pmd))
948                         continue;
949                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
950                 if (ret)
951                         return ret;
952         } while (pmd++, addr = next, addr != end);
953         return 0;
954 }
955
956 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
957                                 unsigned long addr, unsigned long end,
958                                 swp_entry_t entry, struct page *page)
959 {
960         pud_t *pud;
961         unsigned long next;
962         int ret;
963
964         pud = pud_offset(pgd, addr);
965         do {
966                 next = pud_addr_end(addr, end);
967                 if (pud_none_or_clear_bad(pud))
968                         continue;
969                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
970                 if (ret)
971                         return ret;
972         } while (pud++, addr = next, addr != end);
973         return 0;
974 }
975
976 static int unuse_vma(struct vm_area_struct *vma,
977                                 swp_entry_t entry, struct page *page)
978 {
979         pgd_t *pgd;
980         unsigned long addr, end, next;
981         int ret;
982
983         if (page_anon_vma(page)) {
984                 addr = page_address_in_vma(page, vma);
985                 if (addr == -EFAULT)
986                         return 0;
987                 else
988                         end = addr + PAGE_SIZE;
989         } else {
990                 addr = vma->vm_start;
991                 end = vma->vm_end;
992         }
993
994         pgd = pgd_offset(vma->vm_mm, addr);
995         do {
996                 next = pgd_addr_end(addr, end);
997                 if (pgd_none_or_clear_bad(pgd))
998                         continue;
999                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1000                 if (ret)
1001                         return ret;
1002         } while (pgd++, addr = next, addr != end);
1003         return 0;
1004 }
1005
1006 static int unuse_mm(struct mm_struct *mm,
1007                                 swp_entry_t entry, struct page *page)
1008 {
1009         struct vm_area_struct *vma;
1010         int ret = 0;
1011
1012         if (!down_read_trylock(&mm->mmap_sem)) {
1013                 /*
1014                  * Activate page so shrink_inactive_list is unlikely to unmap
1015                  * its ptes while lock is dropped, so swapoff can make progress.
1016                  */
1017                 activate_page(page);
1018                 unlock_page(page);
1019                 down_read(&mm->mmap_sem);
1020                 lock_page(page);
1021         }
1022         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1023                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1024                         break;
1025         }
1026         up_read(&mm->mmap_sem);
1027         return (ret < 0)? ret: 0;
1028 }
1029
1030 /*
1031  * Scan swap_map from current position to next entry still in use.
1032  * Recycle to start on reaching the end, returning 0 when empty.
1033  */
1034 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1035                                         unsigned int prev)
1036 {
1037         unsigned int max = si->max;
1038         unsigned int i = prev;
1039         unsigned char count;
1040
1041         /*
1042          * No need for swap_lock here: we're just looking
1043          * for whether an entry is in use, not modifying it; false
1044          * hits are okay, and sys_swapoff() has already prevented new
1045          * allocations from this area (while holding swap_lock).
1046          */
1047         for (;;) {
1048                 if (++i >= max) {
1049                         if (!prev) {
1050                                 i = 0;
1051                                 break;
1052                         }
1053                         /*
1054                          * No entries in use at top of swap_map,
1055                          * loop back to start and recheck there.
1056                          */
1057                         max = prev + 1;
1058                         prev = 0;
1059                         i = 1;
1060                 }
1061                 count = si->swap_map[i];
1062                 if (count && swap_count(count) != SWAP_MAP_BAD)
1063                         break;
1064         }
1065         return i;
1066 }
1067
1068 /*
1069  * We completely avoid races by reading each swap page in advance,
1070  * and then search for the process using it.  All the necessary
1071  * page table adjustments can then be made atomically.
1072  */
1073 static int try_to_unuse(unsigned int type)
1074 {
1075         struct swap_info_struct *si = swap_info[type];
1076         struct mm_struct *start_mm;
1077         unsigned char *swap_map;
1078         unsigned char swcount;
1079         struct page *page;
1080         swp_entry_t entry;
1081         unsigned int i = 0;
1082         int retval = 0;
1083
1084         /*
1085          * When searching mms for an entry, a good strategy is to
1086          * start at the first mm we freed the previous entry from
1087          * (though actually we don't notice whether we or coincidence
1088          * freed the entry).  Initialize this start_mm with a hold.
1089          *
1090          * A simpler strategy would be to start at the last mm we
1091          * freed the previous entry from; but that would take less
1092          * advantage of mmlist ordering, which clusters forked mms
1093          * together, child after parent.  If we race with dup_mmap(), we
1094          * prefer to resolve parent before child, lest we miss entries
1095          * duplicated after we scanned child: using last mm would invert
1096          * that.
1097          */
1098         start_mm = &init_mm;
1099         atomic_inc(&init_mm.mm_users);
1100
1101         /*
1102          * Keep on scanning until all entries have gone.  Usually,
1103          * one pass through swap_map is enough, but not necessarily:
1104          * there are races when an instance of an entry might be missed.
1105          */
1106         while ((i = find_next_to_unuse(si, i)) != 0) {
1107                 if (signal_pending(current)) {
1108                         retval = -EINTR;
1109                         break;
1110                 }
1111
1112                 /*
1113                  * Get a page for the entry, using the existing swap
1114                  * cache page if there is one.  Otherwise, get a clean
1115                  * page and read the swap into it.
1116                  */
1117                 swap_map = &si->swap_map[i];
1118                 entry = swp_entry(type, i);
1119                 page = read_swap_cache_async(entry,
1120                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1121                 if (!page) {
1122                         /*
1123                          * Either swap_duplicate() failed because entry
1124                          * has been freed independently, and will not be
1125                          * reused since sys_swapoff() already disabled
1126                          * allocation from here, or alloc_page() failed.
1127                          */
1128                         if (!*swap_map)
1129                                 continue;
1130                         retval = -ENOMEM;
1131                         break;
1132                 }
1133
1134                 /*
1135                  * Don't hold on to start_mm if it looks like exiting.
1136                  */
1137                 if (atomic_read(&start_mm->mm_users) == 1) {
1138                         mmput(start_mm);
1139                         start_mm = &init_mm;
1140                         atomic_inc(&init_mm.mm_users);
1141                 }
1142
1143                 /*
1144                  * Wait for and lock page.  When do_swap_page races with
1145                  * try_to_unuse, do_swap_page can handle the fault much
1146                  * faster than try_to_unuse can locate the entry.  This
1147                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1148                  * defer to do_swap_page in such a case - in some tests,
1149                  * do_swap_page and try_to_unuse repeatedly compete.
1150                  */
1151                 wait_on_page_locked(page);
1152                 wait_on_page_writeback(page);
1153                 lock_page(page);
1154                 wait_on_page_writeback(page);
1155
1156                 /*
1157                  * Remove all references to entry.
1158                  */
1159                 swcount = *swap_map;
1160                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1161                         retval = shmem_unuse(entry, page);
1162                         /* page has already been unlocked and released */
1163                         if (retval < 0)
1164                                 break;
1165                         continue;
1166                 }
1167                 if (swap_count(swcount) && start_mm != &init_mm)
1168                         retval = unuse_mm(start_mm, entry, page);
1169
1170                 if (swap_count(*swap_map)) {
1171                         int set_start_mm = (*swap_map >= swcount);
1172                         struct list_head *p = &start_mm->mmlist;
1173                         struct mm_struct *new_start_mm = start_mm;
1174                         struct mm_struct *prev_mm = start_mm;
1175                         struct mm_struct *mm;
1176
1177                         atomic_inc(&new_start_mm->mm_users);
1178                         atomic_inc(&prev_mm->mm_users);
1179                         spin_lock(&mmlist_lock);
1180                         while (swap_count(*swap_map) && !retval &&
1181                                         (p = p->next) != &start_mm->mmlist) {
1182                                 mm = list_entry(p, struct mm_struct, mmlist);
1183                                 if (!atomic_inc_not_zero(&mm->mm_users))
1184                                         continue;
1185                                 spin_unlock(&mmlist_lock);
1186                                 mmput(prev_mm);
1187                                 prev_mm = mm;
1188
1189                                 cond_resched();
1190
1191                                 swcount = *swap_map;
1192                                 if (!swap_count(swcount)) /* any usage ? */
1193                                         ;
1194                                 else if (mm == &init_mm)
1195                                         set_start_mm = 1;
1196                                 else
1197                                         retval = unuse_mm(mm, entry, page);
1198
1199                                 if (set_start_mm && *swap_map < swcount) {
1200                                         mmput(new_start_mm);
1201                                         atomic_inc(&mm->mm_users);
1202                                         new_start_mm = mm;
1203                                         set_start_mm = 0;
1204                                 }
1205                                 spin_lock(&mmlist_lock);
1206                         }
1207                         spin_unlock(&mmlist_lock);
1208                         mmput(prev_mm);
1209                         mmput(start_mm);
1210                         start_mm = new_start_mm;
1211                 }
1212                 if (retval) {
1213                         unlock_page(page);
1214                         page_cache_release(page);
1215                         break;
1216                 }
1217
1218                 /*
1219                  * If a reference remains (rare), we would like to leave
1220                  * the page in the swap cache; but try_to_unmap could
1221                  * then re-duplicate the entry once we drop page lock,
1222                  * so we might loop indefinitely; also, that page could
1223                  * not be swapped out to other storage meanwhile.  So:
1224                  * delete from cache even if there's another reference,
1225                  * after ensuring that the data has been saved to disk -
1226                  * since if the reference remains (rarer), it will be
1227                  * read from disk into another page.  Splitting into two
1228                  * pages would be incorrect if swap supported "shared
1229                  * private" pages, but they are handled by tmpfs files.
1230                  *
1231                  * Given how unuse_vma() targets one particular offset
1232                  * in an anon_vma, once the anon_vma has been determined,
1233                  * this splitting happens to be just what is needed to
1234                  * handle where KSM pages have been swapped out: re-reading
1235                  * is unnecessarily slow, but we can fix that later on.
1236                  */
1237                 if (swap_count(*swap_map) &&
1238                      PageDirty(page) && PageSwapCache(page)) {
1239                         struct writeback_control wbc = {
1240                                 .sync_mode = WB_SYNC_NONE,
1241                         };
1242
1243                         swap_writepage(page, &wbc);
1244                         lock_page(page);
1245                         wait_on_page_writeback(page);
1246                 }
1247
1248                 /*
1249                  * It is conceivable that a racing task removed this page from
1250                  * swap cache just before we acquired the page lock at the top,
1251                  * or while we dropped it in unuse_mm().  The page might even
1252                  * be back in swap cache on another swap area: that we must not
1253                  * delete, since it may not have been written out to swap yet.
1254                  */
1255                 if (PageSwapCache(page) &&
1256                     likely(page_private(page) == entry.val))
1257                         delete_from_swap_cache(page);
1258
1259                 /*
1260                  * So we could skip searching mms once swap count went
1261                  * to 1, we did not mark any present ptes as dirty: must
1262                  * mark page dirty so shrink_page_list will preserve it.
1263                  */
1264                 SetPageDirty(page);
1265                 unlock_page(page);
1266                 page_cache_release(page);
1267
1268                 /*
1269                  * Make sure that we aren't completely killing
1270                  * interactive performance.
1271                  */
1272                 cond_resched();
1273         }
1274
1275         mmput(start_mm);
1276         return retval;
1277 }
1278
1279 /*
1280  * After a successful try_to_unuse, if no swap is now in use, we know
1281  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1282  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1283  * added to the mmlist just after page_duplicate - before would be racy.
1284  */
1285 static void drain_mmlist(void)
1286 {
1287         struct list_head *p, *next;
1288         unsigned int type;
1289
1290         for (type = 0; type < nr_swapfiles; type++)
1291                 if (swap_info[type]->inuse_pages)
1292                         return;
1293         spin_lock(&mmlist_lock);
1294         list_for_each_safe(p, next, &init_mm.mmlist)
1295                 list_del_init(p);
1296         spin_unlock(&mmlist_lock);
1297 }
1298
1299 /*
1300  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1301  * corresponds to page offset for the specified swap entry.
1302  * Note that the type of this function is sector_t, but it returns page offset
1303  * into the bdev, not sector offset.
1304  */
1305 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1306 {
1307         struct swap_info_struct *sis;
1308         struct swap_extent *start_se;
1309         struct swap_extent *se;
1310         pgoff_t offset;
1311
1312         sis = swap_info[swp_type(entry)];
1313         *bdev = sis->bdev;
1314
1315         offset = swp_offset(entry);
1316         start_se = sis->curr_swap_extent;
1317         se = start_se;
1318
1319         for ( ; ; ) {
1320                 struct list_head *lh;
1321
1322                 if (se->start_page <= offset &&
1323                                 offset < (se->start_page + se->nr_pages)) {
1324                         return se->start_block + (offset - se->start_page);
1325                 }
1326                 lh = se->list.next;
1327                 se = list_entry(lh, struct swap_extent, list);
1328                 sis->curr_swap_extent = se;
1329                 BUG_ON(se == start_se);         /* It *must* be present */
1330         }
1331 }
1332
1333 /*
1334  * Returns the page offset into bdev for the specified page's swap entry.
1335  */
1336 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1337 {
1338         swp_entry_t entry;
1339         entry.val = page_private(page);
1340         return map_swap_entry(entry, bdev);
1341 }
1342
1343 /*
1344  * Free all of a swapdev's extent information
1345  */
1346 static void destroy_swap_extents(struct swap_info_struct *sis)
1347 {
1348         while (!list_empty(&sis->first_swap_extent.list)) {
1349                 struct swap_extent *se;
1350
1351                 se = list_entry(sis->first_swap_extent.list.next,
1352                                 struct swap_extent, list);
1353                 list_del(&se->list);
1354                 kfree(se);
1355         }
1356 }
1357
1358 /*
1359  * Add a block range (and the corresponding page range) into this swapdev's
1360  * extent list.  The extent list is kept sorted in page order.
1361  *
1362  * This function rather assumes that it is called in ascending page order.
1363  */
1364 static int
1365 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1366                 unsigned long nr_pages, sector_t start_block)
1367 {
1368         struct swap_extent *se;
1369         struct swap_extent *new_se;
1370         struct list_head *lh;
1371
1372         if (start_page == 0) {
1373                 se = &sis->first_swap_extent;
1374                 sis->curr_swap_extent = se;
1375                 se->start_page = 0;
1376                 se->nr_pages = nr_pages;
1377                 se->start_block = start_block;
1378                 return 1;
1379         } else {
1380                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1381                 se = list_entry(lh, struct swap_extent, list);
1382                 BUG_ON(se->start_page + se->nr_pages != start_page);
1383                 if (se->start_block + se->nr_pages == start_block) {
1384                         /* Merge it */
1385                         se->nr_pages += nr_pages;
1386                         return 0;
1387                 }
1388         }
1389
1390         /*
1391          * No merge.  Insert a new extent, preserving ordering.
1392          */
1393         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1394         if (new_se == NULL)
1395                 return -ENOMEM;
1396         new_se->start_page = start_page;
1397         new_se->nr_pages = nr_pages;
1398         new_se->start_block = start_block;
1399
1400         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1401         return 1;
1402 }
1403
1404 /*
1405  * A `swap extent' is a simple thing which maps a contiguous range of pages
1406  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1407  * is built at swapon time and is then used at swap_writepage/swap_readpage
1408  * time for locating where on disk a page belongs.
1409  *
1410  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1411  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1412  * swap files identically.
1413  *
1414  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1415  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1416  * swapfiles are handled *identically* after swapon time.
1417  *
1418  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1419  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1420  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1421  * requirements, they are simply tossed out - we will never use those blocks
1422  * for swapping.
1423  *
1424  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1425  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1426  * which will scribble on the fs.
1427  *
1428  * The amount of disk space which a single swap extent represents varies.
1429  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1430  * extents in the list.  To avoid much list walking, we cache the previous
1431  * search location in `curr_swap_extent', and start new searches from there.
1432  * This is extremely effective.  The average number of iterations in
1433  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1434  */
1435 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1436 {
1437         struct inode *inode;
1438         unsigned blocks_per_page;
1439         unsigned long page_no;
1440         unsigned blkbits;
1441         sector_t probe_block;
1442         sector_t last_block;
1443         sector_t lowest_block = -1;
1444         sector_t highest_block = 0;
1445         int nr_extents = 0;
1446         int ret;
1447
1448         inode = sis->swap_file->f_mapping->host;
1449         if (S_ISBLK(inode->i_mode)) {
1450                 ret = add_swap_extent(sis, 0, sis->max, 0);
1451                 *span = sis->pages;
1452                 goto out;
1453         }
1454
1455         blkbits = inode->i_blkbits;
1456         blocks_per_page = PAGE_SIZE >> blkbits;
1457
1458         /*
1459          * Map all the blocks into the extent list.  This code doesn't try
1460          * to be very smart.
1461          */
1462         probe_block = 0;
1463         page_no = 0;
1464         last_block = i_size_read(inode) >> blkbits;
1465         while ((probe_block + blocks_per_page) <= last_block &&
1466                         page_no < sis->max) {
1467                 unsigned block_in_page;
1468                 sector_t first_block;
1469
1470                 first_block = bmap(inode, probe_block);
1471                 if (first_block == 0)
1472                         goto bad_bmap;
1473
1474                 /*
1475                  * It must be PAGE_SIZE aligned on-disk
1476                  */
1477                 if (first_block & (blocks_per_page - 1)) {
1478                         probe_block++;
1479                         goto reprobe;
1480                 }
1481
1482                 for (block_in_page = 1; block_in_page < blocks_per_page;
1483                                         block_in_page++) {
1484                         sector_t block;
1485
1486                         block = bmap(inode, probe_block + block_in_page);
1487                         if (block == 0)
1488                                 goto bad_bmap;
1489                         if (block != first_block + block_in_page) {
1490                                 /* Discontiguity */
1491                                 probe_block++;
1492                                 goto reprobe;
1493                         }
1494                 }
1495
1496                 first_block >>= (PAGE_SHIFT - blkbits);
1497                 if (page_no) {  /* exclude the header page */
1498                         if (first_block < lowest_block)
1499                                 lowest_block = first_block;
1500                         if (first_block > highest_block)
1501                                 highest_block = first_block;
1502                 }
1503
1504                 /*
1505                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1506                  */
1507                 ret = add_swap_extent(sis, page_no, 1, first_block);
1508                 if (ret < 0)
1509                         goto out;
1510                 nr_extents += ret;
1511                 page_no++;
1512                 probe_block += blocks_per_page;
1513 reprobe:
1514                 continue;
1515         }
1516         ret = nr_extents;
1517         *span = 1 + highest_block - lowest_block;
1518         if (page_no == 0)
1519                 page_no = 1;    /* force Empty message */
1520         sis->max = page_no;
1521         sis->pages = page_no - 1;
1522         sis->highest_bit = page_no - 1;
1523 out:
1524         return ret;
1525 bad_bmap:
1526         printk(KERN_ERR "swapon: swapfile has holes\n");
1527         ret = -EINVAL;
1528         goto out;
1529 }
1530
1531 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1532 {
1533         struct swap_info_struct *p = NULL;
1534         unsigned char *swap_map;
1535         struct file *swap_file, *victim;
1536         struct address_space *mapping;
1537         struct inode *inode;
1538         char *pathname;
1539         int i, type, prev;
1540         int err;
1541
1542         if (!capable(CAP_SYS_ADMIN))
1543                 return -EPERM;
1544
1545         pathname = getname(specialfile);
1546         err = PTR_ERR(pathname);
1547         if (IS_ERR(pathname))
1548                 goto out;
1549
1550         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1551         putname(pathname);
1552         err = PTR_ERR(victim);
1553         if (IS_ERR(victim))
1554                 goto out;
1555
1556         mapping = victim->f_mapping;
1557         prev = -1;
1558         spin_lock(&swap_lock);
1559         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1560                 p = swap_info[type];
1561                 if (p->flags & SWP_WRITEOK) {
1562                         if (p->swap_file->f_mapping == mapping)
1563                                 break;
1564                 }
1565                 prev = type;
1566         }
1567         if (type < 0) {
1568                 err = -EINVAL;
1569                 spin_unlock(&swap_lock);
1570                 goto out_dput;
1571         }
1572         if (!security_vm_enough_memory(p->pages))
1573                 vm_unacct_memory(p->pages);
1574         else {
1575                 err = -ENOMEM;
1576                 spin_unlock(&swap_lock);
1577                 goto out_dput;
1578         }
1579         if (prev < 0)
1580                 swap_list.head = p->next;
1581         else
1582                 swap_info[prev]->next = p->next;
1583         if (type == swap_list.next) {
1584                 /* just pick something that's safe... */
1585                 swap_list.next = swap_list.head;
1586         }
1587         if (p->prio < 0) {
1588                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1589                         swap_info[i]->prio = p->prio--;
1590                 least_priority++;
1591         }
1592         nr_swap_pages -= p->pages;
1593         total_swap_pages -= p->pages;
1594         p->flags &= ~SWP_WRITEOK;
1595         spin_unlock(&swap_lock);
1596
1597         current->flags |= PF_OOM_ORIGIN;
1598         err = try_to_unuse(type);
1599         current->flags &= ~PF_OOM_ORIGIN;
1600
1601         if (err) {
1602                 /* re-insert swap space back into swap_list */
1603                 spin_lock(&swap_lock);
1604                 if (p->prio < 0)
1605                         p->prio = --least_priority;
1606                 prev = -1;
1607                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1608                         if (p->prio >= swap_info[i]->prio)
1609                                 break;
1610                         prev = i;
1611                 }
1612                 p->next = i;
1613                 if (prev < 0)
1614                         swap_list.head = swap_list.next = type;
1615                 else
1616                         swap_info[prev]->next = type;
1617                 nr_swap_pages += p->pages;
1618                 total_swap_pages += p->pages;
1619                 p->flags |= SWP_WRITEOK;
1620                 spin_unlock(&swap_lock);
1621                 goto out_dput;
1622         }
1623
1624         /* wait for any unplug function to finish */
1625         down_write(&swap_unplug_sem);
1626         up_write(&swap_unplug_sem);
1627
1628         destroy_swap_extents(p);
1629         if (p->flags & SWP_CONTINUED)
1630                 free_swap_count_continuations(p);
1631
1632         mutex_lock(&swapon_mutex);
1633         spin_lock(&swap_lock);
1634         drain_mmlist();
1635
1636         /* wait for anyone still in scan_swap_map */
1637         p->highest_bit = 0;             /* cuts scans short */
1638         while (p->flags >= SWP_SCANNING) {
1639                 spin_unlock(&swap_lock);
1640                 schedule_timeout_uninterruptible(1);
1641                 spin_lock(&swap_lock);
1642         }
1643
1644         swap_file = p->swap_file;
1645         p->swap_file = NULL;
1646         p->max = 0;
1647         swap_map = p->swap_map;
1648         p->swap_map = NULL;
1649         p->flags = 0;
1650         spin_unlock(&swap_lock);
1651         mutex_unlock(&swapon_mutex);
1652         vfree(swap_map);
1653         /* Destroy swap account informatin */
1654         swap_cgroup_swapoff(type);
1655
1656         inode = mapping->host;
1657         if (S_ISBLK(inode->i_mode)) {
1658                 struct block_device *bdev = I_BDEV(inode);
1659                 set_blocksize(bdev, p->old_block_size);
1660                 bd_release(bdev);
1661         } else {
1662                 mutex_lock(&inode->i_mutex);
1663                 inode->i_flags &= ~S_SWAPFILE;
1664                 mutex_unlock(&inode->i_mutex);
1665         }
1666         filp_close(swap_file, NULL);
1667         err = 0;
1668
1669 out_dput:
1670         filp_close(victim, NULL);
1671 out:
1672         return err;
1673 }
1674
1675 #ifdef CONFIG_PROC_FS
1676 /* iterator */
1677 static void *swap_start(struct seq_file *swap, loff_t *pos)
1678 {
1679         struct swap_info_struct *si;
1680         int type;
1681         loff_t l = *pos;
1682
1683         mutex_lock(&swapon_mutex);
1684
1685         if (!l)
1686                 return SEQ_START_TOKEN;
1687
1688         for (type = 0; type < nr_swapfiles; type++) {
1689                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1690                 si = swap_info[type];
1691                 if (!(si->flags & SWP_USED) || !si->swap_map)
1692                         continue;
1693                 if (!--l)
1694                         return si;
1695         }
1696
1697         return NULL;
1698 }
1699
1700 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1701 {
1702         struct swap_info_struct *si = v;
1703         int type;
1704
1705         if (v == SEQ_START_TOKEN)
1706                 type = 0;
1707         else
1708                 type = si->type + 1;
1709
1710         for (; type < nr_swapfiles; type++) {
1711                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1712                 si = swap_info[type];
1713                 if (!(si->flags & SWP_USED) || !si->swap_map)
1714                         continue;
1715                 ++*pos;
1716                 return si;
1717         }
1718
1719         return NULL;
1720 }
1721
1722 static void swap_stop(struct seq_file *swap, void *v)
1723 {
1724         mutex_unlock(&swapon_mutex);
1725 }
1726
1727 static int swap_show(struct seq_file *swap, void *v)
1728 {
1729         struct swap_info_struct *si = v;
1730         struct file *file;
1731         int len;
1732
1733         if (si == SEQ_START_TOKEN) {
1734                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1735                 return 0;
1736         }
1737
1738         file = si->swap_file;
1739         len = seq_path(swap, &file->f_path, " \t\n\\");
1740         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1741                         len < 40 ? 40 - len : 1, " ",
1742                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1743                                 "partition" : "file\t",
1744                         si->pages << (PAGE_SHIFT - 10),
1745                         si->inuse_pages << (PAGE_SHIFT - 10),
1746                         si->prio);
1747         return 0;
1748 }
1749
1750 static const struct seq_operations swaps_op = {
1751         .start =        swap_start,
1752         .next =         swap_next,
1753         .stop =         swap_stop,
1754         .show =         swap_show
1755 };
1756
1757 static int swaps_open(struct inode *inode, struct file *file)
1758 {
1759         return seq_open(file, &swaps_op);
1760 }
1761
1762 static const struct file_operations proc_swaps_operations = {
1763         .open           = swaps_open,
1764         .read           = seq_read,
1765         .llseek         = seq_lseek,
1766         .release        = seq_release,
1767 };
1768
1769 static int __init procswaps_init(void)
1770 {
1771         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1772         return 0;
1773 }
1774 __initcall(procswaps_init);
1775 #endif /* CONFIG_PROC_FS */
1776
1777 #ifdef MAX_SWAPFILES_CHECK
1778 static int __init max_swapfiles_check(void)
1779 {
1780         MAX_SWAPFILES_CHECK();
1781         return 0;
1782 }
1783 late_initcall(max_swapfiles_check);
1784 #endif
1785
1786 /*
1787  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1788  *
1789  * The swapon system call
1790  */
1791 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1792 {
1793         struct swap_info_struct *p;
1794         char *name = NULL;
1795         struct block_device *bdev = NULL;
1796         struct file *swap_file = NULL;
1797         struct address_space *mapping;
1798         unsigned int type;
1799         int i, prev;
1800         int error;
1801         union swap_header *swap_header;
1802         unsigned int nr_good_pages;
1803         int nr_extents = 0;
1804         sector_t span;
1805         unsigned long maxpages;
1806         unsigned long swapfilepages;
1807         unsigned char *swap_map = NULL;
1808         struct page *page = NULL;
1809         struct inode *inode = NULL;
1810         int did_down = 0;
1811
1812         if (!capable(CAP_SYS_ADMIN))
1813                 return -EPERM;
1814
1815         p = kzalloc(sizeof(*p), GFP_KERNEL);
1816         if (!p)
1817                 return -ENOMEM;
1818
1819         spin_lock(&swap_lock);
1820         for (type = 0; type < nr_swapfiles; type++) {
1821                 if (!(swap_info[type]->flags & SWP_USED))
1822                         break;
1823         }
1824         error = -EPERM;
1825         if (type >= MAX_SWAPFILES) {
1826                 spin_unlock(&swap_lock);
1827                 kfree(p);
1828                 goto out;
1829         }
1830         if (type >= nr_swapfiles) {
1831                 p->type = type;
1832                 swap_info[type] = p;
1833                 /*
1834                  * Write swap_info[type] before nr_swapfiles, in case a
1835                  * racing procfs swap_start() or swap_next() is reading them.
1836                  * (We never shrink nr_swapfiles, we never free this entry.)
1837                  */
1838                 smp_wmb();
1839                 nr_swapfiles++;
1840         } else {
1841                 kfree(p);
1842                 p = swap_info[type];
1843                 /*
1844                  * Do not memset this entry: a racing procfs swap_next()
1845                  * would be relying on p->type to remain valid.
1846                  */
1847         }
1848         INIT_LIST_HEAD(&p->first_swap_extent.list);
1849         p->flags = SWP_USED;
1850         p->next = -1;
1851         spin_unlock(&swap_lock);
1852
1853         name = getname(specialfile);
1854         error = PTR_ERR(name);
1855         if (IS_ERR(name)) {
1856                 name = NULL;
1857                 goto bad_swap_2;
1858         }
1859         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1860         error = PTR_ERR(swap_file);
1861         if (IS_ERR(swap_file)) {
1862                 swap_file = NULL;
1863                 goto bad_swap_2;
1864         }
1865
1866         p->swap_file = swap_file;
1867         mapping = swap_file->f_mapping;
1868         inode = mapping->host;
1869
1870         error = -EBUSY;
1871         for (i = 0; i < nr_swapfiles; i++) {
1872                 struct swap_info_struct *q = swap_info[i];
1873
1874                 if (i == type || !q->swap_file)
1875                         continue;
1876                 if (mapping == q->swap_file->f_mapping)
1877                         goto bad_swap;
1878         }
1879
1880         error = -EINVAL;
1881         if (S_ISBLK(inode->i_mode)) {
1882                 bdev = I_BDEV(inode);
1883                 error = bd_claim(bdev, sys_swapon);
1884                 if (error < 0) {
1885                         bdev = NULL;
1886                         error = -EINVAL;
1887                         goto bad_swap;
1888                 }
1889                 p->old_block_size = block_size(bdev);
1890                 error = set_blocksize(bdev, PAGE_SIZE);
1891                 if (error < 0)
1892                         goto bad_swap;
1893                 p->bdev = bdev;
1894                 p->flags |= SWP_BLKDEV;
1895         } else if (S_ISREG(inode->i_mode)) {
1896                 p->bdev = inode->i_sb->s_bdev;
1897                 mutex_lock(&inode->i_mutex);
1898                 did_down = 1;
1899                 if (IS_SWAPFILE(inode)) {
1900                         error = -EBUSY;
1901                         goto bad_swap;
1902                 }
1903         } else {
1904                 goto bad_swap;
1905         }
1906
1907         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1908
1909         /*
1910          * Read the swap header.
1911          */
1912         if (!mapping->a_ops->readpage) {
1913                 error = -EINVAL;
1914                 goto bad_swap;
1915         }
1916         page = read_mapping_page(mapping, 0, swap_file);
1917         if (IS_ERR(page)) {
1918                 error = PTR_ERR(page);
1919                 goto bad_swap;
1920         }
1921         swap_header = kmap(page);
1922
1923         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1924                 printk(KERN_ERR "Unable to find swap-space signature\n");
1925                 error = -EINVAL;
1926                 goto bad_swap;
1927         }
1928
1929         /* swap partition endianess hack... */
1930         if (swab32(swap_header->info.version) == 1) {
1931                 swab32s(&swap_header->info.version);
1932                 swab32s(&swap_header->info.last_page);
1933                 swab32s(&swap_header->info.nr_badpages);
1934                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1935                         swab32s(&swap_header->info.badpages[i]);
1936         }
1937         /* Check the swap header's sub-version */
1938         if (swap_header->info.version != 1) {
1939                 printk(KERN_WARNING
1940                        "Unable to handle swap header version %d\n",
1941                        swap_header->info.version);
1942                 error = -EINVAL;
1943                 goto bad_swap;
1944         }
1945
1946         p->lowest_bit  = 1;
1947         p->cluster_next = 1;
1948         p->cluster_nr = 0;
1949
1950         /*
1951          * Find out how many pages are allowed for a single swap
1952          * device. There are two limiting factors: 1) the number of
1953          * bits for the swap offset in the swp_entry_t type and
1954          * 2) the number of bits in the a swap pte as defined by
1955          * the different architectures. In order to find the
1956          * largest possible bit mask a swap entry with swap type 0
1957          * and swap offset ~0UL is created, encoded to a swap pte,
1958          * decoded to a swp_entry_t again and finally the swap
1959          * offset is extracted. This will mask all the bits from
1960          * the initial ~0UL mask that can't be encoded in either
1961          * the swp_entry_t or the architecture definition of a
1962          * swap pte.
1963          */
1964         maxpages = swp_offset(pte_to_swp_entry(
1965                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1966         if (maxpages > swap_header->info.last_page) {
1967                 maxpages = swap_header->info.last_page + 1;
1968                 /* p->max is an unsigned int: don't overflow it */
1969                 if ((unsigned int)maxpages == 0)
1970                         maxpages = UINT_MAX;
1971         }
1972         p->highest_bit = maxpages - 1;
1973
1974         error = -EINVAL;
1975         if (!maxpages)
1976                 goto bad_swap;
1977         if (swapfilepages && maxpages > swapfilepages) {
1978                 printk(KERN_WARNING
1979                        "Swap area shorter than signature indicates\n");
1980                 goto bad_swap;
1981         }
1982         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1983                 goto bad_swap;
1984         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1985                 goto bad_swap;
1986
1987         /* OK, set up the swap map and apply the bad block list */
1988         swap_map = vmalloc(maxpages);
1989         if (!swap_map) {
1990                 error = -ENOMEM;
1991                 goto bad_swap;
1992         }
1993
1994         memset(swap_map, 0, maxpages);
1995         nr_good_pages = maxpages - 1;   /* omit header page */
1996
1997         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1998                 unsigned int page_nr = swap_header->info.badpages[i];
1999                 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2000                         error = -EINVAL;
2001                         goto bad_swap;
2002                 }
2003                 if (page_nr < maxpages) {
2004                         swap_map[page_nr] = SWAP_MAP_BAD;
2005                         nr_good_pages--;
2006                 }
2007         }
2008
2009         error = swap_cgroup_swapon(type, maxpages);
2010         if (error)
2011                 goto bad_swap;
2012
2013         if (nr_good_pages) {
2014                 swap_map[0] = SWAP_MAP_BAD;
2015                 p->max = maxpages;
2016                 p->pages = nr_good_pages;
2017                 nr_extents = setup_swap_extents(p, &span);
2018                 if (nr_extents < 0) {
2019                         error = nr_extents;
2020                         goto bad_swap;
2021                 }
2022                 nr_good_pages = p->pages;
2023         }
2024         if (!nr_good_pages) {
2025                 printk(KERN_WARNING "Empty swap-file\n");
2026                 error = -EINVAL;
2027                 goto bad_swap;
2028         }
2029
2030         if (p->bdev) {
2031                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2032                         p->flags |= SWP_SOLIDSTATE;
2033                         p->cluster_next = 1 + (random32() % p->highest_bit);
2034                 }
2035                 if (discard_swap(p) == 0)
2036                         p->flags |= SWP_DISCARDABLE;
2037         }
2038
2039         mutex_lock(&swapon_mutex);
2040         spin_lock(&swap_lock);
2041         if (swap_flags & SWAP_FLAG_PREFER)
2042                 p->prio =
2043                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2044         else
2045                 p->prio = --least_priority;
2046         p->swap_map = swap_map;
2047         p->flags |= SWP_WRITEOK;
2048         nr_swap_pages += nr_good_pages;
2049         total_swap_pages += nr_good_pages;
2050
2051         printk(KERN_INFO "Adding %uk swap on %s.  "
2052                         "Priority:%d extents:%d across:%lluk %s%s\n",
2053                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2054                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2055                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2056                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2057
2058         /* insert swap space into swap_list: */
2059         prev = -1;
2060         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2061                 if (p->prio >= swap_info[i]->prio)
2062                         break;
2063                 prev = i;
2064         }
2065         p->next = i;
2066         if (prev < 0)
2067                 swap_list.head = swap_list.next = type;
2068         else
2069                 swap_info[prev]->next = type;
2070         spin_unlock(&swap_lock);
2071         mutex_unlock(&swapon_mutex);
2072         error = 0;
2073         goto out;
2074 bad_swap:
2075         if (bdev) {
2076                 set_blocksize(bdev, p->old_block_size);
2077                 bd_release(bdev);
2078         }
2079         destroy_swap_extents(p);
2080         swap_cgroup_swapoff(type);
2081 bad_swap_2:
2082         spin_lock(&swap_lock);
2083         p->swap_file = NULL;
2084         p->flags = 0;
2085         spin_unlock(&swap_lock);
2086         vfree(swap_map);
2087         if (swap_file)
2088                 filp_close(swap_file, NULL);
2089 out:
2090         if (page && !IS_ERR(page)) {
2091                 kunmap(page);
2092                 page_cache_release(page);
2093         }
2094         if (name)
2095                 putname(name);
2096         if (did_down) {
2097                 if (!error)
2098                         inode->i_flags |= S_SWAPFILE;
2099                 mutex_unlock(&inode->i_mutex);
2100         }
2101         return error;
2102 }
2103
2104 void si_swapinfo(struct sysinfo *val)
2105 {
2106         unsigned int type;
2107         unsigned long nr_to_be_unused = 0;
2108
2109         spin_lock(&swap_lock);
2110         for (type = 0; type < nr_swapfiles; type++) {
2111                 struct swap_info_struct *si = swap_info[type];
2112
2113                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2114                         nr_to_be_unused += si->inuse_pages;
2115         }
2116         val->freeswap = nr_swap_pages + nr_to_be_unused;
2117         val->totalswap = total_swap_pages + nr_to_be_unused;
2118         spin_unlock(&swap_lock);
2119 }
2120
2121 /*
2122  * Verify that a swap entry is valid and increment its swap map count.
2123  *
2124  * Returns error code in following case.
2125  * - success -> 0
2126  * - swp_entry is invalid -> EINVAL
2127  * - swp_entry is migration entry -> EINVAL
2128  * - swap-cache reference is requested but there is already one. -> EEXIST
2129  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2130  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2131  */
2132 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2133 {
2134         struct swap_info_struct *p;
2135         unsigned long offset, type;
2136         unsigned char count;
2137         unsigned char has_cache;
2138         int err = -EINVAL;
2139
2140         if (non_swap_entry(entry))
2141                 goto out;
2142
2143         type = swp_type(entry);
2144         if (type >= nr_swapfiles)
2145                 goto bad_file;
2146         p = swap_info[type];
2147         offset = swp_offset(entry);
2148
2149         spin_lock(&swap_lock);
2150         if (unlikely(offset >= p->max))
2151                 goto unlock_out;
2152
2153         count = p->swap_map[offset];
2154         has_cache = count & SWAP_HAS_CACHE;
2155         count &= ~SWAP_HAS_CACHE;
2156         err = 0;
2157
2158         if (usage == SWAP_HAS_CACHE) {
2159
2160                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2161                 if (!has_cache && count)
2162                         has_cache = SWAP_HAS_CACHE;
2163                 else if (has_cache)             /* someone else added cache */
2164                         err = -EEXIST;
2165                 else                            /* no users remaining */
2166                         err = -ENOENT;
2167
2168         } else if (count || has_cache) {
2169
2170                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2171                         count += usage;
2172                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2173                         err = -EINVAL;
2174                 else if (swap_count_continued(p, offset, count))
2175                         count = COUNT_CONTINUED;
2176                 else
2177                         err = -ENOMEM;
2178         } else
2179                 err = -ENOENT;                  /* unused swap entry */
2180
2181         p->swap_map[offset] = count | has_cache;
2182
2183 unlock_out:
2184         spin_unlock(&swap_lock);
2185 out:
2186         return err;
2187
2188 bad_file:
2189         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2190         goto out;
2191 }
2192
2193 /*
2194  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2195  * (in which case its reference count is never incremented).
2196  */
2197 void swap_shmem_alloc(swp_entry_t entry)
2198 {
2199         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2200 }
2201
2202 /*
2203  * Increase reference count of swap entry by 1.
2204  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2205  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2206  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2207  * might occur if a page table entry has got corrupted.
2208  */
2209 int swap_duplicate(swp_entry_t entry)
2210 {
2211         int err = 0;
2212
2213         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2214                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2215         return err;
2216 }
2217
2218 /*
2219  * @entry: swap entry for which we allocate swap cache.
2220  *
2221  * Called when allocating swap cache for existing swap entry,
2222  * This can return error codes. Returns 0 at success.
2223  * -EBUSY means there is a swap cache.
2224  * Note: return code is different from swap_duplicate().
2225  */
2226 int swapcache_prepare(swp_entry_t entry)
2227 {
2228         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2229 }
2230
2231 /*
2232  * swap_lock prevents swap_map being freed. Don't grab an extra
2233  * reference on the swaphandle, it doesn't matter if it becomes unused.
2234  */
2235 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2236 {
2237         struct swap_info_struct *si;
2238         int our_page_cluster = page_cluster;
2239         pgoff_t target, toff;
2240         pgoff_t base, end;
2241         int nr_pages = 0;
2242
2243         if (!our_page_cluster)  /* no readahead */
2244                 return 0;
2245
2246         si = swap_info[swp_type(entry)];
2247         target = swp_offset(entry);
2248         base = (target >> our_page_cluster) << our_page_cluster;
2249         end = base + (1 << our_page_cluster);
2250         if (!base)              /* first page is swap header */
2251                 base++;
2252
2253         spin_lock(&swap_lock);
2254         if (end > si->max)      /* don't go beyond end of map */
2255                 end = si->max;
2256
2257         /* Count contiguous allocated slots above our target */
2258         for (toff = target; ++toff < end; nr_pages++) {
2259                 /* Don't read in free or bad pages */
2260                 if (!si->swap_map[toff])
2261                         break;
2262                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2263                         break;
2264         }
2265         /* Count contiguous allocated slots below our target */
2266         for (toff = target; --toff >= base; nr_pages++) {
2267                 /* Don't read in free or bad pages */
2268                 if (!si->swap_map[toff])
2269                         break;
2270                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2271                         break;
2272         }
2273         spin_unlock(&swap_lock);
2274
2275         /*
2276          * Indicate starting offset, and return number of pages to get:
2277          * if only 1, say 0, since there's then no readahead to be done.
2278          */
2279         *offset = ++toff;
2280         return nr_pages? ++nr_pages: 0;
2281 }
2282
2283 /*
2284  * add_swap_count_continuation - called when a swap count is duplicated
2285  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2286  * page of the original vmalloc'ed swap_map, to hold the continuation count
2287  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2288  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2289  *
2290  * These continuation pages are seldom referenced: the common paths all work
2291  * on the original swap_map, only referring to a continuation page when the
2292  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2293  *
2294  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2295  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2296  * can be called after dropping locks.
2297  */
2298 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2299 {
2300         struct swap_info_struct *si;
2301         struct page *head;
2302         struct page *page;
2303         struct page *list_page;
2304         pgoff_t offset;
2305         unsigned char count;
2306
2307         /*
2308          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2309          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2310          */
2311         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2312
2313         si = swap_info_get(entry);
2314         if (!si) {
2315                 /*
2316                  * An acceptable race has occurred since the failing
2317                  * __swap_duplicate(): the swap entry has been freed,
2318                  * perhaps even the whole swap_map cleared for swapoff.
2319                  */
2320                 goto outer;
2321         }
2322
2323         offset = swp_offset(entry);
2324         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2325
2326         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2327                 /*
2328                  * The higher the swap count, the more likely it is that tasks
2329                  * will race to add swap count continuation: we need to avoid
2330                  * over-provisioning.
2331                  */
2332                 goto out;
2333         }
2334
2335         if (!page) {
2336                 spin_unlock(&swap_lock);
2337                 return -ENOMEM;
2338         }
2339
2340         /*
2341          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2342          * no architecture is using highmem pages for kernel pagetables: so it
2343          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2344          */
2345         head = vmalloc_to_page(si->swap_map + offset);
2346         offset &= ~PAGE_MASK;
2347
2348         /*
2349          * Page allocation does not initialize the page's lru field,
2350          * but it does always reset its private field.
2351          */
2352         if (!page_private(head)) {
2353                 BUG_ON(count & COUNT_CONTINUED);
2354                 INIT_LIST_HEAD(&head->lru);
2355                 set_page_private(head, SWP_CONTINUED);
2356                 si->flags |= SWP_CONTINUED;
2357         }
2358
2359         list_for_each_entry(list_page, &head->lru, lru) {
2360                 unsigned char *map;
2361
2362                 /*
2363                  * If the previous map said no continuation, but we've found
2364                  * a continuation page, free our allocation and use this one.
2365                  */
2366                 if (!(count & COUNT_CONTINUED))
2367                         goto out;
2368
2369                 map = kmap_atomic(list_page, KM_USER0) + offset;
2370                 count = *map;
2371                 kunmap_atomic(map, KM_USER0);
2372
2373                 /*
2374                  * If this continuation count now has some space in it,
2375                  * free our allocation and use this one.
2376                  */
2377                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2378                         goto out;
2379         }
2380
2381         list_add_tail(&page->lru, &head->lru);
2382         page = NULL;                    /* now it's attached, don't free it */
2383 out:
2384         spin_unlock(&swap_lock);
2385 outer:
2386         if (page)
2387                 __free_page(page);
2388         return 0;
2389 }
2390
2391 /*
2392  * swap_count_continued - when the original swap_map count is incremented
2393  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2394  * into, carry if so, or else fail until a new continuation page is allocated;
2395  * when the original swap_map count is decremented from 0 with continuation,
2396  * borrow from the continuation and report whether it still holds more.
2397  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2398  */
2399 static bool swap_count_continued(struct swap_info_struct *si,
2400                                  pgoff_t offset, unsigned char count)
2401 {
2402         struct page *head;
2403         struct page *page;
2404         unsigned char *map;
2405
2406         head = vmalloc_to_page(si->swap_map + offset);
2407         if (page_private(head) != SWP_CONTINUED) {
2408                 BUG_ON(count & COUNT_CONTINUED);
2409                 return false;           /* need to add count continuation */
2410         }
2411
2412         offset &= ~PAGE_MASK;
2413         page = list_entry(head->lru.next, struct page, lru);
2414         map = kmap_atomic(page, KM_USER0) + offset;
2415
2416         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2417                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2418
2419         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2420                 /*
2421                  * Think of how you add 1 to 999
2422                  */
2423                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2424                         kunmap_atomic(map, KM_USER0);
2425                         page = list_entry(page->lru.next, struct page, lru);
2426                         BUG_ON(page == head);
2427                         map = kmap_atomic(page, KM_USER0) + offset;
2428                 }
2429                 if (*map == SWAP_CONT_MAX) {
2430                         kunmap_atomic(map, KM_USER0);
2431                         page = list_entry(page->lru.next, struct page, lru);
2432                         if (page == head)
2433                                 return false;   /* add count continuation */
2434                         map = kmap_atomic(page, KM_USER0) + offset;
2435 init_map:               *map = 0;               /* we didn't zero the page */
2436                 }
2437                 *map += 1;
2438                 kunmap_atomic(map, KM_USER0);
2439                 page = list_entry(page->lru.prev, struct page, lru);
2440                 while (page != head) {
2441                         map = kmap_atomic(page, KM_USER0) + offset;
2442                         *map = COUNT_CONTINUED;
2443                         kunmap_atomic(map, KM_USER0);
2444                         page = list_entry(page->lru.prev, struct page, lru);
2445                 }
2446                 return true;                    /* incremented */
2447
2448         } else {                                /* decrementing */
2449                 /*
2450                  * Think of how you subtract 1 from 1000
2451                  */
2452                 BUG_ON(count != COUNT_CONTINUED);
2453                 while (*map == COUNT_CONTINUED) {
2454                         kunmap_atomic(map, KM_USER0);
2455                         page = list_entry(page->lru.next, struct page, lru);
2456                         BUG_ON(page == head);
2457                         map = kmap_atomic(page, KM_USER0) + offset;
2458                 }
2459                 BUG_ON(*map == 0);
2460                 *map -= 1;
2461                 if (*map == 0)
2462                         count = 0;
2463                 kunmap_atomic(map, KM_USER0);
2464                 page = list_entry(page->lru.prev, struct page, lru);
2465                 while (page != head) {
2466                         map = kmap_atomic(page, KM_USER0) + offset;
2467                         *map = SWAP_CONT_MAX | count;
2468                         count = COUNT_CONTINUED;
2469                         kunmap_atomic(map, KM_USER0);
2470                         page = list_entry(page->lru.prev, struct page, lru);
2471                 }
2472                 return count == COUNT_CONTINUED;
2473         }
2474 }
2475
2476 /*
2477  * free_swap_count_continuations - swapoff free all the continuation pages
2478  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2479  */
2480 static void free_swap_count_continuations(struct swap_info_struct *si)
2481 {
2482         pgoff_t offset;
2483
2484         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2485                 struct page *head;
2486                 head = vmalloc_to_page(si->swap_map + offset);
2487                 if (page_private(head)) {
2488                         struct list_head *this, *next;
2489                         list_for_each_safe(this, next, &head->lru) {
2490                                 struct page *page;
2491                                 page = list_entry(this, struct page, lru);
2492                                 list_del(this);
2493                                 __free_page(page);
2494                         }
2495                 }
2496         }
2497 }