mm: arch: rename in_gate_area_no_task to in_gate_area_no_mm
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/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 #include <linux/poll.h>
34
35 #include <asm/pgtable.h>
36 #include <asm/tlbflush.h>
37 #include <linux/swapops.h>
38 #include <linux/page_cgroup.h>
39
40 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
41                                  unsigned char);
42 static void free_swap_count_continuations(struct swap_info_struct *);
43 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
44
45 static DEFINE_SPINLOCK(swap_lock);
46 static unsigned int nr_swapfiles;
47 long nr_swap_pages;
48 long total_swap_pages;
49 static int least_priority;
50
51 static const char Bad_file[] = "Bad swap file entry ";
52 static const char Unused_file[] = "Unused swap file entry ";
53 static const char Bad_offset[] = "Bad swap offset entry ";
54 static const char Unused_offset[] = "Unused swap offset entry ";
55
56 static struct swap_list_t swap_list = {-1, -1};
57
58 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
59
60 static DEFINE_MUTEX(swapon_mutex);
61
62 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
63 /* Activity counter to indicate that a swapon or swapoff has occurred */
64 static atomic_t proc_poll_event = ATOMIC_INIT(0);
65
66 static inline unsigned char swap_count(unsigned char ent)
67 {
68         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
69 }
70
71 /* returns 1 if swap entry is freed */
72 static int
73 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
74 {
75         swp_entry_t entry = swp_entry(si->type, offset);
76         struct page *page;
77         int ret = 0;
78
79         page = find_get_page(&swapper_space, entry.val);
80         if (!page)
81                 return 0;
82         /*
83          * This function is called from scan_swap_map() and it's called
84          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
85          * We have to use trylock for avoiding deadlock. This is a special
86          * case and you should use try_to_free_swap() with explicit lock_page()
87          * in usual operations.
88          */
89         if (trylock_page(page)) {
90                 ret = try_to_free_swap(page);
91                 unlock_page(page);
92         }
93         page_cache_release(page);
94         return ret;
95 }
96
97 /*
98  * We need this because the bdev->unplug_fn can sleep and we cannot
99  * hold swap_lock while calling the unplug_fn. And swap_lock
100  * cannot be turned into a mutex.
101  */
102 static DECLARE_RWSEM(swap_unplug_sem);
103
104 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
105 {
106         swp_entry_t entry;
107
108         down_read(&swap_unplug_sem);
109         entry.val = page_private(page);
110         if (PageSwapCache(page)) {
111                 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
112                 struct backing_dev_info *bdi;
113
114                 /*
115                  * If the page is removed from swapcache from under us (with a
116                  * racy try_to_unuse/swapoff) we need an additional reference
117                  * count to avoid reading garbage from page_private(page) above.
118                  * If the WARN_ON triggers during a swapoff it maybe the race
119                  * condition and it's harmless. However if it triggers without
120                  * swapoff it signals a problem.
121                  */
122                 WARN_ON(page_count(page) <= 1);
123
124                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
125                 blk_run_backing_dev(bdi, page);
126         }
127         up_read(&swap_unplug_sem);
128 }
129
130 /*
131  * swapon tell device that all the old swap contents can be discarded,
132  * to allow the swap device to optimize its wear-levelling.
133  */
134 static int discard_swap(struct swap_info_struct *si)
135 {
136         struct swap_extent *se;
137         sector_t start_block;
138         sector_t nr_blocks;
139         int err = 0;
140
141         /* Do not discard the swap header page! */
142         se = &si->first_swap_extent;
143         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
144         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
145         if (nr_blocks) {
146                 err = blkdev_issue_discard(si->bdev, start_block,
147                                 nr_blocks, GFP_KERNEL, 0);
148                 if (err)
149                         return err;
150                 cond_resched();
151         }
152
153         list_for_each_entry(se, &si->first_swap_extent.list, list) {
154                 start_block = se->start_block << (PAGE_SHIFT - 9);
155                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
156
157                 err = blkdev_issue_discard(si->bdev, start_block,
158                                 nr_blocks, GFP_KERNEL, 0);
159                 if (err)
160                         break;
161
162                 cond_resched();
163         }
164         return err;             /* That will often be -EOPNOTSUPP */
165 }
166
167 /*
168  * swap allocation tell device that a cluster of swap can now be discarded,
169  * to allow the swap device to optimize its wear-levelling.
170  */
171 static void discard_swap_cluster(struct swap_info_struct *si,
172                                  pgoff_t start_page, pgoff_t nr_pages)
173 {
174         struct swap_extent *se = si->curr_swap_extent;
175         int found_extent = 0;
176
177         while (nr_pages) {
178                 struct list_head *lh;
179
180                 if (se->start_page <= start_page &&
181                     start_page < se->start_page + se->nr_pages) {
182                         pgoff_t offset = start_page - se->start_page;
183                         sector_t start_block = se->start_block + offset;
184                         sector_t nr_blocks = se->nr_pages - offset;
185
186                         if (nr_blocks > nr_pages)
187                                 nr_blocks = nr_pages;
188                         start_page += nr_blocks;
189                         nr_pages -= nr_blocks;
190
191                         if (!found_extent++)
192                                 si->curr_swap_extent = se;
193
194                         start_block <<= PAGE_SHIFT - 9;
195                         nr_blocks <<= PAGE_SHIFT - 9;
196                         if (blkdev_issue_discard(si->bdev, start_block,
197                                     nr_blocks, GFP_NOIO, 0))
198                                 break;
199                 }
200
201                 lh = se->list.next;
202                 se = list_entry(lh, struct swap_extent, list);
203         }
204 }
205
206 static int wait_for_discard(void *word)
207 {
208         schedule();
209         return 0;
210 }
211
212 #define SWAPFILE_CLUSTER        256
213 #define LATENCY_LIMIT           256
214
215 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
216                                           unsigned char usage)
217 {
218         unsigned long offset;
219         unsigned long scan_base;
220         unsigned long last_in_cluster = 0;
221         int latency_ration = LATENCY_LIMIT;
222         int found_free_cluster = 0;
223
224         /*
225          * We try to cluster swap pages by allocating them sequentially
226          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
227          * way, however, we resort to first-free allocation, starting
228          * a new cluster.  This prevents us from scattering swap pages
229          * all over the entire swap partition, so that we reduce
230          * overall disk seek times between swap pages.  -- sct
231          * But we do now try to find an empty cluster.  -Andrea
232          * And we let swap pages go all over an SSD partition.  Hugh
233          */
234
235         si->flags += SWP_SCANNING;
236         scan_base = offset = si->cluster_next;
237
238         if (unlikely(!si->cluster_nr--)) {
239                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
240                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
241                         goto checks;
242                 }
243                 if (si->flags & SWP_DISCARDABLE) {
244                         /*
245                          * Start range check on racing allocations, in case
246                          * they overlap the cluster we eventually decide on
247                          * (we scan without swap_lock to allow preemption).
248                          * It's hardly conceivable that cluster_nr could be
249                          * wrapped during our scan, but don't depend on it.
250                          */
251                         if (si->lowest_alloc)
252                                 goto checks;
253                         si->lowest_alloc = si->max;
254                         si->highest_alloc = 0;
255                 }
256                 spin_unlock(&swap_lock);
257
258                 /*
259                  * If seek is expensive, start searching for new cluster from
260                  * start of partition, to minimize the span of allocated swap.
261                  * But if seek is cheap, search from our current position, so
262                  * that swap is allocated from all over the partition: if the
263                  * Flash Translation Layer only remaps within limited zones,
264                  * we don't want to wear out the first zone too quickly.
265                  */
266                 if (!(si->flags & SWP_SOLIDSTATE))
267                         scan_base = offset = si->lowest_bit;
268                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
269
270                 /* Locate the first empty (unaligned) cluster */
271                 for (; last_in_cluster <= si->highest_bit; offset++) {
272                         if (si->swap_map[offset])
273                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
274                         else if (offset == last_in_cluster) {
275                                 spin_lock(&swap_lock);
276                                 offset -= SWAPFILE_CLUSTER - 1;
277                                 si->cluster_next = offset;
278                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
279                                 found_free_cluster = 1;
280                                 goto checks;
281                         }
282                         if (unlikely(--latency_ration < 0)) {
283                                 cond_resched();
284                                 latency_ration = LATENCY_LIMIT;
285                         }
286                 }
287
288                 offset = si->lowest_bit;
289                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
290
291                 /* Locate the first empty (unaligned) cluster */
292                 for (; last_in_cluster < scan_base; offset++) {
293                         if (si->swap_map[offset])
294                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
295                         else if (offset == last_in_cluster) {
296                                 spin_lock(&swap_lock);
297                                 offset -= SWAPFILE_CLUSTER - 1;
298                                 si->cluster_next = offset;
299                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
300                                 found_free_cluster = 1;
301                                 goto checks;
302                         }
303                         if (unlikely(--latency_ration < 0)) {
304                                 cond_resched();
305                                 latency_ration = LATENCY_LIMIT;
306                         }
307                 }
308
309                 offset = scan_base;
310                 spin_lock(&swap_lock);
311                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
312                 si->lowest_alloc = 0;
313         }
314
315 checks:
316         if (!(si->flags & SWP_WRITEOK))
317                 goto no_page;
318         if (!si->highest_bit)
319                 goto no_page;
320         if (offset > si->highest_bit)
321                 scan_base = offset = si->lowest_bit;
322
323         /* reuse swap entry of cache-only swap if not busy. */
324         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
325                 int swap_was_freed;
326                 spin_unlock(&swap_lock);
327                 swap_was_freed = __try_to_reclaim_swap(si, offset);
328                 spin_lock(&swap_lock);
329                 /* entry was freed successfully, try to use this again */
330                 if (swap_was_freed)
331                         goto checks;
332                 goto scan; /* check next one */
333         }
334
335         if (si->swap_map[offset])
336                 goto scan;
337
338         if (offset == si->lowest_bit)
339                 si->lowest_bit++;
340         if (offset == si->highest_bit)
341                 si->highest_bit--;
342         si->inuse_pages++;
343         if (si->inuse_pages == si->pages) {
344                 si->lowest_bit = si->max;
345                 si->highest_bit = 0;
346         }
347         si->swap_map[offset] = usage;
348         si->cluster_next = offset + 1;
349         si->flags -= SWP_SCANNING;
350
351         if (si->lowest_alloc) {
352                 /*
353                  * Only set when SWP_DISCARDABLE, and there's a scan
354                  * for a free cluster in progress or just completed.
355                  */
356                 if (found_free_cluster) {
357                         /*
358                          * To optimize wear-levelling, discard the
359                          * old data of the cluster, taking care not to
360                          * discard any of its pages that have already
361                          * been allocated by racing tasks (offset has
362                          * already stepped over any at the beginning).
363                          */
364                         if (offset < si->highest_alloc &&
365                             si->lowest_alloc <= last_in_cluster)
366                                 last_in_cluster = si->lowest_alloc - 1;
367                         si->flags |= SWP_DISCARDING;
368                         spin_unlock(&swap_lock);
369
370                         if (offset < last_in_cluster)
371                                 discard_swap_cluster(si, offset,
372                                         last_in_cluster - offset + 1);
373
374                         spin_lock(&swap_lock);
375                         si->lowest_alloc = 0;
376                         si->flags &= ~SWP_DISCARDING;
377
378                         smp_mb();       /* wake_up_bit advises this */
379                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
380
381                 } else if (si->flags & SWP_DISCARDING) {
382                         /*
383                          * Delay using pages allocated by racing tasks
384                          * until the whole discard has been issued. We
385                          * could defer that delay until swap_writepage,
386                          * but it's easier to keep this self-contained.
387                          */
388                         spin_unlock(&swap_lock);
389                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
390                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
391                         spin_lock(&swap_lock);
392                 } else {
393                         /*
394                          * Note pages allocated by racing tasks while
395                          * scan for a free cluster is in progress, so
396                          * that its final discard can exclude them.
397                          */
398                         if (offset < si->lowest_alloc)
399                                 si->lowest_alloc = offset;
400                         if (offset > si->highest_alloc)
401                                 si->highest_alloc = offset;
402                 }
403         }
404         return offset;
405
406 scan:
407         spin_unlock(&swap_lock);
408         while (++offset <= si->highest_bit) {
409                 if (!si->swap_map[offset]) {
410                         spin_lock(&swap_lock);
411                         goto checks;
412                 }
413                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
414                         spin_lock(&swap_lock);
415                         goto checks;
416                 }
417                 if (unlikely(--latency_ration < 0)) {
418                         cond_resched();
419                         latency_ration = LATENCY_LIMIT;
420                 }
421         }
422         offset = si->lowest_bit;
423         while (++offset < scan_base) {
424                 if (!si->swap_map[offset]) {
425                         spin_lock(&swap_lock);
426                         goto checks;
427                 }
428                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
429                         spin_lock(&swap_lock);
430                         goto checks;
431                 }
432                 if (unlikely(--latency_ration < 0)) {
433                         cond_resched();
434                         latency_ration = LATENCY_LIMIT;
435                 }
436         }
437         spin_lock(&swap_lock);
438
439 no_page:
440         si->flags -= SWP_SCANNING;
441         return 0;
442 }
443
444 swp_entry_t get_swap_page(void)
445 {
446         struct swap_info_struct *si;
447         pgoff_t offset;
448         int type, next;
449         int wrapped = 0;
450
451         spin_lock(&swap_lock);
452         if (nr_swap_pages <= 0)
453                 goto noswap;
454         nr_swap_pages--;
455
456         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
457                 si = swap_info[type];
458                 next = si->next;
459                 if (next < 0 ||
460                     (!wrapped && si->prio != swap_info[next]->prio)) {
461                         next = swap_list.head;
462                         wrapped++;
463                 }
464
465                 if (!si->highest_bit)
466                         continue;
467                 if (!(si->flags & SWP_WRITEOK))
468                         continue;
469
470                 swap_list.next = next;
471                 /* This is called for allocating swap entry for cache */
472                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
473                 if (offset) {
474                         spin_unlock(&swap_lock);
475                         return swp_entry(type, offset);
476                 }
477                 next = swap_list.next;
478         }
479
480         nr_swap_pages++;
481 noswap:
482         spin_unlock(&swap_lock);
483         return (swp_entry_t) {0};
484 }
485
486 /* The only caller of this function is now susupend routine */
487 swp_entry_t get_swap_page_of_type(int type)
488 {
489         struct swap_info_struct *si;
490         pgoff_t offset;
491
492         spin_lock(&swap_lock);
493         si = swap_info[type];
494         if (si && (si->flags & SWP_WRITEOK)) {
495                 nr_swap_pages--;
496                 /* This is called for allocating swap entry, not cache */
497                 offset = scan_swap_map(si, 1);
498                 if (offset) {
499                         spin_unlock(&swap_lock);
500                         return swp_entry(type, offset);
501                 }
502                 nr_swap_pages++;
503         }
504         spin_unlock(&swap_lock);
505         return (swp_entry_t) {0};
506 }
507
508 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
509 {
510         struct swap_info_struct *p;
511         unsigned long offset, type;
512
513         if (!entry.val)
514                 goto out;
515         type = swp_type(entry);
516         if (type >= nr_swapfiles)
517                 goto bad_nofile;
518         p = swap_info[type];
519         if (!(p->flags & SWP_USED))
520                 goto bad_device;
521         offset = swp_offset(entry);
522         if (offset >= p->max)
523                 goto bad_offset;
524         if (!p->swap_map[offset])
525                 goto bad_free;
526         spin_lock(&swap_lock);
527         return p;
528
529 bad_free:
530         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
531         goto out;
532 bad_offset:
533         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
534         goto out;
535 bad_device:
536         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
537         goto out;
538 bad_nofile:
539         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
540 out:
541         return NULL;
542 }
543
544 static unsigned char swap_entry_free(struct swap_info_struct *p,
545                                      swp_entry_t entry, unsigned char usage)
546 {
547         unsigned long offset = swp_offset(entry);
548         unsigned char count;
549         unsigned char has_cache;
550
551         count = p->swap_map[offset];
552         has_cache = count & SWAP_HAS_CACHE;
553         count &= ~SWAP_HAS_CACHE;
554
555         if (usage == SWAP_HAS_CACHE) {
556                 VM_BUG_ON(!has_cache);
557                 has_cache = 0;
558         } else if (count == SWAP_MAP_SHMEM) {
559                 /*
560                  * Or we could insist on shmem.c using a special
561                  * swap_shmem_free() and free_shmem_swap_and_cache()...
562                  */
563                 count = 0;
564         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
565                 if (count == COUNT_CONTINUED) {
566                         if (swap_count_continued(p, offset, count))
567                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
568                         else
569                                 count = SWAP_MAP_MAX;
570                 } else
571                         count--;
572         }
573
574         if (!count)
575                 mem_cgroup_uncharge_swap(entry);
576
577         usage = count | has_cache;
578         p->swap_map[offset] = usage;
579
580         /* free if no reference */
581         if (!usage) {
582                 struct gendisk *disk = p->bdev->bd_disk;
583                 if (offset < p->lowest_bit)
584                         p->lowest_bit = offset;
585                 if (offset > p->highest_bit)
586                         p->highest_bit = offset;
587                 if (swap_list.next >= 0 &&
588                     p->prio > swap_info[swap_list.next]->prio)
589                         swap_list.next = p->type;
590                 nr_swap_pages++;
591                 p->inuse_pages--;
592                 if ((p->flags & SWP_BLKDEV) &&
593                                 disk->fops->swap_slot_free_notify)
594                         disk->fops->swap_slot_free_notify(p->bdev, offset);
595         }
596
597         return usage;
598 }
599
600 /*
601  * Caller has made sure that the swapdevice corresponding to entry
602  * is still around or has not been recycled.
603  */
604 void swap_free(swp_entry_t entry)
605 {
606         struct swap_info_struct *p;
607
608         p = swap_info_get(entry);
609         if (p) {
610                 swap_entry_free(p, entry, 1);
611                 spin_unlock(&swap_lock);
612         }
613 }
614
615 /*
616  * Called after dropping swapcache to decrease refcnt to swap entries.
617  */
618 void swapcache_free(swp_entry_t entry, struct page *page)
619 {
620         struct swap_info_struct *p;
621         unsigned char count;
622
623         p = swap_info_get(entry);
624         if (p) {
625                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
626                 if (page)
627                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
628                 spin_unlock(&swap_lock);
629         }
630 }
631
632 /*
633  * How many references to page are currently swapped out?
634  * This does not give an exact answer when swap count is continued,
635  * but does include the high COUNT_CONTINUED flag to allow for that.
636  */
637 static inline int page_swapcount(struct page *page)
638 {
639         int count = 0;
640         struct swap_info_struct *p;
641         swp_entry_t entry;
642
643         entry.val = page_private(page);
644         p = swap_info_get(entry);
645         if (p) {
646                 count = swap_count(p->swap_map[swp_offset(entry)]);
647                 spin_unlock(&swap_lock);
648         }
649         return count;
650 }
651
652 /*
653  * We can write to an anon page without COW if there are no other references
654  * to it.  And as a side-effect, free up its swap: because the old content
655  * on disk will never be read, and seeking back there to write new content
656  * later would only waste time away from clustering.
657  */
658 int reuse_swap_page(struct page *page)
659 {
660         int count;
661
662         VM_BUG_ON(!PageLocked(page));
663         if (unlikely(PageKsm(page)))
664                 return 0;
665         count = page_mapcount(page);
666         if (count <= 1 && PageSwapCache(page)) {
667                 count += page_swapcount(page);
668                 if (count == 1 && !PageWriteback(page)) {
669                         delete_from_swap_cache(page);
670                         SetPageDirty(page);
671                 }
672         }
673         return count <= 1;
674 }
675
676 /*
677  * If swap is getting full, or if there are no more mappings of this page,
678  * then try_to_free_swap is called to free its swap space.
679  */
680 int try_to_free_swap(struct page *page)
681 {
682         VM_BUG_ON(!PageLocked(page));
683
684         if (!PageSwapCache(page))
685                 return 0;
686         if (PageWriteback(page))
687                 return 0;
688         if (page_swapcount(page))
689                 return 0;
690
691         /*
692          * Once hibernation has begun to create its image of memory,
693          * there's a danger that one of the calls to try_to_free_swap()
694          * - most probably a call from __try_to_reclaim_swap() while
695          * hibernation is allocating its own swap pages for the image,
696          * but conceivably even a call from memory reclaim - will free
697          * the swap from a page which has already been recorded in the
698          * image as a clean swapcache page, and then reuse its swap for
699          * another page of the image.  On waking from hibernation, the
700          * original page might be freed under memory pressure, then
701          * later read back in from swap, now with the wrong data.
702          *
703          * Hibernation clears bits from gfp_allowed_mask to prevent
704          * memory reclaim from writing to disk, so check that here.
705          */
706         if (!(gfp_allowed_mask & __GFP_IO))
707                 return 0;
708
709         delete_from_swap_cache(page);
710         SetPageDirty(page);
711         return 1;
712 }
713
714 /*
715  * Free the swap entry like above, but also try to
716  * free the page cache entry if it is the last user.
717  */
718 int free_swap_and_cache(swp_entry_t entry)
719 {
720         struct swap_info_struct *p;
721         struct page *page = NULL;
722
723         if (non_swap_entry(entry))
724                 return 1;
725
726         p = swap_info_get(entry);
727         if (p) {
728                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
729                         page = find_get_page(&swapper_space, entry.val);
730                         if (page && !trylock_page(page)) {
731                                 page_cache_release(page);
732                                 page = NULL;
733                         }
734                 }
735                 spin_unlock(&swap_lock);
736         }
737         if (page) {
738                 /*
739                  * Not mapped elsewhere, or swap space full? Free it!
740                  * Also recheck PageSwapCache now page is locked (above).
741                  */
742                 if (PageSwapCache(page) && !PageWriteback(page) &&
743                                 (!page_mapped(page) || vm_swap_full())) {
744                         delete_from_swap_cache(page);
745                         SetPageDirty(page);
746                 }
747                 unlock_page(page);
748                 page_cache_release(page);
749         }
750         return p != NULL;
751 }
752
753 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
754 /**
755  * mem_cgroup_count_swap_user - count the user of a swap entry
756  * @ent: the swap entry to be checked
757  * @pagep: the pointer for the swap cache page of the entry to be stored
758  *
759  * Returns the number of the user of the swap entry. The number is valid only
760  * for swaps of anonymous pages.
761  * If the entry is found on swap cache, the page is stored to pagep with
762  * refcount of it being incremented.
763  */
764 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
765 {
766         struct page *page;
767         struct swap_info_struct *p;
768         int count = 0;
769
770         page = find_get_page(&swapper_space, ent.val);
771         if (page)
772                 count += page_mapcount(page);
773         p = swap_info_get(ent);
774         if (p) {
775                 count += swap_count(p->swap_map[swp_offset(ent)]);
776                 spin_unlock(&swap_lock);
777         }
778
779         *pagep = page;
780         return count;
781 }
782 #endif
783
784 #ifdef CONFIG_HIBERNATION
785 /*
786  * Find the swap type that corresponds to given device (if any).
787  *
788  * @offset - number of the PAGE_SIZE-sized block of the device, starting
789  * from 0, in which the swap header is expected to be located.
790  *
791  * This is needed for the suspend to disk (aka swsusp).
792  */
793 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
794 {
795         struct block_device *bdev = NULL;
796         int type;
797
798         if (device)
799                 bdev = bdget(device);
800
801         spin_lock(&swap_lock);
802         for (type = 0; type < nr_swapfiles; type++) {
803                 struct swap_info_struct *sis = swap_info[type];
804
805                 if (!(sis->flags & SWP_WRITEOK))
806                         continue;
807
808                 if (!bdev) {
809                         if (bdev_p)
810                                 *bdev_p = bdgrab(sis->bdev);
811
812                         spin_unlock(&swap_lock);
813                         return type;
814                 }
815                 if (bdev == sis->bdev) {
816                         struct swap_extent *se = &sis->first_swap_extent;
817
818                         if (se->start_block == offset) {
819                                 if (bdev_p)
820                                         *bdev_p = bdgrab(sis->bdev);
821
822                                 spin_unlock(&swap_lock);
823                                 bdput(bdev);
824                                 return type;
825                         }
826                 }
827         }
828         spin_unlock(&swap_lock);
829         if (bdev)
830                 bdput(bdev);
831
832         return -ENODEV;
833 }
834
835 /*
836  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
837  * corresponding to given index in swap_info (swap type).
838  */
839 sector_t swapdev_block(int type, pgoff_t offset)
840 {
841         struct block_device *bdev;
842
843         if ((unsigned int)type >= nr_swapfiles)
844                 return 0;
845         if (!(swap_info[type]->flags & SWP_WRITEOK))
846                 return 0;
847         return map_swap_entry(swp_entry(type, offset), &bdev);
848 }
849
850 /*
851  * Return either the total number of swap pages of given type, or the number
852  * of free pages of that type (depending on @free)
853  *
854  * This is needed for software suspend
855  */
856 unsigned int count_swap_pages(int type, int free)
857 {
858         unsigned int n = 0;
859
860         spin_lock(&swap_lock);
861         if ((unsigned int)type < nr_swapfiles) {
862                 struct swap_info_struct *sis = swap_info[type];
863
864                 if (sis->flags & SWP_WRITEOK) {
865                         n = sis->pages;
866                         if (free)
867                                 n -= sis->inuse_pages;
868                 }
869         }
870         spin_unlock(&swap_lock);
871         return n;
872 }
873 #endif /* CONFIG_HIBERNATION */
874
875 /*
876  * No need to decide whether this PTE shares the swap entry with others,
877  * just let do_wp_page work it out if a write is requested later - to
878  * force COW, vm_page_prot omits write permission from any private vma.
879  */
880 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
881                 unsigned long addr, swp_entry_t entry, struct page *page)
882 {
883         struct mem_cgroup *ptr = NULL;
884         spinlock_t *ptl;
885         pte_t *pte;
886         int ret = 1;
887
888         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
889                 ret = -ENOMEM;
890                 goto out_nolock;
891         }
892
893         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
894         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
895                 if (ret > 0)
896                         mem_cgroup_cancel_charge_swapin(ptr);
897                 ret = 0;
898                 goto out;
899         }
900
901         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
902         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
903         get_page(page);
904         set_pte_at(vma->vm_mm, addr, pte,
905                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
906         page_add_anon_rmap(page, vma, addr);
907         mem_cgroup_commit_charge_swapin(page, ptr);
908         swap_free(entry);
909         /*
910          * Move the page to the active list so it is not
911          * immediately swapped out again after swapon.
912          */
913         activate_page(page);
914 out:
915         pte_unmap_unlock(pte, ptl);
916 out_nolock:
917         return ret;
918 }
919
920 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
921                                 unsigned long addr, unsigned long end,
922                                 swp_entry_t entry, struct page *page)
923 {
924         pte_t swp_pte = swp_entry_to_pte(entry);
925         pte_t *pte;
926         int ret = 0;
927
928         /*
929          * We don't actually need pte lock while scanning for swp_pte: since
930          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
931          * page table while we're scanning; though it could get zapped, and on
932          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
933          * of unmatched parts which look like swp_pte, so unuse_pte must
934          * recheck under pte lock.  Scanning without pte lock lets it be
935          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
936          */
937         pte = pte_offset_map(pmd, addr);
938         do {
939                 /*
940                  * swapoff spends a _lot_ of time in this loop!
941                  * Test inline before going to call unuse_pte.
942                  */
943                 if (unlikely(pte_same(*pte, swp_pte))) {
944                         pte_unmap(pte);
945                         ret = unuse_pte(vma, pmd, addr, entry, page);
946                         if (ret)
947                                 goto out;
948                         pte = pte_offset_map(pmd, addr);
949                 }
950         } while (pte++, addr += PAGE_SIZE, addr != end);
951         pte_unmap(pte - 1);
952 out:
953         return ret;
954 }
955
956 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
957                                 unsigned long addr, unsigned long end,
958                                 swp_entry_t entry, struct page *page)
959 {
960         pmd_t *pmd;
961         unsigned long next;
962         int ret;
963
964         pmd = pmd_offset(pud, addr);
965         do {
966                 next = pmd_addr_end(addr, end);
967                 if (unlikely(pmd_trans_huge(*pmd)))
968                         continue;
969                 if (pmd_none_or_clear_bad(pmd))
970                         continue;
971                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
972                 if (ret)
973                         return ret;
974         } while (pmd++, addr = next, addr != end);
975         return 0;
976 }
977
978 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
979                                 unsigned long addr, unsigned long end,
980                                 swp_entry_t entry, struct page *page)
981 {
982         pud_t *pud;
983         unsigned long next;
984         int ret;
985
986         pud = pud_offset(pgd, addr);
987         do {
988                 next = pud_addr_end(addr, end);
989                 if (pud_none_or_clear_bad(pud))
990                         continue;
991                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
992                 if (ret)
993                         return ret;
994         } while (pud++, addr = next, addr != end);
995         return 0;
996 }
997
998 static int unuse_vma(struct vm_area_struct *vma,
999                                 swp_entry_t entry, struct page *page)
1000 {
1001         pgd_t *pgd;
1002         unsigned long addr, end, next;
1003         int ret;
1004
1005         if (page_anon_vma(page)) {
1006                 addr = page_address_in_vma(page, vma);
1007                 if (addr == -EFAULT)
1008                         return 0;
1009                 else
1010                         end = addr + PAGE_SIZE;
1011         } else {
1012                 addr = vma->vm_start;
1013                 end = vma->vm_end;
1014         }
1015
1016         pgd = pgd_offset(vma->vm_mm, addr);
1017         do {
1018                 next = pgd_addr_end(addr, end);
1019                 if (pgd_none_or_clear_bad(pgd))
1020                         continue;
1021                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1022                 if (ret)
1023                         return ret;
1024         } while (pgd++, addr = next, addr != end);
1025         return 0;
1026 }
1027
1028 static int unuse_mm(struct mm_struct *mm,
1029                                 swp_entry_t entry, struct page *page)
1030 {
1031         struct vm_area_struct *vma;
1032         int ret = 0;
1033
1034         if (!down_read_trylock(&mm->mmap_sem)) {
1035                 /*
1036                  * Activate page so shrink_inactive_list is unlikely to unmap
1037                  * its ptes while lock is dropped, so swapoff can make progress.
1038                  */
1039                 activate_page(page);
1040                 unlock_page(page);
1041                 down_read(&mm->mmap_sem);
1042                 lock_page(page);
1043         }
1044         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1045                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1046                         break;
1047         }
1048         up_read(&mm->mmap_sem);
1049         return (ret < 0)? ret: 0;
1050 }
1051
1052 /*
1053  * Scan swap_map from current position to next entry still in use.
1054  * Recycle to start on reaching the end, returning 0 when empty.
1055  */
1056 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1057                                         unsigned int prev)
1058 {
1059         unsigned int max = si->max;
1060         unsigned int i = prev;
1061         unsigned char count;
1062
1063         /*
1064          * No need for swap_lock here: we're just looking
1065          * for whether an entry is in use, not modifying it; false
1066          * hits are okay, and sys_swapoff() has already prevented new
1067          * allocations from this area (while holding swap_lock).
1068          */
1069         for (;;) {
1070                 if (++i >= max) {
1071                         if (!prev) {
1072                                 i = 0;
1073                                 break;
1074                         }
1075                         /*
1076                          * No entries in use at top of swap_map,
1077                          * loop back to start and recheck there.
1078                          */
1079                         max = prev + 1;
1080                         prev = 0;
1081                         i = 1;
1082                 }
1083                 count = si->swap_map[i];
1084                 if (count && swap_count(count) != SWAP_MAP_BAD)
1085                         break;
1086         }
1087         return i;
1088 }
1089
1090 /*
1091  * We completely avoid races by reading each swap page in advance,
1092  * and then search for the process using it.  All the necessary
1093  * page table adjustments can then be made atomically.
1094  */
1095 static int try_to_unuse(unsigned int type)
1096 {
1097         struct swap_info_struct *si = swap_info[type];
1098         struct mm_struct *start_mm;
1099         unsigned char *swap_map;
1100         unsigned char swcount;
1101         struct page *page;
1102         swp_entry_t entry;
1103         unsigned int i = 0;
1104         int retval = 0;
1105
1106         /*
1107          * When searching mms for an entry, a good strategy is to
1108          * start at the first mm we freed the previous entry from
1109          * (though actually we don't notice whether we or coincidence
1110          * freed the entry).  Initialize this start_mm with a hold.
1111          *
1112          * A simpler strategy would be to start at the last mm we
1113          * freed the previous entry from; but that would take less
1114          * advantage of mmlist ordering, which clusters forked mms
1115          * together, child after parent.  If we race with dup_mmap(), we
1116          * prefer to resolve parent before child, lest we miss entries
1117          * duplicated after we scanned child: using last mm would invert
1118          * that.
1119          */
1120         start_mm = &init_mm;
1121         atomic_inc(&init_mm.mm_users);
1122
1123         /*
1124          * Keep on scanning until all entries have gone.  Usually,
1125          * one pass through swap_map is enough, but not necessarily:
1126          * there are races when an instance of an entry might be missed.
1127          */
1128         while ((i = find_next_to_unuse(si, i)) != 0) {
1129                 if (signal_pending(current)) {
1130                         retval = -EINTR;
1131                         break;
1132                 }
1133
1134                 /*
1135                  * Get a page for the entry, using the existing swap
1136                  * cache page if there is one.  Otherwise, get a clean
1137                  * page and read the swap into it.
1138                  */
1139                 swap_map = &si->swap_map[i];
1140                 entry = swp_entry(type, i);
1141                 page = read_swap_cache_async(entry,
1142                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1143                 if (!page) {
1144                         /*
1145                          * Either swap_duplicate() failed because entry
1146                          * has been freed independently, and will not be
1147                          * reused since sys_swapoff() already disabled
1148                          * allocation from here, or alloc_page() failed.
1149                          */
1150                         if (!*swap_map)
1151                                 continue;
1152                         retval = -ENOMEM;
1153                         break;
1154                 }
1155
1156                 /*
1157                  * Don't hold on to start_mm if it looks like exiting.
1158                  */
1159                 if (atomic_read(&start_mm->mm_users) == 1) {
1160                         mmput(start_mm);
1161                         start_mm = &init_mm;
1162                         atomic_inc(&init_mm.mm_users);
1163                 }
1164
1165                 /*
1166                  * Wait for and lock page.  When do_swap_page races with
1167                  * try_to_unuse, do_swap_page can handle the fault much
1168                  * faster than try_to_unuse can locate the entry.  This
1169                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1170                  * defer to do_swap_page in such a case - in some tests,
1171                  * do_swap_page and try_to_unuse repeatedly compete.
1172                  */
1173                 wait_on_page_locked(page);
1174                 wait_on_page_writeback(page);
1175                 lock_page(page);
1176                 wait_on_page_writeback(page);
1177
1178                 /*
1179                  * Remove all references to entry.
1180                  */
1181                 swcount = *swap_map;
1182                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1183                         retval = shmem_unuse(entry, page);
1184                         /* page has already been unlocked and released */
1185                         if (retval < 0)
1186                                 break;
1187                         continue;
1188                 }
1189                 if (swap_count(swcount) && start_mm != &init_mm)
1190                         retval = unuse_mm(start_mm, entry, page);
1191
1192                 if (swap_count(*swap_map)) {
1193                         int set_start_mm = (*swap_map >= swcount);
1194                         struct list_head *p = &start_mm->mmlist;
1195                         struct mm_struct *new_start_mm = start_mm;
1196                         struct mm_struct *prev_mm = start_mm;
1197                         struct mm_struct *mm;
1198
1199                         atomic_inc(&new_start_mm->mm_users);
1200                         atomic_inc(&prev_mm->mm_users);
1201                         spin_lock(&mmlist_lock);
1202                         while (swap_count(*swap_map) && !retval &&
1203                                         (p = p->next) != &start_mm->mmlist) {
1204                                 mm = list_entry(p, struct mm_struct, mmlist);
1205                                 if (!atomic_inc_not_zero(&mm->mm_users))
1206                                         continue;
1207                                 spin_unlock(&mmlist_lock);
1208                                 mmput(prev_mm);
1209                                 prev_mm = mm;
1210
1211                                 cond_resched();
1212
1213                                 swcount = *swap_map;
1214                                 if (!swap_count(swcount)) /* any usage ? */
1215                                         ;
1216                                 else if (mm == &init_mm)
1217                                         set_start_mm = 1;
1218                                 else
1219                                         retval = unuse_mm(mm, entry, page);
1220
1221                                 if (set_start_mm && *swap_map < swcount) {
1222                                         mmput(new_start_mm);
1223                                         atomic_inc(&mm->mm_users);
1224                                         new_start_mm = mm;
1225                                         set_start_mm = 0;
1226                                 }
1227                                 spin_lock(&mmlist_lock);
1228                         }
1229                         spin_unlock(&mmlist_lock);
1230                         mmput(prev_mm);
1231                         mmput(start_mm);
1232                         start_mm = new_start_mm;
1233                 }
1234                 if (retval) {
1235                         unlock_page(page);
1236                         page_cache_release(page);
1237                         break;
1238                 }
1239
1240                 /*
1241                  * If a reference remains (rare), we would like to leave
1242                  * the page in the swap cache; but try_to_unmap could
1243                  * then re-duplicate the entry once we drop page lock,
1244                  * so we might loop indefinitely; also, that page could
1245                  * not be swapped out to other storage meanwhile.  So:
1246                  * delete from cache even if there's another reference,
1247                  * after ensuring that the data has been saved to disk -
1248                  * since if the reference remains (rarer), it will be
1249                  * read from disk into another page.  Splitting into two
1250                  * pages would be incorrect if swap supported "shared
1251                  * private" pages, but they are handled by tmpfs files.
1252                  *
1253                  * Given how unuse_vma() targets one particular offset
1254                  * in an anon_vma, once the anon_vma has been determined,
1255                  * this splitting happens to be just what is needed to
1256                  * handle where KSM pages have been swapped out: re-reading
1257                  * is unnecessarily slow, but we can fix that later on.
1258                  */
1259                 if (swap_count(*swap_map) &&
1260                      PageDirty(page) && PageSwapCache(page)) {
1261                         struct writeback_control wbc = {
1262                                 .sync_mode = WB_SYNC_NONE,
1263                         };
1264
1265                         swap_writepage(page, &wbc);
1266                         lock_page(page);
1267                         wait_on_page_writeback(page);
1268                 }
1269
1270                 /*
1271                  * It is conceivable that a racing task removed this page from
1272                  * swap cache just before we acquired the page lock at the top,
1273                  * or while we dropped it in unuse_mm().  The page might even
1274                  * be back in swap cache on another swap area: that we must not
1275                  * delete, since it may not have been written out to swap yet.
1276                  */
1277                 if (PageSwapCache(page) &&
1278                     likely(page_private(page) == entry.val))
1279                         delete_from_swap_cache(page);
1280
1281                 /*
1282                  * So we could skip searching mms once swap count went
1283                  * to 1, we did not mark any present ptes as dirty: must
1284                  * mark page dirty so shrink_page_list will preserve it.
1285                  */
1286                 SetPageDirty(page);
1287                 unlock_page(page);
1288                 page_cache_release(page);
1289
1290                 /*
1291                  * Make sure that we aren't completely killing
1292                  * interactive performance.
1293                  */
1294                 cond_resched();
1295         }
1296
1297         mmput(start_mm);
1298         return retval;
1299 }
1300
1301 /*
1302  * After a successful try_to_unuse, if no swap is now in use, we know
1303  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1304  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1305  * added to the mmlist just after page_duplicate - before would be racy.
1306  */
1307 static void drain_mmlist(void)
1308 {
1309         struct list_head *p, *next;
1310         unsigned int type;
1311
1312         for (type = 0; type < nr_swapfiles; type++)
1313                 if (swap_info[type]->inuse_pages)
1314                         return;
1315         spin_lock(&mmlist_lock);
1316         list_for_each_safe(p, next, &init_mm.mmlist)
1317                 list_del_init(p);
1318         spin_unlock(&mmlist_lock);
1319 }
1320
1321 /*
1322  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1323  * corresponds to page offset for the specified swap entry.
1324  * Note that the type of this function is sector_t, but it returns page offset
1325  * into the bdev, not sector offset.
1326  */
1327 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1328 {
1329         struct swap_info_struct *sis;
1330         struct swap_extent *start_se;
1331         struct swap_extent *se;
1332         pgoff_t offset;
1333
1334         sis = swap_info[swp_type(entry)];
1335         *bdev = sis->bdev;
1336
1337         offset = swp_offset(entry);
1338         start_se = sis->curr_swap_extent;
1339         se = start_se;
1340
1341         for ( ; ; ) {
1342                 struct list_head *lh;
1343
1344                 if (se->start_page <= offset &&
1345                                 offset < (se->start_page + se->nr_pages)) {
1346                         return se->start_block + (offset - se->start_page);
1347                 }
1348                 lh = se->list.next;
1349                 se = list_entry(lh, struct swap_extent, list);
1350                 sis->curr_swap_extent = se;
1351                 BUG_ON(se == start_se);         /* It *must* be present */
1352         }
1353 }
1354
1355 /*
1356  * Returns the page offset into bdev for the specified page's swap entry.
1357  */
1358 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1359 {
1360         swp_entry_t entry;
1361         entry.val = page_private(page);
1362         return map_swap_entry(entry, bdev);
1363 }
1364
1365 /*
1366  * Free all of a swapdev's extent information
1367  */
1368 static void destroy_swap_extents(struct swap_info_struct *sis)
1369 {
1370         while (!list_empty(&sis->first_swap_extent.list)) {
1371                 struct swap_extent *se;
1372
1373                 se = list_entry(sis->first_swap_extent.list.next,
1374                                 struct swap_extent, list);
1375                 list_del(&se->list);
1376                 kfree(se);
1377         }
1378 }
1379
1380 /*
1381  * Add a block range (and the corresponding page range) into this swapdev's
1382  * extent list.  The extent list is kept sorted in page order.
1383  *
1384  * This function rather assumes that it is called in ascending page order.
1385  */
1386 static int
1387 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1388                 unsigned long nr_pages, sector_t start_block)
1389 {
1390         struct swap_extent *se;
1391         struct swap_extent *new_se;
1392         struct list_head *lh;
1393
1394         if (start_page == 0) {
1395                 se = &sis->first_swap_extent;
1396                 sis->curr_swap_extent = se;
1397                 se->start_page = 0;
1398                 se->nr_pages = nr_pages;
1399                 se->start_block = start_block;
1400                 return 1;
1401         } else {
1402                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1403                 se = list_entry(lh, struct swap_extent, list);
1404                 BUG_ON(se->start_page + se->nr_pages != start_page);
1405                 if (se->start_block + se->nr_pages == start_block) {
1406                         /* Merge it */
1407                         se->nr_pages += nr_pages;
1408                         return 0;
1409                 }
1410         }
1411
1412         /*
1413          * No merge.  Insert a new extent, preserving ordering.
1414          */
1415         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1416         if (new_se == NULL)
1417                 return -ENOMEM;
1418         new_se->start_page = start_page;
1419         new_se->nr_pages = nr_pages;
1420         new_se->start_block = start_block;
1421
1422         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1423         return 1;
1424 }
1425
1426 /*
1427  * A `swap extent' is a simple thing which maps a contiguous range of pages
1428  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1429  * is built at swapon time and is then used at swap_writepage/swap_readpage
1430  * time for locating where on disk a page belongs.
1431  *
1432  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1433  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1434  * swap files identically.
1435  *
1436  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1437  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1438  * swapfiles are handled *identically* after swapon time.
1439  *
1440  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1441  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1442  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1443  * requirements, they are simply tossed out - we will never use those blocks
1444  * for swapping.
1445  *
1446  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1447  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1448  * which will scribble on the fs.
1449  *
1450  * The amount of disk space which a single swap extent represents varies.
1451  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1452  * extents in the list.  To avoid much list walking, we cache the previous
1453  * search location in `curr_swap_extent', and start new searches from there.
1454  * This is extremely effective.  The average number of iterations in
1455  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1456  */
1457 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1458 {
1459         struct inode *inode;
1460         unsigned blocks_per_page;
1461         unsigned long page_no;
1462         unsigned blkbits;
1463         sector_t probe_block;
1464         sector_t last_block;
1465         sector_t lowest_block = -1;
1466         sector_t highest_block = 0;
1467         int nr_extents = 0;
1468         int ret;
1469
1470         inode = sis->swap_file->f_mapping->host;
1471         if (S_ISBLK(inode->i_mode)) {
1472                 ret = add_swap_extent(sis, 0, sis->max, 0);
1473                 *span = sis->pages;
1474                 goto out;
1475         }
1476
1477         blkbits = inode->i_blkbits;
1478         blocks_per_page = PAGE_SIZE >> blkbits;
1479
1480         /*
1481          * Map all the blocks into the extent list.  This code doesn't try
1482          * to be very smart.
1483          */
1484         probe_block = 0;
1485         page_no = 0;
1486         last_block = i_size_read(inode) >> blkbits;
1487         while ((probe_block + blocks_per_page) <= last_block &&
1488                         page_no < sis->max) {
1489                 unsigned block_in_page;
1490                 sector_t first_block;
1491
1492                 first_block = bmap(inode, probe_block);
1493                 if (first_block == 0)
1494                         goto bad_bmap;
1495
1496                 /*
1497                  * It must be PAGE_SIZE aligned on-disk
1498                  */
1499                 if (first_block & (blocks_per_page - 1)) {
1500                         probe_block++;
1501                         goto reprobe;
1502                 }
1503
1504                 for (block_in_page = 1; block_in_page < blocks_per_page;
1505                                         block_in_page++) {
1506                         sector_t block;
1507
1508                         block = bmap(inode, probe_block + block_in_page);
1509                         if (block == 0)
1510                                 goto bad_bmap;
1511                         if (block != first_block + block_in_page) {
1512                                 /* Discontiguity */
1513                                 probe_block++;
1514                                 goto reprobe;
1515                         }
1516                 }
1517
1518                 first_block >>= (PAGE_SHIFT - blkbits);
1519                 if (page_no) {  /* exclude the header page */
1520                         if (first_block < lowest_block)
1521                                 lowest_block = first_block;
1522                         if (first_block > highest_block)
1523                                 highest_block = first_block;
1524                 }
1525
1526                 /*
1527                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1528                  */
1529                 ret = add_swap_extent(sis, page_no, 1, first_block);
1530                 if (ret < 0)
1531                         goto out;
1532                 nr_extents += ret;
1533                 page_no++;
1534                 probe_block += blocks_per_page;
1535 reprobe:
1536                 continue;
1537         }
1538         ret = nr_extents;
1539         *span = 1 + highest_block - lowest_block;
1540         if (page_no == 0)
1541                 page_no = 1;    /* force Empty message */
1542         sis->max = page_no;
1543         sis->pages = page_no - 1;
1544         sis->highest_bit = page_no - 1;
1545 out:
1546         return ret;
1547 bad_bmap:
1548         printk(KERN_ERR "swapon: swapfile has holes\n");
1549         ret = -EINVAL;
1550         goto out;
1551 }
1552
1553 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1554 {
1555         struct swap_info_struct *p = NULL;
1556         unsigned char *swap_map;
1557         struct file *swap_file, *victim;
1558         struct address_space *mapping;
1559         struct inode *inode;
1560         char *pathname;
1561         int i, type, prev;
1562         int err;
1563
1564         if (!capable(CAP_SYS_ADMIN))
1565                 return -EPERM;
1566
1567         pathname = getname(specialfile);
1568         err = PTR_ERR(pathname);
1569         if (IS_ERR(pathname))
1570                 goto out;
1571
1572         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1573         putname(pathname);
1574         err = PTR_ERR(victim);
1575         if (IS_ERR(victim))
1576                 goto out;
1577
1578         mapping = victim->f_mapping;
1579         prev = -1;
1580         spin_lock(&swap_lock);
1581         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1582                 p = swap_info[type];
1583                 if (p->flags & SWP_WRITEOK) {
1584                         if (p->swap_file->f_mapping == mapping)
1585                                 break;
1586                 }
1587                 prev = type;
1588         }
1589         if (type < 0) {
1590                 err = -EINVAL;
1591                 spin_unlock(&swap_lock);
1592                 goto out_dput;
1593         }
1594         if (!security_vm_enough_memory(p->pages))
1595                 vm_unacct_memory(p->pages);
1596         else {
1597                 err = -ENOMEM;
1598                 spin_unlock(&swap_lock);
1599                 goto out_dput;
1600         }
1601         if (prev < 0)
1602                 swap_list.head = p->next;
1603         else
1604                 swap_info[prev]->next = p->next;
1605         if (type == swap_list.next) {
1606                 /* just pick something that's safe... */
1607                 swap_list.next = swap_list.head;
1608         }
1609         if (p->prio < 0) {
1610                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1611                         swap_info[i]->prio = p->prio--;
1612                 least_priority++;
1613         }
1614         nr_swap_pages -= p->pages;
1615         total_swap_pages -= p->pages;
1616         p->flags &= ~SWP_WRITEOK;
1617         spin_unlock(&swap_lock);
1618
1619         current->flags |= PF_OOM_ORIGIN;
1620         err = try_to_unuse(type);
1621         current->flags &= ~PF_OOM_ORIGIN;
1622
1623         if (err) {
1624                 /* re-insert swap space back into swap_list */
1625                 spin_lock(&swap_lock);
1626                 if (p->prio < 0)
1627                         p->prio = --least_priority;
1628                 prev = -1;
1629                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1630                         if (p->prio >= swap_info[i]->prio)
1631                                 break;
1632                         prev = i;
1633                 }
1634                 p->next = i;
1635                 if (prev < 0)
1636                         swap_list.head = swap_list.next = type;
1637                 else
1638                         swap_info[prev]->next = type;
1639                 nr_swap_pages += p->pages;
1640                 total_swap_pages += p->pages;
1641                 p->flags |= SWP_WRITEOK;
1642                 spin_unlock(&swap_lock);
1643                 goto out_dput;
1644         }
1645
1646         /* wait for any unplug function to finish */
1647         down_write(&swap_unplug_sem);
1648         up_write(&swap_unplug_sem);
1649
1650         destroy_swap_extents(p);
1651         if (p->flags & SWP_CONTINUED)
1652                 free_swap_count_continuations(p);
1653
1654         mutex_lock(&swapon_mutex);
1655         spin_lock(&swap_lock);
1656         drain_mmlist();
1657
1658         /* wait for anyone still in scan_swap_map */
1659         p->highest_bit = 0;             /* cuts scans short */
1660         while (p->flags >= SWP_SCANNING) {
1661                 spin_unlock(&swap_lock);
1662                 schedule_timeout_uninterruptible(1);
1663                 spin_lock(&swap_lock);
1664         }
1665
1666         swap_file = p->swap_file;
1667         p->swap_file = NULL;
1668         p->max = 0;
1669         swap_map = p->swap_map;
1670         p->swap_map = NULL;
1671         p->flags = 0;
1672         spin_unlock(&swap_lock);
1673         mutex_unlock(&swapon_mutex);
1674         vfree(swap_map);
1675         /* Destroy swap account informatin */
1676         swap_cgroup_swapoff(type);
1677
1678         inode = mapping->host;
1679         if (S_ISBLK(inode->i_mode)) {
1680                 struct block_device *bdev = I_BDEV(inode);
1681                 set_blocksize(bdev, p->old_block_size);
1682                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1683         } else {
1684                 mutex_lock(&inode->i_mutex);
1685                 inode->i_flags &= ~S_SWAPFILE;
1686                 mutex_unlock(&inode->i_mutex);
1687         }
1688         filp_close(swap_file, NULL);
1689         err = 0;
1690         atomic_inc(&proc_poll_event);
1691         wake_up_interruptible(&proc_poll_wait);
1692
1693 out_dput:
1694         filp_close(victim, NULL);
1695 out:
1696         return err;
1697 }
1698
1699 #ifdef CONFIG_PROC_FS
1700 struct proc_swaps {
1701         struct seq_file seq;
1702         int event;
1703 };
1704
1705 static unsigned swaps_poll(struct file *file, poll_table *wait)
1706 {
1707         struct proc_swaps *s = file->private_data;
1708
1709         poll_wait(file, &proc_poll_wait, wait);
1710
1711         if (s->event != atomic_read(&proc_poll_event)) {
1712                 s->event = atomic_read(&proc_poll_event);
1713                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1714         }
1715
1716         return POLLIN | POLLRDNORM;
1717 }
1718
1719 /* iterator */
1720 static void *swap_start(struct seq_file *swap, loff_t *pos)
1721 {
1722         struct swap_info_struct *si;
1723         int type;
1724         loff_t l = *pos;
1725
1726         mutex_lock(&swapon_mutex);
1727
1728         if (!l)
1729                 return SEQ_START_TOKEN;
1730
1731         for (type = 0; type < nr_swapfiles; type++) {
1732                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1733                 si = swap_info[type];
1734                 if (!(si->flags & SWP_USED) || !si->swap_map)
1735                         continue;
1736                 if (!--l)
1737                         return si;
1738         }
1739
1740         return NULL;
1741 }
1742
1743 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1744 {
1745         struct swap_info_struct *si = v;
1746         int type;
1747
1748         if (v == SEQ_START_TOKEN)
1749                 type = 0;
1750         else
1751                 type = si->type + 1;
1752
1753         for (; type < nr_swapfiles; type++) {
1754                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1755                 si = swap_info[type];
1756                 if (!(si->flags & SWP_USED) || !si->swap_map)
1757                         continue;
1758                 ++*pos;
1759                 return si;
1760         }
1761
1762         return NULL;
1763 }
1764
1765 static void swap_stop(struct seq_file *swap, void *v)
1766 {
1767         mutex_unlock(&swapon_mutex);
1768 }
1769
1770 static int swap_show(struct seq_file *swap, void *v)
1771 {
1772         struct swap_info_struct *si = v;
1773         struct file *file;
1774         int len;
1775
1776         if (si == SEQ_START_TOKEN) {
1777                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1778                 return 0;
1779         }
1780
1781         file = si->swap_file;
1782         len = seq_path(swap, &file->f_path, " \t\n\\");
1783         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1784                         len < 40 ? 40 - len : 1, " ",
1785                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1786                                 "partition" : "file\t",
1787                         si->pages << (PAGE_SHIFT - 10),
1788                         si->inuse_pages << (PAGE_SHIFT - 10),
1789                         si->prio);
1790         return 0;
1791 }
1792
1793 static const struct seq_operations swaps_op = {
1794         .start =        swap_start,
1795         .next =         swap_next,
1796         .stop =         swap_stop,
1797         .show =         swap_show
1798 };
1799
1800 static int swaps_open(struct inode *inode, struct file *file)
1801 {
1802         struct proc_swaps *s;
1803         int ret;
1804
1805         s = kmalloc(sizeof(struct proc_swaps), GFP_KERNEL);
1806         if (!s)
1807                 return -ENOMEM;
1808
1809         file->private_data = s;
1810
1811         ret = seq_open(file, &swaps_op);
1812         if (ret) {
1813                 kfree(s);
1814                 return ret;
1815         }
1816
1817         s->seq.private = s;
1818         s->event = atomic_read(&proc_poll_event);
1819         return ret;
1820 }
1821
1822 static const struct file_operations proc_swaps_operations = {
1823         .open           = swaps_open,
1824         .read           = seq_read,
1825         .llseek         = seq_lseek,
1826         .release        = seq_release,
1827         .poll           = swaps_poll,
1828 };
1829
1830 static int __init procswaps_init(void)
1831 {
1832         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1833         return 0;
1834 }
1835 __initcall(procswaps_init);
1836 #endif /* CONFIG_PROC_FS */
1837
1838 #ifdef MAX_SWAPFILES_CHECK
1839 static int __init max_swapfiles_check(void)
1840 {
1841         MAX_SWAPFILES_CHECK();
1842         return 0;
1843 }
1844 late_initcall(max_swapfiles_check);
1845 #endif
1846
1847 /*
1848  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1849  *
1850  * The swapon system call
1851  */
1852 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1853 {
1854         struct swap_info_struct *p;
1855         char *name = NULL;
1856         struct block_device *bdev = NULL;
1857         struct file *swap_file = NULL;
1858         struct address_space *mapping;
1859         unsigned int type;
1860         int i, prev;
1861         int error;
1862         union swap_header *swap_header;
1863         unsigned int nr_good_pages;
1864         int nr_extents = 0;
1865         sector_t span;
1866         unsigned long maxpages;
1867         unsigned long swapfilepages;
1868         unsigned char *swap_map = NULL;
1869         struct page *page = NULL;
1870         struct inode *inode = NULL;
1871         int did_down = 0;
1872
1873         if (!capable(CAP_SYS_ADMIN))
1874                 return -EPERM;
1875
1876         p = kzalloc(sizeof(*p), GFP_KERNEL);
1877         if (!p)
1878                 return -ENOMEM;
1879
1880         spin_lock(&swap_lock);
1881         for (type = 0; type < nr_swapfiles; type++) {
1882                 if (!(swap_info[type]->flags & SWP_USED))
1883                         break;
1884         }
1885         error = -EPERM;
1886         if (type >= MAX_SWAPFILES) {
1887                 spin_unlock(&swap_lock);
1888                 kfree(p);
1889                 goto out;
1890         }
1891         if (type >= nr_swapfiles) {
1892                 p->type = type;
1893                 swap_info[type] = p;
1894                 /*
1895                  * Write swap_info[type] before nr_swapfiles, in case a
1896                  * racing procfs swap_start() or swap_next() is reading them.
1897                  * (We never shrink nr_swapfiles, we never free this entry.)
1898                  */
1899                 smp_wmb();
1900                 nr_swapfiles++;
1901         } else {
1902                 kfree(p);
1903                 p = swap_info[type];
1904                 /*
1905                  * Do not memset this entry: a racing procfs swap_next()
1906                  * would be relying on p->type to remain valid.
1907                  */
1908         }
1909         INIT_LIST_HEAD(&p->first_swap_extent.list);
1910         p->flags = SWP_USED;
1911         p->next = -1;
1912         spin_unlock(&swap_lock);
1913
1914         name = getname(specialfile);
1915         error = PTR_ERR(name);
1916         if (IS_ERR(name)) {
1917                 name = NULL;
1918                 goto bad_swap_2;
1919         }
1920         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1921         error = PTR_ERR(swap_file);
1922         if (IS_ERR(swap_file)) {
1923                 swap_file = NULL;
1924                 goto bad_swap_2;
1925         }
1926
1927         p->swap_file = swap_file;
1928         mapping = swap_file->f_mapping;
1929         inode = mapping->host;
1930
1931         error = -EBUSY;
1932         for (i = 0; i < nr_swapfiles; i++) {
1933                 struct swap_info_struct *q = swap_info[i];
1934
1935                 if (i == type || !q->swap_file)
1936                         continue;
1937                 if (mapping == q->swap_file->f_mapping)
1938                         goto bad_swap;
1939         }
1940
1941         error = -EINVAL;
1942         if (S_ISBLK(inode->i_mode)) {
1943                 bdev = bdgrab(I_BDEV(inode));
1944                 error = blkdev_get(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1945                                    sys_swapon);
1946                 if (error < 0) {
1947                         bdev = NULL;
1948                         error = -EINVAL;
1949                         goto bad_swap;
1950                 }
1951                 p->old_block_size = block_size(bdev);
1952                 error = set_blocksize(bdev, PAGE_SIZE);
1953                 if (error < 0)
1954                         goto bad_swap;
1955                 p->bdev = bdev;
1956                 p->flags |= SWP_BLKDEV;
1957         } else if (S_ISREG(inode->i_mode)) {
1958                 p->bdev = inode->i_sb->s_bdev;
1959                 mutex_lock(&inode->i_mutex);
1960                 did_down = 1;
1961                 if (IS_SWAPFILE(inode)) {
1962                         error = -EBUSY;
1963                         goto bad_swap;
1964                 }
1965         } else {
1966                 goto bad_swap;
1967         }
1968
1969         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1970
1971         /*
1972          * Read the swap header.
1973          */
1974         if (!mapping->a_ops->readpage) {
1975                 error = -EINVAL;
1976                 goto bad_swap;
1977         }
1978         page = read_mapping_page(mapping, 0, swap_file);
1979         if (IS_ERR(page)) {
1980                 error = PTR_ERR(page);
1981                 goto bad_swap;
1982         }
1983         swap_header = kmap(page);
1984
1985         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1986                 printk(KERN_ERR "Unable to find swap-space signature\n");
1987                 error = -EINVAL;
1988                 goto bad_swap;
1989         }
1990
1991         /* swap partition endianess hack... */
1992         if (swab32(swap_header->info.version) == 1) {
1993                 swab32s(&swap_header->info.version);
1994                 swab32s(&swap_header->info.last_page);
1995                 swab32s(&swap_header->info.nr_badpages);
1996                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1997                         swab32s(&swap_header->info.badpages[i]);
1998         }
1999         /* Check the swap header's sub-version */
2000         if (swap_header->info.version != 1) {
2001                 printk(KERN_WARNING
2002                        "Unable to handle swap header version %d\n",
2003                        swap_header->info.version);
2004                 error = -EINVAL;
2005                 goto bad_swap;
2006         }
2007
2008         p->lowest_bit  = 1;
2009         p->cluster_next = 1;
2010         p->cluster_nr = 0;
2011
2012         /*
2013          * Find out how many pages are allowed for a single swap
2014          * device. There are two limiting factors: 1) the number of
2015          * bits for the swap offset in the swp_entry_t type and
2016          * 2) the number of bits in the a swap pte as defined by
2017          * the different architectures. In order to find the
2018          * largest possible bit mask a swap entry with swap type 0
2019          * and swap offset ~0UL is created, encoded to a swap pte,
2020          * decoded to a swp_entry_t again and finally the swap
2021          * offset is extracted. This will mask all the bits from
2022          * the initial ~0UL mask that can't be encoded in either
2023          * the swp_entry_t or the architecture definition of a
2024          * swap pte.
2025          */
2026         maxpages = swp_offset(pte_to_swp_entry(
2027                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2028         if (maxpages > swap_header->info.last_page) {
2029                 maxpages = swap_header->info.last_page + 1;
2030                 /* p->max is an unsigned int: don't overflow it */
2031                 if ((unsigned int)maxpages == 0)
2032                         maxpages = UINT_MAX;
2033         }
2034         p->highest_bit = maxpages - 1;
2035
2036         error = -EINVAL;
2037         if (!maxpages)
2038                 goto bad_swap;
2039         if (swapfilepages && maxpages > swapfilepages) {
2040                 printk(KERN_WARNING
2041                        "Swap area shorter than signature indicates\n");
2042                 goto bad_swap;
2043         }
2044         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2045                 goto bad_swap;
2046         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2047                 goto bad_swap;
2048
2049         /* OK, set up the swap map and apply the bad block list */
2050         swap_map = vmalloc(maxpages);
2051         if (!swap_map) {
2052                 error = -ENOMEM;
2053                 goto bad_swap;
2054         }
2055
2056         memset(swap_map, 0, maxpages);
2057         nr_good_pages = maxpages - 1;   /* omit header page */
2058
2059         for (i = 0; i < swap_header->info.nr_badpages; i++) {
2060                 unsigned int page_nr = swap_header->info.badpages[i];
2061                 if (page_nr == 0 || page_nr > swap_header->info.last_page) {
2062                         error = -EINVAL;
2063                         goto bad_swap;
2064                 }
2065                 if (page_nr < maxpages) {
2066                         swap_map[page_nr] = SWAP_MAP_BAD;
2067                         nr_good_pages--;
2068                 }
2069         }
2070
2071         error = swap_cgroup_swapon(type, maxpages);
2072         if (error)
2073                 goto bad_swap;
2074
2075         if (nr_good_pages) {
2076                 swap_map[0] = SWAP_MAP_BAD;
2077                 p->max = maxpages;
2078                 p->pages = nr_good_pages;
2079                 nr_extents = setup_swap_extents(p, &span);
2080                 if (nr_extents < 0) {
2081                         error = nr_extents;
2082                         goto bad_swap;
2083                 }
2084                 nr_good_pages = p->pages;
2085         }
2086         if (!nr_good_pages) {
2087                 printk(KERN_WARNING "Empty swap-file\n");
2088                 error = -EINVAL;
2089                 goto bad_swap;
2090         }
2091
2092         if (p->bdev) {
2093                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2094                         p->flags |= SWP_SOLIDSTATE;
2095                         p->cluster_next = 1 + (random32() % p->highest_bit);
2096                 }
2097                 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2098                         p->flags |= SWP_DISCARDABLE;
2099         }
2100
2101         mutex_lock(&swapon_mutex);
2102         spin_lock(&swap_lock);
2103         if (swap_flags & SWAP_FLAG_PREFER)
2104                 p->prio =
2105                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2106         else
2107                 p->prio = --least_priority;
2108         p->swap_map = swap_map;
2109         p->flags |= SWP_WRITEOK;
2110         nr_swap_pages += nr_good_pages;
2111         total_swap_pages += nr_good_pages;
2112
2113         printk(KERN_INFO "Adding %uk swap on %s.  "
2114                         "Priority:%d extents:%d across:%lluk %s%s\n",
2115                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2116                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2117                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2118                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2119
2120         /* insert swap space into swap_list: */
2121         prev = -1;
2122         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2123                 if (p->prio >= swap_info[i]->prio)
2124                         break;
2125                 prev = i;
2126         }
2127         p->next = i;
2128         if (prev < 0)
2129                 swap_list.head = swap_list.next = type;
2130         else
2131                 swap_info[prev]->next = type;
2132         spin_unlock(&swap_lock);
2133         mutex_unlock(&swapon_mutex);
2134         atomic_inc(&proc_poll_event);
2135         wake_up_interruptible(&proc_poll_wait);
2136
2137         error = 0;
2138         goto out;
2139 bad_swap:
2140         if (bdev) {
2141                 set_blocksize(bdev, p->old_block_size);
2142                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2143         }
2144         destroy_swap_extents(p);
2145         swap_cgroup_swapoff(type);
2146 bad_swap_2:
2147         spin_lock(&swap_lock);
2148         p->swap_file = NULL;
2149         p->flags = 0;
2150         spin_unlock(&swap_lock);
2151         vfree(swap_map);
2152         if (swap_file)
2153                 filp_close(swap_file, NULL);
2154 out:
2155         if (page && !IS_ERR(page)) {
2156                 kunmap(page);
2157                 page_cache_release(page);
2158         }
2159         if (name)
2160                 putname(name);
2161         if (did_down) {
2162                 if (!error)
2163                         inode->i_flags |= S_SWAPFILE;
2164                 mutex_unlock(&inode->i_mutex);
2165         }
2166         return error;
2167 }
2168
2169 void si_swapinfo(struct sysinfo *val)
2170 {
2171         unsigned int type;
2172         unsigned long nr_to_be_unused = 0;
2173
2174         spin_lock(&swap_lock);
2175         for (type = 0; type < nr_swapfiles; type++) {
2176                 struct swap_info_struct *si = swap_info[type];
2177
2178                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2179                         nr_to_be_unused += si->inuse_pages;
2180         }
2181         val->freeswap = nr_swap_pages + nr_to_be_unused;
2182         val->totalswap = total_swap_pages + nr_to_be_unused;
2183         spin_unlock(&swap_lock);
2184 }
2185
2186 /*
2187  * Verify that a swap entry is valid and increment its swap map count.
2188  *
2189  * Returns error code in following case.
2190  * - success -> 0
2191  * - swp_entry is invalid -> EINVAL
2192  * - swp_entry is migration entry -> EINVAL
2193  * - swap-cache reference is requested but there is already one. -> EEXIST
2194  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2195  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2196  */
2197 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2198 {
2199         struct swap_info_struct *p;
2200         unsigned long offset, type;
2201         unsigned char count;
2202         unsigned char has_cache;
2203         int err = -EINVAL;
2204
2205         if (non_swap_entry(entry))
2206                 goto out;
2207
2208         type = swp_type(entry);
2209         if (type >= nr_swapfiles)
2210                 goto bad_file;
2211         p = swap_info[type];
2212         offset = swp_offset(entry);
2213
2214         spin_lock(&swap_lock);
2215         if (unlikely(offset >= p->max))
2216                 goto unlock_out;
2217
2218         count = p->swap_map[offset];
2219         has_cache = count & SWAP_HAS_CACHE;
2220         count &= ~SWAP_HAS_CACHE;
2221         err = 0;
2222
2223         if (usage == SWAP_HAS_CACHE) {
2224
2225                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2226                 if (!has_cache && count)
2227                         has_cache = SWAP_HAS_CACHE;
2228                 else if (has_cache)             /* someone else added cache */
2229                         err = -EEXIST;
2230                 else                            /* no users remaining */
2231                         err = -ENOENT;
2232
2233         } else if (count || has_cache) {
2234
2235                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2236                         count += usage;
2237                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2238                         err = -EINVAL;
2239                 else if (swap_count_continued(p, offset, count))
2240                         count = COUNT_CONTINUED;
2241                 else
2242                         err = -ENOMEM;
2243         } else
2244                 err = -ENOENT;                  /* unused swap entry */
2245
2246         p->swap_map[offset] = count | has_cache;
2247
2248 unlock_out:
2249         spin_unlock(&swap_lock);
2250 out:
2251         return err;
2252
2253 bad_file:
2254         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2255         goto out;
2256 }
2257
2258 /*
2259  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2260  * (in which case its reference count is never incremented).
2261  */
2262 void swap_shmem_alloc(swp_entry_t entry)
2263 {
2264         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2265 }
2266
2267 /*
2268  * Increase reference count of swap entry by 1.
2269  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2270  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2271  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2272  * might occur if a page table entry has got corrupted.
2273  */
2274 int swap_duplicate(swp_entry_t entry)
2275 {
2276         int err = 0;
2277
2278         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2279                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2280         return err;
2281 }
2282
2283 /*
2284  * @entry: swap entry for which we allocate swap cache.
2285  *
2286  * Called when allocating swap cache for existing swap entry,
2287  * This can return error codes. Returns 0 at success.
2288  * -EBUSY means there is a swap cache.
2289  * Note: return code is different from swap_duplicate().
2290  */
2291 int swapcache_prepare(swp_entry_t entry)
2292 {
2293         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2294 }
2295
2296 /*
2297  * swap_lock prevents swap_map being freed. Don't grab an extra
2298  * reference on the swaphandle, it doesn't matter if it becomes unused.
2299  */
2300 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2301 {
2302         struct swap_info_struct *si;
2303         int our_page_cluster = page_cluster;
2304         pgoff_t target, toff;
2305         pgoff_t base, end;
2306         int nr_pages = 0;
2307
2308         if (!our_page_cluster)  /* no readahead */
2309                 return 0;
2310
2311         si = swap_info[swp_type(entry)];
2312         target = swp_offset(entry);
2313         base = (target >> our_page_cluster) << our_page_cluster;
2314         end = base + (1 << our_page_cluster);
2315         if (!base)              /* first page is swap header */
2316                 base++;
2317
2318         spin_lock(&swap_lock);
2319         if (end > si->max)      /* don't go beyond end of map */
2320                 end = si->max;
2321
2322         /* Count contiguous allocated slots above our target */
2323         for (toff = target; ++toff < end; nr_pages++) {
2324                 /* Don't read in free or bad pages */
2325                 if (!si->swap_map[toff])
2326                         break;
2327                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2328                         break;
2329         }
2330         /* Count contiguous allocated slots below our target */
2331         for (toff = target; --toff >= base; nr_pages++) {
2332                 /* Don't read in free or bad pages */
2333                 if (!si->swap_map[toff])
2334                         break;
2335                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2336                         break;
2337         }
2338         spin_unlock(&swap_lock);
2339
2340         /*
2341          * Indicate starting offset, and return number of pages to get:
2342          * if only 1, say 0, since there's then no readahead to be done.
2343          */
2344         *offset = ++toff;
2345         return nr_pages? ++nr_pages: 0;
2346 }
2347
2348 /*
2349  * add_swap_count_continuation - called when a swap count is duplicated
2350  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2351  * page of the original vmalloc'ed swap_map, to hold the continuation count
2352  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2353  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2354  *
2355  * These continuation pages are seldom referenced: the common paths all work
2356  * on the original swap_map, only referring to a continuation page when the
2357  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2358  *
2359  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2360  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2361  * can be called after dropping locks.
2362  */
2363 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2364 {
2365         struct swap_info_struct *si;
2366         struct page *head;
2367         struct page *page;
2368         struct page *list_page;
2369         pgoff_t offset;
2370         unsigned char count;
2371
2372         /*
2373          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2374          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2375          */
2376         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2377
2378         si = swap_info_get(entry);
2379         if (!si) {
2380                 /*
2381                  * An acceptable race has occurred since the failing
2382                  * __swap_duplicate(): the swap entry has been freed,
2383                  * perhaps even the whole swap_map cleared for swapoff.
2384                  */
2385                 goto outer;
2386         }
2387
2388         offset = swp_offset(entry);
2389         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2390
2391         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2392                 /*
2393                  * The higher the swap count, the more likely it is that tasks
2394                  * will race to add swap count continuation: we need to avoid
2395                  * over-provisioning.
2396                  */
2397                 goto out;
2398         }
2399
2400         if (!page) {
2401                 spin_unlock(&swap_lock);
2402                 return -ENOMEM;
2403         }
2404
2405         /*
2406          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2407          * no architecture is using highmem pages for kernel pagetables: so it
2408          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2409          */
2410         head = vmalloc_to_page(si->swap_map + offset);
2411         offset &= ~PAGE_MASK;
2412
2413         /*
2414          * Page allocation does not initialize the page's lru field,
2415          * but it does always reset its private field.
2416          */
2417         if (!page_private(head)) {
2418                 BUG_ON(count & COUNT_CONTINUED);
2419                 INIT_LIST_HEAD(&head->lru);
2420                 set_page_private(head, SWP_CONTINUED);
2421                 si->flags |= SWP_CONTINUED;
2422         }
2423
2424         list_for_each_entry(list_page, &head->lru, lru) {
2425                 unsigned char *map;
2426
2427                 /*
2428                  * If the previous map said no continuation, but we've found
2429                  * a continuation page, free our allocation and use this one.
2430                  */
2431                 if (!(count & COUNT_CONTINUED))
2432                         goto out;
2433
2434                 map = kmap_atomic(list_page, KM_USER0) + offset;
2435                 count = *map;
2436                 kunmap_atomic(map, KM_USER0);
2437
2438                 /*
2439                  * If this continuation count now has some space in it,
2440                  * free our allocation and use this one.
2441                  */
2442                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2443                         goto out;
2444         }
2445
2446         list_add_tail(&page->lru, &head->lru);
2447         page = NULL;                    /* now it's attached, don't free it */
2448 out:
2449         spin_unlock(&swap_lock);
2450 outer:
2451         if (page)
2452                 __free_page(page);
2453         return 0;
2454 }
2455
2456 /*
2457  * swap_count_continued - when the original swap_map count is incremented
2458  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2459  * into, carry if so, or else fail until a new continuation page is allocated;
2460  * when the original swap_map count is decremented from 0 with continuation,
2461  * borrow from the continuation and report whether it still holds more.
2462  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2463  */
2464 static bool swap_count_continued(struct swap_info_struct *si,
2465                                  pgoff_t offset, unsigned char count)
2466 {
2467         struct page *head;
2468         struct page *page;
2469         unsigned char *map;
2470
2471         head = vmalloc_to_page(si->swap_map + offset);
2472         if (page_private(head) != SWP_CONTINUED) {
2473                 BUG_ON(count & COUNT_CONTINUED);
2474                 return false;           /* need to add count continuation */
2475         }
2476
2477         offset &= ~PAGE_MASK;
2478         page = list_entry(head->lru.next, struct page, lru);
2479         map = kmap_atomic(page, KM_USER0) + offset;
2480
2481         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2482                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2483
2484         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2485                 /*
2486                  * Think of how you add 1 to 999
2487                  */
2488                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2489                         kunmap_atomic(map, KM_USER0);
2490                         page = list_entry(page->lru.next, struct page, lru);
2491                         BUG_ON(page == head);
2492                         map = kmap_atomic(page, KM_USER0) + offset;
2493                 }
2494                 if (*map == SWAP_CONT_MAX) {
2495                         kunmap_atomic(map, KM_USER0);
2496                         page = list_entry(page->lru.next, struct page, lru);
2497                         if (page == head)
2498                                 return false;   /* add count continuation */
2499                         map = kmap_atomic(page, KM_USER0) + offset;
2500 init_map:               *map = 0;               /* we didn't zero the page */
2501                 }
2502                 *map += 1;
2503                 kunmap_atomic(map, KM_USER0);
2504                 page = list_entry(page->lru.prev, struct page, lru);
2505                 while (page != head) {
2506                         map = kmap_atomic(page, KM_USER0) + offset;
2507                         *map = COUNT_CONTINUED;
2508                         kunmap_atomic(map, KM_USER0);
2509                         page = list_entry(page->lru.prev, struct page, lru);
2510                 }
2511                 return true;                    /* incremented */
2512
2513         } else {                                /* decrementing */
2514                 /*
2515                  * Think of how you subtract 1 from 1000
2516                  */
2517                 BUG_ON(count != COUNT_CONTINUED);
2518                 while (*map == COUNT_CONTINUED) {
2519                         kunmap_atomic(map, KM_USER0);
2520                         page = list_entry(page->lru.next, struct page, lru);
2521                         BUG_ON(page == head);
2522                         map = kmap_atomic(page, KM_USER0) + offset;
2523                 }
2524                 BUG_ON(*map == 0);
2525                 *map -= 1;
2526                 if (*map == 0)
2527                         count = 0;
2528                 kunmap_atomic(map, KM_USER0);
2529                 page = list_entry(page->lru.prev, struct page, lru);
2530                 while (page != head) {
2531                         map = kmap_atomic(page, KM_USER0) + offset;
2532                         *map = SWAP_CONT_MAX | count;
2533                         count = COUNT_CONTINUED;
2534                         kunmap_atomic(map, KM_USER0);
2535                         page = list_entry(page->lru.prev, struct page, lru);
2536                 }
2537                 return count == COUNT_CONTINUED;
2538         }
2539 }
2540
2541 /*
2542  * free_swap_count_continuations - swapoff free all the continuation pages
2543  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2544  */
2545 static void free_swap_count_continuations(struct swap_info_struct *si)
2546 {
2547         pgoff_t offset;
2548
2549         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2550                 struct page *head;
2551                 head = vmalloc_to_page(si->swap_map + offset);
2552                 if (page_private(head)) {
2553                         struct list_head *this, *next;
2554                         list_for_each_safe(this, next, &head->lru) {
2555                                 struct page *page;
2556                                 page = list_entry(this, struct page, lru);
2557                                 list_del(this);
2558                                 __free_page(page);
2559                         }
2560                 }
2561         }
2562 }