0579d9069b61d780abd75146ad870af8a4639425
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32
33 #include <asm/pgtable.h>
34 #include <asm/tlbflush.h>
35 #include <linux/swapops.h>
36 #include <linux/page_cgroup.h>
37
38 static DEFINE_SPINLOCK(swap_lock);
39 static unsigned int nr_swapfiles;
40 long nr_swap_pages;
41 long total_swap_pages;
42 static int swap_overflow;
43 static int least_priority;
44
45 static const char Bad_file[] = "Bad swap file entry ";
46 static const char Unused_file[] = "Unused swap file entry ";
47 static const char Bad_offset[] = "Bad swap offset entry ";
48 static const char Unused_offset[] = "Unused swap offset entry ";
49
50 static struct swap_list_t swap_list = {-1, -1};
51
52 static struct swap_info_struct swap_info[MAX_SWAPFILES];
53
54 static DEFINE_MUTEX(swapon_mutex);
55
56 /*
57  * We need this because the bdev->unplug_fn can sleep and we cannot
58  * hold swap_lock while calling the unplug_fn. And swap_lock
59  * cannot be turned into a mutex.
60  */
61 static DECLARE_RWSEM(swap_unplug_sem);
62
63 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
64 {
65         swp_entry_t entry;
66
67         down_read(&swap_unplug_sem);
68         entry.val = page_private(page);
69         if (PageSwapCache(page)) {
70                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
71                 struct backing_dev_info *bdi;
72
73                 /*
74                  * If the page is removed from swapcache from under us (with a
75                  * racy try_to_unuse/swapoff) we need an additional reference
76                  * count to avoid reading garbage from page_private(page) above.
77                  * If the WARN_ON triggers during a swapoff it maybe the race
78                  * condition and it's harmless. However if it triggers without
79                  * swapoff it signals a problem.
80                  */
81                 WARN_ON(page_count(page) <= 1);
82
83                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
84                 blk_run_backing_dev(bdi, page);
85         }
86         up_read(&swap_unplug_sem);
87 }
88
89 /*
90  * swapon tell device that all the old swap contents can be discarded,
91  * to allow the swap device to optimize its wear-levelling.
92  */
93 static int discard_swap(struct swap_info_struct *si)
94 {
95         struct swap_extent *se;
96         int err = 0;
97
98         list_for_each_entry(se, &si->extent_list, list) {
99                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
100                 sector_t nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
101
102                 if (se->start_page == 0) {
103                         /* Do not discard the swap header page! */
104                         start_block += 1 << (PAGE_SHIFT - 9);
105                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
106                         if (!nr_blocks)
107                                 continue;
108                 }
109
110                 err = blkdev_issue_discard(si->bdev, start_block,
111                                                 nr_blocks, GFP_KERNEL);
112                 if (err)
113                         break;
114
115                 cond_resched();
116         }
117         return err;             /* That will often be -EOPNOTSUPP */
118 }
119
120 /*
121  * swap allocation tell device that a cluster of swap can now be discarded,
122  * to allow the swap device to optimize its wear-levelling.
123  */
124 static void discard_swap_cluster(struct swap_info_struct *si,
125                                  pgoff_t start_page, pgoff_t nr_pages)
126 {
127         struct swap_extent *se = si->curr_swap_extent;
128         int found_extent = 0;
129
130         while (nr_pages) {
131                 struct list_head *lh;
132
133                 if (se->start_page <= start_page &&
134                     start_page < se->start_page + se->nr_pages) {
135                         pgoff_t offset = start_page - se->start_page;
136                         sector_t start_block = se->start_block + offset;
137                         sector_t nr_blocks = se->nr_pages - offset;
138
139                         if (nr_blocks > nr_pages)
140                                 nr_blocks = nr_pages;
141                         start_page += nr_blocks;
142                         nr_pages -= nr_blocks;
143
144                         if (!found_extent++)
145                                 si->curr_swap_extent = se;
146
147                         start_block <<= PAGE_SHIFT - 9;
148                         nr_blocks <<= PAGE_SHIFT - 9;
149                         if (blkdev_issue_discard(si->bdev, start_block,
150                                                         nr_blocks, GFP_NOIO))
151                                 break;
152                 }
153
154                 lh = se->list.next;
155                 if (lh == &si->extent_list)
156                         lh = lh->next;
157                 se = list_entry(lh, struct swap_extent, list);
158         }
159 }
160
161 static int wait_for_discard(void *word)
162 {
163         schedule();
164         return 0;
165 }
166
167 #define SWAPFILE_CLUSTER        256
168 #define LATENCY_LIMIT           256
169
170 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
171 {
172         unsigned long offset;
173         unsigned long scan_base;
174         unsigned long last_in_cluster = 0;
175         int latency_ration = LATENCY_LIMIT;
176         int found_free_cluster = 0;
177
178         /*
179          * We try to cluster swap pages by allocating them sequentially
180          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
181          * way, however, we resort to first-free allocation, starting
182          * a new cluster.  This prevents us from scattering swap pages
183          * all over the entire swap partition, so that we reduce
184          * overall disk seek times between swap pages.  -- sct
185          * But we do now try to find an empty cluster.  -Andrea
186          * And we let swap pages go all over an SSD partition.  Hugh
187          */
188
189         si->flags += SWP_SCANNING;
190         scan_base = offset = si->cluster_next;
191
192         if (unlikely(!si->cluster_nr--)) {
193                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
194                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
195                         goto checks;
196                 }
197                 if (si->flags & SWP_DISCARDABLE) {
198                         /*
199                          * Start range check on racing allocations, in case
200                          * they overlap the cluster we eventually decide on
201                          * (we scan without swap_lock to allow preemption).
202                          * It's hardly conceivable that cluster_nr could be
203                          * wrapped during our scan, but don't depend on it.
204                          */
205                         if (si->lowest_alloc)
206                                 goto checks;
207                         si->lowest_alloc = si->max;
208                         si->highest_alloc = 0;
209                 }
210                 spin_unlock(&swap_lock);
211
212                 /*
213                  * If seek is expensive, start searching for new cluster from
214                  * start of partition, to minimize the span of allocated swap.
215                  * But if seek is cheap, search from our current position, so
216                  * that swap is allocated from all over the partition: if the
217                  * Flash Translation Layer only remaps within limited zones,
218                  * we don't want to wear out the first zone too quickly.
219                  */
220                 if (!(si->flags & SWP_SOLIDSTATE))
221                         scan_base = offset = si->lowest_bit;
222                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
223
224                 /* Locate the first empty (unaligned) cluster */
225                 for (; last_in_cluster <= si->highest_bit; offset++) {
226                         if (si->swap_map[offset])
227                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
228                         else if (offset == last_in_cluster) {
229                                 spin_lock(&swap_lock);
230                                 offset -= SWAPFILE_CLUSTER - 1;
231                                 si->cluster_next = offset;
232                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
233                                 found_free_cluster = 1;
234                                 goto checks;
235                         }
236                         if (unlikely(--latency_ration < 0)) {
237                                 cond_resched();
238                                 latency_ration = LATENCY_LIMIT;
239                         }
240                 }
241
242                 offset = si->lowest_bit;
243                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
244
245                 /* Locate the first empty (unaligned) cluster */
246                 for (; last_in_cluster < scan_base; offset++) {
247                         if (si->swap_map[offset])
248                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
249                         else if (offset == last_in_cluster) {
250                                 spin_lock(&swap_lock);
251                                 offset -= SWAPFILE_CLUSTER - 1;
252                                 si->cluster_next = offset;
253                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
254                                 found_free_cluster = 1;
255                                 goto checks;
256                         }
257                         if (unlikely(--latency_ration < 0)) {
258                                 cond_resched();
259                                 latency_ration = LATENCY_LIMIT;
260                         }
261                 }
262
263                 offset = scan_base;
264                 spin_lock(&swap_lock);
265                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
266                 si->lowest_alloc = 0;
267         }
268
269 checks:
270         if (!(si->flags & SWP_WRITEOK))
271                 goto no_page;
272         if (!si->highest_bit)
273                 goto no_page;
274         if (offset > si->highest_bit)
275                 scan_base = offset = si->lowest_bit;
276         if (si->swap_map[offset])
277                 goto scan;
278
279         if (offset == si->lowest_bit)
280                 si->lowest_bit++;
281         if (offset == si->highest_bit)
282                 si->highest_bit--;
283         si->inuse_pages++;
284         if (si->inuse_pages == si->pages) {
285                 si->lowest_bit = si->max;
286                 si->highest_bit = 0;
287         }
288         si->swap_map[offset] = 1;
289         si->cluster_next = offset + 1;
290         si->flags -= SWP_SCANNING;
291
292         if (si->lowest_alloc) {
293                 /*
294                  * Only set when SWP_DISCARDABLE, and there's a scan
295                  * for a free cluster in progress or just completed.
296                  */
297                 if (found_free_cluster) {
298                         /*
299                          * To optimize wear-levelling, discard the
300                          * old data of the cluster, taking care not to
301                          * discard any of its pages that have already
302                          * been allocated by racing tasks (offset has
303                          * already stepped over any at the beginning).
304                          */
305                         if (offset < si->highest_alloc &&
306                             si->lowest_alloc <= last_in_cluster)
307                                 last_in_cluster = si->lowest_alloc - 1;
308                         si->flags |= SWP_DISCARDING;
309                         spin_unlock(&swap_lock);
310
311                         if (offset < last_in_cluster)
312                                 discard_swap_cluster(si, offset,
313                                         last_in_cluster - offset + 1);
314
315                         spin_lock(&swap_lock);
316                         si->lowest_alloc = 0;
317                         si->flags &= ~SWP_DISCARDING;
318
319                         smp_mb();       /* wake_up_bit advises this */
320                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
321
322                 } else if (si->flags & SWP_DISCARDING) {
323                         /*
324                          * Delay using pages allocated by racing tasks
325                          * until the whole discard has been issued. We
326                          * could defer that delay until swap_writepage,
327                          * but it's easier to keep this self-contained.
328                          */
329                         spin_unlock(&swap_lock);
330                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
331                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
332                         spin_lock(&swap_lock);
333                 } else {
334                         /*
335                          * Note pages allocated by racing tasks while
336                          * scan for a free cluster is in progress, so
337                          * that its final discard can exclude them.
338                          */
339                         if (offset < si->lowest_alloc)
340                                 si->lowest_alloc = offset;
341                         if (offset > si->highest_alloc)
342                                 si->highest_alloc = offset;
343                 }
344         }
345         return offset;
346
347 scan:
348         spin_unlock(&swap_lock);
349         while (++offset <= si->highest_bit) {
350                 if (!si->swap_map[offset]) {
351                         spin_lock(&swap_lock);
352                         goto checks;
353                 }
354                 if (unlikely(--latency_ration < 0)) {
355                         cond_resched();
356                         latency_ration = LATENCY_LIMIT;
357                 }
358         }
359         offset = si->lowest_bit;
360         while (++offset < scan_base) {
361                 if (!si->swap_map[offset]) {
362                         spin_lock(&swap_lock);
363                         goto checks;
364                 }
365                 if (unlikely(--latency_ration < 0)) {
366                         cond_resched();
367                         latency_ration = LATENCY_LIMIT;
368                 }
369         }
370         spin_lock(&swap_lock);
371
372 no_page:
373         si->flags -= SWP_SCANNING;
374         return 0;
375 }
376
377 swp_entry_t get_swap_page(void)
378 {
379         struct swap_info_struct *si;
380         pgoff_t offset;
381         int type, next;
382         int wrapped = 0;
383
384         spin_lock(&swap_lock);
385         if (nr_swap_pages <= 0)
386                 goto noswap;
387         nr_swap_pages--;
388
389         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
390                 si = swap_info + type;
391                 next = si->next;
392                 if (next < 0 ||
393                     (!wrapped && si->prio != swap_info[next].prio)) {
394                         next = swap_list.head;
395                         wrapped++;
396                 }
397
398                 if (!si->highest_bit)
399                         continue;
400                 if (!(si->flags & SWP_WRITEOK))
401                         continue;
402
403                 swap_list.next = next;
404                 offset = scan_swap_map(si);
405                 if (offset) {
406                         spin_unlock(&swap_lock);
407                         return swp_entry(type, offset);
408                 }
409                 next = swap_list.next;
410         }
411
412         nr_swap_pages++;
413 noswap:
414         spin_unlock(&swap_lock);
415         return (swp_entry_t) {0};
416 }
417
418 swp_entry_t get_swap_page_of_type(int type)
419 {
420         struct swap_info_struct *si;
421         pgoff_t offset;
422
423         spin_lock(&swap_lock);
424         si = swap_info + type;
425         if (si->flags & SWP_WRITEOK) {
426                 nr_swap_pages--;
427                 offset = scan_swap_map(si);
428                 if (offset) {
429                         spin_unlock(&swap_lock);
430                         return swp_entry(type, offset);
431                 }
432                 nr_swap_pages++;
433         }
434         spin_unlock(&swap_lock);
435         return (swp_entry_t) {0};
436 }
437
438 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
439 {
440         struct swap_info_struct * p;
441         unsigned long offset, type;
442
443         if (!entry.val)
444                 goto out;
445         type = swp_type(entry);
446         if (type >= nr_swapfiles)
447                 goto bad_nofile;
448         p = & swap_info[type];
449         if (!(p->flags & SWP_USED))
450                 goto bad_device;
451         offset = swp_offset(entry);
452         if (offset >= p->max)
453                 goto bad_offset;
454         if (!p->swap_map[offset])
455                 goto bad_free;
456         spin_lock(&swap_lock);
457         return p;
458
459 bad_free:
460         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
461         goto out;
462 bad_offset:
463         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
464         goto out;
465 bad_device:
466         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
467         goto out;
468 bad_nofile:
469         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
470 out:
471         return NULL;
472 }
473
474 static int swap_entry_free(struct swap_info_struct *p, swp_entry_t ent)
475 {
476         unsigned long offset = swp_offset(ent);
477         int count = p->swap_map[offset];
478
479         if (count < SWAP_MAP_MAX) {
480                 count--;
481                 p->swap_map[offset] = count;
482                 if (!count) {
483                         if (offset < p->lowest_bit)
484                                 p->lowest_bit = offset;
485                         if (offset > p->highest_bit)
486                                 p->highest_bit = offset;
487                         if (p->prio > swap_info[swap_list.next].prio)
488                                 swap_list.next = p - swap_info;
489                         nr_swap_pages++;
490                         p->inuse_pages--;
491                         mem_cgroup_uncharge_swap(ent);
492                 }
493         }
494         return count;
495 }
496
497 /*
498  * Caller has made sure that the swapdevice corresponding to entry
499  * is still around or has not been recycled.
500  */
501 void swap_free(swp_entry_t entry)
502 {
503         struct swap_info_struct * p;
504
505         p = swap_info_get(entry);
506         if (p) {
507                 swap_entry_free(p, entry);
508                 spin_unlock(&swap_lock);
509         }
510 }
511
512 /*
513  * How many references to page are currently swapped out?
514  */
515 static inline int page_swapcount(struct page *page)
516 {
517         int count = 0;
518         struct swap_info_struct *p;
519         swp_entry_t entry;
520
521         entry.val = page_private(page);
522         p = swap_info_get(entry);
523         if (p) {
524                 /* Subtract the 1 for the swap cache itself */
525                 count = p->swap_map[swp_offset(entry)] - 1;
526                 spin_unlock(&swap_lock);
527         }
528         return count;
529 }
530
531 /*
532  * We can write to an anon page without COW if there are no other references
533  * to it.  And as a side-effect, free up its swap: because the old content
534  * on disk will never be read, and seeking back there to write new content
535  * later would only waste time away from clustering.
536  */
537 int reuse_swap_page(struct page *page)
538 {
539         int count;
540
541         VM_BUG_ON(!PageLocked(page));
542         count = page_mapcount(page);
543         if (count <= 1 && PageSwapCache(page)) {
544                 count += page_swapcount(page);
545                 if (count == 1 && !PageWriteback(page)) {
546                         delete_from_swap_cache(page);
547                         SetPageDirty(page);
548                 }
549         }
550         return count == 1;
551 }
552
553 /*
554  * If swap is getting full, or if there are no more mappings of this page,
555  * then try_to_free_swap is called to free its swap space.
556  */
557 int try_to_free_swap(struct page *page)
558 {
559         VM_BUG_ON(!PageLocked(page));
560
561         if (!PageSwapCache(page))
562                 return 0;
563         if (PageWriteback(page))
564                 return 0;
565         if (page_swapcount(page))
566                 return 0;
567
568         delete_from_swap_cache(page);
569         SetPageDirty(page);
570         return 1;
571 }
572
573 /*
574  * Free the swap entry like above, but also try to
575  * free the page cache entry if it is the last user.
576  */
577 int free_swap_and_cache(swp_entry_t entry)
578 {
579         struct swap_info_struct *p;
580         struct page *page = NULL;
581
582         if (is_migration_entry(entry))
583                 return 1;
584
585         p = swap_info_get(entry);
586         if (p) {
587                 if (swap_entry_free(p, entry) == 1) {
588                         page = find_get_page(&swapper_space, entry.val);
589                         if (page && !trylock_page(page)) {
590                                 page_cache_release(page);
591                                 page = NULL;
592                         }
593                 }
594                 spin_unlock(&swap_lock);
595         }
596         if (page) {
597                 /*
598                  * Not mapped elsewhere, or swap space full? Free it!
599                  * Also recheck PageSwapCache now page is locked (above).
600                  */
601                 if (PageSwapCache(page) && !PageWriteback(page) &&
602                                 (!page_mapped(page) || vm_swap_full())) {
603                         delete_from_swap_cache(page);
604                         SetPageDirty(page);
605                 }
606                 unlock_page(page);
607                 page_cache_release(page);
608         }
609         return p != NULL;
610 }
611
612 #ifdef CONFIG_HIBERNATION
613 /*
614  * Find the swap type that corresponds to given device (if any).
615  *
616  * @offset - number of the PAGE_SIZE-sized block of the device, starting
617  * from 0, in which the swap header is expected to be located.
618  *
619  * This is needed for the suspend to disk (aka swsusp).
620  */
621 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
622 {
623         struct block_device *bdev = NULL;
624         int i;
625
626         if (device)
627                 bdev = bdget(device);
628
629         spin_lock(&swap_lock);
630         for (i = 0; i < nr_swapfiles; i++) {
631                 struct swap_info_struct *sis = swap_info + i;
632
633                 if (!(sis->flags & SWP_WRITEOK))
634                         continue;
635
636                 if (!bdev) {
637                         if (bdev_p)
638                                 *bdev_p = sis->bdev;
639
640                         spin_unlock(&swap_lock);
641                         return i;
642                 }
643                 if (bdev == sis->bdev) {
644                         struct swap_extent *se;
645
646                         se = list_entry(sis->extent_list.next,
647                                         struct swap_extent, list);
648                         if (se->start_block == offset) {
649                                 if (bdev_p)
650                                         *bdev_p = sis->bdev;
651
652                                 spin_unlock(&swap_lock);
653                                 bdput(bdev);
654                                 return i;
655                         }
656                 }
657         }
658         spin_unlock(&swap_lock);
659         if (bdev)
660                 bdput(bdev);
661
662         return -ENODEV;
663 }
664
665 /*
666  * Return either the total number of swap pages of given type, or the number
667  * of free pages of that type (depending on @free)
668  *
669  * This is needed for software suspend
670  */
671 unsigned int count_swap_pages(int type, int free)
672 {
673         unsigned int n = 0;
674
675         if (type < nr_swapfiles) {
676                 spin_lock(&swap_lock);
677                 if (swap_info[type].flags & SWP_WRITEOK) {
678                         n = swap_info[type].pages;
679                         if (free)
680                                 n -= swap_info[type].inuse_pages;
681                 }
682                 spin_unlock(&swap_lock);
683         }
684         return n;
685 }
686 #endif
687
688 /*
689  * No need to decide whether this PTE shares the swap entry with others,
690  * just let do_wp_page work it out if a write is requested later - to
691  * force COW, vm_page_prot omits write permission from any private vma.
692  */
693 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
694                 unsigned long addr, swp_entry_t entry, struct page *page)
695 {
696         struct mem_cgroup *ptr = NULL;
697         spinlock_t *ptl;
698         pte_t *pte;
699         int ret = 1;
700
701         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
702                                         GFP_HIGHUSER_MOVABLE, &ptr))
703                 ret = -ENOMEM;
704
705         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
706         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
707                 if (ret > 0)
708                         mem_cgroup_cancel_charge_swapin(ptr);
709                 ret = 0;
710                 goto out;
711         }
712
713         inc_mm_counter(vma->vm_mm, anon_rss);
714         get_page(page);
715         set_pte_at(vma->vm_mm, addr, pte,
716                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
717         page_add_anon_rmap(page, vma, addr);
718         mem_cgroup_commit_charge_swapin(page, ptr);
719         swap_free(entry);
720         /*
721          * Move the page to the active list so it is not
722          * immediately swapped out again after swapon.
723          */
724         activate_page(page);
725 out:
726         pte_unmap_unlock(pte, ptl);
727         return ret;
728 }
729
730 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
731                                 unsigned long addr, unsigned long end,
732                                 swp_entry_t entry, struct page *page)
733 {
734         pte_t swp_pte = swp_entry_to_pte(entry);
735         pte_t *pte;
736         int ret = 0;
737
738         /*
739          * We don't actually need pte lock while scanning for swp_pte: since
740          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
741          * page table while we're scanning; though it could get zapped, and on
742          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
743          * of unmatched parts which look like swp_pte, so unuse_pte must
744          * recheck under pte lock.  Scanning without pte lock lets it be
745          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
746          */
747         pte = pte_offset_map(pmd, addr);
748         do {
749                 /*
750                  * swapoff spends a _lot_ of time in this loop!
751                  * Test inline before going to call unuse_pte.
752                  */
753                 if (unlikely(pte_same(*pte, swp_pte))) {
754                         pte_unmap(pte);
755                         ret = unuse_pte(vma, pmd, addr, entry, page);
756                         if (ret)
757                                 goto out;
758                         pte = pte_offset_map(pmd, addr);
759                 }
760         } while (pte++, addr += PAGE_SIZE, addr != end);
761         pte_unmap(pte - 1);
762 out:
763         return ret;
764 }
765
766 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
767                                 unsigned long addr, unsigned long end,
768                                 swp_entry_t entry, struct page *page)
769 {
770         pmd_t *pmd;
771         unsigned long next;
772         int ret;
773
774         pmd = pmd_offset(pud, addr);
775         do {
776                 next = pmd_addr_end(addr, end);
777                 if (pmd_none_or_clear_bad(pmd))
778                         continue;
779                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
780                 if (ret)
781                         return ret;
782         } while (pmd++, addr = next, addr != end);
783         return 0;
784 }
785
786 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
787                                 unsigned long addr, unsigned long end,
788                                 swp_entry_t entry, struct page *page)
789 {
790         pud_t *pud;
791         unsigned long next;
792         int ret;
793
794         pud = pud_offset(pgd, addr);
795         do {
796                 next = pud_addr_end(addr, end);
797                 if (pud_none_or_clear_bad(pud))
798                         continue;
799                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
800                 if (ret)
801                         return ret;
802         } while (pud++, addr = next, addr != end);
803         return 0;
804 }
805
806 static int unuse_vma(struct vm_area_struct *vma,
807                                 swp_entry_t entry, struct page *page)
808 {
809         pgd_t *pgd;
810         unsigned long addr, end, next;
811         int ret;
812
813         if (page->mapping) {
814                 addr = page_address_in_vma(page, vma);
815                 if (addr == -EFAULT)
816                         return 0;
817                 else
818                         end = addr + PAGE_SIZE;
819         } else {
820                 addr = vma->vm_start;
821                 end = vma->vm_end;
822         }
823
824         pgd = pgd_offset(vma->vm_mm, addr);
825         do {
826                 next = pgd_addr_end(addr, end);
827                 if (pgd_none_or_clear_bad(pgd))
828                         continue;
829                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
830                 if (ret)
831                         return ret;
832         } while (pgd++, addr = next, addr != end);
833         return 0;
834 }
835
836 static int unuse_mm(struct mm_struct *mm,
837                                 swp_entry_t entry, struct page *page)
838 {
839         struct vm_area_struct *vma;
840         int ret = 0;
841
842         if (!down_read_trylock(&mm->mmap_sem)) {
843                 /*
844                  * Activate page so shrink_inactive_list is unlikely to unmap
845                  * its ptes while lock is dropped, so swapoff can make progress.
846                  */
847                 activate_page(page);
848                 unlock_page(page);
849                 down_read(&mm->mmap_sem);
850                 lock_page(page);
851         }
852         for (vma = mm->mmap; vma; vma = vma->vm_next) {
853                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
854                         break;
855         }
856         up_read(&mm->mmap_sem);
857         return (ret < 0)? ret: 0;
858 }
859
860 /*
861  * Scan swap_map from current position to next entry still in use.
862  * Recycle to start on reaching the end, returning 0 when empty.
863  */
864 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
865                                         unsigned int prev)
866 {
867         unsigned int max = si->max;
868         unsigned int i = prev;
869         int count;
870
871         /*
872          * No need for swap_lock here: we're just looking
873          * for whether an entry is in use, not modifying it; false
874          * hits are okay, and sys_swapoff() has already prevented new
875          * allocations from this area (while holding swap_lock).
876          */
877         for (;;) {
878                 if (++i >= max) {
879                         if (!prev) {
880                                 i = 0;
881                                 break;
882                         }
883                         /*
884                          * No entries in use at top of swap_map,
885                          * loop back to start and recheck there.
886                          */
887                         max = prev + 1;
888                         prev = 0;
889                         i = 1;
890                 }
891                 count = si->swap_map[i];
892                 if (count && count != SWAP_MAP_BAD)
893                         break;
894         }
895         return i;
896 }
897
898 /*
899  * We completely avoid races by reading each swap page in advance,
900  * and then search for the process using it.  All the necessary
901  * page table adjustments can then be made atomically.
902  */
903 static int try_to_unuse(unsigned int type)
904 {
905         struct swap_info_struct * si = &swap_info[type];
906         struct mm_struct *start_mm;
907         unsigned short *swap_map;
908         unsigned short swcount;
909         struct page *page;
910         swp_entry_t entry;
911         unsigned int i = 0;
912         int retval = 0;
913         int reset_overflow = 0;
914         int shmem;
915
916         /*
917          * When searching mms for an entry, a good strategy is to
918          * start at the first mm we freed the previous entry from
919          * (though actually we don't notice whether we or coincidence
920          * freed the entry).  Initialize this start_mm with a hold.
921          *
922          * A simpler strategy would be to start at the last mm we
923          * freed the previous entry from; but that would take less
924          * advantage of mmlist ordering, which clusters forked mms
925          * together, child after parent.  If we race with dup_mmap(), we
926          * prefer to resolve parent before child, lest we miss entries
927          * duplicated after we scanned child: using last mm would invert
928          * that.  Though it's only a serious concern when an overflowed
929          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
930          */
931         start_mm = &init_mm;
932         atomic_inc(&init_mm.mm_users);
933
934         /*
935          * Keep on scanning until all entries have gone.  Usually,
936          * one pass through swap_map is enough, but not necessarily:
937          * there are races when an instance of an entry might be missed.
938          */
939         while ((i = find_next_to_unuse(si, i)) != 0) {
940                 if (signal_pending(current)) {
941                         retval = -EINTR;
942                         break;
943                 }
944
945                 /*
946                  * Get a page for the entry, using the existing swap
947                  * cache page if there is one.  Otherwise, get a clean
948                  * page and read the swap into it.
949                  */
950                 swap_map = &si->swap_map[i];
951                 entry = swp_entry(type, i);
952                 page = read_swap_cache_async(entry,
953                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
954                 if (!page) {
955                         /*
956                          * Either swap_duplicate() failed because entry
957                          * has been freed independently, and will not be
958                          * reused since sys_swapoff() already disabled
959                          * allocation from here, or alloc_page() failed.
960                          */
961                         if (!*swap_map)
962                                 continue;
963                         retval = -ENOMEM;
964                         break;
965                 }
966
967                 /*
968                  * Don't hold on to start_mm if it looks like exiting.
969                  */
970                 if (atomic_read(&start_mm->mm_users) == 1) {
971                         mmput(start_mm);
972                         start_mm = &init_mm;
973                         atomic_inc(&init_mm.mm_users);
974                 }
975
976                 /*
977                  * Wait for and lock page.  When do_swap_page races with
978                  * try_to_unuse, do_swap_page can handle the fault much
979                  * faster than try_to_unuse can locate the entry.  This
980                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
981                  * defer to do_swap_page in such a case - in some tests,
982                  * do_swap_page and try_to_unuse repeatedly compete.
983                  */
984                 wait_on_page_locked(page);
985                 wait_on_page_writeback(page);
986                 lock_page(page);
987                 wait_on_page_writeback(page);
988
989                 /*
990                  * Remove all references to entry.
991                  * Whenever we reach init_mm, there's no address space
992                  * to search, but use it as a reminder to search shmem.
993                  */
994                 shmem = 0;
995                 swcount = *swap_map;
996                 if (swcount > 1) {
997                         if (start_mm == &init_mm)
998                                 shmem = shmem_unuse(entry, page);
999                         else
1000                                 retval = unuse_mm(start_mm, entry, page);
1001                 }
1002                 if (*swap_map > 1) {
1003                         int set_start_mm = (*swap_map >= swcount);
1004                         struct list_head *p = &start_mm->mmlist;
1005                         struct mm_struct *new_start_mm = start_mm;
1006                         struct mm_struct *prev_mm = start_mm;
1007                         struct mm_struct *mm;
1008
1009                         atomic_inc(&new_start_mm->mm_users);
1010                         atomic_inc(&prev_mm->mm_users);
1011                         spin_lock(&mmlist_lock);
1012                         while (*swap_map > 1 && !retval && !shmem &&
1013                                         (p = p->next) != &start_mm->mmlist) {
1014                                 mm = list_entry(p, struct mm_struct, mmlist);
1015                                 if (!atomic_inc_not_zero(&mm->mm_users))
1016                                         continue;
1017                                 spin_unlock(&mmlist_lock);
1018                                 mmput(prev_mm);
1019                                 prev_mm = mm;
1020
1021                                 cond_resched();
1022
1023                                 swcount = *swap_map;
1024                                 if (swcount <= 1)
1025                                         ;
1026                                 else if (mm == &init_mm) {
1027                                         set_start_mm = 1;
1028                                         shmem = shmem_unuse(entry, page);
1029                                 } else
1030                                         retval = unuse_mm(mm, entry, page);
1031                                 if (set_start_mm && *swap_map < swcount) {
1032                                         mmput(new_start_mm);
1033                                         atomic_inc(&mm->mm_users);
1034                                         new_start_mm = mm;
1035                                         set_start_mm = 0;
1036                                 }
1037                                 spin_lock(&mmlist_lock);
1038                         }
1039                         spin_unlock(&mmlist_lock);
1040                         mmput(prev_mm);
1041                         mmput(start_mm);
1042                         start_mm = new_start_mm;
1043                 }
1044                 if (shmem) {
1045                         /* page has already been unlocked and released */
1046                         if (shmem > 0)
1047                                 continue;
1048                         retval = shmem;
1049                         break;
1050                 }
1051                 if (retval) {
1052                         unlock_page(page);
1053                         page_cache_release(page);
1054                         break;
1055                 }
1056
1057                 /*
1058                  * How could swap count reach 0x7fff when the maximum
1059                  * pid is 0x7fff, and there's no way to repeat a swap
1060                  * page within an mm (except in shmem, where it's the
1061                  * shared object which takes the reference count)?
1062                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
1063                  *
1064                  * If that's wrong, then we should worry more about
1065                  * exit_mmap() and do_munmap() cases described above:
1066                  * we might be resetting SWAP_MAP_MAX too early here.
1067                  * We know "Undead"s can happen, they're okay, so don't
1068                  * report them; but do report if we reset SWAP_MAP_MAX.
1069                  */
1070                 if (*swap_map == SWAP_MAP_MAX) {
1071                         spin_lock(&swap_lock);
1072                         *swap_map = 1;
1073                         spin_unlock(&swap_lock);
1074                         reset_overflow = 1;
1075                 }
1076
1077                 /*
1078                  * If a reference remains (rare), we would like to leave
1079                  * the page in the swap cache; but try_to_unmap could
1080                  * then re-duplicate the entry once we drop page lock,
1081                  * so we might loop indefinitely; also, that page could
1082                  * not be swapped out to other storage meanwhile.  So:
1083                  * delete from cache even if there's another reference,
1084                  * after ensuring that the data has been saved to disk -
1085                  * since if the reference remains (rarer), it will be
1086                  * read from disk into another page.  Splitting into two
1087                  * pages would be incorrect if swap supported "shared
1088                  * private" pages, but they are handled by tmpfs files.
1089                  */
1090                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
1091                         struct writeback_control wbc = {
1092                                 .sync_mode = WB_SYNC_NONE,
1093                         };
1094
1095                         swap_writepage(page, &wbc);
1096                         lock_page(page);
1097                         wait_on_page_writeback(page);
1098                 }
1099
1100                 /*
1101                  * It is conceivable that a racing task removed this page from
1102                  * swap cache just before we acquired the page lock at the top,
1103                  * or while we dropped it in unuse_mm().  The page might even
1104                  * be back in swap cache on another swap area: that we must not
1105                  * delete, since it may not have been written out to swap yet.
1106                  */
1107                 if (PageSwapCache(page) &&
1108                     likely(page_private(page) == entry.val))
1109                         delete_from_swap_cache(page);
1110
1111                 /*
1112                  * So we could skip searching mms once swap count went
1113                  * to 1, we did not mark any present ptes as dirty: must
1114                  * mark page dirty so shrink_page_list will preserve it.
1115                  */
1116                 SetPageDirty(page);
1117                 unlock_page(page);
1118                 page_cache_release(page);
1119
1120                 /*
1121                  * Make sure that we aren't completely killing
1122                  * interactive performance.
1123                  */
1124                 cond_resched();
1125         }
1126
1127         mmput(start_mm);
1128         if (reset_overflow) {
1129                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1130                 swap_overflow = 0;
1131         }
1132         return retval;
1133 }
1134
1135 /*
1136  * After a successful try_to_unuse, if no swap is now in use, we know
1137  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1138  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1139  * added to the mmlist just after page_duplicate - before would be racy.
1140  */
1141 static void drain_mmlist(void)
1142 {
1143         struct list_head *p, *next;
1144         unsigned int i;
1145
1146         for (i = 0; i < nr_swapfiles; i++)
1147                 if (swap_info[i].inuse_pages)
1148                         return;
1149         spin_lock(&mmlist_lock);
1150         list_for_each_safe(p, next, &init_mm.mmlist)
1151                 list_del_init(p);
1152         spin_unlock(&mmlist_lock);
1153 }
1154
1155 /*
1156  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1157  * corresponds to page offset `offset'.
1158  */
1159 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1160 {
1161         struct swap_extent *se = sis->curr_swap_extent;
1162         struct swap_extent *start_se = se;
1163
1164         for ( ; ; ) {
1165                 struct list_head *lh;
1166
1167                 if (se->start_page <= offset &&
1168                                 offset < (se->start_page + se->nr_pages)) {
1169                         return se->start_block + (offset - se->start_page);
1170                 }
1171                 lh = se->list.next;
1172                 if (lh == &sis->extent_list)
1173                         lh = lh->next;
1174                 se = list_entry(lh, struct swap_extent, list);
1175                 sis->curr_swap_extent = se;
1176                 BUG_ON(se == start_se);         /* It *must* be present */
1177         }
1178 }
1179
1180 #ifdef CONFIG_HIBERNATION
1181 /*
1182  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1183  * corresponding to given index in swap_info (swap type).
1184  */
1185 sector_t swapdev_block(int swap_type, pgoff_t offset)
1186 {
1187         struct swap_info_struct *sis;
1188
1189         if (swap_type >= nr_swapfiles)
1190                 return 0;
1191
1192         sis = swap_info + swap_type;
1193         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1194 }
1195 #endif /* CONFIG_HIBERNATION */
1196
1197 /*
1198  * Free all of a swapdev's extent information
1199  */
1200 static void destroy_swap_extents(struct swap_info_struct *sis)
1201 {
1202         while (!list_empty(&sis->extent_list)) {
1203                 struct swap_extent *se;
1204
1205                 se = list_entry(sis->extent_list.next,
1206                                 struct swap_extent, list);
1207                 list_del(&se->list);
1208                 kfree(se);
1209         }
1210 }
1211
1212 /*
1213  * Add a block range (and the corresponding page range) into this swapdev's
1214  * extent list.  The extent list is kept sorted in page order.
1215  *
1216  * This function rather assumes that it is called in ascending page order.
1217  */
1218 static int
1219 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1220                 unsigned long nr_pages, sector_t start_block)
1221 {
1222         struct swap_extent *se;
1223         struct swap_extent *new_se;
1224         struct list_head *lh;
1225
1226         lh = sis->extent_list.prev;     /* The highest page extent */
1227         if (lh != &sis->extent_list) {
1228                 se = list_entry(lh, struct swap_extent, list);
1229                 BUG_ON(se->start_page + se->nr_pages != start_page);
1230                 if (se->start_block + se->nr_pages == start_block) {
1231                         /* Merge it */
1232                         se->nr_pages += nr_pages;
1233                         return 0;
1234                 }
1235         }
1236
1237         /*
1238          * No merge.  Insert a new extent, preserving ordering.
1239          */
1240         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1241         if (new_se == NULL)
1242                 return -ENOMEM;
1243         new_se->start_page = start_page;
1244         new_se->nr_pages = nr_pages;
1245         new_se->start_block = start_block;
1246
1247         list_add_tail(&new_se->list, &sis->extent_list);
1248         return 1;
1249 }
1250
1251 /*
1252  * A `swap extent' is a simple thing which maps a contiguous range of pages
1253  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1254  * is built at swapon time and is then used at swap_writepage/swap_readpage
1255  * time for locating where on disk a page belongs.
1256  *
1257  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1258  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1259  * swap files identically.
1260  *
1261  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1262  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1263  * swapfiles are handled *identically* after swapon time.
1264  *
1265  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1266  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1267  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1268  * requirements, they are simply tossed out - we will never use those blocks
1269  * for swapping.
1270  *
1271  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1272  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1273  * which will scribble on the fs.
1274  *
1275  * The amount of disk space which a single swap extent represents varies.
1276  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1277  * extents in the list.  To avoid much list walking, we cache the previous
1278  * search location in `curr_swap_extent', and start new searches from there.
1279  * This is extremely effective.  The average number of iterations in
1280  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1281  */
1282 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1283 {
1284         struct inode *inode;
1285         unsigned blocks_per_page;
1286         unsigned long page_no;
1287         unsigned blkbits;
1288         sector_t probe_block;
1289         sector_t last_block;
1290         sector_t lowest_block = -1;
1291         sector_t highest_block = 0;
1292         int nr_extents = 0;
1293         int ret;
1294
1295         inode = sis->swap_file->f_mapping->host;
1296         if (S_ISBLK(inode->i_mode)) {
1297                 ret = add_swap_extent(sis, 0, sis->max, 0);
1298                 *span = sis->pages;
1299                 goto done;
1300         }
1301
1302         blkbits = inode->i_blkbits;
1303         blocks_per_page = PAGE_SIZE >> blkbits;
1304
1305         /*
1306          * Map all the blocks into the extent list.  This code doesn't try
1307          * to be very smart.
1308          */
1309         probe_block = 0;
1310         page_no = 0;
1311         last_block = i_size_read(inode) >> blkbits;
1312         while ((probe_block + blocks_per_page) <= last_block &&
1313                         page_no < sis->max) {
1314                 unsigned block_in_page;
1315                 sector_t first_block;
1316
1317                 first_block = bmap(inode, probe_block);
1318                 if (first_block == 0)
1319                         goto bad_bmap;
1320
1321                 /*
1322                  * It must be PAGE_SIZE aligned on-disk
1323                  */
1324                 if (first_block & (blocks_per_page - 1)) {
1325                         probe_block++;
1326                         goto reprobe;
1327                 }
1328
1329                 for (block_in_page = 1; block_in_page < blocks_per_page;
1330                                         block_in_page++) {
1331                         sector_t block;
1332
1333                         block = bmap(inode, probe_block + block_in_page);
1334                         if (block == 0)
1335                                 goto bad_bmap;
1336                         if (block != first_block + block_in_page) {
1337                                 /* Discontiguity */
1338                                 probe_block++;
1339                                 goto reprobe;
1340                         }
1341                 }
1342
1343                 first_block >>= (PAGE_SHIFT - blkbits);
1344                 if (page_no) {  /* exclude the header page */
1345                         if (first_block < lowest_block)
1346                                 lowest_block = first_block;
1347                         if (first_block > highest_block)
1348                                 highest_block = first_block;
1349                 }
1350
1351                 /*
1352                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1353                  */
1354                 ret = add_swap_extent(sis, page_no, 1, first_block);
1355                 if (ret < 0)
1356                         goto out;
1357                 nr_extents += ret;
1358                 page_no++;
1359                 probe_block += blocks_per_page;
1360 reprobe:
1361                 continue;
1362         }
1363         ret = nr_extents;
1364         *span = 1 + highest_block - lowest_block;
1365         if (page_no == 0)
1366                 page_no = 1;    /* force Empty message */
1367         sis->max = page_no;
1368         sis->pages = page_no - 1;
1369         sis->highest_bit = page_no - 1;
1370 done:
1371         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1372                                         struct swap_extent, list);
1373         goto out;
1374 bad_bmap:
1375         printk(KERN_ERR "swapon: swapfile has holes\n");
1376         ret = -EINVAL;
1377 out:
1378         return ret;
1379 }
1380
1381 asmlinkage long sys_swapoff(const char __user * specialfile)
1382 {
1383         struct swap_info_struct * p = NULL;
1384         unsigned short *swap_map;
1385         struct file *swap_file, *victim;
1386         struct address_space *mapping;
1387         struct inode *inode;
1388         char * pathname;
1389         int i, type, prev;
1390         int err;
1391
1392         if (!capable(CAP_SYS_ADMIN))
1393                 return -EPERM;
1394
1395         pathname = getname(specialfile);
1396         err = PTR_ERR(pathname);
1397         if (IS_ERR(pathname))
1398                 goto out;
1399
1400         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1401         putname(pathname);
1402         err = PTR_ERR(victim);
1403         if (IS_ERR(victim))
1404                 goto out;
1405
1406         mapping = victim->f_mapping;
1407         prev = -1;
1408         spin_lock(&swap_lock);
1409         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1410                 p = swap_info + type;
1411                 if (p->flags & SWP_WRITEOK) {
1412                         if (p->swap_file->f_mapping == mapping)
1413                                 break;
1414                 }
1415                 prev = type;
1416         }
1417         if (type < 0) {
1418                 err = -EINVAL;
1419                 spin_unlock(&swap_lock);
1420                 goto out_dput;
1421         }
1422         if (!security_vm_enough_memory(p->pages))
1423                 vm_unacct_memory(p->pages);
1424         else {
1425                 err = -ENOMEM;
1426                 spin_unlock(&swap_lock);
1427                 goto out_dput;
1428         }
1429         if (prev < 0) {
1430                 swap_list.head = p->next;
1431         } else {
1432                 swap_info[prev].next = p->next;
1433         }
1434         if (type == swap_list.next) {
1435                 /* just pick something that's safe... */
1436                 swap_list.next = swap_list.head;
1437         }
1438         if (p->prio < 0) {
1439                 for (i = p->next; i >= 0; i = swap_info[i].next)
1440                         swap_info[i].prio = p->prio--;
1441                 least_priority++;
1442         }
1443         nr_swap_pages -= p->pages;
1444         total_swap_pages -= p->pages;
1445         p->flags &= ~SWP_WRITEOK;
1446         spin_unlock(&swap_lock);
1447
1448         current->flags |= PF_SWAPOFF;
1449         err = try_to_unuse(type);
1450         current->flags &= ~PF_SWAPOFF;
1451
1452         if (err) {
1453                 /* re-insert swap space back into swap_list */
1454                 spin_lock(&swap_lock);
1455                 if (p->prio < 0)
1456                         p->prio = --least_priority;
1457                 prev = -1;
1458                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1459                         if (p->prio >= swap_info[i].prio)
1460                                 break;
1461                         prev = i;
1462                 }
1463                 p->next = i;
1464                 if (prev < 0)
1465                         swap_list.head = swap_list.next = p - swap_info;
1466                 else
1467                         swap_info[prev].next = p - swap_info;
1468                 nr_swap_pages += p->pages;
1469                 total_swap_pages += p->pages;
1470                 p->flags |= SWP_WRITEOK;
1471                 spin_unlock(&swap_lock);
1472                 goto out_dput;
1473         }
1474
1475         /* wait for any unplug function to finish */
1476         down_write(&swap_unplug_sem);
1477         up_write(&swap_unplug_sem);
1478
1479         destroy_swap_extents(p);
1480         mutex_lock(&swapon_mutex);
1481         spin_lock(&swap_lock);
1482         drain_mmlist();
1483
1484         /* wait for anyone still in scan_swap_map */
1485         p->highest_bit = 0;             /* cuts scans short */
1486         while (p->flags >= SWP_SCANNING) {
1487                 spin_unlock(&swap_lock);
1488                 schedule_timeout_uninterruptible(1);
1489                 spin_lock(&swap_lock);
1490         }
1491
1492         swap_file = p->swap_file;
1493         p->swap_file = NULL;
1494         p->max = 0;
1495         swap_map = p->swap_map;
1496         p->swap_map = NULL;
1497         p->flags = 0;
1498         spin_unlock(&swap_lock);
1499         mutex_unlock(&swapon_mutex);
1500         vfree(swap_map);
1501         /* Destroy swap account informatin */
1502         swap_cgroup_swapoff(type);
1503
1504         inode = mapping->host;
1505         if (S_ISBLK(inode->i_mode)) {
1506                 struct block_device *bdev = I_BDEV(inode);
1507                 set_blocksize(bdev, p->old_block_size);
1508                 bd_release(bdev);
1509         } else {
1510                 mutex_lock(&inode->i_mutex);
1511                 inode->i_flags &= ~S_SWAPFILE;
1512                 mutex_unlock(&inode->i_mutex);
1513         }
1514         filp_close(swap_file, NULL);
1515         err = 0;
1516
1517 out_dput:
1518         filp_close(victim, NULL);
1519 out:
1520         return err;
1521 }
1522
1523 #ifdef CONFIG_PROC_FS
1524 /* iterator */
1525 static void *swap_start(struct seq_file *swap, loff_t *pos)
1526 {
1527         struct swap_info_struct *ptr = swap_info;
1528         int i;
1529         loff_t l = *pos;
1530
1531         mutex_lock(&swapon_mutex);
1532
1533         if (!l)
1534                 return SEQ_START_TOKEN;
1535
1536         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1537                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1538                         continue;
1539                 if (!--l)
1540                         return ptr;
1541         }
1542
1543         return NULL;
1544 }
1545
1546 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1547 {
1548         struct swap_info_struct *ptr;
1549         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1550
1551         if (v == SEQ_START_TOKEN)
1552                 ptr = swap_info;
1553         else {
1554                 ptr = v;
1555                 ptr++;
1556         }
1557
1558         for (; ptr < endptr; ptr++) {
1559                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1560                         continue;
1561                 ++*pos;
1562                 return ptr;
1563         }
1564
1565         return NULL;
1566 }
1567
1568 static void swap_stop(struct seq_file *swap, void *v)
1569 {
1570         mutex_unlock(&swapon_mutex);
1571 }
1572
1573 static int swap_show(struct seq_file *swap, void *v)
1574 {
1575         struct swap_info_struct *ptr = v;
1576         struct file *file;
1577         int len;
1578
1579         if (ptr == SEQ_START_TOKEN) {
1580                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1581                 return 0;
1582         }
1583
1584         file = ptr->swap_file;
1585         len = seq_path(swap, &file->f_path, " \t\n\\");
1586         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1587                         len < 40 ? 40 - len : 1, " ",
1588                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1589                                 "partition" : "file\t",
1590                         ptr->pages << (PAGE_SHIFT - 10),
1591                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1592                         ptr->prio);
1593         return 0;
1594 }
1595
1596 static const struct seq_operations swaps_op = {
1597         .start =        swap_start,
1598         .next =         swap_next,
1599         .stop =         swap_stop,
1600         .show =         swap_show
1601 };
1602
1603 static int swaps_open(struct inode *inode, struct file *file)
1604 {
1605         return seq_open(file, &swaps_op);
1606 }
1607
1608 static const struct file_operations proc_swaps_operations = {
1609         .open           = swaps_open,
1610         .read           = seq_read,
1611         .llseek         = seq_lseek,
1612         .release        = seq_release,
1613 };
1614
1615 static int __init procswaps_init(void)
1616 {
1617         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1618         return 0;
1619 }
1620 __initcall(procswaps_init);
1621 #endif /* CONFIG_PROC_FS */
1622
1623 #ifdef MAX_SWAPFILES_CHECK
1624 static int __init max_swapfiles_check(void)
1625 {
1626         MAX_SWAPFILES_CHECK();
1627         return 0;
1628 }
1629 late_initcall(max_swapfiles_check);
1630 #endif
1631
1632 /*
1633  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1634  *
1635  * The swapon system call
1636  */
1637 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1638 {
1639         struct swap_info_struct * p;
1640         char *name = NULL;
1641         struct block_device *bdev = NULL;
1642         struct file *swap_file = NULL;
1643         struct address_space *mapping;
1644         unsigned int type;
1645         int i, prev;
1646         int error;
1647         union swap_header *swap_header = NULL;
1648         unsigned int nr_good_pages = 0;
1649         int nr_extents = 0;
1650         sector_t span;
1651         unsigned long maxpages = 1;
1652         unsigned long swapfilepages;
1653         unsigned short *swap_map = NULL;
1654         struct page *page = NULL;
1655         struct inode *inode = NULL;
1656         int did_down = 0;
1657
1658         if (!capable(CAP_SYS_ADMIN))
1659                 return -EPERM;
1660         spin_lock(&swap_lock);
1661         p = swap_info;
1662         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1663                 if (!(p->flags & SWP_USED))
1664                         break;
1665         error = -EPERM;
1666         if (type >= MAX_SWAPFILES) {
1667                 spin_unlock(&swap_lock);
1668                 goto out;
1669         }
1670         if (type >= nr_swapfiles)
1671                 nr_swapfiles = type+1;
1672         memset(p, 0, sizeof(*p));
1673         INIT_LIST_HEAD(&p->extent_list);
1674         p->flags = SWP_USED;
1675         p->next = -1;
1676         spin_unlock(&swap_lock);
1677         name = getname(specialfile);
1678         error = PTR_ERR(name);
1679         if (IS_ERR(name)) {
1680                 name = NULL;
1681                 goto bad_swap_2;
1682         }
1683         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1684         error = PTR_ERR(swap_file);
1685         if (IS_ERR(swap_file)) {
1686                 swap_file = NULL;
1687                 goto bad_swap_2;
1688         }
1689
1690         p->swap_file = swap_file;
1691         mapping = swap_file->f_mapping;
1692         inode = mapping->host;
1693
1694         error = -EBUSY;
1695         for (i = 0; i < nr_swapfiles; i++) {
1696                 struct swap_info_struct *q = &swap_info[i];
1697
1698                 if (i == type || !q->swap_file)
1699                         continue;
1700                 if (mapping == q->swap_file->f_mapping)
1701                         goto bad_swap;
1702         }
1703
1704         error = -EINVAL;
1705         if (S_ISBLK(inode->i_mode)) {
1706                 bdev = I_BDEV(inode);
1707                 error = bd_claim(bdev, sys_swapon);
1708                 if (error < 0) {
1709                         bdev = NULL;
1710                         error = -EINVAL;
1711                         goto bad_swap;
1712                 }
1713                 p->old_block_size = block_size(bdev);
1714                 error = set_blocksize(bdev, PAGE_SIZE);
1715                 if (error < 0)
1716                         goto bad_swap;
1717                 p->bdev = bdev;
1718         } else if (S_ISREG(inode->i_mode)) {
1719                 p->bdev = inode->i_sb->s_bdev;
1720                 mutex_lock(&inode->i_mutex);
1721                 did_down = 1;
1722                 if (IS_SWAPFILE(inode)) {
1723                         error = -EBUSY;
1724                         goto bad_swap;
1725                 }
1726         } else {
1727                 goto bad_swap;
1728         }
1729
1730         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1731
1732         /*
1733          * Read the swap header.
1734          */
1735         if (!mapping->a_ops->readpage) {
1736                 error = -EINVAL;
1737                 goto bad_swap;
1738         }
1739         page = read_mapping_page(mapping, 0, swap_file);
1740         if (IS_ERR(page)) {
1741                 error = PTR_ERR(page);
1742                 goto bad_swap;
1743         }
1744         swap_header = kmap(page);
1745
1746         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1747                 printk(KERN_ERR "Unable to find swap-space signature\n");
1748                 error = -EINVAL;
1749                 goto bad_swap;
1750         }
1751
1752         /* swap partition endianess hack... */
1753         if (swab32(swap_header->info.version) == 1) {
1754                 swab32s(&swap_header->info.version);
1755                 swab32s(&swap_header->info.last_page);
1756                 swab32s(&swap_header->info.nr_badpages);
1757                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1758                         swab32s(&swap_header->info.badpages[i]);
1759         }
1760         /* Check the swap header's sub-version */
1761         if (swap_header->info.version != 1) {
1762                 printk(KERN_WARNING
1763                        "Unable to handle swap header version %d\n",
1764                        swap_header->info.version);
1765                 error = -EINVAL;
1766                 goto bad_swap;
1767         }
1768
1769         p->lowest_bit  = 1;
1770         p->cluster_next = 1;
1771
1772         /*
1773          * Find out how many pages are allowed for a single swap
1774          * device. There are two limiting factors: 1) the number of
1775          * bits for the swap offset in the swp_entry_t type and
1776          * 2) the number of bits in the a swap pte as defined by
1777          * the different architectures. In order to find the
1778          * largest possible bit mask a swap entry with swap type 0
1779          * and swap offset ~0UL is created, encoded to a swap pte,
1780          * decoded to a swp_entry_t again and finally the swap
1781          * offset is extracted. This will mask all the bits from
1782          * the initial ~0UL mask that can't be encoded in either
1783          * the swp_entry_t or the architecture definition of a
1784          * swap pte.
1785          */
1786         maxpages = swp_offset(pte_to_swp_entry(
1787                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1788         if (maxpages > swap_header->info.last_page)
1789                 maxpages = swap_header->info.last_page;
1790         p->highest_bit = maxpages - 1;
1791
1792         error = -EINVAL;
1793         if (!maxpages)
1794                 goto bad_swap;
1795         if (swapfilepages && maxpages > swapfilepages) {
1796                 printk(KERN_WARNING
1797                        "Swap area shorter than signature indicates\n");
1798                 goto bad_swap;
1799         }
1800         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1801                 goto bad_swap;
1802         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1803                 goto bad_swap;
1804
1805         /* OK, set up the swap map and apply the bad block list */
1806         swap_map = vmalloc(maxpages * sizeof(short));
1807         if (!swap_map) {
1808                 error = -ENOMEM;
1809                 goto bad_swap;
1810         }
1811
1812         memset(swap_map, 0, maxpages * sizeof(short));
1813         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1814                 int page_nr = swap_header->info.badpages[i];
1815                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1816                         error = -EINVAL;
1817                         goto bad_swap;
1818                 }
1819                 swap_map[page_nr] = SWAP_MAP_BAD;
1820         }
1821
1822         error = swap_cgroup_swapon(type, maxpages);
1823         if (error)
1824                 goto bad_swap;
1825
1826         nr_good_pages = swap_header->info.last_page -
1827                         swap_header->info.nr_badpages -
1828                         1 /* header page */;
1829
1830         if (nr_good_pages) {
1831                 swap_map[0] = SWAP_MAP_BAD;
1832                 p->max = maxpages;
1833                 p->pages = nr_good_pages;
1834                 nr_extents = setup_swap_extents(p, &span);
1835                 if (nr_extents < 0) {
1836                         error = nr_extents;
1837                         goto bad_swap;
1838                 }
1839                 nr_good_pages = p->pages;
1840         }
1841         if (!nr_good_pages) {
1842                 printk(KERN_WARNING "Empty swap-file\n");
1843                 error = -EINVAL;
1844                 goto bad_swap;
1845         }
1846
1847         if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1848                 p->flags |= SWP_SOLIDSTATE;
1849                 p->cluster_next = 1 + (random32() % p->highest_bit);
1850         }
1851         if (discard_swap(p) == 0)
1852                 p->flags |= SWP_DISCARDABLE;
1853
1854         mutex_lock(&swapon_mutex);
1855         spin_lock(&swap_lock);
1856         if (swap_flags & SWAP_FLAG_PREFER)
1857                 p->prio =
1858                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1859         else
1860                 p->prio = --least_priority;
1861         p->swap_map = swap_map;
1862         p->flags |= SWP_WRITEOK;
1863         nr_swap_pages += nr_good_pages;
1864         total_swap_pages += nr_good_pages;
1865
1866         printk(KERN_INFO "Adding %uk swap on %s.  "
1867                         "Priority:%d extents:%d across:%lluk %s%s\n",
1868                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1869                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1870                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1871                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1872
1873         /* insert swap space into swap_list: */
1874         prev = -1;
1875         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1876                 if (p->prio >= swap_info[i].prio) {
1877                         break;
1878                 }
1879                 prev = i;
1880         }
1881         p->next = i;
1882         if (prev < 0) {
1883                 swap_list.head = swap_list.next = p - swap_info;
1884         } else {
1885                 swap_info[prev].next = p - swap_info;
1886         }
1887         spin_unlock(&swap_lock);
1888         mutex_unlock(&swapon_mutex);
1889         error = 0;
1890         goto out;
1891 bad_swap:
1892         if (bdev) {
1893                 set_blocksize(bdev, p->old_block_size);
1894                 bd_release(bdev);
1895         }
1896         destroy_swap_extents(p);
1897         swap_cgroup_swapoff(type);
1898 bad_swap_2:
1899         spin_lock(&swap_lock);
1900         p->swap_file = NULL;
1901         p->flags = 0;
1902         spin_unlock(&swap_lock);
1903         vfree(swap_map);
1904         if (swap_file)
1905                 filp_close(swap_file, NULL);
1906 out:
1907         if (page && !IS_ERR(page)) {
1908                 kunmap(page);
1909                 page_cache_release(page);
1910         }
1911         if (name)
1912                 putname(name);
1913         if (did_down) {
1914                 if (!error)
1915                         inode->i_flags |= S_SWAPFILE;
1916                 mutex_unlock(&inode->i_mutex);
1917         }
1918         return error;
1919 }
1920
1921 void si_swapinfo(struct sysinfo *val)
1922 {
1923         unsigned int i;
1924         unsigned long nr_to_be_unused = 0;
1925
1926         spin_lock(&swap_lock);
1927         for (i = 0; i < nr_swapfiles; i++) {
1928                 if (!(swap_info[i].flags & SWP_USED) ||
1929                      (swap_info[i].flags & SWP_WRITEOK))
1930                         continue;
1931                 nr_to_be_unused += swap_info[i].inuse_pages;
1932         }
1933         val->freeswap = nr_swap_pages + nr_to_be_unused;
1934         val->totalswap = total_swap_pages + nr_to_be_unused;
1935         spin_unlock(&swap_lock);
1936 }
1937
1938 /*
1939  * Verify that a swap entry is valid and increment its swap map count.
1940  *
1941  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1942  * "permanent", but will be reclaimed by the next swapoff.
1943  */
1944 int swap_duplicate(swp_entry_t entry)
1945 {
1946         struct swap_info_struct * p;
1947         unsigned long offset, type;
1948         int result = 0;
1949
1950         if (is_migration_entry(entry))
1951                 return 1;
1952
1953         type = swp_type(entry);
1954         if (type >= nr_swapfiles)
1955                 goto bad_file;
1956         p = type + swap_info;
1957         offset = swp_offset(entry);
1958
1959         spin_lock(&swap_lock);
1960         if (offset < p->max && p->swap_map[offset]) {
1961                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1962                         p->swap_map[offset]++;
1963                         result = 1;
1964                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1965                         if (swap_overflow++ < 5)
1966                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1967                         p->swap_map[offset] = SWAP_MAP_MAX;
1968                         result = 1;
1969                 }
1970         }
1971         spin_unlock(&swap_lock);
1972 out:
1973         return result;
1974
1975 bad_file:
1976         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1977         goto out;
1978 }
1979
1980 struct swap_info_struct *
1981 get_swap_info_struct(unsigned type)
1982 {
1983         return &swap_info[type];
1984 }
1985
1986 /*
1987  * swap_lock prevents swap_map being freed. Don't grab an extra
1988  * reference on the swaphandle, it doesn't matter if it becomes unused.
1989  */
1990 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1991 {
1992         struct swap_info_struct *si;
1993         int our_page_cluster = page_cluster;
1994         pgoff_t target, toff;
1995         pgoff_t base, end;
1996         int nr_pages = 0;
1997
1998         if (!our_page_cluster)  /* no readahead */
1999                 return 0;
2000
2001         si = &swap_info[swp_type(entry)];
2002         target = swp_offset(entry);
2003         base = (target >> our_page_cluster) << our_page_cluster;
2004         end = base + (1 << our_page_cluster);
2005         if (!base)              /* first page is swap header */
2006                 base++;
2007
2008         spin_lock(&swap_lock);
2009         if (end > si->max)      /* don't go beyond end of map */
2010                 end = si->max;
2011
2012         /* Count contiguous allocated slots above our target */
2013         for (toff = target; ++toff < end; nr_pages++) {
2014                 /* Don't read in free or bad pages */
2015                 if (!si->swap_map[toff])
2016                         break;
2017                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2018                         break;
2019         }
2020         /* Count contiguous allocated slots below our target */
2021         for (toff = target; --toff >= base; nr_pages++) {
2022                 /* Don't read in free or bad pages */
2023                 if (!si->swap_map[toff])
2024                         break;
2025                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2026                         break;
2027         }
2028         spin_unlock(&swap_lock);
2029
2030         /*
2031          * Indicate starting offset, and return number of pages to get:
2032          * if only 1, say 0, since there's then no readahead to be done.
2033          */
2034         *offset = ++toff;
2035         return nr_pages? ++nr_pages: 0;
2036 }