ca75b9e7c09f6efd864fe78f45ec95627bef903d
[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/writeback.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/init.h>
23 #include <linux/module.h>
24 #include <linux/rmap.h>
25 #include <linux/security.h>
26 #include <linux/backing-dev.h>
27 #include <linux/mutex.h>
28 #include <linux/capability.h>
29 #include <linux/syscalls.h>
30 #include <linux/memcontrol.h>
31
32 #include <asm/pgtable.h>
33 #include <asm/tlbflush.h>
34 #include <linux/swapops.h>
35
36 static DEFINE_SPINLOCK(swap_lock);
37 static unsigned int nr_swapfiles;
38 long nr_swap_pages;
39 long total_swap_pages;
40 static int swap_overflow;
41 static int least_priority;
42
43 static const char Bad_file[] = "Bad swap file entry ";
44 static const char Unused_file[] = "Unused swap file entry ";
45 static const char Bad_offset[] = "Bad swap offset entry ";
46 static const char Unused_offset[] = "Unused swap offset entry ";
47
48 static struct swap_list_t swap_list = {-1, -1};
49
50 static struct swap_info_struct swap_info[MAX_SWAPFILES];
51
52 static DEFINE_MUTEX(swapon_mutex);
53
54 /*
55  * We need this because the bdev->unplug_fn can sleep and we cannot
56  * hold swap_lock while calling the unplug_fn. And swap_lock
57  * cannot be turned into a mutex.
58  */
59 static DECLARE_RWSEM(swap_unplug_sem);
60
61 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
62 {
63         swp_entry_t entry;
64
65         down_read(&swap_unplug_sem);
66         entry.val = page_private(page);
67         if (PageSwapCache(page)) {
68                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
69                 struct backing_dev_info *bdi;
70
71                 /*
72                  * If the page is removed from swapcache from under us (with a
73                  * racy try_to_unuse/swapoff) we need an additional reference
74                  * count to avoid reading garbage from page_private(page) above.
75                  * If the WARN_ON triggers during a swapoff it maybe the race
76                  * condition and it's harmless. However if it triggers without
77                  * swapoff it signals a problem.
78                  */
79                 WARN_ON(page_count(page) <= 1);
80
81                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
82                 blk_run_backing_dev(bdi, page);
83         }
84         up_read(&swap_unplug_sem);
85 }
86
87 /*
88  * swapon tell device that all the old swap contents can be discarded,
89  * to allow the swap device to optimize its wear-levelling.
90  */
91 static int discard_swap(struct swap_info_struct *si)
92 {
93         struct swap_extent *se;
94         int err = 0;
95
96         list_for_each_entry(se, &si->extent_list, list) {
97                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
98                 pgoff_t nr_blocks = se->nr_pages << (PAGE_SHIFT - 9);
99
100                 if (se->start_page == 0) {
101                         /* Do not discard the swap header page! */
102                         start_block += 1 << (PAGE_SHIFT - 9);
103                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
104                         if (!nr_blocks)
105                                 continue;
106                 }
107
108                 err = blkdev_issue_discard(si->bdev, start_block,
109                                                 nr_blocks, GFP_KERNEL);
110                 if (err)
111                         break;
112
113                 cond_resched();
114         }
115         return err;             /* That will often be -EOPNOTSUPP */
116 }
117
118 /*
119  * swap allocation tell device that a cluster of swap can now be discarded,
120  * to allow the swap device to optimize its wear-levelling.
121  */
122 static void discard_swap_cluster(struct swap_info_struct *si,
123                                  pgoff_t start_page, pgoff_t nr_pages)
124 {
125         struct swap_extent *se = si->curr_swap_extent;
126         int found_extent = 0;
127
128         while (nr_pages) {
129                 struct list_head *lh;
130
131                 if (se->start_page <= start_page &&
132                     start_page < se->start_page + se->nr_pages) {
133                         pgoff_t offset = start_page - se->start_page;
134                         sector_t start_block = se->start_block + offset;
135                         pgoff_t nr_blocks = se->nr_pages - offset;
136
137                         if (nr_blocks > nr_pages)
138                                 nr_blocks = nr_pages;
139                         start_page += nr_blocks;
140                         nr_pages -= nr_blocks;
141
142                         if (!found_extent++)
143                                 si->curr_swap_extent = se;
144
145                         start_block <<= PAGE_SHIFT - 9;
146                         nr_blocks <<= PAGE_SHIFT - 9;
147                         if (blkdev_issue_discard(si->bdev, start_block,
148                                                         nr_blocks, GFP_NOIO))
149                                 break;
150                 }
151
152                 lh = se->list.next;
153                 if (lh == &si->extent_list)
154                         lh = lh->next;
155                 se = list_entry(lh, struct swap_extent, list);
156         }
157 }
158
159 static int wait_for_discard(void *word)
160 {
161         schedule();
162         return 0;
163 }
164
165 #define SWAPFILE_CLUSTER        256
166 #define LATENCY_LIMIT           256
167
168 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
169 {
170         unsigned long offset;
171         unsigned long last_in_cluster = 0;
172         int latency_ration = LATENCY_LIMIT;
173         int found_free_cluster = 0;
174
175         /*
176          * We try to cluster swap pages by allocating them sequentially
177          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
178          * way, however, we resort to first-free allocation, starting
179          * a new cluster.  This prevents us from scattering swap pages
180          * all over the entire swap partition, so that we reduce
181          * overall disk seek times between swap pages.  -- sct
182          * But we do now try to find an empty cluster.  -Andrea
183          */
184
185         si->flags += SWP_SCANNING;
186         offset = si->cluster_next;
187
188         if (unlikely(!si->cluster_nr--)) {
189                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
190                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
191                         goto checks;
192                 }
193                 if (si->flags & SWP_DISCARDABLE) {
194                         /*
195                          * Start range check on racing allocations, in case
196                          * they overlap the cluster we eventually decide on
197                          * (we scan without swap_lock to allow preemption).
198                          * It's hardly conceivable that cluster_nr could be
199                          * wrapped during our scan, but don't depend on it.
200                          */
201                         if (si->lowest_alloc)
202                                 goto checks;
203                         si->lowest_alloc = si->max;
204                         si->highest_alloc = 0;
205                 }
206                 spin_unlock(&swap_lock);
207
208                 offset = si->lowest_bit;
209                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
210
211                 /* Locate the first empty (unaligned) cluster */
212                 for (; last_in_cluster <= si->highest_bit; offset++) {
213                         if (si->swap_map[offset])
214                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
215                         else if (offset == last_in_cluster) {
216                                 spin_lock(&swap_lock);
217                                 offset -= SWAPFILE_CLUSTER - 1;
218                                 si->cluster_next = offset;
219                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
220                                 found_free_cluster = 1;
221                                 goto checks;
222                         }
223                         if (unlikely(--latency_ration < 0)) {
224                                 cond_resched();
225                                 latency_ration = LATENCY_LIMIT;
226                         }
227                 }
228
229                 offset = si->lowest_bit;
230                 spin_lock(&swap_lock);
231                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
232                 si->lowest_alloc = 0;
233         }
234
235 checks:
236         if (!(si->flags & SWP_WRITEOK))
237                 goto no_page;
238         if (!si->highest_bit)
239                 goto no_page;
240         if (offset > si->highest_bit)
241                 offset = si->lowest_bit;
242         if (si->swap_map[offset])
243                 goto scan;
244
245         if (offset == si->lowest_bit)
246                 si->lowest_bit++;
247         if (offset == si->highest_bit)
248                 si->highest_bit--;
249         si->inuse_pages++;
250         if (si->inuse_pages == si->pages) {
251                 si->lowest_bit = si->max;
252                 si->highest_bit = 0;
253         }
254         si->swap_map[offset] = 1;
255         si->cluster_next = offset + 1;
256         si->flags -= SWP_SCANNING;
257
258         if (si->lowest_alloc) {
259                 /*
260                  * Only set when SWP_DISCARDABLE, and there's a scan
261                  * for a free cluster in progress or just completed.
262                  */
263                 if (found_free_cluster) {
264                         /*
265                          * To optimize wear-levelling, discard the
266                          * old data of the cluster, taking care not to
267                          * discard any of its pages that have already
268                          * been allocated by racing tasks (offset has
269                          * already stepped over any at the beginning).
270                          */
271                         if (offset < si->highest_alloc &&
272                             si->lowest_alloc <= last_in_cluster)
273                                 last_in_cluster = si->lowest_alloc - 1;
274                         si->flags |= SWP_DISCARDING;
275                         spin_unlock(&swap_lock);
276
277                         if (offset < last_in_cluster)
278                                 discard_swap_cluster(si, offset,
279                                         last_in_cluster - offset + 1);
280
281                         spin_lock(&swap_lock);
282                         si->lowest_alloc = 0;
283                         si->flags &= ~SWP_DISCARDING;
284
285                         smp_mb();       /* wake_up_bit advises this */
286                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
287
288                 } else if (si->flags & SWP_DISCARDING) {
289                         /*
290                          * Delay using pages allocated by racing tasks
291                          * until the whole discard has been issued. We
292                          * could defer that delay until swap_writepage,
293                          * but it's easier to keep this self-contained.
294                          */
295                         spin_unlock(&swap_lock);
296                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
297                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
298                         spin_lock(&swap_lock);
299                 } else {
300                         /*
301                          * Note pages allocated by racing tasks while
302                          * scan for a free cluster is in progress, so
303                          * that its final discard can exclude them.
304                          */
305                         if (offset < si->lowest_alloc)
306                                 si->lowest_alloc = offset;
307                         if (offset > si->highest_alloc)
308                                 si->highest_alloc = offset;
309                 }
310         }
311         return offset;
312
313 scan:
314         spin_unlock(&swap_lock);
315         while (++offset <= si->highest_bit) {
316                 if (!si->swap_map[offset]) {
317                         spin_lock(&swap_lock);
318                         goto checks;
319                 }
320                 if (unlikely(--latency_ration < 0)) {
321                         cond_resched();
322                         latency_ration = LATENCY_LIMIT;
323                 }
324         }
325         spin_lock(&swap_lock);
326         goto checks;
327
328 no_page:
329         si->flags -= SWP_SCANNING;
330         return 0;
331 }
332
333 swp_entry_t get_swap_page(void)
334 {
335         struct swap_info_struct *si;
336         pgoff_t offset;
337         int type, next;
338         int wrapped = 0;
339
340         spin_lock(&swap_lock);
341         if (nr_swap_pages <= 0)
342                 goto noswap;
343         nr_swap_pages--;
344
345         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
346                 si = swap_info + type;
347                 next = si->next;
348                 if (next < 0 ||
349                     (!wrapped && si->prio != swap_info[next].prio)) {
350                         next = swap_list.head;
351                         wrapped++;
352                 }
353
354                 if (!si->highest_bit)
355                         continue;
356                 if (!(si->flags & SWP_WRITEOK))
357                         continue;
358
359                 swap_list.next = next;
360                 offset = scan_swap_map(si);
361                 if (offset) {
362                         spin_unlock(&swap_lock);
363                         return swp_entry(type, offset);
364                 }
365                 next = swap_list.next;
366         }
367
368         nr_swap_pages++;
369 noswap:
370         spin_unlock(&swap_lock);
371         return (swp_entry_t) {0};
372 }
373
374 swp_entry_t get_swap_page_of_type(int type)
375 {
376         struct swap_info_struct *si;
377         pgoff_t offset;
378
379         spin_lock(&swap_lock);
380         si = swap_info + type;
381         if (si->flags & SWP_WRITEOK) {
382                 nr_swap_pages--;
383                 offset = scan_swap_map(si);
384                 if (offset) {
385                         spin_unlock(&swap_lock);
386                         return swp_entry(type, offset);
387                 }
388                 nr_swap_pages++;
389         }
390         spin_unlock(&swap_lock);
391         return (swp_entry_t) {0};
392 }
393
394 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
395 {
396         struct swap_info_struct * p;
397         unsigned long offset, type;
398
399         if (!entry.val)
400                 goto out;
401         type = swp_type(entry);
402         if (type >= nr_swapfiles)
403                 goto bad_nofile;
404         p = & swap_info[type];
405         if (!(p->flags & SWP_USED))
406                 goto bad_device;
407         offset = swp_offset(entry);
408         if (offset >= p->max)
409                 goto bad_offset;
410         if (!p->swap_map[offset])
411                 goto bad_free;
412         spin_lock(&swap_lock);
413         return p;
414
415 bad_free:
416         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
417         goto out;
418 bad_offset:
419         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
420         goto out;
421 bad_device:
422         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
423         goto out;
424 bad_nofile:
425         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
426 out:
427         return NULL;
428 }
429
430 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
431 {
432         int count = p->swap_map[offset];
433
434         if (count < SWAP_MAP_MAX) {
435                 count--;
436                 p->swap_map[offset] = count;
437                 if (!count) {
438                         if (offset < p->lowest_bit)
439                                 p->lowest_bit = offset;
440                         if (offset > p->highest_bit)
441                                 p->highest_bit = offset;
442                         if (p->prio > swap_info[swap_list.next].prio)
443                                 swap_list.next = p - swap_info;
444                         nr_swap_pages++;
445                         p->inuse_pages--;
446                 }
447         }
448         return count;
449 }
450
451 /*
452  * Caller has made sure that the swapdevice corresponding to entry
453  * is still around or has not been recycled.
454  */
455 void swap_free(swp_entry_t entry)
456 {
457         struct swap_info_struct * p;
458
459         p = swap_info_get(entry);
460         if (p) {
461                 swap_entry_free(p, swp_offset(entry));
462                 spin_unlock(&swap_lock);
463         }
464 }
465
466 /*
467  * How many references to page are currently swapped out?
468  */
469 static inline int page_swapcount(struct page *page)
470 {
471         int count = 0;
472         struct swap_info_struct *p;
473         swp_entry_t entry;
474
475         entry.val = page_private(page);
476         p = swap_info_get(entry);
477         if (p) {
478                 /* Subtract the 1 for the swap cache itself */
479                 count = p->swap_map[swp_offset(entry)] - 1;
480                 spin_unlock(&swap_lock);
481         }
482         return count;
483 }
484
485 /*
486  * We can write to an anon page without COW if there are no other references
487  * to it.  And as a side-effect, free up its swap: because the old content
488  * on disk will never be read, and seeking back there to write new content
489  * later would only waste time away from clustering.
490  */
491 int reuse_swap_page(struct page *page)
492 {
493         int count;
494
495         VM_BUG_ON(!PageLocked(page));
496         count = page_mapcount(page);
497         if (count <= 1 && PageSwapCache(page)) {
498                 count += page_swapcount(page);
499                 if (count == 1 && !PageWriteback(page)) {
500                         delete_from_swap_cache(page);
501                         SetPageDirty(page);
502                 }
503         }
504         return count == 1;
505 }
506
507 /*
508  * If swap is getting full, or if there are no more mappings of this page,
509  * then try_to_free_swap is called to free its swap space.
510  */
511 int try_to_free_swap(struct page *page)
512 {
513         VM_BUG_ON(!PageLocked(page));
514
515         if (!PageSwapCache(page))
516                 return 0;
517         if (PageWriteback(page))
518                 return 0;
519         if (page_swapcount(page))
520                 return 0;
521
522         delete_from_swap_cache(page);
523         SetPageDirty(page);
524         return 1;
525 }
526
527 /*
528  * Free the swap entry like above, but also try to
529  * free the page cache entry if it is the last user.
530  */
531 void free_swap_and_cache(swp_entry_t entry)
532 {
533         struct swap_info_struct * p;
534         struct page *page = NULL;
535
536         if (is_migration_entry(entry))
537                 return;
538
539         p = swap_info_get(entry);
540         if (p) {
541                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
542                         page = find_get_page(&swapper_space, entry.val);
543                         if (page && !trylock_page(page)) {
544                                 page_cache_release(page);
545                                 page = NULL;
546                         }
547                 }
548                 spin_unlock(&swap_lock);
549         }
550         if (page) {
551                 /*
552                  * Not mapped elsewhere, or swap space full? Free it!
553                  * Also recheck PageSwapCache now page is locked (above).
554                  */
555                 if (PageSwapCache(page) && !PageWriteback(page) &&
556                                 (!page_mapped(page) || vm_swap_full())) {
557                         delete_from_swap_cache(page);
558                         SetPageDirty(page);
559                 }
560                 unlock_page(page);
561                 page_cache_release(page);
562         }
563 }
564
565 #ifdef CONFIG_HIBERNATION
566 /*
567  * Find the swap type that corresponds to given device (if any).
568  *
569  * @offset - number of the PAGE_SIZE-sized block of the device, starting
570  * from 0, in which the swap header is expected to be located.
571  *
572  * This is needed for the suspend to disk (aka swsusp).
573  */
574 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
575 {
576         struct block_device *bdev = NULL;
577         int i;
578
579         if (device)
580                 bdev = bdget(device);
581
582         spin_lock(&swap_lock);
583         for (i = 0; i < nr_swapfiles; i++) {
584                 struct swap_info_struct *sis = swap_info + i;
585
586                 if (!(sis->flags & SWP_WRITEOK))
587                         continue;
588
589                 if (!bdev) {
590                         if (bdev_p)
591                                 *bdev_p = sis->bdev;
592
593                         spin_unlock(&swap_lock);
594                         return i;
595                 }
596                 if (bdev == sis->bdev) {
597                         struct swap_extent *se;
598
599                         se = list_entry(sis->extent_list.next,
600                                         struct swap_extent, list);
601                         if (se->start_block == offset) {
602                                 if (bdev_p)
603                                         *bdev_p = sis->bdev;
604
605                                 spin_unlock(&swap_lock);
606                                 bdput(bdev);
607                                 return i;
608                         }
609                 }
610         }
611         spin_unlock(&swap_lock);
612         if (bdev)
613                 bdput(bdev);
614
615         return -ENODEV;
616 }
617
618 /*
619  * Return either the total number of swap pages of given type, or the number
620  * of free pages of that type (depending on @free)
621  *
622  * This is needed for software suspend
623  */
624 unsigned int count_swap_pages(int type, int free)
625 {
626         unsigned int n = 0;
627
628         if (type < nr_swapfiles) {
629                 spin_lock(&swap_lock);
630                 if (swap_info[type].flags & SWP_WRITEOK) {
631                         n = swap_info[type].pages;
632                         if (free)
633                                 n -= swap_info[type].inuse_pages;
634                 }
635                 spin_unlock(&swap_lock);
636         }
637         return n;
638 }
639 #endif
640
641 /*
642  * No need to decide whether this PTE shares the swap entry with others,
643  * just let do_wp_page work it out if a write is requested later - to
644  * force COW, vm_page_prot omits write permission from any private vma.
645  */
646 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
647                 unsigned long addr, swp_entry_t entry, struct page *page)
648 {
649         spinlock_t *ptl;
650         pte_t *pte;
651         int ret = 1;
652
653         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
654                 ret = -ENOMEM;
655
656         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
657         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
658                 if (ret > 0)
659                         mem_cgroup_uncharge_page(page);
660                 ret = 0;
661                 goto out;
662         }
663
664         inc_mm_counter(vma->vm_mm, anon_rss);
665         get_page(page);
666         set_pte_at(vma->vm_mm, addr, pte,
667                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
668         page_add_anon_rmap(page, vma, addr);
669         swap_free(entry);
670         /*
671          * Move the page to the active list so it is not
672          * immediately swapped out again after swapon.
673          */
674         activate_page(page);
675 out:
676         pte_unmap_unlock(pte, ptl);
677         return ret;
678 }
679
680 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
681                                 unsigned long addr, unsigned long end,
682                                 swp_entry_t entry, struct page *page)
683 {
684         pte_t swp_pte = swp_entry_to_pte(entry);
685         pte_t *pte;
686         int ret = 0;
687
688         /*
689          * We don't actually need pte lock while scanning for swp_pte: since
690          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
691          * page table while we're scanning; though it could get zapped, and on
692          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
693          * of unmatched parts which look like swp_pte, so unuse_pte must
694          * recheck under pte lock.  Scanning without pte lock lets it be
695          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
696          */
697         pte = pte_offset_map(pmd, addr);
698         do {
699                 /*
700                  * swapoff spends a _lot_ of time in this loop!
701                  * Test inline before going to call unuse_pte.
702                  */
703                 if (unlikely(pte_same(*pte, swp_pte))) {
704                         pte_unmap(pte);
705                         ret = unuse_pte(vma, pmd, addr, entry, page);
706                         if (ret)
707                                 goto out;
708                         pte = pte_offset_map(pmd, addr);
709                 }
710         } while (pte++, addr += PAGE_SIZE, addr != end);
711         pte_unmap(pte - 1);
712 out:
713         return ret;
714 }
715
716 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
717                                 unsigned long addr, unsigned long end,
718                                 swp_entry_t entry, struct page *page)
719 {
720         pmd_t *pmd;
721         unsigned long next;
722         int ret;
723
724         pmd = pmd_offset(pud, addr);
725         do {
726                 next = pmd_addr_end(addr, end);
727                 if (pmd_none_or_clear_bad(pmd))
728                         continue;
729                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
730                 if (ret)
731                         return ret;
732         } while (pmd++, addr = next, addr != end);
733         return 0;
734 }
735
736 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
737                                 unsigned long addr, unsigned long end,
738                                 swp_entry_t entry, struct page *page)
739 {
740         pud_t *pud;
741         unsigned long next;
742         int ret;
743
744         pud = pud_offset(pgd, addr);
745         do {
746                 next = pud_addr_end(addr, end);
747                 if (pud_none_or_clear_bad(pud))
748                         continue;
749                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
750                 if (ret)
751                         return ret;
752         } while (pud++, addr = next, addr != end);
753         return 0;
754 }
755
756 static int unuse_vma(struct vm_area_struct *vma,
757                                 swp_entry_t entry, struct page *page)
758 {
759         pgd_t *pgd;
760         unsigned long addr, end, next;
761         int ret;
762
763         if (page->mapping) {
764                 addr = page_address_in_vma(page, vma);
765                 if (addr == -EFAULT)
766                         return 0;
767                 else
768                         end = addr + PAGE_SIZE;
769         } else {
770                 addr = vma->vm_start;
771                 end = vma->vm_end;
772         }
773
774         pgd = pgd_offset(vma->vm_mm, addr);
775         do {
776                 next = pgd_addr_end(addr, end);
777                 if (pgd_none_or_clear_bad(pgd))
778                         continue;
779                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
780                 if (ret)
781                         return ret;
782         } while (pgd++, addr = next, addr != end);
783         return 0;
784 }
785
786 static int unuse_mm(struct mm_struct *mm,
787                                 swp_entry_t entry, struct page *page)
788 {
789         struct vm_area_struct *vma;
790         int ret = 0;
791
792         if (!down_read_trylock(&mm->mmap_sem)) {
793                 /*
794                  * Activate page so shrink_inactive_list is unlikely to unmap
795                  * its ptes while lock is dropped, so swapoff can make progress.
796                  */
797                 activate_page(page);
798                 unlock_page(page);
799                 down_read(&mm->mmap_sem);
800                 lock_page(page);
801         }
802         for (vma = mm->mmap; vma; vma = vma->vm_next) {
803                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
804                         break;
805         }
806         up_read(&mm->mmap_sem);
807         return (ret < 0)? ret: 0;
808 }
809
810 /*
811  * Scan swap_map from current position to next entry still in use.
812  * Recycle to start on reaching the end, returning 0 when empty.
813  */
814 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
815                                         unsigned int prev)
816 {
817         unsigned int max = si->max;
818         unsigned int i = prev;
819         int count;
820
821         /*
822          * No need for swap_lock here: we're just looking
823          * for whether an entry is in use, not modifying it; false
824          * hits are okay, and sys_swapoff() has already prevented new
825          * allocations from this area (while holding swap_lock).
826          */
827         for (;;) {
828                 if (++i >= max) {
829                         if (!prev) {
830                                 i = 0;
831                                 break;
832                         }
833                         /*
834                          * No entries in use at top of swap_map,
835                          * loop back to start and recheck there.
836                          */
837                         max = prev + 1;
838                         prev = 0;
839                         i = 1;
840                 }
841                 count = si->swap_map[i];
842                 if (count && count != SWAP_MAP_BAD)
843                         break;
844         }
845         return i;
846 }
847
848 /*
849  * We completely avoid races by reading each swap page in advance,
850  * and then search for the process using it.  All the necessary
851  * page table adjustments can then be made atomically.
852  */
853 static int try_to_unuse(unsigned int type)
854 {
855         struct swap_info_struct * si = &swap_info[type];
856         struct mm_struct *start_mm;
857         unsigned short *swap_map;
858         unsigned short swcount;
859         struct page *page;
860         swp_entry_t entry;
861         unsigned int i = 0;
862         int retval = 0;
863         int reset_overflow = 0;
864         int shmem;
865
866         /*
867          * When searching mms for an entry, a good strategy is to
868          * start at the first mm we freed the previous entry from
869          * (though actually we don't notice whether we or coincidence
870          * freed the entry).  Initialize this start_mm with a hold.
871          *
872          * A simpler strategy would be to start at the last mm we
873          * freed the previous entry from; but that would take less
874          * advantage of mmlist ordering, which clusters forked mms
875          * together, child after parent.  If we race with dup_mmap(), we
876          * prefer to resolve parent before child, lest we miss entries
877          * duplicated after we scanned child: using last mm would invert
878          * that.  Though it's only a serious concern when an overflowed
879          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
880          */
881         start_mm = &init_mm;
882         atomic_inc(&init_mm.mm_users);
883
884         /*
885          * Keep on scanning until all entries have gone.  Usually,
886          * one pass through swap_map is enough, but not necessarily:
887          * there are races when an instance of an entry might be missed.
888          */
889         while ((i = find_next_to_unuse(si, i)) != 0) {
890                 if (signal_pending(current)) {
891                         retval = -EINTR;
892                         break;
893                 }
894
895                 /*
896                  * Get a page for the entry, using the existing swap
897                  * cache page if there is one.  Otherwise, get a clean
898                  * page and read the swap into it.
899                  */
900                 swap_map = &si->swap_map[i];
901                 entry = swp_entry(type, i);
902                 page = read_swap_cache_async(entry,
903                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
904                 if (!page) {
905                         /*
906                          * Either swap_duplicate() failed because entry
907                          * has been freed independently, and will not be
908                          * reused since sys_swapoff() already disabled
909                          * allocation from here, or alloc_page() failed.
910                          */
911                         if (!*swap_map)
912                                 continue;
913                         retval = -ENOMEM;
914                         break;
915                 }
916
917                 /*
918                  * Don't hold on to start_mm if it looks like exiting.
919                  */
920                 if (atomic_read(&start_mm->mm_users) == 1) {
921                         mmput(start_mm);
922                         start_mm = &init_mm;
923                         atomic_inc(&init_mm.mm_users);
924                 }
925
926                 /*
927                  * Wait for and lock page.  When do_swap_page races with
928                  * try_to_unuse, do_swap_page can handle the fault much
929                  * faster than try_to_unuse can locate the entry.  This
930                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
931                  * defer to do_swap_page in such a case - in some tests,
932                  * do_swap_page and try_to_unuse repeatedly compete.
933                  */
934                 wait_on_page_locked(page);
935                 wait_on_page_writeback(page);
936                 lock_page(page);
937                 wait_on_page_writeback(page);
938
939                 /*
940                  * Remove all references to entry.
941                  * Whenever we reach init_mm, there's no address space
942                  * to search, but use it as a reminder to search shmem.
943                  */
944                 shmem = 0;
945                 swcount = *swap_map;
946                 if (swcount > 1) {
947                         if (start_mm == &init_mm)
948                                 shmem = shmem_unuse(entry, page);
949                         else
950                                 retval = unuse_mm(start_mm, entry, page);
951                 }
952                 if (*swap_map > 1) {
953                         int set_start_mm = (*swap_map >= swcount);
954                         struct list_head *p = &start_mm->mmlist;
955                         struct mm_struct *new_start_mm = start_mm;
956                         struct mm_struct *prev_mm = start_mm;
957                         struct mm_struct *mm;
958
959                         atomic_inc(&new_start_mm->mm_users);
960                         atomic_inc(&prev_mm->mm_users);
961                         spin_lock(&mmlist_lock);
962                         while (*swap_map > 1 && !retval && !shmem &&
963                                         (p = p->next) != &start_mm->mmlist) {
964                                 mm = list_entry(p, struct mm_struct, mmlist);
965                                 if (!atomic_inc_not_zero(&mm->mm_users))
966                                         continue;
967                                 spin_unlock(&mmlist_lock);
968                                 mmput(prev_mm);
969                                 prev_mm = mm;
970
971                                 cond_resched();
972
973                                 swcount = *swap_map;
974                                 if (swcount <= 1)
975                                         ;
976                                 else if (mm == &init_mm) {
977                                         set_start_mm = 1;
978                                         shmem = shmem_unuse(entry, page);
979                                 } else
980                                         retval = unuse_mm(mm, entry, page);
981                                 if (set_start_mm && *swap_map < swcount) {
982                                         mmput(new_start_mm);
983                                         atomic_inc(&mm->mm_users);
984                                         new_start_mm = mm;
985                                         set_start_mm = 0;
986                                 }
987                                 spin_lock(&mmlist_lock);
988                         }
989                         spin_unlock(&mmlist_lock);
990                         mmput(prev_mm);
991                         mmput(start_mm);
992                         start_mm = new_start_mm;
993                 }
994                 if (shmem) {
995                         /* page has already been unlocked and released */
996                         if (shmem > 0)
997                                 continue;
998                         retval = shmem;
999                         break;
1000                 }
1001                 if (retval) {
1002                         unlock_page(page);
1003                         page_cache_release(page);
1004                         break;
1005                 }
1006
1007                 /*
1008                  * How could swap count reach 0x7fff when the maximum
1009                  * pid is 0x7fff, and there's no way to repeat a swap
1010                  * page within an mm (except in shmem, where it's the
1011                  * shared object which takes the reference count)?
1012                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
1013                  *
1014                  * If that's wrong, then we should worry more about
1015                  * exit_mmap() and do_munmap() cases described above:
1016                  * we might be resetting SWAP_MAP_MAX too early here.
1017                  * We know "Undead"s can happen, they're okay, so don't
1018                  * report them; but do report if we reset SWAP_MAP_MAX.
1019                  */
1020                 if (*swap_map == SWAP_MAP_MAX) {
1021                         spin_lock(&swap_lock);
1022                         *swap_map = 1;
1023                         spin_unlock(&swap_lock);
1024                         reset_overflow = 1;
1025                 }
1026
1027                 /*
1028                  * If a reference remains (rare), we would like to leave
1029                  * the page in the swap cache; but try_to_unmap could
1030                  * then re-duplicate the entry once we drop page lock,
1031                  * so we might loop indefinitely; also, that page could
1032                  * not be swapped out to other storage meanwhile.  So:
1033                  * delete from cache even if there's another reference,
1034                  * after ensuring that the data has been saved to disk -
1035                  * since if the reference remains (rarer), it will be
1036                  * read from disk into another page.  Splitting into two
1037                  * pages would be incorrect if swap supported "shared
1038                  * private" pages, but they are handled by tmpfs files.
1039                  */
1040                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
1041                         struct writeback_control wbc = {
1042                                 .sync_mode = WB_SYNC_NONE,
1043                         };
1044
1045                         swap_writepage(page, &wbc);
1046                         lock_page(page);
1047                         wait_on_page_writeback(page);
1048                 }
1049
1050                 /*
1051                  * It is conceivable that a racing task removed this page from
1052                  * swap cache just before we acquired the page lock at the top,
1053                  * or while we dropped it in unuse_mm().  The page might even
1054                  * be back in swap cache on another swap area: that we must not
1055                  * delete, since it may not have been written out to swap yet.
1056                  */
1057                 if (PageSwapCache(page) &&
1058                     likely(page_private(page) == entry.val))
1059                         delete_from_swap_cache(page);
1060
1061                 /*
1062                  * So we could skip searching mms once swap count went
1063                  * to 1, we did not mark any present ptes as dirty: must
1064                  * mark page dirty so shrink_page_list will preserve it.
1065                  */
1066                 SetPageDirty(page);
1067                 unlock_page(page);
1068                 page_cache_release(page);
1069
1070                 /*
1071                  * Make sure that we aren't completely killing
1072                  * interactive performance.
1073                  */
1074                 cond_resched();
1075         }
1076
1077         mmput(start_mm);
1078         if (reset_overflow) {
1079                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1080                 swap_overflow = 0;
1081         }
1082         return retval;
1083 }
1084
1085 /*
1086  * After a successful try_to_unuse, if no swap is now in use, we know
1087  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1088  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1089  * added to the mmlist just after page_duplicate - before would be racy.
1090  */
1091 static void drain_mmlist(void)
1092 {
1093         struct list_head *p, *next;
1094         unsigned int i;
1095
1096         for (i = 0; i < nr_swapfiles; i++)
1097                 if (swap_info[i].inuse_pages)
1098                         return;
1099         spin_lock(&mmlist_lock);
1100         list_for_each_safe(p, next, &init_mm.mmlist)
1101                 list_del_init(p);
1102         spin_unlock(&mmlist_lock);
1103 }
1104
1105 /*
1106  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1107  * corresponds to page offset `offset'.
1108  */
1109 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1110 {
1111         struct swap_extent *se = sis->curr_swap_extent;
1112         struct swap_extent *start_se = se;
1113
1114         for ( ; ; ) {
1115                 struct list_head *lh;
1116
1117                 if (se->start_page <= offset &&
1118                                 offset < (se->start_page + se->nr_pages)) {
1119                         return se->start_block + (offset - se->start_page);
1120                 }
1121                 lh = se->list.next;
1122                 if (lh == &sis->extent_list)
1123                         lh = lh->next;
1124                 se = list_entry(lh, struct swap_extent, list);
1125                 sis->curr_swap_extent = se;
1126                 BUG_ON(se == start_se);         /* It *must* be present */
1127         }
1128 }
1129
1130 #ifdef CONFIG_HIBERNATION
1131 /*
1132  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1133  * corresponding to given index in swap_info (swap type).
1134  */
1135 sector_t swapdev_block(int swap_type, pgoff_t offset)
1136 {
1137         struct swap_info_struct *sis;
1138
1139         if (swap_type >= nr_swapfiles)
1140                 return 0;
1141
1142         sis = swap_info + swap_type;
1143         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1144 }
1145 #endif /* CONFIG_HIBERNATION */
1146
1147 /*
1148  * Free all of a swapdev's extent information
1149  */
1150 static void destroy_swap_extents(struct swap_info_struct *sis)
1151 {
1152         while (!list_empty(&sis->extent_list)) {
1153                 struct swap_extent *se;
1154
1155                 se = list_entry(sis->extent_list.next,
1156                                 struct swap_extent, list);
1157                 list_del(&se->list);
1158                 kfree(se);
1159         }
1160 }
1161
1162 /*
1163  * Add a block range (and the corresponding page range) into this swapdev's
1164  * extent list.  The extent list is kept sorted in page order.
1165  *
1166  * This function rather assumes that it is called in ascending page order.
1167  */
1168 static int
1169 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1170                 unsigned long nr_pages, sector_t start_block)
1171 {
1172         struct swap_extent *se;
1173         struct swap_extent *new_se;
1174         struct list_head *lh;
1175
1176         lh = sis->extent_list.prev;     /* The highest page extent */
1177         if (lh != &sis->extent_list) {
1178                 se = list_entry(lh, struct swap_extent, list);
1179                 BUG_ON(se->start_page + se->nr_pages != start_page);
1180                 if (se->start_block + se->nr_pages == start_block) {
1181                         /* Merge it */
1182                         se->nr_pages += nr_pages;
1183                         return 0;
1184                 }
1185         }
1186
1187         /*
1188          * No merge.  Insert a new extent, preserving ordering.
1189          */
1190         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1191         if (new_se == NULL)
1192                 return -ENOMEM;
1193         new_se->start_page = start_page;
1194         new_se->nr_pages = nr_pages;
1195         new_se->start_block = start_block;
1196
1197         list_add_tail(&new_se->list, &sis->extent_list);
1198         return 1;
1199 }
1200
1201 /*
1202  * A `swap extent' is a simple thing which maps a contiguous range of pages
1203  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1204  * is built at swapon time and is then used at swap_writepage/swap_readpage
1205  * time for locating where on disk a page belongs.
1206  *
1207  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1208  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1209  * swap files identically.
1210  *
1211  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1212  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1213  * swapfiles are handled *identically* after swapon time.
1214  *
1215  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1216  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1217  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1218  * requirements, they are simply tossed out - we will never use those blocks
1219  * for swapping.
1220  *
1221  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1222  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1223  * which will scribble on the fs.
1224  *
1225  * The amount of disk space which a single swap extent represents varies.
1226  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1227  * extents in the list.  To avoid much list walking, we cache the previous
1228  * search location in `curr_swap_extent', and start new searches from there.
1229  * This is extremely effective.  The average number of iterations in
1230  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1231  */
1232 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1233 {
1234         struct inode *inode;
1235         unsigned blocks_per_page;
1236         unsigned long page_no;
1237         unsigned blkbits;
1238         sector_t probe_block;
1239         sector_t last_block;
1240         sector_t lowest_block = -1;
1241         sector_t highest_block = 0;
1242         int nr_extents = 0;
1243         int ret;
1244
1245         inode = sis->swap_file->f_mapping->host;
1246         if (S_ISBLK(inode->i_mode)) {
1247                 ret = add_swap_extent(sis, 0, sis->max, 0);
1248                 *span = sis->pages;
1249                 goto done;
1250         }
1251
1252         blkbits = inode->i_blkbits;
1253         blocks_per_page = PAGE_SIZE >> blkbits;
1254
1255         /*
1256          * Map all the blocks into the extent list.  This code doesn't try
1257          * to be very smart.
1258          */
1259         probe_block = 0;
1260         page_no = 0;
1261         last_block = i_size_read(inode) >> blkbits;
1262         while ((probe_block + blocks_per_page) <= last_block &&
1263                         page_no < sis->max) {
1264                 unsigned block_in_page;
1265                 sector_t first_block;
1266
1267                 first_block = bmap(inode, probe_block);
1268                 if (first_block == 0)
1269                         goto bad_bmap;
1270
1271                 /*
1272                  * It must be PAGE_SIZE aligned on-disk
1273                  */
1274                 if (first_block & (blocks_per_page - 1)) {
1275                         probe_block++;
1276                         goto reprobe;
1277                 }
1278
1279                 for (block_in_page = 1; block_in_page < blocks_per_page;
1280                                         block_in_page++) {
1281                         sector_t block;
1282
1283                         block = bmap(inode, probe_block + block_in_page);
1284                         if (block == 0)
1285                                 goto bad_bmap;
1286                         if (block != first_block + block_in_page) {
1287                                 /* Discontiguity */
1288                                 probe_block++;
1289                                 goto reprobe;
1290                         }
1291                 }
1292
1293                 first_block >>= (PAGE_SHIFT - blkbits);
1294                 if (page_no) {  /* exclude the header page */
1295                         if (first_block < lowest_block)
1296                                 lowest_block = first_block;
1297                         if (first_block > highest_block)
1298                                 highest_block = first_block;
1299                 }
1300
1301                 /*
1302                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1303                  */
1304                 ret = add_swap_extent(sis, page_no, 1, first_block);
1305                 if (ret < 0)
1306                         goto out;
1307                 nr_extents += ret;
1308                 page_no++;
1309                 probe_block += blocks_per_page;
1310 reprobe:
1311                 continue;
1312         }
1313         ret = nr_extents;
1314         *span = 1 + highest_block - lowest_block;
1315         if (page_no == 0)
1316                 page_no = 1;    /* force Empty message */
1317         sis->max = page_no;
1318         sis->pages = page_no - 1;
1319         sis->highest_bit = page_no - 1;
1320 done:
1321         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1322                                         struct swap_extent, list);
1323         goto out;
1324 bad_bmap:
1325         printk(KERN_ERR "swapon: swapfile has holes\n");
1326         ret = -EINVAL;
1327 out:
1328         return ret;
1329 }
1330
1331 #if 0   /* We don't need this yet */
1332 #include <linux/backing-dev.h>
1333 int page_queue_congested(struct page *page)
1334 {
1335         struct backing_dev_info *bdi;
1336
1337         VM_BUG_ON(!PageLocked(page));   /* It pins the swap_info_struct */
1338
1339         if (PageSwapCache(page)) {
1340                 swp_entry_t entry = { .val = page_private(page) };
1341                 struct swap_info_struct *sis;
1342
1343                 sis = get_swap_info_struct(swp_type(entry));
1344                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1345         } else
1346                 bdi = page->mapping->backing_dev_info;
1347         return bdi_write_congested(bdi);
1348 }
1349 #endif
1350
1351 asmlinkage long sys_swapoff(const char __user * specialfile)
1352 {
1353         struct swap_info_struct * p = NULL;
1354         unsigned short *swap_map;
1355         struct file *swap_file, *victim;
1356         struct address_space *mapping;
1357         struct inode *inode;
1358         char * pathname;
1359         int i, type, prev;
1360         int err;
1361
1362         if (!capable(CAP_SYS_ADMIN))
1363                 return -EPERM;
1364
1365         pathname = getname(specialfile);
1366         err = PTR_ERR(pathname);
1367         if (IS_ERR(pathname))
1368                 goto out;
1369
1370         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1371         putname(pathname);
1372         err = PTR_ERR(victim);
1373         if (IS_ERR(victim))
1374                 goto out;
1375
1376         mapping = victim->f_mapping;
1377         prev = -1;
1378         spin_lock(&swap_lock);
1379         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1380                 p = swap_info + type;
1381                 if (p->flags & SWP_WRITEOK) {
1382                         if (p->swap_file->f_mapping == mapping)
1383                                 break;
1384                 }
1385                 prev = type;
1386         }
1387         if (type < 0) {
1388                 err = -EINVAL;
1389                 spin_unlock(&swap_lock);
1390                 goto out_dput;
1391         }
1392         if (!security_vm_enough_memory(p->pages))
1393                 vm_unacct_memory(p->pages);
1394         else {
1395                 err = -ENOMEM;
1396                 spin_unlock(&swap_lock);
1397                 goto out_dput;
1398         }
1399         if (prev < 0) {
1400                 swap_list.head = p->next;
1401         } else {
1402                 swap_info[prev].next = p->next;
1403         }
1404         if (type == swap_list.next) {
1405                 /* just pick something that's safe... */
1406                 swap_list.next = swap_list.head;
1407         }
1408         if (p->prio < 0) {
1409                 for (i = p->next; i >= 0; i = swap_info[i].next)
1410                         swap_info[i].prio = p->prio--;
1411                 least_priority++;
1412         }
1413         nr_swap_pages -= p->pages;
1414         total_swap_pages -= p->pages;
1415         p->flags &= ~SWP_WRITEOK;
1416         spin_unlock(&swap_lock);
1417
1418         current->flags |= PF_SWAPOFF;
1419         err = try_to_unuse(type);
1420         current->flags &= ~PF_SWAPOFF;
1421
1422         if (err) {
1423                 /* re-insert swap space back into swap_list */
1424                 spin_lock(&swap_lock);
1425                 if (p->prio < 0)
1426                         p->prio = --least_priority;
1427                 prev = -1;
1428                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1429                         if (p->prio >= swap_info[i].prio)
1430                                 break;
1431                         prev = i;
1432                 }
1433                 p->next = i;
1434                 if (prev < 0)
1435                         swap_list.head = swap_list.next = p - swap_info;
1436                 else
1437                         swap_info[prev].next = p - swap_info;
1438                 nr_swap_pages += p->pages;
1439                 total_swap_pages += p->pages;
1440                 p->flags |= SWP_WRITEOK;
1441                 spin_unlock(&swap_lock);
1442                 goto out_dput;
1443         }
1444
1445         /* wait for any unplug function to finish */
1446         down_write(&swap_unplug_sem);
1447         up_write(&swap_unplug_sem);
1448
1449         destroy_swap_extents(p);
1450         mutex_lock(&swapon_mutex);
1451         spin_lock(&swap_lock);
1452         drain_mmlist();
1453
1454         /* wait for anyone still in scan_swap_map */
1455         p->highest_bit = 0;             /* cuts scans short */
1456         while (p->flags >= SWP_SCANNING) {
1457                 spin_unlock(&swap_lock);
1458                 schedule_timeout_uninterruptible(1);
1459                 spin_lock(&swap_lock);
1460         }
1461
1462         swap_file = p->swap_file;
1463         p->swap_file = NULL;
1464         p->max = 0;
1465         swap_map = p->swap_map;
1466         p->swap_map = NULL;
1467         p->flags = 0;
1468         spin_unlock(&swap_lock);
1469         mutex_unlock(&swapon_mutex);
1470         vfree(swap_map);
1471         inode = mapping->host;
1472         if (S_ISBLK(inode->i_mode)) {
1473                 struct block_device *bdev = I_BDEV(inode);
1474                 set_blocksize(bdev, p->old_block_size);
1475                 bd_release(bdev);
1476         } else {
1477                 mutex_lock(&inode->i_mutex);
1478                 inode->i_flags &= ~S_SWAPFILE;
1479                 mutex_unlock(&inode->i_mutex);
1480         }
1481         filp_close(swap_file, NULL);
1482         err = 0;
1483
1484 out_dput:
1485         filp_close(victim, NULL);
1486 out:
1487         return err;
1488 }
1489
1490 #ifdef CONFIG_PROC_FS
1491 /* iterator */
1492 static void *swap_start(struct seq_file *swap, loff_t *pos)
1493 {
1494         struct swap_info_struct *ptr = swap_info;
1495         int i;
1496         loff_t l = *pos;
1497
1498         mutex_lock(&swapon_mutex);
1499
1500         if (!l)
1501                 return SEQ_START_TOKEN;
1502
1503         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1504                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1505                         continue;
1506                 if (!--l)
1507                         return ptr;
1508         }
1509
1510         return NULL;
1511 }
1512
1513 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1514 {
1515         struct swap_info_struct *ptr;
1516         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1517
1518         if (v == SEQ_START_TOKEN)
1519                 ptr = swap_info;
1520         else {
1521                 ptr = v;
1522                 ptr++;
1523         }
1524
1525         for (; ptr < endptr; ptr++) {
1526                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1527                         continue;
1528                 ++*pos;
1529                 return ptr;
1530         }
1531
1532         return NULL;
1533 }
1534
1535 static void swap_stop(struct seq_file *swap, void *v)
1536 {
1537         mutex_unlock(&swapon_mutex);
1538 }
1539
1540 static int swap_show(struct seq_file *swap, void *v)
1541 {
1542         struct swap_info_struct *ptr = v;
1543         struct file *file;
1544         int len;
1545
1546         if (ptr == SEQ_START_TOKEN) {
1547                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1548                 return 0;
1549         }
1550
1551         file = ptr->swap_file;
1552         len = seq_path(swap, &file->f_path, " \t\n\\");
1553         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1554                         len < 40 ? 40 - len : 1, " ",
1555                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1556                                 "partition" : "file\t",
1557                         ptr->pages << (PAGE_SHIFT - 10),
1558                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1559                         ptr->prio);
1560         return 0;
1561 }
1562
1563 static const struct seq_operations swaps_op = {
1564         .start =        swap_start,
1565         .next =         swap_next,
1566         .stop =         swap_stop,
1567         .show =         swap_show
1568 };
1569
1570 static int swaps_open(struct inode *inode, struct file *file)
1571 {
1572         return seq_open(file, &swaps_op);
1573 }
1574
1575 static const struct file_operations proc_swaps_operations = {
1576         .open           = swaps_open,
1577         .read           = seq_read,
1578         .llseek         = seq_lseek,
1579         .release        = seq_release,
1580 };
1581
1582 static int __init procswaps_init(void)
1583 {
1584         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1585         return 0;
1586 }
1587 __initcall(procswaps_init);
1588 #endif /* CONFIG_PROC_FS */
1589
1590 #ifdef MAX_SWAPFILES_CHECK
1591 static int __init max_swapfiles_check(void)
1592 {
1593         MAX_SWAPFILES_CHECK();
1594         return 0;
1595 }
1596 late_initcall(max_swapfiles_check);
1597 #endif
1598
1599 /*
1600  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1601  *
1602  * The swapon system call
1603  */
1604 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1605 {
1606         struct swap_info_struct * p;
1607         char *name = NULL;
1608         struct block_device *bdev = NULL;
1609         struct file *swap_file = NULL;
1610         struct address_space *mapping;
1611         unsigned int type;
1612         int i, prev;
1613         int error;
1614         union swap_header *swap_header = NULL;
1615         unsigned int nr_good_pages = 0;
1616         int nr_extents = 0;
1617         sector_t span;
1618         unsigned long maxpages = 1;
1619         unsigned long swapfilepages;
1620         unsigned short *swap_map = NULL;
1621         struct page *page = NULL;
1622         struct inode *inode = NULL;
1623         int did_down = 0;
1624
1625         if (!capable(CAP_SYS_ADMIN))
1626                 return -EPERM;
1627         spin_lock(&swap_lock);
1628         p = swap_info;
1629         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1630                 if (!(p->flags & SWP_USED))
1631                         break;
1632         error = -EPERM;
1633         if (type >= MAX_SWAPFILES) {
1634                 spin_unlock(&swap_lock);
1635                 goto out;
1636         }
1637         if (type >= nr_swapfiles)
1638                 nr_swapfiles = type+1;
1639         memset(p, 0, sizeof(*p));
1640         INIT_LIST_HEAD(&p->extent_list);
1641         p->flags = SWP_USED;
1642         p->next = -1;
1643         spin_unlock(&swap_lock);
1644         name = getname(specialfile);
1645         error = PTR_ERR(name);
1646         if (IS_ERR(name)) {
1647                 name = NULL;
1648                 goto bad_swap_2;
1649         }
1650         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1651         error = PTR_ERR(swap_file);
1652         if (IS_ERR(swap_file)) {
1653                 swap_file = NULL;
1654                 goto bad_swap_2;
1655         }
1656
1657         p->swap_file = swap_file;
1658         mapping = swap_file->f_mapping;
1659         inode = mapping->host;
1660
1661         error = -EBUSY;
1662         for (i = 0; i < nr_swapfiles; i++) {
1663                 struct swap_info_struct *q = &swap_info[i];
1664
1665                 if (i == type || !q->swap_file)
1666                         continue;
1667                 if (mapping == q->swap_file->f_mapping)
1668                         goto bad_swap;
1669         }
1670
1671         error = -EINVAL;
1672         if (S_ISBLK(inode->i_mode)) {
1673                 bdev = I_BDEV(inode);
1674                 error = bd_claim(bdev, sys_swapon);
1675                 if (error < 0) {
1676                         bdev = NULL;
1677                         error = -EINVAL;
1678                         goto bad_swap;
1679                 }
1680                 p->old_block_size = block_size(bdev);
1681                 error = set_blocksize(bdev, PAGE_SIZE);
1682                 if (error < 0)
1683                         goto bad_swap;
1684                 p->bdev = bdev;
1685         } else if (S_ISREG(inode->i_mode)) {
1686                 p->bdev = inode->i_sb->s_bdev;
1687                 mutex_lock(&inode->i_mutex);
1688                 did_down = 1;
1689                 if (IS_SWAPFILE(inode)) {
1690                         error = -EBUSY;
1691                         goto bad_swap;
1692                 }
1693         } else {
1694                 goto bad_swap;
1695         }
1696
1697         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1698
1699         /*
1700          * Read the swap header.
1701          */
1702         if (!mapping->a_ops->readpage) {
1703                 error = -EINVAL;
1704                 goto bad_swap;
1705         }
1706         page = read_mapping_page(mapping, 0, swap_file);
1707         if (IS_ERR(page)) {
1708                 error = PTR_ERR(page);
1709                 goto bad_swap;
1710         }
1711         swap_header = kmap(page);
1712
1713         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1714                 printk(KERN_ERR "Unable to find swap-space signature\n");
1715                 error = -EINVAL;
1716                 goto bad_swap;
1717         }
1718
1719         /* swap partition endianess hack... */
1720         if (swab32(swap_header->info.version) == 1) {
1721                 swab32s(&swap_header->info.version);
1722                 swab32s(&swap_header->info.last_page);
1723                 swab32s(&swap_header->info.nr_badpages);
1724                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1725                         swab32s(&swap_header->info.badpages[i]);
1726         }
1727         /* Check the swap header's sub-version */
1728         if (swap_header->info.version != 1) {
1729                 printk(KERN_WARNING
1730                        "Unable to handle swap header version %d\n",
1731                        swap_header->info.version);
1732                 error = -EINVAL;
1733                 goto bad_swap;
1734         }
1735
1736         p->lowest_bit  = 1;
1737         p->cluster_next = 1;
1738
1739         /*
1740          * Find out how many pages are allowed for a single swap
1741          * device. There are two limiting factors: 1) the number of
1742          * bits for the swap offset in the swp_entry_t type and
1743          * 2) the number of bits in the a swap pte as defined by
1744          * the different architectures. In order to find the
1745          * largest possible bit mask a swap entry with swap type 0
1746          * and swap offset ~0UL is created, encoded to a swap pte,
1747          * decoded to a swp_entry_t again and finally the swap
1748          * offset is extracted. This will mask all the bits from
1749          * the initial ~0UL mask that can't be encoded in either
1750          * the swp_entry_t or the architecture definition of a
1751          * swap pte.
1752          */
1753         maxpages = swp_offset(pte_to_swp_entry(
1754                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1755         if (maxpages > swap_header->info.last_page)
1756                 maxpages = swap_header->info.last_page;
1757         p->highest_bit = maxpages - 1;
1758
1759         error = -EINVAL;
1760         if (!maxpages)
1761                 goto bad_swap;
1762         if (swapfilepages && maxpages > swapfilepages) {
1763                 printk(KERN_WARNING
1764                        "Swap area shorter than signature indicates\n");
1765                 goto bad_swap;
1766         }
1767         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1768                 goto bad_swap;
1769         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1770                 goto bad_swap;
1771
1772         /* OK, set up the swap map and apply the bad block list */
1773         swap_map = vmalloc(maxpages * sizeof(short));
1774         if (!swap_map) {
1775                 error = -ENOMEM;
1776                 goto bad_swap;
1777         }
1778
1779         memset(swap_map, 0, maxpages * sizeof(short));
1780         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1781                 int page_nr = swap_header->info.badpages[i];
1782                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1783                         error = -EINVAL;
1784                         goto bad_swap;
1785                 }
1786                 swap_map[page_nr] = SWAP_MAP_BAD;
1787         }
1788         nr_good_pages = swap_header->info.last_page -
1789                         swap_header->info.nr_badpages -
1790                         1 /* header page */;
1791
1792         if (nr_good_pages) {
1793                 swap_map[0] = SWAP_MAP_BAD;
1794                 p->max = maxpages;
1795                 p->pages = nr_good_pages;
1796                 nr_extents = setup_swap_extents(p, &span);
1797                 if (nr_extents < 0) {
1798                         error = nr_extents;
1799                         goto bad_swap;
1800                 }
1801                 nr_good_pages = p->pages;
1802         }
1803         if (!nr_good_pages) {
1804                 printk(KERN_WARNING "Empty swap-file\n");
1805                 error = -EINVAL;
1806                 goto bad_swap;
1807         }
1808
1809         if (discard_swap(p) == 0)
1810                 p->flags |= SWP_DISCARDABLE;
1811
1812         mutex_lock(&swapon_mutex);
1813         spin_lock(&swap_lock);
1814         if (swap_flags & SWAP_FLAG_PREFER)
1815                 p->prio =
1816                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1817         else
1818                 p->prio = --least_priority;
1819         p->swap_map = swap_map;
1820         p->flags |= SWP_WRITEOK;
1821         nr_swap_pages += nr_good_pages;
1822         total_swap_pages += nr_good_pages;
1823
1824         printk(KERN_INFO "Adding %uk swap on %s.  "
1825                         "Priority:%d extents:%d across:%lluk%s\n",
1826                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1827                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1828                 (p->flags & SWP_DISCARDABLE) ? " D" : "");
1829
1830         /* insert swap space into swap_list: */
1831         prev = -1;
1832         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1833                 if (p->prio >= swap_info[i].prio) {
1834                         break;
1835                 }
1836                 prev = i;
1837         }
1838         p->next = i;
1839         if (prev < 0) {
1840                 swap_list.head = swap_list.next = p - swap_info;
1841         } else {
1842                 swap_info[prev].next = p - swap_info;
1843         }
1844         spin_unlock(&swap_lock);
1845         mutex_unlock(&swapon_mutex);
1846         error = 0;
1847         goto out;
1848 bad_swap:
1849         if (bdev) {
1850                 set_blocksize(bdev, p->old_block_size);
1851                 bd_release(bdev);
1852         }
1853         destroy_swap_extents(p);
1854 bad_swap_2:
1855         spin_lock(&swap_lock);
1856         p->swap_file = NULL;
1857         p->flags = 0;
1858         spin_unlock(&swap_lock);
1859         vfree(swap_map);
1860         if (swap_file)
1861                 filp_close(swap_file, NULL);
1862 out:
1863         if (page && !IS_ERR(page)) {
1864                 kunmap(page);
1865                 page_cache_release(page);
1866         }
1867         if (name)
1868                 putname(name);
1869         if (did_down) {
1870                 if (!error)
1871                         inode->i_flags |= S_SWAPFILE;
1872                 mutex_unlock(&inode->i_mutex);
1873         }
1874         return error;
1875 }
1876
1877 void si_swapinfo(struct sysinfo *val)
1878 {
1879         unsigned int i;
1880         unsigned long nr_to_be_unused = 0;
1881
1882         spin_lock(&swap_lock);
1883         for (i = 0; i < nr_swapfiles; i++) {
1884                 if (!(swap_info[i].flags & SWP_USED) ||
1885                      (swap_info[i].flags & SWP_WRITEOK))
1886                         continue;
1887                 nr_to_be_unused += swap_info[i].inuse_pages;
1888         }
1889         val->freeswap = nr_swap_pages + nr_to_be_unused;
1890         val->totalswap = total_swap_pages + nr_to_be_unused;
1891         spin_unlock(&swap_lock);
1892 }
1893
1894 /*
1895  * Verify that a swap entry is valid and increment its swap map count.
1896  *
1897  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1898  * "permanent", but will be reclaimed by the next swapoff.
1899  */
1900 int swap_duplicate(swp_entry_t entry)
1901 {
1902         struct swap_info_struct * p;
1903         unsigned long offset, type;
1904         int result = 0;
1905
1906         if (is_migration_entry(entry))
1907                 return 1;
1908
1909         type = swp_type(entry);
1910         if (type >= nr_swapfiles)
1911                 goto bad_file;
1912         p = type + swap_info;
1913         offset = swp_offset(entry);
1914
1915         spin_lock(&swap_lock);
1916         if (offset < p->max && p->swap_map[offset]) {
1917                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1918                         p->swap_map[offset]++;
1919                         result = 1;
1920                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1921                         if (swap_overflow++ < 5)
1922                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1923                         p->swap_map[offset] = SWAP_MAP_MAX;
1924                         result = 1;
1925                 }
1926         }
1927         spin_unlock(&swap_lock);
1928 out:
1929         return result;
1930
1931 bad_file:
1932         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1933         goto out;
1934 }
1935
1936 struct swap_info_struct *
1937 get_swap_info_struct(unsigned type)
1938 {
1939         return &swap_info[type];
1940 }
1941
1942 /*
1943  * swap_lock prevents swap_map being freed. Don't grab an extra
1944  * reference on the swaphandle, it doesn't matter if it becomes unused.
1945  */
1946 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1947 {
1948         struct swap_info_struct *si;
1949         int our_page_cluster = page_cluster;
1950         pgoff_t target, toff;
1951         pgoff_t base, end;
1952         int nr_pages = 0;
1953
1954         if (!our_page_cluster)  /* no readahead */
1955                 return 0;
1956
1957         si = &swap_info[swp_type(entry)];
1958         target = swp_offset(entry);
1959         base = (target >> our_page_cluster) << our_page_cluster;
1960         end = base + (1 << our_page_cluster);
1961         if (!base)              /* first page is swap header */
1962                 base++;
1963
1964         spin_lock(&swap_lock);
1965         if (end > si->max)      /* don't go beyond end of map */
1966                 end = si->max;
1967
1968         /* Count contiguous allocated slots above our target */
1969         for (toff = target; ++toff < end; nr_pages++) {
1970                 /* Don't read in free or bad pages */
1971                 if (!si->swap_map[toff])
1972                         break;
1973                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1974                         break;
1975         }
1976         /* Count contiguous allocated slots below our target */
1977         for (toff = target; --toff >= base; nr_pages++) {
1978                 /* Don't read in free or bad pages */
1979                 if (!si->swap_map[toff])
1980                         break;
1981                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1982                         break;
1983         }
1984         spin_unlock(&swap_lock);
1985
1986         /*
1987          * Indicate starting offset, and return number of pages to get:
1988          * if only 1, say 0, since there's then no readahead to be done.
1989          */
1990         *offset = ++toff;
1991         return nr_pages? ++nr_pages: 0;
1992 }