swapfile: swapon randomize if nonrot
[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
37 static DEFINE_SPINLOCK(swap_lock);
38 static unsigned int nr_swapfiles;
39 long nr_swap_pages;
40 long total_swap_pages;
41 static int swap_overflow;
42 static int least_priority;
43
44 static const char Bad_file[] = "Bad swap file entry ";
45 static const char Unused_file[] = "Unused swap file entry ";
46 static const char Bad_offset[] = "Bad swap offset entry ";
47 static const char Unused_offset[] = "Unused swap offset entry ";
48
49 static struct swap_list_t swap_list = {-1, -1};
50
51 static struct swap_info_struct swap_info[MAX_SWAPFILES];
52
53 static DEFINE_MUTEX(swapon_mutex);
54
55 /*
56  * We need this because the bdev->unplug_fn can sleep and we cannot
57  * hold swap_lock while calling the unplug_fn. And swap_lock
58  * cannot be turned into a mutex.
59  */
60 static DECLARE_RWSEM(swap_unplug_sem);
61
62 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
63 {
64         swp_entry_t entry;
65
66         down_read(&swap_unplug_sem);
67         entry.val = page_private(page);
68         if (PageSwapCache(page)) {
69                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
70                 struct backing_dev_info *bdi;
71
72                 /*
73                  * If the page is removed from swapcache from under us (with a
74                  * racy try_to_unuse/swapoff) we need an additional reference
75                  * count to avoid reading garbage from page_private(page) above.
76                  * If the WARN_ON triggers during a swapoff it maybe the race
77                  * condition and it's harmless. However if it triggers without
78                  * swapoff it signals a problem.
79                  */
80                 WARN_ON(page_count(page) <= 1);
81
82                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
83                 blk_run_backing_dev(bdi, page);
84         }
85         up_read(&swap_unplug_sem);
86 }
87
88 /*
89  * swapon tell device that all the old swap contents can be discarded,
90  * to allow the swap device to optimize its wear-levelling.
91  */
92 static int discard_swap(struct swap_info_struct *si)
93 {
94         struct swap_extent *se;
95         int err = 0;
96
97         list_for_each_entry(se, &si->extent_list, list) {
98                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
99                 pgoff_t nr_blocks = se->nr_pages << (PAGE_SHIFT - 9);
100
101                 if (se->start_page == 0) {
102                         /* Do not discard the swap header page! */
103                         start_block += 1 << (PAGE_SHIFT - 9);
104                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
105                         if (!nr_blocks)
106                                 continue;
107                 }
108
109                 err = blkdev_issue_discard(si->bdev, start_block,
110                                                 nr_blocks, GFP_KERNEL);
111                 if (err)
112                         break;
113
114                 cond_resched();
115         }
116         return err;             /* That will often be -EOPNOTSUPP */
117 }
118
119 /*
120  * swap allocation tell device that a cluster of swap can now be discarded,
121  * to allow the swap device to optimize its wear-levelling.
122  */
123 static void discard_swap_cluster(struct swap_info_struct *si,
124                                  pgoff_t start_page, pgoff_t nr_pages)
125 {
126         struct swap_extent *se = si->curr_swap_extent;
127         int found_extent = 0;
128
129         while (nr_pages) {
130                 struct list_head *lh;
131
132                 if (se->start_page <= start_page &&
133                     start_page < se->start_page + se->nr_pages) {
134                         pgoff_t offset = start_page - se->start_page;
135                         sector_t start_block = se->start_block + offset;
136                         pgoff_t nr_blocks = se->nr_pages - offset;
137
138                         if (nr_blocks > nr_pages)
139                                 nr_blocks = nr_pages;
140                         start_page += nr_blocks;
141                         nr_pages -= nr_blocks;
142
143                         if (!found_extent++)
144                                 si->curr_swap_extent = se;
145
146                         start_block <<= PAGE_SHIFT - 9;
147                         nr_blocks <<= PAGE_SHIFT - 9;
148                         if (blkdev_issue_discard(si->bdev, start_block,
149                                                         nr_blocks, GFP_NOIO))
150                                 break;
151                 }
152
153                 lh = se->list.next;
154                 if (lh == &si->extent_list)
155                         lh = lh->next;
156                 se = list_entry(lh, struct swap_extent, list);
157         }
158 }
159
160 static int wait_for_discard(void *word)
161 {
162         schedule();
163         return 0;
164 }
165
166 #define SWAPFILE_CLUSTER        256
167 #define LATENCY_LIMIT           256
168
169 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
170 {
171         unsigned long offset;
172         unsigned long last_in_cluster = 0;
173         int latency_ration = LATENCY_LIMIT;
174         int found_free_cluster = 0;
175
176         /*
177          * We try to cluster swap pages by allocating them sequentially
178          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
179          * way, however, we resort to first-free allocation, starting
180          * a new cluster.  This prevents us from scattering swap pages
181          * all over the entire swap partition, so that we reduce
182          * overall disk seek times between swap pages.  -- sct
183          * But we do now try to find an empty cluster.  -Andrea
184          */
185
186         si->flags += SWP_SCANNING;
187         offset = si->cluster_next;
188
189         if (unlikely(!si->cluster_nr--)) {
190                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
191                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
192                         goto checks;
193                 }
194                 if (si->flags & SWP_DISCARDABLE) {
195                         /*
196                          * Start range check on racing allocations, in case
197                          * they overlap the cluster we eventually decide on
198                          * (we scan without swap_lock to allow preemption).
199                          * It's hardly conceivable that cluster_nr could be
200                          * wrapped during our scan, but don't depend on it.
201                          */
202                         if (si->lowest_alloc)
203                                 goto checks;
204                         si->lowest_alloc = si->max;
205                         si->highest_alloc = 0;
206                 }
207                 spin_unlock(&swap_lock);
208
209                 offset = si->lowest_bit;
210                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
211
212                 /* Locate the first empty (unaligned) cluster */
213                 for (; last_in_cluster <= si->highest_bit; offset++) {
214                         if (si->swap_map[offset])
215                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
216                         else if (offset == last_in_cluster) {
217                                 spin_lock(&swap_lock);
218                                 offset -= SWAPFILE_CLUSTER - 1;
219                                 si->cluster_next = offset;
220                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
221                                 found_free_cluster = 1;
222                                 goto checks;
223                         }
224                         if (unlikely(--latency_ration < 0)) {
225                                 cond_resched();
226                                 latency_ration = LATENCY_LIMIT;
227                         }
228                 }
229
230                 offset = si->lowest_bit;
231                 spin_lock(&swap_lock);
232                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
233                 si->lowest_alloc = 0;
234         }
235
236 checks:
237         if (!(si->flags & SWP_WRITEOK))
238                 goto no_page;
239         if (!si->highest_bit)
240                 goto no_page;
241         if (offset > si->highest_bit)
242                 offset = si->lowest_bit;
243         if (si->swap_map[offset])
244                 goto scan;
245
246         if (offset == si->lowest_bit)
247                 si->lowest_bit++;
248         if (offset == si->highest_bit)
249                 si->highest_bit--;
250         si->inuse_pages++;
251         if (si->inuse_pages == si->pages) {
252                 si->lowest_bit = si->max;
253                 si->highest_bit = 0;
254         }
255         si->swap_map[offset] = 1;
256         si->cluster_next = offset + 1;
257         si->flags -= SWP_SCANNING;
258
259         if (si->lowest_alloc) {
260                 /*
261                  * Only set when SWP_DISCARDABLE, and there's a scan
262                  * for a free cluster in progress or just completed.
263                  */
264                 if (found_free_cluster) {
265                         /*
266                          * To optimize wear-levelling, discard the
267                          * old data of the cluster, taking care not to
268                          * discard any of its pages that have already
269                          * been allocated by racing tasks (offset has
270                          * already stepped over any at the beginning).
271                          */
272                         if (offset < si->highest_alloc &&
273                             si->lowest_alloc <= last_in_cluster)
274                                 last_in_cluster = si->lowest_alloc - 1;
275                         si->flags |= SWP_DISCARDING;
276                         spin_unlock(&swap_lock);
277
278                         if (offset < last_in_cluster)
279                                 discard_swap_cluster(si, offset,
280                                         last_in_cluster - offset + 1);
281
282                         spin_lock(&swap_lock);
283                         si->lowest_alloc = 0;
284                         si->flags &= ~SWP_DISCARDING;
285
286                         smp_mb();       /* wake_up_bit advises this */
287                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
288
289                 } else if (si->flags & SWP_DISCARDING) {
290                         /*
291                          * Delay using pages allocated by racing tasks
292                          * until the whole discard has been issued. We
293                          * could defer that delay until swap_writepage,
294                          * but it's easier to keep this self-contained.
295                          */
296                         spin_unlock(&swap_lock);
297                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
298                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
299                         spin_lock(&swap_lock);
300                 } else {
301                         /*
302                          * Note pages allocated by racing tasks while
303                          * scan for a free cluster is in progress, so
304                          * that its final discard can exclude them.
305                          */
306                         if (offset < si->lowest_alloc)
307                                 si->lowest_alloc = offset;
308                         if (offset > si->highest_alloc)
309                                 si->highest_alloc = offset;
310                 }
311         }
312         return offset;
313
314 scan:
315         spin_unlock(&swap_lock);
316         while (++offset <= si->highest_bit) {
317                 if (!si->swap_map[offset]) {
318                         spin_lock(&swap_lock);
319                         goto checks;
320                 }
321                 if (unlikely(--latency_ration < 0)) {
322                         cond_resched();
323                         latency_ration = LATENCY_LIMIT;
324                 }
325         }
326         spin_lock(&swap_lock);
327         goto checks;
328
329 no_page:
330         si->flags -= SWP_SCANNING;
331         return 0;
332 }
333
334 swp_entry_t get_swap_page(void)
335 {
336         struct swap_info_struct *si;
337         pgoff_t offset;
338         int type, next;
339         int wrapped = 0;
340
341         spin_lock(&swap_lock);
342         if (nr_swap_pages <= 0)
343                 goto noswap;
344         nr_swap_pages--;
345
346         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
347                 si = swap_info + type;
348                 next = si->next;
349                 if (next < 0 ||
350                     (!wrapped && si->prio != swap_info[next].prio)) {
351                         next = swap_list.head;
352                         wrapped++;
353                 }
354
355                 if (!si->highest_bit)
356                         continue;
357                 if (!(si->flags & SWP_WRITEOK))
358                         continue;
359
360                 swap_list.next = next;
361                 offset = scan_swap_map(si);
362                 if (offset) {
363                         spin_unlock(&swap_lock);
364                         return swp_entry(type, offset);
365                 }
366                 next = swap_list.next;
367         }
368
369         nr_swap_pages++;
370 noswap:
371         spin_unlock(&swap_lock);
372         return (swp_entry_t) {0};
373 }
374
375 swp_entry_t get_swap_page_of_type(int type)
376 {
377         struct swap_info_struct *si;
378         pgoff_t offset;
379
380         spin_lock(&swap_lock);
381         si = swap_info + type;
382         if (si->flags & SWP_WRITEOK) {
383                 nr_swap_pages--;
384                 offset = scan_swap_map(si);
385                 if (offset) {
386                         spin_unlock(&swap_lock);
387                         return swp_entry(type, offset);
388                 }
389                 nr_swap_pages++;
390         }
391         spin_unlock(&swap_lock);
392         return (swp_entry_t) {0};
393 }
394
395 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
396 {
397         struct swap_info_struct * p;
398         unsigned long offset, type;
399
400         if (!entry.val)
401                 goto out;
402         type = swp_type(entry);
403         if (type >= nr_swapfiles)
404                 goto bad_nofile;
405         p = & swap_info[type];
406         if (!(p->flags & SWP_USED))
407                 goto bad_device;
408         offset = swp_offset(entry);
409         if (offset >= p->max)
410                 goto bad_offset;
411         if (!p->swap_map[offset])
412                 goto bad_free;
413         spin_lock(&swap_lock);
414         return p;
415
416 bad_free:
417         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
418         goto out;
419 bad_offset:
420         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
421         goto out;
422 bad_device:
423         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
424         goto out;
425 bad_nofile:
426         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
427 out:
428         return NULL;
429 }
430
431 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
432 {
433         int count = p->swap_map[offset];
434
435         if (count < SWAP_MAP_MAX) {
436                 count--;
437                 p->swap_map[offset] = count;
438                 if (!count) {
439                         if (offset < p->lowest_bit)
440                                 p->lowest_bit = offset;
441                         if (offset > p->highest_bit)
442                                 p->highest_bit = offset;
443                         if (p->prio > swap_info[swap_list.next].prio)
444                                 swap_list.next = p - swap_info;
445                         nr_swap_pages++;
446                         p->inuse_pages--;
447                 }
448         }
449         return count;
450 }
451
452 /*
453  * Caller has made sure that the swapdevice corresponding to entry
454  * is still around or has not been recycled.
455  */
456 void swap_free(swp_entry_t entry)
457 {
458         struct swap_info_struct * p;
459
460         p = swap_info_get(entry);
461         if (p) {
462                 swap_entry_free(p, swp_offset(entry));
463                 spin_unlock(&swap_lock);
464         }
465 }
466
467 /*
468  * How many references to page are currently swapped out?
469  */
470 static inline int page_swapcount(struct page *page)
471 {
472         int count = 0;
473         struct swap_info_struct *p;
474         swp_entry_t entry;
475
476         entry.val = page_private(page);
477         p = swap_info_get(entry);
478         if (p) {
479                 /* Subtract the 1 for the swap cache itself */
480                 count = p->swap_map[swp_offset(entry)] - 1;
481                 spin_unlock(&swap_lock);
482         }
483         return count;
484 }
485
486 /*
487  * We can write to an anon page without COW if there are no other references
488  * to it.  And as a side-effect, free up its swap: because the old content
489  * on disk will never be read, and seeking back there to write new content
490  * later would only waste time away from clustering.
491  */
492 int reuse_swap_page(struct page *page)
493 {
494         int count;
495
496         VM_BUG_ON(!PageLocked(page));
497         count = page_mapcount(page);
498         if (count <= 1 && PageSwapCache(page)) {
499                 count += page_swapcount(page);
500                 if (count == 1 && !PageWriteback(page)) {
501                         delete_from_swap_cache(page);
502                         SetPageDirty(page);
503                 }
504         }
505         return count == 1;
506 }
507
508 /*
509  * If swap is getting full, or if there are no more mappings of this page,
510  * then try_to_free_swap is called to free its swap space.
511  */
512 int try_to_free_swap(struct page *page)
513 {
514         VM_BUG_ON(!PageLocked(page));
515
516         if (!PageSwapCache(page))
517                 return 0;
518         if (PageWriteback(page))
519                 return 0;
520         if (page_swapcount(page))
521                 return 0;
522
523         delete_from_swap_cache(page);
524         SetPageDirty(page);
525         return 1;
526 }
527
528 /*
529  * Free the swap entry like above, but also try to
530  * free the page cache entry if it is the last user.
531  */
532 void free_swap_and_cache(swp_entry_t entry)
533 {
534         struct swap_info_struct * p;
535         struct page *page = NULL;
536
537         if (is_migration_entry(entry))
538                 return;
539
540         p = swap_info_get(entry);
541         if (p) {
542                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
543                         page = find_get_page(&swapper_space, entry.val);
544                         if (page && !trylock_page(page)) {
545                                 page_cache_release(page);
546                                 page = NULL;
547                         }
548                 }
549                 spin_unlock(&swap_lock);
550         }
551         if (page) {
552                 /*
553                  * Not mapped elsewhere, or swap space full? Free it!
554                  * Also recheck PageSwapCache now page is locked (above).
555                  */
556                 if (PageSwapCache(page) && !PageWriteback(page) &&
557                                 (!page_mapped(page) || vm_swap_full())) {
558                         delete_from_swap_cache(page);
559                         SetPageDirty(page);
560                 }
561                 unlock_page(page);
562                 page_cache_release(page);
563         }
564 }
565
566 #ifdef CONFIG_HIBERNATION
567 /*
568  * Find the swap type that corresponds to given device (if any).
569  *
570  * @offset - number of the PAGE_SIZE-sized block of the device, starting
571  * from 0, in which the swap header is expected to be located.
572  *
573  * This is needed for the suspend to disk (aka swsusp).
574  */
575 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
576 {
577         struct block_device *bdev = NULL;
578         int i;
579
580         if (device)
581                 bdev = bdget(device);
582
583         spin_lock(&swap_lock);
584         for (i = 0; i < nr_swapfiles; i++) {
585                 struct swap_info_struct *sis = swap_info + i;
586
587                 if (!(sis->flags & SWP_WRITEOK))
588                         continue;
589
590                 if (!bdev) {
591                         if (bdev_p)
592                                 *bdev_p = sis->bdev;
593
594                         spin_unlock(&swap_lock);
595                         return i;
596                 }
597                 if (bdev == sis->bdev) {
598                         struct swap_extent *se;
599
600                         se = list_entry(sis->extent_list.next,
601                                         struct swap_extent, list);
602                         if (se->start_block == offset) {
603                                 if (bdev_p)
604                                         *bdev_p = sis->bdev;
605
606                                 spin_unlock(&swap_lock);
607                                 bdput(bdev);
608                                 return i;
609                         }
610                 }
611         }
612         spin_unlock(&swap_lock);
613         if (bdev)
614                 bdput(bdev);
615
616         return -ENODEV;
617 }
618
619 /*
620  * Return either the total number of swap pages of given type, or the number
621  * of free pages of that type (depending on @free)
622  *
623  * This is needed for software suspend
624  */
625 unsigned int count_swap_pages(int type, int free)
626 {
627         unsigned int n = 0;
628
629         if (type < nr_swapfiles) {
630                 spin_lock(&swap_lock);
631                 if (swap_info[type].flags & SWP_WRITEOK) {
632                         n = swap_info[type].pages;
633                         if (free)
634                                 n -= swap_info[type].inuse_pages;
635                 }
636                 spin_unlock(&swap_lock);
637         }
638         return n;
639 }
640 #endif
641
642 /*
643  * No need to decide whether this PTE shares the swap entry with others,
644  * just let do_wp_page work it out if a write is requested later - to
645  * force COW, vm_page_prot omits write permission from any private vma.
646  */
647 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
648                 unsigned long addr, swp_entry_t entry, struct page *page)
649 {
650         spinlock_t *ptl;
651         pte_t *pte;
652         int ret = 1;
653
654         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
655                 ret = -ENOMEM;
656
657         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
658         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
659                 if (ret > 0)
660                         mem_cgroup_uncharge_page(page);
661                 ret = 0;
662                 goto out;
663         }
664
665         inc_mm_counter(vma->vm_mm, anon_rss);
666         get_page(page);
667         set_pte_at(vma->vm_mm, addr, pte,
668                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
669         page_add_anon_rmap(page, vma, addr);
670         swap_free(entry);
671         /*
672          * Move the page to the active list so it is not
673          * immediately swapped out again after swapon.
674          */
675         activate_page(page);
676 out:
677         pte_unmap_unlock(pte, ptl);
678         return ret;
679 }
680
681 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
682                                 unsigned long addr, unsigned long end,
683                                 swp_entry_t entry, struct page *page)
684 {
685         pte_t swp_pte = swp_entry_to_pte(entry);
686         pte_t *pte;
687         int ret = 0;
688
689         /*
690          * We don't actually need pte lock while scanning for swp_pte: since
691          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
692          * page table while we're scanning; though it could get zapped, and on
693          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
694          * of unmatched parts which look like swp_pte, so unuse_pte must
695          * recheck under pte lock.  Scanning without pte lock lets it be
696          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
697          */
698         pte = pte_offset_map(pmd, addr);
699         do {
700                 /*
701                  * swapoff spends a _lot_ of time in this loop!
702                  * Test inline before going to call unuse_pte.
703                  */
704                 if (unlikely(pte_same(*pte, swp_pte))) {
705                         pte_unmap(pte);
706                         ret = unuse_pte(vma, pmd, addr, entry, page);
707                         if (ret)
708                                 goto out;
709                         pte = pte_offset_map(pmd, addr);
710                 }
711         } while (pte++, addr += PAGE_SIZE, addr != end);
712         pte_unmap(pte - 1);
713 out:
714         return ret;
715 }
716
717 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
718                                 unsigned long addr, unsigned long end,
719                                 swp_entry_t entry, struct page *page)
720 {
721         pmd_t *pmd;
722         unsigned long next;
723         int ret;
724
725         pmd = pmd_offset(pud, addr);
726         do {
727                 next = pmd_addr_end(addr, end);
728                 if (pmd_none_or_clear_bad(pmd))
729                         continue;
730                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
731                 if (ret)
732                         return ret;
733         } while (pmd++, addr = next, addr != end);
734         return 0;
735 }
736
737 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
738                                 unsigned long addr, unsigned long end,
739                                 swp_entry_t entry, struct page *page)
740 {
741         pud_t *pud;
742         unsigned long next;
743         int ret;
744
745         pud = pud_offset(pgd, addr);
746         do {
747                 next = pud_addr_end(addr, end);
748                 if (pud_none_or_clear_bad(pud))
749                         continue;
750                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
751                 if (ret)
752                         return ret;
753         } while (pud++, addr = next, addr != end);
754         return 0;
755 }
756
757 static int unuse_vma(struct vm_area_struct *vma,
758                                 swp_entry_t entry, struct page *page)
759 {
760         pgd_t *pgd;
761         unsigned long addr, end, next;
762         int ret;
763
764         if (page->mapping) {
765                 addr = page_address_in_vma(page, vma);
766                 if (addr == -EFAULT)
767                         return 0;
768                 else
769                         end = addr + PAGE_SIZE;
770         } else {
771                 addr = vma->vm_start;
772                 end = vma->vm_end;
773         }
774
775         pgd = pgd_offset(vma->vm_mm, addr);
776         do {
777                 next = pgd_addr_end(addr, end);
778                 if (pgd_none_or_clear_bad(pgd))
779                         continue;
780                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
781                 if (ret)
782                         return ret;
783         } while (pgd++, addr = next, addr != end);
784         return 0;
785 }
786
787 static int unuse_mm(struct mm_struct *mm,
788                                 swp_entry_t entry, struct page *page)
789 {
790         struct vm_area_struct *vma;
791         int ret = 0;
792
793         if (!down_read_trylock(&mm->mmap_sem)) {
794                 /*
795                  * Activate page so shrink_inactive_list is unlikely to unmap
796                  * its ptes while lock is dropped, so swapoff can make progress.
797                  */
798                 activate_page(page);
799                 unlock_page(page);
800                 down_read(&mm->mmap_sem);
801                 lock_page(page);
802         }
803         for (vma = mm->mmap; vma; vma = vma->vm_next) {
804                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
805                         break;
806         }
807         up_read(&mm->mmap_sem);
808         return (ret < 0)? ret: 0;
809 }
810
811 /*
812  * Scan swap_map from current position to next entry still in use.
813  * Recycle to start on reaching the end, returning 0 when empty.
814  */
815 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
816                                         unsigned int prev)
817 {
818         unsigned int max = si->max;
819         unsigned int i = prev;
820         int count;
821
822         /*
823          * No need for swap_lock here: we're just looking
824          * for whether an entry is in use, not modifying it; false
825          * hits are okay, and sys_swapoff() has already prevented new
826          * allocations from this area (while holding swap_lock).
827          */
828         for (;;) {
829                 if (++i >= max) {
830                         if (!prev) {
831                                 i = 0;
832                                 break;
833                         }
834                         /*
835                          * No entries in use at top of swap_map,
836                          * loop back to start and recheck there.
837                          */
838                         max = prev + 1;
839                         prev = 0;
840                         i = 1;
841                 }
842                 count = si->swap_map[i];
843                 if (count && count != SWAP_MAP_BAD)
844                         break;
845         }
846         return i;
847 }
848
849 /*
850  * We completely avoid races by reading each swap page in advance,
851  * and then search for the process using it.  All the necessary
852  * page table adjustments can then be made atomically.
853  */
854 static int try_to_unuse(unsigned int type)
855 {
856         struct swap_info_struct * si = &swap_info[type];
857         struct mm_struct *start_mm;
858         unsigned short *swap_map;
859         unsigned short swcount;
860         struct page *page;
861         swp_entry_t entry;
862         unsigned int i = 0;
863         int retval = 0;
864         int reset_overflow = 0;
865         int shmem;
866
867         /*
868          * When searching mms for an entry, a good strategy is to
869          * start at the first mm we freed the previous entry from
870          * (though actually we don't notice whether we or coincidence
871          * freed the entry).  Initialize this start_mm with a hold.
872          *
873          * A simpler strategy would be to start at the last mm we
874          * freed the previous entry from; but that would take less
875          * advantage of mmlist ordering, which clusters forked mms
876          * together, child after parent.  If we race with dup_mmap(), we
877          * prefer to resolve parent before child, lest we miss entries
878          * duplicated after we scanned child: using last mm would invert
879          * that.  Though it's only a serious concern when an overflowed
880          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
881          */
882         start_mm = &init_mm;
883         atomic_inc(&init_mm.mm_users);
884
885         /*
886          * Keep on scanning until all entries have gone.  Usually,
887          * one pass through swap_map is enough, but not necessarily:
888          * there are races when an instance of an entry might be missed.
889          */
890         while ((i = find_next_to_unuse(si, i)) != 0) {
891                 if (signal_pending(current)) {
892                         retval = -EINTR;
893                         break;
894                 }
895
896                 /*
897                  * Get a page for the entry, using the existing swap
898                  * cache page if there is one.  Otherwise, get a clean
899                  * page and read the swap into it.
900                  */
901                 swap_map = &si->swap_map[i];
902                 entry = swp_entry(type, i);
903                 page = read_swap_cache_async(entry,
904                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
905                 if (!page) {
906                         /*
907                          * Either swap_duplicate() failed because entry
908                          * has been freed independently, and will not be
909                          * reused since sys_swapoff() already disabled
910                          * allocation from here, or alloc_page() failed.
911                          */
912                         if (!*swap_map)
913                                 continue;
914                         retval = -ENOMEM;
915                         break;
916                 }
917
918                 /*
919                  * Don't hold on to start_mm if it looks like exiting.
920                  */
921                 if (atomic_read(&start_mm->mm_users) == 1) {
922                         mmput(start_mm);
923                         start_mm = &init_mm;
924                         atomic_inc(&init_mm.mm_users);
925                 }
926
927                 /*
928                  * Wait for and lock page.  When do_swap_page races with
929                  * try_to_unuse, do_swap_page can handle the fault much
930                  * faster than try_to_unuse can locate the entry.  This
931                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
932                  * defer to do_swap_page in such a case - in some tests,
933                  * do_swap_page and try_to_unuse repeatedly compete.
934                  */
935                 wait_on_page_locked(page);
936                 wait_on_page_writeback(page);
937                 lock_page(page);
938                 wait_on_page_writeback(page);
939
940                 /*
941                  * Remove all references to entry.
942                  * Whenever we reach init_mm, there's no address space
943                  * to search, but use it as a reminder to search shmem.
944                  */
945                 shmem = 0;
946                 swcount = *swap_map;
947                 if (swcount > 1) {
948                         if (start_mm == &init_mm)
949                                 shmem = shmem_unuse(entry, page);
950                         else
951                                 retval = unuse_mm(start_mm, entry, page);
952                 }
953                 if (*swap_map > 1) {
954                         int set_start_mm = (*swap_map >= swcount);
955                         struct list_head *p = &start_mm->mmlist;
956                         struct mm_struct *new_start_mm = start_mm;
957                         struct mm_struct *prev_mm = start_mm;
958                         struct mm_struct *mm;
959
960                         atomic_inc(&new_start_mm->mm_users);
961                         atomic_inc(&prev_mm->mm_users);
962                         spin_lock(&mmlist_lock);
963                         while (*swap_map > 1 && !retval && !shmem &&
964                                         (p = p->next) != &start_mm->mmlist) {
965                                 mm = list_entry(p, struct mm_struct, mmlist);
966                                 if (!atomic_inc_not_zero(&mm->mm_users))
967                                         continue;
968                                 spin_unlock(&mmlist_lock);
969                                 mmput(prev_mm);
970                                 prev_mm = mm;
971
972                                 cond_resched();
973
974                                 swcount = *swap_map;
975                                 if (swcount <= 1)
976                                         ;
977                                 else if (mm == &init_mm) {
978                                         set_start_mm = 1;
979                                         shmem = shmem_unuse(entry, page);
980                                 } else
981                                         retval = unuse_mm(mm, entry, page);
982                                 if (set_start_mm && *swap_map < swcount) {
983                                         mmput(new_start_mm);
984                                         atomic_inc(&mm->mm_users);
985                                         new_start_mm = mm;
986                                         set_start_mm = 0;
987                                 }
988                                 spin_lock(&mmlist_lock);
989                         }
990                         spin_unlock(&mmlist_lock);
991                         mmput(prev_mm);
992                         mmput(start_mm);
993                         start_mm = new_start_mm;
994                 }
995                 if (shmem) {
996                         /* page has already been unlocked and released */
997                         if (shmem > 0)
998                                 continue;
999                         retval = shmem;
1000                         break;
1001                 }
1002                 if (retval) {
1003                         unlock_page(page);
1004                         page_cache_release(page);
1005                         break;
1006                 }
1007
1008                 /*
1009                  * How could swap count reach 0x7fff when the maximum
1010                  * pid is 0x7fff, and there's no way to repeat a swap
1011                  * page within an mm (except in shmem, where it's the
1012                  * shared object which takes the reference count)?
1013                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
1014                  *
1015                  * If that's wrong, then we should worry more about
1016                  * exit_mmap() and do_munmap() cases described above:
1017                  * we might be resetting SWAP_MAP_MAX too early here.
1018                  * We know "Undead"s can happen, they're okay, so don't
1019                  * report them; but do report if we reset SWAP_MAP_MAX.
1020                  */
1021                 if (*swap_map == SWAP_MAP_MAX) {
1022                         spin_lock(&swap_lock);
1023                         *swap_map = 1;
1024                         spin_unlock(&swap_lock);
1025                         reset_overflow = 1;
1026                 }
1027
1028                 /*
1029                  * If a reference remains (rare), we would like to leave
1030                  * the page in the swap cache; but try_to_unmap could
1031                  * then re-duplicate the entry once we drop page lock,
1032                  * so we might loop indefinitely; also, that page could
1033                  * not be swapped out to other storage meanwhile.  So:
1034                  * delete from cache even if there's another reference,
1035                  * after ensuring that the data has been saved to disk -
1036                  * since if the reference remains (rarer), it will be
1037                  * read from disk into another page.  Splitting into two
1038                  * pages would be incorrect if swap supported "shared
1039                  * private" pages, but they are handled by tmpfs files.
1040                  */
1041                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
1042                         struct writeback_control wbc = {
1043                                 .sync_mode = WB_SYNC_NONE,
1044                         };
1045
1046                         swap_writepage(page, &wbc);
1047                         lock_page(page);
1048                         wait_on_page_writeback(page);
1049                 }
1050
1051                 /*
1052                  * It is conceivable that a racing task removed this page from
1053                  * swap cache just before we acquired the page lock at the top,
1054                  * or while we dropped it in unuse_mm().  The page might even
1055                  * be back in swap cache on another swap area: that we must not
1056                  * delete, since it may not have been written out to swap yet.
1057                  */
1058                 if (PageSwapCache(page) &&
1059                     likely(page_private(page) == entry.val))
1060                         delete_from_swap_cache(page);
1061
1062                 /*
1063                  * So we could skip searching mms once swap count went
1064                  * to 1, we did not mark any present ptes as dirty: must
1065                  * mark page dirty so shrink_page_list will preserve it.
1066                  */
1067                 SetPageDirty(page);
1068                 unlock_page(page);
1069                 page_cache_release(page);
1070
1071                 /*
1072                  * Make sure that we aren't completely killing
1073                  * interactive performance.
1074                  */
1075                 cond_resched();
1076         }
1077
1078         mmput(start_mm);
1079         if (reset_overflow) {
1080                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1081                 swap_overflow = 0;
1082         }
1083         return retval;
1084 }
1085
1086 /*
1087  * After a successful try_to_unuse, if no swap is now in use, we know
1088  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1089  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1090  * added to the mmlist just after page_duplicate - before would be racy.
1091  */
1092 static void drain_mmlist(void)
1093 {
1094         struct list_head *p, *next;
1095         unsigned int i;
1096
1097         for (i = 0; i < nr_swapfiles; i++)
1098                 if (swap_info[i].inuse_pages)
1099                         return;
1100         spin_lock(&mmlist_lock);
1101         list_for_each_safe(p, next, &init_mm.mmlist)
1102                 list_del_init(p);
1103         spin_unlock(&mmlist_lock);
1104 }
1105
1106 /*
1107  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1108  * corresponds to page offset `offset'.
1109  */
1110 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1111 {
1112         struct swap_extent *se = sis->curr_swap_extent;
1113         struct swap_extent *start_se = se;
1114
1115         for ( ; ; ) {
1116                 struct list_head *lh;
1117
1118                 if (se->start_page <= offset &&
1119                                 offset < (se->start_page + se->nr_pages)) {
1120                         return se->start_block + (offset - se->start_page);
1121                 }
1122                 lh = se->list.next;
1123                 if (lh == &sis->extent_list)
1124                         lh = lh->next;
1125                 se = list_entry(lh, struct swap_extent, list);
1126                 sis->curr_swap_extent = se;
1127                 BUG_ON(se == start_se);         /* It *must* be present */
1128         }
1129 }
1130
1131 #ifdef CONFIG_HIBERNATION
1132 /*
1133  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1134  * corresponding to given index in swap_info (swap type).
1135  */
1136 sector_t swapdev_block(int swap_type, pgoff_t offset)
1137 {
1138         struct swap_info_struct *sis;
1139
1140         if (swap_type >= nr_swapfiles)
1141                 return 0;
1142
1143         sis = swap_info + swap_type;
1144         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1145 }
1146 #endif /* CONFIG_HIBERNATION */
1147
1148 /*
1149  * Free all of a swapdev's extent information
1150  */
1151 static void destroy_swap_extents(struct swap_info_struct *sis)
1152 {
1153         while (!list_empty(&sis->extent_list)) {
1154                 struct swap_extent *se;
1155
1156                 se = list_entry(sis->extent_list.next,
1157                                 struct swap_extent, list);
1158                 list_del(&se->list);
1159                 kfree(se);
1160         }
1161 }
1162
1163 /*
1164  * Add a block range (and the corresponding page range) into this swapdev's
1165  * extent list.  The extent list is kept sorted in page order.
1166  *
1167  * This function rather assumes that it is called in ascending page order.
1168  */
1169 static int
1170 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1171                 unsigned long nr_pages, sector_t start_block)
1172 {
1173         struct swap_extent *se;
1174         struct swap_extent *new_se;
1175         struct list_head *lh;
1176
1177         lh = sis->extent_list.prev;     /* The highest page extent */
1178         if (lh != &sis->extent_list) {
1179                 se = list_entry(lh, struct swap_extent, list);
1180                 BUG_ON(se->start_page + se->nr_pages != start_page);
1181                 if (se->start_block + se->nr_pages == start_block) {
1182                         /* Merge it */
1183                         se->nr_pages += nr_pages;
1184                         return 0;
1185                 }
1186         }
1187
1188         /*
1189          * No merge.  Insert a new extent, preserving ordering.
1190          */
1191         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1192         if (new_se == NULL)
1193                 return -ENOMEM;
1194         new_se->start_page = start_page;
1195         new_se->nr_pages = nr_pages;
1196         new_se->start_block = start_block;
1197
1198         list_add_tail(&new_se->list, &sis->extent_list);
1199         return 1;
1200 }
1201
1202 /*
1203  * A `swap extent' is a simple thing which maps a contiguous range of pages
1204  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1205  * is built at swapon time and is then used at swap_writepage/swap_readpage
1206  * time for locating where on disk a page belongs.
1207  *
1208  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1209  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1210  * swap files identically.
1211  *
1212  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1213  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1214  * swapfiles are handled *identically* after swapon time.
1215  *
1216  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1217  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1218  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1219  * requirements, they are simply tossed out - we will never use those blocks
1220  * for swapping.
1221  *
1222  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1223  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1224  * which will scribble on the fs.
1225  *
1226  * The amount of disk space which a single swap extent represents varies.
1227  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1228  * extents in the list.  To avoid much list walking, we cache the previous
1229  * search location in `curr_swap_extent', and start new searches from there.
1230  * This is extremely effective.  The average number of iterations in
1231  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1232  */
1233 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1234 {
1235         struct inode *inode;
1236         unsigned blocks_per_page;
1237         unsigned long page_no;
1238         unsigned blkbits;
1239         sector_t probe_block;
1240         sector_t last_block;
1241         sector_t lowest_block = -1;
1242         sector_t highest_block = 0;
1243         int nr_extents = 0;
1244         int ret;
1245
1246         inode = sis->swap_file->f_mapping->host;
1247         if (S_ISBLK(inode->i_mode)) {
1248                 ret = add_swap_extent(sis, 0, sis->max, 0);
1249                 *span = sis->pages;
1250                 goto done;
1251         }
1252
1253         blkbits = inode->i_blkbits;
1254         blocks_per_page = PAGE_SIZE >> blkbits;
1255
1256         /*
1257          * Map all the blocks into the extent list.  This code doesn't try
1258          * to be very smart.
1259          */
1260         probe_block = 0;
1261         page_no = 0;
1262         last_block = i_size_read(inode) >> blkbits;
1263         while ((probe_block + blocks_per_page) <= last_block &&
1264                         page_no < sis->max) {
1265                 unsigned block_in_page;
1266                 sector_t first_block;
1267
1268                 first_block = bmap(inode, probe_block);
1269                 if (first_block == 0)
1270                         goto bad_bmap;
1271
1272                 /*
1273                  * It must be PAGE_SIZE aligned on-disk
1274                  */
1275                 if (first_block & (blocks_per_page - 1)) {
1276                         probe_block++;
1277                         goto reprobe;
1278                 }
1279
1280                 for (block_in_page = 1; block_in_page < blocks_per_page;
1281                                         block_in_page++) {
1282                         sector_t block;
1283
1284                         block = bmap(inode, probe_block + block_in_page);
1285                         if (block == 0)
1286                                 goto bad_bmap;
1287                         if (block != first_block + block_in_page) {
1288                                 /* Discontiguity */
1289                                 probe_block++;
1290                                 goto reprobe;
1291                         }
1292                 }
1293
1294                 first_block >>= (PAGE_SHIFT - blkbits);
1295                 if (page_no) {  /* exclude the header page */
1296                         if (first_block < lowest_block)
1297                                 lowest_block = first_block;
1298                         if (first_block > highest_block)
1299                                 highest_block = first_block;
1300                 }
1301
1302                 /*
1303                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1304                  */
1305                 ret = add_swap_extent(sis, page_no, 1, first_block);
1306                 if (ret < 0)
1307                         goto out;
1308                 nr_extents += ret;
1309                 page_no++;
1310                 probe_block += blocks_per_page;
1311 reprobe:
1312                 continue;
1313         }
1314         ret = nr_extents;
1315         *span = 1 + highest_block - lowest_block;
1316         if (page_no == 0)
1317                 page_no = 1;    /* force Empty message */
1318         sis->max = page_no;
1319         sis->pages = page_no - 1;
1320         sis->highest_bit = page_no - 1;
1321 done:
1322         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1323                                         struct swap_extent, list);
1324         goto out;
1325 bad_bmap:
1326         printk(KERN_ERR "swapon: swapfile has holes\n");
1327         ret = -EINVAL;
1328 out:
1329         return ret;
1330 }
1331
1332 #if 0   /* We don't need this yet */
1333 #include <linux/backing-dev.h>
1334 int page_queue_congested(struct page *page)
1335 {
1336         struct backing_dev_info *bdi;
1337
1338         VM_BUG_ON(!PageLocked(page));   /* It pins the swap_info_struct */
1339
1340         if (PageSwapCache(page)) {
1341                 swp_entry_t entry = { .val = page_private(page) };
1342                 struct swap_info_struct *sis;
1343
1344                 sis = get_swap_info_struct(swp_type(entry));
1345                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1346         } else
1347                 bdi = page->mapping->backing_dev_info;
1348         return bdi_write_congested(bdi);
1349 }
1350 #endif
1351
1352 asmlinkage long sys_swapoff(const char __user * specialfile)
1353 {
1354         struct swap_info_struct * p = NULL;
1355         unsigned short *swap_map;
1356         struct file *swap_file, *victim;
1357         struct address_space *mapping;
1358         struct inode *inode;
1359         char * pathname;
1360         int i, type, prev;
1361         int err;
1362
1363         if (!capable(CAP_SYS_ADMIN))
1364                 return -EPERM;
1365
1366         pathname = getname(specialfile);
1367         err = PTR_ERR(pathname);
1368         if (IS_ERR(pathname))
1369                 goto out;
1370
1371         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1372         putname(pathname);
1373         err = PTR_ERR(victim);
1374         if (IS_ERR(victim))
1375                 goto out;
1376
1377         mapping = victim->f_mapping;
1378         prev = -1;
1379         spin_lock(&swap_lock);
1380         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1381                 p = swap_info + type;
1382                 if (p->flags & SWP_WRITEOK) {
1383                         if (p->swap_file->f_mapping == mapping)
1384                                 break;
1385                 }
1386                 prev = type;
1387         }
1388         if (type < 0) {
1389                 err = -EINVAL;
1390                 spin_unlock(&swap_lock);
1391                 goto out_dput;
1392         }
1393         if (!security_vm_enough_memory(p->pages))
1394                 vm_unacct_memory(p->pages);
1395         else {
1396                 err = -ENOMEM;
1397                 spin_unlock(&swap_lock);
1398                 goto out_dput;
1399         }
1400         if (prev < 0) {
1401                 swap_list.head = p->next;
1402         } else {
1403                 swap_info[prev].next = p->next;
1404         }
1405         if (type == swap_list.next) {
1406                 /* just pick something that's safe... */
1407                 swap_list.next = swap_list.head;
1408         }
1409         if (p->prio < 0) {
1410                 for (i = p->next; i >= 0; i = swap_info[i].next)
1411                         swap_info[i].prio = p->prio--;
1412                 least_priority++;
1413         }
1414         nr_swap_pages -= p->pages;
1415         total_swap_pages -= p->pages;
1416         p->flags &= ~SWP_WRITEOK;
1417         spin_unlock(&swap_lock);
1418
1419         current->flags |= PF_SWAPOFF;
1420         err = try_to_unuse(type);
1421         current->flags &= ~PF_SWAPOFF;
1422
1423         if (err) {
1424                 /* re-insert swap space back into swap_list */
1425                 spin_lock(&swap_lock);
1426                 if (p->prio < 0)
1427                         p->prio = --least_priority;
1428                 prev = -1;
1429                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1430                         if (p->prio >= swap_info[i].prio)
1431                                 break;
1432                         prev = i;
1433                 }
1434                 p->next = i;
1435                 if (prev < 0)
1436                         swap_list.head = swap_list.next = p - swap_info;
1437                 else
1438                         swap_info[prev].next = p - swap_info;
1439                 nr_swap_pages += p->pages;
1440                 total_swap_pages += p->pages;
1441                 p->flags |= SWP_WRITEOK;
1442                 spin_unlock(&swap_lock);
1443                 goto out_dput;
1444         }
1445
1446         /* wait for any unplug function to finish */
1447         down_write(&swap_unplug_sem);
1448         up_write(&swap_unplug_sem);
1449
1450         destroy_swap_extents(p);
1451         mutex_lock(&swapon_mutex);
1452         spin_lock(&swap_lock);
1453         drain_mmlist();
1454
1455         /* wait for anyone still in scan_swap_map */
1456         p->highest_bit = 0;             /* cuts scans short */
1457         while (p->flags >= SWP_SCANNING) {
1458                 spin_unlock(&swap_lock);
1459                 schedule_timeout_uninterruptible(1);
1460                 spin_lock(&swap_lock);
1461         }
1462
1463         swap_file = p->swap_file;
1464         p->swap_file = NULL;
1465         p->max = 0;
1466         swap_map = p->swap_map;
1467         p->swap_map = NULL;
1468         p->flags = 0;
1469         spin_unlock(&swap_lock);
1470         mutex_unlock(&swapon_mutex);
1471         vfree(swap_map);
1472         inode = mapping->host;
1473         if (S_ISBLK(inode->i_mode)) {
1474                 struct block_device *bdev = I_BDEV(inode);
1475                 set_blocksize(bdev, p->old_block_size);
1476                 bd_release(bdev);
1477         } else {
1478                 mutex_lock(&inode->i_mutex);
1479                 inode->i_flags &= ~S_SWAPFILE;
1480                 mutex_unlock(&inode->i_mutex);
1481         }
1482         filp_close(swap_file, NULL);
1483         err = 0;
1484
1485 out_dput:
1486         filp_close(victim, NULL);
1487 out:
1488         return err;
1489 }
1490
1491 #ifdef CONFIG_PROC_FS
1492 /* iterator */
1493 static void *swap_start(struct seq_file *swap, loff_t *pos)
1494 {
1495         struct swap_info_struct *ptr = swap_info;
1496         int i;
1497         loff_t l = *pos;
1498
1499         mutex_lock(&swapon_mutex);
1500
1501         if (!l)
1502                 return SEQ_START_TOKEN;
1503
1504         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1505                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1506                         continue;
1507                 if (!--l)
1508                         return ptr;
1509         }
1510
1511         return NULL;
1512 }
1513
1514 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1515 {
1516         struct swap_info_struct *ptr;
1517         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1518
1519         if (v == SEQ_START_TOKEN)
1520                 ptr = swap_info;
1521         else {
1522                 ptr = v;
1523                 ptr++;
1524         }
1525
1526         for (; ptr < endptr; ptr++) {
1527                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1528                         continue;
1529                 ++*pos;
1530                 return ptr;
1531         }
1532
1533         return NULL;
1534 }
1535
1536 static void swap_stop(struct seq_file *swap, void *v)
1537 {
1538         mutex_unlock(&swapon_mutex);
1539 }
1540
1541 static int swap_show(struct seq_file *swap, void *v)
1542 {
1543         struct swap_info_struct *ptr = v;
1544         struct file *file;
1545         int len;
1546
1547         if (ptr == SEQ_START_TOKEN) {
1548                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1549                 return 0;
1550         }
1551
1552         file = ptr->swap_file;
1553         len = seq_path(swap, &file->f_path, " \t\n\\");
1554         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1555                         len < 40 ? 40 - len : 1, " ",
1556                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1557                                 "partition" : "file\t",
1558                         ptr->pages << (PAGE_SHIFT - 10),
1559                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1560                         ptr->prio);
1561         return 0;
1562 }
1563
1564 static const struct seq_operations swaps_op = {
1565         .start =        swap_start,
1566         .next =         swap_next,
1567         .stop =         swap_stop,
1568         .show =         swap_show
1569 };
1570
1571 static int swaps_open(struct inode *inode, struct file *file)
1572 {
1573         return seq_open(file, &swaps_op);
1574 }
1575
1576 static const struct file_operations proc_swaps_operations = {
1577         .open           = swaps_open,
1578         .read           = seq_read,
1579         .llseek         = seq_lseek,
1580         .release        = seq_release,
1581 };
1582
1583 static int __init procswaps_init(void)
1584 {
1585         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1586         return 0;
1587 }
1588 __initcall(procswaps_init);
1589 #endif /* CONFIG_PROC_FS */
1590
1591 #ifdef MAX_SWAPFILES_CHECK
1592 static int __init max_swapfiles_check(void)
1593 {
1594         MAX_SWAPFILES_CHECK();
1595         return 0;
1596 }
1597 late_initcall(max_swapfiles_check);
1598 #endif
1599
1600 /*
1601  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1602  *
1603  * The swapon system call
1604  */
1605 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1606 {
1607         struct swap_info_struct * p;
1608         char *name = NULL;
1609         struct block_device *bdev = NULL;
1610         struct file *swap_file = NULL;
1611         struct address_space *mapping;
1612         unsigned int type;
1613         int i, prev;
1614         int error;
1615         union swap_header *swap_header = NULL;
1616         unsigned int nr_good_pages = 0;
1617         int nr_extents = 0;
1618         sector_t span;
1619         unsigned long maxpages = 1;
1620         unsigned long swapfilepages;
1621         unsigned short *swap_map = NULL;
1622         struct page *page = NULL;
1623         struct inode *inode = NULL;
1624         int did_down = 0;
1625
1626         if (!capable(CAP_SYS_ADMIN))
1627                 return -EPERM;
1628         spin_lock(&swap_lock);
1629         p = swap_info;
1630         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1631                 if (!(p->flags & SWP_USED))
1632                         break;
1633         error = -EPERM;
1634         if (type >= MAX_SWAPFILES) {
1635                 spin_unlock(&swap_lock);
1636                 goto out;
1637         }
1638         if (type >= nr_swapfiles)
1639                 nr_swapfiles = type+1;
1640         memset(p, 0, sizeof(*p));
1641         INIT_LIST_HEAD(&p->extent_list);
1642         p->flags = SWP_USED;
1643         p->next = -1;
1644         spin_unlock(&swap_lock);
1645         name = getname(specialfile);
1646         error = PTR_ERR(name);
1647         if (IS_ERR(name)) {
1648                 name = NULL;
1649                 goto bad_swap_2;
1650         }
1651         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1652         error = PTR_ERR(swap_file);
1653         if (IS_ERR(swap_file)) {
1654                 swap_file = NULL;
1655                 goto bad_swap_2;
1656         }
1657
1658         p->swap_file = swap_file;
1659         mapping = swap_file->f_mapping;
1660         inode = mapping->host;
1661
1662         error = -EBUSY;
1663         for (i = 0; i < nr_swapfiles; i++) {
1664                 struct swap_info_struct *q = &swap_info[i];
1665
1666                 if (i == type || !q->swap_file)
1667                         continue;
1668                 if (mapping == q->swap_file->f_mapping)
1669                         goto bad_swap;
1670         }
1671
1672         error = -EINVAL;
1673         if (S_ISBLK(inode->i_mode)) {
1674                 bdev = I_BDEV(inode);
1675                 error = bd_claim(bdev, sys_swapon);
1676                 if (error < 0) {
1677                         bdev = NULL;
1678                         error = -EINVAL;
1679                         goto bad_swap;
1680                 }
1681                 p->old_block_size = block_size(bdev);
1682                 error = set_blocksize(bdev, PAGE_SIZE);
1683                 if (error < 0)
1684                         goto bad_swap;
1685                 p->bdev = bdev;
1686         } else if (S_ISREG(inode->i_mode)) {
1687                 p->bdev = inode->i_sb->s_bdev;
1688                 mutex_lock(&inode->i_mutex);
1689                 did_down = 1;
1690                 if (IS_SWAPFILE(inode)) {
1691                         error = -EBUSY;
1692                         goto bad_swap;
1693                 }
1694         } else {
1695                 goto bad_swap;
1696         }
1697
1698         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1699
1700         /*
1701          * Read the swap header.
1702          */
1703         if (!mapping->a_ops->readpage) {
1704                 error = -EINVAL;
1705                 goto bad_swap;
1706         }
1707         page = read_mapping_page(mapping, 0, swap_file);
1708         if (IS_ERR(page)) {
1709                 error = PTR_ERR(page);
1710                 goto bad_swap;
1711         }
1712         swap_header = kmap(page);
1713
1714         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1715                 printk(KERN_ERR "Unable to find swap-space signature\n");
1716                 error = -EINVAL;
1717                 goto bad_swap;
1718         }
1719
1720         /* swap partition endianess hack... */
1721         if (swab32(swap_header->info.version) == 1) {
1722                 swab32s(&swap_header->info.version);
1723                 swab32s(&swap_header->info.last_page);
1724                 swab32s(&swap_header->info.nr_badpages);
1725                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1726                         swab32s(&swap_header->info.badpages[i]);
1727         }
1728         /* Check the swap header's sub-version */
1729         if (swap_header->info.version != 1) {
1730                 printk(KERN_WARNING
1731                        "Unable to handle swap header version %d\n",
1732                        swap_header->info.version);
1733                 error = -EINVAL;
1734                 goto bad_swap;
1735         }
1736
1737         p->lowest_bit  = 1;
1738         p->cluster_next = 1;
1739
1740         /*
1741          * Find out how many pages are allowed for a single swap
1742          * device. There are two limiting factors: 1) the number of
1743          * bits for the swap offset in the swp_entry_t type and
1744          * 2) the number of bits in the a swap pte as defined by
1745          * the different architectures. In order to find the
1746          * largest possible bit mask a swap entry with swap type 0
1747          * and swap offset ~0UL is created, encoded to a swap pte,
1748          * decoded to a swp_entry_t again and finally the swap
1749          * offset is extracted. This will mask all the bits from
1750          * the initial ~0UL mask that can't be encoded in either
1751          * the swp_entry_t or the architecture definition of a
1752          * swap pte.
1753          */
1754         maxpages = swp_offset(pte_to_swp_entry(
1755                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1756         if (maxpages > swap_header->info.last_page)
1757                 maxpages = swap_header->info.last_page;
1758         p->highest_bit = maxpages - 1;
1759
1760         error = -EINVAL;
1761         if (!maxpages)
1762                 goto bad_swap;
1763         if (swapfilepages && maxpages > swapfilepages) {
1764                 printk(KERN_WARNING
1765                        "Swap area shorter than signature indicates\n");
1766                 goto bad_swap;
1767         }
1768         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1769                 goto bad_swap;
1770         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1771                 goto bad_swap;
1772
1773         /* OK, set up the swap map and apply the bad block list */
1774         swap_map = vmalloc(maxpages * sizeof(short));
1775         if (!swap_map) {
1776                 error = -ENOMEM;
1777                 goto bad_swap;
1778         }
1779
1780         memset(swap_map, 0, maxpages * sizeof(short));
1781         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1782                 int page_nr = swap_header->info.badpages[i];
1783                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1784                         error = -EINVAL;
1785                         goto bad_swap;
1786                 }
1787                 swap_map[page_nr] = SWAP_MAP_BAD;
1788         }
1789         nr_good_pages = swap_header->info.last_page -
1790                         swap_header->info.nr_badpages -
1791                         1 /* header page */;
1792
1793         if (nr_good_pages) {
1794                 swap_map[0] = SWAP_MAP_BAD;
1795                 p->max = maxpages;
1796                 p->pages = nr_good_pages;
1797                 nr_extents = setup_swap_extents(p, &span);
1798                 if (nr_extents < 0) {
1799                         error = nr_extents;
1800                         goto bad_swap;
1801                 }
1802                 nr_good_pages = p->pages;
1803         }
1804         if (!nr_good_pages) {
1805                 printk(KERN_WARNING "Empty swap-file\n");
1806                 error = -EINVAL;
1807                 goto bad_swap;
1808         }
1809
1810         if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1811                 p->flags |= SWP_SOLIDSTATE;
1812                 srandom32((u32)get_seconds());
1813                 p->cluster_next = 1 + (random32() % p->highest_bit);
1814         }
1815         if (discard_swap(p) == 0)
1816                 p->flags |= SWP_DISCARDABLE;
1817
1818         mutex_lock(&swapon_mutex);
1819         spin_lock(&swap_lock);
1820         if (swap_flags & SWAP_FLAG_PREFER)
1821                 p->prio =
1822                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1823         else
1824                 p->prio = --least_priority;
1825         p->swap_map = swap_map;
1826         p->flags |= SWP_WRITEOK;
1827         nr_swap_pages += nr_good_pages;
1828         total_swap_pages += nr_good_pages;
1829
1830         printk(KERN_INFO "Adding %uk swap on %s.  "
1831                         "Priority:%d extents:%d across:%lluk %s%s\n",
1832                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1833                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1834                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1835                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1836
1837         /* insert swap space into swap_list: */
1838         prev = -1;
1839         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1840                 if (p->prio >= swap_info[i].prio) {
1841                         break;
1842                 }
1843                 prev = i;
1844         }
1845         p->next = i;
1846         if (prev < 0) {
1847                 swap_list.head = swap_list.next = p - swap_info;
1848         } else {
1849                 swap_info[prev].next = p - swap_info;
1850         }
1851         spin_unlock(&swap_lock);
1852         mutex_unlock(&swapon_mutex);
1853         error = 0;
1854         goto out;
1855 bad_swap:
1856         if (bdev) {
1857                 set_blocksize(bdev, p->old_block_size);
1858                 bd_release(bdev);
1859         }
1860         destroy_swap_extents(p);
1861 bad_swap_2:
1862         spin_lock(&swap_lock);
1863         p->swap_file = NULL;
1864         p->flags = 0;
1865         spin_unlock(&swap_lock);
1866         vfree(swap_map);
1867         if (swap_file)
1868                 filp_close(swap_file, NULL);
1869 out:
1870         if (page && !IS_ERR(page)) {
1871                 kunmap(page);
1872                 page_cache_release(page);
1873         }
1874         if (name)
1875                 putname(name);
1876         if (did_down) {
1877                 if (!error)
1878                         inode->i_flags |= S_SWAPFILE;
1879                 mutex_unlock(&inode->i_mutex);
1880         }
1881         return error;
1882 }
1883
1884 void si_swapinfo(struct sysinfo *val)
1885 {
1886         unsigned int i;
1887         unsigned long nr_to_be_unused = 0;
1888
1889         spin_lock(&swap_lock);
1890         for (i = 0; i < nr_swapfiles; i++) {
1891                 if (!(swap_info[i].flags & SWP_USED) ||
1892                      (swap_info[i].flags & SWP_WRITEOK))
1893                         continue;
1894                 nr_to_be_unused += swap_info[i].inuse_pages;
1895         }
1896         val->freeswap = nr_swap_pages + nr_to_be_unused;
1897         val->totalswap = total_swap_pages + nr_to_be_unused;
1898         spin_unlock(&swap_lock);
1899 }
1900
1901 /*
1902  * Verify that a swap entry is valid and increment its swap map count.
1903  *
1904  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1905  * "permanent", but will be reclaimed by the next swapoff.
1906  */
1907 int swap_duplicate(swp_entry_t entry)
1908 {
1909         struct swap_info_struct * p;
1910         unsigned long offset, type;
1911         int result = 0;
1912
1913         if (is_migration_entry(entry))
1914                 return 1;
1915
1916         type = swp_type(entry);
1917         if (type >= nr_swapfiles)
1918                 goto bad_file;
1919         p = type + swap_info;
1920         offset = swp_offset(entry);
1921
1922         spin_lock(&swap_lock);
1923         if (offset < p->max && p->swap_map[offset]) {
1924                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1925                         p->swap_map[offset]++;
1926                         result = 1;
1927                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1928                         if (swap_overflow++ < 5)
1929                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1930                         p->swap_map[offset] = SWAP_MAP_MAX;
1931                         result = 1;
1932                 }
1933         }
1934         spin_unlock(&swap_lock);
1935 out:
1936         return result;
1937
1938 bad_file:
1939         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1940         goto out;
1941 }
1942
1943 struct swap_info_struct *
1944 get_swap_info_struct(unsigned type)
1945 {
1946         return &swap_info[type];
1947 }
1948
1949 /*
1950  * swap_lock prevents swap_map being freed. Don't grab an extra
1951  * reference on the swaphandle, it doesn't matter if it becomes unused.
1952  */
1953 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1954 {
1955         struct swap_info_struct *si;
1956         int our_page_cluster = page_cluster;
1957         pgoff_t target, toff;
1958         pgoff_t base, end;
1959         int nr_pages = 0;
1960
1961         if (!our_page_cluster)  /* no readahead */
1962                 return 0;
1963
1964         si = &swap_info[swp_type(entry)];
1965         target = swp_offset(entry);
1966         base = (target >> our_page_cluster) << our_page_cluster;
1967         end = base + (1 << our_page_cluster);
1968         if (!base)              /* first page is swap header */
1969                 base++;
1970
1971         spin_lock(&swap_lock);
1972         if (end > si->max)      /* don't go beyond end of map */
1973                 end = si->max;
1974
1975         /* Count contiguous allocated slots above our target */
1976         for (toff = target; ++toff < end; nr_pages++) {
1977                 /* Don't read in free or bad pages */
1978                 if (!si->swap_map[toff])
1979                         break;
1980                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1981                         break;
1982         }
1983         /* Count contiguous allocated slots below our target */
1984         for (toff = target; --toff >= base; nr_pages++) {
1985                 /* Don't read in free or bad pages */
1986                 if (!si->swap_map[toff])
1987                         break;
1988                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1989                         break;
1990         }
1991         spin_unlock(&swap_lock);
1992
1993         /*
1994          * Indicate starting offset, and return number of pages to get:
1995          * if only 1, say 0, since there's then no readahead to be done.
1996          */
1997         *offset = ++toff;
1998         return nr_pages? ++nr_pages: 0;
1999 }