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