mm: add swap cache interface for swap reference
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32
33 #include <asm/pgtable.h>
34 #include <asm/tlbflush.h>
35 #include <linux/swapops.h>
36 #include <linux/page_cgroup.h>
37
38 static DEFINE_SPINLOCK(swap_lock);
39 static unsigned int nr_swapfiles;
40 long nr_swap_pages;
41 long total_swap_pages;
42 static int swap_overflow;
43 static int least_priority;
44
45 static const char Bad_file[] = "Bad swap file entry ";
46 static const char Unused_file[] = "Unused swap file entry ";
47 static const char Bad_offset[] = "Bad swap offset entry ";
48 static const char Unused_offset[] = "Unused swap offset entry ";
49
50 static struct swap_list_t swap_list = {-1, -1};
51
52 static struct swap_info_struct swap_info[MAX_SWAPFILES];
53
54 static DEFINE_MUTEX(swapon_mutex);
55
56 /*
57  * We need this because the bdev->unplug_fn can sleep and we cannot
58  * hold swap_lock while calling the unplug_fn. And swap_lock
59  * cannot be turned into a mutex.
60  */
61 static DECLARE_RWSEM(swap_unplug_sem);
62
63 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
64 {
65         swp_entry_t entry;
66
67         down_read(&swap_unplug_sem);
68         entry.val = page_private(page);
69         if (PageSwapCache(page)) {
70                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
71                 struct backing_dev_info *bdi;
72
73                 /*
74                  * If the page is removed from swapcache from under us (with a
75                  * racy try_to_unuse/swapoff) we need an additional reference
76                  * count to avoid reading garbage from page_private(page) above.
77                  * If the WARN_ON triggers during a swapoff it maybe the race
78                  * condition and it's harmless. However if it triggers without
79                  * swapoff it signals a problem.
80                  */
81                 WARN_ON(page_count(page) <= 1);
82
83                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
84                 blk_run_backing_dev(bdi, page);
85         }
86         up_read(&swap_unplug_sem);
87 }
88
89 /*
90  * swapon tell device that all the old swap contents can be discarded,
91  * to allow the swap device to optimize its wear-levelling.
92  */
93 static int discard_swap(struct swap_info_struct *si)
94 {
95         struct swap_extent *se;
96         int err = 0;
97
98         list_for_each_entry(se, &si->extent_list, list) {
99                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
100                 sector_t nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
101
102                 if (se->start_page == 0) {
103                         /* Do not discard the swap header page! */
104                         start_block += 1 << (PAGE_SHIFT - 9);
105                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
106                         if (!nr_blocks)
107                                 continue;
108                 }
109
110                 err = blkdev_issue_discard(si->bdev, start_block,
111                                                 nr_blocks, GFP_KERNEL);
112                 if (err)
113                         break;
114
115                 cond_resched();
116         }
117         return err;             /* That will often be -EOPNOTSUPP */
118 }
119
120 /*
121  * swap allocation tell device that a cluster of swap can now be discarded,
122  * to allow the swap device to optimize its wear-levelling.
123  */
124 static void discard_swap_cluster(struct swap_info_struct *si,
125                                  pgoff_t start_page, pgoff_t nr_pages)
126 {
127         struct swap_extent *se = si->curr_swap_extent;
128         int found_extent = 0;
129
130         while (nr_pages) {
131                 struct list_head *lh;
132
133                 if (se->start_page <= start_page &&
134                     start_page < se->start_page + se->nr_pages) {
135                         pgoff_t offset = start_page - se->start_page;
136                         sector_t start_block = se->start_block + offset;
137                         sector_t nr_blocks = se->nr_pages - offset;
138
139                         if (nr_blocks > nr_pages)
140                                 nr_blocks = nr_pages;
141                         start_page += nr_blocks;
142                         nr_pages -= nr_blocks;
143
144                         if (!found_extent++)
145                                 si->curr_swap_extent = se;
146
147                         start_block <<= PAGE_SHIFT - 9;
148                         nr_blocks <<= PAGE_SHIFT - 9;
149                         if (blkdev_issue_discard(si->bdev, start_block,
150                                                         nr_blocks, GFP_NOIO))
151                                 break;
152                 }
153
154                 lh = se->list.next;
155                 if (lh == &si->extent_list)
156                         lh = lh->next;
157                 se = list_entry(lh, struct swap_extent, list);
158         }
159 }
160
161 static int wait_for_discard(void *word)
162 {
163         schedule();
164         return 0;
165 }
166
167 #define SWAPFILE_CLUSTER        256
168 #define LATENCY_LIMIT           256
169
170 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
171 {
172         unsigned long offset;
173         unsigned long scan_base;
174         unsigned long last_in_cluster = 0;
175         int latency_ration = LATENCY_LIMIT;
176         int found_free_cluster = 0;
177
178         /*
179          * We try to cluster swap pages by allocating them sequentially
180          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
181          * way, however, we resort to first-free allocation, starting
182          * a new cluster.  This prevents us from scattering swap pages
183          * all over the entire swap partition, so that we reduce
184          * overall disk seek times between swap pages.  -- sct
185          * But we do now try to find an empty cluster.  -Andrea
186          * And we let swap pages go all over an SSD partition.  Hugh
187          */
188
189         si->flags += SWP_SCANNING;
190         scan_base = offset = si->cluster_next;
191
192         if (unlikely(!si->cluster_nr--)) {
193                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
194                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
195                         goto checks;
196                 }
197                 if (si->flags & SWP_DISCARDABLE) {
198                         /*
199                          * Start range check on racing allocations, in case
200                          * they overlap the cluster we eventually decide on
201                          * (we scan without swap_lock to allow preemption).
202                          * It's hardly conceivable that cluster_nr could be
203                          * wrapped during our scan, but don't depend on it.
204                          */
205                         if (si->lowest_alloc)
206                                 goto checks;
207                         si->lowest_alloc = si->max;
208                         si->highest_alloc = 0;
209                 }
210                 spin_unlock(&swap_lock);
211
212                 /*
213                  * If seek is expensive, start searching for new cluster from
214                  * start of partition, to minimize the span of allocated swap.
215                  * But if seek is cheap, search from our current position, so
216                  * that swap is allocated from all over the partition: if the
217                  * Flash Translation Layer only remaps within limited zones,
218                  * we don't want to wear out the first zone too quickly.
219                  */
220                 if (!(si->flags & SWP_SOLIDSTATE))
221                         scan_base = offset = si->lowest_bit;
222                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
223
224                 /* Locate the first empty (unaligned) cluster */
225                 for (; last_in_cluster <= si->highest_bit; offset++) {
226                         if (si->swap_map[offset])
227                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
228                         else if (offset == last_in_cluster) {
229                                 spin_lock(&swap_lock);
230                                 offset -= SWAPFILE_CLUSTER - 1;
231                                 si->cluster_next = offset;
232                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
233                                 found_free_cluster = 1;
234                                 goto checks;
235                         }
236                         if (unlikely(--latency_ration < 0)) {
237                                 cond_resched();
238                                 latency_ration = LATENCY_LIMIT;
239                         }
240                 }
241
242                 offset = si->lowest_bit;
243                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
244
245                 /* Locate the first empty (unaligned) cluster */
246                 for (; last_in_cluster < scan_base; offset++) {
247                         if (si->swap_map[offset])
248                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
249                         else if (offset == last_in_cluster) {
250                                 spin_lock(&swap_lock);
251                                 offset -= SWAPFILE_CLUSTER - 1;
252                                 si->cluster_next = offset;
253                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
254                                 found_free_cluster = 1;
255                                 goto checks;
256                         }
257                         if (unlikely(--latency_ration < 0)) {
258                                 cond_resched();
259                                 latency_ration = LATENCY_LIMIT;
260                         }
261                 }
262
263                 offset = scan_base;
264                 spin_lock(&swap_lock);
265                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
266                 si->lowest_alloc = 0;
267         }
268
269 checks:
270         if (!(si->flags & SWP_WRITEOK))
271                 goto no_page;
272         if (!si->highest_bit)
273                 goto no_page;
274         if (offset > si->highest_bit)
275                 scan_base = offset = si->lowest_bit;
276         if (si->swap_map[offset])
277                 goto scan;
278
279         if (offset == si->lowest_bit)
280                 si->lowest_bit++;
281         if (offset == si->highest_bit)
282                 si->highest_bit--;
283         si->inuse_pages++;
284         if (si->inuse_pages == si->pages) {
285                 si->lowest_bit = si->max;
286                 si->highest_bit = 0;
287         }
288         si->swap_map[offset] = 1;
289         si->cluster_next = offset + 1;
290         si->flags -= SWP_SCANNING;
291
292         if (si->lowest_alloc) {
293                 /*
294                  * Only set when SWP_DISCARDABLE, and there's a scan
295                  * for a free cluster in progress or just completed.
296                  */
297                 if (found_free_cluster) {
298                         /*
299                          * To optimize wear-levelling, discard the
300                          * old data of the cluster, taking care not to
301                          * discard any of its pages that have already
302                          * been allocated by racing tasks (offset has
303                          * already stepped over any at the beginning).
304                          */
305                         if (offset < si->highest_alloc &&
306                             si->lowest_alloc <= last_in_cluster)
307                                 last_in_cluster = si->lowest_alloc - 1;
308                         si->flags |= SWP_DISCARDING;
309                         spin_unlock(&swap_lock);
310
311                         if (offset < last_in_cluster)
312                                 discard_swap_cluster(si, offset,
313                                         last_in_cluster - offset + 1);
314
315                         spin_lock(&swap_lock);
316                         si->lowest_alloc = 0;
317                         si->flags &= ~SWP_DISCARDING;
318
319                         smp_mb();       /* wake_up_bit advises this */
320                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
321
322                 } else if (si->flags & SWP_DISCARDING) {
323                         /*
324                          * Delay using pages allocated by racing tasks
325                          * until the whole discard has been issued. We
326                          * could defer that delay until swap_writepage,
327                          * but it's easier to keep this self-contained.
328                          */
329                         spin_unlock(&swap_lock);
330                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
331                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
332                         spin_lock(&swap_lock);
333                 } else {
334                         /*
335                          * Note pages allocated by racing tasks while
336                          * scan for a free cluster is in progress, so
337                          * that its final discard can exclude them.
338                          */
339                         if (offset < si->lowest_alloc)
340                                 si->lowest_alloc = offset;
341                         if (offset > si->highest_alloc)
342                                 si->highest_alloc = offset;
343                 }
344         }
345         return offset;
346
347 scan:
348         spin_unlock(&swap_lock);
349         while (++offset <= si->highest_bit) {
350                 if (!si->swap_map[offset]) {
351                         spin_lock(&swap_lock);
352                         goto checks;
353                 }
354                 if (unlikely(--latency_ration < 0)) {
355                         cond_resched();
356                         latency_ration = LATENCY_LIMIT;
357                 }
358         }
359         offset = si->lowest_bit;
360         while (++offset < scan_base) {
361                 if (!si->swap_map[offset]) {
362                         spin_lock(&swap_lock);
363                         goto checks;
364                 }
365                 if (unlikely(--latency_ration < 0)) {
366                         cond_resched();
367                         latency_ration = LATENCY_LIMIT;
368                 }
369         }
370         spin_lock(&swap_lock);
371
372 no_page:
373         si->flags -= SWP_SCANNING;
374         return 0;
375 }
376
377 swp_entry_t get_swap_page(void)
378 {
379         struct swap_info_struct *si;
380         pgoff_t offset;
381         int type, next;
382         int wrapped = 0;
383
384         spin_lock(&swap_lock);
385         if (nr_swap_pages <= 0)
386                 goto noswap;
387         nr_swap_pages--;
388
389         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
390                 si = swap_info + type;
391                 next = si->next;
392                 if (next < 0 ||
393                     (!wrapped && si->prio != swap_info[next].prio)) {
394                         next = swap_list.head;
395                         wrapped++;
396                 }
397
398                 if (!si->highest_bit)
399                         continue;
400                 if (!(si->flags & SWP_WRITEOK))
401                         continue;
402
403                 swap_list.next = next;
404                 offset = scan_swap_map(si);
405                 if (offset) {
406                         spin_unlock(&swap_lock);
407                         return swp_entry(type, offset);
408                 }
409                 next = swap_list.next;
410         }
411
412         nr_swap_pages++;
413 noswap:
414         spin_unlock(&swap_lock);
415         return (swp_entry_t) {0};
416 }
417
418 swp_entry_t get_swap_page_of_type(int type)
419 {
420         struct swap_info_struct *si;
421         pgoff_t offset;
422
423         spin_lock(&swap_lock);
424         si = swap_info + type;
425         if (si->flags & SWP_WRITEOK) {
426                 nr_swap_pages--;
427                 offset = scan_swap_map(si);
428                 if (offset) {
429                         spin_unlock(&swap_lock);
430                         return swp_entry(type, offset);
431                 }
432                 nr_swap_pages++;
433         }
434         spin_unlock(&swap_lock);
435         return (swp_entry_t) {0};
436 }
437
438 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
439 {
440         struct swap_info_struct * p;
441         unsigned long offset, type;
442
443         if (!entry.val)
444                 goto out;
445         type = swp_type(entry);
446         if (type >= nr_swapfiles)
447                 goto bad_nofile;
448         p = & swap_info[type];
449         if (!(p->flags & SWP_USED))
450                 goto bad_device;
451         offset = swp_offset(entry);
452         if (offset >= p->max)
453                 goto bad_offset;
454         if (!p->swap_map[offset])
455                 goto bad_free;
456         spin_lock(&swap_lock);
457         return p;
458
459 bad_free:
460         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
461         goto out;
462 bad_offset:
463         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
464         goto out;
465 bad_device:
466         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
467         goto out;
468 bad_nofile:
469         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
470 out:
471         return NULL;
472 }
473
474 static int swap_entry_free(struct swap_info_struct *p, swp_entry_t ent)
475 {
476         unsigned long offset = swp_offset(ent);
477         int count = p->swap_map[offset];
478
479         if (count < SWAP_MAP_MAX) {
480                 count--;
481                 p->swap_map[offset] = count;
482                 if (!count) {
483                         if (offset < p->lowest_bit)
484                                 p->lowest_bit = offset;
485                         if (offset > p->highest_bit)
486                                 p->highest_bit = offset;
487                         if (p->prio > swap_info[swap_list.next].prio)
488                                 swap_list.next = p - swap_info;
489                         nr_swap_pages++;
490                         p->inuse_pages--;
491                         mem_cgroup_uncharge_swap(ent);
492                 }
493         }
494         return count;
495 }
496
497 /*
498  * Caller has made sure that the swapdevice corresponding to entry
499  * is still around or has not been recycled.
500  */
501 void swap_free(swp_entry_t entry)
502 {
503         struct swap_info_struct * p;
504
505         p = swap_info_get(entry);
506         if (p) {
507                 swap_entry_free(p, entry);
508                 spin_unlock(&swap_lock);
509         }
510 }
511
512 /*
513  * Called after dropping swapcache to decrease refcnt to swap entries.
514  */
515 void swapcache_free(swp_entry_t entry, struct page *page)
516 {
517         if (page)
518                 mem_cgroup_uncharge_swapcache(page, entry);
519         return swap_free(entry);
520 }
521
522 /*
523  * How many references to page are currently swapped out?
524  */
525 static inline int page_swapcount(struct page *page)
526 {
527         int count = 0;
528         struct swap_info_struct *p;
529         swp_entry_t entry;
530
531         entry.val = page_private(page);
532         p = swap_info_get(entry);
533         if (p) {
534                 /* Subtract the 1 for the swap cache itself */
535                 count = p->swap_map[swp_offset(entry)] - 1;
536                 spin_unlock(&swap_lock);
537         }
538         return count;
539 }
540
541 /*
542  * We can write to an anon page without COW if there are no other references
543  * to it.  And as a side-effect, free up its swap: because the old content
544  * on disk will never be read, and seeking back there to write new content
545  * later would only waste time away from clustering.
546  */
547 int reuse_swap_page(struct page *page)
548 {
549         int count;
550
551         VM_BUG_ON(!PageLocked(page));
552         count = page_mapcount(page);
553         if (count <= 1 && PageSwapCache(page)) {
554                 count += page_swapcount(page);
555                 if (count == 1 && !PageWriteback(page)) {
556                         delete_from_swap_cache(page);
557                         SetPageDirty(page);
558                 }
559         }
560         return count == 1;
561 }
562
563 /*
564  * If swap is getting full, or if there are no more mappings of this page,
565  * then try_to_free_swap is called to free its swap space.
566  */
567 int try_to_free_swap(struct page *page)
568 {
569         VM_BUG_ON(!PageLocked(page));
570
571         if (!PageSwapCache(page))
572                 return 0;
573         if (PageWriteback(page))
574                 return 0;
575         if (page_swapcount(page))
576                 return 0;
577
578         delete_from_swap_cache(page);
579         SetPageDirty(page);
580         return 1;
581 }
582
583 /*
584  * Free the swap entry like above, but also try to
585  * free the page cache entry if it is the last user.
586  */
587 int free_swap_and_cache(swp_entry_t entry)
588 {
589         struct swap_info_struct *p;
590         struct page *page = NULL;
591
592         if (is_migration_entry(entry))
593                 return 1;
594
595         p = swap_info_get(entry);
596         if (p) {
597                 if (swap_entry_free(p, entry) == 1) {
598                         page = find_get_page(&swapper_space, entry.val);
599                         if (page && !trylock_page(page)) {
600                                 page_cache_release(page);
601                                 page = NULL;
602                         }
603                 }
604                 spin_unlock(&swap_lock);
605         }
606         if (page) {
607                 /*
608                  * Not mapped elsewhere, or swap space full? Free it!
609                  * Also recheck PageSwapCache now page is locked (above).
610                  */
611                 if (PageSwapCache(page) && !PageWriteback(page) &&
612                                 (!page_mapped(page) || vm_swap_full())) {
613                         delete_from_swap_cache(page);
614                         SetPageDirty(page);
615                 }
616                 unlock_page(page);
617                 page_cache_release(page);
618         }
619         return p != NULL;
620 }
621
622 #ifdef CONFIG_HIBERNATION
623 /*
624  * Find the swap type that corresponds to given device (if any).
625  *
626  * @offset - number of the PAGE_SIZE-sized block of the device, starting
627  * from 0, in which the swap header is expected to be located.
628  *
629  * This is needed for the suspend to disk (aka swsusp).
630  */
631 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
632 {
633         struct block_device *bdev = NULL;
634         int i;
635
636         if (device)
637                 bdev = bdget(device);
638
639         spin_lock(&swap_lock);
640         for (i = 0; i < nr_swapfiles; i++) {
641                 struct swap_info_struct *sis = swap_info + i;
642
643                 if (!(sis->flags & SWP_WRITEOK))
644                         continue;
645
646                 if (!bdev) {
647                         if (bdev_p)
648                                 *bdev_p = bdget(sis->bdev->bd_dev);
649
650                         spin_unlock(&swap_lock);
651                         return i;
652                 }
653                 if (bdev == sis->bdev) {
654                         struct swap_extent *se;
655
656                         se = list_entry(sis->extent_list.next,
657                                         struct swap_extent, list);
658                         if (se->start_block == offset) {
659                                 if (bdev_p)
660                                         *bdev_p = bdget(sis->bdev->bd_dev);
661
662                                 spin_unlock(&swap_lock);
663                                 bdput(bdev);
664                                 return i;
665                         }
666                 }
667         }
668         spin_unlock(&swap_lock);
669         if (bdev)
670                 bdput(bdev);
671
672         return -ENODEV;
673 }
674
675 /*
676  * Return either the total number of swap pages of given type, or the number
677  * of free pages of that type (depending on @free)
678  *
679  * This is needed for software suspend
680  */
681 unsigned int count_swap_pages(int type, int free)
682 {
683         unsigned int n = 0;
684
685         if (type < nr_swapfiles) {
686                 spin_lock(&swap_lock);
687                 if (swap_info[type].flags & SWP_WRITEOK) {
688                         n = swap_info[type].pages;
689                         if (free)
690                                 n -= swap_info[type].inuse_pages;
691                 }
692                 spin_unlock(&swap_lock);
693         }
694         return n;
695 }
696 #endif
697
698 /*
699  * No need to decide whether this PTE shares the swap entry with others,
700  * just let do_wp_page work it out if a write is requested later - to
701  * force COW, vm_page_prot omits write permission from any private vma.
702  */
703 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
704                 unsigned long addr, swp_entry_t entry, struct page *page)
705 {
706         struct mem_cgroup *ptr = NULL;
707         spinlock_t *ptl;
708         pte_t *pte;
709         int ret = 1;
710
711         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
712                 ret = -ENOMEM;
713                 goto out_nolock;
714         }
715
716         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
717         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
718                 if (ret > 0)
719                         mem_cgroup_cancel_charge_swapin(ptr);
720                 ret = 0;
721                 goto out;
722         }
723
724         inc_mm_counter(vma->vm_mm, anon_rss);
725         get_page(page);
726         set_pte_at(vma->vm_mm, addr, pte,
727                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
728         page_add_anon_rmap(page, vma, addr);
729         mem_cgroup_commit_charge_swapin(page, ptr);
730         swap_free(entry);
731         /*
732          * Move the page to the active list so it is not
733          * immediately swapped out again after swapon.
734          */
735         activate_page(page);
736 out:
737         pte_unmap_unlock(pte, ptl);
738 out_nolock:
739         return ret;
740 }
741
742 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
743                                 unsigned long addr, unsigned long end,
744                                 swp_entry_t entry, struct page *page)
745 {
746         pte_t swp_pte = swp_entry_to_pte(entry);
747         pte_t *pte;
748         int ret = 0;
749
750         /*
751          * We don't actually need pte lock while scanning for swp_pte: since
752          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
753          * page table while we're scanning; though it could get zapped, and on
754          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
755          * of unmatched parts which look like swp_pte, so unuse_pte must
756          * recheck under pte lock.  Scanning without pte lock lets it be
757          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
758          */
759         pte = pte_offset_map(pmd, addr);
760         do {
761                 /*
762                  * swapoff spends a _lot_ of time in this loop!
763                  * Test inline before going to call unuse_pte.
764                  */
765                 if (unlikely(pte_same(*pte, swp_pte))) {
766                         pte_unmap(pte);
767                         ret = unuse_pte(vma, pmd, addr, entry, page);
768                         if (ret)
769                                 goto out;
770                         pte = pte_offset_map(pmd, addr);
771                 }
772         } while (pte++, addr += PAGE_SIZE, addr != end);
773         pte_unmap(pte - 1);
774 out:
775         return ret;
776 }
777
778 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
779                                 unsigned long addr, unsigned long end,
780                                 swp_entry_t entry, struct page *page)
781 {
782         pmd_t *pmd;
783         unsigned long next;
784         int ret;
785
786         pmd = pmd_offset(pud, addr);
787         do {
788                 next = pmd_addr_end(addr, end);
789                 if (pmd_none_or_clear_bad(pmd))
790                         continue;
791                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
792                 if (ret)
793                         return ret;
794         } while (pmd++, addr = next, addr != end);
795         return 0;
796 }
797
798 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
799                                 unsigned long addr, unsigned long end,
800                                 swp_entry_t entry, struct page *page)
801 {
802         pud_t *pud;
803         unsigned long next;
804         int ret;
805
806         pud = pud_offset(pgd, addr);
807         do {
808                 next = pud_addr_end(addr, end);
809                 if (pud_none_or_clear_bad(pud))
810                         continue;
811                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
812                 if (ret)
813                         return ret;
814         } while (pud++, addr = next, addr != end);
815         return 0;
816 }
817
818 static int unuse_vma(struct vm_area_struct *vma,
819                                 swp_entry_t entry, struct page *page)
820 {
821         pgd_t *pgd;
822         unsigned long addr, end, next;
823         int ret;
824
825         if (page->mapping) {
826                 addr = page_address_in_vma(page, vma);
827                 if (addr == -EFAULT)
828                         return 0;
829                 else
830                         end = addr + PAGE_SIZE;
831         } else {
832                 addr = vma->vm_start;
833                 end = vma->vm_end;
834         }
835
836         pgd = pgd_offset(vma->vm_mm, addr);
837         do {
838                 next = pgd_addr_end(addr, end);
839                 if (pgd_none_or_clear_bad(pgd))
840                         continue;
841                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
842                 if (ret)
843                         return ret;
844         } while (pgd++, addr = next, addr != end);
845         return 0;
846 }
847
848 static int unuse_mm(struct mm_struct *mm,
849                                 swp_entry_t entry, struct page *page)
850 {
851         struct vm_area_struct *vma;
852         int ret = 0;
853
854         if (!down_read_trylock(&mm->mmap_sem)) {
855                 /*
856                  * Activate page so shrink_inactive_list is unlikely to unmap
857                  * its ptes while lock is dropped, so swapoff can make progress.
858                  */
859                 activate_page(page);
860                 unlock_page(page);
861                 down_read(&mm->mmap_sem);
862                 lock_page(page);
863         }
864         for (vma = mm->mmap; vma; vma = vma->vm_next) {
865                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
866                         break;
867         }
868         up_read(&mm->mmap_sem);
869         return (ret < 0)? ret: 0;
870 }
871
872 /*
873  * Scan swap_map from current position to next entry still in use.
874  * Recycle to start on reaching the end, returning 0 when empty.
875  */
876 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
877                                         unsigned int prev)
878 {
879         unsigned int max = si->max;
880         unsigned int i = prev;
881         int count;
882
883         /*
884          * No need for swap_lock here: we're just looking
885          * for whether an entry is in use, not modifying it; false
886          * hits are okay, and sys_swapoff() has already prevented new
887          * allocations from this area (while holding swap_lock).
888          */
889         for (;;) {
890                 if (++i >= max) {
891                         if (!prev) {
892                                 i = 0;
893                                 break;
894                         }
895                         /*
896                          * No entries in use at top of swap_map,
897                          * loop back to start and recheck there.
898                          */
899                         max = prev + 1;
900                         prev = 0;
901                         i = 1;
902                 }
903                 count = si->swap_map[i];
904                 if (count && count != SWAP_MAP_BAD)
905                         break;
906         }
907         return i;
908 }
909
910 /*
911  * We completely avoid races by reading each swap page in advance,
912  * and then search for the process using it.  All the necessary
913  * page table adjustments can then be made atomically.
914  */
915 static int try_to_unuse(unsigned int type)
916 {
917         struct swap_info_struct * si = &swap_info[type];
918         struct mm_struct *start_mm;
919         unsigned short *swap_map;
920         unsigned short swcount;
921         struct page *page;
922         swp_entry_t entry;
923         unsigned int i = 0;
924         int retval = 0;
925         int reset_overflow = 0;
926         int shmem;
927
928         /*
929          * When searching mms for an entry, a good strategy is to
930          * start at the first mm we freed the previous entry from
931          * (though actually we don't notice whether we or coincidence
932          * freed the entry).  Initialize this start_mm with a hold.
933          *
934          * A simpler strategy would be to start at the last mm we
935          * freed the previous entry from; but that would take less
936          * advantage of mmlist ordering, which clusters forked mms
937          * together, child after parent.  If we race with dup_mmap(), we
938          * prefer to resolve parent before child, lest we miss entries
939          * duplicated after we scanned child: using last mm would invert
940          * that.  Though it's only a serious concern when an overflowed
941          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
942          */
943         start_mm = &init_mm;
944         atomic_inc(&init_mm.mm_users);
945
946         /*
947          * Keep on scanning until all entries have gone.  Usually,
948          * one pass through swap_map is enough, but not necessarily:
949          * there are races when an instance of an entry might be missed.
950          */
951         while ((i = find_next_to_unuse(si, i)) != 0) {
952                 if (signal_pending(current)) {
953                         retval = -EINTR;
954                         break;
955                 }
956
957                 /*
958                  * Get a page for the entry, using the existing swap
959                  * cache page if there is one.  Otherwise, get a clean
960                  * page and read the swap into it.
961                  */
962                 swap_map = &si->swap_map[i];
963                 entry = swp_entry(type, i);
964                 page = read_swap_cache_async(entry,
965                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
966                 if (!page) {
967                         /*
968                          * Either swap_duplicate() failed because entry
969                          * has been freed independently, and will not be
970                          * reused since sys_swapoff() already disabled
971                          * allocation from here, or alloc_page() failed.
972                          */
973                         if (!*swap_map)
974                                 continue;
975                         retval = -ENOMEM;
976                         break;
977                 }
978
979                 /*
980                  * Don't hold on to start_mm if it looks like exiting.
981                  */
982                 if (atomic_read(&start_mm->mm_users) == 1) {
983                         mmput(start_mm);
984                         start_mm = &init_mm;
985                         atomic_inc(&init_mm.mm_users);
986                 }
987
988                 /*
989                  * Wait for and lock page.  When do_swap_page races with
990                  * try_to_unuse, do_swap_page can handle the fault much
991                  * faster than try_to_unuse can locate the entry.  This
992                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
993                  * defer to do_swap_page in such a case - in some tests,
994                  * do_swap_page and try_to_unuse repeatedly compete.
995                  */
996                 wait_on_page_locked(page);
997                 wait_on_page_writeback(page);
998                 lock_page(page);
999                 wait_on_page_writeback(page);
1000
1001                 /*
1002                  * Remove all references to entry.
1003                  * Whenever we reach init_mm, there's no address space
1004                  * to search, but use it as a reminder to search shmem.
1005                  */
1006                 shmem = 0;
1007                 swcount = *swap_map;
1008                 if (swcount > 1) {
1009                         if (start_mm == &init_mm)
1010                                 shmem = shmem_unuse(entry, page);
1011                         else
1012                                 retval = unuse_mm(start_mm, entry, page);
1013                 }
1014                 if (*swap_map > 1) {
1015                         int set_start_mm = (*swap_map >= swcount);
1016                         struct list_head *p = &start_mm->mmlist;
1017                         struct mm_struct *new_start_mm = start_mm;
1018                         struct mm_struct *prev_mm = start_mm;
1019                         struct mm_struct *mm;
1020
1021                         atomic_inc(&new_start_mm->mm_users);
1022                         atomic_inc(&prev_mm->mm_users);
1023                         spin_lock(&mmlist_lock);
1024                         while (*swap_map > 1 && !retval && !shmem &&
1025                                         (p = p->next) != &start_mm->mmlist) {
1026                                 mm = list_entry(p, struct mm_struct, mmlist);
1027                                 if (!atomic_inc_not_zero(&mm->mm_users))
1028                                         continue;
1029                                 spin_unlock(&mmlist_lock);
1030                                 mmput(prev_mm);
1031                                 prev_mm = mm;
1032
1033                                 cond_resched();
1034
1035                                 swcount = *swap_map;
1036                                 if (swcount <= 1)
1037                                         ;
1038                                 else if (mm == &init_mm) {
1039                                         set_start_mm = 1;
1040                                         shmem = shmem_unuse(entry, page);
1041                                 } else
1042                                         retval = unuse_mm(mm, entry, page);
1043                                 if (set_start_mm && *swap_map < swcount) {
1044                                         mmput(new_start_mm);
1045                                         atomic_inc(&mm->mm_users);
1046                                         new_start_mm = mm;
1047                                         set_start_mm = 0;
1048                                 }
1049                                 spin_lock(&mmlist_lock);
1050                         }
1051                         spin_unlock(&mmlist_lock);
1052                         mmput(prev_mm);
1053                         mmput(start_mm);
1054                         start_mm = new_start_mm;
1055                 }
1056                 if (shmem) {
1057                         /* page has already been unlocked and released */
1058                         if (shmem > 0)
1059                                 continue;
1060                         retval = shmem;
1061                         break;
1062                 }
1063                 if (retval) {
1064                         unlock_page(page);
1065                         page_cache_release(page);
1066                         break;
1067                 }
1068
1069                 /*
1070                  * How could swap count reach 0x7fff when the maximum
1071                  * pid is 0x7fff, and there's no way to repeat a swap
1072                  * page within an mm (except in shmem, where it's the
1073                  * shared object which takes the reference count)?
1074                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
1075                  *
1076                  * If that's wrong, then we should worry more about
1077                  * exit_mmap() and do_munmap() cases described above:
1078                  * we might be resetting SWAP_MAP_MAX too early here.
1079                  * We know "Undead"s can happen, they're okay, so don't
1080                  * report them; but do report if we reset SWAP_MAP_MAX.
1081                  */
1082                 if (*swap_map == SWAP_MAP_MAX) {
1083                         spin_lock(&swap_lock);
1084                         *swap_map = 1;
1085                         spin_unlock(&swap_lock);
1086                         reset_overflow = 1;
1087                 }
1088
1089                 /*
1090                  * If a reference remains (rare), we would like to leave
1091                  * the page in the swap cache; but try_to_unmap could
1092                  * then re-duplicate the entry once we drop page lock,
1093                  * so we might loop indefinitely; also, that page could
1094                  * not be swapped out to other storage meanwhile.  So:
1095                  * delete from cache even if there's another reference,
1096                  * after ensuring that the data has been saved to disk -
1097                  * since if the reference remains (rarer), it will be
1098                  * read from disk into another page.  Splitting into two
1099                  * pages would be incorrect if swap supported "shared
1100                  * private" pages, but they are handled by tmpfs files.
1101                  */
1102                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
1103                         struct writeback_control wbc = {
1104                                 .sync_mode = WB_SYNC_NONE,
1105                         };
1106
1107                         swap_writepage(page, &wbc);
1108                         lock_page(page);
1109                         wait_on_page_writeback(page);
1110                 }
1111
1112                 /*
1113                  * It is conceivable that a racing task removed this page from
1114                  * swap cache just before we acquired the page lock at the top,
1115                  * or while we dropped it in unuse_mm().  The page might even
1116                  * be back in swap cache on another swap area: that we must not
1117                  * delete, since it may not have been written out to swap yet.
1118                  */
1119                 if (PageSwapCache(page) &&
1120                     likely(page_private(page) == entry.val))
1121                         delete_from_swap_cache(page);
1122
1123                 /*
1124                  * So we could skip searching mms once swap count went
1125                  * to 1, we did not mark any present ptes as dirty: must
1126                  * mark page dirty so shrink_page_list will preserve it.
1127                  */
1128                 SetPageDirty(page);
1129                 unlock_page(page);
1130                 page_cache_release(page);
1131
1132                 /*
1133                  * Make sure that we aren't completely killing
1134                  * interactive performance.
1135                  */
1136                 cond_resched();
1137         }
1138
1139         mmput(start_mm);
1140         if (reset_overflow) {
1141                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1142                 swap_overflow = 0;
1143         }
1144         return retval;
1145 }
1146
1147 /*
1148  * After a successful try_to_unuse, if no swap is now in use, we know
1149  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1150  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1151  * added to the mmlist just after page_duplicate - before would be racy.
1152  */
1153 static void drain_mmlist(void)
1154 {
1155         struct list_head *p, *next;
1156         unsigned int i;
1157
1158         for (i = 0; i < nr_swapfiles; i++)
1159                 if (swap_info[i].inuse_pages)
1160                         return;
1161         spin_lock(&mmlist_lock);
1162         list_for_each_safe(p, next, &init_mm.mmlist)
1163                 list_del_init(p);
1164         spin_unlock(&mmlist_lock);
1165 }
1166
1167 /*
1168  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1169  * corresponds to page offset `offset'.
1170  */
1171 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1172 {
1173         struct swap_extent *se = sis->curr_swap_extent;
1174         struct swap_extent *start_se = se;
1175
1176         for ( ; ; ) {
1177                 struct list_head *lh;
1178
1179                 if (se->start_page <= offset &&
1180                                 offset < (se->start_page + se->nr_pages)) {
1181                         return se->start_block + (offset - se->start_page);
1182                 }
1183                 lh = se->list.next;
1184                 if (lh == &sis->extent_list)
1185                         lh = lh->next;
1186                 se = list_entry(lh, struct swap_extent, list);
1187                 sis->curr_swap_extent = se;
1188                 BUG_ON(se == start_se);         /* It *must* be present */
1189         }
1190 }
1191
1192 #ifdef CONFIG_HIBERNATION
1193 /*
1194  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1195  * corresponding to given index in swap_info (swap type).
1196  */
1197 sector_t swapdev_block(int swap_type, pgoff_t offset)
1198 {
1199         struct swap_info_struct *sis;
1200
1201         if (swap_type >= nr_swapfiles)
1202                 return 0;
1203
1204         sis = swap_info + swap_type;
1205         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1206 }
1207 #endif /* CONFIG_HIBERNATION */
1208
1209 /*
1210  * Free all of a swapdev's extent information
1211  */
1212 static void destroy_swap_extents(struct swap_info_struct *sis)
1213 {
1214         while (!list_empty(&sis->extent_list)) {
1215                 struct swap_extent *se;
1216
1217                 se = list_entry(sis->extent_list.next,
1218                                 struct swap_extent, list);
1219                 list_del(&se->list);
1220                 kfree(se);
1221         }
1222 }
1223
1224 /*
1225  * Add a block range (and the corresponding page range) into this swapdev's
1226  * extent list.  The extent list is kept sorted in page order.
1227  *
1228  * This function rather assumes that it is called in ascending page order.
1229  */
1230 static int
1231 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1232                 unsigned long nr_pages, sector_t start_block)
1233 {
1234         struct swap_extent *se;
1235         struct swap_extent *new_se;
1236         struct list_head *lh;
1237
1238         lh = sis->extent_list.prev;     /* The highest page extent */
1239         if (lh != &sis->extent_list) {
1240                 se = list_entry(lh, struct swap_extent, list);
1241                 BUG_ON(se->start_page + se->nr_pages != start_page);
1242                 if (se->start_block + se->nr_pages == start_block) {
1243                         /* Merge it */
1244                         se->nr_pages += nr_pages;
1245                         return 0;
1246                 }
1247         }
1248
1249         /*
1250          * No merge.  Insert a new extent, preserving ordering.
1251          */
1252         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1253         if (new_se == NULL)
1254                 return -ENOMEM;
1255         new_se->start_page = start_page;
1256         new_se->nr_pages = nr_pages;
1257         new_se->start_block = start_block;
1258
1259         list_add_tail(&new_se->list, &sis->extent_list);
1260         return 1;
1261 }
1262
1263 /*
1264  * A `swap extent' is a simple thing which maps a contiguous range of pages
1265  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1266  * is built at swapon time and is then used at swap_writepage/swap_readpage
1267  * time for locating where on disk a page belongs.
1268  *
1269  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1270  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1271  * swap files identically.
1272  *
1273  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1274  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1275  * swapfiles are handled *identically* after swapon time.
1276  *
1277  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1278  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1279  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1280  * requirements, they are simply tossed out - we will never use those blocks
1281  * for swapping.
1282  *
1283  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1284  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1285  * which will scribble on the fs.
1286  *
1287  * The amount of disk space which a single swap extent represents varies.
1288  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1289  * extents in the list.  To avoid much list walking, we cache the previous
1290  * search location in `curr_swap_extent', and start new searches from there.
1291  * This is extremely effective.  The average number of iterations in
1292  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1293  */
1294 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1295 {
1296         struct inode *inode;
1297         unsigned blocks_per_page;
1298         unsigned long page_no;
1299         unsigned blkbits;
1300         sector_t probe_block;
1301         sector_t last_block;
1302         sector_t lowest_block = -1;
1303         sector_t highest_block = 0;
1304         int nr_extents = 0;
1305         int ret;
1306
1307         inode = sis->swap_file->f_mapping->host;
1308         if (S_ISBLK(inode->i_mode)) {
1309                 ret = add_swap_extent(sis, 0, sis->max, 0);
1310                 *span = sis->pages;
1311                 goto done;
1312         }
1313
1314         blkbits = inode->i_blkbits;
1315         blocks_per_page = PAGE_SIZE >> blkbits;
1316
1317         /*
1318          * Map all the blocks into the extent list.  This code doesn't try
1319          * to be very smart.
1320          */
1321         probe_block = 0;
1322         page_no = 0;
1323         last_block = i_size_read(inode) >> blkbits;
1324         while ((probe_block + blocks_per_page) <= last_block &&
1325                         page_no < sis->max) {
1326                 unsigned block_in_page;
1327                 sector_t first_block;
1328
1329                 first_block = bmap(inode, probe_block);
1330                 if (first_block == 0)
1331                         goto bad_bmap;
1332
1333                 /*
1334                  * It must be PAGE_SIZE aligned on-disk
1335                  */
1336                 if (first_block & (blocks_per_page - 1)) {
1337                         probe_block++;
1338                         goto reprobe;
1339                 }
1340
1341                 for (block_in_page = 1; block_in_page < blocks_per_page;
1342                                         block_in_page++) {
1343                         sector_t block;
1344
1345                         block = bmap(inode, probe_block + block_in_page);
1346                         if (block == 0)
1347                                 goto bad_bmap;
1348                         if (block != first_block + block_in_page) {
1349                                 /* Discontiguity */
1350                                 probe_block++;
1351                                 goto reprobe;
1352                         }
1353                 }
1354
1355                 first_block >>= (PAGE_SHIFT - blkbits);
1356                 if (page_no) {  /* exclude the header page */
1357                         if (first_block < lowest_block)
1358                                 lowest_block = first_block;
1359                         if (first_block > highest_block)
1360                                 highest_block = first_block;
1361                 }
1362
1363                 /*
1364                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1365                  */
1366                 ret = add_swap_extent(sis, page_no, 1, first_block);
1367                 if (ret < 0)
1368                         goto out;
1369                 nr_extents += ret;
1370                 page_no++;
1371                 probe_block += blocks_per_page;
1372 reprobe:
1373                 continue;
1374         }
1375         ret = nr_extents;
1376         *span = 1 + highest_block - lowest_block;
1377         if (page_no == 0)
1378                 page_no = 1;    /* force Empty message */
1379         sis->max = page_no;
1380         sis->pages = page_no - 1;
1381         sis->highest_bit = page_no - 1;
1382 done:
1383         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1384                                         struct swap_extent, list);
1385         goto out;
1386 bad_bmap:
1387         printk(KERN_ERR "swapon: swapfile has holes\n");
1388         ret = -EINVAL;
1389 out:
1390         return ret;
1391 }
1392
1393 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1394 {
1395         struct swap_info_struct * p = NULL;
1396         unsigned short *swap_map;
1397         struct file *swap_file, *victim;
1398         struct address_space *mapping;
1399         struct inode *inode;
1400         char * pathname;
1401         int i, type, prev;
1402         int err;
1403
1404         if (!capable(CAP_SYS_ADMIN))
1405                 return -EPERM;
1406
1407         pathname = getname(specialfile);
1408         err = PTR_ERR(pathname);
1409         if (IS_ERR(pathname))
1410                 goto out;
1411
1412         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1413         putname(pathname);
1414         err = PTR_ERR(victim);
1415         if (IS_ERR(victim))
1416                 goto out;
1417
1418         mapping = victim->f_mapping;
1419         prev = -1;
1420         spin_lock(&swap_lock);
1421         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1422                 p = swap_info + type;
1423                 if (p->flags & SWP_WRITEOK) {
1424                         if (p->swap_file->f_mapping == mapping)
1425                                 break;
1426                 }
1427                 prev = type;
1428         }
1429         if (type < 0) {
1430                 err = -EINVAL;
1431                 spin_unlock(&swap_lock);
1432                 goto out_dput;
1433         }
1434         if (!security_vm_enough_memory(p->pages))
1435                 vm_unacct_memory(p->pages);
1436         else {
1437                 err = -ENOMEM;
1438                 spin_unlock(&swap_lock);
1439                 goto out_dput;
1440         }
1441         if (prev < 0) {
1442                 swap_list.head = p->next;
1443         } else {
1444                 swap_info[prev].next = p->next;
1445         }
1446         if (type == swap_list.next) {
1447                 /* just pick something that's safe... */
1448                 swap_list.next = swap_list.head;
1449         }
1450         if (p->prio < 0) {
1451                 for (i = p->next; i >= 0; i = swap_info[i].next)
1452                         swap_info[i].prio = p->prio--;
1453                 least_priority++;
1454         }
1455         nr_swap_pages -= p->pages;
1456         total_swap_pages -= p->pages;
1457         p->flags &= ~SWP_WRITEOK;
1458         spin_unlock(&swap_lock);
1459
1460         current->flags |= PF_SWAPOFF;
1461         err = try_to_unuse(type);
1462         current->flags &= ~PF_SWAPOFF;
1463
1464         if (err) {
1465                 /* re-insert swap space back into swap_list */
1466                 spin_lock(&swap_lock);
1467                 if (p->prio < 0)
1468                         p->prio = --least_priority;
1469                 prev = -1;
1470                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1471                         if (p->prio >= swap_info[i].prio)
1472                                 break;
1473                         prev = i;
1474                 }
1475                 p->next = i;
1476                 if (prev < 0)
1477                         swap_list.head = swap_list.next = p - swap_info;
1478                 else
1479                         swap_info[prev].next = p - swap_info;
1480                 nr_swap_pages += p->pages;
1481                 total_swap_pages += p->pages;
1482                 p->flags |= SWP_WRITEOK;
1483                 spin_unlock(&swap_lock);
1484                 goto out_dput;
1485         }
1486
1487         /* wait for any unplug function to finish */
1488         down_write(&swap_unplug_sem);
1489         up_write(&swap_unplug_sem);
1490
1491         destroy_swap_extents(p);
1492         mutex_lock(&swapon_mutex);
1493         spin_lock(&swap_lock);
1494         drain_mmlist();
1495
1496         /* wait for anyone still in scan_swap_map */
1497         p->highest_bit = 0;             /* cuts scans short */
1498         while (p->flags >= SWP_SCANNING) {
1499                 spin_unlock(&swap_lock);
1500                 schedule_timeout_uninterruptible(1);
1501                 spin_lock(&swap_lock);
1502         }
1503
1504         swap_file = p->swap_file;
1505         p->swap_file = NULL;
1506         p->max = 0;
1507         swap_map = p->swap_map;
1508         p->swap_map = NULL;
1509         p->flags = 0;
1510         spin_unlock(&swap_lock);
1511         mutex_unlock(&swapon_mutex);
1512         vfree(swap_map);
1513         /* Destroy swap account informatin */
1514         swap_cgroup_swapoff(type);
1515
1516         inode = mapping->host;
1517         if (S_ISBLK(inode->i_mode)) {
1518                 struct block_device *bdev = I_BDEV(inode);
1519                 set_blocksize(bdev, p->old_block_size);
1520                 bd_release(bdev);
1521         } else {
1522                 mutex_lock(&inode->i_mutex);
1523                 inode->i_flags &= ~S_SWAPFILE;
1524                 mutex_unlock(&inode->i_mutex);
1525         }
1526         filp_close(swap_file, NULL);
1527         err = 0;
1528
1529 out_dput:
1530         filp_close(victim, NULL);
1531 out:
1532         return err;
1533 }
1534
1535 #ifdef CONFIG_PROC_FS
1536 /* iterator */
1537 static void *swap_start(struct seq_file *swap, loff_t *pos)
1538 {
1539         struct swap_info_struct *ptr = swap_info;
1540         int i;
1541         loff_t l = *pos;
1542
1543         mutex_lock(&swapon_mutex);
1544
1545         if (!l)
1546                 return SEQ_START_TOKEN;
1547
1548         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1549                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1550                         continue;
1551                 if (!--l)
1552                         return ptr;
1553         }
1554
1555         return NULL;
1556 }
1557
1558 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1559 {
1560         struct swap_info_struct *ptr;
1561         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1562
1563         if (v == SEQ_START_TOKEN)
1564                 ptr = swap_info;
1565         else {
1566                 ptr = v;
1567                 ptr++;
1568         }
1569
1570         for (; ptr < endptr; ptr++) {
1571                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1572                         continue;
1573                 ++*pos;
1574                 return ptr;
1575         }
1576
1577         return NULL;
1578 }
1579
1580 static void swap_stop(struct seq_file *swap, void *v)
1581 {
1582         mutex_unlock(&swapon_mutex);
1583 }
1584
1585 static int swap_show(struct seq_file *swap, void *v)
1586 {
1587         struct swap_info_struct *ptr = v;
1588         struct file *file;
1589         int len;
1590
1591         if (ptr == SEQ_START_TOKEN) {
1592                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1593                 return 0;
1594         }
1595
1596         file = ptr->swap_file;
1597         len = seq_path(swap, &file->f_path, " \t\n\\");
1598         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1599                         len < 40 ? 40 - len : 1, " ",
1600                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1601                                 "partition" : "file\t",
1602                         ptr->pages << (PAGE_SHIFT - 10),
1603                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1604                         ptr->prio);
1605         return 0;
1606 }
1607
1608 static const struct seq_operations swaps_op = {
1609         .start =        swap_start,
1610         .next =         swap_next,
1611         .stop =         swap_stop,
1612         .show =         swap_show
1613 };
1614
1615 static int swaps_open(struct inode *inode, struct file *file)
1616 {
1617         return seq_open(file, &swaps_op);
1618 }
1619
1620 static const struct file_operations proc_swaps_operations = {
1621         .open           = swaps_open,
1622         .read           = seq_read,
1623         .llseek         = seq_lseek,
1624         .release        = seq_release,
1625 };
1626
1627 static int __init procswaps_init(void)
1628 {
1629         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1630         return 0;
1631 }
1632 __initcall(procswaps_init);
1633 #endif /* CONFIG_PROC_FS */
1634
1635 #ifdef MAX_SWAPFILES_CHECK
1636 static int __init max_swapfiles_check(void)
1637 {
1638         MAX_SWAPFILES_CHECK();
1639         return 0;
1640 }
1641 late_initcall(max_swapfiles_check);
1642 #endif
1643
1644 /*
1645  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1646  *
1647  * The swapon system call
1648  */
1649 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1650 {
1651         struct swap_info_struct * p;
1652         char *name = NULL;
1653         struct block_device *bdev = NULL;
1654         struct file *swap_file = NULL;
1655         struct address_space *mapping;
1656         unsigned int type;
1657         int i, prev;
1658         int error;
1659         union swap_header *swap_header = NULL;
1660         unsigned int nr_good_pages = 0;
1661         int nr_extents = 0;
1662         sector_t span;
1663         unsigned long maxpages = 1;
1664         unsigned long swapfilepages;
1665         unsigned short *swap_map = NULL;
1666         struct page *page = NULL;
1667         struct inode *inode = NULL;
1668         int did_down = 0;
1669
1670         if (!capable(CAP_SYS_ADMIN))
1671                 return -EPERM;
1672         spin_lock(&swap_lock);
1673         p = swap_info;
1674         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1675                 if (!(p->flags & SWP_USED))
1676                         break;
1677         error = -EPERM;
1678         if (type >= MAX_SWAPFILES) {
1679                 spin_unlock(&swap_lock);
1680                 goto out;
1681         }
1682         if (type >= nr_swapfiles)
1683                 nr_swapfiles = type+1;
1684         memset(p, 0, sizeof(*p));
1685         INIT_LIST_HEAD(&p->extent_list);
1686         p->flags = SWP_USED;
1687         p->next = -1;
1688         spin_unlock(&swap_lock);
1689         name = getname(specialfile);
1690         error = PTR_ERR(name);
1691         if (IS_ERR(name)) {
1692                 name = NULL;
1693                 goto bad_swap_2;
1694         }
1695         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1696         error = PTR_ERR(swap_file);
1697         if (IS_ERR(swap_file)) {
1698                 swap_file = NULL;
1699                 goto bad_swap_2;
1700         }
1701
1702         p->swap_file = swap_file;
1703         mapping = swap_file->f_mapping;
1704         inode = mapping->host;
1705
1706         error = -EBUSY;
1707         for (i = 0; i < nr_swapfiles; i++) {
1708                 struct swap_info_struct *q = &swap_info[i];
1709
1710                 if (i == type || !q->swap_file)
1711                         continue;
1712                 if (mapping == q->swap_file->f_mapping)
1713                         goto bad_swap;
1714         }
1715
1716         error = -EINVAL;
1717         if (S_ISBLK(inode->i_mode)) {
1718                 bdev = I_BDEV(inode);
1719                 error = bd_claim(bdev, sys_swapon);
1720                 if (error < 0) {
1721                         bdev = NULL;
1722                         error = -EINVAL;
1723                         goto bad_swap;
1724                 }
1725                 p->old_block_size = block_size(bdev);
1726                 error = set_blocksize(bdev, PAGE_SIZE);
1727                 if (error < 0)
1728                         goto bad_swap;
1729                 p->bdev = bdev;
1730         } else if (S_ISREG(inode->i_mode)) {
1731                 p->bdev = inode->i_sb->s_bdev;
1732                 mutex_lock(&inode->i_mutex);
1733                 did_down = 1;
1734                 if (IS_SWAPFILE(inode)) {
1735                         error = -EBUSY;
1736                         goto bad_swap;
1737                 }
1738         } else {
1739                 goto bad_swap;
1740         }
1741
1742         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1743
1744         /*
1745          * Read the swap header.
1746          */
1747         if (!mapping->a_ops->readpage) {
1748                 error = -EINVAL;
1749                 goto bad_swap;
1750         }
1751         page = read_mapping_page(mapping, 0, swap_file);
1752         if (IS_ERR(page)) {
1753                 error = PTR_ERR(page);
1754                 goto bad_swap;
1755         }
1756         swap_header = kmap(page);
1757
1758         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1759                 printk(KERN_ERR "Unable to find swap-space signature\n");
1760                 error = -EINVAL;
1761                 goto bad_swap;
1762         }
1763
1764         /* swap partition endianess hack... */
1765         if (swab32(swap_header->info.version) == 1) {
1766                 swab32s(&swap_header->info.version);
1767                 swab32s(&swap_header->info.last_page);
1768                 swab32s(&swap_header->info.nr_badpages);
1769                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1770                         swab32s(&swap_header->info.badpages[i]);
1771         }
1772         /* Check the swap header's sub-version */
1773         if (swap_header->info.version != 1) {
1774                 printk(KERN_WARNING
1775                        "Unable to handle swap header version %d\n",
1776                        swap_header->info.version);
1777                 error = -EINVAL;
1778                 goto bad_swap;
1779         }
1780
1781         p->lowest_bit  = 1;
1782         p->cluster_next = 1;
1783
1784         /*
1785          * Find out how many pages are allowed for a single swap
1786          * device. There are two limiting factors: 1) the number of
1787          * bits for the swap offset in the swp_entry_t type and
1788          * 2) the number of bits in the a swap pte as defined by
1789          * the different architectures. In order to find the
1790          * largest possible bit mask a swap entry with swap type 0
1791          * and swap offset ~0UL is created, encoded to a swap pte,
1792          * decoded to a swp_entry_t again and finally the swap
1793          * offset is extracted. This will mask all the bits from
1794          * the initial ~0UL mask that can't be encoded in either
1795          * the swp_entry_t or the architecture definition of a
1796          * swap pte.
1797          */
1798         maxpages = swp_offset(pte_to_swp_entry(
1799                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1800         if (maxpages > swap_header->info.last_page)
1801                 maxpages = swap_header->info.last_page;
1802         p->highest_bit = maxpages - 1;
1803
1804         error = -EINVAL;
1805         if (!maxpages)
1806                 goto bad_swap;
1807         if (swapfilepages && maxpages > swapfilepages) {
1808                 printk(KERN_WARNING
1809                        "Swap area shorter than signature indicates\n");
1810                 goto bad_swap;
1811         }
1812         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1813                 goto bad_swap;
1814         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1815                 goto bad_swap;
1816
1817         /* OK, set up the swap map and apply the bad block list */
1818         swap_map = vmalloc(maxpages * sizeof(short));
1819         if (!swap_map) {
1820                 error = -ENOMEM;
1821                 goto bad_swap;
1822         }
1823
1824         memset(swap_map, 0, maxpages * sizeof(short));
1825         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1826                 int page_nr = swap_header->info.badpages[i];
1827                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1828                         error = -EINVAL;
1829                         goto bad_swap;
1830                 }
1831                 swap_map[page_nr] = SWAP_MAP_BAD;
1832         }
1833
1834         error = swap_cgroup_swapon(type, maxpages);
1835         if (error)
1836                 goto bad_swap;
1837
1838         nr_good_pages = swap_header->info.last_page -
1839                         swap_header->info.nr_badpages -
1840                         1 /* header page */;
1841
1842         if (nr_good_pages) {
1843                 swap_map[0] = SWAP_MAP_BAD;
1844                 p->max = maxpages;
1845                 p->pages = nr_good_pages;
1846                 nr_extents = setup_swap_extents(p, &span);
1847                 if (nr_extents < 0) {
1848                         error = nr_extents;
1849                         goto bad_swap;
1850                 }
1851                 nr_good_pages = p->pages;
1852         }
1853         if (!nr_good_pages) {
1854                 printk(KERN_WARNING "Empty swap-file\n");
1855                 error = -EINVAL;
1856                 goto bad_swap;
1857         }
1858
1859         if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1860                 p->flags |= SWP_SOLIDSTATE;
1861                 p->cluster_next = 1 + (random32() % p->highest_bit);
1862         }
1863         if (discard_swap(p) == 0)
1864                 p->flags |= SWP_DISCARDABLE;
1865
1866         mutex_lock(&swapon_mutex);
1867         spin_lock(&swap_lock);
1868         if (swap_flags & SWAP_FLAG_PREFER)
1869                 p->prio =
1870                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1871         else
1872                 p->prio = --least_priority;
1873         p->swap_map = swap_map;
1874         p->flags |= SWP_WRITEOK;
1875         nr_swap_pages += nr_good_pages;
1876         total_swap_pages += nr_good_pages;
1877
1878         printk(KERN_INFO "Adding %uk swap on %s.  "
1879                         "Priority:%d extents:%d across:%lluk %s%s\n",
1880                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1881                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1882                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1883                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1884
1885         /* insert swap space into swap_list: */
1886         prev = -1;
1887         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1888                 if (p->prio >= swap_info[i].prio) {
1889                         break;
1890                 }
1891                 prev = i;
1892         }
1893         p->next = i;
1894         if (prev < 0) {
1895                 swap_list.head = swap_list.next = p - swap_info;
1896         } else {
1897                 swap_info[prev].next = p - swap_info;
1898         }
1899         spin_unlock(&swap_lock);
1900         mutex_unlock(&swapon_mutex);
1901         error = 0;
1902         goto out;
1903 bad_swap:
1904         if (bdev) {
1905                 set_blocksize(bdev, p->old_block_size);
1906                 bd_release(bdev);
1907         }
1908         destroy_swap_extents(p);
1909         swap_cgroup_swapoff(type);
1910 bad_swap_2:
1911         spin_lock(&swap_lock);
1912         p->swap_file = NULL;
1913         p->flags = 0;
1914         spin_unlock(&swap_lock);
1915         vfree(swap_map);
1916         if (swap_file)
1917                 filp_close(swap_file, NULL);
1918 out:
1919         if (page && !IS_ERR(page)) {
1920                 kunmap(page);
1921                 page_cache_release(page);
1922         }
1923         if (name)
1924                 putname(name);
1925         if (did_down) {
1926                 if (!error)
1927                         inode->i_flags |= S_SWAPFILE;
1928                 mutex_unlock(&inode->i_mutex);
1929         }
1930         return error;
1931 }
1932
1933 void si_swapinfo(struct sysinfo *val)
1934 {
1935         unsigned int i;
1936         unsigned long nr_to_be_unused = 0;
1937
1938         spin_lock(&swap_lock);
1939         for (i = 0; i < nr_swapfiles; i++) {
1940                 if (!(swap_info[i].flags & SWP_USED) ||
1941                      (swap_info[i].flags & SWP_WRITEOK))
1942                         continue;
1943                 nr_to_be_unused += swap_info[i].inuse_pages;
1944         }
1945         val->freeswap = nr_swap_pages + nr_to_be_unused;
1946         val->totalswap = total_swap_pages + nr_to_be_unused;
1947         spin_unlock(&swap_lock);
1948 }
1949
1950 /*
1951  * Verify that a swap entry is valid and increment its swap map count.
1952  *
1953  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1954  * "permanent", but will be reclaimed by the next swapoff.
1955  */
1956 int swap_duplicate(swp_entry_t entry)
1957 {
1958         struct swap_info_struct * p;
1959         unsigned long offset, type;
1960         int result = 0;
1961
1962         if (is_migration_entry(entry))
1963                 return 1;
1964
1965         type = swp_type(entry);
1966         if (type >= nr_swapfiles)
1967                 goto bad_file;
1968         p = type + swap_info;
1969         offset = swp_offset(entry);
1970
1971         spin_lock(&swap_lock);
1972         if (offset < p->max && p->swap_map[offset]) {
1973                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1974                         p->swap_map[offset]++;
1975                         result = 1;
1976                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1977                         if (swap_overflow++ < 5)
1978                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1979                         p->swap_map[offset] = SWAP_MAP_MAX;
1980                         result = 1;
1981                 }
1982         }
1983         spin_unlock(&swap_lock);
1984 out:
1985         return result;
1986
1987 bad_file:
1988         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1989         goto out;
1990 }
1991
1992 /*
1993  * Called when allocating swap cache for exising swap entry,
1994  */
1995 int swapcache_prepare(swp_entry_t entry)
1996 {
1997         return swap_duplicate(entry);
1998 }
1999
2000
2001 struct swap_info_struct *
2002 get_swap_info_struct(unsigned type)
2003 {
2004         return &swap_info[type];
2005 }
2006
2007 /*
2008  * swap_lock prevents swap_map being freed. Don't grab an extra
2009  * reference on the swaphandle, it doesn't matter if it becomes unused.
2010  */
2011 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2012 {
2013         struct swap_info_struct *si;
2014         int our_page_cluster = page_cluster;
2015         pgoff_t target, toff;
2016         pgoff_t base, end;
2017         int nr_pages = 0;
2018
2019         if (!our_page_cluster)  /* no readahead */
2020                 return 0;
2021
2022         si = &swap_info[swp_type(entry)];
2023         target = swp_offset(entry);
2024         base = (target >> our_page_cluster) << our_page_cluster;
2025         end = base + (1 << our_page_cluster);
2026         if (!base)              /* first page is swap header */
2027                 base++;
2028
2029         spin_lock(&swap_lock);
2030         if (end > si->max)      /* don't go beyond end of map */
2031                 end = si->max;
2032
2033         /* Count contiguous allocated slots above our target */
2034         for (toff = target; ++toff < end; nr_pages++) {
2035                 /* Don't read in free or bad pages */
2036                 if (!si->swap_map[toff])
2037                         break;
2038                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2039                         break;
2040         }
2041         /* Count contiguous allocated slots below our target */
2042         for (toff = target; --toff >= base; nr_pages++) {
2043                 /* Don't read in free or bad pages */
2044                 if (!si->swap_map[toff])
2045                         break;
2046                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2047                         break;
2048         }
2049         spin_unlock(&swap_lock);
2050
2051         /*
2052          * Indicate starting offset, and return number of pages to get:
2053          * if only 1, say 0, since there's then no readahead to be done.
2054          */
2055         *offset = ++toff;
2056         return nr_pages? ++nr_pages: 0;
2057 }