swapfile: swap allocation cycle if nonrot
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32
33 #include <asm/pgtable.h>
34 #include <asm/tlbflush.h>
35 #include <linux/swapops.h>
36
37 static DEFINE_SPINLOCK(swap_lock);
38 static unsigned int nr_swapfiles;
39 long nr_swap_pages;
40 long total_swap_pages;
41 static int swap_overflow;
42 static int least_priority;
43
44 static const char Bad_file[] = "Bad swap file entry ";
45 static const char Unused_file[] = "Unused swap file entry ";
46 static const char Bad_offset[] = "Bad swap offset entry ";
47 static const char Unused_offset[] = "Unused swap offset entry ";
48
49 static struct swap_list_t swap_list = {-1, -1};
50
51 static struct swap_info_struct swap_info[MAX_SWAPFILES];
52
53 static DEFINE_MUTEX(swapon_mutex);
54
55 /*
56  * We need this because the bdev->unplug_fn can sleep and we cannot
57  * hold swap_lock while calling the unplug_fn. And swap_lock
58  * cannot be turned into a mutex.
59  */
60 static DECLARE_RWSEM(swap_unplug_sem);
61
62 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
63 {
64         swp_entry_t entry;
65
66         down_read(&swap_unplug_sem);
67         entry.val = page_private(page);
68         if (PageSwapCache(page)) {
69                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
70                 struct backing_dev_info *bdi;
71
72                 /*
73                  * If the page is removed from swapcache from under us (with a
74                  * racy try_to_unuse/swapoff) we need an additional reference
75                  * count to avoid reading garbage from page_private(page) above.
76                  * If the WARN_ON triggers during a swapoff it maybe the race
77                  * condition and it's harmless. However if it triggers without
78                  * swapoff it signals a problem.
79                  */
80                 WARN_ON(page_count(page) <= 1);
81
82                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
83                 blk_run_backing_dev(bdi, page);
84         }
85         up_read(&swap_unplug_sem);
86 }
87
88 /*
89  * swapon tell device that all the old swap contents can be discarded,
90  * to allow the swap device to optimize its wear-levelling.
91  */
92 static int discard_swap(struct swap_info_struct *si)
93 {
94         struct swap_extent *se;
95         int err = 0;
96
97         list_for_each_entry(se, &si->extent_list, list) {
98                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
99                 pgoff_t nr_blocks = se->nr_pages << (PAGE_SHIFT - 9);
100
101                 if (se->start_page == 0) {
102                         /* Do not discard the swap header page! */
103                         start_block += 1 << (PAGE_SHIFT - 9);
104                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
105                         if (!nr_blocks)
106                                 continue;
107                 }
108
109                 err = blkdev_issue_discard(si->bdev, start_block,
110                                                 nr_blocks, GFP_KERNEL);
111                 if (err)
112                         break;
113
114                 cond_resched();
115         }
116         return err;             /* That will often be -EOPNOTSUPP */
117 }
118
119 /*
120  * swap allocation tell device that a cluster of swap can now be discarded,
121  * to allow the swap device to optimize its wear-levelling.
122  */
123 static void discard_swap_cluster(struct swap_info_struct *si,
124                                  pgoff_t start_page, pgoff_t nr_pages)
125 {
126         struct swap_extent *se = si->curr_swap_extent;
127         int found_extent = 0;
128
129         while (nr_pages) {
130                 struct list_head *lh;
131
132                 if (se->start_page <= start_page &&
133                     start_page < se->start_page + se->nr_pages) {
134                         pgoff_t offset = start_page - se->start_page;
135                         sector_t start_block = se->start_block + offset;
136                         pgoff_t nr_blocks = se->nr_pages - offset;
137
138                         if (nr_blocks > nr_pages)
139                                 nr_blocks = nr_pages;
140                         start_page += nr_blocks;
141                         nr_pages -= nr_blocks;
142
143                         if (!found_extent++)
144                                 si->curr_swap_extent = se;
145
146                         start_block <<= PAGE_SHIFT - 9;
147                         nr_blocks <<= PAGE_SHIFT - 9;
148                         if (blkdev_issue_discard(si->bdev, start_block,
149                                                         nr_blocks, GFP_NOIO))
150                                 break;
151                 }
152
153                 lh = se->list.next;
154                 if (lh == &si->extent_list)
155                         lh = lh->next;
156                 se = list_entry(lh, struct swap_extent, list);
157         }
158 }
159
160 static int wait_for_discard(void *word)
161 {
162         schedule();
163         return 0;
164 }
165
166 #define SWAPFILE_CLUSTER        256
167 #define LATENCY_LIMIT           256
168
169 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
170 {
171         unsigned long offset;
172         unsigned long 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 void 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;
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 }
607
608 #ifdef CONFIG_HIBERNATION
609 /*
610  * Find the swap type that corresponds to given device (if any).
611  *
612  * @offset - number of the PAGE_SIZE-sized block of the device, starting
613  * from 0, in which the swap header is expected to be located.
614  *
615  * This is needed for the suspend to disk (aka swsusp).
616  */
617 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
618 {
619         struct block_device *bdev = NULL;
620         int i;
621
622         if (device)
623                 bdev = bdget(device);
624
625         spin_lock(&swap_lock);
626         for (i = 0; i < nr_swapfiles; i++) {
627                 struct swap_info_struct *sis = swap_info + i;
628
629                 if (!(sis->flags & SWP_WRITEOK))
630                         continue;
631
632                 if (!bdev) {
633                         if (bdev_p)
634                                 *bdev_p = sis->bdev;
635
636                         spin_unlock(&swap_lock);
637                         return i;
638                 }
639                 if (bdev == sis->bdev) {
640                         struct swap_extent *se;
641
642                         se = list_entry(sis->extent_list.next,
643                                         struct swap_extent, list);
644                         if (se->start_block == offset) {
645                                 if (bdev_p)
646                                         *bdev_p = sis->bdev;
647
648                                 spin_unlock(&swap_lock);
649                                 bdput(bdev);
650                                 return i;
651                         }
652                 }
653         }
654         spin_unlock(&swap_lock);
655         if (bdev)
656                 bdput(bdev);
657
658         return -ENODEV;
659 }
660
661 /*
662  * Return either the total number of swap pages of given type, or the number
663  * of free pages of that type (depending on @free)
664  *
665  * This is needed for software suspend
666  */
667 unsigned int count_swap_pages(int type, int free)
668 {
669         unsigned int n = 0;
670
671         if (type < nr_swapfiles) {
672                 spin_lock(&swap_lock);
673                 if (swap_info[type].flags & SWP_WRITEOK) {
674                         n = swap_info[type].pages;
675                         if (free)
676                                 n -= swap_info[type].inuse_pages;
677                 }
678                 spin_unlock(&swap_lock);
679         }
680         return n;
681 }
682 #endif
683
684 /*
685  * No need to decide whether this PTE shares the swap entry with others,
686  * just let do_wp_page work it out if a write is requested later - to
687  * force COW, vm_page_prot omits write permission from any private vma.
688  */
689 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
690                 unsigned long addr, swp_entry_t entry, struct page *page)
691 {
692         spinlock_t *ptl;
693         pte_t *pte;
694         int ret = 1;
695
696         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
697                 ret = -ENOMEM;
698
699         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
700         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
701                 if (ret > 0)
702                         mem_cgroup_uncharge_page(page);
703                 ret = 0;
704                 goto out;
705         }
706
707         inc_mm_counter(vma->vm_mm, anon_rss);
708         get_page(page);
709         set_pte_at(vma->vm_mm, addr, pte,
710                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
711         page_add_anon_rmap(page, vma, addr);
712         swap_free(entry);
713         /*
714          * Move the page to the active list so it is not
715          * immediately swapped out again after swapon.
716          */
717         activate_page(page);
718 out:
719         pte_unmap_unlock(pte, ptl);
720         return ret;
721 }
722
723 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
724                                 unsigned long addr, unsigned long end,
725                                 swp_entry_t entry, struct page *page)
726 {
727         pte_t swp_pte = swp_entry_to_pte(entry);
728         pte_t *pte;
729         int ret = 0;
730
731         /*
732          * We don't actually need pte lock while scanning for swp_pte: since
733          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
734          * page table while we're scanning; though it could get zapped, and on
735          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
736          * of unmatched parts which look like swp_pte, so unuse_pte must
737          * recheck under pte lock.  Scanning without pte lock lets it be
738          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
739          */
740         pte = pte_offset_map(pmd, addr);
741         do {
742                 /*
743                  * swapoff spends a _lot_ of time in this loop!
744                  * Test inline before going to call unuse_pte.
745                  */
746                 if (unlikely(pte_same(*pte, swp_pte))) {
747                         pte_unmap(pte);
748                         ret = unuse_pte(vma, pmd, addr, entry, page);
749                         if (ret)
750                                 goto out;
751                         pte = pte_offset_map(pmd, addr);
752                 }
753         } while (pte++, addr += PAGE_SIZE, addr != end);
754         pte_unmap(pte - 1);
755 out:
756         return ret;
757 }
758
759 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
760                                 unsigned long addr, unsigned long end,
761                                 swp_entry_t entry, struct page *page)
762 {
763         pmd_t *pmd;
764         unsigned long next;
765         int ret;
766
767         pmd = pmd_offset(pud, addr);
768         do {
769                 next = pmd_addr_end(addr, end);
770                 if (pmd_none_or_clear_bad(pmd))
771                         continue;
772                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
773                 if (ret)
774                         return ret;
775         } while (pmd++, addr = next, addr != end);
776         return 0;
777 }
778
779 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
780                                 unsigned long addr, unsigned long end,
781                                 swp_entry_t entry, struct page *page)
782 {
783         pud_t *pud;
784         unsigned long next;
785         int ret;
786
787         pud = pud_offset(pgd, addr);
788         do {
789                 next = pud_addr_end(addr, end);
790                 if (pud_none_or_clear_bad(pud))
791                         continue;
792                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
793                 if (ret)
794                         return ret;
795         } while (pud++, addr = next, addr != end);
796         return 0;
797 }
798
799 static int unuse_vma(struct vm_area_struct *vma,
800                                 swp_entry_t entry, struct page *page)
801 {
802         pgd_t *pgd;
803         unsigned long addr, end, next;
804         int ret;
805
806         if (page->mapping) {
807                 addr = page_address_in_vma(page, vma);
808                 if (addr == -EFAULT)
809                         return 0;
810                 else
811                         end = addr + PAGE_SIZE;
812         } else {
813                 addr = vma->vm_start;
814                 end = vma->vm_end;
815         }
816
817         pgd = pgd_offset(vma->vm_mm, addr);
818         do {
819                 next = pgd_addr_end(addr, end);
820                 if (pgd_none_or_clear_bad(pgd))
821                         continue;
822                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
823                 if (ret)
824                         return ret;
825         } while (pgd++, addr = next, addr != end);
826         return 0;
827 }
828
829 static int unuse_mm(struct mm_struct *mm,
830                                 swp_entry_t entry, struct page *page)
831 {
832         struct vm_area_struct *vma;
833         int ret = 0;
834
835         if (!down_read_trylock(&mm->mmap_sem)) {
836                 /*
837                  * Activate page so shrink_inactive_list is unlikely to unmap
838                  * its ptes while lock is dropped, so swapoff can make progress.
839                  */
840                 activate_page(page);
841                 unlock_page(page);
842                 down_read(&mm->mmap_sem);
843                 lock_page(page);
844         }
845         for (vma = mm->mmap; vma; vma = vma->vm_next) {
846                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
847                         break;
848         }
849         up_read(&mm->mmap_sem);
850         return (ret < 0)? ret: 0;
851 }
852
853 /*
854  * Scan swap_map from current position to next entry still in use.
855  * Recycle to start on reaching the end, returning 0 when empty.
856  */
857 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
858                                         unsigned int prev)
859 {
860         unsigned int max = si->max;
861         unsigned int i = prev;
862         int count;
863
864         /*
865          * No need for swap_lock here: we're just looking
866          * for whether an entry is in use, not modifying it; false
867          * hits are okay, and sys_swapoff() has already prevented new
868          * allocations from this area (while holding swap_lock).
869          */
870         for (;;) {
871                 if (++i >= max) {
872                         if (!prev) {
873                                 i = 0;
874                                 break;
875                         }
876                         /*
877                          * No entries in use at top of swap_map,
878                          * loop back to start and recheck there.
879                          */
880                         max = prev + 1;
881                         prev = 0;
882                         i = 1;
883                 }
884                 count = si->swap_map[i];
885                 if (count && count != SWAP_MAP_BAD)
886                         break;
887         }
888         return i;
889 }
890
891 /*
892  * We completely avoid races by reading each swap page in advance,
893  * and then search for the process using it.  All the necessary
894  * page table adjustments can then be made atomically.
895  */
896 static int try_to_unuse(unsigned int type)
897 {
898         struct swap_info_struct * si = &swap_info[type];
899         struct mm_struct *start_mm;
900         unsigned short *swap_map;
901         unsigned short swcount;
902         struct page *page;
903         swp_entry_t entry;
904         unsigned int i = 0;
905         int retval = 0;
906         int reset_overflow = 0;
907         int shmem;
908
909         /*
910          * When searching mms for an entry, a good strategy is to
911          * start at the first mm we freed the previous entry from
912          * (though actually we don't notice whether we or coincidence
913          * freed the entry).  Initialize this start_mm with a hold.
914          *
915          * A simpler strategy would be to start at the last mm we
916          * freed the previous entry from; but that would take less
917          * advantage of mmlist ordering, which clusters forked mms
918          * together, child after parent.  If we race with dup_mmap(), we
919          * prefer to resolve parent before child, lest we miss entries
920          * duplicated after we scanned child: using last mm would invert
921          * that.  Though it's only a serious concern when an overflowed
922          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
923          */
924         start_mm = &init_mm;
925         atomic_inc(&init_mm.mm_users);
926
927         /*
928          * Keep on scanning until all entries have gone.  Usually,
929          * one pass through swap_map is enough, but not necessarily:
930          * there are races when an instance of an entry might be missed.
931          */
932         while ((i = find_next_to_unuse(si, i)) != 0) {
933                 if (signal_pending(current)) {
934                         retval = -EINTR;
935                         break;
936                 }
937
938                 /*
939                  * Get a page for the entry, using the existing swap
940                  * cache page if there is one.  Otherwise, get a clean
941                  * page and read the swap into it.
942                  */
943                 swap_map = &si->swap_map[i];
944                 entry = swp_entry(type, i);
945                 page = read_swap_cache_async(entry,
946                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
947                 if (!page) {
948                         /*
949                          * Either swap_duplicate() failed because entry
950                          * has been freed independently, and will not be
951                          * reused since sys_swapoff() already disabled
952                          * allocation from here, or alloc_page() failed.
953                          */
954                         if (!*swap_map)
955                                 continue;
956                         retval = -ENOMEM;
957                         break;
958                 }
959
960                 /*
961                  * Don't hold on to start_mm if it looks like exiting.
962                  */
963                 if (atomic_read(&start_mm->mm_users) == 1) {
964                         mmput(start_mm);
965                         start_mm = &init_mm;
966                         atomic_inc(&init_mm.mm_users);
967                 }
968
969                 /*
970                  * Wait for and lock page.  When do_swap_page races with
971                  * try_to_unuse, do_swap_page can handle the fault much
972                  * faster than try_to_unuse can locate the entry.  This
973                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
974                  * defer to do_swap_page in such a case - in some tests,
975                  * do_swap_page and try_to_unuse repeatedly compete.
976                  */
977                 wait_on_page_locked(page);
978                 wait_on_page_writeback(page);
979                 lock_page(page);
980                 wait_on_page_writeback(page);
981
982                 /*
983                  * Remove all references to entry.
984                  * Whenever we reach init_mm, there's no address space
985                  * to search, but use it as a reminder to search shmem.
986                  */
987                 shmem = 0;
988                 swcount = *swap_map;
989                 if (swcount > 1) {
990                         if (start_mm == &init_mm)
991                                 shmem = shmem_unuse(entry, page);
992                         else
993                                 retval = unuse_mm(start_mm, entry, page);
994                 }
995                 if (*swap_map > 1) {
996                         int set_start_mm = (*swap_map >= swcount);
997                         struct list_head *p = &start_mm->mmlist;
998                         struct mm_struct *new_start_mm = start_mm;
999                         struct mm_struct *prev_mm = start_mm;
1000                         struct mm_struct *mm;
1001
1002                         atomic_inc(&new_start_mm->mm_users);
1003                         atomic_inc(&prev_mm->mm_users);
1004                         spin_lock(&mmlist_lock);
1005                         while (*swap_map > 1 && !retval && !shmem &&
1006                                         (p = p->next) != &start_mm->mmlist) {
1007                                 mm = list_entry(p, struct mm_struct, mmlist);
1008                                 if (!atomic_inc_not_zero(&mm->mm_users))
1009                                         continue;
1010                                 spin_unlock(&mmlist_lock);
1011                                 mmput(prev_mm);
1012                                 prev_mm = mm;
1013
1014                                 cond_resched();
1015
1016                                 swcount = *swap_map;
1017                                 if (swcount <= 1)
1018                                         ;
1019                                 else if (mm == &init_mm) {
1020                                         set_start_mm = 1;
1021                                         shmem = shmem_unuse(entry, page);
1022                                 } else
1023                                         retval = unuse_mm(mm, entry, page);
1024                                 if (set_start_mm && *swap_map < swcount) {
1025                                         mmput(new_start_mm);
1026                                         atomic_inc(&mm->mm_users);
1027                                         new_start_mm = mm;
1028                                         set_start_mm = 0;
1029                                 }
1030                                 spin_lock(&mmlist_lock);
1031                         }
1032                         spin_unlock(&mmlist_lock);
1033                         mmput(prev_mm);
1034                         mmput(start_mm);
1035                         start_mm = new_start_mm;
1036                 }
1037                 if (shmem) {
1038                         /* page has already been unlocked and released */
1039                         if (shmem > 0)
1040                                 continue;
1041                         retval = shmem;
1042                         break;
1043                 }
1044                 if (retval) {
1045                         unlock_page(page);
1046                         page_cache_release(page);
1047                         break;
1048                 }
1049
1050                 /*
1051                  * How could swap count reach 0x7fff when the maximum
1052                  * pid is 0x7fff, and there's no way to repeat a swap
1053                  * page within an mm (except in shmem, where it's the
1054                  * shared object which takes the reference count)?
1055                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
1056                  *
1057                  * If that's wrong, then we should worry more about
1058                  * exit_mmap() and do_munmap() cases described above:
1059                  * we might be resetting SWAP_MAP_MAX too early here.
1060                  * We know "Undead"s can happen, they're okay, so don't
1061                  * report them; but do report if we reset SWAP_MAP_MAX.
1062                  */
1063                 if (*swap_map == SWAP_MAP_MAX) {
1064                         spin_lock(&swap_lock);
1065                         *swap_map = 1;
1066                         spin_unlock(&swap_lock);
1067                         reset_overflow = 1;
1068                 }
1069
1070                 /*
1071                  * If a reference remains (rare), we would like to leave
1072                  * the page in the swap cache; but try_to_unmap could
1073                  * then re-duplicate the entry once we drop page lock,
1074                  * so we might loop indefinitely; also, that page could
1075                  * not be swapped out to other storage meanwhile.  So:
1076                  * delete from cache even if there's another reference,
1077                  * after ensuring that the data has been saved to disk -
1078                  * since if the reference remains (rarer), it will be
1079                  * read from disk into another page.  Splitting into two
1080                  * pages would be incorrect if swap supported "shared
1081                  * private" pages, but they are handled by tmpfs files.
1082                  */
1083                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
1084                         struct writeback_control wbc = {
1085                                 .sync_mode = WB_SYNC_NONE,
1086                         };
1087
1088                         swap_writepage(page, &wbc);
1089                         lock_page(page);
1090                         wait_on_page_writeback(page);
1091                 }
1092
1093                 /*
1094                  * It is conceivable that a racing task removed this page from
1095                  * swap cache just before we acquired the page lock at the top,
1096                  * or while we dropped it in unuse_mm().  The page might even
1097                  * be back in swap cache on another swap area: that we must not
1098                  * delete, since it may not have been written out to swap yet.
1099                  */
1100                 if (PageSwapCache(page) &&
1101                     likely(page_private(page) == entry.val))
1102                         delete_from_swap_cache(page);
1103
1104                 /*
1105                  * So we could skip searching mms once swap count went
1106                  * to 1, we did not mark any present ptes as dirty: must
1107                  * mark page dirty so shrink_page_list will preserve it.
1108                  */
1109                 SetPageDirty(page);
1110                 unlock_page(page);
1111                 page_cache_release(page);
1112
1113                 /*
1114                  * Make sure that we aren't completely killing
1115                  * interactive performance.
1116                  */
1117                 cond_resched();
1118         }
1119
1120         mmput(start_mm);
1121         if (reset_overflow) {
1122                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1123                 swap_overflow = 0;
1124         }
1125         return retval;
1126 }
1127
1128 /*
1129  * After a successful try_to_unuse, if no swap is now in use, we know
1130  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1131  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1132  * added to the mmlist just after page_duplicate - before would be racy.
1133  */
1134 static void drain_mmlist(void)
1135 {
1136         struct list_head *p, *next;
1137         unsigned int i;
1138
1139         for (i = 0; i < nr_swapfiles; i++)
1140                 if (swap_info[i].inuse_pages)
1141                         return;
1142         spin_lock(&mmlist_lock);
1143         list_for_each_safe(p, next, &init_mm.mmlist)
1144                 list_del_init(p);
1145         spin_unlock(&mmlist_lock);
1146 }
1147
1148 /*
1149  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1150  * corresponds to page offset `offset'.
1151  */
1152 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1153 {
1154         struct swap_extent *se = sis->curr_swap_extent;
1155         struct swap_extent *start_se = se;
1156
1157         for ( ; ; ) {
1158                 struct list_head *lh;
1159
1160                 if (se->start_page <= offset &&
1161                                 offset < (se->start_page + se->nr_pages)) {
1162                         return se->start_block + (offset - se->start_page);
1163                 }
1164                 lh = se->list.next;
1165                 if (lh == &sis->extent_list)
1166                         lh = lh->next;
1167                 se = list_entry(lh, struct swap_extent, list);
1168                 sis->curr_swap_extent = se;
1169                 BUG_ON(se == start_se);         /* It *must* be present */
1170         }
1171 }
1172
1173 #ifdef CONFIG_HIBERNATION
1174 /*
1175  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1176  * corresponding to given index in swap_info (swap type).
1177  */
1178 sector_t swapdev_block(int swap_type, pgoff_t offset)
1179 {
1180         struct swap_info_struct *sis;
1181
1182         if (swap_type >= nr_swapfiles)
1183                 return 0;
1184
1185         sis = swap_info + swap_type;
1186         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1187 }
1188 #endif /* CONFIG_HIBERNATION */
1189
1190 /*
1191  * Free all of a swapdev's extent information
1192  */
1193 static void destroy_swap_extents(struct swap_info_struct *sis)
1194 {
1195         while (!list_empty(&sis->extent_list)) {
1196                 struct swap_extent *se;
1197
1198                 se = list_entry(sis->extent_list.next,
1199                                 struct swap_extent, list);
1200                 list_del(&se->list);
1201                 kfree(se);
1202         }
1203 }
1204
1205 /*
1206  * Add a block range (and the corresponding page range) into this swapdev's
1207  * extent list.  The extent list is kept sorted in page order.
1208  *
1209  * This function rather assumes that it is called in ascending page order.
1210  */
1211 static int
1212 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1213                 unsigned long nr_pages, sector_t start_block)
1214 {
1215         struct swap_extent *se;
1216         struct swap_extent *new_se;
1217         struct list_head *lh;
1218
1219         lh = sis->extent_list.prev;     /* The highest page extent */
1220         if (lh != &sis->extent_list) {
1221                 se = list_entry(lh, struct swap_extent, list);
1222                 BUG_ON(se->start_page + se->nr_pages != start_page);
1223                 if (se->start_block + se->nr_pages == start_block) {
1224                         /* Merge it */
1225                         se->nr_pages += nr_pages;
1226                         return 0;
1227                 }
1228         }
1229
1230         /*
1231          * No merge.  Insert a new extent, preserving ordering.
1232          */
1233         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1234         if (new_se == NULL)
1235                 return -ENOMEM;
1236         new_se->start_page = start_page;
1237         new_se->nr_pages = nr_pages;
1238         new_se->start_block = start_block;
1239
1240         list_add_tail(&new_se->list, &sis->extent_list);
1241         return 1;
1242 }
1243
1244 /*
1245  * A `swap extent' is a simple thing which maps a contiguous range of pages
1246  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1247  * is built at swapon time and is then used at swap_writepage/swap_readpage
1248  * time for locating where on disk a page belongs.
1249  *
1250  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1251  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1252  * swap files identically.
1253  *
1254  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1255  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1256  * swapfiles are handled *identically* after swapon time.
1257  *
1258  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1259  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1260  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1261  * requirements, they are simply tossed out - we will never use those blocks
1262  * for swapping.
1263  *
1264  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1265  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1266  * which will scribble on the fs.
1267  *
1268  * The amount of disk space which a single swap extent represents varies.
1269  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1270  * extents in the list.  To avoid much list walking, we cache the previous
1271  * search location in `curr_swap_extent', and start new searches from there.
1272  * This is extremely effective.  The average number of iterations in
1273  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1274  */
1275 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1276 {
1277         struct inode *inode;
1278         unsigned blocks_per_page;
1279         unsigned long page_no;
1280         unsigned blkbits;
1281         sector_t probe_block;
1282         sector_t last_block;
1283         sector_t lowest_block = -1;
1284         sector_t highest_block = 0;
1285         int nr_extents = 0;
1286         int ret;
1287
1288         inode = sis->swap_file->f_mapping->host;
1289         if (S_ISBLK(inode->i_mode)) {
1290                 ret = add_swap_extent(sis, 0, sis->max, 0);
1291                 *span = sis->pages;
1292                 goto done;
1293         }
1294
1295         blkbits = inode->i_blkbits;
1296         blocks_per_page = PAGE_SIZE >> blkbits;
1297
1298         /*
1299          * Map all the blocks into the extent list.  This code doesn't try
1300          * to be very smart.
1301          */
1302         probe_block = 0;
1303         page_no = 0;
1304         last_block = i_size_read(inode) >> blkbits;
1305         while ((probe_block + blocks_per_page) <= last_block &&
1306                         page_no < sis->max) {
1307                 unsigned block_in_page;
1308                 sector_t first_block;
1309
1310                 first_block = bmap(inode, probe_block);
1311                 if (first_block == 0)
1312                         goto bad_bmap;
1313
1314                 /*
1315                  * It must be PAGE_SIZE aligned on-disk
1316                  */
1317                 if (first_block & (blocks_per_page - 1)) {
1318                         probe_block++;
1319                         goto reprobe;
1320                 }
1321
1322                 for (block_in_page = 1; block_in_page < blocks_per_page;
1323                                         block_in_page++) {
1324                         sector_t block;
1325
1326                         block = bmap(inode, probe_block + block_in_page);
1327                         if (block == 0)
1328                                 goto bad_bmap;
1329                         if (block != first_block + block_in_page) {
1330                                 /* Discontiguity */
1331                                 probe_block++;
1332                                 goto reprobe;
1333                         }
1334                 }
1335
1336                 first_block >>= (PAGE_SHIFT - blkbits);
1337                 if (page_no) {  /* exclude the header page */
1338                         if (first_block < lowest_block)
1339                                 lowest_block = first_block;
1340                         if (first_block > highest_block)
1341                                 highest_block = first_block;
1342                 }
1343
1344                 /*
1345                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1346                  */
1347                 ret = add_swap_extent(sis, page_no, 1, first_block);
1348                 if (ret < 0)
1349                         goto out;
1350                 nr_extents += ret;
1351                 page_no++;
1352                 probe_block += blocks_per_page;
1353 reprobe:
1354                 continue;
1355         }
1356         ret = nr_extents;
1357         *span = 1 + highest_block - lowest_block;
1358         if (page_no == 0)
1359                 page_no = 1;    /* force Empty message */
1360         sis->max = page_no;
1361         sis->pages = page_no - 1;
1362         sis->highest_bit = page_no - 1;
1363 done:
1364         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1365                                         struct swap_extent, list);
1366         goto out;
1367 bad_bmap:
1368         printk(KERN_ERR "swapon: swapfile has holes\n");
1369         ret = -EINVAL;
1370 out:
1371         return ret;
1372 }
1373
1374 #if 0   /* We don't need this yet */
1375 #include <linux/backing-dev.h>
1376 int page_queue_congested(struct page *page)
1377 {
1378         struct backing_dev_info *bdi;
1379
1380         VM_BUG_ON(!PageLocked(page));   /* It pins the swap_info_struct */
1381
1382         if (PageSwapCache(page)) {
1383                 swp_entry_t entry = { .val = page_private(page) };
1384                 struct swap_info_struct *sis;
1385
1386                 sis = get_swap_info_struct(swp_type(entry));
1387                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1388         } else
1389                 bdi = page->mapping->backing_dev_info;
1390         return bdi_write_congested(bdi);
1391 }
1392 #endif
1393
1394 asmlinkage long sys_swapoff(const char __user * specialfile)
1395 {
1396         struct swap_info_struct * p = NULL;
1397         unsigned short *swap_map;
1398         struct file *swap_file, *victim;
1399         struct address_space *mapping;
1400         struct inode *inode;
1401         char * pathname;
1402         int i, type, prev;
1403         int err;
1404
1405         if (!capable(CAP_SYS_ADMIN))
1406                 return -EPERM;
1407
1408         pathname = getname(specialfile);
1409         err = PTR_ERR(pathname);
1410         if (IS_ERR(pathname))
1411                 goto out;
1412
1413         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1414         putname(pathname);
1415         err = PTR_ERR(victim);
1416         if (IS_ERR(victim))
1417                 goto out;
1418
1419         mapping = victim->f_mapping;
1420         prev = -1;
1421         spin_lock(&swap_lock);
1422         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1423                 p = swap_info + type;
1424                 if (p->flags & SWP_WRITEOK) {
1425                         if (p->swap_file->f_mapping == mapping)
1426                                 break;
1427                 }
1428                 prev = type;
1429         }
1430         if (type < 0) {
1431                 err = -EINVAL;
1432                 spin_unlock(&swap_lock);
1433                 goto out_dput;
1434         }
1435         if (!security_vm_enough_memory(p->pages))
1436                 vm_unacct_memory(p->pages);
1437         else {
1438                 err = -ENOMEM;
1439                 spin_unlock(&swap_lock);
1440                 goto out_dput;
1441         }
1442         if (prev < 0) {
1443                 swap_list.head = p->next;
1444         } else {
1445                 swap_info[prev].next = p->next;
1446         }
1447         if (type == swap_list.next) {
1448                 /* just pick something that's safe... */
1449                 swap_list.next = swap_list.head;
1450         }
1451         if (p->prio < 0) {
1452                 for (i = p->next; i >= 0; i = swap_info[i].next)
1453                         swap_info[i].prio = p->prio--;
1454                 least_priority++;
1455         }
1456         nr_swap_pages -= p->pages;
1457         total_swap_pages -= p->pages;
1458         p->flags &= ~SWP_WRITEOK;
1459         spin_unlock(&swap_lock);
1460
1461         current->flags |= PF_SWAPOFF;
1462         err = try_to_unuse(type);
1463         current->flags &= ~PF_SWAPOFF;
1464
1465         if (err) {
1466                 /* re-insert swap space back into swap_list */
1467                 spin_lock(&swap_lock);
1468                 if (p->prio < 0)
1469                         p->prio = --least_priority;
1470                 prev = -1;
1471                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1472                         if (p->prio >= swap_info[i].prio)
1473                                 break;
1474                         prev = i;
1475                 }
1476                 p->next = i;
1477                 if (prev < 0)
1478                         swap_list.head = swap_list.next = p - swap_info;
1479                 else
1480                         swap_info[prev].next = p - swap_info;
1481                 nr_swap_pages += p->pages;
1482                 total_swap_pages += p->pages;
1483                 p->flags |= SWP_WRITEOK;
1484                 spin_unlock(&swap_lock);
1485                 goto out_dput;
1486         }
1487
1488         /* wait for any unplug function to finish */
1489         down_write(&swap_unplug_sem);
1490         up_write(&swap_unplug_sem);
1491
1492         destroy_swap_extents(p);
1493         mutex_lock(&swapon_mutex);
1494         spin_lock(&swap_lock);
1495         drain_mmlist();
1496
1497         /* wait for anyone still in scan_swap_map */
1498         p->highest_bit = 0;             /* cuts scans short */
1499         while (p->flags >= SWP_SCANNING) {
1500                 spin_unlock(&swap_lock);
1501                 schedule_timeout_uninterruptible(1);
1502                 spin_lock(&swap_lock);
1503         }
1504
1505         swap_file = p->swap_file;
1506         p->swap_file = NULL;
1507         p->max = 0;
1508         swap_map = p->swap_map;
1509         p->swap_map = NULL;
1510         p->flags = 0;
1511         spin_unlock(&swap_lock);
1512         mutex_unlock(&swapon_mutex);
1513         vfree(swap_map);
1514         inode = mapping->host;
1515         if (S_ISBLK(inode->i_mode)) {
1516                 struct block_device *bdev = I_BDEV(inode);
1517                 set_blocksize(bdev, p->old_block_size);
1518                 bd_release(bdev);
1519         } else {
1520                 mutex_lock(&inode->i_mutex);
1521                 inode->i_flags &= ~S_SWAPFILE;
1522                 mutex_unlock(&inode->i_mutex);
1523         }
1524         filp_close(swap_file, NULL);
1525         err = 0;
1526
1527 out_dput:
1528         filp_close(victim, NULL);
1529 out:
1530         return err;
1531 }
1532
1533 #ifdef CONFIG_PROC_FS
1534 /* iterator */
1535 static void *swap_start(struct seq_file *swap, loff_t *pos)
1536 {
1537         struct swap_info_struct *ptr = swap_info;
1538         int i;
1539         loff_t l = *pos;
1540
1541         mutex_lock(&swapon_mutex);
1542
1543         if (!l)
1544                 return SEQ_START_TOKEN;
1545
1546         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1547                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1548                         continue;
1549                 if (!--l)
1550                         return ptr;
1551         }
1552
1553         return NULL;
1554 }
1555
1556 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1557 {
1558         struct swap_info_struct *ptr;
1559         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1560
1561         if (v == SEQ_START_TOKEN)
1562                 ptr = swap_info;
1563         else {
1564                 ptr = v;
1565                 ptr++;
1566         }
1567
1568         for (; ptr < endptr; ptr++) {
1569                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1570                         continue;
1571                 ++*pos;
1572                 return ptr;
1573         }
1574
1575         return NULL;
1576 }
1577
1578 static void swap_stop(struct seq_file *swap, void *v)
1579 {
1580         mutex_unlock(&swapon_mutex);
1581 }
1582
1583 static int swap_show(struct seq_file *swap, void *v)
1584 {
1585         struct swap_info_struct *ptr = v;
1586         struct file *file;
1587         int len;
1588
1589         if (ptr == SEQ_START_TOKEN) {
1590                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1591                 return 0;
1592         }
1593
1594         file = ptr->swap_file;
1595         len = seq_path(swap, &file->f_path, " \t\n\\");
1596         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1597                         len < 40 ? 40 - len : 1, " ",
1598                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1599                                 "partition" : "file\t",
1600                         ptr->pages << (PAGE_SHIFT - 10),
1601                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1602                         ptr->prio);
1603         return 0;
1604 }
1605
1606 static const struct seq_operations swaps_op = {
1607         .start =        swap_start,
1608         .next =         swap_next,
1609         .stop =         swap_stop,
1610         .show =         swap_show
1611 };
1612
1613 static int swaps_open(struct inode *inode, struct file *file)
1614 {
1615         return seq_open(file, &swaps_op);
1616 }
1617
1618 static const struct file_operations proc_swaps_operations = {
1619         .open           = swaps_open,
1620         .read           = seq_read,
1621         .llseek         = seq_lseek,
1622         .release        = seq_release,
1623 };
1624
1625 static int __init procswaps_init(void)
1626 {
1627         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1628         return 0;
1629 }
1630 __initcall(procswaps_init);
1631 #endif /* CONFIG_PROC_FS */
1632
1633 #ifdef MAX_SWAPFILES_CHECK
1634 static int __init max_swapfiles_check(void)
1635 {
1636         MAX_SWAPFILES_CHECK();
1637         return 0;
1638 }
1639 late_initcall(max_swapfiles_check);
1640 #endif
1641
1642 /*
1643  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1644  *
1645  * The swapon system call
1646  */
1647 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1648 {
1649         struct swap_info_struct * p;
1650         char *name = NULL;
1651         struct block_device *bdev = NULL;
1652         struct file *swap_file = NULL;
1653         struct address_space *mapping;
1654         unsigned int type;
1655         int i, prev;
1656         int error;
1657         union swap_header *swap_header = NULL;
1658         unsigned int nr_good_pages = 0;
1659         int nr_extents = 0;
1660         sector_t span;
1661         unsigned long maxpages = 1;
1662         unsigned long swapfilepages;
1663         unsigned short *swap_map = NULL;
1664         struct page *page = NULL;
1665         struct inode *inode = NULL;
1666         int did_down = 0;
1667
1668         if (!capable(CAP_SYS_ADMIN))
1669                 return -EPERM;
1670         spin_lock(&swap_lock);
1671         p = swap_info;
1672         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1673                 if (!(p->flags & SWP_USED))
1674                         break;
1675         error = -EPERM;
1676         if (type >= MAX_SWAPFILES) {
1677                 spin_unlock(&swap_lock);
1678                 goto out;
1679         }
1680         if (type >= nr_swapfiles)
1681                 nr_swapfiles = type+1;
1682         memset(p, 0, sizeof(*p));
1683         INIT_LIST_HEAD(&p->extent_list);
1684         p->flags = SWP_USED;
1685         p->next = -1;
1686         spin_unlock(&swap_lock);
1687         name = getname(specialfile);
1688         error = PTR_ERR(name);
1689         if (IS_ERR(name)) {
1690                 name = NULL;
1691                 goto bad_swap_2;
1692         }
1693         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1694         error = PTR_ERR(swap_file);
1695         if (IS_ERR(swap_file)) {
1696                 swap_file = NULL;
1697                 goto bad_swap_2;
1698         }
1699
1700         p->swap_file = swap_file;
1701         mapping = swap_file->f_mapping;
1702         inode = mapping->host;
1703
1704         error = -EBUSY;
1705         for (i = 0; i < nr_swapfiles; i++) {
1706                 struct swap_info_struct *q = &swap_info[i];
1707
1708                 if (i == type || !q->swap_file)
1709                         continue;
1710                 if (mapping == q->swap_file->f_mapping)
1711                         goto bad_swap;
1712         }
1713
1714         error = -EINVAL;
1715         if (S_ISBLK(inode->i_mode)) {
1716                 bdev = I_BDEV(inode);
1717                 error = bd_claim(bdev, sys_swapon);
1718                 if (error < 0) {
1719                         bdev = NULL;
1720                         error = -EINVAL;
1721                         goto bad_swap;
1722                 }
1723                 p->old_block_size = block_size(bdev);
1724                 error = set_blocksize(bdev, PAGE_SIZE);
1725                 if (error < 0)
1726                         goto bad_swap;
1727                 p->bdev = bdev;
1728         } else if (S_ISREG(inode->i_mode)) {
1729                 p->bdev = inode->i_sb->s_bdev;
1730                 mutex_lock(&inode->i_mutex);
1731                 did_down = 1;
1732                 if (IS_SWAPFILE(inode)) {
1733                         error = -EBUSY;
1734                         goto bad_swap;
1735                 }
1736         } else {
1737                 goto bad_swap;
1738         }
1739
1740         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1741
1742         /*
1743          * Read the swap header.
1744          */
1745         if (!mapping->a_ops->readpage) {
1746                 error = -EINVAL;
1747                 goto bad_swap;
1748         }
1749         page = read_mapping_page(mapping, 0, swap_file);
1750         if (IS_ERR(page)) {
1751                 error = PTR_ERR(page);
1752                 goto bad_swap;
1753         }
1754         swap_header = kmap(page);
1755
1756         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1757                 printk(KERN_ERR "Unable to find swap-space signature\n");
1758                 error = -EINVAL;
1759                 goto bad_swap;
1760         }
1761
1762         /* swap partition endianess hack... */
1763         if (swab32(swap_header->info.version) == 1) {
1764                 swab32s(&swap_header->info.version);
1765                 swab32s(&swap_header->info.last_page);
1766                 swab32s(&swap_header->info.nr_badpages);
1767                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1768                         swab32s(&swap_header->info.badpages[i]);
1769         }
1770         /* Check the swap header's sub-version */
1771         if (swap_header->info.version != 1) {
1772                 printk(KERN_WARNING
1773                        "Unable to handle swap header version %d\n",
1774                        swap_header->info.version);
1775                 error = -EINVAL;
1776                 goto bad_swap;
1777         }
1778
1779         p->lowest_bit  = 1;
1780         p->cluster_next = 1;
1781
1782         /*
1783          * Find out how many pages are allowed for a single swap
1784          * device. There are two limiting factors: 1) the number of
1785          * bits for the swap offset in the swp_entry_t type and
1786          * 2) the number of bits in the a swap pte as defined by
1787          * the different architectures. In order to find the
1788          * largest possible bit mask a swap entry with swap type 0
1789          * and swap offset ~0UL is created, encoded to a swap pte,
1790          * decoded to a swp_entry_t again and finally the swap
1791          * offset is extracted. This will mask all the bits from
1792          * the initial ~0UL mask that can't be encoded in either
1793          * the swp_entry_t or the architecture definition of a
1794          * swap pte.
1795          */
1796         maxpages = swp_offset(pte_to_swp_entry(
1797                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1798         if (maxpages > swap_header->info.last_page)
1799                 maxpages = swap_header->info.last_page;
1800         p->highest_bit = maxpages - 1;
1801
1802         error = -EINVAL;
1803         if (!maxpages)
1804                 goto bad_swap;
1805         if (swapfilepages && maxpages > swapfilepages) {
1806                 printk(KERN_WARNING
1807                        "Swap area shorter than signature indicates\n");
1808                 goto bad_swap;
1809         }
1810         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1811                 goto bad_swap;
1812         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1813                 goto bad_swap;
1814
1815         /* OK, set up the swap map and apply the bad block list */
1816         swap_map = vmalloc(maxpages * sizeof(short));
1817         if (!swap_map) {
1818                 error = -ENOMEM;
1819                 goto bad_swap;
1820         }
1821
1822         memset(swap_map, 0, maxpages * sizeof(short));
1823         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1824                 int page_nr = swap_header->info.badpages[i];
1825                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1826                         error = -EINVAL;
1827                         goto bad_swap;
1828                 }
1829                 swap_map[page_nr] = SWAP_MAP_BAD;
1830         }
1831         nr_good_pages = swap_header->info.last_page -
1832                         swap_header->info.nr_badpages -
1833                         1 /* header page */;
1834
1835         if (nr_good_pages) {
1836                 swap_map[0] = SWAP_MAP_BAD;
1837                 p->max = maxpages;
1838                 p->pages = nr_good_pages;
1839                 nr_extents = setup_swap_extents(p, &span);
1840                 if (nr_extents < 0) {
1841                         error = nr_extents;
1842                         goto bad_swap;
1843                 }
1844                 nr_good_pages = p->pages;
1845         }
1846         if (!nr_good_pages) {
1847                 printk(KERN_WARNING "Empty swap-file\n");
1848                 error = -EINVAL;
1849                 goto bad_swap;
1850         }
1851
1852         if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1853                 p->flags |= SWP_SOLIDSTATE;
1854                 srandom32((u32)get_seconds());
1855                 p->cluster_next = 1 + (random32() % p->highest_bit);
1856         }
1857         if (discard_swap(p) == 0)
1858                 p->flags |= SWP_DISCARDABLE;
1859
1860         mutex_lock(&swapon_mutex);
1861         spin_lock(&swap_lock);
1862         if (swap_flags & SWAP_FLAG_PREFER)
1863                 p->prio =
1864                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1865         else
1866                 p->prio = --least_priority;
1867         p->swap_map = swap_map;
1868         p->flags |= SWP_WRITEOK;
1869         nr_swap_pages += nr_good_pages;
1870         total_swap_pages += nr_good_pages;
1871
1872         printk(KERN_INFO "Adding %uk swap on %s.  "
1873                         "Priority:%d extents:%d across:%lluk %s%s\n",
1874                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1875                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1876                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1877                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1878
1879         /* insert swap space into swap_list: */
1880         prev = -1;
1881         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1882                 if (p->prio >= swap_info[i].prio) {
1883                         break;
1884                 }
1885                 prev = i;
1886         }
1887         p->next = i;
1888         if (prev < 0) {
1889                 swap_list.head = swap_list.next = p - swap_info;
1890         } else {
1891                 swap_info[prev].next = p - swap_info;
1892         }
1893         spin_unlock(&swap_lock);
1894         mutex_unlock(&swapon_mutex);
1895         error = 0;
1896         goto out;
1897 bad_swap:
1898         if (bdev) {
1899                 set_blocksize(bdev, p->old_block_size);
1900                 bd_release(bdev);
1901         }
1902         destroy_swap_extents(p);
1903 bad_swap_2:
1904         spin_lock(&swap_lock);
1905         p->swap_file = NULL;
1906         p->flags = 0;
1907         spin_unlock(&swap_lock);
1908         vfree(swap_map);
1909         if (swap_file)
1910                 filp_close(swap_file, NULL);
1911 out:
1912         if (page && !IS_ERR(page)) {
1913                 kunmap(page);
1914                 page_cache_release(page);
1915         }
1916         if (name)
1917                 putname(name);
1918         if (did_down) {
1919                 if (!error)
1920                         inode->i_flags |= S_SWAPFILE;
1921                 mutex_unlock(&inode->i_mutex);
1922         }
1923         return error;
1924 }
1925
1926 void si_swapinfo(struct sysinfo *val)
1927 {
1928         unsigned int i;
1929         unsigned long nr_to_be_unused = 0;
1930
1931         spin_lock(&swap_lock);
1932         for (i = 0; i < nr_swapfiles; i++) {
1933                 if (!(swap_info[i].flags & SWP_USED) ||
1934                      (swap_info[i].flags & SWP_WRITEOK))
1935                         continue;
1936                 nr_to_be_unused += swap_info[i].inuse_pages;
1937         }
1938         val->freeswap = nr_swap_pages + nr_to_be_unused;
1939         val->totalswap = total_swap_pages + nr_to_be_unused;
1940         spin_unlock(&swap_lock);
1941 }
1942
1943 /*
1944  * Verify that a swap entry is valid and increment its swap map count.
1945  *
1946  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1947  * "permanent", but will be reclaimed by the next swapoff.
1948  */
1949 int swap_duplicate(swp_entry_t entry)
1950 {
1951         struct swap_info_struct * p;
1952         unsigned long offset, type;
1953         int result = 0;
1954
1955         if (is_migration_entry(entry))
1956                 return 1;
1957
1958         type = swp_type(entry);
1959         if (type >= nr_swapfiles)
1960                 goto bad_file;
1961         p = type + swap_info;
1962         offset = swp_offset(entry);
1963
1964         spin_lock(&swap_lock);
1965         if (offset < p->max && p->swap_map[offset]) {
1966                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1967                         p->swap_map[offset]++;
1968                         result = 1;
1969                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1970                         if (swap_overflow++ < 5)
1971                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1972                         p->swap_map[offset] = SWAP_MAP_MAX;
1973                         result = 1;
1974                 }
1975         }
1976         spin_unlock(&swap_lock);
1977 out:
1978         return result;
1979
1980 bad_file:
1981         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1982         goto out;
1983 }
1984
1985 struct swap_info_struct *
1986 get_swap_info_struct(unsigned type)
1987 {
1988         return &swap_info[type];
1989 }
1990
1991 /*
1992  * swap_lock prevents swap_map being freed. Don't grab an extra
1993  * reference on the swaphandle, it doesn't matter if it becomes unused.
1994  */
1995 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1996 {
1997         struct swap_info_struct *si;
1998         int our_page_cluster = page_cluster;
1999         pgoff_t target, toff;
2000         pgoff_t base, end;
2001         int nr_pages = 0;
2002
2003         if (!our_page_cluster)  /* no readahead */
2004                 return 0;
2005
2006         si = &swap_info[swp_type(entry)];
2007         target = swp_offset(entry);
2008         base = (target >> our_page_cluster) << our_page_cluster;
2009         end = base + (1 << our_page_cluster);
2010         if (!base)              /* first page is swap header */
2011                 base++;
2012
2013         spin_lock(&swap_lock);
2014         if (end > si->max)      /* don't go beyond end of map */
2015                 end = si->max;
2016
2017         /* Count contiguous allocated slots above our target */
2018         for (toff = target; ++toff < end; nr_pages++) {
2019                 /* Don't read in free or bad pages */
2020                 if (!si->swap_map[toff])
2021                         break;
2022                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2023                         break;
2024         }
2025         /* Count contiguous allocated slots below our target */
2026         for (toff = target; --toff >= base; nr_pages++) {
2027                 /* Don't read in free or bad pages */
2028                 if (!si->swap_map[toff])
2029                         break;
2030                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2031                         break;
2032         }
2033         spin_unlock(&swap_lock);
2034
2035         /*
2036          * Indicate starting offset, and return number of pages to get:
2037          * if only 1, say 0, since there's then no readahead to be done.
2038          */
2039         *offset = ++toff;
2040         return nr_pages? ++nr_pages: 0;
2041 }