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